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 /* Check Conv(all_one) & Const = all_one */
1082 ir_tarval *one = get_mode_all_one(convopmode);
1083 ir_tarval *conv = tarval_convert_to(one, mode);
1084 ir_tarval *and = tarval_and(conv, tv);
1086 if (tarval_is_all_one(and)) {
1087 /* Conv(X) & Const = X */
1089 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1096 if (tarval_is_all_one(tv)) {
1098 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1102 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1105 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1110 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1113 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1118 } /* equivalent_node_And */
1121 * Try to remove useless Conv's:
1123 static ir_node *equivalent_node_Conv(ir_node *n)
1126 ir_node *a = get_Conv_op(n);
1128 ir_mode *n_mode = get_irn_mode(n);
1129 ir_mode *a_mode = get_irn_mode(a);
1132 if (n_mode == a_mode) { /* No Conv necessary */
1133 if (get_Conv_strict(n)) {
1136 /* neither Minus nor Confirm change the precision,
1137 so we can "look-through" */
1140 p = get_Minus_op(p);
1141 } else if (is_Confirm(p)) {
1142 p = get_Confirm_value(p);
1148 if (is_Conv(p) && get_Conv_strict(p)) {
1149 /* we known already, that a_mode == n_mode, and neither
1150 Minus change the mode, so the second Conv
1152 assert(get_irn_mode(p) == n_mode);
1154 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1158 ir_node *pred = get_Proj_pred(p);
1159 if (is_Load(pred)) {
1160 /* Loads always return with the exact precision of n_mode */
1161 assert(get_Load_mode(pred) == n_mode);
1163 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1166 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1167 pred = get_Proj_pred(pred);
1168 if (is_Start(pred)) {
1169 /* Arguments always return with the exact precision,
1170 as strictConv's are place before Call -- if the
1171 caller was compiled with the same setting.
1172 Otherwise, the semantics is probably still right. */
1173 assert(get_irn_mode(p) == n_mode);
1175 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1181 /* special case: the immediate predecessor is also a Conv */
1182 if (! get_Conv_strict(a)) {
1183 /* first one is not strict, kick it */
1185 a_mode = get_irn_mode(a);
1189 /* else both are strict conv, second is superfluous */
1191 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1196 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1199 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1200 ir_node *b = get_Conv_op(a);
1201 ir_mode *b_mode = get_irn_mode(b);
1203 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1204 /* both are strict conv */
1205 if (smaller_mode(a_mode, n_mode)) {
1206 /* both are strict, but the first is smaller, so
1207 the second cannot remove more precision, remove the
1209 set_Conv_strict(n, 0);
1212 if (n_mode == b_mode) {
1213 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1214 if (n_mode == mode_b) {
1215 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1216 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1218 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1219 if (values_in_mode(b_mode, a_mode)) {
1220 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1221 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1226 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1227 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1228 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1229 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1231 if (float_mantissa >= int_mantissa) {
1233 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1238 if (smaller_mode(b_mode, a_mode)) {
1239 if (get_Conv_strict(n))
1240 set_Conv_strict(b, 1);
1241 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1242 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1249 } /* equivalent_node_Conv */
1252 * - fold Phi-nodes, iff they have only one predecessor except
1255 static ir_node *equivalent_node_Phi(ir_node *n)
1260 ir_node *first_val = NULL; /* to shutup gcc */
1262 if (!get_opt_optimize() &&
1263 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1266 n_preds = get_Phi_n_preds(n);
1268 /* Phi of dead Region without predecessors. */
1272 /* Find first non-self-referencing input */
1273 for (i = 0; i < n_preds; ++i) {
1274 first_val = get_Phi_pred(n, i);
1275 /* not self pointer */
1276 if (first_val != n) {
1277 /* then found first value. */
1282 /* search for rest of inputs, determine if any of these
1283 are non-self-referencing */
1284 while (++i < n_preds) {
1285 ir_node *scnd_val = get_Phi_pred(n, i);
1286 if (scnd_val != n && scnd_val != first_val) {
1291 if (i >= n_preds && !is_Dummy(first_val)) {
1292 /* Fold, if no multiple distinct non-self-referencing inputs */
1294 DBG_OPT_PHI(oldn, n);
1297 } /* equivalent_node_Phi */
1300 * Optimize Proj(Tuple).
1302 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1304 ir_node *oldn = proj;
1305 ir_node *tuple = get_Proj_pred(proj);
1307 /* Remove the Tuple/Proj combination. */
1308 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1309 DBG_OPT_TUPLE(oldn, tuple, proj);
1312 } /* equivalent_node_Proj_Tuple */
1315 * Optimize a / 1 = a.
1317 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1319 ir_node *oldn = proj;
1320 ir_node *div = get_Proj_pred(proj);
1321 ir_node *b = get_Div_right(div);
1322 ir_tarval *tb = value_of(b);
1324 /* Div is not commutative. */
1325 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1326 switch (get_Proj_proj(proj)) {
1328 proj = get_Div_mem(div);
1329 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1333 proj = get_Div_left(div);
1334 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1338 /* we cannot replace the exception Proj's here, this is done in
1339 transform_node_Proj_Div() */
1344 } /* equivalent_node_Proj_Div */
1347 * Optimize CopyB(mem, x, x) into a Nop.
1349 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1351 ir_node *oldn = proj;
1352 ir_node *copyb = get_Proj_pred(proj);
1353 ir_node *a = get_CopyB_dst(copyb);
1354 ir_node *b = get_CopyB_src(copyb);
1357 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1358 switch (get_Proj_proj(proj)) {
1360 proj = get_CopyB_mem(copyb);
1361 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1366 } /* equivalent_node_Proj_CopyB */
1369 * Optimize Bounds(idx, idx, upper) into idx.
1371 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1373 ir_node *oldn = proj;
1374 ir_node *bound = get_Proj_pred(proj);
1375 ir_node *idx = get_Bound_index(bound);
1376 ir_node *pred = skip_Proj(idx);
1379 if (idx == get_Bound_lower(bound))
1381 else if (is_Bound(pred)) {
1383 * idx was Bounds checked previously, it is still valid if
1384 * lower <= pred_lower && pred_upper <= upper.
1386 ir_node *lower = get_Bound_lower(bound);
1387 ir_node *upper = get_Bound_upper(bound);
1388 if (get_Bound_lower(pred) == lower &&
1389 get_Bound_upper(pred) == upper) {
1391 * One could expect that we simply return the previous
1392 * Bound here. However, this would be wrong, as we could
1393 * add an exception Proj to a new location then.
1394 * So, we must turn in into a tuple.
1400 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1401 switch (get_Proj_proj(proj)) {
1403 DBG_OPT_EXC_REM(proj);
1404 proj = get_Bound_mem(bound);
1408 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1411 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1416 } /* equivalent_node_Proj_Bound */
1419 * Does all optimizations on nodes that must be done on its Projs
1420 * because of creating new nodes.
1422 static ir_node *equivalent_node_Proj(ir_node *proj)
1424 ir_node *n = get_Proj_pred(proj);
1425 if (n->op->ops.equivalent_node_Proj)
1426 return n->op->ops.equivalent_node_Proj(proj);
1428 } /* equivalent_node_Proj */
1433 static ir_node *equivalent_node_Id(ir_node *n)
1441 DBG_OPT_ID(oldn, n);
1443 } /* equivalent_node_Id */
1448 static ir_node *equivalent_node_Mux(ir_node *n)
1450 ir_node *oldn = n, *sel = get_Mux_sel(n);
1452 ir_tarval *ts = value_of(sel);
1454 /* Mux(true, f, t) == t */
1455 if (ts == tarval_b_true) {
1456 n = get_Mux_true(n);
1457 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1460 /* Mux(false, f, t) == f */
1461 if (ts == tarval_b_false) {
1462 n = get_Mux_false(n);
1463 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1466 n_t = get_Mux_true(n);
1467 n_f = get_Mux_false(n);
1469 /* Mux(v, x, T) == x */
1470 if (is_Unknown(n_f)) {
1472 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1475 /* Mux(v, T, x) == x */
1476 if (is_Unknown(n_t)) {
1478 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1482 /* Mux(v, x, x) == x */
1485 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1488 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1489 ir_relation relation = get_Cmp_relation(sel);
1490 ir_node *f = get_Mux_false(n);
1491 ir_node *t = get_Mux_true(n);
1494 * Note further that these optimization work even for floating point
1495 * with NaN's because -NaN == NaN.
1496 * However, if +0 and -0 is handled differently, we cannot use the first one.
1498 ir_node *const cmp_l = get_Cmp_left(sel);
1499 ir_node *const cmp_r = get_Cmp_right(sel);
1502 case ir_relation_equal:
1503 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1504 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1506 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1511 case ir_relation_less_greater:
1512 case ir_relation_unordered_less_greater:
1513 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1514 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1516 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1525 * Note: normalization puts the constant on the right side,
1526 * so we check only one case.
1528 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1529 /* Mux(t CMP 0, X, t) */
1530 if (is_Minus(f) && get_Minus_op(f) == t) {
1531 /* Mux(t CMP 0, -t, t) */
1532 if (relation == ir_relation_equal) {
1533 /* Mux(t == 0, -t, t) ==> -t */
1535 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1536 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1537 /* Mux(t != 0, -t, t) ==> t */
1539 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1549 * Remove Confirm nodes if setting is on.
1550 * Replace Confirms(x, '=', Constlike) by Constlike.
1552 static ir_node *equivalent_node_Confirm(ir_node *n)
1554 ir_node *pred = get_Confirm_value(n);
1555 ir_relation relation = get_Confirm_relation(n);
1557 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1559 * rare case: two identical Confirms one after another,
1560 * replace the second one with the first.
1563 pred = get_Confirm_value(n);
1569 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1570 * perform no actual computation, as, e.g., the Id nodes. It does not create
1571 * new nodes. It is therefore safe to free n if the node returned is not n.
1572 * If a node returns a Tuple we can not just skip it. If the size of the
1573 * in array fits, we transform n into a tuple (e.g., Div).
1575 ir_node *equivalent_node(ir_node *n)
1577 if (n->op->ops.equivalent_node)
1578 return n->op->ops.equivalent_node(n);
1580 } /* equivalent_node */
1583 * Sets the default equivalent node operation for an ir_op_ops.
1585 * @param code the opcode for the default operation
1586 * @param ops the operations initialized
1591 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1595 ops->equivalent_node = equivalent_node_##a; \
1597 #define CASE_PROJ(a) \
1599 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1633 } /* firm_set_default_equivalent_node */
1636 * Returns non-zero if a node is a Phi node
1637 * with all predecessors constant.
1639 static int is_const_Phi(ir_node *n)
1643 if (! is_Phi(n) || get_irn_arity(n) == 0)
1645 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1646 if (! is_Const(get_irn_n(n, i)))
1650 } /* is_const_Phi */
1652 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1653 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1656 * in reality eval_func should be tarval (*eval_func)() but incomplete
1657 * declarations are bad style and generate noisy warnings
1659 typedef void (*eval_func)(void);
1662 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1664 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1666 if (eval == (eval_func) tarval_sub) {
1667 tarval_sub_type func = (tarval_sub_type)eval;
1669 return func(a, b, mode);
1671 tarval_binop_type func = (tarval_binop_type)eval;
1678 * Apply an evaluator on a binop with a constant operators (and one Phi).
1680 * @param phi the Phi node
1681 * @param other the other operand
1682 * @param eval an evaluator function
1683 * @param mode the mode of the result, may be different from the mode of the Phi!
1684 * @param left if non-zero, other is the left operand, else the right
1686 * @return a new Phi node if the conversion was successful, NULL else
1688 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1694 int i, n = get_irn_arity(phi);
1696 NEW_ARR_A(void *, res, n);
1698 for (i = 0; i < n; ++i) {
1699 pred = get_irn_n(phi, i);
1700 tv = get_Const_tarval(pred);
1701 tv = do_eval(eval, other, tv, mode);
1703 if (tv == tarval_bad) {
1704 /* folding failed, bad */
1710 for (i = 0; i < n; ++i) {
1711 pred = get_irn_n(phi, i);
1712 tv = get_Const_tarval(pred);
1713 tv = do_eval(eval, tv, other, mode);
1715 if (tv == tarval_bad) {
1716 /* folding failed, bad */
1722 irg = get_irn_irg(phi);
1723 for (i = 0; i < n; ++i) {
1724 pred = get_irn_n(phi, i);
1725 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1727 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1728 } /* apply_binop_on_phi */
1731 * Apply an evaluator on a binop with two constant Phi.
1733 * @param a the left Phi node
1734 * @param b the right Phi node
1735 * @param eval an evaluator function
1736 * @param mode the mode of the result, may be different from the mode of the Phi!
1738 * @return a new Phi node if the conversion was successful, NULL else
1740 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1742 ir_tarval *tv_l, *tv_r, *tv;
1748 if (get_nodes_block(a) != get_nodes_block(b))
1751 n = get_irn_arity(a);
1752 NEW_ARR_A(void *, res, n);
1754 for (i = 0; i < n; ++i) {
1755 pred = get_irn_n(a, i);
1756 tv_l = get_Const_tarval(pred);
1757 pred = get_irn_n(b, i);
1758 tv_r = get_Const_tarval(pred);
1759 tv = do_eval(eval, tv_l, tv_r, mode);
1761 if (tv == tarval_bad) {
1762 /* folding failed, bad */
1767 irg = get_irn_irg(a);
1768 for (i = 0; i < n; ++i) {
1769 pred = get_irn_n(a, i);
1770 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1772 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1773 } /* apply_binop_on_2_phis */
1776 * Apply an evaluator on a unop with a constant operator (a Phi).
1778 * @param phi the Phi node
1779 * @param eval an evaluator function
1781 * @return a new Phi node if the conversion was successful, NULL else
1783 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1790 int i, n = get_irn_arity(phi);
1792 NEW_ARR_A(void *, res, n);
1793 for (i = 0; i < n; ++i) {
1794 pred = get_irn_n(phi, i);
1795 tv = get_Const_tarval(pred);
1798 if (tv == tarval_bad) {
1799 /* folding failed, bad */
1804 mode = get_irn_mode(phi);
1805 irg = get_irn_irg(phi);
1806 for (i = 0; i < n; ++i) {
1807 pred = get_irn_n(phi, i);
1808 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1810 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1811 } /* apply_unop_on_phi */
1814 * Apply a conversion on a constant operator (a Phi).
1816 * @param phi the Phi node
1818 * @return a new Phi node if the conversion was successful, NULL else
1820 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1826 int i, n = get_irn_arity(phi);
1828 NEW_ARR_A(void *, res, n);
1829 for (i = 0; i < n; ++i) {
1830 pred = get_irn_n(phi, i);
1831 tv = get_Const_tarval(pred);
1832 tv = tarval_convert_to(tv, mode);
1834 if (tv == tarval_bad) {
1835 /* folding failed, bad */
1840 irg = get_irn_irg(phi);
1841 for (i = 0; i < n; ++i) {
1842 pred = get_irn_n(phi, i);
1843 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1845 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1846 } /* apply_conv_on_phi */
1849 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1850 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1851 * If possible, remove the Conv's.
1853 static ir_node *transform_node_AddSub(ir_node *n)
1855 ir_mode *mode = get_irn_mode(n);
1857 if (mode_is_reference(mode)) {
1858 ir_node *left = get_binop_left(n);
1859 ir_node *right = get_binop_right(n);
1860 unsigned ref_bits = get_mode_size_bits(mode);
1862 if (is_Conv(left)) {
1863 ir_mode *lmode = get_irn_mode(left);
1864 unsigned bits = get_mode_size_bits(lmode);
1866 if (ref_bits == bits &&
1867 mode_is_int(lmode) &&
1868 get_mode_arithmetic(lmode) == irma_twos_complement) {
1869 ir_node *pre = get_Conv_op(left);
1870 ir_mode *pre_mode = get_irn_mode(pre);
1872 if (mode_is_int(pre_mode) &&
1873 get_mode_size_bits(pre_mode) == bits &&
1874 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1875 /* ok, this conv just changes to sign, moreover the calculation
1876 * is done with same number of bits as our address mode, so
1877 * we can ignore the conv as address calculation can be viewed
1878 * as either signed or unsigned
1880 set_binop_left(n, pre);
1885 if (is_Conv(right)) {
1886 ir_mode *rmode = get_irn_mode(right);
1887 unsigned bits = get_mode_size_bits(rmode);
1889 if (ref_bits == bits &&
1890 mode_is_int(rmode) &&
1891 get_mode_arithmetic(rmode) == irma_twos_complement) {
1892 ir_node *pre = get_Conv_op(right);
1893 ir_mode *pre_mode = get_irn_mode(pre);
1895 if (mode_is_int(pre_mode) &&
1896 get_mode_size_bits(pre_mode) == bits &&
1897 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1898 /* ok, this conv just changes to sign, moreover the calculation
1899 * is done with same number of bits as our address mode, so
1900 * we can ignore the conv as address calculation can be viewed
1901 * as either signed or unsigned
1903 set_binop_right(n, pre);
1908 /* let address arithmetic use unsigned modes */
1909 if (is_Const(right)) {
1910 ir_mode *rmode = get_irn_mode(right);
1912 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1913 /* convert a AddP(P, *s) into AddP(P, *u) */
1914 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1916 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
1917 set_binop_right(n, pre);
1923 } /* transform_node_AddSub */
1925 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1928 if (is_Const(b) && is_const_Phi(a)) { \
1929 /* check for Op(Phi, Const) */ \
1930 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1932 else if (is_Const(a) && is_const_Phi(b)) { \
1933 /* check for Op(Const, Phi) */ \
1934 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1936 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1937 /* check for Op(Phi, Phi) */ \
1938 c = apply_binop_on_2_phis(a, b, eval, mode); \
1941 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1946 #define HANDLE_UNOP_PHI(eval, a, c) \
1949 if (is_const_Phi(a)) { \
1950 /* check for Op(Phi) */ \
1951 c = apply_unop_on_phi(a, eval); \
1953 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1960 * Do the AddSub optimization, then Transform
1961 * Constant folding on Phi
1962 * Add(a,a) -> Mul(a, 2)
1963 * Add(Mul(a, x), a) -> Mul(a, x+1)
1964 * if the mode is integer or float.
