Merge from mainline (gomp-merge-2005-02-26).
[official-gcc.git] / gcc / loop-iv.c
blob04f53b88d3ef5f21d7be37993dbc0ef3db97feaa
1 /* Rtl-level induction variable analysis.
2 Copyright (C) 2004 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 2, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to the Free
18 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
19 02111-1307, USA. */
21 /* This is just a very simplistic analysis of induction variables of the loop.
22 The major use is for determining the number of iterations of a loop for
23 loop unrolling, doloop optimization and branch prediction. For this we
24 are only interested in bivs and a fairly limited set of givs that are
25 needed in the exit condition. We also only compute the iv information on
26 demand.
28 The interesting registers are determined. A register is interesting if
30 -- it is set only in the blocks that dominate the latch of the current loop
31 -- all its sets are simple -- i.e. in the form we understand
33 We also number the insns sequentially in each basic block. For a use of the
34 interesting reg, it is now easy to find a reaching definition (there may be
35 only one).
37 Induction variable is then simply analyzed by walking the use-def
38 chains.
40 Usage:
42 iv_analysis_loop_init (loop);
43 insn = iv_get_reaching_def (where, reg);
44 if (iv_analyze (insn, reg, &iv))
46 ...
48 iv_analysis_done (); */
50 #include "config.h"
51 #include "system.h"
52 #include "coretypes.h"
53 #include "tm.h"
54 #include "rtl.h"
55 #include "hard-reg-set.h"
56 #include "obstack.h"
57 #include "basic-block.h"
58 #include "cfgloop.h"
59 #include "expr.h"
60 #include "output.h"
62 /* The insn information. */
64 struct insn_info
66 /* Id of the insn. */
67 unsigned luid;
69 /* The previous definition of the register defined by the single
70 set in the insn. */
71 rtx prev_def;
73 /* The description of the iv. */
74 struct rtx_iv iv;
77 static struct insn_info *insn_info;
79 /* The last definition of register. */
81 static rtx *last_def;
83 /* The bivs. */
85 static struct rtx_iv *bivs;
87 /* Maximal insn number for that there is place in insn_info array. */
89 static unsigned max_insn_no;
91 /* Maximal register number for that there is place in bivs and last_def
92 arrays. */
94 static unsigned max_reg_no;
96 /* Dumps information about IV to FILE. */
98 extern void dump_iv_info (FILE *, struct rtx_iv *);
99 void
100 dump_iv_info (FILE *file, struct rtx_iv *iv)
102 if (!iv->base)
104 fprintf (file, "not simple");
105 return;
108 if (iv->step == const0_rtx
109 && !iv->first_special)
110 fprintf (file, "invariant ");
112 print_rtl (file, iv->base);
113 if (iv->step != const0_rtx)
115 fprintf (file, " + ");
116 print_rtl (file, iv->step);
117 fprintf (file, " * iteration");
119 fprintf (file, " (in %s)", GET_MODE_NAME (iv->mode));
121 if (iv->mode != iv->extend_mode)
122 fprintf (file, " %s to %s",
123 rtx_name[iv->extend],
124 GET_MODE_NAME (iv->extend_mode));
126 if (iv->mult != const1_rtx)
128 fprintf (file, " * ");
129 print_rtl (file, iv->mult);
131 if (iv->delta != const0_rtx)
133 fprintf (file, " + ");
134 print_rtl (file, iv->delta);
136 if (iv->first_special)
137 fprintf (file, " (first special)");
140 /* Assigns luids to insns in basic block BB. */
142 static void
143 assign_luids (basic_block bb)
145 unsigned i = 0, uid;
146 rtx insn;
148 FOR_BB_INSNS (bb, insn)
150 uid = INSN_UID (insn);
151 insn_info[uid].luid = i++;
152 insn_info[uid].prev_def = NULL_RTX;
153 insn_info[uid].iv.analysed = false;
157 /* Generates a subreg to get the least significant part of EXPR (in mode
158 INNER_MODE) to OUTER_MODE. */
161 lowpart_subreg (enum machine_mode outer_mode, rtx expr,
162 enum machine_mode inner_mode)
164 return simplify_gen_subreg (outer_mode, expr, inner_mode,
165 subreg_lowpart_offset (outer_mode, inner_mode));
168 /* Checks whether REG is a well-behaved register. */
170 static bool
171 simple_reg_p (rtx reg)
173 unsigned r;
175 if (GET_CODE (reg) == SUBREG)
177 if (!subreg_lowpart_p (reg))
178 return false;
179 reg = SUBREG_REG (reg);
182 if (!REG_P (reg))
183 return false;
185 r = REGNO (reg);
186 if (HARD_REGISTER_NUM_P (r))
187 return false;
189 if (GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT)
190 return false;
192 if (last_def[r] == const0_rtx)
193 return false;
195 return true;
198 /* Checks whether assignment LHS = RHS is simple enough for us to process. */
200 static bool
201 simple_set_p (rtx lhs, rtx rhs)
203 rtx op0, op1;
205 if (!REG_P (lhs)
206 || !simple_reg_p (lhs))
207 return false;
209 if (CONSTANT_P (rhs))
210 return true;
212 switch (GET_CODE (rhs))
214 case SUBREG:
215 case REG:
216 return simple_reg_p (rhs);
218 case SIGN_EXTEND:
219 case ZERO_EXTEND:
220 case NEG:
221 return simple_reg_p (XEXP (rhs, 0));
223 case PLUS:
224 case MINUS:
225 case MULT:
226 case ASHIFT:
227 op0 = XEXP (rhs, 0);
228 op1 = XEXP (rhs, 1);
230 if (!simple_reg_p (op0)
231 && !CONSTANT_P (op0))
232 return false;
234 if (!simple_reg_p (op1)
235 && !CONSTANT_P (op1))
236 return false;
238 if (GET_CODE (rhs) == MULT
239 && !CONSTANT_P (op0)
240 && !CONSTANT_P (op1))
241 return false;
243 if (GET_CODE (rhs) == ASHIFT
244 && CONSTANT_P (op0))
245 return false;
247 return true;
249 default:
250 return false;
254 /* Mark single SET in INSN. */
256 static rtx
257 mark_single_set (rtx insn, rtx set)
259 rtx def = SET_DEST (set), src;
260 unsigned regno, uid;
262 src = find_reg_equal_equiv_note (insn);
263 if (src)
264 src = XEXP (src, 0);
265 else
266 src = SET_SRC (set);
268 if (!simple_set_p (SET_DEST (set), src))
269 return NULL_RTX;
271 regno = REGNO (def);
272 uid = INSN_UID (insn);
274 bivs[regno].analysed = false;
275 insn_info[uid].prev_def = last_def[regno];
276 last_def[regno] = insn;
278 return def;
281 /* Invalidate register REG unless it is equal to EXCEPT. */
283 static void
284 kill_sets (rtx reg, rtx by ATTRIBUTE_UNUSED, void *except)
286 if (GET_CODE (reg) == SUBREG)
287 reg = SUBREG_REG (reg);
288 if (!REG_P (reg))
289 return;
290 if (reg == except)
291 return;
293 last_def[REGNO (reg)] = const0_rtx;
296 /* Marks sets in basic block BB. If DOM is true, BB dominates the loop
297 latch. */
299 static void
300 mark_sets (basic_block bb, bool dom)
302 rtx insn, set, def;
304 FOR_BB_INSNS (bb, insn)
306 if (!INSN_P (insn))
307 continue;
309 if (dom
310 && (set = single_set (insn)))
311 def = mark_single_set (insn, set);
312 else
313 def = NULL_RTX;
315 note_stores (PATTERN (insn), kill_sets, def);
319 /* Prepare the data for an induction variable analysis of a LOOP. */
321 void
322 iv_analysis_loop_init (struct loop *loop)
324 basic_block *body = get_loop_body_in_dom_order (loop);
325 unsigned b;
327 if ((unsigned) get_max_uid () >= max_insn_no)
329 /* Add some reserve for insns and registers produced in optimizations. */
330 max_insn_no = get_max_uid () + 100;
331 if (insn_info)
332 free (insn_info);
333 insn_info = xmalloc (max_insn_no * sizeof (struct insn_info));
336 if ((unsigned) max_reg_num () >= max_reg_no)
338 max_reg_no = max_reg_num () + 100;
339 if (last_def)
340 free (last_def);
341 last_def = xmalloc (max_reg_no * sizeof (rtx));
342 if (bivs)
343 free (bivs);
344 bivs = xmalloc (max_reg_no * sizeof (struct rtx_iv));
347 memset (last_def, 0, max_reg_num () * sizeof (rtx));
349 for (b = 0; b < loop->num_nodes; b++)
351 assign_luids (body[b]);
352 mark_sets (body[b], just_once_each_iteration_p (loop, body[b]));
355 free (body);
358 /* Gets definition of REG reaching the INSN. If REG is not simple, const0_rtx
359 is returned. If INSN is before the first def in the loop, NULL_RTX is
360 returned. */
363 iv_get_reaching_def (rtx insn, rtx reg)
