| blob | c6de53b663a2f1cf466c8e16b8b99a74b8e9f484 |
1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 93, 94, 95, 97, 1998 Free Software Foundation, Inc.
3 Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* Try to unroll a loop, and split induction variables.
24 Loops for which the number of iterations can be calculated exactly are
25 handled specially. If the number of iterations times the insn_count is
26 less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
27 Otherwise, we try to unroll the loop a number of times modulo the number
28 of iterations, so that only one exit test will be needed. It is unrolled
29 a number of times approximately equal to MAX_UNROLLED_INSNS divided by
30 the insn count.
32 Otherwise, if the number of iterations can be calculated exactly at
33 run time, and the loop is always entered at the top, then we try to
34 precondition the loop. That is, at run time, calculate how many times
35 the loop will execute, and then execute the loop body a few times so
36 that the remaining iterations will be some multiple of 4 (or 2 if the
37 loop is large). Then fall through to a loop unrolled 4 (or 2) times,
38 with only one exit test needed at the end of the loop.
40 Otherwise, if the number of iterations can not be calculated exactly,
41 not even at run time, then we still unroll the loop a number of times
42 approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
43 but there must be an exit test after each copy of the loop body.
45 For each induction variable, which is dead outside the loop (replaceable)
46 or for which we can easily calculate the final value, if we can easily
47 calculate its value at each place where it is set as a function of the
48 current loop unroll count and the variable's value at loop entry, then
49 the induction variable is split into `N' different variables, one for
50 each copy of the loop body. One variable is live across the backward
51 branch, and the others are all calculated as a function of this variable.
52 This helps eliminate data dependencies, and leads to further opportunities
53 for cse. */
55 /* Possible improvements follow: */
57 /* ??? Add an extra pass somewhere to determine whether unrolling will
58 give any benefit. E.g. after generating all unrolled insns, compute the
59 cost of all insns and compare against cost of insns in rolled loop.
61 - On traditional architectures, unrolling a non-constant bound loop
62 is a win if there is a giv whose only use is in memory addresses, the
63 memory addresses can be split, and hence giv increments can be
64 eliminated.
65 - It is also a win if the loop is executed many times, and preconditioning
66 can be performed for the loop.
67 Add code to check for these and similar cases. */
69 /* ??? Improve control of which loops get unrolled. Could use profiling
70 info to only unroll the most commonly executed loops. Perhaps have
71 a user specifyable option to control the amount of code expansion,
72 or the percent of loops to consider for unrolling. Etc. */
74 /* ??? Look at the register copies inside the loop to see if they form a
75 simple permutation. If so, iterate the permutation until it gets back to
76 the start state. This is how many times we should unroll the loop, for
77 best results, because then all register copies can be eliminated.
78 For example, the lisp nreverse function should be unrolled 3 times
79 while (this)
80 {
81 next = this->cdr;
82 this->cdr = prev;
83 prev = this;
84 this = next;
85 }
87 ??? The number of times to unroll the loop may also be based on data
88 references in the loop. For example, if we have a loop that references
89 x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
91 /* ??? Add some simple linear equation solving capability so that we can
92 determine the number of loop iterations for more complex loops.
93 For example, consider this loop from gdb
94 #define SWAP_TARGET_AND_HOST(buffer,len)
95 {
96 char tmp;
97 char *p = (char *) buffer;
98 char *q = ((char *) buffer) + len - 1;
99 int iterations = (len + 1) >> 1;
100 int i;
101 for (p; p < q; p++, q--;)
102 {
103 tmp = *q;
104 *q = *p;
105 *p = tmp;
106 }
107 }
108 Note that:
109 start value = p = &buffer + current_iteration
110 end value = q = &buffer + len - 1 - current_iteration
111 Given the loop exit test of "p < q", then there must be "q - p" iterations,
112 set equal to zero and solve for number of iterations:
113 q - p = len - 1 - 2*current_iteration = 0
114 current_iteration = (len - 1) / 2
115 Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
116 iterations of this loop. */
118 /* ??? Currently, no labels are marked as loop invariant when doing loop
119 unrolling. This is because an insn inside the loop, that loads the address
120 of a label inside the loop into a register, could be moved outside the loop
121 by the invariant code motion pass if labels were invariant. If the loop
122 is subsequently unrolled, the code will be wrong because each unrolled
123 body of the loop will use the same address, whereas each actually needs a
124 different address. A case where this happens is when a loop containing
125 a switch statement is unrolled.
127 It would be better to let labels be considered invariant. When we
128 unroll loops here, check to see if any insns using a label local to the
129 loop were moved before the loop. If so, then correct the problem, by
130 moving the insn back into the loop, or perhaps replicate the insn before
131 the loop, one copy for each time the loop is unrolled. */
133 /* The prime factors looked for when trying to unroll a loop by some
134 number which is modulo the total number of iterations. Just checking
135 for these 4 prime factors will find at least one factor for 75% of
136 all numbers theoretically. Practically speaking, this will succeed
137 almost all of the time since loops are generally a multiple of 2
138 and/or 5. */
140 #define NUM_FACTORS 4
145 /* Describes the different types of loop unrolling performed. */
161 /* This controls which loops are unrolled, and by how much we unroll
162 them. */
164 #ifndef MAX_UNROLLED_INSNS
165 #define MAX_UNROLLED_INSNS 100
166 #endif
168 /* Indexed by register number, if non-zero, then it contains a pointer
169 to a struct induction for a DEST_REG giv which has been combined with
170 one of more address givs. This is needed because whenever such a DEST_REG
171 giv is modified, we must modify the value of all split address givs
172 that were combined with this DEST_REG giv. */
176 /* Indexed by register number, if this is a splittable induction variable,
177 then this will hold the current value of the register, which depends on the
178 iteration number. */
182 /* Indexed by register number, if this is a splittable induction variable,
183 then this will hold the number of instructions in the loop that modify
184 the induction variable. Used to ensure that only the last insn modifying
185 a split iv will update the original iv of the dest. */
189 /* Values describing the current loop's iteration variable. These are set up
190 by loop_iterations, and used by precondition_loop_p. */
198 /* Forward declarations. */
217 /* Try to unroll one loop and split induction variables in the loop.
219 The loop is described by the arguments LOOP_END, INSN_COUNT, and
220 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
221 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
222 indicates whether information generated in the strength reduction pass
223 is available.
225 This function is intended to be called from within `strength_reduce'
226 in loop.c. */
228 void
230 strength_reduce_p)
231 rtx loop_end;
233 rtx loop_start;
234 rtx end_insert_before;
236 {
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;
260 /* Don't bother unrolling huge loops. Since the minimum factor is
261 two, loops greater than one half of MAX_UNROLLED_INSNS will never
262 be unrolled. */
264 {
268 }
270 /* When emitting debugger info, we can't unroll loops with unequal numbers
271 of block_beg and block_end notes, because that would unbalance the block
272 structure of the function. This can happen as a result of the
273 "if (foo) bar; else break;" optimization in jump.c. */
274 /* ??? Gcc has a general policy that -g is never supposed to change the code
275 that the compiler emits, so we must disable this optimization always,
276 even if debug info is not being output. This is rare, so this should
277 not be a significant performance problem. */
280 {
285 {
287 {
289 block_begins++;
291 block_ends++;
292 }
293 }
296 {
301 }
302 }
304 /* Determine type of unroll to perform. Depends on the number of iterations
305 and the size of the loop. */
307 /* If there is no strength reduce info, then set loop_n_iterations to zero.
