| blob | 690200296eb083e40d96a3f4eea34506fc45198b |
1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995, 1997, 1998, 1999, 2000, 2001,
3 2002, 2003, 2004
4 Free Software Foundation, Inc.
5 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 2, or (at your option) any later
12 version.
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING. If not, write to the Free
21 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
22 02111-1307, USA. */
24 /* Try to unroll a loop, and split induction variables.
26 Loops for which the number of iterations can be calculated exactly are
27 handled specially. If the number of iterations times the insn_count is
28 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
29 Otherwise, we try to unroll the loop a number of times modulo the number
30 of iterations, so that only one exit test will be needed. It is unrolled
31 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
32 the insn count.
34 Otherwise, if the number of iterations can be calculated exactly at
35 run time, and the loop is always entered at the top, then we try to
36 precondition the loop. That is, at run time, calculate how many times
37 the loop will execute, and then execute the loop body a few times so
38 that the remaining iterations will be some multiple of 4 (or 2 if the
39 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
40 with only one exit test needed at the end of the loop.
42 Otherwise, if the number of iterations can not be calculated exactly,
43 not even at run time, then we still unroll the loop a number of times
44 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
45 but there must be an exit test after each copy of the loop body.
47 For each induction variable, which is dead outside the loop (replaceable)
48 or for which we can easily calculate the final value, if we can easily
49 calculate its value at each place where it is set as a function of the
50 current loop unroll count and the variable's value at loop entry, then
51 the induction variable is split into `N' different variables, one for
52 each copy of the loop body. One variable is live across the backward
53 branch, and the others are all calculated as a function of this variable.
54 This helps eliminate data dependencies, and leads to further opportunities
55 for cse. */
57 /* Possible improvements follow: */
59 /* ??? Add an extra pass somewhere to determine whether unrolling will
60 give any benefit. E.g. after generating all unrolled insns, compute the
61 cost of all insns and compare against cost of insns in rolled loop.
63 - On traditional architectures, unrolling a non-constant bound loop
64 is a win if there is a giv whose only use is in memory addresses, the
65 memory addresses can be split, and hence giv increments can be
66 eliminated.
67 - It is also a win if the loop is executed many times, and preconditioning
68 can be performed for the loop.
69 Add code to check for these and similar cases. */
71 /* ??? Improve control of which loops get unrolled. Could use profiling
72 info to only unroll the most commonly executed loops. Perhaps have
73 a user specifiable option to control the amount of code expansion,
74 or the percent of loops to consider for unrolling. Etc. */
76 /* ??? Look at the register copies inside the loop to see if they form a
77 simple permutation. If so, iterate the permutation until it gets back to
78 the start state. This is how many times we should unroll the loop, for
79 best results, because then all register copies can be eliminated.
80 For example, the lisp nreverse function should be unrolled 3 times
81 while (this)
82 {
83 next = this->cdr;
84 this->cdr = prev;
85 prev = this;
86 this = next;
87 }
89 ??? The number of times to unroll the loop may also be based on data
90 references in the loop. For example, if we have a loop that references
91 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
93 /* ??? Add some simple linear equation solving capability so that we can
94 determine the number of loop iterations for more complex loops.
95 For example, consider this loop from gdb
96 #define SWAP_TARGET_AND_HOST(buffer,len)
97 {
98 char tmp;
99 char *p = (char *) buffer;
100 char *q = ((char *) buffer) + len - 1;
101 int iterations = (len + 1) >> 1;
102 int i;
103 for (p; p < q; p++, q--;)
104 {
105 tmp = *q;
106 *q = *p;
107 *p = tmp;
108 }
109 }
110 Note that:
111 start value = p = &buffer + current_iteration
112 end value = q = &buffer + len - 1 - current_iteration
113 Given the loop exit test of "p < q", then there must be "q - p" iterations,
114 set equal to zero and solve for number of iterations:
115 q - p = len - 1 - 2*current_iteration = 0
116 current_iteration = (len - 1) / 2
117 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
118 iterations of this loop. */
120 /* ??? Currently, no labels are marked as loop invariant when doing loop
121 unrolling. This is because an insn inside the loop, that loads the address
122 of a label inside the loop into a register, could be moved outside the loop
123 by the invariant code motion pass if labels were invariant. If the loop
124 is subsequently unrolled, the code will be wrong because each unrolled
125 body of the loop will use the same address, whereas each actually needs a
126 different address. A case where this happens is when a loop containing
127 a switch statement is unrolled.
129 It would be better to let labels be considered invariant. When we
130 unroll loops here, check to see if any insns using a label local to the
131 loop were moved before the loop. If so, then correct the problem, by
132 moving the insn back into the loop, or perhaps replicate the insn before
133 the loop, one copy for each time the loop is unrolled. */
156 /* The prime factors looked for when trying to unroll a loop by some
157 number which is modulo the total number of iterations. Just checking
158 for these 4 prime factors will find at least one factor for 75% of
159 all numbers theoretically. Practically speaking, this will succeed
160 almost all of the time since loops are generally a multiple of 2
161 and/or 5. */
163 #define NUM_FACTORS 4
168 /* Describes the different types of loop unrolling performed. */
170 enum unroll_types
171 {
172 UNROLL_COMPLETELY,
173 UNROLL_MODULO,
174 UNROLL_NAIVE
175 };
177 /* Indexed by register number, if nonzero, then it contains a pointer
178 to a struct induction for a DEST_REG giv which has been combined with
179 one of more address givs. This is needed because whenever such a DEST_REG
180 giv is modified, we must modify the value of all split address givs
181 that were combined with this DEST_REG giv. */
185 /* Indexed by register number, if this is a splittable induction variable,
186 then this will hold the current value of the register, which depends on the
187 iteration number. */
191 /* Indexed by register number, if this is a splittable induction variable,
192 then this will hold the number of instructions in the loop that modify
193 the induction variable. Used to ensure that only the last insn modifying
194 a split iv will update the original iv of the dest. */
198 /* Forward declarations. */
221 /* Try to unroll one loop and split induction variables in the loop.
223 The loop is described by the arguments LOOP and INSN_COUNT.
224 STRENGTH_REDUCTION_P indicates whether information generated in the
225 strength reduction pass is available.
227 This function is intended to be called from within `strength_reduce'
228 in loop.c. */
230 void
232 {
242 rtx insert_before;
249 rtx start_label;
254 rtx safety_label;
255 /* This points to the last real insn in the loop, which should be either
256 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
257 jumps). */
258 rtx last_loop_insn;
262 /* Don't bother unrolling huge loops. Since the minimum factor is
263 two, loops greater than one half of MAX_UNROLLED_INSNS will never
264 be unrolled. */
266 {
270 }
272 /* Determine type of unroll to perform. Depends on the number of iterations
273 and the size of the loop. */
275 /* If there is no strength reduce info, then set
276 loop_info->n_iterations to zero. This can happen if
277 strength_reduce can't find any bivs in the loop. A value of zero
278 indicates that the number of iterations could not be calculated. */
287 /* Find and save a pointer to the last nonnote insn in the loop. */
291 /* Calculate how many times to unroll the loop. Indicate whether or
292 not the loop is being completely unrolled. */
295 {
296 /* Handle the case where the loop begins with an unconditional
297 jump to the loop condition. Make sure to delete the jump
298 insn, otherwise the loop body will never execute. */
300 /* FIXME this actually checks for a jump to the continue point, which
301 is not the same as the condition in a for loop. As a result, this
302 optimization fails for most for loops. We should really use flow
303 information rather than instruction pattern matching. */
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 }
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
334 /* Remove the loop notes since this is no longer a loop. */
345 }
346 }
349 /* Avoid overflow in the next expression. */
352 {
355 }
357 {
358 /* Try to factor the number of iterations. Don't bother with the
359 general case, only using 2, 3, 5, and 7 will get 75% of all
360 numbers theoretically, and almost all in practice. */
368 {
371 }
373 /* Start with the larger factors first so that we generally
374 get lots of unrolling. */
380 {
382 {
385 }
386 else
388 }
390 /* If we couldn't find any factors, then unroll as in the normal
391 case. */
393 {
396 }
397 else
399 }
401 /* Default case, calculate number of times to unroll loop based on its
402 size. */
404 {
409 else
411 }
413 /* Now we know how many times to unroll the loop. */
419 {
420 /* Loops of these types can start with jump down to the exit condition
421 in rare circumstances.