365 unsigned regno, luid, auid;
366 rtx ainsn;
367 basic_block bb, abb;
369 if (GET_CODE (reg) == SUBREG)
371 if (!subreg_lowpart_p (reg))
372 return const0_rtx;
373 reg = SUBREG_REG (reg);
375 if (!REG_P (reg))
376 return NULL_RTX;
378 regno = REGNO (reg);
379 if (!last_def[regno]
380 || last_def[regno] == const0_rtx)
381 return last_def[regno];
383 bb = BLOCK_FOR_INSN (insn);
384 luid = insn_info[INSN_UID (insn)].luid;
386 ainsn = last_def[regno];
387 while (1)
389 abb = BLOCK_FOR_INSN (ainsn);
391 if (dominated_by_p (CDI_DOMINATORS, bb, abb))
392 break;
394 auid = INSN_UID (ainsn);
395 ainsn = insn_info[auid].prev_def;
397 if (!ainsn)
398 return NULL_RTX;
401 while (1)
403 abb = BLOCK_FOR_INSN (ainsn);
404 if (abb != bb)
405 return ainsn;
407 auid = INSN_UID (ainsn);
408 if (luid > insn_info[auid].luid)
409 return ainsn;
411 ainsn = insn_info[auid].prev_def;
412 if (!ainsn)
413 return NULL_RTX;
417 /* Sets IV to invariant CST in MODE. Always returns true (just for
418 consistency with other iv manipulation functions that may fail). */
420 static bool
421 iv_constant (struct rtx_iv *iv, rtx cst, enum machine_mode mode)
423 if (mode == VOIDmode)
424 mode = GET_MODE (cst);
426 iv->analysed = true;
427 iv->mode = mode;
428 iv->base = cst;
429 iv->step = const0_rtx;
430 iv->first_special = false;
431 iv->extend = UNKNOWN;
432 iv->extend_mode = iv->mode;
433 iv->delta = const0_rtx;
434 iv->mult = const1_rtx;
436 return true;
439 /* Evaluates application of subreg to MODE on IV. */
441 static bool
442 iv_subreg (struct rtx_iv *iv, enum machine_mode mode)
444 /* If iv is invariant, just calculate the new value. */
445 if (iv->step == const0_rtx
446 && !iv->first_special)
448 rtx val = get_iv_value (iv, const0_rtx);
449 val = lowpart_subreg (mode, val, iv->extend_mode);
451 iv->base = val;
452 iv->extend = UNKNOWN;
453 iv->mode = iv->extend_mode = mode;
454 iv->delta = const0_rtx;
455 iv->mult = const1_rtx;
456 return true;
459 if (iv->extend_mode == mode)
460 return true;
462 if (GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (iv->mode))
463 return false;
465 iv->extend = UNKNOWN;
466 iv->mode = mode;
468 iv->base = simplify_gen_binary (PLUS, iv->extend_mode, iv->delta,
469 simplify_gen_binary (MULT, iv->extend_mode,
470 iv->base, iv->mult));
471 iv->step = simplify_gen_binary (MULT, iv->extend_mode, iv->step, iv->mult);
472 iv->mult = const1_rtx;
473 iv->delta = const0_rtx;
474 iv->first_special = false;
476 return true;
479 /* Evaluates application of EXTEND to MODE on IV. */
481 static bool
482 iv_extend (struct rtx_iv *iv, enum rtx_code extend, enum machine_mode mode)
484 /* If iv is invariant, just calculate the new value. */
485 if (iv->step == const0_rtx
486 && !iv->first_special)
488 rtx val = get_iv_value (iv, const0_rtx);
489 val = simplify_gen_unary (extend, mode, val, iv->extend_mode);
491 iv->base = val;
492 iv->extend = UNKNOWN;
493 iv->mode = iv->extend_mode = mode;
494 iv->delta = const0_rtx;
495 iv->mult = const1_rtx;
496 return true;
499 if (mode != iv->extend_mode)
500 return false;
502 if (iv->extend != UNKNOWN
503 && iv->extend != extend)
504 return false;
506 iv->extend = extend;
508 return true;
511 /* Evaluates negation of IV. */
513 static bool
514 iv_neg (struct rtx_iv *iv)
516 if (iv->extend == UNKNOWN)
518 iv->base = simplify_gen_unary (NEG, iv->extend_mode,
519 iv->base, iv->extend_mode);
520 iv->step = simplify_gen_unary (NEG, iv->extend_mode,
521 iv->step, iv->extend_mode);
523 else
525 iv->delta = simplify_gen_unary (NEG, iv->extend_mode,
526 iv->delta, iv->extend_mode);
527 iv->mult = simplify_gen_unary (NEG, iv->extend_mode,
528 iv->mult, iv->extend_mode);
531 return true;
534 /* Evaluates addition or subtraction (according to OP) of IV1 to IV0. */
536 static bool
537 iv_add (struct rtx_iv *iv0, struct rtx_iv *iv1, enum rtx_code op)
539 enum machine_mode mode;
540 rtx arg;
542 /* Extend the constant to extend_mode of the other operand if necessary. */
543 if (iv0->extend == UNKNOWN
544 && iv0->mode == iv0->extend_mode
545 && iv0->step == const0_rtx
546 && GET_MODE_SIZE (iv0->extend_mode) < GET_MODE_SIZE (iv1->extend_mode))
548 iv0->extend_mode = iv1->extend_mode;
549 iv0->base = simplify_gen_unary (ZERO_EXTEND, iv0->extend_mode,
550 iv0->base, iv0->mode);
552 if (iv1->extend == UNKNOWN
553 && iv1->mode == iv1->extend_mode
554 && iv1->step == const0_rtx
555 && GET_MODE_SIZE (iv1->extend_mode) < GET_MODE_SIZE (iv0->extend_mode))
557 iv1->extend_mode = iv0->extend_mode;
558 iv1->base = simplify_gen_unary (ZERO_EXTEND, iv1->extend_mode,
559 iv1->base, iv1->mode);
562 mode = iv0->extend_mode;
563 if (mode != iv1->extend_mode)
564 return false;
566 if (iv0->extend == UNKNOWN && iv1->extend == UNKNOWN)
568 if (iv0->mode != iv1->mode)
569 return false;
571 iv0->base = simplify_gen_binary (op, mode, iv0->base, iv1->base);
572 iv0->step = simplify_gen_binary (op, mode, iv0->step, iv1->step);
574 return true;
577 /* Handle addition of constant. */
578 if (iv1->extend == UNKNOWN
579 && iv1->mode == mode
580 && iv1->step == const0_rtx)
582 iv0->delta = simplify_gen_binary (op, mode, iv0->delta, iv1->base);
583 return true;
586 if (iv0->extend == UNKNOWN
587 && iv0->mode == mode
588 && iv0->step == const0_rtx)
590 arg = iv0->base;
591 *iv0 = *iv1;
592 if (op == MINUS
593 && !iv_neg (iv0))
594 return false;
596 iv0->delta = simplify_gen_binary (PLUS, mode, iv0->delta, arg);
597 return true;
600 return false;
603 /* Evaluates multiplication of IV by constant CST. */
605 static bool
606 iv_mult (struct rtx_iv *iv, rtx mby)
608 enum machine_mode mode = iv->extend_mode;
610 if (GET_MODE (mby) != VOIDmode
611 && GET_MODE (mby) != mode)
612 return false;
614 if (iv->extend == UNKNOWN)
616 iv->base = simplify_gen_binary (MULT, mode, iv->base, mby);
617 iv->step = simplify_gen_binary (MULT, mode, iv->step, mby);
619 else
621 iv->delta = simplify_gen_binary (MULT, mode, iv->delta, mby);
622 iv->mult = simplify_gen_binary (MULT, mode, iv->mult, mby);
625 return true;
628 /* Evaluates shift of IV by constant CST. */
630 static bool
631 iv_shift (struct rtx_iv *iv, rtx mby)
633 enum machine_mode mode = iv->extend_mode;
635 if (GET_MODE (mby) != VOIDmode
636 && GET_MODE (mby) != mode)
637 return false;
639 if (iv->extend == UNKNOWN)
641 iv->base = simplify_gen_binary (ASHIFT, mode, iv->base, mby);
642 iv->step = simplify_gen_binary (ASHIFT, mode, iv->step, mby);
644 else
646 iv->delta = simplify_gen_binary (ASHIFT, mode, iv->delta, mby);
647 iv->mult = simplify_gen_binary (ASHIFT, mode, iv->mult, mby);
650 return true;
653 /* The recursive part of get_biv_step. Gets the value of the single value
654 defined in INSN wrto initial value of REG inside loop, in shape described
655 at get_biv_step. */
657 static bool
658 get_biv_step_1 (rtx insn, rtx reg,
659 rtx *inner_step, enum machine_mode *inner_mode,
660 enum rtx_code *extend, enum machine_mode outer_mode,
661 rtx *outer_step)
663 rtx set, lhs, rhs, op0 = NULL_RTX, op1 = NULL_RTX;
664 rtx next, nextr, def_insn, tmp;
665 enum rtx_code code;
667 set = single_set (insn);
668 rhs = find_reg_equal_equiv_note (insn);
669 if (rhs)
670 rhs = XEXP (rhs, 0);
671 else
672 rhs = SET_SRC (set);
673 lhs = SET_DEST (set);
675 code = GET_CODE (rhs);
676 switch (code)
678 case SUBREG:
679 case REG:
680 next = rhs;
681 break;
683 case PLUS:
684 case MINUS:
685 op0 = XEXP (rhs, 0);
686 op1 = XEXP (rhs, 1);
688 if (code == PLUS && CONSTANT_P (op0))
690 tmp = op0; op0 = op1; op1 = tmp;
693 if (!simple_reg_p (op0)
694 || !CONSTANT_P (op1))
695 return false;
697 if (GET_MODE (rhs) != outer_mode)
699 /* ppc64 uses expressions like
701 (set x:SI (plus:SI (subreg:SI y:DI) 1)).
