2 * Copyright (C) 1995-2008 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
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Returns the tarval of a Const node or tarval_bad for all other nodes.
59 static tarval *default_value_of(const ir_node *n) {
61 return get_Const_tarval(n); /* might return tarval_bad */
66 value_of_func value_of_ptr = default_value_of;
68 /* * Set a new value_of function. */
69 void set_value_of_func(value_of_func func) {
73 value_of_ptr = default_value_of;
77 * Return the value of a Constant.
79 static tarval *computed_value_Const(const ir_node *n) {
80 return get_Const_tarval(n);
81 } /* computed_value_Const */
84 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
86 static tarval *computed_value_SymConst(const ir_node *n) {
90 switch (get_SymConst_kind(n)) {
91 case symconst_type_size:
92 type = get_SymConst_type(n);
93 if (get_type_state(type) == layout_fixed)
94 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
96 case symconst_type_align:
97 type = get_SymConst_type(n);
98 if (get_type_state(type) == layout_fixed)
99 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
101 case symconst_ofs_ent:
102 ent = get_SymConst_entity(n);
103 type = get_entity_owner(ent);
104 if (get_type_state(type) == layout_fixed)
105 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
111 } /* computed_value_SymConst */
114 * Return the value of an Add.
116 static tarval *computed_value_Add(const ir_node *n) {
117 ir_node *a = get_Add_left(n);
118 ir_node *b = get_Add_right(n);
120 tarval *ta = value_of(a);
121 tarval *tb = value_of(b);
123 if ((ta != tarval_bad) && (tb != tarval_bad))
124 return tarval_add(ta, tb);
127 } /* computed_value_Add */
130 * Return the value of a Sub.
131 * Special case: a - a
133 static tarval *computed_value_Sub(const ir_node *n) {
134 ir_mode *mode = get_irn_mode(n);
135 ir_node *a = get_Sub_left(n);
136 ir_node *b = get_Sub_right(n);
141 if (a == b && !is_Bad(a))
142 return get_mode_null(mode);
147 if ((ta != tarval_bad) && (tb != tarval_bad))
148 return tarval_sub(ta, tb, mode);
151 } /* computed_value_Sub */
154 * Return the value of a Carry.
155 * Special : a op 0, 0 op b
157 static tarval *computed_value_Carry(const ir_node *n) {
158 ir_node *a = get_binop_left(n);
159 ir_node *b = get_binop_right(n);
160 ir_mode *m = get_irn_mode(n);
162 tarval *ta = value_of(a);
163 tarval *tb = value_of(b);
165 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
169 if (tarval_is_null(ta) || tarval_is_null(tb))
170 return get_mode_null(m);
173 } /* computed_value_Carry */
176 * Return the value of a Borrow.
179 static tarval *computed_value_Borrow(const ir_node *n) {
180 ir_node *a = get_binop_left(n);
181 ir_node *b = get_binop_right(n);
182 ir_mode *m = get_irn_mode(n);
184 tarval *ta = value_of(a);
185 tarval *tb = value_of(b);
187 if ((ta != tarval_bad) && (tb != tarval_bad)) {
188 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
189 } else if (tarval_is_null(ta)) {
190 return get_mode_null(m);
193 } /* computed_value_Borrow */
196 * Return the value of an unary Minus.
198 static tarval *computed_value_Minus(const ir_node *n) {
199 ir_node *a = get_Minus_op(n);
200 tarval *ta = value_of(a);
202 if (ta != tarval_bad)
203 return tarval_neg(ta);
206 } /* computed_value_Minus */
209 * Return the value of a Mul.
211 static tarval *computed_value_Mul(const ir_node *n) {
212 ir_node *a = get_Mul_left(n);
213 ir_node *b = get_Mul_right(n);
216 tarval *ta = value_of(a);
217 tarval *tb = value_of(b);
219 mode = get_irn_mode(n);
220 if (mode != get_irn_mode(a)) {
221 /* n * n = 2n bit multiplication */
222 ta = tarval_convert_to(ta, mode);
223 tb = tarval_convert_to(tb, mode);
226 if (ta != tarval_bad && tb != tarval_bad) {
227 return tarval_mul(ta, tb);
229 /* a*0 = 0 or 0*b = 0 */
230 if (ta == get_mode_null(mode))
232 if (tb == get_mode_null(mode))
236 } /* computed_value_Mul */
239 * Return the value of an Abs.
241 static tarval *computed_value_Abs(const ir_node *n) {
242 ir_node *a = get_Abs_op(n);
243 tarval *ta = value_of(a);
245 if (ta != tarval_bad)
246 return tarval_abs(ta);
249 } /* computed_value_Abs */
252 * Return the value of an And.
253 * Special case: a & 0, 0 & b
255 static tarval *computed_value_And(const ir_node *n) {
256 ir_node *a = get_And_left(n);
257 ir_node *b = get_And_right(n);
259 tarval *ta = value_of(a);
260 tarval *tb = value_of(b);
262 if ((ta != tarval_bad) && (tb != tarval_bad)) {
263 return tarval_and (ta, tb);
265 if (tarval_is_null(ta)) return ta;
266 if (tarval_is_null(tb)) return tb;
269 } /* computed_value_And */
272 * Return the value of an Or.
273 * Special case: a | 1...1, 1...1 | b
275 static tarval *computed_value_Or(const ir_node *n) {
276 ir_node *a = get_Or_left(n);
277 ir_node *b = get_Or_right(n);
279 tarval *ta = value_of(a);
280 tarval *tb = value_of(b);
282 if ((ta != tarval_bad) && (tb != tarval_bad)) {
283 return tarval_or (ta, tb);
285 if (tarval_is_all_one(ta)) return ta;
286 if (tarval_is_all_one(tb)) return tb;
289 } /* computed_value_Or */
292 * Return the value of an Eor.
294 static tarval *computed_value_Eor(const ir_node *n) {
295 ir_node *a = get_Eor_left(n);
296 ir_node *b = get_Eor_right(n);
301 return get_mode_null(get_irn_mode(n));
306 if ((ta != tarval_bad) && (tb != tarval_bad)) {
307 return tarval_eor (ta, tb);
310 } /* computed_value_Eor */
313 * Return the value of a Not.
315 static tarval *computed_value_Not(const ir_node *n) {
316 ir_node *a = get_Not_op(n);
317 tarval *ta = value_of(a);
319 if (ta != tarval_bad)
320 return tarval_not(ta);
323 } /* computed_value_Not */
326 * Return the value of a Shl.
328 static tarval *computed_value_Shl(const ir_node *n) {
329 ir_node *a = get_Shl_left(n);
330 ir_node *b = get_Shl_right(n);
332 tarval *ta = value_of(a);
333 tarval *tb = value_of(b);
335 if ((ta != tarval_bad) && (tb != tarval_bad)) {
336 return tarval_shl (ta, tb);
339 } /* computed_value_Shl */
342 * Return the value of a Shr.
344 static tarval *computed_value_Shr(const ir_node *n) {
345 ir_node *a = get_Shr_left(n);
346 ir_node *b = get_Shr_right(n);
348 tarval *ta = value_of(a);
349 tarval *tb = value_of(b);
351 if ((ta != tarval_bad) && (tb != tarval_bad)) {
352 return tarval_shr (ta, tb);
355 } /* computed_value_Shr */
358 * Return the value of a Shrs.
360 static tarval *computed_value_Shrs(const ir_node *n) {
361 ir_node *a = get_Shrs_left(n);
362 ir_node *b = get_Shrs_right(n);
364 tarval *ta = value_of(a);
365 tarval *tb = value_of(b);
367 if ((ta != tarval_bad) && (tb != tarval_bad)) {
368 return tarval_shrs (ta, tb);
371 } /* computed_value_Shrs */
374 * Return the value of a Rotl.
376 static tarval *computed_value_Rotl(const ir_node *n) {
377 ir_node *a = get_Rotl_left(n);
378 ir_node *b = get_Rotl_right(n);
380 tarval *ta = value_of(a);
381 tarval *tb = value_of(b);
383 if ((ta != tarval_bad) && (tb != tarval_bad)) {
384 return tarval_rotl(ta, tb);
387 } /* computed_value_Rotl */
390 * Return the value of a Conv.
392 static tarval *computed_value_Conv(const ir_node *n) {
393 ir_node *a = get_Conv_op(n);
394 tarval *ta = value_of(a);
396 if (ta != tarval_bad)
397 return tarval_convert_to(ta, get_irn_mode(n));
400 } /* computed_value_Conv */
403 * Calculate the value of a Mux: can be evaluated, if the
404 * sel and the right input are known.
406 static tarval *computed_value_Mux(const ir_node *n) {
407 ir_node *sel = get_Mux_sel(n);
408 tarval *ts = value_of(sel);
410 if (ts == get_tarval_b_true()) {
411 ir_node *v = get_Mux_true(n);
414 else if (ts == get_tarval_b_false()) {
415 ir_node *v = get_Mux_false(n);
419 } /* computed_value_Mux */
422 * Calculate the value of a Confirm: can be evaluated,
423 * if it has the form Confirm(x, '=', Const).
425 static tarval *computed_value_Confirm(const ir_node *n) {
427 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
428 * Do NOT optimize them away (CondEval wants them), so wait until
429 * remove_confirm is activated.
431 if (get_opt_remove_confirm()) {
432 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
433 tarval *tv = value_of(get_Confirm_bound(n));
434 if (tv != tarval_bad)
438 return value_of(get_Confirm_value(n));
439 } /* computed_value_Confirm */
442 * Return the value of a Proj(Cmp).
444 * This performs a first step of unreachable code elimination.
445 * Proj can not be computed, but folding a Cmp above the Proj here is
446 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
448 * There are several case where we can evaluate a Cmp node, see later.
450 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
451 ir_node *a = get_Proj_pred(n);
452 ir_node *aa = get_Cmp_left(a);
453 ir_node *ab = get_Cmp_right(a);
454 long proj_nr = get_Proj_proj(n);
457 * BEWARE: a == a is NOT always True for floating Point values, as
458 * NaN != NaN is defined, so we must check this here.
461 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
464 /* This is a trick with the bits used for encoding the Cmp
465 Proj numbers, the following statement is not the same:
466 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
467 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
470 tarval *taa = value_of(aa);
471 tarval *tab = value_of(ab);
472 ir_mode *mode = get_irn_mode(aa);
475 * The predecessors of Cmp are target values. We can evaluate
478 if ((taa != tarval_bad) && (tab != tarval_bad)) {
479 /* strange checks... */
480 pn_Cmp flags = tarval_cmp(taa, tab);
481 if (flags != pn_Cmp_False) {
482 return new_tarval_from_long (proj_nr & flags, mode_b);
485 /* for integer values, we can check against MIN/MAX */
486 else if (mode_is_int(mode)) {
487 /* MIN <=/> x. This results in true/false. */
488 if (taa == get_mode_min(mode)) {
489 /* a compare with the MIN value */
490 if (proj_nr == pn_Cmp_Le)
491 return get_tarval_b_true();
492 else if (proj_nr == pn_Cmp_Gt)
493 return get_tarval_b_false();
495 /* x >=/< MIN. This results in true/false. */
497 if (tab == get_mode_min(mode)) {
498 /* a compare with the MIN value */
499 if (proj_nr == pn_Cmp_Ge)
500 return get_tarval_b_true();
501 else if (proj_nr == pn_Cmp_Lt)
502 return get_tarval_b_false();
504 /* MAX >=/< x. This results in true/false. */
505 else if (taa == get_mode_max(mode)) {
506 if (proj_nr == pn_Cmp_Ge)
507 return get_tarval_b_true();
508 else if (proj_nr == pn_Cmp_Lt)
509 return get_tarval_b_false();
511 /* x <=/> MAX. This results in true/false. */
512 else if (tab == get_mode_max(mode)) {
513 if (proj_nr == pn_Cmp_Le)
514 return get_tarval_b_true();
515 else if (proj_nr == pn_Cmp_Gt)
516 return get_tarval_b_false();
520 * The predecessors are Allocs or (void*)(0) constants. Allocs never
521 * return NULL, they raise an exception. Therefore we can predict
525 ir_node *aaa = skip_Proj(aa);
526 ir_node *aba = skip_Proj(ab);
528 if ( ( (/* aa is ProjP and aaa is Alloc */
530 && mode_is_reference(get_irn_mode(aa))
532 && ( (/* ab is NULL */
533 mode_is_reference(get_irn_mode(ab))
534 && tarval_is_null(tab))
535 || (/* ab is other Alloc */
537 && mode_is_reference(get_irn_mode(ab))
540 || (/* aa is NULL and aba is Alloc */
541 mode_is_reference(get_irn_mode(aa))
542 && tarval_is_null(taa)
544 && mode_is_reference(get_irn_mode(ab))
547 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
550 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
551 } /* computed_value_Proj_Cmp */
554 * Return the value of a floating point Quot.
556 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
557 tarval *ta = value_of(a);
558 tarval *tb = value_of(b);
560 /* cannot optimize 0 / b = 0 because of NaN */
561 if (ta != tarval_bad && tb != tarval_bad)
562 return tarval_quo(ta, tb);
564 } /* do_computed_value_Quot */
567 * Calculate the value of an integer Div of two nodes.
568 * Special case: 0 / b
570 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
571 tarval *ta = value_of(a);
573 const ir_node *dummy;
575 /* Compute c1 / c2 or 0 / a, a != 0 */
576 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
577 return ta; /* 0 / b == 0 */
579 if (ta != tarval_bad && tb != tarval_bad)
580 return tarval_div(ta, tb);
582 } /* do_computed_value_Div */
585 * Calculate the value of an integer Mod of two nodes.
586 * Special case: a % 1
588 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
589 tarval *ta = value_of(a);
590 tarval *tb = value_of(b);
592 /* Compute a % 1 or c1 % c2 */
593 if (tarval_is_one(tb))
594 return get_mode_null(get_irn_mode(a));
595 if (ta != tarval_bad && tb != tarval_bad)
596 return tarval_mod(ta, tb);
598 } /* do_computed_value_Mod */
601 * Return the value of a Proj(DivMod).
603 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
604 long proj_nr = get_Proj_proj(n);
606 /* compute either the Div or the Mod part */
607 if (proj_nr == pn_DivMod_res_div) {
608 const ir_node *a = get_Proj_pred(n);
609 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
610 } else if (proj_nr == pn_DivMod_res_mod) {
611 const ir_node *a = get_Proj_pred(n);
612 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
615 } /* computed_value_Proj_DivMod */
618 * Return the value of a Proj(Div).
620 static tarval *computed_value_Proj_Div(const ir_node *n) {
621 long proj_nr = get_Proj_proj(n);
623 if (proj_nr == pn_Div_res) {
624 const ir_node *a = get_Proj_pred(n);
625 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
628 } /* computed_value_Proj_Div */
631 * Return the value of a Proj(Mod).
633 static tarval *computed_value_Proj_Mod(const ir_node *n) {
634 long proj_nr = get_Proj_proj(n);
636 if (proj_nr == pn_Mod_res) {
637 const ir_node *a = get_Proj_pred(n);
638 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
641 } /* computed_value_Proj_Mod */
644 * Return the value of a Proj(Quot).
646 static tarval *computed_value_Proj_Quot(const ir_node *n) {
647 long proj_nr = get_Proj_proj(n);
649 if (proj_nr == pn_Quot_res) {
650 const ir_node *a = get_Proj_pred(n);
651 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
654 } /* computed_value_Proj_Quot */
657 * Return the value of a Proj.
659 static tarval *computed_value_Proj(const ir_node *proj) {
660 ir_node *n = get_Proj_pred(proj);
662 if (n->op->ops.computed_value_Proj != NULL)
663 return n->op->ops.computed_value_Proj(proj);
665 } /* computed_value_Proj */
668 * If the parameter n can be computed, return its value, else tarval_bad.
669 * Performs constant folding.
671 * @param n The node this should be evaluated
673 tarval *computed_value(const ir_node *n) {
674 if (n->op->ops.computed_value)
675 return n->op->ops.computed_value(n);
677 } /* computed_value */
680 * Set the default computed_value evaluator in an ir_op_ops.
682 * @param code the opcode for the default operation
683 * @param ops the operations initialized
688 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
692 ops->computed_value = computed_value_##a; \
694 #define CASE_PROJ(a) \
696 ops->computed_value_Proj = computed_value_Proj_##a; \
733 } /* firm_set_default_computed_value */
736 * Returns a equivalent block for another block.
737 * If the block has only one predecessor, this is
738 * the equivalent one. If the only predecessor of a block is
739 * the block itself, this is a dead block.
741 * If both predecessors of a block are the branches of a binary
742 * Cond, the equivalent block is Cond's block.
744 * If all predecessors of a block are bad or lies in a dead
745 * block, the current block is dead as well.
747 * Note, that blocks are NEVER turned into Bad's, instead
748 * the dead_block flag is set. So, never test for is_Bad(block),
749 * always use is_dead_Block(block).
751 static ir_node *equivalent_node_Block(ir_node *n)
756 /* don't optimize dead blocks */
757 if (is_Block_dead(n))
760 n_preds = get_Block_n_cfgpreds(n);
762 /* The Block constructor does not call optimize, but mature_immBlock()
763 calls the optimization. */
764 assert(get_Block_matured(n));
766 /* Straightening: a single entry Block following a single exit Block
767 can be merged, if it is not the Start block. */
768 /* !!! Beware, all Phi-nodes of n must have been optimized away.
769 This should be true, as the block is matured before optimize is called.
770 But what about Phi-cycles with the Phi0/Id that could not be resolved?
771 Remaining Phi nodes are just Ids. */
773 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
776 ir_node *predblock = get_nodes_block(pred);
777 if (predblock == oldn) {
778 /* Jmp jumps into the block it is in -- deal self cycle. */
779 n = set_Block_dead(n);
780 DBG_OPT_DEAD_BLOCK(oldn, n);
781 } else if (get_opt_control_flow_straightening()) {
783 DBG_OPT_STG(oldn, n);
785 } else if (is_Cond(pred)) {
786 ir_node *predblock = get_nodes_block(pred);
787 if (predblock == oldn) {
788 /* Jmp jumps into the block it is in -- deal self cycle. */
789 n = set_Block_dead(n);
790 DBG_OPT_DEAD_BLOCK(oldn, n);
793 } else if ((n_preds == 2) &&
794 (get_opt_control_flow_weak_simplification())) {
795 /* Test whether Cond jumps twice to this block
796 * The more general case which more than 2 predecessors is handles
797 * in optimize_cf(), we handle only this special case for speed here.