308 This can happen if strength_reduce can't find any bivs in the loop.
309 A value of zero indicates that the number of iterations could not be
310 calculated. */
316 {
320 }
322 /* Find and save a pointer to the last nonnote insn in the loop. */
326 /* Calculate how many times to unroll the loop. Indicate whether or
327 not the loop is being completely unrolled. */
330 {
331 /* If number of iterations is exactly 1, then eliminate the compare and
332 branch at the end of the loop since they will never be taken.
333 Then return, since no other action is needed here. */
335 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
336 don't do anything. */
339 {
340 /* Delete the jump insn. This will delete the barrier also. */
342 }
344 {
345 #ifdef HAVE_cc0
346 /* The immediately preceding insn is a compare which must be
347 deleted. */
350 #else
351 /* The immediately preceding insn may not be the compare, so don't
352 delete it. */
354 #endif
355 }
357 }
360 {
363 }
365 {
366 /* Try to factor the number of iterations. Don't bother with the
367 general case, only using 2, 3, 5, and 7 will get 75% of all
368 numbers theoretically, and almost all in practice. */
376 {
379 }
381 /* Start with the larger factors first so that we generally
382 get lots of unrolling. */
388 {
390 {
393 }
394 else
396 }
398 /* If we couldn't find any factors, then unroll as in the normal
399 case. */
401 {
405 }
406 else
408 }
411 /* Default case, calculate number of times to unroll loop based on its
412 size. */
414 {
419 else
423 }
425 /* Now we know how many times to unroll the loop. */
433 {
434 /* Loops of these types should never start with a jump down to
435 the exit condition test. For now, check for this case just to
436 be sure. UNROLL_NAIVE loops can be of this form, this case is
437 handled below. */
443 }
446 {
447 /* Completely unrolling the loop: Delete the compare and branch at
448 the end (the last two instructions). This delete must done at the
449 very end of loop unrolling, to avoid problems with calls to
450 back_branch_in_range_p, which is called by find_splittable_regs.
451 All increments of splittable bivs/givs are changed to load constant
452 instructions. */
456 /* Set insert_before to the instruction immediately after the JUMP_INSN
457 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
458 the loop will be correctly handled by copy_loop_body. */
461 /* Set copy_end to the insn before the jump at the end of the loop. */
465 {
466 #ifdef HAVE_cc0
467 /* The instruction immediately before the JUMP_INSN is a compare
468 instruction which we do not want to copy. */
470 #else
471 /* The instruction immediately before the JUMP_INSN may not be the
472 compare, so we must copy it. */
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 {
507 #ifdef HAVE_cc0
508 /* The instruction immediately before the JUMP_INSN is a compare
509 instruction which we do not want to copy or delete. */
512 #else
513 /* The instruction immediately before the JUMP_INSN may not be the
514 compare, so we must copy it. */
517 #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 }
638 {
639 /* In this case, we must copy the jump and barrier, because they will
640 not be converted to jumps to an immediately following label. */
644 }
649 {
650 /* ??? The loop ends with a conditional branch that does not branch back
651 to the loop start label. In this case, we must emit an unconditional
652 branch to the loop exit after emitting the final branch.
653 copy_loop_body does not have support for this currently, so we
654 give up. It doesn't seem worthwhile to unroll anyways since
655 unrolling would increase the number of branch instructions
656 executed. */
661 }
663 /* Allocate a translation table for the labels and insn numbers.
664 They will be filled in as we copy the insns in the loop. */
673 /* Allocate the label map. */
676 {
681 }
682 else
685 /* Search the loop and mark all local labels, i.e. the ones which have to
686 be distinct labels when copied. For all labels which might be
687 non-local, set their label_map entries to point to themselves.
688 If they happen to be local their label_map entries will be overwritten
689 before the loop body is copied. The label_map entries for local labels
690 will be set to a different value each time the loop body is copied. */
693 {
694 rtx note;
699 {
706 {
710 rtx label;
713 {
717 label);
718 }
719 }
720 }
724 }
726 /* Allocate space for the insn map. */
730 /* Set this to zero, to indicate that we are doing loop unrolling,
731 not function inlining. */
734 /* The register and constant maps depend on the number of registers
735 present, so the final maps can't be created until after
736 find_splittable_regs is called. However, they are needed for
737 preconditioning, so we create temporary maps when preconditioning
738 is performed. */
740 /* The preconditioning code may allocate two new pseudo registers. */
743 /* Allocate and zero out the splittable_regs and addr_combined_regs
744 arrays. These must be zeroed here because they will be used if
745 loop preconditioning is performed, and must be zero for that case.
747 It is safe to do this here, since the extra registers created by the
748 preconditioning code and find_splittable_regs will never be used
749 to access the splittable_regs[] and addr_combined_regs[] arrays. */
755 addr_combined_regs
758 /* We must limit it to max_reg_before_loop, because only these pseudo
759 registers have valid regno_first_uid info. Any register created after
760 that is unlikely to be local to the loop anyways. */
764 /* Mark all local registers, i.e. the ones which are referenced only
765 inside the loop. */
767 {
771 /* If a register is used in the jump insn, we must not duplicate it
772 since it will also be used outside the loop. */
774 copy_end_luid--;
775 /* If copy_start points to the NOTE that starts the loop, then we must
776 use the next luid, because invariant pseudo-regs moved out of the loop
777 have their lifetimes modified to start here, but they are not safe
778 to duplicate. */
780 copy_start_luid++;
782 /* If a pseudo's lifetime is entirely contained within this loop, then we
783 can use a different pseudo in each unrolled copy of the loop. This
784 results in better code. */
790 {
791 /* However, we must also check for loop-carried dependencies.
792 If the value the pseudo has at the end of iteration X is
793 used by iteration X+1, then we can not use a different pseudo
794 for each unrolled copy of the loop. */
795 /* A pseudo is safe if regno_first_uid is a set, and this
796 set dominates all instructions from regno_first_uid to
797 regno_last_uid. */
798 /* ??? This check is simplistic. We would get better code if
799 this check was more sophisticated. */
805 {
808 else
810 j);
811 }
812 }
813 }
815 /* If this loop requires exit tests when unrolled, check to see if we
816 can precondition the loop so as to make the exit tests unnecessary.
