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 /* Make types visible to allow most efficient access */
57 static bool is_Or_Eor_Add(const ir_node *node)
59 if (is_Or(node) || is_Eor(node) || is_Add(node)) {
60 ir_node *left = get_binop_left(node);
61 ir_node *right = get_binop_right(node);
62 vrp_attr *vrp_left = vrp_get_info(left);
63 vrp_attr *vrp_right = vrp_get_info(right);
64 if (vrp_left != NULL && vrp_right != NULL) {
66 = tarval_and(vrp_left->bits_not_set, vrp_right->bits_not_set);
67 return tarval_is_null(vrp_val);
74 * Returns the tarval of a Const node or tarval_bad for all other nodes.
76 static ir_tarval *default_value_of(const ir_node *n)
79 return get_Const_tarval(n); /* might return tarval_bad */
84 value_of_func value_of_ptr = default_value_of;
86 /* * Set a new value_of function. */
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 * Return the value of a Cmp.
622 * The basic idea here is to determine which relations are possible and which
623 * one are definitely impossible.
625 static ir_tarval *computed_value_Cmp(const ir_node *cmp)
627 /* we can't construct Constb after lowering mode_b nodes */
628 if (is_irg_state(get_irn_irg(cmp), IR_GRAPH_STATE_MODEB_LOWERED))
631 return compute_cmp(cmp);
635 * Calculate the value of an integer Div.
636 * Special case: 0 / b
638 static ir_tarval *do_computed_value_Div(const ir_node *div)
640 const ir_node *a = get_Div_left(div);
641 const ir_node *b = get_Div_right(div);
642 const ir_mode *mode = get_Div_resmode(div);
643 ir_tarval *ta = value_of(a);
645 const ir_node *dummy;
647 /* cannot optimize 0 / b = 0 because of NaN */
648 if (!mode_is_float(mode)) {
649 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
650 return ta; /* 0 / b == 0 if b != 0 */
653 if (ta != tarval_bad && tb != tarval_bad)
654 return tarval_div(ta, tb);
659 * Calculate the value of an integer Mod of two nodes.
660 * Special case: a % 1
662 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
664 ir_tarval *ta = value_of(a);
665 ir_tarval *tb = value_of(b);
667 /* Compute a % 1 or c1 % c2 */
668 if (tarval_is_one(tb))
669 return get_mode_null(get_irn_mode(a));
670 if (ta != tarval_bad && tb != tarval_bad)
671 return tarval_mod(ta, tb);
676 * Return the value of a Proj(Div).
678 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
680 long proj_nr = get_Proj_proj(n);
681 if (proj_nr != pn_Div_res)
684 return do_computed_value_Div(get_Proj_pred(n));
688 * Return the value of a Proj(Mod).
690 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
692 long proj_nr = get_Proj_proj(n);
694 if (proj_nr == pn_Mod_res) {
695 const ir_node *mod = get_Proj_pred(n);
696 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
702 * Return the value of a Proj.
704 static ir_tarval *computed_value_Proj(const ir_node *proj)
706 ir_node *n = get_Proj_pred(proj);
708 if (n->op->ops.computed_value_Proj != NULL)
709 return n->op->ops.computed_value_Proj(proj);
714 * If the parameter n can be computed, return its value, else tarval_bad.
715 * Performs constant folding.
717 * @param n The node this should be evaluated
719 ir_tarval *computed_value(const ir_node *n)
721 vrp_attr *vrp = vrp_get_info(n);
722 if (vrp != NULL && vrp->bits_set == vrp->bits_not_set)
723 return vrp->bits_set;
725 if (n->op->ops.computed_value)
726 return n->op->ops.computed_value(n);
730 void firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
734 ops->computed_value = computed_value_##a; \
736 #define CASE_PROJ(a) \
738 ops->computed_value_Proj = computed_value_Proj_##a; \
774 * Optimize operations that are commutative and have neutral 0,
775 * so a op 0 = 0 op a = a.
777 static ir_node *equivalent_node_neutral_zero(ir_node *n)
781 ir_node *a = get_binop_left(n);
782 ir_node *b = get_binop_right(n);
787 /* After running compute_node there is only one constant predecessor.
788 Find this predecessors value and remember the other node: */
789 if ((tv = value_of(a)) != tarval_bad) {
791 } else if ((tv = value_of(b)) != tarval_bad) {
796 /* If this predecessors constant value is zero, the operation is
797 * unnecessary. Remove it.
799 * Beware: If n is a Add, the mode of on and n might be different
800 * which happens in this rare construction: NULL + 3.
801 * Then, a Conv would be needed which we cannot include here.
803 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
806 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
813 * Eor is commutative and has neutral 0.
815 static ir_node *equivalent_node_Eor(ir_node *n)
821 n = equivalent_node_neutral_zero(n);
822 if (n != oldn) return n;
825 b = get_Eor_right(n);
827 if (is_Eor(a) || is_Or_Eor_Add(a)) {
828 ir_node *aa = get_binop_left(a);
829 ir_node *ab = get_binop_right(a);
832 /* (a ^ b) ^ a -> b */
834 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
836 } else if (ab == b) {
837 /* (a ^ b) ^ b -> a */
839 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
843 if (is_Eor(b) || is_Or_Eor_Add(b)) {
844 ir_node *ba = get_binop_left(b);
845 ir_node *bb = get_binop_right(b);
848 /* a ^ (a ^ b) -> b */
850 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
852 } else if (bb == a) {
853 /* a ^ (b ^ a) -> b */
855 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
863 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
865 * The second one looks strange, but this construct
866 * is used heavily in the LCC sources :-).
868 * Beware: The Mode of an Add may be different than the mode of its
869 * predecessors, so we could not return a predecessors in all cases.
871 static ir_node *equivalent_node_Add(ir_node *n)
874 ir_node *left, *right;
875 ir_mode *mode = get_irn_mode(n);
877 n = equivalent_node_neutral_zero(n);
881 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
882 if (mode_is_float(mode)) {
883 ir_graph *irg = get_irn_irg(n);
884 if (get_irg_fp_model(irg) & fp_strict_algebraic)
888 left = get_Add_left(n);
889 right = get_Add_right(n);
892 if (get_Sub_right(left) == right) {
895 n = get_Sub_left(left);
896 if (mode == get_irn_mode(n)) {
897 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
903 if (get_Sub_right(right) == left) {
906 n = get_Sub_left(right);
907 if (mode == get_irn_mode(n)) {
908 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
917 * optimize operations that are not commutative but have neutral 0 on left,
920 static ir_node *equivalent_node_left_zero(ir_node *n)
924 ir_node *a = get_binop_left(n);
925 ir_node *b = get_binop_right(n);
926 ir_tarval *tb = value_of(b);
928 if (tarval_is_null(tb)) {
931 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
936 #define equivalent_node_Shl equivalent_node_left_zero
937 #define equivalent_node_Shr equivalent_node_left_zero
938 #define equivalent_node_Shrs equivalent_node_left_zero
939 #define equivalent_node_Rotl equivalent_node_left_zero
942 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
944 * The second one looks strange, but this construct
945 * is used heavily in the LCC sources :-).
947 * Beware: The Mode of a Sub may be different than the mode of its
948 * predecessors, so we could not return a predecessors in all cases.
950 static ir_node *equivalent_node_Sub(ir_node *n)
954 ir_mode *mode = get_irn_mode(n);
957 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
958 if (mode_is_float(mode)) {
959 ir_graph *irg = get_irn_irg(n);
960 if (get_irg_fp_model(irg) & fp_strict_algebraic)
964 b = get_Sub_right(n);
967 /* Beware: modes might be different */
968 if (tarval_is_null(tb)) {
969 ir_node *a = get_Sub_left(n);
970 if (mode == get_irn_mode(a)) {
973 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
981 * Optimize an "self-inverse unary op", i.e. op(op(n)) = n.
984 * -(-a) == a, but might overflow two times.
985 * We handle it anyway here but the better way would be a
986 * flag. This would be needed for Pascal for instance.
988 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
991 ir_node *pred = get_unop_op(n);
993 /* optimize symmetric unop */
994 if (get_irn_op(pred) == get_irn_op(n)) {
995 n = get_unop_op(pred);
996 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1001 /** Optimize Not(Not(x)) == x. */
1002 #define equivalent_node_Not equivalent_node_idempotent_unop
1004 /** -(-x) == x ??? Is this possible or can --x raise an
1005 out of bounds exception if min =! max? */
1006 #define equivalent_node_Minus equivalent_node_idempotent_unop
1009 * Optimize a * 1 = 1 * a = a.
1011 static ir_node *equivalent_node_Mul(ir_node *n)
1014 ir_node *a = get_Mul_left(n);
1016 /* we can handle here only the n * n = n bit cases */
1017 if (get_irn_mode(n) == get_irn_mode(a)) {
1018 ir_node *b = get_Mul_right(n);
1022 * Mul is commutative and has again an other neutral element.
1023 * Constants are place right, so check this case first.
1026 if (tarval_is_one(tv)) {
1028 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1031 if (tarval_is_one(tv)) {
1033 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1041 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1043 static ir_node *equivalent_node_Or(ir_node *n)
1047 ir_node *a = get_Or_left(n);
1048 ir_node *b = get_Or_right(n);
1052 n = a; /* idempotence */
1053 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1056 /* constants are normalized to right, check this side first */
1058 if (tarval_is_null(tv)) {
1060 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1064 if (tarval_is_null(tv)) {
1066 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1074 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1076 static ir_node *equivalent_node_And(ir_node *n)
1080 ir_node *a = get_And_left(n);
1081 ir_node *b = get_And_right(n);
1085 n = a; /* idempotence */
1086 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1089 /* constants are normalized to right, check this side first */
1091 if (tarval_is_all_one(tv)) {
1093 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1096 if (tv != get_tarval_bad()) {
1097 ir_mode *mode = get_irn_mode(n);
1098 if (!mode_is_signed(mode) && is_Conv(a)) {
1099 ir_node *convop = get_Conv_op(a);
1100 ir_mode *convopmode = get_irn_mode(convop);
1101 if (!mode_is_signed(convopmode)) {
1102 /* Check Conv(all_one) & Const = all_one */
1103 ir_tarval *one = get_mode_all_one(convopmode);
1104 ir_tarval *conv = tarval_convert_to(one, mode);
1105 ir_tarval *and = tarval_and(conv, tv);
1107 if (tarval_is_all_one(and)) {
1108 /* Conv(X) & Const = X */
1110 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1117 if (tarval_is_all_one(tv)) {
1119 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1123 if ((is_Or(a) || is_Or_Eor_Add(a))
1124 && (b == get_binop_left(a) || b == get_binop_right(a))) {
1126 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1130 if ((is_Or(b) || is_Or_Eor_Add(b))
1131 && (a == get_binop_left(b) || a == get_binop_right(b))) {
1133 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1140 * Try to remove useless Conv's:
1142 static ir_node *equivalent_node_Conv(ir_node *n)
1145 ir_node *a = get_Conv_op(n);
1147 ir_mode *n_mode = get_irn_mode(n);
1148 ir_mode *a_mode = get_irn_mode(a);
1151 if (n_mode == a_mode) { /* No Conv necessary */
1152 if (get_Conv_strict(n)) {
1155 /* neither Minus nor Confirm change the precision,
1156 so we can "look-through" */
1159 p = get_Minus_op(p);
1160 } else if (is_Confirm(p)) {
1161 p = get_Confirm_value(p);
1167 if (is_Conv(p) && get_Conv_strict(p)) {
1168 /* we known already, that a_mode == n_mode, and neither
1169 Minus change the mode, so the second Conv
1171 assert(get_irn_mode(p) == n_mode);
1173 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1177 ir_node *pred = get_Proj_pred(p);
1178 if (is_Load(pred)) {
1179 /* Loads always return with the exact precision of n_mode */
1180 assert(get_Load_mode(pred) == n_mode);
1182 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1185 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1186 pred = get_Proj_pred(pred);
1187 if (is_Start(pred)) {
1188 /* Arguments always return with the exact precision,
1189 as strictConv's are place before Call -- if the
1190 caller was compiled with the same setting.
1191 Otherwise, the semantics is probably still right. */
1192 assert(get_irn_mode(p) == n_mode);
1194 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1200 /* special case: the immediate predecessor is also a Conv */
1201 if (! get_Conv_strict(a)) {
1202 /* first one is not strict, kick it */
1204 a_mode = get_irn_mode(a);
1208 /* else both are strict conv, second is superfluous */
1210 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1215 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1218 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1219 ir_node *b = get_Conv_op(a);
1220 ir_mode *b_mode = get_irn_mode(b);
1222 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1223 /* both are strict conv */
1224 if (smaller_mode(a_mode, n_mode)) {
1225 /* both are strict, but the first is smaller, so
1226 the second cannot remove more precision, remove the
1228 set_Conv_strict(n, 0);
1231 if (n_mode == b_mode) {
1232 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1233 if (n_mode == mode_b) {
1234 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1235 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1237 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1238 if (values_in_mode(b_mode, a_mode)) {
1239 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1240 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1245 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1246 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1247 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1248 unsigned float_mantissa = get_mode_mantissa_size(a_mode);
1250 if (float_mantissa >= int_mantissa) {
1252 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1257 if (smaller_mode(b_mode, a_mode)) {
1258 if (get_Conv_strict(n))
1259 set_Conv_strict(b, 1);
1260 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1261 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1271 * - fold Phi-nodes, iff they have only one predecessor except
1274 static ir_node *equivalent_node_Phi(ir_node *n)
1279 ir_node *first_val = NULL; /* to shutup gcc */
1281 if (!get_opt_optimize() &&
1282 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1285 n_preds = get_Phi_n_preds(n);
1287 /* Phi of dead Region without predecessors. */
1291 /* Find first non-self-referencing input */
1292 for (i = 0; i < n_preds; ++i) {
1293 first_val = get_Phi_pred(n, i);
1294 /* not self pointer */
1295 if (first_val != n) {
1296 /* then found first value. */
1301 /* search for rest of inputs, determine if any of these
1302 are non-self-referencing */
1303 while (++i < n_preds) {
1304 ir_node *scnd_val = get_Phi_pred(n, i);
1305 if (scnd_val != n && scnd_val != first_val) {
1310 if (i >= n_preds && !is_Dummy(first_val)) {
1311 /* Fold, if no multiple distinct non-self-referencing inputs */
1313 DBG_OPT_PHI(oldn, n);
1319 * Optimize Proj(Tuple).
1321 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1323 ir_node *oldn = proj;
1324 ir_node *tuple = get_Proj_pred(proj);
1326 /* Remove the Tuple/Proj combination. */
1327 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1328 DBG_OPT_TUPLE(oldn, tuple, proj);
1334 * Optimize a / 1 = a.
1336 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1338 ir_node *oldn = proj;
1339 ir_node *div = get_Proj_pred(proj);
1340 ir_node *b = get_Div_right(div);
1341 ir_tarval *tb = value_of(b);
1343 /* Div is not commutative. */
1344 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1345 switch (get_Proj_proj(proj)) {
1347 proj = get_Div_mem(div);
1348 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1352 proj = get_Div_left(div);
1353 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1357 /* we cannot replace the exception Proj's here, this is done in
1358 transform_node_Proj_Div() */
1366 * Optimize CopyB(mem, x, x) into a Nop.
1368 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1370 ir_node *oldn = proj;
1371 ir_node *copyb = get_Proj_pred(proj);
1372 ir_node *a = get_CopyB_dst(copyb);
1373 ir_node *b = get_CopyB_src(copyb);
1376 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1377 switch (get_Proj_proj(proj)) {
1379 proj = get_CopyB_mem(copyb);
1380 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1388 * Optimize Bounds(idx, idx, upper) into idx.
1390 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1392 ir_node *oldn = proj;
1393 ir_node *bound = get_Proj_pred(proj);
1394 ir_node *idx = get_Bound_index(bound);
1395 ir_node *pred = skip_Proj(idx);
1398 if (idx == get_Bound_lower(bound))
1400 else if (is_Bound(pred)) {
1402 * idx was Bounds checked previously, it is still valid if
1403 * lower <= pred_lower && pred_upper <= upper.
1405 ir_node *lower = get_Bound_lower(bound);
1406 ir_node *upper = get_Bound_upper(bound);
1407 if (get_Bound_lower(pred) == lower &&
1408 get_Bound_upper(pred) == upper) {
1410 * One could expect that we simply return the previous
1411 * Bound here. However, this would be wrong, as we could
1412 * add an exception Proj to a new location then.
1413 * So, we must turn in into a tuple.
1419 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1420 switch (get_Proj_proj(proj)) {
1422 DBG_OPT_EXC_REM(proj);
1423 proj = get_Bound_mem(bound);
1427 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1430 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1438 * Does all optimizations on nodes that must be done on its Projs
1439 * because of creating new nodes.
1441 static ir_node *equivalent_node_Proj(ir_node *proj)
1443 ir_node *n = get_Proj_pred(proj);
1444 if (n->op->ops.equivalent_node_Proj)
1445 return n->op->ops.equivalent_node_Proj(proj);
1452 static ir_node *equivalent_node_Id(ir_node *n)
1460 DBG_OPT_ID(oldn, n);
1467 static ir_node *equivalent_node_Mux(ir_node *n)
1469 ir_node *oldn = n, *sel = get_Mux_sel(n);
1471 ir_tarval *ts = value_of(sel);
1473 if (ts == tarval_bad && is_Cmp(sel)) {
1474 /* try again with a direct call to compute_cmp, as we don't care
1475 * about the MODEB_LOWERED flag here */
1476 ts = compute_cmp(sel);
1479 /* Mux(true, f, t) == t */
1480 if (ts == tarval_b_true) {
1481 n = get_Mux_true(n);
1482 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1485 /* Mux(false, f, t) == f */
1486 if (ts == tarval_b_false) {
1487 n = get_Mux_false(n);
1488 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1491 n_t = get_Mux_true(n);
1492 n_f = get_Mux_false(n);
1494 /* Mux(v, x, T) == x */
1495 if (is_Unknown(n_f)) {
1497 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1500 /* Mux(v, T, x) == x */
1501 if (is_Unknown(n_t)) {
1503 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1507 /* Mux(v, x, x) == x */
1510 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1513 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1514 ir_relation relation = get_Cmp_relation(sel);
1515 ir_node *f = get_Mux_false(n);
1516 ir_node *t = get_Mux_true(n);
1519 * Note further that these optimization work even for floating point
1520 * with NaN's because -NaN == NaN.
1521 * However, if +0 and -0 is handled differently, we cannot use the first one.
1523 ir_node *const cmp_l = get_Cmp_left(sel);
1524 ir_node *const cmp_r = get_Cmp_right(sel);
1527 case ir_relation_equal:
1528 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1529 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1531 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1536 case ir_relation_less_greater:
1537 case ir_relation_unordered_less_greater:
1538 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1539 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1541 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1550 * Note: normalization puts the constant on the right side,
1551 * so we check only one case.
1553 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1554 /* Mux(t CMP 0, X, t) */
1555 if (is_Minus(f) && get_Minus_op(f) == t) {
1556 /* Mux(t CMP 0, -t, t) */
1557 if (relation == ir_relation_equal) {
1558 /* Mux(t == 0, -t, t) ==> -t */
1560 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1561 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1562 /* Mux(t != 0, -t, t) ==> t */
1564 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1574 * Remove Confirm nodes if setting is on.
1575 * Replace Confirms(x, '=', Constlike) by Constlike.
