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 && get_irn_mode(a) == mode_b) {
2793 /* It's a boolean Cond, branching on a boolean constant.
2794 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2795 ir_node *blk = get_nodes_block(n);
2796 jmp = new_r_Jmp(blk);
2797 turn_into_tuple(n, pn_Cond_max+1);
2798 if (ta == tarval_b_true) {
2799 set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
2800 set_Tuple_pred(n, pn_Cond_true, jmp);
2802 set_Tuple_pred(n, pn_Cond_false, jmp);
2803 set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X));
2805 /* We might generate an endless loop, so keep it alive. */
2806 add_End_keepalive(get_irg_end(irg), blk);
2807 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
2810 } /* transform_node_Cond */
2813 * Prototype of a recursive transform function
2814 * for bitwise distributive transformations.
2816 typedef ir_node* (*recursive_transform)(ir_node *n);
2819 * makes use of distributive laws for and, or, eor
2820 * and(a OP c, b OP c) -> and(a, b) OP c
2821 * note, might return a different op than n
2823 static ir_node *transform_bitwise_distributive(ir_node *n,
2824 recursive_transform trans_func)
2827 ir_node *a = get_binop_left(n);
2828 ir_node *b = get_binop_right(n);
2829 ir_op *op = get_irn_op(a);
2830 ir_op *op_root = get_irn_op(n);
2832 if (op != get_irn_op(b))
2835 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2836 if (op == op_Conv) {
2837 ir_node *a_op = get_Conv_op(a);
2838 ir_node *b_op = get_Conv_op(b);
2839 ir_mode *a_mode = get_irn_mode(a_op);
2840 ir_mode *b_mode = get_irn_mode(b_op);
2841 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2842 ir_node *blk = get_nodes_block(n);
2845 set_binop_left(n, a_op);
2846 set_binop_right(n, b_op);
2847 set_irn_mode(n, a_mode);
2849 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2851 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2857 /* nothing to gain here */
2861 if (op == op_Shrs || op == op_Shr || op == op_Shl
2862 || op == op_And || op == op_Or || op == op_Eor) {
2863 ir_node *a_left = get_binop_left(a);
2864 ir_node *a_right = get_binop_right(a);
2865 ir_node *b_left = get_binop_left(b);
2866 ir_node *b_right = get_binop_right(b);
2868 ir_node *op1 = NULL;
2869 ir_node *op2 = NULL;
2871 if (is_op_commutative(op)) {
2872 if (a_left == b_left) {
2876 } else if (a_left == b_right) {
2880 } else if (a_right == b_left) {
2886 if (a_right == b_right) {
2893 /* (a sop c) & (b sop c) => (a & b) sop c */
2894 ir_node *blk = get_nodes_block(n);
2896 ir_node *new_n = exact_copy(n);
2897 set_binop_left(new_n, op1);
2898 set_binop_right(new_n, op2);
2899 new_n = trans_func(new_n);
2901 if (op_root == op_Eor && op == op_Or) {
2902 dbg_info *dbgi = get_irn_dbg_info(n);
2903 ir_mode *mode = get_irn_mode(c);
2905 c = new_rd_Not(dbgi, blk, c, mode);
2906 n = new_rd_And(dbgi, blk, new_n, c, mode);
2909 set_nodes_block(n, blk);
2910 set_binop_left(n, new_n);
2911 set_binop_right(n, c);
2915 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2924 * Create a 0 constant of given mode.
2926 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
2928 ir_tarval *tv = get_mode_null(mode);
2929 ir_node *cnst = new_r_Const(irg, tv);
2934 static bool is_shiftop(const ir_node *n)
2936 return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
2940 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
2942 * - and, or, xor instead of &
2943 * - Shl, Shr, Shrs, rotl instead of >>
2944 * (with a special case for Or/Xor + Shrs)
2946 * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
2948 static ir_node *transform_node_shift_bitop(ir_node *n)
2950 ir_graph *irg = get_irn_irg(n);
2951 ir_node *right = get_binop_right(n);
2952 ir_mode *mode = get_irn_mode(n);
2954 ir_node *bitop_left;
2955 ir_node *bitop_right;
2964 ir_tarval *tv_shift;
2966 if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
2969 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
2971 if (!is_Const(right))
2974 left = get_binop_left(n);
2975 op_left = get_irn_op(left);
2976 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
2979 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
2980 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
2981 /* TODO: test if sign bit is affectes */
2985 bitop_right = get_binop_right(left);
2986 if (!is_Const(bitop_right))
2989 bitop_left = get_binop_left(left);
2991 block = get_nodes_block(n);
2992 dbgi = get_irn_dbg_info(n);
2993 tv1 = get_Const_tarval(bitop_right);
2994 tv2 = get_Const_tarval(right);
2996 assert(get_tarval_mode(tv1) == mode);
2999 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
3000 tv_shift = tarval_shl(tv1, tv2);
3001 } else if (is_Shr(n)) {
3002 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
3003 tv_shift = tarval_shr(tv1, tv2);
3004 } else if (is_Shrs(n)) {
3005 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
3006 tv_shift = tarval_shrs(tv1, tv2);
3009 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
3010 tv_shift = tarval_rotl(tv1, tv2);
3013 assert(get_tarval_mode(tv_shift) == mode);
3014 irg = get_irn_irg(n);
3015 new_const = new_r_Const(irg, tv_shift);
3017 if (op_left == op_And) {
3018 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
3019 } else if (op_left == op_Or) {
3020 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
3022 assert(op_left == op_Eor);
3023 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
3030 * normalisation: (x >> c1) & c2 to (x & (c2<<c1)) >> c1
3032 * - and, or, xor instead of &
3033 * - Shl, Shr, Shrs, rotl instead of >>
3034 * (with a special case for Or/Xor + Shrs)
3036 * This normalisation is usually good for the backend since << C can often be
3037 * matched as address-mode.
3039 static ir_node *transform_node_bitop_shift(ir_node *n)
3041 ir_graph *irg = get_irn_irg(n);
3042 ir_node *left = get_binop_left(n);
3043 ir_node *right = get_binop_right(n);
3044 ir_mode *mode = get_irn_mode(n);
3045 ir_node *shift_left;
3046 ir_node *shift_right;
3048 dbg_info *dbg_bitop;
3049 dbg_info *dbg_shift;
3055 ir_tarval *tv_bitop;
3057 if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
3060 assert(is_And(n) || is_Or(n) || is_Eor(n));
3061 if (!is_Const(right) || !is_shiftop(left))
3064 shift_left = get_binop_left(left);
3065 shift_right = get_binop_right(left);
3066 if (!is_Const(shift_right))
3069 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
3070 if (is_Shrs(left)) {
3071 /* TODO this could be improved */
3075 irg = get_irn_irg(n);
3076 block = get_nodes_block(n);
3077 dbg_bitop = get_irn_dbg_info(n);
3078 dbg_shift = get_irn_dbg_info(left);
3079 tv1 = get_Const_tarval(shift_right);
3080 tv2 = get_Const_tarval(right);
3081 assert(get_tarval_mode(tv2) == mode);
3084 tv_bitop = tarval_shr(tv2, tv1);
3085 } else if (is_Shr(left)) {
3086 if (is_Or(n) || is_Eor(n)) {
3088 * TODO this can be improved by checking whether
3089 * the left shift produces an overflow
3093 tv_bitop = tarval_shl(tv2, tv1);
3095 assert(is_Rotl(left));
3096 tv_bitop = tarval_rotl(tv2, tarval_neg(tv1));
3098 new_const = new_r_Const(irg, tv_bitop);
3101 new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode);
3102 } else if (is_Or(n)) {
3103 new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode);
3106 new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode);
3110 new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode);
3111 } else if (is_Shr(left)) {
3112 new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode);
3114 assert(is_Rotl(left));
3115 new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode);
3124 static ir_node *transform_node_And(ir_node *n)
3126 ir_node *c, *oldn = n;
3127 ir_node *a = get_And_left(n);
3128 ir_node *b = get_And_right(n);
3130 vrp_attr *a_vrp, *b_vrp;
3132 if (is_Cmp(a) && is_Cmp(b)) {
3133 ir_node *a_left = get_Cmp_left(a);
3134 ir_node *a_right = get_Cmp_right(a);
3135 ir_node *b_left = get_Cmp_left(b);
3136 ir_node *b_right = get_Cmp_right(b);
3137 ir_relation a_relation = get_Cmp_relation(a);
3138 ir_relation b_relation = get_Cmp_relation(b);
3139 /* we can combine the relations of two compares with the same
3141 if (a_left == b_left && b_left == b_right) {
3142 dbg_info *dbgi = get_irn_dbg_info(n);
3143 ir_node *block = get_nodes_block(n);
3144 ir_relation new_relation = a_relation & b_relation;
3145 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3147 /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */
3148 if (a_relation == b_relation && a_relation == ir_relation_equal
3149 && !mode_is_float(get_irn_mode(a_left))
3150 && !mode_is_float(get_irn_mode(b_left))) {
3151 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) {
3152 dbg_info *dbgi = get_irn_dbg_info(n);
3153 ir_node *block = get_nodes_block(n);
3154 ir_mode *a_mode = get_irn_mode(a_left);
3155 ir_mode *b_mode = get_irn_mode(b_left);
3156 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3157 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3158 ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode);
3159 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
3160 ir_graph *irg = get_irn_irg(n);
3161 ir_node *zero = create_zero_const(irg, b_mode);
3162 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3164 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
3165 dbg_info *dbgi = get_irn_dbg_info(n);
3166 ir_node *block = get_nodes_block(n);
3167 ir_mode *a_mode = get_irn_mode(a_left);
3168 ir_mode *b_mode = get_irn_mode(b_left);
3169 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
3170 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
3171 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
3172 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
3173 ir_graph *irg = get_irn_irg(n);
3174 ir_node *zero = create_zero_const(irg, a_mode);
3175 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3180 mode = get_irn_mode(n);
3181 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3185 ir_node *op = get_Not_op(b);
3187 ir_node *ba = get_And_left(op);
3188 ir_node *bb = get_And_right(op);
3190 /* it's enough to test the following cases due to normalization! */
3191 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3192 /* (a|b) & ~(a&b) = a^b */
3193 ir_node *block = get_nodes_block(n);
3195 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3196 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3204 ir_node *op = get_Not_op(a);
3206 ir_node *aa = get_And_left(op);
3207 ir_node *ab = get_And_right(op);
3209 /* it's enough to test the following cases due to normalization! */
3210 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3211 /* (a|b) & ~(a&b) = a^b */
3212 ir_node *block = get_nodes_block(n);
3214 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3215 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3222 ir_node *al = get_Eor_left(a);
3223 ir_node *ar = get_Eor_right(a);
3226 /* (b ^ a) & b -> ~a & b */
3227 dbg_info *dbg = get_irn_dbg_info(n);
3228 ir_node *block = get_nodes_block(n);
3230 ar = new_rd_Not(dbg, block, ar, mode);
