* varasm.c (assemble_start_function): Remove reset of in_section.
[official-gcc.git] / gcc / loop-iv.c
blob09ed73c6324287ac566e1fe403d066fb37f4ef0f
1 /* Rtl-level induction variable analysis.
2 Copyright (C) 2004, 2005 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, 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);
674 code = GET_CODE (rhs);
675 switch (code)
677 case SUBREG:
678 case REG:
679 next = rhs;
680 break;
682 case PLUS:
683 case MINUS:
684 op0 = XEXP (rhs, 0);
685 op1 = XEXP (rhs, 1);
687 if (code == PLUS && CONSTANT_P (op0))
689 tmp = op0; op0 = op1; op1 = tmp;
692 if (!simple_reg_p (op0)
693 || !CONSTANT_P (op1))
694 return false;
696 if (GET_MODE (rhs) != outer_mode)
698 /* ppc64 uses expressions like
700 (set x:SI (plus:SI (subreg:SI y:DI) 1)).
702 this is equivalent to
704 (set x':DI (plus:DI y:DI 1))
705 (set x:SI (subreg:SI (x':DI)). */
706 if (GET_CODE (op0) != SUBREG)
707 return false;
708 if (GET_MODE (SUBREG_REG (op0)) != outer_mode)
709 return false;
712 next = op0;
713 break;
715 case SIGN_EXTEND:
716 case ZERO_EXTEND:
717 if (GET_MODE (rhs) != outer_mode)
718 return false;
720 op0 = XEXP (rhs, 0);
721 if (!simple_reg_p (op0))
722 return false;
724 next = op0;
725 break;
727 default:
728 return false;
731 if (GET_CODE (next) == SUBREG)
733 if (!subreg_lowpart_p (next))
734 return false;
736 nextr = SUBREG_REG (next);
737 if (GET_MODE (nextr) != outer_mode)
738 return false;
740 else
741 nextr = next;
743 def_insn = iv_get_reaching_def (insn, nextr);
744 if (def_insn == const0_rtx)
745 return false;
747 if (!def_insn)
749 if (!rtx_equal_p (nextr, reg))
750 return false;
752 *inner_step = const0_rtx;
753 *extend = UNKNOWN;
754 *inner_mode = outer_mode;
755 *outer_step = const0_rtx;
757 else if (!get_biv_step_1 (def_insn, reg,
758 inner_step, inner_mode, extend, outer_mode,
759 outer_step))
760 return false;
762 if (GET_CODE (next) == SUBREG)
764 enum machine_mode amode = GET_MODE (next);
766 if (GET_MODE_SIZE (amode) > GET_MODE_SIZE (*inner_mode))
767 return false;
769 *inner_mode = amode;
770 *inner_step = simplify_gen_binary (PLUS, outer_mode,
771 *inner_step, *outer_step);
772 *outer_step = const0_rtx;
773 *extend = UNKNOWN;
776 switch (code)
778 case REG:
779 case SUBREG:
780 break;
782 case PLUS:
783 case MINUS:
784 if (*inner_mode == outer_mode
785 /* See comment in previous switch. */
786 || GET_MODE (rhs) != outer_mode)
787 *inner_step = simplify_gen_binary (code, outer_mode,
788 *inner_step, op1);
789 else
790 *outer_step = simplify_gen_binary (code, outer_mode,
791 *outer_step, op1);
792 break;
794 case SIGN_EXTEND:
795 case ZERO_EXTEND:
796 gcc_assert (GET_MODE (op0) == *inner_mode
797 && *extend == UNKNOWN
798 && *outer_step == const0_rtx);
800 *extend = code;
801 break;
803 default:
804 gcc_unreachable ();
807 return true;
810 /* Gets the operation on register REG inside loop, in shape
812 OUTER_STEP + EXTEND_{OUTER_MODE} (SUBREG_{INNER_MODE} (REG + INNER_STEP))
814 If the operation cannot be described in this shape, return false. */
816 static bool
817 get_biv_step (rtx reg, rtx *inner_step, enum machine_mode *inner_mode,
818 enum rtx_code *extend, enum machine_mode *outer_mode,
819 rtx *outer_step)
821 *outer_mode = GET_MODE (reg);
823 if (!get_biv_step_1 (last_def[REGNO (reg)], reg,
824 inner_step, inner_mode, extend, *outer_mode,
825 outer_step))
826 return false;
828 gcc_assert ((*inner_mode == *outer_mode) != (*extend != UNKNOWN));
829 gcc_assert (*inner_mode != *outer_mode || *outer_step == const0_rtx);
831 return true;
834 /* Determines whether DEF is a biv and if so, stores its description
835 to *IV. */
837 static bool
838 iv_analyze_biv (rtx def, struct rtx_iv *iv)
840 unsigned regno;
841 rtx inner_step, outer_step;
842 enum machine_mode inner_mode, outer_mode;
843 enum rtx_code extend;
845 if (dump_file)
847 fprintf (dump_file, "Analysing ");
848 print_rtl (dump_file, def);
849 fprintf (dump_file, " for bivness.\n");
852 if (!REG_P (def))
854 if (!CONSTANT_P (def))
855 return false;
857 return iv_constant (iv, def, VOIDmode);
860 regno = REGNO (def);
861 if (last_def[regno] == const0_rtx)
863 if (dump_file)
864 fprintf (dump_file, " not simple.\n");
865 return false;
868 if (last_def[regno] && bivs[regno].analysed)
870 if (dump_file)
871 fprintf (dump_file, " already analysed.\n");
873 *iv = bivs[regno];
874 return iv->base != NULL_RTX;
877 if (!last_def[regno])
879 iv_constant (iv, def, VOIDmode);
880 goto end;
883 iv->analysed = true;
884 if (!get_biv_step (def, &inner_step, &inner_mode, &extend,
885 &outer_mode, &outer_step))
887 iv->base = NULL_RTX;
888 goto end;
891 /* Loop transforms base to es (base + inner_step) + outer_step,
892 where es means extend of subreg between inner_mode and outer_mode.
