| blob | a795dcbc202c313e2f73765276fd42a36412256b |
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. */
217 /* Try to unroll one loop and split induction variables in the loop.
219 The loop is described by the arguments LOOP and INSN_COUNT.
220 STRENGTH_REDUCTION_P indicates whether information generated in the
221 strength reduction pass is available.
223 This function is intended to be called from within `strength_reduce'
224 in loop.c. */
226 void
231 {
241 rtx insert_before;
248 rtx start_label;
253 rtx safety_label;
254 /* This points to the last real insn in the loop, which should be either
255 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
256 jumps). */
257 rtx last_loop_insn;
261 /* Don't bother unrolling huge loops. Since the minimum factor is
262 two, loops greater than one half of MAX_UNROLLED_INSNS will never
263 be unrolled. */
265 {
269 }
271 /* Determine type of unroll to perform. Depends on the number of iterations
272 and the size of the loop. */
274 /* If there is no strength reduce info, then set
275 loop_info->n_iterations to zero. This can happen if
276 strength_reduce can't find any bivs in the loop. A value of zero
277 indicates that the number of iterations could not be calculated. */
283 {
288 }
290 /* Find and save a pointer to the last nonnote insn in the loop. */
294 /* Calculate how many times to unroll the loop. Indicate whether or
295 not the loop is being completely unrolled. */
298 {
299 /* Handle the case where the loop begins with an unconditional
300 jump to the loop condition. Make sure to delete the jump
301 insn, otherwise the loop body will never execute. */
307 /* If number of iterations is exactly 1, then eliminate the compare and
308 branch at the end of the loop since they will never be taken.
309 Then return, since no other action is needed here. */
311 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
312 don't do anything. */
315 {
316 /* Delete the jump insn. This will delete the barrier also. */
318 }
320 {
321 #ifdef HAVE_cc0
323 #endif
325 #ifdef HAVE_cc0
326 /* The immediately preceding insn may be a compare which must be
327 deleted. */
330 #endif
331 }
333 /* Remove the loop notes since this is no longer a loop. */
344 }
346 /* Avoid overflow in the next expression. */
349 {
352 }
354 {
355 /* Try to factor the number of iterations. Don't bother with the
356 general case, only using 2, 3, 5, and 7 will get 75% of all
357 numbers theoretically, and almost all in practice. */
365 {
368 }
370 /* Start with the larger factors first so that we generally
371 get lots of unrolling. */
377 {
379 {
382 }
383 else
385 }
387 /* If we couldn't find any factors, then unroll as in the normal
388 case. */
390 {
393 }
394 else
396 }
398 /* Default case, calculate number of times to unroll loop based on its
399 size. */
401 {
406 else
408 }
410 /* Now we know how many times to unroll the loop. */
416 {
417 /* Loops of these types can start with jump down to the exit condition
418 in rare circumstances.
420 Consider a pair of nested loops where the inner loop is part
421 of the exit code for the outer loop.
423 In this case jump.c will not duplicate the exit test for the outer
424 loop, so it will start with a jump to the exit code.
426 Then consider if the inner loop turns out to iterate once and
427 only once. We will end up deleting the jumps associated with
428 the inner loop. However, the loop notes are not removed from
429 the instruction stream.
431 And finally assume that we can compute the number of iterations
432 for the outer loop.
434 In this case unroll may want to unroll the outer loop even though
435 it starts with a jump to the outer loop's exit code.
437 We could try to optimize this case, but it hardly seems worth it.
438 Just return without unrolling the loop in such cases. */
445 }
448 {
449 /* Completely unrolling the loop: Delete the compare and branch at
450 the end (the last two instructions). This delete must done at the
451 very end of loop unrolling, to avoid problems with calls to
452 back_branch_in_range_p, which is called by find_splittable_regs.
453 All increments of splittable bivs/givs are changed to load constant
454 instructions. */
458 /* Set insert_before to the instruction immediately after the JUMP_INSN
459 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
460 the loop will be correctly handled by copy_loop_body. */
463 /* Set copy_end to the insn before the jump at the end of the loop. */
467 {
469 #ifdef HAVE_cc0
470 /* The instruction immediately before the JUMP_INSN may be a compare
471 instruction which we do not want to copy. */
474 #endif
475 }
476 else
477 {
478 /* We currently can't unroll a loop if it doesn't end with a
479 JUMP_INSN. There would need to be a mechanism that recognizes
480 this case, and then inserts a jump after each loop body, which
481 jumps to after the last loop body. */
486 }
487 }
489 {
490 /* Partially unrolling the loop: The compare and branch at the end
491 (the last two instructions) must remain. Don't copy the compare
492 and branch instructions at the end of the loop. Insert the unrolled
493 code immediately before the compare/branch at the end so that the
494 code will fall through to them as before. */
498 /* Set insert_before to the jump insn at the end of the loop.
499 Set copy_end to before the jump insn at the end of the loop. */
501 {
504 }
506 {
508 #ifdef HAVE_cc0
509 /* The instruction immediately before the JUMP_INSN may be a compare
510 instruction which we do not want to copy or delete. */
513 #endif
515 }
516 else
517 {
518 /* We currently can't unroll a loop if it doesn't end with a
519 JUMP_INSN. There would need to be a mechanism that recognizes
520 this case, and then inserts a jump after each loop body, which
521 jumps to after the last loop body. */
526 }
527 }
528 else
529 {
530 /* Normal case: Must copy the compare and branch instructions at the
531 end of the loop. */
534 {
535 /* Loop ends with an unconditional jump and a barrier.
536 Handle this like above, don't copy jump and barrier.
537 This is not strictly necessary, but doing so prevents generating
538 unconditional jumps to an immediately following label.
540 This will be corrected below if the target of this jump is
541 not the start_label. */
545 }
547 {
548 /* Set insert_before to immediately after the JUMP_INSN, so that
549 NOTEs at the end of the loop will be correctly handled by
550 copy_loop_body. */
553 }
554 else
555 {
556 /* We currently can't unroll a loop if it doesn't end with a
557 JUMP_INSN. There would need to be a mechanism that recognizes
558 this case, and then inserts a jump after each loop body, which
559 jumps to after the last loop body. */
564 }
566 /* If copying exit test branches because they can not be eliminated,
567 then must convert the fall through case of the branch to a jump past
568 the end of the loop. Create a label to emit after the loop and save
569 it for later use. Do not use the label after the loop, if any, since
570 it might be used by insns outside the loop, or there might be insns
571 added before it later by final_[bg]iv_value which must be after
572 the real exit label. */
580 {
581 /* The loop starts with a jump down to the exit condition test.
582 Start copying the loop after the barrier following this
583 jump insn. */
586 /* Splitting induction variables doesn't work when the loop is
587 entered via a jump to the bottom, because then we end up doing
588 a comparison against a new register for a split variable, but
589 we did not execute the set insn for the new register because
590 it was skipped over. */
595 }
596 else
598 }
600 /* This should always be the first label in the loop. */
602 /* There may be a line number note and/or a loop continue note here. */
606 {
607 /* This can happen as a result of jump threading. If the first insns in
608 the loop test the same condition as the loop's backward jump, or the
609 opposite condition, then the backward jump will be modified to point
610 to elsewhere, and the loop's start label is deleted.
612 This case currently can not be handled by the loop unrolling code. */
618 }
620 {
621 /* The jump optimization pass must have combined the original start label
622 with a named label for a goto. We can't unroll this case because
623 jumps which go to the named label must be handled differently than
624 jumps to the loop start, and it is impossible to differentiate them
625 in this case. */
630 }
636 {
637 /* In this case, we must copy the jump and barrier, because they will
638 not be converted to jumps to an immediately following label. */
642 }
647 {
648 /* ??? The loop ends with a conditional branch that does not branch back
649 to the loop start label. In this case, we must emit an unconditional
650 branch to the loop exit after emitting the final branch.