817 Just like variable splitting, this is not safe if the loop is entered
818 via a jump to the bottom. Also, can not do this if no strength
819 reduce info, because precondition_loop_p uses this info. */
821 /* Must copy the loop body for preconditioning before the following
822 find_splittable_regs call since that will emit insns which need to
823 be after the preconditioned loop copies, but immediately before the
824 unrolled loop copies. */
826 /* Also, it is not safe to split induction variables for the preconditioned
827 copies of the loop body. If we split induction variables, then the code
828 assumes that each induction variable can be represented as a function
829 of its initial value and the loop iteration number. This is not true
830 in this case, because the last preconditioned copy of the loop body
831 could be any iteration from the first up to the `unroll_number-1'th,
832 depending on the initial value of the iteration variable. Therefore
833 we can not split induction variables here, because we can not calculate
834 their value. Hence, this code must occur before find_splittable_regs
835 is called. */
838 {
843 {
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 /* Decide what mode to do these calculations in. Choose the larger
878 of final_value's mode and initial_value's mode, or a full-word if
879 both are constants. */
882 {
886 }
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. */
902 OPTAB_LIB_WIDEN);
904 /* Now calculate (diff % (unroll * abs (increment))) by using an
905 and instruction. */
910 /* Now emit a sequence of branches to jump to the proper precond
911 loop entry point. */
917 /* Check for the case where the initial value is greater than or
918 equal to the final value. In that case, we want to execute
919 exactly one loop iteration. The code below will fail for this
920 case. This check does not apply if the loop has a NE
921 comparison at the end. */
924 {
929 else
933 }
935 /* Assuming the unroll_number is 4, and the increment is 2, then
936 for a negative increment: for a positive increment:
937 diff = 0,1 precond 0 diff = 0,7 precond 0
938 diff = 2,3 precond 3 diff = 1,2 precond 1
939 diff = 4,5 precond 2 diff = 3,4 precond 2
940 diff = 6,7 precond 1 diff = 5,6 precond 3 */
942 /* We only need to emit (unroll_number - 1) branches here, the
943 last case just falls through to the following code. */
945 /* ??? This would give better code if we emitted a tree of branches
946 instead of the current linear list of branches. */
949 {
953 /* For negative increments, must invert the constant compared
954 against, except when comparing against zero. */
956 {
959 }
961 {
964 }
965 else
966 {
969 }
978 else
982 }
984 /* If the increment is greater than one, then we need another branch,
985 to handle other cases equivalent to 0. */
987 /* ??? This should be merged into the code above somehow to help
988 simplify the code here, and reduce the number of branches emitted.
989 For the negative increment case, the branch here could easily
990 be merged with the `0' case branch above. For the positive
991 increment case, it is not clear how this can be simplified. */
994 {
999 {
1002 }
1003 else
1004 {
1007 }
1014 else
1018 }
1024 /* Only the last copy of the loop body here needs the exit
1025 test, so set copy_end to exclude the compare/branch here,
1026 and then reset it inside the loop when get to the last
1027 copy. */
1032 {
1033 #ifdef HAVE_cc0
1034 /* The immediately preceding insn is a compare which we do not
1035 want to copy. */
1037 #else
1038 /* The immediately preceding insn may not be a compare, so we
1039 must copy it. */
1041 #endif
1042 }
1043 else
1047 {
1063 {
1067 }
1068 /* The last copy needs the compare/branch insns at the end,
1069 so reset copy_end here if the loop ends with a conditional
1070 branch. */
1073 {
1076 else
1078 }
1080 /* None of the copies are the `last_iteration', so just
1081 pass zero for that parameter. */
1085 }
1089 {
1092 }
1093 else
1094 {
1095 #ifdef HAVE_cc0
1096 /* The immediately preceding insn is a compare which we do not
1097 want to copy. */
1100 #else
1101 /* The immediately preceding insn may not be a compare, so we
1102 must copy it. */
1105 #endif
1106 }
1108 /* Set unroll type to MODULO now. */
1112 #ifdef HAIFA
1113 /* Fix the initial value for the loop as needed. */
1117 #endif
1118 }
1119 }
1121 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1122 the loop unless all loops are being unrolled. */
1124 {
1128 }
1130 /* At this point, we are guaranteed to unroll the loop. */
1132 /* Keep track of the unroll factor for each loop. */
1135 else
1139 /* For each biv and giv, determine whether it can be safely split into
1140 a different variable for each unrolled copy of the loop body.
1141 We precalculate and save this info here, since computing it is
1142 expensive.
1144 Do this before deleting any instructions from the loop, so that
1145 back_branch_in_range_p will work correctly. */
1149 else
1153 /* find_splittable_regs may have created some new registers, so must
1154 reallocate the reg_map with the new larger size, and must realloc
1155 the constant maps also. */
1162 /* Space is needed in some of the map for new registers, so new_maxregnum
1163 is an (over)estimate of how many registers will exist at the end. */
1166 /* Must realloc space for the constant maps, because the number of registers
1167 may have changed. */
1176 /* Search the list of bivs and givs to find ones which need to be remapped
1177 when split, and set their reg_map entry appropriately. */
1180 {
1183 #if 0
1184 /* Currently, non-reduced/final-value givs are never split. */
1188 #endif
1189 }
1191 /* Use our current register alignment and pointer flags. */
1195 /* If the loop is being partially unrolled, and the iteration variables
1196 are being split, and are being renamed for the split, then must fix up
1197 the compare/jump instruction at the end of the loop to refer to the new
1198 registers. This compare isn't copied, so the registers used in it
1199 will never be replaced if it isn't done here. */
1202 {
1206 }
1208 /* For unroll_number - 1 times, make a copy of each instruction
1209 between copy_start and copy_end, and insert these new instructions
1210 before the end of the loop. */
1213 {
1225 {
1229 }
1231 /* If loop starts with a branch to the test, then fix it so that
1232 it points to the test of the first unrolled copy of the loop. */
1234 {
1239 CODE_LABEL_NUMBER
1243 /* Set the jump label so that it can be used by later loop unrolling
1244 passes. */
1247 }
1252 }
1254 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1255 insn to be deleted. This prevents any runaway delete_insn call from
1256 more insns that it should, as it always stops at a CODE_LABEL. */
1258 /* Delete the compare and branch at the end of the loop if completely
1259 unrolling the loop. Deleting the backward branch at the end also
1260 deletes the code label at the start of the loop. This is done at
1261 the very end to avoid problems with back_branch_in_range_p. */
1265 else
1268 /* Delete all of the original loop instructions. Don't delete the
1269 LOOP_BEG note, or the first code label in the loop. */
1273 {
1276 else
1278 }
1280 /* Can now delete the 'safety' label emitted to protect us from runaway
1281 delete_insn calls. */
1286 /* If exit_label exists, emit it after the loop. Doing the emit here
1287 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1288 This is needed so that mostly_true_jump in reorg.c will treat jumps
1289 to this loop end label correctly, i.e. predict that they are usually
1290 not taken. */
1293 }
1294 \f
1295 /* Return true if the loop can be safely, and profitably, preconditioned
1296 so that the unrolled copies of the loop body don't need exit tests.
1298 This only works if final_value, initial_value and increment can be
1299 determined, and if increment is a constant power of 2.
1300 If increment is not a power of 2, then the preconditioning modulo
1301 operation would require a real modulo instead of a boolean AND, and this
1302 is not considered `profitable'. */
1304 /* ??? If the loop is known to be executed very many times, or the machine
1305 has a very cheap divide instruction, then preconditioning is a win even
1306 when the increment is not a power of 2. Use RTX_COST to compute
1307 whether divide is cheap. */
1309 static int
1311 loop_end)
1314 {
1317 {
1323 {
1325 loop_dump_stream);
1327 loop_n_iterations);
1329 }
1331 }
1334 {
1339 }
1341 {
1346 }
1348 {
1353 }
1356 {
1361 }
1363 /* Unsigned_compare and compare_dir can be ignored here, since they do
1364 not matter for preconditioning. */
1367 {
1372 }
1374 /* Must ensure that final_value is invariant, so call invariant_p to
1375 check. Before doing so, must check regno against max_reg_before_loop
1376 to make sure that the register is in the range covered by invariant_p.
1377 If it isn't, then it is most likely a biv/giv which by definition are
1378 not invariant. */
1384 {
1389 }
1391 /* Fail for floating point values, since the caller of this function
1392 does not have code to deal with them. */
1395 {
1400 }
1402 /* Now set initial_value to be the iteration_var, since that may be a
1403 simpler expression, and is guaranteed to be correct if all of the
1404 above tests succeed.
