| blob | 4470c5b82e1d537890d68e33b0654b00bd61e6af |
1 /* Try to unroll loops, and split induction variables.
2 Copyright (C) 1992, 1993, 1994, 1995 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. */
156 #include <stdio.h>
159 /* This controls which loops are unrolled, and by how much we unroll
160 them. */
162 #ifndef MAX_UNROLLED_INSNS
163 #define MAX_UNROLLED_INSNS 100
164 #endif
166 /* Indexed by register number, if non-zero, then it contains a pointer
167 to a struct induction for a DEST_REG giv which has been combined with
168 one of more address givs. This is needed because whenever such a DEST_REG
169 giv is modified, we must modify the value of all split address givs
170 that were combined with this DEST_REG giv. */
174 /* Indexed by register number, if this is a splittable induction variable,
175 then this will hold the current value of the register, which depends on the
176 iteration number. */
180 /* Indexed by register number, if this is a splittable induction variable,
181 then this will hold the number of instructions in the loop that modify
182 the induction variable. Used to ensure that only the last insn modifying
183 a split iv will update the original iv of the dest. */
187 /* Values describing the current loop's iteration variable. These are set up
188 by loop_iterations, and used by precondition_loop_p. */
195 /* Forward declarations. */
213 /* Try to unroll one loop and split induction variables in the loop.
215 The loop is described by the arguments LOOP_END, INSN_COUNT, and
216 LOOP_START. END_INSERT_BEFORE indicates where insns should be added
217 which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
218 indicates whether information generated in the strength reduction pass
219 is available.
221 This function is intended to be called from within `strength_reduce'
222 in loop.c. */
224 void
226 strength_reduce_p)
227 rtx loop_end;
229 rtx loop_start;
230 rtx end_insert_before;
232 {
238 rtx insert_before;
245 rtx start_label;
250 rtx safety_label;
251 /* This points to the last real insn in the loop, which should be either
252 a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
253 jumps). */
254 rtx last_loop_insn;
256 /* Don't bother unrolling huge loops. Since the minimum factor is
257 two, loops greater than one half of MAX_UNROLLED_INSNS will never
258 be unrolled. */
260 {
264 }
266 /* When emitting debugger info, we can't unroll loops with unequal numbers
267 of block_beg and block_end notes, because that would unbalance the block
268 structure of the function. This can happen as a result of the
269 "if (foo) bar; else break;" optimization in jump.c. */
272 {
277 {
279 {
281 block_begins++;
283 block_ends++;
284 }
285 }
288 {
293 }
294 }
296 /* Determine type of unroll to perform. Depends on the number of iterations
297 and the size of the loop. */
299 /* If there is no strength reduce info, then set loop_n_iterations to zero.
300 This can happen if strength_reduce can't find any bivs in the loop.
301 A value of zero indicates that the number of iterations could not be
302 calculated. */
311 /* Find and save a pointer to the last nonnote insn in the loop. */
315 /* Calculate how many times to unroll the loop. Indicate whether or
316 not the loop is being completely unrolled. */
319 {
320 /* If number of iterations is exactly 1, then eliminate the compare and
321 branch at the end of the loop since they will never be taken.
322 Then return, since no other action is needed here. */
324 /* If the last instruction is not a BARRIER or a JUMP_INSN, then
325 don't do anything. */
328 {
329 /* Delete the jump insn. This will delete the barrier also. */
331 }
333 {
334 #ifdef HAVE_cc0
335 /* The immediately preceding insn is a compare which must be
336 deleted. */
339 #else
340 /* The immediately preceding insn may not be the compare, so don't
341 delete it. */
343 #endif
344 }
346 }
349 {
352 }
354 {
355 /* Try to factor the number of iterations. Don't bother with the
356 general case, only using 2, 3, 5, and 7 will get 75% of all
357 numbers theoretically, and almost all in practice. */
365 {
368 }
370 /* Start with the larger factors first so that we generally
371 get lots of unrolling. */
377 {
379 {
382 }
383 else
385 }
387 /* If we couldn't find any factors, then unroll as in the normal
388 case. */
390 {
394 }
395 else
397 }
400 /* Default case, calculate number of times to unroll loop based on its
401 size. */
403 {
408 else
412 }
414 /* Now we know how many times to unroll the loop. */
422 {
423 /* Loops of these types should never start with a jump down to
424 the exit condition test. For now, check for this case just to
425 be sure. UNROLL_NAIVE loops can be of this form, this case is
426 handled below. */
432 }
435 {
436 /* Completely unrolling the loop: Delete the compare and branch at
437 the end (the last two instructions). This delete must done at the
438 very end of loop unrolling, to avoid problems with calls to
439 back_branch_in_range_p, which is called by find_splittable_regs.
440 All increments of splittable bivs/givs are changed to load constant
441 instructions. */
445 /* Set insert_before to the instruction immediately after the JUMP_INSN
446 (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
447 the loop will be correctly handled by copy_loop_body. */
450 /* Set copy_end to the insn before the jump at the end of the loop. */
454 {
455 #ifdef HAVE_cc0
456 /* The instruction immediately before the JUMP_INSN is a compare
457 instruction which we do not want to copy. */
459 #else
460 /* The instruction immediately before the JUMP_INSN may not be the
461 compare, so we must copy it. */
463 #endif
464 }
465 else
466 {
467 /* We currently can't unroll a loop if it doesn't end with a
468 JUMP_INSN. There would need to be a mechanism that recognizes
469 this case, and then inserts a jump after each loop body, which
470 jumps to after the last loop body. */
475 }
476 }
478 {
479 /* Partially unrolling the loop: The compare and branch at the end
480 (the last two instructions) must remain. Don't copy the compare
481 and branch instructions at the end of the loop. Insert the unrolled
482 code immediately before the compare/branch at the end so that the
483 code will fall through to them as before. */
487 /* Set insert_before to the jump insn at the end of the loop.
488 Set copy_end to before the jump insn at the end of the loop. */
490 {
493 }
495 {
496 #ifdef HAVE_cc0
497 /* The instruction immediately before the JUMP_INSN is a compare
498 instruction which we do not want to copy or delete. */
501 #else
502 /* The instruction immediately before the JUMP_INSN may not be the
503 compare, so we must copy it. */
506 #endif
507 }
508 else
509 {
510 /* We currently can't unroll a loop if it doesn't end with a
511 JUMP_INSN. There would need to be a mechanism that recognizes
512 this case, and then inserts a jump after each loop body, which
513 jumps to after the last loop body. */
518 }
519 }
520 else
521 {
522 /* Normal case: Must copy the compare and branch instructions at the
523 end of the loop. */
526 {
527 /* Loop ends with an unconditional jump and a barrier.
528 Handle this like above, don't copy jump and barrier.
529 This is not strictly necessary, but doing so prevents generating
530 unconditional jumps to an immediately following label.
