1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 1998 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 };
152 #include "insn-config.h"
153 #include "integrate.h"
160 /* This controls which loops are unrolled, and by how much we unroll
163 #ifndef MAX_UNROLLED_INSNS
164 #define MAX_UNROLLED_INSNS 100
167 /* Indexed by register number, if non-zero, then it contains a pointer
168 to a struct induction for a DEST_REG giv which has been combined with
169 one of more address givs. This is needed because whenever such a DEST_REG
170 giv is modified, we must modify the value of all split address givs
171 that were combined with this DEST_REG giv. */
173 static struct induction **addr_combined_regs;
175 /* Indexed by register number, if this is a splittable induction variable,
176 then this will hold the current value of the register, which depends on the
179 static rtx *splittable_regs;
181 /* Indexed by register number, if this is a splittable induction variable,
182 then this will hold the number of instructions in the loop that modify
183 the induction variable. Used to ensure that only the last insn modifying
184 a split iv will update the original iv of the dest. */
186 static int *splittable_regs_updates;
188 /* Values describing the current loop's iteration variable. These are set up
189 by loop_iterations, and used by precondition_loop_p. */
191 static rtx loop_iteration_var;
192 static rtx loop_initial_value;
193 static rtx loop_increment;
194 static rtx loop_final_value;
195 static enum rtx_code loop_comparison_code;
197 /* Forward declarations. */
199 static void init_reg_map PROTO((struct inline_remap *, int));
200 static int precondition_loop_p PROTO((rtx *, rtx *, rtx *, rtx, rtx));
201 static rtx calculate_giv_inc PROTO((rtx, rtx, int));
202 static rtx initial_reg_note_copy PROTO((rtx, struct inline_remap *));
203 static void final_reg_note_copy PROTO((rtx, struct inline_remap *));
204 static void copy_loop_body PROTO((rtx, rtx, struct inline_remap *, rtx, int,
205 enum unroll_types, rtx, rtx, rtx, rtx));
206 void iteration_info PROTO((rtx, rtx *, rtx *, rtx, rtx));
207 static rtx approx_final_value PROTO((enum rtx_code, rtx, int *, int *));
208 static int find_splittable_regs PROTO((enum unroll_types, rtx, rtx, rtx, int));
209 static int find_splittable_givs PROTO((struct iv_class *,enum unroll_types,
210 rtx, rtx, rtx, int));
211 static int reg_dead_after_loop PROTO((rtx, rtx, rtx));
212 static rtx fold_rtx_mult_add PROTO((rtx, rtx, rtx, enum machine_mode));
213 static rtx remap_split_bivs PROTO((rtx));
215 /* Try to unroll one loop and split induction variables in the loop.
217 The loop is described by the arguments LOOP_END, INSN_COUNT, and
218 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
219 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
220 indicates whether information generated in the strength reduction pass
223 This function is intended to be called from within `strength_reduce'
227 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
232 rtx end_insert_before;
233 int strength_reduce_p;
236 int unroll_number = 1;
237 rtx copy_start, copy_end;
238 rtx insn, sequence, pattern, tem;
239 int max_labelno, max_insnno;
241 struct inline_remap *map;
249 int splitting_not_safe = 0;
250 enum unroll_types unroll_type;
251 int loop_preconditioned = 0;
253 /* This points to the last real insn in the loop, which should be either
254 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
258 /* Don't bother unrolling huge loops. Since the minimum factor is
259 two, loops greater than one half of MAX_UNROLLED_INSNS will never
261 if (insn_count > MAX_UNROLLED_INSNS / 2)
263 if (loop_dump_stream)
264 fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
268 /* When emitting debugger info, we can't unroll loops with unequal numbers
269 of block_beg and block_end notes, because that would unbalance the block
270 structure of the function. This can happen as a result of the
271 "if (foo) bar; else break;" optimization in jump.c. */
272 /* ??? Gcc has a general policy that -g is never supposed to change the code
273 that the compiler emits, so we must disable this optimization always,
274 even if debug info is not being output. This is rare, so this should
275 not be a significant performance problem. */
277 if (1 /* write_symbols != NO_DEBUG */)
279 int block_begins = 0;
282 for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn))
284 if (GET_CODE (insn) == NOTE)
286 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG)
288 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END)
293 if (block_begins != block_ends)
295 if (loop_dump_stream)
296 fprintf (loop_dump_stream,
297 "Unrolling failure: Unbalanced block notes.\n");
302 /* Determine type of unroll to perform. Depends on the number of iterations
303 and the size of the loop. */
305 /* If there is no strength reduce info, then set loop_n_iterations to zero.
306 This can happen if strength_reduce can't find any bivs in the loop.
307 A value of zero indicates that the number of iterations could not be
310 if (! strength_reduce_p)
311 loop_n_iterations = 0;
313 if (loop_dump_stream && loop_n_iterations > 0)
314 fprintf (loop_dump_stream,
315 "Loop unrolling: %d iterations.\n", loop_n_iterations);
317 /* Find and save a pointer to the last nonnote insn in the loop. */
319 last_loop_insn = prev_nonnote_insn (loop_end);
321 /* Calculate how many times to unroll the loop. Indicate whether or
322 not the loop is being completely unrolled. */
324 if (loop_n_iterations == 1)
326 /* If number of iterations is exactly 1, then eliminate the compare and
327 branch at the end of the loop since they will never be taken.
328 Then return, since no other action is needed here. */
330 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
331 don't do anything. */
333 if (GET_CODE (last_loop_insn) == BARRIER)
335 /* Delete the jump insn. This will delete the barrier also. */
336 delete_insn (PREV_INSN (last_loop_insn));
338 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
341 /* The immediately preceding insn is a compare which must be
343 delete_insn (last_loop_insn);
344 delete_insn (PREV_INSN (last_loop_insn));
346 /* The immediately preceding insn may not be the compare, so don't
348 delete_insn (last_loop_insn);
353 else if (loop_n_iterations > 0
354 && loop_n_iterations * insn_count < MAX_UNROLLED_INSNS)
356 unroll_number = loop_n_iterations;
357 unroll_type = UNROLL_COMPLETELY;
359 else if (loop_n_iterations > 0)
361 /* Try to factor the number of iterations. Don't bother with the
362 general case, only using 2, 3, 5, and 7 will get 75% of all
363 numbers theoretically, and almost all in practice. */
365 for (i = 0; i < NUM_FACTORS; i++)
366 factors[i].count = 0;
368 temp = loop_n_iterations;
369 for (i = NUM_FACTORS - 1; i >= 0; i--)
370 while (temp % factors[i].factor == 0)
373 temp = temp / factors[i].factor;
376 /* Start with the larger factors first so that we generally
377 get lots of unrolling. */
381 for (i = 3; i >= 0; i--)
382 while (factors[i].count--)
384 if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
386 unroll_number *= factors[i].factor;
387 temp *= factors[i].factor;
393 /* If we couldn't find any factors, then unroll as in the normal
395 if (unroll_number == 1)
397 if (loop_dump_stream)
398 fprintf (loop_dump_stream,
399 "Loop unrolling: No factors found.\n");
402 unroll_type = UNROLL_MODULO;
406 /* Default case, calculate number of times to unroll loop based on its
408 if (unroll_number == 1)
410 if (8 * insn_count < MAX_UNROLLED_INSNS)
412 else if (4 * insn_count < MAX_UNROLLED_INSNS)
417 unroll_type = UNROLL_NAIVE;
420 /* Now we know how many times to unroll the loop. */
422 if (loop_dump_stream)
423 fprintf (loop_dump_stream,
424 "Unrolling loop %d times.\n", unroll_number);
427 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
429 /* Loops of these types should never start with a jump down to
430 the exit condition test. For now, check for this case just to
431 be sure. UNROLL_NAIVE loops can be of this form, this case is
434 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
435 insn = NEXT_INSN (insn);
436 if (GET_CODE (insn) == JUMP_INSN)
440 if (unroll_type == UNROLL_COMPLETELY)
442 /* Completely unrolling the loop: Delete the compare and branch at
443 the end (the last two instructions). This delete must done at the
444 very end of loop unrolling, to avoid problems with calls to
445 back_branch_in_range_p, which is called by find_splittable_regs.
446 All increments of splittable bivs/givs are changed to load constant
449 copy_start = loop_start;
451 /* Set insert_before to the instruction immediately after the JUMP_INSN
452 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
453 the loop will be correctly handled by copy_loop_body. */
454 insert_before = NEXT_INSN (last_loop_insn);
456 /* Set copy_end to the insn before the jump at the end of the loop. */
457 if (GET_CODE (last_loop_insn) == BARRIER)
458 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
459 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
462 /* The instruction immediately before the JUMP_INSN is a compare
463 instruction which we do not want to copy. */
464 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
466 /* The instruction immediately before the JUMP_INSN may not be the
467 compare, so we must copy it. */
468 copy_end = PREV_INSN (last_loop_insn);
473 /* We currently can't unroll a loop if it doesn't end with a
474 JUMP_INSN. There would need to be a mechanism that recognizes
475 this case, and then inserts a jump after each loop body, which
476 jumps to after the last loop body. */
477 if (loop_dump_stream)
478 fprintf (loop_dump_stream,
479 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
483 else if (unroll_type == UNROLL_MODULO)
485 /* Partially unrolling the loop: The compare and branch at the end
486 (the last two instructions) must remain. Don't copy the compare
487 and branch instructions at the end of the loop. Insert the unrolled
488 code immediately before the compare/branch at the end so that the
489 code will fall through to them as before. */
491 copy_start = loop_start;
493 /* Set insert_before to the jump insn at the end of the loop.
494 Set copy_end to before the jump insn at the end of the loop. */
495 if (GET_CODE (last_loop_insn) == BARRIER)
497 insert_before = PREV_INSN (last_loop_insn);
498 copy_end = PREV_INSN (insert_before);
500 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
503 /* The instruction immediately before the JUMP_INSN is a compare
504 instruction which we do not want to copy or delete. */
505 insert_before = PREV_INSN (last_loop_insn);
506 copy_end = PREV_INSN (insert_before);
508 /* The instruction immediately before the JUMP_INSN may not be the
509 compare, so we must copy it. */
510 insert_before = last_loop_insn;
511 copy_end = PREV_INSN (last_loop_insn);
516 /* We currently can't unroll a loop if it doesn't end with a
517 JUMP_INSN. There would need to be a mechanism that recognizes
518 this case, and then inserts a jump after each loop body, which
519 jumps to after the last loop body. */
520 if (loop_dump_stream)
521 fprintf (loop_dump_stream,
522 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
528 /* Normal case: Must copy the compare and branch instructions at the
531 if (GET_CODE (last_loop_insn) == BARRIER)
533 /* Loop ends with an unconditional jump and a barrier.
534 Handle this like above, don't copy jump and barrier.
535 This is not strictly necessary, but doing so prevents generating
536 unconditional jumps to an immediately following label.
538 This will be corrected below if the target of this jump is
539 not the start_label. */
541 insert_before = PREV_INSN (last_loop_insn);
542 copy_end = PREV_INSN (insert_before);
544 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
546 /* Set insert_before to immediately after the JUMP_INSN, so that
547 NOTEs at the end of the loop will be correctly handled by
549 insert_before = NEXT_INSN (last_loop_insn);
550 copy_end = last_loop_insn;
554 /* We currently can't unroll a loop if it doesn't end with a
555 JUMP_INSN. There would need to be a mechanism that recognizes
556 this case, and then inserts a jump after each loop body, which
557 jumps to after the last loop body. */
558 if (loop_dump_stream)
559 fprintf (loop_dump_stream,
560 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
564 /* If copying exit test branches because they can not be eliminated,
565 then must convert the fall through case of the branch to a jump past
566 the end of the loop. Create a label to emit after the loop and save
567 it for later use. Do not use the label after the loop, if any, since
568 it might be used by insns outside the loop, or there might be insns
569 added before it later by final_[bg]iv_value which must be after
570 the real exit label. */
571 exit_label = gen_label_rtx ();
574 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
575 insn = NEXT_INSN (insn);
577 if (GET_CODE (insn) == JUMP_INSN)
579 /* The loop starts with a jump down to the exit condition test.
580 Start copying the loop after the barrier following this
582 copy_start = NEXT_INSN (insn);
584 /* Splitting induction variables doesn't work when the loop is
585 entered via a jump to the bottom, because then we end up doing
586 a comparison against a new register for a split variable, but
587 we did not execute the set insn for the new register because
588 it was skipped over. */
589 splitting_not_safe = 1;
590 if (loop_dump_stream)
591 fprintf (loop_dump_stream,
592 "Splitting not safe, because loop not entered at top.\n");
595 copy_start = loop_start;
598 /* This should always be the first label in the loop. */
599 start_label = NEXT_INSN (copy_start);
600 /* There may be a line number note and/or a loop continue note here. */
601 while (GET_CODE (start_label) == NOTE)
602 start_label = NEXT_INSN (start_label);
603 if (GET_CODE (start_label) != CODE_LABEL)
605 /* This can happen as a result of jump threading. If the first insns in
606 the loop test the same condition as the loop's backward jump, or the
607 opposite condition, then the backward jump will be modified to point
608 to elsewhere, and the loop's start label is deleted.