1965 * Transform Add(a,-b) into Sub(a,b).
1966 * Reassociation might fold this further.
1968 static ir_node *transform_node_Add(ir_node *n)
1971 ir_node *a, *b, *c, *oldn = n;
1972 vrp_attr *a_vrp, *b_vrp;
1974 n = transform_node_AddSub(n);
1976 a = get_Add_left(n);
1977 b = get_Add_right(n);
1979 mode = get_irn_mode(n);
1981 if (mode_is_reference(mode)) {
1982 ir_mode *lmode = get_irn_mode(a);
1984 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
1985 /* an Add(a, NULL) is a hidden Conv */
1986 dbg_info *dbg = get_irn_dbg_info(n);
1987 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
1991 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
1993 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1994 if (mode_is_float(mode)) {
1995 ir_graph *irg = get_irn_irg(n);
1996 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2000 if (mode_is_num(mode)) {
2001 ir_graph *irg = get_irn_irg(n);
2002 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2003 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2004 && a == b && mode_is_int(mode)) {
2005 ir_node *block = get_nodes_block(n);
2008 get_irn_dbg_info(n),
2011 new_r_Const_long(irg, mode, 2),
2013 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2018 get_irn_dbg_info(n),
2023 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2028 get_irn_dbg_info(n),
2033 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2036 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2037 /* Here we rely on constants be on the RIGHT side */
2039 ir_node *op = get_Not_op(a);
2041 if (is_Const(b) && is_Const_one(b)) {
2043 ir_node *blk = get_nodes_block(n);
2044 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2045 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2052 a_vrp = vrp_get_info(a);
2053 b_vrp = vrp_get_info(b);
2055 if (a_vrp && b_vrp) {
2056 ir_tarval *vrp_val = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2058 if (tarval_is_null(vrp_val)) {
2059 dbg_info *dbgi = get_irn_dbg_info(n);
2060 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2064 } /* transform_node_Add */
2067 * returns -cnst or NULL if impossible
2069 static ir_node *const_negate(ir_node *cnst)
2071 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2072 dbg_info *dbgi = get_irn_dbg_info(cnst);
2073 ir_graph *irg = get_irn_irg(cnst);
2074 if (tv == tarval_bad) return NULL;
2075 return new_rd_Const(dbgi, irg, tv);
2079 * Do the AddSub optimization, then Transform
2080 * Constant folding on Phi
2081 * Sub(0,a) -> Minus(a)
2082 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2083 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2084 * Sub(Add(a, x), x) -> a
2085 * Sub(x, Add(x, a)) -> -a
2086 * Sub(x, Const) -> Add(x, -Const)
2088 static ir_node *transform_node_Sub(ir_node *n)
2094 n = transform_node_AddSub(n);
2096 a = get_Sub_left(n);
2097 b = get_Sub_right(n);
2099 mode = get_irn_mode(n);
2101 if (mode_is_int(mode)) {
2102 ir_mode *lmode = get_irn_mode(a);
2104 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2105 /* a Sub(a, NULL) is a hidden Conv */
2106 dbg_info *dbg = get_irn_dbg_info(n);
2107 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2108 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2112 if (mode == lmode &&
2113 get_mode_arithmetic(mode) == irma_twos_complement &&
2115 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2117 dbg_info *dbg = get_irn_dbg_info(n);
2118 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2119 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2125 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2127 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2128 if (mode_is_float(mode)) {
2129 ir_graph *irg = get_irn_irg(n);
2130 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2134 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2135 /* a - C -> a + (-C) */
2136 ir_node *cnst = const_negate(b);
2138 ir_node *block = get_nodes_block(n);
2139 dbg_info *dbgi = get_irn_dbg_info(n);
2141 n = new_rd_Add(dbgi, block, a, cnst, mode);
2142 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2147 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2148 dbg_info *dbg = get_irn_dbg_info(n);
2149 ir_node *block = get_nodes_block(n);
2150 ir_node *left = get_Minus_op(a);
2151 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2153 n = new_rd_Minus(dbg, block, add, mode);
2154 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2156 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2157 dbg_info *dbg = get_irn_dbg_info(n);
2158 ir_node *block = get_nodes_block(n);
2159 ir_node *right = get_Minus_op(b);
2161 n = new_rd_Add(dbg, block, a, right, mode);
2162 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2164 } else if (is_Sub(b)) {
2165 /* a - (b - c) -> a + (c - b)
2166 * -> (a - b) + c iff (b - c) is a pointer */
2167 dbg_info *s_dbg = get_irn_dbg_info(b);
2168 ir_node *s_left = get_Sub_left(b);
2169 ir_node *s_right = get_Sub_right(b);
2170 ir_mode *s_mode = get_irn_mode(b);
2171 if (mode_is_reference(s_mode)) {
2172 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2173 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2174 dbg_info *a_dbg = get_irn_dbg_info(n);
2177 s_right = new_r_Conv(lowest_block, s_right, mode);
2178 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2180 ir_node *s_block = get_nodes_block(b);
2181 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2182 dbg_info *a_dbg = get_irn_dbg_info(n);
2183 ir_node *a_block = get_nodes_block(n);
2185 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2187 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2190 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2191 ir_node *m_right = get_Mul_right(b);
2192 if (is_Const(m_right)) {
2193 ir_node *cnst2 = const_negate(m_right);
2194 if (cnst2 != NULL) {
2195 dbg_info *m_dbg = get_irn_dbg_info(b);
2196 ir_node *m_block = get_nodes_block(b);
2197 ir_node *m_left = get_Mul_left(b);
2198 ir_mode *m_mode = get_irn_mode(b);
2199 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2200 dbg_info *a_dbg = get_irn_dbg_info(n);
2201 ir_node *a_block = get_nodes_block(n);
2203 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2204 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2211 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2212 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2214 get_irn_dbg_info(n),
2218 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2222 if (mode_wrap_around(mode)) {
2223 ir_node *left = get_Add_left(a);
2224 ir_node *right = get_Add_right(a);
2226 /* FIXME: Does the Conv's work only for two complement or generally? */
2228 if (mode != get_irn_mode(right)) {
2229 /* This Sub is an effective Cast */
2230 right = new_r_Conv(get_nodes_block(n), right, mode);
2233 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2235 } else if (right == b) {
2236 if (mode != get_irn_mode(left)) {
2237 /* This Sub is an effective Cast */
2238 left = new_r_Conv(get_nodes_block(n), left, mode);
2241 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2247 if (mode_wrap_around(mode)) {
2248 ir_node *left = get_Add_left(b);
2249 ir_node *right = get_Add_right(b);
2251 /* FIXME: Does the Conv's work only for two complement or generally? */
2253 ir_mode *r_mode = get_irn_mode(right);
2255 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2256 if (mode != r_mode) {
2257 /* This Sub is an effective Cast */
2258 n = new_r_Conv(get_nodes_block(n), n, mode);
2260 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2262 } else if (right == a) {
2263 ir_mode *l_mode = get_irn_mode(left);
2265 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2266 if (mode != l_mode) {
2267 /* This Sub is an effective Cast */
2268 n = new_r_Conv(get_nodes_block(n), n, mode);
2270 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2275 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2276 ir_mode *mode = get_irn_mode(a);
2278 if (mode == get_irn_mode(b)) {
2280 ir_node *op_a = get_Conv_op(a);
2281 ir_node *op_b = get_Conv_op(b);
2283 /* check if it's allowed to skip the conv */
2284 ma = get_irn_mode(op_a);
2285 mb = get_irn_mode(op_b);
2287 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2288 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2291 set_Sub_right(n, b);
2297 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2298 if (!is_reassoc_running() && is_Mul(a)) {
2299 ir_node *ma = get_Mul_left(a);
2300 ir_node *mb = get_Mul_right(a);
2303 ir_node *blk = get_nodes_block(n);
2304 ir_graph *irg = get_irn_irg(n);
2306 get_irn_dbg_info(n),
2310 get_irn_dbg_info(n),
2313 new_r_Const(irg, get_mode_one(mode)),
2316 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2318 } else if (mb == b) {
2319 ir_node *blk = get_nodes_block(n);
2320 ir_graph *irg = get_irn_irg(n);
2322 get_irn_dbg_info(n),
2326 get_irn_dbg_info(n),
2329 new_r_Const(irg, get_mode_one(mode)),
2332 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2336 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2337 ir_node *x = get_Sub_left(a);
2338 ir_node *y = get_Sub_right(a);
2339 ir_node *blk = get_nodes_block(n);
2340 ir_mode *m_b = get_irn_mode(b);
2341 ir_mode *m_y = get_irn_mode(y);
2345 /* Determine the right mode for the Add. */
2348 else if (mode_is_reference(m_b))
2350 else if (mode_is_reference(m_y))
2354 * Both modes are different but none is reference,
2355 * happens for instance in SubP(SubP(P, Iu), Is).
2356 * We have two possibilities here: Cast or ignore.
2357 * Currently we ignore this case.
2362 add = new_r_Add(blk, y, b, add_mode);
2364 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2365 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2369 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2370 /* c - ~X = X + (c+1) */
2371 if (is_Const(a) && is_Not(b)) {
2372 ir_tarval *tv = get_Const_tarval(a);
2374 tv = tarval_add(tv, get_mode_one(mode));
2375 if (tv != tarval_bad) {
2376 ir_node *blk = get_nodes_block(n);
2377 ir_graph *irg = get_irn_irg(n);
2378 ir_node *c = new_r_Const(irg, tv);
2379 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2380 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2384 /* x-(x&y) = x & ~y */
2386 ir_node *and_left = get_And_left(b);
2387 ir_node *and_right = get_And_right(b);
2388 if (and_right == a) {
2389 ir_node *tmp = and_left;
2390 and_left = and_right;
2393 if (and_left == a) {
2394 dbg_info *dbgi = get_irn_dbg_info(n);
2395 ir_node *block = get_nodes_block(n);
2396 ir_mode *mode = get_irn_mode(n);
2397 ir_node *notn = new_rd_Not(dbgi, block, and_right, mode);
2398 ir_node *and = new_rd_And(dbgi, block, a, notn, mode);
2404 } /* transform_node_Sub */
2407 * Several transformation done on n*n=2n bits mul.
2408 * These transformations must be done here because new nodes may be produced.
2410 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2413 ir_node *a = get_Mul_left(n);
2414 ir_node *b = get_Mul_right(n);
2415 ir_tarval *ta = value_of(a);
2416 ir_tarval *tb = value_of(b);
2417 ir_mode *smode = get_irn_mode(a);
2419 if (ta == get_mode_one(smode)) {
2420 /* (L)1 * (L)b = (L)b */
2421 ir_node *blk = get_nodes_block(n);
2422 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2423 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2426 else if (ta == get_mode_minus_one(smode)) {
2427 /* (L)-1 * (L)b = (L)b */
2428 ir_node *blk = get_nodes_block(n);
2429 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2430 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2431 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2434 if (tb == get_mode_one(smode)) {
2435 /* (L)a * (L)1 = (L)a */
2436 ir_node *blk = get_irn_n(a, -1);
2437 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2438 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2441 else if (tb == get_mode_minus_one(smode)) {
2442 /* (L)a * (L)-1 = (L)-a */
2443 ir_node *blk = get_nodes_block(n);
2444 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2445 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2446 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2453 * Transform Mul(a,-1) into -a.
2454 * Do constant evaluation of Phi nodes.
2455 * Do architecture dependent optimizations on Mul nodes
2457 static ir_node *transform_node_Mul(ir_node *n)
2459 ir_node *c, *oldn = n;
2460 ir_mode *mode = get_irn_mode(n);
2461 ir_node *a = get_Mul_left(n);
2462 ir_node *b = get_Mul_right(n);
2464 if (is_Bad(a) || is_Bad(b))
2467 if (mode != get_irn_mode(a))
2468 return transform_node_Mul2n(n, mode);
2470 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2472 if (mode_is_signed(mode)) {
2475 if (value_of(a) == get_mode_minus_one(mode))
2477 else if (value_of(b) == get_mode_minus_one(mode))
2480 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2481 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2486 if (is_Const(b)) { /* (-a) * const -> a * -const */
2487 ir_node *cnst = const_negate(b);
2489 dbg_info *dbgi = get_irn_dbg_info(n);
2490 ir_node *block = get_nodes_block(n);
2491 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2492 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2495 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2496 dbg_info *dbgi = get_irn_dbg_info(n);
2497 ir_node *block = get_nodes_block(n);
2498 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2499 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2501 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2502 ir_node *sub_l = get_Sub_left(b);
2503 ir_node *sub_r = get_Sub_right(b);
2504 dbg_info *dbgi = get_irn_dbg_info(n);
2505 ir_node *block = get_nodes_block(n);
2506 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2507 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2508 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2511 } else if (is_Minus(b)) {
2512 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2513 ir_node *sub_l = get_Sub_left(a);
2514 ir_node *sub_r = get_Sub_right(a);
2515 dbg_info *dbgi = get_irn_dbg_info(n);
2516 ir_node *block = get_nodes_block(n);
2517 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2518 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2519 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2522 } else if (is_Shl(a)) {
2523 ir_node *const shl_l = get_Shl_left(a);
2524 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2525 /* (1 << x) * b -> b << x */
2526 dbg_info *const dbgi = get_irn_dbg_info(n);
2527 ir_node *const block = get_nodes_block(n);
2528 ir_node *const shl_r = get_Shl_right(a);
2529 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2530 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2533 } else if (is_Shl(b)) {
2534 ir_node *const shl_l = get_Shl_left(b);
2535 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2536 /* a * (1 << x) -> a << x */
2537 dbg_info *const dbgi = get_irn_dbg_info(n);
2538 ir_node *const block = get_nodes_block(n);
2539 ir_node *const shl_r = get_Shl_right(b);
2540 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2541 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2545 if (get_mode_arithmetic(mode) == irma_ieee754) {
2547 ir_tarval *tv = get_Const_tarval(a);
2548 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2549 && !tarval_is_negative(tv)) {
2550 /* 2.0 * b = b + b */
2551 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2552 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2556 else if (is_Const(b)) {
2557 ir_tarval *tv = get_Const_tarval(b);
2558 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2559 && !tarval_is_negative(tv)) {
2560 /* a * 2.0 = a + a */
2561 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2562 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2567 return arch_dep_replace_mul_with_shifts(n);
2568 } /* transform_node_Mul */
2571 * Transform a Div Node.
2573 static ir_node *transform_node_Div(ir_node *n)
2575 ir_mode *mode = get_Div_resmode(n);
2576 ir_node *a = get_Div_left(n);
2577 ir_node *b = get_Div_right(n);
2579 const ir_node *dummy;
2581 if (mode_is_int(mode)) {
2582 if (is_Const(b) && is_const_Phi(a)) {
2583 /* check for Div(Phi, Const) */
2584 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2586 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2589 } else if (is_Const(a) && is_const_Phi(b)) {
2590 /* check for Div(Const, Phi) */
2591 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2593 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2596 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2597 /* check for Div(Phi, Phi) */
2598 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2600 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2605 if (a == b && value_not_zero(a, &dummy)) {
2606 ir_graph *irg = get_irn_irg(n);
2607 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2608 value = new_r_Const(irg, get_mode_one(mode));
2609 DBG_OPT_CSTEVAL(n, value);
2612 if (mode_is_signed(mode) && is_Const(b)) {
2613 ir_tarval *tv = get_Const_tarval(b);
2615 if (tv == get_mode_minus_one(mode)) {
2617 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2618 DBG_OPT_CSTEVAL(n, value);
2622 /* Try architecture dependent optimization */
2623 value = arch_dep_replace_div_by_const(n);
2626 assert(mode_is_float(mode));
2628 /* Optimize x/c to x*(1/c) */
2629 if (get_mode_arithmetic(mode) == irma_ieee754) {
2630 ir_tarval *tv = value_of(b);
2632 if (tv != tarval_bad) {
2633 int rem = tarval_fp_ops_enabled();
2636 * Floating point constant folding might be disabled here to
2638 * However, as we check for exact result, doing it is safe.