799 ir_node *a = get_Block_cfgpred(n, 0);
800 ir_node *b = get_Block_cfgpred(n, 1);
802 if (is_Proj(a) && is_Proj(b)) {
803 ir_node *cond = get_Proj_pred(a);
805 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
806 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
807 /* Also a single entry Block following a single exit Block. Phis have
808 twice the same operand and will be optimized away. */
809 n = get_nodes_block(cond);
810 DBG_OPT_IFSIM1(oldn, a, b, n);
813 } else if (get_opt_unreachable_code() &&
814 (n != get_irg_start_block(current_ir_graph)) &&
815 (n != get_irg_end_block(current_ir_graph)) ) {
818 /* If all inputs are dead, this block is dead too, except if it is
819 the start or end block. This is one step of unreachable code
821 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
822 ir_node *pred = get_Block_cfgpred(n, i);
825 if (is_Bad(pred)) continue;
826 pred_blk = get_nodes_block(skip_Proj(pred));
828 if (is_Block_dead(pred_blk)) continue;
831 /* really found a living input */
836 n = set_Block_dead(n);
837 DBG_OPT_DEAD_BLOCK(oldn, n);
842 } /* equivalent_node_Block */
845 * Returns a equivalent node for a Jmp, a Bad :-)
846 * Of course this only happens if the Block of the Jmp is dead.
848 static ir_node *equivalent_node_Jmp(ir_node *n) {
851 /* unreachable code elimination */
852 if (is_Block_dead(get_nodes_block(n))) {
853 n = get_irg_bad(current_ir_graph);
854 DBG_OPT_DEAD_BLOCK(oldn, n);
857 } /* equivalent_node_Jmp */
859 /** Raise is handled in the same way as Jmp. */
860 #define equivalent_node_Raise equivalent_node_Jmp
863 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
864 See transform_node_Proj_Cond(). */
867 * Optimize operations that are commutative and have neutral 0,
868 * so a op 0 = 0 op a = a.
870 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
873 ir_node *a = get_binop_left(n);
874 ir_node *b = get_binop_right(n);
879 /* After running compute_node there is only one constant predecessor.
880 Find this predecessors value and remember the other node: */
881 if ((tv = value_of(a)) != tarval_bad) {
883 } else if ((tv = value_of(b)) != tarval_bad) {
888 /* If this predecessors constant value is zero, the operation is
889 * unnecessary. Remove it.
891 * Beware: If n is a Add, the mode of on and n might be different
892 * which happens in this rare construction: NULL + 3.
893 * Then, a Conv would be needed which we cannot include here.
895 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
898 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
902 } /* equivalent_node_neutral_zero */
905 * Eor is commutative and has neutral 0.
907 static ir_node *equivalent_node_Eor(ir_node *n) {
912 n = equivalent_node_neutral_zero(n);
913 if (n != oldn) return n;
916 b = get_Eor_right(n);
919 ir_node *aa = get_Eor_left(a);
920 ir_node *ab = get_Eor_right(a);
923 /* (a ^ b) ^ a -> b */
925 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
927 } else if (ab == b) {
928 /* (a ^ b) ^ b -> a */
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
935 ir_node *ba = get_Eor_left(b);
936 ir_node *bb = get_Eor_right(b);
939 /* a ^ (a ^ b) -> b */
941 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
943 } else if (bb == a) {
944 /* a ^ (b ^ a) -> b */
946 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
954 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
956 * The second one looks strange, but this construct
957 * is used heavily in the LCC sources :-).
959 * Beware: The Mode of an Add may be different than the mode of its
960 * predecessors, so we could not return a predecessors in all cases.
962 static ir_node *equivalent_node_Add(ir_node *n) {
964 ir_node *left, *right;
965 ir_mode *mode = get_irn_mode(n);
967 n = equivalent_node_neutral_zero(n);
971 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
972 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
975 left = get_Add_left(n);
976 right = get_Add_right(n);
979 if (get_Sub_right(left) == right) {
982 n = get_Sub_left(left);
983 if (mode == get_irn_mode(n)) {
984 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
990 if (get_Sub_right(right) == left) {
993 n = get_Sub_left(right);
994 if (mode == get_irn_mode(n)) {
995 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1001 } /* equivalent_node_Add */
1004 * optimize operations that are not commutative but have neutral 0 on left,
1007 static ir_node *equivalent_node_left_zero(ir_node *n) {
1010 ir_node *a = get_binop_left(n);
1011 ir_node *b = get_binop_right(n);
1012 tarval *tb = value_of(b);
1014 if (tarval_is_null(tb)) {
1017 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1020 } /* equivalent_node_left_zero */
1022 #define equivalent_node_Shl equivalent_node_left_zero
1023 #define equivalent_node_Shr equivalent_node_left_zero
1024 #define equivalent_node_Shrs equivalent_node_left_zero
1025 #define equivalent_node_Rotl equivalent_node_left_zero
1028 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1030 * The second one looks strange, but this construct
1031 * is used heavily in the LCC sources :-).
1033 * Beware: The Mode of a Sub may be different than the mode of its
1034 * predecessors, so we could not return a predecessors in all cases.
1036 static ir_node *equivalent_node_Sub(ir_node *n) {
1039 ir_mode *mode = get_irn_mode(n);
1042 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1043 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1046 b = get_Sub_right(n);
1049 /* Beware: modes might be different */
1050 if (tarval_is_null(tb)) {
1051 ir_node *a = get_Sub_left(n);
1052 if (mode == get_irn_mode(a)) {
1055 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1059 } /* equivalent_node_Sub */
1063 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1066 * -(-a) == a, but might overflow two times.
1067 * We handle it anyway here but the better way would be a
1068 * flag. This would be needed for Pascal for instance.
1070 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1072 ir_node *pred = get_unop_op(n);
1074 /* optimize symmetric unop */
1075 if (get_irn_op(pred) == get_irn_op(n)) {
1076 n = get_unop_op(pred);
1077 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1080 } /* equivalent_node_idempotent_unop */
1082 /** Optimize Not(Not(x)) == x. */
1083 #define equivalent_node_Not equivalent_node_idempotent_unop
1085 /** -(-x) == x ??? Is this possible or can --x raise an
1086 out of bounds exception if min =! max? */
1087 #define equivalent_node_Minus equivalent_node_idempotent_unop
1090 * Optimize a * 1 = 1 * a = a.
1092 static ir_node *equivalent_node_Mul(ir_node *n) {
1094 ir_node *a = get_Mul_left(n);
1096 /* we can handle here only the n * n = n bit cases */
1097 if (get_irn_mode(n) == get_irn_mode(a)) {
1098 ir_node *b = get_Mul_right(n);
1102 * Mul is commutative and has again an other neutral element.
1103 * Constants are place right, so check this case first.
1106 if (tarval_is_one(tv)) {
1108 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1111 if (tarval_is_one(tv)) {
1113 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1118 } /* equivalent_node_Mul */
1121 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1123 static ir_node *equivalent_node_Or(ir_node *n) {
1126 ir_node *a = get_Or_left(n);
1127 ir_node *b = get_Or_right(n);
1131 n = a; /* Or has it's own neutral element */
1132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1135 /* constants are cormalized to right, check this site first */
1137 if (tarval_is_null(tv)) {
1139 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1143 if (tarval_is_null(tv)) {
1145 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1150 } /* equivalent_node_Or */
1153 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1155 static ir_node *equivalent_node_And(ir_node *n) {
1158 ir_node *a = get_And_left(n);
1159 ir_node *b = get_And_right(n);
1163 n = a; /* And has it's own neutral element */
1164 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1167 /* constants are cormalized to right, check this site first */
1169 if (tarval_is_all_one(tv)) {
1171 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1175 if (tarval_is_all_one(tv)) {
1177 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1181 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1184 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1189 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1192 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1197 } /* equivalent_node_And */
1200 * Try to remove useless Conv's:
1202 static ir_node *equivalent_node_Conv(ir_node *n) {
1204 ir_node *a = get_Conv_op(n);
1206 ir_mode *n_mode = get_irn_mode(n);
1207 ir_mode *a_mode = get_irn_mode(a);
1210 if (n_mode == a_mode) { /* No Conv necessary */
1211 if (get_Conv_strict(n)) {
1212 /* special case: the predecessor might be a also a Conv */
1214 if (! get_Conv_strict(a)) {
1215 /* first one is not strict, kick it */
1217 a_mode = get_irn_mode(a);
1221 /* else both are strict conv, second is superfluous */
1222 } else if (is_Proj(a)) {
1223 ir_node *pred = get_Proj_pred(a);
1224 if (is_Load(pred)) {
1225 /* loads always return with the exact precision of n_mode */
1226 assert(get_Load_mode(pred) == n_mode);
1229 /* leave strict floating point Conv's */
1232 /* leave strict floating point Conv's */
1237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1238 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1239 ir_node *b = get_Conv_op(a);
1240 ir_mode *b_mode = get_irn_mode(b);
1242 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1243 /* both are strict conv */
1244 if (smaller_mode(a_mode, n_mode)) {
1245 /* both are strict, but the first is smaller, so
1246 the second cannot remove more precision, remove the
1248 set_Conv_strict(n, 0);
1251 if (n_mode == b_mode) {
1252 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1253 if (n_mode == mode_b) {
1254 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1255 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1256 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1257 if (smaller_mode(b_mode, a_mode)) {
1258 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1259 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1263 if (mode_is_int(n_mode) && mode_is_float(a_mode)) {
1264 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1265 size_t int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1266 size_t float_mantissa;
1267 /* FIXME There is no way to get the mantissa size of a mode */
1268 switch (get_mode_size_bits(a_mode)) {
1269 case 32: float_mantissa = 23 + 1; break; // + 1 for implicit 1
1270 case 64: float_mantissa = 52 + 1; break;
1271 case 80: float_mantissa = 64 + 1; break;
1272 default: float_mantissa = 0; break;
1274 if (float_mantissa != 0 && float_mantissa >= int_mantissa) {
1276 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1281 if (smaller_mode(b_mode, a_mode)) {
1282 if (get_Conv_strict(n))
1283 set_Conv_strict(b, 1);
1284 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1285 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1291 } /* equivalent_node_Conv */
1294 * A Cast may be removed if the type of the previous node
1295 * is already the type of the Cast.
1297 static ir_node *equivalent_node_Cast(ir_node *n) {
1299 ir_node *pred = get_Cast_op(n);
1301 if (get_irn_type(pred) == get_Cast_type(n)) {
1303 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1306 } /* equivalent_node_Cast */
1309 * - fold Phi-nodes, iff they have only one predecessor except
1312 static ir_node *equivalent_node_Phi(ir_node *n) {
1317 ir_node *first_val = NULL; /* to shutup gcc */
1319 if (!get_opt_normalize()) return n;
1321 n_preds = get_Phi_n_preds(n);
1323 block = get_nodes_block(n);
1324 if (is_Block_dead(block)) /* Control dead */
1325 return get_irg_bad(current_ir_graph);
1327 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1329 /* Find first non-self-referencing input */
1330 for (i = 0; i < n_preds; ++i) {
1331 first_val = get_Phi_pred(n, i);
1332 if ( (first_val != n) /* not self pointer */
1334 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1335 * predecessors. Then, Phi nodes in dead code might be removed, causing
1336 * nodes pointing to themself (Add's for instance).
1337 * This is really bad and causes endless recursions in several
1338 * code pathes, so we do NOT optimize such a code.
1339 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1340 * (and bad Phi predecessors), so live code is optimized later.
1342 && (! is_Bad(get_Block_cfgpred(block, i)))
1344 ) { /* value not dead */
1345 break; /* then found first value. */
1350 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1351 return get_irg_bad(current_ir_graph);
1354 /* search for rest of inputs, determine if any of these
1355 are non-self-referencing */
1356 while (++i < n_preds) {
1357 ir_node *scnd_val = get_Phi_pred(n, i);
1358 if ( (scnd_val != n)
1359 && (scnd_val != first_val)
1362 && (! is_Bad(get_Block_cfgpred(block, i)))
1370 /* Fold, if no multiple distinct non-self-referencing inputs */
1372 DBG_OPT_PHI(oldn, n);
1375 } /* equivalent_node_Phi */
1378 * Several optimizations:
1379 * - fold Sync-nodes, iff they have only one predecessor except
1382 static ir_node *equivalent_node_Sync(ir_node *n) {
1383 int arity = get_Sync_n_preds(n);
1386 for (i = 0; i < arity;) {
1387 ir_node *pred = get_Sync_pred(n, i);
1390 /* Remove Bad predecessors */
1397 /* Remove duplicate predecessors */
1403 if (get_Sync_pred(n, j) == pred) {
1411 if (arity == 0) return get_irg_bad(current_ir_graph);
1412 if (arity == 1) return get_Sync_pred(n, 0);
1414 } /* equivalent_node_Sync */
1417 * Optimize Proj(Tuple).
1419 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1420 ir_node *oldn = proj;
1421 ir_node *tuple = get_Proj_pred(proj);
1423 /* Remove the Tuple/Proj combination. */
1424 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1425 DBG_OPT_TUPLE(oldn, tuple, proj);
1428 } /* equivalent_node_Proj_Tuple */
1431 * Optimize a / 1 = a.
1433 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1434 ir_node *oldn = proj;
1435 ir_node *div = get_Proj_pred(proj);
1436 ir_node *b = get_Div_right(div);
1437 tarval *tb = value_of(b);
1439 /* Div is not commutative. */
1440 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1441 switch (get_Proj_proj(proj)) {
1443 proj = get_Div_mem(div);
1444 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1448 proj = get_Div_left(div);
1449 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1453 /* we cannot replace the exception Proj's here, this is done in
1454 transform_node_Proj_Div() */
1459 } /* equivalent_node_Proj_Div */
1462 * Optimize a / 1.0 = a.
1464 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1465 ir_node *oldn = proj;
1466 ir_node *quot = get_Proj_pred(proj);
1467 ir_node *b = get_Quot_right(quot);
1468 tarval *tb = value_of(b);
1470 /* Div is not commutative. */
1471 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1472 switch (get_Proj_proj(proj)) {
1474 proj = get_Quot_mem(quot);
1475 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1479 proj = get_Quot_left(quot);
1480 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1484 /* we cannot replace the exception Proj's here, this is done in
1485 transform_node_Proj_Quot() */
1490 } /* equivalent_node_Proj_Quot */
1493 * Optimize a / 1 = a.
1495 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1496 ir_node *oldn = proj;
1497 ir_node *divmod = get_Proj_pred(proj);
1498 ir_node *b = get_DivMod_right(divmod);
1499 tarval *tb = value_of(b);
1501 /* Div is not commutative. */
1502 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1503 switch (get_Proj_proj(proj)) {
1505 proj = get_DivMod_mem(divmod);
1506 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1509 case pn_DivMod_res_div:
1510 proj = get_DivMod_left(divmod);
1511 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1515 /* we cannot replace the exception Proj's here, this is done in
1516 transform_node_Proj_DivMod().
1517 Note further that the pn_DivMod_res_div case is handled in
1518 computed_value_Proj(). */
1523 } /* equivalent_node_Proj_DivMod */
1526 * Optimize CopyB(mem, x, x) into a Nop.
1528 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1529 ir_node *oldn = proj;
1530 ir_node *copyb = get_Proj_pred(proj);
1531 ir_node *a = get_CopyB_dst(copyb);
1532 ir_node *b = get_CopyB_src(copyb);
1535 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1536 switch (get_Proj_proj(proj)) {
1537 case pn_CopyB_M_regular:
1538 proj = get_CopyB_mem(copyb);
1539 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1542 case pn_CopyB_M_except:
1543 case pn_CopyB_X_except:
1544 DBG_OPT_EXC_REM(proj);
1545 proj = get_irg_bad(current_ir_graph);
1550 } /* equivalent_node_Proj_CopyB */
1553 * Optimize Bounds(idx, idx, upper) into idx.
1555 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1556 ir_node *oldn = proj;
1557 ir_node *bound = get_Proj_pred(proj);
1558 ir_node *idx = get_Bound_index(bound);
1559 ir_node *pred = skip_Proj(idx);
1562 if (idx == get_Bound_lower(bound))
1564 else if (is_Bound(pred)) {
1566 * idx was Bounds checked in the same MacroBlock previously,
1567 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1569 ir_node *lower = get_Bound_lower(bound);
1570 ir_node *upper = get_Bound_upper(bound);
1571 if (get_Bound_lower(pred) == lower &&
1572 get_Bound_upper(pred) == upper &&
1573 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1575 * One could expect that we simply return the previous
1576 * Bound here. However, this would be wrong, as we could
1577 * add an exception Proj to a new location then.
1578 * So, we must turn in into a tuple.
1584 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1585 switch (get_Proj_proj(proj)) {
1587 DBG_OPT_EXC_REM(proj);
1588 proj = get_Bound_mem(bound);
1590 case pn_Bound_X_except:
1591 DBG_OPT_EXC_REM(proj);
1592 proj = get_irg_bad(current_ir_graph);
1596 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1599 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1604 } /* equivalent_node_Proj_Bound */
1607 * Optimize an Exception Proj(Load) with a non-null address.
1609 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1610 if (get_opt_ldst_only_null_ptr_exceptions()) {
1611 if (get_irn_mode(proj) == mode_X) {
1612 ir_node *load = get_Proj_pred(proj);
1614 /* get the Load address */
1615 const ir_node *addr = get_Load_ptr(load);
1616 const ir_node *confirm;
1618 if (value_not_null(addr, &confirm)) {
1619 if (get_Proj_proj(proj) == pn_Load_X_except) {
1620 DBG_OPT_EXC_REM(proj);
1621 return get_irg_bad(current_ir_graph);
1627 } /* equivalent_node_Proj_Load */
1630 * Optimize an Exception Proj(Store) with a non-null address.
1632 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1633 if (get_opt_ldst_only_null_ptr_exceptions()) {
1634 if (get_irn_mode(proj) == mode_X) {
1635 ir_node *store = get_Proj_pred(proj);
1637 /* get the load/store address */
1638 const ir_node *addr = get_Store_ptr(store);
1639 const ir_node *confirm;
1641 if (value_not_null(addr, &confirm)) {
1642 if (get_Proj_proj(proj) == pn_Store_X_except) {
1643 DBG_OPT_EXC_REM(proj);
1644 return get_irg_bad(current_ir_graph);
1650 } /* equivalent_node_Proj_Store */
1653 * Does all optimizations on nodes that must be done on it's Proj's
1654 * because of creating new nodes.