3231 n = new_rd_And(dbg, block, ar, b, mode);
3232 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3236 /* (a ^ b) & b -> ~a & b */
3237 dbg_info *dbg = get_irn_dbg_info(n);
3238 ir_node *block = get_nodes_block(n);
3240 al = new_rd_Not(dbg, block, al, mode);
3241 n = new_rd_And(dbg, block, al, b, mode);
3242 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3247 ir_node *bl = get_Eor_left(b);
3248 ir_node *br = get_Eor_right(b);
3251 /* a & (a ^ b) -> a & ~b */
3252 dbg_info *dbg = get_irn_dbg_info(n);
3253 ir_node *block = get_nodes_block(n);
3255 br = new_rd_Not(dbg, block, br, mode);
3256 n = new_rd_And(dbg, block, br, a, mode);
3257 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3261 /* a & (b ^ a) -> a & ~b */
3262 dbg_info *dbg = get_irn_dbg_info(n);
3263 ir_node *block = get_nodes_block(n);
3265 bl = new_rd_Not(dbg, block, bl, mode);
3266 n = new_rd_And(dbg, block, bl, a, mode);
3267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3271 if (is_Not(a) && is_Not(b)) {
3272 /* ~a & ~b = ~(a|b) */
3273 ir_node *block = get_nodes_block(n);
3274 ir_mode *mode = get_irn_mode(n);
3278 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3279 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3284 b_vrp = vrp_get_info(b);
3285 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3286 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3292 a_vrp = vrp_get_info(a);
3293 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3294 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3298 n = transform_bitwise_distributive(n, transform_node_And);
3300 n = transform_node_bitop_shift(n);
3303 } /* transform_node_And */
3305 /* the order of the values is important! */
3306 typedef enum const_class {
3312 static const_class classify_const(const ir_node* n)
3314 if (is_Const(n)) return const_const;
3315 if (is_irn_constlike(n)) return const_like;
3320 * Determines whether r is more constlike or has a larger index (in that order)
3323 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3325 const const_class l_order = classify_const(l);
3326 const const_class r_order = classify_const(r);
3328 l_order > r_order ||
3329 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3335 static ir_node *transform_node_Eor(ir_node *n)
3337 ir_node *c, *oldn = n;
3338 ir_node *a = get_Eor_left(n);
3339 ir_node *b = get_Eor_right(n);
3340 ir_mode *mode = get_irn_mode(n);
3342 /* we can combine the relations of two compares with the same operands */
3343 if (is_Cmp(a) && is_Cmp(b)) {
3344 ir_node *a_left = get_Cmp_left(a);
3345 ir_node *a_right = get_Cmp_left(a);
3346 ir_node *b_left = get_Cmp_left(b);
3347 ir_node *b_right = get_Cmp_right(b);
3348 if (a_left == b_left && b_left == b_right) {
3349 dbg_info *dbgi = get_irn_dbg_info(n);
3350 ir_node *block = get_nodes_block(n);
3351 ir_relation a_relation = get_Cmp_relation(a);
3352 ir_relation b_relation = get_Cmp_relation(b);
3353 ir_relation new_relation = a_relation ^ b_relation;
3354 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3358 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3360 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3361 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3362 dbg_info *dbg = get_irn_dbg_info(n);
3363 ir_node *block = get_nodes_block(n);
3364 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3365 ir_node *new_left = get_Not_op(a);
3366 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3367 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3369 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3370 dbg_info *dbg = get_irn_dbg_info(n);
3371 ir_node *block = get_nodes_block(n);
3372 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3373 ir_node *new_right = get_Not_op(b);
3374 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3375 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3379 /* x ^ 1...1 -> ~1 */
3380 if (is_Const(b) && is_Const_all_one(b)) {
3381 n = new_r_Not(get_nodes_block(n), a, mode);
3382 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3386 n = transform_bitwise_distributive(n, transform_node_Eor);
3388 n = transform_node_bitop_shift(n);
3391 } /* transform_node_Eor */
3396 static ir_node *transform_node_Not(ir_node *n)
3398 ir_node *c, *oldn = n;
3399 ir_node *a = get_Not_op(n);
3400 ir_mode *mode = get_irn_mode(n);
3402 HANDLE_UNOP_PHI(tarval_not,a,c);
3404 /* check for a boolean Not */
3406 dbg_info *dbgi = get_irn_dbg_info(a);
3407 ir_node *block = get_nodes_block(a);
3408 ir_relation relation = get_Cmp_relation(a);
3409 relation = get_negated_relation(relation);
3410 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3411 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3415 /* normalize ~(a ^ b) => a ^ ~b */
3417 dbg_info *dbg = get_irn_dbg_info(n);
3418 ir_node *block = get_nodes_block(n);
3419 ir_node *eor_right = get_Eor_right(a);
3420 ir_node *eor_left = get_Eor_left(a);
3421 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3422 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3426 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3427 if (is_Minus(a)) { /* ~-x -> x + -1 */
3428 dbg_info *dbg = get_irn_dbg_info(n);
3429 ir_graph *irg = get_irn_irg(n);
3430 ir_node *block = get_nodes_block(n);
3431 ir_node *add_l = get_Minus_op(a);
3432 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3433 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3434 } else if (is_Add(a)) {
3435 ir_node *add_r = get_Add_right(a);
3436 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3437 /* ~(x + -1) = -x */
3438 ir_node *op = get_Add_left(a);
3439 ir_node *blk = get_nodes_block(n);
3440 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3449 * Transform a Minus.
3453 * -(a >>u (size-1)) = a >>s (size-1)
3454 * -(a >>s (size-1)) = a >>u (size-1)
3455 * -(a * const) -> a * -const
3457 static ir_node *transform_node_Minus(ir_node *n)
3459 ir_node *c, *oldn = n;
3460 ir_node *a = get_Minus_op(n);
3463 HANDLE_UNOP_PHI(tarval_neg,a,c);
3465 mode = get_irn_mode(a);
3466 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3467 /* the following rules are only to twos-complement */
3470 ir_node *op = get_Not_op(a);
3471 ir_tarval *tv = get_mode_one(mode);
3472 ir_node *blk = get_nodes_block(n);
3473 ir_graph *irg = get_irn_irg(blk);
3474 ir_node *c = new_r_Const(irg, tv);
3475 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3476 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3480 ir_node *c = get_Shr_right(a);
3483 ir_tarval *tv = get_Const_tarval(c);
3485 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3486 /* -(a >>u (size-1)) = a >>s (size-1) */
3487 ir_node *v = get_Shr_left(a);
3489 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3490 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3496 ir_node *c = get_Shrs_right(a);
3499 ir_tarval *tv = get_Const_tarval(c);
3501 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3502 /* -(a >>s (size-1)) = a >>u (size-1) */
3503 ir_node *v = get_Shrs_left(a);
3505 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3506 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3513 /* - (a-b) = b - a */
3514 ir_node *la = get_Sub_left(a);
3515 ir_node *ra = get_Sub_right(a);
3516 ir_node *blk = get_nodes_block(n);
3518 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3519 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3523 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3524 ir_node *mul_l = get_Mul_left(a);
3525 ir_node *mul_r = get_Mul_right(a);
3526 ir_tarval *tv = value_of(mul_r);
3527 if (tv != tarval_bad) {
3528 tv = tarval_neg(tv);
3529 if (tv != tarval_bad) {
3530 ir_graph *irg = get_irn_irg(n);
3531 ir_node *cnst = new_r_Const(irg, tv);
3532 dbg_info *dbg = get_irn_dbg_info(a);
3533 ir_node *block = get_nodes_block(a);
3534 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3535 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3542 } /* transform_node_Minus */
3545 * Transform a Proj(Load) with a non-null address.
3547 static ir_node *transform_node_Proj_Load(ir_node *proj)
3549 if (get_opt_ldst_only_null_ptr_exceptions()) {
3550 if (get_irn_mode(proj) == mode_X) {
3551 ir_node *load = get_Proj_pred(proj);
3553 /* get the Load address */
3554 const ir_node *addr = get_Load_ptr(load);
3555 const ir_node *confirm;
3557 if (value_not_null(addr, &confirm)) {
3558 if (confirm == NULL) {
3559 /* this node may float if it did not depend on a Confirm */
3560 set_irn_pinned(load, op_pin_state_floats);
3562 if (get_Proj_proj(proj) == pn_Load_X_except) {
3563 ir_graph *irg = get_irn_irg(proj);
3564 DBG_OPT_EXC_REM(proj);
3565 return new_r_Bad(irg, mode_X);
3567 ir_node *blk = get_nodes_block(load);
3568 return new_r_Jmp(blk);
3574 } /* transform_node_Proj_Load */
3577 * Transform a Proj(Store) with a non-null address.
3579 static ir_node *transform_node_Proj_Store(ir_node *proj)
3581 if (get_opt_ldst_only_null_ptr_exceptions()) {
3582 if (get_irn_mode(proj) == mode_X) {
3583 ir_node *store = get_Proj_pred(proj);
3585 /* get the load/store address */
3586 const ir_node *addr = get_Store_ptr(store);
3587 const ir_node *confirm;
3589 if (value_not_null(addr, &confirm)) {
3590 if (confirm == NULL) {
3591 /* this node may float if it did not depend on a Confirm */
3592 set_irn_pinned(store, op_pin_state_floats);
3594 if (get_Proj_proj(proj) == pn_Store_X_except) {
3595 ir_graph *irg = get_irn_irg(proj);
3596 DBG_OPT_EXC_REM(proj);
3597 return new_r_Bad(irg, mode_X);
3599 ir_node *blk = get_nodes_block(store);
3600 return new_r_Jmp(blk);
3606 } /* transform_node_Proj_Store */
3609 * Transform a Proj(Div) with a non-zero value.
3610 * Removes the exceptions and routes the memory to the NoMem node.
3612 static ir_node *transform_node_Proj_Div(ir_node *proj)
3614 ir_node *div = get_Proj_pred(proj);
3615 ir_node *b = get_Div_right(div);
3616 ir_node *res, *new_mem;
3617 const ir_node *confirm;
3620 if (value_not_zero(b, &confirm)) {
3621 /* div(x, y) && y != 0 */
3622 if (confirm == NULL) {
3623 /* we are sure we have a Const != 0 */
3624 new_mem = get_Div_mem(div);
3625 new_mem = skip_Pin(new_mem);
3626 set_Div_mem(div, new_mem);
3627 set_irn_pinned(div, op_pin_state_floats);
3630 proj_nr = get_Proj_proj(proj);
3632 case pn_Div_X_regular:
3633 return new_r_Jmp(get_nodes_block(div));
3635 case pn_Div_X_except: {
3636 ir_graph *irg = get_irn_irg(proj);
3637 /* we found an exception handler, remove it */
3638 DBG_OPT_EXC_REM(proj);
3639 return new_r_Bad(irg, mode_X);
3643 ir_graph *irg = get_irn_irg(proj);
3644 res = get_Div_mem(div);
3645 new_mem = get_irg_no_mem(irg);
3648 /* This node can only float up to the Confirm block */
3649 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3651 set_irn_pinned(div, op_pin_state_floats);
3652 /* this is a Div without exception, we can remove the memory edge */
3653 set_Div_mem(div, new_mem);
3659 } /* transform_node_Proj_Div */
3662 * Transform a Proj(Mod) with a non-zero value.
3663 * Removes the exceptions and routes the memory to the NoMem node.
