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
;
194 static enum rtx_code loop_comparison_code
;
196 /* Forward declarations. */
198 static void init_reg_map
PROTO((struct inline_remap
*, int));
199 static int precondition_loop_p
PROTO((rtx
*, rtx
*, rtx
*, rtx
, rtx
));
200 static rtx calculate_giv_inc
PROTO((rtx
, rtx
, int));
201 static rtx initial_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
202 static void final_reg_note_copy
PROTO((rtx
, struct inline_remap
*));
203 static void copy_loop_body
PROTO((rtx
, rtx
, struct inline_remap
*, rtx
, int,
204 enum unroll_types
, rtx
, rtx
, rtx
, rtx
));
205 static void iteration_info
PROTO((rtx
, rtx
*, rtx
*, rtx
, rtx
));
206 static rtx approx_final_value
PROTO((enum rtx_code
, rtx
, int *, int *));
207 static int find_splittable_regs
PROTO((enum unroll_types
, rtx
, rtx
, rtx
, int));
208 static int find_splittable_givs
PROTO((struct iv_class
*,enum unroll_types
,
209 rtx
, rtx
, rtx
, int));
210 static int reg_dead_after_loop
PROTO((rtx
, rtx
, rtx
));
211 static rtx fold_rtx_mult_add
PROTO((rtx
, rtx
, rtx
, enum machine_mode
));
212 static rtx remap_split_bivs
PROTO((rtx
));
214 /* Try to unroll one loop and split induction variables in the loop.
216 The loop is described by the arguments LOOP_END, INSN_COUNT, and
217 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
218 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
219 indicates whether information generated in the strength reduction pass
222 This function is intended to be called from within `strength_reduce'
226 unroll_loop (loop_end
, insn_count
, loop_start
, end_insert_before
,
231 rtx end_insert_before
;
232 int strength_reduce_p
;
235 int unroll_number
= 1;
236 rtx copy_start
, copy_end
;
237 rtx insn
, copy
, sequence
, pattern
, tem
;
238 int max_labelno
, max_insnno
;
240 struct inline_remap
*map
;
248 int splitting_not_safe
= 0;
249 enum unroll_types unroll_type
;
250 int loop_preconditioned
= 0;
252 /* This points to the last real insn in the loop, which should be either
253 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
257 /* Don't bother unrolling huge loops. Since the minimum factor is
258 two, loops greater than one half of MAX_UNROLLED_INSNS will never
260 if (insn_count
> MAX_UNROLLED_INSNS
/ 2)
262 if (loop_dump_stream
)
263 fprintf (loop_dump_stream
, "Unrolling failure: Loop too big.\n");
267 /* When emitting debugger info, we can't unroll loops with unequal numbers
268 of block_beg and block_end notes, because that would unbalance the block
269 structure of the function. This can happen as a result of the
270 "if (foo) bar; else break;" optimization in jump.c. */
272 if (write_symbols
!= NO_DEBUG
)
274 int block_begins
= 0;
277 for (insn
= loop_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
279 if (GET_CODE (insn
) == NOTE
)
281 if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_BEG
)
283 else if (NOTE_LINE_NUMBER (insn
) == NOTE_INSN_BLOCK_END
)
288 if (block_begins
!= block_ends
)
290 if (loop_dump_stream
)
291 fprintf (loop_dump_stream
,
292 "Unrolling failure: Unbalanced block notes.\n");
297 /* Determine type of unroll to perform. Depends on the number of iterations
298 and the size of the loop. */
300 /* If there is no strength reduce info, then set loop_n_iterations to zero.
301 This can happen if strength_reduce can't find any bivs in the loop.
302 A value of zero indicates that the number of iterations could not be
305 if (! strength_reduce_p
)
306 loop_n_iterations
= 0;
308 if (loop_dump_stream
&& loop_n_iterations
> 0)
309 fprintf (loop_dump_stream
,
310 "Loop unrolling: %d iterations.\n", loop_n_iterations
);
312 /* Find and save a pointer to the last nonnote insn in the loop. */
314 last_loop_insn
= prev_nonnote_insn (loop_end
);
316 /* Calculate how many times to unroll the loop. Indicate whether or
317 not the loop is being completely unrolled. */
319 if (loop_n_iterations
== 1)
321 /* If number of iterations is exactly 1, then eliminate the compare and
322 branch at the end of the loop since they will never be taken.
323 Then return, since no other action is needed here. */
325 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
326 don't do anything. */
328 if (GET_CODE (last_loop_insn
) == BARRIER
)
330 /* Delete the jump insn. This will delete the barrier also. */
331 delete_insn (PREV_INSN (last_loop_insn
));
333 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
336 /* The immediately preceding insn is a compare which must be
338 delete_insn (last_loop_insn
);
339 delete_insn (PREV_INSN (last_loop_insn
));
341 /* The immediately preceding insn may not be the compare, so don't
343 delete_insn (last_loop_insn
);
348 else if (loop_n_iterations
> 0
349 && loop_n_iterations
* insn_count
< MAX_UNROLLED_INSNS
)
351 unroll_number
= loop_n_iterations
;
352 unroll_type
= UNROLL_COMPLETELY
;
354 else if (loop_n_iterations
> 0)
356 /* Try to factor the number of iterations. Don't bother with the
357 general case, only using 2, 3, 5, and 7 will get 75% of all
358 numbers theoretically, and almost all in practice. */
360 for (i
= 0; i
< NUM_FACTORS
; i
++)
361 factors
[i
].count
= 0;
363 temp
= loop_n_iterations
;
364 for (i
= NUM_FACTORS
- 1; i
>= 0; i
--)
365 while (temp
% factors
[i
].factor
== 0)
368 temp
= temp
/ factors
[i
].factor
;
371 /* Start with the larger factors first so that we generally
372 get lots of unrolling. */
376 for (i
= 3; i
>= 0; i
--)
377 while (factors
[i
].count
--)
379 if (temp
* factors
[i
].factor
< MAX_UNROLLED_INSNS
)
381 unroll_number
*= factors
[i
].factor
;
382 temp
*= factors
[i
].factor
;
388 /* If we couldn't find any factors, then unroll as in the normal
390 if (unroll_number
== 1)
392 if (loop_dump_stream
)
393 fprintf (loop_dump_stream
,
394 "Loop unrolling: No factors found.\n");
397 unroll_type
= UNROLL_MODULO
;
401 /* Default case, calculate number of times to unroll loop based on its
403 if (unroll_number
== 1)
405 if (8 * insn_count
< MAX_UNROLLED_INSNS
)
407 else if (4 * insn_count
< MAX_UNROLLED_INSNS
)
412 unroll_type
= UNROLL_NAIVE
;
415 /* Now we know how many times to unroll the loop. */
417 if (loop_dump_stream
)
418 fprintf (loop_dump_stream
,
419 "Unrolling loop %d times.\n", unroll_number
);
422 if (unroll_type
== UNROLL_COMPLETELY
|| unroll_type
== UNROLL_MODULO
)
424 /* Loops of these types should never start with a jump down to
425 the exit condition test. For now, check for this case just to
426 be sure. UNROLL_NAIVE loops can be of this form, this case is
429 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
430 insn
= NEXT_INSN (insn
);
431 if (GET_CODE (insn
) == JUMP_INSN
)
435 if (unroll_type
== UNROLL_COMPLETELY
)
437 /* Completely unrolling the loop: Delete the compare and branch at
438 the end (the last two instructions). This delete must done at the
439 very end of loop unrolling, to avoid problems with calls to
440 back_branch_in_range_p, which is called by find_splittable_regs.
441 All increments of splittable bivs/givs are changed to load constant
444 copy_start
= loop_start
;
446 /* Set insert_before to the instruction immediately after the JUMP_INSN
447 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
448 the loop will be correctly handled by copy_loop_body. */
449 insert_before
= NEXT_INSN (last_loop_insn
);
451 /* Set copy_end to the insn before the jump at the end of the loop. */
452 if (GET_CODE (last_loop_insn
) == BARRIER
)
453 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
454 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
457 /* The instruction immediately before the JUMP_INSN is a compare
458 instruction which we do not want to copy. */
459 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
461 /* The instruction immediately before the JUMP_INSN may not be the
462 compare, so we must copy it. */
463 copy_end
= PREV_INSN (last_loop_insn
);
468 /* We currently can't unroll a loop if it doesn't end with a
469 JUMP_INSN. There would need to be a mechanism that recognizes
470 this case, and then inserts a jump after each loop body, which
471 jumps to after the last loop body. */
472 if (loop_dump_stream
)
473 fprintf (loop_dump_stream
,
474 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
478 else if (unroll_type
== UNROLL_MODULO
)
480 /* Partially unrolling the loop: The compare and branch at the end
481 (the last two instructions) must remain. Don't copy the compare
482 and branch instructions at the end of the loop. Insert the unrolled
483 code immediately before the compare/branch at the end so that the
484 code will fall through to them as before. */
486 copy_start
= loop_start
;
488 /* Set insert_before to the jump insn at the end of the loop.
489 Set copy_end to before the jump insn at the end of the loop. */
490 if (GET_CODE (last_loop_insn
) == BARRIER
)
492 insert_before
= PREV_INSN (last_loop_insn
);
493 copy_end
= PREV_INSN (insert_before
);
495 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
498 /* The instruction immediately before the JUMP_INSN is a compare
499 instruction which we do not want to copy or delete. */
500 insert_before
= PREV_INSN (last_loop_insn
);
501 copy_end
= PREV_INSN (insert_before
);
503 /* The instruction immediately before the JUMP_INSN may not be the
504 compare, so we must copy it. */
505 insert_before
= last_loop_insn
;
506 copy_end
= PREV_INSN (last_loop_insn
);
511 /* We currently can't unroll a loop if it doesn't end with a
512 JUMP_INSN. There would need to be a mechanism that recognizes
513 this case, and then inserts a jump after each loop body, which
514 jumps to after the last loop body. */
515 if (loop_dump_stream
)
516 fprintf (loop_dump_stream
,
517 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
523 /* Normal case: Must copy the compare and branch instructions at the
526 if (GET_CODE (last_loop_insn
) == BARRIER
)
528 /* Loop ends with an unconditional jump and a barrier.
529 Handle this like above, don't copy jump and barrier.
530 This is not strictly necessary, but doing so prevents generating
531 unconditional jumps to an immediately following label.
533 This will be corrected below if the target of this jump is
534 not the start_label. */
536 insert_before
= PREV_INSN (last_loop_insn
);
537 copy_end
= PREV_INSN (insert_before
);
539 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
541 /* Set insert_before to immediately after the JUMP_INSN, so that
542 NOTEs at the end of the loop will be correctly handled by
544 insert_before
= NEXT_INSN (last_loop_insn
);
545 copy_end
= last_loop_insn
;
549 /* We currently can't unroll a loop if it doesn't end with a
550 JUMP_INSN. There would need to be a mechanism that recognizes
551 this case, and then inserts a jump after each loop body, which
552 jumps to after the last loop body. */
553 if (loop_dump_stream
)
554 fprintf (loop_dump_stream
,
555 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
559 /* If copying exit test branches because they can not be eliminated,
560 then must convert the fall through case of the branch to a jump past
561 the end of the loop. Create a label to emit after the loop and save
562 it for later use. Do not use the label after the loop, if any, since
563 it might be used by insns outside the loop, or there might be insns
564 added before it later by final_[bg]iv_value which must be after
565 the real exit label. */
566 exit_label
= gen_label_rtx ();
569 while (GET_CODE (insn
) != CODE_LABEL
&& GET_CODE (insn
) != JUMP_INSN
)
570 insn
= NEXT_INSN (insn
);
572 if (GET_CODE (insn
) == JUMP_INSN
)
574 /* The loop starts with a jump down to the exit condition test.
575 Start copying the loop after the barrier following this
577 copy_start
= NEXT_INSN (insn
);
579 /* Splitting induction variables doesn't work when the loop is
580 entered via a jump to the bottom, because then we end up doing
581 a comparison against a new register for a split variable, but
582 we did not execute the set insn for the new register because
583 it was skipped over. */
584 splitting_not_safe
= 1;
585 if (loop_dump_stream
)
586 fprintf (loop_dump_stream
,
587 "Splitting not safe, because loop not entered at top.\n");
590 copy_start
= loop_start
;
593 /* This should always be the first label in the loop. */
594 start_label
= NEXT_INSN (copy_start
);
595 /* There may be a line number note and/or a loop continue note here. */
596 while (GET_CODE (start_label
) == NOTE
)
597 start_label
= NEXT_INSN (start_label
);
598 if (GET_CODE (start_label
) != CODE_LABEL
)
600 /* This can happen as a result of jump threading. If the first insns in
601 the loop test the same condition as the loop's backward jump, or the
602 opposite condition, then the backward jump will be modified to point
603 to elsewhere, and the loop's start label is deleted.
605 This case currently can not be handled by the loop unrolling code. */
607 if (loop_dump_stream
)
608 fprintf (loop_dump_stream
,
609 "Unrolling failure: unknown insns between BEG note and loop label.\n");
612 if (LABEL_NAME (start_label
))
614 /* The jump optimization pass must have combined the original start label
615 with a named label for a goto. We can't unroll this case because
616 jumps which go to the named label must be handled differently than
617 jumps to the loop start, and it is impossible to differentiate them
619 if (loop_dump_stream
)
620 fprintf (loop_dump_stream
,
621 "Unrolling failure: loop start label is gone\n");
625 if (unroll_type
== UNROLL_NAIVE
626 && GET_CODE (last_loop_insn
) == BARRIER
627 && start_label
!= JUMP_LABEL (PREV_INSN (last_loop_insn
)))
629 /* In this case, we must copy the jump and barrier, because they will
630 not be converted to jumps to an immediately following label. */
632 insert_before
= NEXT_INSN (last_loop_insn
);
633 copy_end
= last_loop_insn
;
636 /* Allocate a translation table for the labels and insn numbers.
