1 /* Perform doloop optimizations
2 Copyright (C) 1999, 2000 Free Software Foundation, Inc.
3 Contributed by Michael P. Hayes (m.hayes@elec.canterbury.ac.nz)
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)
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. */
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
34 /* This module is used to modify loops with a determinable number of
35 iterations to use special low-overhead looping instructions.
37 It first validates whether the loop is well behaved and has a
38 determinable number of iterations (either at compile or run-time).
39 It then modifies the loop to use a low-overhead looping pattern as
42 1. A pseudo register is allocated as the loop iteration counter.
44 2. The number of loop iterations is calculated and is stored
47 3. At the end of the loop, the jump insn is replaced by the
48 doloop_end pattern. The compare must remain because it might be
49 used elsewhere. If the loop-variable or condition register are
50 used elsewhere, they will be eliminated by flow.
52 4. An optional doloop_begin pattern is inserted at the top of the
57 #ifdef HAVE_doloop_end
59 static rtx doloop_condition_get
61 static unsigned HOST_WIDE_INT doloop_iterations_max
62 PARAMS ((const struct loop_info *, enum machine_mode, int));
63 static int doloop_valid_p
64 PARAMS ((const struct loop *, rtx));
65 static int doloop_modify
66 PARAMS ((const struct loop *, rtx, rtx, rtx, rtx, rtx));
67 static int doloop_modify_runtime
68 PARAMS ((const struct loop *, rtx, rtx, rtx, enum machine_mode, rtx));
71 /* Return the loop termination condition for PATTERN or zero
72 if it is not a decrement and branch jump insn. */
74 doloop_condition_get (pattern)
82 /* The canonical doloop pattern we expect is:
84 (parallel [(set (pc) (if_then_else (condition)
87 (set (reg) (plus (reg) (const_int -1)))
88 (additional clobbers and uses)])
90 Some machines (IA-64) make the decrement conditional on
91 the condition as well, so we don't bother verifying the
92 actual decrement. In summary, the branch must be the
93 first entry of the parallel (also required by jump.c),
94 and the second entry of the parallel must be a set of
95 the loop counter register. */
97 if (GET_CODE (pattern) != PARALLEL)
100 cmp = XVECEXP (pattern, 0, 0);
101 inc = XVECEXP (pattern, 0, 1);
103 /* Check for (set (reg) (something)). */
104 if (GET_CODE (inc) != SET || ! REG_P (SET_DEST (inc)))
107 /* Extract loop counter register. */
108 reg = SET_DEST (inc);
110 /* Check for (set (pc) (if_then_else (condition)
113 if (GET_CODE (cmp) != SET
114 || SET_DEST (cmp) != pc_rtx
115 || GET_CODE (SET_SRC (cmp)) != IF_THEN_ELSE
116 || GET_CODE (XEXP (SET_SRC (cmp), 1)) != LABEL_REF
117 || XEXP (SET_SRC (cmp), 2) != pc_rtx)
120 /* Extract loop termination condition. */
121 condition = XEXP (SET_SRC (cmp), 0);
123 if ((GET_CODE (condition) != GE && GET_CODE (condition) != NE)
124 || GET_CODE (XEXP (condition, 1)) != CONST_INT)
127 if (XEXP (condition, 0) == reg)
130 if (GET_CODE (XEXP (condition, 0)) == PLUS
131 && XEXP (XEXP (condition, 0), 0) == reg)
134 /* ??? If a machine uses a funny comparison, we could return a
135 canonicalised form here. */
141 /* Return an estimate of the maximum number of loop iterations for the
142 loop specified by LOOP or zero if the loop is not normal.