2641 tarval_enable_fp_ops(1);
2642 tv = tarval_div(get_mode_one(mode), tv);
2643 tarval_enable_fp_ops(rem);
2645 /* Do the transformation if the result is either exact or we are
2646 not using strict rules. */
2647 if (tv != tarval_bad &&
2648 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2649 ir_node *block = get_nodes_block(n);
2650 ir_graph *irg = get_irn_irg(block);
2651 ir_node *c = new_r_Const(irg, tv);
2652 dbg_info *dbgi = get_irn_dbg_info(n);
2653 value = new_rd_Mul(dbgi, block, a, c, mode);
2666 /* Turn Div into a tuple (mem, jmp, bad, value) */
2667 mem = get_Div_mem(n);
2668 blk = get_nodes_block(n);
2669 irg = get_irn_irg(blk);
2671 /* skip a potential Pin */
2672 mem = skip_Pin(mem);
2673 turn_into_tuple(n, pn_Div_max+1);
2674 set_Tuple_pred(n, pn_Div_M, mem);
2675 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2676 set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X));
2677 set_Tuple_pred(n, pn_Div_res, value);
2680 } /* transform_node_Div */
2683 * Transform a Mod node.
2685 static ir_node *transform_node_Mod(ir_node *n)
2687 ir_mode *mode = get_Mod_resmode(n);
2688 ir_node *a = get_Mod_left(n);
2689 ir_node *b = get_Mod_right(n);
2694 if (is_Const(b) && is_const_Phi(a)) {
2695 /* check for Div(Phi, Const) */
2696 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2698 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2702 else if (is_Const(a) && is_const_Phi(b)) {
2703 /* check for Div(Const, Phi) */
2704 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2706 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2710 else if (is_const_Phi(a) && is_const_Phi(b)) {
2711 /* check for Div(Phi, Phi) */
2712 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2714 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2721 irg = get_irn_irg(n);
2722 if (tv != tarval_bad) {
2723 value = new_r_Const(irg, tv);
2725 DBG_OPT_CSTEVAL(n, value);
2728 ir_node *a = get_Mod_left(n);
2729 ir_node *b = get_Mod_right(n);
2730 const ir_node *dummy;
2732 if (a == b && value_not_zero(a, &dummy)) {
2733 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2734 value = new_r_Const(irg, get_mode_null(mode));
2735 DBG_OPT_CSTEVAL(n, value);
2738 if (mode_is_signed(mode) && is_Const(b)) {
2739 ir_tarval *tv = get_Const_tarval(b);
2741 if (tv == get_mode_minus_one(mode)) {
2743 value = new_r_Const(irg, get_mode_null(mode));
2744 DBG_OPT_CSTEVAL(n, value);
2748 /* Try architecture dependent optimization */
2749 value = arch_dep_replace_mod_by_const(n);
2758 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2759 mem = get_Mod_mem(n);
2760 blk = get_nodes_block(n);
2761 irg = get_irn_irg(blk);
2763 /* skip a potential Pin */
2764 mem = skip_Pin(mem);
2765 turn_into_tuple(n, pn_Mod_max+1);
2766 set_Tuple_pred(n, pn_Mod_M, mem);
2767 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2768 set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X));
2769 set_Tuple_pred(n, pn_Mod_res, value);
2772 } /* transform_node_Mod */
2775 * Transform a Cond node.
2777 * Replace the Cond by a Jmp if it branches on a constant
2780 static ir_node *transform_node_Cond(ir_node *n)
2783 ir_node *a = get_Cond_selector(n);
2784 ir_tarval *ta = value_of(a);
2785 ir_graph *irg = get_irn_irg(n);
2788 /* we need block info which is not available in floating irgs */
2789 if (get_irg_pinned(irg) == op_pin_state_floats)
2792 if ((ta != tarval_bad) &&
2793 (get_irn_mode(a) == mode_b) &&
2794 (get_opt_unreachable_code())) {
2795 /* It's a boolean Cond, branching on a boolean constant.
2796 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2797 ir_node *blk = get_nodes_block(n);
2798 jmp = new_r_Jmp(blk);
2799 turn_into_tuple(n, pn_Cond_max+1);
2800 if (ta == tarval_b_true) {
2801 set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
2802 set_Tuple_pred(n, pn_Cond_true, jmp);
2804 set_Tuple_pred(n, pn_Cond_false, jmp);
2805 set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X));
2807 /* We might generate an endless loop, so keep it alive. */
2808 add_End_keepalive(get_irg_end(irg), blk);
2811 } /* transform_node_Cond */
2814 * Prototype of a recursive transform function
2815 * for bitwise distributive transformations.
2817 typedef ir_node* (*recursive_transform)(ir_node *n);
2820 * makes use of distributive laws for and, or, eor
2821 * and(a OP c, b OP c) -> and(a, b) OP c
2822 * note, might return a different op than n
2824 static ir_node *transform_bitwise_distributive(ir_node *n,
2825 recursive_transform trans_func)
2828 ir_node *a = get_binop_left(n);
2829 ir_node *b = get_binop_right(n);
2830 ir_op *op = get_irn_op(a);
2831 ir_op *op_root = get_irn_op(n);
2833 if (op != get_irn_op(b))
2836 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2837 if (op == op_Conv) {
2838 ir_node *a_op = get_Conv_op(a);
2839 ir_node *b_op = get_Conv_op(b);
2840 ir_mode *a_mode = get_irn_mode(a_op);
2841 ir_mode *b_mode = get_irn_mode(b_op);
2842 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2843 ir_node *blk = get_nodes_block(n);
2846 set_binop_left(n, a_op);
2847 set_binop_right(n, b_op);
2848 set_irn_mode(n, a_mode);
2850 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2852 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2858 /* nothing to gain here */
2862 if (op == op_Shrs || op == op_Shr || op == op_Shl
2863 || op == op_And || op == op_Or || op == op_Eor) {
2864 ir_node *a_left = get_binop_left(a);
2865 ir_node *a_right = get_binop_right(a);
2866 ir_node *b_left = get_binop_left(b);
2867 ir_node *b_right = get_binop_right(b);
2869 ir_node *op1 = NULL;
2870 ir_node *op2 = NULL;
2872 if (is_op_commutative(op)) {
2873 if (a_left == b_left) {
2877 } else if (a_left == b_right) {
2881 } else if (a_right == b_left) {
2887 if (a_right == b_right) {
2894 /* (a sop c) & (b sop c) => (a & b) sop c */
2895 ir_node *blk = get_nodes_block(n);
2897 ir_node *new_n = exact_copy(n);
2898 set_binop_left(new_n, op1);
2899 set_binop_right(new_n, op2);
2900 new_n = trans_func(new_n);
2902 if (op_root == op_Eor && op == op_Or) {
2903 dbg_info *dbgi = get_irn_dbg_info(n);
2904 ir_mode *mode = get_irn_mode(c);
2906 c = new_rd_Not(dbgi, blk, c, mode);
2907 n = new_rd_And(dbgi, blk, new_n, c, mode);
2910 set_nodes_block(n, blk);
2911 set_binop_left(n, new_n);
2912 set_binop_right(n, c);
2916 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2925 * Create a 0 constant of given mode.
2927 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
2929 ir_tarval *tv = get_mode_null(mode);
2930 ir_node *cnst = new_r_Const(irg, tv);
2935 static bool is_shiftop(const ir_node *n)
2937 return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
2941 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
2943 * - and, or, xor instead of &
2944 * - Shl, Shr, Shrs, rotl instead of >>
2945 * (with a special case for Or/Xor + Shrs)
2947 * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
2949 static ir_node *transform_node_shift_bitop(ir_node *n)
2951 ir_graph *irg = get_irn_irg(n);
2952 ir_node *right = get_binop_right(n);
2953 ir_mode *mode = get_irn_mode(n);
2955 ir_node *bitop_left;
2956 ir_node *bitop_right;
2965 ir_tarval *tv_shift;
2967 if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
2970 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
2972 if (!is_Const(right))
2975 left = get_binop_left(n);
2976 op_left = get_irn_op(left);
2977 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
2980 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
2981 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
2982 /* TODO: test if sign bit is affectes */
2986 bitop_right = get_binop_right(left);
2987 if (!is_Const(bitop_right))
2990 bitop_left = get_binop_left(left);
2992 block = get_nodes_block(n);
2993 dbgi = get_irn_dbg_info(n);
2994 tv1 = get_Const_tarval(bitop_right);
2995 tv2 = get_Const_tarval(right);
2997 assert(get_tarval_mode(tv1) == mode);
3000 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
3001 tv_shift = tarval_shl(tv1, tv2);
3002 } else if (is_Shr(n)) {
3003 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
3004 tv_shift = tarval_shr(tv1, tv2);
3005 } else if (is_Shrs(n)) {
3006 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
3007 tv_shift = tarval_shrs(tv1, tv2);
3010 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
3011 tv_shift = tarval_rotl(tv1, tv2);
3014 assert(get_tarval_mode(tv_shift) == mode);
3015 irg = get_irn_irg(n);
3016 new_const = new_r_Const(irg, tv_shift);
3018 if (op_left == op_And) {
3019 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
3020 } else if (op_left == op_Or) {
3021 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
3023 assert(op_left == op_Eor);
3024 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
3031 * normalisation: (x >> c1) & c2 to (x & (c2<<c1)) >> c1
3033 * - and, or, xor instead of &
3034 * - Shl, Shr, Shrs, rotl instead of >>
3035 * (with a special case for Or/Xor + Shrs)
3037 * This normalisation is usually good for the backend since << C can often be
3038 * matched as address-mode.
3040 static ir_node *transform_node_bitop_shift(ir_node *n)
3042 ir_graph *irg = get_irn_irg(n);
3043 ir_node *left = get_binop_left(n);
3044 ir_node *right = get_binop_right(n);
3045 ir_mode *mode = get_irn_mode(n);
3046 ir_node *shift_left;
3047 ir_node *shift_right;
3049 dbg_info *dbg_bitop;
3050 dbg_info *dbg_shift;
3056 ir_tarval *tv_bitop;
3058 if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
3061 assert(is_And(n) || is_Or(n) || is_Eor(n));
3062 if (!is_Const(right) || !is_shiftop(left))
3065 shift_left = get_binop_left(left);
3066 shift_right = get_binop_right(left);
3067 if (!is_Const(shift_right))
3070 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
3071 if (is_Shrs(left)) {
3072 /* TODO this could be improved */
3076 irg = get_irn_irg(n);
3077 block = get_nodes_block(n);
3078 dbg_bitop = get_irn_dbg_info(n);
3079 dbg_shift = get_irn_dbg_info(left);
3080 tv1 = get_Const_tarval(shift_right);
3081 tv2 = get_Const_tarval(right);
3082 assert(get_tarval_mode(tv2) == mode);
3085 tv_bitop = tarval_shr(tv2, tv1);
3086 } else if (is_Shr(left)) {
3087 if (is_Or(n) || is_Eor(n)) {
3089 * TODO this can be improved by checking whether
3090 * the left shift produces an overflow
3094 tv_bitop = tarval_shl(tv2, tv1);
3096 assert(is_Rotl(left));
3097 tv_bitop = tarval_rotl(tv2, tarval_neg(tv1));
3099 new_const = new_r_Const(irg, tv_bitop);
3102 new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode);
3103 } else if (is_Or(n)) {
3104 new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode);
3107 new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode);
3111 new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode);
3112 } else if (is_Shr(left)) {
3113 new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode);
3115 assert(is_Rotl(left));
3116 new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode);
3125 static ir_node *transform_node_And(ir_node *n)
3127 ir_node *c, *oldn = n;
3128 ir_node *a = get_And_left(n);
3129 ir_node *b = get_And_right(n);
3131 vrp_attr *a_vrp, *b_vrp;
3133 if (is_Cmp(a) && is_Cmp(b)) {
3134 ir_node *a_left = get_Cmp_left(a);
3135 ir_node *a_right = get_Cmp_right(a);
3136 ir_node *b_left = get_Cmp_left(b);
3137 ir_node *b_right = get_Cmp_right(b);
3138 ir_relation a_relation = get_Cmp_relation(a);
3139 ir_relation b_relation = get_Cmp_relation(b);
3140 /* we can combine the relations of two compares with the same
3142 if (a_left == b_left && b_left == b_right) {
3143 dbg_info *dbgi = get_irn_dbg_info(n);
3144 ir_node *block = get_nodes_block(n);
3145 ir_relation new_relation = a_relation & b_relation;
3146 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3148 /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */
3149 if (a_relation == b_relation && a_relation == ir_relation_equal
3150 && !mode_is_float(get_irn_mode(a_left))
3151 && !mode_is_float(get_irn_mode(b_left))) {
3152 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
3153 dbg_info *dbgi = get_irn_dbg_info(n);
3154 ir_node *block = get_nodes_block(n);
3155 ir_mode *a_mode = get_irn_mode(a_left);
3156 ir_mode *b_mode = get_irn_mode(b_left);
3157 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3158 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3159 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
3160 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
3161 ir_graph *irg = get_irn_irg(n);
3162 ir_node *zero = create_zero_const(irg, b_mode);
3163 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3165 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
3166 dbg_info *dbgi = get_irn_dbg_info(n);
3167 ir_node *block = get_nodes_block(n);
3168 ir_mode *a_mode = get_irn_mode(a_left);
3169 ir_mode *b_mode = get_irn_mode(b_left);
3170 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3171 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3172 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
3173 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
3174 ir_graph *irg = get_irn_irg(n);
3175 ir_node *zero = create_zero_const(irg, a_mode);
3176 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3181 mode = get_irn_mode(n);
3182 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3186 ir_node *op = get_Not_op(b);
3188 ir_node *ba = get_And_left(op);
3189 ir_node *bb = get_And_right(op);
3191 /* it's enough to test the following cases due to normalization! */
3192 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
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, ba, bb, mode);
3197 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3205 ir_node *op = get_Not_op(a);
3207 ir_node *aa = get_And_left(op);
3208 ir_node *ab = get_And_right(op);
3210 /* it's enough to test the following cases due to normalization! */
3211 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3212 /* (a|b) & ~(a&b) = a^b */
3213 ir_node *block = get_nodes_block(n);
3215 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3216 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3223 ir_node *al = get_Eor_left(a);
3224 ir_node *ar = get_Eor_right(a);
3227 /* (b ^ a) & b -> ~a & b */
3228 dbg_info *dbg = get_irn_dbg_info(n);
3229 ir_node *block = get_nodes_block(n);
3231 ar = new_rd_Not(dbg, block, ar, mode);
3232 n = new_rd_And(dbg, block, ar, b, mode);
3233 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3237 /* (a ^ b) & b -> ~a & b */
3238 dbg_info *dbg = get_irn_dbg_info(n);
3239 ir_node *block = get_nodes_block(n);
3241 al = new_rd_Not(dbg, block, al, mode);
3242 n = new_rd_And(dbg, block, al, b, mode);
3243 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3248 ir_node *bl = get_Eor_left(b);
3249 ir_node *br = get_Eor_right(b);
3252 /* a & (a ^ b) -> a & ~b */
3253 dbg_info *dbg = get_irn_dbg_info(n);
3254 ir_node *block = get_nodes_block(n);
3256 br = new_rd_Not(dbg, block, br, mode);
3257 n = new_rd_And(dbg, block, br, a, mode);
3258 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3262 /* a & (b ^ a) -> a & ~b */
3263 dbg_info *dbg = get_irn_dbg_info(n);
3264 ir_node *block = get_nodes_block(n);
3266 bl = new_rd_Not(dbg, block, bl, mode);
3267 n = new_rd_And(dbg, block, bl, a, mode);
3268 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3272 if (is_Not(a) && is_Not(b)) {
3273 /* ~a & ~b = ~(a|b) */
3274 ir_node *block = get_nodes_block(n);
3275 ir_mode *mode = get_irn_mode(n);
3279 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3280 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3281 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3285 b_vrp = vrp_get_info(b);
3286 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3287 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3293 a_vrp = vrp_get_info(a);
3294 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3295 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3299 n = transform_bitwise_distributive(n, transform_node_And);
3301 n = transform_node_bitop_shift(n);
3304 } /* transform_node_And */
3306 /* the order of the values is important! */
3307 typedef enum const_class {
3313 static const_class classify_const(const ir_node* n)
3315 if (is_Const(n)) return const_const;
3316 if (is_irn_constlike(n)) return const_like;
3321 * Determines whether r is more constlike or has a larger index (in that order)
3324 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3326 const const_class l_order = classify_const(l);
3327 const const_class r_order = classify_const(r);
3329 l_order > r_order ||
3330 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3336 static ir_node *transform_node_Eor(ir_node *n)
3338 ir_node *c, *oldn = n;
3339 ir_node *a = get_Eor_left(n);
3340 ir_node *b = get_Eor_right(n);
3341 ir_mode *mode = get_irn_mode(n);
3343 /* we can combine the relations of two compares with the same operands */
3344 if (is_Cmp(a) && is_Cmp(b)) {
3345 ir_node *a_left = get_Cmp_left(a);
3346 ir_node *a_right = get_Cmp_left(a);
3347 ir_node *b_left = get_Cmp_left(b);
3348 ir_node *b_right = get_Cmp_right(b);
3349 if (a_left == b_left && b_left == b_right) {
3350 dbg_info *dbgi = get_irn_dbg_info(n);
3351 ir_node *block = get_nodes_block(n);
3352 ir_relation a_relation = get_Cmp_relation(a);
3353 ir_relation b_relation = get_Cmp_relation(b);
3354 ir_relation new_relation = a_relation ^ b_relation;
3355 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3359 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3361 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3362 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3363 dbg_info *dbg = get_irn_dbg_info(n);
3364 ir_node *block = get_nodes_block(n);
3365 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3366 ir_node *new_left = get_Not_op(a);
3367 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3368 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3370 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3371 dbg_info *dbg = get_irn_dbg_info(n);
3372 ir_node *block = get_nodes_block(n);
3373 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3374 ir_node *new_right = get_Not_op(b);
3375 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3376 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3380 /* x ^ 1...1 -> ~1 */
3381 if (is_Const(b) && is_Const_all_one(b)) {
3382 n = new_r_Not(get_nodes_block(n), a, mode);
3383 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3387 n = transform_bitwise_distributive(n, transform_node_Eor);
3389 n = transform_node_bitop_shift(n);
3392 } /* transform_node_Eor */
3397 static ir_node *transform_node_Not(ir_node *n)
3399 ir_node *c, *oldn = n;
3400 ir_node *a = get_Not_op(n);
3401 ir_mode *mode = get_irn_mode(n);
3403 HANDLE_UNOP_PHI(tarval_not,a,c);
3405 /* check for a boolean Not */
3407 dbg_info *dbgi = get_irn_dbg_info(a);
3408 ir_node *block = get_nodes_block(a);
3409 ir_relation relation = get_Cmp_relation(a);
3410 relation = get_negated_relation(relation);
3411 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3412 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3416 /* normalize ~(a ^ b) => a ^ ~b */
3418 dbg_info *dbg = get_irn_dbg_info(n);
3419 ir_node *block = get_nodes_block(n);
3420 ir_node *eor_right = get_Eor_right(a);
3421 ir_node *eor_left = get_Eor_left(a);
3422 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3423 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3427 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3428 if (is_Minus(a)) { /* ~-x -> x + -1 */
3429 dbg_info *dbg = get_irn_dbg_info(n);
3430 ir_graph *irg = get_irn_irg(n);
3431 ir_node *block = get_nodes_block(n);
3432 ir_node *add_l = get_Minus_op(a);
3433 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3434 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3435 } else if (is_Add(a)) {
3436 ir_node *add_r = get_Add_right(a);
3437 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3438 /* ~(x + -1) = -x */
3439 ir_node *op = get_Add_left(a);
3440 ir_node *blk = get_nodes_block(n);
3441 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3450 * Transform a Minus.