610 This case currently can not be handled by the loop unrolling code. */
612 if (loop_dump_stream)
613 fprintf (loop_dump_stream,
614 "Unrolling failure: unknown insns between BEG note and loop label.\n");
617 if (LABEL_NAME (start_label))
619 /* The jump optimization pass must have combined the original start label
620 with a named label for a goto. We can't unroll this case because
621 jumps which go to the named label must be handled differently than
622 jumps to the loop start, and it is impossible to differentiate them
624 if (loop_dump_stream)
625 fprintf (loop_dump_stream,
626 "Unrolling failure: loop start label is gone\n");
630 if (unroll_type == UNROLL_NAIVE
631 && GET_CODE (last_loop_insn) == BARRIER
632 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
634 /* In this case, we must copy the jump and barrier, because they will
635 not be converted to jumps to an immediately following label. */
637 insert_before = NEXT_INSN (last_loop_insn);
638 copy_end = last_loop_insn;
641 if (unroll_type == UNROLL_NAIVE
642 && GET_CODE (last_loop_insn) == JUMP_INSN
643 && start_label != JUMP_LABEL (last_loop_insn))
645 /* ??? The loop ends with a conditional branch that does not branch back
646 to the loop start label. In this case, we must emit an unconditional
647 branch to the loop exit after emitting the final branch.
648 copy_loop_body does not have support for this currently, so we
649 give up. It doesn't seem worthwhile to unroll anyways since
650 unrolling would increase the number of branch instructions
652 if (loop_dump_stream)
653 fprintf (loop_dump_stream,
654 "Unrolling failure: final conditional branch not to loop start\n");
658 /* Allocate a translation table for the labels and insn numbers.
659 They will be filled in as we copy the insns in the loop. */
661 max_labelno = max_label_num ();
662 max_insnno = get_max_uid ();
664 map = (struct inline_remap *) alloca (sizeof (struct inline_remap));
666 map->integrating = 0;
668 /* Allocate the label map. */
672 map->label_map = (rtx *) alloca (max_labelno * sizeof (rtx));
674 local_label = (char *) alloca (max_labelno);
675 bzero (local_label, max_labelno);
680 /* Search the loop and mark all local labels, i.e. the ones which have to
681 be distinct labels when copied. For all labels which might be
682 non-local, set their label_map entries to point to themselves.
683 If they happen to be local their label_map entries will be overwritten
684 before the loop body is copied. The label_map entries for local labels
685 will be set to a different value each time the loop body is copied. */
687 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
689 if (GET_CODE (insn) == CODE_LABEL)
690 local_label[CODE_LABEL_NUMBER (insn)] = 1;
691 else if (GET_CODE (insn) == JUMP_INSN)
693 if (JUMP_LABEL (insn))
694 set_label_in_map (map,
695 CODE_LABEL_NUMBER (JUMP_LABEL (insn)),
697 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
698 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
700 rtx pat = PATTERN (insn);
701 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
702 int len = XVECLEN (pat, diff_vec_p);
705 for (i = 0; i < len; i++)
707 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
708 set_label_in_map (map,
709 CODE_LABEL_NUMBER (label),
716 /* Allocate space for the insn map. */
718 map->insn_map = (rtx *) alloca (max_insnno * sizeof (rtx));
720 /* Set this to zero, to indicate that we are doing loop unrolling,
721 not function inlining. */
722 map->inline_target = 0;
724 /* The register and constant maps depend on the number of registers
725 present, so the final maps can't be created until after
726 find_splittable_regs is called. However, they are needed for
727 preconditioning, so we create temporary maps when preconditioning
730 /* The preconditioning code may allocate two new pseudo registers. */
731 maxregnum = max_reg_num ();
733 /* Allocate and zero out the splittable_regs and addr_combined_regs
734 arrays. These must be zeroed here because they will be used if
735 loop preconditioning is performed, and must be zero for that case.
737 It is safe to do this here, since the extra registers created by the
738 preconditioning code and find_splittable_regs will never be used
739 to access the splittable_regs[] and addr_combined_regs[] arrays. */
741 splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx));
742 bzero ((char *) splittable_regs, maxregnum * sizeof (rtx));
743 splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int));
744 bzero ((char *) splittable_regs_updates, maxregnum * sizeof (int));
746 = (struct induction **) alloca (maxregnum * sizeof (struct induction *));
747 bzero ((char *) addr_combined_regs, maxregnum * sizeof (struct induction *));
748 /* We must limit it to max_reg_before_loop, because only these pseudo
749 registers have valid regno_first_uid info. Any register created after
750 that is unlikely to be local to the loop anyways. */
751 local_regno = (char *) alloca (max_reg_before_loop);
752 bzero (local_regno, max_reg_before_loop);
754 /* Mark all local registers, i.e. the ones which are referenced only
756 if (INSN_UID (copy_end) < max_uid_for_loop)
758 int copy_start_luid = INSN_LUID (copy_start);
759 int copy_end_luid = INSN_LUID (copy_end);
761 /* If a register is used in the jump insn, we must not duplicate it
762 since it will also be used outside the loop. */
763 if (GET_CODE (copy_end) == JUMP_INSN)
765 /* If copy_start points to the NOTE that starts the loop, then we must
766 use the next luid, because invariant pseudo-regs moved out of the loop
767 have their lifetimes modified to start here, but they are not safe
769 if (copy_start == loop_start)
772 /* If a pseudo's lifetime is entirely contained within this loop, then we
773 can use a different pseudo in each unrolled copy of the loop. This
774 results in better code. */
775 for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; ++j)
776 if (REGNO_FIRST_UID (j) > 0 && REGNO_FIRST_UID (j) <= max_uid_for_loop
777 && uid_luid[REGNO_FIRST_UID (j)] >= copy_start_luid
778 && REGNO_LAST_UID (j) > 0 && REGNO_LAST_UID (j) <= max_uid_for_loop
779 && uid_luid[REGNO_LAST_UID (j)] <= copy_end_luid)
781 /* However, we must also check for loop-carried dependencies.
782 If the value the pseudo has at the end of iteration X is
783 used by iteration X+1, then we can not use a different pseudo
784 for each unrolled copy of the loop. */
785 /* A pseudo is safe if regno_first_uid is a set, and this
786 set dominates all instructions from regno_first_uid to
788 /* ??? This check is simplistic. We would get better code if
789 this check was more sophisticated. */
790 if (set_dominates_use (j, REGNO_FIRST_UID (j), REGNO_LAST_UID (j),
791 copy_start, copy_end))
794 if (loop_dump_stream)
797 fprintf (loop_dump_stream, "Marked reg %d as local\n", j);
799 fprintf (loop_dump_stream, "Did not mark reg %d as local\n",
805 /* If this loop requires exit tests when unrolled, check to see if we
806 can precondition the loop so as to make the exit tests unnecessary.
807 Just like variable splitting, this is not safe if the loop is entered
808 via a jump to the bottom. Also, can not do this if no strength
809 reduce info, because precondition_loop_p uses this info. */
811 /* Must copy the loop body for preconditioning before the following
812 find_splittable_regs call since that will emit insns which need to
813 be after the preconditioned loop copies, but immediately before the
814 unrolled loop copies. */
816 /* Also, it is not safe to split induction variables for the preconditioned
817 copies of the loop body. If we split induction variables, then the code
818 assumes that each induction variable can be represented as a function
819 of its initial value and the loop iteration number. This is not true
820 in this case, because the last preconditioned copy of the loop body
821 could be any iteration from the first up to the `unroll_number-1'th,
822 depending on the initial value of the iteration variable. Therefore
823 we can not split induction variables here, because we can not calculate
824 their value. Hence, this code must occur before find_splittable_regs
827 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
829 rtx initial_value, final_value, increment;
831 if (precondition_loop_p (&initial_value, &final_value, &increment,
832 loop_start, loop_end))
835 enum machine_mode mode;
837 int abs_inc, neg_inc;
839 map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
841 map->const_equiv_map = (rtx *) alloca (maxregnum * sizeof (rtx));
842 map->const_age_map = (unsigned *) alloca (maxregnum
843 * sizeof (unsigned));
844 map->const_equiv_map_size = maxregnum;
845 global_const_equiv_map = map->const_equiv_map;
846 global_const_equiv_map_size = maxregnum;
848 init_reg_map (map, maxregnum);
850 /* Limit loop unrolling to 4, since this will make 7 copies of
852 if (unroll_number > 4)
855 /* Save the absolute value of the increment, and also whether or
856 not it is negative. */
858 abs_inc = INTVAL (increment);
867 /* Decide what mode to do these calculations in. Choose the larger
868 of final_value's mode and initial_value's mode, or a full-word if
869 both are constants. */
870 mode = GET_MODE (final_value);
871 if (mode == VOIDmode)
873 mode = GET_MODE (initial_value);
874 if (mode == VOIDmode)
877 else if (mode != GET_MODE (initial_value)
878 && (GET_MODE_SIZE (mode)
879 < GET_MODE_SIZE (GET_MODE (initial_value))))
880 mode = GET_MODE (initial_value);
882 /* Calculate the difference between the final and initial values.
883 Final value may be a (plus (reg x) (const_int 1)) rtx.
884 Let the following cse pass simplify this if initial value is
887 We must copy the final and initial values here to avoid
888 improperly shared rtl. */
890 diff = expand_binop (mode, sub_optab, copy_rtx (final_value),
891 copy_rtx (initial_value), NULL_RTX, 0,
894 /* Now calculate (diff % (unroll * abs (increment))) by using an
896 diff = expand_binop (GET_MODE (diff), and_optab, diff,
897 GEN_INT (unroll_number * abs_inc - 1),
898 NULL_RTX, 0, OPTAB_LIB_WIDEN);
900 /* Now emit a sequence of branches to jump to the proper precond
903 labels = (rtx *) alloca (sizeof (rtx) * unroll_number);
904 for (i = 0; i < unroll_number; i++)
905 labels[i] = gen_label_rtx ();
907 /* Check for the case where the initial value is greater than or
908 equal to the final value. In that case, we want to execute
909 exactly one loop iteration. The code below will fail for this
910 case. This check does not apply if the loop has a NE
911 comparison at the end. */
913 if (loop_comparison_code != NE)
915 emit_cmp_insn (initial_value, final_value, neg_inc ? LE : GE,
916 NULL_RTX, mode, 0, 0);
918 emit_jump_insn (gen_ble (labels[1]));
920 emit_jump_insn (gen_bge (labels[1]));
921 JUMP_LABEL (get_last_insn ()) = labels[1];
922 LABEL_NUSES (labels[1])++;
925 /* Assuming the unroll_number is 4, and the increment is 2, then
926 for a negative increment: for a positive increment:
927 diff = 0,1 precond 0 diff = 0,7 precond 0
928 diff = 2,3 precond 3 diff = 1,2 precond 1
929 diff = 4,5 precond 2 diff = 3,4 precond 2
930 diff = 6,7 precond 1 diff = 5,6 precond 3 */
932 /* We only need to emit (unroll_number - 1) branches here, the
933 last case just falls through to the following code. */
935 /* ??? This would give better code if we emitted a tree of branches
936 instead of the current linear list of branches. */
938 for (i = 0; i < unroll_number - 1; i++)
941 enum rtx_code cmp_code;
943 /* For negative increments, must invert the constant compared
944 against, except when comparing against zero. */
952 cmp_const = unroll_number - i;
961 emit_cmp_insn (diff, GEN_INT (abs_inc * cmp_const),
962 cmp_code, NULL_RTX, mode, 0, 0);
965 emit_jump_insn (gen_beq (labels[i]));
967 emit_jump_insn (gen_bge (labels[i]));
969 emit_jump_insn (gen_ble (labels[i]));
970 JUMP_LABEL (get_last_insn ()) = labels[i];
971 LABEL_NUSES (labels[i])++;
974 /* If the increment is greater than one, then we need another branch,
975 to handle other cases equivalent to 0. */
977 /* ??? This should be merged into the code above somehow to help
978 simplify the code here, and reduce the number of branches emitted.
