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
32 #include "irgraph_t.h"
33 #include "iredges_t.h"
39 #include "iroptimize.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
52 #include "firm_types.h"
53 #include "bitfiddle.h"
56 /* Make types visible to allow most efficient access */
60 * Returns the tarval of a Const node or tarval_bad for all other nodes.
62 static ir_tarval *default_value_of(const ir_node *n)
65 return get_Const_tarval(n); /* might return tarval_bad */
70 value_of_func value_of_ptr = default_value_of;
72 /* * Set a new value_of function. */
73 void set_value_of_func(value_of_func func)
78 value_of_ptr = default_value_of;
82 * Return the value of a Constant.
84 static ir_tarval *computed_value_Const(const ir_node *n)
86 return get_Const_tarval(n);
87 } /* computed_value_Const */
90 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
92 static ir_tarval *computed_value_SymConst(const ir_node *n)
97 switch (get_SymConst_kind(n)) {
98 case symconst_type_size:
99 type = get_SymConst_type(n);
100 if (get_type_state(type) == layout_fixed)
101 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
103 case symconst_type_align:
104 type = get_SymConst_type(n);
105 if (get_type_state(type) == layout_fixed)
106 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
108 case symconst_ofs_ent:
109 ent = get_SymConst_entity(n);
110 type = get_entity_owner(ent);
111 if (get_type_state(type) == layout_fixed)
112 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
118 } /* computed_value_SymConst */
121 * Return the value of an Add.
123 static ir_tarval *computed_value_Add(const ir_node *n)
125 ir_node *a = get_Add_left(n);
126 ir_node *b = get_Add_right(n);
128 ir_tarval *ta = value_of(a);
129 ir_tarval *tb = value_of(b);
131 if ((ta != tarval_bad) && (tb != tarval_bad))
132 return tarval_add(ta, tb);
135 if ((is_Not(a) && get_Not_op(a) == b)
136 || (is_Not(b) && get_Not_op(b) == a)) {
137 return get_mode_all_one(get_irn_mode(n));
141 } /* computed_value_Add */
144 * Return the value of a Sub.
145 * Special case: a - a
147 static ir_tarval *computed_value_Sub(const ir_node *n)
149 ir_mode *mode = get_irn_mode(n);
150 ir_node *a = get_Sub_left(n);
151 ir_node *b = get_Sub_right(n);
156 if (! mode_is_float(mode)) {
159 return get_mode_null(mode);
165 if ((ta != tarval_bad) && (tb != tarval_bad))
166 return tarval_sub(ta, tb, mode);
169 } /* computed_value_Sub */
172 * Return the value of a Carry.
173 * Special : a op 0, 0 op b
175 static ir_tarval *computed_value_Carry(const ir_node *n)
177 ir_node *a = get_binop_left(n);
178 ir_node *b = get_binop_right(n);
179 ir_mode *m = get_irn_mode(n);
180 ir_tarval *ta = value_of(a);
181 ir_tarval *tb = value_of(b);
183 if ((ta != tarval_bad) && (tb != tarval_bad)) {
185 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
187 if (tarval_is_null(ta) || tarval_is_null(tb))
188 return get_mode_null(m);
191 } /* computed_value_Carry */
194 * Return the value of a Borrow.
197 static ir_tarval *computed_value_Borrow(const ir_node *n)
199 ir_node *a = get_binop_left(n);
200 ir_node *b = get_binop_right(n);
201 ir_mode *m = get_irn_mode(n);
202 ir_tarval *ta = value_of(a);
203 ir_tarval *tb = value_of(b);
205 if ((ta != tarval_bad) && (tb != tarval_bad)) {
206 return tarval_cmp(ta, tb) == ir_relation_less ? get_mode_one(m) : get_mode_null(m);
207 } else if (tarval_is_null(ta)) {
208 return get_mode_null(m);
211 } /* computed_value_Borrow */
214 * Return the value of an unary Minus.
216 static ir_tarval *computed_value_Minus(const ir_node *n)
218 ir_node *a = get_Minus_op(n);
219 ir_tarval *ta = value_of(a);
221 if (ta != tarval_bad)
222 return tarval_neg(ta);
225 } /* computed_value_Minus */
228 * Return the value of a Mul.
230 static ir_tarval *computed_value_Mul(const ir_node *n)
232 ir_node *a = get_Mul_left(n);
233 ir_node *b = get_Mul_right(n);
234 ir_tarval *ta = value_of(a);
235 ir_tarval *tb = value_of(b);
238 mode = get_irn_mode(n);
239 if (mode != get_irn_mode(a)) {
240 /* n * n = 2n bit multiplication */
241 ta = tarval_convert_to(ta, mode);
242 tb = tarval_convert_to(tb, mode);
245 if (ta != tarval_bad && tb != tarval_bad) {
246 return tarval_mul(ta, tb);
248 /* a * 0 != 0 if a == NaN or a == Inf */
249 if (!mode_is_float(mode)) {
250 /* a*0 = 0 or 0*b = 0 */
251 if (ta == get_mode_null(mode))
253 if (tb == get_mode_null(mode))
258 } /* computed_value_Mul */
261 * Return the value of an And.
262 * Special case: a & 0, 0 & b
264 static ir_tarval *computed_value_And(const ir_node *n)
266 ir_node *a = get_And_left(n);
267 ir_node *b = get_And_right(n);
268 ir_tarval *ta = value_of(a);
269 ir_tarval *tb = value_of(b);
271 if ((ta != tarval_bad) && (tb != tarval_bad)) {
272 return tarval_and (ta, tb);
275 if (tarval_is_null(ta)) return ta;
276 if (tarval_is_null(tb)) return tb;
279 if ((is_Not(a) && get_Not_op(a) == b)
280 || (is_Not(b) && get_Not_op(b) == a)) {
281 return get_mode_null(get_irn_mode(n));
285 } /* computed_value_And */
288 * Return the value of an Or.
289 * Special case: a | 1...1, 1...1 | b
291 static ir_tarval *computed_value_Or(const ir_node *n)
293 ir_node *a = get_Or_left(n);
294 ir_node *b = get_Or_right(n);
295 ir_tarval *ta = value_of(a);
296 ir_tarval *tb = value_of(b);
298 if ((ta != tarval_bad) && (tb != tarval_bad)) {
299 return tarval_or (ta, tb);
302 if (tarval_is_all_one(ta)) return ta;
303 if (tarval_is_all_one(tb)) return tb;
306 if ((is_Not(a) && get_Not_op(a) == b)
307 || (is_Not(b) && get_Not_op(b) == a)) {
308 return get_mode_all_one(get_irn_mode(n));
311 } /* computed_value_Or */
314 * Return the value of an Eor.
316 static ir_tarval *computed_value_Eor(const ir_node *n)
318 ir_node *a = get_Eor_left(n);
319 ir_node *b = get_Eor_right(n);
324 return get_mode_null(get_irn_mode(n));
326 if ((is_Not(a) && get_Not_op(a) == b)
327 || (is_Not(b) && get_Not_op(b) == a)) {
328 return get_mode_all_one(get_irn_mode(n));
334 if ((ta != tarval_bad) && (tb != tarval_bad)) {
335 return tarval_eor(ta, tb);
338 } /* computed_value_Eor */
341 * Return the value of a Not.
343 static ir_tarval *computed_value_Not(const ir_node *n)
345 ir_node *a = get_Not_op(n);
346 ir_tarval *ta = value_of(a);
348 if (ta != tarval_bad)
349 return tarval_not(ta);
352 } /* computed_value_Not */
355 * Tests whether a shift shifts more bits than available in the mode
357 static bool is_oversize_shift(const ir_node *n)
359 ir_node *count = get_binop_right(n);
360 ir_mode *mode = get_irn_mode(n);
361 ir_tarval *tv = value_of(count);
364 if (tv == tarval_bad)
366 if (!tarval_is_long(tv))
368 shiftval = get_tarval_long(tv);
369 modulo_shift = get_mode_modulo_shift(mode);
370 if (shiftval < 0 || (modulo_shift > 0 && shiftval >= modulo_shift))
373 return shiftval >= (long)get_mode_size_bits(mode);
377 * Return the value of a Shl.
379 static ir_tarval *computed_value_Shl(const ir_node *n)
381 ir_node *a = get_Shl_left(n);
382 ir_node *b = get_Shl_right(n);
384 ir_tarval *ta = value_of(a);
385 ir_tarval *tb = value_of(b);
387 if ((ta != tarval_bad) && (tb != tarval_bad)) {
388 return tarval_shl(ta, tb);
391 if (is_oversize_shift(n))
392 return get_mode_null(get_irn_mode(n));
395 } /* computed_value_Shl */
398 * Return the value of a Shr.
400 static ir_tarval *computed_value_Shr(const ir_node *n)
402 ir_node *a = get_Shr_left(n);
403 ir_node *b = get_Shr_right(n);
405 ir_tarval *ta = value_of(a);
406 ir_tarval *tb = value_of(b);
408 if ((ta != tarval_bad) && (tb != tarval_bad)) {
409 return tarval_shr(ta, tb);
411 if (is_oversize_shift(n))
412 return get_mode_null(get_irn_mode(n));
415 } /* computed_value_Shr */
418 * Return the value of a Shrs.
420 static ir_tarval *computed_value_Shrs(const ir_node *n)
422 ir_node *a = get_Shrs_left(n);
423 ir_node *b = get_Shrs_right(n);
425 ir_tarval *ta = value_of(a);
426 ir_tarval *tb = value_of(b);
428 if ((ta != tarval_bad) && (tb != tarval_bad)) {
429 return tarval_shrs(ta, tb);
432 } /* computed_value_Shrs */
435 * Return the value of a Rotl.
437 static ir_tarval *computed_value_Rotl(const ir_node *n)
439 ir_node *a = get_Rotl_left(n);
440 ir_node *b = get_Rotl_right(n);
442 ir_tarval *ta = value_of(a);
443 ir_tarval *tb = value_of(b);
445 if ((ta != tarval_bad) && (tb != tarval_bad)) {
446 return tarval_rotl(ta, tb);
449 } /* computed_value_Rotl */
451 bool ir_zero_when_converted(const ir_node *node, ir_mode *dest_mode)
453 ir_mode *mode = get_irn_mode(node);
454 if (get_mode_arithmetic(mode) != irma_twos_complement
455 || get_mode_arithmetic(dest_mode) != irma_twos_complement)
459 ir_node *count = get_Shl_right(node);
460 if (is_Const(count)) {
461 ir_tarval *tv = get_Const_tarval(count);
462 if (tarval_is_long(tv)) {
463 long shiftval = get_tarval_long(tv);
464 long destbits = get_mode_size_bits(dest_mode);
465 if (shiftval >= destbits
466 && shiftval < (long)get_mode_modulo_shift(mode))
472 ir_node *right = get_And_right(node);
473 if (is_Const(right)) {
474 ir_tarval *tv = get_Const_tarval(right);
475 ir_tarval *conved = tarval_convert_to(tv, dest_mode);
476 return tarval_is_null(conved);
483 * Return the value of a Conv.
485 static ir_tarval *computed_value_Conv(const ir_node *n)
487 ir_node *a = get_Conv_op(n);
488 ir_tarval *ta = value_of(a);
489 ir_mode *mode = get_irn_mode(n);
491 if (ta != tarval_bad)
492 return tarval_convert_to(ta, get_irn_mode(n));
494 if (ir_zero_when_converted(a, mode))
495 return get_mode_null(mode);
498 } /* computed_value_Conv */
501 * Calculate the value of a Mux: can be evaluated, if the
502 * sel and the right input are known.
504 static ir_tarval *computed_value_Mux(const ir_node *n)
506 ir_node *sel = get_Mux_sel(n);
507 ir_tarval *ts = value_of(sel);
509 if (ts == get_tarval_b_true()) {
510 ir_node *v = get_Mux_true(n);
513 else if (ts == get_tarval_b_false()) {
514 ir_node *v = get_Mux_false(n);
518 } /* computed_value_Mux */
521 * Calculate the value of a Confirm: can be evaluated,
522 * if it has the form Confirm(x, '=', Const).
524 static ir_tarval *computed_value_Confirm(const ir_node *n)
526 if (get_Confirm_relation(n) == ir_relation_equal) {
527 ir_tarval *tv = value_of(get_Confirm_bound(n));
528 if (tv != tarval_bad)
531 return value_of(get_Confirm_value(n));
532 } /* computed_value_Confirm */
535 * gives a (conservative) estimation of possible relation when comparing
538 ir_relation ir_get_possible_cmp_relations(const ir_node *left,
539 const ir_node *right)
541 ir_relation possible = ir_relation_true;
542 ir_tarval *tv_l = value_of(left);
543 ir_tarval *tv_r = value_of(right);
544 ir_mode *mode = get_irn_mode(left);
545 ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode);
546 ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode);
548 /* both values known - evaluate them */
549 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
550 possible = tarval_cmp(tv_l, tv_r);
551 /* we can return now, won't get any better */
554 /* a == a is never less or greater (but might be equal or unordered) */
556 possible &= ~ir_relation_less_greater;
557 /* unordered results only happen for float compares */
558 if (!mode_is_float(mode))
559 possible &= ~ir_relation_unordered;
560 /* values can never be less than the least representable number or
561 * greater than the greatest representable number */
563 possible &= ~ir_relation_greater;
565 possible &= ~ir_relation_less;
567 possible &= ~ir_relation_greater;
569 possible &= ~ir_relation_less;
570 /* maybe vrp can tell us more */
571 possible &= vrp_cmp(left, right);
572 /* Alloc nodes never return null (but throw an exception) */
573 if (is_Alloc(left) && tarval_is_null(tv_r))
574 possible &= ~ir_relation_equal;
580 * Return the value of a Cmp.
582 * The basic idea here is to determine which relations are possible and which
583 * one are definitely impossible.
585 static ir_tarval *computed_value_Cmp(const ir_node *cmp)
587 ir_node *left = get_Cmp_left(cmp);
588 ir_node *right = get_Cmp_right(cmp);
589 ir_relation possible = ir_get_possible_cmp_relations(left, right);
590 ir_relation relation = get_Cmp_relation(cmp);
592 /* if none of the requested relations is possible, return false */
593 if ((possible & relation) == ir_relation_false)
594 return tarval_b_false;
595 /* if possible relations are a subset of the requested ones return true */
596 if ((possible & ~relation) == ir_relation_false)
597 return tarval_b_true;
599 return computed_value_Cmp_Confirm(cmp, left, right, relation);
603 * Calculate the value of an integer Div.
604 * Special case: 0 / b
606 static ir_tarval *do_computed_value_Div(const ir_node *div)
608 const ir_node *a = get_Div_left(div);
609 const ir_node *b = get_Div_right(div);
610 const ir_mode *mode = get_Div_resmode(div);
611 ir_tarval *ta = value_of(a);
613 const ir_node *dummy;
615 /* cannot optimize 0 / b = 0 because of NaN */
616 if (!mode_is_float(mode)) {
617 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
618 return ta; /* 0 / b == 0 if b != 0 */
621 if (ta != tarval_bad && tb != tarval_bad)
622 return tarval_div(ta, tb);
624 } /* do_computed_value_Div */
627 * Calculate the value of an integer Mod of two nodes.
628 * Special case: a % 1
630 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
632 ir_tarval *ta = value_of(a);
633 ir_tarval *tb = value_of(b);
635 /* Compute a % 1 or c1 % c2 */
636 if (tarval_is_one(tb))
637 return get_mode_null(get_irn_mode(a));
638 if (ta != tarval_bad && tb != tarval_bad)
639 return tarval_mod(ta, tb);
641 } /* do_computed_value_Mod */
644 * Return the value of a Proj(Div).
646 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
648 long proj_nr = get_Proj_proj(n);
649 if (proj_nr != pn_Div_res)
652 return do_computed_value_Div(get_Proj_pred(n));
653 } /* computed_value_Proj_Div */
656 * Return the value of a Proj(Mod).
658 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
660 long proj_nr = get_Proj_proj(n);
662 if (proj_nr == pn_Mod_res) {
663 const ir_node *mod = get_Proj_pred(n);
664 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
667 } /* computed_value_Proj_Mod */
670 * Return the value of a Proj.
672 static ir_tarval *computed_value_Proj(const ir_node *proj)
674 ir_node *n = get_Proj_pred(proj);
676 if (n->op->ops.computed_value_Proj != NULL)
677 return n->op->ops.computed_value_Proj(proj);
679 } /* computed_value_Proj */
682 * If the parameter n can be computed, return its value, else tarval_bad.
683 * Performs constant folding.
685 * @param n The node this should be evaluated
687 ir_tarval *computed_value(const ir_node *n)
689 vrp_attr *vrp = vrp_get_info(n);
690 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == ir_relation_equal) {
691 return vrp->bits_set;
693 if (n->op->ops.computed_value)
694 return n->op->ops.computed_value(n);
696 } /* computed_value */
699 * Set the default computed_value evaluator in an ir_op_ops.
701 * @param code the opcode for the default operation
702 * @param ops the operations initialized
707 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
711 ops->computed_value = computed_value_##a; \
713 #define CASE_PROJ(a) \
715 ops->computed_value_Proj = computed_value_Proj_##a; \
750 } /* firm_set_default_computed_value */
753 * Optimize operations that are commutative and have neutral 0,
754 * so a op 0 = 0 op a = a.
756 static ir_node *equivalent_node_neutral_zero(ir_node *n)
760 ir_node *a = get_binop_left(n);
761 ir_node *b = get_binop_right(n);
766 /* After running compute_node there is only one constant predecessor.
767 Find this predecessors value and remember the other node: */
768 if ((tv = value_of(a)) != tarval_bad) {
770 } else if ((tv = value_of(b)) != tarval_bad) {
775 /* If this predecessors constant value is zero, the operation is
776 * unnecessary. Remove it.
778 * Beware: If n is a Add, the mode of on and n might be different
779 * which happens in this rare construction: NULL + 3.
780 * Then, a Conv would be needed which we cannot include here.
782 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
785 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
789 } /* equivalent_node_neutral_zero */
792 * Eor is commutative and has neutral 0.
794 static ir_node *equivalent_node_Eor(ir_node *n)
800 n = equivalent_node_neutral_zero(n);
801 if (n != oldn) return n;
804 b = get_Eor_right(n);
807 ir_node *aa = get_Eor_left(a);
808 ir_node *ab = get_Eor_right(a);
811 /* (a ^ b) ^ a -> b */
813 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
815 } else if (ab == b) {
816 /* (a ^ b) ^ b -> a */
818 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
823 ir_node *ba = get_Eor_left(b);
824 ir_node *bb = get_Eor_right(b);
827 /* a ^ (a ^ b) -> b */
829 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
831 } else if (bb == a) {
832 /* a ^ (b ^ a) -> b */
834 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
842 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
844 * The second one looks strange, but this construct
845 * is used heavily in the LCC sources :-).
847 * Beware: The Mode of an Add may be different than the mode of its
848 * predecessors, so we could not return a predecessors in all cases.