3454 * -(a >>u (size-1)) = a >>s (size-1)
3455 * -(a >>s (size-1)) = a >>u (size-1)
3456 * -(a * const) -> a * -const
3458 static ir_node *transform_node_Minus(ir_node *n)
3460 ir_node *c, *oldn = n;
3461 ir_node *a = get_Minus_op(n);
3464 HANDLE_UNOP_PHI(tarval_neg,a,c);
3466 mode = get_irn_mode(a);
3467 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3468 /* the following rules are only to twos-complement */
3471 ir_node *op = get_Not_op(a);
3472 ir_tarval *tv = get_mode_one(mode);
3473 ir_node *blk = get_nodes_block(n);
3474 ir_graph *irg = get_irn_irg(blk);
3475 ir_node *c = new_r_Const(irg, tv);
3476 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3477 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3481 ir_node *c = get_Shr_right(a);
3484 ir_tarval *tv = get_Const_tarval(c);
3486 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3487 /* -(a >>u (size-1)) = a >>s (size-1) */
3488 ir_node *v = get_Shr_left(a);
3490 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3491 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3497 ir_node *c = get_Shrs_right(a);
3500 ir_tarval *tv = get_Const_tarval(c);
3502 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3503 /* -(a >>s (size-1)) = a >>u (size-1) */
3504 ir_node *v = get_Shrs_left(a);
3506 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3507 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3514 /* - (a-b) = b - a */
3515 ir_node *la = get_Sub_left(a);
3516 ir_node *ra = get_Sub_right(a);
3517 ir_node *blk = get_nodes_block(n);
3519 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3520 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3524 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3525 ir_node *mul_l = get_Mul_left(a);
3526 ir_node *mul_r = get_Mul_right(a);
3527 ir_tarval *tv = value_of(mul_r);
3528 if (tv != tarval_bad) {
3529 tv = tarval_neg(tv);
3530 if (tv != tarval_bad) {
3531 ir_graph *irg = get_irn_irg(n);
3532 ir_node *cnst = new_r_Const(irg, tv);
3533 dbg_info *dbg = get_irn_dbg_info(a);
3534 ir_node *block = get_nodes_block(a);
3535 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3536 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3543 } /* transform_node_Minus */
3546 * Transform a Proj(Load) with a non-null address.
3548 static ir_node *transform_node_Proj_Load(ir_node *proj)
3550 if (get_opt_ldst_only_null_ptr_exceptions()) {
3551 if (get_irn_mode(proj) == mode_X) {
3552 ir_node *load = get_Proj_pred(proj);
3554 /* get the Load address */
3555 const ir_node *addr = get_Load_ptr(load);
3556 const ir_node *confirm;
3558 if (value_not_null(addr, &confirm)) {
3559 if (confirm == NULL) {
3560 /* this node may float if it did not depend on a Confirm */
3561 set_irn_pinned(load, op_pin_state_floats);
3563 if (get_Proj_proj(proj) == pn_Load_X_except) {
3564 ir_graph *irg = get_irn_irg(proj);
3565 DBG_OPT_EXC_REM(proj);
3566 return new_r_Bad(irg, mode_X);
3568 ir_node *blk = get_nodes_block(load);
3569 return new_r_Jmp(blk);
3575 } /* transform_node_Proj_Load */
3578 * Transform a Proj(Store) with a non-null address.
3580 static ir_node *transform_node_Proj_Store(ir_node *proj)
3582 if (get_opt_ldst_only_null_ptr_exceptions()) {
3583 if (get_irn_mode(proj) == mode_X) {
3584 ir_node *store = get_Proj_pred(proj);
3586 /* get the load/store address */
3587 const ir_node *addr = get_Store_ptr(store);
3588 const ir_node *confirm;
3590 if (value_not_null(addr, &confirm)) {
3591 if (confirm == NULL) {
3592 /* this node may float if it did not depend on a Confirm */
3593 set_irn_pinned(store, op_pin_state_floats);
3595 if (get_Proj_proj(proj) == pn_Store_X_except) {
3596 ir_graph *irg = get_irn_irg(proj);
3597 DBG_OPT_EXC_REM(proj);
3598 return new_r_Bad(irg, mode_X);
3600 ir_node *blk = get_nodes_block(store);
3601 return new_r_Jmp(blk);
3607 } /* transform_node_Proj_Store */
3610 * Transform a Proj(Div) with a non-zero value.
3611 * Removes the exceptions and routes the memory to the NoMem node.
3613 static ir_node *transform_node_Proj_Div(ir_node *proj)
3615 ir_node *div = get_Proj_pred(proj);
3616 ir_node *b = get_Div_right(div);
3617 ir_node *res, *new_mem;
3618 const ir_node *confirm;
3621 if (value_not_zero(b, &confirm)) {
3622 /* div(x, y) && y != 0 */
3623 if (confirm == NULL) {
3624 /* we are sure we have a Const != 0 */
3625 new_mem = get_Div_mem(div);
3626 new_mem = skip_Pin(new_mem);
3627 set_Div_mem(div, new_mem);
3628 set_irn_pinned(div, op_pin_state_floats);
3631 proj_nr = get_Proj_proj(proj);
3633 case pn_Div_X_regular:
3634 return new_r_Jmp(get_nodes_block(div));
3636 case pn_Div_X_except: {
3637 ir_graph *irg = get_irn_irg(proj);
3638 /* we found an exception handler, remove it */
3639 DBG_OPT_EXC_REM(proj);
3640 return new_r_Bad(irg, mode_X);
3644 ir_graph *irg = get_irn_irg(proj);
3645 res = get_Div_mem(div);
3646 new_mem = get_irg_no_mem(irg);
3649 /* This node can only float up to the Confirm block */
3650 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3652 set_irn_pinned(div, op_pin_state_floats);
3653 /* this is a Div without exception, we can remove the memory edge */
3654 set_Div_mem(div, new_mem);
3660 } /* transform_node_Proj_Div */
3663 * Transform a Proj(Mod) with a non-zero value.
3664 * Removes the exceptions and routes the memory to the NoMem node.
3666 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3668 ir_node *mod = get_Proj_pred(proj);
3669 ir_node *b = get_Mod_right(mod);
3670 ir_node *res, *new_mem;
3671 const ir_node *confirm;
3674 if (value_not_zero(b, &confirm)) {
3675 /* mod(x, y) && y != 0 */
3676 proj_nr = get_Proj_proj(proj);
3678 if (confirm == NULL) {
3679 /* we are sure we have a Const != 0 */
3680 new_mem = get_Mod_mem(mod);
3681 new_mem = skip_Pin(new_mem);
3682 set_Mod_mem(mod, new_mem);
3683 set_irn_pinned(mod, op_pin_state_floats);
3688 case pn_Mod_X_regular:
3689 return new_r_Jmp(get_irn_n(mod, -1));
3691 case pn_Mod_X_except: {
3692 ir_graph *irg = get_irn_irg(proj);
3693 /* we found an exception handler, remove it */
3694 DBG_OPT_EXC_REM(proj);
3695 return new_r_Bad(irg, mode_X);
3699 ir_graph *irg = get_irn_irg(proj);
3700 res = get_Mod_mem(mod);
3701 new_mem = get_irg_no_mem(irg);
3704 /* This node can only float up to the Confirm block */
3705 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3707 /* this is a Mod without exception, we can remove the memory edge */
3708 set_Mod_mem(mod, new_mem);
3712 if (get_Mod_left(mod) == b) {
3713 /* a % a = 0 if a != 0 */
3714 ir_graph *irg = get_irn_irg(proj);
3715 ir_mode *mode = get_irn_mode(proj);
3716 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3718 DBG_OPT_CSTEVAL(mod, res);
3724 } /* transform_node_Proj_Mod */
3727 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3729 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3731 ir_node *n = get_Proj_pred(proj);
3732 ir_node *b = get_Cond_selector(n);
3734 if (!get_opt_unreachable_code())
3737 if (mode_is_int(get_irn_mode(b))) {
3738 ir_tarval *tb = value_of(b);
3740 if (tb != tarval_bad) {
3741 /* we have a constant switch */
3742 long num = get_Proj_proj(proj);
3744 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3745 if (get_tarval_long(tb) == num) {
3746 /* Do NOT create a jump here, or we will have 2 control flow ops
3747 * in a block. This case is optimized away in optimize_cf(). */
3750 ir_graph *irg = get_irn_irg(proj);
3751 /* this case will NEVER be taken, kill it */
3752 return new_r_Bad(irg, mode_X);
3756 long num = get_Proj_proj(proj);
3757 vrp_attr *b_vrp = vrp_get_info(b);
3758 if (num != get_Cond_default_proj(n) && b_vrp) {
3759 /* Try handling with vrp data. We only remove dead parts. */
3760 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3762 if (b_vrp->range_type == VRP_RANGE) {
3763 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3764 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3766 if ((cmp_result & ir_relation_greater) == cmp_result
3767 && (cmp_result2 & ir_relation_less) == cmp_result2) {
3768 ir_graph *irg = get_irn_irg(proj);
3769 return new_r_Bad(irg, mode_X);
3771 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3772 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3773 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3775 if ((cmp_result & ir_relation_less_equal) == cmp_result
3776 && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) {
3777 ir_graph *irg = get_irn_irg(proj);
3778 return new_r_Bad(irg, mode_X);
3783 tarval_and( b_vrp->bits_set, tp),
3785 ) == ir_relation_equal)) {
3786 ir_graph *irg = get_irn_irg(proj);
3787 return new_r_Bad(irg, mode_X);
3793 tarval_not(b_vrp->bits_not_set)),
3794 tarval_not(b_vrp->bits_not_set))
3795 == ir_relation_equal)) {
3796 ir_graph *irg = get_irn_irg(proj);
3797 return new_r_Bad(irg, mode_X);
3806 * return true if the operation returns a value with exactly 1 bit set
3808 static bool is_single_bit(const ir_node *node)
3810 /* a first implementation, could be extended with vrp and others... */
3812 ir_node *shl_l = get_Shl_left(node);
3813 ir_mode *mode = get_irn_mode(node);
3814 int modulo = get_mode_modulo_shift(mode);
3815 /* this works if we shift a 1 and we have modulo shift */
3816 if (is_Const(shl_l) && is_Const_one(shl_l)
3817 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3820 } else if (is_Const(node)) {
3821 ir_tarval *tv = get_Const_tarval(node);
3822 return tarval_is_single_bit(tv);
3828 * Normalizes and optimizes Cmp nodes.
3830 static ir_node *transform_node_Cmp(ir_node *n)
3832 ir_node *left = get_Cmp_left(n);
3833 ir_node *right = get_Cmp_right(n);
3834 ir_mode *mode = get_irn_mode(left);
3835 ir_tarval *tv = NULL;
3836 bool changed = false;
3837 bool changedc = false;
3838 ir_relation relation = get_Cmp_relation(n);
3839 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3841 /* mask out impossible relations */
3842 ir_relation new_relation = relation & possible;
3843 if (new_relation != relation) {
3844 relation = new_relation;
3848 /* Remove unnecessary conversions */
3849 if (is_Conv(left) && is_Conv(right)) {
3850 ir_node *op_left = get_Conv_op(left);
3851 ir_node *op_right = get_Conv_op(right);
3852 ir_mode *mode_left = get_irn_mode(op_left);
3853 ir_mode *mode_right = get_irn_mode(op_right);
3855 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3856 && mode_left != mode_b && mode_right != mode_b) {
3857 ir_node *block = get_nodes_block(n);
3859 if (mode_left == mode_right) {
3863 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3864 } else if (smaller_mode(mode_left, mode_right)) {
3865 left = new_r_Conv(block, op_left, mode_right);
3868 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3869 } else if (smaller_mode(mode_right, mode_left)) {
3871 right = new_r_Conv(block, op_right, mode_left);
3873 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3875 mode = get_irn_mode(left);
3878 if (is_Conv(left) && is_Const(right)) {
3879 ir_node *op_left = get_Conv_op(left);
3880 ir_mode *mode_left = get_irn_mode(op_left);
3881 if (smaller_mode(mode_left, mode) && mode_left != mode_b) {
3882 ir_tarval *tv = get_Const_tarval(right);
3883 tarval_int_overflow_mode_t last_mode
3884 = tarval_get_integer_overflow_mode();
3886 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
3887 new_tv = tarval_convert_to(tv, mode_left);
3888 tarval_set_integer_overflow_mode(last_mode);
3889 if (new_tv != tarval_bad) {
3890 ir_graph *irg = get_irn_irg(n);
3892 right = new_r_Const(irg, new_tv);
3893 mode = get_irn_mode(left);
3895 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3901 * Optimize -a CMP -b into b CMP a.
3902 * This works only for modes where unary Minus cannot Overflow.
3903 * Note that two-complement integers can Overflow so it will NOT work.
