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 /* If a pseudo's lifetime is entirely contained within this loop, then we
747 can use a different pseudo in each unrolled copy of the loop. This
748 results in better code. */
749 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
750 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
751 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
752 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
753 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
755 /* However, we must also check for loop-carried dependencies.
756 If the value the pseudo has at the end of iteration X is
757 used by iteration X+1, then we can not use a different pseudo
758 for each unrolled copy of the loop. */
759 /* A pseudo is safe if regno_first_uid is a set, and this
760 set dominates all instructions from regno_first_uid to
762 /* ??? This check is simplistic. We would get better code if
763 this check was more sophisticated. */
764 if (set_dominates_use (j
, regno_first_uid
[j
], regno_last_uid
[j
],
765 copy_start
, copy_end
))
768 if (loop_dump_stream
)
771 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
773 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
779 /* If this loop requires exit tests when unrolled, check to see if we
780 can precondition the loop so as to make the exit tests unnecessary.
781 Just like variable splitting, this is not safe if the loop is entered
782 via a jump to the bottom. Also, can not do this if no strength
783 reduce info, because precondition_loop_p uses this info. */
785 /* Must copy the loop body for preconditioning before the following
786 find_splittable_regs call since that will emit insns which need to
787 be after the preconditioned loop copies, but immediately before the
788 unrolled loop copies. */
790 /* Also, it is not safe to split induction variables for the preconditioned
791 copies of the loop body. If we split induction variables, then the code
792 assumes that each induction variable can be represented as a function
793 of its initial value and the loop iteration number. This is not true
794 in this case, because the last preconditioned copy of the loop body
795 could be any iteration from the first up to the `unroll_number-1'th,
796 depending on the initial value of the iteration variable. Therefore
797 we can not split induction variables here, because we can not calculate
798 their value. Hence, this code must occur before find_splittable_regs
801 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
803 rtx initial_value
, final_value
, increment
;
805 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
806 loop_start
, loop_end
))
808 register rtx diff
, temp
;
809 enum machine_mode mode
;
811 int abs_inc
, neg_inc
;
813 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
815 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
816 map
->const_age_map
= (unsigned *) alloca (maxregnum
817 * sizeof (unsigned));
818 map
->const_equiv_map_size
= maxregnum
;
819 global_const_equiv_map
= map
->const_equiv_map
;
820 global_const_equiv_map_size
= maxregnum
;
822 init_reg_map (map
, maxregnum
);
824 /* Limit loop unrolling to 4, since this will make 7 copies of
826 if (unroll_number
> 4)
829 /* Save the absolute value of the increment, and also whether or
830 not it is negative. */
832 abs_inc
= INTVAL (increment
);
841 /* Decide what mode to do these calculations in. Choose the larger
842 of final_value's mode and initial_value's mode, or a full-word if
843 both are constants. */
844 mode
= GET_MODE (final_value
);
845 if (mode
== VOIDmode
)
847 mode
= GET_MODE (initial_value
);
848 if (mode
== VOIDmode
)
851 else if (mode
!= GET_MODE (initial_value
)
852 && (GET_MODE_SIZE (mode
)
853 < GET_MODE_SIZE (GET_MODE (initial_value
))))
854 mode
= GET_MODE (initial_value
);
856 /* Calculate the difference between the final and initial values.
857 Final value may be a (plus (reg x) (const_int 1)) rtx.
858 Let the following cse pass simplify this if initial value is
861 We must copy the final and initial values here to avoid
862 improperly shared rtl. */
864 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
865 copy_rtx (initial_value
), NULL_RTX
, 0,
868 /* Now calculate (diff % (unroll * abs (increment))) by using an
870 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
871 GEN_INT (unroll_number
* abs_inc
- 1),
872 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
874 /* Now emit a sequence of branches to jump to the proper precond
877 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
878 for (i
= 0; i
< unroll_number
; i
++)
879 labels
[i
] = gen_label_rtx ();
881 /* Check for the case where the initial value is greater than or equal
882 to the final value. In that case, we want to execute exactly
883 one loop iteration. The code below will fail for this case. */
885 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
886 NULL_RTX
, mode
, 0, 0);
888 emit_jump_insn (gen_ble (labels
[1]));
890 emit_jump_insn (gen_bge (labels
[1]));
891 JUMP_LABEL (get_last_insn ()) = labels
[1];
892 LABEL_NUSES (labels
[1])++;
894 /* Assuming the unroll_number is 4, and the increment is 2, then
895 for a negative increment: for a positive increment:
896 diff = 0,1 precond 0 diff = 0,7 precond 0
897 diff = 2,3 precond 3 diff = 1,2 precond 1
898 diff = 4,5 precond 2 diff = 3,4 precond 2
899 diff = 6,7 precond 1 diff = 5,6 precond 3 */
901 /* We only need to emit (unroll_number - 1) branches here, the
902 last case just falls through to the following code. */
904 /* ??? This would give better code if we emitted a tree of branches
905 instead of the current linear list of branches. */
907 for (i
= 0; i
< unroll_number
- 1; i
++)
910 enum rtx_code cmp_code
;
912 /* For negative increments, must invert the constant compared
913 against, except when comparing against zero. */
921 cmp_const
= unroll_number
- i
;
930 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
931 cmp_code
, NULL_RTX
, mode
, 0, 0);
934 emit_jump_insn (gen_beq (labels
[i
]));
936 emit_jump_insn (gen_bge (labels
[i
]));
938 emit_jump_insn (gen_ble (labels
[i
]));
939 JUMP_LABEL (get_last_insn ()) = labels
[i
];
940 LABEL_NUSES (labels
[i
])++;
943 /* If the increment is greater than one, then we need another branch,
944 to handle other cases equivalent to 0. */
946 /* ??? This should be merged into the code above somehow to help
947 simplify the code here, and reduce the number of branches emitted.
948 For the negative increment case, the branch here could easily
949 be merged with the `0' case branch above. For the positive
950 increment case, it is not clear how this can be simplified. */
955 enum rtx_code cmp_code
;
959 cmp_const
= abs_inc
- 1;
964 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
968 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
972 emit_jump_insn (gen_ble (labels
[0]));
974 emit_jump_insn (gen_bge (labels
[0]));
975 JUMP_LABEL (get_last_insn ()) = labels
[0];
976 LABEL_NUSES (labels
[0])++;
979 sequence
= gen_sequence ();
981 emit_insn_before (sequence
, loop_start
);
983 /* Only the last copy of the loop body here needs the exit
984 test, so set copy_end to exclude the compare/branch here,
985 and then reset it inside the loop when get to the last
988 if (GET_CODE (last_loop_insn
) == BARRIER
)
989 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
990 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
993 /* The immediately preceding insn is a compare which we do not
995 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
997 /* The immediately preceding insn may not be a compare, so we
999 copy_end
= PREV_INSN (last_loop_insn
);
1005 for (i
= 1; i
< unroll_number
; i
++)
1007 emit_label_after (labels
[unroll_number
- i
],
1008 PREV_INSN (loop_start
));
1010 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1011 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1012 bzero ((char *) map
->const_age_map
,
1013 maxregnum
* sizeof (unsigned));
1016 for (j
= 0; j
< max_labelno
; j
++)
1018 map
->label_map
[j
] = gen_label_rtx ();
1020 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1022 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1024 /* The last copy needs the compare/branch insns at the end,
1025 so reset copy_end here if the loop ends with a conditional
1028 if (i
== unroll_number
- 1)
1030 if (GET_CODE (last_loop_insn
) == BARRIER
)
1031 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1033 copy_end
= last_loop_insn
;
1036 /* None of the copies are the `last_iteration', so just
1037 pass zero for that parameter. */
1038 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1039 unroll_type
, start_label
, loop_end
,
1040 loop_start
, copy_end
);
1042 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1044 if (GET_CODE (last_loop_insn
) == BARRIER
)
1046 insert_before
= PREV_INSN (last_loop_insn
);
1047 copy_end
= PREV_INSN (insert_before
);
1052 /* The immediately preceding insn is a compare which we do not
1054 insert_before
= PREV_INSN (last_loop_insn
);
1055 copy_end
= PREV_INSN (insert_before
);
1057 /* The immediately preceding insn may not be a compare, so we
1059 insert_before
= last_loop_insn
;
1060 copy_end
= PREV_INSN (last_loop_insn
);
1064 /* Set unroll type to MODULO now. */
1065 unroll_type
= UNROLL_MODULO
;
1066 loop_preconditioned
= 1;
1070 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1071 the loop unless all loops are being unrolled. */
1072 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1074 if (loop_dump_stream
)
1075 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1079 /* At this point, we are guaranteed to unroll the loop. */
1081 /* For each biv and giv, determine whether it can be safely split into
1082 a different variable for each unrolled copy of the loop body.