532 This will be corrected below if the target of this jump is
533 not the start_label. */
537 }
539 {
540 /* Set insert_before to immediately after the JUMP_INSN, so that
541 NOTEs at the end of the loop will be correctly handled by
542 copy_loop_body. */
545 }
546 else
547 {
548 /* We currently can't unroll a loop if it doesn't end with a
549 JUMP_INSN. There would need to be a mechanism that recognizes
550 this case, and then inserts a jump after each loop body, which
551 jumps to after the last loop body. */
556 }
558 /* If copying exit test branches because they can not be eliminated,
559 then must convert the fall through case of the branch to a jump past
560 the end of the loop. Create a label to emit after the loop and save
561 it for later use. Do not use the label after the loop, if any, since
562 it might be used by insns outside the loop, or there might be insns
563 added before it later by final_[bg]iv_value which must be after
564 the real exit label. */
572 {
573 /* The loop starts with a jump down to the exit condition test.
574 Start copying the loop after the barrier following this
575 jump insn. */
578 /* Splitting induction variables doesn't work when the loop is
579 entered via a jump to the bottom, because then we end up doing
580 a comparison against a new register for a split variable, but
581 we did not execute the set insn for the new register because
582 it was skipped over. */
587 }
588 else
590 }
592 /* This should always be the first label in the loop. */
594 /* There may be a line number note and/or a loop continue note here. */
598 {
599 /* This can happen as a result of jump threading. If the first insns in
600 the loop test the same condition as the loop's backward jump, or the
601 opposite condition, then the backward jump will be modified to point
602 to elsewhere, and the loop's start label is deleted.
604 This case currently can not be handled by the loop unrolling code. */
610 }
612 {
613 /* The jump optimization pass must have combined the original start label
614 with a named label for a goto. We can't unroll this case because
615 jumps which go to the named label must be handled differently than
616 jumps to the loop start, and it is impossible to differentiate them
617 in this case. */
622 }
627 {
628 /* In this case, we must copy the jump and barrier, because they will
629 not be converted to jumps to an immediately following label. */
633 }
635 /* Allocate a translation table for the labels and insn numbers.
636 They will be filled in as we copy the insns in the loop. */
645 /* Allocate the label map. */
648 {
653 }
654 else
657 /* Search the loop and mark all local labels, i.e. the ones which have to
658 be distinct labels when copied. For all labels which might be
659 non-local, set their label_map entries to point to themselves.
660 If they happen to be local their label_map entries will be overwritten
661 before the loop body is copied. The label_map entries for local labels
662 will be set to a different value each time the loop body is copied. */
665 {
669 {
675 {
679 rtx label;
682 {
685 }
686 }
687 }
688 }
690 /* Allocate space for the insn map. */
694 /* Set this to zero, to indicate that we are doing loop unrolling,
695 not function inlining. */
698 /* The register and constant maps depend on the number of registers
699 present, so the final maps can't be created until after
700 find_splittable_regs is called. However, they are needed for
701 preconditioning, so we create temporary maps when preconditioning
702 is performed. */
704 /* The preconditioning code may allocate two new pseudo registers. */
707 /* Allocate and zero out the splittable_regs and addr_combined_regs
708 arrays. These must be zeroed here because they will be used if
709 loop preconditioning is performed, and must be zero for that case.
711 It is safe to do this here, since the extra registers created by the
712 preconditioning code and find_splittable_regs will never be used
713 to access the splittable_regs[] and addr_combined_regs[] arrays. */
719 addr_combined_regs
722 /* We must limit it to max_reg_before_loop, because only these pseudo
723 registers have valid regno_first_uid info. Any register created after
724 that is unlikely to be local to the loop anyways. */
728 /* Mark all local registers, i.e. the ones which are referenced only
729 inside the loop. */
731 {
735 /* If a register is used in the jump insn, we must not duplicate it
736 since it will also be used outside the loop. */
738 copy_end_luid--;
739 /* If copy_start points to the NOTE that starts the loop, then we must
740 use the next luid, because invariant pseudo-regs moved out of the loop
741 have their lifetimes modified to start here, but they are not safe
742 to duplicate. */
744 copy_start_luid++;
752 }
754 /* If this loop requires exit tests when unrolled, check to see if we
755 can precondition the loop so as to make the exit tests unnecessary.
756 Just like variable splitting, this is not safe if the loop is entered
757 via a jump to the bottom. Also, can not do this if no strength
758 reduce info, because precondition_loop_p uses this info. */
760 /* Must copy the loop body for preconditioning before the following
761 find_splittable_regs call since that will emit insns which need to
762 be after the preconditioned loop copies, but immediately before the
763 unrolled loop copies. */
765 /* Also, it is not safe to split induction variables for the preconditioned
766 copies of the loop body. If we split induction variables, then the code
767 assumes that each induction variable can be represented as a function
768 of its initial value and the loop iteration number. This is not true
769 in this case, because the last preconditioned copy of the loop body
770 could be any iteration from the first up to the `unroll_number-1'th,
771 depending on the initial value of the iteration variable. Therefore
772 we can not split induction variables here, because we can not calculate
773 their value. Hence, this code must occur before find_splittable_regs
774 is called. */
777 {
782 {
799 /* Limit loop unrolling to 4, since this will make 7 copies of
800 the loop body. */
804 /* Save the absolute value of the increment, and also whether or
805 not it is negative. */
809 {
812 }
816 /* Decide what mode to do these calculations in. Choose the larger
817 of final_value's mode and initial_value's mode, or a full-word if
818 both are constants. */
821 {
825 }
831 /* Calculate the difference between the final and initial values.
832 Final value may be a (plus (reg x) (const_int 1)) rtx.
833 Let the following cse pass simplify this if initial value is
834 a constant.
836 We must copy the final and initial values here to avoid
837 improperly shared rtl. */
841 OPTAB_LIB_WIDEN);
843 /* Now calculate (diff % (unroll * abs (increment))) by using an
844 and instruction. */
849 /* Now emit a sequence of branches to jump to the proper precond
850 loop entry point. */
856 /* Check for the case where the initial value is greater than or equal
857 to the final value. In that case, we want to execute exactly
858 one loop iteration. The code below will fail for this case. */
864 else
869 /* Assuming the unroll_number is 4, and the increment is 2, then
870 for a negative increment: for a positive increment:
871 diff = 0,1 precond 0 diff = 0,7 precond 0
872 diff = 2,3 precond 3 diff = 1,2 precond 1
873 diff = 4,5 precond 2 diff = 3,4 precond 2
874 diff = 6,7 precond 1 diff = 5,6 precond 3 */
876 /* We only need to emit (unroll_number - 1) branches here, the
877 last case just falls through to the following code. */
879 /* ??? This would give better code if we emitted a tree of branches
880 instead of the current linear list of branches. */
883 {
887 /* For negative increments, must invert the constant compared
888 against, except when comparing against zero. */
890 {
893 }
895 {
898 }
899 else
900 {
903 }
912 else
916 }
918 /* If the increment is greater than one, then we need another branch,
919 to handle other cases equivalent to 0. */
921 /* ??? This should be merged into the code above somehow to help
922 simplify the code here, and reduce the number of branches emitted.