893 The corresponding induction variable is
895 es ((base - outer_step) + i * (inner_step + outer_step)) + outer_step */
897 iv->base = simplify_gen_binary (MINUS, outer_mode, def, outer_step);
898 iv->step = simplify_gen_binary (PLUS, outer_mode, inner_step, outer_step);
899 iv->mode = inner_mode;
900 iv->extend_mode = outer_mode;
901 iv->extend = extend;
902 iv->mult = const1_rtx;
903 iv->delta = outer_step;
904 iv->first_special = inner_mode != outer_mode;
906 end:
907 if (dump_file)
909 fprintf (dump_file, " ");
910 dump_iv_info (dump_file, iv);
911 fprintf (dump_file, "\n");
914 bivs[regno] = *iv;
916 return iv->base != NULL_RTX;
919 /* Analyzes operand OP of INSN and stores the result to *IV. */
921 static bool
922 iv_analyze_op (rtx insn, rtx op, struct rtx_iv *iv)
924 rtx def_insn;
925 unsigned regno;
926 bool inv = CONSTANT_P (op);
928 if (dump_file)
930 fprintf (dump_file, "Analysing operand ");
931 print_rtl (dump_file, op);
932 fprintf (dump_file, " of insn ");
933 print_rtl_single (dump_file, insn);
936 if (GET_CODE (op) == SUBREG)
938 if (!subreg_lowpart_p (op))
939 return false;
941 if (!iv_analyze_op (insn, SUBREG_REG (op), iv))
942 return false;
944 return iv_subreg (iv, GET_MODE (op));
947 if (!inv)
949 regno = REGNO (op);
950 if (!last_def[regno])
951 inv = true;
952 else if (last_def[regno] == const0_rtx)
954 if (dump_file)
955 fprintf (dump_file, " not simple.\n");
956 return false;
960 if (inv)
962 iv_constant (iv, op, VOIDmode);
964 if (dump_file)
966 fprintf (dump_file, " ");
967 dump_iv_info (dump_file, iv);
968 fprintf (dump_file, "\n");
970 return true;
973 def_insn = iv_get_reaching_def (insn, op);
974 if (def_insn == const0_rtx)
976 if (dump_file)
977 fprintf (dump_file, " not simple.\n");
978 return false;
981 return iv_analyze (def_insn, op, iv);
984 /* Analyzes iv DEF defined in INSN and stores the result to *IV. */
986 bool
987 iv_analyze (rtx insn, rtx def, struct rtx_iv *iv)
989 unsigned uid;
990 rtx set, rhs, mby = NULL_RTX, tmp;
991 rtx op0 = NULL_RTX, op1 = NULL_RTX;
992 struct rtx_iv iv0, iv1;
993 enum machine_mode amode;
994 enum rtx_code code;
996 if (insn == const0_rtx)
997 return false;
999 if (GET_CODE (def) == SUBREG)
1001 if (!subreg_lowpart_p (def))
1002 return false;
1004 if (!iv_analyze (insn, SUBREG_REG (def), iv))
1005 return false;
1007 return iv_subreg (iv, GET_MODE (def));
1010 if (!insn)
1011 return iv_analyze_biv (def, iv);
1013 if (dump_file)
1015 fprintf (dump_file, "Analysing def of ");
1016 print_rtl (dump_file, def);
1017 fprintf (dump_file, " in insn ");
1018 print_rtl_single (dump_file, insn);
1021 uid = INSN_UID (insn);
1022 if (insn_info[uid].iv.analysed)
1024 if (dump_file)
1025 fprintf (dump_file, " already analysed.\n");
1026 *iv = insn_info[uid].iv;
1027 return iv->base != NULL_RTX;
1030 iv->mode = VOIDmode;
1031 iv->base = NULL_RTX;
1032 iv->step = NULL_RTX;
1034 set = single_set (insn);
1035 rhs = find_reg_equal_equiv_note (insn);
1036 if (rhs)
1037 rhs = XEXP (rhs, 0);
1038 else
1039 rhs = SET_SRC (set);
1040 code = GET_CODE (rhs);
1042 if (CONSTANT_P (rhs))
1044 op0 = rhs;
1045 amode = GET_MODE (def);
1047 else
1049 switch (code)
1051 case SUBREG:
1052 if (!subreg_lowpart_p (rhs))
1053 goto end;
1054 op0 = rhs;
1055 break;
1057 case REG:
1058 op0 = rhs;
1059 break;
1061 case SIGN_EXTEND:
1062 case ZERO_EXTEND:
1063 case NEG:
1064 op0 = XEXP (rhs, 0);
1065 break;
1067 case PLUS:
1068 case MINUS:
1069 op0 = XEXP (rhs, 0);
1070 op1 = XEXP (rhs, 1);
1071 break;
1073 case MULT:
1074 op0 = XEXP (rhs, 0);
1075 mby = XEXP (rhs, 1);
1076 if (!CONSTANT_P (mby))
1078 gcc_assert (CONSTANT_P (op0));
1079 tmp = op0;
1080 op0 = mby;
1081 mby = tmp;
1083 break;
1085 case ASHIFT:
1086 gcc_assert (!CONSTANT_P (XEXP (rhs, 0)));
1087 op0 = XEXP (rhs, 0);
1088 mby = XEXP (rhs, 1);
1089 break;
1091 default:
1092 gcc_unreachable ();
1095 amode = GET_MODE (rhs);
1098 if (op0)
1100 if (!iv_analyze_op (insn, op0, &iv0))
1101 goto end;
1103 if (iv0.mode == VOIDmode)
1105 iv0.mode = amode;
1106 iv0.extend_mode = amode;
1110 if (op1)
1112 if (!iv_analyze_op (insn, op1, &iv1))
1113 goto end;
1115 if (iv1.mode == VOIDmode)
1117 iv1.mode = amode;
1118 iv1.extend_mode = amode;
1122 switch (code)
1124 case SIGN_EXTEND:
1125 case ZERO_EXTEND:
1126 if (!iv_extend (&iv0, code, amode))
1127 goto end;
1128 break;
1130 case NEG:
1131 if (!iv_neg (&iv0))
1132 goto end;
1133 break;
1135 case PLUS:
1136 case MINUS:
1137 if (!iv_add (&iv0, &iv1, code))
1138 goto end;
1139 break;
1141 case MULT:
1142 if (!iv_mult (&iv0, mby))
1143 goto end;
1144 break;
1146 case ASHIFT:
1147 if (!iv_shift (&iv0, mby))
1148 goto end;
1149 break;
1151 default:
1152 break;
1155 *iv = iv0;
1157 end:
1158 iv->analysed = true;
1159 insn_info[uid].iv = *iv;
1161 if (dump_file)
1163 print_rtl (dump_file, def);
1164 fprintf (dump_file, " in insn ");
1165 print_rtl_single (dump_file, insn);
1166 fprintf (dump_file, " is ");
1167 dump_iv_info (dump_file, iv);
1168 fprintf (dump_file, "\n");
1171 return iv->base != NULL_RTX;
1174 /* Checks whether definition of register REG in INSN a basic induction
1175 variable. IV analysis must have been initialized (via a call to
1176 iv_analysis_loop_init) for this function to produce a result. */
1178 bool
1179 biv_p (rtx insn, rtx reg)
1181 struct rtx_iv iv;
1183 if (!REG_P (reg))
1184 return false;
1186 if (last_def[REGNO (reg)] != insn)
1187 return false;
1189 return iv_analyze_biv (reg, &iv);
1192 /* Calculates value of IV at ITERATION-th iteration. */
1195 get_iv_value (struct rtx_iv *iv, rtx iteration)
1197 rtx val;
1199 /* We would need to generate some if_then_else patterns, and so far
1200 it is not needed anywhere. */
1201 gcc_assert (!iv->first_special);
1203 if (iv->step != const0_rtx && iteration != const0_rtx)
1204 val = simplify_gen_binary (PLUS, iv->extend_mode, iv->base,
1205 simplify_gen_binary (MULT, iv->extend_mode,