143 MODE is the mode of the iteration count and NONNEG is non-zero if
144 the the iteration count has been proved to be non-negative. */
145 static unsigned HOST_WIDE_INT
146 doloop_iterations_max (loop_info, mode, nonneg)
147 const struct loop_info *loop_info;
148 enum machine_mode mode;
151 unsigned HOST_WIDE_INT n_iterations_max;
155 HOST_WIDE_INT abs_inc;
159 abs_inc = INTVAL (loop_info->increment);
168 code = swap_condition (loop_info->comparison_code);
169 min_value = loop_info->final_equiv_value;
170 max_value = loop_info->initial_equiv_value;
174 code = loop_info->comparison_code;
175 min_value = loop_info->initial_equiv_value;
176 max_value = loop_info->final_equiv_value;
179 /* Since the loop has a VTOP, we know that the initial test will be
180 true and thus the value of max_value should be greater than the
181 value of min_value. Thus the difference should always be positive
182 and the code must be LT, LE, LTU, LEU, or NE. Otherwise the loop is
183 not normal, e.g., `for (i = 0; i < 10; i--)'. */
189 unsigned HOST_WIDE_INT umax;
190 unsigned HOST_WIDE_INT umin;
192 if (GET_CODE (min_value) == CONST_INT)
193 umin = INTVAL (min_value);
197 if (GET_CODE (max_value) == CONST_INT)
198 umax = INTVAL (max_value);
200 umax = ((unsigned)2 << (GET_MODE_BITSIZE (mode) - 1)) - 1;
202 n_iterations_max = umax - umin;
212 if (GET_CODE (min_value) == CONST_INT)
213 smin = INTVAL (min_value);
215 smin = -((unsigned)1 << (GET_MODE_BITSIZE (mode) - 1));
217 if (GET_CODE (max_value) == CONST_INT)
218 smax = INTVAL (max_value);
220 smax = ((unsigned)1 << (GET_MODE_BITSIZE (mode) - 1)) - 1;
222 n_iterations_max = smax - smin;
227 if (GET_CODE (min_value) == CONST_INT
228 && GET_CODE (max_value) == CONST_INT)
229 n_iterations_max = INTVAL (max_value) - INTVAL (min_value);
231 /* We need to conservatively assume that we might have the maximum
232 number of iterations without any additional knowledge. */
233 n_iterations_max = ((unsigned)2 << (GET_MODE_BITSIZE (mode) - 1)) - 1;
240 n_iterations_max /= abs_inc;
242 /* If we know that the iteration count is non-negative then adjust
243 n_iterations_max if it is so large that it appears negative. */
245 && n_iterations_max > ((unsigned)1 << (GET_MODE_BITSIZE (mode) - 1)))
246 n_iterations_max = ((unsigned)1 << (GET_MODE_BITSIZE (mode) - 1)) - 1;
248 return n_iterations_max;
252 /* Return non-zero if the loop specified by LOOP is suitable for
253 the use of special low-overhead looping instructions. */
255 doloop_valid_p (loop, jump_insn)
256 const struct loop *loop;
259 const struct loop_info *loop_info = LOOP_INFO (loop);
261 /* The loop must have a conditional jump at the end. */
262 if (! any_condjump_p (jump_insn)
263 || ! onlyjump_p (jump_insn))
265 if (loop_dump_stream)
266 fprintf (loop_dump_stream,
267 "Doloop: Invalid jump at loop end.\n");
271 /* Give up if a loop has been completely unrolled. */
272 if (loop_info->n_iterations == loop_info->unroll_number)
274 if (loop_dump_stream)
275 fprintf (loop_dump_stream,
276 "Doloop: Loop completely unrolled.\n");
280 /* The loop must have a single exit target. A break or return
281 statement within a loop will generate multiple loop exits.