703 this is equivalent to
705 (set x':DI (plus:DI y:DI 1))
706 (set x:SI (subreg:SI (x':DI)). */
707 if (GET_CODE (op0) != SUBREG)
708 return false;
709 if (GET_MODE (SUBREG_REG (op0)) != outer_mode)
710 return false;
713 next = op0;
714 break;
716 case SIGN_EXTEND:
717 case ZERO_EXTEND:
718 if (GET_MODE (rhs) != outer_mode)
719 return false;
721 op0 = XEXP (rhs, 0);
722 if (!simple_reg_p (op0))
723 return false;
725 next = op0;
726 break;
728 default:
729 return false;
732 if (GET_CODE (next) == SUBREG)
734 if (!subreg_lowpart_p (next))
735 return false;
737 nextr = SUBREG_REG (next);
738 if (GET_MODE (nextr) != outer_mode)
739 return false;
741 else
742 nextr = next;
744 def_insn = iv_get_reaching_def (insn, nextr);
745 if (def_insn == const0_rtx)
746 return false;
748 if (!def_insn)
750 if (!rtx_equal_p (nextr, reg))
751 return false;
753 *inner_step = const0_rtx;
754 *extend = UNKNOWN;
755 *inner_mode = outer_mode;
756 *outer_step = const0_rtx;
758 else if (!get_biv_step_1 (def_insn, reg,
759 inner_step, inner_mode, extend, outer_mode,
760 outer_step))
761 return false;
763 if (GET_CODE (next) == SUBREG)
765 enum machine_mode amode = GET_MODE (next);
767 if (GET_MODE_SIZE (amode) > GET_MODE_SIZE (*inner_mode))
768 return false;
770 *inner_mode = amode;
771 *inner_step = simplify_gen_binary (PLUS, outer_mode,
772 *inner_step, *outer_step);
773 *outer_step = const0_rtx;
774 *extend = UNKNOWN;
777 switch (code)
779 case REG:
780 case SUBREG:
781 break;
783 case PLUS:
784 case MINUS:
785 if (*inner_mode == outer_mode
786 /* See comment in previous switch. */
787 || GET_MODE (rhs) != outer_mode)
788 *inner_step = simplify_gen_binary (code, outer_mode,
789 *inner_step, op1);
790 else
791 *outer_step = simplify_gen_binary (code, outer_mode,
792 *outer_step, op1);
793 break;
795 case SIGN_EXTEND:
796 case ZERO_EXTEND:
797 if (GET_MODE (op0) != *inner_mode
798 || *extend != UNKNOWN
799 || *outer_step != const0_rtx)
800 abort ();
802 *extend = code;
803 break;
805 default:
806 abort ();
809 return true;
812 /* Gets the operation on register REG inside loop, in shape
814 OUTER_STEP + EXTEND_{OUTER_MODE} (SUBREG_{INNER_MODE} (REG + INNER_STEP))
816 If the operation cannot be described in this shape, return false. */
818 static bool
819 get_biv_step (rtx reg, rtx *inner_step, enum machine_mode *inner_mode,
820 enum rtx_code *extend, enum machine_mode *outer_mode,
821 rtx *outer_step)
823 *outer_mode = GET_MODE (reg);
825 if (!get_biv_step_1 (last_def[REGNO (reg)], reg,
826 inner_step, inner_mode, extend, *outer_mode,
827 outer_step))
828 return false;
830 if (*inner_mode != *outer_mode
831 && *extend == UNKNOWN)
832 abort ();
834 if (*inner_mode == *outer_mode
835 && *extend != UNKNOWN)
836 abort ();
838 if (*inner_mode == *outer_mode
839 && *outer_step != const0_rtx)
840 abort ();
842 return true;
845 /* Determines whether DEF is a biv and if so, stores its description
846 to *IV. */
848 static bool
849 iv_analyze_biv (rtx def, struct rtx_iv *iv)
851 unsigned regno;
852 rtx inner_step, outer_step;
853 enum machine_mode inner_mode, outer_mode;
854 enum rtx_code extend;
856 if (dump_file)
858 fprintf (dump_file, "Analysing ");
859 print_rtl (dump_file, def);
860 fprintf (dump_file, " for bivness.\n");
863 if (!REG_P (def))
865 if (!CONSTANT_P (def))
866 return false;
868 return iv_constant (iv, def, VOIDmode);
871 regno = REGNO (def);
872 if (last_def[regno] == const0_rtx)
874 if (dump_file)
875 fprintf (dump_file, " not simple.\n");
876 return false;
879 if (last_def[regno] && bivs[regno].analysed)
881 if (dump_file)
882 fprintf (dump_file, " already analysed.\n");
884 *iv = bivs[regno];
885 return iv->base != NULL_RTX;
888 if (!last_def[regno])
890 iv_constant (iv, def, VOIDmode);
891 goto end;
894 iv->analysed = true;
895 if (!get_biv_step (def, &inner_step, &inner_mode, &extend,
896 &outer_mode, &outer_step))
898 iv->base = NULL_RTX;
899 goto end;
902 /* Loop transforms base to es (base + inner_step) + outer_step,
903 where es means extend of subreg between inner_mode and outer_mode.
904 The corresponding induction variable is
906 es ((base - outer_step) + i * (inner_step + outer_step)) + outer_step */
908 iv->base = simplify_gen_binary (MINUS, outer_mode, def, outer_step);
909 iv->step = simplify_gen_binary (PLUS, outer_mode, inner_step, outer_step);
910 iv->mode = inner_mode;
911 iv->extend_mode = outer_mode;
912 iv->extend = extend;
913 iv->mult = const1_rtx;
914 iv->delta = outer_step;
915 iv->first_special = inner_mode != outer_mode;
917 end:
918 if (dump_file)
920 fprintf (dump_file, " ");
921 dump_iv_info (dump_file, iv);
922 fprintf (dump_file, "\n");
925 bivs[regno] = *iv;
927 return iv->base != NULL_RTX;
930 /* Analyzes operand OP of INSN and stores the result to *IV. */
932 static bool
933 iv_analyze_op (rtx insn, rtx op, struct rtx_iv *iv)
935 rtx def_insn;
936 unsigned regno;
937 bool inv = CONSTANT_P (op);
939 if (dump_file)
941 fprintf (dump_file, "Analysing operand ");
942 print_rtl (dump_file, op);
943 fprintf (dump_file, " of insn ");
944 print_rtl_single (dump_file, insn);
947 if (GET_CODE (op) == SUBREG)
949 if (!subreg_lowpart_p (op))
950 return false;
952 if (!iv_analyze_op (insn, SUBREG_REG (op), iv))
953 return false;
955 return iv_subreg (iv, GET_MODE (op));
958 if (!inv)
960 regno = REGNO (op);
961 if (!last_def[regno])
962 inv = true;
963 else if (last_def[regno] == const0_rtx)
965 if (dump_file)
966 fprintf (dump_file, " not simple.\n");
967 return false;
971 if (inv)
973 iv_constant (iv, op, VOIDmode);
975 if (dump_file)
977 fprintf (dump_file, " ");
978 dump_iv_info (dump_file, iv);
979 fprintf (dump_file, "\n");
981 return true;
984 def_insn = iv_get_reaching_def (insn, op);
985 if (def_insn == const0_rtx)
987 if (dump_file)
988 fprintf (dump_file, " not simple.\n");
989 return false;
992 return iv_analyze (def_insn, op, iv);
995 /* Analyzes iv DEF defined in INSN and stores the result to *IV. */
997 bool
998 iv_analyze (rtx insn, rtx def, struct rtx_iv *iv)
1000 unsigned uid;
1001 rtx set, rhs, mby = NULL_RTX, tmp;
1002 rtx op0 = NULL_RTX, op1 = NULL_RTX;
1003 struct rtx_iv iv0, iv1;
1004 enum machine_mode amode;
1005 enum rtx_code code;
1007 if (insn == const0_rtx)
1008 return false;
1010 if (GET_CODE (def) == SUBREG)
1012 if (!subreg_lowpart_p (def))
1013 return false;
1015 if (!iv_analyze (insn, SUBREG_REG (def), iv))
1016 return false;
1018 return iv_subreg (iv, GET_MODE (def));
1021 if (!insn)
1022 return iv_analyze_biv (def, iv);
1024 if (dump_file)
1026 fprintf (dump_file, "Analysing def of ");
1027 print_rtl (dump_file, def);
1028 fprintf (dump_file, " in insn ");
1029 print_rtl_single (dump_file, insn);
1032 uid = INSN_UID (insn);
1033 if (insn_info[uid].iv.analysed)
1035 if (dump_file)
1036 fprintf (dump_file, " already analysed.\n");
1037 *iv = insn_info[uid].iv;
1038 return iv->base != NULL_RTX;
1041 iv->mode = VOIDmode;
1042 iv->base = NULL_RTX;
1043 iv->step = NULL_RTX;
1045 set = single_set (insn);
1046 rhs = find_reg_equal_equiv_note (insn);
1047 if (rhs)
1048 rhs = XEXP (rhs, 0);
1049 else
1050 rhs = SET_SRC (set);
1051 code = GET_CODE (rhs);
1053 if (CONSTANT_P (rhs))
1055 op0 = rhs;
1056 amode = GET_MODE (def);
1058 else
1060 switch (code)
1062 case SUBREG:
1063 if (!subreg_lowpart_p (rhs))
1064 goto end;
1065 op0 = rhs;
1066 break;
1068 case REG:
1069 op0 = rhs;
1070 break;
1072 case SIGN_EXTEND:
1073 case ZERO_EXTEND:
1074 case NEG:
1075 op0 = XEXP (rhs, 0);
1076 break;
1078 case PLUS:
1079 case MINUS:
1080 op0 = XEXP (rhs, 0);
1081 op1 = XEXP (rhs, 1);
1082 break;
1084 case MULT:
1085 op0 = XEXP (rhs, 0);
1086 mby = XEXP (rhs, 1);
1087 if (!CONSTANT_P (mby))
1089 if (!CONSTANT_P (op0))
1090 abort ();
1091 tmp = op0;
1092 op0 = mby;
1093 mby = tmp;
1095 break;
1097 case ASHIFT:
1098 if (CONSTANT_P (XEXP (rhs, 0)))
1099 abort ();
1100 op0 = XEXP (rhs, 0);
1101 mby = XEXP (rhs, 1);
1102 break;
1104 default:
1105 abort ();
1108 amode = GET_MODE (rhs);