423 Consider a pair of nested loops where the inner loop is part
424 of the exit code for the outer loop.
426 In this case jump.c will not duplicate the exit test for the outer
427 loop, so it will start with a jump to the exit code.
429 Then consider if the inner loop turns out to iterate once and
430 only once. We will end up deleting the jumps associated with
431 the inner loop. However, the loop notes are not removed from
432 the instruction stream.
434 And finally assume that we can compute the number of iterations
435 for the outer loop.
437 In this case unroll may want to unroll the outer loop even though
438 it starts with a jump to the outer loop's exit code.
440 We could try to optimize this case, but it hardly seems worth it.
441 Just return without unrolling the loop in such cases. */
448 }
451 {
452 /* Completely unrolling the loop: Delete the compare and branch at
453 the end (the last two instructions). This delete must done at the
454 very end of loop unrolling, to avoid problems with calls to
455 back_branch_in_range_p, which is called by find_splittable_regs.
456 All increments of splittable bivs/givs are changed to load constant
457 instructions. */
461 /* Set insert_before to the instruction immediately after the JUMP_INSN
462 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
463 the loop will be correctly handled by copy_loop_body. */
466 /* Set copy_end to the insn before the jump at the end of the loop. */
470 {
472 #ifdef HAVE_cc0
473 /* The instruction immediately before the JUMP_INSN may be a compare
474 instruction which we do not want to copy. */
477 #endif
478 }
479 else
480 {
481 /* We currently can't unroll a loop if it doesn't end with a
482 JUMP_INSN. There would need to be a mechanism that recognizes
483 this case, and then inserts a jump after each loop body, which
484 jumps to after the last loop body. */
489 }
490 }
492 {
493 /* Partially unrolling the loop: The compare and branch at the end
494 (the last two instructions) must remain. Don't copy the compare
495 and branch instructions at the end of the loop. Insert the unrolled
496 code immediately before the compare/branch at the end so that the
497 code will fall through to them as before. */
501 /* Set insert_before to the jump insn at the end of the loop.
502 Set copy_end to before the jump insn at the end of the loop. */
504 {
507 }
509 {
511 #ifdef HAVE_cc0
512 /* The instruction immediately before the JUMP_INSN may be a compare
513 instruction which we do not want to copy or delete. */
516 #endif
518 }
519 else
520 {
521 /* We currently can't unroll a loop if it doesn't end with a
522 JUMP_INSN. There would need to be a mechanism that recognizes
523 this case, and then inserts a jump after each loop body, which
524 jumps to after the last loop body. */
529 }
530 }
531 else
532 {
533 /* Normal case: Must copy the compare and branch instructions at the
534 end of the loop. */
537 {
538 /* Loop ends with an unconditional jump and a barrier.
539 Handle this like above, don't copy jump and barrier.
540 This is not strictly necessary, but doing so prevents generating
541 unconditional jumps to an immediately following label.
543 This will be corrected below if the target of this jump is
544 not the start_label. */
548 }
550 {
551 /* Set insert_before to immediately after the JUMP_INSN, so that
552 NOTEs at the end of the loop will be correctly handled by
553 copy_loop_body. */
556 }
557 else
558 {
559 /* We currently can't unroll a loop if it doesn't end with a
560 JUMP_INSN. There would need to be a mechanism that recognizes
561 this case, and then inserts a jump after each loop body, which
562 jumps to after the last loop body. */
567 }
569 /* If copying exit test branches because they can not be eliminated,
570 then must convert the fall through case of the branch to a jump past
571 the end of the loop. Create a label to emit after the loop and save
572 it for later use. Do not use the label after the loop, if any, since
573 it might be used by insns outside the loop, or there might be insns
574 added before it later by final_[bg]iv_value which must be after
575 the real exit label. */
583 {
584 /* The loop starts with a jump down to the exit condition test.
585 Start copying the loop after the barrier following this
586 jump insn. */
589 /* Splitting induction variables doesn't work when the loop is
590 entered via a jump to the bottom, because then we end up doing
591 a comparison against a new register for a split variable, but
592 we did not execute the set insn for the new register because
593 it was skipped over. */
598 }
599 else
601 }
603 /* This should always be the first label in the loop. */
605 /* There may be a line number note and/or a loop continue note here. */
609 {
610 /* This can happen as a result of jump threading. If the first insns in
611 the loop test the same condition as the loop's backward jump, or the
612 opposite condition, then the backward jump will be modified to point
613 to elsewhere, and the loop's start label is deleted.
615 This case currently can not be handled by the loop unrolling code. */
621 }
623 {
624 /* The jump optimization pass must have combined the original start label
625 with a named label for a goto. We can't unroll this case because
626 jumps which go to the named label must be handled differently than
627 jumps to the loop start, and it is impossible to differentiate them
628 in this case. */
633 }
639 {
640 /* In this case, we must copy the jump and barrier, because they will
641 not be converted to jumps to an immediately following label. */
645 }
650 {
651 /* ??? The loop ends with a conditional branch that does not branch back
652 to the loop start label. In this case, we must emit an unconditional
653 branch to the loop exit after emitting the final branch.
654 copy_loop_body does not have support for this currently, so we
655 give up. It doesn't seem worthwhile to unroll anyways since
656 unrolling would increase the number of branch instructions
657 executed. */
662 }
664 /* Allocate a translation table for the labels and insn numbers.
665 They will be filled in as we copy the insns in the loop. */
670 /* Various paths through the unroll code may reach the "egress" label
671 without initializing fields within the map structure.
673 To be safe, we use xcalloc to zero the memory. */
676 /* Allocate the label map. */
679 {
682 }
684 /* Search the loop and mark all local labels, i.e. the ones which have to
685 be distinct labels when copied. For all labels which might be
686 non-local, set their label_map entries to point to themselves.
687 If they happen to be local their label_map entries will be overwritten
688 before the loop body is copied. The label_map entries for local labels
689 will be set to a different value each time the loop body is copied. */
692 {
693 rtx note;
698 {
705 {
709 rtx label;
712 {
715 }
716 }
717 }
721 }
723 /* Allocate space for the insn map. */
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. */
752 /* Mark all local registers, i.e. the ones which are referenced only
753 inside the loop. */
755 {
759 /* If a register is used in the jump insn, we must not duplicate it
760 since it will also be used outside the loop. */
762 copy_end_luid--;
764 /* If we have a target that uses cc0, then we also must not duplicate
765 the insn that sets cc0 before the jump insn, if one is present. */
766 #ifdef HAVE_cc0
769 copy_end_luid--;
770 #endif
772 /* If copy_start points to the NOTE that starts the loop, then we must
773 use the next luid, because invariant pseudo-regs moved out of the loop
774 have their lifetimes modified to start here, but they are not safe
775 to duplicate. */
777 copy_start_luid++;
779 /* If a pseudo's lifetime is entirely contained within this loop, then we
780 can use a different pseudo in each unrolled copy of the loop. This
781 results in better code. */
782 /* We must limit the generic test to max_reg_before_loop, because only
783 these pseudo registers have valid regno_first_uid info. */
789 {
790 /* However, we must also check for loop-carried dependencies.