3905 if (!mode_overflow_on_unary_Minus(mode) &&
3906 is_Minus(left) && is_Minus(right)) {
3907 left = get_Minus_op(left);
3908 right = get_Minus_op(right);
3909 relation = get_inversed_relation(relation);
3911 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3914 /* remove operation on both sides if possible */
3915 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3917 * The following operations are NOT safe for floating point operations, for instance
3918 * 1.0 + inf == 2.0 + inf, =/=> x == y
3920 if (mode_is_int(mode)) {
3921 unsigned lop = get_irn_opcode(left);
3923 if (lop == get_irn_opcode(right)) {
3924 ir_node *ll, *lr, *rl, *rr;
3926 /* same operation on both sides, try to remove */
3930 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3931 left = get_unop_op(left);
3932 right = get_unop_op(right);
3934 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3937 ll = get_Add_left(left);
3938 lr = get_Add_right(left);
3939 rl = get_Add_left(right);
3940 rr = get_Add_right(right);
3943 /* X + a CMP X + b ==> a CMP b */
3947 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3948 } else if (ll == rr) {
3949 /* X + a CMP b + X ==> a CMP b */
3953 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3954 } else if (lr == rl) {
3955 /* a + X CMP X + b ==> a CMP b */
3959 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3960 } else if (lr == rr) {
3961 /* a + X CMP b + X ==> a CMP b */
3965 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3969 ll = get_Sub_left(left);
3970 lr = get_Sub_right(left);
3971 rl = get_Sub_left(right);
3972 rr = get_Sub_right(right);
3975 /* X - a CMP X - b ==> a CMP b */
3979 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3980 } else if (lr == rr) {
3981 /* a - X CMP b - X ==> a CMP b */
3985 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3989 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3990 /* a ROTL X CMP b ROTL X ==> a CMP b */
3991 left = get_Rotl_left(left);
3992 right = get_Rotl_left(right);
3994 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4002 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4003 if (is_Add(left) || is_Sub(left)) {
4004 ir_node *ll = get_binop_left(left);
4005 ir_node *lr = get_binop_right(left);
4007 if (lr == right && is_Add(left)) {
4013 ir_graph *irg = get_irn_irg(n);
4015 right = create_zero_const(irg, mode);
4017 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4020 if (is_Add(right) || is_Sub(right)) {
4021 ir_node *rl = get_binop_left(right);
4022 ir_node *rr = get_binop_right(right);
4024 if (rr == left && is_Add(right)) {
4030 ir_graph *irg = get_irn_irg(n);
4032 right = create_zero_const(irg, mode);
4034 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4038 if (is_And(left) && is_Const(right)) {
4039 ir_node *ll = get_binop_left(left);
4040 ir_node *lr = get_binop_right(left);
4041 if (is_Shr(ll) && is_Const(lr)) {
4042 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4043 ir_node *block = get_nodes_block(n);
4044 ir_mode *mode = get_irn_mode(left);
4046 ir_node *llr = get_Shr_right(ll);
4047 if (is_Const(llr)) {
4048 dbg_info *dbg = get_irn_dbg_info(left);
4049 ir_graph *irg = get_irn_irg(left);
4051 ir_tarval *c1 = get_Const_tarval(llr);
4052 ir_tarval *c2 = get_Const_tarval(lr);
4053 ir_tarval *c3 = get_Const_tarval(right);
4054 ir_tarval *mask = tarval_shl(c2, c1);
4055 ir_tarval *value = tarval_shl(c3, c1);
4057 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4058 right = new_r_Const(irg, value);
4063 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4065 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4066 right = get_Eor_right(left);
4067 left = get_Eor_left(left);
4070 } /* mode_is_int(...) */
4073 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4074 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
4075 if (mode_is_int(mode) && is_And(left)
4076 && (relation == ir_relation_equal
4077 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4078 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
4079 ir_node *and0 = get_And_left(left);
4080 ir_node *and1 = get_And_right(left);
4081 if (and1 == right) {
4082 ir_node *tmp = and0;
4086 if (and0 == right && is_single_bit(and0)) {
4087 ir_graph *irg = get_irn_irg(n);
4089 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
4090 right = create_zero_const(irg, mode);
4095 /* replace mode_b compares with ands/ors */
4096 if (mode == mode_b) {
4097 ir_node *block = get_nodes_block(n);
4101 case ir_relation_less_equal:
4102 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4104 case ir_relation_less:
4105 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4107 case ir_relation_greater_equal:
4108 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4110 case ir_relation_greater:
4111 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4113 case ir_relation_less_greater:
4114 bres = new_r_Eor(block, left, right, mode_b);
4116 case ir_relation_equal:
4117 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4120 #ifdef DEBUG_libfirm
4121 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4126 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4132 * First step: normalize the compare op
4133 * by placing the constant on the right side
4134 * or moving the lower address node to the left.
4136 if (!operands_are_normalized(left, right)) {
4141 relation = get_inversed_relation(relation);
4146 * Second step: Try to reduce the magnitude
4147 * of a constant. This may help to generate better code
4148 * later and may help to normalize more compares.
4149 * Of course this is only possible for integer values.
4151 tv = value_of(right);
4152 if (tv != tarval_bad) {
4153 ir_mode *mode = get_irn_mode(right);
4155 /* TODO extend to arbitrary constants */
4156 if (is_Conv(left) && tarval_is_null(tv)) {
4157 ir_node *op = get_Conv_op(left);
4158 ir_mode *op_mode = get_irn_mode(op);
4161 * UpConv(x) REL 0 ==> x REL 0
4162 * Don't do this for float values as it's unclear whether it is a
4163 * win. (on the other side it makes detection/creation of fabs hard)
4165 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4166 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4167 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4168 !mode_is_float(mode)) {
4169 tv = get_mode_null(op_mode);
4173 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4177 if (tv != tarval_bad) {
4178 /* the following optimization is possible on modes without Overflow
4179 * on Unary Minus or on == and !=:
4180 * -a CMP c ==> a swap(CMP) -c
4182 * Beware: for two-complement Overflow may occur, so only == and != can
4183 * be optimized, see this:
4184 * -MININT < 0 =/=> MININT > 0 !!!
4186 if (is_Minus(left) &&
4187 (!mode_overflow_on_unary_Minus(mode) ||
4188 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4189 tv = tarval_neg(tv);
4191 if (tv != tarval_bad) {
4192 left = get_Minus_op(left);
4193 relation = get_inversed_relation(relation);
4195 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4197 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4198 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4199 tv = tarval_not(tv);
4201 if (tv != tarval_bad) {
4202 left = get_Not_op(left);
4204 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4208 /* for integer modes, we have more */
4209 if (mode_is_int(mode)) {
4210 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4211 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4212 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4213 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4215 if (tv != tarval_bad) {
4216 relation ^= ir_relation_equal;
4218 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4221 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4222 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4223 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4224 tv = tarval_add(tv, get_mode_one(mode));
4226 if (tv != tarval_bad) {
4227 relation ^= ir_relation_equal;
4229 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4233 /* the following reassociations work only for == and != */
4234 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4235 if (tv != tarval_bad) {
4236 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4238 ir_node *c1 = get_Sub_right(left);
4239 ir_tarval *tv2 = value_of(c1);
4241 if (tv2 != tarval_bad) {
4242 tv2 = tarval_add(tv, value_of(c1));
4244 if (tv2 != tarval_bad) {
4245 left = get_Sub_left(left);
4248 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4252 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4253 else if (is_Add(left)) {
4254 ir_node *a_l = get_Add_left(left);
4255 ir_node *a_r = get_Add_right(left);
4259 if (is_Const(a_l)) {
4261 tv2 = value_of(a_l);
4264 tv2 = value_of(a_r);
4267 if (tv2 != tarval_bad) {
4268 tv2 = tarval_sub(tv, tv2, NULL);
4270 if (tv2 != tarval_bad) {
4274 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4278 /* -a == c ==> a == -c, -a != c ==> a != -c */
4279 else if (is_Minus(left)) {
4280 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4282 if (tv2 != tarval_bad) {
4283 left = get_Minus_op(left);
4286 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4293 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4294 switch (get_irn_opcode(left)) {
4298 c1 = get_And_right(left);
4301 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4302 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4304 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4306 /* TODO: move to constant evaluation */
4307 ir_graph *irg = get_irn_irg(n);
4308 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4309 c1 = new_r_Const(irg, tv);
4310 DBG_OPT_CSTEVAL(n, c1);
4314 if (tarval_is_single_bit(tv)) {
4316 * optimization for AND:
4318 * And(x, C) == C ==> And(x, C) != 0
4319 * And(x, C) != C ==> And(X, C) == 0
4321 * if C is a single Bit constant.
4324 /* check for Constant's match. We have check hare the tarvals,
4325 because our const might be changed */
4326 if (get_Const_tarval(c1) == tv) {
4327 /* fine: do the transformation */
4328 tv = get_mode_null(get_tarval_mode(tv));
4329 relation ^= ir_relation_less_equal_greater;
4331 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4337 c1 = get_Or_right(left);
4338 if (is_Const(c1) && tarval_is_null(tv)) {
4340 * Or(x, C) == 0 && C != 0 ==> FALSE
4341 * Or(x, C) != 0 && C != 0 ==> TRUE
4343 if (! tarval_is_null(get_Const_tarval(c1))) {
4344 /* TODO: move to constant evaluation */
4345 ir_graph *irg = get_irn_irg(n);
4346 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4347 c1 = new_r_Const(irg, tv);
4348 DBG_OPT_CSTEVAL(n, c1);
4355 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4357 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4360 c1 = get_Shl_right(left);
4362 ir_graph *irg = get_irn_irg(c1);
4363 ir_tarval *tv1 = get_Const_tarval(c1);
4364 ir_mode *mode = get_irn_mode(left);
4365 ir_tarval *minus1 = get_mode_all_one(mode);
4366 ir_tarval *amask = tarval_shr(minus1, tv1);
4367 ir_tarval *cmask = tarval_shl(minus1, tv1);
4370 if (tarval_and(tv, cmask) != tv) {
4371 /* condition not met */
4372 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4373 c1 = new_r_Const(irg, tv);
4374 DBG_OPT_CSTEVAL(n, c1);
4377 sl = get_Shl_left(left);
4378 blk = get_nodes_block(n);
4379 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4380 tv = tarval_shr(tv, tv1);
4382 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4387 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4389 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4392 c1 = get_Shr_right(left);
4394 ir_graph *irg = get_irn_irg(c1);
4395 ir_tarval *tv1 = get_Const_tarval(c1);
4396 ir_mode *mode = get_irn_mode(left);
4397 ir_tarval *minus1 = get_mode_all_one(mode);
4398 ir_tarval *amask = tarval_shl(minus1, tv1);
4399 ir_tarval *cmask = tarval_shr(minus1, tv1);
4402 if (tarval_and(tv, cmask) != tv) {
4403 /* condition not met */
4404 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4405 c1 = new_r_Const(irg, tv);
4406 DBG_OPT_CSTEVAL(n, c1);
4409 sl = get_Shr_left(left);
4410 blk = get_nodes_block(n);
4411 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4412 tv = tarval_shl(tv, tv1);
4414 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4419 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4421 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4424 c1 = get_Shrs_right(left);
4426 ir_graph *irg = get_irn_irg(c1);
4427 ir_tarval *tv1 = get_Const_tarval(c1);
4428 ir_mode *mode = get_irn_mode(left);
4429 ir_tarval *minus1 = get_mode_all_one(mode);
4430 ir_tarval *amask = tarval_shl(minus1, tv1);
4431 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4434 cond = tarval_sub(cond, tv1, NULL);
4435 cond = tarval_shrs(tv, cond);
4437 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4438 /* condition not met */
4439 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4440 c1 = new_r_Const(irg, tv);
4441 DBG_OPT_CSTEVAL(n, c1);
4444 sl = get_Shrs_left(left);
4445 blk = get_nodes_block(n);
4446 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4447 tv = tarval_shl(tv, tv1);
4449 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4454 } /* tarval != bad */
4457 if (changedc) { /* need a new Const */
4458 ir_graph *irg = get_irn_irg(n);
4459 right = new_r_Const(irg, tv);
4463 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4464 ir_node *op = get_Proj_pred(left);
4466 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4467 ir_node *c = get_binop_right(op);
4470 ir_tarval *tv = get_Const_tarval(c);
4472 if (tarval_is_single_bit(tv)) {
4473 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4474 ir_node *v = get_binop_left(op);
4475 ir_node *blk = get_irn_n(op, -1);
4476 ir_graph *irg = get_irn_irg(op);
4477 ir_mode *mode = get_irn_mode(v);
4479 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4480 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4482 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4489 dbg_info *dbgi = get_irn_dbg_info(n);
4490 ir_node *block = get_nodes_block(n);
4492 /* create a new compare */
4493 n = new_rd_Cmp(dbgi, block, left, right, relation);
4500 * Optimize CopyB(mem, x, x) into a Nop.
4502 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4504 ir_node *copyb = get_Proj_pred(proj);
4505 ir_node *a = get_CopyB_dst(copyb);
4506 ir_node *b = get_CopyB_src(copyb);
4509 switch (get_Proj_proj(proj)) {
4510 case pn_CopyB_X_regular:
4511 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4512 DBG_OPT_EXC_REM(proj);
4513 proj = new_r_Jmp(get_nodes_block(copyb));
4515 case pn_CopyB_X_except: {
4516 ir_graph *irg = get_irn_irg(proj);
4517 DBG_OPT_EXC_REM(proj);
4518 proj = new_r_Bad(irg, mode_X);
4526 } /* transform_node_Proj_CopyB */
4529 * Optimize Bounds(idx, idx, upper) into idx.
4531 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4533 ir_node *oldn = proj;
4534 ir_node *bound = get_Proj_pred(proj);
4535 ir_node *idx = get_Bound_index(bound);
4536 ir_node *pred = skip_Proj(idx);
4539 if (idx == get_Bound_lower(bound))
4541 else if (is_Bound(pred)) {
4543 * idx was Bounds checked previously, it is still valid if
4544 * lower <= pred_lower && pred_upper <= upper.
4546 ir_node *lower = get_Bound_lower(bound);
4547 ir_node *upper = get_Bound_upper(bound);
4548 if (get_Bound_lower(pred) == lower &&
4549 get_Bound_upper(pred) == upper) {
4551 * One could expect that we simply return the previous
4552 * Bound here. However, this would be wrong, as we could
4553 * add an exception Proj to a new location then.
4554 * So, we must turn in into a tuple.
4560 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4561 switch (get_Proj_proj(proj)) {
4563 DBG_OPT_EXC_REM(proj);
4564 proj = get_Bound_mem(bound);
4566 case pn_Bound_X_except:
4567 DBG_OPT_EXC_REM(proj);
4568 proj = new_r_Bad(get_irn_irg(proj), mode_X);
4572 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4574 case pn_Bound_X_regular:
4575 DBG_OPT_EXC_REM(proj);
4576 proj = new_r_Jmp(get_nodes_block(bound));
4583 } /* transform_node_Proj_Bound */
4586 * Does all optimizations on nodes that must be done on its Projs
4587 * because of creating new nodes.
4589 static ir_node *transform_node_Proj(ir_node *proj)
4591 ir_node *n = get_Proj_pred(proj);
4593 if (n->op->ops.transform_node_Proj)
4594 return n->op->ops.transform_node_Proj(proj);
4596 } /* transform_node_Proj */
4598 static bool is_block_unreachable(const ir_node *block)
4600 const ir_graph *irg = get_irn_irg(block);
4601 if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK))
4603 return get_Block_dom_depth(block) < 0;
4606 static ir_node *transform_node_Block(ir_node *block)
4608 ir_graph *irg = get_irn_irg(block);
4609 int arity = get_irn_arity(block);
4610 ir_node *bad = NULL;
4613 if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK))
4616 for (i = 0; i < arity; ++i) {
4617 ir_node *pred = get_Block_cfgpred(block, i);
4618 ir_node *pred_block = get_nodes_block(pred);
4619 if (!is_Bad(pred) && !is_block_unreachable(pred_block))
4622 bad = new_r_Bad(irg, mode_X);
4623 set_irn_n(block, i, bad);
4629 static ir_node *transform_node_Phi(ir_node *phi)
4631 int n = get_irn_arity(phi);
4632 ir_mode *mode = get_irn_mode(phi);
4633 ir_node *block = get_nodes_block(phi);
4634 ir_graph *irg = get_irn_irg(phi);
4635 ir_node *bad = NULL;
4638 /* Set phi-operands for bad-block inputs to bad */
4639 for (i = 0; i < n; ++i) {
4640 ir_node *pred = get_Block_cfgpred(block, i);
4641 if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) {
4643 bad = new_r_Bad(irg, mode);
4644 set_irn_n(phi, i, bad);
4648 /* Move Pin nodes down through Phi nodes. */
4649 if (mode == mode_M) {
4650 n = get_irn_arity(phi);
4652 /* Beware of Phi0 */
4657 NEW_ARR_A(ir_node *, in, n);
4659 for (i = 0; i < n; ++i) {
4660 ir_node *pred = get_irn_n(phi, i);
4665 in[i] = get_Pin_op(pred);
4668 /* Move the Pin nodes "behind" the Phi. */
4669 block = get_irn_n(phi, -1);
4670 new_phi = new_r_Phi(block, n, in, mode_M);
4671 return new_r_Pin(block, new_phi);
4674 /* Move Confirms down through Phi nodes. */
4675 else if (mode_is_reference(mode)) {
4676 n = get_irn_arity(phi);
4678 /* Beware of Phi0 */
4680 ir_node *pred = get_irn_n(phi, 0);
4681 ir_node *bound, *new_phi, *block, **in;
4682 ir_relation relation;
4684 if (! is_Confirm(pred))
4687 bound = get_Confirm_bound(pred);
4688 relation = get_Confirm_relation(pred);
4690 NEW_ARR_A(ir_node *, in, n);
4691 in[0] = get_Confirm_value(pred);
4693 for (i = 1; i < n; ++i) {
4694 pred = get_irn_n(phi, i);
4696 if (! is_Confirm(pred) ||
4697 get_Confirm_bound(pred) != bound ||
4698 get_Confirm_relation(pred) != relation)
4700 in[i] = get_Confirm_value(pred);
4702 /* move the Confirm nodes "behind" the Phi */
4703 block = get_irn_n(phi, -1);
4704 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4705 return new_r_Confirm(block, new_phi, bound, relation);
4712 * Returns the operands of a commutative bin-op, if one operand is
4713 * a const, it is returned as the second one.