3665 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3667 ir_node *mod = get_Proj_pred(proj);
3668 ir_node *b = get_Mod_right(mod);
3669 ir_node *res, *new_mem;
3670 const ir_node *confirm;
3673 if (value_not_zero(b, &confirm)) {
3674 /* mod(x, y) && y != 0 */
3675 proj_nr = get_Proj_proj(proj);
3677 if (confirm == NULL) {
3678 /* we are sure we have a Const != 0 */
3679 new_mem = get_Mod_mem(mod);
3680 new_mem = skip_Pin(new_mem);
3681 set_Mod_mem(mod, new_mem);
3682 set_irn_pinned(mod, op_pin_state_floats);
3687 case pn_Mod_X_regular:
3688 return new_r_Jmp(get_irn_n(mod, -1));
3690 case pn_Mod_X_except: {
3691 ir_graph *irg = get_irn_irg(proj);
3692 /* we found an exception handler, remove it */
3693 DBG_OPT_EXC_REM(proj);
3694 return new_r_Bad(irg, mode_X);
3698 ir_graph *irg = get_irn_irg(proj);
3699 res = get_Mod_mem(mod);
3700 new_mem = get_irg_no_mem(irg);
3703 /* This node can only float up to the Confirm block */
3704 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3706 /* this is a Mod without exception, we can remove the memory edge */
3707 set_Mod_mem(mod, new_mem);
3711 if (get_Mod_left(mod) == b) {
3712 /* a % a = 0 if a != 0 */
3713 ir_graph *irg = get_irn_irg(proj);
3714 ir_mode *mode = get_irn_mode(proj);
3715 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3717 DBG_OPT_CSTEVAL(mod, res);
3723 } /* transform_node_Proj_Mod */
3726 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3728 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3730 ir_node *n = get_Proj_pred(proj);
3731 ir_node *b = get_Cond_selector(n);
3733 if (mode_is_int(get_irn_mode(b))) {
3734 ir_tarval *tb = value_of(b);
3736 if (tb != tarval_bad) {
3737 /* we have a constant switch */
3738 long num = get_Proj_proj(proj);
3740 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3741 if (get_tarval_long(tb) == num) {
3742 /* Do NOT create a jump here, or we will have 2 control flow ops
3743 * in a block. This case is optimized away in optimize_cf(). */
3746 ir_graph *irg = get_irn_irg(proj);
3747 /* this case will NEVER be taken, kill it */
3748 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
3749 return new_r_Bad(irg, mode_X);
3753 long num = get_Proj_proj(proj);
3754 vrp_attr *b_vrp = vrp_get_info(b);
3755 if (num != get_Cond_default_proj(n) && b_vrp) {
3756 /* Try handling with vrp data. We only remove dead parts. */
3757 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3759 if (b_vrp->range_type == VRP_RANGE) {
3760 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3761 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3763 if ((cmp_result & ir_relation_greater) == cmp_result
3764 && (cmp_result2 & ir_relation_less) == cmp_result2) {
3765 ir_graph *irg = get_irn_irg(proj);
3766 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
3767 return new_r_Bad(irg, mode_X);
3769 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3770 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3771 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3773 if ((cmp_result & ir_relation_less_equal) == cmp_result
3774 && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) {
3775 ir_graph *irg = get_irn_irg(proj);
3776 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
3777 return new_r_Bad(irg, mode_X);
3782 tarval_and( b_vrp->bits_set, tp),
3784 ) == ir_relation_equal)) {
3785 ir_graph *irg = get_irn_irg(proj);
3786 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
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 clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE);
3798 return new_r_Bad(irg, mode_X);
3807 * return true if the operation returns a value with exactly 1 bit set
3809 static bool is_single_bit(const ir_node *node)
3811 /* a first implementation, could be extended with vrp and others... */
3813 ir_node *shl_l = get_Shl_left(node);
3814 ir_mode *mode = get_irn_mode(node);
3815 int modulo = get_mode_modulo_shift(mode);
3816 /* this works if we shift a 1 and we have modulo shift */
3817 if (is_Const(shl_l) && is_Const_one(shl_l)
3818 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3821 } else if (is_Const(node)) {
3822 ir_tarval *tv = get_Const_tarval(node);
3823 return tarval_is_single_bit(tv);
3829 * Normalizes and optimizes Cmp nodes.
3831 static ir_node *transform_node_Cmp(ir_node *n)
3833 ir_node *left = get_Cmp_left(n);
3834 ir_node *right = get_Cmp_right(n);
3835 ir_mode *mode = get_irn_mode(left);
3836 ir_tarval *tv = NULL;
3837 bool changed = false;
3838 bool changedc = false;
3839 ir_relation relation = get_Cmp_relation(n);
3840 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3842 /* mask out impossible relations */
3843 ir_relation new_relation = relation & possible;
3844 if (new_relation != relation) {
3845 relation = new_relation;
3849 /* Remove unnecessary conversions */
3850 if (is_Conv(left) && is_Conv(right)) {
3851 ir_node *op_left = get_Conv_op(left);
3852 ir_node *op_right = get_Conv_op(right);
3853 ir_mode *mode_left = get_irn_mode(op_left);
3854 ir_mode *mode_right = get_irn_mode(op_right);
3856 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3857 && mode_left != mode_b && mode_right != mode_b) {
3858 ir_node *block = get_nodes_block(n);
3860 if (mode_left == mode_right) {
3864 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3865 } else if (smaller_mode(mode_left, mode_right)) {
3866 left = new_r_Conv(block, op_left, mode_right);
3869 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3870 } else if (smaller_mode(mode_right, mode_left)) {
3872 right = new_r_Conv(block, op_right, mode_left);
3874 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3876 mode = get_irn_mode(left);
3879 if (is_Conv(left) && is_Const(right)) {
3880 ir_node *op_left = get_Conv_op(left);
3881 ir_mode *mode_left = get_irn_mode(op_left);
3882 if (smaller_mode(mode_left, mode) && mode_left != mode_b) {
3883 ir_tarval *tv = get_Const_tarval(right);
3884 tarval_int_overflow_mode_t last_mode
3885 = tarval_get_integer_overflow_mode();
3887 tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD);
3888 new_tv = tarval_convert_to(tv, mode_left);
3889 tarval_set_integer_overflow_mode(last_mode);
3890 if (new_tv != tarval_bad) {
3891 ir_graph *irg = get_irn_irg(n);
3893 right = new_r_Const(irg, new_tv);
3894 mode = get_irn_mode(left);
3896 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3902 * Optimize -a CMP -b into b CMP a.
3903 * This works only for modes where unary Minus cannot Overflow.
3904 * Note that two-complement integers can Overflow so it will NOT work.
3906 if (!mode_overflow_on_unary_Minus(mode) &&
3907 is_Minus(left) && is_Minus(right)) {
3908 left = get_Minus_op(left);
3909 right = get_Minus_op(right);
3910 relation = get_inversed_relation(relation);
3912 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3915 /* remove operation on both sides if possible */
3916 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3918 * The following operations are NOT safe for floating point operations, for instance
3919 * 1.0 + inf == 2.0 + inf, =/=> x == y
3921 if (mode_is_int(mode)) {
3922 unsigned lop = get_irn_opcode(left);
3924 if (lop == get_irn_opcode(right)) {
3925 ir_node *ll, *lr, *rl, *rr;
3927 /* same operation on both sides, try to remove */
3931 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3932 left = get_unop_op(left);
3933 right = get_unop_op(right);
3935 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3938 ll = get_Add_left(left);
3939 lr = get_Add_right(left);
3940 rl = get_Add_left(right);
3941 rr = get_Add_right(right);
3944 /* X + a CMP X + b ==> a CMP b */
3948 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3949 } else if (ll == rr) {
3950 /* X + a CMP b + X ==> a CMP b */
3954 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3955 } else if (lr == rl) {
3956 /* a + X CMP X + b ==> a CMP b */
3960 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3961 } else if (lr == rr) {
3962 /* a + X CMP b + X ==> a CMP b */
3966 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3970 ll = get_Sub_left(left);
3971 lr = get_Sub_right(left);
3972 rl = get_Sub_left(right);
3973 rr = get_Sub_right(right);
3976 /* X - a CMP X - b ==> a CMP b */
3980 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3981 } else if (lr == rr) {
3982 /* a - X CMP b - X ==> a CMP b */
3986 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3990 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3991 /* a ROTL X CMP b ROTL X ==> a CMP b */
3992 left = get_Rotl_left(left);
3993 right = get_Rotl_left(right);
3995 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4003 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4004 if (is_Add(left) || is_Sub(left)) {
4005 ir_node *ll = get_binop_left(left);
4006 ir_node *lr = get_binop_right(left);
4008 if (lr == right && is_Add(left)) {
4014 ir_graph *irg = get_irn_irg(n);
4016 right = create_zero_const(irg, mode);
4018 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4021 if (is_Add(right) || is_Sub(right)) {
4022 ir_node *rl = get_binop_left(right);
4023 ir_node *rr = get_binop_right(right);
4025 if (rr == left && is_Add(right)) {
4031 ir_graph *irg = get_irn_irg(n);
4033 right = create_zero_const(irg, mode);
4035 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4039 if (is_And(left) && is_Const(right)) {
4040 ir_node *ll = get_binop_left(left);
4041 ir_node *lr = get_binop_right(left);
4042 if (is_Shr(ll) && is_Const(lr)) {
4043 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4044 ir_node *block = get_nodes_block(n);
4045 ir_mode *mode = get_irn_mode(left);
4047 ir_node *llr = get_Shr_right(ll);
4048 if (is_Const(llr)) {
4049 dbg_info *dbg = get_irn_dbg_info(left);
4050 ir_graph *irg = get_irn_irg(left);
4052 ir_tarval *c1 = get_Const_tarval(llr);
4053 ir_tarval *c2 = get_Const_tarval(lr);
4054 ir_tarval *c3 = get_Const_tarval(right);
4055 ir_tarval *mask = tarval_shl(c2, c1);
4056 ir_tarval *value = tarval_shl(c3, c1);
4058 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4059 right = new_r_Const(irg, value);
4064 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4066 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4067 right = get_Eor_right(left);
4068 left = get_Eor_left(left);
4071 } /* mode_is_int(...) */
4074 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4075 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
4076 if (mode_is_int(mode) && is_And(left)
4077 && (relation == ir_relation_equal
4078 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
4079 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
4080 ir_node *and0 = get_And_left(left);
4081 ir_node *and1 = get_And_right(left);
4082 if (and1 == right) {
4083 ir_node *tmp = and0;
4087 if (and0 == right && is_single_bit(and0)) {
4088 ir_graph *irg = get_irn_irg(n);
4090 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
4091 right = create_zero_const(irg, mode);
4096 /* replace mode_b compares with ands/ors */
4097 if (mode == mode_b) {
4098 ir_node *block = get_nodes_block(n);
4102 case ir_relation_less_equal:
4103 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
4105 case ir_relation_less:
4106 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
4108 case ir_relation_greater_equal:
4109 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
4111 case ir_relation_greater:
4112 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
4114 case ir_relation_less_greater:
4115 bres = new_r_Eor(block, left, right, mode_b);
4117 case ir_relation_equal:
4118 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4121 #ifdef DEBUG_libfirm
4122 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4127 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4133 * First step: normalize the compare op
4134 * by placing the constant on the right side
4135 * or moving the lower address node to the left.
4137 if (!operands_are_normalized(left, right)) {
4142 relation = get_inversed_relation(relation);
4147 * Second step: Try to reduce the magnitude
4148 * of a constant. This may help to generate better code
4149 * later and may help to normalize more compares.
4150 * Of course this is only possible for integer values.
4152 tv = value_of(right);
4153 if (tv != tarval_bad) {
4154 ir_mode *mode = get_irn_mode(right);
4156 /* TODO extend to arbitrary constants */
4157 if (is_Conv(left) && tarval_is_null(tv)) {
4158 ir_node *op = get_Conv_op(left);
4159 ir_mode *op_mode = get_irn_mode(op);
4162 * UpConv(x) REL 0 ==> x REL 0
4163 * Don't do this for float values as it's unclear whether it is a
4164 * win. (on the other side it makes detection/creation of fabs hard)
4166 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4167 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4168 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4169 !mode_is_float(mode)) {
4170 tv = get_mode_null(op_mode);
4174 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4178 if (tv != tarval_bad) {
4179 /* the following optimization is possible on modes without Overflow
4180 * on Unary Minus or on == and !=:
4181 * -a CMP c ==> a swap(CMP) -c
4183 * Beware: for two-complement Overflow may occur, so only == and != can
4184 * be optimized, see this:
4185 * -MININT < 0 =/=> MININT > 0 !!!
4187 if (is_Minus(left) &&
4188 (!mode_overflow_on_unary_Minus(mode) ||
4189 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4190 tv = tarval_neg(tv);
4192 if (tv != tarval_bad) {
4193 left = get_Minus_op(left);
4194 relation = get_inversed_relation(relation);
4196 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4198 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4199 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4200 tv = tarval_not(tv);
4202 if (tv != tarval_bad) {
4203 left = get_Not_op(left);
4205 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4209 /* for integer modes, we have more */
4210 if (mode_is_int(mode)) {
4211 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4212 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4213 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4214 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4216 if (tv != tarval_bad) {
4217 relation ^= ir_relation_equal;
4219 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4222 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4223 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4224 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4225 tv = tarval_add(tv, get_mode_one(mode));
4227 if (tv != tarval_bad) {
4228 relation ^= ir_relation_equal;
4230 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4234 /* the following reassociations work only for == and != */
4235 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4236 if (tv != tarval_bad) {
4237 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4239 ir_node *c1 = get_Sub_right(left);
4240 ir_tarval *tv2 = value_of(c1);
4242 if (tv2 != tarval_bad) {
4243 tv2 = tarval_add(tv, value_of(c1));
4245 if (tv2 != tarval_bad) {
4246 left = get_Sub_left(left);
4249 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4253 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4254 else if (is_Add(left)) {
4255 ir_node *a_l = get_Add_left(left);
4256 ir_node *a_r = get_Add_right(left);
4260 if (is_Const(a_l)) {
4262 tv2 = value_of(a_l);
4265 tv2 = value_of(a_r);
4268 if (tv2 != tarval_bad) {
4269 tv2 = tarval_sub(tv, tv2, NULL);
4271 if (tv2 != tarval_bad) {
4275 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4279 /* -a == c ==> a == -c, -a != c ==> a != -c */
4280 else if (is_Minus(left)) {
4281 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4283 if (tv2 != tarval_bad) {
4284 left = get_Minus_op(left);
4287 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4294 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4295 switch (get_irn_opcode(left)) {
4299 c1 = get_And_right(left);
4302 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4303 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4305 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4307 /* TODO: move to constant evaluation */
4308 ir_graph *irg = get_irn_irg(n);
4309 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4310 c1 = new_r_Const(irg, tv);
4311 DBG_OPT_CSTEVAL(n, c1);
4315 if (tarval_is_single_bit(tv)) {
4317 * optimization for AND:
4319 * And(x, C) == C ==> And(x, C) != 0
4320 * And(x, C) != C ==> And(X, C) == 0
4322 * if C is a single Bit constant.
