1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */
21 /* Try to unroll a loop, and split induction variables.
23 Loops for which the number of iterations can be calculated exactly are
24 handled specially. If the number of iterations times the insn_count is
25 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
26 Otherwise, we try to unroll the loop a number of times modulo the number
27 of iterations, so that only one exit test will be needed. It is unrolled
28 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
31 Otherwise, if the number of iterations can be calculated exactly at
32 run time, and the loop is always entered at the top, then we try to
33 precondition the loop. That is, at run time, calculate how many times
34 the loop will execute, and then execute the loop body a few times so
35 that the remaining iterations will be some multiple of 4 (or 2 if the
36 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
37 with only one exit test needed at the end of the loop.
39 Otherwise, if the number of iterations can not be calculated exactly,
40 not even at run time, then we still unroll the loop a number of times
41 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
42 but there must be an exit test after each copy of the loop body.
44 For each induction variable, which is dead outside the loop (replaceable)
45 or for which we can easily calculate the final value, if we can easily
46 calculate its value at each place where it is set as a function of the
47 current loop unroll count and the variable's value at loop entry, then
48 the induction variable is split into `N' different variables, one for
49 each copy of the loop body. One variable is live across the backward
50 branch, and the others are all calculated as a function of this variable.
51 This helps eliminate data dependencies, and leads to further opportunities
54 /* Possible improvements follow: */
56 /* ??? Add an extra pass somewhere to determine whether unrolling will
57 give any benefit. E.g. after generating all unrolled insns, compute the
58 cost of all insns and compare against cost of insns in rolled loop.
60 - On traditional architectures, unrolling a non-constant bound loop
61 is a win if there is a giv whose only use is in memory addresses, the
62 memory addresses can be split, and hence giv incremenets can be
64 - It is also a win if the loop is executed many times, and preconditioning
65 can be performed for the loop.
66 Add code to check for these and similar cases. */
68 /* ??? Improve control of which loops get unrolled. Could use profiling
69 info to only unroll the most commonly executed loops. Perhaps have
70 a user specifyable option to control the amount of code expansion,
71 or the percent of loops to consider for unrolling. Etc. */
73 /* ??? Look at the register copies inside the loop to see if they form a
74 simple permutation. If so, iterate the permutation until it gets back to
75 the start state. This is how many times we should unroll the loop, for
76 best results, because then all register copies can be eliminated.
77 For example, the lisp nreverse function should be unrolled 3 times
86 ??? The number of times to unroll the loop may also be based on data
87 references in the loop. For example, if we have a loop that references
88 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
90 /* ??? Add some simple linear equation solving capability so that we can
91 determine the number of loop iterations for more complex loops.
92 For example, consider this loop from gdb
93 #define SWAP_TARGET_AND_HOST(buffer,len)
96 char *p = (char *) buffer;
97 char *q = ((char *) buffer) + len - 1;
98 int iterations = (len + 1) >> 1;
100 for (p; p < q; p++, q--;)
108 start value = p = &buffer + current_iteration
109 end value = q = &buffer + len - 1 - current_iteration
110 Given the loop exit test of "p < q", then there must be "q - p" iterations,
111 set equal to zero and solve for number of iterations:
112 q - p = len - 1 - 2*current_iteration = 0
113 current_iteration = (len - 1) / 2
114 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
115 iterations of this loop. */
117 /* ??? Currently, no labels are marked as loop invariant when doing loop
118 unrolling. This is because an insn inside the loop, that loads the address
119 of a label inside the loop into a register, could be moved outside the loop
120 by the invariant code motion pass if labels were invariant. If the loop
121 is subsequently unrolled, the code will be wrong because each unrolled
122 body of the loop will use the same address, whereas each actually needs a
123 different address. A case where this happens is when a loop containing
124 a switch statement is unrolled.
126 It would be better to let labels be considered invariant. When we
127 unroll loops here, check to see if any insns using a label local to the
128 loop were moved before the loop. If so, then correct the problem, by
129 moving the insn back into the loop, or perhaps replicate the insn before
130 the loop, one copy for each time the loop is unrolled. */
132 /* The prime factors looked for when trying to unroll a loop by some
133 number which is modulo the total number of iterations. Just checking
134 for these 4 prime factors will find at least one factor for 75% of
135 all numbers theoretically. Practically speaking, this will succeed
136 almost all of the time since loops are generally a multiple of 2
139 #define NUM_FACTORS 4
141 struct _factor { int factor, count; } factors[NUM_FACTORS]
142 = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
144 /* Describes the different types of loop unrolling performed. */
146 enum unroll_types { UNROLL_COMPLETELY, UNROLL_MODULO, UNROLL_NAIVE };
150 #include "insn-config.h"
151 #include "integrate.h"
158 /* This controls which loops are unrolled, and by how much we unroll
161 #ifndef MAX_UNROLLED_INSNS
162 #define MAX_UNROLLED_INSNS 100
165 /* Indexed by register number, if non-zero, then it contains a pointer
166 to a struct induction for a DEST_REG giv which has been combined with
167 one of more address givs. This is needed because whenever such a DEST_REG
168 giv is modified, we must modify the value of all split address givs
169 that were combined with this DEST_REG giv. */
171 static struct induction **addr_combined_regs;
173 /* Indexed by register number, if this is a splittable induction variable,
174 then this will hold the current value of the register, which depends on the
177 static rtx *splittable_regs;
179 /* Indexed by register number, if this is a splittable induction variable,
180 then this will hold the number of instructions in the loop that modify
181 the induction variable. Used to ensure that only the last insn modifying
182 a split iv will update the original iv of the dest. */
184 static int *splittable_regs_updates;
186 /* Values describing the current loop's iteration variable. These are set up
187 by loop_iterations, and used by precondition_loop_p. */
189 static rtx loop_iteration_var;
190 static rtx loop_initial_value;
191 static rtx loop_increment;
192 static rtx loop_final_value;
194 /* Forward declarations. */
196 static void init_reg_map ();
197 static int precondition_loop_p ();
198 static void copy_loop_body ();
199 static void iteration_info ();
200 static rtx approx_final_value ();
201 static int find_splittable_regs ();
202 static int find_splittable_givs ();
203 static rtx fold_rtx_mult_add ();
205 /* Try to unroll one loop and split induction variables in the loop.
207 The loop is described by the arguments LOOP_END, INSN_COUNT, and
208 LOOP_START. END_INSERT_BEDFORE indicates where insns should be added
209 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
210 indicates whether information generated in the strength reduction pass
213 This function is intended to be called from within `strength_reduce'
217 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
222 rtx end_insert_before;
223 int strength_reduce_p;
226 int unroll_number = 1;
227 rtx copy_start, copy_end;
228 rtx insn, copy, sequence, pattern, tem;
229 int max_labelno, max_insnno;
231 struct inline_remap *map;
239 int splitting_not_safe = 0;
240 enum unroll_types unroll_type;
241 int loop_preconditioned = 0;
243 /* This points to the last real insn in the loop, which should be either
244 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
248 /* Don't bother unrolling huge loops. Since the minimum factor is
249 two, loops greater than one half of MAX_UNROLLED_INSNS will never
251 if (insn_count > MAX_UNROLLED_INSNS / 2)
253 if (loop_dump_stream)
254 fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
258 /* When emitting debugger info, we can't unroll loops with unequal numbers
259 of block_beg and block_end notes, because that would unbalance the block
260 structure of the function. This can happen as a result of the
261 "if (foo) bar; else break;" optimization in jump.c. */
263 if (write_symbols != NO_DEBUG)
265 int block_begins = 0;
268 for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn))
270 if (GET_CODE (insn) == NOTE)
272 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG)
274 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END)
279 if (block_begins != block_ends)
281 if (loop_dump_stream)
282 fprintf (loop_dump_stream,
283 "Unrolling failure: Unbalanced block notes.\n");
288 /* Determine type of unroll to perform. Depends on the number of iterations
289 and the size of the loop. */
291 /* If there is no strength reduce info, then set loop_n_iterations to zero.
292 This can happen if strength_reduce can't find any bivs in the loop.
293 A value of zero indicates that the number of iterations could not be
296 if (! strength_reduce_p)
297 loop_n_iterations = 0;
299 if (loop_dump_stream && loop_n_iterations > 0)
300 fprintf (loop_dump_stream,
301 "Loop unrolling: %d iterations.\n", loop_n_iterations);
303 /* Find and save a pointer to the last nonnote insn in the loop. */
305 last_loop_insn = prev_nonnote_insn (loop_end);
307 /* Calculate how many times to unroll the loop. Indicate whether or
308 not the loop is being completely unrolled. */
310 if (loop_n_iterations == 1)
312 /* If number of iterations is exactly 1, then eliminate the compare and
313 branch at the end of the loop since they will never be taken.
314 Then return, since no other action is needed here. */
316 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
317 don't do anything. */
319 if (GET_CODE (last_loop_insn) == BARRIER)
321 /* Delete the jump insn. This will delete the barrier also. */
322 delete_insn (PREV_INSN (last_loop_insn));
324 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
327 /* The immediately preceeding insn is a compare which must be
329 delete_insn (last_loop_insn);
330 delete_insn (PREV_INSN (last_loop_insn));
332 /* The immediately preceeding insn may not be the compare, so don't
334 delete_insn (last_loop_insn);
339 else if (loop_n_iterations > 0
340 && loop_n_iterations * insn_count < MAX_UNROLLED_INSNS)
342 unroll_number = loop_n_iterations;
343 unroll_type = UNROLL_COMPLETELY;
345 else if (loop_n_iterations > 0)
347 /* Try to factor the number of iterations. Don't bother with the
348 general case, only using 2, 3, 5, and 7 will get 75% of all
349 numbers theoretically, and almost all in practice. */
351 for (i = 0; i < NUM_FACTORS; i++)
352 factors[i].count = 0;
354 temp = loop_n_iterations;
355 for (i = NUM_FACTORS - 1; i >= 0; i--)
356 while (temp % factors[i].factor == 0)
359 temp = temp / factors[i].factor;
362 /* Start with the larger factors first so that we generally
363 get lots of unrolling. */
367 for (i = 3; i >= 0; i--)
368 while (factors[i].count--)
370 if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
372 unroll_number *= factors[i].factor;
373 temp *= factors[i].factor;
379 /* If we couldn't find any factors, then unroll as in the normal
381 if (unroll_number == 1)
383 if (loop_dump_stream)
384 fprintf (loop_dump_stream,
385 "Loop unrolling: No factors found.\n");
388 unroll_type = UNROLL_MODULO;
392 /* Default case, calculate number of times to unroll loop based on its
394 if (unroll_number == 1)
396 if (8 * insn_count < MAX_UNROLLED_INSNS)
398 else if (4 * insn_count < MAX_UNROLLED_INSNS)
403 unroll_type = UNROLL_NAIVE;
406 /* Now we know how many times to unroll the loop. */
408 if (loop_dump_stream)
409 fprintf (loop_dump_stream,
410 "Unrolling loop %d times.\n", unroll_number);
413 if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
415 /* Loops of these types should never start with a jump down to
416 the exit condition test. For now, check for this case just to
417 be sure. UNROLL_NAIVE loops can be of this form, this case is
420 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
421 insn = NEXT_INSN (insn);
422 if (GET_CODE (insn) == JUMP_INSN)
426 if (unroll_type == UNROLL_COMPLETELY)
428 /* Completely unrolling the loop: Delete the compare and branch at
429 the end (the last two instructions). This delete must done at the
430 very end of loop unrolling, to avoid problems with calls to
431 back_branch_in_range_p, which is called by find_splittable_regs.