1206 iv->step, iteration));
1207 else
1208 val = iv->base;
1210 if (iv->extend_mode == iv->mode)
1211 return val;
1213 val = lowpart_subreg (iv->mode, val, iv->extend_mode);
1215 if (iv->extend == UNKNOWN)
1216 return val;
1218 val = simplify_gen_unary (iv->extend, iv->extend_mode, val, iv->mode);
1219 val = simplify_gen_binary (PLUS, iv->extend_mode, iv->delta,
1220 simplify_gen_binary (MULT, iv->extend_mode,
1221 iv->mult, val));
1223 return val;
1226 /* Free the data for an induction variable analysis. */
1228 void
1229 iv_analysis_done (void)
1231 max_insn_no = 0;
1232 max_reg_no = 0;
1233 if (insn_info)
1235 free (insn_info);
1236 insn_info = NULL;
1238 if (last_def)
1240 free (last_def);
1241 last_def = NULL;
1243 if (bivs)
1245 free (bivs);
1246 bivs = NULL;
1250 /* Computes inverse to X modulo (1 << MOD). */
1252 static unsigned HOST_WIDEST_INT
1253 inverse (unsigned HOST_WIDEST_INT x, int mod)
1255 unsigned HOST_WIDEST_INT mask =
1256 ((unsigned HOST_WIDEST_INT) 1 << (mod - 1) << 1) - 1;
1257 unsigned HOST_WIDEST_INT rslt = 1;
1258 int i;
1260 for (i = 0; i < mod - 1; i++)
1262 rslt = (rslt * x) & mask;
1263 x = (x * x) & mask;
1266 return rslt;
1269 /* Tries to estimate the maximum number of iterations. */
1271 static unsigned HOST_WIDEST_INT
1272 determine_max_iter (struct niter_desc *desc)
1274 rtx niter = desc->niter_expr;
1275 rtx mmin, mmax, left, right;
1276 unsigned HOST_WIDEST_INT nmax, inc;
1278 if (GET_CODE (niter) == AND
1279 && GET_CODE (XEXP (niter, 0)) == CONST_INT)
1281 nmax = INTVAL (XEXP (niter, 0));
1282 if (!(nmax & (nmax + 1)))
1284 desc->niter_max = nmax;
1285 return nmax;
1289 get_mode_bounds (desc->mode, desc->signed_p, desc->mode, &mmin, &mmax);
1290 nmax = INTVAL (mmax) - INTVAL (mmin);
1292 if (GET_CODE (niter) == UDIV)
1294 if (GET_CODE (XEXP (niter, 1)) != CONST_INT)
1296 desc->niter_max = nmax;
1297 return nmax;
1299 inc = INTVAL (XEXP (niter, 1));
1300 niter = XEXP (niter, 0);
1302 else
1303 inc = 1;
1305 if (GET_CODE (niter) == PLUS)
1307 left = XEXP (niter, 0);
1308 right = XEXP (niter, 0);
1310 if (GET_CODE (right) == CONST_INT)
1311 right = GEN_INT (-INTVAL (right));
1313 else if (GET_CODE (niter) == MINUS)
1315 left = XEXP (niter, 0);
1316 right = XEXP (niter, 0);
1318 else
1320 left = niter;
1321 right = mmin;
1324 if (GET_CODE (left) == CONST_INT)
1325 mmax = left;
1326 if (GET_CODE (right) == CONST_INT)
1327 mmin = right;
1328 nmax = INTVAL (mmax) - INTVAL (mmin);
1330 desc->niter_max = nmax / inc;
1331 return nmax / inc;
1334 /* Checks whether register *REG is in set ALT. Callback for for_each_rtx. */
1336 static int
1337 altered_reg_used (rtx *reg, void *alt)
1339 if (!REG_P (*reg))
1340 return 0;
1342 return REGNO_REG_SET_P (alt, REGNO (*reg));
1345 /* Marks registers altered by EXPR in set ALT. */
1347 static void
1348 mark_altered (rtx expr, rtx by ATTRIBUTE_UNUSED, void *alt)
1350 if (GET_CODE (expr) == SUBREG)
1351 expr = SUBREG_REG (expr);
1352 if (!REG_P (expr))
1353 return;
1355 SET_REGNO_REG_SET (alt, REGNO (expr));
1358 /* Checks whether RHS is simple enough to process. */
1360 static bool
1361 simple_rhs_p (rtx rhs)
1363 rtx op0, op1;
1365 if (CONSTANT_P (rhs)
1366 || REG_P (rhs))
1367 return true;
1369 switch (GET_CODE (rhs))
1371 case PLUS:
1372 case MINUS:
1373 op0 = XEXP (rhs, 0);
1374 op1 = XEXP (rhs, 1);
1375 /* Allow reg + const sets only. */
1376 if (REG_P (op0) && CONSTANT_P (op1))
1377 return true;
1378 if (REG_P (op1) && CONSTANT_P (op0))
1379 return true;
1381 return false;
1383 default:
1384 return false;
1388 /* Simplifies *EXPR using assignment in INSN. ALTERED is the set of registers
1389 altered so far. */
1391 static void
1392 simplify_using_assignment (rtx insn, rtx *expr, regset altered)
1394 rtx set = single_set (insn);
1395 rtx lhs = NULL_RTX, rhs;
1396 bool ret = false;
1398 if (set)
1400 lhs = SET_DEST (set);
1401 if (!REG_P (lhs)
1402 || altered_reg_used (&lhs, altered))
1403 ret = true;
1405 else
1406 ret = true;
1408 note_stores (PATTERN (insn), mark_altered, altered);
1409 if (CALL_P (insn))
1411 int i;
1413 /* Kill all call clobbered registers. */
1414 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1415 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
1416 SET_REGNO_REG_SET (altered, i);
1419 if (ret)
1420 return;
1422 rhs = find_reg_equal_equiv_note (insn);
1423 if (rhs)
1424 rhs = XEXP (rhs, 0);
1425 else
1426 rhs = SET_SRC (set);
1428 if (!simple_rhs_p (rhs))
1429 return;
1431 if (for_each_rtx (&rhs, altered_reg_used, altered))
1432 return;
1434 *expr = simplify_replace_rtx (*expr, lhs, rhs);
1437 /* Checks whether A implies B. */
1439 static bool
1440 implies_p (rtx a, rtx b)
1442 rtx op0, op1, opb0, opb1, r;
1443 enum machine_mode mode;
1445 if (GET_CODE (a) == EQ)
1447 op0 = XEXP (a, 0);
1448 op1 = XEXP (a, 1);
1450 if (REG_P (op0))
1452 r = simplify_replace_rtx (b, op0, op1);
1453 if (r == const_true_rtx)
1454 return true;
1457 if (REG_P (op1))
1459 r = simplify_replace_rtx (b, op1, op0);
1460 if (r == const_true_rtx)
1461 return true;
1465 /* A < B implies A + 1 <= B. */
1466 if ((GET_CODE (a) == GT || GET_CODE (a) == LT)
1467 && (GET_CODE (b) == GE || GET_CODE (b) == LE))
1469 op0 = XEXP (a, 0);
1470 op1 = XEXP (a, 1);
1471 opb0 = XEXP (b, 0);
1472 opb1 = XEXP (b, 1);
1474 if (GET_CODE (a) == GT)
1476 r = op0;
1477 op0 = op1;
1478 op1 = r;
1481 if (GET_CODE (b) == GE)
1483 r = opb0;
1484 opb0 = opb1;
1485 opb1 = r;
1488 mode = GET_MODE (op0);
1489 if (mode != GET_MODE (opb0))
1490 mode = VOIDmode;
1491 else if (mode == VOIDmode)
1493 mode = GET_MODE (op1);
1494 if (mode != GET_MODE (opb1))
1495 mode = VOIDmode;
1498 if (mode != VOIDmode
1499 && rtx_equal_p (op1, opb1)
1500 && simplify_gen_binary (MINUS, mode, opb0, op0) == const1_rtx)
1501 return true;
1504 return false;
1507 /* Canonicalizes COND so that
1509 (1) Ensure that operands are ordered according to
1510 swap_commutative_operands_p.