1656 static ir_node *equivalent_node_Proj(ir_node *proj) {
1657 ir_node *n = get_Proj_pred(proj);
1659 if (get_irn_mode(proj) == mode_X) {
1660 if (is_Block_dead(get_nodes_block(n))) {
1661 /* Remove dead control flow -- early gigo(). */
1662 return get_irg_bad(current_ir_graph);
1665 if (n->op->ops.equivalent_node_Proj)
1666 return n->op->ops.equivalent_node_Proj(proj);
1668 } /* equivalent_node_Proj */
1673 static ir_node *equivalent_node_Id(ir_node *n) {
1680 DBG_OPT_ID(oldn, n);
1682 } /* equivalent_node_Id */
1687 static ir_node *equivalent_node_Mux(ir_node *n)
1689 ir_node *oldn = n, *sel = get_Mux_sel(n);
1690 tarval *ts = value_of(sel);
1692 /* Mux(true, f, t) == t */
1693 if (ts == tarval_b_true) {
1694 n = get_Mux_true(n);
1695 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1697 /* Mux(false, f, t) == f */
1698 else if (ts == tarval_b_false) {
1699 n = get_Mux_false(n);
1700 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1702 /* Mux(v, x, x) == x */
1703 else if (get_Mux_false(n) == get_Mux_true(n)) {
1704 n = get_Mux_true(n);
1705 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1707 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1708 ir_node *cmp = get_Proj_pred(sel);
1709 long proj_nr = get_Proj_proj(sel);
1710 ir_node *f = get_Mux_false(n);
1711 ir_node *t = get_Mux_true(n);
1714 * Note further that these optimization work even for floating point
1715 * with NaN's because -NaN == NaN.
1716 * However, if +0 and -0 is handled differently, we cannot use the first one.
1719 ir_node *const cmp_l = get_Cmp_left(cmp);
1720 ir_node *const cmp_r = get_Cmp_right(cmp);
1724 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1725 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1727 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1734 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1735 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1737 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1744 * Note: normalization puts the constant on the right side,
1745 * so we check only one case.
1747 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1748 /* Mux(t CMP 0, X, t) */
1749 if (is_Minus(f) && get_Minus_op(f) == t) {
1750 /* Mux(t CMP 0, -t, t) */
1751 if (proj_nr == pn_Cmp_Eq) {
1752 /* Mux(t == 0, -t, t) ==> -t */
1754 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1755 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1756 /* Mux(t != 0, -t, t) ==> t */
1758 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1765 } /* equivalent_node_Mux */
1768 * Remove Confirm nodes if setting is on.
1769 * Replace Confirms(x, '=', Constlike) by Constlike.
1771 static ir_node *equivalent_node_Confirm(ir_node *n) {
1772 ir_node *pred = get_Confirm_value(n);
1773 pn_Cmp pnc = get_Confirm_cmp(n);
1775 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1777 * rare case: two identical Confirms one after another,
1778 * replace the second one with the first.
1781 pred = get_Confirm_value(n);
1782 pnc = get_Confirm_cmp(n);
1784 if (get_opt_remove_confirm())
1785 return get_Confirm_value(n);
1790 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1791 * perform no actual computation, as, e.g., the Id nodes. It does not create
1792 * new nodes. It is therefore safe to free n if the node returned is not n.
1793 * If a node returns a Tuple we can not just skip it. If the size of the
1794 * in array fits, we transform n into a tuple (e.g., Div).
1796 ir_node *equivalent_node(ir_node *n) {
1797 if (n->op->ops.equivalent_node)
1798 return n->op->ops.equivalent_node(n);
1800 } /* equivalent_node */
1803 * Sets the default equivalent node operation for an ir_op_ops.
1805 * @param code the opcode for the default operation
1806 * @param ops the operations initialized
1811 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1815 ops->equivalent_node = equivalent_node_##a; \
1817 #define CASE_PROJ(a) \
1819 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1861 } /* firm_set_default_equivalent_node */
1864 * Returns non-zero if a node is a Phi node
1865 * with all predecessors constant.
1867 static int is_const_Phi(ir_node *n) {
1870 if (! is_Phi(n) || get_irn_arity(n) == 0)
1872 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1873 if (! is_Const(get_irn_n(n, i)))
1876 } /* is_const_Phi */
1878 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1879 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1882 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1884 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1885 if (eval == tarval_sub) {
1886 tarval_sub_type func = (tarval_sub_type)eval;
1888 return func(a, b, mode);
1890 tarval_binop_type func = (tarval_binop_type)eval;
1897 * Apply an evaluator on a binop with a constant operators (and one Phi).
1899 * @param phi the Phi node
1900 * @param other the other operand
1901 * @param eval an evaluator function
1902 * @param mode the mode of the result, may be different from the mode of the Phi!
1903 * @param left if non-zero, other is the left operand, else the right
1905 * @return a new Phi node if the conversion was successful, NULL else
1907 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1912 int i, n = get_irn_arity(phi);
1914 NEW_ARR_A(void *, res, n);
1916 for (i = 0; i < n; ++i) {
1917 pred = get_irn_n(phi, i);
1918 tv = get_Const_tarval(pred);
1919 tv = do_eval(eval, other, tv, mode);
1921 if (tv == tarval_bad) {
1922 /* folding failed, bad */
1928 for (i = 0; i < n; ++i) {
1929 pred = get_irn_n(phi, i);
1930 tv = get_Const_tarval(pred);
1931 tv = do_eval(eval, tv, other, mode);
1933 if (tv == tarval_bad) {
1934 /* folding failed, bad */
1940 irg = current_ir_graph;
1941 for (i = 0; i < n; ++i) {
1942 pred = get_irn_n(phi, i);
1943 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1944 mode, res[i], get_Const_type(pred));
1946 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1947 } /* apply_binop_on_phi */
1950 * Apply an evaluator on a binop with two constant Phi.
1952 * @param a the left Phi node
1953 * @param b the right Phi node
1954 * @param eval an evaluator function
1955 * @param mode the mode of the result, may be different from the mode of the Phi!
1957 * @return a new Phi node if the conversion was successful, NULL else
1959 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1960 tarval *tv_l, *tv_r, *tv;
1966 if (get_nodes_block(a) != get_nodes_block(b))
1969 n = get_irn_arity(a);
1970 NEW_ARR_A(void *, res, n);
1972 for (i = 0; i < n; ++i) {
1973 pred = get_irn_n(a, i);
1974 tv_l = get_Const_tarval(pred);
1975 pred = get_irn_n(b, i);
1976 tv_r = get_Const_tarval(pred);
1977 tv = do_eval(eval, tv_l, tv_r, mode);
1979 if (tv == tarval_bad) {
1980 /* folding failed, bad */
1985 irg = current_ir_graph;
1986 for (i = 0; i < n; ++i) {
1987 pred = get_irn_n(a, i);
1988 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1990 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1991 } /* apply_binop_on_2_phis */
1994 * Apply an evaluator on a unop with a constant operator (a Phi).
1996 * @param phi the Phi node
1997 * @param eval an evaluator function
1999 * @return a new Phi node if the conversion was successful, NULL else
2001 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2007 int i, n = get_irn_arity(phi);
2009 NEW_ARR_A(void *, res, n);
2010 for (i = 0; i < n; ++i) {
2011 pred = get_irn_n(phi, i);
2012 tv = get_Const_tarval(pred);
2015 if (tv == tarval_bad) {
2016 /* folding failed, bad */
2021 mode = get_irn_mode(phi);
2022 irg = current_ir_graph;
2023 for (i = 0; i < n; ++i) {
2024 pred = get_irn_n(phi, i);
2025 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2026 mode, res[i], get_Const_type(pred));
2028 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2029 } /* apply_unop_on_phi */
2032 * Apply a conversion on a constant operator (a Phi).
2034 * @param phi the Phi node
2036 * @return a new Phi node if the conversion was successful, NULL else
2038 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2043 int i, n = get_irn_arity(phi);
2045 NEW_ARR_A(void *, res, n);
2046 for (i = 0; i < n; ++i) {
2047 pred = get_irn_n(phi, i);
2048 tv = get_Const_tarval(pred);
2049 tv = tarval_convert_to(tv, mode);
2051 if (tv == tarval_bad) {
2052 /* folding failed, bad */
2057 irg = current_ir_graph;
2058 for (i = 0; i < n; ++i) {
2059 pred = get_irn_n(phi, i);
2060 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2061 mode, res[i], get_Const_type(pred));
2063 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2064 } /* apply_conv_on_phi */
2067 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2068 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2069 * If possible, remove the Conv's.
2071 static ir_node *transform_node_AddSub(ir_node *n) {
2072 ir_mode *mode = get_irn_mode(n);
2074 if (mode_is_reference(mode)) {
2075 ir_node *left = get_binop_left(n);
2076 ir_node *right = get_binop_right(n);
2077 unsigned ref_bits = get_mode_size_bits(mode);
2079 if (is_Conv(left)) {
2080 ir_mode *lmode = get_irn_mode(left);
2081 unsigned bits = get_mode_size_bits(lmode);
2083 if (ref_bits == bits &&
2084 mode_is_int(lmode) &&
2085 get_mode_arithmetic(lmode) == irma_twos_complement) {
2086 ir_node *pre = get_Conv_op(left);
2087 ir_mode *pre_mode = get_irn_mode(pre);
2089 if (mode_is_int(pre_mode) &&
2090 get_mode_size_bits(pre_mode) == bits &&
2091 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2092 /* ok, this conv just changes to sign, moreover the calculation
2093 * is done with same number of bits as our address mode, so
2094 * we can ignore the conv as address calculation can be viewed
2095 * as either signed or unsigned
2097 set_binop_left(n, pre);
2102 if (is_Conv(right)) {
2103 ir_mode *rmode = get_irn_mode(right);
2104 unsigned bits = get_mode_size_bits(rmode);
2106 if (ref_bits == bits &&
2107 mode_is_int(rmode) &&
2108 get_mode_arithmetic(rmode) == irma_twos_complement) {
2109 ir_node *pre = get_Conv_op(right);
2110 ir_mode *pre_mode = get_irn_mode(pre);
2112 if (mode_is_int(pre_mode) &&
2113 get_mode_size_bits(pre_mode) == bits &&
2114 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2115 /* ok, this conv just changes to sign, moreover the calculation
2116 * is done with same number of bits as our address mode, so
2117 * we can ignore the conv as address calculation can be viewed
2118 * as either signed or unsigned
2120 set_binop_right(n, pre);
2125 /* let address arithmetic use unsigned modes */
2126 if (is_Const(right)) {
2127 ir_mode *rmode = get_irn_mode(right);
2129 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2130 /* convert a AddP(P, *s) into AddP(P, *u) */
2131 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2133 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2134 set_binop_right(n, pre);
2140 } /* transform_node_AddSub */
2142 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2144 if (is_Const(b) && is_const_Phi(a)) { \
2145 /* check for Op(Phi, Const) */ \
2146 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2148 else if (is_Const(a) && is_const_Phi(b)) { \
2149 /* check for Op(Const, Phi) */ \
2150 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2152 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2153 /* check for Op(Phi, Phi) */ \
2154 c = apply_binop_on_2_phis(a, b, eval, mode); \
2157 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2161 #define HANDLE_UNOP_PHI(eval, a, c) \
2163 if (is_const_Phi(a)) { \
2164 /* check for Op(Phi) */ \
2165 c = apply_unop_on_phi(a, eval); \
2167 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2173 * Do the AddSub optimization, then Transform
2174 * Constant folding on Phi
2175 * Add(a,a) -> Mul(a, 2)
2176 * Add(Mul(a, x), a) -> Mul(a, x+1)
2177 * if the mode is integer or float.
2178 * Transform Add(a,-b) into Sub(a,b).
2179 * Reassociation might fold this further.
2181 static ir_node *transform_node_Add(ir_node *n) {
2183 ir_node *a, *b, *c, *oldn = n;
2185 n = transform_node_AddSub(n);
2187 a = get_Add_left(n);
2188 b = get_Add_right(n);
2190 mode = get_irn_mode(n);
2192 if (mode_is_reference(mode)) {
2193 ir_mode *lmode = get_irn_mode(a);
2195 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2196 /* an Add(a, NULL) is a hidden Conv */
2197 dbg_info *dbg = get_irn_dbg_info(n);
2198 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2202 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2204 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2205 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2208 if (mode_is_num(mode)) {
2209 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2210 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2211 ir_node *block = get_nodes_block(n);
2214 get_irn_dbg_info(n),
2218 new_r_Const_long(current_ir_graph, block, mode, 2),
2220 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2225 get_irn_dbg_info(n),
2231 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2236 get_irn_dbg_info(n),
2242 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2245 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2246 /* Here we rely on constants be on the RIGHT side */
2248 ir_node *op = get_Not_op(a);
2250 if (is_Const(b) && is_Const_one(b)) {
2252 ir_node *blk = get_irn_n(n, -1);
2253 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2254 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2259 ir_node *blk = get_irn_n(n, -1);
2260 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2266 ir_node *op = get_Not_op(b);
2270 ir_node *blk = get_irn_n(n, -1);
2271 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2272 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2279 } /* transform_node_Add */
2282 * returns -cnst or NULL if impossible
2284 static ir_node *const_negate(ir_node *cnst) {
2285 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2286 dbg_info *dbgi = get_irn_dbg_info(cnst);
2287 ir_graph *irg = get_irn_irg(cnst);
2288 ir_node *block = get_nodes_block(cnst);
2289 ir_mode *mode = get_irn_mode(cnst);
2290 if (tv == tarval_bad) return NULL;
2291 return new_rd_Const(dbgi, irg, block, mode, tv);
2295 * Do the AddSub optimization, then Transform
2296 * Constant folding on Phi
2297 * Sub(0,a) -> Minus(a)
2298 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2299 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2300 * Sub(Add(a, x), x) -> a
2301 * Sub(x, Add(x, a)) -> -a
2302 * Sub(x, Const) -> Add(x, -Const)
2304 static ir_node *transform_node_Sub(ir_node *n) {
2309 n = transform_node_AddSub(n);
2311 a = get_Sub_left(n);
2312 b = get_Sub_right(n);
2314 mode = get_irn_mode(n);
2316 if (mode_is_int(mode)) {
2317 ir_mode *lmode = get_irn_mode(a);
2319 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2320 /* a Sub(a, NULL) is a hidden Conv */
2321 dbg_info *dbg = get_irn_dbg_info(n);
2322 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2327 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2329 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2330 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2333 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2334 /* a - C -> a + (-C) */
2335 ir_node *cnst = const_negate(b);
2337 ir_node *block = get_nodes_block(n);
2338 dbg_info *dbgi = get_irn_dbg_info(n);
2339 ir_graph *irg = get_irn_irg(n);
2341 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2342 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2347 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2348 ir_graph *irg = current_ir_graph;
2349 dbg_info *dbg = get_irn_dbg_info(n);
2350 ir_node *block = get_nodes_block(n);
2351 ir_node *left = get_Minus_op(a);
2352 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2354 n = new_rd_Minus(dbg, irg, block, add, mode);
2355 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2357 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2358 ir_graph *irg = current_ir_graph;
2359 dbg_info *dbg = get_irn_dbg_info(n);
2360 ir_node *block = get_nodes_block(n);
2361 ir_node *right = get_Minus_op(b);
2363 n = new_rd_Add(dbg, irg, block, a, right, mode);
2364 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2366 } else if (is_Sub(b)) {
2367 /* a - (b - c) -> a + (c - b)
2368 * -> (a - b) + c iff (b - c) is a pointer */
2369 ir_graph *irg = current_ir_graph;
2370 dbg_info *s_dbg = get_irn_dbg_info(b);
2371 ir_node *s_block = get_nodes_block(b);
2372 ir_node *s_left = get_Sub_left(b);
2373 ir_node *s_right = get_Sub_right(b);
2374 ir_mode *s_mode = get_irn_mode(b);
2375 if (s_mode == mode_P) {
2376 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2377 dbg_info *a_dbg = get_irn_dbg_info(n);
2378 ir_node *a_block = get_nodes_block(n);
2381 s_right = new_r_Conv(irg, a_block, s_right, mode);
2382 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2384 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2385 dbg_info *a_dbg = get_irn_dbg_info(n);
2386 ir_node *a_block = get_nodes_block(n);
2388 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2390 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2392 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2393 ir_node *m_right = get_Mul_right(b);
2394 if (is_Const(m_right)) {
2395 ir_node *cnst2 = const_negate(m_right);
2396 if (cnst2 != NULL) {
2397 ir_graph *irg = current_ir_graph;
2398 dbg_info *m_dbg = get_irn_dbg_info(b);
2399 ir_node *m_block = get_nodes_block(b);
2400 ir_node *m_left = get_Mul_left(b);
2401 ir_mode *m_mode = get_irn_mode(b);
2402 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2403 dbg_info *a_dbg = get_irn_dbg_info(n);
2404 ir_node *a_block = get_nodes_block(n);
2406 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2407 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2413 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2414 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2416 get_irn_dbg_info(n),
2421 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2425 if (mode_wrap_around(mode)) {
2426 ir_node *left = get_Add_left(a);
2427 ir_node *right = get_Add_right(a);
2429 /* FIXME: Does the Conv's work only for two complement or generally? */
2431 if (mode != get_irn_mode(right)) {
2432 /* This Sub is an effective Cast */
2433 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2436 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2438 } else if (right == b) {
2439 if (mode != get_irn_mode(left)) {
2440 /* This Sub is an effective Cast */
2441 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2444 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2450 if (mode_wrap_around(mode)) {
2451 ir_node *left = get_Add_left(b);
2452 ir_node *right = get_Add_right(b);
2454 /* FIXME: Does the Conv's work only for two complement or generally? */
2456 ir_mode *r_mode = get_irn_mode(right);
2458 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2459 if (mode != r_mode) {
2460 /* This Sub is an effective Cast */
2461 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2463 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2465 } else if (right == a) {
2466 ir_mode *l_mode = get_irn_mode(left);
2468 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2469 if (mode != l_mode) {
2470 /* This Sub is an effective Cast */
2471 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2473 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2478 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2479 ir_mode *mode = get_irn_mode(a);
2481 if (mode == get_irn_mode(b)) {
2483 ir_node *op_a = get_Conv_op(a);
2484 ir_node *op_b = get_Conv_op(b);
2486 /* check if it's allowed to skip the conv */
2487 ma = get_irn_mode(op_a);
2488 mb = get_irn_mode(op_b);
2490 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2491 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2494 set_Sub_right(n, b);
2500 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2501 if (!is_reassoc_running() && is_Mul(a)) {
2502 ir_node *ma = get_Mul_left(a);
2503 ir_node *mb = get_Mul_right(a);
2506 ir_node *blk = get_irn_n(n, -1);
2508 get_irn_dbg_info(n),
2509 current_ir_graph, blk,
2512 get_irn_dbg_info(n),
2513 current_ir_graph, blk,
2515 new_r_Const_long(current_ir_graph, blk, mode, 1),
2518 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2520 } else if (mb == b) {
2521 ir_node *blk = get_irn_n(n, -1);
2523 get_irn_dbg_info(n),
2524 current_ir_graph, blk,
2527 get_irn_dbg_info(n),
2528 current_ir_graph, blk,
2530 new_r_Const_long(current_ir_graph, blk, mode, 1),
2533 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2537 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2538 ir_node *x = get_Sub_left(a);
2539 ir_node *y = get_Sub_right(a);
2540 ir_node *blk = get_irn_n(n, -1);
2541 ir_mode *m_b = get_irn_mode(b);
2542 ir_mode *m_y = get_irn_mode(y);
2546 /* Determine the right mode for the Add. */
2549 else if (mode_is_reference(m_b))
2551 else if (mode_is_reference(m_y))
2555 * Both modes are different but none is reference,
2556 * happens for instance in SubP(SubP(P, Iu), Is).