923 For the negative increment case, the branch here could easily
924 be merged with the `0' case branch above. For the positive
925 increment case, it is not clear how this can be simplified. */
928 {
933 {
936 }
937 else
938 {
941 }
948 else
952 }
958 /* Only the last copy of the loop body here needs the exit
959 test, so set copy_end to exclude the compare/branch here,
960 and then reset it inside the loop when get to the last
961 copy. */
966 {
967 #ifdef HAVE_cc0
968 /* The immediately preceding insn is a compare which we do not
969 want to copy. */
971 #else
972 /* The immediately preceding insn may not be a compare, so we
973 must copy it. */
975 #endif
976 }
977 else
981 {
999 /* The last copy needs the compare/branch insns at the end,
1000 so reset copy_end here if the loop ends with a conditional
1001 branch. */
1004 {
1007 else
1009 }
1011 /* None of the copies are the `last_iteration', so just
1012 pass zero for that parameter. */
1016 }
1020 {
1023 }
1024 else
1025 {
1026 #ifdef HAVE_cc0
1027 /* The immediately preceding insn is a compare which we do not
1028 want to copy. */
1031 #else
1032 /* The immediately preceding insn may not be a compare, so we
1033 must copy it. */
1036 #endif
1037 }
1039 /* Set unroll type to MODULO now. */
1042 }
1043 }
1045 /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
1046 the loop unless all loops are being unrolled. */
1048 {
1052 }
1054 /* At this point, we are guaranteed to unroll the loop. */
1056 /* For each biv and giv, determine whether it can be safely split into
1057 a different variable for each unrolled copy of the loop body.
1058 We precalculate and save this info here, since computing it is
1059 expensive.
1061 Do this before deleting any instructions from the loop, so that
1062 back_branch_in_range_p will work correctly. */
1066 else
1070 /* find_splittable_regs may have created some new registers, so must
1071 reallocate the reg_map with the new larger size, and must realloc
1072 the constant maps also. */
1079 /* Space is needed in some of the map for new registers, so new_maxregnum
1080 is an (over)estimate of how many registers will exist at the end. */
1083 /* Must realloc space for the constant maps, because the number of registers
1084 may have changed. */
1093 /* Search the list of bivs and givs to find ones which need to be remapped
1094 when split, and set their reg_map entry appropriately. */
1097 {
1100 #if 0
1101 /* Currently, non-reduced/final-value givs are never split. */
1105 #endif
1106 }
1108 /* Use our current register alignment and pointer flags. */
1112 /* If the loop is being partially unrolled, and the iteration variables
1113 are being split, and are being renamed for the split, then must fix up
1114 the compare/jump instruction at the end of the loop to refer to the new
1115 registers. This compare isn't copied, so the registers used in it
1116 will never be replaced if it isn't done here. */
1119 {
1123 }
1125 /* For unroll_number - 1 times, make a copy of each instruction
1126 between copy_start and copy_end, and insert these new instructions
1127 before the end of the loop. */
1130 {
1144 /* If loop starts with a branch to the test, then fix it so that
1145 it points to the test of the first unrolled copy of the loop. */
1147 {
1155 /* Set the jump label so that it can be used by later loop unrolling
1156 passes. */
1159 }
1164 }
1166 /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1167 insn to be deleted. This prevents any runaway delete_insn call from
1168 more insns that it should, as it always stops at a CODE_LABEL. */
1170 /* Delete the compare and branch at the end of the loop if completely
1171 unrolling the loop. Deleting the backward branch at the end also
1172 deletes the code label at the start of the loop. This is done at
1173 the very end to avoid problems with back_branch_in_range_p. */
1177 else
1180 /* Delete all of the original loop instructions. Don't delete the
1181 LOOP_BEG note, or the first code label in the loop. */
1185 {
1188 else
1190 }
1192 /* Can now delete the 'safety' label emitted to protect us from runaway
1193 delete_insn calls. */
1198 /* If exit_label exists, emit it after the loop. Doing the emit here
1199 forces it to have a higher INSN_UID than any insn in the unrolled loop.
1200 This is needed so that mostly_true_jump in reorg.c will treat jumps
1201 to this loop end label correctly, i.e. predict that they are usually
1202 not taken. */
1205 }
1206 \f
1207 /* Return true if the loop can be safely, and profitably, preconditioned
1208 so that the unrolled copies of the loop body don't need exit tests.
1210 This only works if final_value, initial_value and increment can be
1211 determined, and if increment is a constant power of 2.
1212 If increment is not a power of 2, then the preconditioning modulo
1213 operation would require a real modulo instead of a boolean AND, and this
1214 is not considered `profitable'. */
1216 /* ??? If the loop is known to be executed very many times, or the machine
1217 has a very cheap divide instruction, then preconditioning is a win even
1218 when the increment is not a power of 2. Use RTX_COST to compute
1219 whether divide is cheap. */
1221 static int
1223 loop_end)
1226 {
1229 {
1237 loop_n_iterations);
1239 }
1242 {
1247 }
1249 {
1254 }
1256 {
1261 }
1264 {
1269 }
1271 /* Unsigned_compare and compare_dir can be ignored here, since they do
1272 not matter for preconditioning. */
1275 {
1280 }
1282 /* Must ensure that final_value is invariant, so call invariant_p to
1283 check. Before doing so, must check regno against max_reg_before_loop
1284 to make sure that the register is in the range covered by invariant_p.
1285 If it isn't, then it is most likely a biv/giv which by definition are
1286 not invariant. */
1292 {
1297 }
1299 /* Fail for floating point values, since the caller of this function
1300 does not have code to deal with them. */
1303 {
1308 }
1310 /* Now set initial_value to be the iteration_var, since that may be a
1311 simpler expression, and is guaranteed to be correct if all of the
1312 above tests succeed.
1314 We can not use the initial_value as calculated, because it will be
1315 one too small for loops of the form "while (i-- > 0)". We can not
1316 emit code before the loop_skip_over insns to fix this problem as this
1317 will then give a number one too large for loops of the form
1318 "while (--i > 0)".
1320 Note that all loops that reach here are entered at the top, because
1321 this function is not called if the loop starts with a jump. */
1323 /* Fail if loop_iteration_var is not live before loop_start, since we need
1324 to test its value in the preconditioning code. */
1328 {
1333 }
1339 /* Success! */
1343 }
1346 /* All pseudo-registers must be mapped to themselves. Two hard registers
1347 must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1348 REGNUM, to avoid function-inlining specific conversions of these
1349 registers. All other hard regs can not be mapped because they may be
1350 used with different
1351 modes. */
1353 static void
1357 {
1362 /* Just clear the rest of the entries. */
1370 }
1371 \f
1372 /* Strength-reduction will often emit code for optimized biv/givs which
1373 calculates their value in a temporary register, and then copies the result
1374 to the iv. This procedure reconstructs the pattern computing the iv;
1375 verifying that all operands are of the proper form.