1083 We precalculate and save this info here, since computing it is
1086 Do this before deleting any instructions from the loop, so that
1087 back_branch_in_range_p will work correctly. */
1089 if (splitting_not_safe
)
1092 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1093 end_insert_before
, unroll_number
);
1095 /* find_splittable_regs may have created some new registers, so must
1096 reallocate the reg_map with the new larger size, and must realloc
1097 the constant maps also. */
1099 maxregnum
= max_reg_num ();
1100 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1102 init_reg_map (map
, maxregnum
);
1104 /* Space is needed in some of the map for new registers, so new_maxregnum
1105 is an (over)estimate of how many registers will exist at the end. */
1106 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1108 /* Must realloc space for the constant maps, because the number of registers
1109 may have changed. */
1111 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1112 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1114 map
->const_equiv_map_size
= new_maxregnum
;
1115 global_const_equiv_map
= map
->const_equiv_map
;
1116 global_const_equiv_map_size
= new_maxregnum
;
1118 /* Search the list of bivs and givs to find ones which need to be remapped
1119 when split, and set their reg_map entry appropriately. */
1121 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1123 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1124 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1126 /* Currently, non-reduced/final-value givs are never split. */
1127 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1128 if (REGNO (v
->src_reg
) != bl
->regno
)
1129 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1133 /* Use our current register alignment and pointer flags. */
1134 map
->regno_pointer_flag
= regno_pointer_flag
;
1135 map
->regno_pointer_align
= regno_pointer_align
;
1137 /* If the loop is being partially unrolled, and the iteration variables
1138 are being split, and are being renamed for the split, then must fix up
1139 the compare/jump instruction at the end of the loop to refer to the new
1140 registers. This compare isn't copied, so the registers used in it
1141 will never be replaced if it isn't done here. */
1143 if (unroll_type
== UNROLL_MODULO
)
1145 insn
= NEXT_INSN (copy_end
);
1146 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1147 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1150 /* For unroll_number - 1 times, make a copy of each instruction
1151 between copy_start and copy_end, and insert these new instructions
1152 before the end of the loop. */
1154 for (i
= 0; i
< unroll_number
; i
++)
1156 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1157 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1158 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1161 for (j
= 0; j
< max_labelno
; j
++)
1163 map
->label_map
[j
] = gen_label_rtx ();
1165 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1167 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1169 /* If loop starts with a branch to the test, then fix it so that
1170 it points to the test of the first unrolled copy of the loop. */
1171 if (i
== 0 && loop_start
!= copy_start
)
1173 insn
= PREV_INSN (copy_start
);
1174 pattern
= PATTERN (insn
);
1176 tem
= map
->label_map
[CODE_LABEL_NUMBER
1177 (XEXP (SET_SRC (pattern
), 0))];
1178 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1180 /* Set the jump label so that it can be used by later loop unrolling
1182 JUMP_LABEL (insn
) = tem
;
1183 LABEL_NUSES (tem
)++;
1186 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1187 i
== unroll_number
- 1, unroll_type
, start_label
,
1188 loop_end
, insert_before
, insert_before
);
1191 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1192 insn to be deleted. This prevents any runaway delete_insn call from
1193 more insns that it should, as it always stops at a CODE_LABEL. */
1195 /* Delete the compare and branch at the end of the loop if completely
1196 unrolling the loop. Deleting the backward branch at the end also
1197 deletes the code label at the start of the loop. This is done at
1198 the very end to avoid problems with back_branch_in_range_p. */
1200 if (unroll_type
== UNROLL_COMPLETELY
)
1201 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1203 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1205 /* Delete all of the original loop instructions. Don't delete the
1206 LOOP_BEG note, or the first code label in the loop. */
1208 insn
= NEXT_INSN (copy_start
);
1209 while (insn
!= safety_label
)
1211 if (insn
!= start_label
)
1212 insn
= delete_insn (insn
);
1214 insn
= NEXT_INSN (insn
);
1217 /* Can now delete the 'safety' label emitted to protect us from runaway
1218 delete_insn calls. */
1219 if (INSN_DELETED_P (safety_label
))
1221 delete_insn (safety_label
);
1223 /* If exit_label exists, emit it after the loop. Doing the emit here
1224 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1225 This is needed so that mostly_true_jump in reorg.c will treat jumps
1226 to this loop end label correctly, i.e. predict that they are usually
1229 emit_label_after (exit_label
, loop_end
);
1232 /* Return true if the loop can be safely, and profitably, preconditioned
1233 so that the unrolled copies of the loop body don't need exit tests.
1235 This only works if final_value, initial_value and increment can be
1236 determined, and if increment is a constant power of 2.
1237 If increment is not a power of 2, then the preconditioning modulo
1238 operation would require a real modulo instead of a boolean AND, and this
1239 is not considered `profitable'. */
1241 /* ??? If the loop is known to be executed very many times, or the machine
1242 has a very cheap divide instruction, then preconditioning is a win even
1243 when the increment is not a power of 2. Use RTX_COST to compute
1244 whether divide is cheap. */
1247 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1249 rtx
*initial_value
, *final_value
, *increment
;
1250 rtx loop_start
, loop_end
;
1253 if (loop_n_iterations
> 0)
1255 *initial_value
= const0_rtx
;
1256 *increment
= const1_rtx
;
1257 *final_value
= GEN_INT (loop_n_iterations
);
1259 if (loop_dump_stream
)
1260 fprintf (loop_dump_stream
,
1261 "Preconditioning: Success, number of iterations known, %d.\n",
1266 if (loop_initial_value
== 0)
1268 if (loop_dump_stream
)
1269 fprintf (loop_dump_stream
,
1270 "Preconditioning: Could not find initial value.\n");
1273 else if (loop_increment
== 0)
1275 if (loop_dump_stream
)
1276 fprintf (loop_dump_stream
,
1277 "Preconditioning: Could not find increment value.\n");
1280 else if (GET_CODE (loop_increment
) != CONST_INT
)
1282 if (loop_dump_stream
)
1283 fprintf (loop_dump_stream
,
1284 "Preconditioning: Increment not a constant.\n");
1287 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1288 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1290 if (loop_dump_stream
)
1291 fprintf (loop_dump_stream
,
1292 "Preconditioning: Increment not a constant power of 2.\n");
1296 /* Unsigned_compare and compare_dir can be ignored here, since they do
1297 not matter for preconditioning. */
1299 if (loop_final_value
== 0)
1301 if (loop_dump_stream
)
1302 fprintf (loop_dump_stream
,
1303 "Preconditioning: EQ comparison loop.\n");
1307 /* Must ensure that final_value is invariant, so call invariant_p to
1308 check. Before doing so, must check regno against max_reg_before_loop
1309 to make sure that the register is in the range covered by invariant_p.
1310 If it isn't, then it is most likely a biv/giv which by definition are
1312 if ((GET_CODE (loop_final_value
) == REG
1313 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1314 || (GET_CODE (loop_final_value
) == PLUS
1315 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1316 || ! invariant_p (loop_final_value
))
1318 if (loop_dump_stream
)
1319 fprintf (loop_dump_stream
,
1320 "Preconditioning: Final value not invariant.\n");
1324 /* Fail for floating point values, since the caller of this function
1325 does not have code to deal with them. */
1326 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1327 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1329 if (loop_dump_stream
)
1330 fprintf (loop_dump_stream
,
1331 "Preconditioning: Floating point final or initial value.\n");
1335 /* Now set initial_value to be the iteration_var, since that may be a
1336 simpler expression, and is guaranteed to be correct if all of the
1337 above tests succeed.
1339 We can not use the initial_value as calculated, because it will be
1340 one too small for loops of the form "while (i-- > 0)". We can not
1341 emit code before the loop_skip_over insns to fix this problem as this
1342 will then give a number one too large for loops of the form
1345 Note that all loops that reach here are entered at the top, because
1346 this function is not called if the loop starts with a jump. */
1348 /* Fail if loop_iteration_var is not live before loop_start, since we need
1349 to test its value in the preconditioning code. */
1351 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1352 > INSN_LUID (loop_start
))
1354 if (loop_dump_stream
)
1355 fprintf (loop_dump_stream
,
1356 "Preconditioning: Iteration var not live before loop start.\n");
1360 *initial_value
= loop_iteration_var
;
1361 *increment
= loop_increment
;
1362 *final_value
= loop_final_value
;
1365 if (loop_dump_stream
)
1366 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1371 /* All pseudo-registers must be mapped to themselves. Two hard registers
1372 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1373 REGNUM, to avoid function-inlining specific conversions of these
1374 registers. All other hard regs can not be mapped because they may be
1379 init_reg_map (map
, maxregnum
)
1380 struct inline_remap
*map
;
1385 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1386 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1387 /* Just clear the rest of the entries. */
1388 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1389 map
->reg_map
[i
] = 0;
1391 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1392 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1393 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1394 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1397 /* Strength-reduction will often emit code for optimized biv/givs which
1398 calculates their value in a temporary register, and then copies the result
1399 to the iv. This procedure reconstructs the pattern computing the iv;
1400 verifying that all operands are of the proper form.