651 copy_loop_body does not have support for this currently, so we
652 give up. It doesn't seem worthwhile to unroll anyways since
653 unrolling would increase the number of branch instructions
654 executed. */
659 }
661 /* Allocate a translation table for the labels and insn numbers.
662 They will be filled in as we copy the insns in the loop. */
667 /* Various paths through the unroll code may reach the "egress" label
668 without initializing fields within the map structure.
670 To be safe, we use xcalloc to zero the memory. */
673 /* Allocate the label map. */
676 {
679 }
681 /* Search the loop and mark all local labels, i.e. the ones which have to
682 be distinct labels when copied. For all labels which might be
683 non-local, set their label_map entries to point to themselves.
684 If they happen to be local their label_map entries will be overwritten
685 before the loop body is copied. The label_map entries for local labels
686 will be set to a different value each time the loop body is copied. */
689 {
690 rtx note;
695 {
702 {
706 rtx label;
709 {
712 }
713 }
714 }
718 }
720 /* Allocate space for the insn map. */
724 /* Set this to zero, to indicate that we are doing loop unrolling,
725 not function inlining. */
728 /* The register and constant maps depend on the number of registers
729 present, so the final maps can't be created until after
730 find_splittable_regs is called. However, they are needed for
731 preconditioning, so we create temporary maps when preconditioning
732 is performed. */
734 /* The preconditioning code may allocate two new pseudo registers. */
737 /* local_regno is only valid for regnos < max_local_regnum. */
740 /* Allocate and zero out the splittable_regs and addr_combined_regs
741 arrays. These must be zeroed here because they will be used if
742 loop preconditioning is performed, and must be zero for that case.
744 It is safe to do this here, since the extra registers created by the
745 preconditioning code and find_splittable_regs will never be used
746 to access the splittable_regs[] and addr_combined_regs[] arrays. */
750 addr_combined_regs
754 /* Mark all local registers, i.e. the ones which are referenced only
755 inside the loop. */
757 {
761 /* If a register is used in the jump insn, we must not duplicate it
762 since it will also be used outside the loop. */
764 copy_end_luid--;
766 /* If we have a target that uses cc0, then we also must not duplicate
767 the insn that sets cc0 before the jump insn, if one is present. */
768 #ifdef HAVE_cc0
771 copy_end_luid--;
772 #endif
774 /* If copy_start points to the NOTE that starts the loop, then we must
775 use the next luid, because invariant pseudo-regs moved out of the loop
776 have their lifetimes modified to start here, but they are not safe
777 to duplicate. */
779 copy_start_luid++;
781 /* If a pseudo's lifetime is entirely contained within this loop, then we
782 can use a different pseudo in each unrolled copy of the loop. This
783 results in better code. */
784 /* We must limit the generic test to max_reg_before_loop, because only
785 these pseudo registers have valid regno_first_uid info. */
791 {
792 /* However, we must also check for loop-carried dependencies.
793 If the value the pseudo has at the end of iteration X is
794 used by iteration X+1, then we can not use a different pseudo
795 for each unrolled copy of the loop. */
796 /* A pseudo is safe if regno_first_uid is a set, and this
797 set dominates all instructions from regno_first_uid to
798 regno_last_uid. */
799 /* ??? This check is simplistic. We would get better code if
800 this check was more sophisticated. */
806 {
809 else
811 r);
812 }
813 }
814 }
816 /* If this loop requires exit tests when unrolled, check to see if we
817 can precondition the loop so as to make the exit tests unnecessary.
818 Just like variable splitting, this is not safe if the loop is entered
819 via a jump to the bottom. Also, can not do this if no strength
820 reduce info, because precondition_loop_p uses this info. */
822 /* Must copy the loop body for preconditioning before the following
823 find_splittable_regs call since that will emit insns which need to
824 be after the preconditioned loop copies, but immediately before the
825 unrolled loop copies. */
827 /* Also, it is not safe to split induction variables for the preconditioned
828 copies of the loop body. If we split induction variables, then the code
829 assumes that each induction variable can be represented as a function
830 of its initial value and the loop iteration number. This is not true
831 in this case, because the last preconditioned copy of the loop body
832 could be any iteration from the first up to the `unroll_number-1'th,
833 depending on the initial value of the iteration variable. Therefore
834 we can not split induction variables here, because we can not calculate
835 their value. Hence, this code must occur before find_splittable_regs
836 is called. */
839 {
846 {
847 rtx diff;
862 /* Limit loop unrolling to 4, since this will make 7 copies of
863 the loop body. */
867 /* Save the absolute value of the increment, and also whether or
868 not it is negative. */
872 {
875 }
879 /* Final value may have form of (PLUS val1 const1_rtx). We need
880 to convert it into general operand, so compute the real value. */
883 {
888 }
892 /* Calculate the difference between the final and initial values.
893 Final value may be a (plus (reg x) (const_int 1)) rtx.
894 Let the following cse pass simplify this if initial value is
895 a constant.
897 We must copy the final and initial values here to avoid
898 improperly shared rtl.
900 We have to deal with for (i = 0; --i < 6;) type loops.
901 For such loops the real final value is the first time the
902 loop variable overflows, so the diff we calculate is the
903 distance from the overflow value. This is 0 or ~0 for
904 unsigned loops depending on the direction, or INT_MAX,
905 INT_MAX+1 for signed loops. We really do not need the
906 exact value, since we are only interested in the diff
907 modulo the increment, and the increment is a power of 2,
908 so we can pretend that the overflow value is 0/~0. */
913 OPTAB_LIB_WIDEN);
914 else
918 /* Now calculate (diff % (unroll * abs (increment))) by using an
919 and instruction. */
924 /* Now emit a sequence of branches to jump to the proper precond
925 loop entry point. */
931 /* Check for the case where the initial value is greater than or
932 equal to the final value. In that case, we want to execute
933 exactly one loop iteration. The code below will fail for this
934 case. This check does not apply if the loop has a NE
935 comparison at the end. */
938 {
939 rtx incremented_initval;
941 initial_value,
942 increment,
944 OPTAB_LIB_WIDEN);
949 TAKEN);
952 }
954 /* Assuming the unroll_number is 4, and the increment is 2, then
955 for a negative increment: for a positive increment:
956 diff = 0,1 precond 0 diff = 0,7 precond 0
957 diff = 2,3 precond 3 diff = 1,2 precond 1
958 diff = 4,5 precond 2 diff = 3,4 precond 2
959 diff = 6,7 precond 1 diff = 5,6 precond 3 */
961 /* We only need to emit (unroll_number - 1) branches here, the
962 last case just falls through to the following code. */
964 /* ??? This would give better code if we emitted a tree of branches
965 instead of the current linear list of branches. */
968 {
972 /* For negative increments, must invert the constant compared
973 against, except when comparing against zero. */
975 {
978 }
980 {
983 }
984 else
985 {
988 }
996 }
998 /* If the increment is greater than one, then we need another branch,
999 to handle other cases equivalent to 0. */
1001 /* ??? This should be merged into the code above somehow to help
1002 simplify the code here, and reduce the number of branches emitted.
1003 For the negative increment case, the branch here could easily
1004 be merged with the `0' case branch above. For the positive
1005 increment case, it is not clear how this can be simplified. */
1008 {
1013 {
1016 }
1017 else
1018 {
1021 }
1027 }
1033 /* Only the last copy of the loop body here needs the exit
1034 test, so set copy_end to exclude the compare/branch here,
1035 and then reset it inside the loop when get to the last
1036 copy. */
1041 {
1043 #ifdef HAVE_cc0
1044 /* The immediately preceding insn may be a compare which
1045 we do not want to copy. */
1048 #endif
1049 }
1050 else
1054 {
1070 {
1075 }
1076 /* The last copy needs the compare/branch insns at the end,
1077 so reset copy_end here if the loop ends with a conditional
1078 branch. */
1081 {
1084 else
1086 }
1088 /* None of the copies are the `last_iteration', so just
1089 pass zero for that parameter. */
1093 }
1097 {
1100 }
1101 else
1102 {
1104 #ifdef HAVE_cc0
1105 /* The instruction immediately before the JUMP_INSN may
1106 be a compare instruction which we do not want to copy
1107 or delete. */
1110 #endif
1112 }
1114 /* Set unroll type to MODULO now. */
1118 /* Clean up. */
1120 }
1121 }
1123 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1124 the loop unless all loops are being unrolled. */
1126 {
1131 }
1133 /* At this point, we are guaranteed to unroll the loop. */
1135 /* Keep track of the unroll factor for the loop. */
1138 /* For each biv and giv, determine whether it can be safely split into
1139 a different variable for each unrolled copy of the loop body.