1377 The return value is the amount that the giv is incremented by. */
1379 static rtx
1383 {
1384 rtx increment;
1388 retry:
1389 /* Verify that we have an increment insn here. First check for a plus
1390 as the set source. */
1392 {
1393 /* SR sometimes computes the new giv value in a temp, then copies it
1394 to the new_reg. */
1400 /* The last insn emitted is not needed, so delete it to avoid confusing
1401 the second cse pass. This insn sets the giv unnecessarily. */
1403 }
1405 /* Verify that we have a constant as the second operand of the plus. */
1408 {
1409 /* SR sometimes puts the constant in a register, especially if it is
1410 too big to be an add immed operand. */
1414 /* SR may have used LO_SUM to compute the constant if it is too large
1415 for a load immed operand. In this case, the constant is in operand
1416 one of the LO_SUM rtx. */
1421 {
1422 /* The rs6000 port loads some constants with IOR.
1423 The alpha port loads some constants with ASHIFT. */
1429 /* Don't need the last insn anymore. */
1438 else
1440 }
1445 /* The insn loading the constant into a register is no longer needed,
1446 so delete it. */
1448 }
1452 else
1455 /* Check that the source register is the same as the register we expected
1456 to see as the source. If not, something is seriously wrong. */
1459 {
1460 /* Some machines (e.g. the romp), may emit two add instructions for
1461 certain constants, so lets try looking for another add immediately
1462 before this one if we have only seen one add insn so far. */
1465 {
1466 tries++;
1474 }
1477 }
1480 }
1482 /* Copy REG_NOTES, except for insn references, because not all insn_map
1483 entries are valid yet. We do need to copy registers now though, because
1484 the reg_map entries can change during copying. */
1486 static rtx
1488 rtx notes;
1490 {
1491 rtx copy;
1502 /* Don't substitute for these yet. */
1504 else
1510 }
1512 /* Fixup insn references in copied REG_NOTES. */
1514 static void
1516 rtx notes;
1518 {
1519 rtx note;
1524 }
1526 /* Copy each instruction in the loop, substituting from map as appropriate.
1527 This is very similar to a loop in expand_inline_function. */
1529 static void
1532 copy_notes_from)
1535 rtx exit_label;
1539 {
1547 /* If this isn't the last iteration, then map any references to the
1548 start_label to final_label. Final label will then be emitted immediately
1549 after the end of this loop body if it was ever used.
1551 If this is the last iteration, then map references to the start_label
1552 to itself. */
1554 {
1557 }
1558 else
1564 do
1565 {
1571 {
1577 /* Check to see if this is a giv that has been combined with
1578 some split address givs. (Combined in the sense that
1579 `combine_givs' in loop.c has put two givs in the same register.)
1580 In this case, we must search all givs based on the same biv to
1581 find the address givs. Then split the address givs.
1582 Do this before splitting the giv, since that may map the
1583 SET_DEST to a new register. */
1588 {
1596 /* Although the giv_inc amount is not needed here, we must call
1597 calculate_giv_inc here since it might try to delete the
1598 last insn emitted. If we wait until later to call it,
1599 we might accidentally delete insns generated immediately
1600 below by emit_unrolled_add. */
1604 /* Now find all address giv's that were combined with this
1605 giv 'v'. */
1608 {
1611 /* If this DEST_ADDR giv was not split, then ignore it. */
1615 /* Scale this_giv_inc if the multiplicative factors of
1616 the two givs are different. */
1626 {
1627 /* Must emit an insn to increment the split address
1628 giv. Add in the const_adjust field in case there
1629 was a constant eliminated from the address. */
1632 /* tv->dest_reg will be either a bare register,
1633 or else a register plus a constant. */
1636 else
1639 /* Check for shared address givs, and avoid
1640 incrementing the shared pseudo reg more than
1641 once. */
1643 {
1644 /* tv->dest_reg may actually be a (PLUS (REG)
1645 (CONST)) here, so we must call plus_constant
1646 to add the const_adjust amount before calling
1647 emit_unrolled_add below. */
1651 /* The constant could be too large for an add
1652 immediate, so can't directly emit an insn
1653 here. */
1656 }
1658 /* Reset the giv to be just the register again, in case
1659 it is used after the set we have just emitted.
1660 We must subtract the const_adjust factor added in
1661 above. */
1665 }
1666 }
1667 }
1669 /* If this is a setting of a splittable variable, then determine
1670 how to split the variable, create a new set based on this split,
1671 and set up the reg_map so that later uses of the variable will
1672 use the new split variable. */
1679 {
1684 /* Compute the increment value for the giv, if it wasn't
1685 already computed above. */
1693 {
1694 /* Completely unrolling the loop. Set the induction
1695 variable to a known constant value. */
1697 /* The value in splittable_regs may be an invariant
1698 value, so we must use plus_constant here. */
1703 {
1706 }
1707 else
1708 {
1709 /* The splittable_regs value must be a REG or a
1710 CONST_INT, so put the entire value in the giv_src_reg
1711 variable. */
1714 }
1715 }
1716 else
1717 {
1718 /* Partially unrolling loop. Create a new pseudo
1719 register for the iteration variable, and set it to
1720 be a constant plus the original register. Except
1721 on the last iteration, when the result has to
1722 go back into the original iteration var register. */
1724 /* Handle bivs which must be mapped to a new register
1725 when split. This happens for bivs which need their
1726 final value set before loop entry. The new register
1727 for the biv was stored in the biv's first struct
1728 induction entry by find_splittable_regs. */
1732 {
1735 }
1737 #if 0
1738 /* If non-reduced/final-value givs were split, then
1739 this would have to remap those givs also. See
1740 find_splittable_regs. */
1741 #endif
1748 /* Now split the induction variable by changing the dest
1749 of this insn to a new register, and setting its
1750 reg_map entry to point to this new register.
1752 If this is the last iteration, and this is the last insn
1753 that will update the iv, then reuse the original dest,
1754 to ensure that the iv will have the proper value when
1755 the loop exits or repeats.
1757 Using splittable_regs_updates here like this is safe,
1758 because it can only be greater than one if all
1759 instructions modifying the iv are always executed in
1760 order. */
1764 {
1768 }
1769 else
1771 }
1773 /* The constant being added could be too large for an add
1774 immediate, so can't directly emit an insn here. */
1778 }
1779 else
1780 {
1783 }
1786 #ifdef HAVE_cc0
1787 /* If this insn is setting CC0, it may need to look at
1788 the insn that uses CC0 to see what type of insn it is.
1789 In that case, the call to recog via validate_change will
1790 fail. So don't substitute constants here. Instead,
1791 do it when we emit the following insn.
1793 For example, see the pyr.md file. That machine has signed and
1794 unsigned compares. The compare patterns must check the
1795 following branch insn to see which what kind of compare to
1796 emit.
