2 * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
38 #include "iroptimize.h"
40 #include "dbginfo_t.h"
41 #include "iropt_dbg.h"
50 #include "firm_types.h"
51 #include "bitfiddle.h"
54 #include "firmstat_t.h"
58 static bool is_Or_Eor_Add(const ir_node *node)
60 if (is_Or(node) || is_Eor(node) || is_Add(node)) {
61 ir_node *left = get_binop_left(node);
62 ir_node *right = get_binop_right(node);
63 vrp_attr *vrp_left = vrp_get_info(left);
64 vrp_attr *vrp_right = vrp_get_info(right);
65 if (vrp_left != NULL && vrp_right != NULL) {
67 = tarval_and(vrp_left->bits_not_set, vrp_right->bits_not_set);
68 return tarval_is_null(vrp_val);
75 * Returns the tarval of a Const node or tarval_bad for all other nodes.
77 static ir_tarval *default_value_of(const ir_node *n)
80 return get_Const_tarval(n); /* might return tarval_bad */
85 value_of_func value_of_ptr = default_value_of;
87 void set_value_of_func(value_of_func func)
92 value_of_ptr = default_value_of;
96 * Return the value of a Constant.
98 static ir_tarval *computed_value_Const(const ir_node *n)
100 return get_Const_tarval(n);
104 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
106 static ir_tarval *computed_value_SymConst(const ir_node *n)
111 switch (get_SymConst_kind(n)) {
112 case symconst_type_size:
113 type = get_SymConst_type(n);
114 if (get_type_state(type) == layout_fixed)
115 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
117 case symconst_type_align:
118 type = get_SymConst_type(n);
119 if (get_type_state(type) == layout_fixed)
120 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
122 case symconst_ofs_ent:
123 ent = get_SymConst_entity(n);
124 type = get_entity_owner(ent);
125 if (get_type_state(type) == layout_fixed)
126 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
135 * Return the value of an Add.
137 static ir_tarval *computed_value_Add(const ir_node *n)
139 ir_node *a = get_Add_left(n);
140 ir_node *b = get_Add_right(n);
142 ir_tarval *ta = value_of(a);
143 ir_tarval *tb = value_of(b);
145 if ((ta != tarval_bad) && (tb != tarval_bad))
146 return tarval_add(ta, tb);
149 if ((is_Not(a) && get_Not_op(a) == b)
150 || (is_Not(b) && get_Not_op(b) == a)) {
151 return get_mode_all_one(get_irn_mode(n));
158 * Return the value of a Sub.
159 * Special case: a - a
161 static ir_tarval *computed_value_Sub(const ir_node *n)
163 ir_mode *mode = get_irn_mode(n);
164 ir_node *a = get_Sub_left(n);
165 ir_node *b = get_Sub_right(n);
170 if (! mode_is_float(mode)) {
173 return get_mode_null(mode);
179 if ((ta != tarval_bad) && (tb != tarval_bad))
180 return tarval_sub(ta, tb, mode);
186 * Return the value of a Carry.
187 * Special : a op 0, 0 op b
189 static ir_tarval *computed_value_Carry(const ir_node *n)
191 ir_node *a = get_binop_left(n);
192 ir_node *b = get_binop_right(n);
193 ir_mode *m = get_irn_mode(n);
194 ir_tarval *ta = value_of(a);
195 ir_tarval *tb = value_of(b);
197 if ((ta != tarval_bad) && (tb != tarval_bad)) {
199 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
201 if (tarval_is_null(ta) || tarval_is_null(tb))
202 return get_mode_null(m);
208 * Return the value of a Borrow.
211 static ir_tarval *computed_value_Borrow(const ir_node *n)
213 ir_node *a = get_binop_left(n);
214 ir_node *b = get_binop_right(n);
215 ir_mode *m = get_irn_mode(n);
216 ir_tarval *ta = value_of(a);
217 ir_tarval *tb = value_of(b);
219 if ((ta != tarval_bad) && (tb != tarval_bad)) {
220 return tarval_cmp(ta, tb) == ir_relation_less ? get_mode_one(m) : get_mode_null(m);
221 } else if (tarval_is_null(ta)) {
222 return get_mode_null(m);
228 * Return the value of an unary Minus.
230 static ir_tarval *computed_value_Minus(const ir_node *n)
232 ir_node *a = get_Minus_op(n);
233 ir_tarval *ta = value_of(a);
235 if (ta != tarval_bad)
236 return tarval_neg(ta);
242 * Return the value of a Mul.
244 static ir_tarval *computed_value_Mul(const ir_node *n)
246 ir_node *a = get_Mul_left(n);
247 ir_node *b = get_Mul_right(n);
248 ir_tarval *ta = value_of(a);
249 ir_tarval *tb = value_of(b);
252 mode = get_irn_mode(n);
253 if (mode != get_irn_mode(a)) {
254 /* n * n = 2n bit multiplication */
255 ta = tarval_convert_to(ta, mode);
256 tb = tarval_convert_to(tb, mode);
259 if (ta != tarval_bad && tb != tarval_bad) {
260 return tarval_mul(ta, tb);
262 /* a * 0 != 0 if a == NaN or a == Inf */
263 if (!mode_is_float(mode)) {
264 /* a*0 = 0 or 0*b = 0 */
265 if (ta == get_mode_null(mode))
267 if (tb == get_mode_null(mode))
275 * Return the value of an And.
276 * Special case: a & 0, 0 & b
278 static ir_tarval *computed_value_And(const ir_node *n)
280 ir_node *a = get_And_left(n);
281 ir_node *b = get_And_right(n);
282 ir_tarval *ta = value_of(a);
283 ir_tarval *tb = value_of(b);
285 if ((ta != tarval_bad) && (tb != tarval_bad)) {
286 return tarval_and (ta, tb);
289 if (tarval_is_null(ta)) return ta;
290 if (tarval_is_null(tb)) return tb;
293 if ((is_Not(a) && get_Not_op(a) == b)
294 || (is_Not(b) && get_Not_op(b) == a)) {
295 return get_mode_null(get_irn_mode(n));
302 * Return the value of an Or.
303 * Special case: a | 1...1, 1...1 | b
305 static ir_tarval *computed_value_Or(const ir_node *n)
307 ir_node *a = get_Or_left(n);
308 ir_node *b = get_Or_right(n);
309 ir_tarval *ta = value_of(a);
310 ir_tarval *tb = value_of(b);
312 if ((ta != tarval_bad) && (tb != tarval_bad)) {
313 return tarval_or (ta, tb);
316 if (tarval_is_all_one(ta)) return ta;
317 if (tarval_is_all_one(tb)) return tb;
320 if ((is_Not(a) && get_Not_op(a) == b)
321 || (is_Not(b) && get_Not_op(b) == a)) {
322 return get_mode_all_one(get_irn_mode(n));
328 * Return the value of an Eor.
330 static ir_tarval *computed_value_Eor(const ir_node *n)
332 ir_node *a = get_Eor_left(n);
333 ir_node *b = get_Eor_right(n);
338 return get_mode_null(get_irn_mode(n));
340 if ((is_Not(a) && get_Not_op(a) == b)
341 || (is_Not(b) && get_Not_op(b) == a)) {
342 return get_mode_all_one(get_irn_mode(n));
348 if ((ta != tarval_bad) && (tb != tarval_bad)) {
349 return tarval_eor(ta, tb);
355 * Return the value of a Not.
357 static ir_tarval *computed_value_Not(const ir_node *n)
359 ir_node *a = get_Not_op(n);
360 ir_tarval *ta = value_of(a);
362 if (ta != tarval_bad)
363 return tarval_not(ta);
369 * Tests whether a shift shifts more bits than available in the mode
371 static bool is_oversize_shift(const ir_node *n)
373 ir_node *count = get_binop_right(n);
374 ir_mode *mode = get_irn_mode(n);
375 ir_tarval *tv = value_of(count);
378 if (tv == tarval_bad)
380 if (!tarval_is_long(tv))
382 shiftval = get_tarval_long(tv);
383 modulo_shift = get_mode_modulo_shift(mode);
384 if (shiftval < 0 || (modulo_shift > 0 && shiftval >= modulo_shift))
387 return shiftval >= (long)get_mode_size_bits(mode);
391 * Return the value of a Shl.
393 static ir_tarval *computed_value_Shl(const ir_node *n)
395 ir_node *a = get_Shl_left(n);
396 ir_node *b = get_Shl_right(n);
398 ir_tarval *ta = value_of(a);
399 ir_tarval *tb = value_of(b);
401 if ((ta != tarval_bad) && (tb != tarval_bad)) {
402 return tarval_shl(ta, tb);
405 if (is_oversize_shift(n))
406 return get_mode_null(get_irn_mode(n));
412 * Return the value of a Shr.
414 static ir_tarval *computed_value_Shr(const ir_node *n)
416 ir_node *a = get_Shr_left(n);
417 ir_node *b = get_Shr_right(n);
419 ir_tarval *ta = value_of(a);
420 ir_tarval *tb = value_of(b);
422 if ((ta != tarval_bad) && (tb != tarval_bad)) {
423 return tarval_shr(ta, tb);
425 if (is_oversize_shift(n))
426 return get_mode_null(get_irn_mode(n));
432 * Return the value of a Shrs.
434 static ir_tarval *computed_value_Shrs(const ir_node *n)
436 ir_node *a = get_Shrs_left(n);
437 ir_node *b = get_Shrs_right(n);
439 ir_tarval *ta = value_of(a);
440 ir_tarval *tb = value_of(b);
442 if ((ta != tarval_bad) && (tb != tarval_bad)) {
443 return tarval_shrs(ta, tb);
449 * Return the value of a Rotl.
451 static ir_tarval *computed_value_Rotl(const ir_node *n)
453 ir_node *a = get_Rotl_left(n);
454 ir_node *b = get_Rotl_right(n);
456 ir_tarval *ta = value_of(a);
457 ir_tarval *tb = value_of(b);
459 if ((ta != tarval_bad) && (tb != tarval_bad)) {
460 return tarval_rotl(ta, tb);
465 bool ir_zero_when_converted(const ir_node *node, ir_mode *dest_mode)
467 ir_mode *mode = get_irn_mode(node);
468 if (get_mode_arithmetic(mode) != irma_twos_complement
469 || get_mode_arithmetic(dest_mode) != irma_twos_complement)
473 ir_node *count = get_Shl_right(node);
474 if (is_Const(count)) {
475 ir_tarval *tv = get_Const_tarval(count);
476 if (tarval_is_long(tv)) {
477 long shiftval = get_tarval_long(tv);
478 long destbits = get_mode_size_bits(dest_mode);
479 if (shiftval >= destbits
480 && shiftval < (long)get_mode_modulo_shift(mode))
486 ir_node *right = get_And_right(node);
487 if (is_Const(right)) {
488 ir_tarval *tv = get_Const_tarval(right);
489 ir_tarval *conved = tarval_convert_to(tv, dest_mode);
490 return tarval_is_null(conved);
497 * Return the value of a Conv.
499 static ir_tarval *computed_value_Conv(const ir_node *n)
501 ir_node *a = get_Conv_op(n);
502 ir_tarval *ta = value_of(a);
503 ir_mode *mode = get_irn_mode(n);
505 if (ta != tarval_bad)
506 return tarval_convert_to(ta, get_irn_mode(n));
508 if (ir_zero_when_converted(a, mode))
509 return get_mode_null(mode);
515 * Calculate the value of a Mux: can be evaluated, if the
516 * sel and the right input are known.
518 static ir_tarval *computed_value_Mux(const ir_node *n)
520 ir_node *sel = get_Mux_sel(n);
521 ir_tarval *ts = value_of(sel);
523 if (ts == get_tarval_b_true()) {
524 ir_node *v = get_Mux_true(n);
527 else if (ts == get_tarval_b_false()) {
528 ir_node *v = get_Mux_false(n);
535 * Calculate the value of a Confirm: can be evaluated,
536 * if it has the form Confirm(x, '=', Const).
538 static ir_tarval *computed_value_Confirm(const ir_node *n)
540 if (get_Confirm_relation(n) == ir_relation_equal) {
541 ir_tarval *tv = value_of(get_Confirm_bound(n));
542 if (tv != tarval_bad)
545 return value_of(get_Confirm_value(n));
549 * gives a (conservative) estimation of possible relation when comparing
552 ir_relation ir_get_possible_cmp_relations(const ir_node *left,
553 const ir_node *right)
555 ir_relation possible = ir_relation_true;
556 ir_tarval *tv_l = value_of(left);
557 ir_tarval *tv_r = value_of(right);
558 ir_mode *mode = get_irn_mode(left);
559 ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode);
560 ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode);
562 /* both values known - evaluate them */
563 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
564 possible = tarval_cmp(tv_l, tv_r);
565 /* we can return now, won't get any better */
568 /* a == a is never less or greater (but might be equal or unordered) */
570 possible &= ~ir_relation_less_greater;
571 /* unordered results only happen for float compares */
572 if (!mode_is_float(mode))
573 possible &= ~ir_relation_unordered;
574 /* values can never be less than the least representable number or
575 * greater than the greatest representable number */
577 possible &= ~ir_relation_greater;
579 possible &= ~ir_relation_less;
581 possible &= ~ir_relation_greater;
583 possible &= ~ir_relation_less;
584 /* maybe vrp can tell us more */
585 possible &= vrp_cmp(left, right);
586 /* Alloc nodes never return null (but throw an exception) */
587 if (is_Alloc(left) && tarval_is_null(tv_r))
588 possible &= ~ir_relation_equal;
589 /* stuff known through confirm nodes */
590 if (is_Confirm(left) && get_Confirm_bound(left) == right) {
591 possible &= get_Confirm_relation(left);
593 if (is_Confirm(right) && get_Confirm_bound(right) == left) {
594 ir_relation relation = get_Confirm_relation(right);
595 relation = get_inversed_relation(relation);
596 possible &= relation;
602 static ir_tarval *compute_cmp(const ir_node *cmp)
604 ir_node *left = get_Cmp_left(cmp);
605 ir_node *right = get_Cmp_right(cmp);
606 ir_relation possible = ir_get_possible_cmp_relations(left, right);
607 ir_relation relation = get_Cmp_relation(cmp);
609 /* if none of the requested relations is possible, return false */
610 if ((possible & relation) == ir_relation_false)
611 return tarval_b_false;
612 /* if possible relations are a subset of the requested ones return true */
613 if ((possible & ~relation) == ir_relation_false)
614 return tarval_b_true;
616 return computed_value_Cmp_Confirm(cmp, left, right, relation);
620 * some people want to call compute_cmp directly, in this case we have to
621 * test the constant folding flag again
623 static ir_tarval *compute_cmp_ext(const ir_node *cmp)
625 if (!get_opt_constant_folding())
627 return compute_cmp(cmp);
631 * Return the value of a Cmp.
633 * The basic idea here is to determine which relations are possible and which
634 * one are definitely impossible.
636 static ir_tarval *computed_value_Cmp(const ir_node *cmp)
638 /* we can't construct Constb after lowering mode_b nodes */
639 if (irg_is_constrained(get_irn_irg(cmp), IR_GRAPH_CONSTRAINT_MODEB_LOWERED))
642 return compute_cmp(cmp);
646 * Calculate the value of an integer Div.
647 * Special case: 0 / b
649 static ir_tarval *do_computed_value_Div(const ir_node *div)
651 const ir_node *a = get_Div_left(div);
652 const ir_node *b = get_Div_right(div);
653 const ir_mode *mode = get_Div_resmode(div);
654 ir_tarval *ta = value_of(a);
656 const ir_node *dummy;
658 /* cannot optimize 0 / b = 0 because of NaN */
659 if (!mode_is_float(mode)) {
660 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
661 return ta; /* 0 / b == 0 if b != 0 */
664 if (ta != tarval_bad && tb != tarval_bad)
665 return tarval_div(ta, tb);
670 * Calculate the value of an integer Mod of two nodes.
671 * Special case: a % 1
673 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
675 ir_tarval *ta = value_of(a);
676 ir_tarval *tb = value_of(b);
678 /* Compute a % 1 or c1 % c2 */
679 if (tarval_is_one(tb))
680 return get_mode_null(get_irn_mode(a));
681 if (ta != tarval_bad && tb != tarval_bad)
682 return tarval_mod(ta, tb);
687 * Return the value of a Proj(Div).
689 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
691 long proj_nr = get_Proj_proj(n);
692 if (proj_nr != pn_Div_res)
695 return do_computed_value_Div(get_Proj_pred(n));
699 * Return the value of a Proj(Mod).
701 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
703 long proj_nr = get_Proj_proj(n);
705 if (proj_nr == pn_Mod_res) {
706 const ir_node *mod = get_Proj_pred(n);
707 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
713 * Return the value of a Proj.
715 static ir_tarval *computed_value_Proj(const ir_node *proj)
717 ir_node *n = get_Proj_pred(proj);
719 if (n->op->ops.computed_value_Proj != NULL)
720 return n->op->ops.computed_value_Proj(proj);
725 * If the parameter n can be computed, return its value, else tarval_bad.
726 * Performs constant folding.
728 * @param n The node this should be evaluated
730 ir_tarval *computed_value(const ir_node *n)
732 vrp_attr *vrp = vrp_get_info(n);
733 if (vrp != NULL && vrp->bits_set == vrp->bits_not_set)
734 return vrp->bits_set;
736 if (n->op->ops.computed_value)
737 return n->op->ops.computed_value(n);
742 * Optimize operations that are commutative and have neutral 0,
743 * so a op 0 = 0 op a = a.
745 static ir_node *equivalent_node_neutral_zero(ir_node *n)
749 ir_node *a = get_binop_left(n);
750 ir_node *b = get_binop_right(n);
755 /* After running compute_node there is only one constant predecessor.
756 Find this predecessors value and remember the other node: */
757 if ((tv = value_of(a)) != tarval_bad) {
759 } else if ((tv = value_of(b)) != tarval_bad) {
764 /* If this predecessors constant value is zero, the operation is
765 * unnecessary. Remove it.
767 * Beware: If n is a Add, the mode of on and n might be different
768 * which happens in this rare construction: NULL + 3.
769 * Then, a Conv would be needed which we cannot include here.
771 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
774 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
781 * Eor is commutative and has neutral 0.
783 static ir_node *equivalent_node_Eor(ir_node *n)
789 n = equivalent_node_neutral_zero(n);
790 if (n != oldn) return n;
793 b = get_Eor_right(n);
795 if (is_Eor(a) || is_Or_Eor_Add(a)) {
796 ir_node *aa = get_binop_left(a);
797 ir_node *ab = get_binop_right(a);
800 /* (a ^ b) ^ a -> b */
802 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
804 } else if (ab == b) {
805 /* (a ^ b) ^ b -> a */
807 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
811 if (is_Eor(b) || is_Or_Eor_Add(b)) {
812 ir_node *ba = get_binop_left(b);
813 ir_node *bb = get_binop_right(b);
816 /* a ^ (a ^ b) -> b */
818 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
820 } else if (bb == a) {
821 /* a ^ (b ^ a) -> b */
823 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
831 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
833 * The second one looks strange, but this construct
834 * is used heavily in the LCC sources :-).
836 * Beware: The Mode of an Add may be different than the mode of its
837 * predecessors, so we could not return a predecessors in all cases.
839 static ir_node *equivalent_node_Add(ir_node *n)
842 ir_node *left, *right;
843 ir_mode *mode = get_irn_mode(n);
845 n = equivalent_node_neutral_zero(n);
849 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
850 if (mode_is_float(mode)) {
851 ir_graph *irg = get_irn_irg(n);
852 if (get_irg_fp_model(irg) & fp_strict_algebraic)
856 left = get_Add_left(n);
857 right = get_Add_right(n);
860 if (get_Sub_right(left) == right) {
863 n = get_Sub_left(left);
864 if (mode == get_irn_mode(n)) {
865 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
871 if (get_Sub_right(right) == left) {
874 n = get_Sub_left(right);
875 if (mode == get_irn_mode(n)) {
876 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
885 * optimize operations that are not commutative but have neutral 0 on left,
888 static ir_node *equivalent_node_left_zero(ir_node *n)
892 ir_node *a = get_binop_left(n);
893 ir_node *b = get_binop_right(n);
894 ir_tarval *tb = value_of(b);
896 if (tarval_is_null(tb)) {
899 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
905 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
907 * The second one looks strange, but this construct
908 * is used heavily in the LCC sources :-).
910 * Beware: The Mode of a Sub may be different than the mode of its
911 * predecessors, so we could not return a predecessors in all cases.
913 static ir_node *equivalent_node_Sub(ir_node *n)
917 ir_mode *mode = get_irn_mode(n);
920 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
921 if (mode_is_float(mode)) {
922 ir_graph *irg = get_irn_irg(n);
923 if (get_irg_fp_model(irg) & fp_strict_algebraic)
927 b = get_Sub_right(n);
930 /* Beware: modes might be different */
931 if (tarval_is_null(tb)) {
932 ir_node *a = get_Sub_left(n);
933 if (mode == get_irn_mode(a)) {
936 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
944 * Optimize an "self-inverse unary op", i.e. op(op(n)) = n.
947 * -(-a) == a, but might overflow two times.
948 * We handle it anyway here but the better way would be a
949 * flag. This would be needed for Pascal for instance.