1406 We can not use the initial_value as calculated, because it will be
1407 one too small for loops of the form "while (i-- > 0)". We can not
1408 emit code before the loop_skip_over insns to fix this problem as this
1409 will then give a number one too large for loops of the form
1410 "while (--i > 0)".
1412 Note that all loops that reach here are entered at the top, because
1413 this function is not called if the loop starts with a jump. */
1415 /* Fail if loop_iteration_var is not live before loop_start, since we need
1416 to test its value in the preconditioning code. */
1420 {
1425 }
1431 /* Success! */
1435 }
1438 /* All pseudo-registers must be mapped to themselves. Two hard registers
1439 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1440 REGNUM, to avoid function-inlining specific conversions of these
1441 registers. All other hard regs can not be mapped because they may be
1442 used with different
1443 modes. */
1445 static void
1449 {
1454 /* Just clear the rest of the entries. */
1462 }
1463 \f
1464 /* Strength-reduction will often emit code for optimized biv/givs which
1465 calculates their value in a temporary register, and then copies the result
1466 to the iv. This procedure reconstructs the pattern computing the iv;
1467 verifying that all operands are of the proper form.
1469 PATTERN must be the result of single_set.
1470 The return value is the amount that the giv is incremented by. */
1472 static rtx
1476 {
1477 rtx increment;
1481 retry:
1482 /* Verify that we have an increment insn here. First check for a plus
1483 as the set source. */
1485 {
1486 /* SR sometimes computes the new giv value in a temp, then copies it
1487 to the new_reg. */
1493 /* The last insn emitted is not needed, so delete it to avoid confusing
1494 the second cse pass. This insn sets the giv unnecessarily. */
1496 }
1498 /* Verify that we have a constant as the second operand of the plus. */
1501 {
1502 /* SR sometimes puts the constant in a register, especially if it is
1503 too big to be an add immed operand. */
1507 /* SR may have used LO_SUM to compute the constant if it is too large
1508 for a load immed operand. In this case, the constant is in operand
1509 one of the LO_SUM rtx. */
1513 /* Some ports store large constants in memory and add a REG_EQUAL
1514 note to the store insn. */
1516 {
1520 }
1525 {
1526 /* The rs6000 port loads some constants with IOR.
1527 The alpha port loads some constants with ASHIFT and PLUS. */
1533 /* Don't need the last insn anymore. */
1544 else
1546 }
1551 /* The insn loading the constant into a register is no longer needed,
1552 so delete it. */
1554 }
1558 else
1561 /* Check that the source register is the same as the register we expected
1562 to see as the source. If not, something is seriously wrong. */
1565 {
1566 /* Some machines (e.g. the romp), may emit two add instructions for
1567 certain constants, so lets try looking for another add immediately
1568 before this one if we have only seen one add insn so far. */
1571 {
1572 tries++;
1580 }
1583 }
1586 }
1588 /* Copy REG_NOTES, except for insn references, because not all insn_map
1589 entries are valid yet. We do need to copy registers now though, because
1590 the reg_map entries can change during copying. */
1592 static rtx
1594 rtx notes;
1596 {
1597 rtx copy;
1608 /* Don't substitute for these yet. */
1610 else
1616 }
1618 /* Fixup insn references in copied REG_NOTES. */
1620 static void
1622 rtx notes;
1624 {
1625 rtx note;
1630 }
1632 /* Copy each instruction in the loop, substituting from map as appropriate.
1633 This is very similar to a loop in expand_inline_function. */
1635 static void
1638 copy_notes_from)
1641 rtx exit_label;
1645 {
1649 #ifdef HAVE_cc0
1651 #endif
1655 /* If this isn't the last iteration, then map any references to the
1656 start_label to final_label. Final label will then be emitted immediately
1657 after the end of this loop body if it was ever used.
1659 If this is the last iteration, then map references to the start_label
1660 to itself. */
1662 {
1665 final_label);
1666 }
1667 else
1673 do
1674 {
1680 {
1686 /* Check to see if this is a giv that has been combined with
1687 some split address givs. (Combined in the sense that
1688 `combine_givs' in loop.c has put two givs in the same register.)
1689 In this case, we must search all givs based on the same biv to
1690 find the address givs. Then split the address givs.
1691 Do this before splitting the giv, since that may map the
1692 SET_DEST to a new register. */
1697 {
1705 /* Although the giv_inc amount is not needed here, we must call
1706 calculate_giv_inc here since it might try to delete the
1707 last insn emitted. If we wait until later to call it,
1708 we might accidentally delete insns generated immediately
1709 below by emit_unrolled_add. */
1713 /* Now find all address giv's that were combined with this
1714 giv 'v'. */
1717 {
1720 /* If this DEST_ADDR giv was not split, then ignore it. */
1724 /* Scale this_giv_inc if the multiplicative factors of
1725 the two givs are different. */
1735 {
1736 /* Must emit an insn to increment the split address
1737 giv. Add in the const_adjust field in case there
1738 was a constant eliminated from the address. */
1741 /* tv->dest_reg will be either a bare register,
1742 or else a register plus a constant. */
1745 else
1748 /* Check for shared address givs, and avoid
1749 incrementing the shared pseudo reg more than
1750 once. */
1752 {
1753 /* tv->dest_reg may actually be a (PLUS (REG)
1754 (CONST)) here, so we must call plus_constant
1755 to add the const_adjust amount before calling
1756 emit_unrolled_add below. */
1760 /* The constant could be too large for an add
1761 immediate, so can't directly emit an insn
1762 here. */
1765 }
1767 /* Reset the giv to be just the register again, in case
1768 it is used after the set we have just emitted.
1769 We must subtract the const_adjust factor added in
1770 above. */
1774 }
1775 }
1776 }
1778 /* If this is a setting of a splittable variable, then determine
1779 how to split the variable, create a new set based on this split,
1780 and set up the reg_map so that later uses of the variable will
1781 use the new split variable. */
1788 {
1793 /* Compute the increment value for the giv, if it wasn't
1794 already computed above. */
1802 {
1803 /* Completely unrolling the loop. Set the induction
1804 variable to a known constant value. */
1806 /* The value in splittable_regs may be an invariant
1807 value, so we must use plus_constant here. */
1812 {
1815 }
1816 else
1817 {
1818 /* The splittable_regs value must be a REG or a
1819 CONST_INT, so put the entire value in the giv_src_reg
1820 variable. */
1823 }
1824 }
1825 else
1826 {
1827 /* Partially unrolling loop. Create a new pseudo
1828 register for the iteration variable, and set it to
1829 be a constant plus the original register. Except
1830 on the last iteration, when the result has to
1831 go back into the original iteration var register. */
1833 /* Handle bivs which must be mapped to a new register
1834 when split. This happens for bivs which need their
1835 final value set before loop entry. The new register
1836 for the biv was stored in the biv's first struct
1837 induction entry by find_splittable_regs. */
1841 {
1844 }
1846 #if 0
1847 /* If non-reduced/final-value givs were split, then
1848 this would have to remap those givs also. See
1849 find_splittable_regs. */
1850 #endif
1857 /* Now split the induction variable by changing the dest
1858 of this insn to a new register, and setting its
1859 reg_map entry to point to this new register.
1861 If this is the last iteration, and this is the last insn
1862 that will update the iv, then reuse the original dest,
1863 to ensure that the iv will have the proper value when
1864 the loop exits or repeats.