1511 (2) (LE x const) will be replaced with (LT x <const+1>) and similarly
1512 for GE, GEU, and LEU. */
1515 canon_condition (rtx cond)
1517 rtx tem;
1518 rtx op0, op1;
1519 enum rtx_code code;
1520 enum machine_mode mode;
1522 code = GET_CODE (cond);
1523 op0 = XEXP (cond, 0);
1524 op1 = XEXP (cond, 1);
1526 if (swap_commutative_operands_p (op0, op1))
1528 code = swap_condition (code);
1529 tem = op0;
1530 op0 = op1;
1531 op1 = tem;
1534 mode = GET_MODE (op0);
1535 if (mode == VOIDmode)
1536 mode = GET_MODE (op1);
1537 gcc_assert (mode != VOIDmode);
1539 if (GET_CODE (op1) == CONST_INT
1540 && GET_MODE_CLASS (mode) != MODE_CC
1541 && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
1543 HOST_WIDE_INT const_val = INTVAL (op1);
1544 unsigned HOST_WIDE_INT uconst_val = const_val;
1545 unsigned HOST_WIDE_INT max_val
1546 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (mode);
1548 switch (code)
1550 case LE:
1551 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
1552 code = LT, op1 = gen_int_mode (const_val + 1, GET_MODE (op0));
1553 break;
1555 /* When cross-compiling, const_val might be sign-extended from
1556 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
1557 case GE:
1558 if ((HOST_WIDE_INT) (const_val & max_val)
1559 != (((HOST_WIDE_INT) 1
1560 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
1561 code = GT, op1 = gen_int_mode (const_val - 1, mode);
1562 break;
1564 case LEU:
1565 if (uconst_val < max_val)
1566 code = LTU, op1 = gen_int_mode (uconst_val + 1, mode);
1567 break;
1569 case GEU:
1570 if (uconst_val != 0)
1571 code = GTU, op1 = gen_int_mode (uconst_val - 1, mode);
1572 break;
1574 default:
1575 break;
1579 if (op0 != XEXP (cond, 0)
1580 || op1 != XEXP (cond, 1)
1581 || code != GET_CODE (cond)
1582 || GET_MODE (cond) != SImode)
1583 cond = gen_rtx_fmt_ee (code, SImode, op0, op1);
1585 return cond;
1588 /* Tries to use the fact that COND holds to simplify EXPR. ALTERED is the
1589 set of altered regs. */
1591 void
1592 simplify_using_condition (rtx cond, rtx *expr, regset altered)
1594 rtx rev, reve, exp = *expr;
1596 if (!COMPARISON_P (exp))
1597 return;
1599 /* If some register gets altered later, we do not really speak about its
1600 value at the time of comparison. */
1601 if (altered
1602 && for_each_rtx (&cond, altered_reg_used, altered))
1603 return;
1605 rev = reversed_condition (cond);
1606 reve = reversed_condition (exp);
1608 cond = canon_condition (cond);
1609 exp = canon_condition (exp);
1610 if (rev)
1611 rev = canon_condition (rev);
1612 if (reve)
1613 reve = canon_condition (reve);
1615 if (rtx_equal_p (exp, cond))
1617 *expr = const_true_rtx;
1618 return;
1622 if (rev && rtx_equal_p (exp, rev))
1624 *expr = const0_rtx;
1625 return;
1628 if (implies_p (cond, exp))
1630 *expr = const_true_rtx;
1631 return;
1634 if (reve && implies_p (cond, reve))
1636 *expr = const0_rtx;
1637 return;
1640 /* A proof by contradiction. If *EXPR implies (not cond), *EXPR must
1641 be false. */
1642 if (rev && implies_p (exp, rev))
1644 *expr = const0_rtx;
1645 return;
1648 /* Similarly, If (not *EXPR) implies (not cond), *EXPR must be true. */
1649 if (rev && reve && implies_p (reve, rev))
1651 *expr = const_true_rtx;
1652 return;
1655 /* We would like to have some other tests here. TODO. */
1657 return;
1660 /* Use relationship between A and *B to eventually eliminate *B.
1661 OP is the operation we consider. */
1663 static void
1664 eliminate_implied_condition (enum rtx_code op, rtx a, rtx *b)
1666 switch (op)
1668 case AND:
1669 /* If A implies *B, we may replace *B by true. */
1670 if (implies_p (a, *b))
1671 *b = const_true_rtx;
1672 break;
1674 case IOR:
1675 /* If *B implies A, we may replace *B by false. */
1676 if (implies_p (*b, a))
1677 *b = const0_rtx;
1678 break;
1680 default:
1681 gcc_unreachable ();
1685 /* Eliminates the conditions in TAIL that are implied by HEAD. OP is the
1686 operation we consider. */
1688 static void
1689 eliminate_implied_conditions (enum rtx_code op, rtx *head, rtx tail)
1691 rtx elt;
1693 for (elt = tail; elt; elt = XEXP (elt, 1))
1694 eliminate_implied_condition (op, *head, &XEXP (elt, 0));
1695 for (elt = tail; elt; elt = XEXP (elt, 1))
1696 eliminate_implied_condition (op, XEXP (elt, 0), head);
1699 /* Simplifies *EXPR using initial values at the start of the LOOP. If *EXPR
1700 is a list, its elements are assumed to be combined using OP. */
1702 static void
1703 simplify_using_initial_values (struct loop *loop, enum rtx_code op, rtx *expr)
1705 rtx head, tail, insn;
1706 rtx neutral, aggr;
1707 regset altered;
1708 edge e;
1710 if (!*expr)
1711 return;
1713 if (CONSTANT_P (*expr))
1714 return;
1716 if (GET_CODE (*expr) == EXPR_LIST)
1718 head = XEXP (*expr, 0);
1719 tail = XEXP (*expr, 1);
1721 eliminate_implied_conditions (op, &head, tail);
1723 switch (op)
1725 case AND:
1726 neutral = const_true_rtx;
1727 aggr = const0_rtx;
1728 break;
1730 case IOR:
1731 neutral = const0_rtx;
1732 aggr = const_true_rtx;
1733 break;
1735 default:
1736 gcc_unreachable ();
1739 simplify_using_initial_values (loop, UNKNOWN, &head);
1740 if (head == aggr)
1742 XEXP (*expr, 0) = aggr;
1743 XEXP (*expr, 1) = NULL_RTX;
1744 return;
1746 else if (head == neutral)
1748 *expr = tail;
1749 simplify_using_initial_values (loop, op, expr);
1750 return;
1752 simplify_using_initial_values (loop, op, &tail);
1754 if (tail && XEXP (tail, 0) == aggr)
1756 *expr = tail;
1757 return;
1760 XEXP (*expr, 0) = head;
1761 XEXP (*expr, 1) = tail;
1762 return;
1765 gcc_assert (op == UNKNOWN);
1767 e = loop_preheader_edge (loop);
1768 if (e->src == ENTRY_BLOCK_PTR)
1769 return;
1771 altered = ALLOC_REG_SET (&reg_obstack);
1773 while (1)
1775 insn = BB_END (e->src);
1776 if (any_condjump_p (insn))
1778 rtx cond = get_condition (BB_END (e->src), NULL, false, true);
1780 if (cond && (e->flags & EDGE_FALLTHRU))
1781 cond = reversed_condition (cond);
1782 if (cond)
1784 simplify_using_condition (cond, expr, altered);
1785 if (CONSTANT_P (*expr))
1787 FREE_REG_SET (altered);
1788 return;
1793 FOR_BB_INSNS_REVERSE (e->src, insn)
1795 if (!INSN_P (insn))
1796 continue;
1798 simplify_using_assignment (insn, expr, altered);
1799 if (CONSTANT_P (*expr))
1801 FREE_REG_SET (altered);
1802 return;
1806 if (!single_pred_p (e->src)
1807 || single_pred (e->src) == ENTRY_BLOCK_PTR)
1808 break;
1809 e = single_pred_edge (e->src);
1812 FREE_REG_SET (altered);
1815 /* Transforms invariant IV into MODE. Adds assumptions based on the fact
1816 that IV occurs as left operands of comparison COND and its signedness
1817 is SIGNED_P to DESC. */
1819 static void