850 static ir_node *equivalent_node_Add(ir_node *n)
853 ir_node *left, *right;
854 ir_mode *mode = get_irn_mode(n);
856 n = equivalent_node_neutral_zero(n);
860 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
861 if (mode_is_float(mode)) {
862 ir_graph *irg = get_irn_irg(n);
863 if (get_irg_fp_model(irg) & fp_strict_algebraic)
867 left = get_Add_left(n);
868 right = get_Add_right(n);
871 if (get_Sub_right(left) == right) {
874 n = get_Sub_left(left);
875 if (mode == get_irn_mode(n)) {
876 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
882 if (get_Sub_right(right) == left) {
885 n = get_Sub_left(right);
886 if (mode == get_irn_mode(n)) {
887 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
893 } /* equivalent_node_Add */
896 * optimize operations that are not commutative but have neutral 0 on left,
899 static ir_node *equivalent_node_left_zero(ir_node *n)
903 ir_node *a = get_binop_left(n);
904 ir_node *b = get_binop_right(n);
905 ir_tarval *tb = value_of(b);
907 if (tarval_is_null(tb)) {
910 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
913 } /* equivalent_node_left_zero */
915 #define equivalent_node_Shl equivalent_node_left_zero
916 #define equivalent_node_Shr equivalent_node_left_zero
917 #define equivalent_node_Shrs equivalent_node_left_zero
918 #define equivalent_node_Rotl equivalent_node_left_zero
921 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
923 * The second one looks strange, but this construct
924 * is used heavily in the LCC sources :-).
926 * Beware: The Mode of a Sub may be different than the mode of its
927 * predecessors, so we could not return a predecessors in all cases.
929 static ir_node *equivalent_node_Sub(ir_node *n)
933 ir_mode *mode = get_irn_mode(n);
936 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
937 if (mode_is_float(mode)) {
938 ir_graph *irg = get_irn_irg(n);
939 if (get_irg_fp_model(irg) & fp_strict_algebraic)
943 b = get_Sub_right(n);
946 /* Beware: modes might be different */
947 if (tarval_is_null(tb)) {
948 ir_node *a = get_Sub_left(n);
949 if (mode == get_irn_mode(a)) {
952 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
956 } /* equivalent_node_Sub */
960 * Optimize an "self-inverse unary op", i.e. op(op(n)) = n.
963 * -(-a) == a, but might overflow two times.
964 * We handle it anyway here but the better way would be a
965 * flag. This would be needed for Pascal for instance.
967 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
970 ir_node *pred = get_unop_op(n);
972 /* optimize symmetric unop */
973 if (get_irn_op(pred) == get_irn_op(n)) {
974 n = get_unop_op(pred);
975 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
978 } /* equivalent_node_idempotent_unop */
980 /** Optimize Not(Not(x)) == x. */
981 #define equivalent_node_Not equivalent_node_idempotent_unop
983 /** -(-x) == x ??? Is this possible or can --x raise an
984 out of bounds exception if min =! max? */
985 #define equivalent_node_Minus equivalent_node_idempotent_unop
988 * Optimize a * 1 = 1 * a = a.
990 static ir_node *equivalent_node_Mul(ir_node *n)
993 ir_node *a = get_Mul_left(n);
995 /* we can handle here only the n * n = n bit cases */
996 if (get_irn_mode(n) == get_irn_mode(a)) {
997 ir_node *b = get_Mul_right(n);
1001 * Mul is commutative and has again an other neutral element.
1002 * Constants are place right, so check this case first.
1005 if (tarval_is_one(tv)) {
1007 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1010 if (tarval_is_one(tv)) {
1012 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1017 } /* equivalent_node_Mul */
1020 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1022 static ir_node *equivalent_node_Or(ir_node *n)
1026 ir_node *a = get_Or_left(n);
1027 ir_node *b = get_Or_right(n);
1031 n = a; /* idempotence */
1032 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1035 /* constants are normalized to right, check this side first */
1037 if (tarval_is_null(tv)) {
1039 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1043 if (tarval_is_null(tv)) {
1045 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1050 } /* equivalent_node_Or */
1053 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1055 static ir_node *equivalent_node_And(ir_node *n)
1059 ir_node *a = get_And_left(n);
1060 ir_node *b = get_And_right(n);
1064 n = a; /* idempotence */
1065 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1068 /* constants are normalized to right, check this side first */
1070 if (tarval_is_all_one(tv)) {
1072 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1075 if (tv != get_tarval_bad()) {
1076 ir_mode *mode = get_irn_mode(n);
1077 if (!mode_is_signed(mode) && is_Conv(a)) {
1078 ir_node *convop = get_Conv_op(a);
1079 ir_mode *convopmode = get_irn_mode(convop);
1080 if (!mode_is_signed(convopmode)) {
1081 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1082 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1084 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1091 if (tarval_is_all_one(tv)) {
1093 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1097 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1100 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1105 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1108 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1113 } /* equivalent_node_And */
1116 * Try to remove useless Conv's:
1118 static ir_node *equivalent_node_Conv(ir_node *n)
1121 ir_node *a = get_Conv_op(n);
1123 ir_mode *n_mode = get_irn_mode(n);
1124 ir_mode *a_mode = get_irn_mode(a);
1127 if (n_mode == a_mode) { /* No Conv necessary */
1128 if (get_Conv_strict(n)) {
1131 /* neither Minus nor Confirm change the precision,
1132 so we can "look-through" */
1135 p = get_Minus_op(p);
1136 } else if (is_Confirm(p)) {
1137 p = get_Confirm_value(p);
1143 if (is_Conv(p) && get_Conv_strict(p)) {
1144 /* we known already, that a_mode == n_mode, and neither
1145 Minus change the mode, so the second Conv
1147 assert(get_irn_mode(p) == n_mode);
1149 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1153 ir_node *pred = get_Proj_pred(p);
1154 if (is_Load(pred)) {
1155 /* Loads always return with the exact precision of n_mode */
1156 assert(get_Load_mode(pred) == n_mode);
1158 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1161 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1162 pred = get_Proj_pred(pred);
1163 if (is_Start(pred)) {
1164 /* Arguments always return with the exact precision,
1165 as strictConv's are place before Call -- if the
1166 caller was compiled with the same setting.
1167 Otherwise, the semantics is probably still right. */
1168 assert(get_irn_mode(p) == n_mode);
1170 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1176 /* special case: the immediate predecessor is also a Conv */
1177 if (! get_Conv_strict(a)) {
1178 /* first one is not strict, kick it */
1180 a_mode = get_irn_mode(a);
1184 /* else both are strict conv, second is superfluous */
1186 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1191 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1194 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1195 ir_node *b = get_Conv_op(a);
1196 ir_mode *b_mode = get_irn_mode(b);
1198 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1199 /* both are strict conv */
1200 if (smaller_mode(a_mode, n_mode)) {
1201 /* both are strict, but the first is smaller, so
1202 the second cannot remove more precision, remove the
1204 set_Conv_strict(n, 0);
1207 if (n_mode == b_mode) {
1208 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1209 if (n_mode == mode_b) {
1210 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1211 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1213 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1214 if (values_in_mode(b_mode, a_mode)) {
1215 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1216 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1221 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1222 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1223 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1224 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1226 if (float_mantissa >= int_mantissa) {
1228 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1233 if (smaller_mode(b_mode, a_mode)) {
1234 if (get_Conv_strict(n))
1235 set_Conv_strict(b, 1);
1236 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1237 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1244 } /* equivalent_node_Conv */
1247 * - fold Phi-nodes, iff they have only one predecessor except
1250 static ir_node *equivalent_node_Phi(ir_node *n)
1255 ir_node *first_val = NULL; /* to shutup gcc */
1257 if (!get_opt_optimize() &&
1258 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1261 n_preds = get_Phi_n_preds(n);
1263 /* Phi of dead Region without predecessors. */
1267 /* Find first non-self-referencing input */
1268 for (i = 0; i < n_preds; ++i) {
1269 first_val = get_Phi_pred(n, i);
1270 /* not self pointer */
1271 if (first_val != n) {
1272 /* then found first value. */
1277 /* search for rest of inputs, determine if any of these
1278 are non-self-referencing */
1279 while (++i < n_preds) {
1280 ir_node *scnd_val = get_Phi_pred(n, i);
1281 if (scnd_val != n && scnd_val != first_val) {
1286 if (i >= n_preds && !is_Dummy(first_val)) {
1287 /* Fold, if no multiple distinct non-self-referencing inputs */
1289 DBG_OPT_PHI(oldn, n);
1292 } /* equivalent_node_Phi */
1295 * Optimize Proj(Tuple).
1297 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1299 ir_node *oldn = proj;
1300 ir_node *tuple = get_Proj_pred(proj);
1302 /* Remove the Tuple/Proj combination. */
1303 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1304 DBG_OPT_TUPLE(oldn, tuple, proj);
1307 } /* equivalent_node_Proj_Tuple */
1310 * Optimize a / 1 = a.
1312 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1314 ir_node *oldn = proj;
1315 ir_node *div = get_Proj_pred(proj);
1316 ir_node *b = get_Div_right(div);
1317 ir_tarval *tb = value_of(b);
1319 /* Div is not commutative. */
1320 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1321 switch (get_Proj_proj(proj)) {
1323 proj = get_Div_mem(div);
1324 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1328 proj = get_Div_left(div);
1329 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1333 /* we cannot replace the exception Proj's here, this is done in
1334 transform_node_Proj_Div() */
1339 } /* equivalent_node_Proj_Div */
1342 * Optimize CopyB(mem, x, x) into a Nop.
1344 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1346 ir_node *oldn = proj;
1347 ir_node *copyb = get_Proj_pred(proj);
1348 ir_node *a = get_CopyB_dst(copyb);
1349 ir_node *b = get_CopyB_src(copyb);
1352 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1353 switch (get_Proj_proj(proj)) {
1355 proj = get_CopyB_mem(copyb);
1356 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1361 } /* equivalent_node_Proj_CopyB */
1364 * Optimize Bounds(idx, idx, upper) into idx.
1366 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1368 ir_node *oldn = proj;
1369 ir_node *bound = get_Proj_pred(proj);
1370 ir_node *idx = get_Bound_index(bound);
1371 ir_node *pred = skip_Proj(idx);
1374 if (idx == get_Bound_lower(bound))
1376 else if (is_Bound(pred)) {
1378 * idx was Bounds checked previously, it is still valid if
1379 * lower <= pred_lower && pred_upper <= upper.
1381 ir_node *lower = get_Bound_lower(bound);
1382 ir_node *upper = get_Bound_upper(bound);
1383 if (get_Bound_lower(pred) == lower &&
1384 get_Bound_upper(pred) == upper) {
1386 * One could expect that we simply return the previous
1387 * Bound here. However, this would be wrong, as we could
1388 * add an exception Proj to a new location then.
1389 * So, we must turn in into a tuple.
1395 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1396 switch (get_Proj_proj(proj)) {
1398 DBG_OPT_EXC_REM(proj);
1399 proj = get_Bound_mem(bound);
1403 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1406 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1411 } /* equivalent_node_Proj_Bound */
1414 * Does all optimizations on nodes that must be done on its Projs
1415 * because of creating new nodes.
1417 static ir_node *equivalent_node_Proj(ir_node *proj)
1419 ir_node *n = get_Proj_pred(proj);
1420 if (n->op->ops.equivalent_node_Proj)
1421 return n->op->ops.equivalent_node_Proj(proj);
1423 } /* equivalent_node_Proj */
1428 static ir_node *equivalent_node_Id(ir_node *n)
1436 DBG_OPT_ID(oldn, n);
1438 } /* equivalent_node_Id */
1443 static ir_node *equivalent_node_Mux(ir_node *n)
1445 ir_node *oldn = n, *sel = get_Mux_sel(n);
1447 ir_tarval *ts = value_of(sel);
1449 /* Mux(true, f, t) == t */
1450 if (ts == tarval_b_true) {
1451 n = get_Mux_true(n);
1452 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1455 /* Mux(false, f, t) == f */
1456 if (ts == tarval_b_false) {
1457 n = get_Mux_false(n);
1458 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1461 n_t = get_Mux_true(n);
1462 n_f = get_Mux_false(n);
1464 /* Mux(v, x, T) == x */
1465 if (is_Unknown(n_f)) {
1467 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1470 /* Mux(v, T, x) == x */
1471 if (is_Unknown(n_t)) {
1473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1477 /* Mux(v, x, x) == x */
1480 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1483 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1484 ir_relation relation = get_Cmp_relation(sel);
1485 ir_node *f = get_Mux_false(n);
1486 ir_node *t = get_Mux_true(n);
1489 * Note further that these optimization work even for floating point
1490 * with NaN's because -NaN == NaN.
1491 * However, if +0 and -0 is handled differently, we cannot use the first one.
1493 ir_node *const cmp_l = get_Cmp_left(sel);
1494 ir_node *const cmp_r = get_Cmp_right(sel);
1497 case ir_relation_equal:
1498 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1499 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1501 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1506 case ir_relation_less_greater:
1507 case ir_relation_unordered_less_greater:
1508 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1509 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1520 * Note: normalization puts the constant on the right side,
1521 * so we check only one case.
1523 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1524 /* Mux(t CMP 0, X, t) */
1525 if (is_Minus(f) && get_Minus_op(f) == t) {
1526 /* Mux(t CMP 0, -t, t) */
1527 if (relation == ir_relation_equal) {
1528 /* Mux(t == 0, -t, t) ==> -t */
1530 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1531 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1532 /* Mux(t != 0, -t, t) ==> t */
1534 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1544 * Remove Confirm nodes if setting is on.
1545 * Replace Confirms(x, '=', Constlike) by Constlike.
1547 static ir_node *equivalent_node_Confirm(ir_node *n)
1549 ir_node *pred = get_Confirm_value(n);
1550 ir_relation relation = get_Confirm_relation(n);
1552 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1554 * rare case: two identical Confirms one after another,
1555 * replace the second one with the first.
1558 pred = get_Confirm_value(n);
1564 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1565 * perform no actual computation, as, e.g., the Id nodes. It does not create
1566 * new nodes. It is therefore safe to free n if the node returned is not n.
1567 * If a node returns a Tuple we can not just skip it. If the size of the
1568 * in array fits, we transform n into a tuple (e.g., Div).
1570 ir_node *equivalent_node(ir_node *n)
1572 if (n->op->ops.equivalent_node)
1573 return n->op->ops.equivalent_node(n);
1575 } /* equivalent_node */
1578 * Sets the default equivalent node operation for an ir_op_ops.
1580 * @param code the opcode for the default operation
1581 * @param ops the operations initialized
1586 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1590 ops->equivalent_node = equivalent_node_##a; \
1592 #define CASE_PROJ(a) \
1594 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1628 } /* firm_set_default_equivalent_node */
1631 * Returns non-zero if a node is a Phi node
1632 * with all predecessors constant.
1634 static int is_const_Phi(ir_node *n)
1638 if (! is_Phi(n) || get_irn_arity(n) == 0)
1640 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1641 if (! is_Const(get_irn_n(n, i)))
1645 } /* is_const_Phi */
1647 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1648 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1651 * in reality eval_func should be tarval (*eval_func)() but incomplete
1652 * declarations are bad style and generate noisy warnings
1654 typedef void (*eval_func)(void);
1657 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1659 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1661 if (eval == (eval_func) tarval_sub) {
1662 tarval_sub_type func = (tarval_sub_type)eval;
1664 return func(a, b, mode);
1666 tarval_binop_type func = (tarval_binop_type)eval;
1673 * Apply an evaluator on a binop with a constant operators (and one Phi).
1675 * @param phi the Phi node
1676 * @param other the other operand
1677 * @param eval an evaluator function
1678 * @param mode the mode of the result, may be different from the mode of the Phi!
1679 * @param left if non-zero, other is the left operand, else the right
1681 * @return a new Phi node if the conversion was successful, NULL else
1683 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1689 int i, n = get_irn_arity(phi);
1691 NEW_ARR_A(void *, res, n);
1693 for (i = 0; i < n; ++i) {
1694 pred = get_irn_n(phi, i);
1695 tv = get_Const_tarval(pred);
1696 tv = do_eval(eval, other, tv, mode);
1698 if (tv == tarval_bad) {
1699 /* folding failed, bad */
1705 for (i = 0; i < n; ++i) {
1706 pred = get_irn_n(phi, i);
1707 tv = get_Const_tarval(pred);
1708 tv = do_eval(eval, tv, other, mode);
1710 if (tv == tarval_bad) {
1711 /* folding failed, bad */
1717 irg = get_irn_irg(phi);
1718 for (i = 0; i < n; ++i) {
1719 pred = get_irn_n(phi, i);
1720 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1722 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1723 } /* apply_binop_on_phi */
1726 * Apply an evaluator on a binop with two constant Phi.
1728 * @param a the left Phi node
1729 * @param b the right Phi node
1730 * @param eval an evaluator function
1731 * @param mode the mode of the result, may be different from the mode of the Phi!
1733 * @return a new Phi node if the conversion was successful, NULL else
1735 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1737 ir_tarval *tv_l, *tv_r, *tv;
1743 if (get_nodes_block(a) != get_nodes_block(b))
1746 n = get_irn_arity(a);
1747 NEW_ARR_A(void *, res, n);
1749 for (i = 0; i < n; ++i) {
1750 pred = get_irn_n(a, i);
1751 tv_l = get_Const_tarval(pred);
1752 pred = get_irn_n(b, i);
1753 tv_r = get_Const_tarval(pred);
1754 tv = do_eval(eval, tv_l, tv_r, mode);
1756 if (tv == tarval_bad) {
1757 /* folding failed, bad */
1762 irg = get_irn_irg(a);
1763 for (i = 0; i < n; ++i) {
1764 pred = get_irn_n(a, i);
1765 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1767 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1768 } /* apply_binop_on_2_phis */
1771 * Apply an evaluator on a unop with a constant operator (a Phi).
1773 * @param phi the Phi node
1774 * @param eval an evaluator function
1776 * @return a new Phi node if the conversion was successful, NULL else
1778 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1785 int i, n = get_irn_arity(phi);
1787 NEW_ARR_A(void *, res, n);
1788 for (i = 0; i < n; ++i) {
1789 pred = get_irn_n(phi, i);
1790 tv = get_Const_tarval(pred);
1793 if (tv == tarval_bad) {
1794 /* folding failed, bad */
1799 mode = get_irn_mode(phi);
1800 irg = get_irn_irg(phi);
1801 for (i = 0; i < n; ++i) {
1802 pred = get_irn_n(phi, i);
1803 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1805 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1806 } /* apply_unop_on_phi */
1809 * Apply a conversion on a constant operator (a Phi).
1811 * @param phi the Phi node
1813 * @return a new Phi node if the conversion was successful, NULL else
1815 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1821 int i, n = get_irn_arity(phi);
1823 NEW_ARR_A(void *, res, n);
1824 for (i = 0; i < n; ++i) {
1825 pred = get_irn_n(phi, i);
1826 tv = get_Const_tarval(pred);
1827 tv = tarval_convert_to(tv, mode);
1829 if (tv == tarval_bad) {
1830 /* folding failed, bad */
1835 irg = get_irn_irg(phi);
1836 for (i = 0; i < n; ++i) {
1837 pred = get_irn_n(phi, i);
1838 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1840 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1841 } /* apply_conv_on_phi */
1844 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1845 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1846 * If possible, remove the Conv's.