791 If the value the pseudo has at the end of iteration X is
792 used by iteration X+1, then we can not use a different pseudo
793 for each unrolled copy of the loop. */
794 /* A pseudo is safe if regno_first_uid is a set, and this
795 set dominates all instructions from regno_first_uid to
796 regno_last_uid. */
797 /* ??? This check is simplistic. We would get better code if
798 this check was more sophisticated. */
804 {
807 else
809 r);
810 }
811 }
812 }
814 /* If this loop requires exit tests when unrolled, check to see if we
815 can precondition the loop so as to make the exit tests unnecessary.
816 Just like variable splitting, this is not safe if the loop is entered
817 via a jump to the bottom. Also, can not do this if no strength
818 reduce info, because precondition_loop_p uses this info. */
820 /* Must copy the loop body for preconditioning before the following
821 find_splittable_regs call since that will emit insns which need to
822 be after the preconditioned loop copies, but immediately before the
823 unrolled loop copies. */
825 /* Also, it is not safe to split induction variables for the preconditioned
826 copies of the loop body. If we split induction variables, then the code
827 assumes that each induction variable can be represented as a function
828 of its initial value and the loop iteration number. This is not true
829 in this case, because the last preconditioned copy of the loop body
830 could be any iteration from the first up to the `unroll_number-1'th,
831 depending on the initial value of the iteration variable. Therefore
832 we can not split induction variables here, because we can not calculate
833 their value. Hence, this code must occur before find_splittable_regs
834 is called. */
837 {
844 {
860 /* Limit loop unrolling to 4, since this will make 7 copies of
861 the loop body. */
865 /* Save the absolute value of the increment, and also whether or
866 not it is negative. */
870 {
873 }
877 /* We must copy the final and initial values here to avoid
878 improperly shared rtl. */
882 /* Final value may have form of (PLUS val1 const1_rtx). We need
883 to convert it into general operand, so compute the real value. */
889 /* Calculate the difference between the final and initial values.
890 Final value may be a (plus (reg x) (const_int 1)) rtx.
892 We have to deal with for (i = 0; --i < 6;) type loops.
893 For such loops the real final value is the first time the
894 loop variable overflows, so the diff we calculate is the
895 distance from the overflow value. This is 0 or ~0 for
896 unsigned loops depending on the direction, or INT_MAX,
897 INT_MAX+1 for signed loops. We really do not need the
898 exact value, since we are only interested in the diff
899 modulo the increment, and the increment is a power of 2,
900 so we can pretend that the overflow value is 0/~0. */
904 initial_value);
905 else
910 /* Now calculate (diff % (unroll * abs (increment))) by using an
911 and instruction. */
916 /* Now emit a sequence of branches to jump to the proper precond
917 loop entry point. */
923 /* Check for the case where the initial value is greater than or
924 equal to the final value. In that case, we want to execute
925 exactly one loop iteration. The code below will fail for this
926 case. This check does not apply if the loop has a NE
927 comparison at the end. */
930 {
931 rtx incremented_initval;
934 incremented_initval
936 incremented_initval
939 cmp_code = (less_p
944 incremented_initval,
948 }
950 /* Assuming the unroll_number is 4, and the increment is 2, then
951 for a negative increment: for a positive increment:
952 diff = 0,1 precond 0 diff = 0,7 precond 0
953 diff = 2,3 precond 3 diff = 1,2 precond 1
954 diff = 4,5 precond 2 diff = 3,4 precond 2
955 diff = 6,7 precond 1 diff = 5,6 precond 3 */
957 /* We only need to emit (unroll_number - 1) branches here, the
958 last case just falls through to the following code. */
960 /* ??? This would give better code if we emitted a tree of branches
961 instead of the current linear list of branches. */
964 {
968 /* For negative increments, must invert the constant compared
969 against, except when comparing against zero. */
971 {
974 }
976 {
979 }
980 else
981 {
984 }
992 }
994 /* If the increment is greater than one, then we need another branch,
995 to handle other cases equivalent to 0. */
997 /* ??? This should be merged into the code above somehow to help
998 simplify the code here, and reduce the number of branches emitted.
999 For the negative increment case, the branch here could easily
1000 be merged with the `0' case branch above. For the positive
1001 increment case, it is not clear how this can be simplified. */
1004 {
1009 {
1012 }
1013 else
1014 {
1017 }
1021 }
1027 /* Only the last copy of the loop body here needs the exit
1028 test, so set copy_end to exclude the compare/branch here,
1029 and then reset it inside the loop when get to the last
1030 copy. */
1035 {
1037 #ifdef HAVE_cc0
1038 /* The immediately preceding insn may be a compare which
1039 we do not want to copy. */
1042 #endif
1043 }
1044 else
1048 {
1064 {
1069 }
1070 /* The last copy needs the compare/branch insns at the end,
1071 so reset copy_end here if the loop ends with a conditional
1072 branch. */
1075 {
1078 else
1080 }
1082 /* None of the copies are the `last_iteration', so just
1083 pass zero for that parameter. */
1087 }
1091 {
1094 }
1095 else
1096 {
1098 #ifdef HAVE_cc0
1099 /* The instruction immediately before the JUMP_INSN may
1100 be a compare instruction which we do not want to copy
1101 or delete. */
1104 #endif
1106 }
1108 /* Set unroll type to MODULO now. */
1112 /* Clean up. */
1114 }
1115 }
1117 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1118 the loop unless all loops are being unrolled. */
1120 {
1125 }
1127 /* At this point, we are guaranteed to unroll the loop. */
1129 /* Keep track of the unroll factor for the loop. */
1132 /* And whether the loop has been preconditioned. */
1135 /* Remember whether it was preconditioned for the second loop pass. */
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 }
1326 /* A helper function for unroll_loop. Emit a compare and branch to
1327 satisfy (CMP OP1 OP2), but pass this through the simplifier first.
1328 If the branch turned out to be conditional, return it, otherwise
1329 return NULL. */
1331 static rtx
1334 {
1339 {
1349 }
1351 {
1356 }
1359 }
1360 \f
1361 /* Return true if the loop can be safely, and profitably, preconditioned
1362 so that the unrolled copies of the loop body don't need exit tests.
1364 This only works if final_value, initial_value and increment can be
1365 determined, and if increment is a constant power of 2.
1366 If increment is not a power of 2, then the preconditioning modulo
1367 operation would require a real modulo instead of a boolean AND, and this
1368 is not considered `profitable'. */
1370 /* ??? If the loop is known to be executed very many times, or the machine
1371 has a very cheap divide instruction, then preconditioning is a win even
1372 when the increment is not a power of 2. Use RTX_COST to compute
1373 whether divide is cheap.
1374 ??? A divide by constant doesn't actually need a divide, look at
1375 expand_divmod. The reduced cost of this optimized modulo is not
1376 reflected in RTX_COST. */
1378 int
1382 {
1387 {
1389 {
1393 }
1394 else
1395 {
1399 }
1404 "Preconditioning: Success, number of iterations known, "
1408 }
1411 {
1416 }
1418 {
1423 }
1425 {
1430 }
1432 {
1437 }
1440 {
1445 }
1447 /* Unsigned_compare and compare_dir can be ignored here, since they do
1448 not matter for preconditioning. */
1451 {
1456 }
1458 /* Must ensure that final_value is invariant, so call
1459 loop_invariant_p to check. Before doing so, must check regno
1460 against max_reg_before_loop to make sure that the register is in
1461 the range covered by loop_invariant_p. If it isn't, then it is
1462 most likely a biv/giv which by definition are not invariant. */
1468 {
1473 }
1475 /* Fail for floating point values, since the caller of this function
1476 does not have code to deal with them. */
1479 {
1484 }
1486 /* Fail if loop_info->iteration_var is not live before loop_start,
1487 since we need to test its value in the preconditioning code. */
1491 {
1496 }
1498 /* Note that loop_iterations biases the initial value for GIV iterators
1499 such as "while (i-- > 0)" so that we can calculate the number of
1500 iterations just like for BIV iterators.
