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
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
46 #include "opt_confirms.h"
47 #include "opt_polymorphy.h"
51 /* Make types visible to allow most efficient access */
55 * Returns the tarval of a Const node or tarval_bad for all other nodes.
57 static tarval *default_value_of(const ir_node *n) {
59 return get_Const_tarval(n); /* might return tarval_bad */
64 value_of_func value_of_ptr = default_value_of;
66 /* * Set a new value_of function. */
67 void set_value_of_func(value_of_func func) {
71 value_of_ptr = default_value_of;
75 * Return the value of a Constant.
77 static tarval *computed_value_Const(const ir_node *n) {
78 return get_Const_tarval(n);
79 } /* computed_value_Const */
82 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
84 static tarval *computed_value_SymConst(const ir_node *n) {
88 switch (get_SymConst_kind(n)) {
89 case symconst_type_size:
90 type = get_SymConst_type(n);
91 if (get_type_state(type) == layout_fixed)
92 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
94 case symconst_type_align:
95 type = get_SymConst_type(n);
96 if (get_type_state(type) == layout_fixed)
97 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
99 case symconst_ofs_ent:
100 ent = get_SymConst_entity(n);
101 type = get_entity_owner(ent);
102 if (get_type_state(type) == layout_fixed)
103 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
109 } /* computed_value_SymConst */
112 * Return the value of an Add.
114 static tarval *computed_value_Add(const ir_node *n) {
115 ir_node *a = get_Add_left(n);
116 ir_node *b = get_Add_right(n);
118 tarval *ta = value_of(a);
119 tarval *tb = value_of(b);
121 if ((ta != tarval_bad) && (tb != tarval_bad))
122 return tarval_add(ta, tb);
125 } /* computed_value_Add */
128 * Return the value of a Sub.
129 * Special case: a - a
131 static tarval *computed_value_Sub(const ir_node *n) {
132 ir_mode *mode = get_irn_mode(n);
133 ir_node *a = get_Sub_left(n);
134 ir_node *b = get_Sub_right(n);
139 if (a == b && !is_Bad(a))
140 return get_mode_null(mode);
145 if ((ta != tarval_bad) && (tb != tarval_bad))
146 return tarval_sub(ta, tb, mode);
149 } /* computed_value_Sub */
152 * Return the value of a Carry.
153 * Special : a op 0, 0 op b
155 static tarval *computed_value_Carry(const ir_node *n) {
156 ir_node *a = get_binop_left(n);
157 ir_node *b = get_binop_right(n);
158 ir_mode *m = get_irn_mode(n);
160 tarval *ta = value_of(a);
161 tarval *tb = value_of(b);
163 if ((ta != tarval_bad) && (tb != tarval_bad)) {
165 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
167 if (tarval_is_null(ta) || tarval_is_null(tb))
168 return get_mode_null(m);
171 } /* computed_value_Carry */
174 * Return the value of a Borrow.
177 static tarval *computed_value_Borrow(const ir_node *n) {
178 ir_node *a = get_binop_left(n);
179 ir_node *b = get_binop_right(n);
180 ir_mode *m = get_irn_mode(n);
182 tarval *ta = value_of(a);
183 tarval *tb = value_of(b);
185 if ((ta != tarval_bad) && (tb != tarval_bad)) {
186 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
187 } else if (tarval_is_null(ta)) {
188 return get_mode_null(m);
191 } /* computed_value_Borrow */
194 * Return the value of an unary Minus.
196 static tarval *computed_value_Minus(const ir_node *n) {
197 ir_node *a = get_Minus_op(n);
198 tarval *ta = value_of(a);
200 if (ta != tarval_bad)
201 return tarval_neg(ta);
204 } /* computed_value_Minus */
207 * Return the value of a Mul.
209 static tarval *computed_value_Mul(const ir_node *n) {
210 ir_node *a = get_Mul_left(n);
211 ir_node *b = get_Mul_right(n);
214 tarval *ta = value_of(a);
215 tarval *tb = value_of(b);
217 mode = get_irn_mode(n);
218 if (mode != get_irn_mode(a)) {
219 /* n * n = 2n bit multiplication */
220 ta = tarval_convert_to(ta, mode);
221 tb = tarval_convert_to(tb, mode);
224 if (ta != tarval_bad && tb != tarval_bad) {
225 return tarval_mul(ta, tb);
227 /* a*0 = 0 or 0*b = 0 */
228 if (ta == get_mode_null(mode))
230 if (tb == get_mode_null(mode))
234 } /* computed_value_Mul */
237 * Return the value of an Abs.
239 static tarval *computed_value_Abs(const ir_node *n) {
240 ir_node *a = get_Abs_op(n);
241 tarval *ta = value_of(a);
243 if (ta != tarval_bad)
244 return tarval_abs(ta);
247 } /* computed_value_Abs */
250 * Return the value of an And.
251 * Special case: a & 0, 0 & b
253 static tarval *computed_value_And(const ir_node *n) {
254 ir_node *a = get_And_left(n);
255 ir_node *b = get_And_right(n);
257 tarval *ta = value_of(a);
258 tarval *tb = value_of(b);
260 if ((ta != tarval_bad) && (tb != tarval_bad)) {
261 return tarval_and (ta, tb);
263 if (tarval_is_null(ta)) return ta;
264 if (tarval_is_null(tb)) return tb;
267 } /* computed_value_And */
270 * Return the value of an Or.
271 * Special case: a | 1...1, 1...1 | b
273 static tarval *computed_value_Or(const ir_node *n) {
274 ir_node *a = get_Or_left(n);
275 ir_node *b = get_Or_right(n);
277 tarval *ta = value_of(a);
278 tarval *tb = value_of(b);
280 if ((ta != tarval_bad) && (tb != tarval_bad)) {
281 return tarval_or (ta, tb);
283 if (tarval_is_all_one(ta)) return ta;
284 if (tarval_is_all_one(tb)) return tb;
287 } /* computed_value_Or */
290 * Return the value of an Eor.
292 static tarval *computed_value_Eor(const ir_node *n) {
293 ir_node *a = get_Eor_left(n);
294 ir_node *b = get_Eor_right(n);
299 return get_mode_null(get_irn_mode(n));
304 if ((ta != tarval_bad) && (tb != tarval_bad)) {
305 return tarval_eor (ta, tb);
308 } /* computed_value_Eor */
311 * Return the value of a Not.
313 static tarval *computed_value_Not(const ir_node *n) {
314 ir_node *a = get_Not_op(n);
315 tarval *ta = value_of(a);
317 if (ta != tarval_bad)
318 return tarval_not(ta);
321 } /* computed_value_Not */
324 * Return the value of a Shl.
326 static tarval *computed_value_Shl(const ir_node *n) {
327 ir_node *a = get_Shl_left(n);
328 ir_node *b = get_Shl_right(n);
330 tarval *ta = value_of(a);
331 tarval *tb = value_of(b);
333 if ((ta != tarval_bad) && (tb != tarval_bad)) {
334 return tarval_shl (ta, tb);
337 } /* computed_value_Shl */
340 * Return the value of a Shr.
342 static tarval *computed_value_Shr(const ir_node *n) {
343 ir_node *a = get_Shr_left(n);
344 ir_node *b = get_Shr_right(n);
346 tarval *ta = value_of(a);
347 tarval *tb = value_of(b);
349 if ((ta != tarval_bad) && (tb != tarval_bad)) {
350 return tarval_shr (ta, tb);
353 } /* computed_value_Shr */
356 * Return the value of a Shrs.
358 static tarval *computed_value_Shrs(const ir_node *n) {
359 ir_node *a = get_Shrs_left(n);
360 ir_node *b = get_Shrs_right(n);
362 tarval *ta = value_of(a);
363 tarval *tb = value_of(b);
365 if ((ta != tarval_bad) && (tb != tarval_bad)) {
366 return tarval_shrs (ta, tb);
369 } /* computed_value_Shrs */
372 * Return the value of a Rotl.