1111 if (op0)
1113 if (!iv_analyze_op (insn, op0, &iv0))
1114 goto end;
1116 if (iv0.mode == VOIDmode)
1118 iv0.mode = amode;
1119 iv0.extend_mode = amode;
1123 if (op1)
1125 if (!iv_analyze_op (insn, op1, &iv1))
1126 goto end;
1128 if (iv1.mode == VOIDmode)
1130 iv1.mode = amode;
1131 iv1.extend_mode = amode;
1135 switch (code)
1137 case SIGN_EXTEND:
1138 case ZERO_EXTEND:
1139 if (!iv_extend (&iv0, code, amode))
1140 goto end;
1141 break;
1143 case NEG:
1144 if (!iv_neg (&iv0))
1145 goto end;
1146 break;
1148 case PLUS:
1149 case MINUS:
1150 if (!iv_add (&iv0, &iv1, code))
1151 goto end;
1152 break;
1154 case MULT:
1155 if (!iv_mult (&iv0, mby))
1156 goto end;
1157 break;
1159 case ASHIFT:
1160 if (!iv_shift (&iv0, mby))
1161 goto end;
1162 break;
1164 default:
1165 break;
1168 *iv = iv0;
1170 end:
1171 iv->analysed = true;
1172 insn_info[uid].iv = *iv;
1174 if (dump_file)
1176 print_rtl (dump_file, def);
1177 fprintf (dump_file, " in insn ");
1178 print_rtl_single (dump_file, insn);
1179 fprintf (dump_file, " is ");
1180 dump_iv_info (dump_file, iv);
1181 fprintf (dump_file, "\n");
1184 return iv->base != NULL_RTX;
1187 /* Checks whether definition of register REG in INSN a basic induction
1188 variable. IV analysis must have been initialized (via a call to
1189 iv_analysis_loop_init) for this function to produce a result. */
1191 bool
1192 biv_p (rtx insn, rtx reg)
1194 struct rtx_iv iv;
1196 if (!REG_P (reg))
1197 return false;
1199 if (last_def[REGNO (reg)] != insn)
1200 return false;
1202 return iv_analyze_biv (reg, &iv);
1205 /* Calculates value of IV at ITERATION-th iteration. */
1208 get_iv_value (struct rtx_iv *iv, rtx iteration)
1210 rtx val;
1212 /* We would need to generate some if_then_else patterns, and so far
1213 it is not needed anywhere. */
1214 if (iv->first_special)
1215 abort ();
1217 if (iv->step != const0_rtx && iteration != const0_rtx)
1218 val = simplify_gen_binary (PLUS, iv->extend_mode, iv->base,
1219 simplify_gen_binary (MULT, iv->extend_mode,
1220 iv->step, iteration));
1221 else
1222 val = iv->base;
1224 if (iv->extend_mode == iv->mode)
1225 return val;
1227 val = lowpart_subreg (iv->mode, val, iv->extend_mode);
1229 if (iv->extend == UNKNOWN)
1230 return val;
1232 val = simplify_gen_unary (iv->extend, iv->extend_mode, val, iv->mode);
1233 val = simplify_gen_binary (PLUS, iv->extend_mode, iv->delta,
1234 simplify_gen_binary (MULT, iv->extend_mode,
1235 iv->mult, val));
1237 return val;
1240 /* Free the data for an induction variable analysis. */
1242 void
1243 iv_analysis_done (void)
1245 max_insn_no = 0;
1246 max_reg_no = 0;
1247 if (insn_info)
1249 free (insn_info);
1250 insn_info = NULL;
1252 if (last_def)
1254 free (last_def);
1255 last_def = NULL;
1257 if (bivs)
1259 free (bivs);
1260 bivs = NULL;
1264 /* Computes inverse to X modulo (1 << MOD). */
1266 static unsigned HOST_WIDEST_INT
1267 inverse (unsigned HOST_WIDEST_INT x, int mod)
1269 unsigned HOST_WIDEST_INT mask =
1270 ((unsigned HOST_WIDEST_INT) 1 << (mod - 1) << 1) - 1;
1271 unsigned HOST_WIDEST_INT rslt = 1;
1272 int i;
1274 for (i = 0; i < mod - 1; i++)
1276 rslt = (rslt * x) & mask;
1277 x = (x * x) & mask;
1280 return rslt;
1283 /* Tries to estimate the maximum number of iterations. */
1285 static unsigned HOST_WIDEST_INT
1286 determine_max_iter (struct niter_desc *desc)
1288 rtx niter = desc->niter_expr;
1289 rtx mmin, mmax, left, right;
1290 unsigned HOST_WIDEST_INT nmax, inc;
1292 if (GET_CODE (niter) == AND
1293 && GET_CODE (XEXP (niter, 0)) == CONST_INT)
1295 nmax = INTVAL (XEXP (niter, 0));
1296 if (!(nmax & (nmax + 1)))
1298 desc->niter_max = nmax;
1299 return nmax;
1303 get_mode_bounds (desc->mode, desc->signed_p, desc->mode, &mmin, &mmax);
1304 nmax = INTVAL (mmax) - INTVAL (mmin);
1306 if (GET_CODE (niter) == UDIV)
1308 if (GET_CODE (XEXP (niter, 1)) != CONST_INT)
1310 desc->niter_max = nmax;
1311 return nmax;
1313 inc = INTVAL (XEXP (niter, 1));
1314 niter = XEXP (niter, 0);
1316 else
1317 inc = 1;
1319 if (GET_CODE (niter) == PLUS)
1321 left = XEXP (niter, 0);
1322 right = XEXP (niter, 0);
1324 if (GET_CODE (right) == CONST_INT)
1325 right = GEN_INT (-INTVAL (right));
1327 else if (GET_CODE (niter) == MINUS)
1329 left = XEXP (niter, 0);
1330 right = XEXP (niter, 0);
1332 else
1334 left = niter;
1335 right = mmin;
1338 if (GET_CODE (left) == CONST_INT)
1339 mmax = left;
1340 if (GET_CODE (right) == CONST_INT)
1341 mmin = right;
1342 nmax = INTVAL (mmax) - INTVAL (mmin);
1344 desc->niter_max = nmax / inc;
1345 return nmax / inc;
1348 /* Checks whether register *REG is in set ALT. Callback for for_each_rtx. */
1350 static int
1351 altered_reg_used (rtx *reg, void *alt)
1353 if (!REG_P (*reg))
1354 return 0;
1356 return REGNO_REG_SET_P (alt, REGNO (*reg));
1359 /* Marks registers altered by EXPR in set ALT. */
1361 static void
1362 mark_altered (rtx expr, rtx by ATTRIBUTE_UNUSED, void *alt)
1364 if (GET_CODE (expr) == SUBREG)
1365 expr = SUBREG_REG (expr);
1366 if (!REG_P (expr))
1367 return;
1369 SET_REGNO_REG_SET (alt, REGNO (expr));
1372 /* Checks whether RHS is simple enough to process. */
1374 static bool
1375 simple_rhs_p (rtx rhs)
1377 rtx op0, op1;
1379 if (CONSTANT_P (rhs)
1380 || REG_P (rhs))
1381 return true;
1383 switch (GET_CODE (rhs))
1385 case PLUS:
1386 case MINUS:
1387 op0 = XEXP (rhs, 0);
1388 op1 = XEXP (rhs, 1);
1389 /* Allow reg + const sets only. */
1390 if (REG_P (op0) && CONSTANT_P (op1))
1391 return true;
1392 if (REG_P (op1) && CONSTANT_P (op0))
1393 return true;
1395 return false;
1397 default:
1398 return false;
1402 /* Simplifies *EXPR using assignment in INSN. ALTERED is the set of registers
1403 altered so far. */
1405 static void
1406 simplify_using_assignment (rtx insn, rtx *expr, regset altered)
1408 rtx set = single_set (insn);
1409 rtx lhs = NULL_RTX, rhs;
1410 bool ret = false;
1412 if (set)
1414 lhs = SET_DEST (set);
1415 if (!REG_P (lhs)
1416 || altered_reg_used (&lhs, altered))
1417 ret = true;
1419 else
1420 ret = true;
1422 note_stores (PATTERN (insn), mark_altered, altered);
1423 if (CALL_P (insn))
1425 int i;
1427 /* Kill all call clobbered registers. */
1428 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1429 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
1430 SET_REGNO_REG_SET (altered, i);
1433 if (ret)
1434 return;
1436 rhs = find_reg_equal_equiv_note (insn);
1437 if (rhs)
1438 rhs = XEXP (rhs, 0);
1439 else
1440 rhs = SET_SRC (set);
1442 if (!simple_rhs_p (rhs))
1443 return;
1445 if (for_each_rtx (&rhs, altered_reg_used, altered))
1446 return;
1448 *expr = simplify_replace_rtx (*expr, lhs, rhs);
1451 /* Checks whether A implies B. */
1453 static bool
1454 implies_p (rtx a, rtx b)
1456 rtx op0, op1, opb0, opb1, r;
1457 enum machine_mode mode;
1459 if (GET_CODE (a) == EQ)
1461 op0 = XEXP (a, 0);
1462 op1 = XEXP (a, 1);
1464 if (REG_P (op0))
1466 r = simplify_replace_rtx (b, op0, op1);
1467 if (r == const_true_rtx)
1468 return true;
1471 if (REG_P (op1))
1473 r = simplify_replace_rtx (b, op1, op0);
1474 if (r == const_true_rtx)
1475 return true;
1479 /* A < B implies A + 1 <= B. */
1480 if ((GET_CODE (a) == GT || GET_CODE (a) == LT)
1481 && (GET_CODE (b) == GE || GET_CODE (b) == LE))
1483 op0 = XEXP (a, 0);
1484 op1 = XEXP (a, 1);
1485 opb0 = XEXP (b, 0);
1486 opb1 = XEXP (b, 1);
1488 if (GET_CODE (a) == GT)
1490 r = op0;
1491 op0 = op1;
1492 op1 = r;
1495 if (GET_CODE (b) == GE)
1497 r = opb0;
1498 opb0 = opb1;
1499 opb1 = r;
1502 mode = GET_MODE (op0);
1503 if (mode != GET_MODE (opb0))
1504 mode = VOIDmode;
1505 else if (mode == VOIDmode)
1507 mode = GET_MODE (op1);
1508 if (mode != GET_MODE (opb1))
1509 mode = VOIDmode;
1512 if (mode != VOIDmode
1513 && rtx_equal_p (op1, opb1)
1514 && simplify_gen_binary (MINUS, mode, opb0, op0) == const1_rtx)
1515 return true;
1518 return false;
1521 /* Canonicalizes COND so that
1523 (1) Ensure that operands are ordered according to
1524 swap_commutative_operands_p.