979 For the negative increment case, the branch here could easily
980 be merged with the `0' case branch above. For the positive
981 increment case, it is not clear how this can be simplified. */
986 enum rtx_code cmp_code;
990 cmp_const = abs_inc - 1;
995 cmp_const = abs_inc * (unroll_number - 1) + 1;
999 emit_cmp_insn (diff, GEN_INT (cmp_const), cmp_code, NULL_RTX,
1003 emit_jump_insn (gen_ble (labels[0]));
1005 emit_jump_insn (gen_bge (labels[0]));
1006 JUMP_LABEL (get_last_insn ()) = labels[0];
1007 LABEL_NUSES (labels[0])++;
1010 sequence = gen_sequence ();
1012 emit_insn_before (sequence, loop_start);
1014 /* Only the last copy of the loop body here needs the exit
1015 test, so set copy_end to exclude the compare/branch here,
1016 and then reset it inside the loop when get to the last
1019 if (GET_CODE (last_loop_insn) == BARRIER)
1020 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1021 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
1024 /* The immediately preceding insn is a compare which we do not
1026 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1028 /* The immediately preceding insn may not be a compare, so we
1030 copy_end = PREV_INSN (last_loop_insn);
1036 for (i = 1; i < unroll_number; i++)
1038 emit_label_after (labels[unroll_number - i],
1039 PREV_INSN (loop_start));
1041 bzero ((char *) map->insn_map, max_insnno * sizeof (rtx));
1042 bzero ((char *) map->const_equiv_map, maxregnum * sizeof (rtx));
1043 bzero ((char *) map->const_age_map,
1044 maxregnum * sizeof (unsigned));
1047 for (j = 0; j < max_labelno; j++)
1049 set_label_in_map (map, j, gen_label_rtx ());
1051 for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++)
1054 map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j]));
1055 record_base_value (REGNO (map->reg_map[j]),
1058 /* The last copy needs the compare/branch insns at the end,
1059 so reset copy_end here if the loop ends with a conditional
1062 if (i == unroll_number - 1)
1064 if (GET_CODE (last_loop_insn) == BARRIER)
1065 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
1067 copy_end = last_loop_insn;
1070 /* None of the copies are the `last_iteration', so just
1071 pass zero for that parameter. */
1072 copy_loop_body (copy_start, copy_end, map, exit_label, 0,
1073 unroll_type, start_label, loop_end,
1074 loop_start, copy_end);
1076 emit_label_after (labels[0], PREV_INSN (loop_start));
1078 if (GET_CODE (last_loop_insn) == BARRIER)
1080 insert_before = PREV_INSN (last_loop_insn);
1081 copy_end = PREV_INSN (insert_before);
1086 /* The immediately preceding insn is a compare which we do not
1088 insert_before = PREV_INSN (last_loop_insn);
1089 copy_end = PREV_INSN (insert_before);
1091 /* The immediately preceding insn may not be a compare, so we
1093 insert_before = last_loop_insn;
1094 copy_end = PREV_INSN (last_loop_insn);
1098 /* Set unroll type to MODULO now. */
1099 unroll_type = UNROLL_MODULO;
1100 loop_preconditioned = 1;
1103 /* Fix the initial value for the loop as needed. */
1104 if (loop_n_iterations <= 0)
1105 loop_start_value [uid_loop_num [INSN_UID (loop_start)]]
1111 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1112 the loop unless all loops are being unrolled. */
1113 if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
1115 if (loop_dump_stream)
1116 fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n");
1120 /* At this point, we are guaranteed to unroll the loop. */
1122 /* Keep track of the unroll factor for each loop. */
1123 if (unroll_type == UNROLL_COMPLETELY)
1124 loop_unroll_factor [uid_loop_num [INSN_UID (loop_start)]] = -1;
1126 loop_unroll_factor [uid_loop_num [INSN_UID (loop_start)]] = unroll_number;
1129 /* For each biv and giv, determine whether it can be safely split into
1130 a different variable for each unrolled copy of the loop body.
1131 We precalculate and save this info here, since computing it is
1134 Do this before deleting any instructions from the loop, so that
1135 back_branch_in_range_p will work correctly. */
1137 if (splitting_not_safe)
1140 temp = find_splittable_regs (unroll_type, loop_start, loop_end,
1141 end_insert_before, unroll_number);
1143 /* find_splittable_regs may have created some new registers, so must
1144 reallocate the reg_map with the new larger size, and must realloc
1145 the constant maps also. */
1147 maxregnum = max_reg_num ();
1148 map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
1150 init_reg_map (map, maxregnum);
1152 /* Space is needed in some of the map for new registers, so new_maxregnum
1153 is an (over)estimate of how many registers will exist at the end. */
1154 new_maxregnum = maxregnum + (temp * unroll_number * 2);
1156 /* Must realloc space for the constant maps, because the number of registers
1157 may have changed. */
1159 map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx));
1160 map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned));
1162 map->const_equiv_map_size = new_maxregnum;
1163 global_const_equiv_map = map->const_equiv_map;
1164 global_const_equiv_map_size = new_maxregnum;
1166 /* Search the list of bivs and givs to find ones which need to be remapped
1167 when split, and set their reg_map entry appropriately. */
1169 for (bl = loop_iv_list; bl; bl = bl->next)
1171 if (REGNO (bl->biv->src_reg) != bl->regno)
1172 map->reg_map[bl->regno] = bl->biv->src_reg;
1174 /* Currently, non-reduced/final-value givs are never split. */
1175 for (v = bl->giv; v; v = v->next_iv)
1176 if (REGNO (v->src_reg) != bl->regno)
1177 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1181 /* Use our current register alignment and pointer flags. */
1182 map->regno_pointer_flag = regno_pointer_flag;
1183 map->regno_pointer_align = regno_pointer_align;
1185 /* If the loop is being partially unrolled, and the iteration variables
1186 are being split, and are being renamed for the split, then must fix up
1187 the compare/jump instruction at the end of the loop to refer to the new
1188 registers. This compare isn't copied, so the registers used in it
1189 will never be replaced if it isn't done here. */
1191 if (unroll_type == UNROLL_MODULO)
1193 insn = NEXT_INSN (copy_end);
1194 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
1195 PATTERN (insn) = remap_split_bivs (PATTERN (insn));
1198 /* For unroll_number - 1 times, make a copy of each instruction
1199 between copy_start and copy_end, and insert these new instructions
1200 before the end of the loop. */
1202 for (i = 0; i < unroll_number; i++)
1204 bzero ((char *) map->insn_map, max_insnno * sizeof (rtx));
1205 bzero ((char *) map->const_equiv_map, new_maxregnum * sizeof (rtx));
1206 bzero ((char *) map->const_age_map, new_maxregnum * sizeof (unsigned));
1209 for (j = 0; j < max_labelno; j++)
1211 set_label_in_map (map, j, gen_label_rtx ());
1213 for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++)
1216 map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j]));
1217 record_base_value (REGNO (map->reg_map[j]),
1221 /* If loop starts with a branch to the test, then fix it so that
1222 it points to the test of the first unrolled copy of the loop. */
1223 if (i == 0 && loop_start != copy_start)
1225 insn = PREV_INSN (copy_start);
1226 pattern = PATTERN (insn);
1228 tem = get_label_from_map (map,
1230 (XEXP (SET_SRC (pattern), 0)));
1231 SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem);
1233 /* Set the jump label so that it can be used by later loop unrolling
1235 JUMP_LABEL (insn) = tem;
1236 LABEL_NUSES (tem)++;
1239 copy_loop_body (copy_start, copy_end, map, exit_label,
1240 i == unroll_number - 1, unroll_type, start_label,
1241 loop_end, insert_before, insert_before);
1244 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1245 insn to be deleted. This prevents any runaway delete_insn call from
1246 more insns that it should, as it always stops at a CODE_LABEL. */
1248 /* Delete the compare and branch at the end of the loop if completely
1249 unrolling the loop. Deleting the backward branch at the end also
1250 deletes the code label at the start of the loop. This is done at
1251 the very end to avoid problems with back_branch_in_range_p. */
1253 if (unroll_type == UNROLL_COMPLETELY)
1254 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1256 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1258 /* Delete all of the original loop instructions. Don't delete the
1259 LOOP_BEG note, or the first code label in the loop. */
1261 insn = NEXT_INSN (copy_start);
1262 while (insn != safety_label)
1264 if (insn != start_label)
1265 insn = delete_insn (insn);
1267 insn = NEXT_INSN (insn);
1270 /* Can now delete the 'safety' label emitted to protect us from runaway
1271 delete_insn calls. */
1272 if (INSN_DELETED_P (safety_label))
1274 delete_insn (safety_label);
1276 /* If exit_label exists, emit it after the loop. Doing the emit here
1277 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1278 This is needed so that mostly_true_jump in reorg.c will treat jumps
1279 to this loop end label correctly, i.e. predict that they are usually
1282 emit_label_after (exit_label, loop_end);
1285 /* Return true if the loop can be safely, and profitably, preconditioned
1286 so that the unrolled copies of the loop body don't need exit tests.
1288 This only works if final_value, initial_value and increment can be
1289 determined, and if increment is a constant power of 2.
1290 If increment is not a power of 2, then the preconditioning modulo
1291 operation would require a real modulo instead of a boolean AND, and this
1292 is not considered `profitable'. */
1294 /* ??? If the loop is known to be executed very many times, or the machine
1295 has a very cheap divide instruction, then preconditioning is a win even
1296 when the increment is not a power of 2. Use RTX_COST to compute
1297 whether divide is cheap. */
1300 precondition_loop_p (initial_value, final_value, increment, loop_start,
1302 rtx *initial_value, *final_value, *increment;
1303 rtx loop_start, loop_end;
1306 if (loop_n_iterations > 0)
1308 *initial_value = const0_rtx;
1309 *increment = const1_rtx;
1310 *final_value = GEN_INT (loop_n_iterations);
1312 if (loop_dump_stream)
1313 fprintf (loop_dump_stream,
1314 "Preconditioning: Success, number of iterations known, %d.\n",
1319 if (loop_initial_value == 0)
1321 if (loop_dump_stream)
1322 fprintf (loop_dump_stream,
1323 "Preconditioning: Could not find initial value.\n");
1326 else if (loop_increment == 0)
1328 if (loop_dump_stream)
1329 fprintf (loop_dump_stream,
1330 "Preconditioning: Could not find increment value.\n");
1333 else if (GET_CODE (loop_increment) != CONST_INT)
1335 if (loop_dump_stream)
1336 fprintf (loop_dump_stream,
1337 "Preconditioning: Increment not a constant.\n");
1340 else if ((exact_log2 (INTVAL (loop_increment)) < 0)
1341 && (exact_log2 (- INTVAL (loop_increment)) < 0))
1343 if (loop_dump_stream)
1344 fprintf (loop_dump_stream,
1345 "Preconditioning: Increment not a constant power of 2.\n");
1349 /* Unsigned_compare and compare_dir can be ignored here, since they do
1350 not matter for preconditioning. */
1352 if (loop_final_value == 0)
1354 if (loop_dump_stream)
1355 fprintf (loop_dump_stream,
1356 "Preconditioning: EQ comparison loop.\n");
1360 /* Must ensure that final_value is invariant, so call invariant_p to
1361 check. Before doing so, must check regno against max_reg_before_loop
1362 to make sure that the register is in the range covered by invariant_p.
1363 If it isn't, then it is most likely a biv/giv which by definition are
1365 if ((GET_CODE (loop_final_value) == REG
1366 && REGNO (loop_final_value) >= max_reg_before_loop)
1367 || (GET_CODE (loop_final_value) == PLUS
1368 && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop)
1369 || ! invariant_p (loop_final_value))
1371 if (loop_dump_stream)
1372 fprintf (loop_dump_stream,
1373 "Preconditioning: Final value not invariant.\n");
1377 /* Fail for floating point values, since the caller of this function
1378 does not have code to deal with them. */
1379 if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT
1380 || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT)
1382 if (loop_dump_stream)
1383 fprintf (loop_dump_stream,
1384 "Preconditioning: Floating point final or initial value.\n");
1388 /* Now set initial_value to be the iteration_var, since that may be a
1389 simpler expression, and is guaranteed to be correct if all of the
1390 above tests succeed.