4325 /* check for Constant's match. We have check hare the tarvals,
4326 because our const might be changed */
4327 if (get_Const_tarval(c1) == tv) {
4328 /* fine: do the transformation */
4329 tv = get_mode_null(get_tarval_mode(tv));
4330 relation ^= ir_relation_less_equal_greater;
4332 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4338 c1 = get_Or_right(left);
4339 if (is_Const(c1) && tarval_is_null(tv)) {
4341 * Or(x, C) == 0 && C != 0 ==> FALSE
4342 * Or(x, C) != 0 && C != 0 ==> TRUE
4344 if (! tarval_is_null(get_Const_tarval(c1))) {
4345 /* TODO: move to constant evaluation */
4346 ir_graph *irg = get_irn_irg(n);
4347 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4348 c1 = new_r_Const(irg, tv);
4349 DBG_OPT_CSTEVAL(n, c1);
4356 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4358 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4361 c1 = get_Shl_right(left);
4363 ir_graph *irg = get_irn_irg(c1);
4364 ir_tarval *tv1 = get_Const_tarval(c1);
4365 ir_mode *mode = get_irn_mode(left);
4366 ir_tarval *minus1 = get_mode_all_one(mode);
4367 ir_tarval *amask = tarval_shr(minus1, tv1);
4368 ir_tarval *cmask = tarval_shl(minus1, tv1);
4371 if (tarval_and(tv, cmask) != tv) {
4372 /* condition not met */
4373 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4374 c1 = new_r_Const(irg, tv);
4375 DBG_OPT_CSTEVAL(n, c1);
4378 sl = get_Shl_left(left);
4379 blk = get_nodes_block(n);
4380 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4381 tv = tarval_shr(tv, tv1);
4383 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4388 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4390 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4393 c1 = get_Shr_right(left);
4395 ir_graph *irg = get_irn_irg(c1);
4396 ir_tarval *tv1 = get_Const_tarval(c1);
4397 ir_mode *mode = get_irn_mode(left);
4398 ir_tarval *minus1 = get_mode_all_one(mode);
4399 ir_tarval *amask = tarval_shl(minus1, tv1);
4400 ir_tarval *cmask = tarval_shr(minus1, tv1);
4403 if (tarval_and(tv, cmask) != tv) {
4404 /* condition not met */
4405 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4406 c1 = new_r_Const(irg, tv);
4407 DBG_OPT_CSTEVAL(n, c1);
4410 sl = get_Shr_left(left);
4411 blk = get_nodes_block(n);
4412 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4413 tv = tarval_shl(tv, tv1);
4415 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4420 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4422 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4425 c1 = get_Shrs_right(left);
4427 ir_graph *irg = get_irn_irg(c1);
4428 ir_tarval *tv1 = get_Const_tarval(c1);
4429 ir_mode *mode = get_irn_mode(left);
4430 ir_tarval *minus1 = get_mode_all_one(mode);
4431 ir_tarval *amask = tarval_shl(minus1, tv1);
4432 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4435 cond = tarval_sub(cond, tv1, NULL);
4436 cond = tarval_shrs(tv, cond);
4438 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4439 /* condition not met */
4440 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4441 c1 = new_r_Const(irg, tv);
4442 DBG_OPT_CSTEVAL(n, c1);
4445 sl = get_Shrs_left(left);
4446 blk = get_nodes_block(n);
4447 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4448 tv = tarval_shl(tv, tv1);
4450 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4455 } /* tarval != bad */
4458 if (changedc) { /* need a new Const */
4459 ir_graph *irg = get_irn_irg(n);
4460 right = new_r_Const(irg, tv);
4464 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4465 ir_node *op = get_Proj_pred(left);
4467 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4468 ir_node *c = get_binop_right(op);
4471 ir_tarval *tv = get_Const_tarval(c);
4473 if (tarval_is_single_bit(tv)) {
4474 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4475 ir_node *v = get_binop_left(op);
4476 ir_node *blk = get_irn_n(op, -1);
4477 ir_graph *irg = get_irn_irg(op);
4478 ir_mode *mode = get_irn_mode(v);
4480 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4481 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4483 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4490 dbg_info *dbgi = get_irn_dbg_info(n);
4491 ir_node *block = get_nodes_block(n);
4493 /* create a new compare */
4494 n = new_rd_Cmp(dbgi, block, left, right, relation);
4501 * Optimize CopyB(mem, x, x) into a Nop.
4503 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4505 ir_node *copyb = get_Proj_pred(proj);
4506 ir_node *a = get_CopyB_dst(copyb);
4507 ir_node *b = get_CopyB_src(copyb);
4510 switch (get_Proj_proj(proj)) {
4511 case pn_CopyB_X_regular:
4512 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4513 DBG_OPT_EXC_REM(proj);
4514 proj = new_r_Jmp(get_nodes_block(copyb));
4516 case pn_CopyB_X_except: {
4517 ir_graph *irg = get_irn_irg(proj);
4518 DBG_OPT_EXC_REM(proj);
4519 proj = new_r_Bad(irg, mode_X);
4527 } /* transform_node_Proj_CopyB */
4530 * Optimize Bounds(idx, idx, upper) into idx.
4532 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4534 ir_node *oldn = proj;
4535 ir_node *bound = get_Proj_pred(proj);
4536 ir_node *idx = get_Bound_index(bound);
4537 ir_node *pred = skip_Proj(idx);
4540 if (idx == get_Bound_lower(bound))
4542 else if (is_Bound(pred)) {
4544 * idx was Bounds checked previously, it is still valid if
4545 * lower <= pred_lower && pred_upper <= upper.
4547 ir_node *lower = get_Bound_lower(bound);
4548 ir_node *upper = get_Bound_upper(bound);
4549 if (get_Bound_lower(pred) == lower &&
4550 get_Bound_upper(pred) == upper) {
4552 * One could expect that we simply return the previous
4553 * Bound here. However, this would be wrong, as we could
4554 * add an exception Proj to a new location then.
4555 * So, we must turn in into a tuple.
4561 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4562 switch (get_Proj_proj(proj)) {
4564 DBG_OPT_EXC_REM(proj);
4565 proj = get_Bound_mem(bound);
4567 case pn_Bound_X_except:
4568 DBG_OPT_EXC_REM(proj);
4569 proj = new_r_Bad(get_irn_irg(proj), mode_X);
4573 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4575 case pn_Bound_X_regular:
4576 DBG_OPT_EXC_REM(proj);
4577 proj = new_r_Jmp(get_nodes_block(bound));
4584 } /* transform_node_Proj_Bound */
4587 * Does all optimizations on nodes that must be done on its Projs
4588 * because of creating new nodes.
4590 static ir_node *transform_node_Proj(ir_node *proj)
4592 ir_node *n = get_Proj_pred(proj);
4594 if (n->op->ops.transform_node_Proj)
4595 return n->op->ops.transform_node_Proj(proj);
4597 } /* transform_node_Proj */
4600 * Test wether a block is unreachable
4601 * Note: That this only returns true when
4602 * IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE is set.
4603 * This is important, as you easily end up producing invalid constructs in the
4604 * unreachable code when optimizing away edges into the unreachable code.
4605 * So only set this flag when you iterate localopts to the fixpoint.
4606 * When you reach the fixpoint then all unreachable code is dead
4607 * (= can't be reached by firm edges) and you won't see the invalid constructs
4610 static bool is_block_unreachable(const ir_node *block)
4612 const ir_graph *irg = get_irn_irg(block);
4613 if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE))
4615 return get_Block_dom_depth(block) < 0;
4618 static ir_node *transform_node_Block(ir_node *block)
4620 ir_graph *irg = get_irn_irg(block);
4621 int arity = get_irn_arity(block);
4622 ir_node *bad = NULL;
4625 if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE))
4628 for (i = 0; i < arity; ++i) {
4629 ir_node *const pred = get_Block_cfgpred(block, i);
4630 if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred)))
4633 bad = new_r_Bad(irg, mode_X);
4634 set_irn_n(block, i, bad);
4640 static ir_node *transform_node_Phi(ir_node *phi)
4642 int n = get_irn_arity(phi);
4643 ir_mode *mode = get_irn_mode(phi);
4644 ir_node *block = get_nodes_block(phi);
4645 ir_graph *irg = get_irn_irg(phi);
4646 ir_node *bad = NULL;
4649 /* Set phi-operands for bad-block inputs to bad */
4650 for (i = 0; i < n; ++i) {
4651 if (!is_Bad(get_Phi_pred(phi, i))) {
4652 ir_node *pred = get_Block_cfgpred(block, i);
4653 if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) {
4655 bad = new_r_Bad(irg, mode);
4656 set_irn_n(phi, i, bad);
4661 /* Move Pin nodes down through Phi nodes. */
4662 if (mode == mode_M) {
4663 n = get_irn_arity(phi);
4665 /* Beware of Phi0 */
4670 NEW_ARR_A(ir_node *, in, n);
4672 for (i = 0; i < n; ++i) {
4673 ir_node *pred = get_irn_n(phi, i);
4678 in[i] = get_Pin_op(pred);
4681 /* Move the Pin nodes "behind" the Phi. */
4682 block = get_irn_n(phi, -1);
4683 new_phi = new_r_Phi(block, n, in, mode_M);
4684 return new_r_Pin(block, new_phi);
4687 /* Move Confirms down through Phi nodes. */
4688 else if (mode_is_reference(mode)) {
4689 n = get_irn_arity(phi);
4691 /* Beware of Phi0 */
4693 ir_node *pred = get_irn_n(phi, 0);
4694 ir_node *bound, *new_phi, *block, **in;
4695 ir_relation relation;
4697 if (! is_Confirm(pred))
4700 bound = get_Confirm_bound(pred);
4701 relation = get_Confirm_relation(pred);
4703 NEW_ARR_A(ir_node *, in, n);
4704 in[0] = get_Confirm_value(pred);
4706 for (i = 1; i < n; ++i) {
4707 pred = get_irn_n(phi, i);
4709 if (! is_Confirm(pred) ||
4710 get_Confirm_bound(pred) != bound ||
4711 get_Confirm_relation(pred) != relation)
4713 in[i] = get_Confirm_value(pred);
4715 /* move the Confirm nodes "behind" the Phi */
4716 block = get_irn_n(phi, -1);
4717 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4718 return new_r_Confirm(block, new_phi, bound, relation);
4725 * Returns the operands of a commutative bin-op, if one operand is
4726 * a const, it is returned as the second one.
4728 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4730 ir_node *op_a = get_binop_left(binop);
4731 ir_node *op_b = get_binop_right(binop);
4733 assert(is_op_commutative(get_irn_op(binop)));
4735 if (is_Const(op_a)) {
4742 } /* get_comm_Binop_Ops */
4745 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4746 * Such pattern may arise in bitfield stores.