374 static tarval *computed_value_Rotl(const ir_node *n) {
375 ir_node *a = get_Rotl_left(n);
376 ir_node *b = get_Rotl_right(n);
378 tarval *ta = value_of(a);
379 tarval *tb = value_of(b);
381 if ((ta != tarval_bad) && (tb != tarval_bad)) {
382 return tarval_rotl(ta, tb);
385 } /* computed_value_Rotl */
388 * Return the value of a Conv.
390 static tarval *computed_value_Conv(const ir_node *n) {
391 ir_node *a = get_Conv_op(n);
392 tarval *ta = value_of(a);
394 if (ta != tarval_bad)
395 return tarval_convert_to(ta, get_irn_mode(n));
398 } /* computed_value_Conv */
401 * Calculate the value of a Mux: can be evaluated, if the
402 * sel and the right input are known.
404 static tarval *computed_value_Mux(const ir_node *n) {
405 ir_node *sel = get_Mux_sel(n);
406 tarval *ts = value_of(sel);
408 if (ts == get_tarval_b_true()) {
409 ir_node *v = get_Mux_true(n);
412 else if (ts == get_tarval_b_false()) {
413 ir_node *v = get_Mux_false(n);
417 } /* computed_value_Mux */
420 * Calculate the value of a Confirm: can be evaluated,
421 * if it has the form Confirm(x, '=', Const).
423 static tarval *computed_value_Confirm(const ir_node *n) {
425 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
426 * Do NOT optimize them away (CondEval wants them), so wait until
427 * remove_confirm is activated.
429 if (get_opt_remove_confirm()) {
430 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
431 tarval *tv = value_of(get_Confirm_bound(n));
432 if (tv != tarval_bad)
436 return value_of(get_Confirm_value(n));
437 } /* computed_value_Confirm */
440 * Return the value of a Proj(Cmp).
442 * This performs a first step of unreachable code elimination.
443 * Proj can not be computed, but folding a Cmp above the Proj here is
444 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
446 * There are several case where we can evaluate a Cmp node, see later.
448 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
449 ir_node *a = get_Proj_pred(n);
450 ir_node *aa = get_Cmp_left(a);
451 ir_node *ab = get_Cmp_right(a);
452 long proj_nr = get_Proj_proj(n);
455 * BEWARE: a == a is NOT always True for floating Point values, as
456 * NaN != NaN is defined, so we must check this here.
459 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
462 /* This is a trick with the bits used for encoding the Cmp
463 Proj numbers, the following statement is not the same:
464 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
465 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
468 tarval *taa = value_of(aa);
469 tarval *tab = value_of(ab);
470 ir_mode *mode = get_irn_mode(aa);
473 * The predecessors of Cmp are target values. We can evaluate
476 if ((taa != tarval_bad) && (tab != tarval_bad)) {
477 /* strange checks... */
478 pn_Cmp flags = tarval_cmp(taa, tab);
479 if (flags != pn_Cmp_False) {
480 return new_tarval_from_long (proj_nr & flags, mode_b);
483 /* for integer values, we can check against MIN/MAX */
484 else if (mode_is_int(mode)) {
485 /* MIN <=/> x. This results in true/false. */
486 if (taa == get_mode_min(mode)) {
487 /* a compare with the MIN value */
488 if (proj_nr == pn_Cmp_Le)
489 return get_tarval_b_true();
490 else if (proj_nr == pn_Cmp_Gt)
491 return get_tarval_b_false();
493 /* x >=/< MIN. This results in true/false. */
495 if (tab == get_mode_min(mode)) {
496 /* a compare with the MIN value */
497 if (proj_nr == pn_Cmp_Ge)
498 return get_tarval_b_true();
499 else if (proj_nr == pn_Cmp_Lt)
500 return get_tarval_b_false();
502 /* MAX >=/< x. This results in true/false. */
503 else if (taa == get_mode_max(mode)) {
504 if (proj_nr == pn_Cmp_Ge)
505 return get_tarval_b_true();
506 else if (proj_nr == pn_Cmp_Lt)
507 return get_tarval_b_false();
509 /* x <=/> MAX. This results in true/false. */
510 else if (tab == get_mode_max(mode)) {
511 if (proj_nr == pn_Cmp_Le)
512 return get_tarval_b_true();
513 else if (proj_nr == pn_Cmp_Gt)
514 return get_tarval_b_false();
518 * The predecessors are Allocs or (void*)(0) constants. Allocs never
519 * return NULL, they raise an exception. Therefore we can predict
523 ir_node *aaa = skip_Proj(aa);
524 ir_node *aba = skip_Proj(ab);
526 if ( ( (/* aa is ProjP and aaa is Alloc */
528 && mode_is_reference(get_irn_mode(aa))
530 && ( (/* ab is NULL */
531 mode_is_reference(get_irn_mode(ab))
532 && tarval_is_null(tab))
533 || (/* ab is other Alloc */
535 && mode_is_reference(get_irn_mode(ab))
538 || (/* aa is NULL and aba is Alloc */
539 mode_is_reference(get_irn_mode(aa))
540 && tarval_is_null(taa)
542 && mode_is_reference(get_irn_mode(ab))
545 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
548 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
549 } /* computed_value_Proj_Cmp */
552 * Return the value of a floating point Quot.
554 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
555 tarval *ta = value_of(a);
556 tarval *tb = value_of(b);
558 /* cannot optimize 0 / b = 0 because of NaN */
559 if (ta != tarval_bad && tb != tarval_bad)
560 return tarval_quo(ta, tb);
562 } /* do_computed_value_Quot */
565 * Calculate the value of an integer Div of two nodes.
566 * Special case: 0 / b
568 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
569 tarval *ta = value_of(a);
571 const ir_node *dummy;
573 /* Compute c1 / c2 or 0 / a, a != 0 */
574 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
575 return ta; /* 0 / b == 0 */
577 if (ta != tarval_bad && tb != tarval_bad)
578 return tarval_div(ta, tb);
580 } /* do_computed_value_Div */
583 * Calculate the value of an integer Mod of two nodes.
584 * Special case: a % 1
586 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
587 tarval *ta = value_of(a);
588 tarval *tb = value_of(b);
590 /* Compute a % 1 or c1 % c2 */
591 if (tarval_is_one(tb))
592 return get_mode_null(get_irn_mode(a));
593 if (ta != tarval_bad && tb != tarval_bad)
594 return tarval_mod(ta, tb);
596 } /* do_computed_value_Mod */
599 * Return the value of a Proj(DivMod).
601 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
602 long proj_nr = get_Proj_proj(n);
604 /* compute either the Div or the Mod part */
605 if (proj_nr == pn_DivMod_res_div) {
606 const ir_node *a = get_Proj_pred(n);
607 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
608 } else if (proj_nr == pn_DivMod_res_mod) {
609 const ir_node *a = get_Proj_pred(n);
610 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
613 } /* computed_value_Proj_DivMod */
616 * Return the value of a Proj(Div).
618 static tarval *computed_value_Proj_Div(const ir_node *n) {
619 long proj_nr = get_Proj_proj(n);
621 if (proj_nr == pn_Div_res) {
622 const ir_node *a = get_Proj_pred(n);
623 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
626 } /* computed_value_Proj_Div */
629 * Return the value of a Proj(Mod).
631 static tarval *computed_value_Proj_Mod(const ir_node *n) {
632 long proj_nr = get_Proj_proj(n);
634 if (proj_nr == pn_Mod_res) {
635 const ir_node *a = get_Proj_pred(n);
636 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
639 } /* computed_value_Proj_Mod */
642 * Return the value of a Proj(Quot).
644 static tarval *computed_value_Proj_Quot(const ir_node *n) {
645 long proj_nr = get_Proj_proj(n);
647 if (proj_nr == pn_Quot_res) {
648 const ir_node *a = get_Proj_pred(n);
649 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
652 } /* computed_value_Proj_Quot */
655 * Return the value of a Proj.