951 static ir_node *equivalent_node_involution(ir_node *n)
954 ir_node *pred = get_unop_op(n);
955 if (get_irn_op(pred) == get_irn_op(n)) {
956 n = get_unop_op(pred);
957 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_INVOLUTION);
963 * Optimize a * 1 = 1 * a = a.
965 static ir_node *equivalent_node_Mul(ir_node *n)
968 ir_node *a = get_Mul_left(n);
970 /* we can handle here only the n * n = n bit cases */
971 if (get_irn_mode(n) == get_irn_mode(a)) {
972 ir_node *b = get_Mul_right(n);
976 * Mul is commutative and has again an other neutral element.
977 * Constants are place right, so check this case first.
980 if (tarval_is_one(tv)) {
982 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
985 if (tarval_is_one(tv)) {
987 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
995 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
997 static ir_node *equivalent_node_Or(ir_node *n)
1001 ir_node *a = get_Or_left(n);
1002 ir_node *b = get_Or_right(n);
1006 n = a; /* idempotence */
1007 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1010 /* constants are normalized to right, check this side first */
1012 if (tarval_is_null(tv)) {
1014 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1018 if (tarval_is_null(tv)) {
1020 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1028 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1030 static ir_node *equivalent_node_And(ir_node *n)
1034 ir_node *a = get_And_left(n);
1035 ir_node *b = get_And_right(n);
1039 n = a; /* idempotence */
1040 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1043 /* constants are normalized to right, check this side first */
1045 if (tarval_is_all_one(tv)) {
1047 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1050 if (tv != get_tarval_bad()) {
1051 ir_mode *mode = get_irn_mode(n);
1052 if (!mode_is_signed(mode) && is_Conv(a)) {
1053 ir_node *convop = get_Conv_op(a);
1054 ir_mode *convopmode = get_irn_mode(convop);
1055 if (!mode_is_signed(convopmode)) {
1056 /* Check Conv(all_one) & Const = all_one */
1057 ir_tarval *one = get_mode_all_one(convopmode);
1058 ir_tarval *conv = tarval_convert_to(one, mode);
1059 ir_tarval *tand = tarval_and(conv, tv);
1061 if (tarval_is_all_one(tand)) {
1062 /* Conv(X) & Const = X */
1064 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1071 if (tarval_is_all_one(tv)) {
1073 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1077 if ((is_Or(a) || is_Or_Eor_Add(a))
1078 && (b == get_binop_left(a) || b == get_binop_right(a))) {
1080 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1084 if ((is_Or(b) || is_Or_Eor_Add(b))
1085 && (a == get_binop_left(b) || a == get_binop_right(b))) {
1087 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1094 * Try to remove useless Conv's:
1096 static ir_node *equivalent_node_Conv(ir_node *n)
1099 ir_node *a = get_Conv_op(n);
1101 ir_mode *n_mode = get_irn_mode(n);
1102 ir_mode *a_mode = get_irn_mode(a);
1104 if (n_mode == a_mode) { /* No Conv necessary */
1106 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1108 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1109 ir_node *b = get_Conv_op(a);
1110 ir_mode *b_mode = get_irn_mode(b);
1112 if (n_mode == b_mode && values_in_mode(b_mode, a_mode)) {
1114 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1122 * - fold Phi-nodes, iff they have only one predecessor except
1125 static ir_node *equivalent_node_Phi(ir_node *n)
1130 ir_node *first_val = NULL; /* to shutup gcc */
1132 if (!get_opt_optimize() &&
1133 !irg_is_constrained(get_irn_irg(n), IR_GRAPH_CONSTRAINT_CONSTRUCTION))
1136 n_preds = get_Phi_n_preds(n);
1138 /* Phi of dead Region without predecessors. */
1142 /* Find first non-self-referencing input */
1143 for (i = 0; i < n_preds; ++i) {
1144 first_val = get_Phi_pred(n, i);
1145 /* not self pointer */
1146 if (first_val != n) {
1147 /* then found first value. */
1152 /* search for rest of inputs, determine if any of these
1153 are non-self-referencing */
1154 while (++i < n_preds) {
1155 ir_node *scnd_val = get_Phi_pred(n, i);
1156 if (scnd_val != n && scnd_val != first_val) {
1161 if (i >= n_preds && !is_Dummy(first_val)) {
1162 /* Fold, if no multiple distinct non-self-referencing inputs */
1164 DBG_OPT_PHI(oldn, n);
1170 * Optimize Proj(Tuple).
1172 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1174 ir_node *oldn = proj;
1175 ir_node *tuple = get_Proj_pred(proj);
1177 /* Remove the Tuple/Proj combination. */
1178 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1179 DBG_OPT_TUPLE(oldn, tuple, proj);
1185 * Optimize a / 1 = a.
1187 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1189 ir_node *oldn = proj;
1190 ir_node *div = get_Proj_pred(proj);
1191 ir_node *b = get_Div_right(div);
1192 ir_tarval *tb = value_of(b);
1194 /* Div is not commutative. */
1195 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1196 switch (get_Proj_proj(proj)) {
1198 proj = get_Div_mem(div);
1199 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1203 proj = get_Div_left(div);
1204 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1208 /* we cannot replace the exception Proj's here, this is done in
1209 transform_node_Proj_Div() */
1217 * Optimize CopyB(mem, x, x) into a Nop.
1219 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1221 ir_node *oldn = proj;
1222 ir_node *copyb = get_Proj_pred(proj);
1223 ir_node *a = get_CopyB_dst(copyb);
1224 ir_node *b = get_CopyB_src(copyb);
1227 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1228 switch (get_Proj_proj(proj)) {
1230 proj = get_CopyB_mem(copyb);
1231 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1239 * Optimize Bounds(idx, idx, upper) into idx.
1241 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1243 ir_node *oldn = proj;
1244 ir_node *bound = get_Proj_pred(proj);
1245 ir_node *idx = get_Bound_index(bound);
1246 ir_node *pred = skip_Proj(idx);
1249 if (idx == get_Bound_lower(bound))
1251 else if (is_Bound(pred)) {
1253 * idx was Bounds checked previously, it is still valid if
1254 * lower <= pred_lower && pred_upper <= upper.
1256 ir_node *lower = get_Bound_lower(bound);
1257 ir_node *upper = get_Bound_upper(bound);
1258 if (get_Bound_lower(pred) == lower &&
1259 get_Bound_upper(pred) == upper) {
1261 * One could expect that we simply return the previous
1262 * Bound here. However, this would be wrong, as we could
1263 * add an exception Proj to a new location then.
1264 * So, we must turn in into a tuple.
1270 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1271 switch (get_Proj_proj(proj)) {
1273 DBG_OPT_EXC_REM(proj);
1274 proj = get_Bound_mem(bound);
1278 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1281 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1289 * Does all optimizations on nodes that must be done on its Projs
1290 * because of creating new nodes.
1292 static ir_node *equivalent_node_Proj(ir_node *proj)
1294 ir_node *n = get_Proj_pred(proj);
1295 if (n->op->ops.equivalent_node_Proj)
1296 return n->op->ops.equivalent_node_Proj(proj);
1303 static ir_node *equivalent_node_Id(ir_node *n)
1311 DBG_OPT_ID(oldn, n);
1318 static ir_node *equivalent_node_Mux(ir_node *n)
1320 ir_node *oldn = n, *sel = get_Mux_sel(n);
1322 ir_tarval *ts = value_of(sel);
1324 if (ts == tarval_bad && is_Cmp(sel)) {
1325 /* try again with a direct call to compute_cmp, as we don't care
1326 * about the MODEB_LOWERED flag here */
1327 ts = compute_cmp_ext(sel);
1330 /* Mux(true, f, t) == t */
1331 if (ts == tarval_b_true) {
1332 n = get_Mux_true(n);
1333 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1336 /* Mux(false, f, t) == f */
1337 if (ts == tarval_b_false) {
1338 n = get_Mux_false(n);
1339 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1342 n_t = get_Mux_true(n);
1343 n_f = get_Mux_false(n);
1345 /* Mux(v, x, T) == x */
1346 if (is_Unknown(n_f)) {
1348 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1351 /* Mux(v, T, x) == x */
1352 if (is_Unknown(n_t)) {
1354 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1358 /* Mux(v, x, x) == x */
1361 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1364 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1365 ir_relation relation = get_Cmp_relation(sel);
1366 ir_node *f = get_Mux_false(n);
1367 ir_node *t = get_Mux_true(n);
1370 * Note further that these optimization work even for floating point
1371 * with NaN's because -NaN == NaN.
1372 * However, if +0 and -0 is handled differently, we cannot use the first one.
1374 ir_node *const cmp_l = get_Cmp_left(sel);
1375 ir_node *const cmp_r = get_Cmp_right(sel);
1378 case ir_relation_equal:
1379 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1380 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1382 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1387 case ir_relation_less_greater:
1388 case ir_relation_unordered_less_greater:
1389 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1390 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1392 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1401 * Note: normalization puts the constant on the right side,
1402 * so we check only one case.
1404 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1405 /* Mux(t CMP 0, X, t) */
1406 if (is_Minus(f) && get_Minus_op(f) == t) {
1407 /* Mux(t CMP 0, -t, t) */
1408 if (relation == ir_relation_equal) {
1409 /* Mux(t == 0, -t, t) ==> -t */
1411 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1412 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1413 /* Mux(t != 0, -t, t) ==> t */
1415 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1425 * Remove Confirm nodes if setting is on.
1426 * Replace Confirms(x, '=', Constlike) by Constlike.
1428 static ir_node *equivalent_node_Confirm(ir_node *n)
1430 ir_node *pred = get_Confirm_value(n);
1431 ir_relation relation = get_Confirm_relation(n);
1433 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1435 * rare case: two identical Confirms one after another,
1436 * replace the second one with the first.
1439 pred = get_Confirm_value(n);
1445 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1446 * perform no actual computation, as, e.g., the Id nodes. It does not create
1447 * new nodes. It is therefore safe to free n if the node returned is not n.
1448 * If a node returns a Tuple we can not just skip it. If the size of the
1449 * in array fits, we transform n into a tuple (e.g., Div).
1451 ir_node *equivalent_node(ir_node *n)
1453 if (n->op->ops.equivalent_node)
1454 return n->op->ops.equivalent_node(n);
1459 * Returns non-zero if a node is a Phi node
1460 * with all predecessors constant.
1462 static int is_const_Phi(ir_node *n)
1466 if (! is_Phi(n) || get_irn_arity(n) == 0)
1468 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1469 if (! is_Const(get_irn_n(n, i)))
1475 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1476 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1479 * in reality eval_func should be tarval (*eval_func)() but incomplete
1480 * declarations are bad style and generate noisy warnings
1482 typedef void (*eval_func)(void);
1485 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1487 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1489 if (eval == (eval_func) tarval_sub) {
1490 tarval_sub_type func = (tarval_sub_type)eval;
1492 return func(a, b, mode);
1494 tarval_binop_type func = (tarval_binop_type)eval;
1501 * Apply an evaluator on a binop with a constant operators (and one Phi).
1503 * @param phi the Phi node
1504 * @param other the other operand
1505 * @param eval an evaluator function
1506 * @param mode the mode of the result, may be different from the mode of the Phi!
1507 * @param left if non-zero, other is the left operand, else the right
1509 * @return a new Phi node if the conversion was successful, NULL else
1511 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1513 int n = get_irn_arity(phi);
1514 ir_tarval **tvs = ALLOCAN(ir_tarval*, n);
1516 for (int i = 0; i < n; ++i) {
1517 ir_node *pred = get_irn_n(phi, i);
1518 ir_tarval *tv = get_Const_tarval(pred);
1519 tv = do_eval(eval, other, tv, mode);
1521 if (tv == tarval_bad) {
1522 /* folding failed, bad */
1528 for (int i = 0; i < n; ++i) {
1529 ir_node *pred = get_irn_n(phi, i);
1530 ir_tarval *tv = get_Const_tarval(pred);
1531 tv = do_eval(eval, tv, other, mode);
1533 if (tv == tarval_bad) {
1534 /* folding failed, bad */
1540 ir_graph *irg = get_irn_irg(phi);
1541 ir_node **res = ALLOCAN(ir_node*, n);
1542 for (int i = 0; i < n; ++i) {
1543 res[i] = new_r_Const(irg, tvs[i]);
1545 ir_node *block = get_nodes_block(phi);
1546 return new_r_Phi(block, n, res, mode);
1550 * Apply an evaluator on a binop with two constant Phi.
1552 * @param a the left Phi node
1553 * @param b the right Phi node
1554 * @param eval an evaluator function
1555 * @param mode the mode of the result, may be different from the mode of the Phi!
1557 * @return a new Phi node if the conversion was successful, NULL else
1559 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1561 if (get_nodes_block(a) != get_nodes_block(b))
1564 int n = get_irn_arity(a);
1565 ir_tarval **tvs = ALLOCAN(ir_tarval*, n);
1566 for (int i = 0; i < n; ++i) {
1567 ir_node *pred_a = get_irn_n(a, i);
1568 ir_tarval *tv_l = get_Const_tarval(pred_a);
1569 ir_node *pred_b = get_irn_n(b, i);
1570 ir_tarval *tv_r = get_Const_tarval(pred_b);
1571 ir_tarval *tv = do_eval(eval, tv_l, tv_r, mode);
1573 if (tv == tarval_bad) {
1574 /* folding failed, bad */
1579 ir_graph *irg = get_irn_irg(a);
1580 ir_node **res = ALLOCAN(ir_node*, n);
1581 for (int i = 0; i < n; ++i) {
1582 res[i] = new_r_Const(irg, tvs[i]);
1584 ir_node *block = get_nodes_block(a);
1585 return new_r_Phi(block, n, res, mode);
1589 * Apply an evaluator on a unop with a constant operator (a Phi).
1591 * @param phi the Phi node
1592 * @param eval an evaluator function
1594 * @return a new Phi node if the conversion was successful, NULL else
1596 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1598 int n = get_irn_arity(phi);
1599 ir_tarval **tvs = ALLOCAN(ir_tarval*, n);
1600 for (int i = 0; i < n; ++i) {
1601 ir_node *pred = get_irn_n(phi, i);
1602 ir_tarval *tv = get_Const_tarval(pred);
1605 if (tv == tarval_bad) {
1606 /* folding failed, bad */
1611 ir_graph *irg = get_irn_irg(phi);
1612 ir_node **res = ALLOCAN(ir_node*, n);
1613 for (int i = 0; i < n; ++i) {
1614 res[i] = new_r_Const(irg, tvs[i]);
1616 ir_node *block = get_nodes_block(phi);
1617 ir_mode *mode = get_irn_mode(phi);
1618 return new_r_Phi(block, n, res, mode);
1622 * Apply a conversion on a constant operator (a Phi).
1624 * @param phi the Phi node
1626 * @return a new Phi node if the conversion was successful, NULL else
1628 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1630 int n = get_irn_arity(phi);
1631 ir_tarval **tvs = ALLOCAN(ir_tarval*, n);
1632 for (int i = 0; i < n; ++i) {
1633 ir_node *pred = get_irn_n(phi, i);
1634 ir_tarval *tv = get_Const_tarval(pred);
1635 tv = tarval_convert_to(tv, mode);
1637 if (tv == tarval_bad) {
1638 /* folding failed, bad */
1643 ir_graph *irg = get_irn_irg(phi);
1644 ir_node **res = ALLOCAN(ir_node*, n);
1645 for (int i = 0; i < n; ++i) {
1646 res[i] = new_r_Const(irg, tvs[i]);
1648 ir_node *block = get_nodes_block(phi);
1649 return new_r_Phi(block, n, res, mode);
1653 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1654 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1655 * If possible, remove the Conv's.
1657 static ir_node *transform_node_AddSub(ir_node *n)
1659 ir_mode *mode = get_irn_mode(n);
1661 if (mode_is_reference(mode)) {
1662 ir_node *left = get_binop_left(n);
1663 ir_node *right = get_binop_right(n);
1664 unsigned ref_bits = get_mode_size_bits(mode);
1666 if (is_Conv(left)) {
1667 ir_mode *lmode = get_irn_mode(left);
1668 unsigned bits = get_mode_size_bits(lmode);
1670 if (ref_bits == bits &&
1671 mode_is_int(lmode) &&
1672 get_mode_arithmetic(lmode) == irma_twos_complement) {
1673 ir_node *pre = get_Conv_op(left);
1674 ir_mode *pre_mode = get_irn_mode(pre);
1676 if (mode_is_int(pre_mode) &&
1677 get_mode_size_bits(pre_mode) == bits &&
1678 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1679 /* ok, this conv just changes to sign, moreover the calculation
1680 * is done with same number of bits as our address mode, so
1681 * we can ignore the conv as address calculation can be viewed
1682 * as either signed or unsigned
1684 set_binop_left(n, pre);
1689 if (is_Conv(right)) {
1690 ir_mode *rmode = get_irn_mode(right);
1691 unsigned bits = get_mode_size_bits(rmode);
1693 if (ref_bits == bits &&
1694 mode_is_int(rmode) &&
1695 get_mode_arithmetic(rmode) == irma_twos_complement) {
1696 ir_node *pre = get_Conv_op(right);
1697 ir_mode *pre_mode = get_irn_mode(pre);
1699 if (mode_is_int(pre_mode) &&
1700 get_mode_size_bits(pre_mode) == bits &&
1701 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1702 /* ok, this conv just changes to sign, moreover the calculation
1703 * is done with same number of bits as our address mode, so
1704 * we can ignore the conv as address calculation can be viewed
1705 * as either signed or unsigned
1707 set_binop_right(n, pre);
1712 /* let address arithmetic use unsigned modes */
1713 if (is_Const(right)) {
1714 ir_mode *rmode = get_irn_mode(right);
1716 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1717 /* convert a AddP(P, *s) into AddP(P, *u) */
1718 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1720 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
1721 set_binop_right(n, pre);
1729 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1732 if (is_Const(b) && is_const_Phi(a)) { \
1733 /* check for Op(Phi, Const) */ \
1734 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1736 else if (is_Const(a) && is_const_Phi(b)) { \
1737 /* check for Op(Const, Phi) */ \
1738 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1740 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1741 /* check for Op(Phi, Phi) */ \
1742 c = apply_binop_on_2_phis(a, b, eval, mode); \
1745 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1750 #define HANDLE_UNOP_PHI(eval, a, c) \
1753 if (is_const_Phi(a)) { \
1754 /* check for Op(Phi) */ \
1755 c = apply_unop_on_phi(a, eval); \
1757 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1764 * Create a 0 constant of given mode.
1766 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
1768 ir_tarval *tv = get_mode_null(mode);
1769 ir_node *cnst = new_r_Const(irg, tv);
1774 static bool is_shiftop(const ir_node *n)
1776 return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
1779 /* the order of the values is important! */
1780 typedef enum const_class {
1786 static const_class classify_const(const ir_node* n)
1788 if (is_Const(n)) return const_const;
1789 if (is_irn_constlike(n)) return const_like;
1794 * Determines whether r is more constlike or has a larger index (in that order)
1797 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
1799 const const_class l_order = classify_const(l);
1800 const const_class r_order = classify_const(r);
1802 l_order > r_order ||
1803 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
1806 static bool is_cmp_unequal(const ir_node *node)
1808 ir_relation relation = get_Cmp_relation(node);
1809 ir_node *left = get_Cmp_left(node);
1810 ir_node *right = get_Cmp_right(node);
1811 ir_mode *mode = get_irn_mode(left);
1813 if (relation == ir_relation_less_greater)
1816 if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right))
1817 return relation == ir_relation_greater;
1822 * returns true for Cmp(x == 0) or Cmp(x != 0)
1824 static bool is_cmp_equality_zero(const ir_node *node)
1826 ir_relation relation;
1827 ir_node *right = get_Cmp_right(node);
1829 if (!is_Const(right) || !is_Const_null(right))
1831 relation = get_Cmp_relation(node);
1832 return relation == ir_relation_equal
1833 || relation == ir_relation_less_greater
1834 || (!mode_is_signed(get_irn_mode(right))
1835 && relation == ir_relation_greater);
1839 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
1840 * Such pattern may arise in bitfield stores.