432 All increments of splittable bivs/givs are changed to load constant
435 copy_start = loop_start;
437 /* Set insert_before to the instruction immediately after the JUMP_INSN
438 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
439 the loop will be correctly handled by copy_loop_body. */
440 insert_before = NEXT_INSN (last_loop_insn);
442 /* Set copy_end to the insn before the jump at the end of the loop. */
443 if (GET_CODE (last_loop_insn) == BARRIER)
444 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
445 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
448 /* The instruction immediately before the JUMP_INSN is a compare
449 instruction which we do not want to copy. */
450 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
452 /* The instruction immediately before the JUMP_INSN may not be the
453 compare, so we must copy it. */
454 copy_end = PREV_INSN (last_loop_insn);
459 /* We currently can't unroll a loop if it doesn't end with a
460 JUMP_INSN. There would need to be a mechanism that recognizes
461 this case, and then inserts a jump after each loop body, which
462 jumps to after the last loop body. */
463 if (loop_dump_stream)
464 fprintf (loop_dump_stream,
465 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
469 else if (unroll_type == UNROLL_MODULO)
471 /* Partially unrolling the loop: The compare and branch at the end
472 (the last two instructions) must remain. Don't copy the compare
473 and branch instructions at the end of the loop. Insert the unrolled
474 code immediately before the compare/branch at the end so that the
475 code will fall through to them as before. */
477 copy_start = loop_start;
479 /* Set insert_before to the jump insn at the end of the loop.
480 Set copy_end to before the jump insn at the end of the loop. */
481 if (GET_CODE (last_loop_insn) == BARRIER)
483 insert_before = PREV_INSN (last_loop_insn);
484 copy_end = PREV_INSN (insert_before);
486 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
489 /* The instruction immediately before the JUMP_INSN is a compare
490 instruction which we do not want to copy or delete. */
491 insert_before = PREV_INSN (last_loop_insn);
492 copy_end = PREV_INSN (insert_before);
494 /* The instruction immediately before the JUMP_INSN may not be the
495 compare, so we must copy it. */
496 insert_before = last_loop_insn;
497 copy_end = PREV_INSN (last_loop_insn);
502 /* We currently can't unroll a loop if it doesn't end with a
503 JUMP_INSN. There would need to be a mechanism that recognizes
504 this case, and then inserts a jump after each loop body, which
505 jumps to after the last loop body. */
506 if (loop_dump_stream)
507 fprintf (loop_dump_stream,
508 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
514 /* Normal case: Must copy the compare and branch instructions at the
517 if (GET_CODE (last_loop_insn) == BARRIER)
519 /* Loop ends with an unconditional jump and a barrier.
520 Handle this like above, don't copy jump and barrier.
521 This is not strictly necessary, but doing so prevents generating
522 unconditional jumps to an immediately following label.
524 This will be corrected below if the target of this jump is
525 not the start_label. */
527 insert_before = PREV_INSN (last_loop_insn);
528 copy_end = PREV_INSN (insert_before);
530 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
532 /* Set insert_before to immediately after the JUMP_INSN, so that
533 NOTEs at the end of the loop will be correctly handled by
535 insert_before = NEXT_INSN (last_loop_insn);
536 copy_end = last_loop_insn;
540 /* We currently can't unroll a loop if it doesn't end with a
541 JUMP_INSN. There would need to be a mechanism that recognizes
542 this case, and then inserts a jump after each loop body, which
543 jumps to after the last loop body. */
544 if (loop_dump_stream)
545 fprintf (loop_dump_stream,
546 "Unrolling failure: loop does not end with a JUMP_INSN.\n");
550 /* If copying exit test branches because they can not be eliminated,
551 then must convert the fall through case of the branch to a jump past
552 the end of the loop. Create a label to emit after the loop and save
553 it for later use. Do not use the label after the loop, if any, since
554 it might be used by insns outside the loop, or there might be insns
555 added before it later by final_[bg]iv_value which must be after
556 the real exit label. */
557 exit_label = gen_label_rtx ();
560 while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
561 insn = NEXT_INSN (insn);
563 if (GET_CODE (insn) == JUMP_INSN)
565 /* The loop starts with a jump down to the exit condition test.
566 Start copying the loop after the barrier following this
568 copy_start = NEXT_INSN (insn);
570 /* Splitting induction variables doesn't work when the loop is
571 entered via a jump to the bottom, because then we end up doing
572 a comparison against a new register for a split variable, but
573 we did not execute the set insn for the new register because
574 it was skipped over. */
575 splitting_not_safe = 1;
576 if (loop_dump_stream)
577 fprintf (loop_dump_stream,
578 "Splitting not safe, because loop not entered at top.\n");
581 copy_start = loop_start;
584 /* This should always be the first label in the loop. */
585 start_label = NEXT_INSN (copy_start);
586 /* There may be a line number note and/or a loop continue note here. */
587 while (GET_CODE (start_label) == NOTE)
588 start_label = NEXT_INSN (start_label);
589 if (GET_CODE (start_label) != CODE_LABEL)
591 /* This can happen as a result of jump threading. If the first insns in
592 the loop test the same condition as the loop's backward jump, or the
593 opposite condition, then the backward jump will be modified to point
594 to elsewhere, and the loop's start label is deleted.
596 This case currently can not be handled by the loop unrolling code. */
598 if (loop_dump_stream)
599 fprintf (loop_dump_stream,
600 "Unrolling failure: unknown insns between BEG note and loop label.\n");
604 if (unroll_type == UNROLL_NAIVE
605 && GET_CODE (last_loop_insn) == BARRIER
606 && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
608 /* In this case, we must copy the jump and barrier, because they will
609 not be converted to jumps to an immediately following label. */
611 insert_before = NEXT_INSN (last_loop_insn);
612 copy_end = last_loop_insn;
615 /* Allocate a translation table for the labels and insn numbers.
616 They will be filled in as we copy the insns in the loop. */
618 max_labelno = max_label_num ();
619 max_insnno = get_max_uid ();
621 map = (struct inline_remap *) alloca (sizeof (struct inline_remap));
623 /* Allocate the label map. */
627 map->label_map = (rtx *) alloca (max_labelno * sizeof (rtx));
629 local_label = (char *) alloca (max_labelno);
630 bzero (local_label, max_labelno);
635 /* Search the loop and mark all local labels, i.e. the ones which have to
636 be distinct labels when copied. For all labels which might be
637 non-local, set their label_map entries to point to themselves.
638 If they happen to be local their label_map entries will be overwritten
639 before the loop body is copied. The label_map entries for local labels
640 will be set to a different value each time the loop body is copied. */
642 for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
644 if (GET_CODE (insn) == CODE_LABEL)
645 local_label[CODE_LABEL_NUMBER (insn)] = 1;
646 else if (GET_CODE (insn) == JUMP_INSN)
648 if (JUMP_LABEL (insn))
649 map->label_map[CODE_LABEL_NUMBER (JUMP_LABEL (insn))]
651 else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
652 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
654 rtx pat = PATTERN (insn);
655 int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
656 int len = XVECLEN (pat, diff_vec_p);
659 for (i = 0; i < len; i++)
661 label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
662 map->label_map[CODE_LABEL_NUMBER (label)] = label;
668 /* Allocate space for the insn map. */
670 map->insn_map = (rtx *) alloca (max_insnno * sizeof (rtx));
672 /* Set this to zero, to indicate that we are doing loop unrolling,
673 not function inlining. */
674 map->inline_target = 0;
676 /* The register and constant maps depend on the number of registers
677 present, so the final maps can't be created until after
678 find_splittable_regs is called. However, they are needed for
679 preconditioning, so we create temporary maps when preconditioning
682 /* The preconditioning code may allocate two new pseudo registers. */
683 maxregnum = max_reg_num ();
685 /* Allocate and zero out the splittable_regs and addr_combined_regs
686 arrays. These must be zeroed here because they will be used if
687 loop preconditioning is performed, and must be zero for that case.
689 It is safe to do this here, since the extra registers created by the
690 preconditioning code and find_splittable_regs will never be used
691 to accees the splittable_regs[] and addr_combined_regs[] arrays. */
693 splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx));
694 bzero (splittable_regs, maxregnum * sizeof (rtx));
695 splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int));
696 bzero (splittable_regs_updates, maxregnum * sizeof (int));
698 = (struct induction **) alloca (maxregnum * sizeof (struct induction *));
699 bzero (addr_combined_regs, maxregnum * sizeof (struct induction *));
701 /* If this loop requires exit tests when unrolled, check to see if we
702 can precondition the loop so as to make the exit tests unnecessary.
703 Just like variable splitting, this is not safe if the loop is entered
704 via a jump to the bottom. Also, can not do this if no strength
705 reduce info, because precondition_loop_p uses this info. */
707 /* Must copy the loop body for preconditioning before the following
708 find_splittable_regs call since that will emit insns which need to
709 be after the preconditioned loop copies, but immediately before the
710 unrolled loop copies. */
712 /* Also, it is not safe to split induction variables for the preconditioned
713 copies of the loop body. If we split induction variables, then the code
714 assumes that each induction variable can be represented as a function
715 of its initial value and the loop iteration number. This is not true
716 in this case, because the last preconditioned copy of the loop body
717 could be any iteration from the first up to the `unroll_number-1'th,
718 depending on the initial value of the iteration variable. Therefore
719 we can not split induction variables here, because we can not calculate
720 their value. Hence, this code must occur before find_splittable_regs
723 if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
725 rtx initial_value, final_value, increment;
727 if (precondition_loop_p (&initial_value, &final_value, &increment,
728 loop_start, loop_end))
730 register rtx diff, temp;
731 enum machine_mode mode;
733 int abs_inc, neg_inc;
735 map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
737 map->const_equiv_map = (rtx *) alloca (maxregnum * sizeof (rtx));
738 map->const_age_map = (unsigned *) alloca (maxregnum
739 * sizeof (unsigned));
740 map->const_equiv_map_size = maxregnum;
741 global_const_equiv_map = map->const_equiv_map;
743 init_reg_map (map, maxregnum);
745 /* Limit loop unrolling to 4, since this will make 7 copies of
747 if (unroll_number > 4)
750 /* Save the absolute value of the increment, and also whether or
751 not it is negative. */
753 abs_inc = INTVAL (increment);
762 /* Decide what mode to do these calculations in. Choose the larger
763 of final_value's mode and initial_value's mode, or a full-word if
764 both are constants. */
765 mode = GET_MODE (final_value);
766 if (mode == VOIDmode)
768 mode = GET_MODE (initial_value);
769 if (mode == VOIDmode)
772 else if (mode != GET_MODE (initial_value)
773 && (GET_MODE_SIZE (mode)
774 < GET_MODE_SIZE (GET_MODE (initial_value))))
775 mode = GET_MODE (initial_value);
777 /* Calculate the difference between the final and initial values.