1502 Also note that the absolute values of initial_value and
1503 final_value are unimportant as only their difference is used for
1504 calculating the number of loop iterations. */
1509 /* Decide what mode to do these calculations in. Choose the larger
1510 of final_value's mode and initial_value's mode, or a full-word if
1511 both are constants. */
1514 {
1518 }
1524 /* Success! */
1528 }
1530 /* All pseudo-registers must be mapped to themselves. Two hard registers
1531 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1532 REGNUM, to avoid function-inlining specific conversions of these
1533 registers. All other hard regs can not be mapped because they may be
1534 used with different
1535 modes. */
1537 static void
1539 {
1544 /* Just clear the rest of the entries. */
1552 }
1553 \f
1554 /* Strength-reduction will often emit code for optimized biv/givs which
1555 calculates their value in a temporary register, and then copies the result
1556 to the iv. This procedure reconstructs the pattern computing the iv;
1557 verifying that all operands are of the proper form.
1559 PATTERN must be the result of single_set.
1560 The return value is the amount that the giv is incremented by. */
1562 static rtx
1564 {
1565 rtx increment;
1569 retry:
1570 /* Verify that we have an increment insn here. First check for a plus
1571 as the set source. */
1573 {
1574 /* SR sometimes computes the new giv value in a temp, then copies it
1575 to the new_reg. */
1581 /* The last insn emitted is not needed, so delete it to avoid confusing
1582 the second cse pass. This insn sets the giv unnecessarily. */
1584 }
1586 /* Verify that we have a constant as the second operand of the plus. */
1589 {
1590 /* SR sometimes puts the constant in a register, especially if it is
1591 too big to be an add immed operand. */
1594 /* SR may have used LO_SUM to compute the constant if it is too large
1595 for a load immed operand. In this case, the constant is in operand
1596 one of the LO_SUM rtx. */
1600 /* Some ports store large constants in memory and add a REG_EQUAL
1601 note to the store insn. */
1603 {
1607 }
1613 {
1614 /* The rs6000 port loads some constants with IOR.
1615 The alpha port loads some constants with ASHIFT and PLUS.
1616 The sparc64 port loads some constants with LSHIFTRT. */
1622 /* Don't need the last insn anymore. */
1635 else
1637 }
1642 /* The insn loading the constant into a register is no longer needed,
1643 so delete it. */
1645 }
1649 else
1652 /* Check that the source register is the same as the register we expected
1653 to see as the source. If not, something is seriously wrong. */
1656 {
1657 /* Some machines (e.g. the romp), may emit two add instructions for
1658 certain constants, so lets try looking for another add immediately
1659 before this one if we have only seen one add insn so far. */
1662 {
1663 tries++;
1671 }
1674 }
1677 }
1679 /* Copy REG_NOTES, except for insn references, because not all insn_map
1680 entries are valid yet. We do need to copy registers now though, because
1681 the reg_map entries can change during copying. */
1683 static rtx
1685 {
1686 rtx copy;
1697 /* Don't substitute for these yet. */
1699 else
1705 }
1707 /* Fixup insn references in copied REG_NOTES. */
1709 static void
1711 {
1713 {
1717 {
1720 /* If we failed to remap the note, something is awry.
1721 Allow REG_LABEL as it may reference label outside
1722 the unrolled loop. */
1724 {
1727 }
1728 else
1730 }
1733 }
1734 }
1736 /* Copy each instruction in the loop, substituting from map as appropriate.
1737 This is very similar to a loop in expand_inline_function. */
1739 static void
1744 rtx copy_notes_from)
1745 {
1750 #ifdef HAVE_cc0
1752 #endif
1756 /* If this isn't the last iteration, then map any references to the
1757 start_label to final_label. Final label will then be emitted immediately
1758 after the end of this loop body if it was ever used.
1760 If this is the last iteration, then map references to the start_label
1761 to itself. */
1763 {
1766 }
1767 else
1773 do
1774 {
1780 {
1786 /* Check to see if this is a giv that has been combined with
1787 some split address givs. (Combined in the sense that
1788 `combine_givs' in loop.c has put two givs in the same register.)
1789 In this case, we must search all givs based on the same biv to
1790 find the address givs. Then split the address givs.
1791 Do this before splitting the giv, since that may map the
1792 SET_DEST to a new register. */
1797 {
1805 /* Although the giv_inc amount is not needed here, we must call
1806 calculate_giv_inc here since it might try to delete the
1807 last insn emitted. If we wait until later to call it,
1808 we might accidentally delete insns generated immediately
1809 below by emit_unrolled_add. */
1813 /* Now find all address giv's that were combined with this
1814 giv 'v'. */
1817 {
1820 /* If this DEST_ADDR giv was not split, then ignore it. */
1824 /* Scale this_giv_inc if the multiplicative factors of
1825 the two givs are different. */
1835 {
1836 /* Must emit an insn to increment the split address
1837 giv. Add in the const_adjust field in case there
1838 was a constant eliminated from the address. */
1841 /* tv->dest_reg will be either a bare register,
1842 or else a register plus a constant. */
1845 else
1848 /* Check for shared address givs, and avoid
1849 incrementing the shared pseudo reg more than
1850 once. */
1852 {
1853 /* tv->dest_reg may actually be a (PLUS (REG)
1854 (CONST)) here, so we must call plus_constant
1855 to add the const_adjust amount before calling
1856 emit_unrolled_add below. */
1861 {
1862 /* The constant could be too large for an add
1863 immediate, so can't directly emit an insn
1864 here. */
1867 }
1868 }
1870 /* Reset the giv to be just the register again, in case
1871 it is used after the set we have just emitted.
1872 We must subtract the const_adjust factor added in
1873 above. */
1877 }
1878 }
1879 }
1881 /* If this is a setting of a splittable variable, then determine
1882 how to split the variable, create a new set based on this split,
1883 and set up the reg_map so that later uses of the variable will
1884 use the new split variable. */
1891 {
1899 /* Compute the increment value for the giv, if it wasn't
1900 already computed above. */
1907 {
1908 /* Completely unrolling the loop. Set the induction
1909 variable to a known constant value. */
1911 /* The value in splittable_regs may be an invariant
1912 value, so we must use plus_constant here. */
1918 {
1921 }
1922 else
1923 {
1924 /* The splittable_regs value must be a REG or a
1925 CONST_INT, so put the entire value in the giv_src_reg
1926 variable. */
1929 }
1930 }
1931 else
1932 {
1933 /* Partially unrolling loop. Create a new pseudo
1934 register for the iteration variable, and set it to
1935 be a constant plus the original register. Except
1936 on the last iteration, when the result has to
1937 go back into the original iteration var register. */
1939 /* Handle bivs which must be mapped to a new register
1940 when split. This happens for bivs which need their
1941 final value set before loop entry. The new register
1942 for the biv was stored in the biv's first struct
1943 induction entry by find_splittable_regs. */
1947 {
1950 }
1952 #if 0
1953 /* If non-reduced/final-value givs were split, then
1954 this would have to remap those givs also. See
1955 find_splittable_regs. */
1956 #endif
1960 giv_inc,
1964 /* Now split the induction variable by changing the dest
1965 of this insn to a new register, and setting its
1966 reg_map entry to point to this new register.
1968 If this is the last iteration, and this is the last insn
1969 that will update the iv, then reuse the original dest,
1970 to ensure that the iv will have the proper value when
1971 the loop exits or repeats.
1973 Using splittable_regs_updates here like this is safe,
1974 because it can only be greater than one if all
1975 instructions modifying the iv are always executed in
1976 order. */
1980 {
1985 giv_inc == const0_rtx
1986 ? giv_src_reg
1990 }
1991 else
1993 }
1995 /* The constant being added could be too large for an add
1996 immediate, so can't directly emit an insn here. */
2000 }
2001 else
2002 {
2005 }
2009 /* If there is a REG_EQUAL note present whose value
2010 is not loop invariant, then delete it, since it
2011 may cause problems with later optimization passes. */
2016 #ifdef HAVE_cc0
2017 /* If this insn is setting CC0, it may need to look at
2018 the insn that uses CC0 to see what type of insn it is.