1798 If the previous insn set CC0, substitute constants on it as
1799 well. */
1802 else
1803 {
1808 }
1809 #else
1811 #endif
1813 /* Make split induction variable constants `permanent' since we
1814 know there are no backward branches across iteration variable
1815 settings which would invalidate this. */
1817 {
1823 }
1833 {
1834 /* This is a branch to the beginning of the loop; this is the
1835 last insn being copied; and this is not the last iteration.
1836 In this case, we want to change the original fall through
1837 case to be a branch past the end of the loop, and the
1838 original jump label case to fall_through. */
1841 {
1847 }
1848 else
1849 {
1850 rtx jmp;
1852 /* Can't do it by reversing the jump (probably because we
1853 couldn't reverse the conditions), so emit a new
1854 jump_insn after COPY, and redirect the jump around
1855 that. */
1865 }
1866 }
1868 #ifdef HAVE_cc0
1872 #endif
1875 /* Set the jump label of COPY correctly to avoid problems with
1876 later passes of unroll_loop, if INSN had jump label set. */
1878 {
1881 /* Can't use the label_map for every insn, since this may be
1882 the backward branch, and hence the label was not mapped. */
1884 {
1889 {
1892 else
1894 }
1895 }
1899 else
1900 {
1901 /* An unrecognizable jump insn, probably the entry jump
1902 for a switch statement. This label must have been mapped,
1903 so just use the label_map to get the new jump label. */
1906 }
1908 /* If this is a non-local jump, then must increase the label
1909 use count so that the label will not be deleted when the
1910 original jump is deleted. */
1912 }
1915 {
1923 }
1925 /* If this used to be a conditional jump insn but whose branch
1926 direction is now known, we must do something special. */
1928 {
1929 #ifdef HAVE_cc0
1930 /* The previous insn set cc0 for us. So delete it. */
1932 #endif
1934 /* If this is now a no-op, delete it. */
1936 {
1937 /* Don't let delete_insn delete the label referenced here,
1938 because we might possibly need it later for some other
1939 instruction in the loop. */
1946 }
1947 else
1948 /* Otherwise, this is unconditional jump so we must put a
1949 BARRIER after it. We could do some dead code elimination
1950 here, but jump.c will do it just as well. */
1952 }
1960 /* Because the USAGE information potentially contains objects other
1961 than hard registers, we need to copy it. */
1965 #ifdef HAVE_cc0
1969 #endif
1972 /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1978 /* If this is the loop start label, then we don't need to emit a
1979 copy of this label since no one will use it. */
1982 {
1985 }
1993 /* VTOP notes are valid only before the loop exit test. If placed
1994 anywhere else, loop may generate bad code. */
2001 else
2008 }
2011 }
2014 /* Now finish coping the REG_NOTES. */
2016 do
2017 {
2023 }
2026 /* There may be notes between copy_notes_from and loop_end. Emit a copy of
2027 each of these notes here, since there may be some important ones, such as
2028 NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
2029 iteration, because the original notes won't be deleted.
2031 We can't use insert_before here, because when from preconditioning,
2032 insert_before points before the loop. We can't use copy_end, because
2033 there may be insns already inserted after it (which we don't want to
2034 copy) when not from preconditioning code. */
2037 {
2039 {
2043 }
2044 }
2052 }
2053 \f
2054 /* Emit an insn, using the expand_binop to ensure that a valid insn is
2055 emitted. This will correctly handle the case where the increment value
2056 won't fit in the immediate field of a PLUS insns. */
2058 void
2061 {
2062 rtx result;
2069 }
2070 \f
2071 /* Searches the insns between INSN and LOOP_END. Returns 1 if there
2072 is a backward branch in that range that branches to somewhere between
2073 LOOP_START and INSN. Returns 0 otherwise. */
2075 /* ??? This is quadratic algorithm. Could be rewritten to be linear.
2076 In practice, this is not a problem, because this function is seldom called,
2077 and uses a negligible amount of CPU time on average. */
2079 int
2081 rtx insn;
2083 {
2086 /* Stop before we get to the backward branch at the end of the loop. */
2091 /* Check in case insn has been deleted, search forward for first non
2092 deleted insn following it. */
2096 /* Check for the case where insn is the last insn in the loop. */
2101 {
2103 {
2106 /* Search from loop_start to insn, to see if one of them is
2107 the target_insn. We can't use INSN_LUID comparisons here,
2108 since insn may not have an LUID entry. */
2112 }
2113 }
2116 }
2118 /* Try to generate the simplest rtx for the expression
2119 (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
2120 value of giv's. */
2122 static rtx
2126 {
2128 rtx result;
2130 /* The modes must all be the same. This should always be true. For now,
2131 check to make sure. */
2137 /* Ensure that if at least one of mult1/mult2 are constant, then mult2
2138 will be a constant. */
2140 {
2144 }
2150 /* Again, put the constant second. */
2152 {
2156 }
2163 }
2165 /* Searches the list of induction struct's for the biv BL, to try to calculate
2166 the total increment value for one iteration of the loop as a constant.
2168 Returns the increment value as an rtx, simplified as much as possible,
2169 if it can be calculated. Otherwise, returns 0. */
2171 rtx
2175 {
2177 rtx result;
2179 /* For increment, must check every instruction that sets it. Each
2180 instruction must be executed only once each time through the loop.
2181 To verify this, we check that the the insn is always executed, and that
2182 there are no backward branches after the insn that branch to before it.
2183 Also, the insn must have a mult_val of one (to make sure it really is
2184 an increment). */
2188 {
2192 else
2194 }
2197 }
2199 /* Determine the initial value of the iteration variable, and the amount
2200 that it is incremented each loop. Use the tables constructed by
2201 the strength reduction pass to calculate these values.
2203 Initial_value and/or increment are set to zero if their values could not
2204 be calculated. */
2206 static void
2210 {
2214 /* Clear the result values, in case no answer can be found. */
2218 /* The iteration variable can be either a giv or a biv. Check to see
2219 which it is, and compute the variable's initial value, and increment
2220 value if possible. */
2222 /* If this is a new register, can't handle it since we don't have any
2223 reg_iv_type entry for it. */
2225 {
2230 }
2231 /* Reject iteration variables larger than the host long size, since they
2232 could result in a number of iterations greater than the range of our
2233 `unsigned long' variable loop_n_iterations. */
2235 {
2240 }
2242 {
2247 }
2249 {
2250 /* Grab initial value, only useful if it is a constant. */
2255 }
2257 {
2258 #if 1
2259 /* ??? The code below does not work because the incorrect number of
2260 iterations is calculated when the biv is incremented after the giv
2261 is set (which is the usual case). This can probably be accounted
2262 for by biasing the initial_value by subtracting the amount of the
2263 increment that occurs between the giv set and the giv test. However,
2264 a giv as an iterator is very rare, so it does not seem worthwhile
2265 to handle this. */
2266 /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2271 #else
2272 /* Initial value is mult_val times the biv's initial value plus
2273 add_val. Only useful if it is a constant. */
2279 /* Increment value is mult_val times the increment value of the biv. */
2285 #endif
2286 }
2287 else
2288 {
2293 }
2294 }
2296 /* Calculate the approximate final value of the iteration variable
2297 which has an loop exit test with code COMPARISON_CODE and comparison value
2298 of COMPARISON_VALUE. Also returns an indication of whether the comparison
2299 was signed or unsigned, and the direction of the comparison. This info is
2300 needed to calculate the number of loop iterations. */
2302 static rtx
2305 rtx comparison_value;
2308 {
2309 /* Calculate the final value of the induction variable.