1577 static ir_node *equivalent_node_Confirm(ir_node *n)
1579 ir_node *pred = get_Confirm_value(n);
1580 ir_relation relation = get_Confirm_relation(n);
1582 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1584 * rare case: two identical Confirms one after another,
1585 * replace the second one with the first.
1588 pred = get_Confirm_value(n);
1594 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1595 * perform no actual computation, as, e.g., the Id nodes. It does not create
1596 * new nodes. It is therefore safe to free n if the node returned is not n.
1597 * If a node returns a Tuple we can not just skip it. If the size of the
1598 * in array fits, we transform n into a tuple (e.g., Div).
1600 ir_node *equivalent_node(ir_node *n)
1602 if (n->op->ops.equivalent_node)
1603 return n->op->ops.equivalent_node(n);
1607 void firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1611 ops->equivalent_node = equivalent_node_##a; \
1613 #define CASE_PROJ(a) \
1615 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1650 * Returns non-zero if a node is a Phi node
1651 * with all predecessors constant.
1653 static int is_const_Phi(ir_node *n)
1657 if (! is_Phi(n) || get_irn_arity(n) == 0)
1659 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1660 if (! is_Const(get_irn_n(n, i)))
1666 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1667 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1670 * in reality eval_func should be tarval (*eval_func)() but incomplete
1671 * declarations are bad style and generate noisy warnings
1673 typedef void (*eval_func)(void);
1676 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1678 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1680 if (eval == (eval_func) tarval_sub) {
1681 tarval_sub_type func = (tarval_sub_type)eval;
1683 return func(a, b, mode);
1685 tarval_binop_type func = (tarval_binop_type)eval;
1692 * Apply an evaluator on a binop with a constant operators (and one Phi).
1694 * @param phi the Phi node
1695 * @param other the other operand
1696 * @param eval an evaluator function
1697 * @param mode the mode of the result, may be different from the mode of the Phi!
1698 * @param left if non-zero, other is the left operand, else the right
1700 * @return a new Phi node if the conversion was successful, NULL else
1702 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1708 int i, n = get_irn_arity(phi);
1710 NEW_ARR_A(void *, res, n);
1712 for (i = 0; i < n; ++i) {
1713 pred = get_irn_n(phi, i);
1714 tv = get_Const_tarval(pred);
1715 tv = do_eval(eval, other, tv, mode);
1717 if (tv == tarval_bad) {
1718 /* folding failed, bad */
1724 for (i = 0; i < n; ++i) {
1725 pred = get_irn_n(phi, i);
1726 tv = get_Const_tarval(pred);
1727 tv = do_eval(eval, tv, other, mode);
1729 if (tv == tarval_bad) {
1730 /* folding failed, bad */
1736 irg = get_irn_irg(phi);
1737 for (i = 0; i < n; ++i) {
1738 pred = get_irn_n(phi, i);
1739 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1741 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1745 * Apply an evaluator on a binop with two constant Phi.
1747 * @param a the left Phi node
1748 * @param b the right Phi node
1749 * @param eval an evaluator function
1750 * @param mode the mode of the result, may be different from the mode of the Phi!
1752 * @return a new Phi node if the conversion was successful, NULL else
1754 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1756 ir_tarval *tv_l, *tv_r, *tv;
1762 if (get_nodes_block(a) != get_nodes_block(b))
1765 n = get_irn_arity(a);
1766 NEW_ARR_A(void *, res, n);
1768 for (i = 0; i < n; ++i) {
1769 pred = get_irn_n(a, i);
1770 tv_l = get_Const_tarval(pred);
1771 pred = get_irn_n(b, i);
1772 tv_r = get_Const_tarval(pred);
1773 tv = do_eval(eval, tv_l, tv_r, mode);
1775 if (tv == tarval_bad) {
1776 /* folding failed, bad */
1781 irg = get_irn_irg(a);
1782 for (i = 0; i < n; ++i) {
1783 pred = get_irn_n(a, i);
1784 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1786 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1790 * Apply an evaluator on a unop with a constant operator (a Phi).
1792 * @param phi the Phi node
1793 * @param eval an evaluator function
1795 * @return a new Phi node if the conversion was successful, NULL else
1797 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1804 int i, n = get_irn_arity(phi);
1806 NEW_ARR_A(void *, res, n);
1807 for (i = 0; i < n; ++i) {
1808 pred = get_irn_n(phi, i);
1809 tv = get_Const_tarval(pred);
1812 if (tv == tarval_bad) {
1813 /* folding failed, bad */
1818 mode = get_irn_mode(phi);
1819 irg = get_irn_irg(phi);
1820 for (i = 0; i < n; ++i) {
1821 pred = get_irn_n(phi, i);
1822 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1824 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1828 * Apply a conversion on a constant operator (a Phi).
1830 * @param phi the Phi node
1832 * @return a new Phi node if the conversion was successful, NULL else
1834 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1840 int i, n = get_irn_arity(phi);
1842 NEW_ARR_A(void *, res, n);
1843 for (i = 0; i < n; ++i) {
1844 pred = get_irn_n(phi, i);
1845 tv = get_Const_tarval(pred);
1846 tv = tarval_convert_to(tv, mode);
1848 if (tv == tarval_bad) {
1849 /* folding failed, bad */
1854 irg = get_irn_irg(phi);
1855 for (i = 0; i < n; ++i) {
1856 pred = get_irn_n(phi, i);
1857 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1859 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1863 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1864 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1865 * If possible, remove the Conv's.
1867 static ir_node *transform_node_AddSub(ir_node *n)
1869 ir_mode *mode = get_irn_mode(n);
1871 if (mode_is_reference(mode)) {
1872 ir_node *left = get_binop_left(n);
1873 ir_node *right = get_binop_right(n);
1874 unsigned ref_bits = get_mode_size_bits(mode);
1876 if (is_Conv(left)) {
1877 ir_mode *lmode = get_irn_mode(left);
1878 unsigned bits = get_mode_size_bits(lmode);
1880 if (ref_bits == bits &&
1881 mode_is_int(lmode) &&
1882 get_mode_arithmetic(lmode) == irma_twos_complement) {
1883 ir_node *pre = get_Conv_op(left);
1884 ir_mode *pre_mode = get_irn_mode(pre);
1886 if (mode_is_int(pre_mode) &&
1887 get_mode_size_bits(pre_mode) == bits &&
1888 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1889 /* ok, this conv just changes to sign, moreover the calculation
1890 * is done with same number of bits as our address mode, so
1891 * we can ignore the conv as address calculation can be viewed
1892 * as either signed or unsigned
1894 set_binop_left(n, pre);
1899 if (is_Conv(right)) {
1900 ir_mode *rmode = get_irn_mode(right);
1901 unsigned bits = get_mode_size_bits(rmode);
1903 if (ref_bits == bits &&
1904 mode_is_int(rmode) &&
1905 get_mode_arithmetic(rmode) == irma_twos_complement) {
1906 ir_node *pre = get_Conv_op(right);
1907 ir_mode *pre_mode = get_irn_mode(pre);
1909 if (mode_is_int(pre_mode) &&
1910 get_mode_size_bits(pre_mode) == bits &&
1911 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1912 /* ok, this conv just changes to sign, moreover the calculation
1913 * is done with same number of bits as our address mode, so
1914 * we can ignore the conv as address calculation can be viewed
1915 * as either signed or unsigned
1917 set_binop_right(n, pre);
1922 /* let address arithmetic use unsigned modes */
1923 if (is_Const(right)) {
1924 ir_mode *rmode = get_irn_mode(right);
1926 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1927 /* convert a AddP(P, *s) into AddP(P, *u) */
1928 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1930 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
1931 set_binop_right(n, pre);
1939 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1942 if (is_Const(b) && is_const_Phi(a)) { \
1943 /* check for Op(Phi, Const) */ \
1944 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1946 else if (is_Const(a) && is_const_Phi(b)) { \
1947 /* check for Op(Const, Phi) */ \
1948 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1950 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1951 /* check for Op(Phi, Phi) */ \
1952 c = apply_binop_on_2_phis(a, b, eval, mode); \
1955 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1960 #define HANDLE_UNOP_PHI(eval, a, c) \
1963 if (is_const_Phi(a)) { \
1964 /* check for Op(Phi) */ \
1965 c = apply_unop_on_phi(a, eval); \
1967 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1974 * Create a 0 constant of given mode.
1976 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
1978 ir_tarval *tv = get_mode_null(mode);
1979 ir_node *cnst = new_r_Const(irg, tv);
1984 static bool is_shiftop(const ir_node *n)
1986 return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
1989 /* the order of the values is important! */
1990 typedef enum const_class {
1996 static const_class classify_const(const ir_node* n)
1998 if (is_Const(n)) return const_const;
1999 if (is_irn_constlike(n)) return const_like;
2004 * Determines whether r is more constlike or has a larger index (in that order)
2007 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
2009 const const_class l_order = classify_const(l);
2010 const const_class r_order = classify_const(r);
2012 l_order > r_order ||
2013 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
2016 static bool is_cmp_unequal(const ir_node *node)
2018 ir_relation relation = get_Cmp_relation(node);
2019 ir_node *left = get_Cmp_left(node);
2020 ir_node *right = get_Cmp_right(node);
2021 ir_mode *mode = get_irn_mode(left);
2023 if (relation == ir_relation_less_greater)
2026 if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right))
2027 return relation == ir_relation_greater;
2032 * returns true for Cmp(x == 0) or Cmp(x != 0)
2034 static bool is_cmp_equality_zero(const ir_node *node)
2036 ir_relation relation;
2037 ir_node *right = get_Cmp_right(node);
2039 if (!is_Const(right) || !is_Const_null(right))
2041 relation = get_Cmp_relation(node);
2042 return relation == ir_relation_equal
2043 || relation == ir_relation_less_greater
2044 || (!mode_is_signed(get_irn_mode(right))
2045 && relation == ir_relation_greater);
2049 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2050 * Such pattern may arise in bitfield stores.
2052 * value c4 value c4 & c2
2053 * AND c3 AND c1 | c3
2060 * AND c1 ===> OR if (c1 | c2) == 0x111..11
2063 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
2065 ir_node *irn_and, *c1;
2067 ir_node *and_l, *c3;
2068 ir_node *value, *c4;
2069 ir_node *new_and, *new_const, *block;
2070 ir_mode *mode = get_irn_mode(irn_or);
2072 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
2076 irn_and = get_binop_left(irn_or);
2077 c1 = get_binop_right(irn_or);
2078 if (!is_Const(c1) || !is_And(irn_and))
2081 or_l = get_binop_left(irn_and);
2082 c2 = get_binop_right(irn_and);
2086 tv1 = get_Const_tarval(c1);
2087 tv2 = get_Const_tarval(c2);
2089 tv = tarval_or(tv1, tv2);
2090 if (tarval_is_all_one(tv)) {
2091 /* the AND does NOT clear a bit with isn't set by the OR */
2092 set_binop_left(irn_or, or_l);
2093 set_binop_right(irn_or, c1);
2095 /* check for more */
2099 if (!is_Or(or_l) && !is_Or_Eor_Add(or_l))
2102 and_l = get_binop_left(or_l);
2103 c3 = get_binop_right(or_l);
2104 if (!is_Const(c3) || !is_And(and_l))
2107 value = get_binop_left(and_l);
2108 c4 = get_binop_right(and_l);
2112 /* ok, found the pattern, check for conditions */
2113 assert(mode == get_irn_mode(irn_and));
2114 assert(mode == get_irn_mode(or_l));
2115 assert(mode == get_irn_mode(and_l));
2117 tv3 = get_Const_tarval(c3);
2118 tv4 = get_Const_tarval(c4);
2120 tv = tarval_or(tv4, tv2);
2121 if (!tarval_is_all_one(tv)) {
2122 /* have at least one 0 at the same bit position */
2126 if (tv3 != tarval_andnot(tv3, tv4)) {
2127 /* bit in the or_mask is outside the and_mask */
2131 if (tv1 != tarval_andnot(tv1, tv2)) {
2132 /* bit in the or_mask is outside the and_mask */
2136 /* ok, all conditions met */
2137 block = get_irn_n(irn_or, -1);
2138 irg = get_irn_irg(block);
2140 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
2142 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
2144 set_binop_left(irn_or, new_and);
2145 set_binop_right(irn_or, new_const);
2147 /* check for more */
2152 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
2154 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
2156 ir_mode *mode = get_irn_mode(irn_or);
2157 ir_node *shl, *shr, *block;
2158 ir_node *irn, *x, *c1, *c2, *n;
2159 ir_tarval *tv1, *tv2;
2161 /* some backends can't handle rotl */
2162 if (!be_get_backend_param()->support_rotl)
2165 if (! mode_is_int(mode))
2168 shl = get_binop_left(irn_or);
2169 shr = get_binop_right(irn_or);
2178 } else if (!is_Shl(shl)) {
2180 } else if (!is_Shr(shr)) {
2183 x = get_Shl_left(shl);
2184 if (x != get_Shr_left(shr))
2187 c1 = get_Shl_right(shl);
2188 c2 = get_Shr_right(shr);
2189 if (is_Const(c1) && is_Const(c2)) {
2190 tv1 = get_Const_tarval(c1);
2191 if (! tarval_is_long(tv1))
2194 tv2 = get_Const_tarval(c2);
2195 if (! tarval_is_long(tv2))
2198 if (get_tarval_long(tv1) + get_tarval_long(tv2)
2199 != (int) get_mode_size_bits(mode))
2202 /* yet, condition met */
2203 block = get_nodes_block(irn_or);
2205 n = new_r_Rotl(block, x, c1, mode);
2207 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
2211 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
2212 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
2213 if (!ir_is_negated_value(c1, c2)) {
2217 /* yet, condition met */
2218 block = get_nodes_block(irn_or);
2219 n = new_r_Rotl(block, x, c1, mode);
2220 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
2225 * Prototype of a recursive transform function
2226 * for bitwise distributive transformations.
2228 typedef ir_node* (*recursive_transform)(ir_node *n);
2231 * makes use of distributive laws for and, or, eor
2232 * and(a OP c, b OP c) -> and(a, b) OP c
2233 * note, might return a different op than n
2235 static ir_node *transform_bitwise_distributive(ir_node *n,
2236 recursive_transform trans_func)
2239 ir_node *a = get_binop_left(n);
2240 ir_node *b = get_binop_right(n);
2241 ir_op *op = get_irn_op(a);
2242 ir_op *op_root = get_irn_op(n);
2244 if (op != get_irn_op(b))
2247 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2248 if (op == op_Conv) {
2249 ir_node *a_op = get_Conv_op(a);
2250 ir_node *b_op = get_Conv_op(b);
2251 ir_mode *a_mode = get_irn_mode(a_op);
2252 ir_mode *b_mode = get_irn_mode(b_op);
2253 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2254 ir_node *blk = get_nodes_block(n);
2257 set_binop_left(n, a_op);
2258 set_binop_right(n, b_op);
2259 set_irn_mode(n, a_mode);
2261 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2263 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2269 /* nothing to gain here */
2273 if (op == op_Shrs || op == op_Shr || op == op_Shl
2274 || op == op_And || op == op_Or || op == op_Eor) {
2275 ir_node *a_left = get_binop_left(a);
2276 ir_node *a_right = get_binop_right(a);
2277 ir_node *b_left = get_binop_left(b);
2278 ir_node *b_right = get_binop_right(b);
2280 ir_node *op1 = NULL;
2281 ir_node *op2 = NULL;
2283 if (is_op_commutative(op)) {
2284 if (a_left == b_left) {
2288 } else if (a_left == b_right) {
2292 } else if (a_right == b_left) {
2298 if (a_right == b_right) {
2305 /* (a sop c) & (b sop c) => (a & b) sop c */
2306 ir_node *blk = get_nodes_block(n);
2308 ir_node *new_n = exact_copy(n);
2309 set_binop_left(new_n, op1);
2310 set_binop_right(new_n, op2);
2311 new_n = trans_func(new_n);
2313 if (op_root == op_Eor && op == op_Or) {
2314 dbg_info *dbgi = get_irn_dbg_info(n);
2315 ir_mode *mode = get_irn_mode(c);
2317 c = new_rd_Not(dbgi, blk, c, mode);
2318 n = new_rd_And(dbgi, blk, new_n, c, mode);
2321 set_nodes_block(n, blk);
2322 set_binop_left(n, new_n);
2323 set_binop_right(n, c);
2327 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2336 * normalisation: (x >> c1) & c2 to (x & (c2<<c1)) >> c1
2338 * - and, or, xor instead of &
2339 * - Shl, Shr, Shrs, rotl instead of >>
2340 * (with a special case for Or/Xor + Shrs)
2342 * This normalisation is usually good for the backend since << C can often be
2343 * matched as address-mode.
2345 static ir_node *transform_node_bitop_shift(ir_node *n)
2347 ir_graph *irg = get_irn_irg(n);
2348 ir_node *left = get_binop_left(n);
2349 ir_node *right = get_binop_right(n);
2350 ir_mode *mode = get_irn_mode(n);
2351 ir_node *shift_left;
2352 ir_node *shift_right;
2354 dbg_info *dbg_bitop;
2355 dbg_info *dbg_shift;
2361 ir_tarval *tv_bitop;
2363 if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
2366 assert(is_And(n) || is_Or(n) || is_Eor(n) || is_Or_Eor_Add(n));
2367 if (!is_Const(right) || !is_shiftop(left))
2370 shift_left = get_binop_left(left);
2371 shift_right = get_binop_right(left);
2372 if (!is_Const(shift_right))
2375 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
2376 if (is_Shrs(left)) {
2377 /* TODO this could be improved */
2381 irg = get_irn_irg(n);
2382 block = get_nodes_block(n);
2383 dbg_bitop = get_irn_dbg_info(n);
2384 dbg_shift = get_irn_dbg_info(left);
2385 tv1 = get_Const_tarval(shift_right);
2386 tv2 = get_Const_tarval(right);
2387 assert(get_tarval_mode(tv2) == mode);
2390 tv_bitop = tarval_shr(tv2, tv1);
2392 /* Check whether we have lost some bits during the right shift. */
2394 ir_tarval *tv_back_again = tarval_shl(tv_bitop, tv1);
2396 if (tarval_cmp(tv_back_again, tv2) != ir_relation_equal)
2399 } else if (is_Shr(left)) {
2402 * TODO this can be improved by checking whether
2403 * the left shift produces an overflow
2407 tv_bitop = tarval_shl(tv2, tv1);
2409 assert(is_Rotl(left));
2410 tv_bitop = tarval_rotl(tv2, tarval_neg(tv1));
2412 new_const = new_r_Const(irg, tv_bitop);
2415 new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode);
2416 } else if (is_Or(n) || is_Or_Eor_Add(n)) {
2417 new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode);
2420 new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode);
2424 new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode);
2425 } else if (is_Shr(left)) {
2426 new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode);
2428 assert(is_Rotl(left));
2429 new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode);
2435 static bool complement_values(const ir_node *a, const ir_node *b)
2437 if (is_Not(a) && get_Not_op(a) == b)
2439 if (is_Not(b) && get_Not_op(b) == a)
2441 if (is_Const(a) && is_Const(b)) {
2442 ir_tarval *tv_a = get_Const_tarval(a);
2443 ir_tarval *tv_b = get_Const_tarval(b);
2444 return tarval_not(tv_a) == tv_b;
2449 typedef ir_tarval *(tv_fold_binop_func)(ir_tarval *a, ir_tarval *b);
2452 * for associative operations fold:
2453 * op(op(x, c0), c1) to op(x, op(c0, c1)) with constants folded.