2019 In that case, the call to recog via validate_change will
2020 fail. So don't substitute constants here. Instead,
2021 do it when we emit the following insn.
2023 For example, see the pyr.md file. That machine has signed and
2024 unsigned compares. The compare patterns must check the
2025 following branch insn to see which what kind of compare to
2026 emit.
2028 If the previous insn set CC0, substitute constants on it as
2029 well. */
2032 else
2033 {
2038 }
2039 #else
2041 #endif
2043 /* Make split induction variable constants `permanent' since we
2044 know there are no backward branches across iteration variable
2045 settings which would invalidate this. */
2047 {
2054 }
2064 {
2066 CODE_LABEL_NUMBER
2069 }
2072 {
2074 /* This is a branch to the beginning of the loop; this is the
2075 last insn being copied; and this is not the last iteration.
2076 In this case, we want to change the original fall through
2077 case to be a branch past the end of the loop, and the
2078 original jump label case to fall_through. */
2081 {
2082 rtx jmp;
2084 /* Can't do it by reversing the jump (probably because we
2085 couldn't reverse the conditions), so emit a new
2086 jump_insn after COPY, and redirect the jump around
2087 that. */
2096 }
2097 }
2099 #ifdef HAVE_cc0
2103 #endif
2106 /* Set the jump label of COPY correctly to avoid problems with
2107 later passes of unroll_loop, if INSN had jump label set. */
2109 {
2112 /* Can't use the label_map for every insn, since this may be
2113 the backward branch, and hence the label was not mapped. */
2115 {
2120 {
2123 else
2125 }
2126 }
2130 else
2131 {
2132 /* An unrecognizable jump insn, probably the entry jump
2133 for a switch statement. This label must have been mapped,
2134 so just use the label_map to get the new jump label. */
2138 }
2140 /* If this is a non-local jump, then must increase the label
2141 use count so that the label will not be deleted when the
2142 original jump is deleted. */
2144 }
2147 {
2155 }
2157 /* If this used to be a conditional jump insn but whose branch
2158 direction is now known, we must do something special. */
2160 {
2161 #ifdef HAVE_cc0
2162 /* If the previous insn set cc0 for us, delete it. */
2165 #endif
2167 /* If this is now a no-op, delete it. */
2169 {
2172 }
2173 else
2174 /* Otherwise, this is unconditional jump so we must put a
2175 BARRIER after it. We could do some dead code elimination
2176 here, but jump.c will do it just as well. */
2178 }
2189 /* Because the USAGE information potentially contains objects other
2190 than hard registers, we need to copy it. */
2195 #ifdef HAVE_cc0
2199 #endif
2202 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2208 /* If this is the loop start label, then we don't need to emit a
2209 copy of this label since no one will use it. */
2212 {
2216 }
2224 /* VTOP and CONT notes are valid only before the loop exit test.
2225 If placed anywhere else, loop may generate bad code. */
2226 /* BASIC_BLOCK notes exist to stabilize basic block structures with
2227 the associated rtl. We do not want to share the structure in
2228 this new block. */
2235 || (last_iteration
2238 else
2244 }
2247 }
2250 /* Now finish coping the REG_NOTES. */
2252 do
2253 {
2258 }
2261 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2262 each of these notes here, since there may be some important ones, such as
2263 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2264 iteration, because the original notes won't be deleted.
2266 We can't use insert_before here, because when from preconditioning,
2267 insert_before points before the loop. We can't use copy_end, because
2268 there may be insns already inserted after it (which we don't want to
2269 copy) when not from preconditioning code. */
2272 {
2274 {
2275 /* VTOP notes are valid only before the loop exit test.
2276 If placed anywhere else, loop may generate bad code.
2277 Although COPY_NOTES_FROM will be at most one or two (for cc0)
2278 instructions before the last insn in the loop, COPY_NOTES_FROM
2279 can be a NOTE_INSN_LOOP_CONT note if there is no VTOP note,
2280 as in a do .. while loop. */
2287 }
2288 }
2296 }
2297 \f
2298 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2299 emitted. This will correctly handle the case where the increment value
2300 won't fit in the immediate field of a PLUS insns. */
2302 void
2304 {
2305 rtx result;
2312 }
2313 \f
2314 /* Searches the insns between INSN and LOOP->END. Returns 1 if there
2315 is a backward branch in that range that branches to somewhere between
2316 LOOP->START and INSN. Returns 0 otherwise. */
2318 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2319 In practice, this is not a problem, because this function is seldom called,
2320 and uses a negligible amount of CPU time on average. */
2322 int
2324 {
2330 /* Stop before we get to the backward branch at the end of the loop. */
2335 /* Check in case insn has been deleted, search forward for first non
2336 deleted insn following it. */
2340 /* Check for the case where insn is the last insn in the loop. Deal
2341 with the case where INSN was a deleted loop test insn, in which case
2342 it will now be the NOTE_LOOP_END. */
2347 {
2349 {
2352 /* Search from loop_start to insn, to see if one of them is
2353 the target_insn. We can't use INSN_LUID comparisons here,
2354 since insn may not have an LUID entry. */
2358 }
2359 }
2362 }
2364 /* Try to generate the simplest rtx for the expression
2365 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2366 value of giv's. */
2368 static rtx
2370 {
2372 rtx result;
2374 /* The modes must all be the same. This should always be true. For now,
2375 check to make sure. */
2381 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2382 will be a constant. */
2384 {
2388 }
2394 /* Again, put the constant second. */
2396 {
2400 }
2407 }
2409 /* Searches the list of induction struct's for the biv BL, to try to calculate
2410 the total increment value for one iteration of the loop as a constant.
2412 Returns the increment value as an rtx, simplified as much as possible,
2413 if it can be calculated. Otherwise, returns 0. */
2415 rtx
2417 {
2419 rtx result;
2421 /* For increment, must check every instruction that sets it. Each
2422 instruction must be executed only once each time through the loop.
2423 To verify this, we check that the insn is always executed, and that
2424 there are no backward branches after the insn that branch to before it.
2425 Also, the insn must have a mult_val of one (to make sure it really is
2426 an increment). */
2430 {
2434 {
2435 /* If we have already counted it, skip it. */
2440 }
2441 else
2443 }
2446 }
2448 /* For each biv and giv, determine whether it can be safely split into
2449 a different variable for each unrolled copy of the loop body. If it
2450 is safe to split, then indicate that by saving some useful info
2451 in the splittable_regs array.
2453 If the loop is being completely unrolled, then splittable_regs will hold
2454 the current value of the induction variable while the loop is unrolled.
2455 It must be set to the initial value of the induction variable here.
2456 Otherwise, splittable_regs will hold the difference between the current
2457 value of the induction variable and the value the induction variable had
2458 at the top of the loop. It must be set to the value 0 here.