1842 * value c4 value c4 & c2
1843 * AND c3 AND c1 | c3
1850 * AND c1 ===> OR if (c1 | c2) == 0x111..11
1853 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
1855 ir_node *irn_and, *c1;
1857 ir_node *and_l, *c3;
1858 ir_node *value, *c4;
1859 ir_node *new_and, *new_const, *block;
1860 ir_mode *mode = get_irn_mode(irn_or);
1862 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
1866 irn_and = get_binop_left(irn_or);
1867 c1 = get_binop_right(irn_or);
1868 if (!is_Const(c1) || !is_And(irn_and))
1871 or_l = get_binop_left(irn_and);
1872 c2 = get_binop_right(irn_and);
1876 tv1 = get_Const_tarval(c1);
1877 tv2 = get_Const_tarval(c2);
1879 tv = tarval_or(tv1, tv2);
1880 if (tarval_is_all_one(tv)) {
1881 /* the AND does NOT clear a bit with isn't set by the OR */
1882 set_binop_left(irn_or, or_l);
1883 set_binop_right(irn_or, c1);
1885 /* check for more */
1889 if (!is_Or(or_l) && !is_Or_Eor_Add(or_l))
1892 and_l = get_binop_left(or_l);
1893 c3 = get_binop_right(or_l);
1894 if (!is_Const(c3) || !is_And(and_l))
1897 value = get_binop_left(and_l);
1898 c4 = get_binop_right(and_l);
1902 /* ok, found the pattern, check for conditions */
1903 assert(mode == get_irn_mode(irn_and));
1904 assert(mode == get_irn_mode(or_l));
1905 assert(mode == get_irn_mode(and_l));
1907 tv3 = get_Const_tarval(c3);
1908 tv4 = get_Const_tarval(c4);
1910 tv = tarval_or(tv4, tv2);
1911 if (!tarval_is_all_one(tv)) {
1912 /* have at least one 0 at the same bit position */
1916 if (tv3 != tarval_andnot(tv3, tv4)) {
1917 /* bit in the or_mask is outside the and_mask */
1921 if (tv1 != tarval_andnot(tv1, tv2)) {
1922 /* bit in the or_mask is outside the and_mask */
1926 /* ok, all conditions met */
1927 block = get_nodes_block(irn_or);
1928 irg = get_irn_irg(block);
1930 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
1932 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
1934 set_binop_left(irn_or, new_and);
1935 set_binop_right(irn_or, new_const);
1937 /* check for more */
1942 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
1944 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
1946 ir_mode *mode = get_irn_mode(irn_or);
1947 ir_node *shl, *shr, *block;
1948 ir_node *irn, *x, *c1, *c2, *n;
1949 ir_tarval *tv1, *tv2;
1951 /* some backends can't handle rotl */
1952 if (!be_get_backend_param()->support_rotl)
1955 if (! mode_is_int(mode))
1958 shl = get_binop_left(irn_or);
1959 shr = get_binop_right(irn_or);
1968 } else if (!is_Shl(shl)) {
1970 } else if (!is_Shr(shr)) {
1973 x = get_Shl_left(shl);
1974 if (x != get_Shr_left(shr))
1977 c1 = get_Shl_right(shl);
1978 c2 = get_Shr_right(shr);
1979 if (is_Const(c1) && is_Const(c2)) {
1980 tv1 = get_Const_tarval(c1);
1981 if (! tarval_is_long(tv1))
1984 tv2 = get_Const_tarval(c2);
1985 if (! tarval_is_long(tv2))
1988 if (get_tarval_long(tv1) + get_tarval_long(tv2)
1989 != (int) get_mode_size_bits(mode))
1992 /* yet, condition met */
1993 block = get_nodes_block(irn_or);
1995 n = new_r_Rotl(block, x, c1, mode);
1997 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
2001 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
2002 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
2003 if (!ir_is_negated_value(c1, c2)) {
2007 /* yet, condition met */
2008 block = get_nodes_block(irn_or);
2009 n = new_r_Rotl(block, x, c1, mode);
2010 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
2015 * Prototype of a recursive transform function
2016 * for bitwise distributive transformations.
2018 typedef ir_node* (*recursive_transform)(ir_node *n);
2021 * makes use of distributive laws for and, or, eor
2022 * and(a OP c, b OP c) -> and(a, b) OP c
2023 * note, might return a different op than n
2025 static ir_node *transform_bitwise_distributive(ir_node *n,
2026 recursive_transform trans_func)
2029 ir_node *a = get_binop_left(n);
2030 ir_node *b = get_binop_right(n);
2031 ir_op *op = get_irn_op(a);
2032 ir_op *op_root = get_irn_op(n);
2034 if (op != get_irn_op(b))
2037 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2038 if (op == op_Conv) {
2039 ir_node *a_op = get_Conv_op(a);
2040 ir_node *b_op = get_Conv_op(b);
2041 ir_mode *a_mode = get_irn_mode(a_op);
2042 ir_mode *b_mode = get_irn_mode(b_op);
2043 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2044 ir_node *blk = get_nodes_block(n);
2047 set_binop_left(n, a_op);
2048 set_binop_right(n, b_op);
2049 set_irn_mode(n, a_mode);
2051 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2053 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2059 /* nothing to gain here */
2063 if (op == op_Shrs || op == op_Shr || op == op_Shl
2064 || op == op_And || op == op_Or || op == op_Eor) {
2065 ir_node *a_left = get_binop_left(a);
2066 ir_node *a_right = get_binop_right(a);
2067 ir_node *b_left = get_binop_left(b);
2068 ir_node *b_right = get_binop_right(b);
2070 ir_node *op1 = NULL;
2071 ir_node *op2 = NULL;
2073 if (is_op_commutative(op)) {
2074 if (a_left == b_left) {
2078 } else if (a_left == b_right) {
2082 } else if (a_right == b_left) {
2088 if (a_right == b_right) {
2095 /* (a sop c) & (b sop c) => (a & b) sop c */
2096 ir_node *blk = get_nodes_block(n);
2098 ir_node *new_n = exact_copy(n);
2099 set_binop_left(new_n, op1);
2100 set_binop_right(new_n, op2);
2101 new_n = trans_func(new_n);
2103 if (op_root == op_Eor && op == op_Or) {
2104 dbg_info *dbgi = get_irn_dbg_info(n);
2105 ir_mode *mode = get_irn_mode(c);
2107 c = new_rd_Not(dbgi, blk, c, mode);
2108 n = new_rd_And(dbgi, blk, new_n, c, mode);
2111 set_nodes_block(n, blk);
2112 set_binop_left(n, new_n);
2113 set_binop_right(n, c);
2117 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2126 * normalisation: (x >> c1) & c2 to (x & (c2<<c1)) >> c1
2128 * - and, or, xor instead of &
2129 * - Shl, Shr, Shrs, rotl instead of >>
2130 * (with a special case for Or/Xor + Shrs)
2132 * This normalisation is usually good for the backend since << C can often be
2133 * matched as address-mode.
2135 static ir_node *transform_node_bitop_shift(ir_node *n)
2137 ir_graph *irg = get_irn_irg(n);
2138 ir_node *left = get_binop_left(n);
2139 ir_node *right = get_binop_right(n);
2140 ir_mode *mode = get_irn_mode(n);
2141 ir_node *shift_left;
2142 ir_node *shift_right;
2144 dbg_info *dbg_bitop;
2145 dbg_info *dbg_shift;
2151 ir_tarval *tv_bitop;
2153 if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2))
2156 assert(is_And(n) || is_Or(n) || is_Eor(n) || is_Or_Eor_Add(n));
2157 if (!is_Const(right) || !is_shiftop(left))
2160 shift_left = get_binop_left(left);
2161 shift_right = get_binop_right(left);
2162 if (!is_Const(shift_right))
2165 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
2166 if (is_Shrs(left)) {
2167 /* TODO this could be improved */
2171 irg = get_irn_irg(n);
2172 block = get_nodes_block(n);
2173 dbg_bitop = get_irn_dbg_info(n);
2174 dbg_shift = get_irn_dbg_info(left);
2175 tv1 = get_Const_tarval(shift_right);
2176 tv2 = get_Const_tarval(right);
2177 assert(get_tarval_mode(tv2) == mode);
2180 tv_bitop = tarval_shr(tv2, tv1);
2182 /* Check whether we have lost some bits during the right shift. */
2184 ir_tarval *tv_back_again = tarval_shl(tv_bitop, tv1);
2186 if (tarval_cmp(tv_back_again, tv2) != ir_relation_equal)
2189 } else if (is_Shr(left)) {
2192 * TODO this can be improved by checking whether
2193 * the left shift produces an overflow
2197 tv_bitop = tarval_shl(tv2, tv1);
2199 assert(is_Rotl(left));
2200 tv_bitop = tarval_rotl(tv2, tarval_neg(tv1));
2202 new_const = new_r_Const(irg, tv_bitop);
2205 new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode);
2206 } else if (is_Or(n) || is_Or_Eor_Add(n)) {
2207 new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode);
2210 new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode);
2214 new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode);
2215 } else if (is_Shr(left)) {
2216 new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode);
2218 assert(is_Rotl(left));
2219 new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode);
2225 static bool complement_values(const ir_node *a, const ir_node *b)
2227 if (is_Not(a) && get_Not_op(a) == b)
2229 if (is_Not(b) && get_Not_op(b) == a)
2231 if (is_Const(a) && is_Const(b)) {
2232 ir_tarval *tv_a = get_Const_tarval(a);
2233 ir_tarval *tv_b = get_Const_tarval(b);
2234 return tarval_not(tv_a) == tv_b;
2239 typedef ir_tarval *(tv_fold_binop_func)(ir_tarval *a, ir_tarval *b);
2242 * for associative operations fold:
2243 * op(op(x, c0), c1) to op(x, op(c0, c1)) with constants folded.
2244 * This is a "light" version of the reassociation phase
2246 static ir_node *fold_constant_associativity(ir_node *node,
2247 tv_fold_binop_func fold)
2252 ir_node *right = get_binop_right(node);
2253 ir_node *left_right;
2260 if (!is_Const(right))
2263 op = get_irn_op(node);
2264 left = get_binop_left(node);
2265 if (get_irn_op(left) != op)
2268 left_right = get_binop_right(left);
2269 if (!is_Const(left_right))
2272 left_left = get_binop_left(left);
2273 c0 = get_Const_tarval(left_right);
2274 c1 = get_Const_tarval(right);
2275 irg = get_irn_irg(node);
2276 if (get_tarval_mode(c0) != get_tarval_mode(c1))
2278 new_c = fold(c0, c1);
2279 if (new_c == tarval_bad)
2281 new_const = new_r_Const(irg, new_c);
2282 new_node = exact_copy(node);
2283 set_binop_left(new_node, left_left);
2284 set_binop_right(new_node, new_const);
2291 static ir_node *transform_node_Or_(ir_node *n)
2294 ir_node *a = get_binop_left(n);
2295 ir_node *b = get_binop_right(n);
2299 n = fold_constant_associativity(n, tarval_or);
2303 if (is_Not(a) && is_Not(b)) {
2304 /* ~a | ~b = ~(a&b) */
2305 ir_node *block = get_nodes_block(n);
2307 mode = get_irn_mode(n);
2310 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
2311 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
2312 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
2316 /* we can combine the relations of two compares with the same operands */
2317 if (is_Cmp(a) && is_Cmp(b)) {
2318 ir_node *a_left = get_Cmp_left(a);
2319 ir_node *a_right = get_Cmp_right(a);
2320 ir_node *b_left = get_Cmp_left(b);
2321 ir_node *b_right = get_Cmp_right(b);
2322 if (a_left == b_left && b_left == b_right) {
2323 dbg_info *dbgi = get_irn_dbg_info(n);
2324 ir_node *block = get_nodes_block(n);
2325 ir_relation a_relation = get_Cmp_relation(a);
2326 ir_relation b_relation = get_Cmp_relation(b);
2327 ir_relation new_relation = a_relation | b_relation;
2328 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
2330 /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */
2331 if (is_cmp_unequal(a) && is_cmp_unequal(b)
2332 && !mode_is_float(get_irn_mode(a_left))
2333 && !mode_is_float(get_irn_mode(b_left))) {
2334 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
2335 ir_graph *irg = get_irn_irg(n);
2336 dbg_info *dbgi = get_irn_dbg_info(n);
2337 ir_node *block = get_nodes_block(n);
2338 ir_mode *a_mode = get_irn_mode(a_left);
2339 ir_mode *b_mode = get_irn_mode(b_left);
2340 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
2341 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
2342 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
2343 ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode);
2344 ir_node *zero = create_zero_const(irg, b_mode);
2345 return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater);
2347 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
2348 ir_graph *irg = get_irn_irg(n);
2349 dbg_info *dbgi = get_irn_dbg_info(n);
2350 ir_node *block = get_nodes_block(n);
2351 ir_mode *a_mode = get_irn_mode(a_left);
2352 ir_mode *b_mode = get_irn_mode(b_left);
2353 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
2354 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
2355 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
2356 ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode);
2357 ir_node *zero = create_zero_const(irg, a_mode);
2358 return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater);
2363 mode = get_irn_mode(n);
2364 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
2366 n = transform_node_Or_bf_store(n);
2369 n = transform_node_Or_Rotl(n);
2373 n = transform_bitwise_distributive(n, transform_node_Or_);
2376 n = transform_node_bitop_shift(n);
2383 static ir_node *transform_node_Or(ir_node *n)
2385 if (is_Or_Eor_Add(n)) {
2386 dbg_info *dbgi = get_irn_dbg_info(n);
2387 ir_node *block = get_nodes_block(n);
2388 ir_node *left = get_Or_left(n);
2389 ir_node *right = get_Or_right(n);
2390 ir_mode *mode = get_irn_mode(n);
2391 return new_rd_Add(dbgi, block, left, right, mode);
2393 return transform_node_Or_(n);
2399 static ir_node *transform_node_Eor_(ir_node *n)
2402 ir_node *a = get_binop_left(n);
2403 ir_node *b = get_binop_right(n);
2404 ir_mode *mode = get_irn_mode(n);
2407 n = fold_constant_associativity(n, tarval_eor);
2411 /* we can combine the relations of two compares with the same operands */
2412 if (is_Cmp(a) && is_Cmp(b)) {
2413 ir_node *a_left = get_Cmp_left(a);
2414 ir_node *a_right = get_Cmp_left(a);
2415 ir_node *b_left = get_Cmp_left(b);
2416 ir_node *b_right = get_Cmp_right(b);
2417 if (a_left == b_left && b_left == b_right) {
2418 dbg_info *dbgi = get_irn_dbg_info(n);
2419 ir_node *block = get_nodes_block(n);
2420 ir_relation a_relation = get_Cmp_relation(a);
2421 ir_relation b_relation = get_Cmp_relation(b);
2422 ir_relation new_relation = a_relation ^ b_relation;
2423 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
2427 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
2429 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
2430 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
2431 dbg_info *dbg = get_irn_dbg_info(n);
2432 ir_node *block = get_nodes_block(n);
2433 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
2434 ir_node *new_left = get_Not_op(a);
2435 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
2436 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2438 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
2439 dbg_info *dbg = get_irn_dbg_info(n);
2440 ir_node *block = get_nodes_block(n);
2441 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
2442 ir_node *new_right = get_Not_op(b);
2443 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
2444 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2448 /* x ^ 1...1 -> ~1 */
2449 if (is_Const(b) && is_Const_all_one(b)) {
2450 n = new_r_Not(get_nodes_block(n), a, mode);
2451 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2455 n = transform_bitwise_distributive(n, transform_node_Eor_);
2458 n = transform_node_bitop_shift(n);
2465 static ir_node *transform_node_Eor(ir_node *n)
2467 if (is_Or_Eor_Add(n)) {
2468 dbg_info *dbgi = get_irn_dbg_info(n);
2469 ir_node *block = get_nodes_block(n);
2470 ir_node *left = get_Eor_left(n);
2471 ir_node *right = get_Eor_right(n);
2472 ir_mode *mode = get_irn_mode(n);
2473 return new_rd_Add(dbgi, block, left, right, mode);
2475 return transform_node_Eor_(n);
2479 * Do the AddSub optimization, then Transform
2480 * Constant folding on Phi
2481 * Add(a,a) -> Mul(a, 2)
2482 * Add(Mul(a, x), a) -> Mul(a, x+1)
2483 * if the mode is integer or float.
2484 * Transform Add(a,-b) into Sub(a,b).
2485 * Reassociation might fold this further.
2487 static ir_node *transform_node_Add(ir_node *n)
2495 n = fold_constant_associativity(n, tarval_add);
2499 n = transform_node_AddSub(n);
2503 a = get_Add_left(n);
2504 b = get_Add_right(n);
2505 mode = get_irn_mode(n);
2507 if (mode_is_reference(mode)) {
2508 ir_mode *lmode = get_irn_mode(a);
2510 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2511 /* an Add(a, NULL) is a hidden Conv */
2512 dbg_info *dbg = get_irn_dbg_info(n);
2513 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2517 if (is_Const(b) && get_mode_arithmetic(mode) == irma_twos_complement) {
2518 ir_tarval *tv = get_Const_tarval(b);
2519 ir_tarval *min = get_mode_min(mode);
2520 /* if all bits are set, then this has the same effect as a Not.
2521 * Note that the following == gives false for different modes which
2522 * is exactly what we want */
2524 dbg_info *dbgi = get_irn_dbg_info(n);
2525 ir_graph *irg = get_irn_irg(n);
2526 ir_node *block = get_nodes_block(n);
2527 ir_node *cnst = new_r_Const(irg, min);
2528 return new_rd_Eor(dbgi, block, a, cnst, mode);
2532 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2534 /* for FP the following optimizations are only allowed if
2535 * fp_strict_algebraic is disabled */
2536 if (mode_is_float(mode)) {
2537 ir_graph *irg = get_irn_irg(n);
2538 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2542 if (mode_is_num(mode)) {
2543 ir_graph *irg = get_irn_irg(n);
2544 /* the following code leads to endless recursion when Mul are replaced
2545 * by a simple instruction chain */
2546 if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP)
2547 && a == b && mode_is_int(mode)) {
2548 ir_node *block = get_nodes_block(n);
2551 get_irn_dbg_info(n),
2554 new_r_Const_long(irg, mode, 2),
2556 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2561 get_irn_dbg_info(n),
2566 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2571 get_irn_dbg_info(n),
2576 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2579 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2580 /* Here we rely on constants be on the RIGHT side */
2582 ir_node *op = get_Not_op(a);
2584 if (is_Const(b) && is_Const_one(b)) {
2586 ir_node *blk = get_nodes_block(n);
2587 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2588 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2595 if (is_Or_Eor_Add(n)) {
2596 n = transform_node_Or_(n);
2599 n = transform_node_Eor_(n);
2608 * returns -cnst or NULL if impossible
2610 static ir_node *const_negate(ir_node *cnst)
2612 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2613 dbg_info *dbgi = get_irn_dbg_info(cnst);
2614 ir_graph *irg = get_irn_irg(cnst);
2615 if (tv == tarval_bad) return NULL;
2616 return new_rd_Const(dbgi, irg, tv);
2620 * Do the AddSub optimization, then Transform
2621 * Constant folding on Phi
2622 * Sub(0,a) -> Minus(a)
2623 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2624 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2625 * Sub(Add(a, x), x) -> a
2626 * Sub(x, Add(x, a)) -> -a
2627 * Sub(x, Const) -> Add(x, -Const)
2629 static ir_node *transform_node_Sub(ir_node *n)
2635 n = transform_node_AddSub(n);
2637 a = get_Sub_left(n);
2638 b = get_Sub_right(n);
2640 mode = get_irn_mode(n);
2642 if (mode_is_int(mode)) {
2643 ir_mode *lmode = get_irn_mode(a);
2645 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2646 /* a Sub(a, NULL) is a hidden Conv */
2647 dbg_info *dbg = get_irn_dbg_info(n);
2648 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2649 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2653 if (mode == lmode &&
2654 get_mode_arithmetic(mode) == irma_twos_complement &&
2656 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2658 dbg_info *dbg = get_irn_dbg_info(n);
2659 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2660 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2666 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2668 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2669 if (mode_is_float(mode)) {
2670 ir_graph *irg = get_irn_irg(n);
2671 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2675 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2676 /* a - C -> a + (-C) */
2677 ir_node *cnst = const_negate(b);
2679 ir_node *block = get_nodes_block(n);
2680 dbg_info *dbgi = get_irn_dbg_info(n);
2682 n = new_rd_Add(dbgi, block, a, cnst, mode);
2683 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2688 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2689 dbg_info *dbg = get_irn_dbg_info(n);
2690 ir_node *block = get_nodes_block(n);
2691 ir_node *left = get_Minus_op(a);
2692 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2694 n = new_rd_Minus(dbg, block, add, mode);
2695 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2697 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2698 dbg_info *dbg = get_irn_dbg_info(n);
2699 ir_node *block = get_nodes_block(n);
2700 ir_node *right = get_Minus_op(b);
2702 n = new_rd_Add(dbg, block, a, right, mode);
2703 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2705 } else if (is_Sub(b)) {
2706 /* a - (b - c) -> a + (c - b)
2707 * -> (a - b) + c iff (b - c) is a pointer */
2708 dbg_info *s_dbg = get_irn_dbg_info(b);
2709 ir_node *s_left = get_Sub_left(b);
2710 ir_node *s_right = get_Sub_right(b);
2711 ir_mode *s_mode = get_irn_mode(b);
2712 if (mode_is_reference(s_mode)) {
2713 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2714 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2715 dbg_info *a_dbg = get_irn_dbg_info(n);
2718 s_right = new_r_Conv(lowest_block, s_right, mode);
2719 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2721 ir_node *s_block = get_nodes_block(b);
2722 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2723 dbg_info *a_dbg = get_irn_dbg_info(n);
2724 ir_node *a_block = get_nodes_block(n);
2726 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2728 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2731 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2732 ir_node *m_right = get_Mul_right(b);
2733 if (is_Const(m_right)) {
2734 ir_node *cnst2 = const_negate(m_right);
2735 if (cnst2 != NULL) {
2736 dbg_info *m_dbg = get_irn_dbg_info(b);
2737 ir_node *m_block = get_nodes_block(b);
2738 ir_node *m_left = get_Mul_left(b);
2739 ir_mode *m_mode = get_irn_mode(b);
2740 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2741 dbg_info *a_dbg = get_irn_dbg_info(n);
2742 ir_node *a_block = get_nodes_block(n);
2744 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2745 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2752 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2753 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2755 get_irn_dbg_info(n),
2759 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2762 if ((is_Add(a) || is_Or_Eor_Add(a)) && mode_wrap_around(mode)) {
2763 ir_node *left = get_binop_left(a);
2764 ir_node *right = get_binop_right(a);
2766 /* FIXME: Does the Conv's work only for two complement or generally? */
2768 if (mode != get_irn_mode(right)) {
2769 /* This Sub is an effective Cast */
2770 right = new_r_Conv(get_nodes_block(n), right, mode);
2773 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2775 } else if (right == b) {
2776 if (mode != get_irn_mode(left)) {
2777 /* This Sub is an effective Cast */
2778 left = new_r_Conv(get_nodes_block(n), left, mode);
2781 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2785 if ((is_Add(b) || is_Or_Eor_Add(b)) && mode_wrap_around(mode)) {
2786 ir_node *left = get_binop_left(b);
2787 ir_node *right = get_binop_right(b);
2789 /* FIXME: Does the Conv's work only for two complement or generally? */
2791 ir_mode *r_mode = get_irn_mode(right);
2793 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2794 if (mode != r_mode) {
2795 /* This Sub is an effective Cast */
2796 n = new_r_Conv(get_nodes_block(n), n, mode);
2798 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2800 } else if (right == a) {
2801 ir_mode *l_mode = get_irn_mode(left);
2803 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2804 if (mode != l_mode) {
2805 /* This Sub is an effective Cast */
2806 n = new_r_Conv(get_nodes_block(n), n, mode);
2808 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2812 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2813 ir_mode *mode = get_irn_mode(a);
2815 if (mode == get_irn_mode(b)) {
2817 ir_node *op_a = get_Conv_op(a);
2818 ir_node *op_b = get_Conv_op(b);
2820 /* check if it's allowed to skip the conv */
2821 ma = get_irn_mode(op_a);
2822 mb = get_irn_mode(op_b);
2824 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2825 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2828 set_Sub_right(n, b);
2834 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2835 if (!is_reassoc_running() && is_Mul(a)) {
2836 ir_node *ma = get_Mul_left(a);
2837 ir_node *mb = get_Mul_right(a);
2840 ir_node *blk = get_nodes_block(n);
2841 ir_graph *irg = get_irn_irg(n);
2843 get_irn_dbg_info(n),
2847 get_irn_dbg_info(n),
2850 new_r_Const(irg, get_mode_one(mode)),
2853 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2855 } else if (mb == b) {
2856 ir_node *blk = get_nodes_block(n);
2857 ir_graph *irg = get_irn_irg(n);
2859 get_irn_dbg_info(n),
2863 get_irn_dbg_info(n),
2866 new_r_Const(irg, get_mode_one(mode)),
2869 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2873 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2874 ir_node *x = get_Sub_left(a);
2875 ir_node *y = get_Sub_right(a);
2876 ir_node *blk = get_nodes_block(n);
2877 ir_mode *m_b = get_irn_mode(b);
2878 ir_mode *m_y = get_irn_mode(y);
2882 /* Determine the right mode for the Add. */
2885 else if (mode_is_reference(m_b))
2887 else if (mode_is_reference(m_y))
2891 * Both modes are different but none is reference,
2892 * happens for instance in SubP(SubP(P, Iu), Is).