282 Another example of a loop that currently generates multiple exit
283 targets is for (i = 0; i < (foo ? 8 : 4); i++) { }. */
284 if (loop_info->has_multiple_exit_targets || loop->exit_count)
286 if (loop_dump_stream)
287 fprintf (loop_dump_stream,
288 "Doloop: Loop has multiple exit targets.\n");
292 /* An indirect jump may jump out of the loop. */
293 if (loop_info->has_indirect_jump)
295 if (loop_dump_stream)
296 fprintf (loop_dump_stream,
297 "Doloop: Indirect jump in function.\n");
301 /* A called function may clobber any special registers required for
302 low-overhead looping. */
303 if (loop_info->has_call)
305 if (loop_dump_stream)
306 fprintf (loop_dump_stream,
307 "Doloop: Function call in loop.\n");
311 /* Some targets (eg, PPC) use the count register for branch on table
312 instructions. ??? This should be a target specific check. */
313 if (loop_info->has_tablejump)
315 if (loop_dump_stream)
316 fprintf (loop_dump_stream,
317 "Doloop: Computed branch in the loop.\n");
321 if (! loop_info->increment)
323 if (loop_dump_stream)
324 fprintf (loop_dump_stream,
325 "Doloop: Could not determine iteration info.\n");
329 if (GET_CODE (loop_info->increment) != CONST_INT)
331 if (loop_dump_stream)
332 fprintf (loop_dump_stream,
333 "Doloop: Increment not an integer constant.\n");
337 /* There is no guarantee that a NE loop will terminate if the
338 absolute increment is not unity. ??? We could compute this
339 condition at run-time and have a additional jump around the loop
340 to ensure an infinite loop. */
341 if (loop_info->comparison_code == NE
342 && INTVAL (loop_info->increment) != -1
343 && INTVAL (loop_info->increment) != 1)
345 if (loop_dump_stream)
346 fprintf (loop_dump_stream,
347 "Doloop: NE loop with non-unity increment.\n");
351 /* Check for loops that may not terminate under special conditions. */
352 if (! loop_info->n_iterations
353 && ((loop_info->comparison_code == LEU
354 && INTVAL (loop_info->increment) > 0)
355 || (loop_info->comparison_code == GEU
356 && INTVAL (loop_info->increment) < 0)))
358 /* If the comparison is LEU and the comparison value is UINT_MAX
359 then the loop will not terminate. Similarly, if the
360 comparison code is GEU and the initial value is 0, the loop
363 Note that with LE and GE, the loop behaviour can be
364 implementation dependent if an overflow occurs, say between
365 INT_MAX and INT_MAX + 1. We thus don't have to worry about
368 ??? We could compute these conditions at run-time and have a
369 additional jump around the loop to ensure an infinite loop.
370 However, it is very unlikely that this is the intended
371 behaviour of the loop and checking for these rare boundary
372 conditions would pessimize all other code. */
373 if (loop_dump_stream)
374 fprintf (loop_dump_stream,
375 "Doloop: Possible infinite iteration case ignored.\n");
382 /* Modify the loop to use the low-overhead looping insn where LOOP
383 describes the loop, ITERATIONS is an RTX containing the desired
384 number of loop iterations, ITERATIONS_MAX is a CONST_INT specifying
385 the maximum number of loop iterations, and DOLOOP_INSN is the
386 low-overhead looping insn to emit at the end of the loop. This
387 returns non-zero if it was successful. */
389 doloop_modify (loop, iterations, iterations_max,
390 doloop_seq, start_label, condition)
391 const struct loop *loop;
405 jump_insn = prev_nonnote_insn (loop->end);
407 if (loop_dump_stream)
409 fprintf (loop_dump_stream, "Doloop: Inserting doloop pattern (");
410 if (GET_CODE (iterations) == CONST_INT)
411 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
412 INTVAL (iterations));
414 fputs ("runtime", loop_dump_stream);
415 fputs (" iterations).", loop_dump_stream);
418 /* Emit the label that will delimit the top of the loop.
419 This has to be done before the delete_insn call below, to prevent
420 delete_insn from deleting too much. */
421 emit_label_after (start_label, loop->top ? loop->top : loop->start);
422 LABEL_NUSES (start_label)++;
424 /* Discard original jump to continue loop. The original compare
425 result may still be live, so it cannot be discarded explicitly. */
426 delete_insn (jump_insn);
428 counter_reg = XEXP (condition, 0);
429 if (GET_CODE (counter_reg) == PLUS)
430 counter_reg = XEXP (counter_reg, 0);
436 switch (GET_CODE (condition))
439 /* Currently only NE tests against zero and one are supported. */
440 if (XEXP (condition, 1) == const0_rtx)
442 else if (XEXP (condition, 1) != const1_rtx)
447 /* Currently only GE tests against zero are supported. */
448 if (XEXP (condition, 1) != const0_rtx)
451 /* The iteration count needs decrementing for a GE test. */
454 /* Determine if the iteration counter will be non-negative.