4748 * value c4 value c4 & c2
4749 * AND c3 AND c1 | c3
4756 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4759 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4761 ir_node *irn_and, *c1;
4763 ir_node *and_l, *c3;
4764 ir_node *value, *c4;
4765 ir_node *new_and, *new_const, *block;
4766 ir_mode *mode = get_irn_mode(irn_or);
4768 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4772 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4773 if (!is_Const(c1) || !is_And(irn_and))
4776 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4780 tv1 = get_Const_tarval(c1);
4781 tv2 = get_Const_tarval(c2);
4783 tv = tarval_or(tv1, tv2);
4784 if (tarval_is_all_one(tv)) {
4785 /* the AND does NOT clear a bit with isn't set by the OR */
4786 set_Or_left(irn_or, or_l);
4787 set_Or_right(irn_or, c1);
4789 /* check for more */
4796 get_comm_Binop_Ops(or_l, &and_l, &c3);
4797 if (!is_Const(c3) || !is_And(and_l))
4800 get_comm_Binop_Ops(and_l, &value, &c4);
4804 /* ok, found the pattern, check for conditions */
4805 assert(mode == get_irn_mode(irn_and));
4806 assert(mode == get_irn_mode(or_l));
4807 assert(mode == get_irn_mode(and_l));
4809 tv3 = get_Const_tarval(c3);
4810 tv4 = get_Const_tarval(c4);
4812 tv = tarval_or(tv4, tv2);
4813 if (!tarval_is_all_one(tv)) {
4814 /* have at least one 0 at the same bit position */
4818 if (tv3 != tarval_andnot(tv3, tv4)) {
4819 /* bit in the or_mask is outside the and_mask */
4823 if (tv1 != tarval_andnot(tv1, tv2)) {
4824 /* bit in the or_mask is outside the and_mask */
4828 /* ok, all conditions met */
4829 block = get_irn_n(irn_or, -1);
4830 irg = get_irn_irg(block);
4832 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4834 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4836 set_Or_left(irn_or, new_and);
4837 set_Or_right(irn_or, new_const);
4839 /* check for more */
4841 } /* transform_node_Or_bf_store */
4844 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4846 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4848 ir_mode *mode = get_irn_mode(irn_or);
4849 ir_node *shl, *shr, *block;
4850 ir_node *irn, *x, *c1, *c2, *n;
4851 ir_tarval *tv1, *tv2;
4853 /* some backends can't handle rotl */
4854 if (!be_get_backend_param()->support_rotl)
4857 if (! mode_is_int(mode))
4860 shl = get_binop_left(irn_or);
4861 shr = get_binop_right(irn_or);
4870 } else if (!is_Shl(shl)) {
4872 } else if (!is_Shr(shr)) {
4875 x = get_Shl_left(shl);
4876 if (x != get_Shr_left(shr))
4879 c1 = get_Shl_right(shl);
4880 c2 = get_Shr_right(shr);
4881 if (is_Const(c1) && is_Const(c2)) {
4882 tv1 = get_Const_tarval(c1);
4883 if (! tarval_is_long(tv1))
4886 tv2 = get_Const_tarval(c2);
4887 if (! tarval_is_long(tv2))
4890 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4891 != (int) get_mode_size_bits(mode))
4894 /* yet, condition met */
4895 block = get_nodes_block(irn_or);
4897 n = new_r_Rotl(block, x, c1, mode);
4899 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4903 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4904 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4905 if (!ir_is_negated_value(c1, c2)) {
4909 /* yet, condition met */
4910 block = get_nodes_block(irn_or);
4911 n = new_r_Rotl(block, x, c1, mode);
4912 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4914 } /* transform_node_Or_Rotl */
4916 static bool is_cmp_unequal(const ir_node *node)
4918 ir_relation relation = get_Cmp_relation(node);
4919 ir_node *left = get_Cmp_left(node);
4920 ir_node *right = get_Cmp_right(node);
4921 ir_mode *mode = get_irn_mode(left);
4923 if (relation == ir_relation_less_greater)
4926 if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right))
4927 return relation == ir_relation_greater;
4934 static ir_node *transform_node_Or(ir_node *n)
4936 ir_node *c, *oldn = n;
4937 ir_node *a = get_Or_left(n);
4938 ir_node *b = get_Or_right(n);
4941 if (is_Not(a) && is_Not(b)) {
4942 /* ~a | ~b = ~(a&b) */
4943 ir_node *block = get_nodes_block(n);
4945 mode = get_irn_mode(n);
4948 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4949 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4950 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4954 /* we can combine the relations of two compares with the same operands */
4955 if (is_Cmp(a) && is_Cmp(b)) {
4956 ir_node *a_left = get_Cmp_left(a);
4957 ir_node *a_right = get_Cmp_right(a);
4958 ir_node *b_left = get_Cmp_left(b);
4959 ir_node *b_right = get_Cmp_right(b);
4960 if (a_left == b_left && b_left == b_right) {
4961 dbg_info *dbgi = get_irn_dbg_info(n);
4962 ir_node *block = get_nodes_block(n);
4963 ir_relation a_relation = get_Cmp_relation(a);
4964 ir_relation b_relation = get_Cmp_relation(b);
4965 ir_relation new_relation = a_relation | b_relation;
4966 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4968 /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */
4969 if (is_cmp_unequal(a) && is_cmp_unequal(b)
4970 && !mode_is_float(get_irn_mode(a_left))
4971 && !mode_is_float(get_irn_mode(b_left))) {
4972 if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_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, xora, b_mode);
4981 ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode);
4982 ir_node *zero = create_zero_const(irg, b_mode);
4983 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4985 if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) {
4986 ir_graph *irg = get_irn_irg(n);
4987 dbg_info *dbgi = get_irn_dbg_info(n);
4988 ir_node *block = get_nodes_block(n);
4989 ir_mode *a_mode = get_irn_mode(a_left);
4990 ir_mode *b_mode = get_irn_mode(b_left);
4991 ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode);
4992 ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode);
4993 ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode);
4994 ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode);
4995 ir_node *zero = create_zero_const(irg, a_mode);
4996 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
5001 mode = get_irn_mode(n);
5002 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
5004 n = transform_node_Or_bf_store(n);
5005 n = transform_node_Or_Rotl(n);
5009 n = transform_bitwise_distributive(n, transform_node_Or);
5011 n = transform_node_bitop_shift(n);
5014 } /* transform_node_Or */
5018 static ir_node *transform_node(ir_node *n);
5021 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5023 * Should be moved to reassociation?
5025 static ir_node *transform_node_shift(ir_node *n)
5027 ir_node *left, *right;
5029 ir_mode *count_mode;
5030 ir_tarval *tv1, *tv2, *res;
5031 ir_node *in[2], *irn, *block;
5035 left = get_binop_left(n);
5037 /* different operations */
5038 if (get_irn_op(left) != get_irn_op(n))
5041 right = get_binop_right(n);
5042 tv1 = value_of(right);
5043 if (tv1 == tarval_bad)
5046 tv2 = value_of(get_binop_right(left));
5047 if (tv2 == tarval_bad)
5050 count_mode = get_tarval_mode(tv1);
5051 if (get_tarval_mode(tv2) != count_mode) {
5052 /* TODO: search bigger mode or something and convert... */
5056 mode = get_irn_mode(n);
5057 modulo_shf = get_mode_modulo_shift(mode);
5059 if (modulo_shf > 0) {
5060 ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode);
5062 /* I'm not so sure what happens in one complement... */
5063 assert(get_mode_arithmetic(count_mode) == irma_twos_complement);
5064 /* modulo shifts should always be a power of 2 (otherwise modulo_mask
5065 * above will be invalid) */
5066 assert(modulo_shf<=0 || is_po2(modulo_shf));
5068 tv1 = tarval_and(tv1, modulo_mask);
5069 tv2 = tarval_and(tv2, modulo_mask);
5071 res = tarval_add(tv1, tv2);
5072 irg = get_irn_irg(n);
5074 /* beware: a simple replacement works only, if res < modulo shift */
5076 int bits = get_mode_size_bits(mode);
5077 ir_tarval *modulo = new_tarval_from_long(bits, count_mode);
5078 res = tarval_mod(res, modulo);
5080 long bits = get_mode_size_bits(mode);
5081 ir_tarval *mode_size = new_tarval_from_long(bits, count_mode);
5083 /* shifting too much */
5084 if (!(tarval_cmp(res, mode_size) & ir_relation_less)) {
5086 ir_node *block = get_nodes_block(n);
5087 dbg_info *dbgi = get_irn_dbg_info(n);
5088 ir_mode *smode = get_irn_mode(right);
5089 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
5090 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5093 return new_r_Const(irg, get_mode_null(mode));
5097 /* ok, we can replace it */
5098 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5099 block = get_nodes_block(n);
5101 in[0] = get_binop_left(left);
5102 in[1] = new_r_Const(irg, res);
5104 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5106 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5108 return transform_node(irn);
5113 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5115 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5116 * (also with x >>s c1 when c1>=c2)
5118 static ir_node *transform_node_shl_shr(ir_node *n)
5121 ir_node *right = get_binop_right(n);
5131 ir_tarval *tv_shift;
5134 ir_relation relation;
5137 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5139 if (!is_Const(right))
5142 left = get_binop_left(n);
5143 mode = get_irn_mode(n);
5144 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5145 ir_node *shr_right = get_binop_right(left);
5147 if (!is_Const(shr_right))
5150 x = get_binop_left(left);
5151 tv_shr = get_Const_tarval(shr_right);
5152 tv_shl = get_Const_tarval(right);
5154 if (is_Shrs(left)) {
5155 /* shrs variant only allowed if c1 >= c2 */
5156 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5159 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5162 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5164 tv_mask = tarval_shl(tv_mask, tv_shl);
5165 } else if (is_Shr(n) && is_Shl(left)) {
5166 ir_node *shl_right = get_Shl_right(left);
5168 if (!is_Const(shl_right))
5171 x = get_Shl_left(left);
5172 tv_shr = get_Const_tarval(right);
5173 tv_shl = get_Const_tarval(shl_right);
5175 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5176 tv_mask = tarval_shr(tv_mask, tv_shr);
5181 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5182 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5185 assert(tv_mask != tarval_bad);
5186 assert(get_tarval_mode(tv_mask) == mode);
5188 block = get_nodes_block(n);
5189 irg = get_irn_irg(block);
5190 dbgi = get_irn_dbg_info(n);
5192 relation = tarval_cmp(tv_shl, tv_shr);
5193 if (relation == ir_relation_less || relation == ir_relation_equal) {
5194 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5195 new_const = new_r_Const(irg, tv_shift);
5197 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5199 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5202 assert(relation == ir_relation_greater);
5203 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5204 new_const = new_r_Const(irg, tv_shift);
5205 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5208 new_const = new_r_Const(irg, tv_mask);
5209 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5214 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5216 ir_mode *mode = get_tarval_mode(tv);
5217 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5218 return tarval_mod(tv, modulo_tv);
5221 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5222 ir_node *left, ir_node *right, ir_mode *mode);
5225 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5226 * then we can use that to minimize the value of Add(x, const) or
5227 * Sub(Const, x). In particular this often avoids 1 instruction in some
5228 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5229 * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
5231 static ir_node *transform_node_shift_modulo(ir_node *n,
5232 new_shift_func new_shift)
5234 ir_mode *mode = get_irn_mode(n);
5235 int modulo = get_mode_modulo_shift(mode);
5236 ir_node *newop = NULL;
5237 ir_mode *mode_right;
5244 if (get_mode_arithmetic(mode) != irma_twos_complement)
5246 if (!is_po2(modulo))
5249 irg = get_irn_irg(n);
5250 block = get_nodes_block(n);
5251 right = get_binop_right(n);
5252 mode_right = get_irn_mode(right);
5253 if (is_Const(right)) {
5254 ir_tarval *tv = get_Const_tarval(right);
5255 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5260 newop = new_r_Const(irg, tv_mod);
5261 } else if (is_Add(right)) {
5262 ir_node *add_right = get_Add_right(right);
5263 if (is_Const(add_right)) {
5264 ir_tarval *tv = get_Const_tarval(add_right);
5265 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5270 newconst = new_r_Const(irg, tv_mod);
5271 newop = new_r_Add(block, get_Add_left(right), newconst,
5274 } else if (is_Sub(right)) {
5275 ir_node *sub_left = get_Sub_left(right);
5276 if (is_Const(sub_left)) {
5277 ir_tarval *tv = get_Const_tarval(sub_left);
5278 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5283 newconst = new_r_Const(irg, tv_mod);
5284 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5291 if (newop != NULL) {
5292 dbg_info *dbgi = get_irn_dbg_info(n);
5293 ir_node *left = get_binop_left(n);
5294 return new_shift(dbgi, block, left, newop, mode);
5302 static ir_node *transform_node_Shr(ir_node *n)
5304 ir_node *c, *oldn = n;
5305 ir_node *left = get_Shr_left(n);
5306 ir_node *right = get_Shr_right(n);
5307 ir_mode *mode = get_irn_mode(n);
5309 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5310 n = transform_node_shift(n);
5313 n = transform_node_shift_modulo(n, new_rd_Shr);
5315 n = transform_node_shl_shr(n);
5317 n = transform_node_shift_bitop(n);
5320 } /* transform_node_Shr */
5325 static ir_node *transform_node_Shrs(ir_node *n)
5327 ir_node *c, *oldn = n;
5328 ir_node *a = get_Shrs_left(n);
5329 ir_node *b = get_Shrs_right(n);
5330 ir_mode *mode = get_irn_mode(n);
5332 if (is_oversize_shift(n)) {
5333 ir_node *block = get_nodes_block(n);
5334 dbg_info *dbgi = get_irn_dbg_info(n);
5335 ir_mode *cmode = get_irn_mode(b);
5336 long val = get_mode_size_bits(cmode)-1;
5337 ir_graph *irg = get_irn_irg(n);
5338 ir_node *cnst = new_r_Const_long(irg, cmode, val);
5339 return new_rd_Shrs(dbgi, block, a, cnst, mode);
5342 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5343 n = transform_node_shift(n);
5346 n = transform_node_shift_modulo(n, new_rd_Shrs);
5348 n = transform_node_shift_bitop(n);
5351 } /* transform_node_Shrs */
5356 static ir_node *transform_node_Shl(ir_node *n)
5358 ir_node *c, *oldn = n;
5359 ir_node *a = get_Shl_left(n);
5360 ir_node *b = get_Shl_right(n);
5361 ir_mode *mode = get_irn_mode(n);
5363 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5364 n = transform_node_shift(n);
5367 n = transform_node_shift_modulo(n, new_rd_Shl);
5369 n = transform_node_shl_shr(n);
5371 n = transform_node_shift_bitop(n);
5374 } /* transform_node_Shl */
5379 static ir_node *transform_node_Rotl(ir_node *n)
5381 ir_node *c, *oldn = n;
5382 ir_node *a = get_Rotl_left(n);
5383 ir_node *b = get_Rotl_right(n);
5384 ir_mode *mode = get_irn_mode(n);
5386 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5387 n = transform_node_shift(n);
5390 n = transform_node_shift_bitop(n);
5393 } /* transform_node_Rotl */
5398 static ir_node *transform_node_Conv(ir_node *n)
5400 ir_node *c, *oldn = n;
5401 ir_mode *mode = get_irn_mode(n);
5402 ir_node *a = get_Conv_op(n);
5404 if (mode != mode_b && is_const_Phi(a)) {
5405 /* Do NOT optimize mode_b Conv's, this leads to remaining
5406 * Phib nodes later, because the conv_b_lower operation
5407 * is instantly reverted, when it tries to insert a Convb.