2557 * We have two possibilities here: Cast or ignore.
2558 * Currently we ignore this case.
2563 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2565 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2566 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2570 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2571 if (is_Const(a) && is_Not(b)) {
2572 /* c - ~X = X + (c+1) */
2573 tarval *tv = get_Const_tarval(a);
2575 tv = tarval_add(tv, get_mode_one(mode));
2576 if (tv != tarval_bad) {
2577 ir_node *blk = get_irn_n(n, -1);
2578 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2579 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2580 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2586 } /* transform_node_Sub */
2589 * Several transformation done on n*n=2n bits mul.
2590 * These transformations must be done here because new nodes may be produced.
2592 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2594 ir_node *a = get_Mul_left(n);
2595 ir_node *b = get_Mul_right(n);
2596 tarval *ta = value_of(a);
2597 tarval *tb = value_of(b);
2598 ir_mode *smode = get_irn_mode(a);
2600 if (ta == get_mode_one(smode)) {
2601 /* (L)1 * (L)b = (L)b */
2602 ir_node *blk = get_irn_n(n, -1);
2603 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2604 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2607 else if (ta == get_mode_minus_one(smode)) {
2608 /* (L)-1 * (L)b = (L)b */
2609 ir_node *blk = get_irn_n(n, -1);
2610 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2611 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2612 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2615 if (tb == get_mode_one(smode)) {
2616 /* (L)a * (L)1 = (L)a */
2617 ir_node *blk = get_irn_n(a, -1);
2618 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2619 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2622 else if (tb == get_mode_minus_one(smode)) {
2623 /* (L)a * (L)-1 = (L)-a */
2624 ir_node *blk = get_irn_n(n, -1);
2625 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2626 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2627 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2634 * Transform Mul(a,-1) into -a.
2635 * Do constant evaluation of Phi nodes.
2636 * Do architecture dependent optimizations on Mul nodes
2638 static ir_node *transform_node_Mul(ir_node *n) {
2639 ir_node *c, *oldn = n;
2640 ir_mode *mode = get_irn_mode(n);
2641 ir_node *a = get_Mul_left(n);
2642 ir_node *b = get_Mul_right(n);
2644 if (is_Bad(a) || is_Bad(b))
2647 if (mode != get_irn_mode(a))
2648 return transform_node_Mul2n(n, mode);
2650 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2652 if (mode_is_signed(mode)) {
2655 if (value_of(a) == get_mode_minus_one(mode))
2657 else if (value_of(b) == get_mode_minus_one(mode))
2660 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2661 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2666 if (is_Const(b)) { /* (-a) * const -> a * -const */
2667 ir_node *cnst = const_negate(b);
2669 dbg_info *dbgi = get_irn_dbg_info(n);
2670 ir_node *block = get_nodes_block(n);
2671 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2672 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2675 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2676 dbg_info *dbgi = get_irn_dbg_info(n);
2677 ir_node *block = get_nodes_block(n);
2678 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2679 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2681 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2682 ir_node *sub_l = get_Sub_left(b);
2683 ir_node *sub_r = get_Sub_right(b);
2684 dbg_info *dbgi = get_irn_dbg_info(n);
2685 ir_graph *irg = current_ir_graph;
2686 ir_node *block = get_nodes_block(n);
2687 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2688 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2689 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2692 } else if (is_Minus(b)) {
2693 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2694 ir_node *sub_l = get_Sub_left(a);
2695 ir_node *sub_r = get_Sub_right(a);
2696 dbg_info *dbgi = get_irn_dbg_info(n);
2697 ir_graph *irg = current_ir_graph;
2698 ir_node *block = get_nodes_block(n);
2699 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2700 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2701 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2705 if (get_mode_arithmetic(mode) == irma_ieee754) {
2707 tarval *tv = get_Const_tarval(a);
2708 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2709 /* 2.0 * b = b + b */
2710 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2711 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2715 else if (is_Const(b)) {
2716 tarval *tv = get_Const_tarval(b);
2717 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2718 /* a * 2.0 = a + a */
2719 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2720 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2725 return arch_dep_replace_mul_with_shifts(n);
2726 } /* transform_node_Mul */
2729 * Transform a Div Node.
2731 static ir_node *transform_node_Div(ir_node *n) {
2732 ir_mode *mode = get_Div_resmode(n);
2733 ir_node *a = get_Div_left(n);
2734 ir_node *b = get_Div_right(n);
2738 if (is_Const(b) && is_const_Phi(a)) {
2739 /* check for Div(Phi, Const) */
2740 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2742 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2746 else if (is_Const(a) && is_const_Phi(b)) {
2747 /* check for Div(Const, Phi) */
2748 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2750 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2754 else if (is_const_Phi(a) && is_const_Phi(b)) {
2755 /* check for Div(Phi, Phi) */
2756 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2758 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2765 if (tv != tarval_bad) {
2766 value = new_Const(get_tarval_mode(tv), tv);
2768 DBG_OPT_CSTEVAL(n, value);
2771 ir_node *a = get_Div_left(n);
2772 ir_node *b = get_Div_right(n);
2773 const ir_node *dummy;
2775 if (a == b && value_not_zero(a, &dummy)) {
2776 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2777 value = new_Const(mode, get_mode_one(mode));
2778 DBG_OPT_CSTEVAL(n, value);
2781 if (mode_is_signed(mode) && is_Const(b)) {
2782 tarval *tv = get_Const_tarval(b);
2784 if (tv == get_mode_minus_one(mode)) {
2786 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2787 DBG_OPT_CSTEVAL(n, value);
2791 /* Try architecture dependent optimization */
2792 value = arch_dep_replace_div_by_const(n);
2800 /* Turn Div into a tuple (mem, jmp, bad, value) */
2801 mem = get_Div_mem(n);
2802 blk = get_irn_n(n, -1);
2804 /* skip a potential Pin */
2806 mem = get_Pin_op(mem);
2807 turn_into_tuple(n, pn_Div_max);
2808 set_Tuple_pred(n, pn_Div_M, mem);
2809 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2810 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2811 set_Tuple_pred(n, pn_Div_res, value);
2814 } /* transform_node_Div */
2817 * Transform a Mod node.
2819 static ir_node *transform_node_Mod(ir_node *n) {
2820 ir_mode *mode = get_Mod_resmode(n);
2821 ir_node *a = get_Mod_left(n);
2822 ir_node *b = get_Mod_right(n);
2826 if (is_Const(b) && is_const_Phi(a)) {
2827 /* check for Div(Phi, Const) */
2828 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2830 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2834 else if (is_Const(a) && is_const_Phi(b)) {
2835 /* check for Div(Const, Phi) */
2836 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2838 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2842 else if (is_const_Phi(a) && is_const_Phi(b)) {
2843 /* check for Div(Phi, Phi) */
2844 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2846 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2853 if (tv != tarval_bad) {
2854 value = new_Const(get_tarval_mode(tv), tv);
2856 DBG_OPT_CSTEVAL(n, value);
2859 ir_node *a = get_Mod_left(n);
2860 ir_node *b = get_Mod_right(n);
2861 const ir_node *dummy;
2863 if (a == b && value_not_zero(a, &dummy)) {
2864 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2865 value = new_Const(mode, get_mode_null(mode));
2866 DBG_OPT_CSTEVAL(n, value);
2869 if (mode_is_signed(mode) && is_Const(b)) {
2870 tarval *tv = get_Const_tarval(b);
2872 if (tv == get_mode_minus_one(mode)) {
2874 value = new_Const(mode, get_mode_null(mode));
2875 DBG_OPT_CSTEVAL(n, value);
2879 /* Try architecture dependent optimization */
2880 value = arch_dep_replace_mod_by_const(n);
2888 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2889 mem = get_Mod_mem(n);
2890 blk = get_irn_n(n, -1);
2892 /* skip a potential Pin */
2894 mem = get_Pin_op(mem);
2895 turn_into_tuple(n, pn_Mod_max);
2896 set_Tuple_pred(n, pn_Mod_M, mem);
2897 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2898 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2899 set_Tuple_pred(n, pn_Mod_res, value);
2902 } /* transform_node_Mod */
2905 * Transform a DivMod node.
2907 static ir_node *transform_node_DivMod(ir_node *n) {
2908 const ir_node *dummy;
2909 ir_node *a = get_DivMod_left(n);
2910 ir_node *b = get_DivMod_right(n);
2911 ir_mode *mode = get_DivMod_resmode(n);
2916 if (is_Const(b) && is_const_Phi(a)) {
2917 /* check for Div(Phi, Const) */
2918 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2919 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2921 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2922 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2926 else if (is_Const(a) && is_const_Phi(b)) {
2927 /* check for Div(Const, Phi) */
2928 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2929 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2931 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2932 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2936 else if (is_const_Phi(a) && is_const_Phi(b)) {
2937 /* check for Div(Phi, Phi) */
2938 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2939 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2941 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2942 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2949 if (tb != tarval_bad) {
2950 if (tb == get_mode_one(get_tarval_mode(tb))) {
2952 vb = new_Const(mode, get_mode_null(mode));
2953 DBG_OPT_CSTEVAL(n, vb);
2955 } else if (ta != tarval_bad) {
2956 tarval *resa, *resb;
2957 resa = tarval_div(ta, tb);
2958 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2959 Jmp for X result!? */
2960 resb = tarval_mod(ta, tb);
2961 if (resb == tarval_bad) return n; /* Causes exception! */
2962 va = new_Const(mode, resa);
2963 vb = new_Const(mode, resb);
2964 DBG_OPT_CSTEVAL(n, va);
2965 DBG_OPT_CSTEVAL(n, vb);
2967 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2968 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2969 vb = new_Const(mode, get_mode_null(mode));
2970 DBG_OPT_CSTEVAL(n, va);
2971 DBG_OPT_CSTEVAL(n, vb);
2973 } else { /* Try architecture dependent optimization */
2976 arch_dep_replace_divmod_by_const(&va, &vb, n);
2977 evaluated = va != NULL;
2979 } else if (a == b) {
2980 if (value_not_zero(a, &dummy)) {
2982 va = new_Const(mode, get_mode_one(mode));
2983 vb = new_Const(mode, get_mode_null(mode));
2984 DBG_OPT_CSTEVAL(n, va);
2985 DBG_OPT_CSTEVAL(n, vb);
2988 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2991 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2992 /* 0 / non-Const = 0 */
2997 if (evaluated) { /* replace by tuple */
3001 mem = get_DivMod_mem(n);
3002 /* skip a potential Pin */
3004 mem = get_Pin_op(mem);
3006 blk = get_irn_n(n, -1);
3007 turn_into_tuple(n, pn_DivMod_max);
3008 set_Tuple_pred(n, pn_DivMod_M, mem);
3009 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3010 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3011 set_Tuple_pred(n, pn_DivMod_res_div, va);
3012 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3016 } /* transform_node_DivMod */
3019 * Optimize x / c to x * (1/c)
3021 static ir_node *transform_node_Quot(ir_node *n) {
3022 ir_mode *mode = get_Quot_resmode(n);
3025 if (get_mode_arithmetic(mode) == irma_ieee754) {
3026 ir_node *b = get_Quot_right(n);
3029 tarval *tv = get_Const_tarval(b);
3033 * Floating point constant folding might be disabled here to
3035 * However, as we check for exact result, doing it is safe.
3038 rem = tarval_enable_fp_ops(1);
3039 tv = tarval_quo(get_mode_one(mode), tv);
3040 (void)tarval_enable_fp_ops(rem);
3042 /* Do the transformation if the result is either exact or we are not
3043 using strict rules. */
3044 if (tv != tarval_bad &&
3045 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3046 ir_node *blk = get_irn_n(n, -1);
3047 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3048 ir_node *a = get_Quot_left(n);
3049 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3050 ir_node *mem = get_Quot_mem(n);
3052 /* skip a potential Pin */
3054 mem = get_Pin_op(mem);
3055 turn_into_tuple(n, pn_Quot_max);
3056 set_Tuple_pred(n, pn_Quot_M, mem);
3057 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3058 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3059 set_Tuple_pred(n, pn_Quot_res, m);
3060 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3065 } /* transform_node_Quot */
3068 * Optimize Abs(x) into x if x is Confirmed >= 0
3069 * Optimize Abs(x) into -x if x is Confirmed <= 0
3070 * Optimize Abs(-x) int Abs(x)
3072 static ir_node *transform_node_Abs(ir_node *n) {
3073 ir_node *c, *oldn = n;
3074 ir_node *a = get_Abs_op(n);
3077 HANDLE_UNOP_PHI(tarval_abs, a, c);
3079 switch (classify_value_sign(a)) {
3080 case value_classified_negative:
3081 mode = get_irn_mode(n);
3084 * We can replace the Abs by -x here.
3085 * We even could add a new Confirm here
3086 * (if not twos complement)
3088 * Note that -x would create a new node, so we could
3089 * not run it in the equivalent_node() context.
3091 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3092 get_nodes_block(n), a, mode);
3094 DBG_OPT_CONFIRM(oldn, n);
3096 case value_classified_positive:
3097 /* n is positive, Abs is not needed */
3100 DBG_OPT_CONFIRM(oldn, n);
3106 /* Abs(-x) = Abs(x) */
3107 mode = get_irn_mode(n);
3108 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3109 get_nodes_block(n), get_Minus_op(a), mode);
3110 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3114 } /* transform_node_Abs */
3117 * Optimize -a CMP -b into b CMP a.
3118 * This works only for for modes where unary Minus
3120 * Note that two-complement integers can Overflow
3121 * so it will NOT work.
3123 * For == and != can be handled in Proj(Cmp)
3125 static ir_node *transform_node_Cmp(ir_node *n) {
3127 ir_node *left = get_Cmp_left(n);
3128 ir_node *right = get_Cmp_right(n);
3130 if (is_Minus(left) && is_Minus(right) &&
3131 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3132 ir_node *const new_left = get_Minus_op(right);
3133 ir_node *const new_right = get_Minus_op(left);
3134 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3135 get_nodes_block(n), new_left, new_right);
3136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3139 } /* transform_node_Cmp */
3143 * Transform a Cond node.
3145 * Replace the Cond by a Jmp if it branches on a constant
3148 static ir_node *transform_node_Cond(ir_node *n) {
3151 ir_node *a = get_Cond_selector(n);
3152 tarval *ta = value_of(a);
3154 /* we need block info which is not available in floating irgs */
3155 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3158 if ((ta != tarval_bad) &&
3159 (get_irn_mode(a) == mode_b) &&
3160 (get_opt_unreachable_code())) {
3161 /* It's a boolean Cond, branching on a boolean constant.
3162 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3163 ir_node *blk = get_nodes_block(n);
3164 jmp = new_r_Jmp(current_ir_graph, blk);
3165 turn_into_tuple(n, pn_Cond_max);
3166 if (ta == tarval_b_true) {
3167 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3168 set_Tuple_pred(n, pn_Cond_true, jmp);
3170 set_Tuple_pred(n, pn_Cond_false, jmp);
3171 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3173 /* We might generate an endless loop, so keep it alive. */
3174 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3177 } /* transform_node_Cond */
3180 * Prototype of a recursive transform function
3181 * for bitwise distributive transformations.