1140 We precalculate and save this info here, since computing it is
1141 expensive.
1143 Do this before deleting any instructions from the loop, so that
1144 back_branch_in_range_p will work correctly. */
1148 else
1151 /* find_splittable_regs may have created some new registers, so must
1152 reallocate the reg_map with the new larger size, and must realloc
1153 the constant maps also. */
1166 /* Search the list of bivs and givs to find ones which need to be remapped
1167 when split, and set their reg_map entry appropriately. */
1170 {
1173 #if 0
1174 /* Currently, non-reduced/final-value givs are never split. */
1178 #endif
1179 }
1181 /* Use our current register alignment and pointer flags. */
1185 /* If the loop is being partially unrolled, and the iteration variables
1186 are being split, and are being renamed for the split, then must fix up
1187 the compare/jump instruction at the end of the loop to refer to the new
1188 registers. This compare isn't copied, so the registers used in it
1189 will never be replaced if it isn't done here. */
1192 {
1196 }
1198 /* For unroll_number times, make a copy of each instruction
1199 between copy_start and copy_end, and insert these new instructions
1200 before the end of the loop. */
1203 {
1215 {
1219 }
1221 /* If loop starts with a branch to the test, then fix it so that
1222 it points to the test of the first unrolled copy of the loop. */
1224 {
1229 CODE_LABEL_NUMBER
1233 /* Set the jump label so that it can be used by later loop unrolling
1234 passes. */
1237 }
1242 }
1244 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1245 insn to be deleted. This prevents any runaway delete_insn call from
1246 more insns that it should, as it always stops at a CODE_LABEL. */
1248 /* Delete the compare and branch at the end of the loop if completely
1249 unrolling the loop. Deleting the backward branch at the end also
1250 deletes the code label at the start of the loop. This is done at
1251 the very end to avoid problems with back_branch_in_range_p. */
1255 else
1258 /* Delete all of the original loop instructions. Don't delete the
1259 LOOP_BEG note, or the first code label in the loop. */
1263 {
1264 /* ??? Don't delete named code labels. They will be deleted when the
1265 jump that references them is deleted. Otherwise, we end up deleting
1266 them twice, which causes them to completely disappear instead of turn
1267 into NOTE_INSN_DELETED_LABEL notes. This in turn causes aborts in
1268 dwarfout.c/dwarf2out.c. We could perhaps fix the dwarf*out.c files
1269 to handle deleted labels instead. Or perhaps fix DECL_RTL of the
1270 associated LABEL_DECL to point to one of the new label instances. */
1271 /* ??? Likewise, we can't delete a NOTE_INSN_DELETED_LABEL note. */
1277 else
1279 }
1281 /* Can now delete the 'safety' label emitted to protect us from runaway
1282 delete_related_insns calls. */
1287 /* If exit_label exists, emit it after the loop. Doing the emit here
1288 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1289 This is needed so that mostly_true_jump in reorg.c will treat jumps
1290 to this loop end label correctly, i.e. predict that they are usually
1291 not taken. */
1295 egress:
1297 {
1298 /* Remove the loop notes since this is no longer a loop. */
1307 }
1312 {
1315 }
1324 }
1325 \f
1326 /* Return true if the loop can be safely, and profitably, preconditioned
1327 so that the unrolled copies of the loop body don't need exit tests.
1329 This only works if final_value, initial_value and increment can be
1330 determined, and if increment is a constant power of 2.
1331 If increment is not a power of 2, then the preconditioning modulo
1332 operation would require a real modulo instead of a boolean AND, and this
1333 is not considered `profitable'. */
1335 /* ??? If the loop is known to be executed very many times, or the machine
1336 has a very cheap divide instruction, then preconditioning is a win even
1337 when the increment is not a power of 2. Use RTX_COST to compute
1338 whether divide is cheap.
1339 ??? A divide by constant doesn't actually need a divide, look at
1340 expand_divmod. The reduced cost of this optimized modulo is not
1341 reflected in RTX_COST. */
1343 int
1348 {
1353 {
1355 {
1359 }
1360 else
1361 {
1365 }
1369 {
1371 loop_dump_stream);
1375 }
1377 }
1380 {
1385 }
1387 {
1392 }
1394 {
1399 }
1401 {
1406 }
1409 {
1414 }
1416 /* Unsigned_compare and compare_dir can be ignored here, since they do
1417 not matter for preconditioning. */
1420 {
1425 }
1427 /* Must ensure that final_value is invariant, so call
1428 loop_invariant_p to check. Before doing so, must check regno
1429 against max_reg_before_loop to make sure that the register is in
1430 the range covered by loop_invariant_p. If it isn't, then it is
1431 most likely a biv/giv which by definition are not invariant. */
1437 {
1442 }
1444 /* Fail for floating point values, since the caller of this function
1445 does not have code to deal with them. */
1448 {
1453 }
1455 /* Fail if loop_info->iteration_var is not live before loop_start,
1456 since we need to test its value in the preconditioning code. */
1460 {
1465 }
1467 /* Note that loop_iterations biases the initial value for GIV iterators
1468 such as "while (i-- > 0)" so that we can calculate the number of
1469 iterations just like for BIV iterators.
1471 Also note that the absolute values of initial_value and
1472 final_value are unimportant as only their difference is used for
1473 calculating the number of loop iterations. */
1478 /* Decide what mode to do these calculations in. Choose the larger
1479 of final_value's mode and initial_value's mode, or a full-word if
1480 both are constants. */
1483 {
1487 }
1493 /* Success! */
1497 }
1499 /* All pseudo-registers must be mapped to themselves. Two hard registers
1500 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1501 REGNUM, to avoid function-inlining specific conversions of these
1502 registers. All other hard regs can not be mapped because they may be
1503 used with different
1504 modes. */
1506 static void
1510 {
1515 /* Just clear the rest of the entries. */
1523 }
1524 \f
1525 /* Strength-reduction will often emit code for optimized biv/givs which
1526 calculates their value in a temporary register, and then copies the result
1527 to the iv. This procedure reconstructs the pattern computing the iv;
1528 verifying that all operands are of the proper form.
1530 PATTERN must be the result of single_set.
1531 The return value is the amount that the giv is incremented by. */
1533 static rtx
1537 {
1538 rtx increment;
1542 retry:
1543 /* Verify that we have an increment insn here. First check for a plus
1544 as the set source. */
1546 {
1547 /* SR sometimes computes the new giv value in a temp, then copies it
1548 to the new_reg. */
1554 /* The last insn emitted is not needed, so delete it to avoid confusing
1555 the second cse pass. This insn sets the giv unnecessarily. */
1557 }
1559 /* Verify that we have a constant as the second operand of the plus. */
1562 {
1563 /* SR sometimes puts the constant in a register, especially if it is
1564 too big to be an add immed operand. */
1567 /* SR may have used LO_SUM to compute the constant if it is too large
1568 for a load immed operand. In this case, the constant is in operand
1569 one of the LO_SUM rtx. */
1573 /* Some ports store large constants in memory and add a REG_EQUAL
1574 note to the store insn. */
1576 {
1580 }
1585 {
1586 /* The rs6000 port loads some constants with IOR.