2454 * This is a "light" version of the reassociation phase
2456 static ir_node *fold_constant_associativity(ir_node *node,
2457 tv_fold_binop_func fold)
2462 ir_node *right = get_binop_right(node);
2463 ir_node *left_right;
2470 if (!is_Const(right))
2473 op = get_irn_op(node);
2474 left = get_binop_left(node);
2475 if (get_irn_op(left) != op)
2478 left_right = get_binop_right(left);
2479 if (!is_Const(left_right))
2482 left_left = get_binop_left(left);
2483 c0 = get_Const_tarval(left_right);
2484 c1 = get_Const_tarval(right);
2485 irg = get_irn_irg(node);
2486 if (get_tarval_mode(c0) != get_tarval_mode(c1))
2488 new_c = fold(c0, c1);
2489 if (new_c == tarval_bad)
2491 new_const = new_r_Const(irg, new_c);
2492 new_node = exact_copy(node);
2493 set_binop_left(new_node, left_left);
2494 set_binop_right(new_node, new_const);
2501 static ir_node *transform_node_Or_(ir_node *n)
2504 ir_node *a = get_binop_left(n);
2505 ir_node *b = get_binop_right(n);
2509 n = fold_constant_associativity(n, tarval_or);
2513 if (is_Not(a) && is_Not(b)) {
2514 /* ~a | ~b = ~(a&b) */
2515 ir_node *block = get_nodes_block(n);
2517 mode = get_irn_mode(n);
2520 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
2521 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
2522 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
2526 /* we can combine the relations of two compares with the same operands */
2527 if (is_Cmp(a) && is_Cmp(b)) {
2528 ir_node *a_left = get_Cmp_left(a);
2529 ir_node *a_right = get_Cmp_right(a);
2530 ir_node *b_left = get_Cmp_left(b);
2531 ir_node *b_right = get_Cmp_right(b);
2532 if (a_left == b_left && b_left == b_right) {
2533 dbg_info *dbgi = get_irn_dbg_info(n);
2534 ir_node *block = get_nodes_block(n);
2535 ir_relation a_relation = get_Cmp_relation(a);
2536 ir_relation b_relation = get_Cmp_relation(b);
2537 ir_relation new_relation = a_relation | b_relation;
2538 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
2540 /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */
2541 if (is_cmp_unequal(a) && is_cmp_unequal(b)
2542 && !mode_is_float(get_irn_mode(a_left))
2543 && !mode_is_float(get_irn_mode(b_left))) {
2544 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
2545 ir_graph *irg = get_irn_irg(n);
2546 dbg_info *dbgi = get_irn_dbg_info(n);
2547 ir_node *block = get_nodes_block(n);
2548 ir_mode *a_mode = get_irn_mode(a_left);
2549 ir_mode *b_mode = get_irn_mode(b_left);
2550 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
2551 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
2552 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
2553 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
2554 ir_node *zero = create_zero_const(irg, b_mode);
2555 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
2557 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
2558 ir_graph *irg = get_irn_irg(n);
2559 dbg_info *dbgi = get_irn_dbg_info(n);
2560 ir_node *block = get_nodes_block(n);
2561 ir_mode *a_mode = get_irn_mode(a_left);
2562 ir_mode *b_mode = get_irn_mode(b_left);
2563 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
2564 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
2565 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
2566 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
2567 ir_node *zero = create_zero_const(irg, a_mode);
2568 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
2573 mode = get_irn_mode(n);
2574 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
2576 n = transform_node_Or_bf_store(n);
2579 n = transform_node_Or_Rotl(n);
2583 n = transform_bitwise_distributive(n, transform_node_Or_);
2586 n = transform_node_bitop_shift(n);
2593 static ir_node *transform_node_Or(ir_node *n)
2595 if (is_Or_Eor_Add(n)) {
2596 dbg_info *dbgi = get_irn_dbg_info(n);
2597 ir_node *block = get_nodes_block(n);
2598 ir_node *left = get_Or_left(n);
2599 ir_node *right = get_Or_right(n);
2600 ir_mode *mode = get_irn_mode(n);
2601 return new_rd_Add(dbgi, block, left, right, mode);
2603 return transform_node_Or_(n);
2609 static ir_node *transform_node_Eor_(ir_node *n)
2612 ir_node *a = get_binop_left(n);
2613 ir_node *b = get_binop_right(n);
2614 ir_mode *mode = get_irn_mode(n);
2617 n = fold_constant_associativity(n, tarval_eor);
2621 /* we can combine the relations of two compares with the same operands */
2622 if (is_Cmp(a) && is_Cmp(b)) {
2623 ir_node *a_left = get_Cmp_left(a);
2624 ir_node *a_right = get_Cmp_left(a);
2625 ir_node *b_left = get_Cmp_left(b);
2626 ir_node *b_right = get_Cmp_right(b);
2627 if (a_left == b_left && b_left == b_right) {
2628 dbg_info *dbgi = get_irn_dbg_info(n);
2629 ir_node *block = get_nodes_block(n);
2630 ir_relation a_relation = get_Cmp_relation(a);
2631 ir_relation b_relation = get_Cmp_relation(b);
2632 ir_relation new_relation = a_relation ^ b_relation;
2633 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
2637 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
2639 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
2640 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
2641 dbg_info *dbg = get_irn_dbg_info(n);
2642 ir_node *block = get_nodes_block(n);
2643 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
2644 ir_node *new_left = get_Not_op(a);
2645 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
2646 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2648 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
2649 dbg_info *dbg = get_irn_dbg_info(n);
2650 ir_node *block = get_nodes_block(n);
2651 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
2652 ir_node *new_right = get_Not_op(b);
2653 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
2654 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2658 /* x ^ 1...1 -> ~1 */
2659 if (is_Const(b) && is_Const_all_one(b)) {
2660 n = new_r_Not(get_nodes_block(n), a, mode);
2661 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2665 n = transform_bitwise_distributive(n, transform_node_Eor_);
2668 n = transform_node_bitop_shift(n);
2675 static ir_node *transform_node_Eor(ir_node *n)
2677 if (is_Or_Eor_Add(n)) {
2678 dbg_info *dbgi = get_irn_dbg_info(n);
2679 ir_node *block = get_nodes_block(n);
2680 ir_node *left = get_Eor_left(n);
2681 ir_node *right = get_Eor_right(n);
2682 ir_mode *mode = get_irn_mode(n);
2683 return new_rd_Add(dbgi, block, left, right, mode);
2685 return transform_node_Eor_(n);
2689 * Do the AddSub optimization, then Transform
2690 * Constant folding on Phi
2691 * Add(a,a) -> Mul(a, 2)
2692 * Add(Mul(a, x), a) -> Mul(a, x+1)
2693 * if the mode is integer or float.
2694 * Transform Add(a,-b) into Sub(a,b).
2695 * Reassociation might fold this further.
2697 static ir_node *transform_node_Add(ir_node *n)
2705 n = fold_constant_associativity(n, tarval_add);
2709 n = transform_node_AddSub(n);
2713 a = get_Add_left(n);
2714 b = get_Add_right(n);
2715 mode = get_irn_mode(n);
2717 if (mode_is_reference(mode)) {
2718 ir_mode *lmode = get_irn_mode(a);
2720 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2721 /* an Add(a, NULL) is a hidden Conv */
2722 dbg_info *dbg = get_irn_dbg_info(n);
2723 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2727 if (is_Const(b) && get_mode_arithmetic(mode) == irma_twos_complement) {
2728 ir_tarval *tv = get_Const_tarval(b);
2729 ir_tarval *min = get_mode_min(mode);
2730 /* if all bits are set, then this has the same effect as a Not.
2731 * Note that the following == gives false for different modes which
2732 * is exactly what we want */
2734 dbg_info *dbgi = get_irn_dbg_info(n);
2735 ir_graph *irg = get_irn_irg(n);
2736 ir_node *block = get_nodes_block(n);
2737 ir_node *cnst = new_r_Const(irg, min);
2738 return new_rd_Eor(dbgi, block, a, cnst, mode);
2742 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2744 /* for FP the following optimizations are only allowed if
2745 * fp_strict_algebraic is disabled */
2746 if (mode_is_float(mode)) {
2747 ir_graph *irg = get_irn_irg(n);
2748 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2752 if (mode_is_num(mode)) {
2753 ir_graph *irg = get_irn_irg(n);
2754 /* the following code leads to endless recursion when Mul are replaced
2755 * by a simple instruction chain */
2756 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2757 && a == b && mode_is_int(mode)) {
2758 ir_node *block = get_nodes_block(n);
2761 get_irn_dbg_info(n),
2764 new_r_Const_long(irg, mode, 2),
2766 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2771 get_irn_dbg_info(n),
2776 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2781 get_irn_dbg_info(n),
2786 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2789 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2790 /* Here we rely on constants be on the RIGHT side */
2792 ir_node *op = get_Not_op(a);
2794 if (is_Const(b) && is_Const_one(b)) {
2796 ir_node *blk = get_nodes_block(n);
2797 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2798 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2805 if (is_Or_Eor_Add(n)) {
2806 n = transform_node_Or_(n);
2809 n = transform_node_Eor_(n);
2818 * returns -cnst or NULL if impossible
2820 static ir_node *const_negate(ir_node *cnst)
2822 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2823 dbg_info *dbgi = get_irn_dbg_info(cnst);
2824 ir_graph *irg = get_irn_irg(cnst);
2825 if (tv == tarval_bad) return NULL;
2826 return new_rd_Const(dbgi, irg, tv);
2830 * Do the AddSub optimization, then Transform
2831 * Constant folding on Phi
2832 * Sub(0,a) -> Minus(a)
2833 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2834 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2835 * Sub(Add(a, x), x) -> a
2836 * Sub(x, Add(x, a)) -> -a
2837 * Sub(x, Const) -> Add(x, -Const)
2839 static ir_node *transform_node_Sub(ir_node *n)
2845 n = transform_node_AddSub(n);
2847 a = get_Sub_left(n);
2848 b = get_Sub_right(n);
2850 mode = get_irn_mode(n);
2852 if (mode_is_int(mode)) {
2853 ir_mode *lmode = get_irn_mode(a);
2855 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2856 /* a Sub(a, NULL) is a hidden Conv */
2857 dbg_info *dbg = get_irn_dbg_info(n);
2858 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2859 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2863 if (mode == lmode &&
2864 get_mode_arithmetic(mode) == irma_twos_complement &&
2866 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2868 dbg_info *dbg = get_irn_dbg_info(n);
2869 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2870 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2876 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2878 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2879 if (mode_is_float(mode)) {
2880 ir_graph *irg = get_irn_irg(n);
2881 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2885 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2886 /* a - C -> a + (-C) */
2887 ir_node *cnst = const_negate(b);
2889 ir_node *block = get_nodes_block(n);
2890 dbg_info *dbgi = get_irn_dbg_info(n);
2892 n = new_rd_Add(dbgi, block, a, cnst, mode);
2893 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2898 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2899 dbg_info *dbg = get_irn_dbg_info(n);
2900 ir_node *block = get_nodes_block(n);
2901 ir_node *left = get_Minus_op(a);
2902 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2904 n = new_rd_Minus(dbg, block, add, mode);
2905 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2907 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2908 dbg_info *dbg = get_irn_dbg_info(n);
2909 ir_node *block = get_nodes_block(n);
2910 ir_node *right = get_Minus_op(b);
2912 n = new_rd_Add(dbg, block, a, right, mode);
2913 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2915 } else if (is_Sub(b)) {
2916 /* a - (b - c) -> a + (c - b)
2917 * -> (a - b) + c iff (b - c) is a pointer */
2918 dbg_info *s_dbg = get_irn_dbg_info(b);
2919 ir_node *s_left = get_Sub_left(b);
2920 ir_node *s_right = get_Sub_right(b);
2921 ir_mode *s_mode = get_irn_mode(b);
2922 if (mode_is_reference(s_mode)) {
2923 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2924 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2925 dbg_info *a_dbg = get_irn_dbg_info(n);
2928 s_right = new_r_Conv(lowest_block, s_right, mode);
2929 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2931 ir_node *s_block = get_nodes_block(b);
2932 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2933 dbg_info *a_dbg = get_irn_dbg_info(n);
2934 ir_node *a_block = get_nodes_block(n);
2936 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2938 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2941 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2942 ir_node *m_right = get_Mul_right(b);
2943 if (is_Const(m_right)) {
2944 ir_node *cnst2 = const_negate(m_right);
2945 if (cnst2 != NULL) {
2946 dbg_info *m_dbg = get_irn_dbg_info(b);
2947 ir_node *m_block = get_nodes_block(b);
2948 ir_node *m_left = get_Mul_left(b);
2949 ir_mode *m_mode = get_irn_mode(b);
2950 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2951 dbg_info *a_dbg = get_irn_dbg_info(n);
2952 ir_node *a_block = get_nodes_block(n);
2954 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2955 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2962 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2963 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2965 get_irn_dbg_info(n),
2969 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2972 if ((is_Add(a) || is_Or_Eor_Add(a)) && mode_wrap_around(mode)) {
2973 ir_node *left = get_binop_left(a);
2974 ir_node *right = get_binop_right(a);
2976 /* FIXME: Does the Conv's work only for two complement or generally? */
2978 if (mode != get_irn_mode(right)) {
2979 /* This Sub is an effective Cast */
2980 right = new_r_Conv(get_nodes_block(n), right, mode);
2983 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2985 } else if (right == b) {
2986 if (mode != get_irn_mode(left)) {
2987 /* This Sub is an effective Cast */
2988 left = new_r_Conv(get_nodes_block(n), left, mode);
2991 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2995 if ((is_Add(b) || is_Or_Eor_Add(b)) && mode_wrap_around(mode)) {
2996 ir_node *left = get_binop_left(b);
2997 ir_node *right = get_binop_right(b);
2999 /* FIXME: Does the Conv's work only for two complement or generally? */
3001 ir_mode *r_mode = get_irn_mode(right);
3003 n = new_r_Minus(get_nodes_block(n), right, r_mode);
3004 if (mode != r_mode) {
3005 /* This Sub is an effective Cast */
3006 n = new_r_Conv(get_nodes_block(n), n, mode);
3008 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
3010 } else if (right == a) {
3011 ir_mode *l_mode = get_irn_mode(left);
3013 n = new_r_Minus(get_nodes_block(n), left, l_mode);
3014 if (mode != l_mode) {
3015 /* This Sub is an effective Cast */
3016 n = new_r_Conv(get_nodes_block(n), n, mode);
3018 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
3022 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
3023 ir_mode *mode = get_irn_mode(a);
3025 if (mode == get_irn_mode(b)) {
3027 ir_node *op_a = get_Conv_op(a);
3028 ir_node *op_b = get_Conv_op(b);
3030 /* check if it's allowed to skip the conv */
3031 ma = get_irn_mode(op_a);
3032 mb = get_irn_mode(op_b);
3034 if (mode_is_reference(ma) && mode_is_reference(mb)) {
3035 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
3038 set_Sub_right(n, b);
3044 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
3045 if (!is_reassoc_running() && is_Mul(a)) {
3046 ir_node *ma = get_Mul_left(a);
3047 ir_node *mb = get_Mul_right(a);
3050 ir_node *blk = get_nodes_block(n);
3051 ir_graph *irg = get_irn_irg(n);
3053 get_irn_dbg_info(n),
3057 get_irn_dbg_info(n),
3060 new_r_Const(irg, get_mode_one(mode)),
3063 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
3065 } else if (mb == b) {
3066 ir_node *blk = get_nodes_block(n);
3067 ir_graph *irg = get_irn_irg(n);
3069 get_irn_dbg_info(n),
3073 get_irn_dbg_info(n),
3076 new_r_Const(irg, get_mode_one(mode)),
3079 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
3083 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
3084 ir_node *x = get_Sub_left(a);
3085 ir_node *y = get_Sub_right(a);
3086 ir_node *blk = get_nodes_block(n);
3087 ir_mode *m_b = get_irn_mode(b);
3088 ir_mode *m_y = get_irn_mode(y);
3092 /* Determine the right mode for the Add. */
3095 else if (mode_is_reference(m_b))
3097 else if (mode_is_reference(m_y))
3101 * Both modes are different but none is reference,
3102 * happens for instance in SubP(SubP(P, Iu), Is).
3103 * We have two possibilities here: Cast or ignore.
3104 * Currently we ignore this case.
3109 add = new_r_Add(blk, y, b, add_mode);
3111 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
3112 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
3116 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3117 /* c - ~X = X + (c+1) */
3118 if (is_Const(a) && is_Not(b)) {
3119 ir_tarval *tv = get_Const_tarval(a);
3121 tv = tarval_add(tv, get_mode_one(mode));
3122 if (tv != tarval_bad) {
3123 ir_node *blk = get_nodes_block(n);
3124 ir_graph *irg = get_irn_irg(n);
3125 ir_node *c = new_r_Const(irg, tv);
3126 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
3127 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
3131 /* x-(x&y) = x & ~y */
3133 ir_node *and_left = get_And_left(b);
3134 ir_node *and_right = get_And_right(b);
3135 if (and_right == a) {
3136 ir_node *tmp = and_left;
3137 and_left = and_right;
3140 if (and_left == a) {
3141 dbg_info *dbgi = get_irn_dbg_info(n);
3142 ir_node *block = get_nodes_block(n);
3143 ir_mode *mode = get_irn_mode(n);
3144 ir_node *notn = new_rd_Not(dbgi, block, and_right, mode);
3145 ir_node *and = new_rd_And(dbgi, block, a, notn, mode);
3154 * Several transformation done on n*n=2n bits mul.
3155 * These transformations must be done here because new nodes may be produced.
3157 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
3160 ir_node *a = get_Mul_left(n);
3161 ir_node *b = get_Mul_right(n);
3162 ir_tarval *ta = value_of(a);
3163 ir_tarval *tb = value_of(b);
3164 ir_mode *smode = get_irn_mode(a);
3166 if (ta == get_mode_one(smode)) {
3167 /* (L)1 * (L)b = (L)b */
3168 ir_node *blk = get_nodes_block(n);
3169 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
3170 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
3173 else if (ta == get_mode_minus_one(smode)) {
3174 /* (L)-1 * (L)b = (L)b */
3175 ir_node *blk = get_nodes_block(n);
3176 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
3177 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
3178 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3181 if (tb == get_mode_one(smode)) {
3182 /* (L)a * (L)1 = (L)a */
3183 ir_node *blk = get_irn_n(a, -1);
3184 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
3185 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
3188 else if (tb == get_mode_minus_one(smode)) {
3189 /* (L)a * (L)-1 = (L)-a */
3190 ir_node *blk = get_nodes_block(n);
3191 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
3192 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
3193 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3200 * Transform Mul(a,-1) into -a.
3201 * Do constant evaluation of Phi nodes.