3183 typedef ir_node* (*recursive_transform)(ir_node *n);
3186 * makes use of distributive laws for and, or, eor
3187 * and(a OP c, b OP c) -> and(a, b) OP c
3188 * note, might return a different op than n
3190 static ir_node *transform_bitwise_distributive(ir_node *n,
3191 recursive_transform trans_func)
3194 ir_node *a = get_binop_left(n);
3195 ir_node *b = get_binop_right(n);
3196 ir_op *op = get_irn_op(a);
3197 ir_op *op_root = get_irn_op(n);
3199 if(op != get_irn_op(b))
3202 if (op == op_Conv) {
3203 ir_node *a_op = get_Conv_op(a);
3204 ir_node *b_op = get_Conv_op(b);
3205 ir_mode *a_mode = get_irn_mode(a_op);
3206 ir_mode *b_mode = get_irn_mode(b_op);
3207 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3208 ir_node *blk = get_irn_n(n, -1);
3211 set_binop_left(n, a_op);
3212 set_binop_right(n, b_op);
3213 set_irn_mode(n, a_mode);
3215 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3217 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3223 /* nothing to gain here */
3227 if (op == op_Shrs || op == op_Shr || op == op_Shl
3228 || op == op_And || op == op_Or || op == op_Eor) {
3229 ir_node *a_left = get_binop_left(a);
3230 ir_node *a_right = get_binop_right(a);
3231 ir_node *b_left = get_binop_left(b);
3232 ir_node *b_right = get_binop_right(b);
3234 ir_node *op1 = NULL;
3235 ir_node *op2 = NULL;
3237 if (is_op_commutative(op)) {
3238 if (a_left == b_left) {
3242 } else if(a_left == b_right) {
3246 } else if(a_right == b_left) {
3252 if(a_right == b_right) {
3259 /* (a sop c) & (b sop c) => (a & b) sop c */
3260 ir_node *blk = get_irn_n(n, -1);
3262 ir_node *new_n = exact_copy(n);
3263 set_binop_left(new_n, op1);
3264 set_binop_right(new_n, op2);
3265 new_n = trans_func(new_n);
3267 if(op_root == op_Eor && op == op_Or) {
3268 dbg_info *dbgi = get_irn_dbg_info(n);
3269 ir_graph *irg = current_ir_graph;
3270 ir_mode *mode = get_irn_mode(c);
3272 c = new_rd_Not(dbgi, irg, blk, c, mode);
3273 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3276 set_nodes_block(n, blk);
3277 set_binop_left(n, new_n);
3278 set_binop_right(n, c);
3279 add_identities(current_ir_graph->value_table, n);
3282 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3293 static ir_node *transform_node_And(ir_node *n) {
3294 ir_node *c, *oldn = n;
3295 ir_node *a = get_And_left(n);
3296 ir_node *b = get_And_right(n);
3299 mode = get_irn_mode(n);
3300 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3302 /* we can evaluate 2 Projs of the same Cmp */
3303 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3304 ir_node *pred_a = get_Proj_pred(a);
3305 ir_node *pred_b = get_Proj_pred(b);
3306 if (pred_a == pred_b) {
3307 dbg_info *dbgi = get_irn_dbg_info(n);
3308 ir_node *block = get_nodes_block(pred_a);
3309 pn_Cmp pn_a = get_Proj_proj(a);
3310 pn_Cmp pn_b = get_Proj_proj(b);
3311 /* yes, we can simply calculate with pncs */
3312 pn_Cmp new_pnc = pn_a & pn_b;
3314 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3319 ir_node *op = get_Not_op(b);
3321 ir_node *ba = get_And_left(op);
3322 ir_node *bb = get_And_right(op);
3324 /* it's enough to test the following cases due to normalization! */
3325 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3326 /* (a|b) & ~(a&b) = a^b */
3327 ir_node *block = get_nodes_block(n);
3329 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3330 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3338 ir_node *op = get_Not_op(a);
3340 ir_node *aa = get_And_left(op);
3341 ir_node *ab = get_And_right(op);
3343 /* it's enough to test the following cases due to normalization! */
3344 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3345 /* (a|b) & ~(a&b) = a^b */
3346 ir_node *block = get_nodes_block(n);
3348 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3349 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3356 ir_node *al = get_Eor_left(a);
3357 ir_node *ar = get_Eor_right(a);
3360 /* (b ^ a) & b -> ~a & b */
3361 dbg_info *dbg = get_irn_dbg_info(n);
3362 ir_node *block = get_nodes_block(n);
3364 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3365 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3366 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3370 /* (a ^ b) & b -> ~a & b */
3371 dbg_info *dbg = get_irn_dbg_info(n);
3372 ir_node *block = get_nodes_block(n);
3374 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3375 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3376 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3381 ir_node *bl = get_Eor_left(b);
3382 ir_node *br = get_Eor_right(b);
3385 /* a & (a ^ b) -> a & ~b */
3386 dbg_info *dbg = get_irn_dbg_info(n);
3387 ir_node *block = get_nodes_block(n);
3389 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3390 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3391 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3395 /* a & (b ^ a) -> a & ~b */
3396 dbg_info *dbg = get_irn_dbg_info(n);
3397 ir_node *block = get_nodes_block(n);
3399 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3400 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3401 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3405 if (is_Not(a) && is_Not(b)) {
3406 /* ~a & ~b = ~(a|b) */
3407 ir_node *block = get_nodes_block(n);
3408 ir_mode *mode = get_irn_mode(n);
3412 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3413 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3414 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3418 n = transform_bitwise_distributive(n, transform_node_And);
3421 } /* transform_node_And */
3426 static ir_node *transform_node_Eor(ir_node *n) {
3427 ir_node *c, *oldn = n;
3428 ir_node *a = get_Eor_left(n);
3429 ir_node *b = get_Eor_right(n);
3430 ir_mode *mode = get_irn_mode(n);
3432 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3434 /* we can evaluate 2 Projs of the same Cmp */
3435 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3436 ir_node *pred_a = get_Proj_pred(a);
3437 ir_node *pred_b = get_Proj_pred(b);
3438 if(pred_a == pred_b) {
3439 dbg_info *dbgi = get_irn_dbg_info(n);
3440 ir_node *block = get_nodes_block(pred_a);
3441 pn_Cmp pn_a = get_Proj_proj(a);
3442 pn_Cmp pn_b = get_Proj_proj(b);
3443 /* yes, we can simply calculate with pncs */
3444 pn_Cmp new_pnc = pn_a ^ pn_b;
3446 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3453 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3454 mode, get_mode_null(mode));
3455 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3456 } else if (mode == mode_b &&
3458 is_Const(b) && is_Const_one(b) &&
3459 is_Cmp(get_Proj_pred(a))) {
3460 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3461 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3462 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3464 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3465 } else if (is_Const(b)) {
3466 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3467 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3468 ir_node *not_op = get_Not_op(a);
3469 dbg_info *dbg = get_irn_dbg_info(n);
3470 ir_graph *irg = current_ir_graph;
3471 ir_node *block = get_nodes_block(n);
3472 ir_mode *mode = get_irn_mode(n);
3473 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3475 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3476 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3477 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3480 n = transform_bitwise_distributive(n, transform_node_Eor);
3484 } /* transform_node_Eor */
3489 static ir_node *transform_node_Not(ir_node *n) {
3490 ir_node *c, *oldn = n;
3491 ir_node *a = get_Not_op(n);
3492 ir_mode *mode = get_irn_mode(n);
3494 HANDLE_UNOP_PHI(tarval_not,a,c);
3496 /* check for a boolean Not */
3497 if (mode == mode_b &&
3499 is_Cmp(get_Proj_pred(a))) {
3500 /* We negate a Cmp. The Cmp has the negated result anyways! */
3501 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3502 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3503 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3507 ir_node *eor_b = get_Eor_right(a);
3508 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3509 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3510 ir_node *eor_a = get_Eor_left(a);
3511 dbg_info *dbg = get_irn_dbg_info(n);
3512 ir_graph *irg = current_ir_graph;
3513 ir_node *block = get_nodes_block(n);
3514 ir_mode *mode = get_irn_mode(n);
3515 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3519 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3520 if (is_Minus(a)) { /* ~-x -> x + -1 */
3521 dbg_info *dbg = get_irn_dbg_info(n);
3522 ir_graph *irg = current_ir_graph;
3523 ir_node *block = get_nodes_block(n);
3524 ir_node *add_l = get_Minus_op(a);
3525 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3526 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3527 } else if (is_Add(a)) {
3528 ir_node *add_r = get_Add_right(a);
3529 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3530 /* ~(x + -1) = -x */
3531 ir_node *op = get_Add_left(a);
3532 ir_node *blk = get_irn_n(n, -1);
3533 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3534 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3539 } /* transform_node_Not */
3542 * Transform a Minus.
3546 * -(a >>u (size-1)) = a >>s (size-1)
3547 * -(a >>s (size-1)) = a >>u (size-1)
3548 * -(a * const) -> a * -const
3550 static ir_node *transform_node_Minus(ir_node *n) {
3551 ir_node *c, *oldn = n;
3552 ir_node *a = get_Minus_op(n);
3555 HANDLE_UNOP_PHI(tarval_neg,a,c);
3557 mode = get_irn_mode(a);
3558 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3559 /* the following rules are only to twos-complement */
3562 ir_node *op = get_Not_op(a);
3563 tarval *tv = get_mode_one(mode);
3564 ir_node *blk = get_irn_n(n, -1);
3565 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3566 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3567 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3571 ir_node *c = get_Shr_right(a);
3574 tarval *tv = get_Const_tarval(c);
3576 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3577 /* -(a >>u (size-1)) = a >>s (size-1) */
3578 ir_node *v = get_Shr_left(a);
3580 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3581 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3587 ir_node *c = get_Shrs_right(a);
3590 tarval *tv = get_Const_tarval(c);
3592 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3593 /* -(a >>s (size-1)) = a >>u (size-1) */
3594 ir_node *v = get_Shrs_left(a);
3596 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3597 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3604 /* - (a-b) = b - a */
3605 ir_node *la = get_Sub_left(a);
3606 ir_node *ra = get_Sub_right(a);
3607 ir_node *blk = get_irn_n(n, -1);
3609 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3610 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3614 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3615 ir_node *mul_l = get_Mul_left(a);
3616 ir_node *mul_r = get_Mul_right(a);
3617 if (is_Const(mul_r)) {
3618 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3619 if(tv != tarval_bad) {
3620 ir_node *cnst = new_Const(mode, tv);
3621 dbg_info *dbg = get_irn_dbg_info(a);
3622 ir_graph *irg = current_ir_graph;
3623 ir_node *block = get_nodes_block(a);
3624 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3625 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3632 } /* transform_node_Minus */
3635 * Transform a Cast_type(Const) into a new Const_type
3637 static ir_node *transform_node_Cast(ir_node *n) {
3639 ir_node *pred = get_Cast_op(n);
3640 ir_type *tp = get_irn_type(n);
3642 if (is_Const(pred) && get_Const_type(pred) != tp) {
3643 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3644 get_Const_tarval(pred), tp);
3645 DBG_OPT_CSTEVAL(oldn, n);
3646 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3647 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3648 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3649 DBG_OPT_CSTEVAL(oldn, n);
3653 } /* transform_node_Cast */
3656 * Transform a Proj(Load) with a non-null address.
3658 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3659 if (get_opt_ldst_only_null_ptr_exceptions()) {
3660 if (get_irn_mode(proj) == mode_X) {
3661 ir_node *load = get_Proj_pred(proj);
3663 /* get the Load address */
3664 const ir_node *addr = get_Load_ptr(load);
3665 const ir_node *confirm;
3667 if (value_not_null(addr, &confirm)) {
3668 if (confirm == NULL) {
3669 /* this node may float if it did not depend on a Confirm */
3670 set_irn_pinned(load, op_pin_state_floats);
3672 if (get_Proj_proj(proj) == pn_Load_X_except) {
3673 DBG_OPT_EXC_REM(proj);
3674 return get_irg_bad(current_ir_graph);
3676 ir_node *blk = get_nodes_block(load);
3677 return new_r_Jmp(current_ir_graph, blk);
3683 } /* transform_node_Proj_Load */
3686 * Transform a Proj(Store) with a non-null address.
3688 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3689 if (get_opt_ldst_only_null_ptr_exceptions()) {
3690 if (get_irn_mode(proj) == mode_X) {
3691 ir_node *store = get_Proj_pred(proj);
3693 /* get the load/store address */
3694 const ir_node *addr = get_Store_ptr(store);
3695 const ir_node *confirm;
3697 if (value_not_null(addr, &confirm)) {
3698 if (confirm == NULL) {
3699 /* this node may float if it did not depend on a Confirm */
3700 set_irn_pinned(store, op_pin_state_floats);
3702 if (get_Proj_proj(proj) == pn_Store_X_except) {
3703 DBG_OPT_EXC_REM(proj);
3704 return get_irg_bad(current_ir_graph);
3706 ir_node *blk = get_nodes_block(store);
3707 return new_r_Jmp(current_ir_graph, blk);
3713 } /* transform_node_Proj_Store */
3716 * Transform a Proj(Div) with a non-zero value.
3717 * Removes the exceptions and routes the memory to the NoMem node.
3719 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3720 ir_node *div = get_Proj_pred(proj);
3721 ir_node *b = get_Div_right(div);
3722 ir_node *res, *new_mem;
3723 const ir_node *confirm;
3726 if (value_not_zero(b, &confirm)) {
3727 /* div(x, y) && y != 0 */
3728 if (confirm == NULL) {
3729 /* we are sure we have a Const != 0 */
3730 new_mem = get_Div_mem(div);
3731 if (is_Pin(new_mem))
3732 new_mem = get_Pin_op(new_mem);
3733 set_Div_mem(div, new_mem);
3734 set_irn_pinned(div, op_pin_state_floats);
3737 proj_nr = get_Proj_proj(proj);
3739 case pn_Div_X_regular:
3740 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3742 case pn_Div_X_except:
3743 /* we found an exception handler, remove it */
3744 DBG_OPT_EXC_REM(proj);
3748 res = get_Div_mem(div);
3749 new_mem = get_irg_no_mem(current_ir_graph);
3752 /* This node can only float up to the Confirm block */
3753 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3755 set_irn_pinned(div, op_pin_state_floats);
3756 /* this is a Div without exception, we can remove the memory edge */
3757 set_Div_mem(div, new_mem);
3762 } /* transform_node_Proj_Div */
3765 * Transform a Proj(Mod) with a non-zero value.
3766 * Removes the exceptions and routes the memory to the NoMem node.
3768 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3769 ir_node *mod = get_Proj_pred(proj);
3770 ir_node *b = get_Mod_right(mod);
3771 ir_node *res, *new_mem;
3772 const ir_node *confirm;
3775 if (value_not_zero(b, &confirm)) {
3776 /* mod(x, y) && y != 0 */
3777 proj_nr = get_Proj_proj(proj);
3779 if (confirm == NULL) {
3780 /* we are sure we have a Const != 0 */
3781 new_mem = get_Mod_mem(mod);
3782 if (is_Pin(new_mem))
3783 new_mem = get_Pin_op(new_mem);
3784 set_Mod_mem(mod, new_mem);
3785 set_irn_pinned(mod, op_pin_state_floats);
3790 case pn_Mod_X_regular:
3791 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3793 case pn_Mod_X_except:
3794 /* we found an exception handler, remove it */
3795 DBG_OPT_EXC_REM(proj);
3799 res = get_Mod_mem(mod);
3800 new_mem = get_irg_no_mem(current_ir_graph);
3803 /* This node can only float up to the Confirm block */
3804 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3806 /* this is a Mod without exception, we can remove the memory edge */
3807 set_Mod_mem(mod, new_mem);
3810 if (get_Mod_left(mod) == b) {
3811 /* a % a = 0 if a != 0 */
3812 ir_mode *mode = get_irn_mode(proj);
3813 ir_node *res = new_Const(mode, get_mode_null(mode));
3815 DBG_OPT_CSTEVAL(mod, res);
3821 } /* transform_node_Proj_Mod */
3824 * Transform a Proj(DivMod) with a non-zero value.
3825 * Removes the exceptions and routes the memory to the NoMem node.
3827 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3828 ir_node *divmod = get_Proj_pred(proj);
3829 ir_node *b = get_DivMod_right(divmod);
3830 ir_node *res, *new_mem;
3831 const ir_node *confirm;
3834 if (value_not_zero(b, &confirm)) {
3835 /* DivMod(x, y) && y != 0 */
3836 proj_nr = get_Proj_proj(proj);
3838 if (confirm == NULL) {
3839 /* we are sure we have a Const != 0 */
3840 new_mem = get_DivMod_mem(divmod);
3841 if (is_Pin(new_mem))
3842 new_mem = get_Pin_op(new_mem);
3843 set_DivMod_mem(divmod, new_mem);
3844 set_irn_pinned(divmod, op_pin_state_floats);
3849 case pn_DivMod_X_regular:
3850 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3852 case pn_DivMod_X_except:
3853 /* we found an exception handler, remove it */
3854 DBG_OPT_EXC_REM(proj);
3858 res = get_DivMod_mem(divmod);
3859 new_mem = get_irg_no_mem(current_ir_graph);
3862 /* This node can only float up to the Confirm block */
3863 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3865 /* this is a DivMod without exception, we can remove the memory edge */
3866 set_DivMod_mem(divmod, new_mem);
3869 case pn_DivMod_res_mod:
3870 if (get_DivMod_left(divmod) == b) {
3871 /* a % a = 0 if a != 0 */
3872 ir_mode *mode = get_irn_mode(proj);
3873 ir_node *res = new_Const(mode, get_mode_null(mode));
3875 DBG_OPT_CSTEVAL(divmod, res);
3881 } /* transform_node_Proj_DivMod */
3884 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3886 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3887 if (get_opt_unreachable_code()) {
3888 ir_node *n = get_Proj_pred(proj);
3889 ir_node *b = get_Cond_selector(n);
3891 if (mode_is_int(get_irn_mode(b))) {
3892 tarval *tb = value_of(b);
3894 if (tb != tarval_bad) {
3895 /* we have a constant switch */
3896 long num = get_Proj_proj(proj);
3898 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3899 if (get_tarval_long(tb) == num) {
3900 /* Do NOT create a jump here, or we will have 2 control flow ops
3901 * in a block. This case is optimized away in optimize_cf(). */
3904 /* this case will NEVER be taken, kill it */
3905 return get_irg_bad(current_ir_graph);
3912 } /* transform_node_Proj_Cond */
3915 * Create a 0 constant of given mode.
3917 static ir_node *create_zero_const(ir_mode *mode) {
3918 tarval *tv = get_mode_null(mode);
3919 ir_node *cnst = new_Const(mode, tv);
3924 /* the order of the values is important! */
3925 typedef enum const_class {
3931 static const_class classify_const(const ir_node* n)
3933 if (is_Const(n)) return const_const;
3934 if (is_irn_constlike(n)) return const_like;
3939 * Determines whether r is more constlike or has a larger index (in that order)
3942 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3944 const const_class l_order = classify_const(l);
3945 const const_class r_order = classify_const(r);
3947 l_order > r_order ||
3948 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3952 * Normalizes and optimizes Cmp nodes.
3954 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3955 ir_node *n = get_Proj_pred(proj);
3956 ir_node *left = get_Cmp_left(n);
3957 ir_node *right = get_Cmp_right(n);
3960 ir_mode *mode = NULL;
3961 long proj_nr = get_Proj_proj(proj);
3963 /* we can evaluate some cases directly */
3966 return new_Const(mode_b, get_tarval_b_false());
3968 return new_Const(mode_b, get_tarval_b_true());
3970 if (!mode_is_float(get_irn_mode(left)))
3971 return new_Const(mode_b, get_tarval_b_true());
3977 /* remove Casts of both sides */
3978 left = skip_Cast(left);
3979 right = skip_Cast(right);
3981 /* Remove unnecessary conversions */
3982 /* TODO handle constants */
3983 if (is_Conv(left) && is_Conv(right)) {
3984 ir_mode *mode = get_irn_mode(left);
3985 ir_node *op_left = get_Conv_op(left);
3986 ir_node *op_right = get_Conv_op(right);
3987 ir_mode *mode_left = get_irn_mode(op_left);
3988 ir_mode *mode_right = get_irn_mode(op_right);
3990 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3991 && mode_left != mode_b && mode_right != mode_b) {
3992 ir_graph *irg = current_ir_graph;
3993 ir_node *block = get_nodes_block(n);
3995 if (mode_left == mode_right) {
3999 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4000 } else if (smaller_mode(mode_left, mode_right)) {
4001 left = new_r_Conv(irg, block, op_left, mode_right);
4004 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4005 } else if (smaller_mode(mode_right, mode_left)) {
4007 right = new_r_Conv(irg, block, op_right, mode_left);
4009 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4014 /* remove operation on both sides if possible */
4015 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4017 * The following operations are NOT safe for floating point operations, for instance
4018 * 1.0 + inf == 2.0 + inf, =/=> x == y
4020 if (mode_is_int(get_irn_mode(left))) {
4021 unsigned lop = get_irn_opcode(left);
4023 if (lop == get_irn_opcode(right)) {
4024 ir_node *ll, *lr, *rl, *rr;
4026 /* same operation on both sides, try to remove */
4030 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4031 left = get_unop_op(left);
4032 right = get_unop_op(right);
4034 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4037 ll = get_Add_left(left);
4038 lr = get_Add_right(left);
4039 rl = get_Add_left(right);
4040 rr = get_Add_right(right);
4043 /* X + a CMP X + b ==> a CMP b */
4047 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4048 } else if (ll == rr) {
4049 /* X + a CMP b + X ==> a CMP b */
4053 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4054 } else if (lr == rl) {
4055 /* a + X CMP X + b ==> a CMP b */
4059 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4060 } else if (lr == rr) {
4061 /* a + X CMP b + X ==> a CMP b */
4065 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4069 ll = get_Sub_left(left);
4070 lr = get_Sub_right(left);
4071 rl = get_Sub_left(right);
4072 rr = get_Sub_right(right);
4075 /* X - a CMP X - b ==> a CMP b */
4079 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4080 } else if (lr == rr) {
4081 /* a - X CMP b - X ==> a CMP b */
4085 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4089 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4090 /* a ROTL X CMP b ROTL X ==> a CMP b */
4091 left = get_Rotl_left(left);
4092 right = get_Rotl_left(right);
4094 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4102 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4103 if (is_Add(left) || is_Sub(left)) {
4104 ir_node *ll = get_binop_left(left);
4105 ir_node *lr = get_binop_right(left);
4107 if (lr == right && is_Add(left)) {
4114 right = create_zero_const(get_irn_mode(left));
4116 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4119 if (is_Add(right) || is_Sub(right)) {
4120 ir_node *rl = get_binop_left(right);
4121 ir_node *rr = get_binop_right(right);
4123 if (rr == left && is_Add(right)) {
4130 right = create_zero_const(get_irn_mode(left));
4132 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4135 } /* mode_is_int(...) */
4136 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4138 /* replace mode_b compares with ands/ors */
4139 if (get_irn_mode(left) == mode_b) {
4140 ir_graph *irg = current_ir_graph;
4141 ir_node *block = get_nodes_block(n);
4145 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4146 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4147 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4148 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4149 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4150 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4151 default: bres = NULL;
4154 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4160 * First step: normalize the compare op
4161 * by placing the constant on the right side
4162 * or moving the lower address node to the left.