657 static tarval *computed_value_Proj(const ir_node *proj) {
658 ir_node *n = get_Proj_pred(proj);
660 if (n->op->ops.computed_value_Proj != NULL)
661 return n->op->ops.computed_value_Proj(proj);
663 } /* computed_value_Proj */
666 * If the parameter n can be computed, return its value, else tarval_bad.
667 * Performs constant folding.
669 * @param n The node this should be evaluated
671 tarval *computed_value(const ir_node *n) {
672 if (n->op->ops.computed_value)
673 return n->op->ops.computed_value(n);
675 } /* computed_value */
678 * Set the default computed_value evaluator in an ir_op_ops.
680 * @param code the opcode for the default operation
681 * @param ops the operations initialized
686 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
690 ops->computed_value = computed_value_##a; \
692 #define CASE_PROJ(a) \
694 ops->computed_value_Proj = computed_value_Proj_##a; \
731 } /* firm_set_default_computed_value */
734 * Returns a equivalent block for another block.
735 * If the block has only one predecessor, this is
736 * the equivalent one. If the only predecessor of a block is
737 * the block itself, this is a dead block.
739 * If both predecessors of a block are the branches of a binary
740 * Cond, the equivalent block is Cond's block.
742 * If all predecessors of a block are bad or lies in a dead
743 * block, the current block is dead as well.
745 * Note, that blocks are NEVER turned into Bad's, instead
746 * the dead_block flag is set. So, never test for is_Bad(block),
747 * always use is_dead_Block(block).
749 static ir_node *equivalent_node_Block(ir_node *n)
754 /* don't optimize dead blocks */
755 if (is_Block_dead(n))
758 n_preds = get_Block_n_cfgpreds(n);
760 /* The Block constructor does not call optimize, but mature_immBlock()
761 calls the optimization. */
762 assert(get_Block_matured(n));
764 /* Straightening: a single entry Block following a single exit Block
765 can be merged, if it is not the Start block. */
766 /* !!! Beware, all Phi-nodes of n must have been optimized away.
767 This should be true, as the block is matured before optimize is called.
768 But what about Phi-cycles with the Phi0/Id that could not be resolved?
769 Remaining Phi nodes are just Ids. */
771 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
774 ir_node *predblock = get_nodes_block(pred);
775 if (predblock == oldn) {
776 /* Jmp jumps into the block it is in -- deal self cycle. */
777 n = set_Block_dead(n);
778 DBG_OPT_DEAD_BLOCK(oldn, n);
779 } else if (get_opt_control_flow_straightening()) {
781 DBG_OPT_STG(oldn, n);
783 } else if (is_Cond(pred)) {
784 ir_node *predblock = get_nodes_block(pred);
785 if (predblock == oldn) {
786 /* Jmp jumps into the block it is in -- deal self cycle. */
787 n = set_Block_dead(n);
788 DBG_OPT_DEAD_BLOCK(oldn, n);
791 } else if ((n_preds == 2) &&
792 (get_opt_control_flow_weak_simplification())) {
793 /* Test whether Cond jumps twice to this block
794 * The more general case which more than 2 predecessors is handles
795 * in optimize_cf(), we handle only this special case for speed here.
797 ir_node *a = get_Block_cfgpred(n, 0);
798 ir_node *b = get_Block_cfgpred(n, 1);
800 if (is_Proj(a) && is_Proj(b)) {
801 ir_node *cond = get_Proj_pred(a);
803 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
804 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
805 /* Also a single entry Block following a single exit Block. Phis have
806 twice the same operand and will be optimized away. */
807 n = get_nodes_block(cond);
808 DBG_OPT_IFSIM1(oldn, a, b, n);
811 } else if (get_opt_unreachable_code() &&
812 (n != get_irg_start_block(current_ir_graph)) &&
813 (n != get_irg_end_block(current_ir_graph)) ) {
816 /* If all inputs are dead, this block is dead too, except if it is
817 the start or end block. This is one step of unreachable code
819 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
820 ir_node *pred = get_Block_cfgpred(n, i);
823 if (is_Bad(pred)) continue;
824 pred_blk = get_nodes_block(skip_Proj(pred));
826 if (is_Block_dead(pred_blk)) continue;
829 /* really found a living input */
834 n = set_Block_dead(n);
835 DBG_OPT_DEAD_BLOCK(oldn, n);
840 } /* equivalent_node_Block */
843 * Returns a equivalent node for a Jmp, a Bad :-)
844 * Of course this only happens if the Block of the Jmp is dead.
846 static ir_node *equivalent_node_Jmp(ir_node *n) {
849 /* unreachable code elimination */
850 if (is_Block_dead(get_nodes_block(n))) {
851 n = get_irg_bad(current_ir_graph);
852 DBG_OPT_DEAD_BLOCK(oldn, n);
855 } /* equivalent_node_Jmp */
857 /** Raise is handled in the same way as Jmp. */
858 #define equivalent_node_Raise equivalent_node_Jmp
861 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
862 See transform_node_Proj_Cond(). */
865 * Optimize operations that are commutative and have neutral 0,
866 * so a op 0 = 0 op a = a.
868 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
871 ir_node *a = get_binop_left(n);
872 ir_node *b = get_binop_right(n);
877 /* After running compute_node there is only one constant predecessor.
878 Find this predecessors value and remember the other node: */
879 if ((tv = value_of(a)) != tarval_bad) {
881 } else if ((tv = value_of(b)) != tarval_bad) {
886 /* If this predecessors constant value is zero, the operation is
887 * unnecessary. Remove it.
889 * Beware: If n is a Add, the mode of on and n might be different
890 * which happens in this rare construction: NULL + 3.
891 * Then, a Conv would be needed which we cannot include here.
893 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
896 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
900 } /* equivalent_node_neutral_zero */
903 * Eor is commutative and has neutral 0.
905 static ir_node *equivalent_node_Eor(ir_node *n) {
910 n = equivalent_node_neutral_zero(n);
911 if (n != oldn) return n;
914 b = get_Eor_right(n);
917 ir_node *aa = get_Eor_left(a);
918 ir_node *ab = get_Eor_right(a);
921 /* (a ^ b) ^ a -> b */
923 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
925 } else if (ab == b) {
926 /* (a ^ b) ^ b -> a */
928 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
933 ir_node *ba = get_Eor_left(b);
934 ir_node *bb = get_Eor_right(b);
937 /* a ^ (a ^ b) -> b */
939 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
941 } else if (bb == a) {
942 /* a ^ (b ^ a) -> b */
944 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
952 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
954 * The second one looks strange, but this construct
955 * is used heavily in the LCC sources :-).
957 * Beware: The Mode of an Add may be different than the mode of its
958 * predecessors, so we could not return a predecessors in all cases.
960 static ir_node *equivalent_node_Add(ir_node *n) {
962 ir_node *left, *right;
963 ir_mode *mode = get_irn_mode(n);
965 n = equivalent_node_neutral_zero(n);
969 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
970 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
973 left = get_Add_left(n);
974 right = get_Add_right(n);
977 if (get_Sub_right(left) == right) {
980 n = get_Sub_left(left);
981 if (mode == get_irn_mode(n)) {
982 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
988 if (get_Sub_right(right) == left) {
991 n = get_Sub_left(right);
992 if (mode == get_irn_mode(n)) {
993 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
999 } /* equivalent_node_Add */
1002 * optimize operations that are not commutative but have neutral 0 on left,
1005 static ir_node *equivalent_node_left_zero(ir_node *n) {
1008 ir_node *a = get_binop_left(n);
1009 ir_node *b = get_binop_right(n);
1010 tarval *tb = value_of(b);
1012 if (tarval_is_null(tb)) {
1015 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1018 } /* equivalent_node_left_zero */
1020 #define equivalent_node_Shl equivalent_node_left_zero
1021 #define equivalent_node_Shr equivalent_node_left_zero
1022 #define equivalent_node_Shrs equivalent_node_left_zero
1023 #define equivalent_node_Rotl equivalent_node_left_zero
1026 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1028 * The second one looks strange, but this construct
1029 * is used heavily in the LCC sources :-).