1866 Using splittable_regs_updates here like this is safe,
1867 because it can only be greater than one if all
1868 instructions modifying the iv are always executed in
1869 order. */
1873 {
1878 giv_inc == const0_rtx
1879 ? giv_src_reg
1883 }
1884 else
1886 }
1888 /* The constant being added could be too large for an add
1889 immediate, so can't directly emit an insn here. */
1893 }
1894 else
1895 {
1898 }
1901 #ifdef HAVE_cc0
1902 /* If this insn is setting CC0, it may need to look at
1903 the insn that uses CC0 to see what type of insn it is.
1904 In that case, the call to recog via validate_change will
1905 fail. So don't substitute constants here. Instead,
1906 do it when we emit the following insn.
1908 For example, see the pyr.md file. That machine has signed and
1909 unsigned compares. The compare patterns must check the
1910 following branch insn to see which what kind of compare to
1911 emit.
1913 If the previous insn set CC0, substitute constants on it as
1914 well. */
1917 else
1918 {
1923 }
1924 #else
1926 #endif
1928 /* Make split induction variable constants `permanent' since we
1929 know there are no backward branches across iteration variable
1930 settings which would invalidate this. */
1932 {
1938 }
1948 {
1949 /* This is a branch to the beginning of the loop; this is the
1950 last insn being copied; and this is not the last iteration.
1951 In this case, we want to change the original fall through
1952 case to be a branch past the end of the loop, and the
1953 original jump label case to fall_through. */
1956 {
1959 CODE_LABEL_NUMBER
1963 }
1964 else
1965 {
1966 rtx jmp;
1968 /* Can't do it by reversing the jump (probably because we
1969 couldn't reverse the conditions), so emit a new
1970 jump_insn after COPY, and redirect the jump around
1971 that. */
1978 CODE_LABEL_NUMBER
1982 }
1983 }
1985 #ifdef HAVE_cc0
1989 #endif
1992 /* Set the jump label of COPY correctly to avoid problems with
1993 later passes of unroll_loop, if INSN had jump label set. */
1995 {
1998 /* Can't use the label_map for every insn, since this may be
1999 the backward branch, and hence the label was not mapped. */
2001 {
2006 {
2009 else
2011 }
2012 }
2016 else
2017 {
2018 /* An unrecognizable jump insn, probably the entry jump
2019 for a switch statement. This label must have been mapped,
2020 so just use the label_map to get the new jump label. */
2024 }
2026 /* If this is a non-local jump, then must increase the label
2027 use count so that the label will not be deleted when the
2028 original jump is deleted. */
2030 }
2033 {
2041 }
2043 /* If this used to be a conditional jump insn but whose branch
2044 direction is now known, we must do something special. */
2046 {
2047 #ifdef HAVE_cc0
2048 /* The previous insn set cc0 for us. So delete it. */
2050 #endif
2052 /* If this is now a no-op, delete it. */
2054 {
2055 /* Don't let delete_insn delete the label referenced here,
2056 because we might possibly need it later for some other
2057 instruction in the loop. */
2064 }
2065 else
2066 /* Otherwise, this is unconditional jump so we must put a
2067 BARRIER after it. We could do some dead code elimination
2068 here, but jump.c will do it just as well. */
2070 }
2078 /* Because the USAGE information potentially contains objects other
2079 than hard registers, we need to copy it. */
2083 #ifdef HAVE_cc0
2087 #endif
2090 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
2096 /* If this is the loop start label, then we don't need to emit a
2097 copy of this label since no one will use it. */
2100 {
2104 }
2112 /* VTOP notes are valid only before the loop exit test. If placed
2113 anywhere else, loop may generate bad code. */
2120 else
2127 }
2130 }
2133 /* Now finish coping the REG_NOTES. */
2135 do
2136 {
2142 }
2145 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2146 each of these notes here, since there may be some important ones, such as
2147 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2148 iteration, because the original notes won't be deleted.
2150 We can't use insert_before here, because when from preconditioning,
2151 insert_before points before the loop. We can't use copy_end, because
2152 there may be insns already inserted after it (which we don't want to
2153 copy) when not from preconditioning code. */
2156 {
2158 {
2162 }
2163 }
2171 }
2172 \f
2173 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2174 emitted. This will correctly handle the case where the increment value
2175 won't fit in the immediate field of a PLUS insns. */
2177 void
2180 {
2181 rtx result;
2188 }
2189 \f
2190 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2191 is a backward branch in that range that branches to somewhere between
2192 LOOP_START and INSN. Returns 0 otherwise. */
2194 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2195 In practice, this is not a problem, because this function is seldom called,
2196 and uses a negligible amount of CPU time on average. */
2198 int
2200 rtx insn;
2202 {
2206 /* Stop before we get to the backward branch at the end of the loop. */
2211 /* Check in case insn has been deleted, search forward for first non
2212 deleted insn following it. */
2216 /* Check for the case where insn is the last insn in the loop. Deal
2217 with the case where INSN was a deleted loop test insn, in which case
2218 it will now be the NOTE_LOOP_END. */
2223 {
2225 {
2228 /* Search from loop_start to insn, to see if one of them is
2229 the target_insn. We can't use INSN_LUID comparisons here,
2230 since insn may not have an LUID entry. */
2234 }
2235 }
2238 }
2240 /* Try to generate the simplest rtx for the expression
2241 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2242 value of giv's. */
2244 static rtx
2248 {
2250 rtx result;
2252 /* The modes must all be the same. This should always be true. For now,
2253 check to make sure. */
2259 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2260 will be a constant. */
2262 {
2266 }
2272 /* Again, put the constant second. */
2274 {
2278 }
2285 }
2287 /* Searches the list of induction struct's for the biv BL, to try to calculate
2288 the total increment value for one iteration of the loop as a constant.
2290 Returns the increment value as an rtx, simplified as much as possible,
2291 if it can be calculated. Otherwise, returns 0. */
2293 rtx
2297 {
2299 rtx result;
2301 /* For increment, must check every instruction that sets it. Each
2302 instruction must be executed only once each time through the loop.
2303 To verify this, we check that the insn is always executed, and that
2304 there are no backward branches after the insn that branch to before it.
2305 Also, the insn must have a mult_val of one (to make sure it really is
2306 an increment). */
2310 {
2314 else
2316 }
2319 }
2321 /* Determine the initial value of the iteration variable, and the amount
2322 that it is incremented each loop. Use the tables constructed by
2323 the strength reduction pass to calculate these values.