2310 The exact final value depends on the branch operator, and increment sign.
2311 This is only an approximate value. It will be wrong if the iteration
2312 variable is not incremented by one each time through the loop, and
2313 approx final value - start value % increment != 0. */
2317 {
2329 /* Can not calculate a final value for this case. */
2348 }
2349 }
2351 /* For each biv and giv, determine whether it can be safely split into
2352 a different variable for each unrolled copy of the loop body. If it
2353 is safe to split, then indicate that by saving some useful info
2354 in the splittable_regs array.
2356 If the loop is being completely unrolled, then splittable_regs will hold
2357 the current value of the induction variable while the loop is unrolled.
2358 It must be set to the initial value of the induction variable here.
2359 Otherwise, splittable_regs will hold the difference between the current
2360 value of the induction variable and the value the induction variable had
2361 at the top of the loop. It must be set to the value 0 here.
2363 Returns the total number of instructions that set registers that are
2364 splittable. */
2366 /* ?? If the loop is only unrolled twice, then most of the restrictions to
2367 constant values are unnecessary, since we can easily calculate increment
2368 values in this case even if nothing is constant. The increment value
2369 should not involve a multiply however. */
2371 /* ?? Even if the biv/giv increment values aren't constant, it may still
2372 be beneficial to split the variable if the loop is only unrolled a few
2373 times, since multiplies by small integers (1,2,3,4) are very cheap. */
2375 static int
2377 unroll_number)
2380 rtx end_insert_before;
2382 {
2386 rtx biv_final_value;
2391 {
2392 /* Biv_total_increment must return a constant value,
2393 otherwise we can not calculate the split values. */
2399 /* The loop must be unrolled completely, or else have a known number
2400 of iterations and only one exit, or else the biv must be dead
2401 outside the loop, or else the final value must be known. Otherwise,
2402 it is unsafe to split the biv since it may not have the proper
2403 value on loop exit. */
2405 /* loop_number_exit_count is non-zero if the loop has an exit other than
2406 a fall through at the end. */
2422 /* If any of the insns setting the BIV don't do so with a simple
2423 PLUS, we don't know how to split it. */
2431 /* If final value is non-zero, then must emit an instruction which sets
2432 the value of the biv to the proper value. This is done after
2433 handling all of the givs, since some of them may need to use the
2434 biv's value in their initialization code. */
2436 /* This biv is splittable. If completely unrolling the loop, save
2437 the biv's initial value. Otherwise, save the constant zero. */
2440 {
2442 {
2443 /* If the initial value of the biv is itself (i.e. it is too
2444 complicated for strength_reduce to compute), or is a hard
2445 register, or it isn't invariant, then we must create a new
2446 pseudo reg to hold the initial value of the biv. */
2452 {
2456 loop_start);
2463 }
2464 else
2466 }
2467 else
2470 /* Save the number of instructions that modify the biv, so that
2471 we can treat the last one specially. */
2479 }
2481 /* Check every giv that depends on this biv to see whether it is
2482 splittable also. Even if the biv isn't splittable, givs which
2483 depend on it may be splittable if the biv is live outside the
2484 loop, and the givs aren't. */
2489 /* If final value is non-zero, then must emit an instruction which sets
2490 the value of the biv to the proper value. This is done after
2491 handling all of the givs, since some of them may need to use the
2492 biv's value in their initialization code. */
2494 {
2495 /* If the loop has multiple exits, emit the insns before the
2496 loop to ensure that it will always be executed no matter
2497 how the loop exits. Otherwise emit the insn after the loop,
2498 since this is slightly more efficient. */
2501 biv_final_value),
2502 end_insert_before);
2503 else
2504 {
2505 /* Create a new register to hold the value of the biv, and then
2506 set the biv to its final value before the loop start. The biv
2507 is set to its final value before loop start to ensure that
2508 this insn will always be executed, no matter how the loop
2509 exits. */
2512 loop_start);
2514 biv_final_value),
2515 loop_start);
2521 /* Set up the mapping from the original biv register to the new
2522 register. */
2524 }
2525 }
2526 }
2528 }
2530 /* Return 1 if the first and last unrolled copy of the address giv V is valid
2531 for the instruction that is using it. Do not make any changes to that
2532 instruction. */
2534 static int
2537 rtx giv_inc;
2539 {
2545 /* First check to see if either address would fail. */
2550 /* Now put things back the way they were before. This will always
2551 succeed. */
2555 }
2557 /* For every giv based on the biv BL, check to determine whether it is
2558 splittable. This is a subroutine to find_splittable_regs ().
2560 Return the number of instructions that set splittable registers. */
2562 static int
2564 unroll_number)
2568 rtx increment;
2570 {
2572 rtx final_value;
2573 rtx tem;
2576 /* Scan the list of givs, and set the same_insn field when there are
2577 multiple identical givs in the same insn. */
2585 {
2588 /* Only split the giv if it has already been reduced, or if the loop is
2589 being completely unrolled. */
2593 /* The giv can be split if the insn that sets the giv is executed once
2594 and only once on every iteration of the loop. */
2595 /* An address giv can always be split. v->insn is just a use not a set,
2596 and hence it does not matter whether it is always executed. All that
2597 matters is that all the biv increments are always executed, and we
2598 won't reach here if they aren't. */
2604 /* The giv increment value must be a constant. */
2610 /* The loop must be unrolled completely, or else have a known number of
2611 iterations and only one exit, or else the giv must be dead outside
2612 the loop, or else the final value of the giv must be known.