4715 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4717 ir_node *op_a = get_binop_left(binop);
4718 ir_node *op_b = get_binop_right(binop);
4720 assert(is_op_commutative(get_irn_op(binop)));
4722 if (is_Const(op_a)) {
4729 } /* get_comm_Binop_Ops */
4732 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4733 * Such pattern may arise in bitfield stores.
4735 * value c4 value c4 & c2
4736 * AND c3 AND c1 | c3
4743 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4746 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4748 ir_node *irn_and, *c1;
4750 ir_node *and_l, *c3;
4751 ir_node *value, *c4;
4752 ir_node *new_and, *new_const, *block;
4753 ir_mode *mode = get_irn_mode(irn_or);
4755 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4759 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4760 if (!is_Const(c1) || !is_And(irn_and))
4763 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4767 tv1 = get_Const_tarval(c1);
4768 tv2 = get_Const_tarval(c2);
4770 tv = tarval_or(tv1, tv2);
4771 if (tarval_is_all_one(tv)) {
4772 /* the AND does NOT clear a bit with isn't set by the OR */
4773 set_Or_left(irn_or, or_l);
4774 set_Or_right(irn_or, c1);
4776 /* check for more */
4783 get_comm_Binop_Ops(or_l, &and_l, &c3);
4784 if (!is_Const(c3) || !is_And(and_l))
4787 get_comm_Binop_Ops(and_l, &value, &c4);
4791 /* ok, found the pattern, check for conditions */
4792 assert(mode == get_irn_mode(irn_and));
4793 assert(mode == get_irn_mode(or_l));
4794 assert(mode == get_irn_mode(and_l));
4796 tv3 = get_Const_tarval(c3);
4797 tv4 = get_Const_tarval(c4);
4799 tv = tarval_or(tv4, tv2);
4800 if (!tarval_is_all_one(tv)) {
4801 /* have at least one 0 at the same bit position */
4805 if (tv3 != tarval_andnot(tv3, tv4)) {
4806 /* bit in the or_mask is outside the and_mask */
4810 if (tv1 != tarval_andnot(tv1, tv2)) {
4811 /* bit in the or_mask is outside the and_mask */
4815 /* ok, all conditions met */
4816 block = get_irn_n(irn_or, -1);
4817 irg = get_irn_irg(block);
4819 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4821 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4823 set_Or_left(irn_or, new_and);
4824 set_Or_right(irn_or, new_const);
4826 /* check for more */
4828 } /* transform_node_Or_bf_store */
4831 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4833 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4835 ir_mode *mode = get_irn_mode(irn_or);
4836 ir_node *shl, *shr, *block;
4837 ir_node *irn, *x, *c1, *c2, *n;
4838 ir_tarval *tv1, *tv2;
4840 /* some backends can't handle rotl */
4841 if (!be_get_backend_param()->support_rotl)
4844 if (! mode_is_int(mode))
4847 shl = get_binop_left(irn_or);
4848 shr = get_binop_right(irn_or);
4857 } else if (!is_Shl(shl)) {
4859 } else if (!is_Shr(shr)) {
4862 x = get_Shl_left(shl);
4863 if (x != get_Shr_left(shr))
4866 c1 = get_Shl_right(shl);
4867 c2 = get_Shr_right(shr);
4868 if (is_Const(c1) && is_Const(c2)) {
4869 tv1 = get_Const_tarval(c1);
4870 if (! tarval_is_long(tv1))
4873 tv2 = get_Const_tarval(c2);
4874 if (! tarval_is_long(tv2))
4877 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4878 != (int) get_mode_size_bits(mode))
4881 /* yet, condition met */
4882 block = get_nodes_block(irn_or);
4884 n = new_r_Rotl(block, x, c1, mode);
4886 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4890 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4891 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4892 if (!ir_is_negated_value(c1, c2)) {
4896 /* yet, condition met */
4897 block = get_nodes_block(irn_or);
4898 n = new_r_Rotl(block, x, c1, mode);
4899 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4901 } /* transform_node_Or_Rotl */
4903 static bool is_cmp_unequal(const ir_node *node)
4905 ir_relation relation = get_Cmp_relation(node);
4906 ir_node *left = get_Cmp_left(node);
4907 ir_node *right = get_Cmp_right(node);
4908 ir_mode *mode = get_irn_mode(left);
4910 if (relation == ir_relation_less_greater)
4913 if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right))
4914 return relation == ir_relation_greater;
4921 static ir_node *transform_node_Or(ir_node *n)
4923 ir_node *c, *oldn = n;
4924 ir_node *a = get_Or_left(n);
4925 ir_node *b = get_Or_right(n);
4928 if (is_Not(a) && is_Not(b)) {
4929 /* ~a | ~b = ~(a&b) */
4930 ir_node *block = get_nodes_block(n);
4932 mode = get_irn_mode(n);
4935 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4936 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4937 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4941 /* we can combine the relations of two compares with the same operands */
4942 if (is_Cmp(a) && is_Cmp(b)) {
4943 ir_node *a_left = get_Cmp_left(a);
4944 ir_node *a_right = get_Cmp_right(a);
4945 ir_node *b_left = get_Cmp_left(b);
4946 ir_node *b_right = get_Cmp_right(b);
4947 if (a_left == b_left && b_left == b_right) {
4948 dbg_info *dbgi = get_irn_dbg_info(n);
4949 ir_node *block = get_nodes_block(n);
4950 ir_relation a_relation = get_Cmp_relation(a);
4951 ir_relation b_relation = get_Cmp_relation(b);
4952 ir_relation new_relation = a_relation | b_relation;
4953 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4955 /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */
4956 if (is_cmp_unequal(a) && is_cmp_unequal(b)
4957 && !mode_is_float(get_irn_mode(a_left))
4958 && !mode_is_float(get_irn_mode(b_left))) {
4959 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
4960 ir_graph *irg = get_irn_irg(n);
4961 dbg_info *dbgi = get_irn_dbg_info(n);
4962 ir_node *block = get_nodes_block(n);
4963 ir_mode *a_mode = get_irn_mode(a_left);
4964 ir_mode *b_mode = get_irn_mode(b_left);
4965 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
4966 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
4967 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
4968 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
4969 ir_node *zero = create_zero_const(irg, b_mode);
4970 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4972 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
4973 ir_graph *irg = get_irn_irg(n);
4974 dbg_info *dbgi = get_irn_dbg_info(n);
4975 ir_node *block = get_nodes_block(n);
4976 ir_mode *a_mode = get_irn_mode(a_left);
4977 ir_mode *b_mode = get_irn_mode(b_left);
4978 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
4979 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
4980 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
4981 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
4982 ir_node *zero = create_zero_const(irg, a_mode);
4983 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4988 mode = get_irn_mode(n);
4989 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4991 n = transform_node_Or_bf_store(n);
4992 n = transform_node_Or_Rotl(n);
4996 n = transform_bitwise_distributive(n, transform_node_Or);
4998 n = transform_node_bitop_shift(n);
5001 } /* transform_node_Or */
5005 static ir_node *transform_node(ir_node *n);
5008 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5010 * Should be moved to reassociation?
5012 static ir_node *transform_node_shift(ir_node *n)
5014 ir_node *left, *right;
5016 ir_mode *count_mode;
5017 ir_tarval *tv1, *tv2, *res;
5018 ir_node *in[2], *irn, *block;
5022 left = get_binop_left(n);
5024 /* different operations */
5025 if (get_irn_op(left) != get_irn_op(n))
5028 right = get_binop_right(n);
5029 tv1 = value_of(right);
5030 if (tv1 == tarval_bad)
5033 tv2 = value_of(get_binop_right(left));
5034 if (tv2 == tarval_bad)
5037 count_mode = get_tarval_mode(tv1);
5038 if (get_tarval_mode(tv2) != count_mode) {
5039 /* TODO: search bigger mode or something and convert... */
5043 mode = get_irn_mode(n);
5044 modulo_shf = get_mode_modulo_shift(mode);
5046 if (modulo_shf > 0) {
5047 ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode);
5049 /* I'm not so sure what happens in one complement... */
5050 assert(get_mode_arithmetic(count_mode) == irma_twos_complement);
5051 /* modulo shifts should always be a power of 2 (otherwise modulo_mask
5052 * above will be invalid) */
5053 assert(modulo_shf<=0 || is_po2(modulo_shf));
5055 tv1 = tarval_and(tv1, modulo_mask);
5056 tv2 = tarval_and(tv2, modulo_mask);
5058 res = tarval_add(tv1, tv2);
5059 irg = get_irn_irg(n);
5061 /* beware: a simple replacement works only, if res < modulo shift */
5063 int bits = get_mode_size_bits(mode);
5064 ir_tarval *modulo = new_tarval_from_long(bits, count_mode);
5065 res = tarval_mod(res, modulo);
5067 long bits = get_mode_size_bits(mode);
5068 ir_tarval *mode_size = new_tarval_from_long(bits, count_mode);
5070 /* shifting too much */
5071 if (!(tarval_cmp(res, mode_size) & ir_relation_less)) {
5073 ir_node *block = get_nodes_block(n);
5074 dbg_info *dbgi = get_irn_dbg_info(n);
5075 ir_mode *smode = get_irn_mode(right);
5076 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
5077 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5080 return new_r_Const(irg, get_mode_null(mode));
5084 /* ok, we can replace it */
5085 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5086 block = get_nodes_block(n);
5088 in[0] = get_binop_left(left);
5089 in[1] = new_r_Const(irg, res);
5091 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5093 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5095 return transform_node(irn);
5100 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5102 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5103 * (also with x >>s c1 when c1>=c2)
5105 static ir_node *transform_node_shl_shr(ir_node *n)
5108 ir_node *right = get_binop_right(n);
5118 ir_tarval *tv_shift;
5121 ir_relation relation;
5124 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5126 if (!is_Const(right))
5129 left = get_binop_left(n);
5130 mode = get_irn_mode(n);
5131 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5132 ir_node *shr_right = get_binop_right(left);
5134 if (!is_Const(shr_right))
5137 x = get_binop_left(left);
5138 tv_shr = get_Const_tarval(shr_right);
5139 tv_shl = get_Const_tarval(right);
5141 if (is_Shrs(left)) {
5142 /* shrs variant only allowed if c1 >= c2 */
5143 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5146 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5149 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5151 tv_mask = tarval_shl(tv_mask, tv_shl);
5152 } else if (is_Shr(n) && is_Shl(left)) {
5153 ir_node *shl_right = get_Shl_right(left);
5155 if (!is_Const(shl_right))
5158 x = get_Shl_left(left);
5159 tv_shr = get_Const_tarval(right);
5160 tv_shl = get_Const_tarval(shl_right);
5162 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5163 tv_mask = tarval_shr(tv_mask, tv_shr);
5168 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5169 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5172 assert(tv_mask != tarval_bad);
5173 assert(get_tarval_mode(tv_mask) == mode);
5175 block = get_nodes_block(n);
5176 irg = get_irn_irg(block);
5177 dbgi = get_irn_dbg_info(n);
5179 relation = tarval_cmp(tv_shl, tv_shr);
5180 if (relation == ir_relation_less || relation == ir_relation_equal) {
5181 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5182 new_const = new_r_Const(irg, tv_shift);
5184 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5186 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5189 assert(relation == ir_relation_greater);
5190 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5191 new_const = new_r_Const(irg, tv_shift);
5192 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5195 new_const = new_r_Const(irg, tv_mask);
5196 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5201 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5203 ir_mode *mode = get_tarval_mode(tv);
5204 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5205 return tarval_mod(tv, modulo_tv);
5208 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5209 ir_node *left, ir_node *right, ir_mode *mode);
5212 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5213 * then we can use that to minimize the value of Add(x, const) or
5214 * Sub(Const, x). In particular this often avoids 1 instruction in some
5215 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5216 * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
5218 static ir_node *transform_node_shift_modulo(ir_node *n,
5219 new_shift_func new_shift)
5221 ir_mode *mode = get_irn_mode(n);
5222 int modulo = get_mode_modulo_shift(mode);
5223 ir_node *newop = NULL;
5224 ir_mode *mode_right;
5231 if (get_mode_arithmetic(mode) != irma_twos_complement)
5233 if (!is_po2(modulo))
5236 irg = get_irn_irg(n);
5237 block = get_nodes_block(n);
5238 right = get_binop_right(n);
5239 mode_right = get_irn_mode(right);
5240 if (is_Const(right)) {
5241 ir_tarval *tv = get_Const_tarval(right);
5242 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5247 newop = new_r_Const(irg, tv_mod);
5248 } else if (is_Add(right)) {
5249 ir_node *add_right = get_Add_right(right);
5250 if (is_Const(add_right)) {
5251 ir_tarval *tv = get_Const_tarval(add_right);
5252 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5257 newconst = new_r_Const(irg, tv_mod);
5258 newop = new_r_Add(block, get_Add_left(right), newconst,
5261 } else if (is_Sub(right)) {
5262 ir_node *sub_left = get_Sub_left(right);
5263 if (is_Const(sub_left)) {
5264 ir_tarval *tv = get_Const_tarval(sub_left);
5265 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5270 newconst = new_r_Const(irg, tv_mod);
5271 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5278 if (newop != NULL) {
5279 dbg_info *dbgi = get_irn_dbg_info(n);
5280 ir_node *left = get_binop_left(n);
5281 return new_shift(dbgi, block, left, newop, mode);
5289 static ir_node *transform_node_Shr(ir_node *n)
5291 ir_node *c, *oldn = n;
5292 ir_node *left = get_Shr_left(n);
5293 ir_node *right = get_Shr_right(n);
5294 ir_mode *mode = get_irn_mode(n);
5296 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5297 n = transform_node_shift(n);
5300 n = transform_node_shift_modulo(n, new_rd_Shr);
5302 n = transform_node_shl_shr(n);
5304 n = transform_node_shift_bitop(n);
5307 } /* transform_node_Shr */
5312 static ir_node *transform_node_Shrs(ir_node *n)
5314 ir_node *c, *oldn = n;
5315 ir_node *a = get_Shrs_left(n);
5316 ir_node *b = get_Shrs_right(n);
5317 ir_mode *mode = get_irn_mode(n);
5319 if (is_oversize_shift(n)) {
5320 ir_node *block = get_nodes_block(n);
5321 dbg_info *dbgi = get_irn_dbg_info(n);
5322 ir_mode *cmode = get_irn_mode(b);
5323 long val = get_mode_size_bits(cmode)-1;
5324 ir_graph *irg = get_irn_irg(n);
5325 ir_node *cnst = new_r_Const_long(irg, cmode, val);
5326 return new_rd_Shrs(dbgi, block, a, cnst, mode);
5329 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5330 n = transform_node_shift(n);
5333 n = transform_node_shift_modulo(n, new_rd_Shrs);
5335 n = transform_node_shift_bitop(n);
5338 } /* transform_node_Shrs */
5343 static ir_node *transform_node_Shl(ir_node *n)
5345 ir_node *c, *oldn = n;
5346 ir_node *a = get_Shl_left(n);
5347 ir_node *b = get_Shl_right(n);
5348 ir_mode *mode = get_irn_mode(n);
5350 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5351 n = transform_node_shift(n);
5354 n = transform_node_shift_modulo(n, new_rd_Shl);
5356 n = transform_node_shl_shr(n);
5358 n = transform_node_shift_bitop(n);
5361 } /* transform_node_Shl */
5366 static ir_node *transform_node_Rotl(ir_node *n)
5368 ir_node *c, *oldn = n;
5369 ir_node *a = get_Rotl_left(n);
5370 ir_node *b = get_Rotl_right(n);
5371 ir_mode *mode = get_irn_mode(n);
5373 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5374 n = transform_node_shift(n);
5377 n = transform_node_shift_bitop(n);
5380 } /* transform_node_Rotl */
5385 static ir_node *transform_node_Conv(ir_node *n)
5387 ir_node *c, *oldn = n;
5388 ir_mode *mode = get_irn_mode(n);
5389 ir_node *a = get_Conv_op(n);
5391 if (mode != mode_b && is_const_Phi(a)) {
5392 /* Do NOT optimize mode_b Conv's, this leads to remaining
5393 * Phib nodes later, because the conv_b_lower operation
5394 * is instantly reverted, when it tries to insert a Convb.