1392 We can not use the initial_value as calculated, because it will be
1393 one too small for loops of the form "while (i-- > 0)". We can not
1394 emit code before the loop_skip_over insns to fix this problem as this
1395 will then give a number one too large for loops of the form
1398 Note that all loops that reach here are entered at the top, because
1399 this function is not called if the loop starts with a jump. */
1401 /* Fail if loop_iteration_var is not live before loop_start, since we need
1402 to test its value in the preconditioning code. */
1404 if (uid_luid[REGNO_FIRST_UID (REGNO (loop_iteration_var))]
1405 > INSN_LUID (loop_start))
1407 if (loop_dump_stream)
1408 fprintf (loop_dump_stream,
1409 "Preconditioning: Iteration var not live before loop start.\n");
1413 *initial_value = loop_iteration_var;
1414 *increment = loop_increment;
1415 *final_value = loop_final_value;
1418 if (loop_dump_stream)
1419 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1424 /* All pseudo-registers must be mapped to themselves. Two hard registers
1425 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1426 REGNUM, to avoid function-inlining specific conversions of these
1427 registers. All other hard regs can not be mapped because they may be
1432 init_reg_map (map, maxregnum)
1433 struct inline_remap *map;
1438 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1439 map->reg_map[i] = regno_reg_rtx[i];
1440 /* Just clear the rest of the entries. */
1441 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1442 map->reg_map[i] = 0;
1444 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1445 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1446 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1447 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1450 /* Strength-reduction will often emit code for optimized biv/givs which
1451 calculates their value in a temporary register, and then copies the result
1452 to the iv. This procedure reconstructs the pattern computing the iv;
1453 verifying that all operands are of the proper form.
1455 PATTERN must be the result of single_set.
1456 The return value is the amount that the giv is incremented by. */
1459 calculate_giv_inc (pattern, src_insn, regno)
1460 rtx pattern, src_insn;
1464 rtx increment_total = 0;
1468 /* Verify that we have an increment insn here. First check for a plus
1469 as the set source. */
1470 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1472 /* SR sometimes computes the new giv value in a temp, then copies it
1474 src_insn = PREV_INSN (src_insn);
1475 pattern = PATTERN (src_insn);
1476 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1479 /* The last insn emitted is not needed, so delete it to avoid confusing
1480 the second cse pass. This insn sets the giv unnecessarily. */
1481 delete_insn (get_last_insn ());
1484 /* Verify that we have a constant as the second operand of the plus. */
1485 increment = XEXP (SET_SRC (pattern), 1);
1486 if (GET_CODE (increment) != CONST_INT)
1488 /* SR sometimes puts the constant in a register, especially if it is
1489 too big to be an add immed operand. */
1490 src_insn = PREV_INSN (src_insn);
1491 increment = SET_SRC (PATTERN (src_insn));
1493 /* SR may have used LO_SUM to compute the constant if it is too large
1494 for a load immed operand. In this case, the constant is in operand
1495 one of the LO_SUM rtx. */
1496 if (GET_CODE (increment) == LO_SUM)
1497 increment = XEXP (increment, 1);
1499 /* Some ports store large constants in memory and add a REG_EQUAL
1500 note to the store insn. */
1501 else if (GET_CODE (increment) == MEM)
1503 rtx note = find_reg_note (src_insn, REG_EQUAL, 0);
1505 increment = XEXP (note, 0);
1508 else if (GET_CODE (increment) == IOR
1509 || GET_CODE (increment) == ASHIFT
1510 || GET_CODE (increment) == PLUS)
1512 /* The rs6000 port loads some constants with IOR.
1513 The alpha port loads some constants with ASHIFT and PLUS. */
1514 rtx second_part = XEXP (increment, 1);
1515 enum rtx_code code = GET_CODE (increment);
1517 src_insn = PREV_INSN (src_insn);
1518 increment = SET_SRC (PATTERN (src_insn));
1519 /* Don't need the last insn anymore. */
1520 delete_insn (get_last_insn ());
1522 if (GET_CODE (second_part) != CONST_INT
1523 || GET_CODE (increment) != CONST_INT)
1527 increment = GEN_INT (INTVAL (increment) | INTVAL (second_part));
1528 else if (code == PLUS)
1529 increment = GEN_INT (INTVAL (increment) + INTVAL (second_part));
1531 increment = GEN_INT (INTVAL (increment) << INTVAL (second_part));
1534 if (GET_CODE (increment) != CONST_INT)
1537 /* The insn loading the constant into a register is no longer needed,
1539 delete_insn (get_last_insn ());
1542 if (increment_total)
1543 increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment));
1545 increment_total = increment;
1547 /* Check that the source register is the same as the register we expected
1548 to see as the source. If not, something is seriously wrong. */
1549 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1550 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1552 /* Some machines (e.g. the romp), may emit two add instructions for
1553 certain constants, so lets try looking for another add immediately
1554 before this one if we have only seen one add insn so far. */
1560 src_insn = PREV_INSN (src_insn);
1561 pattern = PATTERN (src_insn);
1563 delete_insn (get_last_insn ());
1571 return increment_total;
1574 /* Copy REG_NOTES, except for insn references, because not all insn_map
1575 entries are valid yet. We do need to copy registers now though, because
1576 the reg_map entries can change during copying. */
1579 initial_reg_note_copy (notes, map)
1581 struct inline_remap *map;
1588 copy = rtx_alloc (GET_CODE (notes));
1589 PUT_MODE (copy, GET_MODE (notes));
1591 if (GET_CODE (notes) == EXPR_LIST)
1592 XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map);
1593 else if (GET_CODE (notes) == INSN_LIST)
1594 /* Don't substitute for these yet. */
1595 XEXP (copy, 0) = XEXP (notes, 0);
1599 XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map);
1604 /* Fixup insn references in copied REG_NOTES. */
1607 final_reg_note_copy (notes, map)
1609 struct inline_remap *map;
1613 for (note = notes; note; note = XEXP (note, 1))
1614 if (GET_CODE (note) == INSN_LIST)
1615 XEXP (note, 0) = map->insn_map[INSN_UID (XEXP (note, 0))];
1618 /* Copy each instruction in the loop, substituting from map as appropriate.
1619 This is very similar to a loop in expand_inline_function. */
1622 copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration,
1623 unroll_type, start_label, loop_end, insert_before,
1625 rtx copy_start, copy_end;
1626 struct inline_remap *map;
1629 enum unroll_types unroll_type;
1630 rtx start_label, loop_end, insert_before, copy_notes_from;
1634 int dest_reg_was_split, i;
1638 rtx final_label = 0;
1639 rtx giv_inc, giv_dest_reg, giv_src_reg;
1641 /* If this isn't the last iteration, then map any references to the
1642 start_label to final_label. Final label will then be emitted immediately
1643 after the end of this loop body if it was ever used.
1645 If this is the last iteration, then map references to the start_label
1647 if (! last_iteration)
1649 final_label = gen_label_rtx ();
1650 set_label_in_map (map, CODE_LABEL_NUMBER (start_label),
1654 set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label);
1661 insn = NEXT_INSN (insn);
1663 map->orig_asm_operands_vector = 0;
1665 switch (GET_CODE (insn))
1668 pattern = PATTERN (insn);
1672 /* Check to see if this is a giv that has been combined with
1673 some split address givs. (Combined in the sense that
1674 `combine_givs' in loop.c has put two givs in the same register.)
1675 In this case, we must search all givs based on the same biv to
1676 find the address givs. Then split the address givs.
1677 Do this before splitting the giv, since that may map the
1678 SET_DEST to a new register. */
1680 if ((set = single_set (insn))
1681 && GET_CODE (SET_DEST (set)) == REG
1682 && addr_combined_regs[REGNO (SET_DEST (set))])
1684 struct iv_class *bl;
1685 struct induction *v, *tv;
1686 int regno = REGNO (SET_DEST (set));
1688 v = addr_combined_regs[REGNO (SET_DEST (set))];
1689 bl = reg_biv_class[REGNO (v->src_reg)];
1691 /* Although the giv_inc amount is not needed here, we must call
1692 calculate_giv_inc here since it might try to delete the
1693 last insn emitted. If we wait until later to call it,
1694 we might accidentally delete insns generated immediately
1695 below by emit_unrolled_add. */
1697 giv_inc = calculate_giv_inc (set, insn, regno);
1699 /* Now find all address giv's that were combined with this
1701 for (tv = bl->giv; tv; tv = tv->next_iv)
1702 if (tv->giv_type == DEST_ADDR && tv->same == v)
1706 /* If this DEST_ADDR giv was not split, then ignore it. */
1707 if (*tv->location != tv->dest_reg)
1710 /* Scale this_giv_inc if the multiplicative factors of
1711 the two givs are different. */
1712 this_giv_inc = INTVAL (giv_inc);
1713 if (tv->mult_val != v->mult_val)
1714 this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1715 * INTVAL (tv->mult_val));
1717 tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1718 *tv->location = tv->dest_reg;
1720 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1722 /* Must emit an insn to increment the split address
1723 giv. Add in the const_adjust field in case there
1724 was a constant eliminated from the address. */
1725 rtx value, dest_reg;
1727 /* tv->dest_reg will be either a bare register,
1728 or else a register plus a constant. */
1729 if (GET_CODE (tv->dest_reg) == REG)
1730 dest_reg = tv->dest_reg;
1732 dest_reg = XEXP (tv->dest_reg, 0);
1734 /* Check for shared address givs, and avoid
1735 incrementing the shared pseudo reg more than
1737 if (! tv->same_insn && ! tv->shared)
1739 /* tv->dest_reg may actually be a (PLUS (REG)
1740 (CONST)) here, so we must call plus_constant
1741 to add the const_adjust amount before calling
1742 emit_unrolled_add below. */
1743 value = plus_constant (tv->dest_reg,
1746 /* The constant could be too large for an add
1747 immediate, so can't directly emit an insn
1749 emit_unrolled_add (dest_reg, XEXP (value, 0),
1753 /* Reset the giv to be just the register again, in case
1754 it is used after the set we have just emitted.
1755 We must subtract the const_adjust factor added in
1757 tv->dest_reg = plus_constant (dest_reg,
1758 - tv->const_adjust);
1759 *tv->location = tv->dest_reg;
1764 /* If this is a setting of a splittable variable, then determine
1765 how to split the variable, create a new set based on this split,
1766 and set up the reg_map so that later uses of the variable will
1767 use the new split variable. */
1769 dest_reg_was_split = 0;
1771 if ((set = single_set (insn))
1772 && GET_CODE (SET_DEST (set)) == REG
1773 && splittable_regs[REGNO (SET_DEST (set))])
1775 int regno = REGNO (SET_DEST (set));
1777 dest_reg_was_split = 1;
1779 /* Compute the increment value for the giv, if it wasn't
1780 already computed above. */
1783 giv_inc = calculate_giv_inc (set, insn, regno);
1784 giv_dest_reg = SET_DEST (set);
1785 giv_src_reg = SET_DEST (set);
1787 if (unroll_type == UNROLL_COMPLETELY)
1789 /* Completely unrolling the loop. Set the induction
1790 variable to a known constant value. */
1792 /* The value in splittable_regs may be an invariant
1793 value, so we must use plus_constant here. */
1794 splittable_regs[regno]
1795 = plus_constant (splittable_regs[regno], INTVAL (giv_inc));
1797 if (GET_CODE (splittable_regs[regno]) == PLUS)
1799 giv_src_reg = XEXP (splittable_regs[regno], 0);
1800 giv_inc = XEXP (splittable_regs[regno], 1);
1804 /* The splittable_regs value must be a REG or a
1805 CONST_INT, so put the entire value in the giv_src_reg
1807 giv_src_reg = splittable_regs[regno];
1808 giv_inc = const0_rtx;
1813 /* Partially unrolling loop. Create a new pseudo
1814 register for the iteration variable, and set it to
1815 be a constant plus the original register. Except
1816 on the last iteration, when the result has to
1817 go back into the original iteration var register. */
1819 /* Handle bivs which must be mapped to a new register
1820 when split. This happens for bivs which need their
1821 final value set before loop entry. The new register
1822 for the biv was stored in the biv's first struct
1823 induction entry by find_splittable_regs. */
1825 if (regno < max_reg_before_loop
1826 && reg_iv_type[regno] == BASIC_INDUCT)
1828 giv_src_reg = reg_biv_class[regno]->biv->src_reg;
1829 giv_dest_reg = giv_src_reg;
1833 /* If non-reduced/final-value givs were split, then
1834 this would have to remap those givs also. See
1835 find_splittable_regs. */
1838 splittable_regs[regno]
1839 = GEN_INT (INTVAL (giv_inc)
1840 + INTVAL (splittable_regs[regno]));
1841 giv_inc = splittable_regs[regno];
1843 /* Now split the induction variable by changing the dest
1844 of this insn to a new register, and setting its
1845 reg_map entry to point to this new register.
1847 If this is the last iteration, and this is the last insn
1848 that will update the iv, then reuse the original dest,
1849 to ensure that the iv will have the proper value when
1850 the loop exits or repeats.