1848 static ir_node *transform_node_AddSub(ir_node *n)
1850 ir_mode *mode = get_irn_mode(n);
1852 if (mode_is_reference(mode)) {
1853 ir_node *left = get_binop_left(n);
1854 ir_node *right = get_binop_right(n);
1855 unsigned ref_bits = get_mode_size_bits(mode);
1857 if (is_Conv(left)) {
1858 ir_mode *lmode = get_irn_mode(left);
1859 unsigned bits = get_mode_size_bits(lmode);
1861 if (ref_bits == bits &&
1862 mode_is_int(lmode) &&
1863 get_mode_arithmetic(lmode) == irma_twos_complement) {
1864 ir_node *pre = get_Conv_op(left);
1865 ir_mode *pre_mode = get_irn_mode(pre);
1867 if (mode_is_int(pre_mode) &&
1868 get_mode_size_bits(pre_mode) == bits &&
1869 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1870 /* ok, this conv just changes to sign, moreover the calculation
1871 * is done with same number of bits as our address mode, so
1872 * we can ignore the conv as address calculation can be viewed
1873 * as either signed or unsigned
1875 set_binop_left(n, pre);
1880 if (is_Conv(right)) {
1881 ir_mode *rmode = get_irn_mode(right);
1882 unsigned bits = get_mode_size_bits(rmode);
1884 if (ref_bits == bits &&
1885 mode_is_int(rmode) &&
1886 get_mode_arithmetic(rmode) == irma_twos_complement) {
1887 ir_node *pre = get_Conv_op(right);
1888 ir_mode *pre_mode = get_irn_mode(pre);
1890 if (mode_is_int(pre_mode) &&
1891 get_mode_size_bits(pre_mode) == bits &&
1892 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1893 /* ok, this conv just changes to sign, moreover the calculation
1894 * is done with same number of bits as our address mode, so
1895 * we can ignore the conv as address calculation can be viewed
1896 * as either signed or unsigned
1898 set_binop_right(n, pre);
1903 /* let address arithmetic use unsigned modes */
1904 if (is_Const(right)) {
1905 ir_mode *rmode = get_irn_mode(right);
1907 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1908 /* convert a AddP(P, *s) into AddP(P, *u) */
1909 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1911 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
1912 set_binop_right(n, pre);
1918 } /* transform_node_AddSub */
1920 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1923 if (is_Const(b) && is_const_Phi(a)) { \
1924 /* check for Op(Phi, Const) */ \
1925 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1927 else if (is_Const(a) && is_const_Phi(b)) { \
1928 /* check for Op(Const, Phi) */ \
1929 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1931 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1932 /* check for Op(Phi, Phi) */ \
1933 c = apply_binop_on_2_phis(a, b, eval, mode); \
1936 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1941 #define HANDLE_UNOP_PHI(eval, a, c) \
1944 if (is_const_Phi(a)) { \
1945 /* check for Op(Phi) */ \
1946 c = apply_unop_on_phi(a, eval); \
1948 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1955 * Do the AddSub optimization, then Transform
1956 * Constant folding on Phi
1957 * Add(a,a) -> Mul(a, 2)
1958 * Add(Mul(a, x), a) -> Mul(a, x+1)
1959 * if the mode is integer or float.
1960 * Transform Add(a,-b) into Sub(a,b).
1961 * Reassociation might fold this further.
1963 static ir_node *transform_node_Add(ir_node *n)
1966 ir_node *a, *b, *c, *oldn = n;
1967 vrp_attr *a_vrp, *b_vrp;
1969 n = transform_node_AddSub(n);
1971 a = get_Add_left(n);
1972 b = get_Add_right(n);
1974 mode = get_irn_mode(n);
1976 if (mode_is_reference(mode)) {
1977 ir_mode *lmode = get_irn_mode(a);
1979 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
1980 /* an Add(a, NULL) is a hidden Conv */
1981 dbg_info *dbg = get_irn_dbg_info(n);
1982 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
1986 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
1988 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1989 if (mode_is_float(mode)) {
1990 ir_graph *irg = get_irn_irg(n);
1991 if (get_irg_fp_model(irg) & fp_strict_algebraic)
1995 if (mode_is_num(mode)) {
1996 ir_graph *irg = get_irn_irg(n);
1997 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
1998 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
1999 && a == b && mode_is_int(mode)) {
2000 ir_node *block = get_nodes_block(n);
2003 get_irn_dbg_info(n),
2006 new_r_Const_long(irg, mode, 2),
2008 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2013 get_irn_dbg_info(n),
2018 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2023 get_irn_dbg_info(n),
2028 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2031 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2032 /* Here we rely on constants be on the RIGHT side */
2034 ir_node *op = get_Not_op(a);
2036 if (is_Const(b) && is_Const_one(b)) {
2038 ir_node *blk = get_nodes_block(n);
2039 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2040 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2047 a_vrp = vrp_get_info(a);
2048 b_vrp = vrp_get_info(b);
2050 if (a_vrp && b_vrp) {
2051 ir_tarval *vrp_val = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2053 if (tarval_is_null(vrp_val)) {
2054 dbg_info *dbgi = get_irn_dbg_info(n);
2055 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2059 } /* transform_node_Add */
2062 * returns -cnst or NULL if impossible
2064 static ir_node *const_negate(ir_node *cnst)
2066 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2067 dbg_info *dbgi = get_irn_dbg_info(cnst);
2068 ir_graph *irg = get_irn_irg(cnst);
2069 if (tv == tarval_bad) return NULL;
2070 return new_rd_Const(dbgi, irg, tv);
2074 * Do the AddSub optimization, then Transform
2075 * Constant folding on Phi
2076 * Sub(0,a) -> Minus(a)
2077 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2078 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2079 * Sub(Add(a, x), x) -> a
2080 * Sub(x, Add(x, a)) -> -a
2081 * Sub(x, Const) -> Add(x, -Const)
2083 static ir_node *transform_node_Sub(ir_node *n)
2089 n = transform_node_AddSub(n);
2091 a = get_Sub_left(n);
2092 b = get_Sub_right(n);
2094 mode = get_irn_mode(n);
2096 if (mode_is_int(mode)) {
2097 ir_mode *lmode = get_irn_mode(a);
2099 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2100 /* a Sub(a, NULL) is a hidden Conv */
2101 dbg_info *dbg = get_irn_dbg_info(n);
2102 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2103 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2107 if (mode == lmode &&
2108 get_mode_arithmetic(mode) == irma_twos_complement &&
2110 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2112 dbg_info *dbg = get_irn_dbg_info(n);
2113 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2114 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2120 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2122 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2123 if (mode_is_float(mode)) {
2124 ir_graph *irg = get_irn_irg(n);
2125 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2129 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2130 /* a - C -> a + (-C) */
2131 ir_node *cnst = const_negate(b);
2133 ir_node *block = get_nodes_block(n);
2134 dbg_info *dbgi = get_irn_dbg_info(n);
2136 n = new_rd_Add(dbgi, block, a, cnst, mode);
2137 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2142 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2143 dbg_info *dbg = get_irn_dbg_info(n);
2144 ir_node *block = get_nodes_block(n);
2145 ir_node *left = get_Minus_op(a);
2146 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2148 n = new_rd_Minus(dbg, block, add, mode);
2149 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2151 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2152 dbg_info *dbg = get_irn_dbg_info(n);
2153 ir_node *block = get_nodes_block(n);
2154 ir_node *right = get_Minus_op(b);
2156 n = new_rd_Add(dbg, block, a, right, mode);
2157 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2159 } else if (is_Sub(b)) {
2160 /* a - (b - c) -> a + (c - b)
2161 * -> (a - b) + c iff (b - c) is a pointer */
2162 dbg_info *s_dbg = get_irn_dbg_info(b);
2163 ir_node *s_left = get_Sub_left(b);
2164 ir_node *s_right = get_Sub_right(b);
2165 ir_mode *s_mode = get_irn_mode(b);
2166 if (mode_is_reference(s_mode)) {
2167 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2168 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2169 dbg_info *a_dbg = get_irn_dbg_info(n);
2172 s_right = new_r_Conv(lowest_block, s_right, mode);
2173 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2175 ir_node *s_block = get_nodes_block(b);
2176 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2177 dbg_info *a_dbg = get_irn_dbg_info(n);
2178 ir_node *a_block = get_nodes_block(n);
2180 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2182 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2185 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2186 ir_node *m_right = get_Mul_right(b);
2187 if (is_Const(m_right)) {
2188 ir_node *cnst2 = const_negate(m_right);
2189 if (cnst2 != NULL) {
2190 dbg_info *m_dbg = get_irn_dbg_info(b);
2191 ir_node *m_block = get_nodes_block(b);
2192 ir_node *m_left = get_Mul_left(b);
2193 ir_mode *m_mode = get_irn_mode(b);
2194 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2195 dbg_info *a_dbg = get_irn_dbg_info(n);
2196 ir_node *a_block = get_nodes_block(n);
2198 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2199 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2206 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2207 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2209 get_irn_dbg_info(n),
2213 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2217 if (mode_wrap_around(mode)) {
2218 ir_node *left = get_Add_left(a);
2219 ir_node *right = get_Add_right(a);
2221 /* FIXME: Does the Conv's work only for two complement or generally? */
2223 if (mode != get_irn_mode(right)) {
2224 /* This Sub is an effective Cast */
2225 right = new_r_Conv(get_nodes_block(n), right, mode);
2228 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2230 } else if (right == b) {
2231 if (mode != get_irn_mode(left)) {
2232 /* This Sub is an effective Cast */
2233 left = new_r_Conv(get_nodes_block(n), left, mode);
2236 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2242 if (mode_wrap_around(mode)) {
2243 ir_node *left = get_Add_left(b);
2244 ir_node *right = get_Add_right(b);
2246 /* FIXME: Does the Conv's work only for two complement or generally? */
2248 ir_mode *r_mode = get_irn_mode(right);
2250 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2251 if (mode != r_mode) {
2252 /* This Sub is an effective Cast */
2253 n = new_r_Conv(get_nodes_block(n), n, mode);
2255 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2257 } else if (right == a) {
2258 ir_mode *l_mode = get_irn_mode(left);
2260 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2261 if (mode != l_mode) {
2262 /* This Sub is an effective Cast */
2263 n = new_r_Conv(get_nodes_block(n), n, mode);
2265 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2270 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2271 ir_mode *mode = get_irn_mode(a);
2273 if (mode == get_irn_mode(b)) {
2275 ir_node *op_a = get_Conv_op(a);
2276 ir_node *op_b = get_Conv_op(b);
2278 /* check if it's allowed to skip the conv */
2279 ma = get_irn_mode(op_a);
2280 mb = get_irn_mode(op_b);
2282 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2283 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2286 set_Sub_right(n, b);
2292 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2293 if (!is_reassoc_running() && is_Mul(a)) {
2294 ir_node *ma = get_Mul_left(a);
2295 ir_node *mb = get_Mul_right(a);
2298 ir_node *blk = get_nodes_block(n);
2299 ir_graph *irg = get_irn_irg(n);
2301 get_irn_dbg_info(n),
2305 get_irn_dbg_info(n),
2308 new_r_Const(irg, get_mode_one(mode)),
2311 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2313 } else if (mb == b) {
2314 ir_node *blk = get_nodes_block(n);
2315 ir_graph *irg = get_irn_irg(n);
2317 get_irn_dbg_info(n),
2321 get_irn_dbg_info(n),
2324 new_r_Const(irg, get_mode_one(mode)),
2327 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2331 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2332 ir_node *x = get_Sub_left(a);
2333 ir_node *y = get_Sub_right(a);
2334 ir_node *blk = get_nodes_block(n);
2335 ir_mode *m_b = get_irn_mode(b);
2336 ir_mode *m_y = get_irn_mode(y);
2340 /* Determine the right mode for the Add. */
2343 else if (mode_is_reference(m_b))
2345 else if (mode_is_reference(m_y))
2349 * Both modes are different but none is reference,
2350 * happens for instance in SubP(SubP(P, Iu), Is).
2351 * We have two possibilities here: Cast or ignore.
2352 * Currently we ignore this case.
2357 add = new_r_Add(blk, y, b, add_mode);
2359 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2360 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2364 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2365 /* c - ~X = X + (c+1) */
2366 if (is_Const(a) && is_Not(b)) {
2367 ir_tarval *tv = get_Const_tarval(a);
2369 tv = tarval_add(tv, get_mode_one(mode));
2370 if (tv != tarval_bad) {
2371 ir_node *blk = get_nodes_block(n);
2372 ir_graph *irg = get_irn_irg(n);
2373 ir_node *c = new_r_Const(irg, tv);
2374 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2375 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2379 /* x-(x&y) = x & ~y */
2381 ir_node *and_left = get_And_left(b);
2382 ir_node *and_right = get_And_right(b);
2383 if (and_right == a) {
2384 ir_node *tmp = and_left;
2385 and_left = and_right;
2388 if (and_left == a) {
2389 dbg_info *dbgi = get_irn_dbg_info(n);
2390 ir_node *block = get_nodes_block(n);
2391 ir_mode *mode = get_irn_mode(n);
2392 ir_node *notn = new_rd_Not(dbgi, block, and_right, mode);
2393 ir_node *and = new_rd_And(dbgi, block, a, notn, mode);
2399 } /* transform_node_Sub */
2402 * Several transformation done on n*n=2n bits mul.
2403 * These transformations must be done here because new nodes may be produced.
2405 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2408 ir_node *a = get_Mul_left(n);
2409 ir_node *b = get_Mul_right(n);
2410 ir_tarval *ta = value_of(a);
2411 ir_tarval *tb = value_of(b);
2412 ir_mode *smode = get_irn_mode(a);
2414 if (ta == get_mode_one(smode)) {
2415 /* (L)1 * (L)b = (L)b */
2416 ir_node *blk = get_nodes_block(n);
2417 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2418 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2421 else if (ta == get_mode_minus_one(smode)) {
2422 /* (L)-1 * (L)b = (L)b */
2423 ir_node *blk = get_nodes_block(n);
2424 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2425 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2426 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2429 if (tb == get_mode_one(smode)) {
2430 /* (L)a * (L)1 = (L)a */
2431 ir_node *blk = get_irn_n(a, -1);
2432 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2433 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2436 else if (tb == get_mode_minus_one(smode)) {
2437 /* (L)a * (L)-1 = (L)-a */
2438 ir_node *blk = get_nodes_block(n);
2439 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2440 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2441 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2448 * Transform Mul(a,-1) into -a.
2449 * Do constant evaluation of Phi nodes.
2450 * Do architecture dependent optimizations on Mul nodes
2452 static ir_node *transform_node_Mul(ir_node *n)
2454 ir_node *c, *oldn = n;
2455 ir_mode *mode = get_irn_mode(n);
2456 ir_node *a = get_Mul_left(n);
2457 ir_node *b = get_Mul_right(n);
2459 if (is_Bad(a) || is_Bad(b))
2462 if (mode != get_irn_mode(a))
2463 return transform_node_Mul2n(n, mode);
2465 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2467 if (mode_is_signed(mode)) {
2470 if (value_of(a) == get_mode_minus_one(mode))
2472 else if (value_of(b) == get_mode_minus_one(mode))
2475 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2476 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2481 if (is_Const(b)) { /* (-a) * const -> a * -const */
2482 ir_node *cnst = const_negate(b);
2484 dbg_info *dbgi = get_irn_dbg_info(n);
2485 ir_node *block = get_nodes_block(n);
2486 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2487 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2490 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2491 dbg_info *dbgi = get_irn_dbg_info(n);
2492 ir_node *block = get_nodes_block(n);
2493 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2494 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2496 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2497 ir_node *sub_l = get_Sub_left(b);
2498 ir_node *sub_r = get_Sub_right(b);
2499 dbg_info *dbgi = get_irn_dbg_info(n);
2500 ir_node *block = get_nodes_block(n);
2501 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2502 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2503 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2506 } else if (is_Minus(b)) {
2507 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2508 ir_node *sub_l = get_Sub_left(a);
2509 ir_node *sub_r = get_Sub_right(a);
2510 dbg_info *dbgi = get_irn_dbg_info(n);
2511 ir_node *block = get_nodes_block(n);
2512 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2513 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2514 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2517 } else if (is_Shl(a)) {
2518 ir_node *const shl_l = get_Shl_left(a);
2519 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2520 /* (1 << x) * b -> b << x */
2521 dbg_info *const dbgi = get_irn_dbg_info(n);
2522 ir_node *const block = get_nodes_block(n);
2523 ir_node *const shl_r = get_Shl_right(a);
2524 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2525 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2528 } else if (is_Shl(b)) {
2529 ir_node *const shl_l = get_Shl_left(b);
2530 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2531 /* a * (1 << x) -> a << x */
2532 dbg_info *const dbgi = get_irn_dbg_info(n);
2533 ir_node *const block = get_nodes_block(n);
2534 ir_node *const shl_r = get_Shl_right(b);
2535 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2536 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2540 if (get_mode_arithmetic(mode) == irma_ieee754) {
2542 ir_tarval *tv = get_Const_tarval(a);
2543 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2544 && !tarval_is_negative(tv)) {
2545 /* 2.0 * b = b + b */
2546 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2547 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2551 else if (is_Const(b)) {
2552 ir_tarval *tv = get_Const_tarval(b);
2553 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2554 && !tarval_is_negative(tv)) {
2555 /* a * 2.0 = a + a */
2556 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2557 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2562 return arch_dep_replace_mul_with_shifts(n);
2563 } /* transform_node_Mul */
2566 * Transform a Div Node.
2568 static ir_node *transform_node_Div(ir_node *n)
2570 ir_mode *mode = get_Div_resmode(n);
2571 ir_node *a = get_Div_left(n);
2572 ir_node *b = get_Div_right(n);
2574 const ir_node *dummy;
2576 if (mode_is_int(mode)) {
2577 if (is_Const(b) && is_const_Phi(a)) {
2578 /* check for Div(Phi, Const) */
2579 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2581 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2584 } else if (is_Const(a) && is_const_Phi(b)) {
2585 /* check for Div(Const, Phi) */
2586 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2588 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2591 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2592 /* check for Div(Phi, Phi) */
2593 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2595 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2600 if (a == b && value_not_zero(a, &dummy)) {
2601 ir_graph *irg = get_irn_irg(n);
2602 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2603 value = new_r_Const(irg, get_mode_one(mode));
2604 DBG_OPT_CSTEVAL(n, value);
2607 if (mode_is_signed(mode) && is_Const(b)) {
2608 ir_tarval *tv = get_Const_tarval(b);
2610 if (tv == get_mode_minus_one(mode)) {
2612 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2613 DBG_OPT_CSTEVAL(n, value);
2617 /* Try architecture dependent optimization */
2618 value = arch_dep_replace_div_by_const(n);
2621 assert(mode_is_float(mode));
2623 /* Optimize x/c to x*(1/c) */
2624 if (get_mode_arithmetic(mode) == irma_ieee754) {
2625 ir_tarval *tv = value_of(b);
2627 if (tv != tarval_bad) {
2628 int rem = tarval_fp_ops_enabled();
2631 * Floating point constant folding might be disabled here to
2633 * However, as we check for exact result, doing it is safe.
2636 tarval_enable_fp_ops(1);
2637 tv = tarval_div(get_mode_one(mode), tv);
2638 tarval_enable_fp_ops(rem);
2640 /* Do the transformation if the result is either exact or we are
2641 not using strict rules. */
2642 if (tv != tarval_bad &&
2643 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2644 ir_node *block = get_nodes_block(n);
2645 ir_graph *irg = get_irn_irg(block);
2646 ir_node *c = new_r_Const(irg, tv);
2647 dbg_info *dbgi = get_irn_dbg_info(n);
2648 value = new_rd_Mul(dbgi, block, a, c, mode);
2661 /* Turn Div into a tuple (mem, jmp, bad, value) */
2662 mem = get_Div_mem(n);
2663 blk = get_nodes_block(n);
2664 irg = get_irn_irg(blk);
2666 /* skip a potential Pin */
2667 mem = skip_Pin(mem);
2668 turn_into_tuple(n, pn_Div_max+1);
2669 set_Tuple_pred(n, pn_Div_M, mem);
2670 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2671 set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X));
2672 set_Tuple_pred(n, pn_Div_res, value);
2675 } /* transform_node_Div */
2678 * Transform a Mod node.