637 They will be filled in as we copy the insns in the loop. */
639 max_labelno
= max_label_num ();
640 max_insnno
= get_max_uid ();
642 map
= (struct inline_remap
*) alloca (sizeof (struct inline_remap
));
644 map
->integrating
= 0;
646 /* Allocate the label map. */
650 map
->label_map
= (rtx
*) alloca (max_labelno
* sizeof (rtx
));
652 local_label
= (char *) alloca (max_labelno
);
653 bzero (local_label
, max_labelno
);
658 /* Search the loop and mark all local labels, i.e. the ones which have to
659 be distinct labels when copied. For all labels which might be
660 non-local, set their label_map entries to point to themselves.
661 If they happen to be local their label_map entries will be overwritten
662 before the loop body is copied. The label_map entries for local labels
663 will be set to a different value each time the loop body is copied. */
665 for (insn
= copy_start
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
667 if (GET_CODE (insn
) == CODE_LABEL
)
668 local_label
[CODE_LABEL_NUMBER (insn
)] = 1;
669 else if (GET_CODE (insn
) == JUMP_INSN
)
671 if (JUMP_LABEL (insn
))
672 map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))]
674 else if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
675 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
677 rtx pat
= PATTERN (insn
);
678 int diff_vec_p
= GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
;
679 int len
= XVECLEN (pat
, diff_vec_p
);
682 for (i
= 0; i
< len
; i
++)
684 label
= XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0);
685 map
->label_map
[CODE_LABEL_NUMBER (label
)] = label
;
691 /* Allocate space for the insn map. */
693 map
->insn_map
= (rtx
*) alloca (max_insnno
* sizeof (rtx
));
695 /* Set this to zero, to indicate that we are doing loop unrolling,
696 not function inlining. */
697 map
->inline_target
= 0;
699 /* The register and constant maps depend on the number of registers
700 present, so the final maps can't be created until after
701 find_splittable_regs is called. However, they are needed for
702 preconditioning, so we create temporary maps when preconditioning
705 /* The preconditioning code may allocate two new pseudo registers. */
706 maxregnum
= max_reg_num ();
708 /* Allocate and zero out the splittable_regs and addr_combined_regs
709 arrays. These must be zeroed here because they will be used if
710 loop preconditioning is performed, and must be zero for that case.
712 It is safe to do this here, since the extra registers created by the
713 preconditioning code and find_splittable_regs will never be used
714 to access the splittable_regs[] and addr_combined_regs[] arrays. */
716 splittable_regs
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
717 bzero ((char *) splittable_regs
, maxregnum
* sizeof (rtx
));
718 splittable_regs_updates
= (int *) alloca (maxregnum
* sizeof (int));
719 bzero ((char *) splittable_regs_updates
, maxregnum
* sizeof (int));
721 = (struct induction
**) alloca (maxregnum
* sizeof (struct induction
*));
722 bzero ((char *) addr_combined_regs
, maxregnum
* sizeof (struct induction
*));
723 /* We must limit it to max_reg_before_loop, because only these pseudo
724 registers have valid regno_first_uid info. Any register created after
725 that is unlikely to be local to the loop anyways. */
726 local_regno
= (char *) alloca (max_reg_before_loop
);
727 bzero (local_regno
, max_reg_before_loop
);
729 /* Mark all local registers, i.e. the ones which are referenced only
731 if (INSN_UID (copy_end
) < max_uid_for_loop
)
733 int copy_start_luid
= INSN_LUID (copy_start
);
734 int copy_end_luid
= INSN_LUID (copy_end
);
736 /* If a register is used in the jump insn, we must not duplicate it
737 since it will also be used outside the loop. */
738 if (GET_CODE (copy_end
) == JUMP_INSN
)
740 /* If copy_start points to the NOTE that starts the loop, then we must
741 use the next luid, because invariant pseudo-regs moved out of the loop
742 have their lifetimes modified to start here, but they are not safe
744 if (copy_start
== loop_start
)
747 /* If a pseudo's lifetime is entirely contained within this loop, then we
748 can use a different pseudo in each unrolled copy of the loop. This
749 results in better code. */
750 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; ++j
)
751 if (regno_first_uid
[j
] > 0 && regno_first_uid
[j
] <= max_uid_for_loop
752 && uid_luid
[regno_first_uid
[j
]] >= copy_start_luid
753 && regno_last_uid
[j
] > 0 && regno_last_uid
[j
] <= max_uid_for_loop
754 && uid_luid
[regno_last_uid
[j
]] <= copy_end_luid
)
756 /* However, we must also check for loop-carried dependencies.
757 If the value the pseudo has at the end of iteration X is
758 used by iteration X+1, then we can not use a different pseudo
759 for each unrolled copy of the loop. */
760 /* A pseudo is safe if regno_first_uid is a set, and this
761 set dominates all instructions from regno_first_uid to
763 /* ??? This check is simplistic. We would get better code if
764 this check was more sophisticated. */
765 if (set_dominates_use (j
, regno_first_uid
[j
], regno_last_uid
[j
],
766 copy_start
, copy_end
))
769 if (loop_dump_stream
)
772 fprintf (loop_dump_stream
, "Marked reg %d as local\n", j
);
774 fprintf (loop_dump_stream
, "Did not mark reg %d as local\n",
780 /* If this loop requires exit tests when unrolled, check to see if we
781 can precondition the loop so as to make the exit tests unnecessary.
782 Just like variable splitting, this is not safe if the loop is entered
783 via a jump to the bottom. Also, can not do this if no strength
784 reduce info, because precondition_loop_p uses this info. */
786 /* Must copy the loop body for preconditioning before the following
787 find_splittable_regs call since that will emit insns which need to
788 be after the preconditioned loop copies, but immediately before the
789 unrolled loop copies. */
791 /* Also, it is not safe to split induction variables for the preconditioned
792 copies of the loop body. If we split induction variables, then the code
793 assumes that each induction variable can be represented as a function
794 of its initial value and the loop iteration number. This is not true
795 in this case, because the last preconditioned copy of the loop body
796 could be any iteration from the first up to the `unroll_number-1'th,
797 depending on the initial value of the iteration variable. Therefore
798 we can not split induction variables here, because we can not calculate
799 their value. Hence, this code must occur before find_splittable_regs
802 if (unroll_type
== UNROLL_NAIVE
&& ! splitting_not_safe
&& strength_reduce_p
)
804 rtx initial_value
, final_value
, increment
;
806 if (precondition_loop_p (&initial_value
, &final_value
, &increment
,
807 loop_start
, loop_end
))
809 register rtx diff
, temp
;
810 enum machine_mode mode
;
812 int abs_inc
, neg_inc
;
814 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
816 map
->const_equiv_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
817 map
->const_age_map
= (unsigned *) alloca (maxregnum
818 * sizeof (unsigned));
819 map
->const_equiv_map_size
= maxregnum
;
820 global_const_equiv_map
= map
->const_equiv_map
;
821 global_const_equiv_map_size
= maxregnum
;
823 init_reg_map (map
, maxregnum
);
825 /* Limit loop unrolling to 4, since this will make 7 copies of
827 if (unroll_number
> 4)
830 /* Save the absolute value of the increment, and also whether or
831 not it is negative. */
833 abs_inc
= INTVAL (increment
);
842 /* Decide what mode to do these calculations in. Choose the larger
843 of final_value's mode and initial_value's mode, or a full-word if
844 both are constants. */
845 mode
= GET_MODE (final_value
);
846 if (mode
== VOIDmode
)
848 mode
= GET_MODE (initial_value
);
849 if (mode
== VOIDmode
)
852 else if (mode
!= GET_MODE (initial_value
)
853 && (GET_MODE_SIZE (mode
)
854 < GET_MODE_SIZE (GET_MODE (initial_value
))))
855 mode
= GET_MODE (initial_value
);
857 /* Calculate the difference between the final and initial values.
858 Final value may be a (plus (reg x) (const_int 1)) rtx.
859 Let the following cse pass simplify this if initial value is
862 We must copy the final and initial values here to avoid
863 improperly shared rtl. */
865 diff
= expand_binop (mode
, sub_optab
, copy_rtx (final_value
),
866 copy_rtx (initial_value
), NULL_RTX
, 0,
869 /* Now calculate (diff % (unroll * abs (increment))) by using an
871 diff
= expand_binop (GET_MODE (diff
), and_optab
, diff
,
872 GEN_INT (unroll_number
* abs_inc
- 1),
873 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
875 /* Now emit a sequence of branches to jump to the proper precond
878 labels
= (rtx
*) alloca (sizeof (rtx
) * unroll_number
);
879 for (i
= 0; i
< unroll_number
; i
++)
880 labels
[i
] = gen_label_rtx ();
882 /* Check for the case where the initial value is greater than or
883 equal to the final value. In that case, we want to execute
884 exactly one loop iteration. The code below will fail for this
885 case. This check does not apply if the loop has a NE
886 comparison at the end. */
888 if (loop_comparison_code
!= NE
)
890 emit_cmp_insn (initial_value
, final_value
, neg_inc
? LE
: GE
,
891 NULL_RTX
, mode
, 0, 0);
893 emit_jump_insn (gen_ble (labels
[1]));
895 emit_jump_insn (gen_bge (labels
[1]));
896 JUMP_LABEL (get_last_insn ()) = labels
[1];
897 LABEL_NUSES (labels
[1])++;
900 /* Assuming the unroll_number is 4, and the increment is 2, then
901 for a negative increment: for a positive increment:
902 diff = 0,1 precond 0 diff = 0,7 precond 0
903 diff = 2,3 precond 3 diff = 1,2 precond 1
904 diff = 4,5 precond 2 diff = 3,4 precond 2
905 diff = 6,7 precond 1 diff = 5,6 precond 3 */
907 /* We only need to emit (unroll_number - 1) branches here, the
908 last case just falls through to the following code. */
910 /* ??? This would give better code if we emitted a tree of branches
911 instead of the current linear list of branches. */
913 for (i
= 0; i
< unroll_number
- 1; i
++)
916 enum rtx_code cmp_code
;
918 /* For negative increments, must invert the constant compared
919 against, except when comparing against zero. */
927 cmp_const
= unroll_number
- i
;
936 emit_cmp_insn (diff
, GEN_INT (abs_inc
* cmp_const
),
937 cmp_code
, NULL_RTX
, mode
, 0, 0);
940 emit_jump_insn (gen_beq (labels
[i
]));
942 emit_jump_insn (gen_bge (labels
[i
]));
944 emit_jump_insn (gen_ble (labels
[i
]));
945 JUMP_LABEL (get_last_insn ()) = labels
[i
];
946 LABEL_NUSES (labels
[i
])++;
949 /* If the increment is greater than one, then we need another branch,
950 to handle other cases equivalent to 0. */
952 /* ??? This should be merged into the code above somehow to help
953 simplify the code here, and reduce the number of branches emitted.
954 For the negative increment case, the branch here could easily
955 be merged with the `0' case branch above. For the positive
956 increment case, it is not clear how this can be simplified. */
961 enum rtx_code cmp_code
;
965 cmp_const
= abs_inc
- 1;
970 cmp_const
= abs_inc
* (unroll_number
- 1) + 1;
974 emit_cmp_insn (diff
, GEN_INT (cmp_const
), cmp_code
, NULL_RTX
,
978 emit_jump_insn (gen_ble (labels
[0]));
980 emit_jump_insn (gen_bge (labels
[0]));
981 JUMP_LABEL (get_last_insn ()) = labels
[0];
982 LABEL_NUSES (labels
[0])++;
985 sequence
= gen_sequence ();
987 emit_insn_before (sequence
, loop_start
);
989 /* Only the last copy of the loop body here needs the exit
990 test, so set copy_end to exclude the compare/branch here,
991 and then reset it inside the loop when get to the last
994 if (GET_CODE (last_loop_insn
) == BARRIER
)
995 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
996 else if (GET_CODE (last_loop_insn
) == JUMP_INSN
)
999 /* The immediately preceding insn is a compare which we do not
1001 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1003 /* The immediately preceding insn may not be a compare, so we
1005 copy_end
= PREV_INSN (last_loop_insn
);
1011 for (i
= 1; i
< unroll_number
; i
++)
1013 emit_label_after (labels
[unroll_number
- i
],
1014 PREV_INSN (loop_start
));
1016 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1017 bzero ((char *) map
->const_equiv_map
, maxregnum
* sizeof (rtx
));
1018 bzero ((char *) map
->const_age_map
,
1019 maxregnum
* sizeof (unsigned));
1022 for (j
= 0; j
< max_labelno
; j
++)
1024 map
->label_map
[j
] = gen_label_rtx ();
1026 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1028 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1030 /* The last copy needs the compare/branch insns at the end,
1031 so reset copy_end here if the loop ends with a conditional
1034 if (i
== unroll_number
- 1)
1036 if (GET_CODE (last_loop_insn
) == BARRIER
)
1037 copy_end
= PREV_INSN (PREV_INSN (last_loop_insn
));
1039 copy_end
= last_loop_insn
;
1042 /* None of the copies are the `last_iteration', so just
1043 pass zero for that parameter. */
1044 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, 0,
1045 unroll_type
, start_label
, loop_end
,
1046 loop_start
, copy_end
);
1048 emit_label_after (labels
[0], PREV_INSN (loop_start
));
1050 if (GET_CODE (last_loop_insn
) == BARRIER
)
1052 insert_before
= PREV_INSN (last_loop_insn
);
1053 copy_end
= PREV_INSN (insert_before
);
1058 /* The immediately preceding insn is a compare which we do not
1060 insert_before
= PREV_INSN (last_loop_insn
);
1061 copy_end
= PREV_INSN (insert_before
);
1063 /* The immediately preceding insn may not be a compare, so we
1065 insert_before
= last_loop_insn
;
1066 copy_end
= PREV_INSN (last_loop_insn
);
1070 /* Set unroll type to MODULO now. */
1071 unroll_type
= UNROLL_MODULO
;
1072 loop_preconditioned
= 1;
1076 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1077 the loop unless all loops are being unrolled. */
1078 if (unroll_type
== UNROLL_NAIVE
&& ! flag_unroll_all_loops
)
1080 if (loop_dump_stream
)
1081 fprintf (loop_dump_stream
, "Unrolling failure: Naive unrolling not being done.\n");
1085 /* At this point, we are guaranteed to unroll the loop. */
1087 /* For each biv and giv, determine whether it can be safely split into
1088 a different variable for each unrolled copy of the loop body.