5396 c = apply_conv_on_phi(a, mode);
5398 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5403 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5404 ir_graph *irg = get_irn_irg(n);
5405 return new_r_Unknown(irg, mode);
5408 if (mode_is_reference(mode) &&
5409 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5411 ir_node *l = get_Add_left(a);
5412 ir_node *r = get_Add_right(a);
5413 dbg_info *dbgi = get_irn_dbg_info(a);
5414 ir_node *block = get_nodes_block(n);
5416 ir_node *lop = get_Conv_op(l);
5417 if (get_irn_mode(lop) == mode) {
5418 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5419 n = new_rd_Add(dbgi, block, lop, r, mode);
5424 ir_node *rop = get_Conv_op(r);
5425 if (get_irn_mode(rop) == mode) {
5426 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5427 n = new_rd_Add(dbgi, block, l, rop, mode);
5434 } /* transform_node_Conv */
5437 * Remove dead blocks and nodes in dead blocks
5438 * in keep alive list. We do not generate a new End node.
5440 static ir_node *transform_node_End(ir_node *n)
5442 int i, j, n_keepalives = get_End_n_keepalives(n);
5445 NEW_ARR_A(ir_node *, in, n_keepalives);
5447 for (i = j = 0; i < n_keepalives; ++i) {
5448 ir_node *ka = get_End_keepalive(n, i);
5450 /* no need to keep Bad */
5453 /* do not keep unreachable code */
5454 block = is_Block(ka) ? ka : get_nodes_block(ka);
5455 if (is_block_unreachable(block))
5459 if (j != n_keepalives)
5460 set_End_keepalives(n, j, in);
5462 } /* transform_node_End */
5464 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5466 if (is_Minus(a) && get_Minus_op(a) == b)
5468 if (is_Minus(b) && get_Minus_op(b) == a)
5470 if (is_Sub(a) && is_Sub(b)) {
5471 ir_node *a_left = get_Sub_left(a);
5472 ir_node *a_right = get_Sub_right(a);
5473 ir_node *b_left = get_Sub_left(b);
5474 ir_node *b_right = get_Sub_right(b);
5476 if (a_left == b_right && a_right == b_left)
5483 static const ir_node *skip_upconv(const ir_node *node)
5485 while (is_Conv(node)) {
5486 ir_mode *mode = get_irn_mode(node);
5487 const ir_node *op = get_Conv_op(node);
5488 ir_mode *op_mode = get_irn_mode(op);
5489 if (!smaller_mode(op_mode, mode))
5496 int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true,
5497 const ir_node *mux_false)
5502 ir_relation relation;
5508 * Note further that these optimization work even for floating point
5509 * with NaN's because -NaN == NaN.
5510 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5513 mode = get_irn_mode(mux_true);
5514 if (mode_honor_signed_zeros(mode))
5517 /* must be <, <=, >=, > */
5518 relation = get_Cmp_relation(sel);
5519 if ((relation & ir_relation_less_greater) == 0)
5522 if (!ir_is_negated_value(mux_true, mux_false))
5525 mux_true = skip_upconv(mux_true);
5526 mux_false = skip_upconv(mux_false);
5528 /* must be x cmp 0 */
5529 cmp_right = get_Cmp_right(sel);
5530 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
5533 cmp_left = get_Cmp_left(sel);
5534 if (cmp_left == mux_false) {
5535 if (relation & ir_relation_less) {
5538 assert(relation & ir_relation_greater);
5541 } else if (cmp_left == mux_true) {
5542 if (relation & ir_relation_less) {
5545 assert(relation & ir_relation_greater);
5553 ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true,
5556 ir_node *cmp_left = get_Cmp_left(sel);
5557 return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
5561 * Optimize a Mux into some simpler cases.
5563 static ir_node *transform_node_Mux(ir_node *n)
5566 ir_node *sel = get_Mux_sel(n);
5567 ir_mode *mode = get_irn_mode(n);
5568 ir_node *t = get_Mux_true(n);
5569 ir_node *f = get_Mux_false(n);
5570 ir_graph *irg = get_irn_irg(n);
5572 /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
5573 if (get_mode_arithmetic(mode) == irma_twos_complement) {
5574 int abs = ir_mux_is_abs(sel, t, f);
5576 dbg_info *dbgi = get_irn_dbg_info(n);
5577 ir_node *block = get_nodes_block(n);
5578 ir_node *op = ir_get_abs_op(sel, t, f);
5579 int bits = get_mode_size_bits(mode);
5580 ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
5581 ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
5582 ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode);
5585 res = new_rd_Sub(dbgi, block, xorn, sext, mode);
5587 res = new_rd_Sub(dbgi, block, sext, xorn, mode);
5593 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5597 ir_node* block = get_nodes_block(n);
5599 ir_node* c1 = get_Mux_sel(t);
5600 ir_node* t1 = get_Mux_true(t);
5601 ir_node* f1 = get_Mux_false(t);
5603 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5604 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5605 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5610 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5611 } else if (f == t1) {
5612 /* Mux(cond0, Mux(cond1, x, y), x) */
5613 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5614 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5615 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5620 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5622 } else if (is_Mux(f)) {
5623 ir_node* block = get_nodes_block(n);
5625 ir_node* c1 = get_Mux_sel(f);
5626 ir_node* t1 = get_Mux_true(f);
5627 ir_node* f1 = get_Mux_false(f);
5629 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5630 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5631 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5636 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5637 } else if (t == f1) {
5638 /* Mux(cond0, x, Mux(cond1, y, x)) */
5639 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5640 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5641 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5646 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5650 /* first normalization step: try to move a constant to the false side,
5651 * 0 preferred on false side too */
5652 if (is_Cmp(sel) && is_Const(t) &&
5653 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5654 dbg_info *seldbgi = get_irn_dbg_info(sel);
5655 ir_node *block = get_nodes_block(sel);
5656 ir_relation relation = get_Cmp_relation(sel);
5661 /* Mux(x, a, b) => Mux(not(x), b, a) */
5662 relation = get_negated_relation(relation);
5663 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5664 get_Cmp_right(sel), relation);
5665 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5668 /* note: after normalization, false can only happen on default */
5669 if (mode == mode_b) {
5670 dbg_info *dbg = get_irn_dbg_info(n);
5671 ir_node *block = get_nodes_block(n);
5674 ir_tarval *tv_t = get_Const_tarval(t);
5675 if (tv_t == tarval_b_true) {
5677 /* Muxb(sel, true, false) = sel */
5678 assert(get_Const_tarval(f) == tarval_b_false);
5679 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5682 /* Muxb(sel, true, x) = Or(sel, x) */
5683 n = new_rd_Or(dbg, block, sel, f, mode_b);
5684 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5688 } else if (is_Const(f)) {
5689 ir_tarval *tv_f = get_Const_tarval(f);
5690 if (tv_f == tarval_b_true) {
5691 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5692 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5693 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5694 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5697 /* Muxb(sel, x, false) = And(sel, x) */
5698 assert(tv_f == tarval_b_false);
5699 n = new_rd_And(dbg, block, sel, t, mode_b);
5700 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5706 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5707 * value to integer. */
5708 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5709 ir_tarval *a = get_Const_tarval(t);
5710 ir_tarval *b = get_Const_tarval(f);
5712 if (tarval_is_one(a) && tarval_is_null(b)) {
5713 ir_node *block = get_nodes_block(n);
5714 ir_node *conv = new_r_Conv(block, sel, mode);
5716 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5718 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5719 ir_node *block = get_nodes_block(n);
5720 ir_node *not_ = new_r_Not(block, sel, mode_b);
5721 ir_node *conv = new_r_Conv(block, not_, mode);
5723 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5729 ir_node *cmp_r = get_Cmp_right(sel);
5730 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5731 ir_node *block = get_nodes_block(n);
5732 ir_node *cmp_l = get_Cmp_left(sel);
5734 if (mode_is_int(mode)) {
5735 ir_relation relation = get_Cmp_relation(sel);
5737 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5738 /* Mux((a & b) != 0, c, 0) */
5739 ir_node *and_r = get_And_right(cmp_l);
5742 if (and_r == t && f == cmp_r) {
5743 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5744 if (relation == ir_relation_less_greater) {
5745 /* Mux((a & 2^C) != 0, 2^C, 0) */
5747 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5749 /* Mux((a & 2^C) == 0, 2^C, 0) */
5750 n = new_rd_Eor(get_irn_dbg_info(n),
5751 block, cmp_l, t, mode);
5752 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5757 if (is_Shl(and_r)) {
5758 ir_node *shl_l = get_Shl_left(and_r);
5759 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5760 if (and_r == t && f == cmp_r) {
5761 if (relation == ir_relation_less_greater) {
5762 /* (a & (1 << n)) != 0, (1 << n), 0) */
5764 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5766 /* (a & (1 << n)) == 0, (1 << n), 0) */
5767 n = new_rd_Eor(get_irn_dbg_info(n),
5768 block, cmp_l, t, mode);
5769 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5775 and_l = get_And_left(cmp_l);
5776 if (is_Shl(and_l)) {
5777 ir_node *shl_l = get_Shl_left(and_l);
5778 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5779 if (and_l == t && f == cmp_r) {
5780 if (relation == ir_relation_less_greater) {
5781 /* ((1 << n) & a) != 0, (1 << n), 0) */
5783 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5785 /* ((1 << n) & a) == 0, (1 << n), 0) */
5786 n = new_rd_Eor(get_irn_dbg_info(n),
5787 block, cmp_l, t, mode);
5788 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5803 * optimize Sync nodes that have other syncs as input we simply add the inputs
5804 * of the other sync to our own inputs
5806 static ir_node *transform_node_Sync(ir_node *n)
5808 int arity = get_Sync_n_preds(n);
5811 for (i = 0; i < arity;) {
5812 ir_node *pred = get_Sync_pred(n, i);
5816 /* Remove Bad predecessors */
5823 /* Remove duplicate predecessors */
5824 for (j = 0; j < i; ++j) {
5825 if (get_Sync_pred(n, j) == pred) {
5834 if (!is_Sync(pred)) {
5842 pred_arity = get_Sync_n_preds(pred);
5843 for (j = 0; j < pred_arity; ++j) {
5844 ir_node *pred_pred = get_Sync_pred(pred, j);
5849 add_irn_n(n, pred_pred);
5853 if (get_Sync_pred(n, k) == pred_pred) break;
5859 ir_graph *irg = get_irn_irg(n);
5860 return new_r_Bad(irg, mode_M);
5863 return get_Sync_pred(n, 0);
5866 /* rehash the sync node */
5871 static ir_node *transform_node_Load(ir_node *n)
5873 /* if our memory predecessor is a load from the same address, then reuse the
5874 * previous result */
5875 ir_node *mem = get_Load_mem(n);
5880 /* don't touch volatile loads */
5881 if (get_Load_volatility(n) == volatility_is_volatile)
5883 mem_pred = get_Proj_pred(mem);
5884 if (is_Load(mem_pred)) {
5885 ir_node *pred_load = mem_pred;
5887 /* conservatively compare the 2 loads. TODO: This could be less strict
5888 * with fixup code in some situations (like smaller/bigger modes) */
5889 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
5891 if (get_Load_mode(pred_load) != get_Load_mode(n))
5893 /* all combinations of aligned/unaligned pred/n should be fine so we do
5894 * not compare the unaligned attribute */
5896 ir_node *block = get_nodes_block(n);
5897 ir_node *jmp = new_r_Jmp(block);
5898 ir_graph *irg = get_irn_irg(n);
5899 ir_node *bad = new_r_Bad(irg, mode_X);
5900 ir_mode *mode = get_Load_mode(n);
5901 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
5902 ir_node *in[] = { mem, res, jmp, bad };
5903 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5906 } else if (is_Store(mem_pred)) {
5907 ir_node *pred_store = mem_pred;
5908 ir_node *value = get_Store_value(pred_store);
5910 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
5912 if (get_irn_mode(value) != get_Load_mode(n))
5914 /* all combinations of aligned/unaligned pred/n should be fine so we do
5915 * not compare the unaligned attribute */
5917 ir_node *block = get_nodes_block(n);
5918 ir_node *jmp = new_r_Jmp(block);
5919 ir_graph *irg = get_irn_irg(n);
5920 ir_node *bad = new_r_Bad(irg, mode_X);
5921 ir_node *res = value;
5922 ir_node *in[] = { mem, res, jmp, bad };
5923 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5932 * optimize a trampoline Call into a direct Call
5934 static ir_node *transform_node_Call(ir_node *call)
5936 ir_node *callee = get_Call_ptr(call);
5937 ir_node *adr, *mem, *res, *bl, **in;
5938 ir_type *ctp, *mtp, *tp;
5942 size_t i, n_res, n_param;
5945 if (! is_Proj(callee))
5947 callee = get_Proj_pred(callee);
5948 if (! is_Builtin(callee))
5950 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5953 mem = get_Call_mem(call);
5955 if (skip_Proj(mem) == callee) {
5956 /* memory is routed to the trampoline, skip */
5957 mem = get_Builtin_mem(callee);
5960 /* build a new call type */
5961 mtp = get_Call_type(call);
5962 tdb = get_type_dbg_info(mtp);
5964 n_res = get_method_n_ress(mtp);
5965 n_param = get_method_n_params(mtp);
5966 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5968 for (i = 0; i < n_res; ++i)
5969 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5971 NEW_ARR_A(ir_node *, in, n_param + 1);
5973 /* FIXME: we don't need a new pointer type in every step */
5974 irg = get_irn_irg(call);
5975 tp = get_irg_frame_type(irg);
5976 tp = new_type_pointer(tp);
5977 set_method_param_type(ctp, 0, tp);
5979 in[0] = get_Builtin_param(callee, 2);
5980 for (i = 0; i < n_param; ++i) {
5981 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5982 in[i + 1] = get_Call_param(call, i);
5984 var = get_method_variadicity(mtp);
5985 set_method_variadicity(ctp, var);
5986 /* When we resolve a trampoline, the function must be called by a this-call */
5987 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5988 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5990 adr = get_Builtin_param(callee, 1);
5992 db = get_irn_dbg_info(call);
5993 bl = get_nodes_block(call);
5995 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5996 if (get_irn_pinned(call) == op_pin_state_floats)
5997 set_irn_pinned(res, op_pin_state_floats);
5999 } /* transform_node_Call */
6002 * Tries several [inplace] [optimizing] transformations and returns an
6003 * equivalent node. The difference to equivalent_node() is that these
6004 * transformations _do_ generate new nodes, and thus the old node must
6005 * not be freed even if the equivalent node isn't the old one.
6007 static ir_node *transform_node(ir_node *n)
6012 * Transform_node is the only "optimizing transformation" that might
6013 * return a node with a different opcode. We iterate HERE until fixpoint
6014 * to get the final result.
6018 if (n->op->ops.transform_node != NULL)
6019 n = n->op->ops.transform_node(n);
6020 } while (oldn != n);
6023 } /* transform_node */
6026 * Sets the default transform node operation for an ir_op_ops.