1402 The return value is the amount that the giv is incremented by. */
1405 calculate_giv_inc (pattern
, src_insn
, regno
)
1406 rtx pattern
, src_insn
;
1410 rtx increment_total
= 0;
1414 /* Verify that we have an increment insn here. First check for a plus
1415 as the set source. */
1416 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1418 /* SR sometimes computes the new giv value in a temp, then copies it
1420 src_insn
= PREV_INSN (src_insn
);
1421 pattern
= PATTERN (src_insn
);
1422 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1425 /* The last insn emitted is not needed, so delete it to avoid confusing
1426 the second cse pass. This insn sets the giv unnecessarily. */
1427 delete_insn (get_last_insn ());
1430 /* Verify that we have a constant as the second operand of the plus. */
1431 increment
= XEXP (SET_SRC (pattern
), 1);
1432 if (GET_CODE (increment
) != CONST_INT
)
1434 /* SR sometimes puts the constant in a register, especially if it is
1435 too big to be an add immed operand. */
1436 src_insn
= PREV_INSN (src_insn
);
1437 increment
= SET_SRC (PATTERN (src_insn
));
1439 /* SR may have used LO_SUM to compute the constant if it is too large
1440 for a load immed operand. In this case, the constant is in operand
1441 one of the LO_SUM rtx. */
1442 if (GET_CODE (increment
) == LO_SUM
)
1443 increment
= XEXP (increment
, 1);
1444 else if (GET_CODE (increment
) == IOR
1445 || GET_CODE (increment
) == ASHIFT
)
1447 /* The rs6000 port loads some constants with IOR.
1448 The alpha port loads some constants with ASHIFT. */
1449 rtx second_part
= XEXP (increment
, 1);
1450 enum rtx_code code
= GET_CODE (increment
);
1452 src_insn
= PREV_INSN (src_insn
);
1453 increment
= SET_SRC (PATTERN (src_insn
));
1454 /* Don't need the last insn anymore. */
1455 delete_insn (get_last_insn ());
1457 if (GET_CODE (second_part
) != CONST_INT
1458 || GET_CODE (increment
) != CONST_INT
)
1462 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1464 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1467 if (GET_CODE (increment
) != CONST_INT
)
1470 /* The insn loading the constant into a register is no longer needed,
1472 delete_insn (get_last_insn ());
1475 if (increment_total
)
1476 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1478 increment_total
= increment
;
1480 /* Check that the source register is the same as the register we expected
1481 to see as the source. If not, something is seriously wrong. */
1482 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1483 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1485 /* Some machines (e.g. the romp), may emit two add instructions for
1486 certain constants, so lets try looking for another add immediately
1487 before this one if we have only seen one add insn so far. */
1493 src_insn
= PREV_INSN (src_insn
);
1494 pattern
= PATTERN (src_insn
);
1496 delete_insn (get_last_insn ());
1504 return increment_total
;
1507 /* Copy REG_NOTES, except for insn references, because not all insn_map
1508 entries are valid yet. We do need to copy registers now though, because
1509 the reg_map entries can change during copying. */
1512 initial_reg_note_copy (notes
, map
)
1514 struct inline_remap
*map
;
1521 copy
= rtx_alloc (GET_CODE (notes
));
1522 PUT_MODE (copy
, GET_MODE (notes
));
1524 if (GET_CODE (notes
) == EXPR_LIST
)
1525 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1526 else if (GET_CODE (notes
) == INSN_LIST
)
1527 /* Don't substitute for these yet. */
1528 XEXP (copy
, 0) = XEXP (notes
, 0);
1532 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1537 /* Fixup insn references in copied REG_NOTES. */
1540 final_reg_note_copy (notes
, map
)
1542 struct inline_remap
*map
;
1546 for (note
= notes
; note
; note
= XEXP (note
, 1))
1547 if (GET_CODE (note
) == INSN_LIST
)
1548 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1551 /* Copy each instruction in the loop, substituting from map as appropriate.
1552 This is very similar to a loop in expand_inline_function. */
1555 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1556 unroll_type
, start_label
, loop_end
, insert_before
,
1558 rtx copy_start
, copy_end
;
1559 struct inline_remap
*map
;
1562 enum unroll_types unroll_type
;
1563 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1567 int dest_reg_was_split
, i
;
1569 rtx final_label
= 0;
1570 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1572 /* If this isn't the last iteration, then map any references to the
1573 start_label to final_label. Final label will then be emitted immediately
1574 after the end of this loop body if it was ever used.
1576 If this is the last iteration, then map references to the start_label
1578 if (! last_iteration
)
1580 final_label
= gen_label_rtx ();
1581 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1584 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1591 insn
= NEXT_INSN (insn
);
1593 map
->orig_asm_operands_vector
= 0;
1595 switch (GET_CODE (insn
))
1598 pattern
= PATTERN (insn
);
1602 /* Check to see if this is a giv that has been combined with
1603 some split address givs. (Combined in the sense that
1604 `combine_givs' in loop.c has put two givs in the same register.)
1605 In this case, we must search all givs based on the same biv to
1606 find the address givs. Then split the address givs.
1607 Do this before splitting the giv, since that may map the
1608 SET_DEST to a new register. */
1610 if (GET_CODE (pattern
) == SET
1611 && GET_CODE (SET_DEST (pattern
)) == REG
1612 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1614 struct iv_class
*bl
;
1615 struct induction
*v
, *tv
;
1616 int regno
= REGNO (SET_DEST (pattern
));
1618 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1619 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1621 /* Although the giv_inc amount is not needed here, we must call
1622 calculate_giv_inc here since it might try to delete the
1623 last insn emitted. If we wait until later to call it,
1624 we might accidentally delete insns generated immediately
1625 below by emit_unrolled_add. */
1627 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1629 /* Now find all address giv's that were combined with this
1631 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1632 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1636 /* If this DEST_ADDR giv was not split, then ignore it. */
1637 if (*tv
->location
!= tv
->dest_reg
)
1640 /* Scale this_giv_inc if the multiplicative factors of
1641 the two givs are different. */
1642 this_giv_inc
= INTVAL (giv_inc
);
1643 if (tv
->mult_val
!= v
->mult_val
)
1644 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1645 * INTVAL (tv
->mult_val
));
1647 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1648 *tv
->location
= tv
->dest_reg
;
1650 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1652 /* Must emit an insn to increment the split address
1653 giv. Add in the const_adjust field in case there
1654 was a constant eliminated from the address. */
1655 rtx value
, dest_reg
;
1657 /* tv->dest_reg will be either a bare register,
1658 or else a register plus a constant. */
1659 if (GET_CODE (tv
->dest_reg
) == REG
)
1660 dest_reg
= tv
->dest_reg
;
1662 dest_reg
= XEXP (tv
->dest_reg
, 0);
1664 /* Check for shared address givs, and avoid
1665 incrementing the shared pseudo reg more than
1667 if (! tv
->same_insn
)
1669 /* tv->dest_reg may actually be a (PLUS (REG)
1670 (CONST)) here, so we must call plus_constant
1671 to add the const_adjust amount before calling
1672 emit_unrolled_add below. */
1673 value
= plus_constant (tv
->dest_reg
,
1676 /* The constant could be too large for an add
1677 immediate, so can't directly emit an insn
1679 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1683 /* Reset the giv to be just the register again, in case
1684 it is used after the set we have just emitted.
1685 We must subtract the const_adjust factor added in
1687 tv
->dest_reg
= plus_constant (dest_reg
,
1688 - tv
->const_adjust
);
1689 *tv
->location
= tv
->dest_reg
;
1694 /* If this is a setting of a splittable variable, then determine
1695 how to split the variable, create a new set based on this split,
1696 and set up the reg_map so that later uses of the variable will
1697 use the new split variable. */
1699 dest_reg_was_split
= 0;
1701 if (GET_CODE (pattern
) == SET
1702 && GET_CODE (SET_DEST (pattern
)) == REG
1703 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1705 int regno
= REGNO (SET_DEST (pattern
));
1707 dest_reg_was_split
= 1;
1709 /* Compute the increment value for the giv, if it wasn't
1710 already computed above. */
1713 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1714 giv_dest_reg
= SET_DEST (pattern
);
1715 giv_src_reg
= SET_DEST (pattern
);
1717 if (unroll_type
== UNROLL_COMPLETELY
)
1719 /* Completely unrolling the loop. Set the induction
1720 variable to a known constant value. */
1722 /* The value in splittable_regs may be an invariant
1723 value, so we must use plus_constant here. */
1724 splittable_regs
[regno
]
1725 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1727 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1729 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1730 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1734 /* The splittable_regs value must be a REG or a
1735 CONST_INT, so put the entire value in the giv_src_reg
1737 giv_src_reg
= splittable_regs
[regno
];
1738 giv_inc
= const0_rtx
;
1743 /* Partially unrolling loop. Create a new pseudo
1744 register for the iteration variable, and set it to
1745 be a constant plus the original register. Except
1746 on the last iteration, when the result has to
1747 go back into the original iteration var register. */
1749 /* Handle bivs which must be mapped to a new register
1750 when split. This happens for bivs which need their
1751 final value set before loop entry. The new register
1752 for the biv was stored in the biv's first struct
1753 induction entry by find_splittable_regs. */
1755 if (regno
< max_reg_before_loop
1756 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1758 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1759 giv_dest_reg
= giv_src_reg
;
1763 /* If non-reduced/final-value givs were split, then
1764 this would have to remap those givs also. See
1765 find_splittable_regs. */
1768 splittable_regs
[regno
]
1769 = GEN_INT (INTVAL (giv_inc
)
1770 + INTVAL (splittable_regs
[regno
]));
1771 giv_inc
= splittable_regs
[regno
];
1773 /* Now split the induction variable by changing the dest
1774 of this insn to a new register, and setting its
1775 reg_map entry to point to this new register.