1089 We precalculate and save this info here, since computing it is
1092 Do this before deleting any instructions from the loop, so that
1093 back_branch_in_range_p will work correctly. */
1095 if (splitting_not_safe
)
1098 temp
= find_splittable_regs (unroll_type
, loop_start
, loop_end
,
1099 end_insert_before
, unroll_number
);
1101 /* find_splittable_regs may have created some new registers, so must
1102 reallocate the reg_map with the new larger size, and must realloc
1103 the constant maps also. */
1105 maxregnum
= max_reg_num ();
1106 map
->reg_map
= (rtx
*) alloca (maxregnum
* sizeof (rtx
));
1108 init_reg_map (map
, maxregnum
);
1110 /* Space is needed in some of the map for new registers, so new_maxregnum
1111 is an (over)estimate of how many registers will exist at the end. */
1112 new_maxregnum
= maxregnum
+ (temp
* unroll_number
* 2);
1114 /* Must realloc space for the constant maps, because the number of registers
1115 may have changed. */
1117 map
->const_equiv_map
= (rtx
*) alloca (new_maxregnum
* sizeof (rtx
));
1118 map
->const_age_map
= (unsigned *) alloca (new_maxregnum
* sizeof (unsigned));
1120 map
->const_equiv_map_size
= new_maxregnum
;
1121 global_const_equiv_map
= map
->const_equiv_map
;
1122 global_const_equiv_map_size
= new_maxregnum
;
1124 /* Search the list of bivs and givs to find ones which need to be remapped
1125 when split, and set their reg_map entry appropriately. */
1127 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
1129 if (REGNO (bl
->biv
->src_reg
) != bl
->regno
)
1130 map
->reg_map
[bl
->regno
] = bl
->biv
->src_reg
;
1132 /* Currently, non-reduced/final-value givs are never split. */
1133 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
1134 if (REGNO (v
->src_reg
) != bl
->regno
)
1135 map
->reg_map
[REGNO (v
->dest_reg
)] = v
->src_reg
;
1139 /* Use our current register alignment and pointer flags. */
1140 map
->regno_pointer_flag
= regno_pointer_flag
;
1141 map
->regno_pointer_align
= regno_pointer_align
;
1143 /* If the loop is being partially unrolled, and the iteration variables
1144 are being split, and are being renamed for the split, then must fix up
1145 the compare/jump instruction at the end of the loop to refer to the new
1146 registers. This compare isn't copied, so the registers used in it
1147 will never be replaced if it isn't done here. */
1149 if (unroll_type
== UNROLL_MODULO
)
1151 insn
= NEXT_INSN (copy_end
);
1152 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
)
1153 PATTERN (insn
) = remap_split_bivs (PATTERN (insn
));
1156 /* For unroll_number - 1 times, make a copy of each instruction
1157 between copy_start and copy_end, and insert these new instructions
1158 before the end of the loop. */
1160 for (i
= 0; i
< unroll_number
; i
++)
1162 bzero ((char *) map
->insn_map
, max_insnno
* sizeof (rtx
));
1163 bzero ((char *) map
->const_equiv_map
, new_maxregnum
* sizeof (rtx
));
1164 bzero ((char *) map
->const_age_map
, new_maxregnum
* sizeof (unsigned));
1167 for (j
= 0; j
< max_labelno
; j
++)
1169 map
->label_map
[j
] = gen_label_rtx ();
1171 for (j
= FIRST_PSEUDO_REGISTER
; j
< max_reg_before_loop
; j
++)
1173 map
->reg_map
[j
] = gen_reg_rtx (GET_MODE (regno_reg_rtx
[j
]));
1175 /* If loop starts with a branch to the test, then fix it so that
1176 it points to the test of the first unrolled copy of the loop. */
1177 if (i
== 0 && loop_start
!= copy_start
)
1179 insn
= PREV_INSN (copy_start
);
1180 pattern
= PATTERN (insn
);
1182 tem
= map
->label_map
[CODE_LABEL_NUMBER
1183 (XEXP (SET_SRC (pattern
), 0))];
1184 SET_SRC (pattern
) = gen_rtx (LABEL_REF
, VOIDmode
, tem
);
1186 /* Set the jump label so that it can be used by later loop unrolling
1188 JUMP_LABEL (insn
) = tem
;
1189 LABEL_NUSES (tem
)++;
1192 copy_loop_body (copy_start
, copy_end
, map
, exit_label
,
1193 i
== unroll_number
- 1, unroll_type
, start_label
,
1194 loop_end
, insert_before
, insert_before
);
1197 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1198 insn to be deleted. This prevents any runaway delete_insn call from
1199 more insns that it should, as it always stops at a CODE_LABEL. */
1201 /* Delete the compare and branch at the end of the loop if completely
1202 unrolling the loop. Deleting the backward branch at the end also
1203 deletes the code label at the start of the loop. This is done at
1204 the very end to avoid problems with back_branch_in_range_p. */
1206 if (unroll_type
== UNROLL_COMPLETELY
)
1207 safety_label
= emit_label_after (gen_label_rtx (), last_loop_insn
);
1209 safety_label
= emit_label_after (gen_label_rtx (), copy_end
);
1211 /* Delete all of the original loop instructions. Don't delete the
1212 LOOP_BEG note, or the first code label in the loop. */
1214 insn
= NEXT_INSN (copy_start
);
1215 while (insn
!= safety_label
)
1217 if (insn
!= start_label
)
1218 insn
= delete_insn (insn
);
1220 insn
= NEXT_INSN (insn
);
1223 /* Can now delete the 'safety' label emitted to protect us from runaway
1224 delete_insn calls. */
1225 if (INSN_DELETED_P (safety_label
))
1227 delete_insn (safety_label
);
1229 /* If exit_label exists, emit it after the loop. Doing the emit here
1230 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1231 This is needed so that mostly_true_jump in reorg.c will treat jumps
1232 to this loop end label correctly, i.e. predict that they are usually
1235 emit_label_after (exit_label
, loop_end
);
1238 /* Return true if the loop can be safely, and profitably, preconditioned
1239 so that the unrolled copies of the loop body don't need exit tests.
1241 This only works if final_value, initial_value and increment can be
1242 determined, and if increment is a constant power of 2.
1243 If increment is not a power of 2, then the preconditioning modulo
1244 operation would require a real modulo instead of a boolean AND, and this
1245 is not considered `profitable'. */
1247 /* ??? If the loop is known to be executed very many times, or the machine
1248 has a very cheap divide instruction, then preconditioning is a win even
1249 when the increment is not a power of 2. Use RTX_COST to compute
1250 whether divide is cheap. */
1253 precondition_loop_p (initial_value
, final_value
, increment
, loop_start
,
1255 rtx
*initial_value
, *final_value
, *increment
;
1256 rtx loop_start
, loop_end
;
1259 if (loop_n_iterations
> 0)
1261 *initial_value
= const0_rtx
;
1262 *increment
= const1_rtx
;
1263 *final_value
= GEN_INT (loop_n_iterations
);
1265 if (loop_dump_stream
)
1266 fprintf (loop_dump_stream
,
1267 "Preconditioning: Success, number of iterations known, %d.\n",
1272 if (loop_initial_value
== 0)
1274 if (loop_dump_stream
)
1275 fprintf (loop_dump_stream
,
1276 "Preconditioning: Could not find initial value.\n");
1279 else if (loop_increment
== 0)
1281 if (loop_dump_stream
)
1282 fprintf (loop_dump_stream
,
1283 "Preconditioning: Could not find increment value.\n");
1286 else if (GET_CODE (loop_increment
) != CONST_INT
)
1288 if (loop_dump_stream
)
1289 fprintf (loop_dump_stream
,
1290 "Preconditioning: Increment not a constant.\n");
1293 else if ((exact_log2 (INTVAL (loop_increment
)) < 0)
1294 && (exact_log2 (- INTVAL (loop_increment
)) < 0))
1296 if (loop_dump_stream
)
1297 fprintf (loop_dump_stream
,
1298 "Preconditioning: Increment not a constant power of 2.\n");
1302 /* Unsigned_compare and compare_dir can be ignored here, since they do
1303 not matter for preconditioning. */
1305 if (loop_final_value
== 0)
1307 if (loop_dump_stream
)
1308 fprintf (loop_dump_stream
,
1309 "Preconditioning: EQ comparison loop.\n");
1313 /* Must ensure that final_value is invariant, so call invariant_p to
1314 check. Before doing so, must check regno against max_reg_before_loop
1315 to make sure that the register is in the range covered by invariant_p.
1316 If it isn't, then it is most likely a biv/giv which by definition are
1318 if ((GET_CODE (loop_final_value
) == REG
1319 && REGNO (loop_final_value
) >= max_reg_before_loop
)
1320 || (GET_CODE (loop_final_value
) == PLUS
1321 && REGNO (XEXP (loop_final_value
, 0)) >= max_reg_before_loop
)
1322 || ! invariant_p (loop_final_value
))
1324 if (loop_dump_stream
)
1325 fprintf (loop_dump_stream
,
1326 "Preconditioning: Final value not invariant.\n");
1330 /* Fail for floating point values, since the caller of this function
1331 does not have code to deal with them. */
1332 if (GET_MODE_CLASS (GET_MODE (loop_final_value
)) == MODE_FLOAT
1333 || GET_MODE_CLASS (GET_MODE (loop_initial_value
)) == MODE_FLOAT
)
1335 if (loop_dump_stream
)
1336 fprintf (loop_dump_stream
,
1337 "Preconditioning: Floating point final or initial value.\n");
1341 /* Now set initial_value to be the iteration_var, since that may be a
1342 simpler expression, and is guaranteed to be correct if all of the
1343 above tests succeed.
1345 We can not use the initial_value as calculated, because it will be
1346 one too small for loops of the form "while (i-- > 0)". We can not
1347 emit code before the loop_skip_over insns to fix this problem as this
1348 will then give a number one too large for loops of the form
1351 Note that all loops that reach here are entered at the top, because
1352 this function is not called if the loop starts with a jump. */
1354 /* Fail if loop_iteration_var is not live before loop_start, since we need
1355 to test its value in the preconditioning code. */
1357 if (uid_luid
[regno_first_uid
[REGNO (loop_iteration_var
)]]
1358 > INSN_LUID (loop_start
))
1360 if (loop_dump_stream
)
1361 fprintf (loop_dump_stream
,
1362 "Preconditioning: Iteration var not live before loop start.\n");
1366 *initial_value
= loop_iteration_var
;
1367 *increment
= loop_increment
;
1368 *final_value
= loop_final_value
;
1371 if (loop_dump_stream
)
1372 fprintf (loop_dump_stream
, "Preconditioning: Successful.\n");
1377 /* All pseudo-registers must be mapped to themselves. Two hard registers
1378 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1379 REGNUM, to avoid function-inlining specific conversions of these
1380 registers. All other hard regs can not be mapped because they may be
1385 init_reg_map (map
, maxregnum
)
1386 struct inline_remap
*map
;
1391 for (i
= maxregnum
- 1; i
> LAST_VIRTUAL_REGISTER
; i
--)
1392 map
->reg_map
[i
] = regno_reg_rtx
[i
];
1393 /* Just clear the rest of the entries. */
1394 for (i
= LAST_VIRTUAL_REGISTER
; i
>= 0; i
--)
1395 map
->reg_map
[i
] = 0;
1397 map
->reg_map
[VIRTUAL_STACK_VARS_REGNUM
]
1398 = regno_reg_rtx
[VIRTUAL_STACK_VARS_REGNUM
];
1399 map
->reg_map
[VIRTUAL_INCOMING_ARGS_REGNUM
]
1400 = regno_reg_rtx
[VIRTUAL_INCOMING_ARGS_REGNUM
];
1403 /* Strength-reduction will often emit code for optimized biv/givs which
1404 calculates their value in a temporary register, and then copies the result
1405 to the iv. This procedure reconstructs the pattern computing the iv;
1406 verifying that all operands are of the proper form.