2893 * We have two possibilities here: Cast or ignore.
2894 * Currently we ignore this case.
2899 add = new_r_Add(blk, y, b, add_mode);
2901 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2902 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2906 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2907 /* c - ~X = X + (c+1) */
2908 if (is_Const(a) && is_Not(b)) {
2909 ir_tarval *tv = get_Const_tarval(a);
2911 tv = tarval_add(tv, get_mode_one(mode));
2912 if (tv != tarval_bad) {
2913 ir_node *blk = get_nodes_block(n);
2914 ir_graph *irg = get_irn_irg(n);
2915 ir_node *c = new_r_Const(irg, tv);
2916 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2917 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2921 /* x-(x&y) = x & ~y */
2923 ir_node *and_left = get_And_left(b);
2924 ir_node *and_right = get_And_right(b);
2925 if (and_right == a) {
2926 ir_node *tmp = and_left;
2927 and_left = and_right;
2930 if (and_left == a) {
2931 dbg_info *dbgi = get_irn_dbg_info(n);
2932 ir_node *block = get_nodes_block(n);
2933 ir_mode *mode = get_irn_mode(n);
2934 ir_node *notn = new_rd_Not(dbgi, block, and_right, mode);
2935 ir_node *andn = new_rd_And(dbgi, block, a, notn, mode);
2944 * Several transformation done on n*n=2n bits mul.
2945 * These transformations must be done here because new nodes may be produced.
2947 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2950 ir_node *a = get_Mul_left(n);
2951 ir_node *b = get_Mul_right(n);
2952 ir_tarval *ta = value_of(a);
2953 ir_tarval *tb = value_of(b);
2954 ir_mode *smode = get_irn_mode(a);
2956 if (ta == get_mode_one(smode)) {
2957 /* (L)1 * (L)b = (L)b */
2958 ir_node *blk = get_nodes_block(n);
2959 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2960 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2963 else if (ta == get_mode_minus_one(smode)) {
2964 /* (L)-1 * (L)b = (L)b */
2965 ir_node *blk = get_nodes_block(n);
2966 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2967 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2968 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2971 if (tb == get_mode_one(smode)) {
2972 /* (L)a * (L)1 = (L)a */
2973 ir_node *blk = get_nodes_block(a);
2974 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2975 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2978 else if (tb == get_mode_minus_one(smode)) {
2979 /* (L)a * (L)-1 = (L)-a */
2980 ir_node *blk = get_nodes_block(n);
2981 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2982 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2983 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2990 * Transform Mul(a,-1) into -a.
2991 * Do constant evaluation of Phi nodes.
2992 * Do architecture dependent optimizations on Mul nodes
2994 static ir_node *transform_node_Mul(ir_node *n)
2996 ir_node *c, *oldn = n;
2997 ir_mode *mode = get_irn_mode(n);
2998 ir_node *a = get_Mul_left(n);
2999 ir_node *b = get_Mul_right(n);
3001 n = fold_constant_associativity(n, tarval_mul);
3005 if (mode != get_irn_mode(a))
3006 return transform_node_Mul2n(n, mode);
3008 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
3010 if (mode_is_signed(mode)) {
3013 if (value_of(a) == get_mode_minus_one(mode))
3015 else if (value_of(b) == get_mode_minus_one(mode))
3018 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
3019 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3024 if (is_Const(b)) { /* (-a) * const -> a * -const */
3025 ir_node *cnst = const_negate(b);
3027 dbg_info *dbgi = get_irn_dbg_info(n);
3028 ir_node *block = get_nodes_block(n);
3029 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
3030 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3033 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
3034 dbg_info *dbgi = get_irn_dbg_info(n);
3035 ir_node *block = get_nodes_block(n);
3036 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
3037 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
3039 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
3040 ir_node *sub_l = get_Sub_left(b);
3041 ir_node *sub_r = get_Sub_right(b);
3042 dbg_info *dbgi = get_irn_dbg_info(n);
3043 ir_node *block = get_nodes_block(n);
3044 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
3045 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
3046 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
3049 } else if (is_Minus(b)) {
3050 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
3051 ir_node *sub_l = get_Sub_left(a);
3052 ir_node *sub_r = get_Sub_right(a);
3053 dbg_info *dbgi = get_irn_dbg_info(n);
3054 ir_node *block = get_nodes_block(n);
3055 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
3056 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
3057 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
3060 } else if (is_Shl(a)) {
3061 ir_node *const shl_l = get_Shl_left(a);
3062 if (is_Const(shl_l) && is_Const_one(shl_l)) {
3063 /* (1 << x) * b -> b << x */
3064 dbg_info *const dbgi = get_irn_dbg_info(n);
3065 ir_node *const block = get_nodes_block(n);
3066 ir_node *const shl_r = get_Shl_right(a);
3067 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
3068 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
3071 } else if (is_Shl(b)) {
3072 ir_node *const shl_l = get_Shl_left(b);
3073 if (is_Const(shl_l) && is_Const_one(shl_l)) {
3074 /* a * (1 << x) -> a << x */
3075 dbg_info *const dbgi = get_irn_dbg_info(n);
3076 ir_node *const block = get_nodes_block(n);
3077 ir_node *const shl_r = get_Shl_right(b);
3078 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
3079 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
3083 if (get_mode_arithmetic(mode) == irma_ieee754
3084 || get_mode_arithmetic(mode) == irma_x86_extended_float) {
3086 ir_tarval *tv = get_Const_tarval(a);
3087 if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv)
3088 && !tarval_is_negative(tv)) {
3089 /* 2.0 * b = b + b */
3090 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
3091 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
3095 else if (is_Const(b)) {
3096 ir_tarval *tv = get_Const_tarval(b);
3097 if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv)
3098 && !tarval_is_negative(tv)) {
3099 /* a * 2.0 = a + a */
3100 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
3101 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
3106 return arch_dep_replace_mul_with_shifts(n);
3110 * Transform a Div Node.
3112 static ir_node *transform_node_Div(ir_node *n)
3114 ir_mode *mode = get_Div_resmode(n);
3115 ir_node *a = get_Div_left(n);
3116 ir_node *b = get_Div_right(n);
3118 const ir_node *dummy;
3120 if (mode_is_int(mode)) {
3121 if (is_Const(b) && is_const_Phi(a)) {
3122 /* check for Div(Phi, Const) */
3123 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3125 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3128 } else if (is_Const(a) && is_const_Phi(b)) {
3129 /* check for Div(Const, Phi) */
3130 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3132 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3135 } else if (is_const_Phi(a) && is_const_Phi(b)) {
3136 /* check for Div(Phi, Phi) */
3137 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3139 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3144 if (a == b && value_not_zero(a, &dummy)) {
3145 ir_graph *irg = get_irn_irg(n);
3146 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
3147 value = new_r_Const(irg, get_mode_one(mode));
3148 DBG_OPT_CSTEVAL(n, value);
3151 if (mode_is_signed(mode) && is_Const(b)) {
3152 ir_tarval *tv = get_Const_tarval(b);
3154 if (tv == get_mode_minus_one(mode)) {
3156 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3157 DBG_OPT_CSTEVAL(n, value);
3161 /* Try architecture dependent optimization */
3162 value = arch_dep_replace_div_by_const(n);
3165 assert(mode_is_float(mode));
3167 /* Optimize x/c to x*(1/c) */
3168 if (get_mode_arithmetic(mode) == irma_ieee754) {
3169 ir_tarval *tv = value_of(b);
3171 if (tv != tarval_bad) {
3172 int rem = tarval_fp_ops_enabled();
3175 * Floating point constant folding might be disabled here to
3177 * However, as we check for exact result, doing it is safe.
3180 tarval_enable_fp_ops(1);
3181 tv = tarval_div(get_mode_one(mode), tv);
3182 tarval_enable_fp_ops(rem);
3184 /* Do the transformation if the result is either exact or we are
3185 not using strict rules. */
3186 if (tv != tarval_bad &&
3187 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
3188 ir_node *block = get_nodes_block(n);
3189 ir_graph *irg = get_irn_irg(block);
3190 ir_node *c = new_r_Const(irg, tv);
3191 dbg_info *dbgi = get_irn_dbg_info(n);
3192 value = new_rd_Mul(dbgi, block, a, c, mode);
3205 /* Turn Div into a tuple (mem, jmp, bad, value) */
3206 mem = get_Div_mem(n);
3207 blk = get_nodes_block(n);
3208 irg = get_irn_irg(blk);
3210 /* skip a potential Pin */
3211 mem = skip_Pin(mem);
3212 turn_into_tuple(n, pn_Div_max+1);
3213 set_Tuple_pred(n, pn_Div_M, mem);
3214 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
3215 set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X));
3216 set_Tuple_pred(n, pn_Div_res, value);
3222 * Transform a Mod node.
3224 static ir_node *transform_node_Mod(ir_node *n)
3226 ir_mode *mode = get_Mod_resmode(n);
3227 ir_node *a = get_Mod_left(n);
3228 ir_node *b = get_Mod_right(n);
3233 if (is_Const(b) && is_const_Phi(a)) {
3234 /* check for Div(Phi, Const) */
3235 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3237 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3241 else if (is_Const(a) && is_const_Phi(b)) {
3242 /* check for Div(Const, Phi) */
3243 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3245 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3249 else if (is_const_Phi(a) && is_const_Phi(b)) {
3250 /* check for Div(Phi, Phi) */
3251 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3253 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3260 irg = get_irn_irg(n);
3261 if (tv != tarval_bad) {
3262 value = new_r_Const(irg, tv);
3264 DBG_OPT_CSTEVAL(n, value);
3267 ir_node *a = get_Mod_left(n);
3268 ir_node *b = get_Mod_right(n);
3269 const ir_node *dummy;
3271 if (a == b && value_not_zero(a, &dummy)) {
3272 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
3273 value = new_r_Const(irg, get_mode_null(mode));
3274 DBG_OPT_CSTEVAL(n, value);
3277 if (mode_is_signed(mode) && is_Const(b)) {
3278 ir_tarval *tv = get_Const_tarval(b);
3280 if (tv == get_mode_minus_one(mode)) {
3282 value = new_r_Const(irg, get_mode_null(mode));
3283 DBG_OPT_CSTEVAL(n, value);
3287 /* Try architecture dependent optimization */
3288 value = arch_dep_replace_mod_by_const(n);
3297 /* Turn Mod into a tuple (mem, jmp, bad, value) */
3298 mem = get_Mod_mem(n);
3299 blk = get_nodes_block(n);
3300 irg = get_irn_irg(blk);
3302 /* skip a potential Pin */
3303 mem = skip_Pin(mem);
3304 turn_into_tuple(n, pn_Mod_max+1);
3305 set_Tuple_pred(n, pn_Mod_M, mem);
3306 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3307 set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X));
3308 set_Tuple_pred(n, pn_Mod_res, value);
3314 * Transform a Cond node.
3316 * Replace the Cond by a Jmp if it branches on a constant
3319 static ir_node *transform_node_Cond(ir_node *n)
3321 ir_node *a = get_Cond_selector(n);
3322 ir_graph *irg = get_irn_irg(n);
3326 /* we need block info which is not available in floating irgs */
3327 if (get_irg_pinned(irg) == op_pin_state_floats)
3331 if (ta == tarval_bad && is_Cmp(a)) {
3332 /* try again with a direct call to compute_cmp, as we don't care
3333 * about the MODEB_LOWERED flag here */
3334 ta = compute_cmp_ext(a);
3337 if (ta != tarval_bad && get_irn_mode(a) == mode_b) {
3338 /* It's a boolean Cond, branching on a boolean constant.
3339 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3340 ir_node *blk = get_nodes_block(n);
3341 jmp = new_r_Jmp(blk);
3342 turn_into_tuple(n, pn_Cond_max+1);
3343 if (ta == tarval_b_true) {
3344 set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
3345 set_Tuple_pred(n, pn_Cond_true, jmp);
3347 set_Tuple_pred(n, pn_Cond_false, jmp);
3348 set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X));
3350 clear_irg_properties(irg, IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE);
3355 static ir_node *transform_node_Switch(ir_node *n)
3357 ir_node *op = get_Switch_selector(n);
3358 ir_tarval *val = value_of(op);
3359 if (val != tarval_bad) {
3360 dbg_info *dbgi = get_irn_dbg_info(n);
3361 ir_graph *irg = get_irn_irg(n);
3362 unsigned n_outs = get_Switch_n_outs(n);
3363 ir_node *block = get_nodes_block(n);
3364 ir_node *bad = new_r_Bad(irg, mode_X);
3365 ir_node **in = XMALLOCN(ir_node*, n_outs);
3366 const ir_switch_table *table = get_Switch_table(n);
3367 size_t n_entries = ir_switch_table_get_n_entries(table);
3371 for (i = 0; i < n_entries; ++i) {
3372 const ir_switch_table_entry *entry
3373 = ir_switch_table_get_entry_const(table, i);
3374 ir_tarval *min = entry->min;
3375 ir_tarval *max = entry->max;
3378 if ((min == max && min == val)
3379 || (tarval_cmp(val, min) != ir_relation_less
3380 && tarval_cmp(val, max) != ir_relation_greater)) {
3385 for (o = 0; o < n_outs; ++o) {
3386 if (o == (unsigned)jmp_pn) {
3387 in[o] = new_rd_Jmp(dbgi, block);
3392 return new_r_Tuple(block, (int)n_outs, in);
3398 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
3400 * - and, or, xor instead of &
3401 * - Shl, Shr, Shrs, rotl instead of >>
3402 * (with a special case for Or/Xor + Shrs)
3404 * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
3406 static ir_node *transform_node_shift_bitop(ir_node *n)
3408 ir_graph *irg = get_irn_irg(n);
3409 ir_node *right = get_binop_right(n);
3410 ir_mode *mode = get_irn_mode(n);
3412 ir_node *bitop_left;
3413 ir_node *bitop_right;
3422 ir_tarval *tv_shift;
3424 if (irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2))
3427 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
3429 if (!is_Const(right))
3432 left = get_binop_left(n);
3433 op_left = get_irn_op(left);
3434 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
3437 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
3438 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
3439 /* TODO: test if sign bit is affectes */
3443 bitop_right = get_binop_right(left);
3444 if (!is_Const(bitop_right))
3447 bitop_left = get_binop_left(left);
3449 block = get_nodes_block(n);
3450 dbgi = get_irn_dbg_info(n);
3451 tv1 = get_Const_tarval(bitop_right);
3452 tv2 = get_Const_tarval(right);
3454 assert(get_tarval_mode(tv1) == mode);
3457 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
3458 tv_shift = tarval_shl(tv1, tv2);
3459 } else if (is_Shr(n)) {
3460 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
3461 tv_shift = tarval_shr(tv1, tv2);
3462 } else if (is_Shrs(n)) {
3463 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
3464 tv_shift = tarval_shrs(tv1, tv2);
3467 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
3468 tv_shift = tarval_rotl(tv1, tv2);
3471 assert(get_tarval_mode(tv_shift) == mode);
3472 irg = get_irn_irg(n);
3473 new_const = new_r_Const(irg, tv_shift);
3475 if (op_left == op_And) {
3476 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
3477 } else if (op_left == op_Or) {
3478 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
3480 assert(op_left == op_Eor);
3481 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
3490 static ir_node *transform_node_And(ir_node *n)
3492 ir_node *c, *oldn = n;
3493 ir_node *a = get_And_left(n);
3494 ir_node *b = get_And_right(n);
3497 n = fold_constant_associativity(n, tarval_and);
3501 if (is_Cmp(a) && is_Cmp(b)) {
3502 ir_node *a_left = get_Cmp_left(a);
3503 ir_node *a_right = get_Cmp_right(a);
3504 ir_node *b_left = get_Cmp_left(b);
3505 ir_node *b_right = get_Cmp_right(b);
3506 ir_relation a_relation = get_Cmp_relation(a);
3507 ir_relation b_relation = get_Cmp_relation(b);
3508 /* we can combine the relations of two compares with the same
3510 if (a_left == b_left && b_left == b_right) {
3511 dbg_info *dbgi = get_irn_dbg_info(n);
3512 ir_node *block = get_nodes_block(n);
3513 ir_relation new_relation = a_relation & b_relation;
3514 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3516 /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */
3517 if (a_relation == b_relation && a_relation == ir_relation_equal
3518 && !mode_is_float(get_irn_mode(a_left))
3519 && !mode_is_float(get_irn_mode(b_left))) {
3520 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
3521 dbg_info *dbgi = get_irn_dbg_info(n);
3522 ir_node *block = get_nodes_block(n);
3523 ir_mode *a_mode = get_irn_mode(a_left);
3524 ir_mode *b_mode = get_irn_mode(b_left);
3525 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3526 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3527 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
3528 ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode);
3529 ir_graph *irg = get_irn_irg(n);
3530 ir_node *zero = create_zero_const(irg, b_mode);
3531 return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal);
3533 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
3534 dbg_info *dbgi = get_irn_dbg_info(n);
3535 ir_node *block = get_nodes_block(n);
3536 ir_mode *a_mode = get_irn_mode(a_left);
3537 ir_mode *b_mode = get_irn_mode(b_left);
3538 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3539 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3540 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
3541 ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode);
3542 ir_graph *irg = get_irn_irg(n);
3543 ir_node *zero = create_zero_const(irg, a_mode);
3544 return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal);
3549 mode = get_irn_mode(n);
3550 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3552 if (is_Or(a) || is_Or_Eor_Add(a)) {
3553 ir_node *or_left = get_binop_left(a);
3554 ir_node *or_right = get_binop_right(a);
3555 if (complement_values(or_left, b)) {
3556 /* (a|b) & ~a => b & ~a */
3557 dbg_info *dbgi = get_irn_dbg_info(n);
3558 ir_node *block = get_nodes_block(n);
3559 return new_rd_And(dbgi, block, or_right, b, mode);
3560 } else if (complement_values(or_right, b)) {
3561 /* (a|b) & ~b => a & ~b */
3562 dbg_info *dbgi = get_irn_dbg_info(n);
3563 ir_node *block = get_nodes_block(n);
3564 return new_rd_And(dbgi, block, or_left, b, mode);
3565 } else if (is_Not(b)) {
3566 ir_node *op = get_Not_op(b);
3568 ir_node *ba = get_And_left(op);
3569 ir_node *bb = get_And_right(op);
3571 /* it's enough to test the following cases due to normalization! */
3572 if (or_left == ba && or_right == bb) {
3573 /* (a|b) & ~(a&b) = a^b */
3574 ir_node *block = get_nodes_block(n);
3576 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3577 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3583 if (is_Or(b) || is_Or_Eor_Add(b)) {
3584 ir_node *or_left = get_binop_left(b);
3585 ir_node *or_right = get_binop_right(b);
3586 if (complement_values(or_left, a)) {
3587 /* (a|b) & ~a => b & ~a */
3588 dbg_info *dbgi = get_irn_dbg_info(n);
3589 ir_node *block = get_nodes_block(n);
3590 return new_rd_And(dbgi, block, or_right, a, mode);
3591 } else if (complement_values(or_right, a)) {
3592 /* (a|b) & ~b => a & ~b */
3593 dbg_info *dbgi = get_irn_dbg_info(n);
3594 ir_node *block = get_nodes_block(n);
3595 return new_rd_And(dbgi, block, or_left, a, mode);
3596 } else if (is_Not(a)) {
3597 ir_node *op = get_Not_op(a);
3599 ir_node *aa = get_And_left(op);
3600 ir_node *ab = get_And_right(op);
3602 /* it's enough to test the following cases due to normalization! */
3603 if (or_left == aa && or_right == ab) {
3604 /* (a|b) & ~(a&b) = a^b */
3605 ir_node *block = get_nodes_block(n);
3607 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3608 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3614 if (is_Eor(a) || is_Or_Eor_Add(a)) {
3615 ir_node *al = get_binop_left(a);
3616 ir_node *ar = get_binop_right(a);
3619 /* (b ^ a) & b -> ~a & b */
3620 dbg_info *dbg = get_irn_dbg_info(n);
3621 ir_node *block = get_nodes_block(n);
3623 ar = new_rd_Not(dbg, block, ar, mode);
3624 n = new_rd_And(dbg, block, ar, b, mode);
3625 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3629 /* (a ^ b) & b -> ~a & b */
3630 dbg_info *dbg = get_irn_dbg_info(n);
3631 ir_node *block = get_nodes_block(n);
3633 al = new_rd_Not(dbg, block, al, mode);
3634 n = new_rd_And(dbg, block, al, b, mode);
3635 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3639 if (is_Eor(b) || is_Or_Eor_Add(b)) {
3640 ir_node *bl = get_binop_left(b);
3641 ir_node *br = get_binop_right(b);
3644 /* a & (a ^ b) -> a & ~b */
3645 dbg_info *dbg = get_irn_dbg_info(n);
3646 ir_node *block = get_nodes_block(n);
3648 br = new_rd_Not(dbg, block, br, mode);
3649 n = new_rd_And(dbg, block, br, a, mode);
3650 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3654 /* a & (b ^ a) -> a & ~b */
3655 dbg_info *dbg = get_irn_dbg_info(n);
3656 ir_node *block = get_nodes_block(n);
3658 bl = new_rd_Not(dbg, block, bl, mode);
3659 n = new_rd_And(dbg, block, bl, a, mode);
3660 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3664 if (is_Not(a) && is_Not(b)) {
3665 /* ~a & ~b = ~(a|b) */
3666 ir_node *block = get_nodes_block(n);
3667 ir_mode *mode = get_irn_mode(n);
3671 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3672 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3673 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3678 vrp_attr *b_vrp = vrp_get_info(b);
3679 ir_tarval *a_val = get_Const_tarval(a);
3680 if (b_vrp != NULL && tarval_or(a_val, b_vrp->bits_not_set) == a_val) {
3686 vrp_attr *a_vrp = vrp_get_info(a);
3687 ir_tarval *b_val = get_Const_tarval(b);
3688 if (a_vrp != NULL && tarval_or(b_val, a_vrp->bits_not_set) == b_val) {
3693 n = transform_bitwise_distributive(n, transform_node_And);
3695 n = transform_node_bitop_shift(n);
3703 static ir_node *transform_node_Not(ir_node *n)
3705 ir_node *c, *oldn = n;
3706 ir_node *a = get_Not_op(n);
3707 ir_mode *mode = get_irn_mode(n);
3709 HANDLE_UNOP_PHI(tarval_not,a,c);
3711 /* check for a boolean Not */
3713 dbg_info *dbgi = get_irn_dbg_info(a);
3714 ir_node *block = get_nodes_block(a);
3715 ir_relation relation = get_Cmp_relation(a);
3716 relation = get_negated_relation(relation);
3717 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3718 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3722 /* normalize ~(a ^ b) => a ^ ~b */
3723 if (is_Eor(a) || is_Or_Eor_Add(a)) {
3724 dbg_info *dbg = get_irn_dbg_info(n);
3725 ir_node *block = get_nodes_block(n);
3726 ir_node *eor_right = get_binop_right(a);
3727 ir_node *eor_left = get_binop_left(a);
3728 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3729 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3733 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3734 if (is_Minus(a)) { /* ~-x -> x + -1 */
3735 dbg_info *dbg = get_irn_dbg_info(n);
3736 ir_graph *irg = get_irn_irg(n);
3737 ir_node *block = get_nodes_block(n);
3738 ir_node *add_l = get_Minus_op(a);
3739 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3740 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3741 } else if (is_Add(a) || is_Or_Eor_Add(a)) {
3742 ir_node *add_r = get_binop_right(a);
3743 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3744 /* ~(x + -1) = -x */
3745 ir_node *op = get_binop_left(a);
3746 ir_node *blk = get_nodes_block(n);
3747 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3748 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3756 * Transform a Minus.