2460 Returns the total number of instructions that set registers that are
2461 splittable. */
2463 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2464 constant values are unnecessary, since we can easily calculate increment
2465 values in this case even if nothing is constant. The increment value
2466 should not involve a multiply however. */
2468 /* ?? Even if the biv/giv increment values aren't constant, it may still
2469 be beneficial to split the variable if the loop is only unrolled a few
2470 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2472 static int
2475 {
2480 rtx biv_final_value;
2485 {
2486 /* Biv_total_increment must return a constant value,
2487 otherwise we can not calculate the split values. */
2493 /* The loop must be unrolled completely, or else have a known number
2494 of iterations and only one exit, or else the biv must be dead
2495 outside the loop, or else the final value must be known. Otherwise,
2496 it is unsafe to split the biv since it may not have the proper
2497 value on loop exit. */
2499 /* loop_number_exit_count is nonzero if the loop has an exit other than
2500 a fall through at the end. */
2515 /* If any of the insns setting the BIV don't do so with a simple
2516 PLUS, we don't know how to split it. */
2524 /* If final value is nonzero, then must emit an instruction which sets
2525 the value of the biv to the proper value. This is done after
2526 handling all of the givs, since some of them may need to use the
2527 biv's value in their initialization code. */
2529 /* This biv is splittable. If completely unrolling the loop, save
2530 the biv's initial value. Otherwise, save the constant zero. */
2533 {
2535 {
2536 /* If the initial value of the biv is itself (i.e. it is too
2537 complicated for strength_reduce to compute), or is a hard
2538 register, or it isn't invariant, then we must create a new
2539 pseudo reg to hold the initial value of the biv. */
2545 {
2558 }
2559 else
2561 }
2562 else
2565 /* Save the number of instructions that modify the biv, so that
2566 we can treat the last one specially. */
2574 }
2576 /* Check every giv that depends on this biv to see whether it is
2577 splittable also. Even if the biv isn't splittable, givs which
2578 depend on it may be splittable if the biv is live outside the
2579 loop, and the givs aren't. */
2582 unroll_number);
2584 /* If final value is nonzero, then must emit an instruction which sets
2585 the value of the biv to the proper value. This is done after
2586 handling all of the givs, since some of them may need to use the
2587 biv's value in their initialization code. */
2589 {
2590 /* If the loop has multiple exits, emit the insns before the
2591 loop to ensure that it will always be executed no matter
2592 how the loop exits. Otherwise emit the insn after the loop,
2593 since this is slightly more efficient. */
2596 biv_final_value));
2597 else
2598 {
2599 /* Create a new register to hold the value of the biv, and then
2600 set the biv to its final value before the loop start. The biv
2601 is set to its final value before loop start to ensure that
2602 this insn will always be executed, no matter how the loop
2603 exits. */
2609 biv_final_value));
2615 /* Set up the mapping from the original biv register to the new
2616 register. */
2618 }
2619 }
2620 }
2622 }
2624 /* For every giv based on the biv BL, check to determine whether it is
2625 splittable. This is a subroutine to find_splittable_regs ().
2627 Return the number of instructions that set splittable registers. */
2629 static int
2633 {
2636 rtx final_value;
2637 rtx tem;
2640 /* Scan the list of givs, and set the same_insn field when there are
2641 multiple identical givs in the same insn. */
2649 {
2652 /* Only split the giv if it has already been reduced, or if the loop is
2653 being completely unrolled. */
2657 /* The giv can be split if the insn that sets the giv is executed once
2658 and only once on every iteration of the loop. */
2659 /* An address giv can always be split. v->insn is just a use not a set,
2660 and hence it does not matter whether it is always executed. All that
2661 matters is that all the biv increments are always executed, and we
2662 won't reach here if they aren't. */
2668 /* The giv increment value must be a constant. */
2674 /* The loop must be unrolled completely, or else have a known number of
2675 iterations and only one exit, or else the giv must be dead outside
2676 the loop, or else the final value of the giv must be known.
2677 Otherwise, it is not safe to split the giv since it may not have the
2678 proper value on loop exit. */
2680 /* The used outside loop test will fail for DEST_ADDR givs. They are
2681 never used outside the loop anyways, so it is always safe to split a
2682 DEST_ADDR giv. */
2688 /* The next part is true if the pseudo is used outside the loop.
2689 We assume that this is true for any pseudo created after loop
2690 starts, because we don't have a reg_n_info entry for them. */
2693 /* Check for the case where the pseudo is set by a shift/add
2694 sequence, in which case the first insn setting the pseudo
2695 is the first insn of the shift/add sequence. */
2699 /* Line above always fails if INSN was moved by loop opt. */
2705 #if 0
2706 /* Currently, non-reduced/final-value givs are never split. */
2707 /* Should emit insns after the loop if possible, as the biv final value
2708 code below does. */
2710 /* If the final value is nonzero, and the giv has not been reduced,
2711 then must emit an instruction to set the final value. */
2713 {
2714 /* Create a new register to hold the value of the giv, and then set
2715 the giv to its final value before the loop start. The giv is set
2716 to its final value before loop start to ensure that this insn
2717 will always be executed, no matter how we exit. */
2727 }
2728 #endif
2730 /* This giv is splittable. If completely unrolling the loop, save the
2731 giv's initial value. Otherwise, save the constant zero for it. */
2734 {
2735 /* It is not safe to use bl->initial_value here, because it may not
2736 be invariant. It is safe to use the initial value stored in
2737 the splittable_regs array if it is set. In rare cases, it won't
2738 be set, so then we do exactly the same thing as
2739 find_splittable_regs does to get a safe value. */
2740 rtx biv_initial_value;
2748 else
2749 {
2755 }
2759 }
2760 else
2764 {
2765 /* If a giv was combined with another giv, then we can only split
2766 this giv if the giv it was combined with was reduced. This
2767 is because the value of v->new_reg is meaningless in this
2768 case. */
2770 {
2775 }
2776 /* If the giv is an address destination, it could be something other
2777 than a simple register, these have to be treated differently. */
2779 {
2780 /* If value is not a constant, register, or register plus
2781 constant, then compute its value into a register before
2782 loop start. This prevents invalid rtx sharing, and should
2783 generate better code. We can use bl->initial_value here
2784 instead of splittable_regs[bl->regno] because this code
2785 is going before the loop start. */
2792 {
2799 }
2802 }
2803 else
2805 }
2806 else
2807 {
2808 #if 0
2809 /* Currently, unreduced giv's can't be split. This is not too much
2810 of a problem since unreduced giv's are not live across loop
2811 iterations anyways. When unrolling a loop completely though,
2812 it makes sense to reduce&split givs when possible, as this will
2813 result in simpler instructions, and will not require that a reg
2814 be live across loop iterations. */
2819 #else
2821 #endif
2822 }
2824 /* Unreduced givs are only updated once by definition. Reduced givs
2825 are updated as many times as their biv is. Mark it so if this is
2826 a splittable register. Don't need to do anything for address givs
2827 where this may not be a register. */
2830 {
2836 }
2838 result++;
2841 {
2848 else
2852 }
2853 }
2856 }
2857 \f
2858 /* Try to prove that the register is dead after the loop exits. Trace every
2859 loop exit looking for an insn that will always be executed, which sets
2860 the register to some value, and appears before the first use of the register
2861 is found. If successful, then return 1, otherwise return 0. */
2863 /* ?? Could be made more intelligent in the handling of jumps, so that
2864 it can search past if statements and other similar structures. */
2866 static int
2868 {
2873 /* In addition to checking all exits of this loop, we must also check
2874 all exits of inner nested loops that would exit this loop. We don't
2875 have any way to identify those, so we just give up if there are any
2876 such inner loop exits. */
2879 label_count++;
2884 /* HACK: Must also search the loop fall through exit, create a label_ref
2885 here which points to the loop->end, and append the loop_number_exit_labels
2886 list to it. */
2891 {
2892 /* Succeed if find an insn which sets the biv or if reach end of
2893 function. Fail if find an insn that uses the biv, or if come to
2894 a conditional jump. */
2898 {
2900 {
2915 {
2919 /* Prevent infinite loop following infinite loops. */
2922 else
2924 }
2925 }
2928 }
2929 }
2931 /* Success, the register is dead on all loop exits. */
2933 }
2935 /* Try to calculate the final value of the biv, the value it will have at
2936 the end of the loop. If we can do it, return that value. */
2938 rtx
2940 {
2944 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2949 /* The final value for reversed bivs must be calculated differently than
2950 for ordinary bivs. In this case, there is already an insn after the
2951 loop which sets this biv's final value (if necessary), and there are
2952 no other loop exits, so we can return any value. */
2954 {
2960 }
2962 /* Try to calculate the final value as initial value + (number of iterations
2963 * increment). For this to work, increment must be invariant, the only
2964 exit from the loop must be the fall through at the bottom (otherwise
2965 it may not have its final value when the loop exits), and the initial
2966 value of the biv must be invariant. */
2971 {
2975 {
2976 /* Can calculate the loop exit value, emit insns after loop
2977 end to calculate this value into a temporary register in
2978 case it is needed later. */
2990 }
2991 }
2993 /* Check to see if the biv is dead at all loop exits. */
2995 {
3002 }
3005 }
3007 /* Try to calculate the final value of the giv, the value it will have at
3008 the end of the loop. If we can do it, return that value. */
3010 rtx
3012 {
3015 rtx insn;
3017 rtx seq;
3023 /* The final value for givs which depend on reversed bivs must be calculated
3024 differently than for ordinary givs. In this case, there is already an
3025 insn after the loop which sets this giv's final value (if necessary),
3026 and there are no other loop exits, so we can return any value. */
3028 {
3034 }
3036 /* Try to calculate the final value as a function of the biv it depends
3037 upon. The only exit from the loop must be the fall through at the bottom
3038 and the insn that sets the giv must be executed on every iteration
3039 (otherwise the giv may not have its final value when the loop exits). */
3041 /* ??? Can calculate the final giv value by subtracting off the
3042 extra biv increments times the giv's mult_val. The loop must have
3043 only one exit for this to work, but the loop iterations does not need
3044 to be known. */
3049 {
3050 /* ?? It is tempting to use the biv's value here since these insns will
3051 be put after the loop, and hence the biv will have its final value
3052 then. However, this fails if the biv is subsequently eliminated.