5409 c = apply_conv_on_phi(a, mode);
5411 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5416 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5417 ir_graph *irg = get_irn_irg(n);
5418 return new_r_Unknown(irg, mode);
5421 if (mode_is_reference(mode) &&
5422 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5424 ir_node *l = get_Add_left(a);
5425 ir_node *r = get_Add_right(a);
5426 dbg_info *dbgi = get_irn_dbg_info(a);
5427 ir_node *block = get_nodes_block(n);
5429 ir_node *lop = get_Conv_op(l);
5430 if (get_irn_mode(lop) == mode) {
5431 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5432 n = new_rd_Add(dbgi, block, lop, r, mode);
5437 ir_node *rop = get_Conv_op(r);
5438 if (get_irn_mode(rop) == mode) {
5439 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5440 n = new_rd_Add(dbgi, block, l, rop, mode);
5447 } /* transform_node_Conv */
5450 * Remove dead blocks and nodes in dead blocks
5451 * in keep alive list. We do not generate a new End node.
5453 static ir_node *transform_node_End(ir_node *n)
5455 int i, j, n_keepalives = get_End_n_keepalives(n);
5458 NEW_ARR_A(ir_node *, in, n_keepalives);
5460 for (i = j = 0; i < n_keepalives; ++i) {
5461 ir_node *ka = get_End_keepalive(n, i);
5463 /* no need to keep Bad */
5466 /* do not keep unreachable code */
5467 block = is_Block(ka) ? ka : get_nodes_block(ka);
5468 if (is_block_unreachable(block))
5472 if (j != n_keepalives)
5473 set_End_keepalives(n, j, in);
5475 } /* transform_node_End */
5477 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5479 if (is_Minus(a) && get_Minus_op(a) == b)
5481 if (is_Minus(b) && get_Minus_op(b) == a)
5483 if (is_Sub(a) && is_Sub(b)) {
5484 ir_node *a_left = get_Sub_left(a);
5485 ir_node *a_right = get_Sub_right(a);
5486 ir_node *b_left = get_Sub_left(b);
5487 ir_node *b_right = get_Sub_right(b);
5489 if (a_left == b_right && a_right == b_left)
5496 static const ir_node *skip_upconv(const ir_node *node)
5498 while (is_Conv(node)) {
5499 ir_mode *mode = get_irn_mode(node);
5500 const ir_node *op = get_Conv_op(node);
5501 ir_mode *op_mode = get_irn_mode(op);
5502 if (!smaller_mode(op_mode, mode))
5509 int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true,
5510 const ir_node *mux_false)
5515 ir_relation relation;
5521 * Note further that these optimization work even for floating point
5522 * with NaN's because -NaN == NaN.
5523 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5526 mode = get_irn_mode(mux_true);
5527 if (mode_honor_signed_zeros(mode))
5530 /* must be <, <=, >=, > */
5531 relation = get_Cmp_relation(sel);
5532 if ((relation & ir_relation_less_greater) == 0)
5535 if (!ir_is_negated_value(mux_true, mux_false))
5538 mux_true = skip_upconv(mux_true);
5539 mux_false = skip_upconv(mux_false);
5541 /* must be x cmp 0 */
5542 cmp_right = get_Cmp_right(sel);
5543 if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
5546 cmp_left = get_Cmp_left(sel);
5547 if (cmp_left == mux_false) {
5548 if (relation & ir_relation_less) {
5551 assert(relation & ir_relation_greater);
5554 } else if (cmp_left == mux_true) {
5555 if (relation & ir_relation_less) {
5558 assert(relation & ir_relation_greater);
5566 ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true,
5569 ir_node *cmp_left = get_Cmp_left(sel);
5570 return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
5574 * Optimize a Mux into some simpler cases.
5576 static ir_node *transform_node_Mux(ir_node *n)
5579 ir_node *sel = get_Mux_sel(n);
5580 ir_mode *mode = get_irn_mode(n);
5581 ir_node *t = get_Mux_true(n);
5582 ir_node *f = get_Mux_false(n);
5583 ir_graph *irg = get_irn_irg(n);
5585 /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
5586 if (get_mode_arithmetic(mode) == irma_twos_complement) {
5587 int abs = ir_mux_is_abs(sel, t, f);
5589 dbg_info *dbgi = get_irn_dbg_info(n);
5590 ir_node *block = get_nodes_block(n);
5591 ir_node *op = ir_get_abs_op(sel, t, f);
5592 int bits = get_mode_size_bits(mode);
5593 ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
5594 ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
5595 ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode);
5598 res = new_rd_Sub(dbgi, block, xorn, sext, mode);
5600 res = new_rd_Sub(dbgi, block, sext, xorn, mode);
5606 /* first normalization step: try to move a constant to the false side,
5607 * 0 preferred on false side too */
5608 if (is_Cmp(sel) && is_Const(t) &&
5609 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5610 dbg_info *seldbgi = get_irn_dbg_info(sel);
5611 ir_node *block = get_nodes_block(sel);
5612 ir_relation relation = get_Cmp_relation(sel);
5617 /* Mux(x, a, b) => Mux(not(x), b, a) */
5618 relation = get_negated_relation(relation);
5619 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5620 get_Cmp_right(sel), relation);
5621 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5624 /* the following optimisations create new mode_b nodes, so only do them
5625 * before mode_b lowering */
5626 if (!is_irg_state(irg, IR_GRAPH_STATE_MODEB_LOWERED)) {
5628 ir_node* block = get_nodes_block(n);
5630 ir_node* c1 = get_Mux_sel(t);
5631 ir_node* t1 = get_Mux_true(t);
5632 ir_node* f1 = get_Mux_false(t);
5634 /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */
5635 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5636 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5641 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5642 } else if (f == t1) {
5643 /* Mux(cond0, Mux(cond1, x, y), x) */
5644 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5645 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5646 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5651 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5653 } else if (is_Mux(f)) {
5654 ir_node* block = get_nodes_block(n);
5656 ir_node* c1 = get_Mux_sel(f);
5657 ir_node* t1 = get_Mux_true(f);
5658 ir_node* f1 = get_Mux_false(f);
5660 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5661 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5662 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5667 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5668 } else if (t == f1) {
5669 /* Mux(cond0, x, Mux(cond1, y, x)) */
5670 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5671 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5672 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5677 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5681 /* note: after normalization, false can only happen on default */
5682 if (mode == mode_b) {
5683 dbg_info *dbg = get_irn_dbg_info(n);
5684 ir_node *block = get_nodes_block(n);
5687 ir_tarval *tv_t = get_Const_tarval(t);
5688 if (tv_t == tarval_b_true) {
5690 /* Muxb(sel, true, false) = sel */
5691 assert(get_Const_tarval(f) == tarval_b_false);
5692 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5695 /* Muxb(sel, true, x) = Or(sel, x) */
5696 n = new_rd_Or(dbg, block, sel, f, mode_b);
5697 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5701 } else if (is_Const(f)) {
5702 ir_tarval *tv_f = get_Const_tarval(f);
5703 if (tv_f == tarval_b_true) {
5704 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5705 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5706 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5707 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5710 /* Muxb(sel, x, false) = And(sel, x) */
5711 assert(tv_f == tarval_b_false);
5712 n = new_rd_And(dbg, block, sel, t, mode_b);
5713 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5719 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5720 * value to integer. */
5721 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5722 ir_tarval *a = get_Const_tarval(t);
5723 ir_tarval *b = get_Const_tarval(f);
5725 if (tarval_is_one(a) && tarval_is_null(b)) {
5726 ir_node *block = get_nodes_block(n);
5727 ir_node *conv = new_r_Conv(block, sel, mode);
5729 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5731 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5732 ir_node *block = get_nodes_block(n);
5733 ir_node *not_ = new_r_Not(block, sel, mode_b);
5734 ir_node *conv = new_r_Conv(block, not_, mode);
5736 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5743 ir_node *cmp_r = get_Cmp_right(sel);
5744 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5745 ir_node *block = get_nodes_block(n);
5746 ir_node *cmp_l = get_Cmp_left(sel);
5748 if (mode_is_int(mode)) {
5749 ir_relation relation = get_Cmp_relation(sel);
5751 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5752 /* Mux((a & b) != 0, c, 0) */
5753 ir_node *and_r = get_And_right(cmp_l);
5756 if (and_r == t && f == cmp_r) {
5757 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5758 if (relation == ir_relation_less_greater) {
5759 /* Mux((a & 2^C) != 0, 2^C, 0) == a & 2^c */
5761 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5763 /* Mux((a & 2^C) == 0, 2^C, 0) == (a & 2^c) xor (2^c) */
5764 n = new_rd_Eor(get_irn_dbg_info(n),
5765 block, cmp_l, t, mode);
5766 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5771 if (is_Shl(and_r)) {
5772 ir_node *shl_l = get_Shl_left(and_r);
5773 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5774 if (and_r == t && f == cmp_r) {
5775 if (relation == ir_relation_less_greater) {
5776 /* (a & (1 << n)) != 0, (1 << n), 0) == a & (1<<n) */
5778 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5780 /* (a & (1 << n)) == 0, (1 << n), 0) == (a & (1<<n)) xor (1<<n) */
5781 n = new_rd_Eor(get_irn_dbg_info(n),
5782 block, cmp_l, t, mode);
5783 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5789 and_l = get_And_left(cmp_l);
5790 if (is_Shl(and_l)) {
5791 ir_node *shl_l = get_Shl_left(and_l);
5792 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5793 if (and_l == t && f == cmp_r) {
5794 if (relation == ir_relation_less_greater) {
5795 /* ((1 << n) & a) != 0, (1 << n), 0) */
5797 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5799 /* ((1 << n) & a) == 0, (1 << n), 0) */
5800 n = new_rd_Eor(get_irn_dbg_info(n),
5801 block, cmp_l, t, mode);
5802 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5817 * optimize Sync nodes that have other syncs as input we simply add the inputs
5818 * of the other sync to our own inputs
5820 static ir_node *transform_node_Sync(ir_node *n)
5822 int arity = get_Sync_n_preds(n);
5825 for (i = 0; i < arity;) {
5826 ir_node *pred = get_Sync_pred(n, i);
5830 /* Remove Bad predecessors */
5837 /* Remove duplicate predecessors */
5838 for (j = 0; j < i; ++j) {
5839 if (get_Sync_pred(n, j) == pred) {
5848 if (!is_Sync(pred)) {
5856 pred_arity = get_Sync_n_preds(pred);
5857 for (j = 0; j < pred_arity; ++j) {
5858 ir_node *pred_pred = get_Sync_pred(pred, j);
5863 add_irn_n(n, pred_pred);
5867 if (get_Sync_pred(n, k) == pred_pred) break;
5873 ir_graph *irg = get_irn_irg(n);
5874 return new_r_Bad(irg, mode_M);
5877 return get_Sync_pred(n, 0);
5880 /* rehash the sync node */
5885 static ir_node *transform_node_Load(ir_node *n)
5887 /* if our memory predecessor is a load from the same address, then reuse the