3760 * -(a >>u (size-1)) = a >>s (size-1)
3761 * -(a >>s (size-1)) = a >>u (size-1)
3762 * -(a * const) -> a * -const
3764 static ir_node *transform_node_Minus(ir_node *n)
3766 ir_node *c, *oldn = n;
3767 ir_node *a = get_Minus_op(n);
3770 HANDLE_UNOP_PHI(tarval_neg,a,c);
3772 mode = get_irn_mode(a);
3773 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3774 /* the following rules are only to twos-complement */
3777 ir_node *op = get_Not_op(a);
3778 ir_tarval *tv = get_mode_one(mode);
3779 ir_node *blk = get_nodes_block(n);
3780 ir_graph *irg = get_irn_irg(blk);
3781 ir_node *c = new_r_Const(irg, tv);
3782 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3783 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3787 ir_node *c = get_Shr_right(a);
3790 ir_tarval *tv = get_Const_tarval(c);
3792 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3793 /* -(a >>u (size-1)) = a >>s (size-1) */
3794 ir_node *v = get_Shr_left(a);
3796 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3797 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3803 ir_node *c = get_Shrs_right(a);
3806 ir_tarval *tv = get_Const_tarval(c);
3808 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3809 /* -(a >>s (size-1)) = a >>u (size-1) */
3810 ir_node *v = get_Shrs_left(a);
3812 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3813 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3820 /* - (a-b) = b - a */
3821 ir_node *la = get_Sub_left(a);
3822 ir_node *ra = get_Sub_right(a);
3823 ir_node *blk = get_nodes_block(n);
3825 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3826 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3830 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3831 ir_node *mul_l = get_Mul_left(a);
3832 ir_node *mul_r = get_Mul_right(a);
3833 ir_tarval *tv = value_of(mul_r);
3834 if (tv != tarval_bad) {
3835 tv = tarval_neg(tv);
3836 if (tv != tarval_bad) {
3837 ir_graph *irg = get_irn_irg(n);
3838 ir_node *cnst = new_r_Const(irg, tv);
3839 dbg_info *dbg = get_irn_dbg_info(a);
3840 ir_node *block = get_nodes_block(a);
3841 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3842 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3852 * Transform a Proj(Load) with a non-null address.
3854 static ir_node *transform_node_Proj_Load(ir_node *proj)
3856 if (get_irn_mode(proj) == mode_X) {
3857 ir_node *load = get_Proj_pred(proj);
3859 /* get the Load address */
3860 const ir_node *addr = get_Load_ptr(load);
3861 const ir_node *confirm;
3863 if (value_not_null(addr, &confirm)) {
3864 if (confirm == NULL) {
3865 /* this node may float if it did not depend on a Confirm */
3866 set_irn_pinned(load, op_pin_state_floats);
3868 if (get_Proj_proj(proj) == pn_Load_X_except) {
3869 ir_graph *irg = get_irn_irg(proj);
3870 DBG_OPT_EXC_REM(proj);
3871 return new_r_Bad(irg, mode_X);
3873 ir_node *blk = get_nodes_block(load);
3874 return new_r_Jmp(blk);
3882 * Transform a Proj(Store) with a non-null address.
3884 static ir_node *transform_node_Proj_Store(ir_node *proj)
3886 if (get_irn_mode(proj) == mode_X) {
3887 ir_node *store = get_Proj_pred(proj);
3889 /* get the load/store address */
3890 const ir_node *addr = get_Store_ptr(store);
3891 const ir_node *confirm;
3893 if (value_not_null(addr, &confirm)) {
3894 if (confirm == NULL) {
3895 /* this node may float if it did not depend on a Confirm */
3896 set_irn_pinned(store, op_pin_state_floats);
3898 if (get_Proj_proj(proj) == pn_Store_X_except) {
3899 ir_graph *irg = get_irn_irg(proj);
3900 DBG_OPT_EXC_REM(proj);
3901 return new_r_Bad(irg, mode_X);
3903 ir_node *blk = get_nodes_block(store);
3904 return new_r_Jmp(blk);
3912 * Transform a Proj(Div) with a non-zero value.
3913 * Removes the exceptions and routes the memory to the NoMem node.
3915 static ir_node *transform_node_Proj_Div(ir_node *proj)
3917 ir_node *div = get_Proj_pred(proj);
3918 ir_node *b = get_Div_right(div);
3919 ir_node *res, *new_mem;
3920 const ir_node *confirm;
3923 if (value_not_zero(b, &confirm)) {
3924 /* div(x, y) && y != 0 */
3925 if (confirm == NULL) {
3926 /* we are sure we have a Const != 0 */
3927 new_mem = get_Div_mem(div);
3928 new_mem = skip_Pin(new_mem);
3929 set_Div_mem(div, new_mem);
3930 set_irn_pinned(div, op_pin_state_floats);
3933 proj_nr = get_Proj_proj(proj);
3935 case pn_Div_X_regular:
3936 return new_r_Jmp(get_nodes_block(div));
3938 case pn_Div_X_except: {
3939 ir_graph *irg = get_irn_irg(proj);
3940 /* we found an exception handler, remove it */
3941 DBG_OPT_EXC_REM(proj);
3942 return new_r_Bad(irg, mode_X);
3946 ir_graph *irg = get_irn_irg(proj);
3947 res = get_Div_mem(div);
3948 new_mem = get_irg_no_mem(irg);
3951 /* This node can only float up to the Confirm block */
3952 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3954 set_irn_pinned(div, op_pin_state_floats);
3955 /* this is a Div without exception, we can remove the memory edge */
3956 set_Div_mem(div, new_mem);
3965 * Transform a Proj(Mod) with a non-zero value.
3966 * Removes the exceptions and routes the memory to the NoMem node.
3968 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3970 ir_node *mod = get_Proj_pred(proj);
3971 ir_node *b = get_Mod_right(mod);
3972 ir_node *res, *new_mem;
3973 const ir_node *confirm;
3976 if (value_not_zero(b, &confirm)) {
3977 /* mod(x, y) && y != 0 */
3978 proj_nr = get_Proj_proj(proj);
3980 if (confirm == NULL) {
3981 /* we are sure we have a Const != 0 */
3982 new_mem = get_Mod_mem(mod);
3983 new_mem = skip_Pin(new_mem);
3984 set_Mod_mem(mod, new_mem);
3985 set_irn_pinned(mod, op_pin_state_floats);
3990 case pn_Mod_X_regular:
3991 return new_r_Jmp(get_nodes_block(mod));
3993 case pn_Mod_X_except: {
3994 ir_graph *irg = get_irn_irg(proj);
3995 /* we found an exception handler, remove it */
3996 DBG_OPT_EXC_REM(proj);
3997 return new_r_Bad(irg, mode_X);
4001 ir_graph *irg = get_irn_irg(proj);
4002 res = get_Mod_mem(mod);
4003 new_mem = get_irg_no_mem(irg);
4006 /* This node can only float up to the Confirm block */
4007 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
4009 /* this is a Mod without exception, we can remove the memory edge */
4010 set_Mod_mem(mod, new_mem);
4014 if (get_Mod_left(mod) == b) {
4015 /* a % a = 0 if a != 0 */
4016 ir_graph *irg = get_irn_irg(proj);
4017 ir_mode *mode = get_irn_mode(proj);
4018 ir_node *res = new_r_Const(irg, get_mode_null(mode));
4020 DBG_OPT_CSTEVAL(mod, res);
4029 * return true if the operation returns a value with exactly 1 bit set
4031 static bool is_single_bit(const ir_node *node)
4033 /* a first implementation, could be extended with vrp and others... */
4035 ir_node *shl_l = get_Shl_left(node);
4036 ir_mode *mode = get_irn_mode(node);
4037 int modulo = get_mode_modulo_shift(mode);
4038 /* this works if we shift a 1 and we have modulo shift */
4039 if (is_Const(shl_l) && is_Const_one(shl_l)
4040 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
4043 } else if (is_Const(node)) {
4044 ir_tarval *tv = get_Const_tarval(node);
4045 return tarval_is_single_bit(tv);
4051 * checks if node just flips a bit in another node and returns that other node
4052 * if so. @p tv should be a value having just 1 bit set
4054 static ir_node *flips_bit(const ir_node *node, ir_tarval *tv)
4057 return get_Not_op(node);
4059 ir_node *right = get_Eor_right(node);
4060 if (is_Const(right)) {
4061 ir_tarval *right_tv = get_Const_tarval(right);
4062 ir_mode *mode = get_irn_mode(node);
4063 if (tarval_and(right_tv, tv) != get_mode_null(mode))
4064 return get_Eor_left(node);
4071 * Normalizes and optimizes Cmp nodes.
4073 static ir_node *transform_node_Cmp(ir_node *n)
4075 ir_node *left = get_Cmp_left(n);
4076 ir_node *right = get_Cmp_right(n);
4077 ir_mode *mode = get_irn_mode(left);
4078 ir_tarval *tv = NULL;
4079 bool changed = false;
4080 bool changedc = false;
4081 ir_relation relation = get_Cmp_relation(n);
4082 ir_relation possible = ir_get_possible_cmp_relations(left, right);
4084 /* mask out impossible relations */
4085 ir_relation new_relation = relation & possible;
4086 if (new_relation != relation) {
4087 relation = new_relation;
4091 /* Remove unnecessary conversions */
4092 if (!mode_is_float(mode)
4093 || be_get_backend_param()->mode_float_arithmetic == NULL) {
4094 if (is_Conv(left) && is_Conv(right)) {
4095 ir_node *op_left = get_Conv_op(left);
4096 ir_node *op_right = get_Conv_op(right);
4097 ir_mode *mode_left = get_irn_mode(op_left);
4098 ir_mode *mode_right = get_irn_mode(op_right);
4100 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4101 && mode_left != mode_b && mode_right != mode_b) {
4102 ir_node *block = get_nodes_block(n);
4104 if (mode_left == mode_right) {
4108 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4109 } else if (smaller_mode(mode_left, mode_right)) {
4110 left = new_r_Conv(block, op_left, mode_right);
4113 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4114 } else if (smaller_mode(mode_right, mode_left)) {
4116 right = new_r_Conv(block, op_right, mode_left);
4118 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4120 mode = get_irn_mode(left);
4123 if (is_Conv(left) && is_Const(right)) {
4124 ir_node *op_left = get_Conv_op(left);
4125 ir_mode *mode_left = get_irn_mode(op_left);
4126 if (smaller_mode(mode_left, mode) && mode_left != mode_b) {
4127 ir_tarval *tv = get_Const_tarval(right);
4128 tarval_int_overflow_mode_t last_mode
4129 = tarval_get_integer_overflow_mode();
4131 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
4132 new_tv = tarval_convert_to(tv, mode_left);
4133 tarval_set_integer_overflow_mode(last_mode);
4134 if (new_tv != tarval_bad) {
4135 ir_graph *irg = get_irn_irg(n);
4137 right = new_r_Const(irg, new_tv);
4138 mode = get_irn_mode(left);
4140 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4147 * Optimize -a CMP -b into b CMP a.
4148 * This works only for modes where unary Minus cannot Overflow.
4149 * Note that two-complement integers can Overflow so it will NOT work.
4151 if (!mode_overflow_on_unary_Minus(mode) &&
4152 is_Minus(left) && is_Minus(right)) {
4153 left = get_Minus_op(left);
4154 right = get_Minus_op(right);
4155 relation = get_inversed_relation(relation);
4157 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4160 /* remove operation on both sides if possible */
4161 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4163 * The following operations are NOT safe for floating point operations, for instance
4164 * 1.0 + inf == 2.0 + inf, =/=> x == y
4166 if (mode_is_int(mode)) {
4167 unsigned lop = get_irn_opcode(left);
4169 if (lop == get_irn_opcode(right)) {
4170 ir_node *ll, *lr, *rl, *rr;
4172 /* same operation on both sides, try to remove */
4176 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4177 left = get_unop_op(left);
4178 right = get_unop_op(right);
4180 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4183 ll = get_Add_left(left);
4184 lr = get_Add_right(left);
4185 rl = get_Add_left(right);
4186 rr = get_Add_right(right);
4189 /* X + a CMP X + b ==> a CMP b */
4193 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4194 } else if (ll == rr) {
4195 /* X + a CMP b + X ==> a CMP b */
4199 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4200 } else if (lr == rl) {
4201 /* a + X CMP X + b ==> a CMP b */
4205 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4206 } else if (lr == rr) {
4207 /* a + X CMP b + X ==> a CMP b */
4211 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4215 ll = get_Sub_left(left);
4216 lr = get_Sub_right(left);
4217 rl = get_Sub_left(right);
4218 rr = get_Sub_right(right);
4221 /* X - a CMP X - b ==> a CMP b */
4225 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4226 } else if (lr == rr) {
4227 /* a - X CMP b - X ==> a CMP b */
4231 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4235 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4236 /* a ROTL X CMP b ROTL X ==> a CMP b */
4237 left = get_Rotl_left(left);
4238 right = get_Rotl_left(right);
4240 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4248 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4249 if (is_Add(left) || is_Sub(left) || is_Or_Eor_Add(left)) {
4250 ir_node *ll = get_binop_left(left);
4251 ir_node *lr = get_binop_right(left);
4253 if (lr == right && (is_Add(left) || is_Or_Eor_Add(left))) {
4259 ir_graph *irg = get_irn_irg(n);
4261 right = create_zero_const(irg, mode);
4263 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4266 if (is_Add(right) || is_Sub(right) || is_Or_Eor_Add(right)) {
4267 ir_node *rl = get_binop_left(right);
4268 ir_node *rr = get_binop_right(right);
4270 if (rr == left && (is_Add(right) || is_Or_Eor_Add(right))) {
4276 ir_graph *irg = get_irn_irg(n);
4278 right = create_zero_const(irg, mode);
4280 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4284 if (is_And(left) && is_Const(right)) {
4285 ir_node *ll = get_binop_left(left);
4286 ir_node *lr = get_binop_right(left);
4287 if (is_Shr(ll) && is_Const(lr)) {
4288 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4289 ir_node *block = get_nodes_block(n);
4290 ir_mode *mode = get_irn_mode(left);
4292 ir_node *llr = get_Shr_right(ll);
4293 if (is_Const(llr)) {
4294 dbg_info *dbg = get_irn_dbg_info(left);
4295 ir_graph *irg = get_irn_irg(left);
4297 ir_tarval *c1 = get_Const_tarval(llr);
4298 ir_tarval *c2 = get_Const_tarval(lr);
4299 ir_tarval *c3 = get_Const_tarval(right);
4300 ir_tarval *mask = tarval_shl(c2, c1);
4301 ir_tarval *value = tarval_shl(c3, c1);
4303 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4304 right = new_r_Const(irg, value);
4309 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4311 if (is_Const(right) && is_Const_null(right) &&
4312 (is_Eor(left) || is_Or_Eor_Add(left))) {
4313 right = get_Eor_right(left);
4314 left = get_Eor_left(left);
4320 if (mode_is_int(mode) && is_And(left)) {
4321 /* a complicated Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4322 * by the simpler Cmp(And(1bit, val), 0) negated pnc */
4323 if (relation == ir_relation_equal
4324 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4325 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less)) {
4326 ir_node *and0 = get_And_left(left);
4327 ir_node *and1 = get_And_right(left);
4328 if (and1 == right) {
4329 ir_node *tmp = and0;
4333 if (and0 == right && is_single_bit(and0)) {
4334 ir_graph *irg = get_irn_irg(n);
4336 relation == ir_relation_equal ? ir_relation_less_greater
4337 : ir_relation_equal;
4338 right = create_zero_const(irg, mode);
4344 if (is_Const(right) && is_Const_null(right) &&
4345 (relation == ir_relation_equal
4346 || (relation == ir_relation_less_greater)
4347 || (!mode_is_signed(mode) && relation == ir_relation_greater))) {
4349 /* instead of flipping the bit before the bit-test operation negate
4351 ir_node *and0 = get_And_left(left);
4352 ir_node *and1 = get_And_right(left);
4353 if (is_Const(and1)) {
4354 ir_tarval *tv = get_Const_tarval(and1);
4355 if (tarval_is_single_bit(tv)) {
4356 ir_node *flipped = flips_bit(and0, tv);
4357 if (flipped != NULL) {
4358 dbg_info *dbgi = get_irn_dbg_info(left);
4359 ir_node *block = get_nodes_block(left);
4360 relation = get_negated_relation(relation);
4361 left = new_rd_And(dbgi, block, flipped, and1, mode);
4370 /* replace mode_b compares with ands/ors */
4371 if (mode == mode_b) {
4372 ir_node *block = get_nodes_block(n);
4376 case ir_relation_less_equal:
4377 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4379 case ir_relation_less:
4380 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4382 case ir_relation_greater_equal:
4383 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4385 case ir_relation_greater:
4386 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4388 case ir_relation_less_greater:
4389 bres = new_r_Eor(block, left, right, mode_b);
4391 case ir_relation_equal:
4392 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4395 #ifdef DEBUG_libfirm
4396 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4401 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4407 * First step: normalize the compare op
4408 * by placing the constant on the right side
4409 * or moving the lower address node to the left.