1852 Using splittable_regs_updates here like this is safe,
1853 because it can only be greater than one if all
1854 instructions modifying the iv are always executed in
1857 if (! last_iteration
1858 || (splittable_regs_updates[regno]-- != 1))
1860 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1862 map->reg_map[regno] = tem;
1863 record_base_value (REGNO (tem), giv_src_reg);
1866 map->reg_map[regno] = giv_src_reg;
1869 /* The constant being added could be too large for an add
1870 immediate, so can't directly emit an insn here. */
1871 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1872 copy = get_last_insn ();
1873 pattern = PATTERN (copy);
1877 pattern = copy_rtx_and_substitute (pattern, map);
1878 copy = emit_insn (pattern);
1880 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
1883 /* If this insn is setting CC0, it may need to look at
1884 the insn that uses CC0 to see what type of insn it is.
1885 In that case, the call to recog via validate_change will
1886 fail. So don't substitute constants here. Instead,
1887 do it when we emit the following insn.
1889 For example, see the pyr.md file. That machine has signed and
1890 unsigned compares. The compare patterns must check the
1891 following branch insn to see which what kind of compare to
1894 If the previous insn set CC0, substitute constants on it as
1896 if (sets_cc0_p (PATTERN (copy)) != 0)
1901 try_constants (cc0_insn, map);
1903 try_constants (copy, map);
1906 try_constants (copy, map);
1909 /* Make split induction variable constants `permanent' since we
1910 know there are no backward branches across iteration variable
1911 settings which would invalidate this. */
1912 if (dest_reg_was_split)
1914 int regno = REGNO (SET_DEST (pattern));
1916 if (regno < map->const_equiv_map_size
1917 && map->const_age_map[regno] == map->const_age)
1918 map->const_age_map[regno] = -1;
1923 pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1924 copy = emit_jump_insn (pattern);
1925 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
1927 if (JUMP_LABEL (insn) == start_label && insn == copy_end
1928 && ! last_iteration)
1930 /* This is a branch to the beginning of the loop; this is the
1931 last insn being copied; and this is not the last iteration.
1932 In this case, we want to change the original fall through
1933 case to be a branch past the end of the loop, and the
1934 original jump label case to fall_through. */
1936 if (invert_exp (pattern, copy))
1938 if (! redirect_exp (&pattern,
1939 get_label_from_map (map,
1941 (JUMP_LABEL (insn))),
1948 rtx lab = gen_label_rtx ();
1949 /* Can't do it by reversing the jump (probably because we
1950 couldn't reverse the conditions), so emit a new
1951 jump_insn after COPY, and redirect the jump around
1953 jmp = emit_jump_insn_after (gen_jump (exit_label), copy);
1954 jmp = emit_barrier_after (jmp);
1955 emit_label_after (lab, jmp);
1956 LABEL_NUSES (lab) = 0;
1957 if (! redirect_exp (&pattern,
1958 get_label_from_map (map,
1960 (JUMP_LABEL (insn))),
1968 try_constants (cc0_insn, map);
1971 try_constants (copy, map);
1973 /* Set the jump label of COPY correctly to avoid problems with
1974 later passes of unroll_loop, if INSN had jump label set. */
1975 if (JUMP_LABEL (insn))
1979 /* Can't use the label_map for every insn, since this may be
1980 the backward branch, and hence the label was not mapped. */
1981 if ((set = single_set (copy)))
1983 tem = SET_SRC (set);
1984 if (GET_CODE (tem) == LABEL_REF)
1985 label = XEXP (tem, 0);
1986 else if (GET_CODE (tem) == IF_THEN_ELSE)
1988 if (XEXP (tem, 1) != pc_rtx)
1989 label = XEXP (XEXP (tem, 1), 0);
1991 label = XEXP (XEXP (tem, 2), 0);
1995 if (label && GET_CODE (label) == CODE_LABEL)
1996 JUMP_LABEL (copy) = label;
1999 /* An unrecognizable jump insn, probably the entry jump
2000 for a switch statement. This label must have been mapped,
2001 so just use the label_map to get the new jump label. */
2003 = get_label_from_map (map,
2004 CODE_LABEL_NUMBER (JUMP_LABEL (insn)));
2007 /* If this is a non-local jump, then must increase the label
2008 use count so that the label will not be deleted when the
2009 original jump is deleted. */
2010 LABEL_NUSES (JUMP_LABEL (copy))++;
2012 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
2013 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
2015 rtx pat = PATTERN (copy);
2016 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
2017 int len = XVECLEN (pat, diff_vec_p);
2020 for (i = 0; i < len; i++)
2021 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
2024 /* If this used to be a conditional jump insn but whose branch
2025 direction is now known, we must do something special. */
2026 if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value)
2029 /* The previous insn set cc0 for us. So delete it. */
2030 delete_insn (PREV_INSN (copy));
2033 /* If this is now a no-op, delete it. */
2034 if (map->last_pc_value == pc_rtx)
2036 /* Don't let delete_insn delete the label referenced here,
2037 because we might possibly need it later for some other
2038 instruction in the loop. */
2039 if (JUMP_LABEL (copy))
2040 LABEL_NUSES (JUMP_LABEL (copy))++;
2042 if (JUMP_LABEL (copy))
2043 LABEL_NUSES (JUMP_LABEL (copy))--;
2047 /* Otherwise, this is unconditional jump so we must put a
2048 BARRIER after it. We could do some dead code elimination
2049 here, but jump.c will do it just as well. */
2055 pattern = copy_rtx_and_substitute (PATTERN (insn), map);
2056 copy = emit_call_insn (pattern);
2057 REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map);
2059 /* Because the USAGE information potentially contains objects other
2060 than hard registers, we need to copy it. */
2061 CALL_INSN_FUNCTION_USAGE (copy)
2062 = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn), map);
2066 try_constants (cc0_insn, map);
2069 try_constants (copy, map);
2071 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2072 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2073 map->const_equiv_map[i] = 0;
2077 /* If this is the loop start label, then we don't need to emit a
2078 copy of this label since no one will use it. */
2080 if (insn != start_label)
2082 copy = emit_label (get_label_from_map (map,
2083 CODE_LABEL_NUMBER (insn)));
2089 copy = emit_barrier ();
2093 /* VTOP notes are valid only before the loop exit test. If placed
2094 anywhere else, loop may generate bad code. */
2096 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
2097 && (NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP
2098 || (last_iteration && unroll_type != UNROLL_COMPLETELY)))
2099 copy = emit_note (NOTE_SOURCE_FILE (insn),
2100 NOTE_LINE_NUMBER (insn));
2110 map->insn_map[INSN_UID (insn)] = copy;
2112 while (insn != copy_end);
2114 /* Now finish coping the REG_NOTES. */
2118 insn = NEXT_INSN (insn);
2119 if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
2120 || GET_CODE (insn) == CALL_INSN)
2121 && map->insn_map[INSN_UID (insn)])
2122 final_reg_note_copy (REG_NOTES (map->insn_map[INSN_UID (insn)]), map);
2124 while (insn != copy_end);
2126 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2127 each of these notes here, since there may be some important ones, such as
2128 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2129 iteration, because the original notes won't be deleted.
2131 We can't use insert_before here, because when from preconditioning,
2132 insert_before points before the loop. We can't use copy_end, because
2133 there may be insns already inserted after it (which we don't want to
2134 copy) when not from preconditioning code. */
2136 if (! last_iteration)
2138 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
2140 if (GET_CODE (insn) == NOTE
2141 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
2142 emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
2146 if (final_label && LABEL_NUSES (final_label) > 0)
2147 emit_label (final_label);
2149 tem = gen_sequence ();
2151 emit_insn_before (tem, insert_before);
2154 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2155 emitted. This will correctly handle the case where the increment value
2156 won't fit in the immediate field of a PLUS insns. */
2159 emit_unrolled_add (dest_reg, src_reg, increment)
2160 rtx dest_reg, src_reg, increment;
2164 result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment,
2165 dest_reg, 0, OPTAB_LIB_WIDEN);
2167 if (dest_reg != result)
2168 emit_move_insn (dest_reg, result);
2171 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2172 is a backward branch in that range that branches to somewhere between
2173 LOOP_START and INSN. Returns 0 otherwise. */
2175 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2176 In practice, this is not a problem, because this function is seldom called,
2177 and uses a negligible amount of CPU time on average. */
2180 back_branch_in_range_p (insn, loop_start, loop_end)
2182 rtx loop_start, loop_end;
2184 rtx p, q, target_insn;
2185 rtx orig_loop_end = loop_end;
2187 /* Stop before we get to the backward branch at the end of the loop. */
2188 loop_end = prev_nonnote_insn (loop_end);
2189 if (GET_CODE (loop_end) == BARRIER)
2190 loop_end = PREV_INSN (loop_end);
2192 /* Check in case insn has been deleted, search forward for first non
2193 deleted insn following it. */
2194 while (INSN_DELETED_P (insn))
2195 insn = NEXT_INSN (insn);
2197 /* Check for the case where insn is the last insn in the loop. Deal
2198 with the case where INSN was a deleted loop test insn, in which case
2199 it will now be the NOTE_LOOP_END. */
2200 if (insn == loop_end || insn == orig_loop_end)
2203 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
2205 if (GET_CODE (p) == JUMP_INSN)
2207 target_insn = JUMP_LABEL (p);
2209 /* Search from loop_start to insn, to see if one of them is
2210 the target_insn. We can't use INSN_LUID comparisons here,
2211 since insn may not have an LUID entry. */
2212 for (q = loop_start; q != insn; q = NEXT_INSN (q))
2213 if (q == target_insn)
2221 /* Try to generate the simplest rtx for the expression
2222 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2226 fold_rtx_mult_add (mult1, mult2, add1, mode)
2227 rtx mult1, mult2, add1;
2228 enum machine_mode mode;
2233 /* The modes must all be the same. This should always be true. For now,
2234 check to make sure. */
2235 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
2236 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
2237 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
2240 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2241 will be a constant. */
2242 if (GET_CODE (mult1) == CONST_INT)
2249 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
2251 mult_res = gen_rtx_MULT (mode, mult1, mult2);
2253 /* Again, put the constant second. */
2254 if (GET_CODE (add1) == CONST_INT)
2261 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
2263 result = gen_rtx_PLUS (mode, add1, mult_res);
2268 /* Searches the list of induction struct's for the biv BL, to try to calculate
2269 the total increment value for one iteration of the loop as a constant.
2271 Returns the increment value as an rtx, simplified as much as possible,
2272 if it can be calculated. Otherwise, returns 0. */
2275 biv_total_increment (bl, loop_start, loop_end)
2276 struct iv_class *bl;
2277 rtx loop_start, loop_end;
2279 struct induction *v;
2282 /* For increment, must check every instruction that sets it. Each
2283 instruction must be executed only once each time through the loop.
2284 To verify this, we check that the the insn is always executed, and that
2285 there are no backward branches after the insn that branch to before it.
2286 Also, the insn must have a mult_val of one (to make sure it really is
2289 result = const0_rtx;
2290 for (v = bl->biv; v; v = v->next_iv)
2292 if (v->always_computable && v->mult_val == const1_rtx
2293 && ! back_branch_in_range_p (v->insn, loop_start, loop_end))
2294 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2302 /* Determine the initial value of the iteration variable, and the amount
2303 that it is incremented each loop. Use the tables constructed by
2304 the strength reduction pass to calculate these values.
2306 Initial_value and/or increment are set to zero if their values could not
2310 iteration_info (iteration_var, initial_value, increment, loop_start, loop_end)
2311 rtx iteration_var, *initial_value, *increment;
2312 rtx loop_start, loop_end;
2314 struct iv_class *bl;
2316 struct induction *v;
2319 /* Clear the result values, in case no answer can be found. */
2323 /* The iteration variable can be either a giv or a biv. Check to see
2324 which it is, and compute the variable's initial value, and increment
2325 value if possible. */
2327 /* If this is a new register, can't handle it since we don't have any
2328 reg_iv_type entry for it. */
2329 if (REGNO (iteration_var) >= max_reg_before_loop)
2331 if (loop_dump_stream)
2332 fprintf (loop_dump_stream,
2333 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2337 /* Reject iteration variables larger than the host wide int size, since they
2338 could result in a number of iterations greater than the range of our
2339 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2340 else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var))
2341 > HOST_BITS_PER_WIDE_INT))
2343 if (loop_dump_stream)
2344 fprintf (loop_dump_stream,
2345 "Loop unrolling: Iteration var rejected because mode too large.\n");
2348 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
2350 if (loop_dump_stream)
2351 fprintf (loop_dump_stream,
2352 "Loop unrolling: Iteration var not an integer.\n");
2355 else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT)
2357 /* Grab initial value, only useful if it is a constant. */
2358 bl = reg_biv_class[REGNO (iteration_var)];
2359 *initial_value = bl->initial_value;
2361 *increment = biv_total_increment (bl, loop_start, loop_end);
2363 else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT)
2366 /* ??? The code below does not work because the incorrect number of
2367 iterations is calculated when the biv is incremented after the giv
2368 is set (which is the usual case). This can probably be accounted
2369 for by biasing the initial_value by subtracting the amount of the
2370 increment that occurs between the giv set and the giv test. However,
2371 a giv as an iterator is very rare, so it does not seem worthwhile
2373 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2374 if (loop_dump_stream)
2375 fprintf (loop_dump_stream,
2376 "Loop unrolling: Giv iterators are not handled.\n");
2379 /* Initial value is mult_val times the biv's initial value plus
2380 add_val. Only useful if it is a constant. */
2381 v = reg_iv_info[REGNO (iteration_var)];
2382 bl = reg_biv_class[REGNO (v->src_reg)];
2383 *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value,
2384 v->add_val, v->mode);
2386 /* Increment value is mult_val times the increment value of the biv. */
2388 *increment = biv_total_increment (bl, loop_start, loop_end);
2390 *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx,
2396 if (loop_dump_stream)
2397 fprintf (loop_dump_stream,
2398 "Loop unrolling: Not basic or general induction var.\n");
2403 /* Calculate the approximate final value of the iteration variable
2404 which has an loop exit test with code COMPARISON_CODE and comparison value
2405 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2406 was signed or unsigned, and the direction of the comparison. This info is
2407 needed to calculate the number of loop iterations. */
2410 approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir)
2411 enum rtx_code comparison_code;
2412 rtx comparison_value;
2416 /* Calculate the final value of the induction variable.