1777 If this is the last iteration, and this is the last insn
1778 that will update the iv, then reuse the original dest,
1779 to ensure that the iv will have the proper value when
1780 the loop exits or repeats.
1782 Using splittable_regs_updates here like this is safe,
1783 because it can only be greater than one if all
1784 instructions modifying the iv are always executed in
1787 if (! last_iteration
1788 || (splittable_regs_updates
[regno
]-- != 1))
1790 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1792 map
->reg_map
[regno
] = tem
;
1795 map
->reg_map
[regno
] = giv_src_reg
;
1798 /* The constant being added could be too large for an add
1799 immediate, so can't directly emit an insn here. */
1800 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1801 copy
= get_last_insn ();
1802 pattern
= PATTERN (copy
);
1806 pattern
= copy_rtx_and_substitute (pattern
, map
);
1807 copy
= emit_insn (pattern
);
1809 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1812 /* If this insn is setting CC0, it may need to look at
1813 the insn that uses CC0 to see what type of insn it is.
1814 In that case, the call to recog via validate_change will
1815 fail. So don't substitute constants here. Instead,
1816 do it when we emit the following insn.
1818 For example, see the pyr.md file. That machine has signed and
1819 unsigned compares. The compare patterns must check the
1820 following branch insn to see which what kind of compare to
1823 If the previous insn set CC0, substitute constants on it as
1825 if (sets_cc0_p (PATTERN (copy
)) != 0)
1830 try_constants (cc0_insn
, map
);
1832 try_constants (copy
, map
);
1835 try_constants (copy
, map
);
1838 /* Make split induction variable constants `permanent' since we
1839 know there are no backward branches across iteration variable
1840 settings which would invalidate this. */
1841 if (dest_reg_was_split
)
1843 int regno
= REGNO (SET_DEST (pattern
));
1845 if (regno
< map
->const_equiv_map_size
1846 && map
->const_age_map
[regno
] == map
->const_age
)
1847 map
->const_age_map
[regno
] = -1;
1852 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1853 copy
= emit_jump_insn (pattern
);
1854 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1856 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1857 && ! last_iteration
)
1859 /* This is a branch to the beginning of the loop; this is the
1860 last insn being copied; and this is not the last iteration.
1861 In this case, we want to change the original fall through
1862 case to be a branch past the end of the loop, and the
1863 original jump label case to fall_through. */
1865 if (invert_exp (pattern
, copy
))
1867 if (! redirect_exp (&pattern
,
1868 map
->label_map
[CODE_LABEL_NUMBER
1869 (JUMP_LABEL (insn
))],
1876 rtx lab
= gen_label_rtx ();
1877 /* Can't do it by reversing the jump (probably because we
1878 couldn't reverse the conditions), so emit a new
1879 jump_insn after COPY, and redirect the jump around
1881 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1882 jmp
= emit_barrier_after (jmp
);
1883 emit_label_after (lab
, jmp
);
1884 LABEL_NUSES (lab
) = 0;
1885 if (! redirect_exp (&pattern
,
1886 map
->label_map
[CODE_LABEL_NUMBER
1887 (JUMP_LABEL (insn
))],
1895 try_constants (cc0_insn
, map
);
1898 try_constants (copy
, map
);
1900 /* Set the jump label of COPY correctly to avoid problems with
1901 later passes of unroll_loop, if INSN had jump label set. */
1902 if (JUMP_LABEL (insn
))
1906 /* Can't use the label_map for every insn, since this may be
1907 the backward branch, and hence the label was not mapped. */
1908 if (GET_CODE (pattern
) == SET
)
1910 tem
= SET_SRC (pattern
);
1911 if (GET_CODE (tem
) == LABEL_REF
)
1912 label
= XEXP (tem
, 0);
1913 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1915 if (XEXP (tem
, 1) != pc_rtx
)
1916 label
= XEXP (XEXP (tem
, 1), 0);
1918 label
= XEXP (XEXP (tem
, 2), 0);
1922 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1923 JUMP_LABEL (copy
) = label
;
1926 /* An unrecognizable jump insn, probably the entry jump
1927 for a switch statement. This label must have been mapped,
1928 so just use the label_map to get the new jump label. */
1930 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1933 /* If this is a non-local jump, then must increase the label
1934 use count so that the label will not be deleted when the
1935 original jump is deleted. */
1936 LABEL_NUSES (JUMP_LABEL (copy
))++;
1938 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1939 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1941 rtx pat
= PATTERN (copy
);
1942 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1943 int len
= XVECLEN (pat
, diff_vec_p
);
1946 for (i
= 0; i
< len
; i
++)
1947 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1950 /* If this used to be a conditional jump insn but whose branch
1951 direction is now known, we must do something special. */
1952 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1955 /* The previous insn set cc0 for us. So delete it. */
1956 delete_insn (PREV_INSN (copy
));
1959 /* If this is now a no-op, delete it. */
1960 if (map
->last_pc_value
== pc_rtx
)
1962 /* Don't let delete_insn delete the label referenced here,
1963 because we might possibly need it later for some other
1964 instruction in the loop. */
1965 if (JUMP_LABEL (copy
))
1966 LABEL_NUSES (JUMP_LABEL (copy
))++;
1968 if (JUMP_LABEL (copy
))
1969 LABEL_NUSES (JUMP_LABEL (copy
))--;
1973 /* Otherwise, this is unconditional jump so we must put a
1974 BARRIER after it. We could do some dead code elimination
1975 here, but jump.c will do it just as well. */
1981 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1982 copy
= emit_call_insn (pattern
);
1983 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1985 /* Because the USAGE information potentially contains objects other
1986 than hard registers, we need to copy it. */
1987 CALL_INSN_FUNCTION_USAGE (copy
) =
1988 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1992 try_constants (cc0_insn
, map
);
1995 try_constants (copy
, map
);
1997 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1998 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
1999 map
->const_equiv_map
[i
] = 0;
2003 /* If this is the loop start label, then we don't need to emit a
2004 copy of this label since no one will use it. */
2006 if (insn
!= start_label
)
2008 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
2014 copy
= emit_barrier ();
2018 /* VTOP notes are valid only before the loop exit test. If placed
2019 anywhere else, loop may generate bad code. */
2021 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2022 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2023 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2024 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2025 NOTE_LINE_NUMBER (insn
));
2035 map
->insn_map
[INSN_UID (insn
)] = copy
;
2037 while (insn
!= copy_end
);
2039 /* Now finish coping the REG_NOTES. */
2043 insn
= NEXT_INSN (insn
);
2044 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2045 || GET_CODE (insn
) == CALL_INSN
)
2046 && map
->insn_map
[INSN_UID (insn
)])
2047 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2049 while (insn
!= copy_end
);
2051 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2052 each of these notes here, since there may be some important ones, such as
2053 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2054 iteration, because the original notes won't be deleted.
2056 We can't use insert_before here, because when from preconditioning,
2057 insert_before points before the loop. We can't use copy_end, because
2058 there may be insns already inserted after it (which we don't want to
2059 copy) when not from preconditioning code. */
2061 if (! last_iteration
)
2063 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2065 if (GET_CODE (insn
) == NOTE
2066 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2067 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2071 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2072 emit_label (final_label
);
2074 tem
= gen_sequence ();
2076 emit_insn_before (tem
, insert_before
);
2079 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2080 emitted. This will correctly handle the case where the increment value
2081 won't fit in the immediate field of a PLUS insns. */
2084 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2085 rtx dest_reg
, src_reg
, increment
;
2089 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2090 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2092 if (dest_reg
!= result
)
2093 emit_move_insn (dest_reg
, result
);
2096 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2097 is a backward branch in that range that branches to somewhere between
2098 LOOP_START and INSN. Returns 0 otherwise. */
2100 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2101 In practice, this is not a problem, because this function is seldom called,
2102 and uses a negligible amount of CPU time on average. */
2105 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2107 rtx loop_start
, loop_end
;
2109 rtx p
, q
, target_insn
;
2111 /* Stop before we get to the backward branch at the end of the loop. */
2112 loop_end
= prev_nonnote_insn (loop_end
);
2113 if (GET_CODE (loop_end
) == BARRIER
)
2114 loop_end
= PREV_INSN (loop_end
);
2116 /* Check in case insn has been deleted, search forward for first non
2117 deleted insn following it. */
2118 while (INSN_DELETED_P (insn
))
2119 insn
= NEXT_INSN (insn
);
2121 /* Check for the case where insn is the last insn in the loop. */
2122 if (insn
== loop_end
)
2125 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2127 if (GET_CODE (p
) == JUMP_INSN
)
2129 target_insn
= JUMP_LABEL (p
);
2131 /* Search from loop_start to insn, to see if one of them is
2132 the target_insn. We can't use INSN_LUID comparisons here,
2133 since insn may not have an LUID entry. */
2134 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2135 if (q
== target_insn
)
2143 /* Try to generate the simplest rtx for the expression
2144 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2148 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2149 rtx mult1
, mult2
, add1
;
2150 enum machine_mode mode
;
2155 /* The modes must all be the same. This should always be true. For now,
2156 check to make sure. */
2157 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2158 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2159 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2162 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2163 will be a constant. */
2164 if (GET_CODE (mult1
) == CONST_INT
)
2171 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2173 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2175 /* Again, put the constant second. */
2176 if (GET_CODE (add1
) == CONST_INT
)
2183 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2185 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2190 /* Searches the list of induction struct's for the biv BL, to try to calculate
2191 the total increment value for one iteration of the loop as a constant.