4411 if (!operands_are_normalized(left, right)) {
4416 relation = get_inversed_relation(relation);
4421 * Second step: Try to reduce the magnitude
4422 * of a constant. This may help to generate better code
4423 * later and may help to normalize more compares.
4424 * Of course this is only possible for integer values.
4426 tv = value_of(right);
4427 if (tv != tarval_bad) {
4428 ir_mode *mode = get_irn_mode(right);
4430 /* cmp(mux(x, cf, ct), c2) can be eliminated:
4431 * cmp(ct,c2) | cmp(cf,c2) | result
4432 * -----------|------------|--------
4433 * true | true | True
4434 * false | false | False
4436 * false | true | not(x)
4439 ir_node *mux_true = get_Mux_true(left);
4440 ir_node *mux_false = get_Mux_false(left);
4441 if (is_Const(mux_true) && is_Const(mux_false)) {
4442 /* we can fold true/false constant separately */
4443 ir_tarval *tv_true = get_Const_tarval(mux_true);
4444 ir_tarval *tv_false = get_Const_tarval(mux_false);
4445 ir_relation r_true = tarval_cmp(tv_true, tv);
4446 ir_relation r_false = tarval_cmp(tv_false, tv);
4447 if (r_true != ir_relation_false
4448 || r_false != ir_relation_false) {
4449 bool rel_true = (r_true & relation) != 0;
4450 bool rel_false = (r_false & relation) != 0;
4451 ir_node *cond = get_Mux_sel(left);
4452 if (rel_true == rel_false) {
4453 relation = rel_true ? ir_relation_true
4454 : ir_relation_false;
4455 } else if (rel_true) {
4458 dbg_info *dbgi = get_irn_dbg_info(n);
4459 ir_node *block = get_nodes_block(n);
4460 ir_node *notn = new_rd_Not(dbgi, block, cond, mode_b);
4467 /* TODO extend to arbitrary constants */
4468 if (is_Conv(left) && tarval_is_null(tv)) {
4469 ir_node *op = get_Conv_op(left);
4470 ir_mode *op_mode = get_irn_mode(op);
4473 * UpConv(x) REL 0 ==> x REL 0
4474 * Don't do this for float values as it's unclear whether it is a
4475 * win. (on the other side it makes detection/creation of fabs hard)
4477 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4478 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4479 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4480 !mode_is_float(mode)) {
4481 tv = get_mode_null(op_mode);
4485 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4489 if (tv != tarval_bad) {
4490 /* the following optimization is possible on modes without Overflow
4491 * on Unary Minus or on == and !=:
4492 * -a CMP c ==> a swap(CMP) -c
4494 * Beware: for two-complement Overflow may occur, so only == and != can
4495 * be optimized, see this:
4496 * -MININT < 0 =/=> MININT > 0 !!!
4498 if (is_Minus(left) &&
4499 (!mode_overflow_on_unary_Minus(mode) ||
4500 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4501 tv = tarval_neg(tv);
4503 if (tv != tarval_bad) {
4504 left = get_Minus_op(left);
4505 relation = get_inversed_relation(relation);
4507 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4509 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4510 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4511 tv = tarval_not(tv);
4513 if (tv != tarval_bad) {
4514 left = get_Not_op(left);
4516 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4520 /* for integer modes, we have more */
4521 if (mode_is_int(mode) && !is_Const(left)) {
4522 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4523 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4524 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4525 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4527 if (tv != tarval_bad) {
4528 relation ^= ir_relation_equal;
4530 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4533 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4534 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4535 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4536 tv = tarval_add(tv, get_mode_one(mode));
4538 if (tv != tarval_bad) {
4539 relation ^= ir_relation_equal;
4541 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4545 /* the following reassociations work only for == and != */
4546 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4547 if (tv != tarval_bad) {
4548 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4550 ir_node *c1 = get_Sub_right(left);
4551 ir_tarval *tv2 = value_of(c1);
4553 if (tv2 != tarval_bad) {
4554 tv2 = tarval_add(tv, value_of(c1));
4556 if (tv2 != tarval_bad) {
4557 left = get_Sub_left(left);
4560 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4564 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4565 else if (is_Add(left) || is_Or_Eor_Add(left)) {
4566 ir_node *a_l = get_binop_left(left);
4567 ir_node *a_r = get_binop_right(left);
4571 if (is_Const(a_l)) {
4573 tv2 = value_of(a_l);
4576 tv2 = value_of(a_r);
4579 if (tv2 != tarval_bad) {
4580 tv2 = tarval_sub(tv, tv2, NULL);
4582 if (tv2 != tarval_bad) {
4586 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4590 /* -a == c ==> a == -c, -a != c ==> a != -c */
4591 else if (is_Minus(left)) {
4592 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4594 if (tv2 != tarval_bad) {
4595 left = get_Minus_op(left);
4598 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4605 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4606 switch (get_irn_opcode(left)) {
4610 c1 = get_And_right(left);
4613 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4614 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4616 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4618 /* TODO: move to constant evaluation */
4619 ir_graph *irg = get_irn_irg(n);
4620 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4621 c1 = new_r_Const(irg, tv);
4622 DBG_OPT_CSTEVAL(n, c1);
4626 if (tarval_is_single_bit(tv)) {
4628 * optimization for AND:
4630 * And(x, C) == C ==> And(x, C) != 0
4631 * And(x, C) != C ==> And(X, C) == 0
4633 * if C is a single Bit constant.
4636 /* check for Constant's match. We have check hare the tarvals,
4637 because our const might be changed */
4638 if (get_Const_tarval(c1) == tv) {
4639 /* fine: do the transformation */
4640 tv = get_mode_null(get_tarval_mode(tv));
4641 relation ^= ir_relation_less_equal_greater;
4643 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4649 c1 = get_Or_right(left);
4650 if (is_Const(c1) && tarval_is_null(tv)) {
4652 * Or(x, C) == 0 && C != 0 ==> FALSE
4653 * Or(x, C) != 0 && C != 0 ==> TRUE
4655 if (! tarval_is_null(get_Const_tarval(c1))) {
4656 /* TODO: move to constant evaluation */
4657 ir_graph *irg = get_irn_irg(n);
4658 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4659 c1 = new_r_Const(irg, tv);
4660 DBG_OPT_CSTEVAL(n, c1);
4667 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4669 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4672 c1 = get_Shl_right(left);
4674 ir_graph *irg = get_irn_irg(c1);
4675 ir_tarval *tv1 = get_Const_tarval(c1);
4676 ir_mode *mode = get_irn_mode(left);
4677 ir_tarval *minus1 = get_mode_all_one(mode);
4678 ir_tarval *amask = tarval_shr(minus1, tv1);
4679 ir_tarval *cmask = tarval_shl(minus1, tv1);
4682 if (tarval_and(tv, cmask) != tv) {
4683 /* condition not met */
4684 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4685 c1 = new_r_Const(irg, tv);
4686 DBG_OPT_CSTEVAL(n, c1);
4689 sl = get_Shl_left(left);
4690 blk = get_nodes_block(n);
4691 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4692 tv = tarval_shr(tv, tv1);
4694 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4699 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4701 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4704 c1 = get_Shr_right(left);
4706 ir_graph *irg = get_irn_irg(c1);
4707 ir_tarval *tv1 = get_Const_tarval(c1);
4708 ir_mode *mode = get_irn_mode(left);
4709 ir_tarval *minus1 = get_mode_all_one(mode);
4710 ir_tarval *amask = tarval_shl(minus1, tv1);
4711 ir_tarval *cmask = tarval_shr(minus1, tv1);
4714 if (tarval_and(tv, cmask) != tv) {
4715 /* condition not met */
4716 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4717 c1 = new_r_Const(irg, tv);
4718 DBG_OPT_CSTEVAL(n, c1);
4721 sl = get_Shr_left(left);
4722 blk = get_nodes_block(n);
4723 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4724 tv = tarval_shl(tv, tv1);
4726 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4731 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4733 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4736 c1 = get_Shrs_right(left);
4738 ir_graph *irg = get_irn_irg(c1);
4739 ir_tarval *tv1 = get_Const_tarval(c1);
4740 ir_mode *mode = get_irn_mode(left);
4741 ir_tarval *minus1 = get_mode_all_one(mode);
4742 ir_tarval *amask = tarval_shl(minus1, tv1);
4743 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4746 cond = tarval_sub(cond, tv1, NULL);
4747 cond = tarval_shrs(tv, cond);
4749 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4750 /* condition not met */
4751 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4752 c1 = new_r_Const(irg, tv);
4753 DBG_OPT_CSTEVAL(n, c1);
4756 sl = get_Shrs_left(left);
4757 blk = get_nodes_block(n);
4758 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4759 tv = tarval_shl(tv, tv1);
4761 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4769 if (changedc) { /* need a new Const */
4770 ir_graph *irg = get_irn_irg(n);
4771 right = new_r_Const(irg, tv);
4775 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4776 ir_node *op = get_Proj_pred(left);
4778 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4779 ir_node *c = get_binop_right(op);
4782 ir_tarval *tv = get_Const_tarval(c);
4784 if (tarval_is_single_bit(tv)) {
4785 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4786 ir_node *v = get_binop_left(op);
4787 ir_node *blk = get_nodes_block(op);
4788 ir_graph *irg = get_irn_irg(op);
4789 ir_mode *mode = get_irn_mode(v);
4791 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4792 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4794 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4801 dbg_info *dbgi = get_irn_dbg_info(n);
4802 ir_node *block = get_nodes_block(n);
4804 /* create a new compare */
4805 n = new_rd_Cmp(dbgi, block, left, right, relation);
4812 * Optimize CopyB(mem, x, x) into a Nop.
4814 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4816 ir_node *copyb = get_Proj_pred(proj);
4817 ir_node *a = get_CopyB_dst(copyb);
4818 ir_node *b = get_CopyB_src(copyb);
4821 switch (get_Proj_proj(proj)) {
4822 case pn_CopyB_X_regular:
4823 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4824 DBG_OPT_EXC_REM(proj);
4825 proj = new_r_Jmp(get_nodes_block(copyb));
4827 case pn_CopyB_X_except: {
4828 ir_graph *irg = get_irn_irg(proj);
4829 DBG_OPT_EXC_REM(proj);
4830 proj = new_r_Bad(irg, mode_X);
4841 * Optimize Bounds(idx, idx, upper) into idx.
4843 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4845 ir_node *oldn = proj;
4846 ir_node *bound = get_Proj_pred(proj);
4847 ir_node *idx = get_Bound_index(bound);
4848 ir_node *pred = skip_Proj(idx);
4851 if (idx == get_Bound_lower(bound))
4853 else if (is_Bound(pred)) {
4855 * idx was Bounds checked previously, it is still valid if
4856 * lower <= pred_lower && pred_upper <= upper.
4858 ir_node *lower = get_Bound_lower(bound);
4859 ir_node *upper = get_Bound_upper(bound);
4860 if (get_Bound_lower(pred) == lower &&
4861 get_Bound_upper(pred) == upper) {
4863 * One could expect that we simply return the previous
4864 * Bound here. However, this would be wrong, as we could
4865 * add an exception Proj to a new location then.
4866 * So, we must turn in into a tuple.
4872 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4873 switch (get_Proj_proj(proj)) {
4875 DBG_OPT_EXC_REM(proj);
4876 proj = get_Bound_mem(bound);
4878 case pn_Bound_X_except:
4879 DBG_OPT_EXC_REM(proj);
4880 proj = new_r_Bad(get_irn_irg(proj), mode_X);
4884 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4886 case pn_Bound_X_regular:
4887 DBG_OPT_EXC_REM(proj);
4888 proj = new_r_Jmp(get_nodes_block(bound));
4898 * Does all optimizations on nodes that must be done on its Projs
4899 * because of creating new nodes.
4901 static ir_node *transform_node_Proj(ir_node *proj)
4903 ir_node *n = get_Proj_pred(proj);
4905 if (n->op->ops.transform_node_Proj)
4906 return n->op->ops.transform_node_Proj(proj);
4911 * Test whether a block is unreachable
4912 * Note: That this only returns true when
4913 * IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE is set.
4914 * This is important, as you easily end up producing invalid constructs in the
4915 * unreachable code when optimizing away edges into the unreachable code.
4916 * So only set this flag when you iterate localopts to the fixpoint.
4917 * When you reach the fixpoint then all unreachable code is dead
4918 * (= can't be reached by firm edges) and you won't see the invalid constructs
4921 static bool is_block_unreachable(const ir_node *block)
4923 const ir_graph *irg = get_irn_irg(block);
4924 if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE))
4926 return get_Block_dom_depth(block) < 0;
4929 static ir_node *transform_node_Block(ir_node *block)
4931 ir_graph *irg = get_irn_irg(block);
4932 int arity = get_irn_arity(block);
4933 ir_node *bad = NULL;
4936 if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE))
4939 for (i = 0; i < arity; ++i) {
4940 ir_node *const pred = get_Block_cfgpred(block, i);
4941 if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred)))
4944 bad = new_r_Bad(irg, mode_X);
4945 set_irn_n(block, i, bad);
4951 static ir_node *transform_node_Phi(ir_node *phi)
4953 int n = get_irn_arity(phi);
4954 ir_mode *mode = get_irn_mode(phi);
4955 ir_node *block = get_nodes_block(phi);
4956 ir_graph *irg = get_irn_irg(phi);
4957 ir_node *bad = NULL;
4960 /* Set phi-operands for bad-block inputs to bad */
4961 for (i = 0; i < n; ++i) {
4962 if (!is_Bad(get_Phi_pred(phi, i))) {
4963 ir_node *pred = get_Block_cfgpred(block, i);
4964 if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) {
4966 bad = new_r_Bad(irg, mode);
4967 set_irn_n(phi, i, bad);
4972 /* Move Pin nodes down through Phi nodes. */
4973 if (mode == mode_M) {
4974 n = get_irn_arity(phi);
4976 /* Beware of Phi0 */
4980 bool has_pin = false;
4982 NEW_ARR_A(ir_node *, in, n);
4984 for (i = 0; i < n; ++i) {
4985 ir_node *pred = get_irn_n(phi, i);
4988 in[i] = get_Pin_op(pred);
4990 } else if (is_Bad(pred)) {
5000 /* Move the Pin nodes "behind" the Phi. */
5001 new_phi = new_r_Phi(block, n, in, mode_M);
5002 return new_r_Pin(block, new_phi);
5005 /* Move Confirms down through Phi nodes. */
5006 else if (mode_is_reference(mode)) {
5007 n = get_irn_arity(phi);
5009 /* Beware of Phi0 */
5011 ir_node *pred = get_irn_n(phi, 0);
5012 ir_node *bound, *new_phi, **in;
5013 ir_relation relation;
5014 bool has_confirm = false;
5016 if (! is_Confirm(pred))
5019 bound = get_Confirm_bound(pred);
5020 relation = get_Confirm_relation(pred);
5022 NEW_ARR_A(ir_node *, in, n);
5023 in[0] = get_Confirm_value(pred);
5025 for (i = 1; i < n; ++i) {
5026 pred = get_irn_n(phi, i);
5028 if (is_Confirm(pred) &&
5029 get_Confirm_bound(pred) == bound &&
5030 get_Confirm_relation(pred) == relation) {
5031 in[i] = get_Confirm_value(pred);
5033 } else if (is_Bad(pred)) {
5043 /* move the Confirm nodes "behind" the Phi */
5044 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
5045 return new_r_Confirm(block, new_phi, bound, relation);
5052 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5054 * Should be moved to reassociation?