2680 static ir_node *transform_node_Mod(ir_node *n)
2682 ir_mode *mode = get_Mod_resmode(n);
2683 ir_node *a = get_Mod_left(n);
2684 ir_node *b = get_Mod_right(n);
2689 if (is_Const(b) && is_const_Phi(a)) {
2690 /* check for Div(Phi, Const) */
2691 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2693 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2697 else if (is_Const(a) && is_const_Phi(b)) {
2698 /* check for Div(Const, Phi) */
2699 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2701 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2705 else if (is_const_Phi(a) && is_const_Phi(b)) {
2706 /* check for Div(Phi, Phi) */
2707 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2709 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2716 irg = get_irn_irg(n);
2717 if (tv != tarval_bad) {
2718 value = new_r_Const(irg, tv);
2720 DBG_OPT_CSTEVAL(n, value);
2723 ir_node *a = get_Mod_left(n);
2724 ir_node *b = get_Mod_right(n);
2725 const ir_node *dummy;
2727 if (a == b && value_not_zero(a, &dummy)) {
2728 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2729 value = new_r_Const(irg, get_mode_null(mode));
2730 DBG_OPT_CSTEVAL(n, value);
2733 if (mode_is_signed(mode) && is_Const(b)) {
2734 ir_tarval *tv = get_Const_tarval(b);
2736 if (tv == get_mode_minus_one(mode)) {
2738 value = new_r_Const(irg, get_mode_null(mode));
2739 DBG_OPT_CSTEVAL(n, value);
2743 /* Try architecture dependent optimization */
2744 value = arch_dep_replace_mod_by_const(n);
2753 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2754 mem = get_Mod_mem(n);
2755 blk = get_nodes_block(n);
2756 irg = get_irn_irg(blk);
2758 /* skip a potential Pin */
2759 mem = skip_Pin(mem);
2760 turn_into_tuple(n, pn_Mod_max+1);
2761 set_Tuple_pred(n, pn_Mod_M, mem);
2762 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2763 set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X));
2764 set_Tuple_pred(n, pn_Mod_res, value);
2767 } /* transform_node_Mod */
2770 * Transform a Cond node.
2772 * Replace the Cond by a Jmp if it branches on a constant
2775 static ir_node *transform_node_Cond(ir_node *n)
2778 ir_node *a = get_Cond_selector(n);
2779 ir_tarval *ta = value_of(a);
2780 ir_graph *irg = get_irn_irg(n);
2783 /* we need block info which is not available in floating irgs */
2784 if (get_irg_pinned(irg) == op_pin_state_floats)
2787 if ((ta != tarval_bad) &&
2788 (get_irn_mode(a) == mode_b) &&
2789 (get_opt_unreachable_code())) {
2790 /* It's a boolean Cond, branching on a boolean constant.
2791 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2792 ir_node *blk = get_nodes_block(n);
2793 jmp = new_r_Jmp(blk);
2794 turn_into_tuple(n, pn_Cond_max+1);
2795 if (ta == tarval_b_true) {
2796 set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
2797 set_Tuple_pred(n, pn_Cond_true, jmp);
2799 set_Tuple_pred(n, pn_Cond_false, jmp);
2800 set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X));
2802 /* We might generate an endless loop, so keep it alive. */
2803 add_End_keepalive(get_irg_end(irg), blk);
2806 } /* transform_node_Cond */
2809 * Prototype of a recursive transform function
2810 * for bitwise distributive transformations.
2812 typedef ir_node* (*recursive_transform)(ir_node *n);
2815 * makes use of distributive laws for and, or, eor
2816 * and(a OP c, b OP c) -> and(a, b) OP c
2817 * note, might return a different op than n
2819 static ir_node *transform_bitwise_distributive(ir_node *n,
2820 recursive_transform trans_func)
2823 ir_node *a = get_binop_left(n);
2824 ir_node *b = get_binop_right(n);
2825 ir_op *op = get_irn_op(a);
2826 ir_op *op_root = get_irn_op(n);
2828 if (op != get_irn_op(b))
2831 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2832 if (op == op_Conv) {
2833 ir_node *a_op = get_Conv_op(a);
2834 ir_node *b_op = get_Conv_op(b);
2835 ir_mode *a_mode = get_irn_mode(a_op);
2836 ir_mode *b_mode = get_irn_mode(b_op);
2837 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2838 ir_node *blk = get_nodes_block(n);
2841 set_binop_left(n, a_op);
2842 set_binop_right(n, b_op);
2843 set_irn_mode(n, a_mode);
2845 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2847 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2853 /* nothing to gain here */
2857 if (op == op_Shrs || op == op_Shr || op == op_Shl
2858 || op == op_And || op == op_Or || op == op_Eor) {
2859 ir_node *a_left = get_binop_left(a);
2860 ir_node *a_right = get_binop_right(a);
2861 ir_node *b_left = get_binop_left(b);
2862 ir_node *b_right = get_binop_right(b);
2864 ir_node *op1 = NULL;
2865 ir_node *op2 = NULL;
2867 if (is_op_commutative(op)) {
2868 if (a_left == b_left) {
2872 } else if (a_left == b_right) {
2876 } else if (a_right == b_left) {
2882 if (a_right == b_right) {
2889 /* (a sop c) & (b sop c) => (a & b) sop c */
2890 ir_node *blk = get_nodes_block(n);
2892 ir_node *new_n = exact_copy(n);
2893 set_binop_left(new_n, op1);
2894 set_binop_right(new_n, op2);
2895 new_n = trans_func(new_n);
2897 if (op_root == op_Eor && op == op_Or) {
2898 dbg_info *dbgi = get_irn_dbg_info(n);
2899 ir_mode *mode = get_irn_mode(c);
2901 c = new_rd_Not(dbgi, blk, c, mode);
2902 n = new_rd_And(dbgi, blk, new_n, c, mode);
2905 set_nodes_block(n, blk);
2906 set_binop_left(n, new_n);
2907 set_binop_right(n, c);
2911 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2920 * Create a 0 constant of given mode.
2922 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
2924 ir_tarval *tv = get_mode_null(mode);
2925 ir_node *cnst = new_r_Const(irg, tv);
2930 static bool is_shiftop(const ir_node *n)
2932 return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
2936 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
2938 * - and, or, xor instead of &
2939 * - Shl, Shr, Shrs, rotl instead of >>
2940 * (with a special case for Or/Xor + Shrs)
2942 * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
2944 static ir_node *transform_node_shift_bitop(ir_node *n)
2946 ir_graph *irg = get_irn_irg(n);
2947 ir_node *right = get_binop_right(n);
2948 ir_mode *mode = get_irn_mode(n);
2950 ir_node *bitop_left;
2951 ir_node *bitop_right;
2960 ir_tarval *tv_shift;
2962 if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
2965 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
2967 if (!is_Const(right))
2970 left = get_binop_left(n);
2971 op_left = get_irn_op(left);
2972 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
2975 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
2976 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
2977 /* TODO: test if sign bit is affectes */
2981 bitop_right = get_binop_right(left);
2982 if (!is_Const(bitop_right))
2985 bitop_left = get_binop_left(left);
2987 block = get_nodes_block(n);
2988 dbgi = get_irn_dbg_info(n);
2989 tv1 = get_Const_tarval(bitop_right);
2990 tv2 = get_Const_tarval(right);
2992 assert(get_tarval_mode(tv1) == mode);
2995 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
2996 tv_shift = tarval_shl(tv1, tv2);
2997 } else if (is_Shr(n)) {
2998 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
2999 tv_shift = tarval_shr(tv1, tv2);
3000 } else if (is_Shrs(n)) {
3001 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
3002 tv_shift = tarval_shrs(tv1, tv2);
3005 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
3006 tv_shift = tarval_rotl(tv1, tv2);
3009 assert(get_tarval_mode(tv_shift) == mode);
3010 irg = get_irn_irg(n);
3011 new_const = new_r_Const(irg, tv_shift);
3013 if (op_left == op_And) {
3014 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
3015 } else if (op_left == op_Or) {
3016 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
3018 assert(op_left == op_Eor);
3019 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
3026 * normalisation: (x >> c1) & c2 to (x & (c2<<c1)) >> c1
3028 * - and, or, xor instead of &
3029 * - Shl, Shr, Shrs, rotl instead of >>
3030 * (with a special case for Or/Xor + Shrs)
3032 * This normalisation is usually good for the backend since << C can often be
3033 * matched as address-mode.
3035 static ir_node *transform_node_bitop_shift(ir_node *n)
3037 ir_graph *irg = get_irn_irg(n);
3038 ir_node *left = get_binop_left(n);
3039 ir_node *right = get_binop_right(n);
3040 ir_mode *mode = get_irn_mode(n);
3041 ir_node *shift_left;
3042 ir_node *shift_right;
3044 dbg_info *dbg_bitop;
3045 dbg_info *dbg_shift;
3051 ir_tarval *tv_bitop;
3053 if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
3056 assert(is_And(n) || is_Or(n) || is_Eor(n));
3057 if (!is_Const(right) || !is_shiftop(left))
3060 shift_left = get_binop_left(left);
3061 shift_right = get_binop_right(left);
3062 if (!is_Const(shift_right))
3065 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
3066 if (is_Shrs(left)) {
3067 /* TODO this could be improved */
3071 irg = get_irn_irg(n);
3072 block = get_nodes_block(n);
3073 dbg_bitop = get_irn_dbg_info(n);
3074 dbg_shift = get_irn_dbg_info(left);
3075 tv1 = get_Const_tarval(shift_right);
3076 tv2 = get_Const_tarval(right);
3077 assert(get_tarval_mode(tv2) == mode);
3080 tv_bitop = tarval_shr(tv2, tv1);
3081 } else if (is_Shr(left)) {
3082 if (is_Or(n) || is_Eor(n)) {
3084 * TODO this can be improved by checking whether
3085 * the left shift produces an overflow
3089 tv_bitop = tarval_shl(tv2, tv1);
3091 assert(is_Rotl(left));
3092 tv_bitop = tarval_rotl(tv2, tarval_neg(tv1));
3094 new_const = new_r_Const(irg, tv_bitop);
3097 new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode);
3098 } else if (is_Or(n)) {
3099 new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode);
3102 new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode);
3106 new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode);
3107 } else if (is_Shr(left)) {
3108 new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode);
3110 assert(is_Rotl(left));
3111 new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode);
3120 static ir_node *transform_node_And(ir_node *n)
3122 ir_node *c, *oldn = n;
3123 ir_node *a = get_And_left(n);
3124 ir_node *b = get_And_right(n);
3126 vrp_attr *a_vrp, *b_vrp;
3128 if (is_Cmp(a) && is_Cmp(b)) {
3129 ir_node *a_left = get_Cmp_left(a);
3130 ir_node *a_right = get_Cmp_right(a);
3131 ir_node *b_left = get_Cmp_left(b);
3132 ir_node *b_right = get_Cmp_right(b);
3133 ir_relation a_relation = get_Cmp_relation(a);
3134 ir_relation b_relation = get_Cmp_relation(b);
3135 /* we can combine the relations of two compares with the same
3137 if (a_left == b_left && b_left == b_right) {
3138 dbg_info *dbgi = get_irn_dbg_info(n);
3139 ir_node *block = get_nodes_block(n);
3140 ir_relation new_relation = a_relation & b_relation;
3141 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3143 /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */
3144 if (a_relation == b_relation && a_relation == ir_relation_equal
3145 && !mode_is_float(get_irn_mode(a_left))
3146 && !mode_is_float(get_irn_mode(b_left))
3147 && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
3148 dbg_info *dbgi = get_irn_dbg_info(n);
3149 ir_node *block = get_nodes_block(n);
3150 ir_mode *a_mode = get_irn_mode(a_left);
3151 ir_mode *b_mode = get_irn_mode(b_left);
3152 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3153 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3154 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
3155 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
3156 ir_graph *irg = get_irn_irg(n);
3157 ir_node *zero = create_zero_const(irg, a_mode);
3158 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3162 mode = get_irn_mode(n);
3163 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3167 ir_node *op = get_Not_op(b);
3169 ir_node *ba = get_And_left(op);
3170 ir_node *bb = get_And_right(op);
3172 /* it's enough to test the following cases due to normalization! */
3173 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3174 /* (a|b) & ~(a&b) = a^b */
3175 ir_node *block = get_nodes_block(n);
3177 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3178 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3186 ir_node *op = get_Not_op(a);
3188 ir_node *aa = get_And_left(op);
3189 ir_node *ab = get_And_right(op);
3191 /* it's enough to test the following cases due to normalization! */
3192 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3193 /* (a|b) & ~(a&b) = a^b */
3194 ir_node *block = get_nodes_block(n);
3196 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3197 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3204 ir_node *al = get_Eor_left(a);
3205 ir_node *ar = get_Eor_right(a);
3208 /* (b ^ a) & b -> ~a & b */
3209 dbg_info *dbg = get_irn_dbg_info(n);
3210 ir_node *block = get_nodes_block(n);
3212 ar = new_rd_Not(dbg, block, ar, mode);
3213 n = new_rd_And(dbg, block, ar, b, mode);
3214 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3218 /* (a ^ b) & b -> ~a & b */
3219 dbg_info *dbg = get_irn_dbg_info(n);
3220 ir_node *block = get_nodes_block(n);
3222 al = new_rd_Not(dbg, block, al, mode);
3223 n = new_rd_And(dbg, block, al, b, mode);
3224 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3229 ir_node *bl = get_Eor_left(b);
3230 ir_node *br = get_Eor_right(b);
3233 /* a & (a ^ b) -> a & ~b */
3234 dbg_info *dbg = get_irn_dbg_info(n);
3235 ir_node *block = get_nodes_block(n);
3237 br = new_rd_Not(dbg, block, br, mode);
3238 n = new_rd_And(dbg, block, br, a, mode);
3239 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3243 /* a & (b ^ a) -> a & ~b */
3244 dbg_info *dbg = get_irn_dbg_info(n);
3245 ir_node *block = get_nodes_block(n);
3247 bl = new_rd_Not(dbg, block, bl, mode);
3248 n = new_rd_And(dbg, block, bl, a, mode);
3249 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3253 if (is_Not(a) && is_Not(b)) {
3254 /* ~a & ~b = ~(a|b) */
3255 ir_node *block = get_nodes_block(n);
3256 ir_mode *mode = get_irn_mode(n);
3260 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3261 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3262 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3266 b_vrp = vrp_get_info(b);
3267 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3268 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3274 a_vrp = vrp_get_info(a);
3275 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3276 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3280 n = transform_bitwise_distributive(n, transform_node_And);
3282 n = transform_node_bitop_shift(n);
3285 } /* transform_node_And */
3287 /* the order of the values is important! */
3288 typedef enum const_class {
3294 static const_class classify_const(const ir_node* n)
3296 if (is_Const(n)) return const_const;
3297 if (is_irn_constlike(n)) return const_like;
3302 * Determines whether r is more constlike or has a larger index (in that order)
3305 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3307 const const_class l_order = classify_const(l);
3308 const const_class r_order = classify_const(r);
3310 l_order > r_order ||
3311 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3317 static ir_node *transform_node_Eor(ir_node *n)
3319 ir_node *c, *oldn = n;
3320 ir_node *a = get_Eor_left(n);
3321 ir_node *b = get_Eor_right(n);
3322 ir_mode *mode = get_irn_mode(n);
3324 /* we can combine the relations of two compares with the same operands */
3325 if (is_Cmp(a) && is_Cmp(b)) {
3326 ir_node *a_left = get_Cmp_left(a);
3327 ir_node *a_right = get_Cmp_left(a);
3328 ir_node *b_left = get_Cmp_left(b);
3329 ir_node *b_right = get_Cmp_right(b);
3330 if (a_left == b_left && b_left == b_right) {
3331 dbg_info *dbgi = get_irn_dbg_info(n);
3332 ir_node *block = get_nodes_block(n);
3333 ir_relation a_relation = get_Cmp_relation(a);
3334 ir_relation b_relation = get_Cmp_relation(b);
3335 ir_relation new_relation = a_relation ^ b_relation;
3336 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3340 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3342 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3343 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3344 dbg_info *dbg = get_irn_dbg_info(n);
3345 ir_node *block = get_nodes_block(n);
3346 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3347 ir_node *new_left = get_Not_op(a);
3348 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3349 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3351 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3352 dbg_info *dbg = get_irn_dbg_info(n);
3353 ir_node *block = get_nodes_block(n);
3354 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3355 ir_node *new_right = get_Not_op(b);
3356 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3357 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3361 /* x ^ 1...1 -> ~1 */
3362 if (is_Const(b) && is_Const_all_one(b)) {
3363 n = new_r_Not(get_nodes_block(n), a, mode);
3364 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3368 n = transform_bitwise_distributive(n, transform_node_Eor);
3370 n = transform_node_bitop_shift(n);
3373 } /* transform_node_Eor */
3378 static ir_node *transform_node_Not(ir_node *n)
3380 ir_node *c, *oldn = n;
3381 ir_node *a = get_Not_op(n);
3382 ir_mode *mode = get_irn_mode(n);
3384 HANDLE_UNOP_PHI(tarval_not,a,c);
3386 /* check for a boolean Not */
3388 dbg_info *dbgi = get_irn_dbg_info(a);
3389 ir_node *block = get_nodes_block(a);
3390 ir_relation relation = get_Cmp_relation(a);
3391 relation = get_negated_relation(relation);
3392 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3397 /* normalize ~(a ^ b) => a ^ ~b */
3399 dbg_info *dbg = get_irn_dbg_info(n);
3400 ir_node *block = get_nodes_block(n);
3401 ir_node *eor_right = get_Eor_right(a);
3402 ir_node *eor_left = get_Eor_left(a);
3403 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3404 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3408 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3409 if (is_Minus(a)) { /* ~-x -> x + -1 */
3410 dbg_info *dbg = get_irn_dbg_info(n);
3411 ir_graph *irg = get_irn_irg(n);
3412 ir_node *block = get_nodes_block(n);
3413 ir_node *add_l = get_Minus_op(a);
3414 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3415 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3416 } else if (is_Add(a)) {
3417 ir_node *add_r = get_Add_right(a);
3418 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3419 /* ~(x + -1) = -x */
3420 ir_node *op = get_Add_left(a);
3421 ir_node *blk = get_nodes_block(n);
3422 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3423 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3431 * Transform a Minus.