778 Final value may be a (plus (reg x) (const_int 1)) rtx.
779 Let the following cse pass simplify this if initial value is
782 We must copy the final and initial values here to avoid
783 improperly shared rtl. */
785 diff = expand_binop (mode, sub_optab, copy_rtx (final_value),
786 copy_rtx (initial_value), 0, 0,
789 /* Now calculate (diff % (unroll * abs (increment))) by using an
791 diff = expand_binop (GET_MODE (diff), and_optab, diff,
792 gen_rtx (CONST_INT, VOIDmode,
793 unroll_number * abs_inc - 1),
794 0, 0, OPTAB_LIB_WIDEN);
796 /* Now emit a sequence of branches to jump to the proper precond
799 labels = (rtx *) alloca (sizeof (rtx) * unroll_number);
800 for (i = 0; i < unroll_number; i++)
801 labels[i] = gen_label_rtx ();
803 /* Assuming the unroll_number is 4, and the increment is 2, then
804 for a negative increment: for a positive increment:
805 diff = 0,1 precond 0 diff = 0,7 precond 0
806 diff = 2,3 precond 3 diff = 1,2 precond 1
807 diff = 4,5 precond 2 diff = 3,4 precond 2
808 diff = 6,7 precond 1 diff = 5,6 precond 3 */
810 /* We only need to emit (unroll_number - 1) branches here, the
811 last case just falls through to the following code. */
813 /* ??? This would give better code if we emitted a tree of branches
814 instead of the current linear list of branches. */
816 for (i = 0; i < unroll_number - 1; i++)
820 /* For negative increments, must invert the constant compared
821 against, except when comparing against zero. */
825 cmp_const = unroll_number - i;
829 emit_cmp_insn (diff, gen_rtx (CONST_INT, VOIDmode,
830 abs_inc * cmp_const),
834 emit_jump_insn (gen_beq (labels[i]));
836 emit_jump_insn (gen_bge (labels[i]));
838 emit_jump_insn (gen_ble (labels[i]));
839 JUMP_LABEL (get_last_insn ()) = labels[i];
840 LABEL_NUSES (labels[i])++;
843 /* If the increment is greater than one, then we need another branch,
844 to handle other cases equivalent to 0. */
846 /* ??? This should be merged into the code above somehow to help
847 simplify the code here, and reduce the number of branches emitted.
848 For the negative increment case, the branch here could easily
849 be merged with the `0' case branch above. For the positive
850 increment case, it is not clear how this can be simplified. */
857 cmp_const = abs_inc - 1;
859 cmp_const = abs_inc * (unroll_number - 1) + 1;
861 emit_cmp_insn (diff, gen_rtx (CONST_INT, VOIDmode, cmp_const),
865 emit_jump_insn (gen_ble (labels[0]));
867 emit_jump_insn (gen_bge (labels[0]));
868 JUMP_LABEL (get_last_insn ()) = labels[0];
869 LABEL_NUSES (labels[0])++;
872 sequence = gen_sequence ();
874 emit_insn_before (sequence, loop_start);
876 /* Only the last copy of the loop body here needs the exit
877 test, so set copy_end to exclude the compare/branch here,
878 and then reset it inside the loop when get to the last
881 if (GET_CODE (last_loop_insn) == BARRIER)
882 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
883 else if (GET_CODE (last_loop_insn) == JUMP_INSN)
886 /* The immediately preceeding insn is a compare which we do not
888 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
890 /* The immediately preceeding insn may not be a compare, so we
892 copy_end = PREV_INSN (last_loop_insn);
898 for (i = 1; i < unroll_number; i++)
900 emit_label_after (labels[unroll_number - i],
901 PREV_INSN (loop_start));
903 bzero (map->insn_map, max_insnno * sizeof (rtx));
904 bzero (map->const_equiv_map, maxregnum * sizeof (rtx));
905 bzero (map->const_age_map, maxregnum * sizeof (unsigned));
908 for (j = 0; j < max_labelno; j++)
910 map->label_map[j] = gen_label_rtx ();
912 /* The last copy needs the compare/branch insns at the end,
913 so reset copy_end here if the loop ends with a conditional
916 if (i == unroll_number - 1)
918 if (GET_CODE (last_loop_insn) == BARRIER)
919 copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
921 copy_end = last_loop_insn;
924 /* None of the copies are the `last_iteration', so just
925 pass zero for that parameter. */
926 copy_loop_body (copy_start, copy_end, map, exit_label, 0,
927 unroll_type, start_label, loop_end,
928 loop_start, copy_end);
930 emit_label_after (labels[0], PREV_INSN (loop_start));
932 if (GET_CODE (last_loop_insn) == BARRIER)
934 insert_before = PREV_INSN (last_loop_insn);
935 copy_end = PREV_INSN (insert_before);
940 /* The immediately preceeding insn is a compare which we do not
942 insert_before = PREV_INSN (last_loop_insn);
943 copy_end = PREV_INSN (insert_before);
945 /* The immediately preceeding insn may not be a compare, so we
947 insert_before = last_loop_insn;
948 copy_end = PREV_INSN (last_loop_insn);
952 /* Set unroll type to MODULO now. */
953 unroll_type = UNROLL_MODULO;
954 loop_preconditioned = 1;
958 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
959 the loop unless all loops are being unrolled. */
960 if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
962 if (loop_dump_stream)
963 fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n");
967 /* At this point, we are guaranteed to unroll the loop. */
969 /* For each biv and giv, determine whether it can be safely split into
970 a different variable for each unrolled copy of the loop body.
971 We precalculate and save this info here, since computing it is
974 Do this before deleting any instructions from the loop, so that
975 back_branch_in_range_p will work correctly. */
977 if (splitting_not_safe)
980 temp = find_splittable_regs (unroll_type, loop_start, loop_end,
981 end_insert_before, unroll_number);
983 /* find_splittable_regs may have created some new registers, so must
984 reallocate the reg_map with the new larger size, and must realloc
985 the constant maps also. */
987 maxregnum = max_reg_num ();
988 map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
990 init_reg_map (map, maxregnum);
992 /* Space is needed in some of the map for new registers, so new_maxregnum
993 is an (over)estimate of how many registers will exist at the end. */
994 new_maxregnum = maxregnum + (temp * unroll_number * 2);
996 /* Must realloc space for the constant maps, because the number of registers
999 map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx));
1000 map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned));
1002 global_const_equiv_map = map->const_equiv_map;
1004 /* Search the list of bivs and givs to find ones which need to be remapped
1005 when split, and set their reg_map entry appropriately. */
1007 for (bl = loop_iv_list; bl; bl = bl->next)
1009 if (REGNO (bl->biv->src_reg) != bl->regno)
1010 map->reg_map[bl->regno] = bl->biv->src_reg;
1012 /* Currently, non-reduced/final-value givs are never split. */
1013 for (v = bl->giv; v; v = v->next_iv)
1014 if (REGNO (v->src_reg) != bl->regno)
1015 map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1019 /* If the loop is being partially unrolled, and the iteration variables
1020 are being split, and are being renamed for the split, then must fix up
1021 the compare instruction at the end of the loop to refer to the new
1022 registers. This compare isn't copied, so the registers used in it
1023 will never be replaced if it isn't done here. */
1025 if (unroll_type == UNROLL_MODULO)
1027 insn = NEXT_INSN (copy_end);
1028 if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SET)
1031 /* If non-reduced/final-value givs were split, then this would also
1032 have to remap those givs. */
1035 tem = SET_SRC (PATTERN (insn));
1036 /* The set source is a register. */
1037 if (GET_CODE (tem) == REG)
1039 if (REGNO (tem) < max_reg_before_loop
1040 && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1041 SET_SRC (PATTERN (insn))
1042 = reg_biv_class[REGNO (tem)]->biv->src_reg;
1046 /* The set source is a compare of some sort. */
1047 tem = XEXP (SET_SRC (PATTERN (insn)), 0);
1048 if (GET_CODE (tem) == REG
1049 && REGNO (tem) < max_reg_before_loop
1050 && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1051 XEXP (SET_SRC (PATTERN (insn)), 0)
1052 = reg_biv_class[REGNO (tem)]->biv->src_reg;
1054 tem = XEXP (SET_SRC (PATTERN (insn)), 1);
1055 if (GET_CODE (tem) == REG
1056 && REGNO (tem) < max_reg_before_loop
1057 && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1058 XEXP (SET_SRC (PATTERN (insn)), 1)
1059 = reg_biv_class[REGNO (tem)]->biv->src_reg;
1064 /* For unroll_number - 1 times, make a copy of each instruction
1065 between copy_start and copy_end, and insert these new instructions
1066 before the end of the loop. */
1068 for (i = 0; i < unroll_number; i++)
1070 bzero (map->insn_map, max_insnno * sizeof (rtx));
1071 bzero (map->const_equiv_map, new_maxregnum * sizeof (rtx));
1072 bzero (map->const_age_map, new_maxregnum * sizeof (unsigned));
1075 for (j = 0; j < max_labelno; j++)
1077 map->label_map[j] = gen_label_rtx ();
1079 /* If loop starts with a branch to the test, then fix it so that
1080 it points to the test of the first unrolled copy of the loop. */
1081 if (i == 0 && loop_start != copy_start)
1083 insn = PREV_INSN (copy_start);
1084 pattern = PATTERN (insn);
1086 tem = map->label_map[CODE_LABEL_NUMBER
1087 (XEXP (SET_SRC (pattern), 0))];
1088 SET_SRC (pattern) = gen_rtx (LABEL_REF, VOIDmode, tem);
1090 /* Set the jump label so that it can be used by later loop unrolling
1092 JUMP_LABEL (insn) = tem;
1093 LABEL_NUSES (tem)++;
1096 copy_loop_body (copy_start, copy_end, map, exit_label,
1097 i == unroll_number - 1, unroll_type, start_label,
1098 loop_end, insert_before, insert_before);
1101 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1102 insn to be deleted. This prevents any runaway delete_insn call from
1103 more insns that it should, as it always stops at a CODE_LABEL. */
1105 /* Delete the compare and branch at the end of the loop if completely
1106 unrolling the loop. Deleting the backward branch at the end also
1107 deletes the code label at the start of the loop. This is done at
1108 the very end to avoid problems with back_branch_in_range_p. */
1110 if (unroll_type == UNROLL_COMPLETELY)
1111 safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1113 safety_label = emit_label_after (gen_label_rtx (), copy_end);
1115 /* Delete all of the original loop instructions. Don't delete the
1116 LOOP_BEG note, or the first code label in the loop. */
1118 insn = NEXT_INSN (copy_start);
1119 while (insn != safety_label)
1121 if (insn != start_label)
1122 insn = delete_insn (insn);
1124 insn = NEXT_INSN (insn);
1127 /* Can now delete the 'safety' label emitted to protect us from runaway
1128 delete_insn calls. */
1129 if (INSN_DELETED_P (safety_label))
1131 delete_insn (safety_label);
1133 /* If exit_label exists, emit it after the loop. Doing the emit here
1134 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1135 This is needed so that mostly_true_jump in reorg.c will treat jumps
1136 to this loop end label correctly, i.e. predict that they are usually
1139 emit_label_after (exit_label, loop_end);
1141 /* If debugging, we must replicate the tree nodes corresponsing to the blocks
1142 inside the loop, so that the original one to one mapping will remain. */
1144 if (write_symbols != NO_DEBUG)
1146 int copies = unroll_number;
1148 if (loop_preconditioned)
1149 copies += unroll_number - 1;
1151 unroll_block_trees (uid_loop_num[INSN_UID (loop_start)], copies);
1155 /* Return true if the loop can be safely, and profitably, preconditioned
1156 so that the unrolled copies of the loop body don't need exit tests.
1158 This only works if final_value, initial_value and increment can be
1159 determined, and if increment is a constant power of 2.