5888 * previous result */
5889 ir_node *mem = get_Load_mem(n);
5894 /* don't touch volatile loads */
5895 if (get_Load_volatility(n) == volatility_is_volatile)
5897 mem_pred = get_Proj_pred(mem);
5898 if (is_Load(mem_pred)) {
5899 ir_node *pred_load = mem_pred;
5901 /* conservatively compare the 2 loads. TODO: This could be less strict
5902 * with fixup code in some situations (like smaller/bigger modes) */
5903 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
5905 if (get_Load_mode(pred_load) != get_Load_mode(n))
5907 /* all combinations of aligned/unaligned pred/n should be fine so we do
5908 * not compare the unaligned attribute */
5910 ir_node *block = get_nodes_block(n);
5911 ir_node *jmp = new_r_Jmp(block);
5912 ir_graph *irg = get_irn_irg(n);
5913 ir_node *bad = new_r_Bad(irg, mode_X);
5914 ir_mode *mode = get_Load_mode(n);
5915 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
5916 ir_node *in[] = { mem, res, jmp, bad };
5917 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5920 } else if (is_Store(mem_pred)) {
5921 ir_node *pred_store = mem_pred;
5922 ir_node *value = get_Store_value(pred_store);
5924 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
5926 if (get_irn_mode(value) != get_Load_mode(n))
5928 /* all combinations of aligned/unaligned pred/n should be fine so we do
5929 * not compare the unaligned attribute */
5931 ir_node *block = get_nodes_block(n);
5932 ir_node *jmp = new_r_Jmp(block);
5933 ir_graph *irg = get_irn_irg(n);
5934 ir_node *bad = new_r_Bad(irg, mode_X);
5935 ir_node *res = value;
5936 ir_node *in[] = { mem, res, jmp, bad };
5937 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5946 * optimize a trampoline Call into a direct Call
5948 static ir_node *transform_node_Call(ir_node *call)
5950 ir_node *callee = get_Call_ptr(call);
5951 ir_node *adr, *mem, *res, *bl, **in;
5952 ir_type *ctp, *mtp, *tp;
5956 size_t i, n_res, n_param;
5959 if (! is_Proj(callee))
5961 callee = get_Proj_pred(callee);
5962 if (! is_Builtin(callee))
5964 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5967 mem = get_Call_mem(call);
5969 if (skip_Proj(mem) == callee) {
5970 /* memory is routed to the trampoline, skip */
5971 mem = get_Builtin_mem(callee);
5974 /* build a new call type */
5975 mtp = get_Call_type(call);
5976 tdb = get_type_dbg_info(mtp);
5978 n_res = get_method_n_ress(mtp);
5979 n_param = get_method_n_params(mtp);
5980 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5982 for (i = 0; i < n_res; ++i)
5983 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5985 NEW_ARR_A(ir_node *, in, n_param + 1);
5987 /* FIXME: we don't need a new pointer type in every step */
5988 irg = get_irn_irg(call);
5989 tp = get_irg_frame_type(irg);
5990 tp = new_type_pointer(tp);
5991 set_method_param_type(ctp, 0, tp);
5993 in[0] = get_Builtin_param(callee, 2);
5994 for (i = 0; i < n_param; ++i) {
5995 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5996 in[i + 1] = get_Call_param(call, i);
5998 var = get_method_variadicity(mtp);
5999 set_method_variadicity(ctp, var);
6000 /* When we resolve a trampoline, the function must be called by a this-call */
6001 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
6002 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
6004 adr = get_Builtin_param(callee, 1);
6006 db = get_irn_dbg_info(call);
6007 bl = get_nodes_block(call);
6009 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
6010 if (get_irn_pinned(call) == op_pin_state_floats)
6011 set_irn_pinned(res, op_pin_state_floats);
6013 } /* transform_node_Call */
6016 * Tries several [inplace] [optimizing] transformations and returns an
6017 * equivalent node. The difference to equivalent_node() is that these
6018 * transformations _do_ generate new nodes, and thus the old node must
6019 * not be freed even if the equivalent node isn't the old one.
6021 static ir_node *transform_node(ir_node *n)
6026 * Transform_node is the only "optimizing transformation" that might
6027 * return a node with a different opcode. We iterate HERE until fixpoint
6028 * to get the final result.
6032 if (n->op->ops.transform_node != NULL)
6033 n = n->op->ops.transform_node(n);
6034 } while (oldn != n);
6037 } /* transform_node */
6040 * Sets the default transform node operation for an ir_op_ops.
6042 * @param code the opcode for the default operation
6043 * @param ops the operations initialized
6048 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
6052 ops->transform_node = transform_node_##a; \
6054 #define CASE_PROJ(a) \
6056 ops->transform_node_Proj = transform_node_Proj_##a; \
6058 #define CASE_PROJ_EX(a) \
6060 ops->transform_node = transform_node_##a; \
6061 ops->transform_node_Proj = transform_node_Proj_##a; \
6102 } /* firm_set_default_transform_node */
6105 /* **************** Common Subexpression Elimination **************** */
6107 /** The size of the hash table used, should estimate the number of nodes
6109 #define N_IR_NODES 512
6111 /** Compares the attributes of two Const nodes. */
6112 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
6114 return get_Const_tarval(a) != get_Const_tarval(b);
6117 /** Compares the attributes of two Proj nodes. */
6118 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
6120 return a->attr.proj.proj != b->attr.proj.proj;
6123 /** Compares the attributes of two Alloc nodes. */
6124 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
6126 const alloc_attr *pa = &a->attr.alloc;
6127 const alloc_attr *pb = &b->attr.alloc;
6128 return (pa->where != pb->where) || (pa->type != pb->type);
6131 /** Compares the attributes of two Free nodes. */
6132 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
6134 const free_attr *pa = &a->attr.free;
6135 const free_attr *pb = &b->attr.free;
6136 return (pa->where != pb->where) || (pa->type != pb->type);
6139 /** Compares the attributes of two SymConst nodes. */
6140 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
6142 const symconst_attr *pa = &a->attr.symc;
6143 const symconst_attr *pb = &b->attr.symc;
6144 return (pa->kind != pb->kind)
6145 || (pa->sym.type_p != pb->sym.type_p);
6148 /** Compares the attributes of two Call nodes. */
6149 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
6151 const call_attr *pa = &a->attr.call;
6152 const call_attr *pb = &b->attr.call;
6153 return (pa->type != pb->type)
6154 || (pa->tail_call != pb->tail_call);
6157 /** Compares the attributes of two Sel nodes. */
6158 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
6160 const ir_entity *a_ent = get_Sel_entity(a);
6161 const ir_entity *b_ent = get_Sel_entity(b);
6162 return a_ent != b_ent;
6165 /** Compares the attributes of two Phi nodes. */
6166 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
6168 /* we can only enter this function if both nodes have the same number of inputs,
6169 hence it is enough to check if one of them is a Phi0 */
6171 /* check the Phi0 pos attribute */
6172 return a->attr.phi.u.pos != b->attr.phi.u.pos;
6177 /** Compares the attributes of two Conv nodes. */
6178 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
6180 return get_Conv_strict(a) != get_Conv_strict(b);
6183 /** Compares the attributes of two Cast nodes. */
6184 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
6186 return get_Cast_type(a) != get_Cast_type(b);
6189 /** Compares the attributes of two Load nodes. */
6190 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
6192 if (get_Load_volatility(a) == volatility_is_volatile ||
6193 get_Load_volatility(b) == volatility_is_volatile)
6194 /* NEVER do CSE on volatile Loads */
6196 /* do not CSE Loads with different alignment. Be conservative. */
6197 if (get_Load_unaligned(a) != get_Load_unaligned(b))
6200 return get_Load_mode(a) != get_Load_mode(b);
6203 /** Compares the attributes of two Store nodes. */
6204 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
6206 /* do not CSE Stores with different alignment. Be conservative. */
6207 if (get_Store_unaligned(a) != get_Store_unaligned(b))
6210 /* NEVER do CSE on volatile Stores */
6211 return (get_Store_volatility(a) == volatility_is_volatile ||
6212 get_Store_volatility(b) == volatility_is_volatile);
6215 /** Compares two exception attributes */
6216 static int node_cmp_exception(const ir_node *a, const ir_node *b)
6218 const except_attr *ea = &a->attr.except;
6219 const except_attr *eb = &b->attr.except;
6221 return ea->pin_state != eb->pin_state;
6224 #define node_cmp_attr_Bound node_cmp_exception
6226 /** Compares the attributes of two Div nodes. */
6227 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
6229 const div_attr *ma = &a->attr.div;
6230 const div_attr *mb = &b->attr.div;
6231 return ma->exc.pin_state != mb->exc.pin_state ||
6232 ma->resmode != mb->resmode ||
6233 ma->no_remainder != mb->no_remainder;
6236 /** Compares the attributes of two Mod nodes. */
6237 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
6239 const mod_attr *ma = &a->attr.mod;
6240 const mod_attr *mb = &b->attr.mod;
6241 return ma->exc.pin_state != mb->exc.pin_state ||
6242 ma->resmode != mb->resmode;
6245 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
6247 const cmp_attr *ma = &a->attr.cmp;
6248 const cmp_attr *mb = &b->attr.cmp;
6249 return ma->relation != mb->relation;
6252 /** Compares the attributes of two Confirm nodes. */
6253 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
6255 const confirm_attr *ma = &a->attr.confirm;
6256 const confirm_attr *mb = &b->attr.confirm;
6257 return ma->relation != mb->relation;
6260 /** Compares the attributes of two Builtin nodes. */
6261 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
6263 /* no need to compare the type, equal kind means equal type */
6264 return get_Builtin_kind(a) != get_Builtin_kind(b);
6267 /** Compares the attributes of two ASM nodes. */
6268 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6271 const ir_asm_constraint *ca;
6272 const ir_asm_constraint *cb;
6275 if (get_ASM_text(a) != get_ASM_text(b))
6278 /* Should we really check the constraints here? Should be better, but is strange. */
6279 n = get_ASM_n_input_constraints(a);
6280 if (n != get_ASM_n_input_constraints(b))
6283 ca = get_ASM_input_constraints(a);
6284 cb = get_ASM_input_constraints(b);
6285 for (i = 0; i < n; ++i) {
6286 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6287 || ca[i].mode != cb[i].mode)
6291 n = get_ASM_n_output_constraints(a);
6292 if (n != get_ASM_n_output_constraints(b))
6295 ca = get_ASM_output_constraints(a);
6296 cb = get_ASM_output_constraints(b);
6297 for (i = 0; i < n; ++i) {
6298 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6299 || ca[i].mode != cb[i].mode)
6303 n = get_ASM_n_clobbers(a);
6304 if (n != get_ASM_n_clobbers(b))
6307 cla = get_ASM_clobbers(a);
6308 clb = get_ASM_clobbers(b);
6309 for (i = 0; i < n; ++i) {
6310 if (cla[i] != clb[i])
6316 /** Compares the inexistent attributes of two Dummy nodes. */
6317 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6325 * Set the default node attribute compare operation for an ir_op_ops.