3053 Perhaps determine whether biv's are eliminable before trying to
3054 determine whether giv's are replaceable so that we can use the
3055 biv value here if it is not eliminable. */
3057 /* We are emitting code after the end of the loop, so we must make
3058 sure that bl->initial_value is still valid then. It will still
3059 be valid if it is invariant. */
3065 {
3066 /* Can calculate the loop exit value of its biv as
3067 (n_iterations * increment) + initial_value */
3069 /* The loop exit value of the giv is then
3070 (final_biv_value - extra increments) * mult_val + add_val.
3071 The extra increments are any increments to the biv which
3072 occur in the loop after the giv's value is calculated.
3073 We must search from the insn that sets the giv to the end
3074 of the loop to calculate this value. */
3076 /* Put the final biv value in tem. */
3082 tem);
3084 /* Subtract off extra increments as we find them. */
3087 {
3092 {
3096 OPTAB_LIB_WIDEN);
3100 }
3101 }
3103 /* Now calculate the giv's final value. */
3112 }
3113 }
3115 /* Replaceable giv's should never reach here. */
3119 /* Check to see if the biv is dead at all loop exits. */
3121 {
3128 }
3131 }
3133 /* Look back before LOOP->START for the insn that sets REG and return
3134 the equivalent constant if there is a REG_EQUAL note otherwise just
3135 the SET_SRC of REG. */
3137 static rtx
3139 {
3142 rtx ret;
3146 {
3151 {
3152 /* We found the last insn before the loop that sets the register.
3153 If it sets the entire register, and has a REG_EQUAL note,
3154 then use the value of the REG_EQUAL note. */
3157 {
3160 /* Only use the REG_EQUAL note if it is a constant.
3161 Other things, divide in particular, will cause
3162 problems later if we use them. */
3166 else
3169 /* We cannot do this if it changes between the
3170 assignment and loop start though. */
3173 }
3175 }
3176 }
3178 }
3180 /* Return a simplified rtx for the expression OP - REG.
3182 REG must appear in OP, and OP must be a register or the sum of a register
3183 and a second term.
3185 Thus, the return value must be const0_rtx or the second term.
3187 The caller is responsible for verifying that REG appears in OP and OP has
3188 the proper form. */
3190 static rtx
3192 {
3196 {
3201 }
3202 /* OP does not contain REG as a term. */
3204 }
3206 /* Find and return register term common to both expressions OP0 and
3207 OP1 or NULL_RTX if no such term exists. Each expression must be a
3208 REG or a PLUS of a REG. */
3210 static rtx
3212 {
3215 {
3216 rtx op00;
3217 rtx op01;
3218 rtx op10;
3219 rtx op11;
3223 else
3228 else
3231 /* Find and return common register term if present. */
3236 }
3238 /* No common register term found. */
3240 }
3242 /* Determine the loop iterator and calculate the number of loop
3243 iterations. Returns the exact number of loop iterations if it can
3244 be calculated, otherwise returns zero. */
3246 unsigned HOST_WIDE_INT
3248 {
3254 HOST_WIDE_INT inc;
3260 rtx last_loop_insn;
3261 rtx reg_term;
3275 /* We used to use prev_nonnote_insn here, but that fails because it might
3276 accidentally get the branch for a contained loop if the branch for this
3277 loop was deleted. We can only trust branches immediately before the
3278 loop_end. */
3281 /* ??? We should probably try harder to find the jump insn
3282 at the end of the loop. The following code assumes that
3283 the last loop insn is a jump to the top of the loop. */
3285 {
3290 }
3292 /* If there is a more than a single jump to the top of the loop
3293 we cannot (easily) determine the iteration count. */
3295 {
3300 }
3302 /* If there are multiple conditionalized loop exit tests, they may jump
3303 back to differing CODE_LABELs. */
3305 {
3308 do
3309 {
3311 {
3312 /* There are some kinds of jumps we can't deal with easily. */
3314 {
3316 fprintf
3320 }
3323 covered by the uid_luid array. */
3325 /* Check if we jump back into the loop body. */
3328 {
3330 fprintf
3334 }
3335 }
3336 }
3338 }
3340 /* Find the iteration variable. If the last insn is a conditional
3341 branch, and the insn before tests a register value, make that the
3342 iteration variable. */
3346 {
3351 }
3353 /* ??? Get_condition may switch position of induction variable and
3354 invariant register when it canonicalizes the comparison. */
3361 {
3366 }
3368 /* The only new registers that are created before loop iterations
3369 are givs made from biv increments or registers created by
3370 load_mems. In the latter case, it is possible that try_copy_prop
3371 will propagate a new pseudo into the old iteration register but
3372 this will be marked by having the REG_USERVAR_P bit set. */
3378 /* Determine the initial value of the iteration variable, and the amount
3379 that it is incremented each loop. Use the tables constructed by
3380 the strength reduction pass to calculate these values. */
3382 /* Clear the result values, in case no answer can be found. */
3386 /* The iteration variable can be either a giv or a biv. Check to see
3387 which it is, and compute the variable's initial value, and increment
3388 value if possible. */
3390 /* If this is a new register, can't handle it since we don't have any
3391 reg_iv_type entry for it. */
3393 {
3398 }
3400 /* Reject iteration variables larger than the host wide int size, since they
3401 could result in a number of iterations greater than the range of our
3402 `unsigned HOST_WIDE_INT' variable loop_info->n_iterations. */
3405 {
3410 }
3412 {
3417 }
3419 /* Try swapping the comparison to identify a suitable iv. */
3424 {
3429 }
3432 {
3436 /* Grab initial value, only useful if it is a constant. */
3440 {
3445 }
3448 }
3450 {
3453 rtx biv_initial_value;
3459 {
3464 }
3468 /* Increment value is mult_val times the increment value of the biv. */
3472 {
3478 /* The caller assumes that one full increment has occurred at the
3479 first loop test. But that's not true when the biv is incremented
3480 after the giv is set (which is the usual case), e.g.:
3481 i = 6; do {;} while (i++ < 9) .