1160 If increment is not a power of 2, then the preconditioning modulo
1161 operation would require a real modulo instead of a boolean AND, and this
1162 is not considered `profitable'. */
1164 /* ??? If the loop is known to be executed very many times, or the machine
1165 has a very cheap divide instruction, then preconditioning is a win even
1166 when the increment is not a power of 2. Use RTX_COST to compute
1167 whether divide is cheap. */
1170 precondition_loop_p (initial_value, final_value, increment, loop_start,
1172 rtx *initial_value, *final_value, *increment;
1173 rtx loop_start, loop_end;
1175 int unsigned_compare, compare_dir;
1177 if (loop_n_iterations > 0)
1179 *initial_value = const0_rtx;
1180 *increment = const1_rtx;
1181 *final_value = gen_rtx (CONST_INT, VOIDmode, loop_n_iterations);
1183 if (loop_dump_stream)
1184 fprintf (loop_dump_stream,
1185 "Preconditioning: Success, number of iterations known, %d.\n",
1190 if (loop_initial_value == 0)
1192 if (loop_dump_stream)
1193 fprintf (loop_dump_stream,
1194 "Preconditioning: Could not find initial value.\n");
1197 else if (loop_increment == 0)
1199 if (loop_dump_stream)
1200 fprintf (loop_dump_stream,
1201 "Preconditioning: Could not find increment value.\n");
1204 else if (GET_CODE (loop_increment) != CONST_INT)
1206 if (loop_dump_stream)
1207 fprintf (loop_dump_stream,
1208 "Preconditioning: Increment not a constant.\n");
1211 else if ((exact_log2 (INTVAL (loop_increment)) < 0)
1212 && (exact_log2 (- INTVAL (loop_increment)) < 0))
1214 if (loop_dump_stream)
1215 fprintf (loop_dump_stream,
1216 "Preconditioning: Increment not a constant power of 2.\n");
1220 /* Unsigned_compare and compare_dir can be ignored here, since they do
1221 not matter for preconditioning. */
1223 if (loop_final_value == 0)
1225 if (loop_dump_stream)
1226 fprintf (loop_dump_stream,
1227 "Preconditioning: EQ comparison loop.\n");
1231 /* Must ensure that final_value is invariant, so call invariant_p to
1232 check. Before doing so, must check regno against max_reg_before_loop
1233 to make sure that the register is in the range convered by invariant_p.
1234 If it isn't, then it is most likely a biv/giv which by definition are
1236 if ((GET_CODE (loop_final_value) == REG
1237 && REGNO (loop_final_value) >= max_reg_before_loop)
1238 || (GET_CODE (loop_final_value) == PLUS
1239 && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop)
1240 || ! invariant_p (loop_final_value))
1242 if (loop_dump_stream)
1243 fprintf (loop_dump_stream,
1244 "Preconditioning: Final value not invariant.\n");
1248 /* Fail for floating point values, since the caller of this function
1249 does not have code to deal with them. */
1250 if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT
1251 || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT)
1253 if (loop_dump_stream)
1254 fprintf (loop_dump_stream,
1255 "Preconditioning: Floating point final or initial value.\n");
1259 /* Now set initial_value to be the iteration_var, since that may be a
1260 simpler expression, and is guaranteed to be correct if all of the
1261 above tests succeed.
1263 We can not use the initial_value as calculated, because it will be
1264 one too small for loops of the form "while (i-- > 0)". We can not
1265 emit code before the loop_skip_over insns to fix this problem as this
1266 will then give a number one too large for loops of the form
1269 Note that all loops that reach here are entered at the top, because
1270 this function is not called if the loop starts with a jump. */
1272 /* Fail if loop_iteration_var is not live before loop_start, since we need
1273 to test its value in the preconditioning code. */
1275 if (uid_luid[regno_first_uid[REGNO (loop_iteration_var)]]
1276 > INSN_LUID (loop_start))
1278 if (loop_dump_stream)
1279 fprintf (loop_dump_stream,
1280 "Preconditioning: Iteration var not live before loop start.\n");
1284 *initial_value = loop_iteration_var;
1285 *increment = loop_increment;
1286 *final_value = loop_final_value;
1289 if (loop_dump_stream)
1290 fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1295 /* All pseudo-registers must be mapped to themselves. Two hard registers
1296 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1297 REGNUM, to avoid function-inlining specific conversions of these
1298 registers. All other hard regs can not be mapped because they may be
1303 init_reg_map (map, maxregnum)
1304 struct inline_remap *map;
1309 for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1310 map->reg_map[i] = regno_reg_rtx[i];
1311 /* Just clear the rest of the entries. */
1312 for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1313 map->reg_map[i] = 0;
1315 map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1316 = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1317 map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1318 = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1321 /* Strength-reduction will often emit code for optimized biv/givs which
1322 calculates their value in a temporary register, and then copies the result
1323 to the iv. This procedure reconstructs the pattern computing the iv;
1324 verifying that all operands are of the proper form.
1326 The return value is the amount that the giv is incremented by. */
1329 calculate_giv_inc (pattern, src_insn, regno)
1330 rtx pattern, src_insn;
1335 /* Verify that we have an increment insn here. First check for a plus
1336 as the set source. */
1337 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1339 /* SR sometimes computes the new giv value in a temp, then copies it
1341 src_insn = PREV_INSN (src_insn);
1342 pattern = PATTERN (src_insn);
1343 if (GET_CODE (SET_SRC (pattern)) != PLUS)
1346 /* The last insn emitted is not needed, so delete it to avoid confusing
1347 the second cse pass. This insn sets the giv unnecessarily. */
1348 delete_insn (get_last_insn ());
1351 /* Verify that we have a constant as the second operand of the plus. */
1352 increment = XEXP (SET_SRC (pattern), 1);
1353 if (GET_CODE (increment) != CONST_INT)
1355 /* SR sometimes puts the constant in a register, especially if it is
1356 too big to be an add immed operand. */
1357 increment = SET_SRC (PATTERN (PREV_INSN (src_insn)));
1359 /* SR may have used LO_SUM to compute the constant if it is too large
1360 for a load immed operand. In this case, the constant is in operand
1361 one of the LO_SUM rtx. */
1362 if (GET_CODE (increment) == LO_SUM)
1363 increment = XEXP (increment, 1);
1365 if (GET_CODE (increment) != CONST_INT)
1368 /* The insn loading the constant into a register is not longer needed,
1370 delete_insn (get_last_insn ());
1373 /* Check that the source register is the same as the dest register. */
1374 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1375 || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1382 /* Copy each instruction in the loop, substituting from map as appropriate.
1383 This is very similar to a loop in expand_inline_function. */
1386 copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration,
1387 unroll_type, start_label, loop_end, insert_before,
1389 rtx copy_start, copy_end;
1390 struct inline_remap *map;
1392 enum unroll_types unroll_type;
1393 rtx start_label, loop_end, insert_before, copy_notes_from;
1397 int dest_reg_was_split, i;
1399 rtx final_label = 0;
1400 rtx giv_inc, giv_dest_reg, giv_src_reg;
1402 /* If this isn't the last iteration, then map any references to the
1403 start_label to final_label. Final label will then be emitted immediately
1404 after the end of this loop body if it was ever used.
1406 If this is the last iteration, then map references to the start_label
1408 if (! last_iteration)
1410 final_label = gen_label_rtx ();
1411 map->label_map[CODE_LABEL_NUMBER (start_label)] = final_label;
1414 map->label_map[CODE_LABEL_NUMBER (start_label)] = start_label;
1421 insn = NEXT_INSN (insn);
1423 map->orig_asm_operands_vector = 0;
1425 switch (GET_CODE (insn))
1428 pattern = PATTERN (insn);
1432 /* Check to see if this is a giv that has been combined with
1433 some split address givs. (Combined in the sense that
1434 `combine_givs' in loop.c has put two givs in the same register.)
1435 In this case, we must search all givs based on the same biv to
1436 find the address givs. Then split the address givs.
1437 Do this before splitting the giv, since that may map the
1438 SET_DEST to a new register. */
1440 if (GET_CODE (pattern) == SET
1441 && GET_CODE (SET_DEST (pattern)) == REG
1442 && addr_combined_regs[REGNO (SET_DEST (pattern))])
1444 struct iv_class *bl;
1445 struct induction *v, *tv;
1446 int regno = REGNO (SET_DEST (pattern));
1448 v = addr_combined_regs[REGNO (SET_DEST (pattern))];
1449 bl = reg_biv_class[REGNO (v->src_reg)];
1451 /* Although the giv_inc amount is not needed here, we must call
1452 calculate_giv_inc here since it might try to delete the
1453 last insn emitted. If we wait until later to call it,
1454 we might accidentally delete insns generated immediately
1455 below by emit_unrolled_add. */
1457 giv_inc = calculate_giv_inc (pattern, insn, regno);
1459 /* Now find all address giv's that were combined with this
1461 for (tv = bl->giv; tv; tv = tv->next_iv)
1462 if (tv->giv_type == DEST_ADDR && tv->same == v)
1464 /* Increment the giv by the amount that was calculated in
1465 find_splittable_givs, and saved in add_val. */
1466 tv->dest_reg = plus_constant (tv->dest_reg,
1467 INTVAL (tv->add_val));
1468 *tv->location = tv->dest_reg;
1470 if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1472 /* Must emit an insn to increment the split address
1473 giv. Add in the const_adjust field in case there
1474 was a constant eliminated from the address. */
1475 rtx value, dest_reg;
1477 /* tv->dest_reg will be either a bare register,
1478 or else a register plus a constant. */
1479 if (GET_CODE (tv->dest_reg) == REG)
1480 dest_reg = tv->dest_reg;
1482 dest_reg = XEXP (tv->dest_reg, 0);
1484 /* tv->dest_reg may actually be a (PLUS (REG) (CONST))
1485 here, so we must call plus_constant to add
1486 the const_adjust amount before calling
1487 emit_unrolled_add below. */
1488 value = plus_constant (tv->dest_reg, tv->const_adjust);
1490 /* The constant could be too large for an add
1491 immediate, so can't directly emit an insn here. */
1492 emit_unrolled_add (dest_reg, XEXP (value, 0),
1495 /* Reset the giv to be just the register again, in case
1496 it is used after the set we have just emitted.