1820 shorten_into_mode (struct rtx_iv *iv, enum machine_mode mode,
1821 enum rtx_code cond, bool signed_p, struct niter_desc *desc)
1823 rtx mmin, mmax, cond_over, cond_under;
1825 get_mode_bounds (mode, signed_p, iv->extend_mode, &mmin, &mmax);
1826 cond_under = simplify_gen_relational (LT, SImode, iv->extend_mode,
1827 iv->base, mmin);
1828 cond_over = simplify_gen_relational (GT, SImode, iv->extend_mode,
1829 iv->base, mmax);
1831 switch (cond)
1833 case LE:
1834 case LT:
1835 case LEU:
1836 case LTU:
1837 if (cond_under != const0_rtx)
1838 desc->infinite =
1839 alloc_EXPR_LIST (0, cond_under, desc->infinite);
1840 if (cond_over != const0_rtx)
1841 desc->noloop_assumptions =
1842 alloc_EXPR_LIST (0, cond_over, desc->noloop_assumptions);
1843 break;
1845 case GE:
1846 case GT:
1847 case GEU:
1848 case GTU:
1849 if (cond_over != const0_rtx)
1850 desc->infinite =
1851 alloc_EXPR_LIST (0, cond_over, desc->infinite);
1852 if (cond_under != const0_rtx)
1853 desc->noloop_assumptions =
1854 alloc_EXPR_LIST (0, cond_under, desc->noloop_assumptions);
1855 break;
1857 case NE:
1858 if (cond_over != const0_rtx)
1859 desc->infinite =
1860 alloc_EXPR_LIST (0, cond_over, desc->infinite);
1861 if (cond_under != const0_rtx)
1862 desc->infinite =
1863 alloc_EXPR_LIST (0, cond_under, desc->infinite);
1864 break;
1866 default:
1867 gcc_unreachable ();
1870 iv->mode = mode;
1871 iv->extend = signed_p ? SIGN_EXTEND : ZERO_EXTEND;
1874 /* Transforms IV0 and IV1 compared by COND so that they are both compared as
1875 subregs of the same mode if possible (sometimes it is necessary to add
1876 some assumptions to DESC). */
1878 static bool
1879 canonicalize_iv_subregs (struct rtx_iv *iv0, struct rtx_iv *iv1,
1880 enum rtx_code cond, struct niter_desc *desc)
1882 enum machine_mode comp_mode;
1883 bool signed_p;
1885 /* If the ivs behave specially in the first iteration, or are
1886 added/multiplied after extending, we ignore them. */
1887 if (iv0->first_special || iv0->mult != const1_rtx || iv0->delta != const0_rtx)
1888 return false;
1889 if (iv1->first_special || iv1->mult != const1_rtx || iv1->delta != const0_rtx)
1890 return false;
1892 /* If there is some extend, it must match signedness of the comparison. */
1893 switch (cond)
1895 case LE:
1896 case LT:
1897 if (iv0->extend == ZERO_EXTEND
1898 || iv1->extend == ZERO_EXTEND)
1899 return false;
1900 signed_p = true;
1901 break;
1903 case LEU:
1904 case LTU:
1905 if (iv0->extend == SIGN_EXTEND
1906 || iv1->extend == SIGN_EXTEND)
1907 return false;
1908 signed_p = false;
1909 break;
1911 case NE:
1912 if (iv0->extend != UNKNOWN
1913 && iv1->extend != UNKNOWN
1914 && iv0->extend != iv1->extend)
1915 return false;
1917 signed_p = false;
1918 if (iv0->extend != UNKNOWN)
1919 signed_p = iv0->extend == SIGN_EXTEND;
1920 if (iv1->extend != UNKNOWN)
1921 signed_p = iv1->extend == SIGN_EXTEND;
1922 break;
1924 default:
1925 gcc_unreachable ();
1928 /* Values of both variables should be computed in the same mode. These
1929 might indeed be different, if we have comparison like
1931 (compare (subreg:SI (iv0)) (subreg:SI (iv1)))
1933 and iv0 and iv1 are both ivs iterating in SI mode, but calculated
1934 in different modes. This does not seem impossible to handle, but
1935 it hardly ever occurs in practice.
1937 The only exception is the case when one of operands is invariant.
1938 For example pentium 3 generates comparisons like
1939 (lt (subreg:HI (reg:SI)) 100). Here we assign HImode to 100, but we
1940 definitely do not want this prevent the optimization. */
1941 comp_mode = iv0->extend_mode;
1942 if (GET_MODE_BITSIZE (comp_mode) < GET_MODE_BITSIZE (iv1->extend_mode))
1943 comp_mode = iv1->extend_mode;
1945 if (iv0->extend_mode != comp_mode)
1947 if (iv0->mode != iv0->extend_mode
1948 || iv0->step != const0_rtx)
1949 return false;
1951 iv0->base = simplify_gen_unary (signed_p ? SIGN_EXTEND : ZERO_EXTEND,
1952 comp_mode, iv0->base, iv0->mode);
1953 iv0->extend_mode = comp_mode;
1956 if (iv1->extend_mode != comp_mode)
1958 if (iv1->mode != iv1->extend_mode
1959 || iv1->step != const0_rtx)
1960 return false;
1962 iv1->base = simplify_gen_unary (signed_p ? SIGN_EXTEND : ZERO_EXTEND,
1963 comp_mode, iv1->base, iv1->mode);
1964 iv1->extend_mode = comp_mode;
1967 /* Check that both ivs belong to a range of a single mode. If one of the
1968 operands is an invariant, we may need to shorten it into the common
1969 mode. */
1970 if (iv0->mode == iv0->extend_mode
1971 && iv0->step == const0_rtx
1972 && iv0->mode != iv1->mode)
1973 shorten_into_mode (iv0, iv1->mode, cond, signed_p, desc);
1975 if (iv1->mode == iv1->extend_mode
1976 && iv1->step == const0_rtx
1977 && iv0->mode != iv1->mode)
1978 shorten_into_mode (iv1, iv0->mode, swap_condition (cond), signed_p, desc);
1980 if (iv0->mode != iv1->mode)
1981 return false;
1983 desc->mode = iv0->mode;
1984 desc->signed_p = signed_p;
1986 return true;
1989 /* Computes number of iterations of the CONDITION in INSN in LOOP and stores
1990 the result into DESC. Very similar to determine_number_of_iterations
1991 (basically its rtl version), complicated by things like subregs. */
1993 static void
1994 iv_number_of_iterations (struct loop *loop, rtx insn, rtx condition,
1995 struct niter_desc *desc)
1997 rtx op0, op1, delta, step, bound, may_xform, def_insn, tmp, tmp0, tmp1;
1998 struct rtx_iv iv0, iv1, tmp_iv;
1999 rtx assumption, may_not_xform;
2000 enum rtx_code cond;
2001 enum machine_mode mode, comp_mode;
2002 rtx mmin, mmax, mode_mmin, mode_mmax;
2003 unsigned HOST_WIDEST_INT s, size, d, inv;
2004 HOST_WIDEST_INT up, down, inc, step_val;
2005 int was_sharp = false;
2006 rtx old_niter;
2007 bool step_is_pow2;
2009 /* The meaning of these assumptions is this:
2010 if !assumptions
2011 then the rest of information does not have to be valid
2012 if noloop_assumptions then the loop does not roll
2013 if infinite then this exit is never used */
2015 desc->assumptions = NULL_RTX;
2016 desc->noloop_assumptions = NULL_RTX;
2017 desc->infinite = NULL_RTX;
2018 desc->simple_p = true;
2020 desc->const_iter = false;
2021 desc->niter_expr = NULL_RTX;
2022 desc->niter_max = 0;
2024 cond = GET_CODE (condition);
2025 gcc_assert (COMPARISON_P (condition));
2027 mode = GET_MODE (XEXP (condition, 0));
2028 if (mode == VOIDmode)
2029 mode = GET_MODE (XEXP (condition, 1));
2030 /* The constant comparisons should be folded. */
2031 gcc_assert (mode != VOIDmode);
2033 /* We only handle integers or pointers. */
2034 if (GET_MODE_CLASS (mode) != MODE_INT
2035 && GET_MODE_CLASS (mode) != MODE_PARTIAL_INT)
2036 goto fail;
2038 op0 = XEXP (condition, 0);
2039 def_insn = iv_get_reaching_def (insn, op0);
