| blob | 61d63d8e2a359fc45d48275e98cd939ab83be173 |
1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995, 1997, 1998, 1999, 2000, 2001, 2002
3 Free Software Foundation, Inc.
4 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
21 02111-1307, USA. */
23 /* Try to unroll a loop, and split induction variables.
25 Loops for which the number of iterations can be calculated exactly are
26 handled specially. If the number of iterations times the insn_count is
27 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
28 Otherwise, we try to unroll the loop a number of times modulo the number
29 of iterations, so that only one exit test will be needed. It is unrolled
30 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
31 the insn count.
33 Otherwise, if the number of iterations can be calculated exactly at
34 run time, and the loop is always entered at the top, then we try to
35 precondition the loop. That is, at run time, calculate how many times
36 the loop will execute, and then execute the loop body a few times so
37 that the remaining iterations will be some multiple of 4 (or 2 if the
38 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
39 with only one exit test needed at the end of the loop.
41 Otherwise, if the number of iterations can not be calculated exactly,
42 not even at run time, then we still unroll the loop a number of times
43 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
44 but there must be an exit test after each copy of the loop body.
46 For each induction variable, which is dead outside the loop (replaceable)
47 or for which we can easily calculate the final value, if we can easily
48 calculate its value at each place where it is set as a function of the
49 current loop unroll count and the variable's value at loop entry, then
50 the induction variable is split into `N' different variables, one for
51 each copy of the loop body. One variable is live across the backward
52 branch, and the others are all calculated as a function of this variable.
53 This helps eliminate data dependencies, and leads to further opportunities
54 for cse. */
56 /* Possible improvements follow: */
58 /* ??? Add an extra pass somewhere to determine whether unrolling will
59 give any benefit. E.g. after generating all unrolled insns, compute the
60 cost of all insns and compare against cost of insns in rolled loop.
62 - On traditional architectures, unrolling a non-constant bound loop
63 is a win if there is a giv whose only use is in memory addresses, the
64 memory addresses can be split, and hence giv increments can be
65 eliminated.
66 - It is also a win if the loop is executed many times, and preconditioning
67 can be performed for the loop.
68 Add code to check for these and similar cases. */
70 /* ??? Improve control of which loops get unrolled. Could use profiling
71 info to only unroll the most commonly executed loops. Perhaps have
72 a user specifyable option to control the amount of code expansion,
73 or the percent of loops to consider for unrolling. Etc. */
75 /* ??? Look at the register copies inside the loop to see if they form a
76 simple permutation. If so, iterate the permutation until it gets back to
77 the start state. This is how many times we should unroll the loop, for
78 best results, because then all register copies can be eliminated.
79 For example, the lisp nreverse function should be unrolled 3 times
80 while (this)
81 {
82 next = this->cdr;
83 this->cdr = prev;
84 prev = this;
85 this = next;
86 }
88 ??? The number of times to unroll the loop may also be based on data
89 references in the loop. For example, if we have a loop that references
90 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
92 /* ??? Add some simple linear equation solving capability so that we can
93 determine the number of loop iterations for more complex loops.
94 For example, consider this loop from gdb
95 #define SWAP_TARGET_AND_HOST(buffer,len)
96 {
97 char tmp;
98 char *p = (char *) buffer;
99 char *q = ((char *) buffer) + len - 1;
100 int iterations = (len + 1) >> 1;
101 int i;
102 for (p; p < q; p++, q--;)
103 {
104 tmp = *q;
105 *q = *p;
106 *p = tmp;
107 }
108 }
109 Note that:
110 start value = p = &buffer + current_iteration
111 end value = q = &buffer + len - 1 - current_iteration
112 Given the loop exit test of "p < q", then there must be "q - p" iterations,
113 set equal to zero and solve for number of iterations:
114 q - p = len - 1 - 2*current_iteration = 0
115 current_iteration = (len - 1) / 2
116 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
117 iterations of this loop. */
119 /* ??? Currently, no labels are marked as loop invariant when doing loop
120 unrolling. This is because an insn inside the loop, that loads the address
121 of a label inside the loop into a register, could be moved outside the loop
122 by the invariant code motion pass if labels were invariant. If the loop
123 is subsequently unrolled, the code will be wrong because each unrolled
124 body of the loop will use the same address, whereas each actually needs a
125 different address. A case where this happens is when a loop containing
126 a switch statement is unrolled.
128 It would be better to let labels be considered invariant. When we
129 unroll loops here, check to see if any insns using a label local to the
130 loop were moved before the loop. If so, then correct the problem, by
131 moving the insn back into the loop, or perhaps replicate the insn before
132 the loop, one copy for each time the loop is unrolled. */
152 /* The prime factors looked for when trying to unroll a loop by some
153 number which is modulo the total number of iterations. Just checking
154 for these 4 prime factors will find at least one factor for 75% of
155 all numbers theoretically. Practically speaking, this will succeed
156 almost all of the time since loops are generally a multiple of 2
157 and/or 5. */
159 #define NUM_FACTORS 4
164 /* Describes the different types of loop unrolling performed. */
166 enum unroll_types
167 {
168 UNROLL_COMPLETELY,
169 UNROLL_MODULO,
170 UNROLL_NAIVE
171 };
173 /* Indexed by register number, if non-zero, then it contains a pointer
174 to a struct induction for a DEST_REG giv which has been combined with
175 one of more address givs. This is needed because whenever such a DEST_REG
176 giv is modified, we must modify the value of all split address givs
177 that were combined with this DEST_REG giv. */
181 /* Indexed by register number, if this is a splittable induction variable,
182 then this will hold the current value of the register, which depends on the
183 iteration number. */
187 /* Indexed by register number, if this is a splittable induction variable,
188 then this will hold the number of instructions in the loop that modify
189 the induction variable. Used to ensure that only the last insn modifying
190 a split iv will update the original iv of the dest. */
194 /* Forward declarations. */
216 /* Try to unroll one loop and split induction variables in the loop.
218 The loop is described by the arguments LOOP and INSN_COUNT.
219 STRENGTH_REDUCTION_P indicates whether information generated in the
220 strength reduction pass is available.
222 This function is intended to be called from within `strength_reduce'
223 in loop.c. */
225 void
230 {
240 rtx insert_before;
247 rtx start_label;
252 rtx safety_label;
253 /* This points to the last real insn in the loop, which should be either
254 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
255 jumps). */
256 rtx last_loop_insn;
260 /* Don't bother unrolling huge loops. Since the minimum factor is
261 two, loops greater than one half of MAX_UNROLLED_INSNS will never
262 be unrolled. */
264 {
268 }
270 /* Determine type of unroll to perform. Depends on the number of iterations
271 and the size of the loop. */
273 /* If there is no strength reduce info, then set
274 loop_info->n_iterations to zero. This can happen if
275 strength_reduce can't find any bivs in the loop. A value of zero
276 indicates that the number of iterations could not be calculated. */
282 {
287 }
289 /* Find and save a pointer to the last nonnote insn in the loop. */
293 /* Calculate how many times to unroll the loop. Indicate whether or
294 not the loop is being completely unrolled. */
297 {
298 /* Handle the case where the loop begins with an unconditional
299 jump to the loop condition. Make sure to delete the jump
300 insn, otherwise the loop body will never execute. */
306 /* If number of iterations is exactly 1, then eliminate the compare and
307 branch at the end of the loop since they will never be taken.
308 Then return, since no other action is needed here. */
310 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
311 don't do anything. */
314 {
315 /* Delete the jump insn. This will delete the barrier also. */
317 }
319 {
320 #ifdef HAVE_cc0
322 #endif
324 #ifdef HAVE_cc0
325 /* The immediately preceding insn may be a compare which must be
326 deleted. */
329 #endif
330 }
332 /* Remove the loop notes since this is no longer a loop. */
343 }
345 /* Avoid overflow in the next expression. */
348 {
351 }
353 {
354 /* Try to factor the number of iterations. Don't bother with the
355 general case, only using 2, 3, 5, and 7 will get 75% of all
356 numbers theoretically, and almost all in practice. */
364 {
367 }
369 /* Start with the larger factors first so that we generally
370 get lots of unrolling. */
376 {
378 {
381 }
382 else
384 }
386 /* If we couldn't find any factors, then unroll as in the normal
387 case. */
389 {
392 }
393 else
395 }
397 /* Default case, calculate number of times to unroll loop based on its
398 size. */
400 {
405 else
407 }
409 /* Now we know how many times to unroll the loop. */
415 {
416 /* Loops of these types can start with jump down to the exit condition
417 in rare circumstances.