4164 if (!operands_are_normalized(left, right)) {
4170 proj_nr = get_inversed_pnc(proj_nr);
4175 * Second step: Try to reduce the magnitude
4176 * of a constant. This may help to generate better code
4177 * later and may help to normalize more compares.
4178 * Of course this is only possible for integer values.
4180 if (is_Const(right)) {
4181 mode = get_irn_mode(right);
4182 tv = get_Const_tarval(right);
4184 /* TODO extend to arbitrary constants */
4185 if (is_Conv(left) && tarval_is_null(tv)) {
4186 ir_node *op = get_Conv_op(left);
4187 ir_mode *op_mode = get_irn_mode(op);
4190 * UpConv(x) REL 0 ==> x REL 0
4192 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4193 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4194 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4195 tv = get_mode_null(op_mode);
4199 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4203 if (tv != tarval_bad) {
4204 /* the following optimization is possible on modes without Overflow
4205 * on Unary Minus or on == and !=:
4206 * -a CMP c ==> a swap(CMP) -c
4208 * Beware: for two-complement Overflow may occur, so only == and != can
4209 * be optimized, see this:
4210 * -MININT < 0 =/=> MININT > 0 !!!
4212 if (is_Minus(left) &&
4213 (!mode_overflow_on_unary_Minus(mode) ||
4214 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4215 tv = tarval_neg(tv);
4217 if (tv != tarval_bad) {
4218 left = get_Minus_op(left);
4219 proj_nr = get_inversed_pnc(proj_nr);
4221 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4223 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4224 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4225 tv = tarval_not(tv);
4227 if (tv != tarval_bad) {
4228 left = get_Not_op(left);
4230 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4234 /* for integer modes, we have more */
4235 if (mode_is_int(mode)) {
4236 /* Ne includes Unordered which is not possible on integers.
4237 * However, frontends often use this wrong, so fix it here */
4238 if (proj_nr & pn_Cmp_Uo) {
4239 proj_nr &= ~pn_Cmp_Uo;
4240 set_Proj_proj(proj, proj_nr);
4243 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4244 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4245 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4246 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4248 if (tv != tarval_bad) {
4249 proj_nr ^= pn_Cmp_Eq;
4251 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4254 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4255 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4256 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4257 tv = tarval_add(tv, get_mode_one(mode));
4259 if (tv != tarval_bad) {
4260 proj_nr ^= pn_Cmp_Eq;
4262 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4266 /* the following reassociations work only for == and != */
4267 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4269 #if 0 /* Might be not that good in general */
4270 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4271 if (tarval_is_null(tv) && is_Sub(left)) {
4272 right = get_Sub_right(left);
4273 left = get_Sub_left(left);
4275 tv = value_of(right);
4277 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4281 if (tv != tarval_bad) {
4282 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4284 ir_node *c1 = get_Sub_right(left);
4285 tarval *tv2 = value_of(c1);
4287 if (tv2 != tarval_bad) {
4288 tv2 = tarval_add(tv, value_of(c1));
4290 if (tv2 != tarval_bad) {
4291 left = get_Sub_left(left);
4294 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4298 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4299 else if (is_Add(left)) {
4300 ir_node *a_l = get_Add_left(left);
4301 ir_node *a_r = get_Add_right(left);
4305 if (is_Const(a_l)) {
4307 tv2 = value_of(a_l);
4310 tv2 = value_of(a_r);
4313 if (tv2 != tarval_bad) {
4314 tv2 = tarval_sub(tv, tv2, NULL);
4316 if (tv2 != tarval_bad) {
4320 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4324 /* -a == c ==> a == -c, -a != c ==> a != -c */
4325 else if (is_Minus(left)) {
4326 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4328 if (tv2 != tarval_bad) {
4329 left = get_Minus_op(left);
4332 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4337 /* the following reassociations work only for <= */
4338 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4339 if (tv != tarval_bad) {
4340 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4341 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4347 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4348 switch (get_irn_opcode(left)) {
4352 c1 = get_And_right(left);
4355 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4356 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4358 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4360 /* TODO: move to constant evaluation */
4361 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4362 c1 = new_Const(mode_b, tv);
4363 DBG_OPT_CSTEVAL(proj, c1);
4367 if (tarval_is_single_bit(tv)) {
4369 * optimization for AND:
4371 * And(x, C) == C ==> And(x, C) != 0
4372 * And(x, C) != C ==> And(X, C) == 0
4374 * if C is a single Bit constant.
4377 /* check for Constant's match. We have check hare the tarvals,
4378 because our const might be changed */
4379 if (get_Const_tarval(c1) == tv) {
4380 /* fine: do the transformation */
4381 tv = get_mode_null(get_tarval_mode(tv));
4382 proj_nr ^= pn_Cmp_Leg;
4384 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4390 c1 = get_Or_right(left);
4391 if (is_Const(c1) && tarval_is_null(tv)) {
4393 * Or(x, C) == 0 && C != 0 ==> FALSE
4394 * Or(x, C) != 0 && C != 0 ==> TRUE
4396 if (! tarval_is_null(get_Const_tarval(c1))) {
4397 /* TODO: move to constant evaluation */
4398 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4399 c1 = new_Const(mode_b, tv);
4400 DBG_OPT_CSTEVAL(proj, c1);
4407 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4409 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4412 c1 = get_Shl_right(left);
4414 tarval *tv1 = get_Const_tarval(c1);
4415 ir_mode *mode = get_irn_mode(left);
4416 tarval *minus1 = get_mode_all_one(mode);
4417 tarval *amask = tarval_shr(minus1, tv1);
4418 tarval *cmask = tarval_shl(minus1, tv1);
4421 if (tarval_and(tv, cmask) != tv) {
4422 /* condition not met */
4423 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4424 c1 = new_Const(mode_b, tv);
4425 DBG_OPT_CSTEVAL(proj, c1);
4428 sl = get_Shl_left(left);
4429 blk = get_nodes_block(n);
4430 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4431 tv = tarval_shr(tv, tv1);
4433 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4438 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4440 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4443 c1 = get_Shr_right(left);
4445 tarval *tv1 = get_Const_tarval(c1);
4446 ir_mode *mode = get_irn_mode(left);
4447 tarval *minus1 = get_mode_all_one(mode);
4448 tarval *amask = tarval_shl(minus1, tv1);
4449 tarval *cmask = tarval_shr(minus1, tv1);
4452 if (tarval_and(tv, cmask) != tv) {
4453 /* condition not met */
4454 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4455 c1 = new_Const(mode_b, tv);
4456 DBG_OPT_CSTEVAL(proj, c1);
4459 sl = get_Shr_left(left);
4460 blk = get_nodes_block(n);
4461 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4462 tv = tarval_shl(tv, tv1);
4464 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4469 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4471 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4474 c1 = get_Shrs_right(left);
4476 tarval *tv1 = get_Const_tarval(c1);
4477 ir_mode *mode = get_irn_mode(left);
4478 tarval *minus1 = get_mode_all_one(mode);
4479 tarval *amask = tarval_shl(minus1, tv1);
4480 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4483 cond = tarval_sub(cond, tv1, NULL);
4484 cond = tarval_shrs(tv, cond);
4486 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4487 /* condition not met */
4488 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4489 c1 = new_Const(mode_b, tv);
4490 DBG_OPT_CSTEVAL(proj, c1);
4493 sl = get_Shrs_left(left);
4494 blk = get_nodes_block(n);
4495 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4496 tv = tarval_shl(tv, tv1);
4498 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4503 } /* tarval != bad */
4506 if (changed & 2) /* need a new Const */
4507 right = new_Const(mode, tv);
4509 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4510 ir_node *op = get_Proj_pred(left);
4512 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4513 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4514 ir_node *c = get_binop_right(op);
4517 tarval *tv = get_Const_tarval(c);
4519 if (tarval_is_single_bit(tv)) {
4520 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4521 ir_node *v = get_binop_left(op);
4522 ir_node *blk = get_irn_n(op, -1);
4523 ir_mode *mode = get_irn_mode(v);
4525 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4526 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4528 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4535 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4537 /* create a new compare */
4538 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4539 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4543 } /* transform_node_Proj_Cmp */
4546 * Optimize CopyB(mem, x, x) into a Nop.
4548 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4549 ir_node *copyb = get_Proj_pred(proj);
4550 ir_node *a = get_CopyB_dst(copyb);
4551 ir_node *b = get_CopyB_src(copyb);
4554 switch (get_Proj_proj(proj)) {
4555 case pn_CopyB_X_regular:
4556 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4557 DBG_OPT_EXC_REM(proj);
4558 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4560 case pn_CopyB_M_except:
4561 case pn_CopyB_X_except:
4562 DBG_OPT_EXC_REM(proj);
4563 proj = get_irg_bad(current_ir_graph);
4570 } /* transform_node_Proj_CopyB */
4573 * Optimize Bounds(idx, idx, upper) into idx.
4575 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4576 ir_node *oldn = proj;
4577 ir_node *bound = get_Proj_pred(proj);
4578 ir_node *idx = get_Bound_index(bound);
4579 ir_node *pred = skip_Proj(idx);
4582 if (idx == get_Bound_lower(bound))
4584 else if (is_Bound(pred)) {
4586 * idx was Bounds checked in the same MacroBlock previously,
4587 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4589 ir_node *lower = get_Bound_lower(bound);
4590 ir_node *upper = get_Bound_upper(bound);
4591 if (get_Bound_lower(pred) == lower &&
4592 get_Bound_upper(pred) == upper &&
4593 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4595 * One could expect that we simply return the previous
4596 * Bound here. However, this would be wrong, as we could
4597 * add an exception Proj to a new location then.
4598 * So, we must turn in into a tuple.
4604 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4605 switch (get_Proj_proj(proj)) {
4607 DBG_OPT_EXC_REM(proj);
4608 proj = get_Bound_mem(bound);
4610 case pn_Bound_X_except:
4611 DBG_OPT_EXC_REM(proj);
4612 proj = get_irg_bad(current_ir_graph);
4616 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4618 case pn_Bound_X_regular:
4619 DBG_OPT_EXC_REM(proj);
4620 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4627 } /* transform_node_Proj_Bound */
4630 * Does all optimizations on nodes that must be done on it's Proj's
4631 * because of creating new nodes.
4633 static ir_node *transform_node_Proj(ir_node *proj) {
4634 ir_node *n = get_Proj_pred(proj);
4636 if (n->op->ops.transform_node_Proj)
4637 return n->op->ops.transform_node_Proj(proj);
4639 } /* transform_node_Proj */
4642 * Move Confirms down through Phi nodes.
4644 static ir_node *transform_node_Phi(ir_node *phi) {
4646 ir_mode *mode = get_irn_mode(phi);
4648 if (mode_is_reference(mode)) {
4649 n = get_irn_arity(phi);
4651 /* Beware of Phi0 */
4653 ir_node *pred = get_irn_n(phi, 0);
4654 ir_node *bound, *new_Phi, *block, **in;
4657 if (! is_Confirm(pred))
4660 bound = get_Confirm_bound(pred);
4661 pnc = get_Confirm_cmp(pred);
4663 NEW_ARR_A(ir_node *, in, n);
4664 in[0] = get_Confirm_value(pred);
4666 for (i = 1; i < n; ++i) {
4667 pred = get_irn_n(phi, i);
4669 if (! is_Confirm(pred) ||
4670 get_Confirm_bound(pred) != bound ||
4671 get_Confirm_cmp(pred) != pnc)
4673 in[i] = get_Confirm_value(pred);
4675 /* move the Confirm nodes "behind" the Phi */
4676 block = get_irn_n(phi, -1);
4677 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4678 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4682 } /* transform_node_Phi */
4685 * Returns the operands of a commutative bin-op, if one operand is
4686 * a const, it is returned as the second one.
4688 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4689 ir_node *op_a = get_binop_left(binop);
4690 ir_node *op_b = get_binop_right(binop);
4692 assert(is_op_commutative(get_irn_op(binop)));
4694 if (is_Const(op_a)) {
4701 } /* get_comm_Binop_Ops */
4704 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4705 * Such pattern may arise in bitfield stores.
4707 * value c4 value c4 & c2
4708 * AND c3 AND c1 | c3
4715 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4718 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4721 ir_node *and_l, *c3;
4722 ir_node *value, *c4;
4723 ir_node *new_and, *new_const, *block;
4724 ir_mode *mode = get_irn_mode(or);
4726 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4729 get_comm_Binop_Ops(or, &and, &c1);
4730 if (!is_Const(c1) || !is_And(and))
4733 get_comm_Binop_Ops(and, &or_l, &c2);
4737 tv1 = get_Const_tarval(c1);
4738 tv2 = get_Const_tarval(c2);
4740 tv = tarval_or(tv1, tv2);
4741 if (tarval_is_all_one(tv)) {
4742 /* the AND does NOT clear a bit with isn't set by the OR */
4743 set_Or_left(or, or_l);
4744 set_Or_right(or, c1);
4746 /* check for more */
4753 get_comm_Binop_Ops(or_l, &and_l, &c3);
4754 if (!is_Const(c3) || !is_And(and_l))
4757 get_comm_Binop_Ops(and_l, &value, &c4);
4761 /* ok, found the pattern, check for conditions */
4762 assert(mode == get_irn_mode(and));
4763 assert(mode == get_irn_mode(or_l));
4764 assert(mode == get_irn_mode(and_l));
4766 tv3 = get_Const_tarval(c3);
4767 tv4 = get_Const_tarval(c4);
4769 tv = tarval_or(tv4, tv2);
4770 if (!tarval_is_all_one(tv)) {
4771 /* have at least one 0 at the same bit position */
4775 n_tv4 = tarval_not(tv4);
4776 if (tv3 != tarval_and(tv3, n_tv4)) {
4777 /* bit in the or_mask is outside the and_mask */
4781 n_tv2 = tarval_not(tv2);
4782 if (tv1 != tarval_and(tv1, n_tv2)) {
4783 /* bit in the or_mask is outside the and_mask */
4787 /* ok, all conditions met */
4788 block = get_irn_n(or, -1);
4790 new_and = new_r_And(current_ir_graph, block,
4791 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4793 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4795 set_Or_left(or, new_and);
4796 set_Or_right(or, new_const);
4798 /* check for more */
4800 } /* transform_node_Or_bf_store */
4803 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4805 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4806 ir_mode *mode = get_irn_mode(or);
4807 ir_node *shl, *shr, *block;
4808 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4811 if (! mode_is_int(mode))
4814 shl = get_binop_left(or);
4815 shr = get_binop_right(or);
4824 } else if (!is_Shl(shl)) {
4826 } else if (!is_Shr(shr)) {
4829 x = get_Shl_left(shl);
4830 if (x != get_Shr_left(shr))
4833 c1 = get_Shl_right(shl);
4834 c2 = get_Shr_right(shr);
4835 if (is_Const(c1) && is_Const(c2)) {
4836 tv1 = get_Const_tarval(c1);
4837 if (! tarval_is_long(tv1))
4840 tv2 = get_Const_tarval(c2);
4841 if (! tarval_is_long(tv2))
4844 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4845 != (int) get_mode_size_bits(mode))
4848 /* yet, condition met */
4849 block = get_nodes_block(or);
4851 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4853 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4860 rotval = sub; /* a Rot right is not supported, so use a rot left */
4861 } else if (is_Sub(c2)) {
4867 if (get_Sub_right(sub) != v)
4870 c1 = get_Sub_left(sub);
4874 tv1 = get_Const_tarval(c1);
4875 if (! tarval_is_long(tv1))
4878 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4881 /* yet, condition met */
4882 block = get_nodes_block(or);
4884 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4886 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4888 } /* transform_node_Or_Rotl */
4893 static ir_node *transform_node_Or(ir_node *n) {
4894 ir_node *c, *oldn = n;
4895 ir_node *a = get_Or_left(n);
4896 ir_node *b = get_Or_right(n);
4899 if (is_Not(a) && is_Not(b)) {
4900 /* ~a | ~b = ~(a&b) */
4901 ir_node *block = get_nodes_block(n);
4903 mode = get_irn_mode(n);
4906 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4907 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4908 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4912 /* we can evaluate 2 Projs of the same Cmp */
4913 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4914 ir_node *pred_a = get_Proj_pred(a);
4915 ir_node *pred_b = get_Proj_pred(b);
4916 if (pred_a == pred_b) {
4917 dbg_info *dbgi = get_irn_dbg_info(n);
4918 ir_node *block = get_nodes_block(pred_a);
4919 pn_Cmp pn_a = get_Proj_proj(a);
4920 pn_Cmp pn_b = get_Proj_proj(b);
4921 /* yes, we can simply calculate with pncs */
4922 pn_Cmp new_pnc = pn_a | pn_b;
4924 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4929 mode = get_irn_mode(n);
4930 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4932 n = transform_node_Or_bf_store(n);
4933 n = transform_node_Or_Rotl(n);
4937 n = transform_bitwise_distributive(n, transform_node_Or);
4940 } /* transform_node_Or */
4944 static ir_node *transform_node(ir_node *n);
4947 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4949 * Should be moved to reassociation?