1525 (2) (LE x const) will be replaced with (LT x <const+1>) and similarly
1526 for GE, GEU, and LEU. */
1529 canon_condition (rtx cond)
1531 rtx tem;
1532 rtx op0, op1;
1533 enum rtx_code code;
1534 enum machine_mode mode;
1536 code = GET_CODE (cond);
1537 op0 = XEXP (cond, 0);
1538 op1 = XEXP (cond, 1);
1540 if (swap_commutative_operands_p (op0, op1))
1542 code = swap_condition (code);
1543 tem = op0;
1544 op0 = op1;
1545 op1 = tem;
1548 mode = GET_MODE (op0);
1549 if (mode == VOIDmode)
1550 mode = GET_MODE (op1);
1551 if (mode == VOIDmode)
1552 abort ();
1554 if (GET_CODE (op1) == CONST_INT
1555 && GET_MODE_CLASS (mode) != MODE_CC
1556 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
1558 HOST_WIDE_INT const_val = INTVAL (op1);
1559 unsigned HOST_WIDE_INT uconst_val = const_val;
1560 unsigned HOST_WIDE_INT max_val
1561 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (mode);
1563 switch (code)
1565 case LE:
1566 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
1567 code = LT, op1 = gen_int_mode (const_val + 1, GET_MODE (op0));
1568 break;
1570 /* When cross-compiling, const_val might be sign-extended from
1571 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
1572 case GE:
1573 if ((HOST_WIDE_INT) (const_val & max_val)
1574 != (((HOST_WIDE_INT) 1
1575 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
1576 code = GT, op1 = gen_int_mode (const_val - 1, mode);
1577 break;
1579 case LEU:
1580 if (uconst_val < max_val)
1581 code = LTU, op1 = gen_int_mode (uconst_val + 1, mode);
1582 break;
1584 case GEU:
1585 if (uconst_val != 0)
1586 code = GTU, op1 = gen_int_mode (uconst_val - 1, mode);
1587 break;
1589 default:
1590 break;
1594 if (op0 != XEXP (cond, 0)
1595 || op1 != XEXP (cond, 1)
1596 || code != GET_CODE (cond)
1597 || GET_MODE (cond) != SImode)
1598 cond = gen_rtx_fmt_ee (code, SImode, op0, op1);
1600 return cond;
1603 /* Tries to use the fact that COND holds to simplify EXPR. ALTERED is the
1604 set of altered regs. */
1606 void
1607 simplify_using_condition (rtx cond, rtx *expr, regset altered)
1609 rtx rev, reve, exp = *expr;
1611 if (!COMPARISON_P (exp))
1612 return;
1614 /* If some register gets altered later, we do not really speak about its
1615 value at the time of comparison. */
1616 if (altered
1617 && for_each_rtx (&cond, altered_reg_used, altered))
1618 return;
1620 rev = reversed_condition (cond);
1621 reve = reversed_condition (exp);
1623 cond = canon_condition (cond);
1624 exp = canon_condition (exp);
1625 if (rev)
1626 rev = canon_condition (rev);
1627 if (reve)
1628 reve = canon_condition (reve);
1630 if (rtx_equal_p (exp, cond))
1632 *expr = const_true_rtx;
1633 return;
1637 if (rev && rtx_equal_p (exp, rev))
1639 *expr = const0_rtx;
1640 return;
1643 if (implies_p (cond, exp))
1645 *expr = const_true_rtx;
1646 return;
1649 if (reve && implies_p (cond, reve))
1651 *expr = const0_rtx;
1652 return;
1655 /* A proof by contradiction. If *EXPR implies (not cond), *EXPR must
1656 be false. */
1657 if (rev && implies_p (exp, rev))
1659 *expr = const0_rtx;
1660 return;
1663 /* Similarly, If (not *EXPR) implies (not cond), *EXPR must be true. */
1664 if (rev && reve && implies_p (reve, rev))
1666 *expr = const_true_rtx;
1667 return;
1670 /* We would like to have some other tests here. TODO. */
1672 return;
1675 /* Use relationship between A and *B to eventually eliminate *B.
1676 OP is the operation we consider. */
1678 static void
1679 eliminate_implied_condition (enum rtx_code op, rtx a, rtx *b)
1681 if (op == AND)
1683 /* If A implies *B, we may replace *B by true. */
1684 if (implies_p (a, *b))
1685 *b = const_true_rtx;
1687 else if (op == IOR)
1689 /* If *B implies A, we may replace *B by false. */
1690 if (implies_p (*b, a))
1691 *b = const0_rtx;
1693 else
1694 abort ();
1697 /* Eliminates the conditions in TAIL that are implied by HEAD. OP is the
1698 operation we consider. */
1700 static void
1701 eliminate_implied_conditions (enum rtx_code op, rtx *head, rtx tail)
1703 rtx elt;
1705 for (elt = tail; elt; elt = XEXP (elt, 1))
1706 eliminate_implied_condition (op, *head, &XEXP (elt, 0));
1707 for (elt = tail; elt; elt = XEXP (elt, 1))
1708 eliminate_implied_condition (op, XEXP (elt, 0), head);
1711 /* Simplifies *EXPR using initial values at the start of the LOOP. If *EXPR
1712 is a list, its elements are assumed to be combined using OP. */
1714 static void
1715 simplify_using_initial_values (struct loop *loop, enum rtx_code op, rtx *expr)
1717 rtx head, tail, insn;
1718 rtx neutral, aggr;
1719 regset altered;
1720 edge e;
1722 if (!*expr)
1723 return;
1725 if (CONSTANT_P (*expr))
1726 return;
1728 if (GET_CODE (*expr) == EXPR_LIST)
1730 head = XEXP (*expr, 0);
1731 tail = XEXP (*expr, 1);
1733 eliminate_implied_conditions (op, &head, tail);
1735 if (op == AND)
1737 neutral = const_true_rtx;
1738 aggr = const0_rtx;
1740 else if (op == IOR)
1742 neutral = const0_rtx;
1743 aggr = const_true_rtx;
1745 else
1746 abort ();
1748 simplify_using_initial_values (loop, UNKNOWN, &head);
1749 if (head == aggr)
1751 XEXP (*expr, 0) = aggr;
1752 XEXP (*expr, 1) = NULL_RTX;
1753 return;
1755 else if (head == neutral)
1757 *expr = tail;
1758 simplify_using_initial_values (loop, op, expr);
1759 return;
1761 simplify_using_initial_values (loop, op, &tail);
1763 if (tail && XEXP (tail, 0) == aggr)
1765 *expr = tail;
1766 return;
1769 XEXP (*expr, 0) = head;
1770 XEXP (*expr, 1) = tail;
1771 return;
1774 if (op != UNKNOWN)
1775 abort ();
1777 e = loop_preheader_edge (loop);
1778 if (e->src == ENTRY_BLOCK_PTR)
1779 return;
1781 altered = ALLOC_REG_SET (&reg_obstack);
1783 while (1)
1785 basic_block tmp_bb;
1787 insn = BB_END (e->src);
1788 if (any_condjump_p (insn))
1790 rtx cond = get_condition (BB_END (e->src), NULL, false, true);
1792 if (cond && (e->flags & EDGE_FALLTHRU))
1793 cond = reversed_condition (cond);
1794 if (cond)
1796 simplify_using_condition (cond, expr, altered);
1797 if (CONSTANT_P (*expr))
1799 FREE_REG_SET (altered);
1800 return;
1805 FOR_BB_INSNS_REVERSE (e->src, insn)
1807 if (!INSN_P (insn))
1808 continue;
1810 simplify_using_assignment (insn, expr, altered);
1811 if (CONSTANT_P (*expr))
1813 FREE_REG_SET (altered);
1814 return;
1818 /* This is a bit subtle. Store away e->src in tmp_bb, since we
1819 modify `e' and this can invalidate the subsequent count of
1820 e->src's predecessors by looking at the wrong block. */
1821 tmp_bb = e->src;
1822 e = EDGE_PRED (tmp_bb, 0);
1823 if (EDGE_COUNT (tmp_bb->preds) > 1
1824 || e->src == ENTRY_BLOCK_PTR)
1825 break;
1828 FREE_REG_SET (altered);
1831 /* Transforms invariant IV into MODE. Adds assumptions based on the fact
1832 that IV occurs as left operands of comparison COND and its signedness
1833 is SIGNED_P to DESC. */
1835 static void
1836 shorten_into_mode (struct rtx_iv *iv, enum machine_mode mode,
1837 enum rtx_code cond, bool signed_p, struct niter_desc *desc)
1839 rtx mmin, mmax, cond_over, cond_under;
1841 get_mode_bounds (mode, signed_p, iv->extend_mode, &mmin, &mmax);
1842 cond_under = simplify_gen_relational (LT, SImode, iv->extend_mode,
1843 iv->base, mmin);
1844 cond_over = simplify_gen_relational (GT, SImode, iv->extend_mode,
1845 iv->base, mmax);
1847 switch (cond)
1849 case LE:
1850 case LT:
1851 case LEU:
1852 case LTU:
1853 if (cond_under != const0_rtx)
1854 desc->infinite =
1855 alloc_EXPR_LIST (0, cond_under, desc->infinite);
1856 if (cond_over != const0_rtx)
1857 desc->noloop_assumptions =
1858 alloc_EXPR_LIST (0, cond_over, desc->noloop_assumptions);
1859 break;
1861 case GE:
1862 case GT:
1863 case GEU:
1864 case GTU:
1865 if (cond_over != const0_rtx)
1866 desc->infinite =
1867 alloc_EXPR_LIST (0, cond_over, desc->infinite);
1868 if (cond_under != const0_rtx)
1869 desc->noloop_assumptions =
1870 alloc_EXPR_LIST (0, cond_under, desc->noloop_assumptions);
1871 break;
1873 case NE:
1874 if (cond_over != const0_rtx)
1875 desc->infinite =
1876 alloc_EXPR_LIST (0, cond_over, desc->infinite);
1877 if (cond_under != const0_rtx)
1878 desc->infinite =
1879 alloc_EXPR_LIST (0, cond_under, desc->infinite);
1880 break;
1882 default:
1883 abort ();
1886 iv->mode = mode;
1887 iv->extend = signed_p ? SIGN_EXTEND : ZERO_EXTEND;
1890 /* Transforms IV0 and IV1 compared by COND so that they are both compared as
1891 subregs of the same mode if possible (sometimes it is necessary to add
1892 some assumptions to DESC). */
1894 static bool
1895 canonicalize_iv_subregs (struct rtx_iv *iv0, struct rtx_iv *iv1,
1896 enum rtx_code cond, struct niter_desc *desc)
1898 enum machine_mode comp_mode;
1899 bool signed_p;
1901 /* If the ivs behave specially in the first iteration, or are
1902 added/multiplied after extending, we ignore them. */
1903 if (iv0->first_special || iv0->mult != const1_rtx || iv0->delta != const0_rtx)
1904 return false;
1905 if (iv1->first_special || iv1->mult != const1_rtx || iv1->delta != const0_rtx)
1906 return false;
1908 /* If there is some extend, it must match signedness of the comparison. */
1909 switch (cond)
1911 case LE:
1912 case LT:
1913 if (iv0->extend == ZERO_EXTEND
1914 || iv1->extend == ZERO_EXTEND)
1915 return false;
1916 signed_p = true;
1917 break;
1919 case LEU:
1920 case LTU:
1921 if (iv0->extend == SIGN_EXTEND
1922 || iv1->extend == SIGN_EXTEND)
1923 return false;
1924 signed_p = false;
1925 break;
1927 case NE:
1928 if (iv0->extend != UNKNOWN
1929 && iv1->extend != UNKNOWN
1930 && iv0->extend != iv1->extend)
1931 return false;
1933 signed_p = false;
1934 if (iv0->extend != UNKNOWN)
1935 signed_p = iv0->extend == SIGN_EXTEND;
1936 if (iv1->extend != UNKNOWN)
1937 signed_p = iv1->extend == SIGN_EXTEND;
1938 break;
1940 default:
1941 abort ();
1944 /* Values of both variables should be computed in the same mode. These
1945 might indeed be different, if we have comparison like
1947 (compare (subreg:SI (iv0)) (subreg:SI (iv1)))
1949 and iv0 and iv1 are both ivs iterating in SI mode, but calculated
1950 in different modes. This does not seem impossible to handle, but
1951 it hardly ever occurs in practice.