3202 * Do architecture dependent optimizations on Mul nodes
3204 static ir_node *transform_node_Mul(ir_node *n)
3206 ir_node *c, *oldn = n;
3207 ir_mode *mode = get_irn_mode(n);
3208 ir_node *a = get_Mul_left(n);
3209 ir_node *b = get_Mul_right(n);
3211 n = fold_constant_associativity(n, tarval_mul);
3215 if (mode != get_irn_mode(a))
3216 return transform_node_Mul2n(n, mode);
3218 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
3220 if (mode_is_signed(mode)) {
3223 if (value_of(a) == get_mode_minus_one(mode))
3225 else if (value_of(b) == get_mode_minus_one(mode))
3228 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
3229 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3234 if (is_Const(b)) { /* (-a) * const -> a * -const */
3235 ir_node *cnst = const_negate(b);
3237 dbg_info *dbgi = get_irn_dbg_info(n);
3238 ir_node *block = get_nodes_block(n);
3239 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
3240 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
3243 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
3244 dbg_info *dbgi = get_irn_dbg_info(n);
3245 ir_node *block = get_nodes_block(n);
3246 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
3247 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
3249 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
3250 ir_node *sub_l = get_Sub_left(b);
3251 ir_node *sub_r = get_Sub_right(b);
3252 dbg_info *dbgi = get_irn_dbg_info(n);
3253 ir_node *block = get_nodes_block(n);
3254 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
3255 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
3256 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
3259 } else if (is_Minus(b)) {
3260 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
3261 ir_node *sub_l = get_Sub_left(a);
3262 ir_node *sub_r = get_Sub_right(a);
3263 dbg_info *dbgi = get_irn_dbg_info(n);
3264 ir_node *block = get_nodes_block(n);
3265 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
3266 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
3267 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
3270 } else if (is_Shl(a)) {
3271 ir_node *const shl_l = get_Shl_left(a);
3272 if (is_Const(shl_l) && is_Const_one(shl_l)) {
3273 /* (1 << x) * b -> b << x */
3274 dbg_info *const dbgi = get_irn_dbg_info(n);
3275 ir_node *const block = get_nodes_block(n);
3276 ir_node *const shl_r = get_Shl_right(a);
3277 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
3278 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
3281 } else if (is_Shl(b)) {
3282 ir_node *const shl_l = get_Shl_left(b);
3283 if (is_Const(shl_l) && is_Const_one(shl_l)) {
3284 /* a * (1 << x) -> a << x */
3285 dbg_info *const dbgi = get_irn_dbg_info(n);
3286 ir_node *const block = get_nodes_block(n);
3287 ir_node *const shl_r = get_Shl_right(b);
3288 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
3289 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
3293 if (get_mode_arithmetic(mode) == irma_ieee754
3294 || get_mode_arithmetic(mode) == irma_x86_extended_float) {
3296 ir_tarval *tv = get_Const_tarval(a);
3297 if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv)
3298 && !tarval_is_negative(tv)) {
3299 /* 2.0 * b = b + b */
3300 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
3301 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
3305 else if (is_Const(b)) {
3306 ir_tarval *tv = get_Const_tarval(b);
3307 if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv)
3308 && !tarval_is_negative(tv)) {
3309 /* a * 2.0 = a + a */
3310 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
3311 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
3316 return arch_dep_replace_mul_with_shifts(n);
3320 * Transform a Div Node.
3322 static ir_node *transform_node_Div(ir_node *n)
3324 ir_mode *mode = get_Div_resmode(n);
3325 ir_node *a = get_Div_left(n);
3326 ir_node *b = get_Div_right(n);
3328 const ir_node *dummy;
3330 if (mode_is_int(mode)) {
3331 if (is_Const(b) && is_const_Phi(a)) {
3332 /* check for Div(Phi, Const) */
3333 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3335 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3338 } else if (is_Const(a) && is_const_Phi(b)) {
3339 /* check for Div(Const, Phi) */
3340 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3342 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3345 } else if (is_const_Phi(a) && is_const_Phi(b)) {
3346 /* check for Div(Phi, Phi) */
3347 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3349 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3354 if (a == b && value_not_zero(a, &dummy)) {
3355 ir_graph *irg = get_irn_irg(n);
3356 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
3357 value = new_r_Const(irg, get_mode_one(mode));
3358 DBG_OPT_CSTEVAL(n, value);
3361 if (mode_is_signed(mode) && is_Const(b)) {
3362 ir_tarval *tv = get_Const_tarval(b);
3364 if (tv == get_mode_minus_one(mode)) {
3366 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3367 DBG_OPT_CSTEVAL(n, value);
3371 /* Try architecture dependent optimization */
3372 value = arch_dep_replace_div_by_const(n);
3375 assert(mode_is_float(mode));
3377 /* Optimize x/c to x*(1/c) */
3378 if (get_mode_arithmetic(mode) == irma_ieee754) {
3379 ir_tarval *tv = value_of(b);
3381 if (tv != tarval_bad) {
3382 int rem = tarval_fp_ops_enabled();
3385 * Floating point constant folding might be disabled here to
3387 * However, as we check for exact result, doing it is safe.
3390 tarval_enable_fp_ops(1);
3391 tv = tarval_div(get_mode_one(mode), tv);
3392 tarval_enable_fp_ops(rem);
3394 /* Do the transformation if the result is either exact or we are
3395 not using strict rules. */
3396 if (tv != tarval_bad &&
3397 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
3398 ir_node *block = get_nodes_block(n);
3399 ir_graph *irg = get_irn_irg(block);
3400 ir_node *c = new_r_Const(irg, tv);
3401 dbg_info *dbgi = get_irn_dbg_info(n);
3402 value = new_rd_Mul(dbgi, block, a, c, mode);
3415 /* Turn Div into a tuple (mem, jmp, bad, value) */
3416 mem = get_Div_mem(n);
3417 blk = get_nodes_block(n);
3418 irg = get_irn_irg(blk);
3420 /* skip a potential Pin */
3421 mem = skip_Pin(mem);
3422 turn_into_tuple(n, pn_Div_max+1);
3423 set_Tuple_pred(n, pn_Div_M, mem);
3424 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
3425 set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X));
3426 set_Tuple_pred(n, pn_Div_res, value);
3432 * Transform a Mod node.
3434 static ir_node *transform_node_Mod(ir_node *n)
3436 ir_mode *mode = get_Mod_resmode(n);
3437 ir_node *a = get_Mod_left(n);
3438 ir_node *b = get_Mod_right(n);
3443 if (is_Const(b) && is_const_Phi(a)) {
3444 /* check for Div(Phi, Const) */
3445 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3447 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3451 else if (is_Const(a) && is_const_Phi(b)) {
3452 /* check for Div(Const, Phi) */
3453 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3455 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3459 else if (is_const_Phi(a) && is_const_Phi(b)) {
3460 /* check for Div(Phi, Phi) */
3461 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3463 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
3470 irg = get_irn_irg(n);
3471 if (tv != tarval_bad) {
3472 value = new_r_Const(irg, tv);
3474 DBG_OPT_CSTEVAL(n, value);
3477 ir_node *a = get_Mod_left(n);
3478 ir_node *b = get_Mod_right(n);
3479 const ir_node *dummy;
3481 if (a == b && value_not_zero(a, &dummy)) {
3482 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
3483 value = new_r_Const(irg, get_mode_null(mode));
3484 DBG_OPT_CSTEVAL(n, value);
3487 if (mode_is_signed(mode) && is_Const(b)) {
3488 ir_tarval *tv = get_Const_tarval(b);
3490 if (tv == get_mode_minus_one(mode)) {
3492 value = new_r_Const(irg, get_mode_null(mode));
3493 DBG_OPT_CSTEVAL(n, value);
3497 /* Try architecture dependent optimization */
3498 value = arch_dep_replace_mod_by_const(n);
3507 /* Turn Mod into a tuple (mem, jmp, bad, value) */
3508 mem = get_Mod_mem(n);
3509 blk = get_nodes_block(n);
3510 irg = get_irn_irg(blk);
3512 /* skip a potential Pin */
3513 mem = skip_Pin(mem);
3514 turn_into_tuple(n, pn_Mod_max+1);
3515 set_Tuple_pred(n, pn_Mod_M, mem);
3516 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3517 set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X));
3518 set_Tuple_pred(n, pn_Mod_res, value);
3524 * Transform a Cond node.
3526 * Replace the Cond by a Jmp if it branches on a constant
3529 static ir_node *transform_node_Cond(ir_node *n)
3531 ir_node *a = get_Cond_selector(n);
3532 ir_graph *irg = get_irn_irg(n);
3536 /* we need block info which is not available in floating irgs */
3537 if (get_irg_pinned(irg) == op_pin_state_floats)
3541 if (ta == tarval_bad && is_Cmp(a)) {
3542 /* try again with a direct call to compute_cmp, as we don't care
3543 * about the MODEB_LOWERED flag here */
3544 ta = compute_cmp(a);
3547 if (ta != tarval_bad && get_irn_mode(a) == mode_b) {
3548 /* It's a boolean Cond, branching on a boolean constant.
3549 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3550 ir_node *blk = get_nodes_block(n);
3551 jmp = new_r_Jmp(blk);
3552 turn_into_tuple(n, pn_Cond_max+1);
3553 if (ta == tarval_b_true) {
3554 set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
3555 set_Tuple_pred(n, pn_Cond_true, jmp);
3557 set_Tuple_pred(n, pn_Cond_false, jmp);
3558 set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X));
3560 /* We might generate an endless loop, so keep it alive. */
3561 add_End_keepalive(get_irg_end(irg), blk);
3562 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
3567 static ir_node *transform_node_Switch(ir_node *n)
3569 ir_node *op = get_Switch_selector(n);
3570 ir_tarval *val = value_of(op);
3571 if (val != tarval_bad) {
3572 dbg_info *dbgi = get_irn_dbg_info(n);
3573 ir_graph *irg = get_irn_irg(n);
3574 unsigned n_outs = get_Switch_n_outs(n);
3575 ir_node *block = get_nodes_block(n);
3576 ir_node *bad = new_r_Bad(irg, mode_X);
3577 ir_node **in = XMALLOCN(ir_node*, n_outs);
3578 const ir_switch_table *table = get_Switch_table(n);
3579 size_t n_entries = ir_switch_table_get_n_entries(table);
3583 for (i = 0; i < n_entries; ++i) {
3584 const ir_switch_table_entry *entry
3585 = ir_switch_table_get_entry_const(table, i);
3586 ir_tarval *min = entry->min;
3587 ir_tarval *max = entry->max;
3590 if ((min == max && min == val)
3591 || (tarval_cmp(val, min) != ir_relation_less
3592 && tarval_cmp(val, max) != ir_relation_greater)) {
3597 for (o = 0; o < n_outs; ++o) {
3598 if (o == (unsigned)jmp_pn) {
3599 in[o] = new_rd_Jmp(dbgi, block);
3604 return new_r_Tuple(block, (int)n_outs, in);
3610 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
3612 * - and, or, xor instead of &
3613 * - Shl, Shr, Shrs, rotl instead of >>
3614 * (with a special case for Or/Xor + Shrs)
3616 * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
3618 static ir_node *transform_node_shift_bitop(ir_node *n)
3620 ir_graph *irg = get_irn_irg(n);
3621 ir_node *right = get_binop_right(n);
3622 ir_mode *mode = get_irn_mode(n);
3624 ir_node *bitop_left;
3625 ir_node *bitop_right;
3634 ir_tarval *tv_shift;
3636 if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
3639 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
3641 if (!is_Const(right))
3644 left = get_binop_left(n);
3645 op_left = get_irn_op(left);
3646 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
3649 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
3650 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
3651 /* TODO: test if sign bit is affectes */
3655 bitop_right = get_binop_right(left);
3656 if (!is_Const(bitop_right))
3659 bitop_left = get_binop_left(left);
3661 block = get_nodes_block(n);
3662 dbgi = get_irn_dbg_info(n);
3663 tv1 = get_Const_tarval(bitop_right);
3664 tv2 = get_Const_tarval(right);
3666 assert(get_tarval_mode(tv1) == mode);
3669 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
3670 tv_shift = tarval_shl(tv1, tv2);
3671 } else if (is_Shr(n)) {
3672 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
3673 tv_shift = tarval_shr(tv1, tv2);
3674 } else if (is_Shrs(n)) {
3675 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
3676 tv_shift = tarval_shrs(tv1, tv2);
3679 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
3680 tv_shift = tarval_rotl(tv1, tv2);
3683 assert(get_tarval_mode(tv_shift) == mode);
3684 irg = get_irn_irg(n);
3685 new_const = new_r_Const(irg, tv_shift);
3687 if (op_left == op_And) {
3688 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
3689 } else if (op_left == op_Or) {
3690 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
3692 assert(op_left == op_Eor);
3693 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
3702 static ir_node *transform_node_And(ir_node *n)
3704 ir_node *c, *oldn = n;
3705 ir_node *a = get_And_left(n);
3706 ir_node *b = get_And_right(n);
3709 n = fold_constant_associativity(n, tarval_and);
3713 if (is_Cmp(a) && is_Cmp(b)) {
3714 ir_node *a_left = get_Cmp_left(a);
3715 ir_node *a_right = get_Cmp_right(a);
3716 ir_node *b_left = get_Cmp_left(b);
3717 ir_node *b_right = get_Cmp_right(b);
3718 ir_relation a_relation = get_Cmp_relation(a);
3719 ir_relation b_relation = get_Cmp_relation(b);
3720 /* we can combine the relations of two compares with the same
3722 if (a_left == b_left && b_left == b_right) {
3723 dbg_info *dbgi = get_irn_dbg_info(n);
3724 ir_node *block = get_nodes_block(n);
3725 ir_relation new_relation = a_relation & b_relation;
3726 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3728 /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */
3729 if (a_relation == b_relation && a_relation == ir_relation_equal
3730 && !mode_is_float(get_irn_mode(a_left))
3731 && !mode_is_float(get_irn_mode(b_left))) {
3732 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
3733 dbg_info *dbgi = get_irn_dbg_info(n);
3734 ir_node *block = get_nodes_block(n);
3735 ir_mode *a_mode = get_irn_mode(a_left);
3736 ir_mode *b_mode = get_irn_mode(b_left);
3737 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3738 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3739 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
3740 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
3741 ir_graph *irg = get_irn_irg(n);
3742 ir_node *zero = create_zero_const(irg, b_mode);
3743 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3745 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
3746 dbg_info *dbgi = get_irn_dbg_info(n);
3747 ir_node *block = get_nodes_block(n);
3748 ir_mode *a_mode = get_irn_mode(a_left);
3749 ir_mode *b_mode = get_irn_mode(b_left);
3750 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3751 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3752 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
3753 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
3754 ir_graph *irg = get_irn_irg(n);
3755 ir_node *zero = create_zero_const(irg, a_mode);
3756 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3761 mode = get_irn_mode(n);
3762 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3764 if (is_Or(a) || is_Or_Eor_Add(a)) {
3765 ir_node *or_left = get_binop_left(a);
3766 ir_node *or_right = get_binop_right(a);
3767 if (complement_values(or_left, b)) {
3768 /* (a|b) & ~a => b & ~a */
3769 dbg_info *dbgi = get_irn_dbg_info(n);
3770 ir_node *block = get_nodes_block(n);
3771 return new_rd_And(dbgi, block, or_right, b, mode);
3772 } else if (complement_values(or_right, b)) {
3773 /* (a|b) & ~b => a & ~b */
3774 dbg_info *dbgi = get_irn_dbg_info(n);
3775 ir_node *block = get_nodes_block(n);
3776 return new_rd_And(dbgi, block, or_left, b, mode);
3777 } else if (is_Not(b)) {
3778 ir_node *op = get_Not_op(b);
3780 ir_node *ba = get_And_left(op);
3781 ir_node *bb = get_And_right(op);
3783 /* it's enough to test the following cases due to normalization! */
3784 if (or_left == ba && or_right == bb) {
3785 /* (a|b) & ~(a&b) = a^b */
3786 ir_node *block = get_nodes_block(n);
3788 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3789 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3795 if (is_Or(b) || is_Or_Eor_Add(b)) {
3796 ir_node *or_left = get_binop_left(b);
3797 ir_node *or_right = get_binop_right(b);
3798 if (complement_values(or_left, a)) {
3799 /* (a|b) & ~a => b & ~a */
3800 dbg_info *dbgi = get_irn_dbg_info(n);
3801 ir_node *block = get_nodes_block(n);
3802 return new_rd_And(dbgi, block, or_right, a, mode);
3803 } else if (complement_values(or_right, a)) {
3804 /* (a|b) & ~b => a & ~b */
3805 dbg_info *dbgi = get_irn_dbg_info(n);
3806 ir_node *block = get_nodes_block(n);
3807 return new_rd_And(dbgi, block, or_left, a, mode);
3808 } else if (is_Not(a)) {
3809 ir_node *op = get_Not_op(a);
3811 ir_node *aa = get_And_left(op);
3812 ir_node *ab = get_And_right(op);
3814 /* it's enough to test the following cases due to normalization! */
3815 if (or_left == aa && or_right == ab) {
3816 /* (a|b) & ~(a&b) = a^b */
3817 ir_node *block = get_nodes_block(n);
3819 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3820 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3826 if (is_Eor(a) || is_Or_Eor_Add(a)) {
3827 ir_node *al = get_binop_left(a);
3828 ir_node *ar = get_binop_right(a);
3831 /* (b ^ a) & b -> ~a & b */
3832 dbg_info *dbg = get_irn_dbg_info(n);
3833 ir_node *block = get_nodes_block(n);
3835 ar = new_rd_Not(dbg, block, ar, mode);
3836 n = new_rd_And(dbg, block, ar, b, mode);
3837 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3841 /* (a ^ b) & b -> ~a & b */
3842 dbg_info *dbg = get_irn_dbg_info(n);
3843 ir_node *block = get_nodes_block(n);
3845 al = new_rd_Not(dbg, block, al, mode);
3846 n = new_rd_And(dbg, block, al, b, mode);
3847 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3851 if (is_Eor(b) || is_Or_Eor_Add(b)) {
3852 ir_node *bl = get_binop_left(b);
3853 ir_node *br = get_binop_right(b);
3856 /* a & (a ^ b) -> a & ~b */
3857 dbg_info *dbg = get_irn_dbg_info(n);
3858 ir_node *block = get_nodes_block(n);
3860 br = new_rd_Not(dbg, block, br, mode);
3861 n = new_rd_And(dbg, block, br, a, mode);
3862 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3866 /* a & (b ^ a) -> a & ~b */
3867 dbg_info *dbg = get_irn_dbg_info(n);
3868 ir_node *block = get_nodes_block(n);
3870 bl = new_rd_Not(dbg, block, bl, mode);
3871 n = new_rd_And(dbg, block, bl, a, mode);
3872 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3876 if (is_Not(a) && is_Not(b)) {
3877 /* ~a & ~b = ~(a|b) */
3878 ir_node *block = get_nodes_block(n);
3879 ir_mode *mode = get_irn_mode(n);
3883 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3884 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3885 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3890 vrp_attr *b_vrp = vrp_get_info(b);
3891 ir_tarval *a_val = get_Const_tarval(a);
3892 if (b_vrp != NULL && tarval_or(a_val, b_vrp->bits_not_set) == a_val) {
3898 vrp_attr *a_vrp = vrp_get_info(a);
3899 ir_tarval *b_val = get_Const_tarval(b);
3900 if (a_vrp != NULL && tarval_or(b_val, a_vrp->bits_not_set) == b_val) {
3905 n = transform_bitwise_distributive(n, transform_node_And);
3907 n = transform_node_bitop_shift(n);
3915 static ir_node *transform_node_Not(ir_node *n)
3917 ir_node *c, *oldn = n;
3918 ir_node *a = get_Not_op(n);
3919 ir_mode *mode = get_irn_mode(n);
3921 HANDLE_UNOP_PHI(tarval_not,a,c);
3923 /* check for a boolean Not */
3925 dbg_info *dbgi = get_irn_dbg_info(a);
3926 ir_node *block = get_nodes_block(a);
3927 ir_relation relation = get_Cmp_relation(a);
3928 relation = get_negated_relation(relation);
3929 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3930 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3934 /* normalize ~(a ^ b) => a ^ ~b */
3935 if (is_Eor(a) || is_Or_Eor_Add(a)) {
3936 dbg_info *dbg = get_irn_dbg_info(n);
3937 ir_node *block = get_nodes_block(n);
3938 ir_node *eor_right = get_binop_right(a);
3939 ir_node *eor_left = get_binop_left(a);
3940 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3941 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3945 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3946 if (is_Minus(a)) { /* ~-x -> x + -1 */
3947 dbg_info *dbg = get_irn_dbg_info(n);
3948 ir_graph *irg = get_irn_irg(n);
3949 ir_node *block = get_nodes_block(n);
3950 ir_node *add_l = get_Minus_op(a);
3951 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3952 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3953 } else if (is_Add(a) || is_Or_Eor_Add(a)) {
3954 ir_node *add_r = get_binop_right(a);
3955 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3956 /* ~(x + -1) = -x */
3957 ir_node *op = get_binop_left(a);
3958 ir_node *blk = get_nodes_block(n);
3959 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3960 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3968 * Transform a Minus.