3435 * -(a >>u (size-1)) = a >>s (size-1)
3436 * -(a >>s (size-1)) = a >>u (size-1)
3437 * -(a * const) -> a * -const
3439 static ir_node *transform_node_Minus(ir_node *n)
3441 ir_node *c, *oldn = n;
3442 ir_node *a = get_Minus_op(n);
3445 HANDLE_UNOP_PHI(tarval_neg,a,c);
3447 mode = get_irn_mode(a);
3448 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3449 /* the following rules are only to twos-complement */
3452 ir_node *op = get_Not_op(a);
3453 ir_tarval *tv = get_mode_one(mode);
3454 ir_node *blk = get_nodes_block(n);
3455 ir_graph *irg = get_irn_irg(blk);
3456 ir_node *c = new_r_Const(irg, tv);
3457 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3458 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3462 ir_node *c = get_Shr_right(a);
3465 ir_tarval *tv = get_Const_tarval(c);
3467 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3468 /* -(a >>u (size-1)) = a >>s (size-1) */
3469 ir_node *v = get_Shr_left(a);
3471 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3472 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3478 ir_node *c = get_Shrs_right(a);
3481 ir_tarval *tv = get_Const_tarval(c);
3483 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3484 /* -(a >>s (size-1)) = a >>u (size-1) */
3485 ir_node *v = get_Shrs_left(a);
3487 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3488 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3495 /* - (a-b) = b - a */
3496 ir_node *la = get_Sub_left(a);
3497 ir_node *ra = get_Sub_right(a);
3498 ir_node *blk = get_nodes_block(n);
3500 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3501 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3505 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3506 ir_node *mul_l = get_Mul_left(a);
3507 ir_node *mul_r = get_Mul_right(a);
3508 ir_tarval *tv = value_of(mul_r);
3509 if (tv != tarval_bad) {
3510 tv = tarval_neg(tv);
3511 if (tv != tarval_bad) {
3512 ir_graph *irg = get_irn_irg(n);
3513 ir_node *cnst = new_r_Const(irg, tv);
3514 dbg_info *dbg = get_irn_dbg_info(a);
3515 ir_node *block = get_nodes_block(a);
3516 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3517 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3524 } /* transform_node_Minus */
3527 * Transform a Proj(Load) with a non-null address.
3529 static ir_node *transform_node_Proj_Load(ir_node *proj)
3531 if (get_opt_ldst_only_null_ptr_exceptions()) {
3532 if (get_irn_mode(proj) == mode_X) {
3533 ir_node *load = get_Proj_pred(proj);
3535 /* get the Load address */
3536 const ir_node *addr = get_Load_ptr(load);
3537 const ir_node *confirm;
3539 if (value_not_null(addr, &confirm)) {
3540 if (confirm == NULL) {
3541 /* this node may float if it did not depend on a Confirm */
3542 set_irn_pinned(load, op_pin_state_floats);
3544 if (get_Proj_proj(proj) == pn_Load_X_except) {
3545 ir_graph *irg = get_irn_irg(proj);
3546 DBG_OPT_EXC_REM(proj);
3547 return new_r_Bad(irg, mode_X);
3549 ir_node *blk = get_nodes_block(load);
3550 return new_r_Jmp(blk);
3556 } /* transform_node_Proj_Load */
3559 * Transform a Proj(Store) with a non-null address.
3561 static ir_node *transform_node_Proj_Store(ir_node *proj)
3563 if (get_opt_ldst_only_null_ptr_exceptions()) {
3564 if (get_irn_mode(proj) == mode_X) {
3565 ir_node *store = get_Proj_pred(proj);
3567 /* get the load/store address */
3568 const ir_node *addr = get_Store_ptr(store);
3569 const ir_node *confirm;
3571 if (value_not_null(addr, &confirm)) {
3572 if (confirm == NULL) {
3573 /* this node may float if it did not depend on a Confirm */
3574 set_irn_pinned(store, op_pin_state_floats);
3576 if (get_Proj_proj(proj) == pn_Store_X_except) {
3577 ir_graph *irg = get_irn_irg(proj);
3578 DBG_OPT_EXC_REM(proj);
3579 return new_r_Bad(irg, mode_X);
3581 ir_node *blk = get_nodes_block(store);
3582 return new_r_Jmp(blk);
3588 } /* transform_node_Proj_Store */
3591 * Transform a Proj(Div) with a non-zero value.
3592 * Removes the exceptions and routes the memory to the NoMem node.
3594 static ir_node *transform_node_Proj_Div(ir_node *proj)
3596 ir_node *div = get_Proj_pred(proj);
3597 ir_node *b = get_Div_right(div);
3598 ir_node *res, *new_mem;
3599 const ir_node *confirm;
3602 if (value_not_zero(b, &confirm)) {
3603 /* div(x, y) && y != 0 */
3604 if (confirm == NULL) {
3605 /* we are sure we have a Const != 0 */
3606 new_mem = get_Div_mem(div);
3607 new_mem = skip_Pin(new_mem);
3608 set_Div_mem(div, new_mem);
3609 set_irn_pinned(div, op_pin_state_floats);
3612 proj_nr = get_Proj_proj(proj);
3614 case pn_Div_X_regular:
3615 return new_r_Jmp(get_nodes_block(div));
3617 case pn_Div_X_except: {
3618 ir_graph *irg = get_irn_irg(proj);
3619 /* we found an exception handler, remove it */
3620 DBG_OPT_EXC_REM(proj);
3621 return new_r_Bad(irg, mode_X);
3625 ir_graph *irg = get_irn_irg(proj);
3626 res = get_Div_mem(div);
3627 new_mem = get_irg_no_mem(irg);
3630 /* This node can only float up to the Confirm block */
3631 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3633 set_irn_pinned(div, op_pin_state_floats);
3634 /* this is a Div without exception, we can remove the memory edge */
3635 set_Div_mem(div, new_mem);
3641 } /* transform_node_Proj_Div */
3644 * Transform a Proj(Mod) with a non-zero value.
3645 * Removes the exceptions and routes the memory to the NoMem node.
3647 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3649 ir_node *mod = get_Proj_pred(proj);
3650 ir_node *b = get_Mod_right(mod);
3651 ir_node *res, *new_mem;
3652 const ir_node *confirm;
3655 if (value_not_zero(b, &confirm)) {
3656 /* mod(x, y) && y != 0 */
3657 proj_nr = get_Proj_proj(proj);
3659 if (confirm == NULL) {
3660 /* we are sure we have a Const != 0 */
3661 new_mem = get_Mod_mem(mod);
3662 new_mem = skip_Pin(new_mem);
3663 set_Mod_mem(mod, new_mem);
3664 set_irn_pinned(mod, op_pin_state_floats);
3669 case pn_Mod_X_regular:
3670 return new_r_Jmp(get_irn_n(mod, -1));
3672 case pn_Mod_X_except: {
3673 ir_graph *irg = get_irn_irg(proj);
3674 /* we found an exception handler, remove it */
3675 DBG_OPT_EXC_REM(proj);
3676 return new_r_Bad(irg, mode_X);
3680 ir_graph *irg = get_irn_irg(proj);
3681 res = get_Mod_mem(mod);
3682 new_mem = get_irg_no_mem(irg);
3685 /* This node can only float up to the Confirm block */
3686 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3688 /* this is a Mod without exception, we can remove the memory edge */
3689 set_Mod_mem(mod, new_mem);
3693 if (get_Mod_left(mod) == b) {
3694 /* a % a = 0 if a != 0 */
3695 ir_graph *irg = get_irn_irg(proj);
3696 ir_mode *mode = get_irn_mode(proj);
3697 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3699 DBG_OPT_CSTEVAL(mod, res);
3705 } /* transform_node_Proj_Mod */
3708 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3710 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3712 ir_node *n = get_Proj_pred(proj);
3713 ir_node *b = get_Cond_selector(n);
3715 if (!get_opt_unreachable_code())
3718 if (mode_is_int(get_irn_mode(b))) {
3719 ir_tarval *tb = value_of(b);
3721 if (tb != tarval_bad) {
3722 /* we have a constant switch */
3723 long num = get_Proj_proj(proj);
3725 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3726 if (get_tarval_long(tb) == num) {
3727 /* Do NOT create a jump here, or we will have 2 control flow ops
3728 * in a block. This case is optimized away in optimize_cf(). */
3731 ir_graph *irg = get_irn_irg(proj);
3732 /* this case will NEVER be taken, kill it */
3733 return new_r_Bad(irg, mode_X);
3737 long num = get_Proj_proj(proj);
3738 vrp_attr *b_vrp = vrp_get_info(b);
3739 if (num != get_Cond_default_proj(n) && b_vrp) {
3740 /* Try handling with vrp data. We only remove dead parts. */
3741 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3743 if (b_vrp->range_type == VRP_RANGE) {
3744 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3745 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3747 if ((cmp_result & ir_relation_greater) == cmp_result
3748 && (cmp_result2 & ir_relation_less) == cmp_result2) {
3749 ir_graph *irg = get_irn_irg(proj);
3750 return new_r_Bad(irg, mode_X);
3752 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3753 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3754 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3756 if ((cmp_result & ir_relation_less_equal) == cmp_result
3757 && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) {
3758 ir_graph *irg = get_irn_irg(proj);
3759 return new_r_Bad(irg, mode_X);
3764 tarval_and( b_vrp->bits_set, tp),
3766 ) == ir_relation_equal)) {
3767 ir_graph *irg = get_irn_irg(proj);
3768 return new_r_Bad(irg, mode_X);
3774 tarval_not(b_vrp->bits_not_set)),
3775 tarval_not(b_vrp->bits_not_set))
3776 == ir_relation_equal)) {
3777 ir_graph *irg = get_irn_irg(proj);
3778 return new_r_Bad(irg, mode_X);
3787 * return true if the operation returns a value with exactly 1 bit set
3789 static bool is_single_bit(const ir_node *node)
3791 /* a first implementation, could be extended with vrp and others... */
3793 ir_node *shl_l = get_Shl_left(node);
3794 ir_mode *mode = get_irn_mode(node);
3795 int modulo = get_mode_modulo_shift(mode);
3796 /* this works if we shift a 1 and we have modulo shift */
3797 if (is_Const(shl_l) && is_Const_one(shl_l)
3798 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3801 } else if (is_Const(node)) {
3802 ir_tarval *tv = get_Const_tarval(node);
3803 return tarval_is_single_bit(tv);
3809 * Normalizes and optimizes Cmp nodes.
3811 static ir_node *transform_node_Cmp(ir_node *n)
3813 ir_node *left = get_Cmp_left(n);
3814 ir_node *right = get_Cmp_right(n);
3815 ir_mode *mode = get_irn_mode(left);
3816 ir_tarval *tv = NULL;
3817 bool changed = false;
3818 bool changedc = false;
3819 ir_relation relation = get_Cmp_relation(n);
3820 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3822 /* mask out impossible relations */
3823 ir_relation new_relation = relation & possible;
3824 if (new_relation != relation) {
3825 relation = new_relation;
3829 /* Remove unnecessary conversions */
3830 if (is_Conv(left) && is_Conv(right)) {
3831 ir_node *op_left = get_Conv_op(left);
3832 ir_node *op_right = get_Conv_op(right);
3833 ir_mode *mode_left = get_irn_mode(op_left);
3834 ir_mode *mode_right = get_irn_mode(op_right);
3836 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3837 && mode_left != mode_b && mode_right != mode_b) {
3838 ir_node *block = get_nodes_block(n);
3840 if (mode_left == mode_right) {
3844 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3845 } else if (smaller_mode(mode_left, mode_right)) {
3846 left = new_r_Conv(block, op_left, mode_right);
3849 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3850 } else if (smaller_mode(mode_right, mode_left)) {
3852 right = new_r_Conv(block, op_right, mode_left);
3854 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3856 mode = get_irn_mode(left);
3859 if (is_Conv(left) && is_Const(right)) {
3860 ir_node *op_left = get_Conv_op(left);
3861 ir_mode *mode_left = get_irn_mode(op_left);
3862 if (smaller_mode(mode_left, mode) && mode_left != mode_b) {
3863 ir_tarval *tv = get_Const_tarval(right);
3864 tarval_int_overflow_mode_t last_mode
3865 = tarval_get_integer_overflow_mode();
3867 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
3868 new_tv = tarval_convert_to(tv, mode_left);
3869 tarval_set_integer_overflow_mode(last_mode);
3870 if (new_tv != tarval_bad) {
3871 ir_graph *irg = get_irn_irg(n);
3873 right = new_r_Const(irg, new_tv);
3874 mode = get_irn_mode(left);
3876 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3882 * Optimize -a CMP -b into b CMP a.
3883 * This works only for modes where unary Minus cannot Overflow.
3884 * Note that two-complement integers can Overflow so it will NOT work.
3886 if (!mode_overflow_on_unary_Minus(mode) &&
3887 is_Minus(left) && is_Minus(right)) {
3888 left = get_Minus_op(left);
3889 right = get_Minus_op(right);
3890 relation = get_inversed_relation(relation);
3892 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3895 /* remove operation on both sides if possible */
3896 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3898 * The following operations are NOT safe for floating point operations, for instance
3899 * 1.0 + inf == 2.0 + inf, =/=> x == y
3901 if (mode_is_int(mode)) {
3902 unsigned lop = get_irn_opcode(left);
3904 if (lop == get_irn_opcode(right)) {
3905 ir_node *ll, *lr, *rl, *rr;
3907 /* same operation on both sides, try to remove */
3911 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3912 left = get_unop_op(left);
3913 right = get_unop_op(right);
3915 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3918 ll = get_Add_left(left);
3919 lr = get_Add_right(left);
3920 rl = get_Add_left(right);
3921 rr = get_Add_right(right);
3924 /* X + a CMP X + b ==> a CMP b */
3928 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3929 } else if (ll == rr) {
3930 /* X + a CMP b + X ==> a CMP b */
3934 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3935 } else if (lr == rl) {
3936 /* a + X CMP X + b ==> a CMP b */
3940 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3941 } else if (lr == rr) {
3942 /* a + X CMP b + X ==> a CMP b */
3946 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3950 ll = get_Sub_left(left);
3951 lr = get_Sub_right(left);
3952 rl = get_Sub_left(right);
3953 rr = get_Sub_right(right);
3956 /* X - a CMP X - b ==> a CMP b */
3960 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3961 } else if (lr == rr) {
3962 /* a - X CMP b - X ==> a CMP b */
3966 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3970 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3971 /* a ROTL X CMP b ROTL X ==> a CMP b */
3972 left = get_Rotl_left(left);
3973 right = get_Rotl_left(right);
3975 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3983 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3984 if (is_Add(left) || is_Sub(left)) {
3985 ir_node *ll = get_binop_left(left);
3986 ir_node *lr = get_binop_right(left);
3988 if (lr == right && is_Add(left)) {
3994 ir_graph *irg = get_irn_irg(n);
3996 right = create_zero_const(irg, mode);
3998 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4001 if (is_Add(right) || is_Sub(right)) {
4002 ir_node *rl = get_binop_left(right);
4003 ir_node *rr = get_binop_right(right);
4005 if (rr == left && is_Add(right)) {
4011 ir_graph *irg = get_irn_irg(n);
4013 right = create_zero_const(irg, mode);
4015 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4019 if (is_And(left) && is_Const(right)) {
4020 ir_node *ll = get_binop_left(left);
4021 ir_node *lr = get_binop_right(left);
4022 if (is_Shr(ll) && is_Const(lr)) {
4023 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4024 ir_node *block = get_nodes_block(n);
4025 ir_mode *mode = get_irn_mode(left);
4027 ir_node *llr = get_Shr_right(ll);
4028 if (is_Const(llr)) {
4029 dbg_info *dbg = get_irn_dbg_info(left);
4030 ir_graph *irg = get_irn_irg(left);
4032 ir_tarval *c1 = get_Const_tarval(llr);
4033 ir_tarval *c2 = get_Const_tarval(lr);
4034 ir_tarval *c3 = get_Const_tarval(right);
4035 ir_tarval *mask = tarval_shl(c2, c1);
4036 ir_tarval *value = tarval_shl(c3, c1);
4038 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4039 right = new_r_Const(irg, value);
4044 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4046 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4047 right = get_Eor_right(left);
4048 left = get_Eor_left(left);
4051 } /* mode_is_int(...) */
4054 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4055 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
4056 if (mode_is_int(mode) && is_And(left)
4057 && (relation == ir_relation_equal
4058 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4059 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
4060 ir_node *and0 = get_And_left(left);
4061 ir_node *and1 = get_And_right(left);
4062 if (and1 == right) {
4063 ir_node *tmp = and0;
4067 if (and0 == right && is_single_bit(and0)) {
4068 ir_graph *irg = get_irn_irg(n);
4070 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
4071 right = create_zero_const(irg, mode);
4076 /* replace mode_b compares with ands/ors */
4077 if (mode == mode_b) {
4078 ir_node *block = get_nodes_block(n);
4082 case ir_relation_less_equal:
4083 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4085 case ir_relation_less:
4086 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4088 case ir_relation_greater_equal:
4089 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4091 case ir_relation_greater:
4092 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4094 case ir_relation_less_greater:
4095 bres = new_r_Eor(block, left, right, mode_b);
4097 case ir_relation_equal:
4098 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4101 #ifdef DEBUG_libfirm
4102 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4107 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4113 * First step: normalize the compare op
4114 * by placing the constant on the right side
4115 * or moving the lower address node to the left.
4117 if (!operands_are_normalized(left, right)) {
4122 relation = get_inversed_relation(relation);
4127 * Second step: Try to reduce the magnitude
4128 * of a constant. This may help to generate better code
4129 * later and may help to normalize more compares.
4130 * Of course this is only possible for integer values.
4132 tv = value_of(right);
4133 if (tv != tarval_bad) {
4134 ir_mode *mode = get_irn_mode(right);
4136 /* TODO extend to arbitrary constants */
4137 if (is_Conv(left) && tarval_is_null(tv)) {
4138 ir_node *op = get_Conv_op(left);
4139 ir_mode *op_mode = get_irn_mode(op);
4142 * UpConv(x) REL 0 ==> x REL 0
4143 * Don't do this for float values as it's unclear whether it is a
4144 * win. (on the other side it makes detection/creation of fabs hard)
4146 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4147 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4148 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4149 !mode_is_float(mode)) {
4150 tv = get_mode_null(op_mode);
4154 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4158 if (tv != tarval_bad) {
4159 /* the following optimization is possible on modes without Overflow
4160 * on Unary Minus or on == and !=:
4161 * -a CMP c ==> a swap(CMP) -c
4163 * Beware: for two-complement Overflow may occur, so only == and != can
4164 * be optimized, see this:
4165 * -MININT < 0 =/=> MININT > 0 !!!
4167 if (is_Minus(left) &&
4168 (!mode_overflow_on_unary_Minus(mode) ||
4169 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4170 tv = tarval_neg(tv);
4172 if (tv != tarval_bad) {
4173 left = get_Minus_op(left);
4174 relation = get_inversed_relation(relation);
4176 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4178 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4179 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4180 tv = tarval_not(tv);
4182 if (tv != tarval_bad) {
4183 left = get_Not_op(left);
4185 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4189 /* for integer modes, we have more */
4190 if (mode_is_int(mode)) {
4191 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4192 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4193 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4194 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4196 if (tv != tarval_bad) {
4197 relation ^= ir_relation_equal;
4199 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4202 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4203 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4204 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4205 tv = tarval_add(tv, get_mode_one(mode));
4207 if (tv != tarval_bad) {
4208 relation ^= ir_relation_equal;
4210 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4214 /* the following reassociations work only for == and != */
4215 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4216 if (tv != tarval_bad) {
4217 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4219 ir_node *c1 = get_Sub_right(left);
4220 ir_tarval *tv2 = value_of(c1);
4222 if (tv2 != tarval_bad) {
4223 tv2 = tarval_add(tv, value_of(c1));
4225 if (tv2 != tarval_bad) {
4226 left = get_Sub_left(left);
4229 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4233 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4234 else if (is_Add(left)) {
4235 ir_node *a_l = get_Add_left(left);
4236 ir_node *a_r = get_Add_right(left);
4240 if (is_Const(a_l)) {
4242 tv2 = value_of(a_l);
4245 tv2 = value_of(a_r);
4248 if (tv2 != tarval_bad) {
4249 tv2 = tarval_sub(tv, tv2, NULL);
4251 if (tv2 != tarval_bad) {
4255 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4259 /* -a == c ==> a == -c, -a != c ==> a != -c */
4260 else if (is_Minus(left)) {
4261 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4263 if (tv2 != tarval_bad) {
4264 left = get_Minus_op(left);
4267 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4274 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4275 switch (get_irn_opcode(left)) {
4279 c1 = get_And_right(left);
4282 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4283 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4285 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4287 /* TODO: move to constant evaluation */
4288 ir_graph *irg = get_irn_irg(n);
4289 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4290 c1 = new_r_Const(irg, tv);
4291 DBG_OPT_CSTEVAL(n, c1);
4295 if (tarval_is_single_bit(tv)) {
4297 * optimization for AND:
4299 * And(x, C) == C ==> And(x, C) != 0
4300 * And(x, C) != C ==> And(X, C) == 0
4302 * if C is a single Bit constant.