2193 Returns the increment value as an rtx, simplified as much as possible,
2194 if it can be calculated. Otherwise, returns 0. */
2197 biv_total_increment (bl
, loop_start
, loop_end
)
2198 struct iv_class
*bl
;
2199 rtx loop_start
, loop_end
;
2201 struct induction
*v
;
2204 /* For increment, must check every instruction that sets it. Each
2205 instruction must be executed only once each time through the loop.
2206 To verify this, we check that the the insn is always executed, and that
2207 there are no backward branches after the insn that branch to before it.
2208 Also, the insn must have a mult_val of one (to make sure it really is
2211 result
= const0_rtx
;
2212 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2214 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2215 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2216 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2224 /* Determine the initial value of the iteration variable, and the amount
2225 that it is incremented each loop. Use the tables constructed by
2226 the strength reduction pass to calculate these values.
2228 Initial_value and/or increment are set to zero if their values could not
2232 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2233 rtx iteration_var
, *initial_value
, *increment
;
2234 rtx loop_start
, loop_end
;
2236 struct iv_class
*bl
;
2237 struct induction
*v
, *b
;
2239 /* Clear the result values, in case no answer can be found. */
2243 /* The iteration variable can be either a giv or a biv. Check to see
2244 which it is, and compute the variable's initial value, and increment
2245 value if possible. */
2247 /* If this is a new register, can't handle it since we don't have any
2248 reg_iv_type entry for it. */
2249 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2251 if (loop_dump_stream
)
2252 fprintf (loop_dump_stream
,
2253 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2256 /* Reject iteration variables larger than the host long size, since they
2257 could result in a number of iterations greater than the range of our
2258 `unsigned long' variable loop_n_iterations. */
2259 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2261 if (loop_dump_stream
)
2262 fprintf (loop_dump_stream
,
2263 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2266 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2268 if (loop_dump_stream
)
2269 fprintf (loop_dump_stream
,
2270 "Loop unrolling: Iteration var not an integer.\n");
2273 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2275 /* Grab initial value, only useful if it is a constant. */
2276 bl
= reg_biv_class
[REGNO (iteration_var
)];
2277 *initial_value
= bl
->initial_value
;
2279 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2281 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2284 /* ??? The code below does not work because the incorrect number of
2285 iterations is calculated when the biv is incremented after the giv
2286 is set (which is the usual case). This can probably be accounted
2287 for by biasing the initial_value by subtracting the amount of the
2288 increment that occurs between the giv set and the giv test. However,
2289 a giv as an iterator is very rare, so it does not seem worthwhile
2291 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2292 if (loop_dump_stream
)
2293 fprintf (loop_dump_stream
,
2294 "Loop unrolling: Giv iterators are not handled.\n");
2297 /* Initial value is mult_val times the biv's initial value plus
2298 add_val. Only useful if it is a constant. */
2299 v
= reg_iv_info
[REGNO (iteration_var
)];
2300 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2301 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2302 v
->add_val
, v
->mode
);
2304 /* Increment value is mult_val times the increment value of the biv. */
2306 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2308 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2314 if (loop_dump_stream
)
2315 fprintf (loop_dump_stream
,
2316 "Loop unrolling: Not basic or general induction var.\n");
2321 /* Calculate the approximate final value of the iteration variable
2322 which has an loop exit test with code COMPARISON_CODE and comparison value
2323 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2324 was signed or unsigned, and the direction of the comparison. This info is
2325 needed to calculate the number of loop iterations. */
2328 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2329 enum rtx_code comparison_code
;
2330 rtx comparison_value
;
2334 /* Calculate the final value of the induction variable.
2335 The exact final value depends on the branch operator, and increment sign.
2336 This is only an approximate value. It will be wrong if the iteration
2337 variable is not incremented by one each time through the loop, and
2338 approx final value - start value % increment != 0. */
2341 switch (comparison_code
)
2347 return plus_constant (comparison_value
, 1);
2352 return plus_constant (comparison_value
, -1);
2354 /* Can not calculate a final value for this case. */
2361 return comparison_value
;
2367 return comparison_value
;
2370 return comparison_value
;
2376 /* For each biv and giv, determine whether it can be safely split into
2377 a different variable for each unrolled copy of the loop body. If it
2378 is safe to split, then indicate that by saving some useful info
2379 in the splittable_regs array.
2381 If the loop is being completely unrolled, then splittable_regs will hold
2382 the current value of the induction variable while the loop is unrolled.
2383 It must be set to the initial value of the induction variable here.
2384 Otherwise, splittable_regs will hold the difference between the current
2385 value of the induction variable and the value the induction variable had
2386 at the top of the loop. It must be set to the value 0 here.
2388 Returns the total number of instructions that set registers that are
2391 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2392 constant values are unnecessary, since we can easily calculate increment
2393 values in this case even if nothing is constant. The increment value
2394 should not involve a multiply however. */
2396 /* ?? Even if the biv/giv increment values aren't constant, it may still
2397 be beneficial to split the variable if the loop is only unrolled a few
2398 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2401 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2403 enum unroll_types unroll_type
;
2404 rtx loop_start
, loop_end
;
2405 rtx end_insert_before
;
2408 struct iv_class
*bl
;
2409 struct induction
*v
;
2411 rtx biv_final_value
;
2415 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2417 /* Biv_total_increment must return a constant value,
2418 otherwise we can not calculate the split values. */
2420 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2421 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2424 /* The loop must be unrolled completely, or else have a known number
2425 of iterations and only one exit, or else the biv must be dead
2426 outside the loop, or else the final value must be known. Otherwise,
2427 it is unsafe to split the biv since it may not have the proper
2428 value on loop exit. */
2430 /* loop_number_exit_count is non-zero if the loop has an exit other than
2431 a fall through at the end. */
2434 biv_final_value
= 0;
2435 if (unroll_type
!= UNROLL_COMPLETELY
2436 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2437 || unroll_type
== UNROLL_NAIVE
)
2438 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2440 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2441 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2442 < INSN_LUID (bl
->init_insn
))
2443 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2444 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2447 /* If any of the insns setting the BIV don't do so with a simple
2448 PLUS, we don't know how to split it. */
2449 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2450 if ((tem
= single_set (v
->insn
)) == 0
2451 || GET_CODE (SET_DEST (tem
)) != REG
2452 || REGNO (SET_DEST (tem
)) != bl
->regno
2453 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2456 /* If final value is non-zero, then must emit an instruction which sets
2457 the value of the biv to the proper value. This is done after
2458 handling all of the givs, since some of them may need to use the
2459 biv's value in their initialization code. */
2461 /* This biv is splittable. If completely unrolling the loop, save
2462 the biv's initial value. Otherwise, save the constant zero. */
2464 if (biv_splittable
== 1)
2466 if (unroll_type
== UNROLL_COMPLETELY
)
2468 /* If the initial value of the biv is itself (i.e. it is too
2469 complicated for strength_reduce to compute), or is a hard
2470 register, or it isn't invariant, then we must create a new
2471 pseudo reg to hold the initial value of the biv. */
2473 if (GET_CODE (bl
->initial_value
) == REG
2474 && (REGNO (bl
->initial_value
) == bl
->regno
2475 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2476 || ! invariant_p (bl
->initial_value
)))
2478 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2480 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2483 if (loop_dump_stream
)
2484 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2485 bl
->regno
, REGNO (tem
));
2487 splittable_regs
[bl
->regno
] = tem
;
2490 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2493 splittable_regs
[bl
->regno
] = const0_rtx
;
2495 /* Save the number of instructions that modify the biv, so that
2496 we can treat the last one specially. */
2498 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2499 result
+= bl
->biv_count
;
2501 if (loop_dump_stream
)
2502 fprintf (loop_dump_stream
,
2503 "Biv %d safe to split.\n", bl
->regno
);
2506 /* Check every giv that depends on this biv to see whether it is
2507 splittable also. Even if the biv isn't splittable, givs which
2508 depend on it may be splittable if the biv is live outside the
2509 loop, and the givs aren't. */
2511 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2512 increment
, unroll_number
);
2514 /* If final value is non-zero, then must emit an instruction which sets
2515 the value of the biv to the proper value. This is done after
2516 handling all of the givs, since some of them may need to use the
2517 biv's value in their initialization code. */
2518 if (biv_final_value
)
2520 /* If the loop has multiple exits, emit the insns before the
2521 loop to ensure that it will always be executed no matter
2522 how the loop exits. Otherwise emit the insn after the loop,
2523 since this is slightly more efficient. */
2524 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2525 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2530 /* Create a new register to hold the value of the biv, and then
2531 set the biv to its final value before the loop start. The biv
2532 is set to its final value before loop start to ensure that
2533 this insn will always be executed, no matter how the loop
2535 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2536 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2538 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2542 if (loop_dump_stream
)
2543 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2544 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2546 /* Set up the mapping from the original biv register to the new
2548 bl
->biv
->src_reg
= tem
;
2555 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2556 for the instruction that is using it. Do not make any changes to that
2560 verify_addresses (v
, giv_inc
, unroll_number
)
2561 struct induction
*v
;
2566 rtx orig_addr
= *v
->location
;
2567 rtx last_addr
= plus_constant (v
->dest_reg
,
2568 INTVAL (giv_inc
) * (unroll_number
- 1));
2570 /* First check to see if either address would fail. */
2571 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2572 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2575 /* Now put things back the way they were before. This will always
2577 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2582 /* For every giv based on the biv BL, check to determine whether it is
2583 splittable. This is a subroutine to find_splittable_regs ().