1497 We must subtract the const_adjust factor added in
1499 tv->dest_reg = plus_constant (dest_reg,
1500 - tv->const_adjust);
1501 *tv->location = tv->dest_reg;
1506 /* If this is a setting of a splittable variable, then determine
1507 how to split the variable, create a new set based on this split,
1508 and set up the reg_map so that later uses of the variable will
1509 use the new split variable. */
1511 dest_reg_was_split = 0;
1513 if (GET_CODE (pattern) == SET
1514 && GET_CODE (SET_DEST (pattern)) == REG
1515 && splittable_regs[REGNO (SET_DEST (pattern))])
1517 int regno = REGNO (SET_DEST (pattern));
1519 dest_reg_was_split = 1;
1521 /* Compute the increment value for the giv, if it wasn't
1522 already computed above. */
1525 giv_inc = calculate_giv_inc (pattern, insn, regno);
1526 giv_dest_reg = SET_DEST (pattern);
1527 giv_src_reg = SET_DEST (pattern);
1529 if (unroll_type == UNROLL_COMPLETELY)
1531 /* Completely unrolling the loop. Set the induction
1532 variable to a known constant value. */
1534 /* The value in splittable_regs may be an invariant
1535 value, so we must use plus_constant here. */
1536 splittable_regs[regno]
1537 = plus_constant (splittable_regs[regno], INTVAL (giv_inc));
1539 if (GET_CODE (splittable_regs[regno]) == PLUS)
1541 giv_src_reg = XEXP (splittable_regs[regno], 0);
1542 giv_inc = XEXP (splittable_regs[regno], 1);
1546 /* The splittable_regs value must be a REG or a
1547 CONST_INT, so put the entire value in the giv_src_reg
1549 giv_src_reg = splittable_regs[regno];
1550 giv_inc = const0_rtx;
1555 /* Partially unrolling loop. Create a new pseudo
1556 register for the iteration variable, and set it to
1557 be a constant plus the original register. Except
1558 on the last iteration, when the result has to
1559 go back into the original iteration var register. */
1561 /* Handle bivs which must be mapped to a new register
1562 when split. This happens for bivs which need their
1563 final value set before loop entry. The new register
1564 for the biv was stored in the biv's first struct
1565 induction entry by find_splittable_regs. */
1567 if (regno < max_reg_before_loop
1568 && reg_iv_type[regno] == BASIC_INDUCT)
1570 giv_src_reg = reg_biv_class[regno]->biv->src_reg;
1571 giv_dest_reg = giv_src_reg;
1575 /* If non-reduced/final-value givs were split, then
1576 this would have to remap those givs also. See
1577 find_splittable_regs. */
1580 splittable_regs[regno]
1581 = gen_rtx (CONST_INT, VOIDmode,
1583 + INTVAL (splittable_regs[regno]));
1584 giv_inc = splittable_regs[regno];
1586 /* Now split the induction variable by changing the dest
1587 of this insn to a new register, and setting its
1588 reg_map entry to point to this new register.
1590 If this is the last iteration, and this is the last insn
1591 that will update the iv, then reuse the original dest,
1592 to ensure that the iv will have the proper value when
1593 the loop exits or repeats.
1595 Using splittable_regs_updates here like this is safe,
1596 because it can only be greater than one if all
1597 instructions modifying the iv are always executed in
1600 if (! last_iteration
1601 || (splittable_regs_updates[regno]-- != 1))
1603 tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1605 map->reg_map[regno] = tem;
1608 map->reg_map[regno] = giv_src_reg;
1611 /* The constant being added could be too large for an add
1612 immediate, so can't directly emit an insn here. */
1613 emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1614 copy = get_last_insn ();
1615 pattern = PATTERN (copy);
1619 pattern = copy_rtx_and_substitute (pattern, map);
1620 copy = emit_insn (pattern);
1622 /* REG_NOTES will be copied later. */
1625 /* If this insn is setting CC0, it may need to look at
1626 the insn that uses CC0 to see what type of insn it is.
1627 In that case, the call to recog via validate_change will
1628 fail. So don't substitute constants here. Instead,
1629 do it when we emit the following insn.
1631 For example, see the pyr.md file. That machine has signed and
1632 unsigned compares. The compare patterns must check the
1633 following branch insn to see which what kind of compare to
1636 If the previous insn set CC0, substitute constants on it as
1638 if (sets_cc0_p (copy) != 0)
1643 try_constants (cc0_insn, map);
1645 try_constants (copy, map);
1648 try_constants (copy, map);
1651 /* Make split induction variable constants `permanent' since we
1652 know there are no backward branches across iteration variable
1653 settings which would invalidate this. */
1654 if (dest_reg_was_split)
1656 int regno = REGNO (SET_DEST (pattern));
1658 if (map->const_age_map[regno] == map->const_age)
1659 map->const_age_map[regno] = -1;
1664 if (JUMP_LABEL (insn) == start_label && insn == copy_end
1665 && ! last_iteration)
1667 /* This is a branch to the beginning of the loop; this is the
1668 last insn being copied; and this is not the last iteration.
1669 In this case, we want to change the original fall through
1670 case to be a branch past the end of the loop, and the
1671 original jump label case to fall_through. */
1675 /* Never map the label in this case. */
1676 pattern = copy_rtx (PATTERN (insn));
1678 /* Assume a conditional branch, since the code above
1679 does not let unconditional branches be copied. */
1680 if (! condjump_p (insn))
1683 = (XEXP (SET_SRC (PATTERN (insn)), 2) == pc_rtx) + 1;
1685 /* Set the fall through case to the exit label. Must
1686 create a new label_ref since they can't be shared. */
1687 XEXP (SET_SRC (pattern), fall_through)
1688 = gen_rtx (LABEL_REF, VOIDmode, exit_label);
1690 /* Set the original branch case to fall through. */
1691 XEXP (SET_SRC (pattern), 3 - fall_through)
1695 pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1697 copy = emit_jump_insn (pattern);
1701 try_constants (cc0_insn, map);
1704 try_constants (copy, map);
1706 /* Set the jump label of COPY correctly to avoid problems with
1707 later passes of unroll_loop, if INSN had jump label set. */
1708 if (JUMP_LABEL (insn))
1710 /* Can't use the label_map for every insn, since this may be
1711 the backward branch, and hence the label was not mapped. */
1712 if (GET_CODE (pattern) == SET)
1714 tem = SET_SRC (pattern);
1715 if (GET_CODE (tem) == LABEL_REF)
1716 JUMP_LABEL (copy) = XEXP (tem, 0);
1717 else if (GET_CODE (tem) == IF_THEN_ELSE)
1719 if (XEXP (tem, 1) != pc_rtx)
1720 JUMP_LABEL (copy) = XEXP (XEXP (tem, 1), 0);
1722 JUMP_LABEL (copy) = XEXP (XEXP (tem, 2), 0);
1729 /* An unrecognizable jump insn, probably the entry jump
1730 for a switch statement. This label must have been mapped,
1731 so just use the label_map to get the new jump label. */
1732 JUMP_LABEL (copy) = map->label_map[CODE_LABEL_NUMBER
1733 (JUMP_LABEL (insn))];
1736 /* If this is a non-local jump, then must increase the label
1737 use count so that the label will not be deleted when the
1738 original jump is deleted. */
1739 LABEL_NUSES (JUMP_LABEL (copy))++;
1741 else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
1742 || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
1744 rtx pat = PATTERN (copy);
1745 int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
1746 int len = XVECLEN (pat, diff_vec_p);
1749 for (i = 0; i < len; i++)
1750 LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
1753 /* If this used to be a conditional jump insn but whose branch
1754 direction is now known, we must do something special. */
1755 if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value)
1758 /* The previous insn set cc0 for us. So delete it. */
1759 delete_insn (PREV_INSN (copy));
1762 /* If this is now a no-op, delete it. */
1763 if (map->last_pc_value == pc_rtx)
1769 /* Otherwise, this is unconditional jump so we must put a
1770 BARRIER after it. We could do some dead code elimination
1771 here, but jump.c will do it just as well. */
1777 pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1778 copy = emit_call_insn (pattern);
1782 try_constants (cc0_insn, map);
1785 try_constants (copy, map);
1787 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1788 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1789 map->const_equiv_map[i] = 0;
1793 /* If this is the loop start label, then we don't need to emit a
1794 copy of this label since no one will use it. */
1796 if (insn != start_label)
1798 copy = emit_label (map->label_map[CODE_LABEL_NUMBER (insn)]);
1804 copy = emit_barrier ();
1808 if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
1809 copy = emit_note (NOTE_SOURCE_FILE (insn),
1810 NOTE_LINE_NUMBER (insn));
1820 map->insn_map[INSN_UID (insn)] = copy;
1822 while (insn != copy_end);
1824 /* Now copy the REG_NOTES. */
1828 insn = NEXT_INSN (insn);
1829 if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
1830 || GET_CODE (insn) == CALL_INSN)
1831 && map->insn_map[INSN_UID (insn)])
1832 REG_NOTES (map->insn_map[INSN_UID (insn)])
1833 = copy_rtx_and_substitute (REG_NOTES (insn), map);
1835 while (insn != copy_end);
1837 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1838 each of these notes here, since there may be some important ones, such as
1839 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1840 iteration, because the original notes won't be deleted.
1842 We can't use insert_before here, because when from preconditioning,
1843 insert_before points before the loop. We can't use copy_end, because
1844 there may be insns already inserted after it (which we don't want to
1845 copy) when not from preconditioning code. */
1847 if (! last_iteration)
1849 for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
1851 if (GET_CODE (insn) == NOTE
1852 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
1853 emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
1857 if (final_label && LABEL_NUSES (final_label) > 0)
1858 emit_label (final_label);
1860 tem = gen_sequence ();
1862 emit_insn_before (tem, insert_before);
1865 /* Emit an insn, using the expand_binop to ensure that a valid insn is
1866 emitted. This will correctly handle the case where the increment value
1867 won't fit in the immediate field of a PLUS insns. */
1870 emit_unrolled_add (dest_reg, src_reg, increment)
1871 rtx dest_reg, src_reg, increment;
1875 result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment,
1876 dest_reg, 0, OPTAB_LIB_WIDEN);
1878 if (dest_reg != result)
1879 emit_move_insn (dest_reg, result);
1882 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
1883 is a backward branch in that range that branches to somewhere between
1884 LOOP_START and INSN. Returns 0 otherwise. */
1886 /* ??? This is quadratic algorithm. Could be rewriten to be linear.
1887 In practice, this is not a problem, because this function is seldom called,
1888 and uses a negligible amount of CPU time on average. */
1891 back_branch_in_range_p (insn, loop_start, loop_end)
1893 rtx loop_start, loop_end;
1895 rtx p, q, target_insn;
1897 /* Stop before we get to the backward branch at the end of the loop. */
1898 loop_end = prev_nonnote_insn (loop_end);
1899 if (GET_CODE (loop_end) == BARRIER)
1900 loop_end = PREV_INSN (loop_end);
1902 /* Check in case insn has been deleted, search forward for first non
1903 deleted insn following it. */
1904 while (INSN_DELETED_P (insn))
1905 insn = NEXT_INSN (insn);
1907 /* Check for the case where insn is the last insn in the loop. */
1908 if (insn == loop_end)
1911 for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
1913 if (GET_CODE (p) == JUMP_INSN)
1915 target_insn = JUMP_LABEL (p);
1917 /* Search from loop_start to insn, to see if one of them is
1918 the target_insn. We can't use INSN_LUID comparisons here,
1919 since insn may not have an LUID entry. */
1920 for (q = loop_start; q != insn; q = NEXT_INSN (q))
1921 if (q == target_insn)
1929 /* Try to generate the simplest rtx for the expression
1930 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
1934 fold_rtx_mult_add (mult1, mult2, add1, mode)
1935 rtx mult1, mult2, add1;
1936 enum machine_mode mode;
1941 /* The modes must all be the same. This should always be true. For now,
1942 check to make sure. */
1943 if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
1944 || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
1945 || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
1948 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
1949 will be a constant. */
1950 if (GET_CODE (mult1) == CONST_INT)
1957 mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
1959 mult_res = gen_rtx (MULT, mode, mult1, mult2);
1961 /* Again, put the constant second. */
1962 if (GET_CODE (add1) == CONST_INT)
1969 result = simplify_binary_operation (PLUS, mode, add1, mult_res);
1971 result = gen_rtx (PLUS, mode, add1, mult_res);
1976 /* Searches the list of induction struct's for the biv BL, to try to calculate
1977 the total increment value for one iteration of the loop as a constant.