2325 Initial_value and/or increment are set to zero if their values could not
2326 be calculated. */
2328 void
2332 {
2334 #if 0
2336 #endif
2338 /* Clear the result values, in case no answer can be found. */
2342 /* The iteration variable can be either a giv or a biv. Check to see
2343 which it is, and compute the variable's initial value, and increment
2344 value if possible. */
2346 /* If this is a new register, can't handle it since we don't have any
2347 reg_iv_type entry for it. */
2349 {
2354 }
2356 /* Reject iteration variables larger than the host wide int size, since they
2357 could result in a number of iterations greater than the range of our
2358 `unsigned HOST_WIDE_INT' variable loop_n_iterations. */
2361 {
2366 }
2368 {
2373 }
2375 {
2376 /* Grab initial value, only useful if it is a constant. */
2381 }
2383 {
2384 #if 1
2385 /* ??? The code below does not work because the incorrect number of
2386 iterations is calculated when the biv is incremented after the giv
2387 is set (which is the usual case). This can probably be accounted
2388 for by biasing the initial_value by subtracting the amount of the
2389 increment that occurs between the giv set and the giv test. However,
2390 a giv as an iterator is very rare, so it does not seem worthwhile
2391 to handle this. */
2392 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2397 #else
2398 /* Initial value is mult_val times the biv's initial value plus
2399 add_val. Only useful if it is a constant. */
2405 /* Increment value is mult_val times the increment value of the biv. */
2411 #endif
2412 }
2413 else
2414 {
2419 }
2420 }
2422 /* Calculate the approximate final value of the iteration variable
2423 which has an loop exit test with code COMPARISON_CODE and comparison value
2424 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2425 was signed or unsigned, and the direction of the comparison. This info is
2426 needed to calculate the number of loop iterations. */
2428 static rtx
2431 rtx comparison_value;
2434 {
2435 /* Calculate the final value of the induction variable.
2436 The exact final value depends on the branch operator, and increment sign.
2437 This is only an approximate value. It will be wrong if the iteration
2438 variable is not incremented by one each time through the loop, and
2439 approx final value - start value % increment != 0. */
2443 {
2455 /* Can not calculate a final value for this case. */
2474 }
2475 }
2477 /* For each biv and giv, determine whether it can be safely split into
2478 a different variable for each unrolled copy of the loop body. If it
2479 is safe to split, then indicate that by saving some useful info
2480 in the splittable_regs array.
2482 If the loop is being completely unrolled, then splittable_regs will hold
2483 the current value of the induction variable while the loop is unrolled.
2484 It must be set to the initial value of the induction variable here.
2485 Otherwise, splittable_regs will hold the difference between the current
2486 value of the induction variable and the value the induction variable had
2487 at the top of the loop. It must be set to the value 0 here.
2489 Returns the total number of instructions that set registers that are
2490 splittable. */
2492 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2493 constant values are unnecessary, since we can easily calculate increment
2494 values in this case even if nothing is constant. The increment value
2495 should not involve a multiply however. */
2497 /* ?? Even if the biv/giv increment values aren't constant, it may still
2498 be beneficial to split the variable if the loop is only unrolled a few
2499 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2501 static int
2503 unroll_number)
2506 rtx end_insert_before;
2508 {
2512 rtx biv_final_value;
2517 {
2518 /* Biv_total_increment must return a constant value,
2519 otherwise we can not calculate the split values. */
2525 /* The loop must be unrolled completely, or else have a known number
2526 of iterations and only one exit, or else the biv must be dead
2527 outside the loop, or else the final value must be known. Otherwise,
2528 it is unsafe to split the biv since it may not have the proper
2529 value on loop exit. */
2531 /* loop_number_exit_count is non-zero if the loop has an exit other than
2532 a fall through at the end. */
2548 /* If any of the insns setting the BIV don't do so with a simple
2549 PLUS, we don't know how to split it. */
2557 /* If final value is non-zero, then must emit an instruction which sets
2558 the value of the biv to the proper value. This is done after
2559 handling all of the givs, since some of them may need to use the
2560 biv's value in their initialization code. */
2562 /* This biv is splittable. If completely unrolling the loop, save
2563 the biv's initial value. Otherwise, save the constant zero. */
2566 {
2568 {
2569 /* If the initial value of the biv is itself (i.e. it is too
2570 complicated for strength_reduce to compute), or is a hard
2571 register, or it isn't invariant, then we must create a new
2572 pseudo reg to hold the initial value of the biv. */
2578 {
2583 loop_start);
2590 }
2591 else
2593 }
2594 else
2597 /* Save the number of instructions that modify the biv, so that
2598 we can treat the last one specially. */
2606 }
2608 /* Check every giv that depends on this biv to see whether it is
2609 splittable also. Even if the biv isn't splittable, givs which
2610 depend on it may be splittable if the biv is live outside the
2611 loop, and the givs aren't. */
2616 /* If final value is non-zero, then must emit an instruction which sets
2617 the value of the biv to the proper value. This is done after
2618 handling all of the givs, since some of them may need to use the
2619 biv's value in their initialization code. */
2621 {
2622 /* If the loop has multiple exits, emit the insns before the
2623 loop to ensure that it will always be executed no matter
2624 how the loop exits. Otherwise emit the insn after the loop,
2625 since this is slightly more efficient. */
2628 biv_final_value),
2629 end_insert_before);
2630 else
2631 {
2632 /* Create a new register to hold the value of the biv, and then
2633 set the biv to its final value before the loop start. The biv
2634 is set to its final value before loop start to ensure that
2635 this insn will always be executed, no matter how the loop
2636 exits. */
2641 loop_start);
2643 biv_final_value),
2644 loop_start);
2650 /* Set up the mapping from the original biv register to the new
2651 register. */
2653 }
2654 }
2655 }
2657 }
2659 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2660 for the instruction that is using it. Do not make any changes to that
2661 instruction. */
2663 static int
2666 rtx giv_inc;
2668 {
2674 /* First check to see if either address would fail. Handle the fact
2675 that we have may have a match_dup. */
2680 /* Now put things back the way they were before. This should always
2681 succeed. */
2686 }
2688 /* For every giv based on the biv BL, check to determine whether it is
2689 splittable. This is a subroutine to find_splittable_regs ().
2691 Return the number of instructions that set splittable registers. */
2693 static int
2695 unroll_number)
2699 rtx increment;
2701 {
2703 rtx final_value;
2704 rtx tem;
2707 /* Scan the list of givs, and set the same_insn field when there are
2708 multiple identical givs in the same insn. */
2716 {
2719 /* Only split the giv if it has already been reduced, or if the loop is
2720 being completely unrolled. */
2724 /* The giv can be split if the insn that sets the giv is executed once
2725 and only once on every iteration of the loop. */
2726 /* An address giv can always be split. v->insn is just a use not a set,
2727 and hence it does not matter whether it is always executed. All that
2728 matters is that all the biv increments are always executed, and we
2729 won't reach here if they aren't. */
2735 /* The giv increment value must be a constant. */
2741 /* The loop must be unrolled completely, or else have a known number of
2742 iterations and only one exit, or else the giv must be dead outside
2743 the loop, or else the final value of the giv must be known.
2744 Otherwise, it is not safe to split the giv since it may not have the
2745 proper value on loop exit. */
2747 /* The used outside loop test will fail for DEST_ADDR givs. They are
2748 never used outside the loop anyways, so it is always safe to split a
2749 DEST_ADDR giv. */
2756 /* The next part is true if the pseudo is used outside the loop.