6028 * @param code the opcode for the default operation
6029 * @param ops the operations initialized
6034 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
6038 ops->transform_node = transform_node_##a; \
6040 #define CASE_PROJ(a) \
6042 ops->transform_node_Proj = transform_node_Proj_##a; \
6044 #define CASE_PROJ_EX(a) \
6046 ops->transform_node = transform_node_##a; \
6047 ops->transform_node_Proj = transform_node_Proj_##a; \
6088 } /* firm_set_default_transform_node */
6091 /* **************** Common Subexpression Elimination **************** */
6093 /** The size of the hash table used, should estimate the number of nodes
6095 #define N_IR_NODES 512
6097 /** Compares the attributes of two Const nodes. */
6098 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
6100 return get_Const_tarval(a) != get_Const_tarval(b);
6103 /** Compares the attributes of two Proj nodes. */
6104 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
6106 return a->attr.proj.proj != b->attr.proj.proj;
6109 /** Compares the attributes of two Alloc nodes. */
6110 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
6112 const alloc_attr *pa = &a->attr.alloc;
6113 const alloc_attr *pb = &b->attr.alloc;
6114 return (pa->where != pb->where) || (pa->type != pb->type);
6117 /** Compares the attributes of two Free nodes. */
6118 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
6120 const free_attr *pa = &a->attr.free;
6121 const free_attr *pb = &b->attr.free;
6122 return (pa->where != pb->where) || (pa->type != pb->type);
6125 /** Compares the attributes of two SymConst nodes. */
6126 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
6128 const symconst_attr *pa = &a->attr.symc;
6129 const symconst_attr *pb = &b->attr.symc;
6130 return (pa->kind != pb->kind)
6131 || (pa->sym.type_p != pb->sym.type_p);
6134 /** Compares the attributes of two Call nodes. */
6135 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
6137 const call_attr *pa = &a->attr.call;
6138 const call_attr *pb = &b->attr.call;
6139 return (pa->type != pb->type)
6140 || (pa->tail_call != pb->tail_call);
6143 /** Compares the attributes of two Sel nodes. */
6144 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
6146 const ir_entity *a_ent = get_Sel_entity(a);
6147 const ir_entity *b_ent = get_Sel_entity(b);
6148 return a_ent != b_ent;
6151 /** Compares the attributes of two Phi nodes. */
6152 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
6154 /* we can only enter this function if both nodes have the same number of inputs,
6155 hence it is enough to check if one of them is a Phi0 */
6157 /* check the Phi0 pos attribute */
6158 return a->attr.phi.u.pos != b->attr.phi.u.pos;
6163 /** Compares the attributes of two Conv nodes. */
6164 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
6166 return get_Conv_strict(a) != get_Conv_strict(b);
6169 /** Compares the attributes of two Cast nodes. */
6170 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
6172 return get_Cast_type(a) != get_Cast_type(b);
6175 /** Compares the attributes of two Load nodes. */
6176 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
6178 if (get_Load_volatility(a) == volatility_is_volatile ||
6179 get_Load_volatility(b) == volatility_is_volatile)
6180 /* NEVER do CSE on volatile Loads */
6182 /* do not CSE Loads with different alignment. Be conservative. */
6183 if (get_Load_unaligned(a) != get_Load_unaligned(b))
6186 return get_Load_mode(a) != get_Load_mode(b);
6189 /** Compares the attributes of two Store nodes. */
6190 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
6192 /* do not CSE Stores with different alignment. Be conservative. */
6193 if (get_Store_unaligned(a) != get_Store_unaligned(b))
6196 /* NEVER do CSE on volatile Stores */
6197 return (get_Store_volatility(a) == volatility_is_volatile ||
6198 get_Store_volatility(b) == volatility_is_volatile);
6201 /** Compares two exception attributes */
6202 static int node_cmp_exception(const ir_node *a, const ir_node *b)
6204 const except_attr *ea = &a->attr.except;
6205 const except_attr *eb = &b->attr.except;
6207 return ea->pin_state != eb->pin_state;
6210 #define node_cmp_attr_Bound node_cmp_exception
6212 /** Compares the attributes of two Div nodes. */
6213 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
6215 const div_attr *ma = &a->attr.div;
6216 const div_attr *mb = &b->attr.div;
6217 return ma->exc.pin_state != mb->exc.pin_state ||
6218 ma->resmode != mb->resmode ||
6219 ma->no_remainder != mb->no_remainder;
6222 /** Compares the attributes of two Mod nodes. */
6223 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
6225 const mod_attr *ma = &a->attr.mod;
6226 const mod_attr *mb = &b->attr.mod;
6227 return ma->exc.pin_state != mb->exc.pin_state ||
6228 ma->resmode != mb->resmode;
6231 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
6233 const cmp_attr *ma = &a->attr.cmp;
6234 const cmp_attr *mb = &b->attr.cmp;
6235 return ma->relation != mb->relation;
6238 /** Compares the attributes of two Confirm nodes. */
6239 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
6241 const confirm_attr *ma = &a->attr.confirm;
6242 const confirm_attr *mb = &b->attr.confirm;
6243 return ma->relation != mb->relation;
6246 /** Compares the attributes of two Builtin nodes. */
6247 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
6249 /* no need to compare the type, equal kind means equal type */
6250 return get_Builtin_kind(a) != get_Builtin_kind(b);
6253 /** Compares the attributes of two ASM nodes. */
6254 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6257 const ir_asm_constraint *ca;
6258 const ir_asm_constraint *cb;
6261 if (get_ASM_text(a) != get_ASM_text(b))
6264 /* Should we really check the constraints here? Should be better, but is strange. */
6265 n = get_ASM_n_input_constraints(a);
6266 if (n != get_ASM_n_input_constraints(b))
6269 ca = get_ASM_input_constraints(a);
6270 cb = get_ASM_input_constraints(b);
6271 for (i = 0; i < n; ++i) {
6272 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6273 || ca[i].mode != cb[i].mode)
6277 n = get_ASM_n_output_constraints(a);
6278 if (n != get_ASM_n_output_constraints(b))
6281 ca = get_ASM_output_constraints(a);
6282 cb = get_ASM_output_constraints(b);
6283 for (i = 0; i < n; ++i) {
6284 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6285 || ca[i].mode != cb[i].mode)
6289 n = get_ASM_n_clobbers(a);
6290 if (n != get_ASM_n_clobbers(b))
6293 cla = get_ASM_clobbers(a);
6294 clb = get_ASM_clobbers(b);
6295 for (i = 0; i < n; ++i) {
6296 if (cla[i] != clb[i])
6302 /** Compares the inexistent attributes of two Dummy nodes. */
6303 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6311 * Set the default node attribute compare operation for an ir_op_ops.
6313 * @param code the opcode for the default operation
6314 * @param ops the operations initialized
6319 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6323 ops->node_cmp_attr = node_cmp_attr_##a; \
6355 } /* firm_set_default_node_cmp_attr */
6358 * Compare function for two nodes in the value table. Gets two
6359 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6361 int identities_cmp(const void *elt, const void *key)
6363 ir_node *a = (ir_node *)elt;
6364 ir_node *b = (ir_node *)key;
6367 if (a == b) return 0;
6369 if ((get_irn_op(a) != get_irn_op(b)) ||
6370 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6372 /* compare if a's in and b's in are of equal length */
6373 irn_arity_a = get_irn_arity(a);
6374 if (irn_arity_a != get_irn_arity(b))
6377 /* blocks are never the same */
6381 if (get_irn_pinned(a) == op_pin_state_pinned) {
6382 /* for pinned nodes, the block inputs must be equal */
6383 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6386 ir_node *block_a = get_nodes_block(a);
6387 ir_node *block_b = get_nodes_block(b);
6388 if (! get_opt_global_cse()) {
6389 /* for block-local CSE both nodes must be in the same Block */
6390 if (block_a != block_b)
6393 /* The optimistic approach would be to do nothing here.
6394 * However doing GCSE optimistically produces a lot of partially dead code which appears
6395 * to be worse in practice than the missed opportunities.
6396 * So we use a very conservative variant here and only CSE if 1 value dominates the
6398 if (!block_dominates(block_a, block_b)
6399 && !block_dominates(block_b, block_a))
6404 /* compare a->in[0..ins] with b->in[0..ins] */
6405 for (i = 0; i < irn_arity_a; ++i) {
6406 ir_node *pred_a = get_irn_n(a, i);
6407 ir_node *pred_b = get_irn_n(b, i);
6408 if (pred_a != pred_b) {
6409 /* if both predecessors are CSE neutral they might be different */
6410 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6416 * here, we already now that the nodes are identical except their
6419 if (a->op->ops.node_cmp_attr)
6420 return a->op->ops.node_cmp_attr(a, b);
6423 } /* identities_cmp */
6426 * Calculate a hash value of a node.
6428 * @param node The IR-node
6430 unsigned ir_node_hash(const ir_node *node)
6432 return node->op->ops.hash(node);
6433 } /* ir_node_hash */
6436 void new_identities(ir_graph *irg)
6438 if (irg->value_table != NULL)
6439 del_pset(irg->value_table);
6440 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6441 } /* new_identities */
6443 void del_identities(ir_graph *irg)
6445 if (irg->value_table != NULL)
6446 del_pset(irg->value_table);
6447 } /* del_identities */
6449 /* Normalize a node by putting constants (and operands with larger
6450 * node index) on the right (operator side). */
6451 void ir_normalize_node(ir_node *n)
6453 if (is_op_commutative(get_irn_op(n))) {
6454 ir_node *l = get_binop_left(n);
6455 ir_node *r = get_binop_right(n);
6457 /* For commutative operators perform a OP b == b OP a but keep
6458 * constants on the RIGHT side. This helps greatly in some
6459 * optimizations. Moreover we use the idx number to make the form
6461 if (!operands_are_normalized(l, r)) {
6462 set_binop_left(n, r);
6463 set_binop_right(n, l);
6467 } /* ir_normalize_node */
6470 * Return the canonical node computing the same value as n.
6471 * Looks up the node in a hash table, enters it in the table
6472 * if it isn't there yet.
6474 * @param n the node to look up
6476 * @return a node that computes the same value as n or n if no such
6477 * node could be found
6479 ir_node *identify_remember(ir_node *n)
6481 ir_graph *irg = get_irn_irg(n);
6482 pset *value_table = irg->value_table;
6485 if (value_table == NULL)
6488 ir_normalize_node(n);
6489 /* lookup or insert in hash table with given hash key. */
6490 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6493 /* n is reachable again */
6494 edges_node_revival(nn);
6498 } /* identify_remember */
6501 * During construction we set the op_pin_state_pinned flag in the graph right
6502 * when the optimization is performed. The flag turning on procedure global
6503 * cse could be changed between two allocations. This way we are safe.
6505 * @param n The node to lookup
6507 static inline ir_node *identify_cons(ir_node *n)
6511 n = identify_remember(n);
6512 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6513 ir_graph *irg = get_irn_irg(n);
6514 set_irg_pinned(irg, op_pin_state_floats);
6517 } /* identify_cons */
6519 /* Add a node to the identities value table. */
6520 void add_identities(ir_node *node)
6527 identify_remember(node);
6530 /* Visit each node in the value table of a graph. */
6531 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6534 ir_graph *rem = current_ir_graph;
6536 current_ir_graph = irg;
6537 foreach_pset(irg->value_table, ir_node*, node) {
6540 current_ir_graph = rem;
6541 } /* visit_all_identities */
6544 * These optimizations deallocate nodes from the obstack.
6545 * It can only be called if it is guaranteed that no other nodes
6546 * reference this one, i.e., right after construction of a node.
6548 * @param n The node to optimize
6550 ir_node *optimize_node(ir_node *n)
6553 ir_graph *irg = get_irn_irg(n);
6554 unsigned iro = get_irn_opcode(n);
6557 /* Always optimize Phi nodes: part of the construction. */
6558 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6560 /* constant expression evaluation / constant folding */
6561 if (get_opt_constant_folding()) {
6562 /* neither constants nor Tuple values can be evaluated */
6563 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6564 /* try to evaluate */
6565 tv = computed_value(n);
6566 if (tv != tarval_bad) {
6571 * we MUST copy the node here temporarily, because it's still
6572 * needed for DBG_OPT_CSTEVAL
6574 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6575 oldn = (ir_node*)alloca(node_size);
6577 memcpy(oldn, n, node_size);
6578 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6580 /* ARG, copy the in array, we need it for statistics */
6581 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6583 /* note the inplace edges module */
6584 edges_node_deleted(n);
6586 /* evaluation was successful -- replace the node. */
6587 irg_kill_node(irg, n);
6588 nw = new_r_Const(irg, tv);
6590 DBG_OPT_CSTEVAL(oldn, nw);
6596 /* remove unnecessary nodes */
6597 if (get_opt_algebraic_simplification() ||
6598 (iro == iro_Phi) || /* always optimize these nodes. */
6600 (iro == iro_Proj) ||
6601 (iro == iro_Block) ) /* Flags tested local. */
6602 n = equivalent_node(n);
6604 /* Common Subexpression Elimination.
6606 * Checks whether n is already available.
6607 * The block input is used to distinguish different subexpressions. Right
6608 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6609 * subexpressions within a block.
6612 n = identify_cons(n);
6615 edges_node_deleted(oldn);
6617 /* We found an existing, better node, so we can deallocate the old node. */
6618 irg_kill_node(irg, oldn);
6622 /* Some more constant expression evaluation that does not allow to
6624 iro = get_irn_opcode(n);
6625 if (get_opt_algebraic_simplification() ||
6626 (iro == iro_Cond) ||
6627 (iro == iro_Proj)) /* Flags tested local. */
6628 n = transform_node(n);
6630 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6631 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6633 n = identify_remember(o);
6639 } /* optimize_node */
6643 * These optimizations never deallocate nodes (in place). This can cause dead
6644 * nodes lying on the obstack. Remove these by a dead node elimination,
6645 * i.e., a copying garbage collection.
6647 ir_node *optimize_in_place_2(ir_node *n)
6651 unsigned iro = get_irn_opcode(n);
6653 if (!get_opt_optimize() && !is_Phi(n)) return n;
6655 if (iro == iro_Deleted)
6658 /* constant expression evaluation / constant folding */
6659 if (get_opt_constant_folding()) {
6660 /* neither constants nor Tuple values can be evaluated */
6661 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6662 /* try to evaluate */
6663 tv = computed_value(n);
6664 if (tv != tarval_bad) {
6665 /* evaluation was successful -- replace the node. */
6666 ir_graph *irg = get_irn_irg(n);
6668 n = new_r_Const(irg, tv);
6670 DBG_OPT_CSTEVAL(oldn, n);
6676 /* remove unnecessary nodes */
6677 if (get_opt_constant_folding() ||
6678 (iro == iro_Phi) || /* always optimize these nodes. */
6679 (iro == iro_Id) || /* ... */
6680 (iro == iro_Proj) || /* ... */
6681 (iro == iro_Block) ) /* Flags tested local. */
6682 n = equivalent_node(n);
6684 /** common subexpression elimination **/
6685 /* Checks whether n is already available. */
6686 /* The block input is used to distinguish different subexpressions. Right
6687 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6688 subexpressions within a block. */
6689 if (get_opt_cse()) {
6691 n = identify_remember(o);
6696 /* Some more constant expression evaluation. */
6697 iro = get_irn_opcode(n);
6698 if (get_opt_constant_folding() ||
6699 (iro == iro_Cond) ||
6700 (iro == iro_Proj)) /* Flags tested local. */
6701 n = transform_node(n);
6703 /* Now we can verify the node, as it has no dead inputs any more. */
6706 /* Now we have a legal, useful node. Enter it in hash table for cse.
6707 Blocks should be unique anyways. (Except the successor of start:
6708 is cse with the start block!) */
6709 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6711 n = identify_remember(o);
6717 } /* optimize_in_place_2 */
6720 * Wrapper for external use, set proper status bits after optimization.
6722 ir_node *optimize_in_place(ir_node *n)
6724 ir_graph *irg = get_irn_irg(n);
6725 /* Handle graph state */
6726 assert(get_irg_phase_state(irg) != phase_building);
6728 if (get_opt_global_cse())
6729 set_irg_pinned(irg, op_pin_state_floats);
6731 /* FIXME: Maybe we could also test whether optimizing the node can
6732 change the control graph. */
6733 set_irg_doms_inconsistent(irg);
6734 return optimize_in_place_2(n);
6735 } /* optimize_in_place */
6738 * Calculate a hash value of a Const node.
6740 static unsigned hash_Const(const ir_node *node)
6744 /* special value for const, as they only differ in their tarval. */
6745 h = HASH_PTR(node->attr.con.tarval);
6751 * Calculate a hash value of a SymConst node.
6753 static unsigned hash_SymConst(const ir_node *node)
6757 /* all others are pointers */
6758 h = HASH_PTR(node->attr.symc.sym.type_p);
6761 } /* hash_SymConst */
6764 * Set the default hash operation in an ir_op_ops.
6766 * @param code the opcode for the default operation
6767 * @param ops the operations initialized
6772 static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops)
6776 ops->hash = hash_##a; \
6779 /* hash function already set */
6780 if (ops->hash != NULL)
6787 /* use input/mode default hash if no function was given */
6788 ops->hash = firm_default_hash;
6796 * Sets the default operation for an ir_ops.
6798 ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops)
6800 ops = firm_set_default_hash(code, ops);
6801 ops = firm_set_default_computed_value(code, ops);
6802 ops = firm_set_default_equivalent_node(code, ops);
6803 ops = firm_set_default_transform_node(code, ops);
6804 ops = firm_set_default_node_cmp_attr(code, ops);
6805 ops = firm_set_default_get_type_attr(code, ops);
6806 ops = firm_set_default_get_entity_attr(code, ops);
6809 } /* firm_set_default_operations */