2040 if (!iv_analyze (def_insn, op0, &iv0))
2041 goto fail;
2042 if (iv0.extend_mode == VOIDmode)
2043 iv0.mode = iv0.extend_mode = mode;
2045 op1 = XEXP (condition, 1);
2046 def_insn = iv_get_reaching_def (insn, op1);
2047 if (!iv_analyze (def_insn, op1, &iv1))
2048 goto fail;
2049 if (iv1.extend_mode == VOIDmode)
2050 iv1.mode = iv1.extend_mode = mode;
2052 if (GET_MODE_BITSIZE (iv0.extend_mode) > HOST_BITS_PER_WIDE_INT
2053 || GET_MODE_BITSIZE (iv1.extend_mode) > HOST_BITS_PER_WIDE_INT)
2054 goto fail;
2056 /* Check condition and normalize it. */
2058 switch (cond)
2060 case GE:
2061 case GT:
2062 case GEU:
2063 case GTU:
2064 tmp_iv = iv0; iv0 = iv1; iv1 = tmp_iv;
2065 cond = swap_condition (cond);
2066 break;
2067 case NE:
2068 case LE:
2069 case LEU:
2070 case LT:
2071 case LTU:
2072 break;
2073 default:
2074 goto fail;
2077 /* Handle extends. This is relatively nontrivial, so we only try in some
2078 easy cases, when we can canonicalize the ivs (possibly by adding some
2079 assumptions) to shape subreg (base + i * step). This function also fills
2080 in desc->mode and desc->signed_p. */
2082 if (!canonicalize_iv_subregs (&iv0, &iv1, cond, desc))
2083 goto fail;
2085 comp_mode = iv0.extend_mode;
2086 mode = iv0.mode;
2087 size = GET_MODE_BITSIZE (mode);
2088 get_mode_bounds (mode, (cond == LE || cond == LT), comp_mode, &mmin, &mmax);
2089 mode_mmin = lowpart_subreg (mode, mmin, comp_mode);
2090 mode_mmax = lowpart_subreg (mode, mmax, comp_mode);
2092 if (GET_CODE (iv0.step) != CONST_INT || GET_CODE (iv1.step) != CONST_INT)
2093 goto fail;
2095 /* We can take care of the case of two induction variables chasing each other
2096 if the test is NE. I have never seen a loop using it, but still it is
2097 cool. */
2098 if (iv0.step != const0_rtx && iv1.step != const0_rtx)
2100 if (cond != NE)
2101 goto fail;
2103 iv0.step = simplify_gen_binary (MINUS, comp_mode, iv0.step, iv1.step);
2104 iv1.step = const0_rtx;
2107 /* This is either infinite loop or the one that ends immediately, depending
2108 on initial values. Unswitching should remove this kind of conditions. */
2109 if (iv0.step == const0_rtx && iv1.step == const0_rtx)
2110 goto fail;
2112 if (cond != NE)
2114 if (iv0.step == const0_rtx)
2115 step_val = -INTVAL (iv1.step);
2116 else
2117 step_val = INTVAL (iv0.step);
2119 /* Ignore loops of while (i-- < 10) type. */
2120 if (step_val < 0)
2121 goto fail;
2123 step_is_pow2 = !(step_val & (step_val - 1));
2125 else
2127 /* We do not care about whether the step is power of two in this
2128 case. */
2129 step_is_pow2 = false;
2130 step_val = 0;
2133 /* Some more condition normalization. We must record some assumptions
2134 due to overflows. */
2135 switch (cond)
2137 case LT:
2138 case LTU:
2139 /* We want to take care only of non-sharp relationals; this is easy,
2140 as in cases the overflow would make the transformation unsafe
2141 the loop does not roll. Seemingly it would make more sense to want
2142 to take care of sharp relationals instead, as NE is more similar to
2143 them, but the problem is that here the transformation would be more
2144 difficult due to possibly infinite loops. */
2145 if (iv0.step == const0_rtx)
2147 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2148 assumption = simplify_gen_relational (EQ, SImode, mode, tmp,
2149 mode_mmax);
2150 if (assumption == const_true_rtx)
2151 goto zero_iter;
2152 iv0.base = simplify_gen_binary (PLUS, comp_mode,
2153 iv0.base, const1_rtx);
2155 else
2157 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2158 assumption = simplify_gen_relational (EQ, SImode, mode, tmp,
2159 mode_mmin);
2160 if (assumption == const_true_rtx)
2161 goto zero_iter;
2162 iv1.base = simplify_gen_binary (PLUS, comp_mode,
2163 iv1.base, constm1_rtx);
2166 if (assumption != const0_rtx)
2167 desc->noloop_assumptions =
2168 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2169 cond = (cond == LT) ? LE : LEU;
2171 /* It will be useful to be able to tell the difference once more in
2172 LE -> NE reduction. */
2173 was_sharp = true;
2174 break;
2175 default: ;
2178 /* Take care of trivially infinite loops. */
2179 if (cond != NE)
2181 if (iv0.step == const0_rtx)
2183 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2184 if (rtx_equal_p (tmp, mode_mmin))
2186 desc->infinite =
2187 alloc_EXPR_LIST (0, const_true_rtx, NULL_RTX);
2188 return;
2191 else
2193 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2194 if (rtx_equal_p (tmp, mode_mmax))
2196 desc->infinite =
2197 alloc_EXPR_LIST (0, const_true_rtx, NULL_RTX);
2198 return;
2203 /* If we can we want to take care of NE conditions instead of size
2204 comparisons, as they are much more friendly (most importantly
2205 this takes care of special handling of loops with step 1). We can
2206 do it if we first check that upper bound is greater or equal to
2207 lower bound, their difference is constant c modulo step and that
2208 there is not an overflow. */
2209 if (cond != NE)
2211 if (iv0.step == const0_rtx)
2212 step = simplify_gen_unary (NEG, comp_mode, iv1.step, comp_mode);
2213 else
2214 step = iv0.step;
2215 delta = simplify_gen_binary (MINUS, comp_mode, iv1.base, iv0.base);
2216 delta = lowpart_subreg (mode, delta, comp_mode);
2217 delta = simplify_gen_binary (UMOD, mode, delta, step);
2218 may_xform = const0_rtx;
2219 may_not_xform = const_true_rtx;
2221 if (GET_CODE (delta) == CONST_INT)
2223 if (was_sharp && INTVAL (delta) == INTVAL (step) - 1)
2225 /* A special case. We have transformed condition of type
2226 for (i = 0; i < 4; i += 4)
2227 into
2228 for (i = 0; i <= 3; i += 4)
2229 obviously if the test for overflow during that transformation
2230 passed, we cannot overflow here. Most importantly any
2231 loop with sharp end condition and step 1 falls into this
2232 category, so handling this case specially is definitely
2233 worth the troubles. */
2234 may_xform = const_true_rtx;
2236 else if (iv0.step == const0_rtx)
2238 bound = simplify_gen_binary (PLUS, comp_mode, mmin, step);
2239 bound = simplify_gen_binary (MINUS, comp_mode, bound, delta);
2240 bound = lowpart_subreg (mode, bound, comp_mode);
2241 tmp = lowpart_subreg (mode, iv0.base, comp_mode);
2242 may_xform = simplify_gen_relational (cond, SImode, mode,
2243 bound, tmp);
2244 may_not_xform = simplify_gen_relational (reverse_condition (cond),
2245 SImode, mode,
2246 bound, tmp);
2248 else
2250 bound = simplify_gen_binary (MINUS, comp_mode, mmax, step);
2251 bound = simplify_gen_binary (PLUS, comp_mode, bound, delta);
2252 bound = lowpart_subreg (mode, bound, comp_mode);
2253 tmp = lowpart_subreg (mode, iv1.base, comp_mode);