1408 The return value is the amount that the giv is incremented by. */
1411 calculate_giv_inc (pattern
, src_insn
, regno
)
1412 rtx pattern
, src_insn
;
1416 rtx increment_total
= 0;
1420 /* Verify that we have an increment insn here. First check for a plus
1421 as the set source. */
1422 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1424 /* SR sometimes computes the new giv value in a temp, then copies it
1426 src_insn
= PREV_INSN (src_insn
);
1427 pattern
= PATTERN (src_insn
);
1428 if (GET_CODE (SET_SRC (pattern
)) != PLUS
)
1431 /* The last insn emitted is not needed, so delete it to avoid confusing
1432 the second cse pass. This insn sets the giv unnecessarily. */
1433 delete_insn (get_last_insn ());
1436 /* Verify that we have a constant as the second operand of the plus. */
1437 increment
= XEXP (SET_SRC (pattern
), 1);
1438 if (GET_CODE (increment
) != CONST_INT
)
1440 /* SR sometimes puts the constant in a register, especially if it is
1441 too big to be an add immed operand. */
1442 src_insn
= PREV_INSN (src_insn
);
1443 increment
= SET_SRC (PATTERN (src_insn
));
1445 /* SR may have used LO_SUM to compute the constant if it is too large
1446 for a load immed operand. In this case, the constant is in operand
1447 one of the LO_SUM rtx. */
1448 if (GET_CODE (increment
) == LO_SUM
)
1449 increment
= XEXP (increment
, 1);
1450 else if (GET_CODE (increment
) == IOR
1451 || GET_CODE (increment
) == ASHIFT
)
1453 /* The rs6000 port loads some constants with IOR.
1454 The alpha port loads some constants with ASHIFT. */
1455 rtx second_part
= XEXP (increment
, 1);
1456 enum rtx_code code
= GET_CODE (increment
);
1458 src_insn
= PREV_INSN (src_insn
);
1459 increment
= SET_SRC (PATTERN (src_insn
));
1460 /* Don't need the last insn anymore. */
1461 delete_insn (get_last_insn ());
1463 if (GET_CODE (second_part
) != CONST_INT
1464 || GET_CODE (increment
) != CONST_INT
)
1468 increment
= GEN_INT (INTVAL (increment
) | INTVAL (second_part
));
1470 increment
= GEN_INT (INTVAL (increment
) << INTVAL (second_part
));
1473 if (GET_CODE (increment
) != CONST_INT
)
1476 /* The insn loading the constant into a register is no longer needed,
1478 delete_insn (get_last_insn ());
1481 if (increment_total
)
1482 increment_total
= GEN_INT (INTVAL (increment_total
) + INTVAL (increment
));
1484 increment_total
= increment
;
1486 /* Check that the source register is the same as the register we expected
1487 to see as the source. If not, something is seriously wrong. */
1488 if (GET_CODE (XEXP (SET_SRC (pattern
), 0)) != REG
1489 || REGNO (XEXP (SET_SRC (pattern
), 0)) != regno
)
1491 /* Some machines (e.g. the romp), may emit two add instructions for
1492 certain constants, so lets try looking for another add immediately
1493 before this one if we have only seen one add insn so far. */
1499 src_insn
= PREV_INSN (src_insn
);
1500 pattern
= PATTERN (src_insn
);
1502 delete_insn (get_last_insn ());
1510 return increment_total
;
1513 /* Copy REG_NOTES, except for insn references, because not all insn_map
1514 entries are valid yet. We do need to copy registers now though, because
1515 the reg_map entries can change during copying. */
1518 initial_reg_note_copy (notes
, map
)
1520 struct inline_remap
*map
;
1527 copy
= rtx_alloc (GET_CODE (notes
));
1528 PUT_MODE (copy
, GET_MODE (notes
));
1530 if (GET_CODE (notes
) == EXPR_LIST
)
1531 XEXP (copy
, 0) = copy_rtx_and_substitute (XEXP (notes
, 0), map
);
1532 else if (GET_CODE (notes
) == INSN_LIST
)
1533 /* Don't substitute for these yet. */
1534 XEXP (copy
, 0) = XEXP (notes
, 0);
1538 XEXP (copy
, 1) = initial_reg_note_copy (XEXP (notes
, 1), map
);
1543 /* Fixup insn references in copied REG_NOTES. */
1546 final_reg_note_copy (notes
, map
)
1548 struct inline_remap
*map
;
1552 for (note
= notes
; note
; note
= XEXP (note
, 1))
1553 if (GET_CODE (note
) == INSN_LIST
)
1554 XEXP (note
, 0) = map
->insn_map
[INSN_UID (XEXP (note
, 0))];
1557 /* Copy each instruction in the loop, substituting from map as appropriate.
1558 This is very similar to a loop in expand_inline_function. */
1561 copy_loop_body (copy_start
, copy_end
, map
, exit_label
, last_iteration
,
1562 unroll_type
, start_label
, loop_end
, insert_before
,
1564 rtx copy_start
, copy_end
;
1565 struct inline_remap
*map
;
1568 enum unroll_types unroll_type
;
1569 rtx start_label
, loop_end
, insert_before
, copy_notes_from
;
1573 int dest_reg_was_split
, i
;
1575 rtx final_label
= 0;
1576 rtx giv_inc
, giv_dest_reg
, giv_src_reg
;
1578 /* If this isn't the last iteration, then map any references to the
1579 start_label to final_label. Final label will then be emitted immediately
1580 after the end of this loop body if it was ever used.
1582 If this is the last iteration, then map references to the start_label
1584 if (! last_iteration
)
1586 final_label
= gen_label_rtx ();
1587 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = final_label
;
1590 map
->label_map
[CODE_LABEL_NUMBER (start_label
)] = start_label
;
1597 insn
= NEXT_INSN (insn
);
1599 map
->orig_asm_operands_vector
= 0;
1601 switch (GET_CODE (insn
))
1604 pattern
= PATTERN (insn
);
1608 /* Check to see if this is a giv that has been combined with
1609 some split address givs. (Combined in the sense that
1610 `combine_givs' in loop.c has put two givs in the same register.)
1611 In this case, we must search all givs based on the same biv to
1612 find the address givs. Then split the address givs.
1613 Do this before splitting the giv, since that may map the
1614 SET_DEST to a new register. */
1616 if (GET_CODE (pattern
) == SET
1617 && GET_CODE (SET_DEST (pattern
)) == REG
1618 && addr_combined_regs
[REGNO (SET_DEST (pattern
))])
1620 struct iv_class
*bl
;
1621 struct induction
*v
, *tv
;
1622 int regno
= REGNO (SET_DEST (pattern
));
1624 v
= addr_combined_regs
[REGNO (SET_DEST (pattern
))];
1625 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
1627 /* Although the giv_inc amount is not needed here, we must call
1628 calculate_giv_inc here since it might try to delete the
1629 last insn emitted. If we wait until later to call it,
1630 we might accidentally delete insns generated immediately
1631 below by emit_unrolled_add. */
1633 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1635 /* Now find all address giv's that were combined with this
1637 for (tv
= bl
->giv
; tv
; tv
= tv
->next_iv
)
1638 if (tv
->giv_type
== DEST_ADDR
&& tv
->same
== v
)
1642 /* If this DEST_ADDR giv was not split, then ignore it. */
1643 if (*tv
->location
!= tv
->dest_reg
)
1646 /* Scale this_giv_inc if the multiplicative factors of
1647 the two givs are different. */
1648 this_giv_inc
= INTVAL (giv_inc
);
1649 if (tv
->mult_val
!= v
->mult_val
)
1650 this_giv_inc
= (this_giv_inc
/ INTVAL (v
->mult_val
)
1651 * INTVAL (tv
->mult_val
));
1653 tv
->dest_reg
= plus_constant (tv
->dest_reg
, this_giv_inc
);
1654 *tv
->location
= tv
->dest_reg
;
1656 if (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)
1658 /* Must emit an insn to increment the split address
1659 giv. Add in the const_adjust field in case there
1660 was a constant eliminated from the address. */
1661 rtx value
, dest_reg
;
1663 /* tv->dest_reg will be either a bare register,
1664 or else a register plus a constant. */
1665 if (GET_CODE (tv
->dest_reg
) == REG
)
1666 dest_reg
= tv
->dest_reg
;
1668 dest_reg
= XEXP (tv
->dest_reg
, 0);
1670 /* Check for shared address givs, and avoid
1671 incrementing the shared pseudo reg more than
1673 if (! tv
->same_insn
)
1675 /* tv->dest_reg may actually be a (PLUS (REG)
1676 (CONST)) here, so we must call plus_constant
1677 to add the const_adjust amount before calling
1678 emit_unrolled_add below. */
1679 value
= plus_constant (tv
->dest_reg
,
1682 /* The constant could be too large for an add
1683 immediate, so can't directly emit an insn
1685 emit_unrolled_add (dest_reg
, XEXP (value
, 0),
1689 /* Reset the giv to be just the register again, in case
1690 it is used after the set we have just emitted.
1691 We must subtract the const_adjust factor added in
1693 tv
->dest_reg
= plus_constant (dest_reg
,
1694 - tv
->const_adjust
);
1695 *tv
->location
= tv
->dest_reg
;
1700 /* If this is a setting of a splittable variable, then determine
1701 how to split the variable, create a new set based on this split,
1702 and set up the reg_map so that later uses of the variable will
1703 use the new split variable. */
1705 dest_reg_was_split
= 0;
1707 if (GET_CODE (pattern
) == SET
1708 && GET_CODE (SET_DEST (pattern
)) == REG
1709 && splittable_regs
[REGNO (SET_DEST (pattern
))])
1711 int regno
= REGNO (SET_DEST (pattern
));
1713 dest_reg_was_split
= 1;
1715 /* Compute the increment value for the giv, if it wasn't
1716 already computed above. */
1719 giv_inc
= calculate_giv_inc (pattern
, insn
, regno
);
1720 giv_dest_reg
= SET_DEST (pattern
);
1721 giv_src_reg
= SET_DEST (pattern
);
1723 if (unroll_type
== UNROLL_COMPLETELY
)
1725 /* Completely unrolling the loop. Set the induction
1726 variable to a known constant value. */
1728 /* The value in splittable_regs may be an invariant
1729 value, so we must use plus_constant here. */
1730 splittable_regs
[regno
]
1731 = plus_constant (splittable_regs
[regno
], INTVAL (giv_inc
));
1733 if (GET_CODE (splittable_regs
[regno
]) == PLUS
)
1735 giv_src_reg
= XEXP (splittable_regs
[regno
], 0);
1736 giv_inc
= XEXP (splittable_regs
[regno
], 1);
1740 /* The splittable_regs value must be a REG or a
1741 CONST_INT, so put the entire value in the giv_src_reg
1743 giv_src_reg
= splittable_regs
[regno
];
1744 giv_inc
= const0_rtx
;
1749 /* Partially unrolling loop. Create a new pseudo
1750 register for the iteration variable, and set it to
1751 be a constant plus the original register. Except
1752 on the last iteration, when the result has to
1753 go back into the original iteration var register. */
1755 /* Handle bivs which must be mapped to a new register
1756 when split. This happens for bivs which need their
1757 final value set before loop entry. The new register
1758 for the biv was stored in the biv's first struct
1759 induction entry by find_splittable_regs. */
1761 if (regno
< max_reg_before_loop
1762 && reg_iv_type
[regno
] == BASIC_INDUCT
)
1764 giv_src_reg
= reg_biv_class
[regno
]->biv
->src_reg
;
1765 giv_dest_reg
= giv_src_reg
;
1769 /* If non-reduced/final-value givs were split, then
1770 this would have to remap those givs also. See
1771 find_splittable_regs. */
1774 splittable_regs
[regno
]
1775 = GEN_INT (INTVAL (giv_inc
)
1776 + INTVAL (splittable_regs
[regno
]));
1777 giv_inc
= splittable_regs
[regno
];
1779 /* Now split the induction variable by changing the dest
1780 of this insn to a new register, and setting its
1781 reg_map entry to point to this new register.
1783 If this is the last iteration, and this is the last insn
1784 that will update the iv, then reuse the original dest,
1785 to ensure that the iv will have the proper value when
1786 the loop exits or repeats.
1788 Using splittable_regs_updates here like this is safe,
1789 because it can only be greater than one if all
1790 instructions modifying the iv are always executed in
1793 if (! last_iteration
1794 || (splittable_regs_updates
[regno
]-- != 1))
1796 tem
= gen_reg_rtx (GET_MODE (giv_src_reg
));
1798 map
->reg_map
[regno
] = tem
;
1801 map
->reg_map
[regno
] = giv_src_reg
;
1804 /* The constant being added could be too large for an add
1805 immediate, so can't directly emit an insn here. */
1806 emit_unrolled_add (giv_dest_reg
, giv_src_reg
, giv_inc
);
1807 copy
= get_last_insn ();
1808 pattern
= PATTERN (copy
);
1812 pattern
= copy_rtx_and_substitute (pattern
, map
);
1813 copy
= emit_insn (pattern
);
1815 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1818 /* If this insn is setting CC0, it may need to look at
1819 the insn that uses CC0 to see what type of insn it is.
1820 In that case, the call to recog via validate_change will
1821 fail. So don't substitute constants here. Instead,
1822 do it when we emit the following insn.
1824 For example, see the pyr.md file. That machine has signed and
1825 unsigned compares. The compare patterns must check the
1826 following branch insn to see which what kind of compare to
1829 If the previous insn set CC0, substitute constants on it as
1831 if (sets_cc0_p (PATTERN (copy
)) != 0)
1836 try_constants (cc0_insn
, map
);
1838 try_constants (copy
, map
);
1841 try_constants (copy
, map
);
1844 /* Make split induction variable constants `permanent' since we
1845 know there are no backward branches across iteration variable
1846 settings which would invalidate this. */
1847 if (dest_reg_was_split
)
1849 int regno
= REGNO (SET_DEST (pattern
));
1851 if (regno
< map
->const_equiv_map_size
1852 && map
->const_age_map
[regno
] == map
->const_age
)
1853 map
->const_age_map
[regno
] = -1;
1858 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1859 copy
= emit_jump_insn (pattern
);
1860 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1862 if (JUMP_LABEL (insn
) == start_label
&& insn
== copy_end
1863 && ! last_iteration
)
1865 /* This is a branch to the beginning of the loop; this is the
1866 last insn being copied; and this is not the last iteration.