1031 * Beware: The Mode of a Sub may be different than the mode of its
1032 * predecessors, so we could not return a predecessors in all cases.
1034 static ir_node *equivalent_node_Sub(ir_node *n) {
1037 ir_mode *mode = get_irn_mode(n);
1040 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1041 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1044 b = get_Sub_right(n);
1047 /* Beware: modes might be different */
1048 if (tarval_is_null(tb)) {
1049 ir_node *a = get_Sub_left(n);
1050 if (mode == get_irn_mode(a)) {
1053 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1057 } /* equivalent_node_Sub */
1061 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1064 * -(-a) == a, but might overflow two times.
1065 * We handle it anyway here but the better way would be a
1066 * flag. This would be needed for Pascal for instance.
1068 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1070 ir_node *pred = get_unop_op(n);
1072 /* optimize symmetric unop */
1073 if (get_irn_op(pred) == get_irn_op(n)) {
1074 n = get_unop_op(pred);
1075 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1078 } /* equivalent_node_idempotent_unop */
1080 /** Optimize Not(Not(x)) == x. */
1081 #define equivalent_node_Not equivalent_node_idempotent_unop
1083 /** -(-x) == x ??? Is this possible or can --x raise an
1084 out of bounds exception if min =! max? */
1085 #define equivalent_node_Minus equivalent_node_idempotent_unop
1088 * Optimize a * 1 = 1 * a = a.
1090 static ir_node *equivalent_node_Mul(ir_node *n) {
1092 ir_node *a = get_Mul_left(n);
1094 /* we can handle here only the n * n = n bit cases */
1095 if (get_irn_mode(n) == get_irn_mode(a)) {
1096 ir_node *b = get_Mul_right(n);
1100 * Mul is commutative and has again an other neutral element.
1101 * Constants are place right, so check this case first.
1104 if (tarval_is_one(tv)) {
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1109 if (tarval_is_one(tv)) {
1111 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1116 } /* equivalent_node_Mul */
1119 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1121 static ir_node *equivalent_node_Or(ir_node *n) {
1124 ir_node *a = get_Or_left(n);
1125 ir_node *b = get_Or_right(n);
1129 n = a; /* Or has it's own neutral element */
1130 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1133 /* constants are cormalized to right, check this site first */
1135 if (tarval_is_null(tv)) {
1137 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1141 if (tarval_is_null(tv)) {
1143 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1148 } /* equivalent_node_Or */
1151 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1153 static ir_node *equivalent_node_And(ir_node *n) {
1156 ir_node *a = get_And_left(n);
1157 ir_node *b = get_And_right(n);
1161 n = a; /* And has it's own neutral element */
1162 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1165 /* constants are cormalized to right, check this site first */
1167 if (tarval_is_all_one(tv)) {
1169 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1173 if (tarval_is_all_one(tv)) {
1175 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1179 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1182 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1187 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1190 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1195 } /* equivalent_node_And */
1198 * Try to remove useless Conv's:
1200 static ir_node *equivalent_node_Conv(ir_node *n) {
1202 ir_node *a = get_Conv_op(n);
1204 ir_mode *n_mode = get_irn_mode(n);
1205 ir_mode *a_mode = get_irn_mode(a);
1208 if (n_mode == a_mode) { /* No Conv necessary */
1209 if (get_Conv_strict(n)) {
1210 /* special case: the predecessor might be a also a Conv */
1212 if (! get_Conv_strict(a)) {
1213 /* first one is not strict, kick it */
1215 a_mode = get_irn_mode(a);
1219 /* else both are strict conv, second is superfluous */
1222 ir_node *pred = get_Proj_pred(a);
1223 if (is_Load(pred)) {
1224 /* loads always return with the exact precision of n_mode */
1225 assert(get_Load_mode(pred) == n_mode);
1229 /* leave strict floating point Conv's */
1234 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1235 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1236 ir_node *b = get_Conv_op(a);
1237 ir_mode *b_mode = get_irn_mode(b);
1239 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1240 /* both are strict conv */
1241 if (smaller_mode(a_mode, n_mode)) {
1242 /* both are strict, but the first is smaller, so
1243 the second cannot remove more precision, remove the
1245 set_Conv_strict(n, 0);
1248 if (n_mode == b_mode) {
1249 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1250 if (n_mode == mode_b) {
1251 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1252 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1254 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1255 if (values_in_mode(b_mode, a_mode)) {
1256 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1257 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1262 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1263 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1264 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1265 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1267 if (float_mantissa >= int_mantissa) {
1269 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1274 if (smaller_mode(b_mode, a_mode)) {
1275 if (get_Conv_strict(n))
1276 set_Conv_strict(b, 1);
1277 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1278 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1285 } /* equivalent_node_Conv */
1288 * A Cast may be removed if the type of the previous node
1289 * is already the type of the Cast.
1291 static ir_node *equivalent_node_Cast(ir_node *n) {
1293 ir_node *pred = get_Cast_op(n);
1295 if (get_irn_type(pred) == get_Cast_type(n)) {
1297 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1300 } /* equivalent_node_Cast */
1303 * - fold Phi-nodes, iff they have only one predecessor except
1306 static ir_node *equivalent_node_Phi(ir_node *n) {
1311 ir_node *first_val = NULL; /* to shutup gcc */
1313 if (!get_opt_normalize()) return n;
1315 n_preds = get_Phi_n_preds(n);
1317 block = get_nodes_block(n);
1318 if (is_Block_dead(block)) /* Control dead */
1319 return get_irg_bad(current_ir_graph);
1321 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1323 /* Find first non-self-referencing input */
1324 for (i = 0; i < n_preds; ++i) {
1325 first_val = get_Phi_pred(n, i);
1326 if ( (first_val != n) /* not self pointer */
1328 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1329 * predecessors. Then, Phi nodes in dead code might be removed, causing
1330 * nodes pointing to themself (Add's for instance).
1331 * This is really bad and causes endless recursions in several
1332 * code pathes, so we do NOT optimize such a code.
1333 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1334 * (and bad Phi predecessors), so live code is optimized later.
1336 && (! is_Bad(get_Block_cfgpred(block, i)))
1338 ) { /* value not dead */
1339 break; /* then found first value. */
1344 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1345 return get_irg_bad(current_ir_graph);
1348 /* search for rest of inputs, determine if any of these
1349 are non-self-referencing */
1350 while (++i < n_preds) {
1351 ir_node *scnd_val = get_Phi_pred(n, i);
1352 if ( (scnd_val != n)
1353 && (scnd_val != first_val)
1356 && (! is_Bad(get_Block_cfgpred(block, i)))
1364 /* Fold, if no multiple distinct non-self-referencing inputs */
1366 DBG_OPT_PHI(oldn, n);
1369 } /* equivalent_node_Phi */
1372 * Several optimizations:
1373 * - fold Sync-nodes, iff they have only one predecessor except
1376 static ir_node *equivalent_node_Sync(ir_node *n) {
1377 int arity = get_Sync_n_preds(n);
1380 for (i = 0; i < arity;) {
1381 ir_node *pred = get_Sync_pred(n, i);
1384 /* Remove Bad predecessors */
1391 /* Remove duplicate predecessors */
1397 if (get_Sync_pred(n, j) == pred) {
1405 if (arity == 0) return get_irg_bad(current_ir_graph);
1406 if (arity == 1) return get_Sync_pred(n, 0);
1408 } /* equivalent_node_Sync */
1411 * Optimize Proj(Tuple).
1413 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1414 ir_node *oldn = proj;
1415 ir_node *tuple = get_Proj_pred(proj);
1417 /* Remove the Tuple/Proj combination. */
1418 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1419 DBG_OPT_TUPLE(oldn, tuple, proj);
1422 } /* equivalent_node_Proj_Tuple */
1425 * Optimize a / 1 = a.