419 Consider a pair of nested loops where the inner loop is part
420 of the exit code for the outer loop.
422 In this case jump.c will not duplicate the exit test for the outer
423 loop, so it will start with a jump to the exit code.
425 Then consider if the inner loop turns out to iterate once and
426 only once. We will end up deleting the jumps associated with
427 the inner loop. However, the loop notes are not removed from
428 the instruction stream.
430 And finally assume that we can compute the number of iterations
431 for the outer loop.
433 In this case unroll may want to unroll the outer loop even though
434 it starts with a jump to the outer loop's exit code.
436 We could try to optimize this case, but it hardly seems worth it.
437 Just return without unrolling the loop in such cases. */
444 }
447 {
448 /* Completely unrolling the loop: Delete the compare and branch at
449 the end (the last two instructions). This delete must done at the
450 very end of loop unrolling, to avoid problems with calls to
451 back_branch_in_range_p, which is called by find_splittable_regs.
452 All increments of splittable bivs/givs are changed to load constant
453 instructions. */
457 /* Set insert_before to the instruction immediately after the JUMP_INSN
458 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
459 the loop will be correctly handled by copy_loop_body. */
462 /* Set copy_end to the insn before the jump at the end of the loop. */
466 {
468 #ifdef HAVE_cc0
469 /* The instruction immediately before the JUMP_INSN may be a compare
470 instruction which we do not want to copy. */
473 #endif
474 }
475 else
476 {
477 /* We currently can't unroll a loop if it doesn't end with a
478 JUMP_INSN. There would need to be a mechanism that recognizes
479 this case, and then inserts a jump after each loop body, which
480 jumps to after the last loop body. */
485 }
486 }
488 {
489 /* Partially unrolling the loop: The compare and branch at the end
490 (the last two instructions) must remain. Don't copy the compare
491 and branch instructions at the end of the loop. Insert the unrolled
492 code immediately before the compare/branch at the end so that the
493 code will fall through to them as before. */
497 /* Set insert_before to the jump insn at the end of the loop.
498 Set copy_end to before the jump insn at the end of the loop. */
500 {
503 }
505 {
507 #ifdef HAVE_cc0
508 /* The instruction immediately before the JUMP_INSN may be a compare
509 instruction which we do not want to copy or delete. */
512 #endif
514 }
515 else
516 {
517 /* We currently can't unroll a loop if it doesn't end with a
518 JUMP_INSN. There would need to be a mechanism that recognizes
519 this case, and then inserts a jump after each loop body, which
520 jumps to after the last loop body. */
525 }
526 }
527 else
528 {
529 /* Normal case: Must copy the compare and branch instructions at the
530 end of the loop. */
533 {
534 /* Loop ends with an unconditional jump and a barrier.
535 Handle this like above, don't copy jump and barrier.
536 This is not strictly necessary, but doing so prevents generating
537 unconditional jumps to an immediately following label.
539 This will be corrected below if the target of this jump is
540 not the start_label. */
544 }
546 {
547 /* Set insert_before to immediately after the JUMP_INSN, so that
548 NOTEs at the end of the loop will be correctly handled by
549 copy_loop_body. */
552 }
553 else
554 {
555 /* We currently can't unroll a loop if it doesn't end with a
556 JUMP_INSN. There would need to be a mechanism that recognizes
557 this case, and then inserts a jump after each loop body, which
558 jumps to after the last loop body. */
563 }
565 /* If copying exit test branches because they can not be eliminated,
566 then must convert the fall through case of the branch to a jump past
567 the end of the loop. Create a label to emit after the loop and save
568 it for later use. Do not use the label after the loop, if any, since
569 it might be used by insns outside the loop, or there might be insns
570 added before it later by final_[bg]iv_value which must be after
571 the real exit label. */
579 {
580 /* The loop starts with a jump down to the exit condition test.
581 Start copying the loop after the barrier following this
582 jump insn. */
585 /* Splitting induction variables doesn't work when the loop is
586 entered via a jump to the bottom, because then we end up doing
587 a comparison against a new register for a split variable, but
588 we did not execute the set insn for the new register because
589 it was skipped over. */
594 }
595 else
597 }
599 /* This should always be the first label in the loop. */
601 /* There may be a line number note and/or a loop continue note here. */
605 {
606 /* This can happen as a result of jump threading. If the first insns in
607 the loop test the same condition as the loop's backward jump, or the
608 opposite condition, then the backward jump will be modified to point
609 to elsewhere, and the loop's start label is deleted.
611 This case currently can not be handled by the loop unrolling code. */
617 }
619 {
620 /* The jump optimization pass must have combined the original start label
621 with a named label for a goto. We can't unroll this case because
622 jumps which go to the named label must be handled differently than
623 jumps to the loop start, and it is impossible to differentiate them
624 in this case. */
629 }
635 {
636 /* In this case, we must copy the jump and barrier, because they will
637 not be converted to jumps to an immediately following label. */
641 }
646 {
647 /* ??? The loop ends with a conditional branch that does not branch back
648 to the loop start label. In this case, we must emit an unconditional
649 branch to the loop exit after emitting the final branch.
650 copy_loop_body does not have support for this currently, so we
651 give up. It doesn't seem worthwhile to unroll anyways since
652 unrolling would increase the number of branch instructions
653 executed. */
658 }
660 /* Allocate a translation table for the labels and insn numbers.
661 They will be filled in as we copy the insns in the loop. */
666 /* Various paths through the unroll code may reach the "egress" label
667 without initializing fields within the map structure.
669 To be safe, we use xcalloc to zero the memory. */
672 /* Allocate the label map. */
675 {
678 }
680 /* Search the loop and mark all local labels, i.e. the ones which have to
681 be distinct labels when copied. For all labels which might be
682 non-local, set their label_map entries to point to themselves.
683 If they happen to be local their label_map entries will be overwritten
684 before the loop body is copied. The label_map entries for local labels
685 will be set to a different value each time the loop body is copied. */
688 {
689 rtx note;
694 {
701 {
705 rtx label;
708 {
711 }
712 }
713 }
717 }
719 /* Allocate space for the insn map. */
723 /* Set this to zero, to indicate that we are doing loop unrolling,
724 not function inlining. */
727 /* The register and constant maps depend on the number of registers
728 present, so the final maps can't be created until after
729 find_splittable_regs is called. However, they are needed for
730 preconditioning, so we create temporary maps when preconditioning
731 is performed. */
733 /* The preconditioning code may allocate two new pseudo registers. */
736 /* local_regno is only valid for regnos < max_local_regnum. */
739 /* Allocate and zero out the splittable_regs and addr_combined_regs
740 arrays. These must be zeroed here because they will be used if
741 loop preconditioning is performed, and must be zero for that case.
743 It is safe to do this here, since the extra registers created by the
744 preconditioning code and find_splittable_regs will never be used
745 to access the splittable_regs[] and addr_combined_regs[] arrays. */
749 addr_combined_regs
753 /* Mark all local registers, i.e. the ones which are referenced only
754 inside the loop. */
756 {
760 /* If a register is used in the jump insn, we must not duplicate it
761 since it will also be used outside the loop. */
763 copy_end_luid--;
765 /* If we have a target that uses cc0, then we also must not duplicate
766 the insn that sets cc0 before the jump insn, if one is present. */
767 #ifdef HAVE_cc0
770 copy_end_luid--;
771 #endif
773 /* If copy_start points to the NOTE that starts the loop, then we must
774 use the next luid, because invariant pseudo-regs moved out of the loop
775 have their lifetimes modified to start here, but they are not safe
776 to duplicate. */
778 copy_start_luid++;
780 /* If a pseudo's lifetime is entirely contained within this loop, then we
781 can use a different pseudo in each unrolled copy of the loop. This
782 results in better code. */
783 /* We must limit the generic test to max_reg_before_loop, because only
784 these pseudo registers have valid regno_first_uid info. */
790 {
791 /* However, we must also check for loop-carried dependencies.