1979 Returns the increment value as an rtx, simplified as much as possible,
1980 if it can be calculated. Otherwise, returns 0. */
1983 biv_total_increment (bl, loop_start, loop_end)
1984 struct iv_class *bl;
1985 rtx loop_start, loop_end;
1987 struct induction *v;
1990 /* For increment, must check every instruction that sets it. Each
1991 instruction must be executed only once each time through the loop.
1992 To verify this, we check that the the insn is always executed, and that
1993 there are no backward branches after the insn that branch to before it.
1994 Also, the insn must have a mult_val of one (to make sure it really is
1997 result = const0_rtx;
1998 for (v = bl->biv; v; v = v->next_iv)
2000 if (v->always_computable && v->mult_val == const1_rtx
2001 && ! back_branch_in_range_p (v->insn, loop_start, loop_end))
2002 result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
2010 /* Determine the initial value of the iteration variable, and the amount
2011 that it is incremented each loop. Use the tables constructed by
2012 the strength reduction pass to calculate these values.
2014 Initial_value and/or increment are set to zero if their values could not
2018 iteration_info (iteration_var, initial_value, increment, loop_start, loop_end)
2019 rtx iteration_var, *initial_value, *increment;
2020 rtx loop_start, loop_end;
2022 struct iv_class *bl;
2023 struct induction *v, *b;
2025 /* Clear the result values, in case no answer can be found. */
2029 /* The iteration variable can be either a giv or a biv. Check to see
2030 which it is, and compute the variable's initial value, and increment
2031 value if possible. */
2033 /* If this is a new register, can't handle it since we don't have any
2034 reg_iv_type entry for it. */
2035 if (REGNO (iteration_var) >= max_reg_before_loop)
2037 if (loop_dump_stream)
2038 fprintf (loop_dump_stream,
2039 "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2042 /* Reject iteration variables larger than the host long size, since they
2043 could result in a number of iterations greater than the range of our
2044 `unsigned long' variable loop_n_iterations. */
2045 else if (GET_MODE_BITSIZE (GET_MODE (iteration_var)) > HOST_BITS_PER_LONG)
2047 if (loop_dump_stream)
2048 fprintf (loop_dump_stream,
2049 "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2052 else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
2054 if (loop_dump_stream)
2055 fprintf (loop_dump_stream,
2056 "Loop unrolling: Iteration var not an interger.\n");
2059 else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT)
2061 /* Grab initial value, only useful if it is a constant. */
2062 bl = reg_biv_class[REGNO (iteration_var)];
2063 *initial_value = bl->initial_value;
2065 *increment = biv_total_increment (bl, loop_start, loop_end);
2067 else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT)
2070 /* ??? The code below does not work because the incorrect number of
2071 iterations is calculated when the biv is incremented after the giv
2072 is set (which is the usual case). This can probably be accounted
2073 for by biasing the initial_value by subtracting the amount of the
2074 increment that occurs between the giv set and the giv test. However,
2075 a giv as an iterator is very rare, so it does not seem worthwhile
2077 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2078 if (loop_dump_stream)
2079 fprintf (loop_dump_stream,
2080 "Loop unrolling: Giv iterators are not handled.\n");
2083 /* Initial value is mult_val times the biv's initial value plus
2084 add_val. Only useful if it is a constant. */
2085 v = reg_iv_info[REGNO (iteration_var)];
2086 bl = reg_biv_class[REGNO (v->src_reg)];
2087 *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value,
2088 v->add_val, v->mode);
2090 /* Increment value is mult_val times the increment value of the biv. */
2092 *increment = biv_total_increment (bl, loop_start, loop_end);
2094 *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx,
2100 if (loop_dump_stream)
2101 fprintf (loop_dump_stream,
2102 "Loop unrolling: Not basic or general induction var.\n");
2107 /* Calculate the approximate final value of the iteration variable
2108 which has an loop exit test with code COMPARISON_CODE and comparison value
2109 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2110 was signed or unsigned, and the direction of the comparison. This info is
2111 needed to calculate the number of loop iterations. */
2114 approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir)
2115 enum rtx_code comparison_code;
2116 rtx comparison_value;
2120 /* Calculate the final value of the induction variable.
2121 The exact final value depends on the branch operator, and increment sign.
2122 This is only an approximate value. It will be wrong if the iteration
2123 variable is not incremented by one each time through the loop, and
2124 approx final value - start value % increment != 0. */
2127 switch (comparison_code)
2133 return plus_constant (comparison_value, 1);
2138 return plus_constant (comparison_value, -1);
2140 /* Can not calculate a final value for this case. */
2147 return comparison_value;
2153 return comparison_value;
2156 return comparison_value;
2162 /* For each biv and giv, determine whether it can be safely split into
2163 a different variable for each unrolled copy of the loop body. If it
2164 is safe to split, then indicate that by saving some useful info
2165 in the splittable_regs array.
2167 If the loop is being completely unrolled, then splittable_regs will hold
2168 the current value of the induction variable while the loop is unrolled.
2169 It must be set to the initial value of the induction variable here.
2170 Otherwise, splittable_regs will hold the difference between the current
2171 value of the induction variable and the value the induction variable had
2172 at the top of the loop. It must be set to the value 0 here. */
2174 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2175 constant values are unnecessary, since we can easily calculate increment
2176 values in this case even if nothing is constant. The increment value
2177 should not involve a multiply however. */
2179 /* ?? Even if the biv/giv increment values aren't constant, it may still
2180 be beneficial to split the variable if the loop is only unrolled a few
2181 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2184 find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before,
2186 enum unroll_types unroll_type;
2187 rtx loop_start, loop_end;
2188 rtx end_insert_before;
2191 struct iv_class *bl;
2193 rtx biv_final_value;
2197 for (bl = loop_iv_list; bl; bl = bl->next)
2199 /* Biv_total_increment must return a constant value,
2200 otherwise we can not calculate the split values. */
2202 increment = biv_total_increment (bl, loop_start, loop_end);
2203 if (! increment || GET_CODE (increment) != CONST_INT)
2206 /* The loop must be unrolled completely, or else have a known number
2207 of iterations and only one exit, or else the biv must be dead
2208 outside the loop, or else the final value must be known. Otherwise,
2209 it is unsafe to split the biv since it may not have the proper
2210 value on loop exit. */
2212 /* loop_number_exit_labels is non-zero if the loop has an exit other than
2213 a fall through at the end. */
2216 biv_final_value = 0;
2217 if (unroll_type != UNROLL_COMPLETELY
2218 && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2219 || unroll_type == UNROLL_NAIVE)
2220 && (uid_luid[regno_last_uid[bl->regno]] >= INSN_LUID (loop_end)
2222 || INSN_UID (bl->init_insn) >= max_uid_for_loop
2223 || (uid_luid[regno_first_uid[bl->regno]]
2224 < INSN_LUID (bl->init_insn))
2225 || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2226 && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end)))
2229 /* If final value is non-zero, then must emit an instruction which sets
2230 the value of the biv to the proper value. This is done after
2231 handling all of the givs, since some of them may need to use the
2232 biv's value in their initialization code. */
2234 /* This biv is splittable. If completely unrolling the loop, save
2235 the biv's initial value. Otherwise, save the constant zero. */
2237 if (biv_splittable == 1)
2239 if (unroll_type == UNROLL_COMPLETELY)
2241 /* If the initial value of the biv is itself (i.e. it is too
2242 complicated for strength_reduce to compute), or is a hard
2243 register, then we must create a new psuedo reg to hold the
2244 initial value of the biv. */
2246 if (GET_CODE (bl->initial_value) == REG
2247 && (REGNO (bl->initial_value) == bl->regno
2248 || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER))
2250 rtx tem = gen_reg_rtx (bl->biv->mode);
2252 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2255 if (loop_dump_stream)
2256 fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n",
2257 bl->regno, REGNO (tem));
2259 splittable_regs[bl->regno] = tem;
2262 splittable_regs[bl->regno] = bl->initial_value;
2265 splittable_regs[bl->regno] = const0_rtx;
2267 /* Save the number of instructions that modify the biv, so that
2268 we can treat the last one specially. */
2270 splittable_regs_updates[bl->regno] = bl->biv_count;
2274 if (loop_dump_stream)
2275 fprintf (loop_dump_stream,
2276 "Biv %d safe to split.\n", bl->regno);
2279 /* Check every giv that depends on this biv to see whether it is
2280 splittable also. Even if the biv isn't splittable, givs which
2281 depend on it may be splittable if the biv is live outside the
2282 loop, and the givs aren't. */
2284 result = find_splittable_givs (bl, unroll_type, loop_start, loop_end,
2285 increment, unroll_number, result);
2287 /* If final value is non-zero, then must emit an instruction which sets
2288 the value of the biv to the proper value. This is done after
2289 handling all of the givs, since some of them may need to use the
2290 biv's value in their initialization code. */
2291 if (biv_final_value)
2293 /* If the loop has multiple exits, emit the insns before the
2294 loop to ensure that it will always be executed no matter
2295 how the loop exits. Otherwise emit the insn after the loop,
2296 since this is slightly more efficient. */
2297 if (! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2298 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2303 /* Create a new register to hold the value of the biv, and then
2304 set the biv to its final value before the loop start. The biv
2305 is set to its final value before loop start to ensure that
2306 this insn will always be executed, no matter how the loop
2308 rtx tem = gen_reg_rtx (bl->biv->mode);
2309 emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2311 emit_insn_before (gen_move_insn (bl->biv->src_reg,
2315 if (loop_dump_stream)
2316 fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2317 REGNO (bl->biv->src_reg), REGNO (tem));
2319 /* Set up the mapping from the original biv register to the new
2321 bl->biv->src_reg = tem;
2328 /* For every giv based on the biv BL, check to determine whether it is
2329 splittable. This is a subroutine to find_splittable_regs (). */
2332 find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment,
2333 unroll_number, result)
2334 struct iv_class *bl;
2335 enum unroll_types unroll_type;
2336 rtx loop_start, loop_end;
2338 int unroll_number, result;
2340 struct induction *v;
2344 for (v = bl->giv; v; v = v->next_iv)
2348 /* Only split the giv if it has already been reduced, or if the loop is
2349 being completely unrolled. */
2350 if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2353 /* The giv can be split if the insn that sets the giv is executed once
2354 and only once on every iteration of the loop. */
2355 /* An address giv can always be split. v->insn is just a use not a set,
2356 and hence it does not matter whether it is always executed. All that
2357 matters is that all the biv increments are always executed, and we
2358 won't reach here if they aren't. */
2359 if (v->giv_type != DEST_ADDR
2360 && (! v->always_computable
2361 || back_branch_in_range_p (v->insn, loop_start, loop_end)))
2364 /* The giv increment value must be a constant. */
2365 giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2367 if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2370 /* The loop must be unrolled completely, or else have a known number of
2371 iterations and only one exit, or else the giv must be dead outside
2372 the loop, or else the final value of the giv must be known.