2254 may_xform = simplify_gen_relational (cond, SImode, mode,
2255 tmp, bound);
2256 may_not_xform = simplify_gen_relational (reverse_condition (cond),
2257 SImode, mode,
2258 tmp, bound);
2262 if (may_xform != const0_rtx)
2264 /* We perform the transformation always provided that it is not
2265 completely senseless. This is OK, as we would need this assumption
2266 to determine the number of iterations anyway. */
2267 if (may_xform != const_true_rtx)
2269 /* If the step is a power of two and the final value we have
2270 computed overflows, the cycle is infinite. Otherwise it
2271 is nontrivial to compute the number of iterations. */
2272 if (step_is_pow2)
2273 desc->infinite = alloc_EXPR_LIST (0, may_not_xform,
2274 desc->infinite);
2275 else
2276 desc->assumptions = alloc_EXPR_LIST (0, may_xform,
2277 desc->assumptions);
2280 /* We are going to lose some information about upper bound on
2281 number of iterations in this step, so record the information
2282 here. */
2283 inc = INTVAL (iv0.step) - INTVAL (iv1.step);
2284 if (GET_CODE (iv1.base) == CONST_INT)
2285 up = INTVAL (iv1.base);
2286 else
2287 up = INTVAL (mode_mmax) - inc;
2288 down = INTVAL (GET_CODE (iv0.base) == CONST_INT
2289 ? iv0.base
2290 : mode_mmin);
2291 desc->niter_max = (up - down) / inc + 1;
2293 if (iv0.step == const0_rtx)
2295 iv0.base = simplify_gen_binary (PLUS, comp_mode, iv0.base, delta);
2296 iv0.base = simplify_gen_binary (MINUS, comp_mode, iv0.base, step);
2298 else
2300 iv1.base = simplify_gen_binary (MINUS, comp_mode, iv1.base, delta);
2301 iv1.base = simplify_gen_binary (PLUS, comp_mode, iv1.base, step);
2304 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2305 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2306 assumption = simplify_gen_relational (reverse_condition (cond),
2307 SImode, mode, tmp0, tmp1);
2308 if (assumption == const_true_rtx)
2309 goto zero_iter;
2310 else if (assumption != const0_rtx)
2311 desc->noloop_assumptions =
2312 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2313 cond = NE;
2317 /* Count the number of iterations. */
2318 if (cond == NE)
2320 /* Everything we do here is just arithmetics modulo size of mode. This
2321 makes us able to do more involved computations of number of iterations
2322 than in other cases. First transform the condition into shape
2323 s * i <> c, with s positive. */
2324 iv1.base = simplify_gen_binary (MINUS, comp_mode, iv1.base, iv0.base);
2325 iv0.base = const0_rtx;
2326 iv0.step = simplify_gen_binary (MINUS, comp_mode, iv0.step, iv1.step);
2327 iv1.step = const0_rtx;
2328 if (INTVAL (iv0.step) < 0)
2330 iv0.step = simplify_gen_unary (NEG, comp_mode, iv0.step, mode);
2331 iv1.base = simplify_gen_unary (NEG, comp_mode, iv1.base, mode);
2333 iv0.step = lowpart_subreg (mode, iv0.step, comp_mode);
2335 /* Let nsd (s, size of mode) = d. If d does not divide c, the loop
2336 is infinite. Otherwise, the number of iterations is
2337 (inverse(s/d) * (c/d)) mod (size of mode/d). */
2338 s = INTVAL (iv0.step); d = 1;
2339 while (s % 2 != 1)
2341 s /= 2;
2342 d *= 2;
2343 size--;
2345 bound = GEN_INT (((unsigned HOST_WIDEST_INT) 1 << (size - 1 ) << 1) - 1);
2347 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2348 tmp = simplify_gen_binary (UMOD, mode, tmp1, GEN_INT (d));
2349 assumption = simplify_gen_relational (NE, SImode, mode, tmp, const0_rtx);
2350 desc->infinite = alloc_EXPR_LIST (0, assumption, desc->infinite);
2352 tmp = simplify_gen_binary (UDIV, mode, tmp1, GEN_INT (d));
2353 inv = inverse (s, size);
2354 tmp = simplify_gen_binary (MULT, mode, tmp, gen_int_mode (inv, mode));
2355 desc->niter_expr = simplify_gen_binary (AND, mode, tmp, bound);
2357 else
2359 if (iv1.step == const0_rtx)
2360 /* Condition in shape a + s * i <= b
2361 We must know that b + s does not overflow and a <= b + s and then we
2362 can compute number of iterations as (b + s - a) / s. (It might
2363 seem that we in fact could be more clever about testing the b + s
2364 overflow condition using some information about b - a mod s,
2365 but it was already taken into account during LE -> NE transform). */
2367 step = iv0.step;
2368 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2369 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2371 bound = simplify_gen_binary (MINUS, mode, mode_mmax,
2372 lowpart_subreg (mode, step,
2373 comp_mode));
2374 if (step_is_pow2)
2376 rtx t0, t1;
2378 /* If s is power of 2, we know that the loop is infinite if
2379 a % s <= b % s and b + s overflows. */
2380 assumption = simplify_gen_relational (reverse_condition (cond),
2381 SImode, mode,
2382 tmp1, bound);
2384 t0 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp0), step);
2385 t1 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp1), step);
2386 tmp = simplify_gen_relational (cond, SImode, mode, t0, t1);
2387 assumption = simplify_gen_binary (AND, SImode, assumption, tmp);
2388 desc->infinite =
2389 alloc_EXPR_LIST (0, assumption, desc->infinite);
2391 else
2393 assumption = simplify_gen_relational (cond, SImode, mode,
2394 tmp1, bound);
2395 desc->assumptions =
2396 alloc_EXPR_LIST (0, assumption, desc->assumptions);
2399 tmp = simplify_gen_binary (PLUS, comp_mode, iv1.base, iv0.step);
2400 tmp = lowpart_subreg (mode, tmp, comp_mode);
2401 assumption = simplify_gen_relational (reverse_condition (cond),
2402 SImode, mode, tmp0, tmp);
2404 delta = simplify_gen_binary (PLUS, mode, tmp1, step);
2405 delta = simplify_gen_binary (MINUS, mode, delta, tmp0);
2407 else
2409 /* Condition in shape a <= b - s * i
2410 We must know that a - s does not overflow and a - s <= b and then
2411 we can again compute number of iterations as (b - (a - s)) / s. */
2412 step = simplify_gen_unary (NEG, mode, iv1.step, mode);
2413 tmp0 = lowpart_subreg (mode, iv0.base, comp_mode);
2414 tmp1 = lowpart_subreg (mode, iv1.base, comp_mode);
2416 bound = simplify_gen_binary (MINUS, mode, mode_mmin,
2417 lowpart_subreg (mode, step, comp_mode));
2418 if (step_is_pow2)
2420 rtx t0, t1;
2422 /* If s is power of 2, we know that the loop is infinite if
2423 a % s <= b % s and a - s overflows. */
2424 assumption = simplify_gen_relational (reverse_condition (cond),
2425 SImode, mode,
2426 bound, tmp0);
2428 t0 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp0), step);
2429 t1 = simplify_gen_binary (UMOD, mode, copy_rtx (tmp1), step);
2430 tmp = simplify_gen_relational (cond, SImode, mode, t0, t1);
2431 assumption = simplify_gen_binary (AND, SImode, assumption, tmp);
2432 desc->infinite =
2433 alloc_EXPR_LIST (0, assumption, desc->infinite);
2435 else
2437 assumption = simplify_gen_relational (cond, SImode, mode,