1867 In this case, we want to change the original fall through
1868 case to be a branch past the end of the loop, and the
1869 original jump label case to fall_through. */
1871 if (invert_exp (pattern
, copy
))
1873 if (! redirect_exp (&pattern
,
1874 map
->label_map
[CODE_LABEL_NUMBER
1875 (JUMP_LABEL (insn
))],
1882 rtx lab
= gen_label_rtx ();
1883 /* Can't do it by reversing the jump (probably because we
1884 couldn't reverse the conditions), so emit a new
1885 jump_insn after COPY, and redirect the jump around
1887 jmp
= emit_jump_insn_after (gen_jump (exit_label
), copy
);
1888 jmp
= emit_barrier_after (jmp
);
1889 emit_label_after (lab
, jmp
);
1890 LABEL_NUSES (lab
) = 0;
1891 if (! redirect_exp (&pattern
,
1892 map
->label_map
[CODE_LABEL_NUMBER
1893 (JUMP_LABEL (insn
))],
1901 try_constants (cc0_insn
, map
);
1904 try_constants (copy
, map
);
1906 /* Set the jump label of COPY correctly to avoid problems with
1907 later passes of unroll_loop, if INSN had jump label set. */
1908 if (JUMP_LABEL (insn
))
1912 /* Can't use the label_map for every insn, since this may be
1913 the backward branch, and hence the label was not mapped. */
1914 if (GET_CODE (pattern
) == SET
)
1916 tem
= SET_SRC (pattern
);
1917 if (GET_CODE (tem
) == LABEL_REF
)
1918 label
= XEXP (tem
, 0);
1919 else if (GET_CODE (tem
) == IF_THEN_ELSE
)
1921 if (XEXP (tem
, 1) != pc_rtx
)
1922 label
= XEXP (XEXP (tem
, 1), 0);
1924 label
= XEXP (XEXP (tem
, 2), 0);
1928 if (label
&& GET_CODE (label
) == CODE_LABEL
)
1929 JUMP_LABEL (copy
) = label
;
1932 /* An unrecognizable jump insn, probably the entry jump
1933 for a switch statement. This label must have been mapped,
1934 so just use the label_map to get the new jump label. */
1936 = map
->label_map
[CODE_LABEL_NUMBER (JUMP_LABEL (insn
))];
1939 /* If this is a non-local jump, then must increase the label
1940 use count so that the label will not be deleted when the
1941 original jump is deleted. */
1942 LABEL_NUSES (JUMP_LABEL (copy
))++;
1944 else if (GET_CODE (PATTERN (copy
)) == ADDR_VEC
1945 || GET_CODE (PATTERN (copy
)) == ADDR_DIFF_VEC
)
1947 rtx pat
= PATTERN (copy
);
1948 int diff_vec_p
= GET_CODE (pat
) == ADDR_DIFF_VEC
;
1949 int len
= XVECLEN (pat
, diff_vec_p
);
1952 for (i
= 0; i
< len
; i
++)
1953 LABEL_NUSES (XEXP (XVECEXP (pat
, diff_vec_p
, i
), 0))++;
1956 /* If this used to be a conditional jump insn but whose branch
1957 direction is now known, we must do something special. */
1958 if (condjump_p (insn
) && !simplejump_p (insn
) && map
->last_pc_value
)
1961 /* The previous insn set cc0 for us. So delete it. */
1962 delete_insn (PREV_INSN (copy
));
1965 /* If this is now a no-op, delete it. */
1966 if (map
->last_pc_value
== pc_rtx
)
1968 /* Don't let delete_insn delete the label referenced here,
1969 because we might possibly need it later for some other
1970 instruction in the loop. */
1971 if (JUMP_LABEL (copy
))
1972 LABEL_NUSES (JUMP_LABEL (copy
))++;
1974 if (JUMP_LABEL (copy
))
1975 LABEL_NUSES (JUMP_LABEL (copy
))--;
1979 /* Otherwise, this is unconditional jump so we must put a
1980 BARRIER after it. We could do some dead code elimination
1981 here, but jump.c will do it just as well. */
1987 pattern
= copy_rtx_and_substitute (PATTERN (insn
), map
);
1988 copy
= emit_call_insn (pattern
);
1989 REG_NOTES (copy
) = initial_reg_note_copy (REG_NOTES (insn
), map
);
1991 /* Because the USAGE information potentially contains objects other
1992 than hard registers, we need to copy it. */
1993 CALL_INSN_FUNCTION_USAGE (copy
) =
1994 copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn
), map
);
1998 try_constants (cc0_insn
, map
);
2001 try_constants (copy
, map
);
2003 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2004 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
2005 map
->const_equiv_map
[i
] = 0;
2009 /* If this is the loop start label, then we don't need to emit a
2010 copy of this label since no one will use it. */
2012 if (insn
!= start_label
)
2014 copy
= emit_label (map
->label_map
[CODE_LABEL_NUMBER (insn
)]);
2020 copy
= emit_barrier ();
2024 /* VTOP notes are valid only before the loop exit test. If placed
2025 anywhere else, loop may generate bad code. */
2027 if (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
2028 && (NOTE_LINE_NUMBER (insn
) != NOTE_INSN_LOOP_VTOP
2029 || (last_iteration
&& unroll_type
!= UNROLL_COMPLETELY
)))
2030 copy
= emit_note (NOTE_SOURCE_FILE (insn
),
2031 NOTE_LINE_NUMBER (insn
));
2041 map
->insn_map
[INSN_UID (insn
)] = copy
;
2043 while (insn
!= copy_end
);
2045 /* Now finish coping the REG_NOTES. */
2049 insn
= NEXT_INSN (insn
);
2050 if ((GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == JUMP_INSN
2051 || GET_CODE (insn
) == CALL_INSN
)
2052 && map
->insn_map
[INSN_UID (insn
)])
2053 final_reg_note_copy (REG_NOTES (map
->insn_map
[INSN_UID (insn
)]), map
);
2055 while (insn
!= copy_end
);
2057 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2058 each of these notes here, since there may be some important ones, such as
2059 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2060 iteration, because the original notes won't be deleted.
2062 We can't use insert_before here, because when from preconditioning,
2063 insert_before points before the loop. We can't use copy_end, because
2064 there may be insns already inserted after it (which we don't want to
2065 copy) when not from preconditioning code. */
2067 if (! last_iteration
)
2069 for (insn
= copy_notes_from
; insn
!= loop_end
; insn
= NEXT_INSN (insn
))
2071 if (GET_CODE (insn
) == NOTE
2072 && NOTE_LINE_NUMBER (insn
) != NOTE_INSN_DELETED
)
2073 emit_note (NOTE_SOURCE_FILE (insn
), NOTE_LINE_NUMBER (insn
));
2077 if (final_label
&& LABEL_NUSES (final_label
) > 0)
2078 emit_label (final_label
);
2080 tem
= gen_sequence ();
2082 emit_insn_before (tem
, insert_before
);
2085 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2086 emitted. This will correctly handle the case where the increment value
2087 won't fit in the immediate field of a PLUS insns. */
2090 emit_unrolled_add (dest_reg
, src_reg
, increment
)
2091 rtx dest_reg
, src_reg
, increment
;
2095 result
= expand_binop (GET_MODE (dest_reg
), add_optab
, src_reg
, increment
,
2096 dest_reg
, 0, OPTAB_LIB_WIDEN
);
2098 if (dest_reg
!= result
)
2099 emit_move_insn (dest_reg
, result
);
2102 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2103 is a backward branch in that range that branches to somewhere between
2104 LOOP_START and INSN. Returns 0 otherwise. */
2106 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2107 In practice, this is not a problem, because this function is seldom called,
2108 and uses a negligible amount of CPU time on average. */
2111 back_branch_in_range_p (insn
, loop_start
, loop_end
)
2113 rtx loop_start
, loop_end
;
2115 rtx p
, q
, target_insn
;
2117 /* Stop before we get to the backward branch at the end of the loop. */
2118 loop_end
= prev_nonnote_insn (loop_end
);
2119 if (GET_CODE (loop_end
) == BARRIER
)
2120 loop_end
= PREV_INSN (loop_end
);
2122 /* Check in case insn has been deleted, search forward for first non
2123 deleted insn following it. */
2124 while (INSN_DELETED_P (insn
))
2125 insn
= NEXT_INSN (insn
);
2127 /* Check for the case where insn is the last insn in the loop. */
2128 if (insn
== loop_end
)
2131 for (p
= NEXT_INSN (insn
); p
!= loop_end
; p
= NEXT_INSN (p
))
2133 if (GET_CODE (p
) == JUMP_INSN
)
2135 target_insn
= JUMP_LABEL (p
);
2137 /* Search from loop_start to insn, to see if one of them is
2138 the target_insn. We can't use INSN_LUID comparisons here,
2139 since insn may not have an LUID entry. */
2140 for (q
= loop_start
; q
!= insn
; q
= NEXT_INSN (q
))
2141 if (q
== target_insn
)
2149 /* Try to generate the simplest rtx for the expression
2150 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2154 fold_rtx_mult_add (mult1
, mult2
, add1
, mode
)
2155 rtx mult1
, mult2
, add1
;
2156 enum machine_mode mode
;
2161 /* The modes must all be the same. This should always be true. For now,
2162 check to make sure. */
2163 if ((GET_MODE (mult1
) != mode
&& GET_MODE (mult1
) != VOIDmode
)
2164 || (GET_MODE (mult2
) != mode
&& GET_MODE (mult2
) != VOIDmode
)
2165 || (GET_MODE (add1
) != mode
&& GET_MODE (add1
) != VOIDmode
))
2168 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2169 will be a constant. */
2170 if (GET_CODE (mult1
) == CONST_INT
)
2177 mult_res
= simplify_binary_operation (MULT
, mode
, mult1
, mult2
);
2179 mult_res
= gen_rtx (MULT
, mode
, mult1
, mult2
);
2181 /* Again, put the constant second. */
2182 if (GET_CODE (add1
) == CONST_INT
)
2189 result
= simplify_binary_operation (PLUS
, mode
, add1
, mult_res
);
2191 result
= gen_rtx (PLUS
, mode
, add1
, mult_res
);
2196 /* Searches the list of induction struct's for the biv BL, to try to calculate
2197 the total increment value for one iteration of the loop as a constant.
2199 Returns the increment value as an rtx, simplified as much as possible,
2200 if it can be calculated. Otherwise, returns 0. */
2203 biv_total_increment (bl
, loop_start
, loop_end
)
2204 struct iv_class
*bl
;
2205 rtx loop_start
, loop_end
;
2207 struct induction
*v
;
2210 /* For increment, must check every instruction that sets it. Each
2211 instruction must be executed only once each time through the loop.
2212 To verify this, we check that the the insn is always executed, and that
2213 there are no backward branches after the insn that branch to before it.
2214 Also, the insn must have a mult_val of one (to make sure it really is
2217 result
= const0_rtx
;
2218 for (v
= bl
->biv
; v
; v
= v
->next_iv
)
2220 if (v
->always_computable
&& v
->mult_val
== const1_rtx
2221 && ! back_branch_in_range_p (v
->insn
, loop_start
, loop_end
))
2222 result
= fold_rtx_mult_add (result
, const1_rtx
, v
->add_val
, v
->mode
);
2230 /* Determine the initial value of the iteration variable, and the amount
2231 that it is incremented each loop. Use the tables constructed by
2232 the strength reduction pass to calculate these values.
2234 Initial_value and/or increment are set to zero if their values could not
2238 iteration_info (iteration_var
, initial_value
, increment
, loop_start
, loop_end
)
2239 rtx iteration_var
, *initial_value
, *increment
;
2240 rtx loop_start
, loop_end
;
2242 struct iv_class
*bl
;
2243 struct induction
*v
, *b
;
2245 /* Clear the result values, in case no answer can be found. */
2249 /* The iteration variable can be either a giv or a biv. Check to see
2250 which it is, and compute the variable's initial value, and increment
2251 value if possible. */
2253 /* If this is a new register, can't handle it since we don't have any
2254 reg_iv_type entry for it. */
2255 if (REGNO (iteration_var
) >= max_reg_before_loop
)
2257 if (loop_dump_stream
)
2258 fprintf (loop_dump_stream
,
2259 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2262 /* Reject iteration variables larger than the host long size, since they
2263 could result in a number of iterations greater than the range of our
2264 `unsigned long' variable loop_n_iterations. */
2265 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var
)) > HOST_BITS_PER_LONG
)
2267 if (loop_dump_stream
)
2268 fprintf (loop_dump_stream
,
2269 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2272 else if (GET_MODE_CLASS (GET_MODE (iteration_var
)) != MODE_INT
)
2274 if (loop_dump_stream
)
2275 fprintf (loop_dump_stream
,
2276 "Loop unrolling: Iteration var not an integer.\n");
2279 else if (reg_iv_type
[REGNO (iteration_var
)] == BASIC_INDUCT
)
2281 /* Grab initial value, only useful if it is a constant. */
2282 bl
= reg_biv_class
[REGNO (iteration_var
)];
2283 *initial_value
= bl
->initial_value
;
2285 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2287 else if (reg_iv_type
[REGNO (iteration_var
)] == GENERAL_INDUCT
)
2290 /* ??? The code below does not work because the incorrect number of
2291 iterations is calculated when the biv is incremented after the giv
2292 is set (which is the usual case). This can probably be accounted
2293 for by biasing the initial_value by subtracting the amount of the
2294 increment that occurs between the giv set and the giv test. However,
2295 a giv as an iterator is very rare, so it does not seem worthwhile
2297 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2298 if (loop_dump_stream
)
2299 fprintf (loop_dump_stream
,
2300 "Loop unrolling: Giv iterators are not handled.\n");
2303 /* Initial value is mult_val times the biv's initial value plus
2304 add_val. Only useful if it is a constant. */
2305 v
= reg_iv_info
[REGNO (iteration_var
)];
2306 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
2307 *initial_value
= fold_rtx_mult_add (v
->mult_val
, bl
->initial_value
,
2308 v
->add_val
, v
->mode
);
2310 /* Increment value is mult_val times the increment value of the biv. */
2312 *increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2314 *increment
= fold_rtx_mult_add (v
->mult_val
, *increment
, const0_rtx
,
2320 if (loop_dump_stream
)
2321 fprintf (loop_dump_stream
,
2322 "Loop unrolling: Not basic or general induction var.\n");
2327 /* Calculate the approximate final value of the iteration variable
2328 which has an loop exit test with code COMPARISON_CODE and comparison value
2329 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2330 was signed or unsigned, and the direction of the comparison. This info is
2331 needed to calculate the number of loop iterations. */
2334 approx_final_value (comparison_code
, comparison_value
, unsigned_p
, compare_dir
)
2335 enum rtx_code comparison_code
;
2336 rtx comparison_value
;
2340 /* Calculate the final value of the induction variable.