1587 The alpha port loads some constants with ASHIFT and PLUS. */
1593 /* Don't need the last insn anymore. */
1604 else
1606 }
1611 /* The insn loading the constant into a register is no longer needed,
1612 so delete it. */
1614 }
1618 else
1621 /* Check that the source register is the same as the register we expected
1622 to see as the source. If not, something is seriously wrong. */
1625 {
1626 /* Some machines (e.g. the romp), may emit two add instructions for
1627 certain constants, so lets try looking for another add immediately
1628 before this one if we have only seen one add insn so far. */
1631 {
1632 tries++;
1640 }
1643 }
1646 }
1648 /* Copy REG_NOTES, except for insn references, because not all insn_map
1649 entries are valid yet. We do need to copy registers now though, because
1650 the reg_map entries can change during copying. */
1652 static rtx
1654 rtx notes;
1656 {
1657 rtx copy;
1668 /* Don't substitute for these yet. */
1670 else
1676 }
1678 /* Fixup insn references in copied REG_NOTES. */
1680 static void
1684 {
1686 {
1690 {
1691 /* Sometimes, we have a REG_WAS_0 note that points to a
1692 deleted instruction. In that case, we can just delete the
1693 note. */
1695 {
1698 }
1699 else
1700 {
1703 /* If we failed to remap the note, something is awry.
1704 Allow REG_LABEL as it may reference label outside
1705 the unrolled loop. */
1707 {
1710 }
1711 else
1713 }
1714 }
1717 }
1718 }
1720 /* Copy each instruction in the loop, substituting from map as appropriate.
1721 This is very similar to a loop in expand_inline_function. */
1723 static void
1726 copy_notes_from)
1730 rtx exit_label;
1734 {
1739 #ifdef HAVE_cc0
1741 #endif
1745 /* If this isn't the last iteration, then map any references to the
1746 start_label to final_label. Final label will then be emitted immediately
1747 after the end of this loop body if it was ever used.
1749 If this is the last iteration, then map references to the start_label
1750 to itself. */
1752 {
1755 }
1756 else
1762 do
1763 {
1769 {
1775 /* Check to see if this is a giv that has been combined with
1776 some split address givs. (Combined in the sense that
1777 `combine_givs' in loop.c has put two givs in the same register.)
1778 In this case, we must search all givs based on the same biv to
1779 find the address givs. Then split the address givs.
1780 Do this before splitting the giv, since that may map the
1781 SET_DEST to a new register. */
1786 {
1794 /* Although the giv_inc amount is not needed here, we must call
1795 calculate_giv_inc here since it might try to delete the
1796 last insn emitted. If we wait until later to call it,
1797 we might accidentally delete insns generated immediately
1798 below by emit_unrolled_add. */
1802 /* Now find all address giv's that were combined with this
1803 giv 'v'. */
1806 {
1809 /* If this DEST_ADDR giv was not split, then ignore it. */
1813 /* Scale this_giv_inc if the multiplicative factors of
1814 the two givs are different. */
1824 {
1825 /* Must emit an insn to increment the split address
1826 giv. Add in the const_adjust field in case there
1827 was a constant eliminated from the address. */
1830 /* tv->dest_reg will be either a bare register,
1831 or else a register plus a constant. */
1834 else
1837 /* Check for shared address givs, and avoid
1838 incrementing the shared pseudo reg more than
1839 once. */
1841 {
1842 /* tv->dest_reg may actually be a (PLUS (REG)
1843 (CONST)) here, so we must call plus_constant
1844 to add the const_adjust amount before calling
1845 emit_unrolled_add below. */
1850 {
1851 /* The constant could be too large for an add
1852 immediate, so can't directly emit an insn
1853 here. */
1856 }
1857 }
1859 /* Reset the giv to be just the register again, in case
1860 it is used after the set we have just emitted.
1861 We must subtract the const_adjust factor added in
1862 above. */
1866 }
1867 }
1868 }
1870 /* If this is a setting of a splittable variable, then determine
1871 how to split the variable, create a new set based on this split,
1872 and set up the reg_map so that later uses of the variable will
1873 use the new split variable. */
1880 {
1888 /* Compute the increment value for the giv, if it wasn't
1889 already computed above. */
1896 {
1897 /* Completely unrolling the loop. Set the induction
1898 variable to a known constant value. */
1900 /* The value in splittable_regs may be an invariant
1901 value, so we must use plus_constant here. */
1907 {
1910 }
1911 else
1912 {
1913 /* The splittable_regs value must be a REG or a
1914 CONST_INT, so put the entire value in the giv_src_reg
1915 variable. */
1918 }
1919 }
1920 else
1921 {
1922 /* Partially unrolling loop. Create a new pseudo
1923 register for the iteration variable, and set it to
1924 be a constant plus the original register. Except
1925 on the last iteration, when the result has to
1926 go back into the original iteration var register. */
1928 /* Handle bivs which must be mapped to a new register
1929 when split. This happens for bivs which need their
1930 final value set before loop entry. The new register
1931 for the biv was stored in the biv's first struct
1932 induction entry by find_splittable_regs. */
1936 {
1939 }
1941 #if 0
1942 /* If non-reduced/final-value givs were split, then
1943 this would have to remap those givs also. See
1944 find_splittable_regs. */
1945 #endif
1949 giv_inc,
1953 /* Now split the induction variable by changing the dest
1954 of this insn to a new register, and setting its
1955 reg_map entry to point to this new register.
1957 If this is the last iteration, and this is the last insn
1958 that will update the iv, then reuse the original dest,
1959 to ensure that the iv will have the proper value when
1960 the loop exits or repeats.
1962 Using splittable_regs_updates here like this is safe,
1963 because it can only be greater than one if all
1964 instructions modifying the iv are always executed in
1965 order. */
1969 {
1974 giv_inc == const0_rtx
1975 ? giv_src_reg
1979 }
1980 else
1982 }
1984 /* The constant being added could be too large for an add
1985 immediate, so can't directly emit an insn here. */
1989 }
1990 else
1991 {
1994 }
1998 #ifdef HAVE_cc0
1999 /* If this insn is setting CC0, it may need to look at
2000 the insn that uses CC0 to see what type of insn it is.
2001 In that case, the call to recog via validate_change will
2002 fail. So don't substitute constants here. Instead,
2003 do it when we emit the following insn.
2005 For example, see the pyr.md file. That machine has signed and
2006 unsigned compares. The compare patterns must check the
2007 following branch insn to see which what kind of compare to
2008 emit.
2010 If the previous insn set CC0, substitute constants on it as
2011 well. */
2014 else
2015 {
2020 }
2021 #else
2023 #endif
2025 /* Make split induction variable constants `permanent' since we
2026 know there are no backward branches across iteration variable
2027 settings which would invalidate this. */
2029 {
2036 }
2046 {
2048 CODE_LABEL_NUMBER
2051 }
2054 {
2056 /* This is a branch to the beginning of the loop; this is the
2057 last insn being copied; and this is not the last iteration.
2058 In this case, we want to change the original fall through
2059 case to be a branch past the end of the loop, and the
2060 original jump label case to fall_through. */
2063 {
2064 rtx jmp;
2066 /* Can't do it by reversing the jump (probably because we
2067 couldn't reverse the conditions), so emit a new
2068 jump_insn after COPY, and redirect the jump around
2069 that. */
2078 }
2079 }
2081 #ifdef HAVE_cc0
2085 #endif
2088 /* Set the jump label of COPY correctly to avoid problems with
2089 later passes of unroll_loop, if INSN had jump label set. */
2091 {
2094 /* Can't use the label_map for every insn, since this may be
2095 the backward branch, and hence the label was not mapped. */
2097 {
2102 {
2105 else
2107 }
2108 }
2112 else
2113 {
2114 /* An unrecognizable jump insn, probably the entry jump
2115 for a switch statement. This label must have been mapped,
2116 so just use the label_map to get the new jump label. */
2120 }
2122 /* If this is a non-local jump, then must increase the label
2123 use count so that the label will not be deleted when the
2124 original jump is deleted. */
2126 }
2129 {
2137 }
2139 /* If this used to be a conditional jump insn but whose branch
2140 direction is now known, we must do something special. */
2142 {
2143 #ifdef HAVE_cc0
2144 /* If the previous insn set cc0 for us, delete it. */
2147 #endif
2149 /* If this is now a no-op, delete it. */
2151 {
2154 }
2155 else
2156 /* Otherwise, this is unconditional jump so we must put a
2157 BARRIER after it. We could do some dead code elimination
2158 here, but jump.c will do it just as well. */
2160 }
2170 /* Because the USAGE information potentially contains objects other
2171 than hard registers, we need to copy it. */
2176 #ifdef HAVE_cc0
2180 #endif
2183 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2189 /* If this is the loop start label, then we don't need to emit a
2190 copy of this label since no one will use it. */
2193 {
2197 }
2205 /* VTOP and CONT notes are valid only before the loop exit test.