1427 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1428 ir_node *oldn = proj;
1429 ir_node *div = get_Proj_pred(proj);
1430 ir_node *b = get_Div_right(div);
1431 tarval *tb = value_of(b);
1433 /* Div is not commutative. */
1434 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1435 switch (get_Proj_proj(proj)) {
1437 proj = get_Div_mem(div);
1438 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1442 proj = get_Div_left(div);
1443 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1447 /* we cannot replace the exception Proj's here, this is done in
1448 transform_node_Proj_Div() */
1453 } /* equivalent_node_Proj_Div */
1456 * Optimize a / 1.0 = a.
1458 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1459 ir_node *oldn = proj;
1460 ir_node *quot = get_Proj_pred(proj);
1461 ir_node *b = get_Quot_right(quot);
1462 tarval *tb = value_of(b);
1464 /* Div is not commutative. */
1465 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1466 switch (get_Proj_proj(proj)) {
1468 proj = get_Quot_mem(quot);
1469 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1473 proj = get_Quot_left(quot);
1474 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1478 /* we cannot replace the exception Proj's here, this is done in
1479 transform_node_Proj_Quot() */
1484 } /* equivalent_node_Proj_Quot */
1487 * Optimize a / 1 = a.
1489 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1490 ir_node *oldn = proj;
1491 ir_node *divmod = get_Proj_pred(proj);
1492 ir_node *b = get_DivMod_right(divmod);
1493 tarval *tb = value_of(b);
1495 /* Div is not commutative. */
1496 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1497 switch (get_Proj_proj(proj)) {
1499 proj = get_DivMod_mem(divmod);
1500 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1503 case pn_DivMod_res_div:
1504 proj = get_DivMod_left(divmod);
1505 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1509 /* we cannot replace the exception Proj's here, this is done in
1510 transform_node_Proj_DivMod().
1511 Note further that the pn_DivMod_res_div case is handled in
1512 computed_value_Proj(). */
1517 } /* equivalent_node_Proj_DivMod */
1520 * Optimize CopyB(mem, x, x) into a Nop.
1522 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1523 ir_node *oldn = proj;
1524 ir_node *copyb = get_Proj_pred(proj);
1525 ir_node *a = get_CopyB_dst(copyb);
1526 ir_node *b = get_CopyB_src(copyb);
1529 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1530 switch (get_Proj_proj(proj)) {
1531 case pn_CopyB_M_regular:
1532 proj = get_CopyB_mem(copyb);
1533 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1536 case pn_CopyB_M_except:
1537 case pn_CopyB_X_except:
1538 DBG_OPT_EXC_REM(proj);
1539 proj = get_irg_bad(current_ir_graph);
1544 } /* equivalent_node_Proj_CopyB */
1547 * Optimize Bounds(idx, idx, upper) into idx.
1549 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1550 ir_node *oldn = proj;
1551 ir_node *bound = get_Proj_pred(proj);
1552 ir_node *idx = get_Bound_index(bound);
1553 ir_node *pred = skip_Proj(idx);
1556 if (idx == get_Bound_lower(bound))
1558 else if (is_Bound(pred)) {
1560 * idx was Bounds checked in the same MacroBlock previously,
1561 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1563 ir_node *lower = get_Bound_lower(bound);
1564 ir_node *upper = get_Bound_upper(bound);
1565 if (get_Bound_lower(pred) == lower &&
1566 get_Bound_upper(pred) == upper &&
1567 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1569 * One could expect that we simply return the previous
1570 * Bound here. However, this would be wrong, as we could
1571 * add an exception Proj to a new location then.
1572 * So, we must turn in into a tuple.
1578 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1579 switch (get_Proj_proj(proj)) {
1581 DBG_OPT_EXC_REM(proj);
1582 proj = get_Bound_mem(bound);
1584 case pn_Bound_X_except:
1585 DBG_OPT_EXC_REM(proj);
1586 proj = get_irg_bad(current_ir_graph);
1590 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1593 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1598 } /* equivalent_node_Proj_Bound */
1601 * Optimize an Exception Proj(Load) with a non-null address.
1603 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1604 if (get_opt_ldst_only_null_ptr_exceptions()) {
1605 if (get_irn_mode(proj) == mode_X) {
1606 ir_node *load = get_Proj_pred(proj);
1608 /* get the Load address */
1609 const ir_node *addr = get_Load_ptr(load);
1610 const ir_node *confirm;
1612 if (value_not_null(addr, &confirm)) {
1613 if (get_Proj_proj(proj) == pn_Load_X_except) {
1614 DBG_OPT_EXC_REM(proj);
1615 return get_irg_bad(current_ir_graph);
1621 } /* equivalent_node_Proj_Load */
1624 * Optimize an Exception Proj(Store) with a non-null address.
1626 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1627 if (get_opt_ldst_only_null_ptr_exceptions()) {
1628 if (get_irn_mode(proj) == mode_X) {
1629 ir_node *store = get_Proj_pred(proj);
1631 /* get the load/store address */
1632 const ir_node *addr = get_Store_ptr(store);
1633 const ir_node *confirm;
1635 if (value_not_null(addr, &confirm)) {
1636 if (get_Proj_proj(proj) == pn_Store_X_except) {
1637 DBG_OPT_EXC_REM(proj);
1638 return get_irg_bad(current_ir_graph);
1644 } /* equivalent_node_Proj_Store */
1647 * Does all optimizations on nodes that must be done on it's Proj's
1648 * because of creating new nodes.
1650 static ir_node *equivalent_node_Proj(ir_node *proj) {
1651 ir_node *n = get_Proj_pred(proj);
1653 if (get_irn_mode(proj) == mode_X) {
1654 if (is_Block_dead(get_nodes_block(n))) {
1655 /* Remove dead control flow -- early gigo(). */
1656 return get_irg_bad(current_ir_graph);
1659 if (n->op->ops.equivalent_node_Proj)
1660 return n->op->ops.equivalent_node_Proj(proj);
1662 } /* equivalent_node_Proj */
1667 static ir_node *equivalent_node_Id(ir_node *n) {
1674 DBG_OPT_ID(oldn, n);
1676 } /* equivalent_node_Id */
1681 static ir_node *equivalent_node_Mux(ir_node *n)
1683 ir_node *oldn = n, *sel = get_Mux_sel(n);
1684 tarval *ts = value_of(sel);
1686 /* Mux(true, f, t) == t */
1687 if (ts == tarval_b_true) {
1688 n = get_Mux_true(n);
1689 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1691 /* Mux(false, f, t) == f */
1692 else if (ts == tarval_b_false) {
1693 n = get_Mux_false(n);
1694 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1696 /* Mux(v, x, x) == x */
1697 else if (get_Mux_false(n) == get_Mux_true(n)) {
1698 n = get_Mux_true(n);
1699 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1701 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1702 ir_node *cmp = get_Proj_pred(sel);
1703 long proj_nr = get_Proj_proj(sel);
1704 ir_node *f = get_Mux_false(n);
1705 ir_node *t = get_Mux_true(n);
1708 * Note further that these optimization work even for floating point
1709 * with NaN's because -NaN == NaN.
1710 * However, if +0 and -0 is handled differently, we cannot use the first one.
1713 ir_node *const cmp_l = get_Cmp_left(cmp);
1714 ir_node *const cmp_r = get_Cmp_right(cmp);
1718 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1719 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1721 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1728 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1729 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1731 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1738 * Note: normalization puts the constant on the right side,
1739 * so we check only one case.
1741 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1742 /* Mux(t CMP 0, X, t) */
1743 if (is_Minus(f) && get_Minus_op(f) == t) {
1744 /* Mux(t CMP 0, -t, t) */
1745 if (proj_nr == pn_Cmp_Eq) {
1746 /* Mux(t == 0, -t, t) ==> -t */
1748 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1749 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1750 /* Mux(t != 0, -t, t) ==> t */
1752 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1759 } /* equivalent_node_Mux */
1762 * Remove Confirm nodes if setting is on.