792 If the value the pseudo has at the end of iteration X is
793 used by iteration X+1, then we can not use a different pseudo
794 for each unrolled copy of the loop. */
795 /* A pseudo is safe if regno_first_uid is a set, and this
796 set dominates all instructions from regno_first_uid to
797 regno_last_uid. */
798 /* ??? This check is simplistic. We would get better code if
799 this check was more sophisticated. */
805 {
808 else
810 r);
811 }
812 }
813 }
815 /* If this loop requires exit tests when unrolled, check to see if we
816 can precondition the loop so as to make the exit tests unnecessary.
817 Just like variable splitting, this is not safe if the loop is entered
818 via a jump to the bottom. Also, can not do this if no strength
819 reduce info, because precondition_loop_p uses this info. */
821 /* Must copy the loop body for preconditioning before the following
822 find_splittable_regs call since that will emit insns which need to
823 be after the preconditioned loop copies, but immediately before the
824 unrolled loop copies. */
826 /* Also, it is not safe to split induction variables for the preconditioned
827 copies of the loop body. If we split induction variables, then the code
828 assumes that each induction variable can be represented as a function
829 of its initial value and the loop iteration number. This is not true
830 in this case, because the last preconditioned copy of the loop body
831 could be any iteration from the first up to the `unroll_number-1'th,
832 depending on the initial value of the iteration variable. Therefore
833 we can not split induction variables here, because we can not calculate
834 their value. Hence, this code must occur before find_splittable_regs
835 is called. */
838 {
845 {
846 rtx diff;
861 /* Limit loop unrolling to 4, since this will make 7 copies of
862 the loop body. */
866 /* Save the absolute value of the increment, and also whether or
867 not it is negative. */
871 {
874 }
878 /* Final value may have form of (PLUS val1 const1_rtx). We need
879 to convert it into general operand, so compute the real value. */
882 {
887 }
891 /* Calculate the difference between the final and initial values.
892 Final value may be a (plus (reg x) (const_int 1)) rtx.
893 Let the following cse pass simplify this if initial value is
894 a constant.
896 We must copy the final and initial values here to avoid
897 improperly shared rtl.
899 We have to deal with for (i = 0; --i < 6;) type loops.
900 For such loops the real final value is the first time the
901 loop variable overflows, so the diff we calculate is the
902 distance from the overflow value. This is 0 or ~0 for
903 unsigned loops depending on the direction, or INT_MAX,
904 INT_MAX+1 for signed loops. We really do not need the
905 exact value, since we are only interested in the diff
906 modulo the increment, and the increment is a power of 2,
907 so we can pretend that the overflow value is 0/~0. */
912 OPTAB_LIB_WIDEN);
913 else
917 /* Now calculate (diff % (unroll * abs (increment))) by using an
918 and instruction. */
923 /* Now emit a sequence of branches to jump to the proper precond
924 loop entry point. */
930 /* Check for the case where the initial value is greater than or
931 equal to the final value. In that case, we want to execute
932 exactly one loop iteration. The code below will fail for this
933 case. This check does not apply if the loop has a NE
934 comparison at the end. */
937 {
938 rtx incremented_initval;
940 initial_value,
941 increment,
943 OPTAB_LIB_WIDEN);
948 TAKEN);
951 }
953 /* Assuming the unroll_number is 4, and the increment is 2, then
954 for a negative increment: for a positive increment:
955 diff = 0,1 precond 0 diff = 0,7 precond 0
956 diff = 2,3 precond 3 diff = 1,2 precond 1
957 diff = 4,5 precond 2 diff = 3,4 precond 2
958 diff = 6,7 precond 1 diff = 5,6 precond 3 */
960 /* We only need to emit (unroll_number - 1) branches here, the
961 last case just falls through to the following code. */
963 /* ??? This would give better code if we emitted a tree of branches
964 instead of the current linear list of branches. */
967 {
971 /* For negative increments, must invert the constant compared
972 against, except when comparing against zero. */
974 {
977 }
979 {
982 }
983 else
984 {
987 }
995 }
997 /* If the increment is greater than one, then we need another branch,
998 to handle other cases equivalent to 0. */
1000 /* ??? This should be merged into the code above somehow to help
1001 simplify the code here, and reduce the number of branches emitted.
1002 For the negative increment case, the branch here could easily
1003 be merged with the `0' case branch above. For the positive
1004 increment case, it is not clear how this can be simplified. */
1007 {
1012 {
1015 }
1016 else
1017 {
1020 }
1026 }
1032 /* Only the last copy of the loop body here needs the exit
1033 test, so set copy_end to exclude the compare/branch here,
1034 and then reset it inside the loop when get to the last
1035 copy. */
1040 {
1042 #ifdef HAVE_cc0
1043 /* The immediately preceding insn may be a compare which
1044 we do not want to copy. */
1047 #endif
1048 }
1049 else
1053 {
1069 {
1074 }
1075 /* The last copy needs the compare/branch insns at the end,
1076 so reset copy_end here if the loop ends with a conditional
1077 branch. */
1080 {
1083 else
1085 }
1087 /* None of the copies are the `last_iteration', so just
1088 pass zero for that parameter. */
1092 }
1096 {
1099 }
1100 else
1101 {
1103 #ifdef HAVE_cc0
1104 /* The instruction immediately before the JUMP_INSN may
1105 be a compare instruction which we do not want to copy
1106 or delete. */
1109 #endif
1111 }
1113 /* Set unroll type to MODULO now. */
1117 /* Clean up. */
1119 }
1120 }
1122 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1123 the loop unless all loops are being unrolled. */
1125 {
1130 }
1132 /* At this point, we are guaranteed to unroll the loop. */
1134 /* Keep track of the unroll factor for the loop. */
1137 /* And whether the loop has been preconditioned. */
1140 /* For each biv and giv, determine whether it can be safely split into
1141 a different variable for each unrolled copy of the loop body.
1142 We precalculate and save this info here, since computing it is
1143 expensive.
1145 Do this before deleting any instructions from the loop, so that
1146 back_branch_in_range_p will work correctly. */
1150 else
1153 /* find_splittable_regs may have created some new registers, so must
1154 reallocate the reg_map with the new larger size, and must realloc
1155 the constant maps also. */
1168 /* Search the list of bivs and givs to find ones which need to be remapped
1169 when split, and set their reg_map entry appropriately. */
1172 {
1175 #if 0
1176 /* Currently, non-reduced/final-value givs are never split. */
1180 #endif
1181 }
1183 /* Use our current register alignment and pointer flags. */
1187 /* If the loop is being partially unrolled, and the iteration variables
1188 are being split, and are being renamed for the split, then must fix up
1189 the compare/jump instruction at the end of the loop to refer to the new
1190 registers. This compare isn't copied, so the registers used in it
1191 will never be replaced if it isn't done here. */
1194 {
1198 }
1200 /* For unroll_number times, make a copy of each instruction
1201 between copy_start and copy_end, and insert these new instructions
1202 before the end of the loop. */
1205 {
1217 {
1221 }
1223 /* If loop starts with a branch to the test, then fix it so that
1224 it points to the test of the first unrolled copy of the loop. */
1226 {
1231 CODE_LABEL_NUMBER
1235 /* Set the jump label so that it can be used by later loop unrolling
1236 passes. */
1239 }
1244 }
1246 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1247 insn to be deleted. This prevents any runaway delete_insn call from
1248 more insns that it should, as it always stops at a CODE_LABEL. */
1250 /* Delete the compare and branch at the end of the loop if completely
1251 unrolling the loop. Deleting the backward branch at the end also
1252 deletes the code label at the start of the loop. This is done at
1253 the very end to avoid problems with back_branch_in_range_p. */
1257 else
1260 /* Delete all of the original loop instructions. Don't delete the
1261 LOOP_BEG note, or the first code label in the loop. */
1265 {
1266 /* ??? Don't delete named code labels. They will be deleted when the
1267 jump that references them is deleted. Otherwise, we end up deleting
1268 them twice, which causes them to completely disappear instead of turn
1269 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1270 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1271 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1272 associated LABEL_DECL to point to one of the new label instances. */
1273 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1279 else
1281 }
1283 /* Can now delete the 'safety' label emitted to protect us from runaway
1284 delete_related_insns calls. */
1289 /* If exit_label exists, emit it after the loop. Doing the emit here
1290 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1291 This is needed so that mostly_true_jump in reorg.c will treat jumps
1292 to this loop end label correctly, i.e. predict that they are usually
1293 not taken. */
1297 egress:
1299 {
1300 /* Remove the loop notes since this is no longer a loop. */
1309 }
1314 {
1317 }
1326 }
1327 \f
1328 /* Return true if the loop can be safely, and profitably, preconditioned
1329 so that the unrolled copies of the loop body don't need exit tests.