2585 Return the number of instructions that set splittable registers. */
2588 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2590 struct iv_class
*bl
;
2591 enum unroll_types unroll_type
;
2592 rtx loop_start
, loop_end
;
2596 struct induction
*v
, *v2
;
2601 /* Scan the list of givs, and set the same_insn field when there are
2602 multiple identical givs in the same insn. */
2603 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2604 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2605 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2609 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2613 /* Only split the giv if it has already been reduced, or if the loop is
2614 being completely unrolled. */
2615 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2618 /* The giv can be split if the insn that sets the giv is executed once
2619 and only once on every iteration of the loop. */
2620 /* An address giv can always be split. v->insn is just a use not a set,
2621 and hence it does not matter whether it is always executed. All that
2622 matters is that all the biv increments are always executed, and we
2623 won't reach here if they aren't. */
2624 if (v
->giv_type
!= DEST_ADDR
2625 && (! v
->always_computable
2626 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2629 /* The giv increment value must be a constant. */
2630 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2632 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2635 /* The loop must be unrolled completely, or else have a known number of
2636 iterations and only one exit, or else the giv must be dead outside
2637 the loop, or else the final value of the giv must be known.
2638 Otherwise, it is not safe to split the giv since it may not have the
2639 proper value on loop exit. */
2641 /* The used outside loop test will fail for DEST_ADDR givs. They are
2642 never used outside the loop anyways, so it is always safe to split a
2646 if (unroll_type
!= UNROLL_COMPLETELY
2647 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2648 || unroll_type
== UNROLL_NAIVE
)
2649 && v
->giv_type
!= DEST_ADDR
2650 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2651 /* Check for the case where the pseudo is set by a shift/add
2652 sequence, in which case the first insn setting the pseudo
2653 is the first insn of the shift/add sequence. */
2654 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2655 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2656 != INSN_UID (XEXP (tem
, 0)))))
2657 /* Line above always fails if INSN was moved by loop opt. */
2658 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2659 >= INSN_LUID (loop_end
)))
2660 && ! (final_value
= v
->final_value
))
2664 /* Currently, non-reduced/final-value givs are never split. */
2665 /* Should emit insns after the loop if possible, as the biv final value
2668 /* If the final value is non-zero, and the giv has not been reduced,
2669 then must emit an instruction to set the final value. */
2670 if (final_value
&& !v
->new_reg
)
2672 /* Create a new register to hold the value of the giv, and then set
2673 the giv to its final value before the loop start. The giv is set
2674 to its final value before loop start to ensure that this insn
2675 will always be executed, no matter how we exit. */
2676 tem
= gen_reg_rtx (v
->mode
);
2677 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2678 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2681 if (loop_dump_stream
)
2682 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2683 REGNO (v
->dest_reg
), REGNO (tem
));
2689 /* This giv is splittable. If completely unrolling the loop, save the
2690 giv's initial value. Otherwise, save the constant zero for it. */
2692 if (unroll_type
== UNROLL_COMPLETELY
)
2694 /* It is not safe to use bl->initial_value here, because it may not
2695 be invariant. It is safe to use the initial value stored in
2696 the splittable_regs array if it is set. In rare cases, it won't
2697 be set, so then we do exactly the same thing as
2698 find_splittable_regs does to get a safe value. */
2699 rtx biv_initial_value
;
2701 if (splittable_regs
[bl
->regno
])
2702 biv_initial_value
= splittable_regs
[bl
->regno
];
2703 else if (GET_CODE (bl
->initial_value
) != REG
2704 || (REGNO (bl
->initial_value
) != bl
->regno
2705 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2706 biv_initial_value
= bl
->initial_value
;
2709 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2711 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2713 biv_initial_value
= tem
;
2715 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2716 v
->add_val
, v
->mode
);
2723 /* If a giv was combined with another giv, then we can only split
2724 this giv if the giv it was combined with was reduced. This
2725 is because the value of v->new_reg is meaningless in this
2727 if (v
->same
&& ! v
->same
->new_reg
)
2729 if (loop_dump_stream
)
2730 fprintf (loop_dump_stream
,
2731 "giv combined with unreduced giv not split.\n");
2734 /* If the giv is an address destination, it could be something other
2735 than a simple register, these have to be treated differently. */
2736 else if (v
->giv_type
== DEST_REG
)
2738 /* If value is not a constant, register, or register plus
2739 constant, then compute its value into a register before
2740 loop start. This prevents invalid rtx sharing, and should
2741 generate better code. We can use bl->initial_value here
2742 instead of splittable_regs[bl->regno] because this code
2743 is going before the loop start. */
2744 if (unroll_type
== UNROLL_COMPLETELY
2745 && GET_CODE (value
) != CONST_INT
2746 && GET_CODE (value
) != REG
2747 && (GET_CODE (value
) != PLUS
2748 || GET_CODE (XEXP (value
, 0)) != REG
2749 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2751 rtx tem
= gen_reg_rtx (v
->mode
);
2752 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2753 v
->add_val
, tem
, loop_start
);
2757 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2761 /* Splitting address givs is useful since it will often allow us
2762 to eliminate some increment insns for the base giv as
2765 /* If the addr giv is combined with a dest_reg giv, then all
2766 references to that dest reg will be remapped, which is NOT
2767 what we want for split addr regs. We always create a new
2768 register for the split addr giv, just to be safe. */
2770 /* ??? If there are multiple address givs which have been
2771 combined with the same dest_reg giv, then we may only need
2772 one new register for them. Pulling out constants below will
2773 catch some of the common cases of this. Currently, I leave
2774 the work of simplifying multiple address givs to the
2775 following cse pass. */
2777 /* As a special case, if we have multiple identical address givs
2778 within a single instruction, then we do use a single pseudo
2779 reg for both. This is necessary in case one is a match_dup
2782 v
->const_adjust
= 0;
2786 v
->dest_reg
= v
->same_insn
->dest_reg
;
2787 if (loop_dump_stream
)
2788 fprintf (loop_dump_stream
,
2789 "Sharing address givs in insn %d\n",
2790 INSN_UID (v
->insn
));
2792 else if (unroll_type
!= UNROLL_COMPLETELY
)
2794 /* If not completely unrolling the loop, then create a new
2795 register to hold the split value of the DEST_ADDR giv.
2796 Emit insn to initialize its value before loop start. */
2797 tem
= gen_reg_rtx (v
->mode
);
2799 /* If the address giv has a constant in its new_reg value,
2800 then this constant can be pulled out and put in value,
2801 instead of being part of the initialization code. */
2803 if (GET_CODE (v
->new_reg
) == PLUS
2804 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2807 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2809 /* Only succeed if this will give valid addresses.
2810 Try to validate both the first and the last
2811 address resulting from loop unrolling, if
2812 one fails, then can't do const elim here. */
2813 if (verify_addresses (v
, giv_inc
, unroll_number
))
2815 /* Save the negative of the eliminated const, so
2816 that we can calculate the dest_reg's increment
2818 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2820 v
->new_reg
= XEXP (v
->new_reg
, 0);
2821 if (loop_dump_stream
)
2822 fprintf (loop_dump_stream
,
2823 "Eliminating constant from giv %d\n",
2832 /* If the address hasn't been checked for validity yet, do so
2833 now, and fail completely if either the first or the last
2834 unrolled copy of the address is not a valid address
2835 for the instruction that uses it. */
2836 if (v
->dest_reg
== tem
2837 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2839 if (loop_dump_stream
)
2840 fprintf (loop_dump_stream
,
2841 "Invalid address for giv at insn %d\n",
2842 INSN_UID (v
->insn
));
2846 /* To initialize the new register, just move the value of
2847 new_reg into it. This is not guaranteed to give a valid
2848 instruction on machines with complex addressing modes.