2373 Otherwise, it is not safe to split the giv since it may not have the
2374 proper value on loop exit. */
2376 /* The used outside loop test will fail for DEST_ADDR givs. They are
2377 never used outside the loop anyways, so it is always safe to split a
2381 if (unroll_type != UNROLL_COMPLETELY
2382 && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2383 || unroll_type == UNROLL_NAIVE)
2384 && v->giv_type != DEST_ADDR
2385 && ((regno_first_uid[REGNO (v->dest_reg)] != INSN_UID (v->insn)
2386 /* Check for the case where the pseudo is set by a shift/add
2387 sequence, in which case the first insn setting the pseudo
2388 is the first insn of the shift/add sequence. */
2389 && (! (tem = find_reg_note (v->insn, REG_RETVAL, 0))
2390 || (regno_first_uid[REGNO (v->dest_reg)]
2391 != INSN_UID (XEXP (tem, 0)))))
2392 /* Line above always fails if INSN was moved by loop opt. */
2393 || (uid_luid[regno_last_uid[REGNO (v->dest_reg)]]
2394 >= INSN_LUID (loop_end)))
2395 && ! (final_value = v->final_value))
2399 /* Currently, non-reduced/final-value givs are never split. */
2400 /* Should emit insns after the loop if possible, as the biv final value
2403 /* If the final value is non-zero, and the giv has not been reduced,
2404 then must emit an instruction to set the final value. */
2405 if (final_value && !v->new_reg)
2407 /* Create a new register to hold the value of the giv, and then set
2408 the giv to its final value before the loop start. The giv is set
2409 to its final value before loop start to ensure that this insn
2410 will always be executed, no matter how we exit. */
2411 tem = gen_reg_rtx (v->mode);
2412 emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start);
2413 emit_insn_before (gen_move_insn (v->dest_reg, final_value),
2416 if (loop_dump_stream)
2417 fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2418 REGNO (v->dest_reg), REGNO (tem));
2424 /* This giv is splittable. If completely unrolling the loop, save the
2425 giv's initial value. Otherwise, save the constant zero for it. */
2427 if (unroll_type == UNROLL_COMPLETELY)
2428 /* It is not safe to use bl->initial_value here, because it may not
2429 be invariant. It is safe to use the initial value stored in
2430 the splittable_regs array. */
2431 value = fold_rtx_mult_add (v->mult_val, splittable_regs[bl->regno],
2432 v->add_val, v->mode);
2438 /* If the giv is an address destination, it could be something other
2439 than a simple register, these have to be treated differently. */
2440 if (v->giv_type == DEST_REG)
2441 splittable_regs[REGNO (v->new_reg)] = value;
2443 /* If an addr giv was combined with another addr giv, then we
2444 can only split this giv if the addr giv it was combined with
2445 was reduced. This is because the value of v->new_reg is
2446 meaningless in this case. (There is no problem if it was
2447 combined with a dest_reg giv which wasn't reduced, v->new_reg
2448 is still meaningful in this case.) */
2450 else if (v->same && v->same->giv_type == DEST_ADDR
2451 && ! v->same->new_reg)
2453 if (loop_dump_stream)
2454 fprintf (loop_dump_stream,
2455 "DEST_ADDR giv not split, because combined with unreduced DEST_ADDR giv.\n");
2459 /* Splitting address givs is useful since it will often allow us
2460 to eliminate some increment insns for the base giv as
2463 /* If the addr giv is combined with a dest_reg giv, then all
2464 references to that dest reg will be remapped, which is NOT
2465 what we want for split addr regs. We always create a new
2466 register for the split addr giv, just to be safe. */
2468 /* ??? If there are multiple address givs which have been
2469 combined with the same dest_reg giv, then we may only need
2470 one new register for them. Pulling out constants below will
2471 catch some of the common cases of this. Currently, I leave
2472 the work of simplifying multiple address givs to the
2473 following cse pass. */
2475 v->const_adjust = 0;
2476 if (unroll_type != UNROLL_COMPLETELY)
2478 /* If not completely unrolling the loop, then create a new
2479 register to hold the split value of the DEST_ADDR giv.
2480 Emit insn to initialize its value before loop start. */
2481 tem = gen_reg_rtx (v->mode);
2483 /* If the address giv has a constant in its new_reg value,
2484 then this constant can be pulled out and put in value,
2485 instead of being part of the initialization code. */
2487 if (GET_CODE (v->new_reg) == PLUS
2488 && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT)
2491 = plus_constant (tem, INTVAL (XEXP (v->new_reg,1)));
2493 /* Only succeed if this will give valid addresses.
2494 Try to validate both the first and the last
2495 address resulting from loop unrolling, if
2496 one fails, then can't do const elim here. */
2497 if (memory_address_p (v->mode, v->dest_reg)
2498 && memory_address_p (v->mode,
2499 plus_constant (v->dest_reg,
2501 * (unroll_number - 1))))
2503 /* Save the negative of the eliminated const, so
2504 that we can calculate the dest_reg's increment
2506 v->const_adjust = - INTVAL (XEXP (v->new_reg, 1));
2508 v->new_reg = XEXP (v->new_reg, 0);
2509 if (loop_dump_stream)
2510 fprintf (loop_dump_stream,
2511 "Eliminating constant from giv %d\n",
2520 /* If the address hasn't been checked for validity yet, do so
2521 now, and fail completely if either the first or the last
2522 unrolled copy of the address is not a valid address. */
2523 if (v->dest_reg == tem
2524 && (! memory_address_p (v->mode, v->dest_reg)
2525 || ! memory_address_p (v->mode,
2526 plus_constant (v->dest_reg,
2528 * (unroll_number -1)))))
2530 if (loop_dump_stream)
2531 fprintf (loop_dump_stream,
2532 "Illegal address for giv at insn %d\n",
2533 INSN_UID (v->insn));
2537 /* To initialize the new register, just move the value of
2538 new_reg into it. This is not guaranteed to give a valid
2539 instruction on machines with complex addressing modes.
2540 If we can't recognize it, then delete it and emit insns
2541 to calculate the value from scratch. */
2542 emit_insn_before (gen_rtx (SET, VOIDmode, tem,
2543 copy_rtx (v->new_reg)),
2545 if (! recog_memoized (PREV_INSN (loop_start)))
2547 delete_insn (PREV_INSN (loop_start));
2548 emit_iv_add_mult (bl->initial_value, v->mult_val,
2549 v->add_val, tem, loop_start);
2550 if (loop_dump_stream)
2551 fprintf (loop_dump_stream,
2552 "Illegal init insn, rewritten.\n");
2557 v->dest_reg = value;
2559 /* Check the resulting address for validity, and fail
2560 if the resulting address would be illegal. */
2561 if (! memory_address_p (v->mode, v->dest_reg)
2562 || ! memory_address_p (v->mode,
2563 plus_constant (v->dest_reg,
2565 (unroll_number -1))))
2567 if (loop_dump_stream)
2568 fprintf (loop_dump_stream,
2569 "Illegal address for giv at insn %d\n",
2570 INSN_UID (v->insn));
2575 /* Store the value of dest_reg into the insn. This sharing
2576 will not be a problem as this insn will always be copied
2579 *v->location = v->dest_reg;
2581 /* If this address giv is combined with a dest reg giv, then
2582 save the base giv's induction pointer so that we will be
2583 able to handle this address giv properly. The base giv
2584 itself does not have to be splittable. */
2586 if (v->same && v->same->giv_type == DEST_REG)
2587 addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
2589 if (GET_CODE (v->new_reg) == REG)
2591 /* This giv maybe hasn't been combined with any others.
2592 Make sure that it's giv is marked as splittable here. */
2594 splittable_regs[REGNO (v->new_reg)] = value;
2596 /* Make it appear to depend upon itself, so that the
2597 giv will be properly split in the main loop above. */
2601 addr_combined_regs[REGNO (v->new_reg)] = v;
2605 /* Overwrite the old add_val, which is no longer needed, and
2606 substitute the amount that the giv is incremented on each
2607 iteration. We need to save this somewhere, so we know how
2608 much to increment split DEST_ADDR giv's in copy_loop_body. */
2610 v->add_val = giv_inc;
2612 if (loop_dump_stream)
2613 fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
2619 /* Currently, unreduced giv's can't be split. This is not too much
2620 of a problem since unreduced giv's are not live across loop
2621 iterations anyways. When unrolling a loop completely though,
2622 it makes sense to reduce&split givs when possible, as this will
2623 result in simpler instructions, and will not require that a reg
2624 be live across loop iterations. */
2626 splittable_regs[REGNO (v->dest_reg)] = value;
2627 fprintf (stderr, "Giv %d at insn %d not reduced\n",
2628 REGNO (v->dest_reg), INSN_UID (v->insn));
2634 /* Givs are only updated once by definition. Mark it so if this is
2635 a splittable register. Don't need to do anything for address givs
2636 where this may not be a register. */
2638 if (GET_CODE (v->new_reg) == REG)
2639 splittable_regs_updates[REGNO (v->new_reg)] = 1;
2643 if (loop_dump_stream)
2647 if (GET_CODE (v->dest_reg) == CONST_INT)
2649 else if (GET_CODE (v->dest_reg) != REG)
2650 regnum = REGNO (XEXP (v->dest_reg, 0));
2652 regnum = REGNO (v->dest_reg);
2653 fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
2654 regnum, INSN_UID (v->insn));
2661 /* Try to prove that the register is dead after the loop exits. Trace every
2662 loop exit looking for an insn that will always be executed, which sets
2663 the register to some value, and appears before the first use of the register
2664 is found. If successful, then return 1, otherwise return 0. */
2666 /* ?? Could be made more intelligent in the handling of jumps, so that
2667 it can search past if statements and other similar structures. */
2670 reg_dead_after_loop (reg, loop_start, loop_end)
2671 rtx reg, loop_start, loop_end;
2677 /* HACK: Must also search the loop fall through exit, create a label_ref
2678 here which points to the loop_end, and append the loop_number_exit_labels
2680 label = gen_rtx (LABEL_REF, VOIDmode, loop_end);
2681 LABEL_NEXTREF (label)
2682 = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
2684 for ( ; label; label = LABEL_NEXTREF (label))
2686 /* Succeed if find an insn which sets the biv or if reach end of
2687 function. Fail if find an insn that uses the biv, or if come to
2688 a conditional jump. */
2690 insn = NEXT_INSN (XEXP (label, 0));
2693 code = GET_CODE (insn);
2694 if (GET_RTX_CLASS (code) == 'i')
2698 if (reg_referenced_p (reg, PATTERN (insn)))
2701 set = single_set (insn);
2702 if (set && rtx_equal_p (SET_DEST (set), reg))
2706 if (code == JUMP_INSN)
2708 if (GET_CODE (PATTERN (insn)) == RETURN)
2710 else if (! simplejump_p (insn)
2711 /* Prevent infinite loop following infinite loops. */
2712 || jump_count++ > 20)
2715 insn = JUMP_LABEL (insn);
2718 insn = NEXT_INSN (insn);
2722 /* Success, the register is dead on all loop exits. */
2726 /* Try to calculate the final value of the biv, the value it will have at
2727 the end of the loop. If we can do it, return that value. */
2730 final_biv_value (bl, loop_start, loop_end)
2731 struct iv_class *bl;
2732 rtx loop_start, loop_end;
2736 /* The final value for reversed bivs must be calculated differently than
2737 for ordinary bivs. In this case, there is already an insn after the
2738 loop which sets this biv's final value (if necessary), and there are
2739 no other loop exits, so we can return any value. */
2742 if (loop_dump_stream)
2743 fprintf (loop_dump_stream,
2744 "Final biv value for %d, reversed biv.\n", bl->regno);
2749 /* Try to calculate the final value as initial value + (number of iterations
2750 * increment). For this to work, increment must be invariant, the only
2751 exit from the loop must be the fall through at the bottom (otherwise
2752 it may not have its final value when the loop exits), and the initial
2753 value of the biv must be invariant. */
2755 if (loop_n_iterations != 0
2756 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2757 && invariant_p (bl->initial_value))
2759 increment = biv_total_increment (bl, loop_start, loop_end);
2761 if (increment && invariant_p (increment))
2763 /* Can calculate the loop exit value, emit insns after loop
2764 end to calculate this value into a temporary register in
2765 case it is needed later. */
2767 tem = gen_reg_rtx (bl->biv->mode);
2768 emit_iv_add_mult (increment,
2769 gen_rtx (CONST_INT, VOIDmode, loop_n_iterations),
2770 bl->initial_value, tem, NEXT_INSN (loop_end));
2772 if (loop_dump_stream)
2773 fprintf (loop_dump_stream,
2774 "Final biv value for %d, calculated.\n", bl->regno);
2780 /* Check to see if the biv is dead at all loop exits. */
2781 if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end))
2783 if (loop_dump_stream)
2784 fprintf (loop_dump_stream,
2785 "Final biv value for %d, biv dead after loop exit.\n",
2794 /* Try to calculate the final value of the giv, the value it will have at
2795 the end of the loop. If we can do it, return that value. */
2798 final_giv_value (v, loop_start, loop_end)
2799 struct induction *v;
2800 rtx loop_start, loop_end;
2802 struct iv_class *bl;
2803 rtx reg, insn, pattern;
2808 bl = reg_biv_class[REGNO (v->src_reg)];
2810 /* The final value for givs which depend on reversed bivs must be calculated
2811 differently than for ordinary givs. In this case, there is already an
2812 insn after the loop which sets this giv's final value (if necessary),
2813 and there are no other loop exits, so we can return any value. */
2816 if (loop_dump_stream)
2817 fprintf (loop_dump_stream,
2818 "Final giv value for %d, depends on reversed biv\n",
2819 REGNO (v->dest_reg));
2823 /* Try to calculate the final value as a function of the biv it depends
2824 upon. The only exit from the loop must be the fall through at the bottom
2825 (otherwise it may not have its final value when the loop exits). */
2827 /* ??? Can calculate the final giv value by subtracting off the
2828 extra biv increments times the giv's mult_val. The loop must have
2829 only one exit for this to work, but the loop iterations does not need
2832 if (loop_n_iterations != 0
2833 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2835 /* ?? It is tempting to use the biv's value here since these insns will
2836 be put after the loop, and hence the biv will have its final value
2837 then. However, this fails if the biv is subsequently eliminated.