3972 * -(a >>u (size-1)) = a >>s (size-1)
3973 * -(a >>s (size-1)) = a >>u (size-1)
3974 * -(a * const) -> a * -const
3976 static ir_node *transform_node_Minus(ir_node *n)
3978 ir_node *c, *oldn = n;
3979 ir_node *a = get_Minus_op(n);
3982 HANDLE_UNOP_PHI(tarval_neg,a,c);
3984 mode = get_irn_mode(a);
3985 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3986 /* the following rules are only to twos-complement */
3989 ir_node *op = get_Not_op(a);
3990 ir_tarval *tv = get_mode_one(mode);
3991 ir_node *blk = get_nodes_block(n);
3992 ir_graph *irg = get_irn_irg(blk);
3993 ir_node *c = new_r_Const(irg, tv);
3994 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3995 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3999 ir_node *c = get_Shr_right(a);
4002 ir_tarval *tv = get_Const_tarval(c);
4004 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
4005 /* -(a >>u (size-1)) = a >>s (size-1) */
4006 ir_node *v = get_Shr_left(a);
4008 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
4009 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
4015 ir_node *c = get_Shrs_right(a);
4018 ir_tarval *tv = get_Const_tarval(c);
4020 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
4021 /* -(a >>s (size-1)) = a >>u (size-1) */
4022 ir_node *v = get_Shrs_left(a);
4024 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
4025 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
4032 /* - (a-b) = b - a */
4033 ir_node *la = get_Sub_left(a);
4034 ir_node *ra = get_Sub_right(a);
4035 ir_node *blk = get_nodes_block(n);
4037 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
4038 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
4042 if (is_Mul(a)) { /* -(a * const) -> a * -const */
4043 ir_node *mul_l = get_Mul_left(a);
4044 ir_node *mul_r = get_Mul_right(a);
4045 ir_tarval *tv = value_of(mul_r);
4046 if (tv != tarval_bad) {
4047 tv = tarval_neg(tv);
4048 if (tv != tarval_bad) {
4049 ir_graph *irg = get_irn_irg(n);
4050 ir_node *cnst = new_r_Const(irg, tv);
4051 dbg_info *dbg = get_irn_dbg_info(a);
4052 ir_node *block = get_nodes_block(a);
4053 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
4054 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
4064 * Transform a Proj(Load) with a non-null address.
4066 static ir_node *transform_node_Proj_Load(ir_node *proj)
4068 if (get_irn_mode(proj) == mode_X) {
4069 ir_node *load = get_Proj_pred(proj);
4071 /* get the Load address */
4072 const ir_node *addr = get_Load_ptr(load);
4073 const ir_node *confirm;
4075 if (value_not_null(addr, &confirm)) {
4076 if (confirm == NULL) {
4077 /* this node may float if it did not depend on a Confirm */
4078 set_irn_pinned(load, op_pin_state_floats);
4080 if (get_Proj_proj(proj) == pn_Load_X_except) {
4081 ir_graph *irg = get_irn_irg(proj);
4082 DBG_OPT_EXC_REM(proj);
4083 return new_r_Bad(irg, mode_X);
4085 ir_node *blk = get_nodes_block(load);
4086 return new_r_Jmp(blk);
4094 * Transform a Proj(Store) with a non-null address.
4096 static ir_node *transform_node_Proj_Store(ir_node *proj)
4098 if (get_irn_mode(proj) == mode_X) {
4099 ir_node *store = get_Proj_pred(proj);
4101 /* get the load/store address */
4102 const ir_node *addr = get_Store_ptr(store);
4103 const ir_node *confirm;
4105 if (value_not_null(addr, &confirm)) {
4106 if (confirm == NULL) {
4107 /* this node may float if it did not depend on a Confirm */
4108 set_irn_pinned(store, op_pin_state_floats);
4110 if (get_Proj_proj(proj) == pn_Store_X_except) {
4111 ir_graph *irg = get_irn_irg(proj);
4112 DBG_OPT_EXC_REM(proj);
4113 return new_r_Bad(irg, mode_X);
4115 ir_node *blk = get_nodes_block(store);
4116 return new_r_Jmp(blk);
4124 * Transform a Proj(Div) with a non-zero value.
4125 * Removes the exceptions and routes the memory to the NoMem node.
4127 static ir_node *transform_node_Proj_Div(ir_node *proj)
4129 ir_node *div = get_Proj_pred(proj);
4130 ir_node *b = get_Div_right(div);
4131 ir_node *res, *new_mem;
4132 const ir_node *confirm;
4135 if (value_not_zero(b, &confirm)) {
4136 /* div(x, y) && y != 0 */
4137 if (confirm == NULL) {
4138 /* we are sure we have a Const != 0 */
4139 new_mem = get_Div_mem(div);
4140 new_mem = skip_Pin(new_mem);
4141 set_Div_mem(div, new_mem);
4142 set_irn_pinned(div, op_pin_state_floats);
4145 proj_nr = get_Proj_proj(proj);
4147 case pn_Div_X_regular:
4148 return new_r_Jmp(get_nodes_block(div));
4150 case pn_Div_X_except: {
4151 ir_graph *irg = get_irn_irg(proj);
4152 /* we found an exception handler, remove it */
4153 DBG_OPT_EXC_REM(proj);
4154 return new_r_Bad(irg, mode_X);
4158 ir_graph *irg = get_irn_irg(proj);
4159 res = get_Div_mem(div);
4160 new_mem = get_irg_no_mem(irg);
4163 /* This node can only float up to the Confirm block */
4164 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
4166 set_irn_pinned(div, op_pin_state_floats);
4167 /* this is a Div without exception, we can remove the memory edge */
4168 set_Div_mem(div, new_mem);
4177 * Transform a Proj(Mod) with a non-zero value.
4178 * Removes the exceptions and routes the memory to the NoMem node.
4180 static ir_node *transform_node_Proj_Mod(ir_node *proj)
4182 ir_node *mod = get_Proj_pred(proj);
4183 ir_node *b = get_Mod_right(mod);
4184 ir_node *res, *new_mem;
4185 const ir_node *confirm;
4188 if (value_not_zero(b, &confirm)) {
4189 /* mod(x, y) && y != 0 */
4190 proj_nr = get_Proj_proj(proj);
4192 if (confirm == NULL) {
4193 /* we are sure we have a Const != 0 */
4194 new_mem = get_Mod_mem(mod);
4195 new_mem = skip_Pin(new_mem);
4196 set_Mod_mem(mod, new_mem);
4197 set_irn_pinned(mod, op_pin_state_floats);
4202 case pn_Mod_X_regular:
4203 return new_r_Jmp(get_irn_n(mod, -1));
4205 case pn_Mod_X_except: {
4206 ir_graph *irg = get_irn_irg(proj);
4207 /* we found an exception handler, remove it */
4208 DBG_OPT_EXC_REM(proj);
4209 return new_r_Bad(irg, mode_X);
4213 ir_graph *irg = get_irn_irg(proj);
4214 res = get_Mod_mem(mod);
4215 new_mem = get_irg_no_mem(irg);
4218 /* This node can only float up to the Confirm block */
4219 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
4221 /* this is a Mod without exception, we can remove the memory edge */
4222 set_Mod_mem(mod, new_mem);
4226 if (get_Mod_left(mod) == b) {
4227 /* a % a = 0 if a != 0 */
4228 ir_graph *irg = get_irn_irg(proj);
4229 ir_mode *mode = get_irn_mode(proj);
4230 ir_node *res = new_r_Const(irg, get_mode_null(mode));
4232 DBG_OPT_CSTEVAL(mod, res);
4241 * return true if the operation returns a value with exactly 1 bit set
4243 static bool is_single_bit(const ir_node *node)
4245 /* a first implementation, could be extended with vrp and others... */
4247 ir_node *shl_l = get_Shl_left(node);
4248 ir_mode *mode = get_irn_mode(node);
4249 int modulo = get_mode_modulo_shift(mode);
4250 /* this works if we shift a 1 and we have modulo shift */
4251 if (is_Const(shl_l) && is_Const_one(shl_l)
4252 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
4255 } else if (is_Const(node)) {
4256 ir_tarval *tv = get_Const_tarval(node);
4257 return tarval_is_single_bit(tv);
4263 * checks if node just flips a bit in another node and returns that other node
4264 * if so. @p tv should be a value having just 1 bit set
4266 static ir_node *flips_bit(const ir_node *node, ir_tarval *tv)
4269 return get_Not_op(node);
4271 ir_node *right = get_Eor_right(node);
4272 if (is_Const(right)) {
4273 ir_tarval *right_tv = get_Const_tarval(right);
4274 ir_mode *mode = get_irn_mode(node);
4275 if (tarval_and(right_tv, tv) != get_mode_null(mode))
4276 return get_Eor_left(node);
4283 * Normalizes and optimizes Cmp nodes.
4285 static ir_node *transform_node_Cmp(ir_node *n)
4287 ir_node *left = get_Cmp_left(n);
4288 ir_node *right = get_Cmp_right(n);
4289 ir_mode *mode = get_irn_mode(left);
4290 ir_tarval *tv = NULL;
4291 bool changed = false;
4292 bool changedc = false;
4293 ir_relation relation = get_Cmp_relation(n);
4294 ir_relation possible = ir_get_possible_cmp_relations(left, right);
4296 /* mask out impossible relations */
4297 ir_relation new_relation = relation & possible;
4298 if (new_relation != relation) {
4299 relation = new_relation;
4303 /* Remove unnecessary conversions */
4304 if (is_Conv(left) && is_Conv(right)) {
4305 ir_node *op_left = get_Conv_op(left);
4306 ir_node *op_right = get_Conv_op(right);
4307 ir_mode *mode_left = get_irn_mode(op_left);
4308 ir_mode *mode_right = get_irn_mode(op_right);
4310 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4311 && mode_left != mode_b && mode_right != mode_b) {
4312 ir_node *block = get_nodes_block(n);
4314 if (mode_left == mode_right) {
4318 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4319 } else if (smaller_mode(mode_left, mode_right)) {
4320 left = new_r_Conv(block, op_left, mode_right);
4323 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4324 } else if (smaller_mode(mode_right, mode_left)) {
4326 right = new_r_Conv(block, op_right, mode_left);
4328 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4330 mode = get_irn_mode(left);
4333 if (is_Conv(left) && is_Const(right)) {
4334 ir_node *op_left = get_Conv_op(left);
4335 ir_mode *mode_left = get_irn_mode(op_left);
4336 if (smaller_mode(mode_left, mode) && mode_left != mode_b) {
4337 ir_tarval *tv = get_Const_tarval(right);
4338 tarval_int_overflow_mode_t last_mode
4339 = tarval_get_integer_overflow_mode();
4341 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
4342 new_tv = tarval_convert_to(tv, mode_left);
4343 tarval_set_integer_overflow_mode(last_mode);
4344 if (new_tv != tarval_bad) {
4345 ir_graph *irg = get_irn_irg(n);
4347 right = new_r_Const(irg, new_tv);
4348 mode = get_irn_mode(left);
4350 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4356 * Optimize -a CMP -b into b CMP a.
4357 * This works only for modes where unary Minus cannot Overflow.
4358 * Note that two-complement integers can Overflow so it will NOT work.
4360 if (!mode_overflow_on_unary_Minus(mode) &&
4361 is_Minus(left) && is_Minus(right)) {
4362 left = get_Minus_op(left);
4363 right = get_Minus_op(right);
4364 relation = get_inversed_relation(relation);
4366 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4369 /* remove operation on both sides if possible */
4370 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4372 * The following operations are NOT safe for floating point operations, for instance
4373 * 1.0 + inf == 2.0 + inf, =/=> x == y
4375 if (mode_is_int(mode)) {
4376 unsigned lop = get_irn_opcode(left);
4378 if (lop == get_irn_opcode(right)) {
4379 ir_node *ll, *lr, *rl, *rr;
4381 /* same operation on both sides, try to remove */
4385 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4386 left = get_unop_op(left);
4387 right = get_unop_op(right);
4389 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4392 ll = get_Add_left(left);
4393 lr = get_Add_right(left);
4394 rl = get_Add_left(right);
4395 rr = get_Add_right(right);
4398 /* X + a CMP X + b ==> a CMP b */
4402 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4403 } else if (ll == rr) {
4404 /* X + a CMP b + X ==> a CMP b */
4408 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4409 } else if (lr == rl) {
4410 /* a + X CMP X + b ==> a CMP b */
4414 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4415 } else if (lr == rr) {
4416 /* a + X CMP b + X ==> a CMP b */
4420 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4424 ll = get_Sub_left(left);
4425 lr = get_Sub_right(left);
4426 rl = get_Sub_left(right);
4427 rr = get_Sub_right(right);
4430 /* X - a CMP X - b ==> a CMP b */
4434 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4435 } else if (lr == rr) {
4436 /* a - X CMP b - X ==> a CMP b */
4440 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4444 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4445 /* a ROTL X CMP b ROTL X ==> a CMP b */
4446 left = get_Rotl_left(left);
4447 right = get_Rotl_left(right);
4449 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4457 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4458 if (is_Add(left) || is_Sub(left) || is_Or_Eor_Add(left)) {
4459 ir_node *ll = get_binop_left(left);
4460 ir_node *lr = get_binop_right(left);
4462 if (lr == right && (is_Add(left) || is_Or_Eor_Add(left))) {
4468 ir_graph *irg = get_irn_irg(n);
4470 right = create_zero_const(irg, mode);
4472 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4475 if (is_Add(right) || is_Sub(right) || is_Or_Eor_Add(right)) {
4476 ir_node *rl = get_binop_left(right);
4477 ir_node *rr = get_binop_right(right);
4479 if (rr == left && (is_Add(right) || is_Or_Eor_Add(right))) {
4485 ir_graph *irg = get_irn_irg(n);
4487 right = create_zero_const(irg, mode);
4489 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4493 if (is_And(left) && is_Const(right)) {
4494 ir_node *ll = get_binop_left(left);
4495 ir_node *lr = get_binop_right(left);
4496 if (is_Shr(ll) && is_Const(lr)) {
4497 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4498 ir_node *block = get_nodes_block(n);
4499 ir_mode *mode = get_irn_mode(left);
4501 ir_node *llr = get_Shr_right(ll);
4502 if (is_Const(llr)) {
4503 dbg_info *dbg = get_irn_dbg_info(left);
4504 ir_graph *irg = get_irn_irg(left);
4506 ir_tarval *c1 = get_Const_tarval(llr);
4507 ir_tarval *c2 = get_Const_tarval(lr);
4508 ir_tarval *c3 = get_Const_tarval(right);
4509 ir_tarval *mask = tarval_shl(c2, c1);
4510 ir_tarval *value = tarval_shl(c3, c1);
4512 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4513 right = new_r_Const(irg, value);
4518 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4520 if (is_Const(right) && is_Const_null(right) &&
4521 (is_Eor(left) || is_Or_Eor_Add(left))) {
4522 right = get_Eor_right(left);
4523 left = get_Eor_left(left);
4529 if (mode_is_int(mode) && is_And(left)) {
4530 /* a complicated Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4531 * by the simpler Cmp(And(1bit, val), 0) negated pnc */
4532 if (relation == ir_relation_equal
4533 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4534 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less)) {
4535 ir_node *and0 = get_And_left(left);
4536 ir_node *and1 = get_And_right(left);
4537 if (and1 == right) {
4538 ir_node *tmp = and0;
4542 if (and0 == right && is_single_bit(and0)) {
4543 ir_graph *irg = get_irn_irg(n);
4545 relation == ir_relation_equal ? ir_relation_less_greater
4546 : ir_relation_equal;
4547 right = create_zero_const(irg, mode);
4553 if (is_Const(right) && is_Const_null(right) &&
4554 (relation == ir_relation_equal
4555 || (relation == ir_relation_less_greater)
4556 || (!mode_is_signed(mode) && relation == ir_relation_greater))) {
4558 /* instead of flipping the bit before the bit-test operation negate
4560 ir_node *and0 = get_And_left(left);
4561 ir_node *and1 = get_And_right(left);
4562 if (is_Const(and1)) {
4563 ir_tarval *tv = get_Const_tarval(and1);
4564 if (tarval_is_single_bit(tv)) {
4565 ir_node *flipped = flips_bit(and0, tv);
4566 if (flipped != NULL) {
4567 dbg_info *dbgi = get_irn_dbg_info(left);
4568 ir_node *block = get_nodes_block(left);
4569 relation = get_negated_relation(relation);
4570 left = new_rd_And(dbgi, block, flipped, and1, mode);
4579 /* replace mode_b compares with ands/ors */
4580 if (mode == mode_b) {
4581 ir_node *block = get_nodes_block(n);
4585 case ir_relation_less_equal:
4586 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4588 case ir_relation_less:
4589 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4591 case ir_relation_greater_equal:
4592 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4594 case ir_relation_greater:
4595 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4597 case ir_relation_less_greater:
4598 bres = new_r_Eor(block, left, right, mode_b);
4600 case ir_relation_equal:
4601 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4604 #ifdef DEBUG_libfirm
4605 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4610 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4616 * First step: normalize the compare op
4617 * by placing the constant on the right side
4618 * or moving the lower address node to the left.
4620 if (!operands_are_normalized(left, right)) {
4625 relation = get_inversed_relation(relation);
4630 * Second step: Try to reduce the magnitude
4631 * of a constant. This may help to generate better code
4632 * later and may help to normalize more compares.
4633 * Of course this is only possible for integer values.