1953 The only exception is the case when one of operands is invariant.
1954 For example pentium 3 generates comparisons like
1955 (lt (subreg:HI (reg:SI)) 100). Here we assign HImode to 100, but we
1956 definitely do not want this prevent the optimization. */
1957 comp_mode = iv0->extend_mode;
1958 if (GET_MODE_BITSIZE (comp_mode) < GET_MODE_BITSIZE (iv1->extend_mode))
1959 comp_mode = iv1->extend_mode;
1961 if (iv0->extend_mode != comp_mode)
1963 if (iv0->mode != iv0->extend_mode
1964 || iv0->step != const0_rtx)
1965 return false;
1967 iv0->base = simplify_gen_unary (signed_p ? SIGN_EXTEND : ZERO_EXTEND,
1968 comp_mode, iv0->base, iv0->mode);
1969 iv0->extend_mode = comp_mode;
1972 if (iv1->extend_mode != comp_mode)
1974 if (iv1->mode != iv1->extend_mode
1975 || iv1->step != const0_rtx)
1976 return false;
1978 iv1->base = simplify_gen_unary (signed_p ? SIGN_EXTEND : ZERO_EXTEND,
1979 comp_mode, iv1->base, iv1->mode);
1980 iv1->extend_mode = comp_mode;
1983 /* Check that both ivs belong to a range of a single mode. If one of the
1984 operands is an invariant, we may need to shorten it into the common
1985 mode. */
1986 if (iv0->mode == iv0->extend_mode
1987 && iv0->step == const0_rtx
1988 && iv0->mode != iv1->mode)
1989 shorten_into_mode (iv0, iv1->mode, cond, signed_p, desc);
1991 if (iv1->mode == iv1->extend_mode
1992 && iv1->step == const0_rtx
1993 && iv0->mode != iv1->mode)
1994 shorten_into_mode (iv1, iv0->mode, swap_condition (cond), signed_p, desc);
1996 if (iv0->mode != iv1->mode)
1997 return false;
1999 desc->mode = iv0->mode;
2000 desc->signed_p = signed_p;
2002 return true;
2005 /* Computes number of iterations of the CONDITION in INSN in LOOP and stores
2006 the result into DESC. Very similar to determine_number_of_iterations
2007 (basically its rtl version), complicated by things like subregs. */
2009 static void
2010 iv_number_of_iterations (struct loop *loop, rtx insn, rtx condition,
2011 struct niter_desc *desc)
2013 rtx op0, op1, delta, step, bound, may_xform, def_insn, tmp, tmp0, tmp1;
2014 struct rtx_iv iv0, iv1, tmp_iv;
2015 rtx assumption, may_not_xform;
2016 enum rtx_code cond;
2017 enum machine_mode mode, comp_mode;
2018 rtx mmin, mmax, mode_mmin, mode_mmax;
2019 unsigned HOST_WIDEST_INT s, size, d, inv;
2020 HOST_WIDEST_INT up, down, inc, step_val;
2021 int was_sharp = false;
2022 rtx old_niter;
2023 bool step_is_pow2;
2025 /* The meaning of these assumptions is this:
2026 if !assumptions
2027 then the rest of information does not have to be valid
2028 if noloop_assumptions then the loop does not roll
2029 if infinite then this exit is never used */
2031 desc->assumptions = NULL_RTX;
2032 desc->noloop_assumptions = NULL_RTX;
2033 desc->infinite = NULL_RTX;
2034 desc->simple_p = true;
2036 desc->const_iter = false;
2037 desc->niter_expr = NULL_RTX;
2038 desc->niter_max = 0;
2040 cond = GET_CODE (condition);
2041 if (!COMPARISON_P (condition))
2042 abort ();
2044 mode = GET_MODE (XEXP (condition, 0));
2045 if (mode == VOIDmode)
2046 mode = GET_MODE (XEXP (condition, 1));
2047 /* The constant comparisons should be folded. */
2048 if (mode == VOIDmode)
2049 abort ();
2051 /* We only handle integers or pointers. */
2052 if (GET_MODE_CLASS (mode) != MODE_INT
2053 && GET_MODE_CLASS (mode) != MODE_PARTIAL_INT)
2054 goto fail;
2056 op0 = XEXP (condition, 0);
2057 def_insn = iv_get_reaching_def (insn, op0);
2058 if (!iv_analyze (def_insn, op0, &iv0))
2059 goto fail;
2060 if (iv0.extend_mode == VOIDmode)
2061 iv0.mode = iv0.extend_mode = mode;
2063 op1 = XEXP (condition, 1);
2064 def_insn = iv_get_reaching_def (insn, op1);
2065 if (!iv_analyze (def_insn, op1, &iv1))
2066 goto fail;
2067 if (iv1.extend_mode == VOIDmode)
2068 iv1.mode = iv1.extend_mode = mode;
2070 if (GET_MODE_BITSIZE (iv0.extend_mode) > HOST_BITS_PER_WIDE_INT
2071 || GET_MODE_BITSIZE (iv1.extend_mode) > HOST_BITS_PER_WIDE_INT)
2072 goto fail;
2074 /* Check condition and normalize it. */
2076 switch (cond)
2078 case GE:
2079 case GT:
2080 case GEU:
2081 case GTU:
2082 tmp_iv = iv0; iv0 = iv1; iv1 = tmp_iv;
2083 cond = swap_condition (cond);
2084 break;
2085 case NE:
2086 case LE:
2087 case LEU:
2088 case LT:
2089 case LTU:
2090 break;
2091 default:
2092 goto fail;
2095 /* Handle extends. This is relatively nontrivial, so we only try in some
2096 easy cases, when we can canonicalize the ivs (possibly by adding some
2097 assumptions) to shape subreg (base + i * step). This function also fills
2098 in desc->mode and desc->signed_p. */
2100 if (!canonicalize_iv_subregs (&iv0, &iv1, cond, desc))
2101 goto fail;
2103 comp_mode = iv0.extend_mode;
2104 mode = iv0.mode;
2105 size = GET_MODE_BITSIZE (mode);
2106 get_mode_bounds (mode, (cond == LE || cond == LT), comp_mode, &mmin, &mmax);
2107 mode_mmin = lowpart_subreg (mode, mmin, comp_mode);
2108 mode_mmax = lowpart_subreg (mode, mmax, comp_mode);
2110 if (GET_CODE (iv0.step) != CONST_INT || GET_CODE (iv1.step) != CONST_INT)
2111 goto fail;
2113 /* We can take care of the case of two induction variables chasing each other
2114 if the test is NE. I have never seen a loop using it, but still it is
2115 cool. */
2116 if (iv0.step != const0_rtx && iv1.step != const0_rtx)
2118 if (cond != NE)
2119 goto fail;
2121 iv0.step = simplify_gen_binary (MINUS, comp_mode, iv0.step, iv1.step);
2122 iv1.step = const0_rtx;
2125 /* This is either infinite loop or the one that ends immediately, depending
2126 on initial values. Unswitching should remove this kind of conditions. */
2127 if (iv0.step == const0_rtx && iv1.step == const0_rtx)
2128 goto fail;
2130 if (cond != NE)
2132 if (iv0.step == const0_rtx)
2133 step_val = -INTVAL (iv1.step);
2134 else
2135 step_val = INTVAL (iv0.step);
2137 /* Ignore loops of while (i-- < 10) type. */
2138 if (step_val < 0)
2139 goto fail;
2141 step_is_pow2 = !(step_val & (step_val - 1));
2143 else
2145 /* We do not care about whether the step is power of two in this
2146 case. */
2147 step_is_pow2 = false;
2148 step_val = 0;
2151 /* Some more condition normalization. We must record some assumptions
2152 due to overflows. */
2153 switch (cond)
2155 case LT:
2156 case LTU:
2157 /* We want to take care only of non-sharp relationals; this is easy,
2158 as in cases the overflow would make the transformation unsafe
2159 the loop does not roll. Seemingly it would make more sense to want
2160 to take care of sharp relationals instead, as NE is more similar to
2161 them, but the problem is that here the transformation would be more
2162 difficult due to possibly infinite loops. */
2163 if (iv0.step == const0_rtx)
2165 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2166 assumption = simplify_gen_relational (EQ, SImode, mode, tmp,
2167 mode_mmax);
2168 if (assumption == const_true_rtx)
2169 goto zero_iter;
2170 iv0.base = simplify_gen_binary (PLUS, comp_mode,
2171 iv0.base, const1_rtx);
2173 else
2175 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2176 assumption = simplify_gen_relational (EQ, SImode, mode, tmp,
2177 mode_mmin);
2178 if (assumption == const_true_rtx)
2179 goto zero_iter;
2180 iv1.base = simplify_gen_binary (PLUS, comp_mode,
2181 iv1.base, constm1_rtx);
2184 if (assumption != const0_rtx)
2185 desc->noloop_assumptions =
2186 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2187 cond = (cond == LT) ? LE : LEU;
2189 /* It will be useful to be able to tell the difference once more in
2190 LE -> NE reduction. */
2191 was_sharp = true;
2192 break;
2193 default: ;
2196 /* Take care of trivially infinite loops. */
2197 if (cond != NE)
2199 if (iv0.step == const0_rtx)
2201 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2202 if (rtx_equal_p (tmp, mode_mmin))
2204 desc->infinite =
2205 alloc_EXPR_LIST (0, const_true_rtx, NULL_RTX);
2206 return;
2209 else