4305 /* check for Constant's match. We have check hare the tarvals,
4306 because our const might be changed */
4307 if (get_Const_tarval(c1) == tv) {
4308 /* fine: do the transformation */
4309 tv = get_mode_null(get_tarval_mode(tv));
4310 relation ^= ir_relation_less_equal_greater;
4312 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4318 c1 = get_Or_right(left);
4319 if (is_Const(c1) && tarval_is_null(tv)) {
4321 * Or(x, C) == 0 && C != 0 ==> FALSE
4322 * Or(x, C) != 0 && C != 0 ==> TRUE
4324 if (! tarval_is_null(get_Const_tarval(c1))) {
4325 /* TODO: move to constant evaluation */
4326 ir_graph *irg = get_irn_irg(n);
4327 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4328 c1 = new_r_Const(irg, tv);
4329 DBG_OPT_CSTEVAL(n, c1);
4336 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4338 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4341 c1 = get_Shl_right(left);
4343 ir_graph *irg = get_irn_irg(c1);
4344 ir_tarval *tv1 = get_Const_tarval(c1);
4345 ir_mode *mode = get_irn_mode(left);
4346 ir_tarval *minus1 = get_mode_all_one(mode);
4347 ir_tarval *amask = tarval_shr(minus1, tv1);
4348 ir_tarval *cmask = tarval_shl(minus1, tv1);
4351 if (tarval_and(tv, cmask) != tv) {
4352 /* condition not met */
4353 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4354 c1 = new_r_Const(irg, tv);
4355 DBG_OPT_CSTEVAL(n, c1);
4358 sl = get_Shl_left(left);
4359 blk = get_nodes_block(n);
4360 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4361 tv = tarval_shr(tv, tv1);
4363 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4368 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4370 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4373 c1 = get_Shr_right(left);
4375 ir_graph *irg = get_irn_irg(c1);
4376 ir_tarval *tv1 = get_Const_tarval(c1);
4377 ir_mode *mode = get_irn_mode(left);
4378 ir_tarval *minus1 = get_mode_all_one(mode);
4379 ir_tarval *amask = tarval_shl(minus1, tv1);
4380 ir_tarval *cmask = tarval_shr(minus1, tv1);
4383 if (tarval_and(tv, cmask) != tv) {
4384 /* condition not met */
4385 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4386 c1 = new_r_Const(irg, tv);
4387 DBG_OPT_CSTEVAL(n, c1);
4390 sl = get_Shr_left(left);
4391 blk = get_nodes_block(n);
4392 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4393 tv = tarval_shl(tv, tv1);
4395 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4400 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4402 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4405 c1 = get_Shrs_right(left);
4407 ir_graph *irg = get_irn_irg(c1);
4408 ir_tarval *tv1 = get_Const_tarval(c1);
4409 ir_mode *mode = get_irn_mode(left);
4410 ir_tarval *minus1 = get_mode_all_one(mode);
4411 ir_tarval *amask = tarval_shl(minus1, tv1);
4412 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4415 cond = tarval_sub(cond, tv1, NULL);
4416 cond = tarval_shrs(tv, cond);
4418 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4419 /* condition not met */
4420 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4421 c1 = new_r_Const(irg, tv);
4422 DBG_OPT_CSTEVAL(n, c1);
4425 sl = get_Shrs_left(left);
4426 blk = get_nodes_block(n);
4427 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4428 tv = tarval_shl(tv, tv1);
4430 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4435 } /* tarval != bad */
4438 if (changedc) { /* need a new Const */
4439 ir_graph *irg = get_irn_irg(n);
4440 right = new_r_Const(irg, tv);
4444 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4445 ir_node *op = get_Proj_pred(left);
4447 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4448 ir_node *c = get_binop_right(op);
4451 ir_tarval *tv = get_Const_tarval(c);
4453 if (tarval_is_single_bit(tv)) {
4454 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4455 ir_node *v = get_binop_left(op);
4456 ir_node *blk = get_irn_n(op, -1);
4457 ir_graph *irg = get_irn_irg(op);
4458 ir_mode *mode = get_irn_mode(v);
4460 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4461 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4463 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4470 dbg_info *dbgi = get_irn_dbg_info(n);
4471 ir_node *block = get_nodes_block(n);
4473 /* create a new compare */
4474 n = new_rd_Cmp(dbgi, block, left, right, relation);
4481 * Optimize CopyB(mem, x, x) into a Nop.
4483 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4485 ir_node *copyb = get_Proj_pred(proj);
4486 ir_node *a = get_CopyB_dst(copyb);
4487 ir_node *b = get_CopyB_src(copyb);
4490 switch (get_Proj_proj(proj)) {
4491 case pn_CopyB_X_regular:
4492 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4493 DBG_OPT_EXC_REM(proj);
4494 proj = new_r_Jmp(get_nodes_block(copyb));
4496 case pn_CopyB_X_except: {
4497 ir_graph *irg = get_irn_irg(proj);
4498 DBG_OPT_EXC_REM(proj);
4499 proj = new_r_Bad(irg, mode_X);
4507 } /* transform_node_Proj_CopyB */
4510 * Optimize Bounds(idx, idx, upper) into idx.
4512 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4514 ir_node *oldn = proj;
4515 ir_node *bound = get_Proj_pred(proj);
4516 ir_node *idx = get_Bound_index(bound);
4517 ir_node *pred = skip_Proj(idx);
4520 if (idx == get_Bound_lower(bound))
4522 else if (is_Bound(pred)) {
4524 * idx was Bounds checked previously, it is still valid if
4525 * lower <= pred_lower && pred_upper <= upper.
4527 ir_node *lower = get_Bound_lower(bound);
4528 ir_node *upper = get_Bound_upper(bound);
4529 if (get_Bound_lower(pred) == lower &&
4530 get_Bound_upper(pred) == upper) {
4532 * One could expect that we simply return the previous
4533 * Bound here. However, this would be wrong, as we could
4534 * add an exception Proj to a new location then.
4535 * So, we must turn in into a tuple.
4541 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4542 switch (get_Proj_proj(proj)) {
4544 DBG_OPT_EXC_REM(proj);
4545 proj = get_Bound_mem(bound);
4547 case pn_Bound_X_except:
4548 DBG_OPT_EXC_REM(proj);
4549 proj = new_r_Bad(get_irn_irg(proj), mode_X);
4553 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4555 case pn_Bound_X_regular:
4556 DBG_OPT_EXC_REM(proj);
4557 proj = new_r_Jmp(get_nodes_block(bound));
4564 } /* transform_node_Proj_Bound */
4567 * Does all optimizations on nodes that must be done on its Projs
4568 * because of creating new nodes.
4570 static ir_node *transform_node_Proj(ir_node *proj)
4572 ir_node *n = get_Proj_pred(proj);
4574 if (n->op->ops.transform_node_Proj)
4575 return n->op->ops.transform_node_Proj(proj);
4577 } /* transform_node_Proj */
4579 static bool is_block_unreachable(const ir_node *block)
4581 const ir_graph *irg = get_irn_irg(block);
4582 if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK))
4584 return get_Block_dom_depth(block) < 0;
4587 static ir_node *transform_node_Block(ir_node *block)
4589 ir_graph *irg = get_irn_irg(block);
4590 int arity = get_irn_arity(block);
4591 ir_node *bad = NULL;
4594 if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK))
4597 for (i = 0; i < arity; ++i) {
4598 ir_node *pred = get_Block_cfgpred(block, i);
4599 ir_node *pred_block = get_nodes_block(pred);
4600 if (!is_Bad(pred) && !is_block_unreachable(pred_block))
4603 bad = new_r_Bad(irg, mode_X);
4604 set_irn_n(block, i, bad);
4610 static ir_node *transform_node_Phi(ir_node *phi)
4612 int n = get_irn_arity(phi);
4613 ir_mode *mode = get_irn_mode(phi);
4614 ir_node *block = get_nodes_block(phi);
4615 ir_graph *irg = get_irn_irg(phi);
4616 ir_node *bad = NULL;
4619 /* Set phi-operands for bad-block inputs to bad */
4620 for (i = 0; i < n; ++i) {
4621 ir_node *pred = get_Block_cfgpred(block, i);
4622 if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) {
4624 bad = new_r_Bad(irg, mode);
4625 set_irn_n(phi, i, bad);
4629 /* Move Confirms down through Phi nodes. */
4630 if (mode_is_reference(mode)) {
4631 n = get_irn_arity(phi);
4633 /* Beware of Phi0 */
4635 ir_node *pred = get_irn_n(phi, 0);
4636 ir_node *bound, *new_phi, *block, **in;
4637 ir_relation relation;
4639 if (! is_Confirm(pred))
4642 bound = get_Confirm_bound(pred);
4643 relation = get_Confirm_relation(pred);
4645 NEW_ARR_A(ir_node *, in, n);
4646 in[0] = get_Confirm_value(pred);
4648 for (i = 1; i < n; ++i) {
4649 pred = get_irn_n(phi, i);
4651 if (! is_Confirm(pred) ||
4652 get_Confirm_bound(pred) != bound ||
4653 get_Confirm_relation(pred) != relation)
4655 in[i] = get_Confirm_value(pred);
4657 /* move the Confirm nodes "behind" the Phi */
4658 block = get_irn_n(phi, -1);
4659 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4660 return new_r_Confirm(block, new_phi, bound, relation);
4667 * Returns the operands of a commutative bin-op, if one operand is
4668 * a const, it is returned as the second one.
4670 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4672 ir_node *op_a = get_binop_left(binop);
4673 ir_node *op_b = get_binop_right(binop);
4675 assert(is_op_commutative(get_irn_op(binop)));
4677 if (is_Const(op_a)) {
4684 } /* get_comm_Binop_Ops */
4687 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4688 * Such pattern may arise in bitfield stores.
4690 * value c4 value c4 & c2
4691 * AND c3 AND c1 | c3
4698 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4701 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4703 ir_node *irn_and, *c1;
4705 ir_node *and_l, *c3;
4706 ir_node *value, *c4;
4707 ir_node *new_and, *new_const, *block;
4708 ir_mode *mode = get_irn_mode(irn_or);
4710 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4714 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4715 if (!is_Const(c1) || !is_And(irn_and))
4718 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4722 tv1 = get_Const_tarval(c1);
4723 tv2 = get_Const_tarval(c2);
4725 tv = tarval_or(tv1, tv2);
4726 if (tarval_is_all_one(tv)) {
4727 /* the AND does NOT clear a bit with isn't set by the OR */
4728 set_Or_left(irn_or, or_l);
4729 set_Or_right(irn_or, c1);
4731 /* check for more */
4738 get_comm_Binop_Ops(or_l, &and_l, &c3);
4739 if (!is_Const(c3) || !is_And(and_l))
4742 get_comm_Binop_Ops(and_l, &value, &c4);
4746 /* ok, found the pattern, check for conditions */
4747 assert(mode == get_irn_mode(irn_and));
4748 assert(mode == get_irn_mode(or_l));
4749 assert(mode == get_irn_mode(and_l));
4751 tv3 = get_Const_tarval(c3);
4752 tv4 = get_Const_tarval(c4);
4754 tv = tarval_or(tv4, tv2);
4755 if (!tarval_is_all_one(tv)) {
4756 /* have at least one 0 at the same bit position */
4760 if (tv3 != tarval_andnot(tv3, tv4)) {
4761 /* bit in the or_mask is outside the and_mask */
4765 if (tv1 != tarval_andnot(tv1, tv2)) {
4766 /* bit in the or_mask is outside the and_mask */
4770 /* ok, all conditions met */
4771 block = get_irn_n(irn_or, -1);
4772 irg = get_irn_irg(block);
4774 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4776 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4778 set_Or_left(irn_or, new_and);
4779 set_Or_right(irn_or, new_const);
4781 /* check for more */
4783 } /* transform_node_Or_bf_store */
4786 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4788 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4790 ir_mode *mode = get_irn_mode(irn_or);
4791 ir_node *shl, *shr, *block;
4792 ir_node *irn, *x, *c1, *c2, *n;
4793 ir_tarval *tv1, *tv2;
4795 /* some backends can't handle rotl */
4796 if (!be_get_backend_param()->support_rotl)
4799 if (! mode_is_int(mode))
4802 shl = get_binop_left(irn_or);
4803 shr = get_binop_right(irn_or);
4812 } else if (!is_Shl(shl)) {
4814 } else if (!is_Shr(shr)) {
4817 x = get_Shl_left(shl);
4818 if (x != get_Shr_left(shr))
4821 c1 = get_Shl_right(shl);
4822 c2 = get_Shr_right(shr);
4823 if (is_Const(c1) && is_Const(c2)) {
4824 tv1 = get_Const_tarval(c1);
4825 if (! tarval_is_long(tv1))
4828 tv2 = get_Const_tarval(c2);
4829 if (! tarval_is_long(tv2))
4832 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4833 != (int) get_mode_size_bits(mode))
4836 /* yet, condition met */
4837 block = get_nodes_block(irn_or);
4839 n = new_r_Rotl(block, x, c1, mode);
4841 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4845 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4846 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4847 if (!ir_is_negated_value(c1, c2)) {
4851 /* yet, condition met */
4852 block = get_nodes_block(irn_or);
4853 n = new_r_Rotl(block, x, c1, mode);
4854 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4856 } /* transform_node_Or_Rotl */
4858 static bool is_cmp_unequal(const ir_node *node)
4860 ir_relation relation = get_Cmp_relation(node);
4861 ir_node *left = get_Cmp_left(node);
4862 ir_node *right = get_Cmp_right(node);
4863 ir_mode *mode = get_irn_mode(left);
4865 if (relation == ir_relation_less_greater)
4868 if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right))
4869 return relation == ir_relation_greater;
4876 static ir_node *transform_node_Or(ir_node *n)
4878 ir_node *c, *oldn = n;
4879 ir_node *a = get_Or_left(n);
4880 ir_node *b = get_Or_right(n);
4883 if (is_Not(a) && is_Not(b)) {
4884 /* ~a | ~b = ~(a&b) */
4885 ir_node *block = get_nodes_block(n);
4887 mode = get_irn_mode(n);
4890 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4891 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4892 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4896 /* we can combine the relations of two compares with the same operands */
4897 if (is_Cmp(a) && is_Cmp(b)) {
4898 ir_node *a_left = get_Cmp_left(a);
4899 ir_node *a_right = get_Cmp_left(a);
4900 ir_node *b_left = get_Cmp_left(b);
4901 ir_node *b_right = get_Cmp_right(b);
4902 if (a_left == b_left && b_left == b_right) {
4903 dbg_info *dbgi = get_irn_dbg_info(n);
4904 ir_node *block = get_nodes_block(n);
4905 ir_relation a_relation = get_Cmp_relation(a);
4906 ir_relation b_relation = get_Cmp_relation(b);
4907 ir_relation new_relation = a_relation | b_relation;
4908 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4910 /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */
4911 if (is_cmp_unequal(a) && is_cmp_unequal(b)
4912 && !mode_is_float(get_irn_mode(a_left))
4913 && !mode_is_float(get_irn_mode(b_left))
4914 && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
4915 ir_graph *irg = get_irn_irg(n);
4916 dbg_info *dbgi = get_irn_dbg_info(n);
4917 ir_node *block = get_nodes_block(n);
4918 ir_mode *a_mode = get_irn_mode(a_left);
4919 ir_mode *b_mode = get_irn_mode(b_left);
4920 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
4921 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
4922 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
4923 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
4924 ir_node *zero = create_zero_const(irg, a_mode);
4925 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4929 mode = get_irn_mode(n);
4930 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4932 n = transform_node_Or_bf_store(n);
4933 n = transform_node_Or_Rotl(n);
4937 n = transform_bitwise_distributive(n, transform_node_Or);
4939 n = transform_node_bitop_shift(n);
4942 } /* transform_node_Or */
4946 static ir_node *transform_node(ir_node *n);
4949 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4951 * Should be moved to reassociation?