1763 * Replace Confirms(x, '=', Constlike) by Constlike.
1765 static ir_node *equivalent_node_Confirm(ir_node *n) {
1766 ir_node *pred = get_Confirm_value(n);
1767 pn_Cmp pnc = get_Confirm_cmp(n);
1769 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1771 * rare case: two identical Confirms one after another,
1772 * replace the second one with the first.
1775 pred = get_Confirm_value(n);
1776 pnc = get_Confirm_cmp(n);
1778 if (get_opt_remove_confirm())
1779 return get_Confirm_value(n);
1784 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1785 * perform no actual computation, as, e.g., the Id nodes. It does not create
1786 * new nodes. It is therefore safe to free n if the node returned is not n.
1787 * If a node returns a Tuple we can not just skip it. If the size of the
1788 * in array fits, we transform n into a tuple (e.g., Div).
1790 ir_node *equivalent_node(ir_node *n) {
1791 if (n->op->ops.equivalent_node)
1792 return n->op->ops.equivalent_node(n);
1794 } /* equivalent_node */
1797 * Sets the default equivalent node operation for an ir_op_ops.
1799 * @param code the opcode for the default operation
1800 * @param ops the operations initialized
1805 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1809 ops->equivalent_node = equivalent_node_##a; \
1811 #define CASE_PROJ(a) \
1813 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1855 } /* firm_set_default_equivalent_node */
1858 * Returns non-zero if a node is a Phi node
1859 * with all predecessors constant.
1861 static int is_const_Phi(ir_node *n) {
1864 if (! is_Phi(n) || get_irn_arity(n) == 0)
1866 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1867 if (! is_Const(get_irn_n(n, i)))
1871 } /* is_const_Phi */
1873 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1874 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1877 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1879 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1880 if (eval == tarval_sub) {
1881 tarval_sub_type func = (tarval_sub_type)eval;
1883 return func(a, b, mode);
1885 tarval_binop_type func = (tarval_binop_type)eval;
1892 * Apply an evaluator on a binop with a constant operators (and one Phi).
1894 * @param phi the Phi node
1895 * @param other the other operand
1896 * @param eval an evaluator function
1897 * @param mode the mode of the result, may be different from the mode of the Phi!
1898 * @param left if non-zero, other is the left operand, else the right
1900 * @return a new Phi node if the conversion was successful, NULL else
1902 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1907 int i, n = get_irn_arity(phi);
1909 NEW_ARR_A(void *, res, n);
1911 for (i = 0; i < n; ++i) {
1912 pred = get_irn_n(phi, i);
1913 tv = get_Const_tarval(pred);
1914 tv = do_eval(eval, other, tv, mode);
1916 if (tv == tarval_bad) {
1917 /* folding failed, bad */
1923 for (i = 0; i < n; ++i) {
1924 pred = get_irn_n(phi, i);
1925 tv = get_Const_tarval(pred);
1926 tv = do_eval(eval, tv, other, mode);
1928 if (tv == tarval_bad) {
1929 /* folding failed, bad */
1935 irg = current_ir_graph;
1936 for (i = 0; i < n; ++i) {
1937 pred = get_irn_n(phi, i);
1938 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1939 mode, res[i], get_Const_type(pred));
1941 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1942 } /* apply_binop_on_phi */
1945 * Apply an evaluator on a binop with two constant Phi.
1947 * @param a the left Phi node
1948 * @param b the right Phi node
1949 * @param eval an evaluator function
1950 * @param mode the mode of the result, may be different from the mode of the Phi!
1952 * @return a new Phi node if the conversion was successful, NULL else
1954 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1955 tarval *tv_l, *tv_r, *tv;
1961 if (get_nodes_block(a) != get_nodes_block(b))
1964 n = get_irn_arity(a);
1965 NEW_ARR_A(void *, res, n);
1967 for (i = 0; i < n; ++i) {
1968 pred = get_irn_n(a, i);
1969 tv_l = get_Const_tarval(pred);
1970 pred = get_irn_n(b, i);
1971 tv_r = get_Const_tarval(pred);
1972 tv = do_eval(eval, tv_l, tv_r, mode);
1974 if (tv == tarval_bad) {
1975 /* folding failed, bad */
1980 irg = current_ir_graph;
1981 for (i = 0; i < n; ++i) {
1982 pred = get_irn_n(a, i);
1983 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1985 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1986 } /* apply_binop_on_2_phis */
1989 * Apply an evaluator on a unop with a constant operator (a Phi).
1991 * @param phi the Phi node
1992 * @param eval an evaluator function
1994 * @return a new Phi node if the conversion was successful, NULL else
1996 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2002 int i, n = get_irn_arity(phi);
2004 NEW_ARR_A(void *, res, n);
2005 for (i = 0; i < n; ++i) {
2006 pred = get_irn_n(phi, i);
2007 tv = get_Const_tarval(pred);
2010 if (tv == tarval_bad) {
2011 /* folding failed, bad */
2016 mode = get_irn_mode(phi);
2017 irg = current_ir_graph;
2018 for (i = 0; i < n; ++i) {
2019 pred = get_irn_n(phi, i);
2020 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2021 mode, res[i], get_Const_type(pred));
2023 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2024 } /* apply_unop_on_phi */
2027 * Apply a conversion on a constant operator (a Phi).
2029 * @param phi the Phi node
2031 * @return a new Phi node if the conversion was successful, NULL else
2033 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2038 int i, n = get_irn_arity(phi);
2040 NEW_ARR_A(void *, res, n);
2041 for (i = 0; i < n; ++i) {
2042 pred = get_irn_n(phi, i);
2043 tv = get_Const_tarval(pred);
2044 tv = tarval_convert_to(tv, mode);
2046 if (tv == tarval_bad) {
2047 /* folding failed, bad */
2052 irg = current_ir_graph;
2053 for (i = 0; i < n; ++i) {
2054 pred = get_irn_n(phi, i);
2055 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2056 mode, res[i], get_Const_type(pred));
2058 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2059 } /* apply_conv_on_phi */
2062 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2063 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2064 * If possible, remove the Conv's.
2066 static ir_node *transform_node_AddSub(ir_node *n) {
2067 ir_mode *mode = get_irn_mode(n);
2069 if (mode_is_reference(mode)) {
2070 ir_node *left = get_binop_left(n);
2071 ir_node *right = get_binop_right(n);
2072 unsigned ref_bits = get_mode_size_bits(mode);
2074 if (is_Conv(left)) {
2075 ir_mode *lmode = get_irn_mode(left);
2076 unsigned bits = get_mode_size_bits(lmode);
2078 if (ref_bits == bits &&
2079 mode_is_int(lmode) &&
2080 get_mode_arithmetic(lmode) == irma_twos_complement) {
2081 ir_node *pre = get_Conv_op(left);
2082 ir_mode *pre_mode = get_irn_mode(pre);
2084 if (mode_is_int(pre_mode) &&
2085 get_mode_size_bits(pre_mode) == bits &&
2086 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2087 /* ok, this conv just changes to sign, moreover the calculation
2088 * is done with same number of bits as our address mode, so
2089 * we can ignore the conv as address calculation can be viewed
2090 * as either signed or unsigned
2092 set_binop_left(n, pre);
2097 if (is_Conv(right)) {
2098 ir_mode *rmode = get_irn_mode(right);
2099 unsigned bits = get_mode_size_bits(rmode);
2101 if (ref_bits == bits &&
2102 mode_is_int(rmode) &&
2103 get_mode_arithmetic(rmode) == irma_twos_complement) {
2104 ir_node *pre = get_Conv_op(right);
2105 ir_mode *pre_mode = get_irn_mode(pre);
2107 if (mode_is_int(pre_mode) &&
2108 get_mode_size_bits(pre_mode) == bits &&
2109 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2110 /* ok, this conv just changes to sign, moreover the calculation
2111 * is done with same number of bits as our address mode, so
2112 * we can ignore the conv as address calculation can be viewed
2113 * as either signed or unsigned
2115 set_binop_right(n, pre);
2120 /* let address arithmetic use unsigned modes */
2121 if (is_Const(right)) {
2122 ir_mode *rmode = get_irn_mode(right);
2124 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2125 /* convert a AddP(P, *s) into AddP(P, *u) */
2126 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2128 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2129 set_binop_right(n, pre);
2135 } /* transform_node_AddSub */
2137 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2139 if (is_Const(b) && is_const_Phi(a)) { \
2140 /* check for Op(Phi, Const) */ \
2141 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2143 else if (is_Const(a) && is_const_Phi(b)) { \
2144 /* check for Op(Const, Phi) */ \
2145 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2147 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2148 /* check for Op(Phi, Phi) */ \
2149 c = apply_binop_on_2_phis(a, b, eval, mode); \
2152 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2156 #define HANDLE_UNOP_PHI(eval, a, c) \
2158 if (is_const_Phi(a)) { \
2159 /* check for Op(Phi) */ \
2160 c = apply_unop_on_phi(a, eval); \
2162 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2168 * Do the AddSub optimization, then Transform
2169 * Constant folding on Phi
2170 * Add(a,a) -> Mul(a, 2)
2171 * Add(Mul(a, x), a) -> Mul(a, x+1)
2172 * if the mode is integer or float.