6327 * @param code the opcode for the default operation
6328 * @param ops the operations initialized
6333 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6337 ops->node_cmp_attr = node_cmp_attr_##a; \
6369 } /* firm_set_default_node_cmp_attr */
6372 * Compare function for two nodes in the value table. Gets two
6373 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6375 int identities_cmp(const void *elt, const void *key)
6377 ir_node *a = (ir_node *)elt;
6378 ir_node *b = (ir_node *)key;
6381 if (a == b) return 0;
6383 if ((get_irn_op(a) != get_irn_op(b)) ||
6384 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6386 /* compare if a's in and b's in are of equal length */
6387 irn_arity_a = get_irn_arity(a);
6388 if (irn_arity_a != get_irn_arity(b))
6391 /* blocks are never the same */
6395 if (get_irn_pinned(a) == op_pin_state_pinned) {
6396 /* for pinned nodes, the block inputs must be equal */
6397 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6400 ir_node *block_a = get_nodes_block(a);
6401 ir_node *block_b = get_nodes_block(b);
6402 if (! get_opt_global_cse()) {
6403 /* for block-local CSE both nodes must be in the same Block */
6404 if (block_a != block_b)
6407 /* The optimistic approach would be to do nothing here.
6408 * However doing GCSE optimistically produces a lot of partially dead code which appears
6409 * to be worse in practice than the missed opportunities.
6410 * So we use a very conservative variant here and only CSE if 1 value dominates the
6412 if (!block_dominates(block_a, block_b)
6413 && !block_dominates(block_b, block_a))
6418 /* compare a->in[0..ins] with b->in[0..ins] */
6419 for (i = 0; i < irn_arity_a; ++i) {
6420 ir_node *pred_a = get_irn_n(a, i);
6421 ir_node *pred_b = get_irn_n(b, i);
6422 if (pred_a != pred_b) {
6423 /* if both predecessors are CSE neutral they might be different */
6424 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6430 * here, we already now that the nodes are identical except their
6433 if (a->op->ops.node_cmp_attr)
6434 return a->op->ops.node_cmp_attr(a, b);
6437 } /* identities_cmp */
6440 * Calculate a hash value of a node.
6442 * @param node The IR-node
6444 unsigned ir_node_hash(const ir_node *node)
6446 return node->op->ops.hash(node);
6447 } /* ir_node_hash */
6450 void new_identities(ir_graph *irg)
6452 if (irg->value_table != NULL)
6453 del_pset(irg->value_table);
6454 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6455 } /* new_identities */
6457 void del_identities(ir_graph *irg)
6459 if (irg->value_table != NULL)
6460 del_pset(irg->value_table);
6461 } /* del_identities */
6463 /* Normalize a node by putting constants (and operands with larger
6464 * node index) on the right (operator side). */
6465 void ir_normalize_node(ir_node *n)
6467 if (is_op_commutative(get_irn_op(n))) {
6468 ir_node *l = get_binop_left(n);
6469 ir_node *r = get_binop_right(n);
6471 /* For commutative operators perform a OP b == b OP a but keep
6472 * constants on the RIGHT side. This helps greatly in some
6473 * optimizations. Moreover we use the idx number to make the form
6475 if (!operands_are_normalized(l, r)) {
6476 set_binop_left(n, r);
6477 set_binop_right(n, l);
6481 } /* ir_normalize_node */
6484 * Return the canonical node computing the same value as n.
6485 * Looks up the node in a hash table, enters it in the table
6486 * if it isn't there yet.
6488 * @param n the node to look up
6490 * @return a node that computes the same value as n or n if no such
6491 * node could be found
6493 ir_node *identify_remember(ir_node *n)
6495 ir_graph *irg = get_irn_irg(n);
6496 pset *value_table = irg->value_table;
6499 if (value_table == NULL)
6502 ir_normalize_node(n);
6503 /* lookup or insert in hash table with given hash key. */
6504 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6507 /* n is reachable again */
6508 edges_node_revival(nn);
6512 } /* identify_remember */
6515 * During construction we set the op_pin_state_pinned flag in the graph right
6516 * when the optimization is performed. The flag turning on procedure global
6517 * cse could be changed between two allocations. This way we are safe.
6519 * @param n The node to lookup
6521 static inline ir_node *identify_cons(ir_node *n)
6525 n = identify_remember(n);
6526 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6527 ir_graph *irg = get_irn_irg(n);
6528 set_irg_pinned(irg, op_pin_state_floats);
6531 } /* identify_cons */
6533 /* Add a node to the identities value table. */
6534 void add_identities(ir_node *node)
6541 identify_remember(node);
6544 /* Visit each node in the value table of a graph. */
6545 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6548 ir_graph *rem = current_ir_graph;
6550 current_ir_graph = irg;
6551 foreach_pset(irg->value_table, ir_node*, node) {
6554 current_ir_graph = rem;
6555 } /* visit_all_identities */
6558 * These optimizations deallocate nodes from the obstack.
6559 * It can only be called if it is guaranteed that no other nodes
6560 * reference this one, i.e., right after construction of a node.
6562 * @param n The node to optimize
6564 ir_node *optimize_node(ir_node *n)
6567 ir_graph *irg = get_irn_irg(n);
6568 unsigned iro = get_irn_opcode(n);
6571 /* Always optimize Phi nodes: part of the construction. */
6572 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6574 /* constant expression evaluation / constant folding */
6575 if (get_opt_constant_folding()) {
6576 /* neither constants nor Tuple values can be evaluated */
6577 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6578 /* try to evaluate */
6579 tv = computed_value(n);
6580 if (tv != tarval_bad) {
6585 * we MUST copy the node here temporarily, because it's still
6586 * needed for DBG_OPT_CSTEVAL
6588 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6589 oldn = (ir_node*)alloca(node_size);
6591 memcpy(oldn, n, node_size);
6592 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6594 /* ARG, copy the in array, we need it for statistics */
6595 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6597 /* note the inplace edges module */
6598 edges_node_deleted(n);
6600 /* evaluation was successful -- replace the node. */
6601 irg_kill_node(irg, n);
6602 nw = new_r_Const(irg, tv);
6604 DBG_OPT_CSTEVAL(oldn, nw);
6610 /* remove unnecessary nodes */
6611 if (get_opt_algebraic_simplification() ||
6612 (iro == iro_Phi) || /* always optimize these nodes. */
6614 (iro == iro_Proj) ||
6615 (iro == iro_Block) ) /* Flags tested local. */
6616 n = equivalent_node(n);
6618 /* Common Subexpression Elimination.
6620 * Checks whether n is already available.
6621 * The block input is used to distinguish different subexpressions. Right
6622 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6623 * subexpressions within a block.
6626 n = identify_cons(n);
6629 edges_node_deleted(oldn);
6631 /* We found an existing, better node, so we can deallocate the old node. */
6632 irg_kill_node(irg, oldn);
6636 /* Some more constant expression evaluation that does not allow to
6638 iro = get_irn_opcode(n);
6639 if (get_opt_algebraic_simplification() ||
6640 (iro == iro_Cond) ||
6641 (iro == iro_Proj)) /* Flags tested local. */
6642 n = transform_node(n);
6644 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6645 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6647 n = identify_remember(o);
6653 } /* optimize_node */
6657 * These optimizations never deallocate nodes (in place). This can cause dead
6658 * nodes lying on the obstack. Remove these by a dead node elimination,
6659 * i.e., a copying garbage collection.
6661 ir_node *optimize_in_place_2(ir_node *n)
6665 unsigned iro = get_irn_opcode(n);
6667 if (!get_opt_optimize() && !is_Phi(n)) return n;
6669 if (iro == iro_Deleted)
6672 /* constant expression evaluation / constant folding */
6673 if (get_opt_constant_folding()) {
6674 /* neither constants nor Tuple values can be evaluated */
6675 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6676 /* try to evaluate */
6677 tv = computed_value(n);
6678 if (tv != tarval_bad) {
6679 /* evaluation was successful -- replace the node. */
6680 ir_graph *irg = get_irn_irg(n);
6682 n = new_r_Const(irg, tv);
6684 DBG_OPT_CSTEVAL(oldn, n);
6690 /* remove unnecessary nodes */
6691 if (get_opt_constant_folding() ||
6692 (iro == iro_Phi) || /* always optimize these nodes. */
6693 (iro == iro_Id) || /* ... */
6694 (iro == iro_Proj) || /* ... */
6695 (iro == iro_Block) ) /* Flags tested local. */
6696 n = equivalent_node(n);
6698 /** common subexpression elimination **/
6699 /* Checks whether n is already available. */
6700 /* The block input is used to distinguish different subexpressions. Right
6701 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6702 subexpressions within a block. */
6703 if (get_opt_cse()) {
6705 n = identify_remember(o);
6710 /* Some more constant expression evaluation. */
6711 iro = get_irn_opcode(n);
6712 if (get_opt_constant_folding() ||
6713 (iro == iro_Cond) ||
6714 (iro == iro_Proj)) /* Flags tested local. */
6715 n = transform_node(n);
6717 /* Now we can verify the node, as it has no dead inputs any more. */
6720 /* Now we have a legal, useful node. Enter it in hash table for cse.
6721 Blocks should be unique anyways. (Except the successor of start:
6722 is cse with the start block!) */
6723 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6725 n = identify_remember(o);
6731 } /* optimize_in_place_2 */
6734 * Wrapper for external use, set proper status bits after optimization.
6736 ir_node *optimize_in_place(ir_node *n)
6738 ir_graph *irg = get_irn_irg(n);
6739 /* Handle graph state */
6740 assert(get_irg_phase_state(irg) != phase_building);
6742 if (get_opt_global_cse())
6743 set_irg_pinned(irg, op_pin_state_floats);
6745 /* FIXME: Maybe we could also test whether optimizing the node can
6746 change the control graph. */
6747 clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE);
6748 return optimize_in_place_2(n);
6749 } /* optimize_in_place */
6752 * Calculate a hash value of a Const node.
6754 static unsigned hash_Const(const ir_node *node)
6758 /* special value for const, as they only differ in their tarval. */
6759 h = HASH_PTR(node->attr.con.tarval);
6765 * Calculate a hash value of a SymConst node.
6767 static unsigned hash_SymConst(const ir_node *node)
6771 /* all others are pointers */
6772 h = HASH_PTR(node->attr.symc.sym.type_p);
6775 } /* hash_SymConst */
6778 * Set the default hash operation in an ir_op_ops.
6780 * @param code the opcode for the default operation
6781 * @param ops the operations initialized
6786 static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops)
6790 ops->hash = hash_##a; \
6793 /* hash function already set */
6794 if (ops->hash != NULL)
6801 /* use input/mode default hash if no function was given */
6802 ops->hash = firm_default_hash;
6810 * Sets the default operation for an ir_ops.
6812 ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops)
6814 ops = firm_set_default_hash(code, ops);
6815 ops = firm_set_default_computed_value(code, ops);
6816 ops = firm_set_default_equivalent_node(code, ops);
6817 ops = firm_set_default_transform_node(code, ops);
6818 ops = firm_set_default_node_cmp_attr(code, ops);
6819 ops = firm_set_default_get_type_attr(code, ops);
6820 ops = firm_set_default_get_entity_attr(code, ops);
6823 } /* firm_set_default_operations */