2849 If we can't recognize it, then delete it and emit insns
2850 to calculate the value from scratch. */
2851 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2852 copy_rtx (v
->new_reg
)),
2854 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2858 /* We can't use bl->initial_value to compute the initial
2859 value, because the loop may have been preconditioned.
2860 We must calculate it from NEW_REG. Try using
2861 force_operand instead of emit_iv_add_mult. */
2862 delete_insn (PREV_INSN (loop_start
));
2865 ret
= force_operand (v
->new_reg
, tem
);
2867 emit_move_insn (tem
, ret
);
2868 sequence
= gen_sequence ();
2870 emit_insn_before (sequence
, loop_start
);
2872 if (loop_dump_stream
)
2873 fprintf (loop_dump_stream
,
2874 "Invalid init insn, rewritten.\n");
2879 v
->dest_reg
= value
;
2881 /* Check the resulting address for validity, and fail
2882 if the resulting address would be invalid. */
2883 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2885 if (loop_dump_stream
)
2886 fprintf (loop_dump_stream
,
2887 "Invalid address for giv at insn %d\n",
2888 INSN_UID (v
->insn
));
2893 /* Store the value of dest_reg into the insn. This sharing
2894 will not be a problem as this insn will always be copied
2897 *v
->location
= v
->dest_reg
;
2899 /* If this address giv is combined with a dest reg giv, then
2900 save the base giv's induction pointer so that we will be
2901 able to handle this address giv properly. The base giv
2902 itself does not have to be splittable. */
2904 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2905 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2907 if (GET_CODE (v
->new_reg
) == REG
)
2909 /* This giv maybe hasn't been combined with any others.
2910 Make sure that it's giv is marked as splittable here. */
2912 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2914 /* Make it appear to depend upon itself, so that the
2915 giv will be properly split in the main loop above. */
2919 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2923 if (loop_dump_stream
)
2924 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2930 /* Currently, unreduced giv's can't be split. This is not too much
2931 of a problem since unreduced giv's are not live across loop
2932 iterations anyways. When unrolling a loop completely though,
2933 it makes sense to reduce&split givs when possible, as this will
2934 result in simpler instructions, and will not require that a reg
2935 be live across loop iterations. */
2937 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2938 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2939 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2945 /* Givs are only updated once by definition. Mark it so if this is
2946 a splittable register. Don't need to do anything for address givs
2947 where this may not be a register. */
2949 if (GET_CODE (v
->new_reg
) == REG
)
2950 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2954 if (loop_dump_stream
)
2958 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2960 else if (GET_CODE (v
->dest_reg
) != REG
)
2961 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2963 regnum
= REGNO (v
->dest_reg
);
2964 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2965 regnum
, INSN_UID (v
->insn
));
2972 /* Try to prove that the register is dead after the loop exits. Trace every
2973 loop exit looking for an insn that will always be executed, which sets
2974 the register to some value, and appears before the first use of the register
2975 is found. If successful, then return 1, otherwise return 0. */
2977 /* ?? Could be made more intelligent in the handling of jumps, so that
2978 it can search past if statements and other similar structures. */
2981 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2982 rtx reg
, loop_start
, loop_end
;
2987 int label_count
= 0;
2988 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
2990 /* In addition to checking all exits of this loop, we must also check
2991 all exits of inner nested loops that would exit this loop. We don't
2992 have any way to identify those, so we just give up if there are any
2993 such inner loop exits. */
2995 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
2996 label
= LABEL_NEXTREF (label
))
2999 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3002 /* HACK: Must also search the loop fall through exit, create a label_ref
3003 here which points to the loop_end, and append the loop_number_exit_labels
3005 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
3006 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3008 for ( ; label
; label
= LABEL_NEXTREF (label
))
3010 /* Succeed if find an insn which sets the biv or if reach end of
3011 function. Fail if find an insn that uses the biv, or if come to
3012 a conditional jump. */
3014 insn
= NEXT_INSN (XEXP (label
, 0));
3017 code
= GET_CODE (insn
);
3018 if (GET_RTX_CLASS (code
) == 'i')
3022 if (reg_referenced_p (reg
, PATTERN (insn
)))
3025 set
= single_set (insn
);
3026 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3030 if (code
== JUMP_INSN
)
3032 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3034 else if (! simplejump_p (insn
)
3035 /* Prevent infinite loop following infinite loops. */
3036 || jump_count
++ > 20)
3039 insn
= JUMP_LABEL (insn
);
3042 insn
= NEXT_INSN (insn
);
3046 /* Success, the register is dead on all loop exits. */
3050 /* Try to calculate the final value of the biv, the value it will have at
3051 the end of the loop. If we can do it, return that value. */
3054 final_biv_value (bl
, loop_start
, loop_end
)
3055 struct iv_class
*bl
;
3056 rtx loop_start
, loop_end
;
3060 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3062 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3065 /* The final value for reversed bivs must be calculated differently than
3066 for ordinary bivs. In this case, there is already an insn after the
3067 loop which sets this biv's final value (if necessary), and there are
3068 no other loop exits, so we can return any value. */
3071 if (loop_dump_stream
)
3072 fprintf (loop_dump_stream
,
3073 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3078 /* Try to calculate the final value as initial value + (number of iterations
3079 * increment). For this to work, increment must be invariant, the only
3080 exit from the loop must be the fall through at the bottom (otherwise
3081 it may not have its final value when the loop exits), and the initial
3082 value of the biv must be invariant. */
3084 if (loop_n_iterations
!= 0
3085 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3086 && invariant_p (bl
->initial_value
))
3088 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3090 if (increment
&& invariant_p (increment
))
3092 /* Can calculate the loop exit value, emit insns after loop
3093 end to calculate this value into a temporary register in
3094 case it is needed later. */
3096 tem
= gen_reg_rtx (bl
->biv
->mode
);
3097 /* Make sure loop_end is not the last insn. */
3098 if (NEXT_INSN (loop_end
) == 0)
3099 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3100 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3101 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3103 if (loop_dump_stream
)
3104 fprintf (loop_dump_stream
,
3105 "Final biv value for %d, calculated.\n", bl
->regno
);
3111 /* Check to see if the biv is dead at all loop exits. */
3112 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3114 if (loop_dump_stream
)
3115 fprintf (loop_dump_stream
,
3116 "Final biv value for %d, biv dead after loop exit.\n",
3125 /* Try to calculate the final value of the giv, the value it will have at
3126 the end of the loop. If we can do it, return that value. */
3129 final_giv_value (v
, loop_start
, loop_end
)
3130 struct induction
*v
;
3131 rtx loop_start
, loop_end
;
3133 struct iv_class
*bl
;
3136 rtx insert_before
, seq
;
3138 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3140 /* The final value for givs which depend on reversed bivs must be calculated
3141 differently than for ordinary givs. In this case, there is already an
3142 insn after the loop which sets this giv's final value (if necessary),
3143 and there are no other loop exits, so we can return any value. */
3146 if (loop_dump_stream
)
3147 fprintf (loop_dump_stream
,
3148 "Final giv value for %d, depends on reversed biv\n",
3149 REGNO (v
->dest_reg
));
3153 /* Try to calculate the final value as a function of the biv it depends
3154 upon. The only exit from the loop must be the fall through at the bottom
3155 (otherwise it may not have its final value when the loop exits). */
3157 /* ??? Can calculate the final giv value by subtracting off the
3158 extra biv increments times the giv's mult_val. The loop must have
3159 only one exit for this to work, but the loop iterations does not need
3162 if (loop_n_iterations
!= 0
3163 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3165 /* ?? It is tempting to use the biv's value here since these insns will
3166 be put after the loop, and hence the biv will have its final value
3167 then. However, this fails if the biv is subsequently eliminated.
3168 Perhaps determine whether biv's are eliminable before trying to
3169 determine whether giv's are replaceable so that we can use the
3170 biv value here if it is not eliminable. */
3172 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3174 if (increment
&& invariant_p (increment
))
3176 /* Can calculate the loop exit value of its biv as
3177 (loop_n_iterations * increment) + initial_value */
3179 /* The loop exit value of the giv is then
3180 (final_biv_value - extra increments) * mult_val + add_val.
3181 The extra increments are any increments to the biv which
3182 occur in the loop after the giv's value is calculated.