2417 The exact final value depends on the branch operator, and increment sign.
2418 This is only an approximate value. It will be wrong if the iteration
2419 variable is not incremented by one each time through the loop, and
2420 approx final value - start value % increment != 0. */
2423 switch (comparison_code)
2429 return plus_constant (comparison_value, 1);
2434 return plus_constant (comparison_value, -1);
2436 /* Can not calculate a final value for this case. */
2443 return comparison_value;
2449 return comparison_value;
2452 return comparison_value;
2458 /* For each biv and giv, determine whether it can be safely split into
2459 a different variable for each unrolled copy of the loop body. If it
2460 is safe to split, then indicate that by saving some useful info
2461 in the splittable_regs array.
2463 If the loop is being completely unrolled, then splittable_regs will hold
2464 the current value of the induction variable while the loop is unrolled.
2465 It must be set to the initial value of the induction variable here.
2466 Otherwise, splittable_regs will hold the difference between the current
2467 value of the induction variable and the value the induction variable had
2468 at the top of the loop. It must be set to the value 0 here.
2470 Returns the total number of instructions that set registers that are
2473 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2474 constant values are unnecessary, since we can easily calculate increment
2475 values in this case even if nothing is constant. The increment value
2476 should not involve a multiply however. */
2478 /* ?? Even if the biv/giv increment values aren't constant, it may still
2479 be beneficial to split the variable if the loop is only unrolled a few
2480 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2483 find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before,
2485 enum unroll_types unroll_type;
2486 rtx loop_start, loop_end;
2487 rtx end_insert_before;
2490 struct iv_class *bl;
2491 struct induction *v;
2493 rtx biv_final_value;
2497 for (bl = loop_iv_list; bl; bl = bl->next)
2499 /* Biv_total_increment must return a constant value,
2500 otherwise we can not calculate the split values. */
2502 increment = biv_total_increment (bl, loop_start, loop_end);
2503 if (! increment || GET_CODE (increment) != CONST_INT)
2506 /* The loop must be unrolled completely, or else have a known number
2507 of iterations and only one exit, or else the biv must be dead
2508 outside the loop, or else the final value must be known. Otherwise,
2509 it is unsafe to split the biv since it may not have the proper
2510 value on loop exit. */
2512 /* loop_number_exit_count is non-zero if the loop has an exit other than
2513 a fall through at the end. */
2516 biv_final_value = 0;
2517 if (unroll_type != UNROLL_COMPLETELY
2518 && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
2519 || unroll_type == UNROLL_NAIVE)
2520 && (uid_luid[REGNO_LAST_UID (bl->regno)] >= INSN_LUID (loop_end)
2522 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2523 || (uid_luid[REGNO_FIRST_UID (bl->regno)]
2524 < INSN_LUID (bl->init_insn))
2525 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2526 && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end)))
2529 /* If any of the insns setting the BIV don't do so with a simple
2530 PLUS, we don't know how to split it. */
2531 for (v = bl->biv; biv_splittable && v; v = v->next_iv)
2532 if ((tem = single_set (v->insn)) == 0
2533 || GET_CODE (SET_DEST (tem)) != REG
2534 || REGNO (SET_DEST (tem)) != bl->regno
2535 || GET_CODE (SET_SRC (tem)) != PLUS)
2538 /* If final value is non-zero, then must emit an instruction which sets
2539 the value of the biv to the proper value. This is done after
2540 handling all of the givs, since some of them may need to use the
2541 biv's value in their initialization code. */
2543 /* This biv is splittable. If completely unrolling the loop, save
2544 the biv's initial value. Otherwise, save the constant zero. */
2546 if (biv_splittable == 1)
2548 if (unroll_type == UNROLL_COMPLETELY)
2550 /* If the initial value of the biv is itself (i.e. it is too
2551 complicated for strength_reduce to compute), or is a hard
2552 register, or it isn't invariant, then we must create a new
2553 pseudo reg to hold the initial value of the biv. */
2555 if (GET_CODE (bl->initial_value) == REG
2556 && (REGNO (bl->initial_value) == bl->regno
2557 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER
2558 || ! invariant_p (bl->initial_value)))
2560 rtx tem = gen_reg_rtx (bl->biv->mode);
2562 record_base_value (REGNO (tem), bl->biv->add_val);
2563 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2566 if (loop_dump_stream)
2567 fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n",
2568 bl->regno, REGNO (tem));
2570 splittable_regs[bl->regno] = tem;
2573 splittable_regs[bl->regno] = bl->initial_value;
2576 splittable_regs[bl->regno] = const0_rtx;
2578 /* Save the number of instructions that modify the biv, so that
2579 we can treat the last one specially. */
2581 splittable_regs_updates[bl->regno] = bl->biv_count;
2582 result += bl->biv_count;
2584 if (loop_dump_stream)
2585 fprintf (loop_dump_stream,
2586 "Biv %d safe to split.\n", bl->regno);
2589 /* Check every giv that depends on this biv to see whether it is
2590 splittable also. Even if the biv isn't splittable, givs which
2591 depend on it may be splittable if the biv is live outside the
2592 loop, and the givs aren't. */
2594 result += find_splittable_givs (bl, unroll_type, loop_start, loop_end,
2595 increment, unroll_number);
2597 /* If final value is non-zero, then must emit an instruction which sets
2598 the value of the biv to the proper value. This is done after
2599 handling all of the givs, since some of them may need to use the
2600 biv's value in their initialization code. */
2601 if (biv_final_value)
2603 /* If the loop has multiple exits, emit the insns before the
2604 loop to ensure that it will always be executed no matter
2605 how the loop exits. Otherwise emit the insn after the loop,
2606 since this is slightly more efficient. */
2607 if (! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
2608 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2613 /* Create a new register to hold the value of the biv, and then
2614 set the biv to its final value before the loop start. The biv
2615 is set to its final value before loop start to ensure that
2616 this insn will always be executed, no matter how the loop
2618 rtx tem = gen_reg_rtx (bl->biv->mode);
2619 record_base_value (REGNO (tem), bl->biv->add_val);
2621 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2623 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2627 if (loop_dump_stream)
2628 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2629 REGNO (bl->biv->src_reg), REGNO (tem));
2631 /* Set up the mapping from the original biv register to the new
2633 bl->biv->src_reg = tem;
2640 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2641 for the instruction that is using it. Do not make any changes to that
2645 verify_addresses (v, giv_inc, unroll_number)
2646 struct induction *v;
2651 rtx orig_addr = *v->location;
2652 rtx last_addr = plus_constant (v->dest_reg,
2653 INTVAL (giv_inc) * (unroll_number - 1));
2655 /* First check to see if either address would fail. Handle the fact
2656 that we have may have a match_dup. */
2657 if (! validate_replace_rtx (*v->location, v->dest_reg, v->insn)
2658 || ! validate_replace_rtx (*v->location, last_addr, v->insn))
2661 /* Now put things back the way they were before. This will always
2663 validate_change (v->insn, v->location, orig_addr, 0);
2668 /* For every giv based on the biv BL, check to determine whether it is
2669 splittable. This is a subroutine to find_splittable_regs ().
2671 Return the number of instructions that set splittable registers. */
2674 find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment,
2676 struct iv_class *bl;
2677 enum unroll_types unroll_type;
2678 rtx loop_start, loop_end;
2682 struct induction *v, *v2;
2687 /* Scan the list of givs, and set the same_insn field when there are
2688 multiple identical givs in the same insn. */
2689 for (v = bl->giv; v; v = v->next_iv)
2690 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2691 if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg)
2695 for (v = bl->giv; v; v = v->next_iv)
2699 /* Only split the giv if it has already been reduced, or if the loop is
2700 being completely unrolled. */
2701 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2704 /* The giv can be split if the insn that sets the giv is executed once
2705 and only once on every iteration of the loop. */
2706 /* An address giv can always be split. v->insn is just a use not a set,
2707 and hence it does not matter whether it is always executed. All that
2708 matters is that all the biv increments are always executed, and we
2709 won't reach here if they aren't. */
2710 if (v->giv_type != DEST_ADDR
2711 && (! v->always_computable
2712 || back_branch_in_range_p (v->insn, loop_start, loop_end)))
2715 /* The giv increment value must be a constant. */
2716 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2718 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2721 /* The loop must be unrolled completely, or else have a known number of
2722 iterations and only one exit, or else the giv must be dead outside
2723 the loop, or else the final value of the giv must be known.
2724 Otherwise, it is not safe to split the giv since it may not have the
2725 proper value on loop exit. */
2727 /* The used outside loop test will fail for DEST_ADDR givs. They are
2728 never used outside the loop anyways, so it is always safe to split a
2732 if (unroll_type != UNROLL_COMPLETELY
2733 && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
2734 || unroll_type == UNROLL_NAIVE)
2735 && v->giv_type != DEST_ADDR
2736 /* The next part is true if the pseudo is used outside the loop.
2737 We assume that this is true for any pseudo created after loop
2738 starts, because we don't have a reg_n_info entry for them. */
2739 && (REGNO (v->dest_reg) >= max_reg_before_loop
2740 || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn)
2741 /* Check for the case where the pseudo is set by a shift/add
2742 sequence, in which case the first insn setting the pseudo
2743 is the first insn of the shift/add sequence. */
2744 && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
2745 || (REGNO_FIRST_UID (REGNO (v->dest_reg))
2746 != INSN_UID (XEXP (tem, 0)))))
2747 /* Line above always fails if INSN was moved by loop opt. */
2748 || (uid_luid[REGNO_LAST_UID (REGNO (v->dest_reg))]
2749 >= INSN_LUID (loop_end)))
2750 && ! (final_value = v->final_value))
2754 /* Currently, non-reduced/final-value givs are never split. */
2755 /* Should emit insns after the loop if possible, as the biv final value
2758 /* If the final value is non-zero, and the giv has not been reduced,
2759 then must emit an instruction to set the final value. */
2760 if (final_value && !v->new_reg)
2762 /* Create a new register to hold the value of the giv, and then set
2763 the giv to its final value before the loop start. The giv is set
2764 to its final value before loop start to ensure that this insn
2765 will always be executed, no matter how we exit. */
2766 tem = gen_reg_rtx (v->mode);
2767 emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start);
2768 emit_insn_before (gen_move_insn (v->dest_reg, final_value),
2771 if (loop_dump_stream)
2772 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2773 REGNO (v->dest_reg), REGNO (tem));
2779 /* This giv is splittable. If completely unrolling the loop, save the
2780 giv's initial value. Otherwise, save the constant zero for it. */
2782 if (unroll_type == UNROLL_COMPLETELY)
2784 /* It is not safe to use bl->initial_value here, because it may not
2785 be invariant. It is safe to use the initial value stored in
2786 the splittable_regs array if it is set. In rare cases, it won't
2787 be set, so then we do exactly the same thing as
2788 find_splittable_regs does to get a safe value. */
2789 rtx biv_initial_value;
2791 if (splittable_regs[bl->regno])
2792 biv_initial_value = splittable_regs[bl->regno];
2793 else if (GET_CODE (bl->initial_value) != REG
2794 || (REGNO (bl->initial_value) != bl->regno
2795 && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
2796 biv_initial_value = bl->initial_value;
2799 rtx tem = gen_reg_rtx (bl->biv->mode);
2801 record_base_value (REGNO (tem), bl->biv->add_val);
2802 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2804 biv_initial_value = tem;
2806 value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
2807 v->add_val, v->mode);
2814 /* If a giv was combined with another giv, then we can only split
2815 this giv if the giv it was combined with was reduced. This
2816 is because the value of v->new_reg is meaningless in this
2818 if (v->same && ! v->same->new_reg)
2820 if (loop_dump_stream)
2821 fprintf (loop_dump_stream,
2822 "giv combined with unreduced giv not split.\n");
2825 /* If the giv is an address destination, it could be something other
2826 than a simple register, these have to be treated differently. */
2827 else if (v->giv_type == DEST_REG)
2829 /* If value is not a constant, register, or register plus
2830 constant, then compute its value into a register before
2831 loop start. This prevents invalid rtx sharing, and should
2832 generate better code. We can use bl->initial_value here
2833 instead of splittable_regs[bl->regno] because this code
2834 is going before the loop start. */
2835 if (unroll_type == UNROLL_COMPLETELY
2836 && GET_CODE (value) != CONST_INT
2837 && GET_CODE (value) != REG
2838 && (GET_CODE (value) != PLUS
2839 || GET_CODE (XEXP (value, 0)) != REG
2840 || GET_CODE (XEXP (value, 1)) != CONST_INT))
2842 rtx tem = gen_reg_rtx (v->mode);
2843 record_base_value (REGNO (tem), v->add_val);
2844 emit_iv_add_mult (bl->initial_value, v->mult_val,
2845 v->add_val, tem, loop_start);
2849 splittable_regs[REGNO (v->new_reg)] = value;
2853 /* Splitting address givs is useful since it will often allow us
2854 to eliminate some increment insns for the base giv as
2857 /* If the addr giv is combined with a dest_reg giv, then all
2858 references to that dest reg will be remapped, which is NOT
2859 what we want for split addr regs. We always create a new
2860 register for the split addr giv, just to be safe. */
2862 /* If we have multiple identical address givs within a
2863 single instruction, then use a single pseudo reg for
2864 both. This is necessary in case one is a match_dup
2867 v->const_adjust = 0;
2871 v->dest_reg = v->same_insn->dest_reg;
2872 if (loop_dump_stream)
2873 fprintf (loop_dump_stream,
2874 "Sharing address givs in insn %d\n",
2875 INSN_UID (v->insn));
2877 /* If multiple address GIVs have been combined with the
2878 same dest_reg GIV, do not create a new register for
2880 else if (unroll_type != UNROLL_COMPLETELY
2881 && v->giv_type == DEST_ADDR
2882 && v->same && v->same->giv_type == DEST_ADDR
2883 && v->same->unrolled
2884 /* combine_givs_p may return true for some cases
2885 where the add and mult values are not equal.