4951 static ir_node *transform_node_shift(ir_node *n) {
4952 ir_node *left, *right;
4954 tarval *tv1, *tv2, *res;
4955 ir_node *in[2], *irn, *block;
4957 left = get_binop_left(n);
4959 /* different operations */
4960 if (get_irn_op(left) != get_irn_op(n))
4963 right = get_binop_right(n);
4964 tv1 = value_of(right);
4965 if (tv1 == tarval_bad)
4968 tv2 = value_of(get_binop_right(left));
4969 if (tv2 == tarval_bad)
4972 res = tarval_add(tv1, tv2);
4973 mode = get_irn_mode(n);
4975 /* beware: a simple replacement works only, if res < modulo shift */
4977 int modulo_shf = get_mode_modulo_shift(mode);
4978 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4979 if (modulo_shf > 0) {
4980 tarval *modulo = new_tarval_from_long(modulo_shf,
4981 get_tarval_mode(res));
4983 /* shifting too much */
4984 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4986 ir_graph *irg = get_irn_irg(n);
4987 ir_node *block = get_nodes_block(n);
4988 dbg_info *dbgi = get_irn_dbg_info(n);
4989 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
4990 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
4994 return new_Const(mode, get_mode_null(mode));
4998 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5001 /* ok, we can replace it */
5002 block = get_nodes_block(n);
5004 in[0] = get_binop_left(left);
5005 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5007 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5009 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5011 return transform_node(irn);
5012 } /* transform_node_shift */
5015 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5017 * - and, or, xor instead of &
5018 * - Shl, Shr, Shrs, rotl instead of >>
5019 * (with a special case for Or/Xor + Shrs)
5021 static ir_node *transform_node_bitop_shift(ir_node *n) {
5023 ir_node *right = get_binop_right(n);
5024 ir_mode *mode = get_irn_mode(n);
5025 ir_node *bitop_left;
5026 ir_node *bitop_right;
5038 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5040 if (!is_Const(right))
5043 left = get_binop_left(n);
5044 op_left = get_irn_op(left);
5045 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5048 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5049 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5050 /* TODO: test if sign bit is affectes */
5054 bitop_right = get_binop_right(left);
5055 if (!is_Const(bitop_right))
5058 bitop_left = get_binop_left(left);
5060 irg = get_irn_irg(n);
5061 block = get_nodes_block(n);
5062 dbgi = get_irn_dbg_info(n);
5063 tv1 = get_Const_tarval(bitop_right);
5064 tv2 = get_Const_tarval(right);
5066 assert(get_tarval_mode(tv1) == mode);
5069 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5070 tv_shift = tarval_shl(tv1, tv2);
5071 } else if(is_Shr(n)) {
5072 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5073 tv_shift = tarval_shr(tv1, tv2);
5074 } else if(is_Shrs(n)) {
5075 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5076 tv_shift = tarval_shrs(tv1, tv2);
5079 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5080 tv_shift = tarval_rotl(tv1, tv2);
5083 assert(get_tarval_mode(tv_shift) == mode);
5084 new_const = new_Const(mode, tv_shift);
5086 if (op_left == op_And) {
5087 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5088 } else if(op_left == op_Or) {
5089 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5091 assert(op_left == op_Eor);
5092 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5100 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5102 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5103 * (also with x >>s c1 when c1>=c2)
5105 static ir_node *transform_node_shl_shr(ir_node *n) {
5107 ir_node *right = get_binop_right(n);
5123 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5125 if (!is_Const(right))
5128 left = get_binop_left(n);
5129 mode = get_irn_mode(n);
5130 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5131 ir_node *shr_right = get_binop_right(left);
5133 if (!is_Const(shr_right))
5136 x = get_binop_left(left);
5137 tv_shr = get_Const_tarval(shr_right);
5138 tv_shl = get_Const_tarval(right);
5140 if (is_Shrs(left)) {
5141 /* shrs variant only allowed if c1 >= c2 */
5142 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5145 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5148 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5150 tv_mask = tarval_shl(tv_mask, tv_shl);
5151 } else if(is_Shr(n) && is_Shl(left)) {
5152 ir_node *shl_right = get_Shl_right(left);
5154 if (!is_Const(shl_right))
5157 x = get_Shl_left(left);
5158 tv_shr = get_Const_tarval(right);
5159 tv_shl = get_Const_tarval(shl_right);
5161 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5162 tv_mask = tarval_shr(tv_mask, tv_shr);
5167 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5168 assert(tv_mask != tarval_bad);
5169 assert(get_tarval_mode(tv_mask) == mode);
5171 irg = get_irn_irg(n);
5172 block = get_nodes_block(n);
5173 dbgi = get_irn_dbg_info(n);
5175 pnc = tarval_cmp(tv_shl, tv_shr);
5176 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5177 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5178 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5180 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5182 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5185 assert(pnc == pn_Cmp_Gt);
5186 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5187 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5188 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5191 new_const = new_Const(mode, tv_mask);
5192 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5200 static ir_node *transform_node_Shr(ir_node *n) {
5201 ir_node *c, *oldn = n;
5202 ir_node *left = get_Shr_left(n);
5203 ir_node *right = get_Shr_right(n);
5204 ir_mode *mode = get_irn_mode(n);
5206 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5207 n = transform_node_shift(n);
5210 n = transform_node_shl_shr(n);
5212 n = transform_node_bitop_shift(n);
5215 } /* transform_node_Shr */
5220 static ir_node *transform_node_Shrs(ir_node *n) {
5221 ir_node *c, *oldn = n;
5222 ir_node *a = get_Shrs_left(n);
5223 ir_node *b = get_Shrs_right(n);
5224 ir_mode *mode = get_irn_mode(n);
5226 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5227 n = transform_node_shift(n);
5230 n = transform_node_bitop_shift(n);
5233 } /* transform_node_Shrs */
5238 static ir_node *transform_node_Shl(ir_node *n) {
5239 ir_node *c, *oldn = n;
5240 ir_node *a = get_Shl_left(n);
5241 ir_node *b = get_Shl_right(n);
5242 ir_mode *mode = get_irn_mode(n);
5244 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5245 n = transform_node_shift(n);
5248 n = transform_node_shl_shr(n);
5250 n = transform_node_bitop_shift(n);
5253 } /* transform_node_Shl */
5258 static ir_node *transform_node_Rotl(ir_node *n) {
5259 ir_node *c, *oldn = n;
5260 ir_node *a = get_Rotl_left(n);
5261 ir_node *b = get_Rotl_right(n);
5262 ir_mode *mode = get_irn_mode(n);
5264 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5265 n = transform_node_shift(n);
5268 n = transform_node_bitop_shift(n);
5271 } /* transform_node_Rotl */
5276 static ir_node *transform_node_Conv(ir_node *n) {
5277 ir_node *c, *oldn = n;
5278 ir_node *a = get_Conv_op(n);
5280 if (is_const_Phi(a)) {
5281 c = apply_conv_on_phi(a, get_irn_mode(n));
5283 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5288 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5289 ir_mode *mode = get_irn_mode(n);
5290 return new_r_Unknown(current_ir_graph, mode);
5294 } /* transform_node_Conv */
5297 * Remove dead blocks and nodes in dead blocks
5298 * in keep alive list. We do not generate a new End node.
5300 static ir_node *transform_node_End(ir_node *n) {
5301 int i, j, n_keepalives = get_End_n_keepalives(n);
5304 NEW_ARR_A(ir_node *, in, n_keepalives);
5306 for (i = j = 0; i < n_keepalives; ++i) {
5307 ir_node *ka = get_End_keepalive(n, i);
5309 if (! is_Block_dead(ka)) {
5313 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5316 /* FIXME: beabi need to keep a Proj(M) */
5317 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
5320 if (j != n_keepalives)
5321 set_End_keepalives(n, j, in);
5323 } /* transform_node_End */
5325 /** returns 1 if a == -b */
5326 static int is_negated_value(ir_node *a, ir_node *b) {
5327 if (is_Minus(a) && get_Minus_op(a) == b)
5329 if (is_Minus(b) && get_Minus_op(b) == a)
5331 if (is_Sub(a) && is_Sub(b)) {
5332 ir_node *a_left = get_Sub_left(a);
5333 ir_node *a_right = get_Sub_right(a);
5334 ir_node *b_left = get_Sub_left(b);
5335 ir_node *b_right = get_Sub_right(b);
5337 if (a_left == b_right && a_right == b_left)
5345 * Optimize a Mux into some simpler cases.
5347 static ir_node *transform_node_Mux(ir_node *n) {
5348 ir_node *oldn = n, *sel = get_Mux_sel(n);
5349 ir_mode *mode = get_irn_mode(n);
5350 ir_node *t = get_Mux_true(n);
5351 ir_node *f = get_Mux_false(n);
5352 ir_graph *irg = current_ir_graph;
5354 /* first normalization step: move a possible zero to the false case */
5356 ir_node *cmp = get_Proj_pred(sel);
5359 if (is_Const(t) && is_Const_null(t)) {
5362 /* Mux(x, 0, y) => Mux(x, y, 0) */
5363 pn_Cmp pnc = get_Proj_proj(sel);
5364 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5365 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5366 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5374 /* note: after normalization, false can only happen on default */
5375 if (mode == mode_b) {
5376 dbg_info *dbg = get_irn_dbg_info(n);
5377 ir_node *block = get_nodes_block(n);
5378 ir_graph *irg = current_ir_graph;
5381 tarval *tv_t = get_Const_tarval(t);
5382 if (tv_t == tarval_b_true) {
5384 /* Muxb(sel, true, false) = sel */
5385 assert(get_Const_tarval(f) == tarval_b_false);
5386 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5389 /* Muxb(sel, true, x) = Or(sel, x) */
5390 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5391 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5395 } else if (is_Const(f)) {
5396 tarval *tv_f = get_Const_tarval(f);
5397 if (tv_f == tarval_b_true) {
5398 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5399 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5400 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5401 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5404 /* Muxb(sel, x, false) = And(sel, x) */
5405 assert(tv_f == tarval_b_false);
5406 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5407 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5413 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5414 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5415 tarval *a = get_Const_tarval(t);
5416 tarval *b = get_Const_tarval(f);
5417 tarval *null = get_tarval_null(mode);
5420 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5421 diff = tarval_sub(a, b, NULL);
5424 diff = tarval_sub(b, a, NULL);
5428 if (diff == get_tarval_one(mode) && min != null) {
5429 dbg_info *dbg = get_irn_dbg_info(n);
5430 ir_node *block = get_nodes_block(n);
5431 ir_graph *irg = current_ir_graph;
5432 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5433 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5434 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5435 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5441 ir_node *cmp = get_Proj_pred(sel);
5442 long pn = get_Proj_proj(sel);
5445 * Note: normalization puts the constant on the right side,
5446 * so we check only one case.
5448 * Note further that these optimization work even for floating point
5449 * with NaN's because -NaN == NaN.
5450 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5454 ir_node *cmp_r = get_Cmp_right(cmp);
5455 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5456 ir_node *block = get_nodes_block(n);
5457 ir_node *cmp_l = get_Cmp_left(cmp);
5459 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5462 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5463 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5465 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5466 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5468 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5470 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5471 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5473 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5474 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5476 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5478 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5483 if (mode_is_int(mode)) {
5485 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5486 /* Mux((a & b) != 0, c, 0) */
5487 ir_node *and_r = get_And_right(cmp_l);
5490 if (and_r == t && f == cmp_r) {
5491 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5492 if (pn == pn_Cmp_Lg) {
5493 /* Mux((a & 2^C) != 0, 2^C, 0) */
5496 /* Mux((a & 2^C) == 0, 2^C, 0) */
5497 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5498 block, cmp_l, t, mode);
5503 if (is_Shl(and_r)) {
5504 ir_node *shl_l = get_Shl_left(and_r);
5505 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5506 if (and_r == t && f == cmp_r) {
5507 if (pn == pn_Cmp_Lg) {
5508 /* (a & (1 << n)) != 0, (1 << n), 0) */
5511 /* (a & (1 << n)) == 0, (1 << n), 0) */
5512 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5513 block, cmp_l, t, mode);
5519 and_l = get_And_left(cmp_l);
5520 if (is_Shl(and_l)) {
5521 ir_node *shl_l = get_Shl_left(and_l);
5522 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5523 if (and_l == t && f == cmp_r) {
5524 if (pn == pn_Cmp_Lg) {
5525 /* ((1 << n) & a) != 0, (1 << n), 0) */
5528 /* ((1 << n) & a) == 0, (1 << n), 0) */
5529 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5530 block, cmp_l, t, mode);
5541 return arch_transform_node_Mux(n);
5542 } /* transform_node_Mux */
5545 * optimize Sync nodes that have other syncs as input we simply add the inputs
5546 * of the other sync to our own inputs
5548 static ir_node *transform_node_Sync(ir_node *n) {
5549 int arity = get_Sync_n_preds(n);
5552 for (i = 0; i < arity;) {
5553 ir_node *pred = get_Sync_pred(n, i);
5557 if (!is_Sync(pred)) {
5565 pred_arity = get_Sync_n_preds(pred);
5566 for (j = 0; j < pred_arity; ++j) {
5567 ir_node *pred_pred = get_Sync_pred(pred, j);
5572 add_irn_n(n, pred_pred);
5576 if (get_Sync_pred(n, k) == pred_pred) break;
5581 /* rehash the sync node */
5582 add_identities(current_ir_graph->value_table, n);
5588 * Tries several [inplace] [optimizing] transformations and returns an
5589 * equivalent node. The difference to equivalent_node() is that these
5590 * transformations _do_ generate new nodes, and thus the old node must
5591 * not be freed even if the equivalent node isn't the old one.
5593 static ir_node *transform_node(ir_node *n) {
5597 * Transform_node is the only "optimizing transformation" that might
5598 * return a node with a different opcode. We iterate HERE until fixpoint
5599 * to get the final result.
5603 if (n->op->ops.transform_node)
5604 n = n->op->ops.transform_node(n);
5605 } while (oldn != n);
5608 } /* transform_node */
5611 * Sets the default transform node operation for an ir_op_ops.
5613 * @param code the opcode for the default operation
5614 * @param ops the operations initialized
5619 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5623 ops->transform_node = transform_node_##a; \
5625 #define CASE_PROJ(a) \
5627 ops->transform_node_Proj = transform_node_Proj_##a; \
5629 #define CASE_PROJ_EX(a) \
5631 ops->transform_node = transform_node_##a; \
5632 ops->transform_node_Proj = transform_node_Proj_##a; \
5641 CASE_PROJ_EX(DivMod);
5675 } /* firm_set_default_transform_node */
5678 /* **************** Common Subexpression Elimination **************** */
5680 /** The size of the hash table used, should estimate the number of nodes
5682 #define N_IR_NODES 512
5684 /** Compares the attributes of two Const nodes. */
5685 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5686 return (get_Const_tarval(a) != get_Const_tarval(b))
5687 || (get_Const_type(a) != get_Const_type(b));
5688 } /* node_cmp_attr_Const */
5690 /** Compares the attributes of two Proj nodes. */
5691 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5692 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5693 } /* node_cmp_attr_Proj */
5695 /** Compares the attributes of two Filter nodes. */
5696 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5697 return get_Filter_proj(a) != get_Filter_proj(b);
5698 } /* node_cmp_attr_Filter */
5700 /** Compares the attributes of two Alloc nodes. */
5701 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5702 const alloc_attr *pa = get_irn_alloc_attr(a);
5703 const alloc_attr *pb = get_irn_alloc_attr(b);
5704 return (pa->where != pb->where) || (pa->type != pb->type);
5705 } /* node_cmp_attr_Alloc */
5707 /** Compares the attributes of two Free nodes. */
5708 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5709 const free_attr *pa = get_irn_free_attr(a);
5710 const free_attr *pb = get_irn_free_attr(b);
5711 return (pa->where != pb->where) || (pa->type != pb->type);
5712 } /* node_cmp_attr_Free */
5714 /** Compares the attributes of two SymConst nodes. */
5715 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5716 const symconst_attr *pa = get_irn_symconst_attr(a);
5717 const symconst_attr *pb = get_irn_symconst_attr(b);
5718 return (pa->kind != pb->kind)
5719 || (pa->sym.type_p != pb->sym.type_p)
5720 || (pa->tp != pb->tp);
5721 } /* node_cmp_attr_SymConst */
5723 /** Compares the attributes of two Call nodes. */
5724 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5725 return get_irn_call_attr(a) != get_irn_call_attr(b);
5726 } /* node_cmp_attr_Call */
5728 /** Compares the attributes of two Sel nodes. */
5729 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5730 const ir_entity *a_ent = get_Sel_entity(a);
5731 const ir_entity *b_ent = get_Sel_entity(b);
5733 (a_ent->kind != b_ent->kind) ||
5734 (a_ent->name != b_ent->name) ||
5735 (a_ent->owner != b_ent->owner) ||
5736 (a_ent->ld_name != b_ent->ld_name) ||
5737 (a_ent->type != b_ent->type);
5738 } /* node_cmp_attr_Sel */
5740 /** Compares the attributes of two Phi nodes. */
5741 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5742 /* we can only enter this function if both nodes have the same number of inputs,
5743 hence it is enough to check if one of them is a Phi0 */
5745 /* check the Phi0 pos attribute */
5746 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5749 } /* node_cmp_attr_Phi */
5751 /** Compares the attributes of two Conv nodes. */
5752 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5753 return get_Conv_strict(a) != get_Conv_strict(b);
5754 } /* node_cmp_attr_Conv */
5756 /** Compares the attributes of two Cast nodes. */
5757 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5758 return get_Cast_type(a) != get_Cast_type(b);
5759 } /* node_cmp_attr_Cast */
5761 /** Compares the attributes of two Load nodes. */
5762 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5763 if (get_Load_volatility(a) == volatility_is_volatile ||
5764 get_Load_volatility(b) == volatility_is_volatile)
5765 /* NEVER do CSE on volatile Loads */
5767 /* do not CSE Loads with different alignment. Be conservative. */
5768 if (get_Load_align(a) != get_Load_align(b))
5771 return get_Load_mode(a) != get_Load_mode(b);
5772 } /* node_cmp_attr_Load */
5774 /** Compares the attributes of two Store nodes. */
5775 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5776 /* do not CSE Stores with different alignment. Be conservative. */
5777 if (get_Store_align(a) != get_Store_align(b))
5780 /* NEVER do CSE on volatile Stores */
5781 return (get_Store_volatility(a) == volatility_is_volatile ||
5782 get_Store_volatility(b) == volatility_is_volatile);
5783 } /* node_cmp_attr_Store */
5785 /** Compares two exception attributes */
5786 static int node_cmp_exception(ir_node *a, ir_node *b) {
5787 const except_attr *ea = get_irn_except_attr(a);
5788 const except_attr *eb = get_irn_except_attr(b);
5790 return ea->pin_state != eb->pin_state;
5793 #define node_cmp_attr_Bound node_cmp_exception
5795 /** Compares the attributes of two Div nodes. */
5796 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5797 const divmod_attr *ma = get_irn_divmod_attr(a);
5798 const divmod_attr *mb = get_irn_divmod_attr(b);
5799 return ma->exc.pin_state != mb->exc.pin_state ||
5800 ma->res_mode != mb->res_mode ||
5801 ma->no_remainder != mb->no_remainder;
5802 } /* node_cmp_attr_Div */
5804 /** Compares the attributes of two DivMod nodes. */
5805 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5806 const divmod_attr *ma = get_irn_divmod_attr(a);
5807 const divmod_attr *mb = get_irn_divmod_attr(b);
5808 return ma->exc.pin_state != mb->exc.pin_state ||
5809 ma->res_mode != mb->res_mode;
5810 } /* node_cmp_attr_DivMod */
5812 /** Compares the attributes of two Mod nodes. */
5813 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5814 const divmod_attr *ma = get_irn_divmod_attr(a);
5815 const divmod_attr *mb = get_irn_divmod_attr(b);
5816 return ma->exc.pin_state != mb->exc.pin_state ||
5817 ma->res_mode != mb->res_mode;
5818 } /* node_cmp_attr_Mod */
5820 /** Compares the attributes of two Quot nodes. */
5821 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5822 const divmod_attr *ma = get_irn_divmod_attr(a);
5823 const divmod_attr *mb = get_irn_divmod_attr(b);
5824 return ma->exc.pin_state != mb->exc.pin_state ||
5825 ma->res_mode != mb->res_mode;
5826 } /* node_cmp_attr_Quot */
5828 /** Compares the attributes of two Confirm nodes. */
5829 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5830 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5831 } /* node_cmp_attr_Confirm */
5833 /** Compares the attributes of two ASM nodes. */
5834 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5836 const ir_asm_constraint *ca;
5837 const ir_asm_constraint *cb;
5840 if (get_ASM_text(a) != get_ASM_text(b))
5843 /* Should we really check the constraints here? Should be better, but is strange. */
5844 n = get_ASM_n_input_constraints(a);
5845 if (n != get_ASM_n_input_constraints(b))
5848 ca = get_ASM_input_constraints(a);
5849 cb = get_ASM_input_constraints(b);
5850 for (i = 0; i < n; ++i) {
5851 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5855 n = get_ASM_n_output_constraints(a);
5856 if (n != get_ASM_n_output_constraints(b))
5859 ca = get_ASM_output_constraints(a);
5860 cb = get_ASM_output_constraints(b);
5861 for (i = 0; i < n; ++i) {
5862 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5866 n = get_ASM_n_clobbers(a);
5867 if (n != get_ASM_n_clobbers(b))
5870 cla = get_ASM_clobbers(a);
5871 clb = get_ASM_clobbers(b);
5872 for (i = 0; i < n; ++i) {
5873 if (cla[i] != clb[i])
5877 } /* node_cmp_attr_ASM */
5880 * Set the default node attribute compare operation for an ir_op_ops.