4953 static ir_node *transform_node_shift(ir_node *n)
4955 ir_node *left, *right;
4957 ir_mode *count_mode;
4958 ir_tarval *tv1, *tv2, *res;
4959 ir_node *in[2], *irn, *block;
4963 left = get_binop_left(n);
4965 /* different operations */
4966 if (get_irn_op(left) != get_irn_op(n))
4969 right = get_binop_right(n);
4970 tv1 = value_of(right);
4971 if (tv1 == tarval_bad)
4974 tv2 = value_of(get_binop_right(left));
4975 if (tv2 == tarval_bad)
4978 count_mode = get_tarval_mode(tv1);
4979 if (get_tarval_mode(tv2) != count_mode) {
4980 /* TODO: search bigger mode or something and convert... */
4984 mode = get_irn_mode(n);
4985 modulo_shf = get_mode_modulo_shift(mode);
4987 if (modulo_shf > 0) {
4988 ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode);
4990 /* I'm not so sure what happens in one complement... */
4991 assert(get_mode_arithmetic(count_mode) == irma_twos_complement);
4992 /* modulo shifts should always be a power of 2 (otherwise modulo_mask
4993 * above will be invalid) */
4994 assert(modulo_shf<=0 || is_po2(modulo_shf));
4996 tv1 = tarval_and(tv1, modulo_mask);
4997 tv2 = tarval_and(tv2, modulo_mask);
4999 res = tarval_add(tv1, tv2);
5000 irg = get_irn_irg(n);
5002 /* beware: a simple replacement works only, if res < modulo shift */
5004 int bits = get_mode_size_bits(mode);
5005 ir_tarval *modulo = new_tarval_from_long(bits, count_mode);
5006 res = tarval_mod(res, modulo);
5008 long bits = get_mode_size_bits(mode);
5009 ir_tarval *mode_size = new_tarval_from_long(bits, count_mode);
5011 /* shifting too much */
5012 if (!(tarval_cmp(res, mode_size) & ir_relation_less)) {
5014 ir_node *block = get_nodes_block(n);
5015 dbg_info *dbgi = get_irn_dbg_info(n);
5016 ir_mode *smode = get_irn_mode(right);
5017 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
5018 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5021 return new_r_Const(irg, get_mode_null(mode));
5025 /* ok, we can replace it */
5026 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5027 block = get_nodes_block(n);
5029 in[0] = get_binop_left(left);
5030 in[1] = new_r_Const(irg, res);
5032 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5034 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5036 return transform_node(irn);
5041 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5043 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5044 * (also with x >>s c1 when c1>=c2)
5046 static ir_node *transform_node_shl_shr(ir_node *n)
5049 ir_node *right = get_binop_right(n);
5059 ir_tarval *tv_shift;
5062 ir_relation relation;
5065 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5067 if (!is_Const(right))
5070 left = get_binop_left(n);
5071 mode = get_irn_mode(n);
5072 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5073 ir_node *shr_right = get_binop_right(left);
5075 if (!is_Const(shr_right))
5078 x = get_binop_left(left);
5079 tv_shr = get_Const_tarval(shr_right);
5080 tv_shl = get_Const_tarval(right);
5082 if (is_Shrs(left)) {
5083 /* shrs variant only allowed if c1 >= c2 */
5084 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5087 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5090 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5092 tv_mask = tarval_shl(tv_mask, tv_shl);
5093 } else if (is_Shr(n) && is_Shl(left)) {
5094 ir_node *shl_right = get_Shl_right(left);
5096 if (!is_Const(shl_right))
5099 x = get_Shl_left(left);
5100 tv_shr = get_Const_tarval(right);
5101 tv_shl = get_Const_tarval(shl_right);
5103 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5104 tv_mask = tarval_shr(tv_mask, tv_shr);
5109 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5110 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5113 assert(tv_mask != tarval_bad);
5114 assert(get_tarval_mode(tv_mask) == mode);
5116 block = get_nodes_block(n);
5117 irg = get_irn_irg(block);
5118 dbgi = get_irn_dbg_info(n);
5120 relation = tarval_cmp(tv_shl, tv_shr);
5121 if (relation == ir_relation_less || relation == ir_relation_equal) {
5122 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5123 new_const = new_r_Const(irg, tv_shift);
5125 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5127 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5130 assert(relation == ir_relation_greater);
5131 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5132 new_const = new_r_Const(irg, tv_shift);
5133 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5136 new_const = new_r_Const(irg, tv_mask);
5137 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5142 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5144 ir_mode *mode = get_tarval_mode(tv);
5145 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5146 return tarval_mod(tv, modulo_tv);
5149 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5150 ir_node *left, ir_node *right, ir_mode *mode);
5153 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5154 * then we can use that to minimize the value of Add(x, const) or
5155 * Sub(Const, x). In particular this often avoids 1 instruction in some
5156 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5157 * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
5159 static ir_node *transform_node_shift_modulo(ir_node *n,
5160 new_shift_func new_shift)
5162 ir_mode *mode = get_irn_mode(n);
5163 int modulo = get_mode_modulo_shift(mode);
5164 ir_node *newop = NULL;
5165 ir_mode *mode_right;
5172 if (get_mode_arithmetic(mode) != irma_twos_complement)
5174 if (!is_po2(modulo))
5177 irg = get_irn_irg(n);
5178 block = get_nodes_block(n);
5179 right = get_binop_right(n);
5180 mode_right = get_irn_mode(right);
5181 if (is_Const(right)) {
5182 ir_tarval *tv = get_Const_tarval(right);
5183 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5188 newop = new_r_Const(irg, tv_mod);
5189 } else if (is_Add(right)) {
5190 ir_node *add_right = get_Add_right(right);
5191 if (is_Const(add_right)) {
5192 ir_tarval *tv = get_Const_tarval(add_right);
5193 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5198 newconst = new_r_Const(irg, tv_mod);
5199 newop = new_r_Add(block, get_Add_left(right), newconst,
5202 } else if (is_Sub(right)) {
5203 ir_node *sub_left = get_Sub_left(right);
5204 if (is_Const(sub_left)) {
5205 ir_tarval *tv = get_Const_tarval(sub_left);
5206 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5211 newconst = new_r_Const(irg, tv_mod);
5212 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5219 if (newop != NULL) {
5220 dbg_info *dbgi = get_irn_dbg_info(n);
5221 ir_node *left = get_binop_left(n);
5222 return new_shift(dbgi, block, left, newop, mode);
5230 static ir_node *transform_node_Shr(ir_node *n)
5232 ir_node *c, *oldn = n;
5233 ir_node *left = get_Shr_left(n);
5234 ir_node *right = get_Shr_right(n);
5235 ir_mode *mode = get_irn_mode(n);
5237 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5238 n = transform_node_shift(n);
5241 n = transform_node_shift_modulo(n, new_rd_Shr);
5243 n = transform_node_shl_shr(n);
5245 n = transform_node_shift_bitop(n);
5248 } /* transform_node_Shr */
5253 static ir_node *transform_node_Shrs(ir_node *n)
5255 ir_node *c, *oldn = n;
5256 ir_node *a = get_Shrs_left(n);
5257 ir_node *b = get_Shrs_right(n);
5258 ir_mode *mode = get_irn_mode(n);
5260 if (is_oversize_shift(n)) {
5261 ir_node *block = get_nodes_block(n);
5262 dbg_info *dbgi = get_irn_dbg_info(n);
5263 ir_mode *cmode = get_irn_mode(b);
5264 long val = get_mode_size_bits(cmode)-1;
5265 ir_graph *irg = get_irn_irg(n);
5266 ir_node *cnst = new_r_Const_long(irg, cmode, val);
5267 return new_rd_Shrs(dbgi, block, a, cnst, mode);
5270 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5271 n = transform_node_shift(n);
5274 n = transform_node_shift_modulo(n, new_rd_Shrs);
5276 n = transform_node_shift_bitop(n);
5279 } /* transform_node_Shrs */
5284 static ir_node *transform_node_Shl(ir_node *n)
5286 ir_node *c, *oldn = n;
5287 ir_node *a = get_Shl_left(n);
5288 ir_node *b = get_Shl_right(n);
5289 ir_mode *mode = get_irn_mode(n);
5291 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5292 n = transform_node_shift(n);
5295 n = transform_node_shift_modulo(n, new_rd_Shl);
5297 n = transform_node_shl_shr(n);
5299 n = transform_node_shift_bitop(n);
5302 } /* transform_node_Shl */
5307 static ir_node *transform_node_Rotl(ir_node *n)
5309 ir_node *c, *oldn = n;
5310 ir_node *a = get_Rotl_left(n);
5311 ir_node *b = get_Rotl_right(n);
5312 ir_mode *mode = get_irn_mode(n);
5314 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5315 n = transform_node_shift(n);
5318 n = transform_node_shift_bitop(n);
5321 } /* transform_node_Rotl */
5326 static ir_node *transform_node_Conv(ir_node *n)
5328 ir_node *c, *oldn = n;
5329 ir_mode *mode = get_irn_mode(n);
5330 ir_node *a = get_Conv_op(n);
5332 if (mode != mode_b && is_const_Phi(a)) {
5333 /* Do NOT optimize mode_b Conv's, this leads to remaining
5334 * Phib nodes later, because the conv_b_lower operation
5335 * is instantly reverted, when it tries to insert a Convb.
5337 c = apply_conv_on_phi(a, mode);
5339 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5344 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5345 ir_graph *irg = get_irn_irg(n);
5346 return new_r_Unknown(irg, mode);
5349 if (mode_is_reference(mode) &&
5350 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5352 ir_node *l = get_Add_left(a);
5353 ir_node *r = get_Add_right(a);
5354 dbg_info *dbgi = get_irn_dbg_info(a);
5355 ir_node *block = get_nodes_block(n);
5357 ir_node *lop = get_Conv_op(l);
5358 if (get_irn_mode(lop) == mode) {
5359 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5360 n = new_rd_Add(dbgi, block, lop, r, mode);
5365 ir_node *rop = get_Conv_op(r);
5366 if (get_irn_mode(rop) == mode) {
5367 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5368 n = new_rd_Add(dbgi, block, l, rop, mode);
5375 } /* transform_node_Conv */
5378 * Remove dead blocks and nodes in dead blocks
5379 * in keep alive list. We do not generate a new End node.
5381 static ir_node *transform_node_End(ir_node *n)
5383 int i, j, n_keepalives = get_End_n_keepalives(n);
5386 NEW_ARR_A(ir_node *, in, n_keepalives);
5388 for (i = j = 0; i < n_keepalives; ++i) {
5389 ir_node *ka = get_End_keepalive(n, i);
5391 /* no need to keep Bad */
5394 /* do not keep unreachable code */
5395 block = is_Block(ka) ? ka : get_nodes_block(ka);
5396 if (is_block_unreachable(block))
5400 if (j != n_keepalives)
5401 set_End_keepalives(n, j, in);
5403 } /* transform_node_End */
5405 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5407 if (is_Minus(a) && get_Minus_op(a) == b)
5409 if (is_Minus(b) && get_Minus_op(b) == a)
5411 if (is_Sub(a) && is_Sub(b)) {
5412 ir_node *a_left = get_Sub_left(a);
5413 ir_node *a_right = get_Sub_right(a);
5414 ir_node *b_left = get_Sub_left(b);
5415 ir_node *b_right = get_Sub_right(b);
5417 if (a_left == b_right && a_right == b_left)
5424 static const ir_node *skip_upconv(const ir_node *node)
5426 while (is_Conv(node)) {
5427 ir_mode *mode = get_irn_mode(node);
5428 const ir_node *op = get_Conv_op(node);
5429 ir_mode *op_mode = get_irn_mode(op);
5430 if (!smaller_mode(op_mode, mode))
5437 int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true,
5438 const ir_node *mux_false)
5443 ir_relation relation;
5449 * Note further that these optimization work even for floating point
5450 * with NaN's because -NaN == NaN.
5451 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5454 mode = get_irn_mode(mux_true);
5455 if (mode_honor_signed_zeros(mode))
5458 /* must be <, <=, >=, > */
5459 relation = get_Cmp_relation(sel);
5460 if ((relation & ir_relation_less_greater) == 0)
5463 if (!ir_is_negated_value(mux_true, mux_false))
5466 mux_true = skip_upconv(mux_true);
5467 mux_false = skip_upconv(mux_false);
5469 /* must be x cmp 0 */
5470 cmp_right = get_Cmp_right(sel);
5471 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
5474 cmp_left = get_Cmp_left(sel);
5475 if (cmp_left == mux_false) {
5476 if (relation & ir_relation_less) {
5479 assert(relation & ir_relation_greater);
5482 } else if (cmp_left == mux_true) {
5483 if (relation & ir_relation_less) {
5486 assert(relation & ir_relation_greater);
5494 ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true,
5497 ir_node *cmp_left = get_Cmp_left(sel);
5498 return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
5502 * Optimize a Mux into some simpler cases.
5504 static ir_node *transform_node_Mux(ir_node *n)
5507 ir_node *sel = get_Mux_sel(n);
5508 ir_mode *mode = get_irn_mode(n);
5509 ir_node *t = get_Mux_true(n);
5510 ir_node *f = get_Mux_false(n);
5511 ir_graph *irg = get_irn_irg(n);
5513 /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
5514 if (get_mode_arithmetic(mode) == irma_twos_complement) {
5515 int abs = ir_mux_is_abs(sel, t, f);
5517 dbg_info *dbgi = get_irn_dbg_info(n);
5518 ir_node *block = get_nodes_block(n);
5519 ir_node *op = ir_get_abs_op(sel, t, f);
5520 int bits = get_mode_size_bits(mode);
5521 ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
5522 ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
5523 ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode);
5526 res = new_rd_Sub(dbgi, block, xorn, sext, mode);
5528 res = new_rd_Sub(dbgi, block, sext, xorn, mode);
5534 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5538 ir_node* block = get_nodes_block(n);
5540 ir_node* c1 = get_Mux_sel(t);
5541 ir_node* t1 = get_Mux_true(t);
5542 ir_node* f1 = get_Mux_false(t);
5544 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5545 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5546 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5551 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5552 } else if (f == t1) {
5553 /* Mux(cond0, Mux(cond1, x, y), x) */
5554 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5555 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5556 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5561 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5563 } else if (is_Mux(f)) {
5564 ir_node* block = get_nodes_block(n);
5566 ir_node* c1 = get_Mux_sel(f);
5567 ir_node* t1 = get_Mux_true(f);
5568 ir_node* f1 = get_Mux_false(f);
5570 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5571 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5572 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5577 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5578 } else if (t == f1) {
5579 /* Mux(cond0, x, Mux(cond1, y, x)) */
5580 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5581 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5582 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5587 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5591 /* first normalization step: try to move a constant to the false side,
5592 * 0 preferred on false side too */
5593 if (is_Cmp(sel) && is_Const(t) &&
5594 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5595 dbg_info *seldbgi = get_irn_dbg_info(sel);
5596 ir_node *block = get_nodes_block(sel);
5597 ir_relation relation = get_Cmp_relation(sel);
5602 /* Mux(x, a, b) => Mux(not(x), b, a) */
5603 relation = get_negated_relation(relation);
5604 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5605 get_Cmp_right(sel), relation);
5606 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5609 /* note: after normalization, false can only happen on default */
5610 if (mode == mode_b) {
5611 dbg_info *dbg = get_irn_dbg_info(n);
5612 ir_node *block = get_nodes_block(n);
5615 ir_tarval *tv_t = get_Const_tarval(t);
5616 if (tv_t == tarval_b_true) {
5618 /* Muxb(sel, true, false) = sel */
5619 assert(get_Const_tarval(f) == tarval_b_false);
5620 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5623 /* Muxb(sel, true, x) = Or(sel, x) */
5624 n = new_rd_Or(dbg, block, sel, f, mode_b);
5625 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5629 } else if (is_Const(f)) {
5630 ir_tarval *tv_f = get_Const_tarval(f);
5631 if (tv_f == tarval_b_true) {
5632 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5633 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5634 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5635 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5638 /* Muxb(sel, x, false) = And(sel, x) */
5639 assert(tv_f == tarval_b_false);
5640 n = new_rd_And(dbg, block, sel, t, mode_b);
5641 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5647 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5648 * value to integer. */
5649 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5650 ir_tarval *a = get_Const_tarval(t);
5651 ir_tarval *b = get_Const_tarval(f);
5653 if (tarval_is_one(a) && tarval_is_null(b)) {
5654 ir_node *block = get_nodes_block(n);
5655 ir_node *conv = new_r_Conv(block, sel, mode);
5657 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5659 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5660 ir_node *block = get_nodes_block(n);
5661 ir_node *not_ = new_r_Not(block, sel, mode_b);
5662 ir_node *conv = new_r_Conv(block, not_, mode);
5664 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5670 ir_node *cmp_r = get_Cmp_right(sel);
5671 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5672 ir_node *block = get_nodes_block(n);
5673 ir_node *cmp_l = get_Cmp_left(sel);
5675 if (mode_is_int(mode)) {
5676 ir_relation relation = get_Cmp_relation(sel);
5678 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5679 /* Mux((a & b) != 0, c, 0) */
5680 ir_node *and_r = get_And_right(cmp_l);
5683 if (and_r == t && f == cmp_r) {
5684 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5685 if (relation == ir_relation_less_greater) {
5686 /* Mux((a & 2^C) != 0, 2^C, 0) */
5688 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5690 /* Mux((a & 2^C) == 0, 2^C, 0) */
5691 n = new_rd_Eor(get_irn_dbg_info(n),
5692 block, cmp_l, t, mode);
5693 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5698 if (is_Shl(and_r)) {
5699 ir_node *shl_l = get_Shl_left(and_r);
5700 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5701 if (and_r == t && f == cmp_r) {
5702 if (relation == ir_relation_less_greater) {
5703 /* (a & (1 << n)) != 0, (1 << n), 0) */
5705 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5707 /* (a & (1 << n)) == 0, (1 << n), 0) */
5708 n = new_rd_Eor(get_irn_dbg_info(n),
5709 block, cmp_l, t, mode);
5710 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5716 and_l = get_And_left(cmp_l);
5717 if (is_Shl(and_l)) {
5718 ir_node *shl_l = get_Shl_left(and_l);
5719 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5720 if (and_l == t && f == cmp_r) {
5721 if (relation == ir_relation_less_greater) {
5722 /* ((1 << n) & a) != 0, (1 << n), 0) */
5724 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5726 /* ((1 << n) & a) == 0, (1 << n), 0) */
5727 n = new_rd_Eor(get_irn_dbg_info(n),
5728 block, cmp_l, t, mode);
5729 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5744 * optimize Sync nodes that have other syncs as input we simply add the inputs
5745 * of the other sync to our own inputs
5747 static ir_node *transform_node_Sync(ir_node *n)
5749 int arity = get_Sync_n_preds(n);
5752 for (i = 0; i < arity;) {
5753 ir_node *pred = get_Sync_pred(n, i);
5757 /* Remove Bad predecessors */
5764 /* Remove duplicate predecessors */
5765 for (j = 0; j < i; ++j) {
5766 if (get_Sync_pred(n, j) == pred) {
5775 if (!is_Sync(pred)) {
5783 pred_arity = get_Sync_n_preds(pred);
5784 for (j = 0; j < pred_arity; ++j) {
5785 ir_node *pred_pred = get_Sync_pred(pred, j);
5790 add_irn_n(n, pred_pred);
5794 if (get_Sync_pred(n, k) == pred_pred) break;
5800 ir_graph *irg = get_irn_irg(n);
5801 return new_r_Bad(irg, mode_M);
5804 return get_Sync_pred(n, 0);
5807 /* rehash the sync node */
5812 static ir_node *transform_node_Load(ir_node *n)
5814 /* if our memory predecessor is a load from the same address, then reuse the
5815 * previous result */
5816 ir_node *mem = get_Load_mem(n);
5821 /* don't touch volatile loads */
5822 if (get_Load_volatility(n) == volatility_is_volatile)
5824 mem_pred = get_Proj_pred(mem);
5825 if (is_Load(mem_pred)) {
5826 ir_node *pred_load = mem_pred;
5828 /* conservatively compare the 2 loads. TODO: This could be less strict
5829 * with fixup code in some situations (like smaller/bigger modes) */
5830 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
5832 if (get_Load_mode(pred_load) != get_Load_mode(n))
5834 /* all combinations of aligned/unaligned pred/n should be fine so we do
5835 * not compare the unaligned attribute */
5837 ir_node *block = get_nodes_block(n);
5838 ir_node *jmp = new_r_Jmp(block);
5839 ir_graph *irg = get_irn_irg(n);
5840 ir_node *bad = new_r_Bad(irg, mode_X);
5841 ir_mode *mode = get_Load_mode(n);
5842 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
5843 ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad };
5844 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5847 } else if (is_Store(mem_pred)) {
5848 ir_node *pred_store = mem_pred;
5849 ir_node *value = get_Store_value(pred_store);
5851 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
5853 if (get_irn_mode(value) != get_Load_mode(n))
5855 /* all combinations of aligned/unaligned pred/n should be fine so we do
5856 * not compare the unaligned attribute */
5858 ir_node *block = get_nodes_block(n);
5859 ir_node *jmp = new_r_Jmp(block);
5860 ir_graph *irg = get_irn_irg(n);
5861 ir_node *bad = new_r_Bad(irg, mode_X);
5862 ir_node *res = value;
5863 ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad };
5864 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5873 * optimize a trampoline Call into a direct Call
5875 static ir_node *transform_node_Call(ir_node *call)
5877 ir_node *callee = get_Call_ptr(call);
5878 ir_node *adr, *mem, *res, *bl, **in;
5879 ir_type *ctp, *mtp, *tp;
5883 size_t i, n_res, n_param;
5886 if (! is_Proj(callee))
5888 callee = get_Proj_pred(callee);
5889 if (! is_Builtin(callee))
5891 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5894 mem = get_Call_mem(call);
5896 if (skip_Proj(mem) == callee) {
5897 /* memory is routed to the trampoline, skip */
5898 mem = get_Builtin_mem(callee);
5901 /* build a new call type */
5902 mtp = get_Call_type(call);
5903 tdb = get_type_dbg_info(mtp);
5905 n_res = get_method_n_ress(mtp);
5906 n_param = get_method_n_params(mtp);
5907 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5909 for (i = 0; i < n_res; ++i)
5910 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5912 NEW_ARR_A(ir_node *, in, n_param + 1);
5914 /* FIXME: we don't need a new pointer type in every step */
5915 irg = get_irn_irg(call);
5916 tp = get_irg_frame_type(irg);
5917 tp = new_type_pointer(tp);
5918 set_method_param_type(ctp, 0, tp);
5920 in[0] = get_Builtin_param(callee, 2);
5921 for (i = 0; i < n_param; ++i) {
5922 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5923 in[i + 1] = get_Call_param(call, i);
5925 var = get_method_variadicity(mtp);
5926 set_method_variadicity(ctp, var);
5927 /* When we resolve a trampoline, the function must be called by a this-call */
5928 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5929 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5931 adr = get_Builtin_param(callee, 1);
5933 db = get_irn_dbg_info(call);
5934 bl = get_nodes_block(call);
5936 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5937 if (get_irn_pinned(call) == op_pin_state_floats)
5938 set_irn_pinned(res, op_pin_state_floats);
5940 } /* transform_node_Call */
5943 * Tries several [inplace] [optimizing] transformations and returns an
5944 * equivalent node. The difference to equivalent_node() is that these
5945 * transformations _do_ generate new nodes, and thus the old node must
5946 * not be freed even if the equivalent node isn't the old one.