2438 bound, tmp0);
2439 desc->assumptions =
2440 alloc_EXPR_LIST (0, assumption, desc->assumptions);
2443 tmp = simplify_gen_binary (PLUS, comp_mode, iv0.base, iv1.step);
2444 tmp = lowpart_subreg (mode, tmp, comp_mode);
2445 assumption = simplify_gen_relational (reverse_condition (cond),
2446 SImode, mode,
2447 tmp, tmp1);
2448 delta = simplify_gen_binary (MINUS, mode, tmp0, step);
2449 delta = simplify_gen_binary (MINUS, mode, tmp1, delta);
2451 if (assumption == const_true_rtx)
2452 goto zero_iter;
2453 else if (assumption != const0_rtx)
2454 desc->noloop_assumptions =
2455 alloc_EXPR_LIST (0, assumption, desc->noloop_assumptions);
2456 delta = simplify_gen_binary (UDIV, mode, delta, step);
2457 desc->niter_expr = delta;
2460 old_niter = desc->niter_expr;
2462 simplify_using_initial_values (loop, AND, &desc->assumptions);
2463 if (desc->assumptions
2464 && XEXP (desc->assumptions, 0) == const0_rtx)
2465 goto fail;
2466 simplify_using_initial_values (loop, IOR, &desc->noloop_assumptions);
2467 simplify_using_initial_values (loop, IOR, &desc->infinite);
2468 simplify_using_initial_values (loop, UNKNOWN, &desc->niter_expr);
2470 /* Rerun the simplification. Consider code (created by copying loop headers)
2472 i = 0;
2474 if (0 < n)
2478 i++;
2479 } while (i < n);
2482 The first pass determines that i = 0, the second pass uses it to eliminate
2483 noloop assumption. */
2485 simplify_using_initial_values (loop, AND, &desc->assumptions);
2486 if (desc->assumptions
2487 && XEXP (desc->assumptions, 0) == const0_rtx)
2488 goto fail;
2489 simplify_using_initial_values (loop, IOR, &desc->noloop_assumptions);
2490 simplify_using_initial_values (loop, IOR, &desc->infinite);
2491 simplify_using_initial_values (loop, UNKNOWN, &desc->niter_expr);
2493 if (desc->noloop_assumptions
2494 && XEXP (desc->noloop_assumptions, 0) == const_true_rtx)
2495 goto zero_iter;
2497 if (GET_CODE (desc->niter_expr) == CONST_INT)
2499 unsigned HOST_WIDEST_INT val = INTVAL (desc->niter_expr);
2501 desc->const_iter = true;
2502 desc->niter_max = desc->niter = val & GET_MODE_MASK (desc->mode);
2504 else
2506 if (!desc->niter_max)
2507 desc->niter_max = determine_max_iter (desc);
2509 /* simplify_using_initial_values does a copy propagation on the registers
2510 in the expression for the number of iterations. This prolongs life
2511 ranges of registers and increases register pressure, and usually
2512 brings no gain (and if it happens to do, the cse pass will take care
2513 of it anyway). So prevent this behavior, unless it enabled us to
2514 derive that the number of iterations is a constant. */
2515 desc->niter_expr = old_niter;
2518 return;
2520 fail:
2521 desc->simple_p = false;
2522 return;
2524 zero_iter:
2525 desc->const_iter = true;
2526 desc->niter = 0;
2527 desc->niter_max = 0;
2528 desc->niter_expr = const0_rtx;
2529 return;
2532 /* Checks whether E is a simple exit from LOOP and stores its description
2533 into DESC. */
2535 static void
2536 check_simple_exit (struct loop *loop, edge e, struct niter_desc *desc)
2538 basic_block exit_bb;
2539 rtx condition, at;
2540 edge ein;
2542 exit_bb = e->src;
2543 desc->simple_p = false;
2545 /* It must belong directly to the loop. */
2546 if (exit_bb->loop_father != loop)
2547 return;
2549 /* It must be tested (at least) once during any iteration. */
2550 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, exit_bb))
2551 return;
2553 /* It must end in a simple conditional jump. */
2554 if (!any_condjump_p (BB_END (exit_bb)))
2555 return;
2557 ein = EDGE_SUCC (exit_bb, 0);
2558 if (ein == e)
2559 ein = EDGE_SUCC (exit_bb, 1);
2561 desc->out_edge = e;
2562 desc->in_edge = ein;
2564 /* Test whether the condition is suitable. */
2565 if (!(condition = get_condition (BB_END (ein->src), &at, false, false)))
2566 return;
2568 if (ein->flags & EDGE_FALLTHRU)
2570 condition = reversed_condition (condition);
2571 if (!condition)
2572 return;
2575 /* Check that we are able to determine number of iterations and fill
2576 in information about it. */
2577 iv_number_of_iterations (loop, at, condition, desc);
2580 /* Finds a simple exit of LOOP and stores its description into DESC. */
2582 void
2583 find_simple_exit (struct loop *loop, struct niter_desc *desc)
2585 unsigned i;
2586 basic_block *body;
2587 edge e;
2588 struct niter_desc act;
2589 bool any = false;
2590 edge_iterator ei;
2592 desc->simple_p = false;
2593 body = get_loop_body (loop);
2595 for (i = 0; i < loop->num_nodes; i++)
2597 FOR_EACH_EDGE (e, ei, body[i]->succs)
2599 if (flow_bb_inside_loop_p (loop, e->dest))
2600 continue;
2602 check_simple_exit (loop, e, &act);
2603 if (!act.simple_p)
2604 continue;
2606 /* Prefer constant iterations; the less the better. */
2607 if (!any)
2608 any = true;
2609 else if (!act.const_iter
2610 || (desc->const_iter && act.niter >= desc->niter))
2611 continue;
2612 *desc = act;
2616 if (dump_file)
2618 if (desc->simple_p)
2620 fprintf (dump_file, "Loop %d is simple:\n", loop->num);
2621 fprintf (dump_file, " simple exit %d -> %d\n",
2622 desc->out_edge->src->index,
2623 desc->out_edge->dest->index);
2624 if (desc->assumptions)
2626 fprintf (dump_file, " assumptions: ");
2627 print_rtl (dump_file, desc->assumptions);
2628 fprintf (dump_file, "\n");
2630 if (desc->noloop_assumptions)
2632 fprintf (dump_file, " does not roll if: ");
2633 print_rtl (dump_file, desc->noloop_assumptions);
2634 fprintf (dump_file, "\n");
2636 if (desc->infinite)
2638 fprintf (dump_file, " infinite if: ");
2639 print_rtl (dump_file, desc->infinite);
2640 fprintf (dump_file, "\n");
2643 fprintf (dump_file, " number of iterations: ");
2644 print_rtl (dump_file, desc->niter_expr);
2645 fprintf (dump_file, "\n");
2647 fprintf (dump_file, " upper bound: ");
2648 fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, desc->niter_max);
2649 fprintf (dump_file, "\n");
2651 else
2652 fprintf (dump_file, "Loop %d is not simple.\n", loop->num);
2655 free (body);
2658 /* Creates a simple loop description of LOOP if it was not computed
2659 already. */
2661 struct niter_desc *
2662 get_simple_loop_desc (struct loop *loop)
2664 struct niter_desc *desc = simple_loop_desc (loop);
2666 if (desc)
2667 return desc;
2669 desc = xmalloc (sizeof (struct niter_desc));
2670 iv_analysis_loop_init (loop);
2671 find_simple_exit (loop, desc);
2672 loop->aux = desc;
2674 return desc;
2677 /* Releases simple loop description for LOOP. */
2679 void
2680 free_simple_loop_desc (struct loop *loop)
2682 struct niter_desc *desc = simple_loop_desc (loop);
2684 if (!desc)
2685 return;
2687 free (desc);
2688 loop->aux = NULL;