4635 tv = value_of(right);
4636 if (tv != tarval_bad) {
4637 ir_mode *mode = get_irn_mode(right);
4639 /* cmp(mux(x, cf, ct), c2) can be eliminated:
4640 * cmp(ct,c2) | cmp(cf,c2) | result
4641 * -----------|------------|--------
4642 * true | true | True
4643 * false | false | False
4645 * false | true | not(x)
4648 ir_node *mux_true = get_Mux_true(left);
4649 ir_node *mux_false = get_Mux_false(left);
4650 if (is_Const(mux_true) && is_Const(mux_false)) {
4651 /* we can fold true/false constant separately */
4652 ir_tarval *tv_true = get_Const_tarval(mux_true);
4653 ir_tarval *tv_false = get_Const_tarval(mux_false);
4654 ir_relation r_true = tarval_cmp(tv_true, tv);
4655 ir_relation r_false = tarval_cmp(tv_false, tv);
4656 if (r_true != ir_relation_false
4657 || r_false != ir_relation_false) {
4658 bool rel_true = (r_true & relation) != 0;
4659 bool rel_false = (r_false & relation) != 0;
4660 ir_node *cond = get_Mux_sel(left);
4661 if (rel_true == rel_false) {
4662 relation = rel_true ? ir_relation_true
4663 : ir_relation_false;
4664 } else if (rel_true) {
4667 dbg_info *dbgi = get_irn_dbg_info(n);
4668 ir_node *block = get_nodes_block(n);
4669 ir_node *notn = new_rd_Not(dbgi, block, cond, mode_b);
4676 /* TODO extend to arbitrary constants */
4677 if (is_Conv(left) && tarval_is_null(tv)) {
4678 ir_node *op = get_Conv_op(left);
4679 ir_mode *op_mode = get_irn_mode(op);
4682 * UpConv(x) REL 0 ==> x REL 0
4683 * Don't do this for float values as it's unclear whether it is a
4684 * win. (on the other side it makes detection/creation of fabs hard)
4686 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4687 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4688 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4689 !mode_is_float(mode)) {
4690 tv = get_mode_null(op_mode);
4694 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4698 if (tv != tarval_bad) {
4699 /* the following optimization is possible on modes without Overflow
4700 * on Unary Minus or on == and !=:
4701 * -a CMP c ==> a swap(CMP) -c
4703 * Beware: for two-complement Overflow may occur, so only == and != can
4704 * be optimized, see this:
4705 * -MININT < 0 =/=> MININT > 0 !!!
4707 if (is_Minus(left) &&
4708 (!mode_overflow_on_unary_Minus(mode) ||
4709 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4710 tv = tarval_neg(tv);
4712 if (tv != tarval_bad) {
4713 left = get_Minus_op(left);
4714 relation = get_inversed_relation(relation);
4716 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4718 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4719 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4720 tv = tarval_not(tv);
4722 if (tv != tarval_bad) {
4723 left = get_Not_op(left);
4725 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4729 /* for integer modes, we have more */
4730 if (mode_is_int(mode) && !is_Const(left)) {
4731 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4732 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4733 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4734 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4736 if (tv != tarval_bad) {
4737 relation ^= ir_relation_equal;
4739 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4742 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4743 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4744 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4745 tv = tarval_add(tv, get_mode_one(mode));
4747 if (tv != tarval_bad) {
4748 relation ^= ir_relation_equal;
4750 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4754 /* the following reassociations work only for == and != */
4755 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4756 if (tv != tarval_bad) {
4757 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4759 ir_node *c1 = get_Sub_right(left);
4760 ir_tarval *tv2 = value_of(c1);
4762 if (tv2 != tarval_bad) {
4763 tv2 = tarval_add(tv, value_of(c1));
4765 if (tv2 != tarval_bad) {
4766 left = get_Sub_left(left);
4769 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4773 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4774 else if (is_Add(left) || is_Or_Eor_Add(left)) {
4775 ir_node *a_l = get_binop_left(left);
4776 ir_node *a_r = get_binop_right(left);
4780 if (is_Const(a_l)) {
4782 tv2 = value_of(a_l);
4785 tv2 = value_of(a_r);
4788 if (tv2 != tarval_bad) {
4789 tv2 = tarval_sub(tv, tv2, NULL);
4791 if (tv2 != tarval_bad) {
4795 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4799 /* -a == c ==> a == -c, -a != c ==> a != -c */
4800 else if (is_Minus(left)) {
4801 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4803 if (tv2 != tarval_bad) {
4804 left = get_Minus_op(left);
4807 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4814 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4815 switch (get_irn_opcode(left)) {
4819 c1 = get_And_right(left);
4822 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4823 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4825 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4827 /* TODO: move to constant evaluation */
4828 ir_graph *irg = get_irn_irg(n);
4829 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4830 c1 = new_r_Const(irg, tv);
4831 DBG_OPT_CSTEVAL(n, c1);
4835 if (tarval_is_single_bit(tv)) {
4837 * optimization for AND:
4839 * And(x, C) == C ==> And(x, C) != 0
4840 * And(x, C) != C ==> And(X, C) == 0
4842 * if C is a single Bit constant.
4845 /* check for Constant's match. We have check hare the tarvals,
4846 because our const might be changed */
4847 if (get_Const_tarval(c1) == tv) {
4848 /* fine: do the transformation */
4849 tv = get_mode_null(get_tarval_mode(tv));
4850 relation ^= ir_relation_less_equal_greater;
4852 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4858 c1 = get_Or_right(left);
4859 if (is_Const(c1) && tarval_is_null(tv)) {
4861 * Or(x, C) == 0 && C != 0 ==> FALSE
4862 * Or(x, C) != 0 && C != 0 ==> TRUE
4864 if (! tarval_is_null(get_Const_tarval(c1))) {
4865 /* TODO: move to constant evaluation */
4866 ir_graph *irg = get_irn_irg(n);
4867 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4868 c1 = new_r_Const(irg, tv);
4869 DBG_OPT_CSTEVAL(n, c1);
4876 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4878 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4881 c1 = get_Shl_right(left);
4883 ir_graph *irg = get_irn_irg(c1);
4884 ir_tarval *tv1 = get_Const_tarval(c1);
4885 ir_mode *mode = get_irn_mode(left);
4886 ir_tarval *minus1 = get_mode_all_one(mode);
4887 ir_tarval *amask = tarval_shr(minus1, tv1);
4888 ir_tarval *cmask = tarval_shl(minus1, tv1);
4891 if (tarval_and(tv, cmask) != tv) {
4892 /* condition not met */
4893 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4894 c1 = new_r_Const(irg, tv);
4895 DBG_OPT_CSTEVAL(n, c1);
4898 sl = get_Shl_left(left);
4899 blk = get_nodes_block(n);
4900 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4901 tv = tarval_shr(tv, tv1);
4903 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4908 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4910 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4913 c1 = get_Shr_right(left);
4915 ir_graph *irg = get_irn_irg(c1);
4916 ir_tarval *tv1 = get_Const_tarval(c1);
4917 ir_mode *mode = get_irn_mode(left);
4918 ir_tarval *minus1 = get_mode_all_one(mode);
4919 ir_tarval *amask = tarval_shl(minus1, tv1);
4920 ir_tarval *cmask = tarval_shr(minus1, tv1);
4923 if (tarval_and(tv, cmask) != tv) {
4924 /* condition not met */
4925 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4926 c1 = new_r_Const(irg, tv);
4927 DBG_OPT_CSTEVAL(n, c1);
4930 sl = get_Shr_left(left);
4931 blk = get_nodes_block(n);
4932 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4933 tv = tarval_shl(tv, tv1);
4935 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4940 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4942 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4945 c1 = get_Shrs_right(left);
4947 ir_graph *irg = get_irn_irg(c1);
4948 ir_tarval *tv1 = get_Const_tarval(c1);
4949 ir_mode *mode = get_irn_mode(left);
4950 ir_tarval *minus1 = get_mode_all_one(mode);
4951 ir_tarval *amask = tarval_shl(minus1, tv1);
4952 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4955 cond = tarval_sub(cond, tv1, NULL);
4956 cond = tarval_shrs(tv, cond);
4958 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4959 /* condition not met */
4960 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4961 c1 = new_r_Const(irg, tv);
4962 DBG_OPT_CSTEVAL(n, c1);
4965 sl = get_Shrs_left(left);
4966 blk = get_nodes_block(n);
4967 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4968 tv = tarval_shl(tv, tv1);
4970 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4978 if (changedc) { /* need a new Const */
4979 ir_graph *irg = get_irn_irg(n);
4980 right = new_r_Const(irg, tv);
4984 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4985 ir_node *op = get_Proj_pred(left);
4987 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4988 ir_node *c = get_binop_right(op);
4991 ir_tarval *tv = get_Const_tarval(c);
4993 if (tarval_is_single_bit(tv)) {
4994 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4995 ir_node *v = get_binop_left(op);
4996 ir_node *blk = get_irn_n(op, -1);
4997 ir_graph *irg = get_irn_irg(op);
4998 ir_mode *mode = get_irn_mode(v);
5000 tv = tarval_sub(tv, get_mode_one(mode), NULL);
5001 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
5003 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
5010 dbg_info *dbgi = get_irn_dbg_info(n);
5011 ir_node *block = get_nodes_block(n);
5013 /* create a new compare */
5014 n = new_rd_Cmp(dbgi, block, left, right, relation);
5021 * Optimize CopyB(mem, x, x) into a Nop.
5023 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
5025 ir_node *copyb = get_Proj_pred(proj);
5026 ir_node *a = get_CopyB_dst(copyb);
5027 ir_node *b = get_CopyB_src(copyb);
5030 switch (get_Proj_proj(proj)) {
5031 case pn_CopyB_X_regular:
5032 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
5033 DBG_OPT_EXC_REM(proj);
5034 proj = new_r_Jmp(get_nodes_block(copyb));
5036 case pn_CopyB_X_except: {
5037 ir_graph *irg = get_irn_irg(proj);
5038 DBG_OPT_EXC_REM(proj);
5039 proj = new_r_Bad(irg, mode_X);
5050 * Optimize Bounds(idx, idx, upper) into idx.
5052 static ir_node *transform_node_Proj_Bound(ir_node *proj)
5054 ir_node *oldn = proj;
5055 ir_node *bound = get_Proj_pred(proj);
5056 ir_node *idx = get_Bound_index(bound);
5057 ir_node *pred = skip_Proj(idx);
5060 if (idx == get_Bound_lower(bound))
5062 else if (is_Bound(pred)) {
5064 * idx was Bounds checked previously, it is still valid if
5065 * lower <= pred_lower && pred_upper <= upper.
5067 ir_node *lower = get_Bound_lower(bound);
5068 ir_node *upper = get_Bound_upper(bound);
5069 if (get_Bound_lower(pred) == lower &&
5070 get_Bound_upper(pred) == upper) {
5072 * One could expect that we simply return the previous
5073 * Bound here. However, this would be wrong, as we could
5074 * add an exception Proj to a new location then.
5075 * So, we must turn in into a tuple.
5081 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
5082 switch (get_Proj_proj(proj)) {
5084 DBG_OPT_EXC_REM(proj);
5085 proj = get_Bound_mem(bound);
5087 case pn_Bound_X_except:
5088 DBG_OPT_EXC_REM(proj);
5089 proj = new_r_Bad(get_irn_irg(proj), mode_X);
5093 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
5095 case pn_Bound_X_regular:
5096 DBG_OPT_EXC_REM(proj);
5097 proj = new_r_Jmp(get_nodes_block(bound));
5107 * Does all optimizations on nodes that must be done on its Projs
5108 * because of creating new nodes.
5110 static ir_node *transform_node_Proj(ir_node *proj)
5112 ir_node *n = get_Proj_pred(proj);
5114 if (n->op->ops.transform_node_Proj)
5115 return n->op->ops.transform_node_Proj(proj);
5120 * Test whether a block is unreachable
5121 * Note: That this only returns true when
5122 * IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE is set.
5123 * This is important, as you easily end up producing invalid constructs in the
5124 * unreachable code when optimizing away edges into the unreachable code.
5125 * So only set this flag when you iterate localopts to the fixpoint.
5126 * When you reach the fixpoint then all unreachable code is dead
5127 * (= can't be reached by firm edges) and you won't see the invalid constructs
5130 static bool is_block_unreachable(const ir_node *block)
5132 const ir_graph *irg = get_irn_irg(block);
5133 if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE))
5135 return get_Block_dom_depth(block) < 0;
5138 static ir_node *transform_node_Block(ir_node *block)
5140 ir_graph *irg = get_irn_irg(block);
5141 int arity = get_irn_arity(block);
5142 ir_node *bad = NULL;
5145 if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE))
5148 for (i = 0; i < arity; ++i) {
5149 ir_node *const pred = get_Block_cfgpred(block, i);
5150 if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred)))
5153 bad = new_r_Bad(irg, mode_X);
5154 set_irn_n(block, i, bad);
5160 static ir_node *transform_node_Phi(ir_node *phi)
5162 int n = get_irn_arity(phi);
5163 ir_mode *mode = get_irn_mode(phi);
5164 ir_node *block = get_nodes_block(phi);
5165 ir_graph *irg = get_irn_irg(phi);
5166 ir_node *bad = NULL;
5169 /* Set phi-operands for bad-block inputs to bad */
5170 for (i = 0; i < n; ++i) {
5171 if (!is_Bad(get_Phi_pred(phi, i))) {
5172 ir_node *pred = get_Block_cfgpred(block, i);
5173 if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) {
5175 bad = new_r_Bad(irg, mode);
5176 set_irn_n(phi, i, bad);
5181 /* Move Pin nodes down through Phi nodes. */
5182 if (mode == mode_M) {
5183 n = get_irn_arity(phi);
5185 /* Beware of Phi0 */
5189 bool has_pin = false;
5191 NEW_ARR_A(ir_node *, in, n);
5193 for (i = 0; i < n; ++i) {
5194 ir_node *pred = get_irn_n(phi, i);
5197 in[i] = get_Pin_op(pred);
5199 } else if (is_Bad(pred)) {
5209 /* Move the Pin nodes "behind" the Phi. */
5210 block = get_irn_n(phi, -1);
5211 new_phi = new_r_Phi(block, n, in, mode_M);
5212 return new_r_Pin(block, new_phi);
5215 /* Move Confirms down through Phi nodes. */
5216 else if (mode_is_reference(mode)) {
5217 n = get_irn_arity(phi);
5219 /* Beware of Phi0 */
5221 ir_node *pred = get_irn_n(phi, 0);
5222 ir_node *bound, *new_phi, *block, **in;
5223 ir_relation relation;
5224 bool has_confirm = false;
5226 if (! is_Confirm(pred))
5229 bound = get_Confirm_bound(pred);
5230 relation = get_Confirm_relation(pred);
5232 NEW_ARR_A(ir_node *, in, n);
5233 in[0] = get_Confirm_value(pred);
5235 for (i = 1; i < n; ++i) {
5236 pred = get_irn_n(phi, i);
5238 if (is_Confirm(pred) &&
5239 get_Confirm_bound(pred) == bound &&
5240 get_Confirm_relation(pred) == relation) {
5241 in[i] = get_Confirm_value(pred);
5243 } else if (is_Bad(pred)) {
5253 /* move the Confirm nodes "behind" the Phi */
5254 block = get_irn_n(phi, -1);
5255 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
5256 return new_r_Confirm(block, new_phi, bound, relation);
5263 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5265 * Should be moved to reassociation?
5267 static ir_node *transform_node_shift(ir_node *n)
5269 ir_node *left, *right;
5271 ir_mode *count_mode;
5272 ir_tarval *tv1, *tv2, *res;
5273 ir_node *in[2], *irn, *block;
5277 left = get_binop_left(n);
5279 /* different operations */
5280 if (get_irn_op(left) != get_irn_op(n))
5283 right = get_binop_right(n);
5284 tv1 = value_of(right);
5285 if (tv1 == tarval_bad)
5288 tv2 = value_of(get_binop_right(left));
5289 if (tv2 == tarval_bad)
5292 count_mode = get_tarval_mode(tv1);
5293 if (get_tarval_mode(tv2) != count_mode) {
5294 /* TODO: search bigger mode or something and convert... */
5298 mode = get_irn_mode(n);
5299 modulo_shf = get_mode_modulo_shift(mode);
5301 if (modulo_shf > 0) {
5302 ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode);
5304 /* I'm not so sure what happens in one complement... */
5305 assert(get_mode_arithmetic(count_mode) == irma_twos_complement);
5306 /* modulo shifts should always be a power of 2 (otherwise modulo_mask
5307 * above will be invalid) */
5308 assert(modulo_shf<=0 || is_po2(modulo_shf));
5310 tv1 = tarval_and(tv1, modulo_mask);
5311 tv2 = tarval_and(tv2, modulo_mask);
5313 res = tarval_add(tv1, tv2);
5314 irg = get_irn_irg(n);
5316 /* beware: a simple replacement works only, if res < modulo shift */
5318 int bits = get_mode_size_bits(mode);
5319 ir_tarval *modulo = new_tarval_from_long(bits, count_mode);
5320 res = tarval_mod(res, modulo);
5322 long bits = get_mode_size_bits(mode);
5323 ir_tarval *mode_size = new_tarval_from_long(bits, count_mode);
5325 /* shifting too much */
5326 if (!(tarval_cmp(res, mode_size) & ir_relation_less)) {
5328 ir_node *block = get_nodes_block(n);
5329 dbg_info *dbgi = get_irn_dbg_info(n);
5330 ir_mode *smode = get_irn_mode(right);
5331 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
5332 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5335 return new_r_Const(irg, get_mode_null(mode));
5339 /* ok, we can replace it */
5340 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5341 block = get_nodes_block(n);
5343 in[0] = get_binop_left(left);
5344 in[1] = new_r_Const(irg, res);
5346 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5348 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5355 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5357 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5358 * (also with x >>s c1 when c1>=c2)
5360 static ir_node *transform_node_shl_shr(ir_node *n)
5363 ir_node *right = get_binop_right(n);
5373 ir_tarval *tv_shift;
5376 ir_relation relation;
5379 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5381 if (!is_Const(right))
5384 left = get_binop_left(n);
5385 mode = get_irn_mode(n);
5386 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5387 ir_node *shr_right = get_binop_right(left);
5389 if (!is_Const(shr_right))
5392 x = get_binop_left(left);
5393 tv_shr = get_Const_tarval(shr_right);
5394 tv_shl = get_Const_tarval(right);
5396 if (is_Shrs(left)) {
5397 /* shrs variant only allowed if c1 >= c2 */
5398 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5401 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5404 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5406 tv_mask = tarval_shl(tv_mask, tv_shl);
5407 } else if (is_Shr(n) && is_Shl(left)) {
5408 ir_node *shl_right = get_Shl_right(left);
5410 if (!is_Const(shl_right))
5413 x = get_Shl_left(left);
5414 tv_shr = get_Const_tarval(right);
5415 tv_shl = get_Const_tarval(shl_right);
5417 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5418 tv_mask = tarval_shr(tv_mask, tv_shr);
5423 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5424 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5427 assert(tv_mask != tarval_bad);
5428 assert(get_tarval_mode(tv_mask) == mode);
5430 block = get_nodes_block(n);
5431 irg = get_irn_irg(block);
5432 dbgi = get_irn_dbg_info(n);
5434 relation = tarval_cmp(tv_shl, tv_shr);
5435 if (relation == ir_relation_less || relation == ir_relation_equal) {
5436 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5437 new_const = new_r_Const(irg, tv_shift);
5439 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5441 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5444 assert(relation == ir_relation_greater);
5445 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5446 new_const = new_r_Const(irg, tv_shift);
5447 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5450 new_const = new_r_Const(irg, tv_mask);
5451 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5456 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5458 ir_mode *mode = get_tarval_mode(tv);
5459 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5460 return tarval_mod(tv, modulo_tv);
5463 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5464 ir_node *left, ir_node *right, ir_mode *mode);
5467 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5468 * then we can use that to minimize the value of Add(x, const) or
5469 * Sub(Const, x). In particular this often avoids 1 instruction in some
5470 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5471 * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
5473 static ir_node *transform_node_shift_modulo(ir_node *n,
5474 new_shift_func new_shift)
5476 ir_mode *mode = get_irn_mode(n);
5477 int modulo = get_mode_modulo_shift(mode);
5478 ir_node *newop = NULL;
5479 ir_mode *mode_right;
5486 if (get_mode_arithmetic(mode) != irma_twos_complement)
5488 if (!is_po2(modulo))
5491 irg = get_irn_irg(n);
5492 block = get_nodes_block(n);
5493 right = get_binop_right(n);
5494 mode_right = get_irn_mode(right);
5495 if (is_Const(right)) {
5496 ir_tarval *tv = get_Const_tarval(right);
5497 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5502 newop = new_r_Const(irg, tv_mod);
5503 } else if (is_Add(right) || is_Or_Eor_Add(right)) {
5504 ir_node *add_right = get_binop_right(right);
5505 if (is_Const(add_right)) {
5506 ir_tarval *tv = get_Const_tarval(add_right);
5507 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5512 newconst = new_r_Const(irg, tv_mod);
5513 newop = new_r_Add(block, get_binop_left(right), newconst,
5516 } else if (is_Sub(right)) {
5517 ir_node *sub_left = get_Sub_left(right);
5518 if (is_Const(sub_left)) {
5519 ir_tarval *tv = get_Const_tarval(sub_left);
5520 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5525 newconst = new_r_Const(irg, tv_mod);
5526 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5533 if (newop != NULL) {
5534 dbg_info *dbgi = get_irn_dbg_info(n);
5535 ir_node *left = get_binop_left(n);
5536 return new_shift(dbgi, block, left, newop, mode);
5544 static ir_node *transform_node_Shr(ir_node *n)
5546 ir_node *c, *oldn = n;
5547 ir_node *left = get_Shr_left(n);
5548 ir_node *right = get_Shr_right(n);
5549 ir_mode *mode = get_irn_mode(n);
5551 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5552 n = transform_node_shift(n);
5555 n = transform_node_shift_modulo(n, new_rd_Shr);
5557 n = transform_node_shl_shr(n);
5559 n = transform_node_shift_bitop(n);
5567 static ir_node *transform_node_Shrs(ir_node *n)
5570 ir_node *a = get_Shrs_left(n);
5571 ir_node *b = get_Shrs_right(n);
5572 ir_mode *mode = get_irn_mode(n);
5576 if (is_oversize_shift(n)) {
5577 ir_node *block = get_nodes_block(n);
5578 dbg_info *dbgi = get_irn_dbg_info(n);
5579 ir_mode *cmode = get_irn_mode(b);
5580 long val = get_mode_size_bits(cmode)-1;
5581 ir_graph *irg = get_irn_irg(n);
5582 ir_node *cnst = new_r_Const_long(irg, cmode, val);
5583 return new_rd_Shrs(dbgi, block, a, cnst, mode);
5586 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5587 n = transform_node_shift(n);
5591 n = transform_node_shift_modulo(n, new_rd_Shrs);
5594 n = transform_node_shift_bitop(n);
5598 /* normalisation: use Shr when sign bit is guaranteed to be cleared */
5599 attr = vrp_get_info(a);
5601 unsigned bits = get_mode_size_bits(mode);
5602 ir_tarval *scount = new_tarval_from_long(bits-1, mode_Iu);
5603 ir_tarval *sign = tarval_shl(get_mode_one(mode), scount);
5604 if (tarval_is_null(tarval_and(attr->bits_not_set, sign))) {
5605 dbg_info *dbgi = get_irn_dbg_info(n);
5606 ir_node *block = get_nodes_block(n);
5607 return new_rd_Shr(dbgi, block, a, b, mode);
5617 static ir_node *transform_node_Shl(ir_node *n)
5619 ir_node *c, *oldn = n;
5620 ir_node *a = get_Shl_left(n);
5621 ir_node *b = get_Shl_right(n);
5622 ir_mode *mode = get_irn_mode(n);
5624 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5625 n = transform_node_shift(n);
5628 n = transform_node_shift_modulo(n, new_rd_Shl);
5630 n = transform_node_shl_shr(n);
5632 n = transform_node_shift_bitop(n);
5640 static ir_node *transform_node_Rotl(ir_node *n)
5642 ir_node *c, *oldn = n;
5643 ir_node *a = get_Rotl_left(n);
5644 ir_node *b = get_Rotl_right(n);
5645 ir_mode *mode = get_irn_mode(n);
5647 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5648 n = transform_node_shift(n);
5651 n = transform_node_shift_bitop(n);
5659 static ir_node *transform_node_Conv(ir_node *n)
5661 ir_node *c, *oldn = n;
5662 ir_mode *mode = get_irn_mode(n);
5663 ir_node *a = get_Conv_op(n);
5665 if (mode != mode_b && is_const_Phi(a)) {
5666 /* Do NOT optimize mode_b Conv's, this leads to remaining
5667 * Phib nodes later, because the conv_b_lower operation
5668 * is instantly reverted, when it tries to insert a Convb.