2613 Otherwise, it is not safe to split the giv since it may not have the
2614 proper value on loop exit. */
2616 /* The used outside loop test will fail for DEST_ADDR givs. They are
2617 never used outside the loop anyways, so it is always safe to split a
2618 DEST_ADDR giv. */
2626 /* Check for the case where the pseudo is set by a shift/add
2627 sequence, in which case the first insn setting the pseudo
2628 is the first insn of the shift/add sequence. */
2632 /* Line above always fails if INSN was moved by loop opt. */
2638 #if 0
2639 /* Currently, non-reduced/final-value givs are never split. */
2640 /* Should emit insns after the loop if possible, as the biv final value
2641 code below does. */
2643 /* If the final value is non-zero, and the giv has not been reduced,
2644 then must emit an instruction to set the final value. */
2646 {
2647 /* Create a new register to hold the value of the giv, and then set
2648 the giv to its final value before the loop start. The giv is set
2649 to its final value before loop start to ensure that this insn
2650 will always be executed, no matter how we exit. */
2654 loop_start);
2661 }
2662 #endif
2664 /* This giv is splittable. If completely unrolling the loop, save the
2665 giv's initial value. Otherwise, save the constant zero for it. */
2668 {
2669 /* It is not safe to use bl->initial_value here, because it may not
2670 be invariant. It is safe to use the initial value stored in
2671 the splittable_regs array if it is set. In rare cases, it won't
2672 be set, so then we do exactly the same thing as
2673 find_splittable_regs does to get a safe value. */
2674 rtx biv_initial_value;
2682 else
2683 {
2687 loop_start);
2689 }
2692 }
2693 else
2697 {
2698 /* If a giv was combined with another giv, then we can only split
2699 this giv if the giv it was combined with was reduced. This
2700 is because the value of v->new_reg is meaningless in this
2701 case. */
2703 {
2708 }
2709 /* If the giv is an address destination, it could be something other
2710 than a simple register, these have to be treated differently. */
2712 {
2713 /* If value is not a constant, register, or register plus
2714 constant, then compute its value into a register before
2715 loop start. This prevents invalid rtx sharing, and should
2716 generate better code. We can use bl->initial_value here
2717 instead of splittable_regs[bl->regno] because this code
2718 is going before the loop start. */
2725 {
2730 }
2733 }
2734 else
2735 {
2736 /* Splitting address givs is useful since it will often allow us
2737 to eliminate some increment insns for the base giv as
2738 unnecessary. */
2740 /* If the addr giv is combined with a dest_reg giv, then all
2741 references to that dest reg will be remapped, which is NOT
2742 what we want for split addr regs. We always create a new
2743 register for the split addr giv, just to be safe. */
2745 /* ??? If there are multiple address givs which have been
2746 combined with the same dest_reg giv, then we may only need
2747 one new register for them. Pulling out constants below will
2748 catch some of the common cases of this. Currently, I leave
2749 the work of simplifying multiple address givs to the
2750 following cse pass. */
2752 /* As a special case, if we have multiple identical address givs
2753 within a single instruction, then we do use a single pseudo
2754 reg for both. This is necessary in case one is a match_dup
2755 of the other. */
2760 {
2766 }
2768 {
2769 /* If not completely unrolling the loop, then create a new
2770 register to hold the split value of the DEST_ADDR giv.
2771 Emit insn to initialize its value before loop start. */
2774 /* If the address giv has a constant in its new_reg value,
2775 then this constant can be pulled out and put in value,
2776 instead of being part of the initialization code. */
2780 {
2781 v->dest_reg
2784 /* Only succeed if this will give valid addresses.
2785 Try to validate both the first and the last
2786 address resulting from loop unrolling, if
2787 one fails, then can't do const elim here. */
2789 {
2790 /* Save the negative of the eliminated const, so
2791 that we can calculate the dest_reg's increment
2792 value later. */
2800 }
2801 else
2803 }
2804 else
2807 /* If the address hasn't been checked for validity yet, do so
2808 now, and fail completely if either the first or the last
2809 unrolled copy of the address is not a valid address
2810 for the instruction that uses it. */
2813 {
2819 }
2821 /* To initialize the new register, just move the value of
2822 new_reg into it. This is not guaranteed to give a valid
2823 instruction on machines with complex addressing modes.
2824 If we can't recognize it, then delete it and emit insns
2825 to calculate the value from scratch. */
2828 loop_start);
2830 {
2833 /* We can't use bl->initial_value to compute the initial
2834 value, because the loop may have been preconditioned.
2835 We must calculate it from NEW_REG. Try using
2836 force_operand instead of emit_iv_add_mult. */
2850 }
2851 }
2852 else
2853 {
2856 /* Check the resulting address for validity, and fail
2857 if the resulting address would be invalid. */
2859 {
2865 }
2866 }
2868 /* Store the value of dest_reg into the insn. This sharing
2869 will not be a problem as this insn will always be copied
2870 later. */
2874 /* If this address giv is combined with a dest reg giv, then
2875 save the base giv's induction pointer so that we will be
2876 able to handle this address giv properly. The base giv
2877 itself does not have to be splittable. */
2883 {
2884 /* This giv maybe hasn't been combined with any others.
2885 Make sure that it's giv is marked as splittable here. */
2889 /* Make it appear to depend upon itself, so that the
2890 giv will be properly split in the main loop above. */
2892 {
2895 }
2896 }
2900 }
2901 }
2902 else
2903 {
2904 #if 0
2905 /* Currently, unreduced giv's can't be split. This is not too much
2906 of a problem since unreduced giv's are not live across loop
2907 iterations anyways. When unrolling a loop completely though,
2908 it makes sense to reduce&split givs when possible, as this will
2909 result in simpler instructions, and will not require that a reg
2910 be live across loop iterations. */
2915 #else
2917 #endif
2918 }
2920 /* Givs are only updated once by definition. Mark it so if this is
2921 a splittable register. Don't need to do anything for address givs
2922 where this may not be a register. */
2927 result++;
2930 {
2937 else
2941 }
2942 }
2945 }
2946 \f
2947 /* Try to prove that the register is dead after the loop exits. Trace every
2948 loop exit looking for an insn that will always be executed, which sets
2949 the register to some value, and appears before the first use of the register
2950 is found. If successful, then return 1, otherwise return 0. */
2952 /* ?? Could be made more intelligent in the handling of jumps, so that
2953 it can search past if statements and other similar structures. */
2955 static int
2958 {
2965 /* In addition to checking all exits of this loop, we must also check
2966 all exits of inner nested loops that would exit this loop. We don't
2967 have any way to identify those, so we just give up if there are any
2968 such inner loop exits. */
2972 label_count++;
2977 /* HACK: Must also search the loop fall through exit, create a label_ref
2978 here which points to the loop_end, and append the loop_number_exit_labels
2979 list to it. */
2984 {
2985 /* Succeed if find an insn which sets the biv or if reach end of
2986 function. Fail if find an insn that uses the biv, or if come to
2987 a conditional jump. */
2991 {
2994 {
2995 rtx set;
3003 }
3006 {
3010 /* Prevent infinite loop following infinite loops. */
3013 else
3015 }
3018 }
3019 }
3021 /* Success, the register is dead on all loop exits. */
3023 }
3025 /* Try to calculate the final value of the biv, the value it will have at
3026 the end of the loop. If we can do it, return that value. */
3028 rtx
3032 {
3035 /* ??? This only works for MODE_INT biv's. Reject all others for now. */
3040 /* The final value for reversed bivs must be calculated differently than
3041 for ordinary bivs. In this case, there is already an insn after the
3042 loop which sets this biv's final value (if necessary), and there are
3043 no other loop exits, so we can return any value. */
3045 {
3051 }
3053 /* Try to calculate the final value as initial value + (number of iterations
3054 * increment). For this to work, increment must be invariant, the only
3055 exit from the loop must be the fall through at the bottom (otherwise
3056 it may not have its final value when the loop exits), and the initial
3057 value of the biv must be invariant. */
3062 {
3066 {
3067 /* Can calculate the loop exit value, emit insns after loop
3068 end to calculate this value into a temporary register in
3069 case it is needed later. */
3072 /* Make sure loop_end is not the last insn. */
3083 }
3084 }
3086 /* Check to see if the biv is dead at all loop exits. */
3088 {
3095 }
3098 }
3100 /* Try to calculate the final value of the giv, the value it will have at
3101 the end of the loop. If we can do it, return that value. */
3103 rtx
3107 {
3109 rtx insn;
3115 /* The final value for givs which depend on reversed bivs must be calculated
3116 differently than for ordinary givs. In this case, there is already an
3117 insn after the loop which sets this giv's final value (if necessary),
3118 and there are no other loop exits, so we can return any value. */
3120 {
3126 }
3128 /* Try to calculate the final value as a function of the biv it depends
3129 upon. The only exit from the loop must be the fall through at the bottom
3130 (otherwise it may not have its final value when the loop exits). */
3132 /* ??? Can calculate the final giv value by subtracting off the
3133 extra biv increments times the giv's mult_val. The loop must have
3134 only one exit for this to work, but the loop iterations does not need
3135 to be known. */
3139 {
3140 /* ?? It is tempting to use the biv's value here since these insns will
3141 be put after the loop, and hence the biv will have its final value
3142 then. However, this fails if the biv is subsequently eliminated.