2341 The exact final value depends on the branch operator, and increment sign.
2342 This is only an approximate value. It will be wrong if the iteration
2343 variable is not incremented by one each time through the loop, and
2344 approx final value - start value % increment != 0. */
2347 switch (comparison_code
)
2353 return plus_constant (comparison_value
, 1);
2358 return plus_constant (comparison_value
, -1);
2360 /* Can not calculate a final value for this case. */
2367 return comparison_value
;
2373 return comparison_value
;
2376 return comparison_value
;
2382 /* For each biv and giv, determine whether it can be safely split into
2383 a different variable for each unrolled copy of the loop body. If it
2384 is safe to split, then indicate that by saving some useful info
2385 in the splittable_regs array.
2387 If the loop is being completely unrolled, then splittable_regs will hold
2388 the current value of the induction variable while the loop is unrolled.
2389 It must be set to the initial value of the induction variable here.
2390 Otherwise, splittable_regs will hold the difference between the current
2391 value of the induction variable and the value the induction variable had
2392 at the top of the loop. It must be set to the value 0 here.
2394 Returns the total number of instructions that set registers that are
2397 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2398 constant values are unnecessary, since we can easily calculate increment
2399 values in this case even if nothing is constant. The increment value
2400 should not involve a multiply however. */
2402 /* ?? Even if the biv/giv increment values aren't constant, it may still
2403 be beneficial to split the variable if the loop is only unrolled a few
2404 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2407 find_splittable_regs (unroll_type
, loop_start
, loop_end
, end_insert_before
,
2409 enum unroll_types unroll_type
;
2410 rtx loop_start
, loop_end
;
2411 rtx end_insert_before
;
2414 struct iv_class
*bl
;
2415 struct induction
*v
;
2417 rtx biv_final_value
;
2421 for (bl
= loop_iv_list
; bl
; bl
= bl
->next
)
2423 /* Biv_total_increment must return a constant value,
2424 otherwise we can not calculate the split values. */
2426 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
2427 if (! increment
|| GET_CODE (increment
) != CONST_INT
)
2430 /* The loop must be unrolled completely, or else have a known number
2431 of iterations and only one exit, or else the biv must be dead
2432 outside the loop, or else the final value must be known. Otherwise,
2433 it is unsafe to split the biv since it may not have the proper
2434 value on loop exit. */
2436 /* loop_number_exit_count is non-zero if the loop has an exit other than
2437 a fall through at the end. */
2440 biv_final_value
= 0;
2441 if (unroll_type
!= UNROLL_COMPLETELY
2442 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2443 || unroll_type
== UNROLL_NAIVE
)
2444 && (uid_luid
[regno_last_uid
[bl
->regno
]] >= INSN_LUID (loop_end
)
2446 || INSN_UID (bl
->init_insn
) >= max_uid_for_loop
2447 || (uid_luid
[regno_first_uid
[bl
->regno
]]
2448 < INSN_LUID (bl
->init_insn
))
2449 || reg_mentioned_p (bl
->biv
->dest_reg
, SET_SRC (bl
->init_set
)))
2450 && ! (biv_final_value
= final_biv_value (bl
, loop_start
, loop_end
)))
2453 /* If any of the insns setting the BIV don't do so with a simple
2454 PLUS, we don't know how to split it. */
2455 for (v
= bl
->biv
; biv_splittable
&& v
; v
= v
->next_iv
)
2456 if ((tem
= single_set (v
->insn
)) == 0
2457 || GET_CODE (SET_DEST (tem
)) != REG
2458 || REGNO (SET_DEST (tem
)) != bl
->regno
2459 || GET_CODE (SET_SRC (tem
)) != PLUS
)
2462 /* If final value is non-zero, then must emit an instruction which sets
2463 the value of the biv to the proper value. This is done after
2464 handling all of the givs, since some of them may need to use the
2465 biv's value in their initialization code. */
2467 /* This biv is splittable. If completely unrolling the loop, save
2468 the biv's initial value. Otherwise, save the constant zero. */
2470 if (biv_splittable
== 1)
2472 if (unroll_type
== UNROLL_COMPLETELY
)
2474 /* If the initial value of the biv is itself (i.e. it is too
2475 complicated for strength_reduce to compute), or is a hard
2476 register, or it isn't invariant, then we must create a new
2477 pseudo reg to hold the initial value of the biv. */
2479 if (GET_CODE (bl
->initial_value
) == REG
2480 && (REGNO (bl
->initial_value
) == bl
->regno
2481 || REGNO (bl
->initial_value
) < FIRST_PSEUDO_REGISTER
2482 || ! invariant_p (bl
->initial_value
)))
2484 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2486 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2489 if (loop_dump_stream
)
2490 fprintf (loop_dump_stream
, "Biv %d initial value remapped to %d.\n",
2491 bl
->regno
, REGNO (tem
));
2493 splittable_regs
[bl
->regno
] = tem
;
2496 splittable_regs
[bl
->regno
] = bl
->initial_value
;
2499 splittable_regs
[bl
->regno
] = const0_rtx
;
2501 /* Save the number of instructions that modify the biv, so that
2502 we can treat the last one specially. */
2504 splittable_regs_updates
[bl
->regno
] = bl
->biv_count
;
2505 result
+= bl
->biv_count
;
2507 if (loop_dump_stream
)
2508 fprintf (loop_dump_stream
,
2509 "Biv %d safe to split.\n", bl
->regno
);
2512 /* Check every giv that depends on this biv to see whether it is
2513 splittable also. Even if the biv isn't splittable, givs which
2514 depend on it may be splittable if the biv is live outside the
2515 loop, and the givs aren't. */
2517 result
+= find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
,
2518 increment
, unroll_number
);
2520 /* If final value is non-zero, then must emit an instruction which sets
2521 the value of the biv to the proper value. This is done after
2522 handling all of the givs, since some of them may need to use the
2523 biv's value in their initialization code. */
2524 if (biv_final_value
)
2526 /* If the loop has multiple exits, emit the insns before the
2527 loop to ensure that it will always be executed no matter
2528 how the loop exits. Otherwise emit the insn after the loop,
2529 since this is slightly more efficient. */
2530 if (! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
2531 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2536 /* Create a new register to hold the value of the biv, and then
2537 set the biv to its final value before the loop start. The biv
2538 is set to its final value before loop start to ensure that
2539 this insn will always be executed, no matter how the loop
2541 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2542 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2544 emit_insn_before (gen_move_insn (bl
->biv
->src_reg
,
2548 if (loop_dump_stream
)
2549 fprintf (loop_dump_stream
, "Biv %d mapped to %d for split.\n",
2550 REGNO (bl
->biv
->src_reg
), REGNO (tem
));
2552 /* Set up the mapping from the original biv register to the new
2554 bl
->biv
->src_reg
= tem
;
2561 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2562 for the instruction that is using it. Do not make any changes to that
2566 verify_addresses (v
, giv_inc
, unroll_number
)
2567 struct induction
*v
;
2572 rtx orig_addr
= *v
->location
;
2573 rtx last_addr
= plus_constant (v
->dest_reg
,
2574 INTVAL (giv_inc
) * (unroll_number
- 1));
2576 /* First check to see if either address would fail. */
2577 if (! validate_change (v
->insn
, v
->location
, v
->dest_reg
, 0)
2578 || ! validate_change (v
->insn
, v
->location
, last_addr
, 0))
2581 /* Now put things back the way they were before. This will always
2583 validate_change (v
->insn
, v
->location
, orig_addr
, 0);
2588 /* For every giv based on the biv BL, check to determine whether it is
2589 splittable. This is a subroutine to find_splittable_regs ().
2591 Return the number of instructions that set splittable registers. */
2594 find_splittable_givs (bl
, unroll_type
, loop_start
, loop_end
, increment
,
2596 struct iv_class
*bl
;
2597 enum unroll_types unroll_type
;
2598 rtx loop_start
, loop_end
;
2602 struct induction
*v
, *v2
;
2607 /* Scan the list of givs, and set the same_insn field when there are
2608 multiple identical givs in the same insn. */
2609 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2610 for (v2
= v
->next_iv
; v2
; v2
= v2
->next_iv
)
2611 if (v
->insn
== v2
->insn
&& rtx_equal_p (v
->new_reg
, v2
->new_reg
)
2615 for (v
= bl
->giv
; v
; v
= v
->next_iv
)
2619 /* Only split the giv if it has already been reduced, or if the loop is
2620 being completely unrolled. */
2621 if (unroll_type
!= UNROLL_COMPLETELY
&& v
->ignore
)
2624 /* The giv can be split if the insn that sets the giv is executed once
2625 and only once on every iteration of the loop. */
2626 /* An address giv can always be split. v->insn is just a use not a set,
2627 and hence it does not matter whether it is always executed. All that
2628 matters is that all the biv increments are always executed, and we
2629 won't reach here if they aren't. */
2630 if (v
->giv_type
!= DEST_ADDR
2631 && (! v
->always_computable
2632 || back_branch_in_range_p (v
->insn
, loop_start
, loop_end
)))
2635 /* The giv increment value must be a constant. */
2636 giv_inc
= fold_rtx_mult_add (v
->mult_val
, increment
, const0_rtx
,
2638 if (! giv_inc
|| GET_CODE (giv_inc
) != CONST_INT
)
2641 /* The loop must be unrolled completely, or else have a known number of
2642 iterations and only one exit, or else the giv must be dead outside
2643 the loop, or else the final value of the giv must be known.