455 Note that the maximum value loaded is iterations_max - 1. */
456 if ((unsigned HOST_WIDE_INT) INTVAL (iterations_max)
457 <= ((unsigned)1 << (GET_MODE_BITSIZE (GET_MODE (counter_reg)) - 1)))
461 /* Abort if an invalid doloop pattern has been generated. */
468 if (GET_CODE (count) == CONST_INT)
469 count = GEN_INT (INTVAL (count) - 1);
471 count = expand_simple_binop (GET_MODE (counter_reg), MINUS,
473 0, 0, OPTAB_LIB_WIDEN);
476 /* Insert initialization of the count register into the loop header. */
477 convert_move (counter_reg, count, 1);
478 sequence = gen_sequence ();
480 emit_insn_before (sequence, loop->start);
482 /* Some targets (eg, C4x) need to initialize special looping
484 #ifdef HAVE_doloop_begin
488 init = gen_doloop_begin (counter_reg,
489 GET_CODE (iterations) == CONST_INT
490 ? iterations : const0_rtx, iterations_max,
491 GEN_INT (loop->level));
496 sequence = gen_sequence ();
498 emit_insn_after (sequence, loop->start);
503 /* Insert the new low-overhead looping insn. */
504 emit_jump_insn_before (doloop_seq, loop->end);
505 jump_insn = prev_nonnote_insn (loop->end);
506 JUMP_LABEL (jump_insn) = start_label;
508 /* Add a REG_NONNEG note if the actual or estimated maximum number
509 of iterations is non-negative. */
512 REG_NOTES (jump_insn)
513 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (jump_insn));
519 /* Handle the more complex case, where the bounds are not known at
520 compile time. In this case we generate a run_time calculation of
521 the number of iterations. We rely on the existence of a run-time
522 guard to ensure that the loop executes at least once, i.e.,
523 initial_value obeys the loop comparison condition. If a guard is
524 not present, we emit one. The loop to modify is described by LOOP.
525 ITERATIONS_MAX is a CONST_INT specifying the estimated maximum
526 number of loop iterations. DOLOOP_INSN is the low-overhead looping
527 insn to insert. Returns non-zero if loop successfully modified. */
529 doloop_modify_runtime (loop, iterations_max,
530 doloop_seq, start_label, mode, condition)
531 const struct loop *loop;
535 enum machine_mode mode;
538 const struct loop_info *loop_info = LOOP_INFO (loop);
539 HOST_WIDE_INT abs_inc;
548 enum rtx_code comparison_code;
550 increment = loop_info->increment;
551 initial_value = loop_info->initial_value;
552 final_value = loop_info->final_value;
555 abs_inc = INTVAL (increment);
562 comparison_code = loop_info->comparison_code;
563 unsigned_p = (comparison_code == LTU
564 || comparison_code == LEU
565 || comparison_code == GTU
566 || comparison_code == GEU
567 || comparison_code == NE);
569 /* The number of iterations (prior to any loop unrolling) is given by:
571 n = (abs (final - initial) + abs_inc - 1) / abs_inc.