1331 This only works if final_value, initial_value and increment can be
1332 determined, and if increment is a constant power of 2.
1333 If increment is not a power of 2, then the preconditioning modulo
1334 operation would require a real modulo instead of a boolean AND, and this
1335 is not considered `profitable'. */
1337 /* ??? If the loop is known to be executed very many times, or the machine
1338 has a very cheap divide instruction, then preconditioning is a win even
1339 when the increment is not a power of 2. Use RTX_COST to compute
1340 whether divide is cheap.
1341 ??? A divide by constant doesn't actually need a divide, look at
1342 expand_divmod. The reduced cost of this optimized modulo is not
1343 reflected in RTX_COST. */
1345 int
1350 {
1355 {
1357 {
1361 }
1362 else
1363 {
1367 }
1371 {
1373 loop_dump_stream);
1377 }
1379 }
1382 {
1387 }
1389 {
1394 }
1396 {
1401 }
1403 {
1408 }
1411 {
1416 }
1418 /* Unsigned_compare and compare_dir can be ignored here, since they do
1419 not matter for preconditioning. */
1422 {
1427 }
1429 /* Must ensure that final_value is invariant, so call
1430 loop_invariant_p to check. Before doing so, must check regno
1431 against max_reg_before_loop to make sure that the register is in
1432 the range covered by loop_invariant_p. If it isn't, then it is
1433 most likely a biv/giv which by definition are not invariant. */
1439 {
1444 }
1446 /* Fail for floating point values, since the caller of this function
1447 does not have code to deal with them. */
1450 {
1455 }
1457 /* Fail if loop_info->iteration_var is not live before loop_start,
1458 since we need to test its value in the preconditioning code. */
1462 {
1467 }
1469 /* Note that loop_iterations biases the initial value for GIV iterators
1470 such as "while (i-- > 0)" so that we can calculate the number of
1471 iterations just like for BIV iterators.
1473 Also note that the absolute values of initial_value and
1474 final_value are unimportant as only their difference is used for
1475 calculating the number of loop iterations. */
1480 /* Decide what mode to do these calculations in. Choose the larger
1481 of final_value's mode and initial_value's mode, or a full-word if
1482 both are constants. */
1485 {
1489 }
1495 /* Success! */
1499 }
1501 /* All pseudo-registers must be mapped to themselves. Two hard registers
1502 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1503 REGNUM, to avoid function-inlining specific conversions of these
1504 registers. All other hard regs can not be mapped because they may be
1505 used with different
1506 modes. */
1508 static void
1512 {
1517 /* Just clear the rest of the entries. */
1525 }
1526 \f
1527 /* Strength-reduction will often emit code for optimized biv/givs which
1528 calculates their value in a temporary register, and then copies the result
1529 to the iv. This procedure reconstructs the pattern computing the iv;
1530 verifying that all operands are of the proper form.
1532 PATTERN must be the result of single_set.
1533 The return value is the amount that the giv is incremented by. */
1535 static rtx
1539 {
1540 rtx increment;
1544 retry:
1545 /* Verify that we have an increment insn here. First check for a plus
1546 as the set source. */
1548 {
1549 /* SR sometimes computes the new giv value in a temp, then copies it
1550 to the new_reg. */
1556 /* The last insn emitted is not needed, so delete it to avoid confusing
1557 the second cse pass. This insn sets the giv unnecessarily. */
1559 }
1561 /* Verify that we have a constant as the second operand of the plus. */
1564 {
1565 /* SR sometimes puts the constant in a register, especially if it is
1566 too big to be an add immed operand. */
1569 /* SR may have used LO_SUM to compute the constant if it is too large
1570 for a load immed operand. In this case, the constant is in operand
1571 one of the LO_SUM rtx. */
1575 /* Some ports store large constants in memory and add a REG_EQUAL
1576 note to the store insn. */
1578 {
1582 }
1587 {
1588 /* The rs6000 port loads some constants with IOR.
1589 The alpha port loads some constants with ASHIFT and PLUS. */
1595 /* Don't need the last insn anymore. */
1606 else
1608 }
1613 /* The insn loading the constant into a register is no longer needed,
1614 so delete it. */
1616 }
1620 else
1623 /* Check that the source register is the same as the register we expected
1624 to see as the source. If not, something is seriously wrong. */
1627 {
1628 /* Some machines (e.g. the romp), may emit two add instructions for
1629 certain constants, so lets try looking for another add immediately
1630 before this one if we have only seen one add insn so far. */
1633 {
1634 tries++;
1642 }
1645 }
1648 }
1650 /* Copy REG_NOTES, except for insn references, because not all insn_map
1651 entries are valid yet. We do need to copy registers now though, because
1652 the reg_map entries can change during copying. */
1654 static rtx
1656 rtx notes;
1658 {
1659 rtx copy;
1670 /* Don't substitute for these yet. */
1672 else
1678 }
1680 /* Fixup insn references in copied REG_NOTES. */
1682 static void
1686 {
1688 {
1692 {
1693 /* Sometimes, we have a REG_WAS_0 note that points to a
1694 deleted instruction. In that case, we can just delete the
1695 note. */
1697 {
1700 }
1701 else
1702 {
1705 /* If we failed to remap the note, something is awry.
1706 Allow REG_LABEL as it may reference label outside
1707 the unrolled loop. */
1709 {
1712 }
1713 else
1715 }
1716 }
1719 }
1720 }
1722 /* Copy each instruction in the loop, substituting from map as appropriate.
1723 This is very similar to a loop in expand_inline_function. */
1725 static void
1728 copy_notes_from)
1732 rtx exit_label;
1736 {
1741 #ifdef HAVE_cc0
1743 #endif
1747 /* If this isn't the last iteration, then map any references to the
1748 start_label to final_label. Final label will then be emitted immediately
1749 after the end of this loop body if it was ever used.
1751 If this is the last iteration, then map references to the start_label
1752 to itself. */
1754 {
1757 }
1758 else
1764 do
1765 {
1771 {
1777 /* Check to see if this is a giv that has been combined with
1778 some split address givs. (Combined in the sense that
1779 `combine_givs' in loop.c has put two givs in the same register.)
1780 In this case, we must search all givs based on the same biv to
1781 find the address givs. Then split the address givs.
1782 Do this before splitting the giv, since that may map the
1783 SET_DEST to a new register. */
1788 {
1796 /* Although the giv_inc amount is not needed here, we must call
1797 calculate_giv_inc here since it might try to delete the
1798 last insn emitted. If we wait until later to call it,
1799 we might accidentally delete insns generated immediately
1800 below by emit_unrolled_add. */
1804 /* Now find all address giv's that were combined with this
1805 giv 'v'. */
1808 {
1811 /* If this DEST_ADDR giv was not split, then ignore it. */
1815 /* Scale this_giv_inc if the multiplicative factors of
1816 the two givs are different. */
1826 {
1827 /* Must emit an insn to increment the split address
1828 giv. Add in the const_adjust field in case there
1829 was a constant eliminated from the address. */
1832 /* tv->dest_reg will be either a bare register,
1833 or else a register plus a constant. */
1836 else
1839 /* Check for shared address givs, and avoid
1840 incrementing the shared pseudo reg more than
1841 once. */
1843 {
1844 /* tv->dest_reg may actually be a (PLUS (REG)
1845 (CONST)) here, so we must call plus_constant
1846 to add the const_adjust amount before calling
1847 emit_unrolled_add below. */
1852 {
1853 /* The constant could be too large for an add
1854 immediate, so can't directly emit an insn
1855 here. */
1858 }
1859 }
1861 /* Reset the giv to be just the register again, in case
1862 it is used after the set we have just emitted.