2757 We assume that this is true for any pseudo created after loop
2758 starts, because we don't have a reg_n_info entry for them. */
2761 /* Check for the case where the pseudo is set by a shift/add
2762 sequence, in which case the first insn setting the pseudo
2763 is the first insn of the shift/add sequence. */
2767 /* Line above always fails if INSN was moved by loop opt. */
2773 #if 0
2774 /* Currently, non-reduced/final-value givs are never split. */
2775 /* Should emit insns after the loop if possible, as the biv final value
2776 code below does. */
2778 /* If the final value is non-zero, and the giv has not been reduced,
2779 then must emit an instruction to set the final value. */
2781 {
2782 /* Create a new register to hold the value of the giv, and then set
2783 the giv to its final value before the loop start. The giv is set
2784 to its final value before loop start to ensure that this insn
2785 will always be executed, no matter how we exit. */
2789 loop_start);
2796 }
2797 #endif
2799 /* This giv is splittable. If completely unrolling the loop, save the
2800 giv's initial value. Otherwise, save the constant zero for it. */
2803 {
2804 /* It is not safe to use bl->initial_value here, because it may not
2805 be invariant. It is safe to use the initial value stored in
2806 the splittable_regs array if it is set. In rare cases, it won't
2807 be set, so then we do exactly the same thing as
2808 find_splittable_regs does to get a safe value. */
2809 rtx biv_initial_value;
2817 else
2818 {
2823 loop_start);
2825 }
2828 }
2829 else
2833 {
2834 /* If a giv was combined with another giv, then we can only split
2835 this giv if the giv it was combined with was reduced. This
2836 is because the value of v->new_reg is meaningless in this
2837 case. */
2839 {
2844 }
2845 /* If the giv is an address destination, it could be something other
2846 than a simple register, these have to be treated differently. */
2848 {
2849 /* If value is not a constant, register, or register plus
2850 constant, then compute its value into a register before
2851 loop start. This prevents invalid rtx sharing, and should
2852 generate better code. We can use bl->initial_value here
2853 instead of splittable_regs[bl->regno] because this code
2854 is going before the loop start. */
2861 {
2867 }
2870 }
2871 else
2872 {
2873 /* Splitting address givs is useful since it will often allow us
2874 to eliminate some increment insns for the base giv as
2875 unnecessary. */
2877 /* If the addr giv is combined with a dest_reg giv, then all
2878 references to that dest reg will be remapped, which is NOT
2879 what we want for split addr regs. We always create a new
2880 register for the split addr giv, just to be safe. */
2882 /* If we have multiple identical address givs within a
2883 single instruction, then use a single pseudo reg for
2884 both. This is necessary in case one is a match_dup
2885 of the other. */
2890 {
2896 }
2897 /* If multiple address GIVs have been combined with the
2898 same dest_reg GIV, do not create a new register for
2899 each. */
2904 /* combine_givs_p may return true for some cases
2905 where the add and mult values are not equal.
2906 To share a register here, the values must be
2907 equal. */
2911 {
2914 }
2916 {
2917 /* If not completely unrolling the loop, then create a new
2918 register to hold the split value of the DEST_ADDR giv.
2919 Emit insn to initialize its value before loop start. */
2924 /* If the address giv has a constant in its new_reg value,
2925 then this constant can be pulled out and put in value,
2926 instead of being part of the initialization code. */
2930 {
2931 v->dest_reg
2934 /* Only succeed if this will give valid addresses.
2935 Try to validate both the first and the last
2936 address resulting from loop unrolling, if
2937 one fails, then can't do const elim here. */
2939 {
2940 /* Save the negative of the eliminated const, so
2941 that we can calculate the dest_reg's increment
2942 value later. */
2950 }
2951 else
2953 }
2954 else
2957 /* If the address hasn't been checked for validity yet, do so
2958 now, and fail completely if either the first or the last
2959 unrolled copy of the address is not a valid address
2960 for the instruction that uses it. */
2963 {
2973 }
2975 /* We set this after the address check, to guarantee that
2976 the register will be initialized. */
2979 /* To initialize the new register, just move the value of
2980 new_reg into it. This is not guaranteed to give a valid
2981 instruction on machines with complex addressing modes.
2982 If we can't recognize it, then delete it and emit insns
2983 to calculate the value from scratch. */
2986 loop_start);
2988 {
2991 /* We can't use bl->initial_value to compute the initial
2992 value, because the loop may have been preconditioned.
2993 We must calculate it from NEW_REG. Try using
2994 force_operand instead of emit_iv_add_mult. */
3008 }
3009 }
3010 else
3011 {
3014 /* Check the resulting address for validity, and fail
3015 if the resulting address would be invalid. */
3017 {
3027 }
3028 }
3030 /* Store the value of dest_reg into the insn. This sharing
3031 will not be a problem as this insn will always be copied
3032 later. */
3036 /* If this address giv is combined with a dest reg giv, then
3037 save the base giv's induction pointer so that we will be
3038 able to handle this address giv properly. The base giv
3039 itself does not have to be splittable. */
3045 {
3046 /* This giv maybe hasn't been combined with any others.
3047 Make sure that it's giv is marked as splittable here. */
3051 /* Make it appear to depend upon itself, so that the
3052 giv will be properly split in the main loop above. */
3054 {
3057 }
3058 }
3062 }
3063 }
3064 else
3065 {
3066 #if 0
3067 /* Currently, unreduced giv's can't be split. This is not too much
3068 of a problem since unreduced giv's are not live across loop
3069 iterations anyways. When unrolling a loop completely though,
3070 it makes sense to reduce&split givs when possible, as this will
3071 result in simpler instructions, and will not require that a reg
3072 be live across loop iterations. */
3077 #else
3079 #endif
3080 }
3082 /* Unreduced givs are only updated once by definition. Reduced givs
3083 are updated as many times as their biv is. Mark it so if this is
3084 a splittable register. Don't need to do anything for address givs
3085 where this may not be a register. */
3088 {
3094 }
3096 result++;
3099 {
3106 else
3110 }
3111 }
3114 }
3115 \f
3116 /* Try to prove that the register is dead after the loop exits. Trace every
3117 loop exit looking for an insn that will always be executed, which sets
3118 the register to some value, and appears before the first use of the register
3119 is found. If successful, then return 1, otherwise return 0. */
3121 /* ?? Could be made more intelligent in the handling of jumps, so that
3122 it can search past if statements and other similar structures. */
3124 static int
3127 {
3134 /* In addition to checking all exits of this loop, we must also check
3135 all exits of inner nested loops that would exit this loop. We don't
3136 have any way to identify those, so we just give up if there are any
3137 such inner loop exits. */
3141 label_count++;
3146 /* HACK: Must also search the loop fall through exit, create a label_ref
3147 here which points to the loop_end, and append the loop_number_exit_labels
3148 list to it. */
3153 {
3154 /* Succeed if find an insn which sets the biv or if reach end of
3155 function. Fail if find an insn that uses the biv, or if come to
3156 a conditional jump. */
3160 {
3163 {
3164 rtx set;
3172 }
3175 {
3179 /* Prevent infinite loop following infinite loops. */
3182 else
3184 }
3187 }
3188 }
3190 /* Success, the register is dead on all loop exits. */
3192 }
3194 /* Try to calculate the final value of the biv, the value it will have at
3195 the end of the loop. If we can do it, return that value. */
3197 rtx
3201 {
3204 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3209 /* The final value for reversed bivs must be calculated differently than
3210 for ordinary bivs. In this case, there is already an insn after the
3211 loop which sets this biv's final value (if necessary), and there are
3212 no other loop exits, so we can return any value. */
3214 {
3220 }
3222 /* Try to calculate the final value as initial value + (number of iterations
3223 * increment). For this to work, increment must be invariant, the only
3224 exit from the loop must be the fall through at the bottom (otherwise
3225 it may not have its final value when the loop exits), and the initial
3226 value of the biv must be invariant. */
3231 {
3235 {
3236 /* Can calculate the loop exit value, emit insns after loop
3237 end to calculate this value into a temporary register in
3238 case it is needed later. */
3242 /* Make sure loop_end is not the last insn. */
3253 }
3254 }
3256 /* Check to see if the biv is dead at all loop exits. */
3258 {
3265 }
3268 }
3270 /* Try to calculate the final value of the giv, the value it will have at
3271 the end of the loop. If we can do it, return that value. */
3273 rtx
3277 {
3279 rtx insn;
3285 /* The final value for givs which depend on reversed bivs must be calculated
3286 differently than for ordinary givs. In this case, there is already an
3287 insn after the loop which sets this giv's final value (if necessary),
3288 and there are no other loop exits, so we can return any value. */
3290 {
3296 }
3298 /* Try to calculate the final value as a function of the biv it depends
3299 upon. The only exit from the loop must be the fall through at the bottom
3300 (otherwise it may not have its final value when the loop exits). */
3302 /* ??? Can calculate the final giv value by subtracting off the
3303 extra biv increments times the giv's mult_val. The loop must have
3304 only one exit for this to work, but the loop iterations does not need
3305 to be known. */
3309 {
3310 /* ?? It is tempting to use the biv's value here since these insns will
3311 be put after the loop, and hence the biv will have its final value
3312 then. However, this fails if the biv is subsequently eliminated.