5670 c = apply_conv_on_phi(a, mode);
5672 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5677 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5678 ir_graph *irg = get_irn_irg(n);
5679 return new_r_Unknown(irg, mode);
5682 if (mode_is_reference(mode) &&
5683 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5685 ir_node *l = get_Add_left(a);
5686 ir_node *r = get_Add_right(a);
5687 dbg_info *dbgi = get_irn_dbg_info(a);
5688 ir_node *block = get_nodes_block(n);
5690 ir_node *lop = get_Conv_op(l);
5691 if (get_irn_mode(lop) == mode) {
5692 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5693 n = new_rd_Add(dbgi, block, lop, r, mode);
5698 ir_node *rop = get_Conv_op(r);
5699 if (get_irn_mode(rop) == mode) {
5700 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5701 n = new_rd_Add(dbgi, block, l, rop, mode);
5711 * Remove dead blocks and nodes in dead blocks
5712 * in keep alive list. We do not generate a new End node.
5714 static ir_node *transform_node_End(ir_node *n)
5716 int i, j, n_keepalives = get_End_n_keepalives(n);
5719 NEW_ARR_A(ir_node *, in, n_keepalives);
5721 for (i = j = 0; i < n_keepalives; ++i) {
5722 ir_node *ka = get_End_keepalive(n, i);
5724 /* no need to keep Bad */
5727 /* do not keep unreachable code */
5728 block = is_Block(ka) ? ka : get_nodes_block(ka);
5729 if (is_block_unreachable(block))
5733 if (j != n_keepalives)
5734 set_End_keepalives(n, j, in);
5738 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5740 if (is_Minus(a) && get_Minus_op(a) == b)
5742 if (is_Minus(b) && get_Minus_op(b) == a)
5744 if (is_Sub(a) && is_Sub(b)) {
5745 ir_node *a_left = get_Sub_left(a);
5746 ir_node *a_right = get_Sub_right(a);
5747 ir_node *b_left = get_Sub_left(b);
5748 ir_node *b_right = get_Sub_right(b);
5750 if (a_left == b_right && a_right == b_left)
5757 static const ir_node *skip_upconv(const ir_node *node)
5759 while (is_Conv(node)) {
5760 ir_mode *mode = get_irn_mode(node);
5761 const ir_node *op = get_Conv_op(node);
5762 ir_mode *op_mode = get_irn_mode(op);
5763 if (!smaller_mode(op_mode, mode))
5770 int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_false,
5771 const ir_node *mux_true)
5776 ir_relation relation;
5782 * Note further that these optimization work even for floating point
5783 * with NaN's because -NaN == NaN.
5784 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5787 mode = get_irn_mode(mux_true);
5788 if (mode_honor_signed_zeros(mode))
5791 /* must be <, <=, >=, > */
5792 relation = get_Cmp_relation(sel);
5793 if ((relation & ir_relation_less_greater) == 0)
5796 if (!ir_is_negated_value(mux_true, mux_false))
5799 mux_true = skip_upconv(mux_true);
5800 mux_false = skip_upconv(mux_false);
5802 /* must be x cmp 0 */
5803 cmp_right = get_Cmp_right(sel);
5804 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
5807 cmp_left = get_Cmp_left(sel);
5808 if (cmp_left == mux_false) {
5809 if (relation & ir_relation_less) {
5812 assert(relation & ir_relation_greater);
5815 } else if (cmp_left == mux_true) {
5816 if (relation & ir_relation_less) {
5819 assert(relation & ir_relation_greater);
5827 ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_false,
5830 ir_node *cmp_left = get_Cmp_left(sel);
5831 return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
5834 bool ir_is_optimizable_mux(const ir_node *sel, const ir_node *mux_false,
5835 const ir_node *mux_true)
5837 /* this code should return true each time transform_node_Mux would
5838 * optimize the Mux completely away */
5840 ir_mode *mode = get_irn_mode(mux_false);
5841 if (get_mode_arithmetic(mode) == irma_twos_complement
5842 && ir_mux_is_abs(sel, mux_false, mux_true))
5845 if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) {
5846 const ir_node *cmp_r = get_Cmp_right(sel);
5847 const ir_node *cmp_l = get_Cmp_left(sel);
5848 const ir_node *f = mux_false;
5849 const ir_node *t = mux_true;
5851 if (is_Const(t) && is_Const_null(t)) {
5856 if (is_And(cmp_l) && f == cmp_r) {
5857 ir_node *and_r = get_And_right(cmp_l);
5860 if (and_r == t && is_single_bit(and_r))
5862 and_l = get_And_left(cmp_l);
5863 if (and_l == t && is_single_bit(and_l))
5872 * Optimize a Mux(c, 0, 1) node (sometimes called a "set" instruction)
5874 static ir_node *transform_Mux_set(ir_node *n)
5876 ir_node *cond = get_Mux_sel(n);
5881 ir_relation relation;
5895 left = get_Cmp_left(cond);
5896 mode = get_irn_mode(left);
5897 if (!mode_is_int(mode) && !mode_is_reference(mode))
5899 dest_mode = get_irn_mode(n);
5900 if (!mode_is_int(dest_mode) && !mode_is_reference(dest_mode))
5902 right = get_Cmp_right(cond);
5903 relation = get_Cmp_relation(cond) & ~ir_relation_unordered;
5904 if (get_mode_size_bits(mode) >= get_mode_size_bits(dest_mode)
5905 && !(mode_is_signed(mode) && is_Const(right) && is_Const_null(right)
5906 && relation != ir_relation_greater))
5911 case ir_relation_less:
5912 /* a < b -> (a - b) >> 31 */
5916 case ir_relation_less_equal:
5917 /* a <= b -> ~(a - b) >> 31 */
5922 case ir_relation_greater:
5923 /* a > b -> (b - a) >> 31 */
5927 case ir_relation_greater_equal:
5928 /* a >= b -> ~(a - b) >> 31 */
5937 dbgi = get_irn_dbg_info(n);
5938 block = get_nodes_block(n);
5939 irg = get_irn_irg(block);
5940 bits = get_mode_size_bits(dest_mode);
5941 tv = new_tarval_from_long(bits-1, mode_Iu);
5942 shift_cnt = new_rd_Const(dbgi, irg, tv);
5944 if (mode != dest_mode) {
5945 a = new_rd_Conv(dbgi, block, a, dest_mode);
5946 b = new_rd_Conv(dbgi, block, b, dest_mode);
5949 res = new_rd_Sub(dbgi, block, a, b, dest_mode);
5951 res = new_rd_Not(dbgi, block, res, dest_mode);
5953 res = new_rd_Shr(dbgi, block, res, shift_cnt, dest_mode);
5958 * Optimize a Mux into some simpler cases.
5960 static ir_node *transform_node_Mux(ir_node *n)
5963 ir_node *sel = get_Mux_sel(n);
5964 ir_mode *mode = get_irn_mode(n);
5965 ir_node *t = get_Mux_true(n);
5966 ir_node *f = get_Mux_false(n);
5967 ir_graph *irg = get_irn_irg(n);
5969 /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
5970 if (get_mode_arithmetic(mode) == irma_twos_complement) {
5971 int abs = ir_mux_is_abs(sel, f, t);
5973 dbg_info *dbgi = get_irn_dbg_info(n);
5974 ir_node *block = get_nodes_block(n);
5975 ir_node *op = ir_get_abs_op(sel, f, t);
5976 int bits = get_mode_size_bits(mode);
5977 ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
5978 ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
5979 ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode);
5982 res = new_rd_Sub(dbgi, block, xorn, sext, mode);
5984 res = new_rd_Sub(dbgi, block, sext, xorn, mode);
5990 /* first normalization step: try to move a constant to the false side,
5991 * 0 preferred on false side too */
5992 if (is_Cmp(sel) && is_Const(t) &&
5993 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5994 dbg_info *seldbgi = get_irn_dbg_info(sel);
5995 ir_node *block = get_nodes_block(sel);
5996 ir_relation relation = get_Cmp_relation(sel);
6001 /* Mux(x, a, b) => Mux(not(x), b, a) */
6002 relation = get_negated_relation(relation);
6003 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
6004 get_Cmp_right(sel), relation);
6005 return new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
6008 if (is_Const(f) && is_Const_null(f) && is_Const(t) && is_Const_one(t)) {
6009 n = transform_Mux_set(n);
6014 /* the following optimisations create new mode_b nodes, so only do them
6015 * before mode_b lowering */
6016 if (!is_irg_state(irg, IR_GRAPH_STATE_MODEB_LOWERED)) {
6018 ir_node* block = get_nodes_block(n);
6020 ir_node* c1 = get_Mux_sel(t);
6021 ir_node* t1 = get_Mux_true(t);
6022 ir_node* f1 = get_Mux_false(t);
6024 /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */
6025 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
6026 DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE);
6027 return new_r_Mux(block, and_, f1, t1, mode);
6028 } else if (f == t1) {
6029 /* Mux(cond0, Mux(cond1, x, y), x) */
6030 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
6031 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
6032 DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE);
6033 return new_r_Mux(block, and_, t1, f1, mode);
6035 } else if (is_Mux(f)) {
6036 ir_node* block = get_nodes_block(n);
6038 ir_node* c1 = get_Mux_sel(f);
6039 ir_node* t1 = get_Mux_true(f);
6040 ir_node* f1 = get_Mux_false(f);
6042 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
6043 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
6044 DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE);
6045 return new_r_Mux(block, or_, f1, t1, mode);
6046 } else if (t == f1) {
6047 /* Mux(cond0, x, Mux(cond1, y, x)) */
6048 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
6049 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
6050 DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE);
6051 return new_r_Mux(block, or_, t1, f1, mode);
6055 /* note: after normalization, false can only happen on default */
6056 if (mode == mode_b) {
6057 dbg_info *dbg = get_irn_dbg_info(n);
6058 ir_node *block = get_nodes_block(n);
6061 ir_tarval *tv_t = get_Const_tarval(t);
6062 if (tv_t == tarval_b_true) {
6064 /* Muxb(sel, true, false) = sel */
6065 assert(get_Const_tarval(f) == tarval_b_false);
6066 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
6069 /* Muxb(sel, true, x) = Or(sel, x) */
6070 n = new_rd_Or(dbg, block, sel, f, mode_b);
6071 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
6075 } else if (is_Const(f)) {
6076 ir_tarval *tv_f = get_Const_tarval(f);
6077 if (tv_f == tarval_b_true) {
6078 /* Muxb(sel, x, true) = Or(Not(sel), x) */
6079 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
6080 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
6081 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
6084 /* Muxb(sel, x, false) = And(sel, x) */
6085 assert(tv_f == tarval_b_false);
6086 n = new_rd_And(dbg, block, sel, t, mode_b);
6087 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
6094 if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) {
6095 ir_relation relation = get_Cmp_relation(sel);
6096 ir_node *cmp_r = get_Cmp_right(sel);
6097 ir_node *cmp_l = get_Cmp_left(sel);
6098 ir_node *block = get_nodes_block(n);
6100 if (is_And(cmp_l) && f == cmp_r) {
6101 ir_node *and_r = get_And_right(cmp_l);
6104 if (and_r == t && is_single_bit(and_r)) {
6105 if (relation == ir_relation_equal) {
6106 /* Mux((a & (1<<n)) == 0, (1<<n), 0) == (a&(1<<n)) xor ((1<<n)) */
6107 n = new_rd_Eor(get_irn_dbg_info(n),
6108 block, cmp_l, t, mode);
6109 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
6111 /* Mux((a & (1<<n)) != 0, (1<<n), 0) == a & (1<<n) */
6113 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
6117 and_l = get_And_left(cmp_l);
6118 if (and_l == t && is_single_bit(and_l)) {
6119 if (relation == ir_relation_equal) {
6120 /* ((1 << n) & a) == 0, (1 << n), 0) */
6121 n = new_rd_Eor(get_irn_dbg_info(n),
6122 block, cmp_l, t, mode);
6123 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
6125 /* ((1 << n) & a) != 0, (1 << n), 0) */
6127 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
6138 * optimize Sync nodes that have other syncs as input we simply add the inputs
6139 * of the other sync to our own inputs
6141 static ir_node *transform_node_Sync(ir_node *n)
6143 int arity = get_Sync_n_preds(n);
6146 for (i = 0; i < arity;) {
6147 ir_node *pred = get_Sync_pred(n, i);
6151 /* Remove Bad predecessors */
6158 /* Remove duplicate predecessors */
6159 for (j = 0; j < i; ++j) {
6160 if (get_Sync_pred(n, j) == pred) {
6169 if (!is_Sync(pred)) {
6177 pred_arity = get_Sync_n_preds(pred);
6178 for (j = 0; j < pred_arity; ++j) {
6179 ir_node *pred_pred = get_Sync_pred(pred, j);
6184 add_irn_n(n, pred_pred);
6188 if (get_Sync_pred(n, k) == pred_pred) break;
6194 ir_graph *irg = get_irn_irg(n);
6195 return new_r_Bad(irg, mode_M);
6198 return get_Sync_pred(n, 0);
6201 /* rehash the sync node */
6206 static ir_node *transform_node_Load(ir_node *n)
6208 /* if our memory predecessor is a load from the same address, then reuse the
6209 * previous result */
6210 ir_node *mem = get_Load_mem(n);
6215 /* don't touch volatile loads */
6216 if (get_Load_volatility(n) == volatility_is_volatile)
6218 mem_pred = get_Proj_pred(mem);
6219 if (is_Load(mem_pred)) {
6220 ir_node *pred_load = mem_pred;
6222 /* conservatively compare the 2 loads. TODO: This could be less strict
6223 * with fixup code in some situations (like smaller/bigger modes) */
6224 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
6226 if (get_Load_mode(pred_load) != get_Load_mode(n))
6228 /* all combinations of aligned/unaligned pred/n should be fine so we do
6229 * not compare the unaligned attribute */
6231 ir_node *block = get_nodes_block(n);
6232 ir_node *jmp = new_r_Jmp(block);
6233 ir_graph *irg = get_irn_irg(n);
6234 ir_node *bad = new_r_Bad(irg, mode_X);
6235 ir_mode *mode = get_Load_mode(n);
6236 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
6237 ir_node *in[] = { mem, res, jmp, bad };
6238 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
6241 } else if (is_Store(mem_pred)) {
6242 ir_node *pred_store = mem_pred;
6243 ir_node *value = get_Store_value(pred_store);
6245 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
6247 if (get_irn_mode(value) != get_Load_mode(n))
6249 /* all combinations of aligned/unaligned pred/n should be fine so we do
6250 * not compare the unaligned attribute */
6252 ir_node *block = get_nodes_block(n);
6253 ir_node *jmp = new_r_Jmp(block);
6254 ir_graph *irg = get_irn_irg(n);
6255 ir_node *bad = new_r_Bad(irg, mode_X);
6256 ir_node *res = value;
6257 ir_node *in[] = { mem, res, jmp, bad };
6258 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
6267 * optimize a trampoline Call into a direct Call
6269 static ir_node *transform_node_Call(ir_node *call)
6271 ir_node *callee = get_Call_ptr(call);
6272 ir_node *adr, *mem, *res, *bl, **in;
6273 ir_type *ctp, *mtp, *tp;
6277 size_t i, n_res, n_param;
6280 if (! is_Proj(callee))
6282 callee = get_Proj_pred(callee);
6283 if (! is_Builtin(callee))
6285 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
6288 mem = get_Call_mem(call);
6290 if (skip_Proj(mem) == callee) {
6291 /* memory is routed to the trampoline, skip */
6292 mem = get_Builtin_mem(callee);
6295 /* build a new call type */
6296 mtp = get_Call_type(call);
6297 tdb = get_type_dbg_info(mtp);
6299 n_res = get_method_n_ress(mtp);
6300 n_param = get_method_n_params(mtp);
6301 ctp = new_d_type_method(n_param + 1, n_res, tdb);
6303 for (i = 0; i < n_res; ++i)
6304 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
6306 NEW_ARR_A(ir_node *, in, n_param + 1);
6308 /* FIXME: we don't need a new pointer type in every step */
6309 irg = get_irn_irg(call);
6310 tp = get_irg_frame_type(irg);
6311 tp = new_type_pointer(tp);
6312 set_method_param_type(ctp, 0, tp);
6314 in[0] = get_Builtin_param(callee, 2);
6315 for (i = 0; i < n_param; ++i) {
6316 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
6317 in[i + 1] = get_Call_param(call, i);
6319 var = get_method_variadicity(mtp);
6320 set_method_variadicity(ctp, var);
6321 /* When we resolve a trampoline, the function must be called by a this-call */
6322 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
6323 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
6325 adr = get_Builtin_param(callee, 1);
6327 db = get_irn_dbg_info(call);
6328 bl = get_nodes_block(call);
6330 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
6331 if (get_irn_pinned(call) == op_pin_state_floats)
6332 set_irn_pinned(res, op_pin_state_floats);
6336 void firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
6340 ops->transform_node = transform_node_##a; \
6342 #define CASE_PROJ(a) \
6344 ops->transform_node_Proj = transform_node_Proj_##a; \
6346 #define CASE_PROJ_EX(a) \
6348 ops->transform_node = transform_node_##a; \
6349 ops->transform_node_Proj = transform_node_Proj_##a; \
6391 * Tries several [inplace] [optimizing] transformations and returns an
6392 * equivalent node. The difference to equivalent_node() is that these
6393 * transformations _do_ generate new nodes, and thus the old node must
6394 * not be freed even if the equivalent node isn't the old one.