1863 We must subtract the const_adjust factor added in
1864 above. */
1868 }
1869 }
1870 }
1872 /* If this is a setting of a splittable variable, then determine
1873 how to split the variable, create a new set based on this split,
1874 and set up the reg_map so that later uses of the variable will
1875 use the new split variable. */
1882 {
1890 /* Compute the increment value for the giv, if it wasn't
1891 already computed above. */
1898 {
1899 /* Completely unrolling the loop. Set the induction
1900 variable to a known constant value. */
1902 /* The value in splittable_regs may be an invariant
1903 value, so we must use plus_constant here. */
1909 {
1912 }
1913 else
1914 {
1915 /* The splittable_regs value must be a REG or a
1916 CONST_INT, so put the entire value in the giv_src_reg
1917 variable. */
1920 }
1921 }
1922 else
1923 {
1924 /* Partially unrolling loop. Create a new pseudo
1925 register for the iteration variable, and set it to
1926 be a constant plus the original register. Except
1927 on the last iteration, when the result has to
1928 go back into the original iteration var register. */
1930 /* Handle bivs which must be mapped to a new register
1931 when split. This happens for bivs which need their
1932 final value set before loop entry. The new register
1933 for the biv was stored in the biv's first struct
1934 induction entry by find_splittable_regs. */
1938 {
1941 }
1943 #if 0
1944 /* If non-reduced/final-value givs were split, then
1945 this would have to remap those givs also. See
1946 find_splittable_regs. */
1947 #endif
1951 giv_inc,
1955 /* Now split the induction variable by changing the dest
1956 of this insn to a new register, and setting its
1957 reg_map entry to point to this new register.
1959 If this is the last iteration, and this is the last insn
1960 that will update the iv, then reuse the original dest,
1961 to ensure that the iv will have the proper value when
1962 the loop exits or repeats.
1964 Using splittable_regs_updates here like this is safe,
1965 because it can only be greater than one if all
1966 instructions modifying the iv are always executed in
1967 order. */
1971 {
1976 giv_inc == const0_rtx
1977 ? giv_src_reg
1981 }
1982 else
1984 }
1986 /* The constant being added could be too large for an add
1987 immediate, so can't directly emit an insn here. */
1991 }
1992 else
1993 {
1996 }
2000 #ifdef HAVE_cc0
2001 /* If this insn is setting CC0, it may need to look at
2002 the insn that uses CC0 to see what type of insn it is.
2003 In that case, the call to recog via validate_change will
2004 fail. So don't substitute constants here. Instead,
2005 do it when we emit the following insn.
2007 For example, see the pyr.md file. That machine has signed and
2008 unsigned compares. The compare patterns must check the
2009 following branch insn to see which what kind of compare to
2010 emit.
2012 If the previous insn set CC0, substitute constants on it as
2013 well. */
2016 else
2017 {
2022 }
2023 #else
2025 #endif
2027 /* Make split induction variable constants `permanent' since we
2028 know there are no backward branches across iteration variable
2029 settings which would invalidate this. */
2031 {
2038 }
2048 {
2050 CODE_LABEL_NUMBER
2053 }
2056 {
2058 /* This is a branch to the beginning of the loop; this is the
2059 last insn being copied; and this is not the last iteration.
2060 In this case, we want to change the original fall through
2061 case to be a branch past the end of the loop, and the
2062 original jump label case to fall_through. */
2065 {
2066 rtx jmp;
2068 /* Can't do it by reversing the jump (probably because we
2069 couldn't reverse the conditions), so emit a new
2070 jump_insn after COPY, and redirect the jump around
2071 that. */
2080 }
2081 }
2083 #ifdef HAVE_cc0
2087 #endif
2090 /* Set the jump label of COPY correctly to avoid problems with
2091 later passes of unroll_loop, if INSN had jump label set. */
2093 {
2096 /* Can't use the label_map for every insn, since this may be
2097 the backward branch, and hence the label was not mapped. */
2099 {
2104 {
2107 else
2109 }
2110 }
2114 else
2115 {
2116 /* An unrecognizable jump insn, probably the entry jump
2117 for a switch statement. This label must have been mapped,
2118 so just use the label_map to get the new jump label. */
2122 }
2124 /* If this is a non-local jump, then must increase the label
2125 use count so that the label will not be deleted when the
2126 original jump is deleted. */
2128 }
2131 {
2139 }
2141 /* If this used to be a conditional jump insn but whose branch
2142 direction is now known, we must do something special. */
2144 {
2145 #ifdef HAVE_cc0
2146 /* If the previous insn set cc0 for us, delete it. */
2149 #endif
2151 /* If this is now a no-op, delete it. */
2153 {
2156 }
2157 else
2158 /* Otherwise, this is unconditional jump so we must put a
2159 BARRIER after it. We could do some dead code elimination
2160 here, but jump.c will do it just as well. */
2162 }
2172 /* Because the USAGE information potentially contains objects other
2173 than hard registers, we need to copy it. */
2178 #ifdef HAVE_cc0
2182 #endif
2185 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2191 /* If this is the loop start label, then we don't need to emit a
2192 copy of this label since no one will use it. */
2195 {
2199 }
2207 /* VTOP and CONT notes are valid only before the loop exit test.
2208 If placed anywhere else, loop may generate bad code. */
2209 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2210 the associated rtl. We do not want to share the structure in
2211 this new block. */
2221 else
2227 }
2230 }
2233 /* Now finish coping the REG_NOTES. */
2235 do
2236 {
2242 }
2245 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2246 each of these notes here, since there may be some important ones, such as
2247 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2248 iteration, because the original notes won't be deleted.
2250 We can't use insert_before here, because when from preconditioning,
2251 insert_before points before the loop. We can't use copy_end, because
2252 there may be insns already inserted after it (which we don't want to
2253 copy) when not from preconditioning code. */
2256 {
2258 {
2259 /* VTOP notes are valid only before the loop exit test.
2260 If placed anywhere else, loop may generate bad code.
2261 Although COPY_NOTES_FROM will be at most one or two (for cc0)
2262 instructions before the last insn in the loop, COPY_NOTES_FROM
2263 can be a NOTE_INSN_LOOP_CONT note if there is no VTOP note,
2264 as in a do .. while loop. */
2271 }
2272 }
2280 }
2281 \f
2282 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2283 emitted. This will correctly handle the case where the increment value
2284 won't fit in the immediate field of a PLUS insns. */
2286 void
2289 {
2290 rtx result;
2297 }
2298 \f
2299 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2300 is a backward branch in that range that branches to somewhere between
2301 LOOP->START and INSN. Returns 0 otherwise. */
2303 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2304 In practice, this is not a problem, because this function is seldom called,
2305 and uses a negligible amount of CPU time on average. */
2307 int
2310 rtx insn;
2311 {
2317 /* Stop before we get to the backward branch at the end of the loop. */
2322 /* Check in case insn has been deleted, search forward for first non
2323 deleted insn following it. */
2327 /* Check for the case where insn is the last insn in the loop. Deal
2328 with the case where INSN was a deleted loop test insn, in which case
2329 it will now be the NOTE_LOOP_END. */
2334 {
2336 {
2339 /* Search from loop_start to insn, to see if one of them is
2340 the target_insn. We can't use INSN_LUID comparisons here,
2341 since insn may not have an LUID entry. */
2345 }
2346 }
2349 }
2351 /* Try to generate the simplest rtx for the expression
2352 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2353 value of giv's. */
2355 static rtx
2359 {
2361 rtx result;
2363 /* The modes must all be the same. This should always be true. For now,
2364 check to make sure. */
2370 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2371 will be a constant. */
2373 {
2377 }
2383 /* Again, put the constant second. */
2385 {
2389 }
2396 }
2398 /* Searches the list of induction struct's for the biv BL, to try to calculate
2399 the total increment value for one iteration of the loop as a constant.
2401 Returns the increment value as an rtx, simplified as much as possible,
2402 if it can be calculated. Otherwise, returns 0. */
2404 rtx
2407 {
2409 rtx result;
2411 /* For increment, must check every instruction that sets it. Each
2412 instruction must be executed only once each time through the loop.
2413 To verify this, we check that the insn is always executed, and that
2414 there are no backward branches after the insn that branch to before it.