2886 To share a register here, the values must be
2888 && rtx_equal_p (v->same->mult_val, v->mult_val)
2889 && rtx_equal_p (v->same->add_val, v->add_val))
2892 v->dest_reg = v->same->dest_reg;
2895 else if (unroll_type != UNROLL_COMPLETELY)
2897 /* If not completely unrolling the loop, then create a new
2898 register to hold the split value of the DEST_ADDR giv.
2899 Emit insn to initialize its value before loop start. */
2901 rtx tem = gen_reg_rtx (v->mode);
2902 record_base_value (REGNO (tem), v->add_val);
2905 /* If the address giv has a constant in its new_reg value,
2906 then this constant can be pulled out and put in value,
2907 instead of being part of the initialization code. */
2909 if (GET_CODE (v->new_reg) == PLUS
2910 && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT)
2913 = plus_constant (tem, INTVAL (XEXP (v->new_reg,1)));
2915 /* Only succeed if this will give valid addresses.
2916 Try to validate both the first and the last
2917 address resulting from loop unrolling, if
2918 one fails, then can't do const elim here. */
2919 if (verify_addresses (v, giv_inc, unroll_number))
2921 /* Save the negative of the eliminated const, so
2922 that we can calculate the dest_reg's increment
2924 v->const_adjust = - INTVAL (XEXP (v->new_reg, 1));
2926 v->new_reg = XEXP (v->new_reg, 0);
2927 if (loop_dump_stream)
2928 fprintf (loop_dump_stream,
2929 "Eliminating constant from giv %d\n",
2938 /* If the address hasn't been checked for validity yet, do so
2939 now, and fail completely if either the first or the last
2940 unrolled copy of the address is not a valid address
2941 for the instruction that uses it. */
2942 if (v->dest_reg == tem
2943 && ! verify_addresses (v, giv_inc, unroll_number))
2945 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2946 if (v2->same_insn == v)
2949 if (loop_dump_stream)
2950 fprintf (loop_dump_stream,
2951 "Invalid address for giv at insn %d\n",
2952 INSN_UID (v->insn));
2956 /* To initialize the new register, just move the value of
2957 new_reg into it. This is not guaranteed to give a valid
2958 instruction on machines with complex addressing modes.
2959 If we can't recognize it, then delete it and emit insns
2960 to calculate the value from scratch. */
2961 emit_insn_before (gen_rtx_SET (VOIDmode, tem,
2962 copy_rtx (v->new_reg)),
2964 if (recog_memoized (PREV_INSN (loop_start)) < 0)
2968 /* We can't use bl->initial_value to compute the initial
2969 value, because the loop may have been preconditioned.
2970 We must calculate it from NEW_REG. Try using
2971 force_operand instead of emit_iv_add_mult. */
2972 delete_insn (PREV_INSN (loop_start));
2975 ret = force_operand (v->new_reg, tem);
2977 emit_move_insn (tem, ret);
2978 sequence = gen_sequence ();
2980 emit_insn_before (sequence, loop_start);
2982 if (loop_dump_stream)
2983 fprintf (loop_dump_stream,
2984 "Invalid init insn, rewritten.\n");
2989 v->dest_reg = value;
2991 /* Check the resulting address for validity, and fail
2992 if the resulting address would be invalid. */
2993 if (! verify_addresses (v, giv_inc, unroll_number))
2995 for (v2 = v->next_iv; v2; v2 = v2->next_iv)
2996 if (v2->same_insn == v)
2999 if (loop_dump_stream)
3000 fprintf (loop_dump_stream,
3001 "Invalid address for giv at insn %d\n",
3002 INSN_UID (v->insn));
3007 /* Store the value of dest_reg into the insn. This sharing
3008 will not be a problem as this insn will always be copied
3011 *v->location = v->dest_reg;
3013 /* If this address giv is combined with a dest reg giv, then
3014 save the base giv's induction pointer so that we will be
3015 able to handle this address giv properly. The base giv
3016 itself does not have to be splittable. */
3018 if (v->same && v->same->giv_type == DEST_REG)
3019 addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
3021 if (GET_CODE (v->new_reg) == REG)
3023 /* This giv maybe hasn't been combined with any others.
3024 Make sure that it's giv is marked as splittable here. */
3026 splittable_regs[REGNO (v->new_reg)] = value;
3028 /* Make it appear to depend upon itself, so that the
3029 giv will be properly split in the main loop above. */
3033 addr_combined_regs[REGNO (v->new_reg)] = v;
3037 if (loop_dump_stream)
3038 fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
3044 /* Currently, unreduced giv's can't be split. This is not too much
3045 of a problem since unreduced giv's are not live across loop
3046 iterations anyways. When unrolling a loop completely though,
3047 it makes sense to reduce&split givs when possible, as this will
3048 result in simpler instructions, and will not require that a reg
3049 be live across loop iterations. */
3051 splittable_regs[REGNO (v->dest_reg)] = value;
3052 fprintf (stderr, "Giv %d at insn %d not reduced\n",
3053 REGNO (v->dest_reg), INSN_UID (v->insn));
3059 /* Unreduced givs are only updated once by definition. Reduced givs
3060 are updated as many times as their biv is. Mark it so if this is
3061 a splittable register. Don't need to do anything for address givs
3062 where this may not be a register. */
3064 if (GET_CODE (v->new_reg) == REG)
3068 count = reg_biv_class[REGNO (v->src_reg)]->biv_count;
3070 splittable_regs_updates[REGNO (v->new_reg)] = count;
3075 if (loop_dump_stream)
3079 if (GET_CODE (v->dest_reg) == CONST_INT)
3081 else if (GET_CODE (v->dest_reg) != REG)
3082 regnum = REGNO (XEXP (v->dest_reg, 0));
3084 regnum = REGNO (v->dest_reg);
3085 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
3086 regnum, INSN_UID (v->insn));
3093 /* Try to prove that the register is dead after the loop exits. Trace every
3094 loop exit looking for an insn that will always be executed, which sets
3095 the register to some value, and appears before the first use of the register
3096 is found. If successful, then return 1, otherwise return 0. */
3098 /* ?? Could be made more intelligent in the handling of jumps, so that
3099 it can search past if statements and other similar structures. */
3102 reg_dead_after_loop (reg, loop_start, loop_end)
3103 rtx reg, loop_start, loop_end;
3108 int label_count = 0;
3109 int this_loop_num = uid_loop_num[INSN_UID (loop_start)];
3111 /* In addition to checking all exits of this loop, we must also check
3112 all exits of inner nested loops that would exit this loop. We don't
3113 have any way to identify those, so we just give up if there are any
3114 such inner loop exits. */
3116 for (label = loop_number_exit_labels[this_loop_num]; label;
3117 label = LABEL_NEXTREF (label))
3120 if (label_count != loop_number_exit_count[this_loop_num])
3123 /* HACK: Must also search the loop fall through exit, create a label_ref
3124 here which points to the loop_end, and append the loop_number_exit_labels
3126 label = gen_rtx_LABEL_REF (VOIDmode, loop_end);
3127 LABEL_NEXTREF (label) = loop_number_exit_labels[this_loop_num];
3129 for ( ; label; label = LABEL_NEXTREF (label))
3131 /* Succeed if find an insn which sets the biv or if reach end of
3132 function. Fail if find an insn that uses the biv, or if come to
3133 a conditional jump. */
3135 insn = NEXT_INSN (XEXP (label, 0));
3138 code = GET_CODE (insn);
3139 if (GET_RTX_CLASS (code) == 'i')
3143 if (reg_referenced_p (reg, PATTERN (insn)))
3146 set = single_set (insn);
3147 if (set && rtx_equal_p (SET_DEST (set), reg))
3151 if (code == JUMP_INSN)
3153 if (GET_CODE (PATTERN (insn)) == RETURN)
3155 else if (! simplejump_p (insn)
3156 /* Prevent infinite loop following infinite loops. */
3157 || jump_count++ > 20)
3160 insn = JUMP_LABEL (insn);
3163 insn = NEXT_INSN (insn);
3167 /* Success, the register is dead on all loop exits. */
3171 /* Try to calculate the final value of the biv, the value it will have at
3172 the end of the loop. If we can do it, return that value. */
3175 final_biv_value (bl, loop_start, loop_end)
3176 struct iv_class *bl;
3177 rtx loop_start, loop_end;
3181 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3183 if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
3186 /* The final value for reversed bivs must be calculated differently than
3187 for ordinary bivs. In this case, there is already an insn after the
3188 loop which sets this biv's final value (if necessary), and there are
3189 no other loop exits, so we can return any value. */
3192 if (loop_dump_stream)
3193 fprintf (loop_dump_stream,
3194 "Final biv value for %d, reversed biv.\n", bl->regno);
3199 /* Try to calculate the final value as initial value + (number of iterations
3200 * increment). For this to work, increment must be invariant, the only
3201 exit from the loop must be the fall through at the bottom (otherwise
3202 it may not have its final value when the loop exits), and the initial
3203 value of the biv must be invariant. */
3205 if (loop_n_iterations != 0
3206 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]
3207 && invariant_p (bl->initial_value))
3209 increment = biv_total_increment (bl, loop_start, loop_end);
3211 if (increment && invariant_p (increment))
3213 /* Can calculate the loop exit value, emit insns after loop
3214 end to calculate this value into a temporary register in
3215 case it is needed later. */
3217 tem = gen_reg_rtx (bl->biv->mode);
3218 record_base_value (REGNO (tem), bl->biv->add_val);
3219 /* Make sure loop_end is not the last insn. */
3220 if (NEXT_INSN (loop_end) == 0)
3221 emit_note_after (NOTE_INSN_DELETED, loop_end);
3222 emit_iv_add_mult (increment, GEN_INT (loop_n_iterations),
3223 bl->initial_value, tem, NEXT_INSN (loop_end));
3225 if (loop_dump_stream)
3226 fprintf (loop_dump_stream,
3227 "Final biv value for %d, calculated.\n", bl->regno);
3233 /* Check to see if the biv is dead at all loop exits. */
3234 if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end))
3236 if (loop_dump_stream)
3237 fprintf (loop_dump_stream,
3238 "Final biv value for %d, biv dead after loop exit.\n",
3247 /* Try to calculate the final value of the giv, the value it will have at
3248 the end of the loop. If we can do it, return that value. */
3251 final_giv_value (v, loop_start, loop_end)
3252 struct induction *v;
3253 rtx loop_start, loop_end;
3255 struct iv_class *bl;
3258 rtx insert_before, seq;
3260 bl = reg_biv_class[REGNO (v->src_reg)];
3262 /* The final value for givs which depend on reversed bivs must be calculated
3263 differently than for ordinary givs. In this case, there is already an
3264 insn after the loop which sets this giv's final value (if necessary),
3265 and there are no other loop exits, so we can return any value. */
3268 if (loop_dump_stream)
3269 fprintf (loop_dump_stream,
3270 "Final giv value for %d, depends on reversed biv\n",
3271 REGNO (v->dest_reg));
3275 /* Try to calculate the final value as a function of the biv it depends
3276 upon. The only exit from the loop must be the fall through at the bottom
3277 (otherwise it may not have its final value when the loop exits). */
3279 /* ??? Can calculate the final giv value by subtracting off the
3280 extra biv increments times the giv's mult_val. The loop must have
3281 only one exit for this to work, but the loop iterations does not need
3284 if (loop_n_iterations != 0
3285 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
3287 /* ?? It is tempting to use the biv's value here since these insns will
3288 be put after the loop, and hence the biv will have its final value
3289 then. However, this fails if the biv is subsequently eliminated.