573 However, it is possible for the summation to overflow, and a
576 n = abs (final - initial) / abs_inc;
577 n += (abs (final - initial) % abs_inc) != 0;
579 If the loop has been unrolled, then the loop body has been
580 preconditioned to iterate a multiple of unroll_number times. If
581 abs_inc is != 1, the full calculation is
583 t1 = abs_inc * unroll_number;
584 n = abs (final - initial) / t1;
585 n += (abs (final - initial) % t1) > t1 - abs_inc;
587 The division and modulo operations can be avoided by requiring
588 that the increment is a power of 2 (precondition_loop_p enforces
589 this requirement). Nevertheless, the RTX_COSTS should be checked
590 to see if a fast divmod is available. */
593 /* abs (final - initial) */
594 diff = expand_simple_binop (mode, MINUS,
595 copy_rtx (neg_inc ? initial_value : final_value),
596 copy_rtx (neg_inc ? final_value : initial_value),
597 NULL_RTX, unsigned_p, OPTAB_LIB_WIDEN);
599 if (abs_inc * loop_info->unroll_number != 1)
604 unsigned HOST_WIDE_INT limit;
606 shift_count = exact_log2 (abs_inc * loop_info->unroll_number);
610 /* abs (final - initial) / (abs_inc * unroll_number) */
611 iterations = expand_simple_binop (GET_MODE (diff), LSHIFTRT,
612 diff, GEN_INT (shift_count),
618 /* abs (final - initial) % (abs_inc * unroll_number) */
619 rtx count = GEN_INT (abs_inc * loop_info->unroll_number - 1);
620 extra = expand_simple_binop (GET_MODE (iterations), AND,
621 diff, count, NULL_RTX, 1,
624 /* If (abs (final - initial) % (abs_inc * unroll_number)
625 <= abs_inc * (unroll - 1)),
626 jump past following increment instruction. */
627 label = gen_label_rtx();
628 limit = abs_inc * (loop_info->unroll_number - 1);
629 emit_cmp_and_jump_insns (extra, GEN_INT (limit),
630 limit == 0 ? EQ : LEU, NULL_RTX,
631 GET_MODE (extra), 0, 0, label);
632 JUMP_LABEL (get_last_insn ()) = label;
633 LABEL_NUSES (label)++;
635 /* Increment the iteration count by one. */
636 iterations = expand_simple_binop (GET_MODE (iterations), PLUS,
637 iterations, GEN_INT (1),
647 /* If there is a NOTE_INSN_LOOP_VTOP, we have a `for' or `while'
648 style loop, with a loop exit test at the start. Thus, we can
649 assume that the loop condition was true when the loop was
652 `do-while' loops require special treatment since the exit test is
653 not executed before the start of the loop. We need to determine
654 if the loop will terminate after the first pass and to limit the
655 iteration count to one if necessary. */
660 if (loop_dump_stream)
661 fprintf (loop_dump_stream, "Doloop: Do-while loop.\n");
663 /* A `do-while' loop must iterate at least once. If the
664 iteration count is bogus, we set the iteration count to 1.
665 Note that if the loop has been unrolled, then the loop body
666 is guaranteed to execute at least once. */
667 if (loop_info->unroll_number == 1)
669 /* Emit insns to test if the loop will immediately
670 terminate and to set the iteration count to 1 if true. */
671 label = gen_label_rtx();
672 emit_cmp_and_jump_insns (copy_rtx (initial_value),
673 copy_rtx (loop_info->comparison_value),
674 comparison_code, NULL_RTX, mode, 0, 0,
676 JUMP_LABEL (get_last_insn ()) = label;
677 LABEL_NUSES (label)++;
678 emit_move_insn (iterations, const1_rtx);
683 sequence = gen_sequence ();
685 emit_insn_before (sequence, loop->start);
687 return doloop_modify (loop, iterations, iterations_max, doloop_seq,
688 start_label, condition);
692 /* This is the main entry point. Process loop described by LOOP
693 validating that the loop is suitable for conversion to use a low
694 overhead looping instruction, replacing the jump insn where
695 suitable. We distinguish between loops with compile-time bounds
696 and those with run-time bounds. Information from LOOP is used to
697 compute the number of iterations and to determine whether the loop