2415 Also, the insn must have a mult_val of one (to make sure it really is
2416 an increment). */
2420 {
2425 else
2427 }
2430 }
2432 /* For each biv and giv, determine whether it can be safely split into
2433 a different variable for each unrolled copy of the loop body. If it
2434 is safe to split, then indicate that by saving some useful info
2435 in the splittable_regs array.
2437 If the loop is being completely unrolled, then splittable_regs will hold
2438 the current value of the induction variable while the loop is unrolled.
2439 It must be set to the initial value of the induction variable here.
2440 Otherwise, splittable_regs will hold the difference between the current
2441 value of the induction variable and the value the induction variable had
2442 at the top of the loop. It must be set to the value 0 here.
2444 Returns the total number of instructions that set registers that are
2445 splittable. */
2447 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2448 constant values are unnecessary, since we can easily calculate increment
2449 values in this case even if nothing is constant. The increment value
2450 should not involve a multiply however. */
2452 /* ?? Even if the biv/giv increment values aren't constant, it may still
2453 be beneficial to split the variable if the loop is only unrolled a few
2454 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2456 static int
2461 {
2466 rtx biv_final_value;
2471 {
2472 /* Biv_total_increment must return a constant value,
2473 otherwise we can not calculate the split values. */
2479 /* The loop must be unrolled completely, or else have a known number
2480 of iterations and only one exit, or else the biv must be dead
2481 outside the loop, or else the final value must be known. Otherwise,
2482 it is unsafe to split the biv since it may not have the proper
2483 value on loop exit. */
2485 /* loop_number_exit_count is non-zero if the loop has an exit other than
2486 a fall through at the end. */
2501 /* If any of the insns setting the BIV don't do so with a simple
2502 PLUS, we don't know how to split it. */
2510 /* If final value is non-zero, then must emit an instruction which sets
2511 the value of the biv to the proper value. This is done after
2512 handling all of the givs, since some of them may need to use the
2513 biv's value in their initialization code. */
2515 /* This biv is splittable. If completely unrolling the loop, save
2516 the biv's initial value. Otherwise, save the constant zero. */
2519 {
2521 {
2522 /* If the initial value of the biv is itself (i.e. it is too
2523 complicated for strength_reduce to compute), or is a hard
2524 register, or it isn't invariant, then we must create a new
2525 pseudo reg to hold the initial value of the biv. */
2531 {
2544 }
2545 else
2547 }
2548 else
2551 /* Save the number of instructions that modify the biv, so that
2552 we can treat the last one specially. */
2560 }
2562 /* Check every giv that depends on this biv to see whether it is
2563 splittable also. Even if the biv isn't splittable, givs which
2564 depend on it may be splittable if the biv is live outside the
2565 loop, and the givs aren't. */
2568 unroll_number);
2570 /* If final value is non-zero, then must emit an instruction which sets
2571 the value of the biv to the proper value. This is done after
2572 handling all of the givs, since some of them may need to use the
2573 biv's value in their initialization code. */
2575 {
2576 /* If the loop has multiple exits, emit the insns before the
2577 loop to ensure that it will always be executed no matter
2578 how the loop exits. Otherwise emit the insn after the loop,
2579 since this is slightly more efficient. */
2582 biv_final_value));
2583 else
2584 {
2585 /* Create a new register to hold the value of the biv, and then
2586 set the biv to its final value before the loop start. The biv
2587 is set to its final value before loop start to ensure that
2588 this insn will always be executed, no matter how the loop
2589 exits. */
2595 biv_final_value));
2601 /* Set up the mapping from the original biv register to the new
2602 register. */
2604 }
2605 }
2606 }
2608 }
2610 /* For every giv based on the biv BL, check to determine whether it is
2611 splittable. This is a subroutine to find_splittable_regs ().
2613 Return the number of instructions that set splittable registers. */
2615 static int
2620 rtx increment;
2622 {
2625 rtx final_value;
2626 rtx tem;
2629 /* Scan the list of givs, and set the same_insn field when there are
2630 multiple identical givs in the same insn. */
2638 {
2641 /* Only split the giv if it has already been reduced, or if the loop is
2642 being completely unrolled. */
2646 /* The giv can be split if the insn that sets the giv is executed once
2647 and only once on every iteration of the loop. */
2648 /* An address giv can always be split. v->insn is just a use not a set,
2649 and hence it does not matter whether it is always executed. All that
2650 matters is that all the biv increments are always executed, and we
2651 won't reach here if they aren't. */
2657 /* The giv increment value must be a constant. */
2663 /* The loop must be unrolled completely, or else have a known number of
2664 iterations and only one exit, or else the giv must be dead outside
2665 the loop, or else the final value of the giv must be known.
2666 Otherwise, it is not safe to split the giv since it may not have the
2667 proper value on loop exit. */
2669 /* The used outside loop test will fail for DEST_ADDR givs. They are
2670 never used outside the loop anyways, so it is always safe to split a
2671 DEST_ADDR giv. */
2677 /* The next part is true if the pseudo is used outside the loop.
2678 We assume that this is true for any pseudo created after loop
2679 starts, because we don't have a reg_n_info entry for them. */
2682 /* Check for the case where the pseudo is set by a shift/add
2683 sequence, in which case the first insn setting the pseudo
2684 is the first insn of the shift/add sequence. */
2688 /* Line above always fails if INSN was moved by loop opt. */
2694 #if 0
2695 /* Currently, non-reduced/final-value givs are never split. */
2696 /* Should emit insns after the loop if possible, as the biv final value
2697 code below does. */
2699 /* If the final value is non-zero, and the giv has not been reduced,
2700 then must emit an instruction to set the final value. */
2702 {
2703 /* Create a new register to hold the value of the giv, and then set
2704 the giv to its final value before the loop start. The giv is set
2705 to its final value before loop start to ensure that this insn
2706 will always be executed, no matter how we exit. */
2716 }
2717 #endif
2719 /* This giv is splittable. If completely unrolling the loop, save the
2720 giv's initial value. Otherwise, save the constant zero for it. */
2723 {
2724 /* It is not safe to use bl->initial_value here, because it may not
2725 be invariant. It is safe to use the initial value stored in
2726 the splittable_regs array if it is set. In rare cases, it won't
2727 be set, so then we do exactly the same thing as
2728 find_splittable_regs does to get a safe value. */
2729 rtx biv_initial_value;
2737 else
2738 {
2744 }
2748 }
2749 else
2753 {
2754 /* If a giv was combined with another giv, then we can only split
2755 this giv if the giv it was combined with was reduced. This
2756 is because the value of v->new_reg is meaningless in this
2757 case. */
2759 {
2764 }
2765 /* If the giv is an address destination, it could be something other
2766 than a simple register, these have to be treated differently. */
2768 {
2769 /* If value is not a constant, register, or register plus
2770 constant, then compute its value into a register before
2771 loop start. This prevents invalid rtx sharing, and should
2772 generate better code. We can use bl->initial_value here
2773 instead of splittable_regs[bl->regno] because this code
2774 is going before the loop start. */
2781 {
2787 }
2790 }
2791 else
2793 }
2794 else
2795 {
2796 #if 0
2797 /* Currently, unreduced giv's can't be split. This is not too much
2798 of a problem since unreduced giv's are not live across loop
2799 iterations anyways. When unrolling a loop completely though,
2800 it makes sense to reduce&split givs when possible, as this will
2801 result in simpler instructions, and will not require that a reg
2802 be live across loop iterations. */
2807 #else
2809 #endif
2810 }
2812 /* Unreduced givs are only updated once by definition. Reduced givs
2813 are updated as many times as their biv is. Mark it so if this is
2814 a splittable register. Don't need to do anything for address givs
2815 where this may not be a register. */
2818 {
2824 }
2826 result++;
2829 {
2836 else
2840 }
2841 }
2844 }
2845 \f
2846 /* Try to prove that the register is dead after the loop exits. Trace every
2847 loop exit looking for an insn that will always be executed, which sets
2848 the register to some value, and appears before the first use of the register
2849 is found. If successful, then return 1, otherwise return 0. */
2851 /* ?? Could be made more intelligent in the handling of jumps, so that
2852 it can search past if statements and other similar structures. */