2173 * Transform Add(a,-b) into Sub(a,b).
2174 * Reassociation might fold this further.
2176 static ir_node *transform_node_Add(ir_node *n) {
2178 ir_node *a, *b, *c, *oldn = n;
2180 n = transform_node_AddSub(n);
2182 a = get_Add_left(n);
2183 b = get_Add_right(n);
2185 mode = get_irn_mode(n);
2187 if (mode_is_reference(mode)) {
2188 ir_mode *lmode = get_irn_mode(a);
2190 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2191 /* an Add(a, NULL) is a hidden Conv */
2192 dbg_info *dbg = get_irn_dbg_info(n);
2193 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2197 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2199 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2200 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2203 if (mode_is_num(mode)) {
2204 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2205 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2206 ir_node *block = get_nodes_block(n);
2209 get_irn_dbg_info(n),
2213 new_r_Const_long(current_ir_graph, block, mode, 2),
2215 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2220 get_irn_dbg_info(n),
2226 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2231 get_irn_dbg_info(n),
2237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2240 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2241 /* Here we rely on constants be on the RIGHT side */
2243 ir_node *op = get_Not_op(a);
2245 if (is_Const(b) && is_Const_one(b)) {
2247 ir_node *blk = get_nodes_block(n);
2248 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2249 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2254 ir_node *blk = get_nodes_block(n);
2255 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2256 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2261 ir_node *op = get_Not_op(b);
2265 ir_node *blk = get_nodes_block(n);
2266 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2274 } /* transform_node_Add */
2277 * returns -cnst or NULL if impossible
2279 static ir_node *const_negate(ir_node *cnst) {
2280 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2281 dbg_info *dbgi = get_irn_dbg_info(cnst);
2282 ir_graph *irg = get_irn_irg(cnst);
2283 ir_node *block = get_nodes_block(cnst);
2284 ir_mode *mode = get_irn_mode(cnst);
2285 if (tv == tarval_bad) return NULL;
2286 return new_rd_Const(dbgi, irg, block, mode, tv);
2290 * Do the AddSub optimization, then Transform
2291 * Constant folding on Phi
2292 * Sub(0,a) -> Minus(a)
2293 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2294 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2295 * Sub(Add(a, x), x) -> a
2296 * Sub(x, Add(x, a)) -> -a
2297 * Sub(x, Const) -> Add(x, -Const)
2299 static ir_node *transform_node_Sub(ir_node *n) {
2304 n = transform_node_AddSub(n);
2306 a = get_Sub_left(n);
2307 b = get_Sub_right(n);
2309 mode = get_irn_mode(n);
2311 if (mode_is_int(mode)) {
2312 ir_mode *lmode = get_irn_mode(a);
2314 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2315 /* a Sub(a, NULL) is a hidden Conv */
2316 dbg_info *dbg = get_irn_dbg_info(n);
2317 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2318 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2322 if (mode == lmode &&
2323 get_mode_arithmetic(mode) == irma_twos_complement &&
2325 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2327 dbg_info *dbg = get_irn_dbg_info(n);
2328 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2329 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2335 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2337 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2338 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2341 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2342 /* a - C -> a + (-C) */
2343 ir_node *cnst = const_negate(b);
2345 ir_node *block = get_nodes_block(n);
2346 dbg_info *dbgi = get_irn_dbg_info(n);
2347 ir_graph *irg = get_irn_irg(n);
2349 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2350 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2355 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2356 ir_graph *irg = current_ir_graph;
2357 dbg_info *dbg = get_irn_dbg_info(n);
2358 ir_node *block = get_nodes_block(n);
2359 ir_node *left = get_Minus_op(a);
2360 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2362 n = new_rd_Minus(dbg, irg, block, add, mode);
2363 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2365 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2366 ir_graph *irg = current_ir_graph;
2367 dbg_info *dbg = get_irn_dbg_info(n);
2368 ir_node *block = get_nodes_block(n);
2369 ir_node *right = get_Minus_op(b);
2371 n = new_rd_Add(dbg, irg, block, a, right, mode);
2372 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2374 } else if (is_Sub(b)) {
2375 /* a - (b - c) -> a + (c - b)
2376 * -> (a - b) + c iff (b - c) is a pointer */
2377 ir_graph *irg = current_ir_graph;
2378 dbg_info *s_dbg = get_irn_dbg_info(b);
2379 ir_node *s_block = get_nodes_block(b);
2380 ir_node *s_left = get_Sub_left(b);
2381 ir_node *s_right = get_Sub_right(b);
2382 ir_mode *s_mode = get_irn_mode(b);
2383 if (s_mode == mode_P) {
2384 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2385 dbg_info *a_dbg = get_irn_dbg_info(n);
2386 ir_node *a_block = get_nodes_block(n);
2389 s_right = new_r_Conv(irg, a_block, s_right, mode);
2390 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2392 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2393 dbg_info *a_dbg = get_irn_dbg_info(n);
2394 ir_node *a_block = get_nodes_block(n);
2396 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2398 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2400 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2401 ir_node *m_right = get_Mul_right(b);
2402 if (is_Const(m_right)) {
2403 ir_node *cnst2 = const_negate(m_right);
2404 if (cnst2 != NULL) {
2405 ir_graph *irg = current_ir_graph;
2406 dbg_info *m_dbg = get_irn_dbg_info(b);
2407 ir_node *m_block = get_nodes_block(b);
2408 ir_node *m_left = get_Mul_left(b);
2409 ir_mode *m_mode = get_irn_mode(b);
2410 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2411 dbg_info *a_dbg = get_irn_dbg_info(n);
2412 ir_node *a_block = get_nodes_block(n);
2414 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2415 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2421 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2422 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2424 get_irn_dbg_info(n),
2429 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2433 if (mode_wrap_around(mode)) {
2434 ir_node *left = get_Add_left(a);
2435 ir_node *right = get_Add_right(a);
2437 /* FIXME: Does the Conv's work only for two complement or generally? */
2439 if (mode != get_irn_mode(right)) {
2440 /* This Sub is an effective Cast */
2441 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2444 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2446 } else if (right == b) {
2447 if (mode != get_irn_mode(left)) {
2448 /* This Sub is an effective Cast */
2449 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2452 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2458 if (mode_wrap_around(mode)) {
2459 ir_node *left = get_Add_left(b);
2460 ir_node *right = get_Add_right(b);
2462 /* FIXME: Does the Conv's work only for two complement or generally? */
2464 ir_mode *r_mode = get_irn_mode(right);
2466 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2467 if (mode != r_mode) {
2468 /* This Sub is an effective Cast */
2469 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2471 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2473 } else if (right == a) {
2474 ir_mode *l_mode = get_irn_mode(left);
2476 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2477 if (mode != l_mode) {
2478 /* This Sub is an effective Cast */
2479 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2481 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2486 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2487 ir_mode *mode = get_irn_mode(a);
2489 if (mode == get_irn_mode(b)) {
2491 ir_node *op_a = get_Conv_op(a);
2492 ir_node *op_b = get_Conv_op(b);
2494 /* check if it's allowed to skip the conv */
2495 ma = get_irn_mode(op_a);
2496 mb = get_irn_mode(op_b);
2498 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2499 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2502 set_Sub_right(n, b);
2508 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2509 if (!is_reassoc_running() && is_Mul(a)) {
2510 ir_node *ma = get_Mul_left(a);
2511 ir_node *mb = get_Mul_right(a);
2514 ir_node *blk = get_nodes_block(n);
2516 get_irn_dbg_info(n),
2517 current_ir_graph, blk,
2520 get_irn_dbg_info(n),
2521 current_ir_graph, blk,
2523 new_r_Const_long(current_ir_graph, blk, mode, 1),
2526 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2528 } else if (mb == b) {
2529 ir_node *blk = get_nodes_block(n);
2531 get_irn_dbg_info(n),
2532 current_ir_graph, blk,
2535 get_irn_dbg_info(n),
2536 current_ir_graph, blk,
2538 new_r_Const_long(current_ir_graph, blk, mode, 1),
2541 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2545 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2546 ir_node *x = get_Sub_left(a);
2547 ir_node *y = get_Sub_right(a);
2548 ir_node *blk = get_nodes_block(n);
2549 ir_mode *m_b = get_irn_mode(b);
2550 ir_mode *m_y = get_irn_mode(y);
2554 /* Determine the right mode for the Add. */
2557 else if (mode_is_reference(m_b))
2559 else if (mode_is_reference(m_y))
2563 * Both modes are different but none is reference,
2564 * happens for instance in SubP(SubP(P, Iu), Is).