2206 If placed anywhere else, loop may generate bad code. */
2207 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2208 the associated rtl. We do not want to share the structure in
2209 this new block. */
2219 else
2225 }
2228 }
2231 /* Now finish coping the REG_NOTES. */
2233 do
2234 {
2240 }
2243 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2244 each of these notes here, since there may be some important ones, such as
2245 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2246 iteration, because the original notes won't be deleted.
2248 We can't use insert_before here, because when from preconditioning,
2249 insert_before points before the loop. We can't use copy_end, because
2250 there may be insns already inserted after it (which we don't want to
2251 copy) when not from preconditioning code. */
2254 {
2256 {
2257 /* VTOP notes are valid only before the loop exit test.
2258 If placed anywhere else, loop may generate bad code.
2259 There is no need to test for NOTE_INSN_LOOP_CONT notes
2260 here, since COPY_NOTES_FROM will be at most one or two (for cc0)
2261 instructions before the last insn in the loop, and if the
2262 end test is that short, there will be a VTOP note between
2263 the CONT note and the test. */
2269 }
2270 }
2278 }
2279 \f
2280 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2281 emitted. This will correctly handle the case where the increment value
2282 won't fit in the immediate field of a PLUS insns. */
2284 void
2287 {
2288 rtx result;
2295 }
2296 \f
2297 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2298 is a backward branch in that range that branches to somewhere between
2299 LOOP->START and INSN. Returns 0 otherwise. */
2301 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2302 In practice, this is not a problem, because this function is seldom called,
2303 and uses a negligible amount of CPU time on average. */
2305 int
2308 rtx insn;
2309 {
2315 /* Stop before we get to the backward branch at the end of the loop. */
2320 /* Check in case insn has been deleted, search forward for first non
2321 deleted insn following it. */
2325 /* Check for the case where insn is the last insn in the loop. Deal
2326 with the case where INSN was a deleted loop test insn, in which case
2327 it will now be the NOTE_LOOP_END. */
2332 {
2334 {
2337 /* Search from loop_start to insn, to see if one of them is
2338 the target_insn. We can't use INSN_LUID comparisons here,
2339 since insn may not have an LUID entry. */
2343 }
2344 }
2347 }
2349 /* Try to generate the simplest rtx for the expression
2350 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2351 value of giv's. */
2353 static rtx
2357 {
2359 rtx result;
2361 /* The modes must all be the same. This should always be true. For now,
2362 check to make sure. */
2368 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2369 will be a constant. */
2371 {
2375 }
2381 /* Again, put the constant second. */
2383 {
2387 }
2394 }
2396 /* Searches the list of induction struct's for the biv BL, to try to calculate
2397 the total increment value for one iteration of the loop as a constant.
2399 Returns the increment value as an rtx, simplified as much as possible,
2400 if it can be calculated. Otherwise, returns 0. */
2402 rtx
2405 {
2407 rtx result;
2409 /* For increment, must check every instruction that sets it. Each
2410 instruction must be executed only once each time through the loop.
2411 To verify this, we check that the insn is always executed, and that
2412 there are no backward branches after the insn that branch to before it.
2413 Also, the insn must have a mult_val of one (to make sure it really is
2414 an increment). */
2418 {
2422 else
2424 }
2427 }
2429 /* For each biv and giv, determine whether it can be safely split into
2430 a different variable for each unrolled copy of the loop body. If it
2431 is safe to split, then indicate that by saving some useful info
2432 in the splittable_regs array.
2434 If the loop is being completely unrolled, then splittable_regs will hold
2435 the current value of the induction variable while the loop is unrolled.
2436 It must be set to the initial value of the induction variable here.
2437 Otherwise, splittable_regs will hold the difference between the current
2438 value of the induction variable and the value the induction variable had
2439 at the top of the loop. It must be set to the value 0 here.
2441 Returns the total number of instructions that set registers that are
2442 splittable. */
2444 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2445 constant values are unnecessary, since we can easily calculate increment
2446 values in this case even if nothing is constant. The increment value
2447 should not involve a multiply however. */
2449 /* ?? Even if the biv/giv increment values aren't constant, it may still
2450 be beneficial to split the variable if the loop is only unrolled a few
2451 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2453 static int
2458 {
2463 rtx biv_final_value;
2468 {
2469 /* Biv_total_increment must return a constant value,
2470 otherwise we can not calculate the split values. */
2476 /* The loop must be unrolled completely, or else have a known number
2477 of iterations and only one exit, or else the biv must be dead
2478 outside the loop, or else the final value must be known. Otherwise,
2479 it is unsafe to split the biv since it may not have the proper
2480 value on loop exit. */
2482 /* loop_number_exit_count is non-zero if the loop has an exit other than
2483 a fall through at the end. */
2498 /* If any of the insns setting the BIV don't do so with a simple
2499 PLUS, we don't know how to split it. */
2507 /* If final value is non-zero, then must emit an instruction which sets
2508 the value of the biv to the proper value. This is done after
2509 handling all of the givs, since some of them may need to use the
2510 biv's value in their initialization code. */
2512 /* This biv is splittable. If completely unrolling the loop, save
2513 the biv's initial value. Otherwise, save the constant zero. */
2516 {
2518 {
2519 /* If the initial value of the biv is itself (i.e. it is too
2520 complicated for strength_reduce to compute), or is a hard
2521 register, or it isn't invariant, then we must create a new
2522 pseudo reg to hold the initial value of the biv. */
2528 {
2541 }
2542 else
2544 }
2545 else
2548 /* Save the number of instructions that modify the biv, so that
2549 we can treat the last one specially. */
2557 }
2559 /* Check every giv that depends on this biv to see whether it is
2560 splittable also. Even if the biv isn't splittable, givs which
2561 depend on it may be splittable if the biv is live outside the
2562 loop, and the givs aren't. */
2565 unroll_number);
2567 /* If final value is non-zero, then must emit an instruction which sets
2568 the value of the biv to the proper value. This is done after
2569 handling all of the givs, since some of them may need to use the
2570 biv's value in their initialization code. */
2572 {
2573 /* If the loop has multiple exits, emit the insns before the
2574 loop to ensure that it will always be executed no matter
2575 how the loop exits. Otherwise emit the insn after the loop,
2576 since this is slightly more efficient. */
2579 biv_final_value));
2580 else
2581 {
2582 /* Create a new register to hold the value of the biv, and then
2583 set the biv to its final value before the loop start. The biv
2584 is set to its final value before loop start to ensure that
2585 this insn will always be executed, no matter how the loop
2586 exits. */
2592 biv_final_value));
2598 /* Set up the mapping from the original biv register to the new
2599 register. */
2601 }
2602 }
2603 }
2605 }
2607 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2608 for the instruction that is using it. Do not make any changes to that
2609 instruction. */
2611 static int
2614 rtx giv_inc;
2616 {
2622 /* First check to see if either address would fail. Handle the fact
2623 that we have may have a match_dup. */
2628 /* Now put things back the way they were before. This should always
2629 succeed. */
2634 }
2636 /* For every giv based on the biv BL, check to determine whether it is
2637 splittable. This is a subroutine to find_splittable_regs ().