2644 Otherwise, it is not safe to split the giv since it may not have the
2645 proper value on loop exit. */
2647 /* The used outside loop test will fail for DEST_ADDR givs. They are
2648 never used outside the loop anyways, so it is always safe to split a
2652 if (unroll_type
!= UNROLL_COMPLETELY
2653 && (loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
2654 || unroll_type
== UNROLL_NAIVE
)
2655 && v
->giv_type
!= DEST_ADDR
2656 && ((regno_first_uid
[REGNO (v
->dest_reg
)] != INSN_UID (v
->insn
)
2657 /* Check for the case where the pseudo is set by a shift/add
2658 sequence, in which case the first insn setting the pseudo
2659 is the first insn of the shift/add sequence. */
2660 && (! (tem
= find_reg_note (v
->insn
, REG_RETVAL
, NULL_RTX
))
2661 || (regno_first_uid
[REGNO (v
->dest_reg
)]
2662 != INSN_UID (XEXP (tem
, 0)))))
2663 /* Line above always fails if INSN was moved by loop opt. */
2664 || (uid_luid
[regno_last_uid
[REGNO (v
->dest_reg
)]]
2665 >= INSN_LUID (loop_end
)))
2666 && ! (final_value
= v
->final_value
))
2670 /* Currently, non-reduced/final-value givs are never split. */
2671 /* Should emit insns after the loop if possible, as the biv final value
2674 /* If the final value is non-zero, and the giv has not been reduced,
2675 then must emit an instruction to set the final value. */
2676 if (final_value
&& !v
->new_reg
)
2678 /* Create a new register to hold the value of the giv, and then set
2679 the giv to its final value before the loop start. The giv is set
2680 to its final value before loop start to ensure that this insn
2681 will always be executed, no matter how we exit. */
2682 tem
= gen_reg_rtx (v
->mode
);
2683 emit_insn_before (gen_move_insn (tem
, v
->dest_reg
), loop_start
);
2684 emit_insn_before (gen_move_insn (v
->dest_reg
, final_value
),
2687 if (loop_dump_stream
)
2688 fprintf (loop_dump_stream
, "Giv %d mapped to %d for split.\n",
2689 REGNO (v
->dest_reg
), REGNO (tem
));
2695 /* This giv is splittable. If completely unrolling the loop, save the
2696 giv's initial value. Otherwise, save the constant zero for it. */
2698 if (unroll_type
== UNROLL_COMPLETELY
)
2700 /* It is not safe to use bl->initial_value here, because it may not
2701 be invariant. It is safe to use the initial value stored in
2702 the splittable_regs array if it is set. In rare cases, it won't
2703 be set, so then we do exactly the same thing as
2704 find_splittable_regs does to get a safe value. */
2705 rtx biv_initial_value
;
2707 if (splittable_regs
[bl
->regno
])
2708 biv_initial_value
= splittable_regs
[bl
->regno
];
2709 else if (GET_CODE (bl
->initial_value
) != REG
2710 || (REGNO (bl
->initial_value
) != bl
->regno
2711 && REGNO (bl
->initial_value
) >= FIRST_PSEUDO_REGISTER
))
2712 biv_initial_value
= bl
->initial_value
;
2715 rtx tem
= gen_reg_rtx (bl
->biv
->mode
);
2717 emit_insn_before (gen_move_insn (tem
, bl
->biv
->src_reg
),
2719 biv_initial_value
= tem
;
2721 value
= fold_rtx_mult_add (v
->mult_val
, biv_initial_value
,
2722 v
->add_val
, v
->mode
);
2729 /* If a giv was combined with another giv, then we can only split
2730 this giv if the giv it was combined with was reduced. This
2731 is because the value of v->new_reg is meaningless in this
2733 if (v
->same
&& ! v
->same
->new_reg
)
2735 if (loop_dump_stream
)
2736 fprintf (loop_dump_stream
,
2737 "giv combined with unreduced giv not split.\n");
2740 /* If the giv is an address destination, it could be something other
2741 than a simple register, these have to be treated differently. */
2742 else if (v
->giv_type
== DEST_REG
)
2744 /* If value is not a constant, register, or register plus
2745 constant, then compute its value into a register before
2746 loop start. This prevents invalid rtx sharing, and should
2747 generate better code. We can use bl->initial_value here
2748 instead of splittable_regs[bl->regno] because this code
2749 is going before the loop start. */
2750 if (unroll_type
== UNROLL_COMPLETELY
2751 && GET_CODE (value
) != CONST_INT
2752 && GET_CODE (value
) != REG
2753 && (GET_CODE (value
) != PLUS
2754 || GET_CODE (XEXP (value
, 0)) != REG
2755 || GET_CODE (XEXP (value
, 1)) != CONST_INT
))
2757 rtx tem
= gen_reg_rtx (v
->mode
);
2758 emit_iv_add_mult (bl
->initial_value
, v
->mult_val
,
2759 v
->add_val
, tem
, loop_start
);
2763 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2767 /* Splitting address givs is useful since it will often allow us
2768 to eliminate some increment insns for the base giv as
2771 /* If the addr giv is combined with a dest_reg giv, then all
2772 references to that dest reg will be remapped, which is NOT
2773 what we want for split addr regs. We always create a new
2774 register for the split addr giv, just to be safe. */
2776 /* ??? If there are multiple address givs which have been
2777 combined with the same dest_reg giv, then we may only need
2778 one new register for them. Pulling out constants below will
2779 catch some of the common cases of this. Currently, I leave
2780 the work of simplifying multiple address givs to the
2781 following cse pass. */
2783 /* As a special case, if we have multiple identical address givs
2784 within a single instruction, then we do use a single pseudo
2785 reg for both. This is necessary in case one is a match_dup
2788 v
->const_adjust
= 0;
2792 v
->dest_reg
= v
->same_insn
->dest_reg
;
2793 if (loop_dump_stream
)
2794 fprintf (loop_dump_stream
,
2795 "Sharing address givs in insn %d\n",
2796 INSN_UID (v
->insn
));
2798 else if (unroll_type
!= UNROLL_COMPLETELY
)
2800 /* If not completely unrolling the loop, then create a new
2801 register to hold the split value of the DEST_ADDR giv.
2802 Emit insn to initialize its value before loop start. */
2803 tem
= gen_reg_rtx (v
->mode
);
2805 /* If the address giv has a constant in its new_reg value,
2806 then this constant can be pulled out and put in value,
2807 instead of being part of the initialization code. */
2809 if (GET_CODE (v
->new_reg
) == PLUS
2810 && GET_CODE (XEXP (v
->new_reg
, 1)) == CONST_INT
)
2813 = plus_constant (tem
, INTVAL (XEXP (v
->new_reg
,1)));
2815 /* Only succeed if this will give valid addresses.
2816 Try to validate both the first and the last
2817 address resulting from loop unrolling, if
2818 one fails, then can't do const elim here. */
2819 if (verify_addresses (v
, giv_inc
, unroll_number
))
2821 /* Save the negative of the eliminated const, so
2822 that we can calculate the dest_reg's increment
2824 v
->const_adjust
= - INTVAL (XEXP (v
->new_reg
, 1));
2826 v
->new_reg
= XEXP (v
->new_reg
, 0);
2827 if (loop_dump_stream
)
2828 fprintf (loop_dump_stream
,
2829 "Eliminating constant from giv %d\n",
2838 /* If the address hasn't been checked for validity yet, do so
2839 now, and fail completely if either the first or the last
2840 unrolled copy of the address is not a valid address
2841 for the instruction that uses it. */
2842 if (v
->dest_reg
== tem
2843 && ! verify_addresses (v
, giv_inc
, unroll_number
))
2845 if (loop_dump_stream
)
2846 fprintf (loop_dump_stream
,
2847 "Invalid address for giv at insn %d\n",
2848 INSN_UID (v
->insn
));
2852 /* To initialize the new register, just move the value of
2853 new_reg into it. This is not guaranteed to give a valid
2854 instruction on machines with complex addressing modes.
2855 If we can't recognize it, then delete it and emit insns
2856 to calculate the value from scratch. */
2857 emit_insn_before (gen_rtx (SET
, VOIDmode
, tem
,
2858 copy_rtx (v
->new_reg
)),
2860 if (recog_memoized (PREV_INSN (loop_start
)) < 0)
2864 /* We can't use bl->initial_value to compute the initial
2865 value, because the loop may have been preconditioned.
2866 We must calculate it from NEW_REG. Try using
2867 force_operand instead of emit_iv_add_mult. */
2868 delete_insn (PREV_INSN (loop_start
));
2871 ret
= force_operand (v
->new_reg
, tem
);
2873 emit_move_insn (tem
, ret
);
2874 sequence
= gen_sequence ();
2876 emit_insn_before (sequence
, loop_start
);
2878 if (loop_dump_stream
)
2879 fprintf (loop_dump_stream
,
2880 "Invalid init insn, rewritten.\n");
2885 v
->dest_reg
= value
;
2887 /* Check the resulting address for validity, and fail
2888 if the resulting address would be invalid. */
2889 if (! verify_addresses (v
, giv_inc
, unroll_number
))
2891 if (loop_dump_stream
)
2892 fprintf (loop_dump_stream
,
2893 "Invalid address for giv at insn %d\n",
2894 INSN_UID (v
->insn
));
2899 /* Store the value of dest_reg into the insn. This sharing
2900 will not be a problem as this insn will always be copied
2903 *v
->location
= v
->dest_reg
;
2905 /* If this address giv is combined with a dest reg giv, then
2906 save the base giv's induction pointer so that we will be
2907 able to handle this address giv properly. The base giv
2908 itself does not have to be splittable. */
2910 if (v
->same
&& v
->same
->giv_type
== DEST_REG
)
2911 addr_combined_regs
[REGNO (v
->same
->new_reg
)] = v
->same
;
2913 if (GET_CODE (v
->new_reg
) == REG
)
2915 /* This giv maybe hasn't been combined with any others.
2916 Make sure that it's giv is marked as splittable here. */
2918 splittable_regs
[REGNO (v
->new_reg
)] = value
;
2920 /* Make it appear to depend upon itself, so that the
2921 giv will be properly split in the main loop above. */
2925 addr_combined_regs
[REGNO (v
->new_reg
)] = v
;
2929 if (loop_dump_stream
)
2930 fprintf (loop_dump_stream
, "DEST_ADDR giv being split.\n");
2936 /* Currently, unreduced giv's can't be split. This is not too much
2937 of a problem since unreduced giv's are not live across loop
2938 iterations anyways. When unrolling a loop completely though,
2939 it makes sense to reduce&split givs when possible, as this will
2940 result in simpler instructions, and will not require that a reg
2941 be live across loop iterations. */
2943 splittable_regs
[REGNO (v
->dest_reg
)] = value
;
2944 fprintf (stderr
, "Giv %d at insn %d not reduced\n",
2945 REGNO (v
->dest_reg
), INSN_UID (v
->insn
));
2951 /* Givs are only updated once by definition. Mark it so if this is
2952 a splittable register. Don't need to do anything for address givs
2953 where this may not be a register. */
2955 if (GET_CODE (v
->new_reg
) == REG
)
2956 splittable_regs_updates
[REGNO (v
->new_reg
)] = 1;
2960 if (loop_dump_stream
)
2964 if (GET_CODE (v
->dest_reg
) == CONST_INT
)
2966 else if (GET_CODE (v
->dest_reg
) != REG
)
2967 regnum
= REGNO (XEXP (v
->dest_reg
, 0));
2969 regnum
= REGNO (v
->dest_reg
);
2970 fprintf (loop_dump_stream
, "Giv %d at insn %d safe to split.\n",
2971 regnum
, INSN_UID (v
->insn
));
2978 /* Try to prove that the register is dead after the loop exits. Trace every
2979 loop exit looking for an insn that will always be executed, which sets
2980 the register to some value, and appears before the first use of the register
2981 is found. If successful, then return 1, otherwise return 0. */
2983 /* ?? Could be made more intelligent in the handling of jumps, so that
2984 it can search past if statements and other similar structures. */
2987 reg_dead_after_loop (reg
, loop_start
, loop_end
)
2988 rtx reg
, loop_start
, loop_end
;
2993 int label_count
= 0;
2994 int this_loop_num
= uid_loop_num
[INSN_UID (loop_start
)];
2996 /* In addition to checking all exits of this loop, we must also check
2997 all exits of inner nested loops that would exit this loop. We don't
2998 have any way to identify those, so we just give up if there are any
2999 such inner loop exits. */
3001 for (label
= loop_number_exit_labels
[this_loop_num
]; label
;
3002 label
= LABEL_NEXTREF (label
))
3005 if (label_count
!= loop_number_exit_count
[this_loop_num
])
3008 /* HACK: Must also search the loop fall through exit, create a label_ref
3009 here which points to the loop_end, and append the loop_number_exit_labels
3011 label
= gen_rtx (LABEL_REF
, VOIDmode
, loop_end
);
3012 LABEL_NEXTREF (label
) = loop_number_exit_labels
[this_loop_num
];
3014 for ( ; label
; label
= LABEL_NEXTREF (label
))
3016 /* Succeed if find an insn which sets the biv or if reach end of
3017 function. Fail if find an insn that uses the biv, or if come to
3018 a conditional jump. */
3020 insn
= NEXT_INSN (XEXP (label
, 0));
3023 code
= GET_CODE (insn
);
3024 if (GET_RTX_CLASS (code
) == 'i')
3028 if (reg_referenced_p (reg
, PATTERN (insn
)))
3031 set
= single_set (insn
);
3032 if (set
&& rtx_equal_p (SET_DEST (set
), reg
))
3036 if (code
== JUMP_INSN
)
3038 if (GET_CODE (PATTERN (insn
)) == RETURN
)
3040 else if (! simplejump_p (insn
)
3041 /* Prevent infinite loop following infinite loops. */
3042 || jump_count
++ > 20)
3045 insn
= JUMP_LABEL (insn
);
3048 insn
= NEXT_INSN (insn
);
3052 /* Success, the register is dead on all loop exits. */
3056 /* Try to calculate the final value of the biv, the value it will have at
3057 the end of the loop. If we can do it, return that value. */
3060 final_biv_value (bl
, loop_start
, loop_end
)
3061 struct iv_class
*bl
;
3062 rtx loop_start
, loop_end
;
3066 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3068 if (GET_MODE_CLASS (bl
->biv
->mode
) != MODE_INT
)
3071 /* The final value for reversed bivs must be calculated differently than
3072 for ordinary bivs. In this case, there is already an insn after the
3073 loop which sets this biv's final value (if necessary), and there are
3074 no other loop exits, so we can return any value. */
3077 if (loop_dump_stream
)
3078 fprintf (loop_dump_stream
,
3079 "Final biv value for %d, reversed biv.\n", bl
->regno
);
3084 /* Try to calculate the final value as initial value + (number of iterations
3085 * increment). For this to work, increment must be invariant, the only
3086 exit from the loop must be the fall through at the bottom (otherwise
3087 it may not have its final value when the loop exits), and the initial
3088 value of the biv must be invariant. */
3090 if (loop_n_iterations
!= 0
3091 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]]
3092 && invariant_p (bl
->initial_value
))
3094 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3096 if (increment
&& invariant_p (increment
))
3098 /* Can calculate the loop exit value, emit insns after loop
3099 end to calculate this value into a temporary register in
3100 case it is needed later. */
3102 tem
= gen_reg_rtx (bl
->biv
->mode
);
3103 /* Make sure loop_end is not the last insn. */
3104 if (NEXT_INSN (loop_end
) == 0)
3105 emit_note_after (NOTE_INSN_DELETED
, loop_end
);
3106 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3107 bl
->initial_value
, tem
, NEXT_INSN (loop_end
));
3109 if (loop_dump_stream
)
3110 fprintf (loop_dump_stream
,
3111 "Final biv value for %d, calculated.\n", bl
->regno
);
3117 /* Check to see if the biv is dead at all loop exits. */
3118 if (reg_dead_after_loop (bl
->biv
->src_reg
, loop_start
, loop_end
))
3120 if (loop_dump_stream
)
3121 fprintf (loop_dump_stream
,
3122 "Final biv value for %d, biv dead after loop exit.\n",
3131 /* Try to calculate the final value of the giv, the value it will have at
3132 the end of the loop. If we can do it, return that value. */
3135 final_giv_value (v
, loop_start
, loop_end
)
3136 struct induction
*v
;
3137 rtx loop_start
, loop_end
;
3139 struct iv_class
*bl
;
3142 rtx insert_before
, seq
;
3144 bl
= reg_biv_class
[REGNO (v
->src_reg
)];
3146 /* The final value for givs which depend on reversed bivs must be calculated
3147 differently than for ordinary givs. In this case, there is already an
3148 insn after the loop which sets this giv's final value (if necessary),
3149 and there are no other loop exits, so we can return any value. */
3152 if (loop_dump_stream
)
3153 fprintf (loop_dump_stream
,
3154 "Final giv value for %d, depends on reversed biv\n",
3155 REGNO (v
->dest_reg
));
3159 /* Try to calculate the final value as a function of the biv it depends
3160 upon. The only exit from the loop must be the fall through at the bottom
3161 (otherwise it may not have its final value when the loop exits). */
3163 /* ??? Can calculate the final giv value by subtracting off the
3164 extra biv increments times the giv's mult_val. The loop must have
3165 only one exit for this to work, but the loop iterations does not need
3168 if (loop_n_iterations
!= 0
3169 && ! loop_number_exit_count
[uid_loop_num
[INSN_UID (loop_start
)]])
3171 /* ?? It is tempting to use the biv's value here since these insns will
3172 be put after the loop, and hence the biv will have its final value
3173 then. However, this fails if the biv is subsequently eliminated.