3143 Perhaps determine whether biv's are eliminable before trying to
3144 determine whether giv's are replaceable so that we can use the
3145 biv value here if it is not eliminable. */
3150 {
3151 /* Can calculate the loop exit value of its biv as
3152 (loop_n_iterations * increment) + initial_value */
3154 /* The loop exit value of the giv is then
3155 (final_biv_value - extra increments) * mult_val + add_val.
3156 The extra increments are any increments to the biv which
3157 occur in the loop after the giv's value is calculated.
3158 We must search from the insn that sets the giv to the end
3159 of the loop to calculate this value. */
3163 /* Put the final biv value in tem. */
3168 /* Subtract off extra increments as we find them. */
3171 {
3176 {
3180 OPTAB_LIB_WIDEN);
3184 }
3185 }
3187 /* Now calculate the giv's final value. */
3189 insert_before);
3197 }
3198 }
3200 /* Replaceable giv's should never reach here. */
3204 /* Check to see if the biv is dead at all loop exits. */
3206 {
3213 }
3216 }
3219 /* Calculate the number of loop iterations. Returns the exact number of loop
3220 iterations if it can be calculated, otherwise returns zero. */
3222 unsigned HOST_WIDE_INT
3225 {
3229 HOST_WIDE_INT i;
3233 rtx last_loop_insn;
3235 /* First find the iteration variable. If the last insn is a conditional
3236 branch, and the insn before tests a register value, make that the
3237 iteration variable. */
3244 /* We used to use pren_nonnote_insn here, but that fails because it might
3245 accidentally get the branch for a contained loop if the branch for this
3246 loop was deleted. We can only trust branches immediately before the
3247 loop_end. */
3252 {
3257 }
3259 /* ??? Get_condition may switch position of induction variable and
3260 invariant register when it canonicalizes the comparison. */
3267 {
3272 }
3274 /* Loop iterations is always called before any new registers are created
3275 now, so this should never occur. */
3283 /* iteration_info already printed a message. */
3286 /* If the comparison value is an invariant register, then try to find
3287 its value from the insns before the start of the loop. */
3290 {
3294 {
3300 {
3301 /* We found the last insn before the loop that sets the register.
3302 If it sets the entire register, and has a REG_EQUAL note,
3303 then use the value of the REG_EQUAL note. */
3306 {
3309 /* Only use the REG_EQUAL note if it is a constant.
3310 Other things, divide in particular, will cause
3311 problems later if we use them. */
3315 }
3317 }
3318 }
3319 }
3324 /* Save the calculated values describing this loop's bounds, in case
3325 precondition_loop_p will need them later. These values can not be
3326 recalculated inside precondition_loop_p because strength reduction
3327 optimizations may obscure the loop's structure. */
3335 {
3340 }
3342 {
3347 }
3349 {
3354 }
3356 {
3361 }
3363 {
3368 }
3370 /* ?? Final value and initial value do not have to be constants.
3371 Only their difference has to be constant. When the iteration variable
3372 is an array address, the final value and initial value might both
3373 be addresses with the same base but different constant offsets.
3374 Final value must be invariant for this to work.
3376 To do this, need some way to find the values of registers which are
3377 invariant. */
3379 /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3381 final_larger
3386 else
3394 else
3397 /* There are 27 different cases: compare_dir = -1, 0, 1;
3398 final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3399 There are 4 normal cases, 4 reverse cases (where the iteration variable
3400 will overflow before the loop exits), 4 infinite loop cases, and 15
3401 immediate exit (0 or 1 iteration depending on loop type) cases.
3402 Only try to optimize the normal cases. */
3404 /* (compare_dir/final_larger/increment_dir)
3405 Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3406 Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3407 Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3408 Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3410 /* ?? If the meaning of reverse loops (where the iteration variable
3411 will overflow before the loop exits) is undefined, then could
3412 eliminate all of these special checks, and just always assume
3413 the loops are normal/immediate/infinite. Note that this means
3414 the sign of increment_dir does not have to be known. Also,
3415 since it does not really hurt if immediate exit loops or infinite loops
3416 are optimized, then that case could be ignored also, and hence all
3417 loops can be optimized.
3419 According to ANSI Spec, the reverse loop case result is undefined,
3420 because the action on overflow is undefined.
3422 See also the special test for NE loops below. */
3426 /* Normal case. */
3427 ;
3428 else
3429 {
3434 }
3436 /* Calculate the number of iterations, final_value is only an approximation,
3437 so correct for that. Note that tempu and loop_n_iterations are
3438 unsigned, because they can be as large as 2^n - 1. */
3444 {
3447 }
3448 else
3451 /* For NE tests, make sure that the iteration variable won't miss the
3452 final value. If tempu mod i is not zero, then the iteration variable
3453 will overflow before the loop exits, and we can not calculate the
3454 number of iterations. */
3459 }
3461 /* Replace uses of split bivs with their split pseudo register. This is
3462 for original instructions which remain after loop unrolling without
3463 copying. */
3465 static rtx
3467 rtx x;
3468 {
3478 {
3490 #if 0
3491 /* If non-reduced/final-value givs were split, then this would also
3492 have to remap those givs also. */
3493 #endif
3497 }
3501 {
3505 {
3509 }
3510 }
3512 }