2639 Return the number of instructions that set splittable registers. */
2641 static int
2646 rtx increment;
2648 {
2651 rtx final_value;
2652 rtx tem;
2655 /* Scan the list of givs, and set the same_insn field when there are
2656 multiple identical givs in the same insn. */
2664 {
2667 /* Only split the giv if it has already been reduced, or if the loop is
2668 being completely unrolled. */
2672 /* The giv can be split if the insn that sets the giv is executed once
2673 and only once on every iteration of the loop. */
2674 /* An address giv can always be split. v->insn is just a use not a set,
2675 and hence it does not matter whether it is always executed. All that
2676 matters is that all the biv increments are always executed, and we
2677 won't reach here if they aren't. */
2683 /* The giv increment value must be a constant. */
2689 /* The loop must be unrolled completely, or else have a known number of
2690 iterations and only one exit, or else the giv must be dead outside
2691 the loop, or else the final value of the giv must be known.
2692 Otherwise, it is not safe to split the giv since it may not have the
2693 proper value on loop exit. */
2695 /* The used outside loop test will fail for DEST_ADDR givs. They are
2696 never used outside the loop anyways, so it is always safe to split a
2697 DEST_ADDR giv. */
2703 /* The next part is true if the pseudo is used outside the loop.
2704 We assume that this is true for any pseudo created after loop
2705 starts, because we don't have a reg_n_info entry for them. */
2708 /* Check for the case where the pseudo is set by a shift/add
2709 sequence, in which case the first insn setting the pseudo
2710 is the first insn of the shift/add sequence. */
2714 /* Line above always fails if INSN was moved by loop opt. */
2720 #if 0
2721 /* Currently, non-reduced/final-value givs are never split. */
2722 /* Should emit insns after the loop if possible, as the biv final value
2723 code below does. */
2725 /* If the final value is non-zero, and the giv has not been reduced,
2726 then must emit an instruction to set the final value. */
2728 {
2729 /* Create a new register to hold the value of the giv, and then set
2730 the giv to its final value before the loop start. The giv is set
2731 to its final value before loop start to ensure that this insn
2732 will always be executed, no matter how we exit. */
2742 }
2743 #endif
2745 /* This giv is splittable. If completely unrolling the loop, save the
2746 giv's initial value. Otherwise, save the constant zero for it. */
2749 {
2750 /* It is not safe to use bl->initial_value here, because it may not
2751 be invariant. It is safe to use the initial value stored in
2752 the splittable_regs array if it is set. In rare cases, it won't
2753 be set, so then we do exactly the same thing as
2754 find_splittable_regs does to get a safe value. */
2755 rtx biv_initial_value;
2763 else
2764 {
2770 }
2774 }
2775 else
2779 {
2780 /* If a giv was combined with another giv, then we can only split
2781 this giv if the giv it was combined with was reduced. This
2782 is because the value of v->new_reg is meaningless in this
2783 case. */
2785 {
2790 }
2791 /* If the giv is an address destination, it could be something other
2792 than a simple register, these have to be treated differently. */
2794 {
2795 /* If value is not a constant, register, or register plus
2796 constant, then compute its value into a register before
2797 loop start. This prevents invalid rtx sharing, and should
2798 generate better code. We can use bl->initial_value here
2799 instead of splittable_regs[bl->regno] because this code
2800 is going before the loop start. */
2807 {
2813 }
2816 }
2817 else
2818 {
2819 /* Splitting address givs is useful since it will often allow us
2820 to eliminate some increment insns for the base giv as
2821 unnecessary. */
2823 /* If the addr giv is combined with a dest_reg giv, then all
2824 references to that dest reg will be remapped, which is NOT
2825 what we want for split addr regs. We always create a new
2826 register for the split addr giv, just to be safe. */
2828 /* If we have multiple identical address givs within a
2829 single instruction, then use a single pseudo reg for
2830 both. This is necessary in case one is a match_dup
2831 of the other. */
2836 {
2842 }
2843 /* If multiple address GIVs have been combined with the
2844 same dest_reg GIV, do not create a new register for
2845 each. */
2850 /* combine_givs_p may return true for some cases
2851 where the add and mult values are not equal.
2852 To share a register here, the values must be
2853 equal. */
2856 /* If the memory references have different modes,
2857 then the address may not be valid and we must
2858 not share registers. */
2860 {
2863 }
2865 {
2866 /* If not completely unrolling the loop, then create a new
2867 register to hold the split value of the DEST_ADDR giv.
2868 Emit insn to initialize its value before loop start. */
2875 /* If the address giv has a constant in its new_reg value,
2876 then this constant can be pulled out and put in value,
2877 instead of being part of the initialization code. */
2881 {
2882 v->dest_reg
2885 /* Only succeed if this will give valid addresses.
2886 Try to validate both the first and the last
2887 address resulting from loop unrolling, if
2888 one fails, then can't do const elim here. */
2890 {
2891 /* Save the negative of the eliminated const, so
2892 that we can calculate the dest_reg's increment
2893 value later. */
2901 }
2902 else
2904 }
2905 else
2908 /* If the address hasn't been checked for validity yet, do so
2909 now, and fail completely if either the first or the last
2910 unrolled copy of the address is not a valid address
2911 for the instruction that uses it. */
2914 {
2924 }
2929 /* We set this after the address check, to guarantee that
2930 the register will be initialized. */
2933 /* To initialize the new register, just move the value of
2934 new_reg into it. This is not guaranteed to give a valid
2935 instruction on machines with complex addressing modes.
2936 If we can't recognize it, then delete it and emit insns
2937 to calculate the value from scratch. */
2941 {
2944 /* We can't use bl->initial_value to compute the initial
2945 value, because the loop may have been preconditioned.
2946 We must calculate it from NEW_REG. */
2960 }
2961 }
2962 else
2963 {
2966 /* Check the resulting address for validity, and fail
2967 if the resulting address would be invalid. */
2969 {
2979 }
2980 }
2982 /* Store the value of dest_reg into the insn. This sharing
2983 will not be a problem as this insn will always be copied
2984 later. */
2988 /* If this address giv is combined with a dest reg giv, then
2989 save the base giv's induction pointer so that we will be
2990 able to handle this address giv properly. The base giv
2991 itself does not have to be splittable. */
2997 {
2998 /* This giv maybe hasn't been combined with any others.