3174 Perhaps determine whether biv's are eliminable before trying to
3175 determine whether giv's are replaceable so that we can use the
3176 biv value here if it is not eliminable. */
3178 increment
= biv_total_increment (bl
, loop_start
, loop_end
);
3180 if (increment
&& invariant_p (increment
))
3182 /* Can calculate the loop exit value of its biv as
3183 (loop_n_iterations * increment) + initial_value */
3185 /* The loop exit value of the giv is then
3186 (final_biv_value - extra increments) * mult_val + add_val.
3187 The extra increments are any increments to the biv which
3188 occur in the loop after the giv's value is calculated.
3189 We must search from the insn that sets the giv to the end
3190 of the loop to calculate this value. */
3192 insert_before
= NEXT_INSN (loop_end
);
3194 /* Put the final biv value in tem. */
3195 tem
= gen_reg_rtx (bl
->biv
->mode
);
3196 emit_iv_add_mult (increment
, GEN_INT (loop_n_iterations
),
3197 bl
->initial_value
, tem
, insert_before
);
3199 /* Subtract off extra increments as we find them. */
3200 for (insn
= NEXT_INSN (v
->insn
); insn
!= loop_end
;
3201 insn
= NEXT_INSN (insn
))
3203 struct induction
*biv
;
3205 for (biv
= bl
->biv
; biv
; biv
= biv
->next_iv
)
3206 if (biv
->insn
== insn
)
3209 tem
= expand_binop (GET_MODE (tem
), sub_optab
, tem
,
3210 biv
->add_val
, NULL_RTX
, 0,
3212 seq
= gen_sequence ();
3214 emit_insn_before (seq
, insert_before
);
3218 /* Now calculate the giv's final value. */
3219 emit_iv_add_mult (tem
, v
->mult_val
, v
->add_val
, tem
,
3222 if (loop_dump_stream
)
3223 fprintf (loop_dump_stream
,
3224 "Final giv value for %d, calc from biv's value.\n",
3225 REGNO (v
->dest_reg
));
3231 /* Replaceable giv's should never reach here. */
3235 /* Check to see if the biv is dead at all loop exits. */
3236 if (reg_dead_after_loop (v
->dest_reg
, loop_start
, loop_end
))
3238 if (loop_dump_stream
)
3239 fprintf (loop_dump_stream
,
3240 "Final giv value for %d, giv dead after loop exit.\n",
3241 REGNO (v
->dest_reg
));
3250 /* Calculate the number of loop iterations. Returns the exact number of loop
3251 iterations if it can be calculated, otherwise returns zero. */
3253 unsigned HOST_WIDE_INT
3254 loop_iterations (loop_start
, loop_end
)
3255 rtx loop_start
, loop_end
;
3257 rtx comparison
, comparison_value
;
3258 rtx iteration_var
, initial_value
, increment
, final_value
;
3259 enum rtx_code comparison_code
;
3262 int unsigned_compare
, compare_dir
, final_larger
;
3263 unsigned long tempu
;
3266 /* First find the iteration variable. If the last insn is a conditional
3267 branch, and the insn before tests a register value, make that the
3268 iteration variable. */
3270 loop_initial_value
= 0;
3272 loop_final_value
= 0;
3273 loop_iteration_var
= 0;
3275 /* We used to use pren_nonnote_insn here, but that fails because it might
3276 accidentally get the branch for a contained loop if the branch for this
3277 loop was deleted. We can only trust branches immediately before the
3279 last_loop_insn
= PREV_INSN (loop_end
);
3281 comparison
= get_condition_for_loop (last_loop_insn
);
3282 if (comparison
== 0)
3284 if (loop_dump_stream
)
3285 fprintf (loop_dump_stream
,
3286 "Loop unrolling: No final conditional branch found.\n");
3290 /* ??? Get_condition may switch position of induction variable and
3291 invariant register when it canonicalizes the comparison. */
3293 comparison_code
= GET_CODE (comparison
);
3294 iteration_var
= XEXP (comparison
, 0);
3295 comparison_value
= XEXP (comparison
, 1);
3297 if (GET_CODE (iteration_var
) != REG
)
3299 if (loop_dump_stream
)
3300 fprintf (loop_dump_stream
,
3301 "Loop unrolling: Comparison not against register.\n");
3305 /* Loop iterations is always called before any new registers are created
3306 now, so this should never occur. */
3308 if (REGNO (iteration_var
) >= max_reg_before_loop
)
3311 iteration_info (iteration_var
, &initial_value
, &increment
,
3312 loop_start
, loop_end
);
3313 if (initial_value
== 0)
3314 /* iteration_info already printed a message. */
3317 /* If the comparison value is an invariant register, then try to find
3318 its value from the insns before the start of the loop. */
3320 if (GET_CODE (comparison_value
) == REG
&& invariant_p (comparison_value
))
3324 for (insn
= PREV_INSN (loop_start
); insn
; insn
= PREV_INSN (insn
))
3326 if (GET_CODE (insn
) == CODE_LABEL
)
3329 else if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i'
3330 && reg_set_p (comparison_value
, insn
))
3332 /* We found the last insn before the loop that sets the register.
3333 If it sets the entire register, and has a REG_EQUAL note,
3334 then use the value of the REG_EQUAL note. */
3335 if ((set
= single_set (insn
))
3336 && (SET_DEST (set
) == comparison_value
))
3338 rtx note
= find_reg_note (insn
, REG_EQUAL
, NULL_RTX
);
3340 /* Only use the REG_EQUAL note if it is a constant.
3341 Other things, divide in particular, will cause
3342 problems later if we use them. */
3343 if (note
&& GET_CODE (XEXP (note
, 0)) != EXPR_LIST
3344 && CONSTANT_P (XEXP (note
, 0)))
3345 comparison_value
= XEXP (note
, 0);
3352 final_value
= approx_final_value (comparison_code
, comparison_value
,
3353 &unsigned_compare
, &compare_dir
);
3355 /* Save the calculated values describing this loop's bounds, in case
3356 precondition_loop_p will need them later. These values can not be
3357 recalculated inside precondition_loop_p because strength reduction
3358 optimizations may obscure the loop's structure. */
3360 loop_iteration_var
= iteration_var
;
3361 loop_initial_value
= initial_value
;
3362 loop_increment
= increment
;
3363 loop_final_value
= final_value
;
3364 loop_comparison_code
= comparison_code
;
3368 if (loop_dump_stream
)
3369 fprintf (loop_dump_stream
,
3370 "Loop unrolling: Increment value can't be calculated.\n");
3373 else if (GET_CODE (increment
) != CONST_INT
)
3375 if (loop_dump_stream
)
3376 fprintf (loop_dump_stream
,
3377 "Loop unrolling: Increment value not constant.\n");
3380 else if (GET_CODE (initial_value
) != CONST_INT
)
3382 if (loop_dump_stream
)
3383 fprintf (loop_dump_stream
,
3384 "Loop unrolling: Initial value not constant.\n");
3387 else if (final_value
== 0)
3389 if (loop_dump_stream
)
3390 fprintf (loop_dump_stream
,
3391 "Loop unrolling: EQ comparison loop.\n");
3394 else if (GET_CODE (final_value
) != CONST_INT
)
3396 if (loop_dump_stream
)
3397 fprintf (loop_dump_stream
,
3398 "Loop unrolling: Final value not constant.\n");
3402 /* ?? Final value and initial value do not have to be constants.
3403 Only their difference has to be constant. When the iteration variable
3404 is an array address, the final value and initial value might both
3405 be addresses with the same base but different constant offsets.
3406 Final value must be invariant for this to work.
3408 To do this, need some way to find the values of registers which are
3411 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3412 if (unsigned_compare
)
3414 = ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3415 > (unsigned HOST_WIDE_INT
) INTVAL (initial_value
))
3416 - ((unsigned HOST_WIDE_INT
) INTVAL (final_value
)
3417 < (unsigned HOST_WIDE_INT
) INTVAL (initial_value
));
3419 final_larger
= (INTVAL (final_value
) > INTVAL (initial_value
))
3420 - (INTVAL (final_value
) < INTVAL (initial_value
));
3422 if (INTVAL (increment
) > 0)
3424 else if (INTVAL (increment
) == 0)
3429 /* There are 27 different cases: compare_dir = -1, 0, 1;
3430 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3431 There are 4 normal cases, 4 reverse cases (where the iteration variable
3432 will overflow before the loop exits), 4 infinite loop cases, and 15
3433 immediate exit (0 or 1 iteration depending on loop type) cases.
3434 Only try to optimize the normal cases. */
3436 /* (compare_dir/final_larger/increment_dir)
3437 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3438 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3439 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3440 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3442 /* ?? If the meaning of reverse loops (where the iteration variable
3443 will overflow before the loop exits) is undefined, then could
3444 eliminate all of these special checks, and just always assume
3445 the loops are normal/immediate/infinite. Note that this means
3446 the sign of increment_dir does not have to be known. Also,
3447 since it does not really hurt if immediate exit loops or infinite loops
3448 are optimized, then that case could be ignored also, and hence all
3449 loops can be optimized.
3451 According to ANSI Spec, the reverse loop case result is undefined,
3452 because the action on overflow is undefined.
3454 See also the special test for NE loops below. */
3456 if (final_larger
== increment_dir
&& final_larger
!= 0
3457 && (final_larger
== compare_dir
|| compare_dir
== 0))
3462 if (loop_dump_stream
)
3463 fprintf (loop_dump_stream
,
3464 "Loop unrolling: Not normal loop.\n");
3468 /* Calculate the number of iterations, final_value is only an approximation,
3469 so correct for that. Note that tempu and loop_n_iterations are
3470 unsigned, because they can be as large as 2^n - 1. */
3472 i
= INTVAL (increment
);
3474 tempu
= INTVAL (final_value
) - INTVAL (initial_value
);
3477 tempu
= INTVAL (initial_value
) - INTVAL (final_value
);
3483 /* For NE tests, make sure that the iteration variable won't miss the
3484 final value. If tempu mod i is not zero, then the iteration variable
3485 will overflow before the loop exits, and we can not calculate the
3486 number of iterations. */
3487 if (compare_dir
== 0 && (tempu
% i
) != 0)
3490 return tempu
/ i
+ ((tempu
% i
) != 0);
3493 /* Replace uses of split bivs with their split pseudo register. This is
3494 for original instructions which remain after loop unrolling without
3498 remap_split_bivs (x
)
3501 register enum rtx_code code
;
3508 code
= GET_CODE (x
);
3523 /* If non-reduced/final-value givs were split, then this would also
3524 have to remap those givs also. */
3526 if (REGNO (x
) < max_reg_before_loop
3527 && reg_iv_type
[REGNO (x
)] == BASIC_INDUCT
)
3528 return reg_biv_class
[REGNO (x
)]->biv
->src_reg
;
3531 fmt
= GET_RTX_FORMAT (code
);
3532 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
3535 XEXP (x
, i
) = remap_split_bivs (XEXP (x
, i
));
3539 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3540 XVECEXP (x
, i
, j
) = remap_split_bivs (XVECEXP (x
, i
, j
));
3546 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3547 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3548 return 0. COPY_START is where we can start looking for the insns
3549 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3552 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3553 must dominate LAST_UID.
3555 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3556 may not dominate LAST_UID.
3558 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3559 must dominate LAST_UID. */
3562 set_dominates_use (regno
, first_uid
, last_uid
, copy_start
, copy_end
)
3569 int passed_jump
= 0;
3570 rtx p
= NEXT_INSN (copy_start
);
3572 while (INSN_UID (p
) != first_uid
)
3574 if (GET_CODE (p
) == JUMP_INSN
)
3576 /* Could not find FIRST_UID. */
3582 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3583 if (GET_RTX_CLASS (GET_CODE (p
)) != 'i'
3584 || ! dead_or_set_regno_p (p
, regno
))
3587 /* FIRST_UID is always executed. */
3588 if (passed_jump
== 0)
3591 while (INSN_UID (p
) != last_uid
)
3593 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3594 can not be sure that FIRST_UID dominates LAST_UID. */
3595 if (GET_CODE (p
) == CODE_LABEL
)
3600 /* FIRST_UID is always executed if LAST_UID is executed. */