5056 static ir_node *transform_node_shift(ir_node *n)
5058 ir_node *left, *right;
5060 ir_mode *count_mode;
5061 ir_tarval *tv1, *tv2, *res;
5062 ir_node *in[2], *irn, *block;
5066 left = get_binop_left(n);
5068 /* different operations */
5069 if (get_irn_op(left) != get_irn_op(n))
5072 right = get_binop_right(n);
5073 tv1 = value_of(right);
5074 if (tv1 == tarval_bad)
5077 tv2 = value_of(get_binop_right(left));
5078 if (tv2 == tarval_bad)
5081 count_mode = get_tarval_mode(tv1);
5082 if (get_tarval_mode(tv2) != count_mode) {
5083 /* TODO: search bigger mode or something and convert... */
5087 mode = get_irn_mode(n);
5088 modulo_shf = get_mode_modulo_shift(mode);
5090 if (modulo_shf > 0) {
5091 ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode);
5093 /* I'm not so sure what happens in one complement... */
5094 assert(get_mode_arithmetic(count_mode) == irma_twos_complement);
5095 /* modulo shifts should always be a power of 2 (otherwise modulo_mask
5096 * above will be invalid) */
5097 assert(modulo_shf<=0 || is_po2(modulo_shf));
5099 tv1 = tarval_and(tv1, modulo_mask);
5100 tv2 = tarval_and(tv2, modulo_mask);
5102 res = tarval_add(tv1, tv2);
5103 irg = get_irn_irg(n);
5105 /* beware: a simple replacement works only, if res < modulo shift */
5107 int bits = get_mode_size_bits(mode);
5108 ir_tarval *modulo = new_tarval_from_long(bits, count_mode);
5109 res = tarval_mod(res, modulo);
5111 long bits = get_mode_size_bits(mode);
5112 ir_tarval *mode_size = new_tarval_from_long(bits, count_mode);
5114 /* shifting too much */
5115 if (!(tarval_cmp(res, mode_size) & ir_relation_less)) {
5117 ir_node *block = get_nodes_block(n);
5118 dbg_info *dbgi = get_irn_dbg_info(n);
5119 ir_mode *smode = get_irn_mode(right);
5120 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
5121 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5124 return new_r_Const(irg, get_mode_null(mode));
5128 /* ok, we can replace it */
5129 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5130 block = get_nodes_block(n);
5132 in[0] = get_binop_left(left);
5133 in[1] = new_r_Const(irg, res);
5135 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5137 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5144 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5146 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5147 * (also with x >>s c1 when c1>=c2)
5149 static ir_node *transform_node_shl_shr(ir_node *n)
5152 ir_node *right = get_binop_right(n);
5162 ir_tarval *tv_shift;
5165 ir_relation relation;
5168 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5170 if (!is_Const(right))
5173 left = get_binop_left(n);
5174 mode = get_irn_mode(n);
5175 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5176 ir_node *shr_right = get_binop_right(left);
5178 if (!is_Const(shr_right))
5181 x = get_binop_left(left);
5182 tv_shr = get_Const_tarval(shr_right);
5183 tv_shl = get_Const_tarval(right);
5185 if (is_Shrs(left)) {
5186 /* shrs variant only allowed if c1 >= c2 */
5187 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5190 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5193 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5195 tv_mask = tarval_shl(tv_mask, tv_shl);
5196 } else if (is_Shr(n) && is_Shl(left)) {
5197 ir_node *shl_right = get_Shl_right(left);
5199 if (!is_Const(shl_right))
5202 x = get_Shl_left(left);
5203 tv_shr = get_Const_tarval(right);
5204 tv_shl = get_Const_tarval(shl_right);
5206 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5207 tv_mask = tarval_shr(tv_mask, tv_shr);
5212 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5213 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5216 assert(tv_mask != tarval_bad);
5217 assert(get_tarval_mode(tv_mask) == mode);
5219 block = get_nodes_block(n);
5220 irg = get_irn_irg(block);
5221 dbgi = get_irn_dbg_info(n);
5223 relation = tarval_cmp(tv_shl, tv_shr);
5224 if (relation == ir_relation_less || relation == ir_relation_equal) {
5225 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5226 new_const = new_r_Const(irg, tv_shift);
5228 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5230 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5233 assert(relation == ir_relation_greater);
5234 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5235 new_const = new_r_Const(irg, tv_shift);
5236 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5239 new_const = new_r_Const(irg, tv_mask);
5240 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5245 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5247 ir_mode *mode = get_tarval_mode(tv);
5248 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5249 return tarval_mod(tv, modulo_tv);
5252 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5253 ir_node *left, ir_node *right, ir_mode *mode);
5256 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5257 * then we can use that to minimize the value of Add(x, const) or
5258 * Sub(Const, x). In particular this often avoids 1 instruction in some
5259 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5260 * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
5262 static ir_node *transform_node_shift_modulo(ir_node *n,
5263 new_shift_func new_shift)
5265 ir_mode *mode = get_irn_mode(n);
5266 int modulo = get_mode_modulo_shift(mode);
5267 ir_node *newop = NULL;
5268 ir_mode *mode_right;
5275 if (get_mode_arithmetic(mode) != irma_twos_complement)
5277 if (!is_po2(modulo))
5280 irg = get_irn_irg(n);
5281 block = get_nodes_block(n);
5282 right = get_binop_right(n);
5283 mode_right = get_irn_mode(right);
5284 if (is_Const(right)) {
5285 ir_tarval *tv = get_Const_tarval(right);
5286 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5291 newop = new_r_Const(irg, tv_mod);
5292 } else if (is_Add(right) || is_Or_Eor_Add(right)) {
5293 ir_node *add_right = get_binop_right(right);
5294 if (is_Const(add_right)) {
5295 ir_tarval *tv = get_Const_tarval(add_right);
5296 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5301 newconst = new_r_Const(irg, tv_mod);
5302 newop = new_r_Add(block, get_binop_left(right), newconst,
5305 } else if (is_Sub(right)) {
5306 ir_node *sub_left = get_Sub_left(right);
5307 if (is_Const(sub_left)) {
5308 ir_tarval *tv = get_Const_tarval(sub_left);
5309 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5314 newconst = new_r_Const(irg, tv_mod);
5315 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5322 if (newop != NULL) {
5323 dbg_info *dbgi = get_irn_dbg_info(n);
5324 ir_node *left = get_binop_left(n);
5325 return new_shift(dbgi, block, left, newop, mode);
5333 static ir_node *transform_node_Shr(ir_node *n)
5335 ir_node *c, *oldn = n;
5336 ir_node *left = get_Shr_left(n);
5337 ir_node *right = get_Shr_right(n);
5338 ir_mode *mode = get_irn_mode(n);
5340 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5341 n = transform_node_shift(n);
5344 n = transform_node_shift_modulo(n, new_rd_Shr);
5346 n = transform_node_shl_shr(n);
5348 n = transform_node_shift_bitop(n);
5356 static ir_node *transform_node_Shrs(ir_node *n)
5359 ir_node *a = get_Shrs_left(n);
5360 ir_node *b = get_Shrs_right(n);
5361 ir_mode *mode = get_irn_mode(n);
5365 if (is_oversize_shift(n)) {
5366 ir_node *block = get_nodes_block(n);
5367 dbg_info *dbgi = get_irn_dbg_info(n);
5368 ir_mode *cmode = get_irn_mode(b);
5369 long val = get_mode_size_bits(cmode)-1;
5370 ir_graph *irg = get_irn_irg(n);
5371 ir_node *cnst = new_r_Const_long(irg, cmode, val);
5372 return new_rd_Shrs(dbgi, block, a, cnst, mode);
5375 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5376 n = transform_node_shift(n);
5380 n = transform_node_shift_modulo(n, new_rd_Shrs);
5383 n = transform_node_shift_bitop(n);
5387 /* normalisation: use Shr when sign bit is guaranteed to be cleared */
5388 attr = vrp_get_info(a);
5390 unsigned bits = get_mode_size_bits(mode);
5391 ir_tarval *scount = new_tarval_from_long(bits-1, mode_Iu);
5392 ir_tarval *sign = tarval_shl(get_mode_one(mode), scount);
5393 if (tarval_is_null(tarval_and(attr->bits_not_set, sign))) {
5394 dbg_info *dbgi = get_irn_dbg_info(n);
5395 ir_node *block = get_nodes_block(n);
5396 return new_rd_Shr(dbgi, block, a, b, mode);
5406 static ir_node *transform_node_Shl(ir_node *n)
5408 ir_node *c, *oldn = n;
5409 ir_node *a = get_Shl_left(n);
5410 ir_node *b = get_Shl_right(n);
5411 ir_mode *mode = get_irn_mode(n);
5413 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5414 n = transform_node_shift(n);
5417 n = transform_node_shift_modulo(n, new_rd_Shl);
5419 n = transform_node_shl_shr(n);
5421 n = transform_node_shift_bitop(n);
5429 static ir_node *transform_node_Rotl(ir_node *n)
5431 ir_node *c, *oldn = n;
5432 ir_node *a = get_Rotl_left(n);
5433 ir_node *b = get_Rotl_right(n);
5434 ir_mode *mode = get_irn_mode(n);
5436 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5437 n = transform_node_shift(n);
5440 n = transform_node_shift_bitop(n);
5446 * returns mode size for may_leave_out_middle_mode
5448 static unsigned get_significand_size(ir_mode *mode)
5450 const ir_mode_arithmetic arithmetic = get_mode_arithmetic(mode);
5451 switch (arithmetic) {
5453 case irma_x86_extended_float:
5454 return get_mode_mantissa_size(mode) + 1;
5455 case irma_twos_complement:
5456 return get_mode_size_bits(mode);
5458 panic("Conv node with irma_none mode?");
5460 panic("unexpected mode_arithmetic in get_significand_size");
5464 * Returns true if a conversion from mode @p m0 to @p m1 has the same effect
5465 * as converting from @p m0 to @p m1 and then to @p m2.
5466 * Classifying the 3 modes as the big(b), middle(m) and small(s) mode this
5467 * gives the following truth table:
5468 * s -> b -> m : true
5469 * s -> m -> b : !signed(s) || signed(m)
5470 * m -> b -> s : true
5471 * m -> s -> b : false
5472 * b -> s -> m : false
5473 * b -> m -> s : true
5475 * s -> b -> b : true
5476 * s -> s -> b : false
5478 * additional float constraints:
5480 * F -> I -> I: signedness of Is must match
5481 * I -> F -> I: signedness of Is must match
5482 * I -> I -> F: signedness of Is must match
5486 * at least 1 float involved: signedness must match
5488 bool may_leave_out_middle_conv(ir_mode *m0, ir_mode *m1, ir_mode *m2)
5490 int n_floats = mode_is_float(m0) + mode_is_float(m1) + mode_is_float(m2);
5491 if (n_floats == 1) {
5493 int n_signed = mode_is_signed(m0) + mode_is_signed(m1)
5494 + mode_is_signed(m2);
5495 /* we assume that float modes are always signed */
5496 if ((n_signed & 1) != 1)
5499 /* because overflow gives strange results we don't touch this case */
5502 } else if (n_floats == 2 && !mode_is_float(m1)) {
5506 unsigned size0 = get_significand_size(m0);
5507 unsigned size1 = get_significand_size(m1);
5508 unsigned size2 = get_significand_size(m2);
5509 if (size1 < size2 && size0 >= size1)
5513 return !mode_is_signed(m0) || mode_is_signed(m1);
5519 static ir_node *transform_node_Conv(ir_node *n)
5521 ir_node *c, *oldn = n;
5522 ir_mode *mode = get_irn_mode(n);
5523 ir_node *a = get_Conv_op(n);
5526 ir_mode *a_mode = get_irn_mode(a);
5527 ir_node *b = get_Conv_op(a);
5528 ir_mode *b_mode = get_irn_mode(b);
5529 if (may_leave_out_middle_conv(b_mode, a_mode, mode)) {
5530 dbg_info *dbgi = get_irn_dbg_info(n);
5531 ir_node *block = get_nodes_block(n);
5532 return new_rd_Conv(dbgi, block, b, mode);
5536 if (mode != mode_b && is_const_Phi(a)) {
5537 /* Do NOT optimize mode_b Conv's, this leads to remaining
5538 * Phib nodes later, because the conv_b_lower operation
5539 * is instantly reverted, when it tries to insert a Convb.
5541 c = apply_conv_on_phi(a, mode);
5543 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5548 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5549 ir_graph *irg = get_irn_irg(n);
5550 return new_r_Unknown(irg, mode);
5553 if (mode_is_reference(mode) &&
5554 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5556 ir_node *l = get_Add_left(a);
5557 ir_node *r = get_Add_right(a);
5558 dbg_info *dbgi = get_irn_dbg_info(a);
5559 ir_node *block = get_nodes_block(n);
5561 ir_node *lop = get_Conv_op(l);
5562 if (get_irn_mode(lop) == mode) {
5563 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5564 n = new_rd_Add(dbgi, block, lop, r, mode);
5569 ir_node *rop = get_Conv_op(r);
5570 if (get_irn_mode(rop) == mode) {
5571 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5572 n = new_rd_Add(dbgi, block, l, rop, mode);
5582 * Remove dead blocks and nodes in dead blocks
5583 * in keep alive list. We do not generate a new End node.
5585 static ir_node *transform_node_End(ir_node *n)
5587 int i, j, n_keepalives = get_End_n_keepalives(n);
5590 NEW_ARR_A(ir_node *, in, n_keepalives);
5592 for (i = j = 0; i < n_keepalives; ++i) {
5593 ir_node *ka = get_End_keepalive(n, i);
5595 /* no need to keep Bad */
5598 /* do not keep unreachable code */
5599 block = is_Block(ka) ? ka : get_nodes_block(ka);
5600 if (is_block_unreachable(block))
5604 if (j != n_keepalives)
5605 set_End_keepalives(n, j, in);
5609 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5611 if (is_Minus(a) && get_Minus_op(a) == b)
5613 if (is_Minus(b) && get_Minus_op(b) == a)
5615 if (is_Sub(a) && is_Sub(b)) {
5616 ir_node *a_left = get_Sub_left(a);
5617 ir_node *a_right = get_Sub_right(a);
5618 ir_node *b_left = get_Sub_left(b);
5619 ir_node *b_right = get_Sub_right(b);
5621 if (a_left == b_right && a_right == b_left)
5628 static const ir_node *skip_upconv(const ir_node *node)
5630 while (is_Conv(node)) {
5631 ir_mode *mode = get_irn_mode(node);
5632 const ir_node *op = get_Conv_op(node);
5633 ir_mode *op_mode = get_irn_mode(op);
5634 if (!smaller_mode(op_mode, mode))
5641 int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_false,
5642 const ir_node *mux_true)
5647 ir_relation relation;
5653 * Note further that these optimization work even for floating point
5654 * with NaN's because -NaN == NaN.
5655 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5658 mode = get_irn_mode(mux_true);
5659 if (mode_honor_signed_zeros(mode))
5662 /* must be <, <=, >=, > */
5663 relation = get_Cmp_relation(sel);
5664 if ((relation & ir_relation_less_greater) == 0)
5667 if (!ir_is_negated_value(mux_true, mux_false))
5670 mux_true = skip_upconv(mux_true);
5671 mux_false = skip_upconv(mux_false);
5673 /* must be x cmp 0 */
5674 cmp_right = get_Cmp_right(sel);
5675 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
5678 cmp_left = get_Cmp_left(sel);
5679 if (cmp_left == mux_false) {
5680 if (relation & ir_relation_less) {
5683 assert(relation & ir_relation_greater);
5686 } else if (cmp_left == mux_true) {
5687 if (relation & ir_relation_less) {
5690 assert(relation & ir_relation_greater);
5698 ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_false,
5701 ir_node *cmp_left = get_Cmp_left(sel);
5702 return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
5705 bool ir_is_optimizable_mux(const ir_node *sel, const ir_node *mux_false,
5706 const ir_node *mux_true)
5708 /* this code should return true each time transform_node_Mux would
5709 * optimize the Mux completely away */
5711 ir_mode *mode = get_irn_mode(mux_false);
5712 if (get_mode_arithmetic(mode) == irma_twos_complement
5713 && ir_mux_is_abs(sel, mux_false, mux_true))
5716 if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) {
5717 const ir_node *cmp_r = get_Cmp_right(sel);
5718 const ir_node *cmp_l = get_Cmp_left(sel);
5719 const ir_node *f = mux_false;
5720 const ir_node *t = mux_true;
5722 if (is_Const(t) && is_Const_null(t)) {
5727 if (is_And(cmp_l) && f == cmp_r) {
5728 ir_node *and_r = get_And_right(cmp_l);
5731 if (and_r == t && is_single_bit(and_r))
5733 and_l = get_And_left(cmp_l);
5734 if (and_l == t && is_single_bit(and_l))
5743 * Optimize a Mux(c, 0, 1) node (sometimes called a "set" instruction)
5745 static ir_node *transform_Mux_set(ir_node *n)
5747 ir_node *cond = get_Mux_sel(n);
5752 ir_relation relation;
5766 left = get_Cmp_left(cond);
5767 mode = get_irn_mode(left);
5768 if (!mode_is_int(mode) && !mode_is_reference(mode))
5770 dest_mode = get_irn_mode(n);
5771 if (!mode_is_int(dest_mode) && !mode_is_reference(dest_mode))
5773 right = get_Cmp_right(cond);
5774 relation = get_Cmp_relation(cond) & ~ir_relation_unordered;
5775 if (get_mode_size_bits(mode) >= get_mode_size_bits(dest_mode)
5776 && !(mode_is_signed(mode) && is_Const(right) && is_Const_null(right)
5777 && relation != ir_relation_greater))
5782 case ir_relation_less:
5783 /* a < b -> (a - b) >> 31 */
5787 case ir_relation_less_equal:
5788 /* a <= b -> ~(a - b) >> 31 */
5793 case ir_relation_greater:
5794 /* a > b -> (b - a) >> 31 */
5798 case ir_relation_greater_equal:
5799 /* a >= b -> ~(a - b) >> 31 */
5808 dbgi = get_irn_dbg_info(n);
5809 block = get_nodes_block(n);
5810 irg = get_irn_irg(block);
5811 bits = get_mode_size_bits(dest_mode);
5812 tv = new_tarval_from_long(bits-1, mode_Iu);
5813 shift_cnt = new_rd_Const(dbgi, irg, tv);
5815 if (mode != dest_mode) {
5816 a = new_rd_Conv(dbgi, block, a, dest_mode);
5817 b = new_rd_Conv(dbgi, block, b, dest_mode);
5820 res = new_rd_Sub(dbgi, block, a, b, dest_mode);
5822 res = new_rd_Not(dbgi, block, res, dest_mode);
5824 res = new_rd_Shr(dbgi, block, res, shift_cnt, dest_mode);
5829 * Optimize a Mux into some simpler cases.
5831 static ir_node *transform_node_Mux(ir_node *n)
5834 ir_node *sel = get_Mux_sel(n);
5835 ir_mode *mode = get_irn_mode(n);
5836 ir_node *t = get_Mux_true(n);
5837 ir_node *f = get_Mux_false(n);
5838 ir_graph *irg = get_irn_irg(n);
5840 /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
5841 if (get_mode_arithmetic(mode) == irma_twos_complement) {
5842 int abs = ir_mux_is_abs(sel, f, t);
5844 dbg_info *dbgi = get_irn_dbg_info(n);
5845 ir_node *block = get_nodes_block(n);
5846 ir_node *op = ir_get_abs_op(sel, f, t);
5847 int bits = get_mode_size_bits(mode);
5848 ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
5849 ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
5850 ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode);
5853 res = new_rd_Sub(dbgi, block, xorn, sext, mode);
5855 res = new_rd_Sub(dbgi, block, sext, xorn, mode);
5861 /* first normalization step: try to move a constant to the false side,
5862 * 0 preferred on false side too */
5863 if (is_Cmp(sel) && is_Const(t) &&
5864 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5865 dbg_info *seldbgi = get_irn_dbg_info(sel);
5866 ir_node *block = get_nodes_block(sel);
5867 ir_relation relation = get_Cmp_relation(sel);
5872 /* Mux(x, a, b) => Mux(not(x), b, a) */
5873 relation = get_negated_relation(relation);
5874 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5875 get_Cmp_right(sel), relation);
5876 return new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5879 if (is_Const(f) && is_Const_null(f) && is_Const(t) && is_Const_one(t)) {
5880 n = transform_Mux_set(n);
5885 /* the following optimisations create new mode_b nodes, so only do them
5886 * before mode_b lowering */
5887 if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_MODEB_LOWERED)) {
5889 ir_node* block = get_nodes_block(n);
5891 ir_node* c1 = get_Mux_sel(t);
5892 ir_node* t1 = get_Mux_true(t);
5893 ir_node* f1 = get_Mux_false(t);
5895 /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */
5896 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5897 DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE);
5898 return new_r_Mux(block, and_, f1, t1, mode);
5899 } else if (f == t1) {
5900 /* Mux(cond0, Mux(cond1, x, y), x) */
5901 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5902 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5903 DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE);
5904 return new_r_Mux(block, and_, t1, f1, mode);
5906 } else if (is_Mux(f)) {
5907 ir_node* block = get_nodes_block(n);
5909 ir_node* c1 = get_Mux_sel(f);
5910 ir_node* t1 = get_Mux_true(f);
5911 ir_node* f1 = get_Mux_false(f);
5913 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5914 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5915 DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE);
5916 return new_r_Mux(block, or_, f1, t1, mode);
5917 } else if (t == f1) {
5918 /* Mux(cond0, x, Mux(cond1, y, x)) */
5919 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5920 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5921 DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE);
5922 return new_r_Mux(block, or_, t1, f1, mode);
5926 /* note: after normalization, false can only happen on default */
5927 if (mode == mode_b) {
5928 dbg_info *dbg = get_irn_dbg_info(n);
5929 ir_node *block = get_nodes_block(n);
5932 ir_tarval *tv_t = get_Const_tarval(t);
5933 if (tv_t == tarval_b_true) {
5935 /* Muxb(sel, true, false) = sel */
5936 assert(get_Const_tarval(f) == tarval_b_false);
5937 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5940 /* Muxb(sel, true, x) = Or(sel, x) */
5941 n = new_rd_Or(dbg, block, sel, f, mode_b);
5942 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5946 } else if (is_Const(f)) {
5947 ir_tarval *tv_f = get_Const_tarval(f);
5948 if (tv_f == tarval_b_true) {
5949 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5950 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5951 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5952 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5955 /* Muxb(sel, x, false) = And(sel, x) */
5956 assert(tv_f == tarval_b_false);
5957 n = new_rd_And(dbg, block, sel, t, mode_b);
5958 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5965 if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) {
5966 ir_relation relation = get_Cmp_relation(sel);
5967 ir_node *cmp_r = get_Cmp_right(sel);
5968 ir_node *cmp_l = get_Cmp_left(sel);
5969 ir_node *block = get_nodes_block(n);
5971 if (is_And(cmp_l) && f == cmp_r) {
5972 ir_node *and_r = get_And_right(cmp_l);
5975 if (and_r == t && is_single_bit(and_r)) {
5976 if (relation == ir_relation_equal) {
5977 /* Mux((a & (1<<n)) == 0, (1<<n), 0) == (a&(1<<n)) xor ((1<<n)) */
5978 n = new_rd_Eor(get_irn_dbg_info(n),
5979 block, cmp_l, t, mode);
5980 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5982 /* Mux((a & (1<<n)) != 0, (1<<n), 0) == a & (1<<n) */
5984 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5988 and_l = get_And_left(cmp_l);
5989 if (and_l == t && is_single_bit(and_l)) {
5990 if (relation == ir_relation_equal) {
5991 /* ((1 << n) & a) == 0, (1 << n), 0) */
5992 n = new_rd_Eor(get_irn_dbg_info(n),
5993 block, cmp_l, t, mode);
5994 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5996 /* ((1 << n) & a) != 0, (1 << n), 0) */
5998 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
6009 * optimize Sync nodes that have other syncs as input we simply add the inputs
6010 * of the other sync to our own inputs
6012 static ir_node *transform_node_Sync(ir_node *n)
6014 int arity = get_Sync_n_preds(n);
6017 for (i = 0; i < arity;) {
6018 ir_node *pred = get_Sync_pred(n, i);
6022 /* Remove Bad predecessors */
6029 /* Remove duplicate predecessors */
6030 for (j = 0; j < i; ++j) {
6031 if (get_Sync_pred(n, j) == pred) {
6040 if (!is_Sync(pred)) {
6048 pred_arity = get_Sync_n_preds(pred);
6049 for (j = 0; j < pred_arity; ++j) {
6050 ir_node *pred_pred = get_Sync_pred(pred, j);
6055 add_irn_n(n, pred_pred);
6059 if (get_Sync_pred(n, k) == pred_pred)
6066 ir_graph *irg = get_irn_irg(n);
6067 return new_r_Bad(irg, mode_M);
6070 return get_Sync_pred(n, 0);
6073 /* rehash the sync node */
6078 static ir_node *create_load_replacement_tuple(ir_node *n, ir_node *mem,
6081 ir_node *block = get_nodes_block(n);
6082 ir_graph *irg = get_irn_irg(n);
6083 ir_node *in[pn_Load_max+1];
6085 in[pn_Load_M] = mem;
6086 in[pn_Load_res] = res;
6087 if (ir_throws_exception(n)) {
6088 in[pn_Load_X_regular] = new_r_Jmp(block);
6089 in[pn_Load_X_except] = new_r_Bad(irg, mode_X);
6091 assert(pn_Load_max == 4);
6093 ir_node *tuple = new_r_Tuple(block, n_in, in);
6097 static ir_node *transform_node_Load(ir_node *n)
6099 /* don't touch volatile loads */
6100 if (get_Load_volatility(n) == volatility_is_volatile)
6103 ir_node *ptr = get_Load_ptr(n);
6104 const ir_node *confirm;
6105 if (value_not_zero(ptr, &confirm) && confirm == NULL) {
6106 set_irn_pinned(n, op_pin_state_floats);
6109 /* if our memory predecessor is a load from the same address, then reuse the
6110 * previous result */
6111 ir_node *mem = get_Load_mem(n);
6114 ir_node *mem_pred = get_Proj_pred(mem);
6115 if (is_Load(mem_pred)) {
6116 ir_node *pred_load = mem_pred;
6118 /* conservatively compare the 2 loads. TODO: This could be less strict
6119 * with fixup code in some situations (like smaller/bigger modes) */
6120 if (get_Load_ptr(pred_load) != ptr)
6122 if (get_Load_mode(pred_load) != get_Load_mode(n))
6124 /* all combinations of aligned/unaligned pred/n should be fine so we do
6125 * not compare the unaligned attribute */
6126 ir_mode *mode = get_Load_mode(n);
6127 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
6128 return create_load_replacement_tuple(n, mem, res);
6129 } else if (is_Store(mem_pred)) {
6130 ir_node *pred_store = mem_pred;
6131 ir_node *value = get_Store_value(pred_store);
6133 if (get_Store_ptr(pred_store) != ptr)
6135 if (get_irn_mode(value) != get_Load_mode(n))
6137 /* all combinations of aligned/unaligned pred/n should be fine so we do
6138 * not compare the unaligned attribute */
6139 return create_load_replacement_tuple(n, mem, value);
6145 static ir_node *transform_node_Store(ir_node *n)
6147 /* don't touch volatile stores */
6148 if (get_Store_volatility(n) == volatility_is_volatile)
6151 ir_node *ptr = get_Store_ptr(n);
6152 const ir_node *confirm;
6153 if (value_not_zero(ptr, &confirm) && confirm == NULL) {
6154 set_irn_pinned(n, op_pin_state_floats);
6160 * optimize a trampoline Call into a direct Call
6162 static ir_node *transform_node_Call(ir_node *call)
6164 ir_node *callee = get_Call_ptr(call);
6165 ir_node *adr, *mem, *res, *bl, **in;
6166 ir_type *ctp, *mtp, *tp;
6170 size_t i, n_res, n_param;
6173 if (! is_Proj(callee))
6175 callee = get_Proj_pred(callee);
6176 if (! is_Builtin(callee))
6178 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
6181 mem = get_Call_mem(call);
6183 if (skip_Proj(mem) == callee) {
6184 /* memory is routed to the trampoline, skip */
6185 mem = get_Builtin_mem(callee);
6188 /* build a new call type */
6189 mtp = get_Call_type(call);
6190 tdb = get_type_dbg_info(mtp);
6192 n_res = get_method_n_ress(mtp);
6193 n_param = get_method_n_params(mtp);
6194 ctp = new_d_type_method(n_param + 1, n_res, tdb);
6196 for (i = 0; i < n_res; ++i)
6197 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
6199 NEW_ARR_A(ir_node *, in, n_param + 1);
6201 /* FIXME: we don't need a new pointer type in every step */
6202 irg = get_irn_irg(call);
6203 tp = get_irg_frame_type(irg);
6204 tp = new_type_pointer(tp);
6205 set_method_param_type(ctp, 0, tp);
6207 in[0] = get_Builtin_param(callee, 2);
6208 for (i = 0; i < n_param; ++i) {
6209 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
6210 in[i + 1] = get_Call_param(call, i);
6212 var = get_method_variadicity(mtp);
6213 set_method_variadicity(ctp, var);
6214 /* When we resolve a trampoline, the function must be called by a this-call */
6215 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
6216 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
6218 adr = get_Builtin_param(callee, 1);
6220 db = get_irn_dbg_info(call);
6221 bl = get_nodes_block(call);
6223 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
6224 if (get_irn_pinned(call) == op_pin_state_floats)
6225 set_irn_pinned(res, op_pin_state_floats);
6230 * Tries several [inplace] [optimizing] transformations and returns an
6231 * equivalent node. The difference to equivalent_node() is that these
6232 * transformations _do_ generate new nodes, and thus the old node must
6233 * not be freed even if the equivalent node isn't the old one.