2999 Make sure that it's giv is marked as splittable here. */
3003 /* Make it appear to depend upon itself, so that the
3004 giv will be properly split in the main loop above. */
3006 {
3009 }
3010 }
3014 }
3015 }
3016 else
3017 {
3018 #if 0
3019 /* Currently, unreduced giv's can't be split. This is not too much
3020 of a problem since unreduced giv's are not live across loop
3021 iterations anyways. When unrolling a loop completely though,
3022 it makes sense to reduce&split givs when possible, as this will
3023 result in simpler instructions, and will not require that a reg
3024 be live across loop iterations. */
3029 #else
3031 #endif
3032 }
3034 /* Unreduced givs are only updated once by definition. Reduced givs
3035 are updated as many times as their biv is. Mark it so if this is
3036 a splittable register. Don't need to do anything for address givs
3037 where this may not be a register. */
3040 {
3046 }
3048 result++;
3051 {
3058 else
3062 }
3063 }
3066 }
3067 \f
3068 /* Try to prove that the register is dead after the loop exits. Trace every
3069 loop exit looking for an insn that will always be executed, which sets
3070 the register to some value, and appears before the first use of the register
3071 is found. If successful, then return 1, otherwise return 0. */
3073 /* ?? Could be made more intelligent in the handling of jumps, so that
3074 it can search past if statements and other similar structures. */
3076 static int
3079 rtx reg;
3080 {
3086 /* In addition to checking all exits of this loop, we must also check
3087 all exits of inner nested loops that would exit this loop. We don't
3088 have any way to identify those, so we just give up if there are any
3089 such inner loop exits. */
3092 label_count++;
3097 /* HACK: Must also search the loop fall through exit, create a label_ref
3098 here which points to the loop->end, and append the loop_number_exit_labels
3099 list to it. */
3104 {
3105 /* Succeed if find an insn which sets the biv or if reach end of
3106 function. Fail if find an insn that uses the biv, or if come to
3107 a conditional jump. */
3111 {
3114 {
3115 rtx set;
3123 }
3126 {
3130 /* Prevent infinite loop following infinite loops. */
3133 else
3135 }
3138 }
3139 }
3141 /* Success, the register is dead on all loop exits. */
3143 }
3145 /* Try to calculate the final value of the biv, the value it will have at
3146 the end of the loop. If we can do it, return that value. */
3148 rtx
3152 {
3156 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3161 /* The final value for reversed bivs must be calculated differently than
3162 for ordinary bivs. In this case, there is already an insn after the
3163 loop which sets this biv's final value (if necessary), and there are
3164 no other loop exits, so we can return any value. */
3166 {
3172 }
3174 /* Try to calculate the final value as initial value + (number of iterations
3175 * increment). For this to work, increment must be invariant, the only
3176 exit from the loop must be the fall through at the bottom (otherwise
3177 it may not have its final value when the loop exits), and the initial
3178 value of the biv must be invariant. */
3183 {
3187 {
3188 /* Can calculate the loop exit value, emit insns after loop
3189 end to calculate this value into a temporary register in
3190 case it is needed later. */
3202 }
3203 }
3205 /* Check to see if the biv is dead at all loop exits. */
3207 {
3214 }
3217 }
3219 /* Try to calculate the final value of the giv, the value it will have at
3220 the end of the loop. If we can do it, return that value. */
3222 rtx
3226 {
3229 rtx insn;
3231 rtx seq;
3237 /* The final value for givs which depend on reversed bivs must be calculated
3238 differently than for ordinary givs. In this case, there is already an
3239 insn after the loop which sets this giv's final value (if necessary),
3240 and there are no other loop exits, so we can return any value. */
3242 {
3248 }
3250 /* Try to calculate the final value as a function of the biv it depends
3251 upon. The only exit from the loop must be the fall through at the bottom
3252 and the insn that sets the giv must be executed on every iteration
3253 (otherwise the giv may not have its final value when the loop exits). */
3255 /* ??? Can calculate the final giv value by subtracting off the
3256 extra biv increments times the giv's mult_val. The loop must have
3257 only one exit for this to work, but the loop iterations does not need
3258 to be known. */
3263 {
3264 /* ?? It is tempting to use the biv's value here since these insns will
3265 be put after the loop, and hence the biv will have its final value
3266 then. However, this fails if the biv is subsequently eliminated.
3267 Perhaps determine whether biv's are eliminable before trying to
3268 determine whether giv's are replaceable so that we can use the
3269 biv value here if it is not eliminable. */
3271 /* We are emitting code after the end of the loop, so we must make
3272 sure that bl->initial_value is still valid then. It will still
3273 be valid if it is invariant. */
3279 {
3280 /* Can calculate the loop exit value of its biv as
3281 (n_iterations * increment) + initial_value */
3283 /* The loop exit value of the giv is then
3284 (final_biv_value - extra increments) * mult_val + add_val.
3285 The extra increments are any increments to the biv which
3286 occur in the loop after the giv's value is calculated.
3287 We must search from the insn that sets the giv to the end
3288 of the loop to calculate this value. */
3290 /* Put the final biv value in tem. */
3296 tem);
3298 /* Subtract off extra increments as we find them. */
3301 {
3306 {
3310 OPTAB_LIB_WIDEN);
3314 }
3315 }
3317 /* Now calculate the giv's final value. */
3326 }
3327 }
3329 /* Replaceable giv's should never reach here. */
3333 /* Check to see if the biv is dead at all loop exits. */
3335 {
3342 }
3345 }
3347 /* Look back before LOOP->START for the insn that sets REG and return
3348 the equivalent constant if there is a REG_EQUAL note otherwise just
3349 the SET_SRC of REG. */
3351 static rtx
3354 rtx reg;
3355 {
3358 rtx ret;
3362 {
3367 {
3368 /* We found the last insn before the loop that sets the register.
3369 If it sets the entire register, and has a REG_EQUAL note,
3370 then use the value of the REG_EQUAL note. */
3373 {
3376 /* Only use the REG_EQUAL note if it is a constant.
3377 Other things, divide in particular, will cause
3378 problems later if we use them. */
3382 else
3385 /* We cannot do this if it changes between the
3386 assignment and loop start though. */
3389 }
3391 }
3392 }
3394 }
3396 /* Return a simplified rtx for the expression OP - REG.
3398 REG must appear in OP, and OP must be a register or the sum of a register
3399 and a second term.
3401 Thus, the return value must be const0_rtx or the second term.
3403 The caller is responsible for verifying that REG appears in OP and OP has
3404 the proper form. */
3406 static rtx
3409 {
3413 {
3418 }
3419 /* OP does not contain REG as a term. */
3421 }
3423 /* Find and return register term common to both expressions OP0 and
3424 OP1 or NULL_RTX if no such term exists. Each expression must be a
3425 REG or a PLUS of a REG. */
3427 static rtx
3430 {
3433 {
3434 rtx op00;
3435 rtx op01;
3436 rtx op10;
3437 rtx op11;
3441 else
3446 else
3449 /* Find and return common register term if present. */
3454 }
3456 /* No common register term found. */
3458 }
3460 /* Determine the loop iterator and calculate the number of loop
3461 iterations. Returns the exact number of loop iterations if it can
3462 be calculated, otherwise returns zero. */
3464 unsigned HOST_WIDE_INT
3467 {
3473 HOST_WIDE_INT inc;
3479 rtx last_loop_insn;
3480 rtx reg_term;
3494 /* We used to use prev_nonnote_insn here, but that fails because it might
3495 accidentally get the branch for a contained loop if the branch for this
3496 loop was deleted. We can only trust branches immediately before the
3497 loop_end. */
3500 /* ??? We should probably try harder to find the jump insn
3501 at the end of the loop. The following code assumes that
3502 the last loop insn is a jump to the top of the loop. */
3504 {
3509 }
3511 /* If there is a more than a single jump to the top of the loop
3512 we cannot (easily) determine the iteration count. */
3514 {
3519 }
3521 /* If there are multiple conditionalized loop exit tests, they may jump
3522 back to differing CODE_LABELs. */
3524 {
3527 do
3528 {
3530 {
3531 /* There are some kinds of jumps we can't deal with easily. */
3533 {
3535 fprintf
3539 }
3542 covered by the uid_luid array. */
3544 /* Check if we jump back into the loop body. */
3547 {
3549 fprintf
3553 }
3554 }
3555 }
3557 }
3559 /* Find the iteration variable. If the last insn is a conditional
3560 branch, and the insn before tests a register value, make that the
3561 iteration variable. */
3565 {
3570 }
3572 /* ??? Get_condition may switch position of induction variable and
3573 invariant register when it canonicalizes the comparison. */
3580 {
3585 }
3587 /* The only new registers that are created before loop iterations
3588 are givs made from biv increments or registers created by
3589 load_mems. In the latter case, it is possible that try_copy_prop
3590 will propagate a new pseudo into the old iteration register but
3591 this will be marked by having the REG_USERVAR_P bit set. */
3597 /* Determine the initial value of the iteration variable, and the amount
3598 that it is incremented each loop. Use the tables constructed by
3599 the strength reduction pass to calculate these values. */
3601 /* Clear the result values, in case no answer can be found. */
3605 /* The iteration variable can be either a giv or a biv. Check to see
3606 which it is, and compute the variable's initial value, and increment
3607 value if possible. */
3609 /* If this is a new register, can't handle it since we don't have any
3610 reg_iv_type entry for it. */
3612 {
3617 }
3619 /* Reject iteration variables larger than the host wide int size, since they
3620 could result in a number of iterations greater than the range of our
3621 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3624 {
3629 }
3631 {
3636 }
3638 {
3642 /* Grab initial value, only useful if it is a constant. */
3646 {
3651 }
3654 }
3656 {
3659 rtx biv_initial_value;
3665 {
3670 }
3674 /* Increment value is mult_val times the increment value of the biv. */
3678 {
3684 /* The caller assumes that one full increment has occurred at the
3685 first loop test. But that's not true when the biv is incremented
3686 after the giv is set (which is the usual case), e.g.:
3687 i = 6; do {;} while (i++ < 9) .