3183 We must search from the insn that sets the giv to the end
3184 of the loop to calculate this value. */
3186 insert_before
= NEXT_INSN (loop_end
);
3188 /* Put the final biv value in tem. */
3189 tem
= gen_reg_rtx (bl
->biv
->mode
);
3190 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3191 bl
->initial_value
, tem
, insert_before
);
3193 /* Subtract off extra increments as we find them. */
3194 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3195 insn
= NEXT_INSN (insn
))
3197 struct induction
*biv
;
3199 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3200 if (biv
->insn
== insn
)
3203 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3204 biv
->add_val
, NULL_RTX
, 0,
3206 seq
= gen_sequence ();
3208 emit_insn_before (seq
, insert_before
);
3212 /* Now calculate the giv's final value. */
3213 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3216 if (loop_dump_stream
)
3217 fprintf (loop_dump_stream
,
3218 "Final giv value for %d, calc from biv's value.\n",
3219 REGNO (v
->dest_reg
));
3225 /* Replaceable giv's should never reach here. */
3229 /* Check to see if the biv is dead at all loop exits. */
3230 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3232 if (loop_dump_stream
)
3233 fprintf (loop_dump_stream
,
3234 "Final giv value for %d, giv dead after loop exit.\n",
3235 REGNO (v
->dest_reg
));
3244 /* Calculate the number of loop iterations. Returns the exact number of loop
3245 iterations if it can be calculated, otherwise returns zero. */
3247 unsigned HOST_WIDE_INT
3248 loop_iterations (loop_start
, loop_end
)
3249 rtx loop_start
, loop_end
;
3251 rtx comparison
, comparison_value
;
3252 rtx iteration_var
, initial_value
, increment
, final_value
;
3253 enum rtx_code comparison_code
;
3256 int unsigned_compare
, compare_dir
, final_larger
;
3257 unsigned long tempu
;
3260 /* First find the iteration variable. If the last insn is a conditional
3261 branch, and the insn before tests a register value, make that the
3262 iteration variable. */
3264 loop_initial_value
= 0;
3266 loop_final_value
= 0;
3267 loop_iteration_var
= 0;
3269 /* We used to use pren_nonnote_insn here, but that fails because it might
3270 accidentally get the branch for a contained loop if the branch for this
3271 loop was deleted. We can only trust branches immediately before the
3273 last_loop_insn
= PREV_INSN (loop_end
);
3275 comparison
= get_condition_for_loop (last_loop_insn
);
3276 if (comparison
== 0)
3278 if (loop_dump_stream
)
3279 fprintf (loop_dump_stream
,
3280 "Loop unrolling: No final conditional branch found.\n");
3284 /* ??? Get_condition may switch position of induction variable and
3285 invariant register when it canonicalizes the comparison. */
3287 comparison_code
= GET_CODE (comparison
);
3288 iteration_var
= XEXP (comparison
, 0);
3289 comparison_value
= XEXP (comparison
, 1);
3291 if (GET_CODE (iteration_var
) != REG
)
3293 if (loop_dump_stream
)
3294 fprintf (loop_dump_stream
,
3295 "Loop unrolling: Comparison not against register.\n");
3299 /* Loop iterations is always called before any new registers are created
3300 now, so this should never occur. */
3302 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3305 iteration_info (iteration_var
, &initial_value
, &increment
,
3306 loop_start
, loop_end
);
3307 if (initial_value
== 0)
3308 /* iteration_info already printed a message. */
3311 /* If the comparison value is an invariant register, then try to find
3312 its value from the insns before the start of the loop. */
3314 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3318 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3320 if (GET_CODE (insn
) == CODE_LABEL
)
3323 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3324 && reg_set_p (comparison_value
, insn
))
3326 /* We found the last insn before the loop that sets the register.
3327 If it sets the entire register, and has a REG_EQUAL note,
3328 then use the value of the REG_EQUAL note. */
3329 if ((set
= single_set (insn
))
3330 && (SET_DEST (set
) == comparison_value
))
3332 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3334 /* Only use the REG_EQUAL note if it is a constant.
3335 Other things, divide in particular, will cause
3336 problems later if we use them. */
3337 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3338 && CONSTANT_P (XEXP (note
, 0)))
3339 comparison_value
= XEXP (note
, 0);
3346 final_value
= approx_final_value (comparison_code
, comparison_value
,
3347 &unsigned_compare
, &compare_dir
);
3349 /* Save the calculated values describing this loop's bounds, in case
3350 precondition_loop_p will need them later. These values can not be
3351 recalculated inside precondition_loop_p because strength reduction
3352 optimizations may obscure the loop's structure. */
3354 loop_iteration_var
= iteration_var
;
3355 loop_initial_value
= initial_value
;
3356 loop_increment
= increment
;
3357 loop_final_value
= final_value
;
3361 if (loop_dump_stream
)
3362 fprintf (loop_dump_stream
,
3363 "Loop unrolling: Increment value can't be calculated.\n");
3366 else if (GET_CODE (increment
) != CONST_INT
)
3368 if (loop_dump_stream
)
3369 fprintf (loop_dump_stream
,
3370 "Loop unrolling: Increment value not constant.\n");
3373 else if (GET_CODE (initial_value
) != CONST_INT
)
3375 if (loop_dump_stream
)
3376 fprintf (loop_dump_stream
,
3377 "Loop unrolling: Initial value not constant.\n");
3380 else if (final_value
== 0)
3382 if (loop_dump_stream
)
3383 fprintf (loop_dump_stream
,
3384 "Loop unrolling: EQ comparison loop.\n");
3387 else if (GET_CODE (final_value
) != CONST_INT
)
3389 if (loop_dump_stream
)
3390 fprintf (loop_dump_stream
,
3391 "Loop unrolling: Final value not constant.\n");
3395 /* ?? Final value and initial value do not have to be constants.
3396 Only their difference has to be constant. When the iteration variable
3397 is an array address, the final value and initial value might both
3398 be addresses with the same base but different constant offsets.
3399 Final value must be invariant for this to work.
3401 To do this, need some way to find the values of registers which are
3404 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3405 if (unsigned_compare
)
3407 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3408 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3409 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3410 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3412 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3413 - (INTVAL (final_value
) < INTVAL (initial_value
));
3415 if (INTVAL (increment
) > 0)
3417 else if (INTVAL (increment
) == 0)
3422 /* There are 27 different cases: compare_dir = -1, 0, 1;
3423 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3424 There are 4 normal cases, 4 reverse cases (where the iteration variable
3425 will overflow before the loop exits), 4 infinite loop cases, and 15
3426 immediate exit (0 or 1 iteration depending on loop type) cases.
3427 Only try to optimize the normal cases. */
3429 /* (compare_dir/final_larger/increment_dir)
3430 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3431 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3432 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3433 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3435 /* ?? If the meaning of reverse loops (where the iteration variable
3436 will overflow before the loop exits) is undefined, then could
3437 eliminate all of these special checks, and just always assume
3438 the loops are normal/immediate/infinite. Note that this means
3439 the sign of increment_dir does not have to be known. Also,
3440 since it does not really hurt if immediate exit loops or infinite loops
3441 are optimized, then that case could be ignored also, and hence all
3442 loops can be optimized.
3444 According to ANSI Spec, the reverse loop case result is undefined,
3445 because the action on overflow is undefined.
3447 See also the special test for NE loops below. */
3449 if (final_larger
== increment_dir
&& final_larger
!= 0
3450 && (final_larger
== compare_dir
|| compare_dir
== 0))
3455 if (loop_dump_stream
)
3456 fprintf (loop_dump_stream
,
3457 "Loop unrolling: Not normal loop.\n");
3461 /* Calculate the number of iterations, final_value is only an approximation,
3462 so correct for that. Note that tempu and loop_n_iterations are
3463 unsigned, because they can be as large as 2^n - 1. */
3465 i
= INTVAL (increment
);
3467 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3470 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3476 /* For NE tests, make sure that the iteration variable won't miss the
3477 final value. If tempu mod i is not zero, then the iteration variable
3478 will overflow before the loop exits, and we can not calculate the
3479 number of iterations. */
3480 if (compare_dir
== 0 && (tempu
% i
) != 0)
3483 return tempu
/ i
+ ((tempu
% i
) != 0);
3486 /* Replace uses of split bivs with their split pseudo register. This is
3487 for original instructions which remain after loop unrolling without
3491 remap_split_bivs (x
)
3494 register enum rtx_code code
;
3501 code
= GET_CODE (x
);
3516 /* If non-reduced/final-value givs were split, then this would also
3517 have to remap those givs also. */
3519 if (REGNO (x
) < max_reg_before_loop
3520 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3521 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3524 fmt
= GET_RTX_FORMAT (code
);
3525 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3528 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3532 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3533 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3539 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3540 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3541 return 0. COPY_START is where we can start looking for the insns
3542 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3545 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3546 must dominate LAST_UID.
3548 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3549 may not dominate LAST_UID.
3551 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3552 must dominate LAST_UID. */
3555 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3562 int passed_jump
= 0;
3563 rtx p
= NEXT_INSN (copy_start
);
3565 while (INSN_UID (p
) != first_uid
)
3567 if (GET_CODE (p
) == JUMP_INSN
)
3569 /* Could not find FIRST_UID. */
3575 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3576 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3577 || ! dead_or_set_regno_p (p
, regno
))
3580 /* FIRST_UID is always executed. */
3581 if (passed_jump
== 0)
3584 while (INSN_UID (p
) != last_uid
)
3586 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3587 can not be sure that FIRST_UID dominates LAST_UID. */
3588 if (GET_CODE (p
) == CODE_LABEL
)
3593 /* FIRST_UID is always executed if LAST_UID is executed. */