2838 Perhaps determine whether biv's are eliminable before trying to
2839 determine whether giv's are replaceable so that we can use the
2840 biv value here if it is not eliminable. */
2842 increment = biv_total_increment (bl, loop_start, loop_end);
2844 if (increment && invariant_p (increment))
2846 /* Can calculate the loop exit value of its biv as
2847 (loop_n_iterations * increment) + initial_value */
2849 /* The loop exit value of the giv is then
2850 (final_biv_value - extra increments) * mult_val + add_val.
2851 The extra increments are any increments to the biv which
2852 occur in the loop after the giv's value is calculated.
2853 We must search from the insn that sets the giv to the end
2854 of the loop to calculate this value. */
2856 insert_before = NEXT_INSN (loop_end);
2858 /* Put the final biv value in tem. */
2859 tem = gen_reg_rtx (bl->biv->mode);
2860 emit_iv_add_mult (increment,
2861 gen_rtx (CONST_INT, VOIDmode, loop_n_iterations),
2862 bl->initial_value, tem, insert_before);
2864 /* Subtract off extra increments as we find them. */
2865 for (insn = NEXT_INSN (v->insn); insn != loop_end;
2866 insn = NEXT_INSN (insn))
2868 if (GET_CODE (insn) == INSN
2869 && GET_CODE (PATTERN (insn)) == SET
2870 && SET_DEST (PATTERN (insn)) == v->src_reg)
2872 pattern = PATTERN (insn);
2873 if (GET_CODE (SET_SRC (pattern)) != PLUS)
2875 /* Sometimes a biv is computed in a temp reg,
2876 and then copied into the biv reg. */
2877 pattern = PATTERN (PREV_INSN (insn));
2878 if (GET_CODE (SET_SRC (pattern)) != PLUS)
2881 if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
2882 || REGNO (XEXP (SET_SRC (pattern), 0)) != bl->regno)
2885 tem = expand_binop (GET_MODE (tem), sub_optab, tem,
2886 XEXP (SET_SRC (pattern), 1), 0, 0,
2891 /* Now calculate the giv's final value. */
2892 emit_iv_add_mult (tem, v->mult_val, v->add_val, tem,
2895 if (loop_dump_stream)
2896 fprintf (loop_dump_stream,
2897 "Final giv value for %d, calc from biv's value.\n",
2898 REGNO (v->dest_reg));
2904 /* Replaceable giv's should never reach here. */
2908 /* Check to see if the biv is dead at all loop exits. */
2909 if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end))
2911 if (loop_dump_stream)
2912 fprintf (loop_dump_stream,
2913 "Final giv value for %d, giv dead after loop exit.\n",
2914 REGNO (v->dest_reg));
2923 /* Calculate the number of loop iterations. Returns the exact number of loop
2924 iterations if it can be calculated, otherwise retusns zero. */
2927 loop_iterations (loop_start, loop_end)
2928 rtx loop_start, loop_end;
2930 rtx comparison, comparison_value;
2931 rtx iteration_var, initial_value, increment, final_value;
2932 enum rtx_code comparison_code;
2933 int i, increment_dir;
2934 int unsigned_compare, compare_dir, final_larger;
2935 unsigned long tempu;
2938 /* First find the iteration variable. If the last insn is a conditional
2939 branch, and the insn before tests a register value, make that the
2940 iteration variable. */
2942 loop_initial_value = 0;
2944 loop_final_value = 0;
2945 loop_iteration_var = 0;
2947 last_loop_insn = prev_nonnote_insn (loop_end);
2949 comparison = get_condition_for_loop (last_loop_insn);
2950 if (comparison == 0)
2952 if (loop_dump_stream)
2953 fprintf (loop_dump_stream,
2954 "Loop unrolling: No final conditional branch found.\n");
2958 /* ??? Get_condition may switch position of induction variable and
2959 invariant register when it canonicalizes the comparison. */
2961 comparison_code = GET_CODE (comparison);
2962 iteration_var = XEXP (comparison, 0);
2963 comparison_value = XEXP (comparison, 1);
2965 if (GET_CODE (iteration_var) != REG)
2967 if (loop_dump_stream)
2968 fprintf (loop_dump_stream,
2969 "Loop unrolling: Comparison not against register.\n");
2973 /* Loop iterations is always called before any new registers are created
2974 now, so this should never occur. */
2976 if (REGNO (iteration_var) >= max_reg_before_loop)
2979 iteration_info (iteration_var, &initial_value, &increment,
2980 loop_start, loop_end);
2981 if (initial_value == 0)
2982 /* iteration_info already printed a message. */
2987 if (loop_dump_stream)
2988 fprintf (loop_dump_stream,
2989 "Loop unrolling: Increment value can't be calculated.\n");
2992 if (GET_CODE (increment) != CONST_INT)
2994 if (loop_dump_stream)
2995 fprintf (loop_dump_stream,
2996 "Loop unrolling: Increment value not constant.\n");
2999 if (GET_CODE (initial_value) != CONST_INT)
3001 if (loop_dump_stream)
3002 fprintf (loop_dump_stream,
3003 "Loop unrolling: Initial value not constant.\n");
3007 /* If the comparison value is an invariant register, then try to find
3008 its value from the insns before the start of the loop. */
3010 if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value))
3014 for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn))
3016 if (GET_CODE (insn) == CODE_LABEL)
3019 else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
3020 && (set = single_set (insn))
3021 && (SET_DEST (set) == comparison_value))
3023 rtx note = find_reg_note (insn, REG_EQUAL, 0);
3025 if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST)
3026 comparison_value = XEXP (note, 0);
3033 final_value = approx_final_value (comparison_code, comparison_value,
3034 &unsigned_compare, &compare_dir);
3036 /* Save the calculated values describing this loop's bounds, in case
3037 precondition_loop_p will need them later. These values can not be
3038 recalculated inside precondition_loop_p because strength reduction
3039 optimizations may obscure the loop's structure. */
3041 loop_iteration_var = iteration_var;
3042 loop_initial_value = initial_value;
3043 loop_increment = increment;
3044 loop_final_value = final_value;
3046 if (final_value == 0)
3048 if (loop_dump_stream)
3049 fprintf (loop_dump_stream,
3050 "Loop unrolling: EQ comparison loop.\n");
3053 else if (GET_CODE (final_value) != CONST_INT)
3055 if (loop_dump_stream)
3056 fprintf (loop_dump_stream,
3057 "Loop unrolling: Final value not constant.\n");
3061 /* ?? Final value and initial value do not have to be constants.
3062 Only their difference has to be constant. When the iteration variable
3063 is an array address, the final value and initial value might both
3064 be addresses with the same base but different constant offsets.
3065 Final value must be invariant for this to work.
3067 To do this, need someway to find the values of registers which are
3070 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3071 if (unsigned_compare)
3073 = ((unsigned) INTVAL (final_value) > (unsigned) INTVAL (initial_value)) -
3074 ((unsigned) INTVAL (final_value) < (unsigned) INTVAL (initial_value));
3076 final_larger = (INTVAL (final_value) > INTVAL (initial_value)) -
3077 (INTVAL (final_value) < INTVAL (initial_value));
3079 if (INTVAL (increment) > 0)
3081 else if (INTVAL (increment) == 0)
3086 /* There are 27 different cases: compare_dir = -1, 0, 1;
3087 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3088 There are 4 normal cases, 4 reverse cases (where the iteration variable
3089 will overflow before the loop exits), 4 infinite loop cases, and 15
3090 immediate exit (0 or 1 iteration depending on loop type) cases.
3091 Only try to optimize the normal cases. */
3093 /* (compare_dir/final_larger/increment_dir)
3094 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3095 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3096 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3097 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3099 /* ?? If the meaning of reverse loops (where the iteration variable
3100 will overflow before the loop exits) is undefined, then could
3101 eliminate all of these special checks, and just always assume
3102 the loops are normal/immediate/infinite. Note that this means
3103 the sign of increment_dir does not have to be known. Also,
3104 since it does not really hurt if immediate exit loops or infinite loops
3105 are optimized, then that case could be ignored also, and hence all
3106 loops can be optimized.
3108 According to ANSI Spec, the reverse loop case result is undefined,
3109 because the action on overflow is undefined.
3111 See also the special test for NE loops below. */
3113 if (final_larger == increment_dir && final_larger != 0
3114 && (final_larger == compare_dir || compare_dir == 0))
3119 if (loop_dump_stream)
3120 fprintf (loop_dump_stream,
3121 "Loop unrolling: Not normal loop.\n");
3125 /* Calculate the number of iterations, final_value is only an approximation,
3126 so correct for that. Note that tempu and loop_n_iterations are
3127 unsigned, because they can be as large as 2^n - 1. */
3129 i = INTVAL (increment);
3131 tempu = INTVAL (final_value) - INTVAL (initial_value);
3134 tempu = INTVAL (initial_value) - INTVAL (final_value);
3140 /* For NE tests, make sure that the iteration variable won't miss the
3141 final value. If tempu mod i is not zero, then the iteration variable
3142 will overflow before the loop exits, and we can not calculate the
3143 number of iterations. */
3144 if (compare_dir == 0 && (tempu % i) != 0)
3147 return tempu / i + ((tempu % i) != 0);