5882 * @param code the opcode for the default operation
5883 * @param ops the operations initialized
5888 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5892 ops->node_cmp_attr = node_cmp_attr_##a; \
5923 } /* firm_set_default_node_cmp_attr */
5926 * Compare function for two nodes in the value table. Gets two
5927 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5929 int identities_cmp(const void *elt, const void *key) {
5930 ir_node *a = (ir_node *)elt;
5931 ir_node *b = (ir_node *)key;
5934 if (a == b) return 0;
5936 if ((get_irn_op(a) != get_irn_op(b)) ||
5937 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5939 /* compare if a's in and b's in are of equal length */
5940 irn_arity_a = get_irn_intra_arity(a);
5941 if (irn_arity_a != get_irn_intra_arity(b))
5944 if (get_irn_pinned(a) == op_pin_state_pinned) {
5945 /* for pinned nodes, the block inputs must be equal */
5946 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5948 } else if (! get_opt_global_cse()) {
5949 /* for block-local CSE both nodes must be in the same MacroBlock */
5950 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5954 /* compare a->in[0..ins] with b->in[0..ins] */
5955 for (i = 0; i < irn_arity_a; i++)
5956 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5960 * here, we already now that the nodes are identical except their
5963 if (a->op->ops.node_cmp_attr)
5964 return a->op->ops.node_cmp_attr(a, b);
5967 } /* identities_cmp */
5970 * Calculate a hash value of a node.
5972 * @param node The IR-node
5974 unsigned ir_node_hash(const ir_node *node) {
5975 return node->op->ops.hash(node);
5976 } /* ir_node_hash */
5979 pset *new_identities(void) {
5980 return new_pset(identities_cmp, N_IR_NODES);
5981 } /* new_identities */
5983 void del_identities(pset *value_table) {
5984 del_pset(value_table);
5985 } /* del_identities */
5988 * Normalize a node by putting constants (and operands with larger
5989 * node index) on the right (operator side).
5991 * @param n The node to normalize
5993 static void normalize_node(ir_node *n) {
5994 if (is_op_commutative(get_irn_op(n))) {
5995 ir_node *l = get_binop_left(n);
5996 ir_node *r = get_binop_right(n);
5998 /* For commutative operators perform a OP b == b OP a but keep
5999 * constants on the RIGHT side. This helps greatly in some
6000 * optimizations. Moreover we use the idx number to make the form
6002 if (!operands_are_normalized(l, r)) {
6003 set_binop_left(n, r);
6004 set_binop_right(n, l);
6007 } /* normalize_node */
6010 * Update the nodes after a match in the value table. If both nodes have
6011 * the same MacroBlock but different Blocks, we must ensure that the node
6012 * with the dominating Block (the node that is near to the MacroBlock header
6013 * is stored in the table.
6014 * Because a MacroBlock has only one "non-exception" flow, we don't need
6015 * dominance info here: We known, that one block must dominate the other and
6016 * following the only block input will allow to find it.
6018 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6019 ir_node *known_blk, *new_block, *block, *mbh;
6021 if (get_opt_global_cse()) {
6022 /* Block inputs are meaning less */
6025 known_blk = get_irn_n(known_irn, -1);
6026 new_block = get_irn_n(new_ir_node, -1);
6027 if (known_blk == new_block) {
6028 /* already in the same block */
6032 * We expect the typical case when we built the graph. In that case, the
6033 * known_irn is already the upper one, so checking this should be faster.
6036 mbh = get_Block_MacroBlock(new_block);
6038 if (block == known_blk) {
6039 /* ok, we have found it: known_block dominates new_block as expected */
6044 * We have reached the MacroBlock header NOT founding
6045 * the known_block. new_block must dominate known_block.
6048 set_irn_n(known_irn, -1, new_block);
6051 assert(get_Block_n_cfgpreds(block) == 1);
6052 block = get_Block_cfgpred_block(block, 0);
6054 } /* update_value_table */
6057 * Return the canonical node computing the same value as n.
6058 * Looks up the node in a hash table, enters it in the table
6059 * if it isn't there yet.
6061 * @param value_table the HashSet containing all nodes in the
6063 * @param n the node to look up
6065 * @return a node that computes the same value as n or n if no such
6066 * node could be found
6068 ir_node *identify_remember(pset *value_table, ir_node *n) {
6071 if (!value_table) return n;
6074 /* lookup or insert in hash table with given hash key. */
6075 o = pset_insert(value_table, n, ir_node_hash(n));
6078 update_known_irn(o, n);
6083 } /* identify_remember */
6086 * During construction we set the op_pin_state_pinned flag in the graph right when the
6087 * optimization is performed. The flag turning on procedure global cse could
6088 * be changed between two allocations. This way we are safe.
6090 * @param value_table The value table
6091 * @param n The node to lookup
6093 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
6096 n = identify_remember(value_table, n);
6097 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6098 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6100 } /* identify_cons */
6102 /* Add a node to the identities value table. */
6103 void add_identities(pset *value_table, ir_node *node) {
6104 if (get_opt_cse() && is_no_Block(node))
6105 identify_remember(value_table, node);
6106 } /* add_identities */
6108 /* Visit each node in the value table of a graph. */
6109 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6111 ir_graph *rem = current_ir_graph;
6113 current_ir_graph = irg;
6114 foreach_pset(irg->value_table, node)
6116 current_ir_graph = rem;
6117 } /* visit_all_identities */
6120 * Garbage in, garbage out. If a node has a dead input, i.e., the
6121 * Bad node is input to the node, return the Bad node.
6123 static ir_node *gigo(ir_node *node) {
6125 ir_op *op = get_irn_op(node);
6127 /* remove garbage blocks by looking at control flow that leaves the block
6128 and replacing the control flow by Bad. */
6129 if (get_irn_mode(node) == mode_X) {
6130 ir_node *block = get_nodes_block(skip_Proj(node));
6132 /* Don't optimize nodes in immature blocks. */
6133 if (!get_Block_matured(block))
6135 /* Don't optimize End, may have Bads. */
6136 if (op == op_End) return node;
6138 if (is_Block(block)) {
6139 if (is_Block_dead(block)) {
6140 /* control flow from dead block is dead */
6144 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6145 if (!is_Bad(get_irn_n(block, i)))
6149 ir_graph *irg = get_irn_irg(block);
6150 /* the start block is never dead */
6151 if (block != get_irg_start_block(irg)
6152 && block != get_irg_end_block(irg)) {
6154 * Do NOT kill control flow without setting
6155 * the block to dead of bad things can happen:
6156 * We get a Block that is not reachable be irg_block_walk()
6157 * but can be found by irg_walk()!
6159 set_Block_dead(block);
6166 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6167 blocks predecessors is dead. */
6168 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6169 irn_arity = get_irn_arity(node);
6172 * Beware: we can only read the block of a non-floating node.
6174 if (is_irn_pinned_in_irg(node) &&
6175 is_Block_dead(get_nodes_block(skip_Proj(node))))
6178 for (i = 0; i < irn_arity; i++) {
6179 ir_node *pred = get_irn_n(node, i);
6184 /* Propagating Unknowns here seems to be a bad idea, because
6185 sometimes we need a node as a input and did not want that
6187 However, it might be useful to move this into a later phase
6188 (if you think that optimizing such code is useful). */
6189 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6190 return new_Unknown(get_irn_mode(node));
6195 /* With this code we violate the agreement that local_optimize
6196 only leaves Bads in Block, Phi and Tuple nodes. */
6197 /* If Block has only Bads as predecessors it's garbage. */
6198 /* If Phi has only Bads as predecessors it's garbage. */
6199 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6200 irn_arity = get_irn_arity(node);
6201 for (i = 0; i < irn_arity; i++) {
6202 if (!is_Bad(get_irn_n(node, i))) break;
6204 if (i == irn_arity) node = new_Bad();
6211 * These optimizations deallocate nodes from the obstack.
6212 * It can only be called if it is guaranteed that no other nodes
6213 * reference this one, i.e., right after construction of a node.
6215 * @param n The node to optimize
6217 * current_ir_graph must be set to the graph of the node!
6219 ir_node *optimize_node(ir_node *n) {
6222 ir_opcode iro = get_irn_opcode(n);
6224 /* Always optimize Phi nodes: part of the construction. */
6225 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6227 /* constant expression evaluation / constant folding */
6228 if (get_opt_constant_folding()) {
6229 /* neither constants nor Tuple values can be evaluated */
6230 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6231 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6232 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6233 /* try to evaluate */
6234 tv = computed_value(n);
6235 if (tv != tarval_bad) {
6237 ir_type *old_tp = get_irn_type(n);
6238 int i, arity = get_irn_arity(n);
6242 * Try to recover the type of the new expression.
6244 for (i = 0; i < arity && !old_tp; ++i)
6245 old_tp = get_irn_type(get_irn_n(n, i));
6248 * we MUST copy the node here temporary, because it's still needed
6249 * for DBG_OPT_CSTEVAL
6251 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6252 oldn = alloca(node_size);
6254 memcpy(oldn, n, node_size);
6255 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6257 /* ARG, copy the in array, we need it for statistics */
6258 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6260 /* note the inplace edges module */
6261 edges_node_deleted(n, current_ir_graph);
6263 /* evaluation was successful -- replace the node. */
6264 irg_kill_node(current_ir_graph, n);
6265 nw = new_Const(get_tarval_mode(tv), tv);
6267 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6268 set_Const_type(nw, old_tp);
6269 DBG_OPT_CSTEVAL(oldn, nw);
6270 tarval_enable_fp_ops(old_fp_mode);
6273 tarval_enable_fp_ops(old_fp_mode);
6277 /* remove unnecessary nodes */
6278 if (get_opt_algebraic_simplification() ||
6279 (iro == iro_Phi) || /* always optimize these nodes. */
6281 (iro == iro_Proj) ||
6282 (iro == iro_Block) ) /* Flags tested local. */
6283 n = equivalent_node(n);
6285 /* Common Subexpression Elimination.
6287 * Checks whether n is already available.
6288 * The block input is used to distinguish different subexpressions. Right
6289 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6290 * subexpressions within a block.
6293 n = identify_cons(current_ir_graph->value_table, n);
6296 edges_node_deleted(oldn, current_ir_graph);
6298 /* We found an existing, better node, so we can deallocate the old node. */
6299 irg_kill_node(current_ir_graph, oldn);
6303 /* Some more constant expression evaluation that does not allow to
6305 iro = get_irn_opcode(n);
6306 if (get_opt_algebraic_simplification() ||
6307 (iro == iro_Cond) ||
6308 (iro == iro_Proj)) /* Flags tested local. */
6309 n = transform_node(n);
6311 /* Remove nodes with dead (Bad) input.
6312 Run always for transformation induced Bads. */
6315 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6316 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6317 n = identify_remember(current_ir_graph->value_table, n);
6321 } /* optimize_node */
6325 * These optimizations never deallocate nodes (in place). This can cause dead
6326 * nodes lying on the obstack. Remove these by a dead node elimination,
6327 * i.e., a copying garbage collection.
6329 ir_node *optimize_in_place_2(ir_node *n) {
6332 ir_opcode iro = get_irn_opcode(n);
6334 if (!get_opt_optimize() && !is_Phi(n)) return n;
6336 /* constant expression evaluation / constant folding */
6337 if (get_opt_constant_folding()) {
6338 /* neither constants nor Tuple values can be evaluated */
6339 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6340 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6341 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6342 /* try to evaluate */
6343 tv = computed_value(n);
6344 if (tv != tarval_bad) {
6345 /* evaluation was successful -- replace the node. */
6346 ir_type *old_tp = get_irn_type(n);
6347 int i, arity = get_irn_arity(n);
6350 * Try to recover the type of the new expression.
6352 for (i = 0; i < arity && !old_tp; ++i)
6353 old_tp = get_irn_type(get_irn_n(n, i));
6355 n = new_Const(get_tarval_mode(tv), tv);
6357 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6358 set_Const_type(n, old_tp);
6360 DBG_OPT_CSTEVAL(oldn, n);
6361 tarval_enable_fp_ops(old_fp_mode);
6364 tarval_enable_fp_ops(old_fp_mode);
6368 /* remove unnecessary nodes */
6369 if (get_opt_constant_folding() ||
6370 (iro == iro_Phi) || /* always optimize these nodes. */
6371 (iro == iro_Id) || /* ... */
6372 (iro == iro_Proj) || /* ... */
6373 (iro == iro_Block) ) /* Flags tested local. */
6374 n = equivalent_node(n);
6376 /** common subexpression elimination **/
6377 /* Checks whether n is already available. */
6378 /* The block input is used to distinguish different subexpressions. Right
6379 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6380 subexpressions within a block. */
6381 if (get_opt_cse()) {
6382 n = identify_remember(current_ir_graph->value_table, n);
6385 /* Some more constant expression evaluation. */
6386 iro = get_irn_opcode(n);
6387 if (get_opt_constant_folding() ||
6388 (iro == iro_Cond) ||
6389 (iro == iro_Proj)) /* Flags tested local. */
6390 n = transform_node(n);
6392 /* Remove nodes with dead (Bad) input.
6393 Run always for transformation induced Bads. */
6396 /* Now we can verify the node, as it has no dead inputs any more. */
6399 /* Now we have a legal, useful node. Enter it in hash table for cse.
6400 Blocks should be unique anyways. (Except the successor of start:
6401 is cse with the start block!) */
6402 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
6403 n = identify_remember(current_ir_graph->value_table, n);
6406 } /* optimize_in_place_2 */
6409 * Wrapper for external use, set proper status bits after optimization.
6411 ir_node *optimize_in_place(ir_node *n) {
6412 /* Handle graph state */
6413 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6415 if (get_opt_global_cse())
6416 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6417 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6418 set_irg_outs_inconsistent(current_ir_graph);
6420 /* FIXME: Maybe we could also test whether optimizing the node can
6421 change the control graph. */
6422 set_irg_doms_inconsistent(current_ir_graph);
6423 return optimize_in_place_2(n);
6424 } /* optimize_in_place */
6427 * Calculate a hash value of a Const node.
6429 static unsigned hash_Const(const ir_node *node) {
6432 /* special value for const, as they only differ in their tarval. */
6433 h = HASH_PTR(node->attr.con.tv);
6434 h = 9*h + HASH_PTR(get_irn_mode(node));
6440 * Calculate a hash value of a SymConst node.
6442 static unsigned hash_SymConst(const ir_node *node) {
6445 /* special value for const, as they only differ in their symbol. */
6446 h = HASH_PTR(node->attr.symc.sym.type_p);
6447 h = 9*h + HASH_PTR(get_irn_mode(node));
6450 } /* hash_SymConst */
6453 * Set the default hash operation in an ir_op_ops.
6455 * @param code the opcode for the default operation
6456 * @param ops the operations initialized
6461 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6465 ops->hash = hash_##a; \
6468 /* hash function already set */
6469 if (ops->hash != NULL)
6476 /* use input/mode default hash if no function was given */
6477 ops->hash = firm_default_hash;
6485 * Sets the default operation for an ir_ops.
6487 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6488 ops = firm_set_default_hash(code, ops);
6489 ops = firm_set_default_computed_value(code, ops);
6490 ops = firm_set_default_equivalent_node(code, ops);
6491 ops = firm_set_default_transform_node(code, ops);
6492 ops = firm_set_default_node_cmp_attr(code, ops);
6493 ops = firm_set_default_get_type(code, ops);
6494 ops = firm_set_default_get_type_attr(code, ops);
6495 ops = firm_set_default_get_entity_attr(code, ops);
6498 } /* firm_set_default_operations */