2211 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2212 if (rtx_equal_p (tmp, mode_mmax))
2214 desc->infinite =
2215 alloc_EXPR_LIST (0, const_true_rtx, NULL_RTX);
2216 return;
2221 /* If we can we want to take care of NE conditions instead of size
2222 comparisons, as they are much more friendly (most importantly
2223 this takes care of special handling of loops with step 1). We can
2224 do it if we first check that upper bound is greater or equal to
2225 lower bound, their difference is constant c modulo step and that
2226 there is not an overflow. */
2227 if (cond != NE)
2229 if (iv0.step == const0_rtx)
2230 step = simplify_gen_unary (NEG, comp_mode, iv1.step, comp_mode);
2231 else
2232 step = iv0.step;
2233 delta = simplify_gen_binary (MINUS, comp_mode, iv1.base, iv0.base);
2234 delta = lowpart_subreg (mode, delta, comp_mode);
2235 delta = simplify_gen_binary (UMOD, mode, delta, step);
2236 may_xform = const0_rtx;
2237 may_not_xform = const_true_rtx;
2239 if (GET_CODE (delta) == CONST_INT)
2241 if (was_sharp && INTVAL (delta) == INTVAL (step) - 1)
2243 /* A special case. We have transformed condition of type
2244 for (i = 0; i < 4; i += 4)
2245 into
2246 for (i = 0; i <= 3; i += 4)
2247 obviously if the test for overflow during that transformation
2248 passed, we cannot overflow here. Most importantly any
2249 loop with sharp end condition and step 1 falls into this
2250 category, so handling this case specially is definitely
2251 worth the troubles. */
2252 may_xform = const_true_rtx;
2254 else if (iv0.step == const0_rtx)
2256 bound = simplify_gen_binary (PLUS, comp_mode, mmin, step);
2257 bound = simplify_gen_binary (MINUS, comp_mode, bound, delta);
2258 bound = lowpart_subreg (mode, bound, comp_mode);
2259 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2260 may_xform = simplify_gen_relational (cond, SImode, mode,
2261 bound, tmp);
2262 may_not_xform = simplify_gen_relational (reverse_condition (cond),
2263 SImode, mode,
2264 bound, tmp);
2266 else
2268 bound = simplify_gen_binary (MINUS, comp_mode, mmax, step);
2269 bound = simplify_gen_binary (PLUS, comp_mode, bound, delta);
2270 bound = lowpart_subreg (mode, bound, comp_mode);
2271 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2272 may_xform = simplify_gen_relational (cond, SImode, mode,
2273 tmp, bound);
2274 may_not_xform = simplify_gen_relational (reverse_condition (cond),
2275 SImode, mode,
2276 tmp, bound);
2280 if (may_xform != const0_rtx)
2282 /* We perform the transformation always provided that it is not
2283 completely senseless. This is OK, as we would need this assumption
2284 to determine the number of iterations anyway. */
2285 if (may_xform != const_true_rtx)
2287 /* If the step is a power of two and the final value we have
2288 computed overflows, the cycle is infinite. Otherwise it
2289 is nontrivial to compute the number of iterations. */
2290 if (step_is_pow2)
2291 desc->infinite = alloc_EXPR_LIST (0, may_not_xform,
2292 desc->infinite);
2293 else
2294 desc->assumptions = alloc_EXPR_LIST (0, may_xform,
2295 desc->assumptions);
2298 /* We are going to lose some information about upper bound on
2299 number of iterations in this step, so record the information
2300 here. */
2301 inc = INTVAL (iv0.step) - INTVAL (iv1.step);
2302 if (GET_CODE (iv1.base) == CONST_INT)
2303 up = INTVAL (iv1.base);
2304 else
2305 up = INTVAL (mode_mmax) - inc;
2306 down = INTVAL (GET_CODE (iv0.base) == CONST_INT
2307 ? iv0.base
2308 : mode_mmin);
2309 desc->niter_max = (up - down) / inc + 1;
2311 if (iv0.step == const0_rtx)
2313 iv0.base = simplify_gen_binary (PLUS, comp_mode, iv0.base, delta);
2314 iv0.base = simplify_gen_binary (MINUS, comp_mode, iv0.base, step);
2316 else
2318 iv1.base = simplify_gen_binary (MINUS, comp_mode, iv1.base, delta);
2319 iv1.base = simplify_gen_binary (PLUS, comp_mode, iv1.base, step);
2322 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2323 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2324 assumption = simplify_gen_relational (reverse_condition (cond),
2325 SImode, mode, tmp0, tmp1);
2326 if (assumption == const_true_rtx)
2327 goto zero_iter;
2328 else if (assumption != const0_rtx)
2329 desc->noloop_assumptions =
2330 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2331 cond = NE;
2335 /* Count the number of iterations. */
2336 if (cond == NE)
2338 /* Everything we do here is just arithmetics modulo size of mode. This
2339 makes us able to do more involved computations of number of iterations
2340 than in other cases. First transform the condition into shape
2341 s * i <> c, with s positive. */
2342 iv1.base = simplify_gen_binary (MINUS, comp_mode, iv1.base, iv0.base);
2343 iv0.base = const0_rtx;
2344 iv0.step = simplify_gen_binary (MINUS, comp_mode, iv0.step, iv1.step);
2345 iv1.step = const0_rtx;
2346 if (INTVAL (iv0.step) < 0)
2348 iv0.step = simplify_gen_unary (NEG, comp_mode, iv0.step, mode);
2349 iv1.base = simplify_gen_unary (NEG, comp_mode, iv1.base, mode);
2351 iv0.step = lowpart_subreg (mode, iv0.step, comp_mode);
2353 /* Let nsd (s, size of mode) = d. If d does not divide c, the loop
2354 is infinite. Otherwise, the number of iterations is
2355 (inverse(s/d) * (c/d)) mod (size of mode/d). */
2356 s = INTVAL (iv0.step); d = 1;
2357 while (s % 2 != 1)
2359 s /= 2;
2360 d *= 2;
2361 size--;
2363 bound = GEN_INT (((unsigned HOST_WIDEST_INT) 1 << (size - 1 ) << 1) - 1);
2365 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2366 tmp = simplify_gen_binary (UMOD, mode, tmp1, GEN_INT (d));
2367 assumption = simplify_gen_relational (NE, SImode, mode, tmp, const0_rtx);
2368 desc->infinite = alloc_EXPR_LIST (0, assumption, desc->infinite);
2370 tmp = simplify_gen_binary (UDIV, mode, tmp1, GEN_INT (d));
2371 inv = inverse (s, size);
2372 inv = trunc_int_for_mode (inv, mode);
2373 tmp = simplify_gen_binary (MULT, mode, tmp, GEN_INT (inv));
2374 desc->niter_expr = simplify_gen_binary (AND, mode, tmp, bound);
2376 else
2378 if (iv1.step == const0_rtx)
2379 /* Condition in shape a + s * i <= b
2380 We must know that b + s does not overflow and a <= b + s and then we
2381 can compute number of iterations as (b + s - a) / s. (It might
2382 seem that we in fact could be more clever about testing the b + s
2383 overflow condition using some information about b - a mod s,
2384 but it was already taken into account during LE -> NE transform). */
2386 step = iv0.step;
2387 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2388 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2390 bound = simplify_gen_binary (MINUS, mode, mode_mmax,
2391 lowpart_subreg (mode, step,
2392 comp_mode));
2393 if (step_is_pow2)
2395 rtx t0, t1;
2397 /* If s is power of 2, we know that the loop is infinite if
2398 a % s <= b % s and b + s overflows. */
2399 assumption = simplify_gen_relational (reverse_condition (cond),
2400 SImode, mode,
2401 tmp1, bound);
2403 t0 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp0), step);
2404 t1 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp1), step);
2405 tmp = simplify_gen_relational (cond, SImode, mode, t0, t1);
2406 assumption = simplify_gen_binary (AND, SImode, assumption, tmp);
2407 desc->infinite =
2408 alloc_EXPR_LIST (0, assumption, desc->infinite);
2410 else
2412 assumption = simplify_gen_relational (cond, SImode, mode,
2413 tmp1, bound);
2414 desc->assumptions =
2415 alloc_EXPR_LIST (0, assumption, desc->assumptions);
2418 tmp = simplify_gen_binary (PLUS, comp_mode, iv1.base, iv0.step);
2419 tmp = lowpart_subreg (mode, tmp, comp_mode);
2420 assumption = simplify_gen_relational (reverse_condition (cond),
2421 SImode, mode, tmp0, tmp);
2423 delta = simplify_gen_binary (PLUS, mode, tmp1, step);
2424 delta = simplify_gen_binary (MINUS, mode, delta, tmp0);