6235 static ir_node *transform_node(ir_node *n)
6241 iro = get_irn_opcode_(n);
6242 /* constant expression evaluation / constant folding */
6243 if (get_opt_constant_folding()) {
6244 /* neither constants nor Tuple values can be evaluated */
6245 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6246 /* try to evaluate */
6247 ir_tarval *tv = computed_value(n);
6248 if (tv != tarval_bad) {
6249 /* evaluation was successful -- replace the node. */
6250 ir_graph *irg = get_irn_irg(n);
6252 n = new_r_Const(irg, tv);
6254 DBG_OPT_CSTEVAL(old_n, n);
6260 /* remove unnecessary nodes */
6261 if (get_opt_constant_folding() ||
6262 (iro == iro_Phi) || /* always optimize these nodes. */
6263 (iro == iro_Id) || /* ... */
6264 (iro == iro_Proj) || /* ... */
6265 (iro == iro_Block)) { /* Flags tested local. */
6266 n = equivalent_node(n);
6271 /* Some more constant expression evaluation. */
6272 if (get_opt_algebraic_simplification() ||
6273 (iro == iro_Cond) ||
6274 (iro == iro_Proj)) { /* Flags tested local. */
6275 if (n->op->ops.transform_node != NULL) {
6276 n = n->op->ops.transform_node(n);
6286 static void register_computed_value_func(ir_op *op, computed_value_func func)
6288 assert(op->ops.computed_value == NULL || op->ops.computed_value == func);
6289 op->ops.computed_value = func;
6292 static void register_computed_value_func_proj(ir_op *op,
6293 computed_value_func func)
6295 assert(op->ops.computed_value_Proj == NULL
6296 || op->ops.computed_value_Proj == func);
6297 op->ops.computed_value_Proj = func;
6300 static void register_equivalent_node_func(ir_op *op, equivalent_node_func func)
6302 assert(op->ops.equivalent_node == NULL || op->ops.equivalent_node == func);
6303 op->ops.equivalent_node = func;
6306 static void register_equivalent_node_func_proj(ir_op *op,
6307 equivalent_node_func func)
6309 assert(op->ops.equivalent_node_Proj == NULL
6310 || op->ops.equivalent_node_Proj == func);
6311 op->ops.equivalent_node_Proj = func;
6314 static void register_transform_node_func(ir_op *op, transform_node_func func)
6316 assert(op->ops.transform_node == NULL || op->ops.transform_node == func);
6317 op->ops.transform_node = func;
6320 static void register_transform_node_func_proj(ir_op *op,
6321 transform_node_func func)
6323 assert(op->ops.transform_node_Proj == NULL
6324 || op->ops.transform_node_Proj == func);
6325 op->ops.transform_node_Proj = func;
6328 void ir_register_opt_node_ops(void)
6330 register_computed_value_func(op_Add, computed_value_Add);
6331 register_computed_value_func(op_And, computed_value_And);
6332 register_computed_value_func(op_Borrow, computed_value_Borrow);
6333 register_computed_value_func(op_Carry, computed_value_Carry);
6334 register_computed_value_func(op_Cmp, computed_value_Cmp);
6335 register_computed_value_func(op_Confirm, computed_value_Confirm);
6336 register_computed_value_func(op_Const, computed_value_Const);
6337 register_computed_value_func(op_Conv, computed_value_Conv);
6338 register_computed_value_func(op_Eor, computed_value_Eor);
6339 register_computed_value_func(op_Minus, computed_value_Minus);
6340 register_computed_value_func(op_Mul, computed_value_Mul);
6341 register_computed_value_func(op_Mux, computed_value_Mux);
6342 register_computed_value_func(op_Not, computed_value_Not);
6343 register_computed_value_func(op_Or, computed_value_Or);
6344 register_computed_value_func(op_Proj, computed_value_Proj);
6345 register_computed_value_func(op_Rotl, computed_value_Rotl);
6346 register_computed_value_func(op_Shl, computed_value_Shl);
6347 register_computed_value_func(op_Shr, computed_value_Shr);
6348 register_computed_value_func(op_Shrs, computed_value_Shrs);
6349 register_computed_value_func(op_Sub, computed_value_Sub);
6350 register_computed_value_func(op_SymConst, computed_value_SymConst);
6351 register_computed_value_func_proj(op_Div, computed_value_Proj_Div);
6352 register_computed_value_func_proj(op_Mod, computed_value_Proj_Mod);
6354 register_equivalent_node_func(op_Add, equivalent_node_Add);
6355 register_equivalent_node_func(op_And, equivalent_node_And);
6356 register_equivalent_node_func(op_Confirm, equivalent_node_Confirm);
6357 register_equivalent_node_func(op_Conv, equivalent_node_Conv);
6358 register_equivalent_node_func(op_Eor, equivalent_node_Eor);
6359 register_equivalent_node_func(op_Id, equivalent_node_Id);
6360 register_equivalent_node_func(op_Minus, equivalent_node_involution);
6361 register_equivalent_node_func(op_Mul, equivalent_node_Mul);
6362 register_equivalent_node_func(op_Mux, equivalent_node_Mux);
6363 register_equivalent_node_func(op_Not, equivalent_node_involution);
6364 register_equivalent_node_func(op_Or, equivalent_node_Or);
6365 register_equivalent_node_func(op_Phi, equivalent_node_Phi);
6366 register_equivalent_node_func(op_Proj, equivalent_node_Proj);
6367 register_equivalent_node_func(op_Rotl, equivalent_node_left_zero);
6368 register_equivalent_node_func(op_Shl, equivalent_node_left_zero);
6369 register_equivalent_node_func(op_Shr, equivalent_node_left_zero);
6370 register_equivalent_node_func(op_Shrs, equivalent_node_left_zero);
6371 register_equivalent_node_func(op_Sub, equivalent_node_Sub);
6372 register_equivalent_node_func_proj(op_Bound, equivalent_node_Proj_Bound);
6373 register_equivalent_node_func_proj(op_CopyB, equivalent_node_Proj_CopyB);
6374 register_equivalent_node_func_proj(op_Div, equivalent_node_Proj_Div);
6375 register_equivalent_node_func_proj(op_Tuple, equivalent_node_Proj_Tuple);
6377 register_transform_node_func(op_Add, transform_node_Add);
6378 register_transform_node_func(op_And, transform_node_And);
6379 register_transform_node_func(op_Block, transform_node_Block);
6380 register_transform_node_func(op_Call, transform_node_Call);
6381 register_transform_node_func(op_Cmp, transform_node_Cmp);
6382 register_transform_node_func(op_Cond, transform_node_Cond);
6383 register_transform_node_func(op_Conv, transform_node_Conv);
6384 register_transform_node_func(op_Div, transform_node_Div);
6385 register_transform_node_func(op_End, transform_node_End);
6386 register_transform_node_func(op_Eor, transform_node_Eor);
6387 register_transform_node_func(op_Load, transform_node_Load);
6388 register_transform_node_func(op_Minus, transform_node_Minus);
6389 register_transform_node_func(op_Mod, transform_node_Mod);
6390 register_transform_node_func(op_Mul, transform_node_Mul);
6391 register_transform_node_func(op_Mux, transform_node_Mux);
6392 register_transform_node_func(op_Not, transform_node_Not);
6393 register_transform_node_func(op_Or, transform_node_Or);
6394 register_transform_node_func(op_Phi, transform_node_Phi);
6395 register_transform_node_func(op_Proj, transform_node_Proj);
6396 register_transform_node_func(op_Rotl, transform_node_Rotl);
6397 register_transform_node_func(op_Shl, transform_node_Shl);
6398 register_transform_node_func(op_Shrs, transform_node_Shrs);
6399 register_transform_node_func(op_Shr, transform_node_Shr);
6400 register_transform_node_func(op_Store, transform_node_Store);
6401 register_transform_node_func(op_Sub, transform_node_Sub);
6402 register_transform_node_func(op_Switch, transform_node_Switch);
6403 register_transform_node_func(op_Sync, transform_node_Sync);
6404 register_transform_node_func_proj(op_Bound, transform_node_Proj_Bound);
6405 register_transform_node_func_proj(op_CopyB, transform_node_Proj_CopyB);
6406 register_transform_node_func_proj(op_Div, transform_node_Proj_Div);
6407 register_transform_node_func_proj(op_Load, transform_node_Proj_Load);
6408 register_transform_node_func_proj(op_Mod, transform_node_Proj_Mod);
6409 register_transform_node_func_proj(op_Store, transform_node_Proj_Store);
6412 /* **************** Common Subexpression Elimination **************** */
6414 /** The size of the hash table used, should estimate the number of nodes
6416 #define N_IR_NODES 512
6418 int identities_cmp(const void *elt, const void *key)
6420 ir_node *a = (ir_node *)elt;
6421 ir_node *b = (ir_node *)key;
6424 if (a == b) return 0;
6426 if ((get_irn_op(a) != get_irn_op(b)) ||
6427 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6429 /* compare if a's in and b's in are of equal length */
6430 irn_arity_a = get_irn_arity(a);
6431 if (irn_arity_a != get_irn_arity(b))
6434 /* blocks are never the same */
6438 if (get_irn_pinned(a) == op_pin_state_pinned) {
6439 /* for pinned nodes, the block inputs must be equal */
6440 if (get_nodes_block(a) != get_nodes_block(b))
6443 ir_node *block_a = get_nodes_block(a);
6444 ir_node *block_b = get_nodes_block(b);
6445 if (! get_opt_global_cse()) {
6446 /* for block-local CSE both nodes must be in the same Block */
6447 if (block_a != block_b)
6450 /* The optimistic approach would be to do nothing here.
6451 * However doing GCSE optimistically produces a lot of partially dead code which appears
6452 * to be worse in practice than the missed opportunities.
6453 * So we use a very conservative variant here and only CSE if 1 value dominates the
6455 if (!block_dominates(block_a, block_b)
6456 && !block_dominates(block_b, block_a))
6458 /* respect the workaround rule: do not move nodes which are only
6459 * held by keepalive edges */
6460 if (only_used_by_keepalive(a) || only_used_by_keepalive(b))
6465 /* compare a->in[0..ins] with b->in[0..ins] */
6466 for (i = 0; i < irn_arity_a; ++i) {
6467 ir_node *pred_a = get_irn_n(a, i);
6468 ir_node *pred_b = get_irn_n(b, i);
6469 if (pred_a != pred_b) {
6470 /* if both predecessors are CSE neutral they might be different */
6471 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6477 * here, we already now that the nodes are identical except their
6480 if (a->op->ops.node_cmp_attr)
6481 return a->op->ops.node_cmp_attr(a, b);
6486 unsigned ir_node_hash(const ir_node *node)
6488 return node->op->ops.hash(node);
6491 void new_identities(ir_graph *irg)
6493 if (irg->value_table != NULL)
6494 del_pset(irg->value_table);
6495 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6498 void del_identities(ir_graph *irg)
6500 if (irg->value_table != NULL)
6501 del_pset(irg->value_table);
6504 static int cmp_node_nr(const void *a, const void *b)
6506 ir_node **p1 = (ir_node**)a;
6507 ir_node **p2 = (ir_node**)b;
6508 long n1 = get_irn_node_nr(*p1);
6509 long n2 = get_irn_node_nr(*p2);
6510 return (n1>n2) - (n1<n2);
6513 void ir_normalize_node(ir_node *n)
6515 if (is_op_commutative(get_irn_op(n))) {
6516 ir_node *l = get_binop_left(n);
6517 ir_node *r = get_binop_right(n);
6519 /* For commutative operators perform a OP b == b OP a but keep
6520 * constants on the RIGHT side. This helps greatly in some
6521 * optimizations. Moreover we use the idx number to make the form
6523 if (!operands_are_normalized(l, r)) {
6524 set_binop_left(n, r);
6525 set_binop_right(n, l);
6528 } else if (is_Sync(n)) {
6529 /* we assume that most of the time the inputs of a Sync node are already
6530 * sorted, so check this first as a shortcut */
6531 bool ins_sorted = true;
6532 int arity = get_irn_arity(n);
6533 const ir_node *last = get_irn_n(n, 0);
6535 for (i = 1; i < arity; ++i) {
6536 const ir_node *node = get_irn_n(n, i);
6537 if (get_irn_node_nr(node) < get_irn_node_nr(last)) {
6545 ir_node **ins = get_irn_in(n)+1;
6546 ir_node **new_ins = XMALLOCN(ir_node*, arity);
6547 memcpy(new_ins, ins, arity*sizeof(ins[0]));
6548 qsort(new_ins, arity, sizeof(new_ins[0]), cmp_node_nr);
6549 set_irn_in(n, arity, new_ins);
6555 ir_node *identify_remember(ir_node *n)
6557 ir_graph *irg = get_irn_irg(n);
6558 pset *value_table = irg->value_table;
6561 if (value_table == NULL)
6564 ir_normalize_node(n);
6565 /* lookup or insert in hash table with given hash key. */
6566 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6569 /* n is reachable again */
6570 edges_node_revival(nn);
6577 * During construction we set the op_pin_state_pinned flag in the graph right
6578 * when the optimization is performed. The flag turning on procedure global
6579 * cse could be changed between two allocations. This way we are safe.
6581 * @param n The node to lookup
6583 static inline ir_node *identify_cons(ir_node *n)
6587 n = identify_remember(n);
6588 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6589 ir_graph *irg = get_irn_irg(n);
6590 set_irg_pinned(irg, op_pin_state_floats);
6595 void add_identities(ir_node *node)
6602 identify_remember(node);
6605 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6607 ir_graph *rem = current_ir_graph;
6609 current_ir_graph = irg;
6610 foreach_pset(irg->value_table, ir_node, node) {
6613 current_ir_graph = rem;
6616 ir_node *optimize_node(ir_node *n)
6619 ir_graph *irg = get_irn_irg(n);
6620 unsigned iro = get_irn_opcode(n);
6623 /* Always optimize Phi nodes: part of the construction. */
6624 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6626 /* constant expression evaluation / constant folding */
6627 if (get_opt_constant_folding()) {
6628 /* neither constants nor Tuple values can be evaluated */
6629 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6630 /* try to evaluate */
6631 tv = computed_value(n);
6632 if (tv != tarval_bad) {
6637 * we MUST copy the node here temporarily, because it's still
6638 * needed for DBG_OPT_CSTEVAL
6640 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6641 oldn = (ir_node*)alloca(node_size);
6643 memcpy(oldn, n, node_size);
6644 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6646 /* ARG, copy the in array, we need it for statistics */
6647 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6649 /* note the inplace edges module */
6650 edges_node_deleted(n);
6652 /* evaluation was successful -- replace the node. */
6653 irg_kill_node(irg, n);
6654 nw = new_r_Const(irg, tv);
6656 DBG_OPT_CSTEVAL(oldn, nw);
6662 /* remove unnecessary nodes */
6663 if (get_opt_algebraic_simplification() ||
6664 (iro == iro_Phi) || /* always optimize these nodes. */
6666 (iro == iro_Proj) ||
6667 (iro == iro_Block) ) /* Flags tested local. */
6668 n = equivalent_node(n);
6670 /* Common Subexpression Elimination.
6672 * Checks whether n is already available.
6673 * The block input is used to distinguish different subexpressions. Right
6674 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6675 * subexpressions within a block.
6678 n = identify_cons(n);
6681 edges_node_deleted(oldn);
6683 /* We found an existing, better node, so we can deallocate the old node. */
6684 irg_kill_node(irg, oldn);
6688 /* Some more constant expression evaluation that does not allow to
6690 iro = get_irn_opcode(n);
6691 if (get_opt_algebraic_simplification() ||
6692 (iro == iro_Cond) ||
6693 (iro == iro_Proj)) { /* Flags tested local. */
6694 n = transform_node(n);
6697 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6698 if (get_opt_cse()) {
6700 n = identify_remember(o);
6708 ir_node *optimize_in_place_2(ir_node *n)
6710 if (!get_opt_optimize() && !is_Phi(n)) return n;
6715 /** common subexpression elimination **/
6716 /* Checks whether n is already available. */
6717 /* The block input is used to distinguish different subexpressions.
6718 * Right now all nodes are op_pin_state_pinned to blocks, i.e., the cse
6719 * only finds common subexpressions within a block. */
6720 if (get_opt_cse()) {
6722 n = identify_remember(n);
6725 /* we have another existing node now, we do not optimize it here */
6730 n = transform_node(n);
6732 /* Now we can verify the node, as it has no dead inputs any more. */
6735 /* Now we have a legal, useful node. Enter it in hash table for cse.
6737 * Note: This is only necessary because some of the optimisations
6738 * operate in-place (set_XXX_bla, turn_into_tuple, ...) which is considered
6739 * bad practice and should be fixed sometime.
6741 if (get_opt_cse()) {
6743 n = identify_remember(o);
6751 ir_node *optimize_in_place(ir_node *n)
6753 ir_graph *irg = get_irn_irg(n);
6755 if (get_opt_global_cse())
6756 set_irg_pinned(irg, op_pin_state_floats);
6758 /* FIXME: Maybe we could also test whether optimizing the node can
6759 change the control graph. */
6760 clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
6761 return optimize_in_place_2(n);