6396 static ir_node *transform_node(ir_node *n)
6402 iro = get_irn_opcode_(n);
6403 /* constant expression evaluation / constant folding */
6404 if (get_opt_constant_folding()) {
6405 /* neither constants nor Tuple values can be evaluated */
6406 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6407 /* try to evaluate */
6408 ir_tarval *tv = computed_value(n);
6409 if (tv != tarval_bad) {
6410 /* evaluation was successful -- replace the node. */
6411 ir_graph *irg = get_irn_irg(n);
6413 n = new_r_Const(irg, tv);
6415 DBG_OPT_CSTEVAL(old_n, n);
6421 /* remove unnecessary nodes */
6422 if (get_opt_constant_folding() ||
6423 (iro == iro_Phi) || /* always optimize these nodes. */
6424 (iro == iro_Id) || /* ... */
6425 (iro == iro_Proj) || /* ... */
6426 (iro == iro_Block)) { /* Flags tested local. */
6427 n = equivalent_node(n);
6432 /* Some more constant expression evaluation. */
6433 if (get_opt_algebraic_simplification() ||
6434 (iro == iro_Cond) ||
6435 (iro == iro_Proj)) { /* Flags tested local. */
6436 if (n->op->ops.transform_node != NULL) {
6437 n = n->op->ops.transform_node(n);
6447 /* **************** Common Subexpression Elimination **************** */
6449 /** The size of the hash table used, should estimate the number of nodes
6451 #define N_IR_NODES 512
6453 /** Compares two exception attributes */
6454 static int node_cmp_exception(const ir_node *a, const ir_node *b)
6456 const except_attr *ea = &a->attr.except;
6457 const except_attr *eb = &b->attr.except;
6458 return ea->pin_state != eb->pin_state;
6461 /** Compares the attributes of two Const nodes. */
6462 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
6464 return get_Const_tarval(a) != get_Const_tarval(b);
6467 /** Compares the attributes of two Proj nodes. */
6468 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
6470 return a->attr.proj.proj != b->attr.proj.proj;
6473 /** Compares the attributes of two Alloc nodes. */
6474 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
6476 const alloc_attr *pa = &a->attr.alloc;
6477 const alloc_attr *pb = &b->attr.alloc;
6478 if (pa->where != pb->where || pa->type != pb->type)
6480 return node_cmp_exception(a, b);
6483 /** Compares the attributes of two Free nodes. */
6484 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
6486 const free_attr *pa = &a->attr.free;
6487 const free_attr *pb = &b->attr.free;
6488 return (pa->where != pb->where) || (pa->type != pb->type);
6491 /** Compares the attributes of two SymConst nodes. */
6492 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
6494 const symconst_attr *pa = &a->attr.symc;
6495 const symconst_attr *pb = &b->attr.symc;
6496 return (pa->kind != pb->kind)
6497 || (pa->sym.type_p != pb->sym.type_p);
6500 /** Compares the attributes of two Call nodes. */
6501 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
6503 const call_attr *pa = &a->attr.call;
6504 const call_attr *pb = &b->attr.call;
6505 if (pa->type != pb->type)
6507 return node_cmp_exception(a, b);
6510 /** Compares the attributes of two Sel nodes. */
6511 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
6513 const ir_entity *a_ent = get_Sel_entity(a);
6514 const ir_entity *b_ent = get_Sel_entity(b);
6515 return a_ent != b_ent;
6518 /** Compares the attributes of two Phi nodes. */
6519 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
6521 /* we can only enter this function if both nodes have the same number of inputs,
6522 hence it is enough to check if one of them is a Phi0 */
6524 /* check the Phi0 pos attribute */
6525 return a->attr.phi.u.pos != b->attr.phi.u.pos;
6530 /** Compares the attributes of two Conv nodes. */
6531 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
6533 return get_Conv_strict(a) != get_Conv_strict(b);
6536 /** Compares the attributes of two Cast nodes. */
6537 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
6539 return get_Cast_type(a) != get_Cast_type(b);
6542 /** Compares the attributes of two Load nodes. */
6543 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
6545 if (get_Load_volatility(a) == volatility_is_volatile ||
6546 get_Load_volatility(b) == volatility_is_volatile)
6547 /* NEVER do CSE on volatile Loads */
6549 /* do not CSE Loads with different alignment. Be conservative. */
6550 if (get_Load_unaligned(a) != get_Load_unaligned(b))
6552 if (get_Load_mode(a) != get_Load_mode(b))
6554 return node_cmp_exception(a, b);
6557 /** Compares the attributes of two Store nodes. */
6558 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
6560 /* do not CSE Stores with different alignment. Be conservative. */
6561 if (get_Store_unaligned(a) != get_Store_unaligned(b))
6563 /* NEVER do CSE on volatile Stores */
6564 if (get_Store_volatility(a) == volatility_is_volatile ||
6565 get_Store_volatility(b) == volatility_is_volatile)
6567 return node_cmp_exception(a, b);
6570 static int node_cmp_attr_CopyB(const ir_node *a, const ir_node *b)
6572 if (get_CopyB_type(a) != get_CopyB_type(b))
6575 return node_cmp_exception(a, b);
6578 static int node_cmp_attr_Bound(const ir_node *a, const ir_node *b)
6580 return node_cmp_exception(a, b);
6583 /** Compares the attributes of two Div nodes. */
6584 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
6586 const div_attr *ma = &a->attr.div;
6587 const div_attr *mb = &b->attr.div;
6588 if (ma->resmode != mb->resmode || ma->no_remainder != mb->no_remainder)
6590 return node_cmp_exception(a, b);
6593 /** Compares the attributes of two Mod nodes. */
6594 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
6596 const mod_attr *ma = &a->attr.mod;
6597 const mod_attr *mb = &b->attr.mod;
6598 if (ma->resmode != mb->resmode)
6600 return node_cmp_exception(a, b);
6603 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
6605 const cmp_attr *ma = &a->attr.cmp;
6606 const cmp_attr *mb = &b->attr.cmp;
6607 return ma->relation != mb->relation;
6610 /** Compares the attributes of two Confirm nodes. */
6611 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
6613 const confirm_attr *ma = &a->attr.confirm;
6614 const confirm_attr *mb = &b->attr.confirm;
6615 return ma->relation != mb->relation;
6618 /** Compares the attributes of two Builtin nodes. */
6619 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
6621 if (get_Builtin_kind(a) != get_Builtin_kind(b))
6623 if (get_Builtin_type(a) != get_Builtin_type(b))
6625 return node_cmp_exception(a, b);
6628 /** Compares the attributes of two ASM nodes. */
6629 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6633 const ir_asm_constraint *ca;
6634 const ir_asm_constraint *cb;
6637 if (get_ASM_text(a) != get_ASM_text(b))
6640 /* Should we really check the constraints here? Should be better, but is strange. */
6641 n = get_ASM_n_input_constraints(a);
6642 if (n != get_ASM_n_input_constraints(b))
6645 ca = get_ASM_input_constraints(a);
6646 cb = get_ASM_input_constraints(b);
6647 for (i = 0; i < n; ++i) {
6648 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6649 || ca[i].mode != cb[i].mode)
6653 n = get_ASM_n_output_constraints(a);
6654 if (n != get_ASM_n_output_constraints(b))
6657 ca = get_ASM_output_constraints(a);
6658 cb = get_ASM_output_constraints(b);
6659 for (i = 0; i < n; ++i) {
6660 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6661 || ca[i].mode != cb[i].mode)
6665 n = get_ASM_n_clobbers(a);
6666 if (n != get_ASM_n_clobbers(b))
6669 cla = get_ASM_clobbers(a);
6670 clb = get_ASM_clobbers(b);
6671 for (i = 0; i < n; ++i) {
6672 if (cla[i] != clb[i])
6676 return node_cmp_exception(a, b);
6679 /** Compares the inexistent attributes of two Dummy nodes. */
6680 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6684 /* Dummy nodes never equal by definition */
6688 static int node_cmp_attr_InstOf(const ir_node *a, const ir_node *b)
6690 if (get_InstOf_type(a) != get_InstOf_type(b))
6692 return node_cmp_exception(a, b);
6695 void firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6699 ops->node_cmp_attr = node_cmp_attr_##a; \
6733 * Compare function for two nodes in the value table. Gets two
6734 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6736 int identities_cmp(const void *elt, const void *key)
6738 ir_node *a = (ir_node *)elt;
6739 ir_node *b = (ir_node *)key;
6742 if (a == b) return 0;
6744 if ((get_irn_op(a) != get_irn_op(b)) ||
6745 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6747 /* compare if a's in and b's in are of equal length */
6748 irn_arity_a = get_irn_arity(a);
6749 if (irn_arity_a != get_irn_arity(b))
6752 /* blocks are never the same */
6756 if (get_irn_pinned(a) == op_pin_state_pinned) {
6757 /* for pinned nodes, the block inputs must be equal */
6758 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6761 ir_node *block_a = get_nodes_block(a);
6762 ir_node *block_b = get_nodes_block(b);
6763 if (! get_opt_global_cse()) {
6764 /* for block-local CSE both nodes must be in the same Block */
6765 if (block_a != block_b)
6768 /* The optimistic approach would be to do nothing here.
6769 * However doing GCSE optimistically produces a lot of partially dead code which appears
6770 * to be worse in practice than the missed opportunities.
6771 * So we use a very conservative variant here and only CSE if 1 value dominates the
6773 if (!block_dominates(block_a, block_b)
6774 && !block_dominates(block_b, block_a))
6779 /* compare a->in[0..ins] with b->in[0..ins] */
6780 for (i = 0; i < irn_arity_a; ++i) {
6781 ir_node *pred_a = get_irn_n(a, i);
6782 ir_node *pred_b = get_irn_n(b, i);
6783 if (pred_a != pred_b) {
6784 /* if both predecessors are CSE neutral they might be different */
6785 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6791 * here, we already now that the nodes are identical except their
6794 if (a->op->ops.node_cmp_attr)
6795 return a->op->ops.node_cmp_attr(a, b);
6801 * Calculate a hash value of a node.
6803 * @param node The IR-node
6805 unsigned ir_node_hash(const ir_node *node)
6807 return node->op->ops.hash(node);
6811 void new_identities(ir_graph *irg)
6813 if (irg->value_table != NULL)
6814 del_pset(irg->value_table);
6815 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6818 void del_identities(ir_graph *irg)
6820 if (irg->value_table != NULL)
6821 del_pset(irg->value_table);
6824 /* Normalize a node by putting constants (and operands with larger
6825 * node index) on the right (operator side). */
6826 void ir_normalize_node(ir_node *n)
6828 if (is_op_commutative(get_irn_op(n))) {
6829 ir_node *l = get_binop_left(n);
6830 ir_node *r = get_binop_right(n);
6832 /* For commutative operators perform a OP b == b OP a but keep
6833 * constants on the RIGHT side. This helps greatly in some
6834 * optimizations. Moreover we use the idx number to make the form
6836 if (!operands_are_normalized(l, r)) {
6837 set_binop_left(n, r);
6838 set_binop_right(n, l);
6845 * Return the canonical node computing the same value as n.
6846 * Looks up the node in a hash table, enters it in the table
6847 * if it isn't there yet.
6849 * @param n the node to look up
6851 * @return a node that computes the same value as n or n if no such
6852 * node could be found
6854 ir_node *identify_remember(ir_node *n)
6856 ir_graph *irg = get_irn_irg(n);
6857 pset *value_table = irg->value_table;
6860 if (value_table == NULL)
6863 ir_normalize_node(n);
6864 /* lookup or insert in hash table with given hash key. */
6865 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6868 /* n is reachable again */
6869 edges_node_revival(nn);
6876 * During construction we set the op_pin_state_pinned flag in the graph right
6877 * when the optimization is performed. The flag turning on procedure global
6878 * cse could be changed between two allocations. This way we are safe.
6880 * @param n The node to lookup
6882 static inline ir_node *identify_cons(ir_node *n)
6886 n = identify_remember(n);
6887 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6888 ir_graph *irg = get_irn_irg(n);
6889 set_irg_pinned(irg, op_pin_state_floats);
6894 /* Add a node to the identities value table. */
6895 void add_identities(ir_node *node)
6902 identify_remember(node);
6905 /* Visit each node in the value table of a graph. */
6906 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6909 ir_graph *rem = current_ir_graph;
6911 current_ir_graph = irg;
6912 foreach_pset(irg->value_table, ir_node*, node) {
6915 current_ir_graph = rem;
6919 * These optimizations deallocate nodes from the obstack.
6920 * It can only be called if it is guaranteed that no other nodes
6921 * reference this one, i.e., right after construction of a node.
6923 * @param n The node to optimize
6925 ir_node *optimize_node(ir_node *n)
6928 ir_graph *irg = get_irn_irg(n);
6929 unsigned iro = get_irn_opcode(n);
6932 /* Always optimize Phi nodes: part of the construction. */
6933 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6935 /* constant expression evaluation / constant folding */
6936 if (get_opt_constant_folding()) {
6937 /* neither constants nor Tuple values can be evaluated */
6938 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6939 /* try to evaluate */
6940 tv = computed_value(n);
6941 if (tv != tarval_bad) {
6946 * we MUST copy the node here temporarily, because it's still
6947 * needed for DBG_OPT_CSTEVAL
6949 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6950 oldn = (ir_node*)alloca(node_size);
6952 memcpy(oldn, n, node_size);
6953 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6955 /* ARG, copy the in array, we need it for statistics */
6956 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6958 /* note the inplace edges module */
6959 edges_node_deleted(n);
6961 /* evaluation was successful -- replace the node. */
6962 irg_kill_node(irg, n);
6963 nw = new_r_Const(irg, tv);
6965 DBG_OPT_CSTEVAL(oldn, nw);
6971 /* remove unnecessary nodes */
6972 if (get_opt_algebraic_simplification() ||
6973 (iro == iro_Phi) || /* always optimize these nodes. */
6975 (iro == iro_Proj) ||
6976 (iro == iro_Block) ) /* Flags tested local. */
6977 n = equivalent_node(n);
6979 /* Common Subexpression Elimination.
6981 * Checks whether n is already available.
6982 * The block input is used to distinguish different subexpressions. Right
6983 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6984 * subexpressions within a block.
6987 n = identify_cons(n);
6990 edges_node_deleted(oldn);
6992 /* We found an existing, better node, so we can deallocate the old node. */
6993 irg_kill_node(irg, oldn);
6997 /* Some more constant expression evaluation that does not allow to
6999 iro = get_irn_opcode(n);
7000 if (get_opt_algebraic_simplification() ||
7001 (iro == iro_Cond) ||
7002 (iro == iro_Proj)) { /* Flags tested local. */
7003 n = transform_node(n);
7006 /* Now we have a legal, useful node. Enter it in hash table for CSE */
7007 if (get_opt_cse()) {
7009 n = identify_remember(o);
7019 * These optimizations never deallocate nodes (in place). This can cause dead
7020 * nodes lying on the obstack. Remove these by a dead node elimination,
7021 * i.e., a copying garbage collection.
7023 ir_node *optimize_in_place_2(ir_node *n)
7025 if (!get_opt_optimize() && !is_Phi(n)) return n;
7030 /** common subexpression elimination **/
7031 /* Checks whether n is already available. */
7032 /* The block input is used to distinguish different subexpressions.
7033 * Right now all nodes are op_pin_state_pinned to blocks, i.e., the cse
7034 * only finds common subexpressions within a block. */
7035 if (get_opt_cse()) {
7037 n = identify_remember(n);
7040 /* we have another existing node now, we do not optimize it here */
7045 n = transform_node(n);
7047 /* Now we can verify the node, as it has no dead inputs any more. */
7050 /* Now we have a legal, useful node. Enter it in hash table for cse.
7051 * Blocks should be unique anyways. (Except the successor of start:
7052 * is cse with the start block!)
7054 * Note: This is only necessary because some of the optimisations
7055 * operate in-place (set_XXX_bla, turn_into_tuple, ...) which is considered
7056 * bad practice and should be fixed sometime.
7058 if (get_opt_cse()) {
7060 n = identify_remember(o);
7069 * Wrapper for external use, set proper status bits after optimization.
7071 ir_node *optimize_in_place(ir_node *n)
7073 ir_graph *irg = get_irn_irg(n);
7074 /* Handle graph state */
7075 assert(get_irg_phase_state(irg) != phase_building);
7077 if (get_opt_global_cse())
7078 set_irg_pinned(irg, op_pin_state_floats);
7080 /* FIXME: Maybe we could also test whether optimizing the node can
7081 change the control graph. */
7082 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE);
7083 return optimize_in_place_2(n);
7087 * Calculate a hash value of a Const node.
7089 static unsigned hash_Const(const ir_node *node)
7093 /* special value for const, as they only differ in their tarval. */
7094 h = HASH_PTR(node->attr.con.tarval);
7100 * Calculate a hash value of a SymConst node.
7102 static unsigned hash_SymConst(const ir_node *node)
7106 /* all others are pointers */
7107 h = HASH_PTR(node->attr.symc.sym.type_p);
7112 void firm_set_default_hash(unsigned code, ir_op_ops *ops)
7116 ops->hash = hash_##a; \
7119 /* hash function already set */
7120 if (ops->hash != NULL)
7127 /* use input/mode default hash if no function was given */
7128 ops->hash = firm_default_hash;