3290 Perhaps determine whether biv's are eliminable before trying to
3291 determine whether giv's are replaceable so that we can use the
3292 biv value here if it is not eliminable. */
3294 /* We are emitting code after the end of the loop, so we must make
3295 sure that bl->initial_value is still valid then. It will still
3296 be valid if it is invariant. */
3298 increment = biv_total_increment (bl, loop_start, loop_end);
3300 if (increment && invariant_p (increment)
3301 && invariant_p (bl->initial_value))
3303 /* Can calculate the loop exit value of its biv as
3304 (loop_n_iterations * increment) + initial_value */
3306 /* The loop exit value of the giv is then
3307 (final_biv_value - extra increments) * mult_val + add_val.
3308 The extra increments are any increments to the biv which
3309 occur in the loop after the giv's value is calculated.
3310 We must search from the insn that sets the giv to the end
3311 of the loop to calculate this value. */
3313 insert_before = NEXT_INSN (loop_end);
3315 /* Put the final biv value in tem. */
3316 tem = gen_reg_rtx (bl->biv->mode);
3317 record_base_value (REGNO (tem), bl->biv->add_val);
3318 emit_iv_add_mult (increment, GEN_INT (loop_n_iterations),
3319 bl->initial_value, tem, insert_before);
3321 /* Subtract off extra increments as we find them. */
3322 for (insn = NEXT_INSN (v->insn); insn != loop_end;
3323 insn = NEXT_INSN (insn))
3325 struct induction *biv;
3327 for (biv = bl->biv; biv; biv = biv->next_iv)
3328 if (biv->insn == insn)
3331 tem = expand_binop (GET_MODE (tem), sub_optab, tem,
3332 biv->add_val, NULL_RTX, 0,
3334 seq = gen_sequence ();
3336 emit_insn_before (seq, insert_before);
3340 /* Now calculate the giv's final value. */
3341 emit_iv_add_mult (tem, v->mult_val, v->add_val, tem,
3344 if (loop_dump_stream)
3345 fprintf (loop_dump_stream,
3346 "Final giv value for %d, calc from biv's value.\n",
3347 REGNO (v->dest_reg));
3353 /* Replaceable giv's should never reach here. */
3357 /* Check to see if the biv is dead at all loop exits. */
3358 if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end))
3360 if (loop_dump_stream)
3361 fprintf (loop_dump_stream,
3362 "Final giv value for %d, giv dead after loop exit.\n",
3363 REGNO (v->dest_reg));
3372 /* Calculate the number of loop iterations. Returns the exact number of loop
3373 iterations if it can be calculated, otherwise returns zero. */
3375 unsigned HOST_WIDE_INT
3376 loop_iterations (loop_start, loop_end)
3377 rtx loop_start, loop_end;
3379 rtx comparison, comparison_value;
3380 rtx iteration_var, initial_value, increment, final_value;
3381 enum rtx_code comparison_code;
3384 int unsigned_compare, compare_dir, final_larger;
3385 unsigned long tempu;
3388 /* First find the iteration variable. If the last insn is a conditional
3389 branch, and the insn before tests a register value, make that the
3390 iteration variable. */
3392 loop_initial_value = 0;
3394 loop_final_value = 0;
3395 loop_iteration_var = 0;
3397 /* We used to use pren_nonnote_insn here, but that fails because it might
3398 accidentally get the branch for a contained loop if the branch for this
3399 loop was deleted. We can only trust branches immediately before the
3401 last_loop_insn = PREV_INSN (loop_end);
3403 comparison = get_condition_for_loop (last_loop_insn);
3404 if (comparison == 0)
3406 if (loop_dump_stream)
3407 fprintf (loop_dump_stream,
3408 "Loop unrolling: No final conditional branch found.\n");
3412 /* ??? Get_condition may switch position of induction variable and
3413 invariant register when it canonicalizes the comparison. */
3415 comparison_code = GET_CODE (comparison);
3416 iteration_var = XEXP (comparison, 0);
3417 comparison_value = XEXP (comparison, 1);
3419 if (GET_CODE (iteration_var) != REG)
3421 if (loop_dump_stream)
3422 fprintf (loop_dump_stream,
3423 "Loop unrolling: Comparison not against register.\n");
3427 /* Loop iterations is always called before any new registers are created
3428 now, so this should never occur. */
3430 if (REGNO (iteration_var) >= max_reg_before_loop)
3433 iteration_info (iteration_var, &initial_value, &increment,
3434 loop_start, loop_end);
3435 if (initial_value == 0)
3436 /* iteration_info already printed a message. */
3439 /* If the comparison value is an invariant register, then try to find
3440 its value from the insns before the start of the loop. */
3442 if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value))
3446 for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn))
3448 if (GET_CODE (insn) == CODE_LABEL)
3451 else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
3452 && reg_set_p (comparison_value, insn))
3454 /* We found the last insn before the loop that sets the register.
3455 If it sets the entire register, and has a REG_EQUAL note,
3456 then use the value of the REG_EQUAL note. */
3457 if ((set = single_set (insn))
3458 && (SET_DEST (set) == comparison_value))
3460 rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3462 /* Only use the REG_EQUAL note if it is a constant.
3463 Other things, divide in particular, will cause
3464 problems later if we use them. */
3465 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST
3466 && CONSTANT_P (XEXP (note, 0)))
3467 comparison_value = XEXP (note, 0);
3474 final_value = approx_final_value (comparison_code, comparison_value,
3475 &unsigned_compare, &compare_dir);
3477 /* Save the calculated values describing this loop's bounds, in case
3478 precondition_loop_p will need them later. These values can not be
3479 recalculated inside precondition_loop_p because strength reduction
3480 optimizations may obscure the loop's structure. */
3482 loop_iteration_var = iteration_var;
3483 loop_initial_value = initial_value;
3484 loop_increment = increment;
3485 loop_final_value = final_value;
3486 loop_comparison_code = comparison_code;
3490 if (loop_dump_stream)
3491 fprintf (loop_dump_stream,
3492 "Loop unrolling: Increment value can't be calculated.\n");
3495 else if (GET_CODE (increment) != CONST_INT)
3497 if (loop_dump_stream)
3498 fprintf (loop_dump_stream,
3499 "Loop unrolling: Increment value not constant.\n");
3502 else if (GET_CODE (initial_value) != CONST_INT)
3504 if (loop_dump_stream)
3505 fprintf (loop_dump_stream,
3506 "Loop unrolling: Initial value not constant.\n");
3509 else if (final_value == 0)
3511 if (loop_dump_stream)
3512 fprintf (loop_dump_stream,
3513 "Loop unrolling: EQ comparison loop.\n");
3516 else if (GET_CODE (final_value) != CONST_INT)
3518 if (loop_dump_stream)
3519 fprintf (loop_dump_stream,
3520 "Loop unrolling: Final value not constant.\n");
3524 /* ?? Final value and initial value do not have to be constants.
3525 Only their difference has to be constant. When the iteration variable
3526 is an array address, the final value and initial value might both
3527 be addresses with the same base but different constant offsets.
3528 Final value must be invariant for this to work.
3530 To do this, need some way to find the values of registers which are
3533 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3534 if (unsigned_compare)
3536 = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3537 > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
3538 - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
3539 < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
3541 final_larger = (INTVAL (final_value) > INTVAL (initial_value))
3542 - (INTVAL (final_value) < INTVAL (initial_value));
3544 if (INTVAL (increment) > 0)
3546 else if (INTVAL (increment) == 0)
3551 /* There are 27 different cases: compare_dir = -1, 0, 1;
3552 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3553 There are 4 normal cases, 4 reverse cases (where the iteration variable
3554 will overflow before the loop exits), 4 infinite loop cases, and 15
3555 immediate exit (0 or 1 iteration depending on loop type) cases.
3556 Only try to optimize the normal cases. */
3558 /* (compare_dir/final_larger/increment_dir)
3559 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3560 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3561 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3562 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3564 /* ?? If the meaning of reverse loops (where the iteration variable
3565 will overflow before the loop exits) is undefined, then could
3566 eliminate all of these special checks, and just always assume
3567 the loops are normal/immediate/infinite. Note that this means
3568 the sign of increment_dir does not have to be known. Also,
3569 since it does not really hurt if immediate exit loops or infinite loops
3570 are optimized, then that case could be ignored also, and hence all
3571 loops can be optimized.
3573 According to ANSI Spec, the reverse loop case result is undefined,
3574 because the action on overflow is undefined.
3576 See also the special test for NE loops below. */
3578 if (final_larger == increment_dir && final_larger != 0
3579 && (final_larger == compare_dir || compare_dir == 0))
3584 if (loop_dump_stream)
3585 fprintf (loop_dump_stream,
3586 "Loop unrolling: Not normal loop.\n");
3590 /* Calculate the number of iterations, final_value is only an approximation,
3591 so correct for that. Note that tempu and loop_n_iterations are
3592 unsigned, because they can be as large as 2^n - 1. */
3594 i = INTVAL (increment);
3596 tempu = INTVAL (final_value) - INTVAL (initial_value);
3599 tempu = INTVAL (initial_value) - INTVAL (final_value);
3605 /* For NE tests, make sure that the iteration variable won't miss the
3606 final value. If tempu mod i is not zero, then the iteration variable
3607 will overflow before the loop exits, and we can not calculate the
3608 number of iterations. */
3609 if (compare_dir == 0 && (tempu % i) != 0)
3612 return tempu / i + ((tempu % i) != 0);
3615 /* Replace uses of split bivs with their split pseudo register. This is
3616 for original instructions which remain after loop unrolling without
3620 remap_split_bivs (x)
3623 register enum rtx_code code;
3630 code = GET_CODE (x);
3645 /* If non-reduced/final-value givs were split, then this would also
3646 have to remap those givs also. */
3648 if (REGNO (x) < max_reg_before_loop
3649 && reg_iv_type[REGNO (x)] == BASIC_INDUCT)
3650 return reg_biv_class[REGNO (x)]->biv->src_reg;
3657 fmt = GET_RTX_FORMAT (code);
3658 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3661 XEXP (x, i) = remap_split_bivs (XEXP (x, i));
3665 for (j = 0; j < XVECLEN (x, i); j++)
3666 XVECEXP (x, i, j) = remap_split_bivs (XVECEXP (x, i, j));
3672 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3673 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3674 return 0. COPY_START is where we can start looking for the insns
3675 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3678 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3679 must dominate LAST_UID.
3681 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3682 may not dominate LAST_UID.
3684 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3685 must dominate LAST_UID. */
3688 set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end)
3695 int passed_jump = 0;
3696 rtx p = NEXT_INSN (copy_start);
3698 while (INSN_UID (p) != first_uid)
3700 if (GET_CODE (p) == JUMP_INSN)
3702 /* Could not find FIRST_UID. */
3708 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3709 if (GET_RTX_CLASS (GET_CODE (p)) != 'i'
3710 || ! dead_or_set_regno_p (p, regno))
3713 /* FIRST_UID is always executed. */
3714 if (passed_jump == 0)
3717 while (INSN_UID (p) != last_uid)
3719 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3720 can not be sure that FIRST_UID dominates LAST_UID. */
3721 if (GET_CODE (p) == CODE_LABEL)
3723 /* Could not find LAST_UID, but we reached the end of the loop, so
3725 else if (p == copy_end)
3730 /* FIRST_UID is always executed if LAST_UID is executed. */