3688 Therefore, we bias the initial value by subtracting the amount of
3689 the increment that occurs between the giv set and the giv test. */
3691 {
3693 {
3695 {
3701 }
3704 }
3705 }
3706 }
3712 /* Initial value is mult_val times the biv's initial value plus
3713 add_val. Only useful if it is a constant. */
3715 initial_value
3719 }
3720 else
3721 {
3726 }
3734 {
3748 /* Cannot determine loop iterations with this case. */
3765 }
3767 /* If the comparison value is an invariant register, then try to find
3768 its value from the insns before the start of the loop. */
3773 {
3776 /* If we don't get an invariant final value, we are better
3777 off with the original register. */
3780 }
3782 /* Calculate the approximate final value of the induction variable
3783 (on the last successful iteration). The exact final value
3784 depends on the branch operator, and increment sign. It will be
3785 wrong if the iteration variable is not incremented by one each
3786 time through the loop and (comparison_value + off_by_one -
3787 initial_value) % increment != 0.
3788 ??? Note that the final_value may overflow and thus final_larger
3789 will be bogus. A potentially infinite loop will be classified
3790 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3794 /* Save the calculated values describing this loop's bounds, in case
3795 precondition_loop_p will need them later. These values can not be
3796 recalculated inside precondition_loop_p because strength reduction
3797 optimizations may obscure the loop's structure.
3799 These values are only required by precondition_loop_p and insert_bct
3800 whenever the number of iterations cannot be computed at compile time.
3801 Only the difference between final_value and initial_value is
3802 important. Note that final_value is only approximate. */
3811 /* Try to determine the iteration count for loops such
3812 as (for i = init; i < init + const; i++). When running the
3813 loop optimization twice, the first pass often converts simple
3814 loops into this form. */
3817 {
3818 rtx reg1;
3819 rtx reg2;
3820 rtx const2;
3825 else
3828 /* Check for initial_value = reg1, final_value = reg2 + const2,
3829 where reg1 != reg2. */
3831 {
3832 rtx temp;
3834 /* Find what reg1 is equivalent to. Hopefully it will
3835 either be reg2 or reg2 plus a constant. */
3840 else
3841 {
3842 /* Find what reg2 is equivalent to. Hopefully it will
3843 either be reg1 or reg1 plus a constant. Let's ignore
3844 the latter case for now since it is not so common. */
3852 }
3853 }
3855 {
3856 rtx temp;
3858 /* When running the loop optimizer twice, check_dbra_loop
3859 further obfuscates reversible loops of the form:
3860 for (i = init; i < init + const; i++). We often end up with
3861 final_value = 0, initial_value = temp, temp = temp2 - init,
3862 where temp2 = init + const. If the loop has a vtop we
3863 can replace initial_value with const. */
3868 {
3874 }
3875 }
3876 }
3878 /* If have initial_value = reg + const1 and final_value = reg +
3879 const2, then replace initial_value with const1 and final_value
3880 with const2. This should be safe since we are protected by the
3881 initial comparison before entering the loop if we have a vtop.
3882 For example, a + b < a + c is not equivalent to b < c for all a
3883 when using modulo arithmetic.
3885 ??? Without a vtop we could still perform the optimization if we check
3886 the initial and final values carefully. */
3889 {
3892 }
3897 /* For EQ comparison loops, we don't have a valid final value.
3898 Check this now so that we won't leave an invalid value if we
3899 return early for any other reason. */
3904 {
3909 }
3912 {
3913 /* If we have a REG, check to see if REG holds a constant value. */
3914 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3915 clear if it is worthwhile to try to handle such RTL. */
3920 {
3922 {
3927 }
3929 }
3931 }
3934 {
3936 {
3941 }
3943 }
3945 {
3949 }
3951 {
3953 {
3958 }
3960 }
3962 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3964 final_larger
3969 else
3977 else
3980 /* There are 27 different cases: compare_dir = -1, 0, 1;
3981 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3982 There are 4 normal cases, 4 reverse cases (where the iteration variable
3983 will overflow before the loop exits), 4 infinite loop cases, and 15
3984 immediate exit (0 or 1 iteration depending on loop type) cases.
3985 Only try to optimize the normal cases. */
3987 /* (compare_dir/final_larger/increment_dir)
3988 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3989 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3990 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3991 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3993 /* ?? If the meaning of reverse loops (where the iteration variable
3994 will overflow before the loop exits) is undefined, then could
3995 eliminate all of these special checks, and just always assume
3996 the loops are normal/immediate/infinite. Note that this means
3997 the sign of increment_dir does not have to be known. Also,
3998 since it does not really hurt if immediate exit loops or infinite loops
3999 are optimized, then that case could be ignored also, and hence all
4000 loops can be optimized.
4002 According to ANSI Spec, the reverse loop case result is undefined,
4003 because the action on overflow is undefined.
4005 See also the special test for NE loops below. */
4009 /* Normal case. */
4010 ;
4011 else
4012 {
4016 }
4018 /* Calculate the number of iterations, final_value is only an approximation,
4019 so correct for that. Note that abs_diff and n_iterations are
4020 unsigned, because they can be as large as 2^n - 1. */
4024 {
4027 }
4029 {
4032 }
4033 else
4036 /* Given that iteration_var is going to iterate over its own mode,
4037 not HOST_WIDE_INT, disregard higher bits that might have come
4038 into the picture due to sign extension of initial and final
4039 values. */
4044 /* For NE tests, make sure that the iteration variable won't miss
4045 the final value. If abs_diff mod abs_incr is not zero, then the
4046 iteration variable will overflow before the loop exits, and we
4047 can not calculate the number of iterations. */
4051 /* Note that the number of iterations could be calculated using
4052 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
4053 handle potential overflow of the summation. */
4056 }
4058 /* Replace uses of split bivs with their split pseudo register. This is
4059 for original instructions which remain after loop unrolling without
4060 copying. */
4062 static rtx
4065 rtx x;
4066 {
4077 {
4089 #if 0
4090 /* If non-reduced/final-value givs were split, then this would also
4091 have to remap those givs also. */
4092 #endif
4100 }
4104 {
4108 {
4112 }
4113 }
4115 }
4117 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
4118 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
4119 return 0. COPY_START is where we can start looking for the insns
4120 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
4121 insns.
4123 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
4124 must dominate LAST_UID.
4126 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4127 may not dominate LAST_UID.
4129 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
4130 must dominate LAST_UID. */
4132 int
4137 rtx copy_start;
4138 rtx copy_end;
4139 {
4144 {
4147 /* Could not find FIRST_UID. */
4151 }
4153 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
4157 /* FIRST_UID is always executed. */
4162 {
4163 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
4164 can not be sure that FIRST_UID dominates LAST_UID. */
4167 /* Could not find LAST_UID, but we reached the end of the loop, so
4168 it must be safe. */
4172 }
4174 /* FIRST_UID is always executed if LAST_UID is executed. */
4176 }
4178 /* This routine is called when the number of iterations for the unrolled
4179 loop is one. The goal is to identify a loop that begins with an
4180 unconditional branch to the loop continuation note (or a label just after).
4181 In this case, the unconditional branch that starts the loop needs to be
4182 deleted so that we execute the single iteration. */
4184 static rtx
4186 rtx loop_start;
4187 rtx loop_cont;
4188 {
4191 /* See if loop start, or the next insn is an unconditional jump. */
4202 /* Examine insn after loop continuation note. Return if not a label. */
4207 /* Return the loop start if the branch label matches the code label. */
4210 else
4212 }