2426 else
2428 /* Condition in shape a <= b - s * i
2429 We must know that a - s does not overflow and a - s <= b and then
2430 we can again compute number of iterations as (b - (a - s)) / s. */
2431 step = simplify_gen_unary (NEG, mode, iv1.step, mode);
2432 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2433 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2435 bound = simplify_gen_binary (MINUS, mode, mode_mmin,
2436 lowpart_subreg (mode, step, comp_mode));
2437 if (step_is_pow2)
2439 rtx t0, t1;
2441 /* If s is power of 2, we know that the loop is infinite if
2442 a % s <= b % s and a - s overflows. */
2443 assumption = simplify_gen_relational (reverse_condition (cond),
2444 SImode, mode,
2445 bound, tmp0);
2447 t0 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp0), step);
2448 t1 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp1), step);
2449 tmp = simplify_gen_relational (cond, SImode, mode, t0, t1);
2450 assumption = simplify_gen_binary (AND, SImode, assumption, tmp);
2451 desc->infinite =
2452 alloc_EXPR_LIST (0, assumption, desc->infinite);
2454 else
2456 assumption = simplify_gen_relational (cond, SImode, mode,
2457 bound, tmp0);
2458 desc->assumptions =
2459 alloc_EXPR_LIST (0, assumption, desc->assumptions);
2462 tmp = simplify_gen_binary (PLUS, comp_mode, iv0.base, iv1.step);
2463 tmp = lowpart_subreg (mode, tmp, comp_mode);
2464 assumption = simplify_gen_relational (reverse_condition (cond),
2465 SImode, mode,
2466 tmp, tmp1);
2467 delta = simplify_gen_binary (MINUS, mode, tmp0, step);
2468 delta = simplify_gen_binary (MINUS, mode, tmp1, delta);
2470 if (assumption == const_true_rtx)
2471 goto zero_iter;
2472 else if (assumption != const0_rtx)
2473 desc->noloop_assumptions =
2474 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2475 delta = simplify_gen_binary (UDIV, mode, delta, step);
2476 desc->niter_expr = delta;
2479 old_niter = desc->niter_expr;
2481 simplify_using_initial_values (loop, AND, &desc->assumptions);
2482 if (desc->assumptions
2483 && XEXP (desc->assumptions, 0) == const0_rtx)
2484 goto fail;
2485 simplify_using_initial_values (loop, IOR, &desc->noloop_assumptions);
2486 simplify_using_initial_values (loop, IOR, &desc->infinite);
2487 simplify_using_initial_values (loop, UNKNOWN, &desc->niter_expr);
2489 /* Rerun the simplification. Consider code (created by copying loop headers)
2491 i = 0;
2493 if (0 < n)
2497 i++;
2498 } while (i < n);
2501 The first pass determines that i = 0, the second pass uses it to eliminate
2502 noloop assumption. */
2504 simplify_using_initial_values (loop, AND, &desc->assumptions);
2505 if (desc->assumptions
2506 && XEXP (desc->assumptions, 0) == const0_rtx)
2507 goto fail;
2508 simplify_using_initial_values (loop, IOR, &desc->noloop_assumptions);
2509 simplify_using_initial_values (loop, IOR, &desc->infinite);
2510 simplify_using_initial_values (loop, UNKNOWN, &desc->niter_expr);
2512 if (desc->noloop_assumptions
2513 && XEXP (desc->noloop_assumptions, 0) == const_true_rtx)
2514 goto zero_iter;
2516 if (GET_CODE (desc->niter_expr) == CONST_INT)
2518 unsigned HOST_WIDEST_INT val = INTVAL (desc->niter_expr);
2520 desc->const_iter = true;
2521 desc->niter_max = desc->niter = val & GET_MODE_MASK (desc->mode);
2523 else
2525 if (!desc->niter_max)
2526 desc->niter_max = determine_max_iter (desc);
2528 /* simplify_using_initial_values does a copy propagation on the registers
2529 in the expression for the number of iterations. This prolongs life
2530 ranges of registers and increases register pressure, and usually
2531 brings no gain (and if it happens to do, the cse pass will take care
2532 of it anyway). So prevent this behavior, unless it enabled us to
2533 derive that the number of iterations is a constant. */
2534 desc->niter_expr = old_niter;
2537 return;
2539 fail:
2540 desc->simple_p = false;
2541 return;
2543 zero_iter:
2544 desc->const_iter = true;
2545 desc->niter = 0;
2546 desc->niter_max = 0;
2547 desc->niter_expr = const0_rtx;
2548 return;
2551 /* Checks whether E is a simple exit from LOOP and stores its description
2552 into DESC. */
2554 static void
2555 check_simple_exit (struct loop *loop, edge e, struct niter_desc *desc)
2557 basic_block exit_bb;
2558 rtx condition, at;
2559 edge ein;
2561 exit_bb = e->src;
2562 desc->simple_p = false;
2564 /* It must belong directly to the loop. */
2565 if (exit_bb->loop_father != loop)
2566 return;
2568 /* It must be tested (at least) once during any iteration. */
2569 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, exit_bb))
2570 return;
2572 /* It must end in a simple conditional jump. */
2573 if (!any_condjump_p (BB_END (exit_bb)))
2574 return;
2576 ein = EDGE_SUCC (exit_bb, 0);
2577 if (ein == e)
2578 ein = EDGE_SUCC (exit_bb, 1);
2580 desc->out_edge = e;
2581 desc->in_edge = ein;
2583 /* Test whether the condition is suitable. */
2584 if (!(condition = get_condition (BB_END (ein->src), &at, false, false)))
2585 return;
2587 if (ein->flags & EDGE_FALLTHRU)
2589 condition = reversed_condition (condition);
2590 if (!condition)
2591 return;
2594 /* Check that we are able to determine number of iterations and fill
2595 in information about it. */
2596 iv_number_of_iterations (loop, at, condition, desc);
2599 /* Finds a simple exit of LOOP and stores its description into DESC. */
2601 void
2602 find_simple_exit (struct loop *loop, struct niter_desc *desc)
2604 unsigned i;
2605 basic_block *body;
2606 edge e;
2607 struct niter_desc act;
2608 bool any = false;
2609 edge_iterator ei;
2611 desc->simple_p = false;
2612 body = get_loop_body (loop);
2614 for (i = 0; i < loop->num_nodes; i++)
2616 FOR_EACH_EDGE (e, ei, body[i]->succs)
2618 if (flow_bb_inside_loop_p (loop, e->dest))
2619 continue;
2621 check_simple_exit (loop, e, &act);
2622 if (!act.simple_p)
2623 continue;
2625 /* Prefer constant iterations; the less the better. */
2626 if (!any)
2627 any = true;
2628 else if (!act.const_iter
2629 || (desc->const_iter && act.niter >= desc->niter))
2630 continue;
2631 *desc = act;
2635 if (dump_file)
2637 if (desc->simple_p)
2639 fprintf (dump_file, "Loop %d is simple:\n", loop->num);
2640 fprintf (dump_file, " simple exit %d -> %d\n",
2641 desc->out_edge->src->index,
2642 desc->out_edge->dest->index);
2643 if (desc->assumptions)
2645 fprintf (dump_file, " assumptions: ");
2646 print_rtl (dump_file, desc->assumptions);
2647 fprintf (dump_file, "\n");
2649 if (desc->noloop_assumptions)
2651 fprintf (dump_file, " does not roll if: ");
2652 print_rtl (dump_file, desc->noloop_assumptions);
2653 fprintf (dump_file, "\n");
2655 if (desc->infinite)
2657 fprintf (dump_file, " infinite if: ");
2658 print_rtl (dump_file, desc->infinite);
2659 fprintf (dump_file, "\n");
2662 fprintf (dump_file, " number of iterations: ");
2663 print_rtl (dump_file, desc->niter_expr);
2664 fprintf (dump_file, "\n");
2666 fprintf (dump_file, " upper bound: ");
2667 fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, desc->niter_max);
2668 fprintf (dump_file, "\n");
2670 else
2671 fprintf (dump_file, "Loop %d is not simple.\n", loop->num);
2674 free (body);
2677 /* Creates a simple loop description of LOOP if it was not computed
2678 already. */
2680 struct niter_desc *
2681 get_simple_loop_desc (struct loop *loop)
2683 struct niter_desc *desc = simple_loop_desc (loop);
2685 if (desc)
2686 return desc;
2688 desc = xmalloc (sizeof (struct niter_desc));
2689 iv_analysis_loop_init (loop);
2690 find_simple_exit (loop, desc);
2691 loop->aux = desc;
2693 return desc;
2696 /* Releases simple loop description for LOOP. */
2698 void
2699 free_simple_loop_desc (struct loop *loop)
2701 struct niter_desc *desc = simple_loop_desc (loop);
2703 if (!desc)
2704 return;
2706 free (desc);
2707 loop->aux = NULL;