3313 Perhaps determine whether biv's are eliminable before trying to
3314 determine whether giv's are replaceable so that we can use the
3315 biv value here if it is not eliminable. */
3317 /* We are emitting code after the end of the loop, so we must make
3318 sure that bl->initial_value is still valid then. It will still
3319 be valid if it is invariant. */
3325 {
3326 /* Can calculate the loop exit value of its biv as
3327 (loop_n_iterations * increment) + initial_value */
3329 /* The loop exit value of the giv is then
3330 (final_biv_value - extra increments) * mult_val + add_val.
3331 The extra increments are any increments to the biv which
3332 occur in the loop after the giv's value is calculated.
3333 We must search from the insn that sets the giv to the end
3334 of the loop to calculate this value. */
3338 /* Put the final biv value in tem. */
3344 /* Subtract off extra increments as we find them. */
3347 {
3352 {
3356 OPTAB_LIB_WIDEN);
3360 }
3361 }
3363 /* Now calculate the giv's final value. */
3365 insert_before);
3373 }
3374 }
3376 /* Replaceable giv's should never reach here. */
3380 /* Check to see if the biv is dead at all loop exits. */
3382 {
3389 }
3392 }
3395 /* Calculate the number of loop iterations. Returns the exact number of loop
3396 iterations if it can be calculated, otherwise returns zero. */
3398 unsigned HOST_WIDE_INT
3401 {
3405 HOST_WIDE_INT i;
3409 rtx last_loop_insn;
3411 /* First find the iteration variable. If the last insn is a conditional
3412 branch, and the insn before tests a register value, make that the
3413 iteration variable. */
3420 /* We used to use pren_nonnote_insn here, but that fails because it might
3421 accidentally get the branch for a contained loop if the branch for this
3422 loop was deleted. We can only trust branches immediately before the
3423 loop_end. */
3428 {
3433 }
3435 /* ??? Get_condition may switch position of induction variable and
3436 invariant register when it canonicalizes the comparison. */
3443 {
3448 }
3450 /* Loop iterations is always called before any new registers are created
3451 now, so this should never occur. */
3459 /* iteration_info already printed a message. */
3462 /* If the comparison value is an invariant register, then try to find
3463 its value from the insns before the start of the loop. */
3466 {
3470 {
3476 {
3477 /* We found the last insn before the loop that sets the register.
3478 If it sets the entire register, and has a REG_EQUAL note,
3479 then use the value of the REG_EQUAL note. */
3482 {
3485 /* Only use the REG_EQUAL note if it is a constant.
3486 Other things, divide in particular, will cause
3487 problems later if we use them. */
3491 }
3493 }
3494 }
3495 }
3500 /* Save the calculated values describing this loop's bounds, in case
3501 precondition_loop_p will need them later. These values can not be
3502 recalculated inside precondition_loop_p because strength reduction
3503 optimizations may obscure the loop's structure. */
3512 {
3517 }
3519 {
3524 }
3526 {
3531 }
3533 {
3538 }
3540 {
3545 }
3547 /* ?? Final value and initial value do not have to be constants.
3548 Only their difference has to be constant. When the iteration variable
3549 is an array address, the final value and initial value might both
3550 be addresses with the same base but different constant offsets.
3551 Final value must be invariant for this to work.
3553 To do this, need some way to find the values of registers which are
3554 invariant. */
3556 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3558 final_larger
3563 else
3571 else
3574 /* There are 27 different cases: compare_dir = -1, 0, 1;
3575 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3576 There are 4 normal cases, 4 reverse cases (where the iteration variable
3577 will overflow before the loop exits), 4 infinite loop cases, and 15
3578 immediate exit (0 or 1 iteration depending on loop type) cases.
3579 Only try to optimize the normal cases. */
3581 /* (compare_dir/final_larger/increment_dir)
3582 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3583 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3584 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3585 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3587 /* ?? If the meaning of reverse loops (where the iteration variable
3588 will overflow before the loop exits) is undefined, then could
3589 eliminate all of these special checks, and just always assume
3590 the loops are normal/immediate/infinite. Note that this means
3591 the sign of increment_dir does not have to be known. Also,
3592 since it does not really hurt if immediate exit loops or infinite loops
3593 are optimized, then that case could be ignored also, and hence all
3594 loops can be optimized.
3596 According to ANSI Spec, the reverse loop case result is undefined,
3597 because the action on overflow is undefined.
3599 See also the special test for NE loops below. */
3603 /* Normal case. */
3604 ;
3605 else
3606 {
3611 }
3613 /* Calculate the number of iterations, final_value is only an approximation,
3614 so correct for that. Note that tempu and loop_n_iterations are
3615 unsigned, because they can be as large as 2^n - 1. */
3621 {
3624 }
3625 else
3628 /* For NE tests, make sure that the iteration variable won't miss the
3629 final value. If tempu mod i is not zero, then the iteration variable
3630 will overflow before the loop exits, and we can not calculate the
3631 number of iterations. */
3636 }
3638 /* Replace uses of split bivs with their split pseudo register. This is
3639 for original instructions which remain after loop unrolling without
3640 copying. */
3642 static rtx
3644 rtx x;
3645 {
3655 {
3667 #if 0
3668 /* If non-reduced/final-value givs were split, then this would also
3669 have to remap those givs also. */
3670 #endif
3678 }
3682 {
3686 {
3690 }
3691 }
3693 }
3695 /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g.
3696 FIST_UID is always executed if LAST_UID is), then return 1. Otherwise
3697 return 0. COPY_START is where we can start looking for the insns
3698 FIRST_UID and LAST_UID. COPY_END is where we stop looking for these
3699 insns.
3701 If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID
3702 must dominate LAST_UID.
3704 If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3705 may not dominate LAST_UID.
3707 If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID
3708 must dominate LAST_UID. */
3710 int
3715 rtx copy_start;
3716 rtx copy_end;
3717 {
3722 {
3725 /* Could not find FIRST_UID. */
3729 }
3731 /* Verify that FIRST_UID is an insn that entirely sets REGNO. */
3736 /* FIRST_UID is always executed. */
3741 {
3742 /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we
3743 can not be sure that FIRST_UID dominates LAST_UID. */
3746 /* Could not find LAST_UID, but we reached the end of the loop, so
3747 it must be safe. */
3751 }
3753 /* FIRST_UID is always executed if LAST_UID is executed. */
3755 }