698 is a candidate for this optimization. Returns non-zero if loop
699 successfully modified. */
701 doloop_optimize (loop)
702 const struct loop *loop;
704 struct loop_info *loop_info = LOOP_INFO (loop);
709 enum machine_mode mode;
710 unsigned HOST_WIDE_INT n_iterations;
711 unsigned HOST_WIDE_INT n_iterations_max;
712 rtx doloop_seq, doloop_pat, doloop_reg;
718 if (loop_dump_stream)
719 fprintf (loop_dump_stream,
720 "Doloop: Processing loop %d, enclosed levels %d.\n",
721 loop->num, loop->level);
723 jump_insn = prev_nonnote_insn (loop->end);
725 /* Check that loop is a candidate for a low-overhead looping insn. */
726 if (! doloop_valid_p (loop, jump_insn))
729 /* Determine if the loop can be safely, and profitably,
730 preconditioned. While we don't precondition the loop in a loop
731 unrolling sense, this test ensures that the loop is well behaved
732 and that the increment is a constant integer. */
733 if (! precondition_loop_p (loop, &initial_value, &final_value,
736 if (loop_dump_stream)
737 fprintf (loop_dump_stream,
738 "Doloop: Cannot precondition loop.\n");
742 /* Determine or estimate the maximum number of loop iterations. */
743 n_iterations = loop_info->n_iterations;
746 /* This is the simple case where the initial and final loop
747 values are constants. */
748 n_iterations_max = n_iterations;
752 int nonneg = find_reg_note (jump_insn, REG_NONNEG, 0) != 0;
754 /* This is the harder case where the initial and final loop
755 values may not be constants. */
756 n_iterations_max = doloop_iterations_max (loop_info, mode, nonneg);
758 if (! n_iterations_max)
760 /* We have something like `for (i = 0; i < 10; i--)'. */
761 if (loop_dump_stream)
762 fprintf (loop_dump_stream,
763 "Doloop: Not normal loop.\n");
768 /* Account for loop unrolling in the iteration count. This will
769 have no effect if loop_iterations could not determine the number
771 n_iterations /= loop_info->unroll_number;
772 n_iterations_max /= loop_info->unroll_number;
774 if (n_iterations && n_iterations < 3)
776 if (loop_dump_stream)
777 fprintf (loop_dump_stream,
778 "Doloop: Too few iterations (%ld) to be profitable.\n",
779 (long int) n_iterations);
783 iterations = GEN_INT (n_iterations);
784 iterations_max = GEN_INT (n_iterations_max);
786 /* Generate looping insn. If the pattern FAILs then give up trying
787 to modify the loop since there is some aspect the back-end does
789 start_label = gen_label_rtx ();
790 doloop_reg = gen_reg_rtx (mode);
791 doloop_seq = gen_doloop_end (doloop_reg, iterations, iterations_max,
792 GEN_INT (loop->level), start_label);
793 if (! doloop_seq && mode != word_mode)
795 PUT_MODE (doloop_reg, word_mode);
796 doloop_seq = gen_doloop_end (doloop_reg, iterations, iterations_max,
797 GEN_INT (loop->level), start_label);
801 if (loop_dump_stream)
802 fprintf (loop_dump_stream,
803 "Doloop: Target unwilling to use doloop pattern!\n");
807 /* A raw define_insn may yield a plain pattern. If a sequence
808 was involved, the last must be the jump instruction. */
809 if (GET_CODE (doloop_seq) == SEQUENCE)
811 doloop_pat = XVECEXP (doloop_seq, 0, XVECLEN (doloop_seq, 0) - 1);
812 if (GET_CODE (doloop_pat) == JUMP_INSN)
813 doloop_pat = PATTERN (doloop_pat);
815 doloop_pat = NULL_RTX;
818 doloop_pat = doloop_seq;
821 || ! (condition = doloop_condition_get (doloop_pat)))
823 if (loop_dump_stream)
824 fprintf (loop_dump_stream,
825 "Doloop: Unrecognizable doloop pattern!\n");
829 if (n_iterations != 0)
830 /* Handle the simpler case, where we know the iteration count at
832 return doloop_modify (loop, iterations, iterations_max, doloop_seq,
833 start_label, condition);
835 /* Handle the harder case, where we must add additional runtime tests. */
836 return doloop_modify_runtime (loop, iterations_max, doloop_seq,
837 start_label, mode, condition);
840 #endif /* HAVE_doloop_end */