5948 static ir_node *transform_node(ir_node *n)
5953 * Transform_node is the only "optimizing transformation" that might
5954 * return a node with a different opcode. We iterate HERE until fixpoint
5955 * to get the final result.
5959 if (n->op->ops.transform_node != NULL)
5960 n = n->op->ops.transform_node(n);
5961 } while (oldn != n);
5964 } /* transform_node */
5967 * Sets the default transform node operation for an ir_op_ops.
5969 * @param code the opcode for the default operation
5970 * @param ops the operations initialized
5975 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5979 ops->transform_node = transform_node_##a; \
5981 #define CASE_PROJ(a) \
5983 ops->transform_node_Proj = transform_node_Proj_##a; \
5985 #define CASE_PROJ_EX(a) \
5987 ops->transform_node = transform_node_##a; \
5988 ops->transform_node_Proj = transform_node_Proj_##a; \
6029 } /* firm_set_default_transform_node */
6032 /* **************** Common Subexpression Elimination **************** */
6034 /** The size of the hash table used, should estimate the number of nodes
6036 #define N_IR_NODES 512
6038 /** Compares the attributes of two Const nodes. */
6039 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
6041 return get_Const_tarval(a) != get_Const_tarval(b);
6044 /** Compares the attributes of two Proj nodes. */
6045 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
6047 return a->attr.proj.proj != b->attr.proj.proj;
6050 /** Compares the attributes of two Alloc nodes. */
6051 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
6053 const alloc_attr *pa = &a->attr.alloc;
6054 const alloc_attr *pb = &b->attr.alloc;
6055 return (pa->where != pb->where) || (pa->type != pb->type);
6058 /** Compares the attributes of two Free nodes. */
6059 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
6061 const free_attr *pa = &a->attr.free;
6062 const free_attr *pb = &b->attr.free;
6063 return (pa->where != pb->where) || (pa->type != pb->type);
6066 /** Compares the attributes of two SymConst nodes. */
6067 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
6069 const symconst_attr *pa = &a->attr.symc;
6070 const symconst_attr *pb = &b->attr.symc;
6071 return (pa->kind != pb->kind)
6072 || (pa->sym.type_p != pb->sym.type_p);
6075 /** Compares the attributes of two Call nodes. */
6076 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
6078 const call_attr *pa = &a->attr.call;
6079 const call_attr *pb = &b->attr.call;
6080 return (pa->type != pb->type)
6081 || (pa->tail_call != pb->tail_call);
6084 /** Compares the attributes of two Sel nodes. */
6085 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
6087 const ir_entity *a_ent = get_Sel_entity(a);
6088 const ir_entity *b_ent = get_Sel_entity(b);
6089 return a_ent != b_ent;
6092 /** Compares the attributes of two Phi nodes. */
6093 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
6095 /* we can only enter this function if both nodes have the same number of inputs,
6096 hence it is enough to check if one of them is a Phi0 */
6098 /* check the Phi0 pos attribute */
6099 return a->attr.phi.u.pos != b->attr.phi.u.pos;
6104 /** Compares the attributes of two Conv nodes. */
6105 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
6107 return get_Conv_strict(a) != get_Conv_strict(b);
6110 /** Compares the attributes of two Cast nodes. */
6111 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
6113 return get_Cast_type(a) != get_Cast_type(b);
6116 /** Compares the attributes of two Load nodes. */
6117 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
6119 if (get_Load_volatility(a) == volatility_is_volatile ||
6120 get_Load_volatility(b) == volatility_is_volatile)
6121 /* NEVER do CSE on volatile Loads */
6123 /* do not CSE Loads with different alignment. Be conservative. */
6124 if (get_Load_unaligned(a) != get_Load_unaligned(b))
6127 return get_Load_mode(a) != get_Load_mode(b);
6130 /** Compares the attributes of two Store nodes. */
6131 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
6133 /* do not CSE Stores with different alignment. Be conservative. */
6134 if (get_Store_unaligned(a) != get_Store_unaligned(b))
6137 /* NEVER do CSE on volatile Stores */
6138 return (get_Store_volatility(a) == volatility_is_volatile ||
6139 get_Store_volatility(b) == volatility_is_volatile);
6142 /** Compares two exception attributes */
6143 static int node_cmp_exception(const ir_node *a, const ir_node *b)
6145 const except_attr *ea = &a->attr.except;
6146 const except_attr *eb = &b->attr.except;
6148 return ea->pin_state != eb->pin_state;
6151 #define node_cmp_attr_Bound node_cmp_exception
6153 /** Compares the attributes of two Div nodes. */
6154 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
6156 const div_attr *ma = &a->attr.div;
6157 const div_attr *mb = &b->attr.div;
6158 return ma->exc.pin_state != mb->exc.pin_state ||
6159 ma->resmode != mb->resmode ||
6160 ma->no_remainder != mb->no_remainder;
6163 /** Compares the attributes of two Mod nodes. */
6164 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
6166 const mod_attr *ma = &a->attr.mod;
6167 const mod_attr *mb = &b->attr.mod;
6168 return ma->exc.pin_state != mb->exc.pin_state ||
6169 ma->resmode != mb->resmode;
6172 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
6174 const cmp_attr *ma = &a->attr.cmp;
6175 const cmp_attr *mb = &b->attr.cmp;
6176 return ma->relation != mb->relation;
6179 /** Compares the attributes of two Confirm nodes. */
6180 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
6182 const confirm_attr *ma = &a->attr.confirm;
6183 const confirm_attr *mb = &b->attr.confirm;
6184 return ma->relation != mb->relation;
6187 /** Compares the attributes of two Builtin nodes. */
6188 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
6190 /* no need to compare the type, equal kind means equal type */
6191 return get_Builtin_kind(a) != get_Builtin_kind(b);
6194 /** Compares the attributes of two ASM nodes. */
6195 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6198 const ir_asm_constraint *ca;
6199 const ir_asm_constraint *cb;
6202 if (get_ASM_text(a) != get_ASM_text(b))
6205 /* Should we really check the constraints here? Should be better, but is strange. */
6206 n = get_ASM_n_input_constraints(a);
6207 if (n != get_ASM_n_input_constraints(b))
6210 ca = get_ASM_input_constraints(a);
6211 cb = get_ASM_input_constraints(b);
6212 for (i = 0; i < n; ++i) {
6213 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6214 || ca[i].mode != cb[i].mode)
6218 n = get_ASM_n_output_constraints(a);
6219 if (n != get_ASM_n_output_constraints(b))
6222 ca = get_ASM_output_constraints(a);
6223 cb = get_ASM_output_constraints(b);
6224 for (i = 0; i < n; ++i) {
6225 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6226 || ca[i].mode != cb[i].mode)
6230 n = get_ASM_n_clobbers(a);
6231 if (n != get_ASM_n_clobbers(b))
6234 cla = get_ASM_clobbers(a);
6235 clb = get_ASM_clobbers(b);
6236 for (i = 0; i < n; ++i) {
6237 if (cla[i] != clb[i])
6243 /** Compares the inexistent attributes of two Dummy nodes. */
6244 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6252 * Set the default node attribute compare operation for an ir_op_ops.
6254 * @param code the opcode for the default operation
6255 * @param ops the operations initialized
6260 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6264 ops->node_cmp_attr = node_cmp_attr_##a; \
6296 } /* firm_set_default_node_cmp_attr */
6299 * Compare function for two nodes in the value table. Gets two
6300 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6302 int identities_cmp(const void *elt, const void *key)
6304 ir_node *a = (ir_node *)elt;
6305 ir_node *b = (ir_node *)key;
6308 if (a == b) return 0;
6310 if ((get_irn_op(a) != get_irn_op(b)) ||
6311 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6313 /* compare if a's in and b's in are of equal length */
6314 irn_arity_a = get_irn_arity(a);
6315 if (irn_arity_a != get_irn_arity(b))
6318 /* blocks are never the same */
6322 if (get_irn_pinned(a) == op_pin_state_pinned) {
6323 /* for pinned nodes, the block inputs must be equal */
6324 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6327 ir_node *block_a = get_nodes_block(a);
6328 ir_node *block_b = get_nodes_block(b);
6329 if (! get_opt_global_cse()) {
6330 /* for block-local CSE both nodes must be in the same Block */
6331 if (block_a != block_b)
6334 /* The optimistic approach would be to do nothing here.
6335 * However doing GCSE optimistically produces a lot of partially dead code which appears
6336 * to be worse in practice than the missed opportunities.
6337 * So we use a very conservative variant here and only CSE if 1 value dominates the
6339 if (!block_dominates(block_a, block_b)
6340 && !block_dominates(block_b, block_a))
6345 /* compare a->in[0..ins] with b->in[0..ins] */
6346 for (i = 0; i < irn_arity_a; ++i) {
6347 ir_node *pred_a = get_irn_n(a, i);
6348 ir_node *pred_b = get_irn_n(b, i);
6349 if (pred_a != pred_b) {
6350 /* if both predecessors are CSE neutral they might be different */
6351 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6357 * here, we already now that the nodes are identical except their
6360 if (a->op->ops.node_cmp_attr)
6361 return a->op->ops.node_cmp_attr(a, b);
6364 } /* identities_cmp */
6367 * Calculate a hash value of a node.
6369 * @param node The IR-node
6371 unsigned ir_node_hash(const ir_node *node)
6373 return node->op->ops.hash(node);
6374 } /* ir_node_hash */
6377 void new_identities(ir_graph *irg)
6379 if (irg->value_table != NULL)
6380 del_pset(irg->value_table);
6381 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6382 } /* new_identities */
6384 void del_identities(ir_graph *irg)
6386 if (irg->value_table != NULL)
6387 del_pset(irg->value_table);
6388 } /* del_identities */
6390 /* Normalize a node by putting constants (and operands with larger
6391 * node index) on the right (operator side). */
6392 void ir_normalize_node(ir_node *n)
6394 if (is_op_commutative(get_irn_op(n))) {
6395 ir_node *l = get_binop_left(n);
6396 ir_node *r = get_binop_right(n);
6398 /* For commutative operators perform a OP b == b OP a but keep
6399 * constants on the RIGHT side. This helps greatly in some
6400 * optimizations. Moreover we use the idx number to make the form
6402 if (!operands_are_normalized(l, r)) {
6403 set_binop_left(n, r);
6404 set_binop_right(n, l);
6408 } /* ir_normalize_node */
6411 * Return the canonical node computing the same value as n.
6412 * Looks up the node in a hash table, enters it in the table
6413 * if it isn't there yet.
6415 * @param n the node to look up
6417 * @return a node that computes the same value as n or n if no such
6418 * node could be found
6420 ir_node *identify_remember(ir_node *n)
6422 ir_graph *irg = get_irn_irg(n);
6423 pset *value_table = irg->value_table;
6426 if (value_table == NULL)
6429 ir_normalize_node(n);
6430 /* lookup or insert in hash table with given hash key. */
6431 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6434 /* n is reachable again */
6435 edges_node_revival(nn);
6439 } /* identify_remember */
6442 * During construction we set the op_pin_state_pinned flag in the graph right
6443 * when the optimization is performed. The flag turning on procedure global
6444 * cse could be changed between two allocations. This way we are safe.
6446 * @param n The node to lookup
6448 static inline ir_node *identify_cons(ir_node *n)
6452 n = identify_remember(n);
6453 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6454 ir_graph *irg = get_irn_irg(n);
6455 set_irg_pinned(irg, op_pin_state_floats);
6458 } /* identify_cons */
6460 /* Add a node to the identities value table. */
6461 void add_identities(ir_node *node)
6468 identify_remember(node);
6471 /* Visit each node in the value table of a graph. */
6472 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6475 ir_graph *rem = current_ir_graph;
6477 current_ir_graph = irg;
6478 foreach_pset(irg->value_table, ir_node*, node) {
6481 current_ir_graph = rem;
6482 } /* visit_all_identities */
6485 * These optimizations deallocate nodes from the obstack.
6486 * It can only be called if it is guaranteed that no other nodes
6487 * reference this one, i.e., right after construction of a node.
6489 * @param n The node to optimize
6491 ir_node *optimize_node(ir_node *n)
6494 ir_graph *irg = get_irn_irg(n);
6495 unsigned iro = get_irn_opcode(n);
6498 /* Always optimize Phi nodes: part of the construction. */
6499 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6501 /* constant expression evaluation / constant folding */
6502 if (get_opt_constant_folding()) {
6503 /* neither constants nor Tuple values can be evaluated */
6504 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6505 /* try to evaluate */
6506 tv = computed_value(n);
6507 if (tv != tarval_bad) {
6512 * we MUST copy the node here temporarily, because it's still
6513 * needed for DBG_OPT_CSTEVAL
6515 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6516 oldn = (ir_node*)alloca(node_size);
6518 memcpy(oldn, n, node_size);
6519 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6521 /* ARG, copy the in array, we need it for statistics */
6522 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6524 /* note the inplace edges module */
6525 edges_node_deleted(n);
6527 /* evaluation was successful -- replace the node. */
6528 irg_kill_node(irg, n);
6529 nw = new_r_Const(irg, tv);
6531 DBG_OPT_CSTEVAL(oldn, nw);
6537 /* remove unnecessary nodes */
6538 if (get_opt_algebraic_simplification() ||
6539 (iro == iro_Phi) || /* always optimize these nodes. */
6541 (iro == iro_Proj) ||
6542 (iro == iro_Block) ) /* Flags tested local. */
6543 n = equivalent_node(n);
6545 /* Common Subexpression Elimination.
6547 * Checks whether n is already available.
6548 * The block input is used to distinguish different subexpressions. Right
6549 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6550 * subexpressions within a block.
6553 n = identify_cons(n);
6556 edges_node_deleted(oldn);
6558 /* We found an existing, better node, so we can deallocate the old node. */
6559 irg_kill_node(irg, oldn);
6563 /* Some more constant expression evaluation that does not allow to
6565 iro = get_irn_opcode(n);
6566 if (get_opt_algebraic_simplification() ||
6567 (iro == iro_Cond) ||
6568 (iro == iro_Proj)) /* Flags tested local. */
6569 n = transform_node(n);
6571 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6572 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6574 n = identify_remember(o);
6580 } /* optimize_node */
6584 * These optimizations never deallocate nodes (in place). This can cause dead
6585 * nodes lying on the obstack. Remove these by a dead node elimination,
6586 * i.e., a copying garbage collection.
6588 ir_node *optimize_in_place_2(ir_node *n)
6592 unsigned iro = get_irn_opcode(n);
6594 if (!get_opt_optimize() && !is_Phi(n)) return n;
6596 if (iro == iro_Deleted)
6599 /* constant expression evaluation / constant folding */
6600 if (get_opt_constant_folding()) {
6601 /* neither constants nor Tuple values can be evaluated */
6602 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6603 /* try to evaluate */
6604 tv = computed_value(n);
6605 if (tv != tarval_bad) {
6606 /* evaluation was successful -- replace the node. */
6607 ir_graph *irg = get_irn_irg(n);
6609 n = new_r_Const(irg, tv);
6611 DBG_OPT_CSTEVAL(oldn, n);
6617 /* remove unnecessary nodes */
6618 if (get_opt_constant_folding() ||
6619 (iro == iro_Phi) || /* always optimize these nodes. */
6620 (iro == iro_Id) || /* ... */
6621 (iro == iro_Proj) || /* ... */
6622 (iro == iro_Block) ) /* Flags tested local. */
6623 n = equivalent_node(n);
6625 /** common subexpression elimination **/
6626 /* Checks whether n is already available. */
6627 /* The block input is used to distinguish different subexpressions. Right
6628 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6629 subexpressions within a block. */
6630 if (get_opt_cse()) {
6632 n = identify_remember(o);
6637 /* Some more constant expression evaluation. */
6638 iro = get_irn_opcode(n);
6639 if (get_opt_constant_folding() ||
6640 (iro == iro_Cond) ||
6641 (iro == iro_Proj)) /* Flags tested local. */
6642 n = transform_node(n);
6644 /* Now we can verify the node, as it has no dead inputs any more. */
6647 /* Now we have a legal, useful node. Enter it in hash table for cse.
6648 Blocks should be unique anyways. (Except the successor of start:
6649 is cse with the start block!) */
6650 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6652 n = identify_remember(o);
6658 } /* optimize_in_place_2 */
6661 * Wrapper for external use, set proper status bits after optimization.
6663 ir_node *optimize_in_place(ir_node *n)
6665 ir_graph *irg = get_irn_irg(n);
6666 /* Handle graph state */
6667 assert(get_irg_phase_state(irg) != phase_building);
6669 if (get_opt_global_cse())
6670 set_irg_pinned(irg, op_pin_state_floats);
6672 /* FIXME: Maybe we could also test whether optimizing the node can
6673 change the control graph. */
6674 set_irg_doms_inconsistent(irg);
6675 return optimize_in_place_2(n);
6676 } /* optimize_in_place */
6679 * Calculate a hash value of a Const node.
6681 static unsigned hash_Const(const ir_node *node)
6685 /* special value for const, as they only differ in their tarval. */
6686 h = HASH_PTR(node->attr.con.tarval);
6692 * Calculate a hash value of a SymConst node.
6694 static unsigned hash_SymConst(const ir_node *node)
6698 /* all others are pointers */
6699 h = HASH_PTR(node->attr.symc.sym.type_p);
6702 } /* hash_SymConst */
6705 * Set the default hash operation in an ir_op_ops.
6707 * @param code the opcode for the default operation
6708 * @param ops the operations initialized
6713 static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops)
6717 ops->hash = hash_##a; \
6720 /* hash function already set */
6721 if (ops->hash != NULL)
6728 /* use input/mode default hash if no function was given */
6729 ops->hash = firm_default_hash;
6737 * Sets the default operation for an ir_ops.
6739 ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops)
6741 ops = firm_set_default_hash(code, ops);
6742 ops = firm_set_default_computed_value(code, ops);
6743 ops = firm_set_default_equivalent_node(code, ops);
6744 ops = firm_set_default_transform_node(code, ops);
6745 ops = firm_set_default_node_cmp_attr(code, ops);
6746 ops = firm_set_default_get_type_attr(code, ops);
6747 ops = firm_set_default_get_entity_attr(code, ops);
6750 } /* firm_set_default_operations */