3482 Therefore, we bias the initial value by subtracting the amount of
3483 the increment that occurs between the giv set and the giv test. */
3485 {
3487 {
3489 {
3495 }
3497 /* If we have already counted it, skip it. */
3502 }
3503 }
3504 }
3510 /* Initial value is mult_val times the biv's initial value plus
3511 add_val. Only useful if it is a constant. */
3513 initial_value
3517 }
3518 else
3519 {
3524 }
3532 {
3546 /* Cannot determine loop iterations with this case. */
3564 }
3566 /* If the comparison value is an invariant register, then try to find
3567 its value from the insns before the start of the loop. */
3572 {
3575 /* If we don't get an invariant final value, we are better
3576 off with the original register. */
3579 }
3581 /* Calculate the approximate final value of the induction variable
3582 (on the last successful iteration). The exact final value
3583 depends on the branch operator, and increment sign. It will be
3584 wrong if the iteration variable is not incremented by one each
3585 time through the loop and (comparison_value + off_by_one -
3586 initial_value) % increment != 0.
3587 ??? Note that the final_value may overflow and thus final_larger
3588 will be bogus. A potentially infinite loop will be classified
3589 as immediate, e.g. for (i = 0x7ffffff0; i <= 0x7fffffff; i++) */
3593 /* Save the calculated values describing this loop's bounds, in case
3594 precondition_loop_p will need them later. These values can not be
3595 recalculated inside precondition_loop_p because strength reduction
3596 optimizations may obscure the loop's structure.
3598 These values are only required by precondition_loop_p and insert_bct
3599 whenever the number of iterations cannot be computed at compile time.
3600 Only the difference between final_value and initial_value is
3601 important. Note that final_value is only approximate. */
3610 /* Try to determine the iteration count for loops such
3611 as (for i = init; i < init + const; i++). When running the
3612 loop optimization twice, the first pass often converts simple
3613 loops into this form. */
3616 {
3617 rtx reg1;
3618 rtx reg2;
3619 rtx const2;
3624 else
3627 /* Check for initial_value = reg1, final_value = reg2 + const2,
3628 where reg1 != reg2. */
3630 {
3631 rtx temp;
3633 /* Find what reg1 is equivalent to. Hopefully it will
3634 either be reg2 or reg2 plus a constant. */
3640 {
3641 /* Find what reg2 is equivalent to. Hopefully it will
3642 either be reg1 or reg1 plus a constant. Let's ignore
3643 the latter case for now since it is not so common. */
3651 }
3652 }
3654 {
3655 rtx temp;
3657 /* When running the loop optimizer twice, check_dbra_loop
3658 further obfuscates reversible loops of the form:
3659 for (i = init; i < init + const; i++). We often end up with
3660 final_value = 0, initial_value = temp, temp = temp2 - init,
3661 where temp2 = init + const. If the loop has a vtop we
3662 can replace initial_value with const. */
3667 {
3673 }
3674 }
3675 }
3677 /* If have initial_value = reg + const1 and final_value = reg +
3678 const2, then replace initial_value with const1 and final_value
3679 with const2. This should be safe since we are protected by the
3680 initial comparison before entering the loop if we have a vtop.
3681 For example, a + b < a + c is not equivalent to b < c for all a
3682 when using modulo arithmetic.
3684 ??? Without a vtop we could still perform the optimization if we check
3685 the initial and final values carefully. */
3688 {
3691 }
3696 /* For EQ comparison loops, we don't have a valid final value.
3697 Check this now so that we won't leave an invalid value if we
3698 return early for any other reason. */
3703 {
3708 }
3711 {
3712 /* If we have a REG, check to see if REG holds a constant value. */
3713 /* ??? Other RTL, such as (neg (reg)) is possible here, but it isn't
3714 clear if it is worthwhile to try to handle such RTL. */
3719 {
3721 {
3726 }
3728 }
3730 }
3733 {
3735 {
3740 }
3742 }
3744 {
3746 {
3751 }
3753 }
3755 {
3756 rtx inc_once;
3765 {
3766 /* The iterator value once through the loop is equal to the
3767 comparison value. Either we have an infinite loop, or
3768 we'll loop twice. */
3772 }
3773 else
3777 loop_info->final_value
3781 else
3782 loop_info->final_value
3785 loop_info->final_equiv_value
3790 }
3792 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3794 final_larger
3799 else
3807 else
3810 /* There are 27 different cases: compare_dir = -1, 0, 1;
3811 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3812 There are 4 normal cases, 4 reverse cases (where the iteration variable
3813 will overflow before the loop exits), 4 infinite loop cases, and 15
3814 immediate exit (0 or 1 iteration depending on loop type) cases.
3815 Only try to optimize the normal cases. */
3817 /* (compare_dir/final_larger/increment_dir)
3818 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3819 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3820 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3821 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3823 /* ?? If the meaning of reverse loops (where the iteration variable
3824 will overflow before the loop exits) is undefined, then could
3825 eliminate all of these special checks, and just always assume
3826 the loops are normal/immediate/infinite. Note that this means
3827 the sign of increment_dir does not have to be known. Also,
3828 since it does not really hurt if immediate exit loops or infinite loops
3829 are optimized, then that case could be ignored also, and hence all
3830 loops can be optimized.
3832 According to ANSI Spec, the reverse loop case result is undefined,
3833 because the action on overflow is undefined.
3835 See also the special test for NE loops below. */
3839 /* Normal case. */
3840 ;
3841 else
3842 {
3846 }
3848 /* Calculate the number of iterations, final_value is only an approximation,
3849 so correct for that. Note that abs_diff and n_iterations are
3850 unsigned, because they can be as large as 2^n - 1. */
3854 {
3857 }
3859 {
3862 }
3863 else
3866 /* Given that iteration_var is going to iterate over its own mode,
3867 not HOST_WIDE_INT, disregard higher bits that might have come
3868 into the picture due to sign extension of initial and final
3869 values. */
3874 /* For NE tests, make sure that the iteration variable won't miss
3875 the final value. If abs_diff mod abs_incr is not zero, then the
3876 iteration variable will overflow before the loop exits, and we
3877 can not calculate the number of iterations. */
3881 /* Note that the number of iterations could be calculated using
3882 (abs_diff + abs_inc - 1) / abs_inc, provided care was taken to
3883 handle potential overflow of the summation. */
3886 }
3888 /* Replace uses of split bivs with their split pseudo register. This is
3889 for original instructions which remain after loop unrolling without
3890 copying. */
3892 static rtx
3894 {
3905 {
3917 #if 0
3918 /* If non-reduced/final-value givs were split, then this would also
3919 have to remap those givs also. */
3920 #endif
3928 }
3932 {
3936 {
3940 }
3941 }
3943 }
3945 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3946 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3947 return 0. COPY_START is where we can start looking for the insns
3948 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3949 insns.
3951 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3952 must dominate LAST_UID.
3954 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3955 may not dominate LAST_UID.
3957 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3958 must dominate LAST_UID. */
3960 int
3962 rtx copy_end)
3963 {
3968 {
3971 /* Could not find FIRST_UID. */
3975 }
3977 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3981 /* FIRST_UID is always executed. */
3986 {
3987 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3988 can not be sure that FIRST_UID dominates LAST_UID. */
3991 /* Could not find LAST_UID, but we reached the end of the loop, so
3992 it must be safe. */
3996 }
3998 /* FIRST_UID is always executed if LAST_UID is executed. */
4000 }
4002 /* This routine is called when the number of iterations for the unrolled
4003 loop is one. The goal is to identify a loop that begins with an
4004 unconditional branch to the loop continuation note (or a label just after).
4005 In this case, the unconditional branch that starts the loop needs to be
4006 deleted so that we execute the single iteration. */
4008 static rtx
4010 {
4013 /* See if loop start, or the next insn is an unconditional jump. */
4024 /* Examine insn after loop continuation note. Return if not a label. */
4029 /* Return the loop start if the branch label matches the code label. */
4032 else
4034 }