2854 static int
2857 rtx reg;
2858 {
2864 /* In addition to checking all exits of this loop, we must also check
2865 all exits of inner nested loops that would exit this loop. We don't
2866 have any way to identify those, so we just give up if there are any
2867 such inner loop exits. */
2870 label_count++;
2875 /* HACK: Must also search the loop fall through exit, create a label_ref
2876 here which points to the loop->end, and append the loop_number_exit_labels
2877 list to it. */
2882 {
2883 /* Succeed if find an insn which sets the biv or if reach end of
2884 function. Fail if find an insn that uses the biv, or if come to
2885 a conditional jump. */
2889 {
2892 {
2893 rtx set;
2901 }
2904 {
2908 /* Prevent infinite loop following infinite loops. */
2911 else
2913 }
2916 }
2917 }
2919 /* Success, the register is dead on all loop exits. */
2921 }
2923 /* Try to calculate the final value of the biv, the value it will have at
2924 the end of the loop. If we can do it, return that value. */
2926 rtx
2930 {
2934 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2939 /* The final value for reversed bivs must be calculated differently than
2940 for ordinary bivs. In this case, there is already an insn after the
2941 loop which sets this biv's final value (if necessary), and there are
2942 no other loop exits, so we can return any value. */
2944 {
2950 }
2952 /* Try to calculate the final value as initial value + (number of iterations
2953 * increment). For this to work, increment must be invariant, the only
2954 exit from the loop must be the fall through at the bottom (otherwise
2955 it may not have its final value when the loop exits), and the initial
2956 value of the biv must be invariant. */
2961 {
2965 {
2966 /* Can calculate the loop exit value, emit insns after loop
2967 end to calculate this value into a temporary register in
2968 case it is needed later. */
2980 }
2981 }
2983 /* Check to see if the biv is dead at all loop exits. */
2985 {
2992 }
2995 }
2997 /* Try to calculate the final value of the giv, the value it will have at
2998 the end of the loop. If we can do it, return that value. */
3000 rtx
3004 {
3007 rtx insn;
3009 rtx seq;
3015 /* The final value for givs which depend on reversed bivs must be calculated
3016 differently than for ordinary givs. In this case, there is already an
3017 insn after the loop which sets this giv's final value (if necessary),
3018 and there are no other loop exits, so we can return any value. */
3020 {
3026 }
3028 /* Try to calculate the final value as a function of the biv it depends
3029 upon. The only exit from the loop must be the fall through at the bottom
3030 and the insn that sets the giv must be executed on every iteration
3031 (otherwise the giv may not have its final value when the loop exits). */
3033 /* ??? Can calculate the final giv value by subtracting off the
3034 extra biv increments times the giv's mult_val. The loop must have
3035 only one exit for this to work, but the loop iterations does not need
3036 to be known. */
3041 {
3042 /* ?? It is tempting to use the biv's value here since these insns will
3043 be put after the loop, and hence the biv will have its final value
3044 then. However, this fails if the biv is subsequently eliminated.
3045 Perhaps determine whether biv's are eliminable before trying to
3046 determine whether giv's are replaceable so that we can use the
3047 biv value here if it is not eliminable. */
3049 /* We are emitting code after the end of the loop, so we must make
3050 sure that bl->initial_value is still valid then. It will still
3051 be valid if it is invariant. */
3057 {
3058 /* Can calculate the loop exit value of its biv as
3059 (n_iterations * increment) + initial_value */
3061 /* The loop exit value of the giv is then
3062 (final_biv_value - extra increments) * mult_val + add_val.
3063 The extra increments are any increments to the biv which
3064 occur in the loop after the giv's value is calculated.
3065 We must search from the insn that sets the giv to the end
3066 of the loop to calculate this value. */
3068 /* Put the final biv value in tem. */
3074 tem);
3076 /* Subtract off extra increments as we find them. */
3079 {
3084 {
3088 OPTAB_LIB_WIDEN);
3092 }
3093 }
3095 /* Now calculate the giv's final value. */
3104 }
3105 }
3107 /* Replaceable giv's should never reach here. */
3111 /* Check to see if the biv is dead at all loop exits. */
3113 {
3120 }
3123 }
3125 /* Look back before LOOP->START for the insn that sets REG and return
3126 the equivalent constant if there is a REG_EQUAL note otherwise just
3127 the SET_SRC of REG. */
3129 static rtx
3132 rtx reg;
3133 {
3136 rtx ret;
3140 {
3145 {
3146 /* We found the last insn before the loop that sets the register.
3147 If it sets the entire register, and has a REG_EQUAL note,
3148 then use the value of the REG_EQUAL note. */
3151 {
3154 /* Only use the REG_EQUAL note if it is a constant.
3155 Other things, divide in particular, will cause
3156 problems later if we use them. */
3160 else
3163 /* We cannot do this if it changes between the
3164 assignment and loop start though. */
3167 }
3169 }
3170 }
3172 }
3174 /* Return a simplified rtx for the expression OP - REG.
3176 REG must appear in OP, and OP must be a register or the sum of a register
3177 and a second term.
3179 Thus, the return value must be const0_rtx or the second term.
3181 The caller is responsible for verifying that REG appears in OP and OP has
3182 the proper form. */
3184 static rtx
3187 {
3191 {
3196 }
3197 /* OP does not contain REG as a term. */
3199 }
3201 /* Find and return register term common to both expressions OP0 and
3202 OP1 or NULL_RTX if no such term exists. Each expression must be a
3203 REG or a PLUS of a REG. */
3205 static rtx
3208 {
3211 {
3212 rtx op00;
3213 rtx op01;
3214 rtx op10;
3215 rtx op11;
3219 else
3224 else
3227 /* Find and return common register term if present. */
3232 }
3234 /* No common register term found. */
3236 }
3238 /* Determine the loop iterator and calculate the number of loop
3239 iterations. Returns the exact number of loop iterations if it can
3240 be calculated, otherwise returns zero. */
3242 unsigned HOST_WIDE_INT
3245 {
3251 HOST_WIDE_INT inc;
3257 rtx last_loop_insn;
3258 rtx reg_term;
3272 /* We used to use prev_nonnote_insn here, but that fails because it might
3273 accidentally get the branch for a contained loop if the branch for this
3274 loop was deleted. We can only trust branches immediately before the
3275 loop_end. */
3278 /* ??? We should probably try harder to find the jump insn
3279 at the end of the loop. The following code assumes that
3280 the last loop insn is a jump to the top of the loop. */
3282 {
3287 }
3289 /* If there is a more than a single jump to the top of the loop
3290 we cannot (easily) determine the iteration count. */
3292 {
3297 }
3299 /* If there are multiple conditionalized loop exit tests, they may jump
3300 back to differing CODE_LABELs. */
3302 {
3305 do
3306 {
3308 {
3309 /* There are some kinds of jumps we can't deal with easily. */
3311 {
3313 fprintf
3317 }
3320 covered by the uid_luid array. */
3322 /* Check if we jump back into the loop body. */
3325 {
3327 fprintf
3331 }
3332 }
3333 }
3335 }
3337 /* Find the iteration variable. If the last insn is a conditional
3338 branch, and the insn before tests a register value, make that the
3339 iteration variable. */
3343 {
3348 }
3350 /* ??? Get_condition may switch position of induction variable and
3351 invariant register when it canonicalizes the comparison. */
3358 {
3363 }
3365 /* The only new registers that are created before loop iterations
3366 are givs made from biv increments or registers created by
3367 load_mems. In the latter case, it is possible that try_copy_prop
3368 will propagate a new pseudo into the old iteration register but
3369 this will be marked by having the REG_USERVAR_P bit set. */
3375 /* Determine the initial value of the iteration variable, and the amount
3376 that it is incremented each loop. Use the tables constructed by
3377 the strength reduction pass to calculate these values. */
3379 /* Clear the result values, in case no answer can be found. */
3383 /* The iteration variable can be either a giv or a biv. Check to see
3384 which it is, and compute the variable's initial value, and increment
3385 value if possible. */
3387 /* If this is a new register, can't handle it since we don't have any
3388 reg_iv_type entry for it. */
3390 {
3395 }
3397 /* Reject iteration variables larger than the host wide int size, since they
3398 could result in a number of iterations greater than the range of our
3399 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3402 {
3407 }
3409 {
3414 }
3416 {
3420 /* Grab initial value, only useful if it is a constant. */
3424 {
3429 }
3432 }
3434 {
3437 rtx biv_initial_value;
3443 {
3448 }
3452 /* Increment value is mult_val times the increment value of the biv. */
3456 {
3462 /* The caller assumes that one full increment has occurred at the
3463 first loop test. But that's not true when the biv is incremented
3464 after the giv is set (which is the usual case), e.g.:
3465 i = 6; do {;} while (i++ < 9) .
3466 Therefore, we bias the initial value by subtracting the amount of
3467 the increment that occurs between the giv set and the giv test. */
3469 {
3471 {
3473 {
3479 }
3482 }
3483 }
3484 }
3490 /* Initial value is mult_val times the biv's initial value plus
3491 add_val. Only useful if it is a constant. */
3493 initial_value
3497 }
3498 else
3499 {
3504 }
3512 {
3526 /* Cannot determine loop iterations with this case. */
3543 }
3545 /* If the comparison value is an invariant register, then try to find
3546 its value from the insns before the start of the loop. */
3551 {
3554 /* If we don't get an invariant final value, we are better
3555 off with the original register. */
3558 }
3560 /* Calculate the approximate final value of the induction variable
3561 (on the last successful iteration). The exact final value
3562 depends on the branch operator, and increment sign. It will be
3563 wrong if the iteration variable is not incremented by one each
3564 time through the loop and (comparison_value + off_by_one -
3565 initial_value) % increment != 0.
3566 ??? Note that the final_value may overflow and thus final_larger
3567 will be bogus. A potentially infinite loop will be classified
3568 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3572 /* Save the calculated values describing this loop's bounds, in case
3573 precondition_loop_p will need them later. These values can not be
3574 recalculated inside precondition_loop_p because strength reduction
3575 optimizations may obscure the loop's structure.
3577 These values are only required by precondition_loop_p and insert_bct
3578 whenever the number of iterations cannot be computed at compile time.
3579 Only the difference between final_value and initial_value is
3580 important. Note that final_value is only approximate. */
3589 /* Try to determine the iteration count for loops such
3590 as (for i = init; i < init + const; i++). When running the
3591 loop optimization twice, the first pass often converts simple
3592 loops into this form. */
3595 {
3596 rtx reg1;
3597 rtx reg2;
3598 rtx const2;
3603 else
3606 /* Check for initial_value = reg1, final_value = reg2 + const2,
3607 where reg1 != reg2. */
3609 {
3610 rtx temp;
3612 /* Find what reg1 is equivalent to. Hopefully it will
3613 either be reg2 or reg2 plus a constant. */
3618 else
3619 {
3620 /* Find what reg2 is equivalent to. Hopefully it will
3621 either be reg1 or reg1 plus a constant. Let's ignore
3622 the latter case for now since it is not so common. */
3630 }
3631 }
3633 {
3634 rtx temp;
3636 /* When running the loop optimizer twice, check_dbra_loop
3637 further obfuscates reversible loops of the form:
3638 for (i = init; i < init + const; i++). We often end up with
3639 final_value = 0, initial_value = temp, temp = temp2 - init,
3640 where temp2 = init + const. If the loop has a vtop we
3641 can replace initial_value with const. */
3646 {
3652 }
3653 }
3654 }
3656 /* If have initial_value = reg + const1 and final_value = reg +
3657 const2, then replace initial_value with const1 and final_value
3658 with const2. This should be safe since we are protected by the
3659 initial comparison before entering the loop if we have a vtop.
3660 For example, a + b < a + c is not equivalent to b < c for all a
3661 when using modulo arithmetic.
3663 ??? Without a vtop we could still perform the optimization if we check
3664 the initial and final values carefully. */
3667 {
3670 }
3675 /* For EQ comparison loops, we don't have a valid final value.
3676 Check this now so that we won't leave an invalid value if we
3677 return early for any other reason. */
3682 {
3687 }
3690 {
3691 /* If we have a REG, check to see if REG holds a constant value. */
3692 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3693 clear if it is worthwhile to try to handle such RTL. */
3698 {
3700 {
3705 }
3707 }
3709 }
3712 {
3714 {
3719 }
3721 }
3723 {
3725 {
3730 }
3732 }
3734 {
3735 rtx inc_once;
3744 {
3745 /* The iterator value once through the loop is equal to the
3746 comparision value. Either we have an infinite loop, or
3747 we'll loop twice. */
3751 }
3752 else
3756 loop_info->final_value
3760 else
3761 loop_info->final_value
3764 loop_info->final_equiv_value
3769 }
3771 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3773 final_larger
3778 else
3786 else
3789 /* There are 27 different cases: compare_dir = -1, 0, 1;
3790 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3791 There are 4 normal cases, 4 reverse cases (where the iteration variable
3792 will overflow before the loop exits), 4 infinite loop cases, and 15
3793 immediate exit (0 or 1 iteration depending on loop type) cases.
3794 Only try to optimize the normal cases. */
3796 /* (compare_dir/final_larger/increment_dir)
3797 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3798 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3799 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3800 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3802 /* ?? If the meaning of reverse loops (where the iteration variable
3803 will overflow before the loop exits) is undefined, then could
3804 eliminate all of these special checks, and just always assume
3805 the loops are normal/immediate/infinite. Note that this means
3806 the sign of increment_dir does not have to be known. Also,
3807 since it does not really hurt if immediate exit loops or infinite loops
3808 are optimized, then that case could be ignored also, and hence all
3809 loops can be optimized.
3811 According to ANSI Spec, the reverse loop case result is undefined,
3812 because the action on overflow is undefined.
3814 See also the special test for NE loops below. */
3818 /* Normal case. */
3819 ;
3820 else
3821 {
3825 }
3827 /* Calculate the number of iterations, final_value is only an approximation,
3828 so correct for that. Note that abs_diff and n_iterations are
3829 unsigned, because they can be as large as 2^n - 1. */
3833 {
3836 }
3838 {
3841 }
3842 else
3845 /* Given that iteration_var is going to iterate over its own mode,
3846 not HOST_WIDE_INT, disregard higher bits that might have come
3847 into the picture due to sign extension of initial and final
3848 values. */
3853 /* For NE tests, make sure that the iteration variable won't miss
3854 the final value. If abs_diff mod abs_incr is not zero, then the
3855 iteration variable will overflow before the loop exits, and we
3856 can not calculate the number of iterations. */
3860 /* Note that the number of iterations could be calculated using
3861 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3862 handle potential overflow of the summation. */
3865 }
3867 /* Replace uses of split bivs with their split pseudo register. This is
3868 for original instructions which remain after loop unrolling without
3869 copying. */
3871 static rtx
3874 rtx x;
3875 {
3886 {
3898 #if 0
3899 /* If non-reduced/final-value givs were split, then this would also
3900 have to remap those givs also. */
3901 #endif
3909 }
3913 {
3917 {
3921 }
3922 }
3924 }
3926 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3927 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3928 return 0. COPY_START is where we can start looking for the insns
3929 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3930 insns.
3932 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3933 must dominate LAST_UID.
3935 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3936 may not dominate LAST_UID.
3938 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3939 must dominate LAST_UID. */
3941 int
3946 rtx copy_start;
3947 rtx copy_end;
3948 {
3953 {
3956 /* Could not find FIRST_UID. */
3960 }
3962 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3966 /* FIRST_UID is always executed. */
3971 {
3972 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3973 can not be sure that FIRST_UID dominates LAST_UID. */
3976 /* Could not find LAST_UID, but we reached the end of the loop, so
3977 it must be safe. */
3981 }
3983 /* FIRST_UID is always executed if LAST_UID is executed. */
3985 }
3987 /* This routine is called when the number of iterations for the unrolled
3988 loop is one. The goal is to identify a loop that begins with an
3989 unconditional branch to the loop continuation note (or a label just after).
3990 In this case, the unconditional branch that starts the loop needs to be
3991 deleted so that we execute the single iteration. */
3993 static rtx
3995 rtx loop_start;
3996 rtx loop_cont;
3997 {
4000 /* See if loop start, or the next insn is an unconditional jump. */
4011 /* Examine insn after loop continuation note. Return if not a label. */
4016 /* Return the loop start if the branch label matches the code label. */
4019 else
4021 }