2565 * We have two possibilities here: Cast or ignore.
2566 * Currently we ignore this case.
2571 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2573 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2574 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2578 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2579 if (is_Const(a) && is_Not(b)) {
2580 /* c - ~X = X + (c+1) */
2581 tarval *tv = get_Const_tarval(a);
2583 tv = tarval_add(tv, get_mode_one(mode));
2584 if (tv != tarval_bad) {
2585 ir_node *blk = get_nodes_block(n);
2586 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2587 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2588 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2594 } /* transform_node_Sub */
2597 * Several transformation done on n*n=2n bits mul.
2598 * These transformations must be done here because new nodes may be produced.
2600 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2602 ir_node *a = get_Mul_left(n);
2603 ir_node *b = get_Mul_right(n);
2604 tarval *ta = value_of(a);
2605 tarval *tb = value_of(b);
2606 ir_mode *smode = get_irn_mode(a);
2608 if (ta == get_mode_one(smode)) {
2609 /* (L)1 * (L)b = (L)b */
2610 ir_node *blk = get_nodes_block(n);
2611 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2612 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2615 else if (ta == get_mode_minus_one(smode)) {
2616 /* (L)-1 * (L)b = (L)b */
2617 ir_node *blk = get_nodes_block(n);
2618 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2619 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2620 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2623 if (tb == get_mode_one(smode)) {
2624 /* (L)a * (L)1 = (L)a */
2625 ir_node *blk = get_irn_n(a, -1);
2626 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2627 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2630 else if (tb == get_mode_minus_one(smode)) {
2631 /* (L)a * (L)-1 = (L)-a */
2632 ir_node *blk = get_nodes_block(n);
2633 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2634 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2635 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2642 * Transform Mul(a,-1) into -a.
2643 * Do constant evaluation of Phi nodes.
2644 * Do architecture dependent optimizations on Mul nodes
2646 static ir_node *transform_node_Mul(ir_node *n) {
2647 ir_node *c, *oldn = n;
2648 ir_mode *mode = get_irn_mode(n);
2649 ir_node *a = get_Mul_left(n);
2650 ir_node *b = get_Mul_right(n);
2652 if (is_Bad(a) || is_Bad(b))
2655 if (mode != get_irn_mode(a))
2656 return transform_node_Mul2n(n, mode);
2658 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2660 if (mode_is_signed(mode)) {
2663 if (value_of(a) == get_mode_minus_one(mode))
2665 else if (value_of(b) == get_mode_minus_one(mode))
2668 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2669 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2674 if (is_Const(b)) { /* (-a) * const -> a * -const */
2675 ir_node *cnst = const_negate(b);
2677 dbg_info *dbgi = get_irn_dbg_info(n);
2678 ir_node *block = get_nodes_block(n);
2679 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2680 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2683 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2684 dbg_info *dbgi = get_irn_dbg_info(n);
2685 ir_node *block = get_nodes_block(n);
2686 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2687 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2689 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2690 ir_node *sub_l = get_Sub_left(b);
2691 ir_node *sub_r = get_Sub_right(b);
2692 dbg_info *dbgi = get_irn_dbg_info(n);
2693 ir_graph *irg = current_ir_graph;
2694 ir_node *block = get_nodes_block(n);
2695 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2696 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2697 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2700 } else if (is_Minus(b)) {
2701 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2702 ir_node *sub_l = get_Sub_left(a);
2703 ir_node *sub_r = get_Sub_right(a);
2704 dbg_info *dbgi = get_irn_dbg_info(n);
2705 ir_graph *irg = current_ir_graph;
2706 ir_node *block = get_nodes_block(n);
2707 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2708 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2709 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2713 if (get_mode_arithmetic(mode) == irma_ieee754) {
2715 tarval *tv = get_Const_tarval(a);
2716 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2717 && !tarval_is_negative(tv)) {
2718 /* 2.0 * b = b + b */
2719 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2720 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2724 else if (is_Const(b)) {
2725 tarval *tv = get_Const_tarval(b);
2726 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2727 && !tarval_is_negative(tv)) {
2728 /* a * 2.0 = a + a */
2729 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2730 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2735 return arch_dep_replace_mul_with_shifts(n);
2736 } /* transform_node_Mul */
2739 * Transform a Div Node.
2741 static ir_node *transform_node_Div(ir_node *n) {
2742 ir_mode *mode = get_Div_resmode(n);
2743 ir_node *a = get_Div_left(n);
2744 ir_node *b = get_Div_right(n);
2748 if (is_Const(b) && is_const_Phi(a)) {
2749 /* check for Div(Phi, Const) */
2750 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2752 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2756 else if (is_Const(a) && is_const_Phi(b)) {
2757 /* check for Div(Const, Phi) */
2758 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2760 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2764 else if (is_const_Phi(a) && is_const_Phi(b)) {
2765 /* check for Div(Phi, Phi) */
2766 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2768 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2775 if (tv != tarval_bad) {
2776 value = new_Const(get_tarval_mode(tv), tv);
2778 DBG_OPT_CSTEVAL(n, value);
2781 ir_node *a = get_Div_left(n);
2782 ir_node *b = get_Div_right(n);
2783 const ir_node *dummy;
2785 if (a == b && value_not_zero(a, &dummy)) {
2786 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2787 value = new_Const(mode, get_mode_one(mode));
2788 DBG_OPT_CSTEVAL(n, value);
2791 if (mode_is_signed(mode) && is_Const(b)) {
2792 tarval *tv = get_Const_tarval(b);
2794 if (tv == get_mode_minus_one(mode)) {
2796 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2797 DBG_OPT_CSTEVAL(n, value);
2801 /* Try architecture dependent optimization */
2802 value = arch_dep_replace_div_by_const(n);
2810 /* Turn Div into a tuple (mem, jmp, bad, value) */
2811 mem = get_Div_mem(n);
2812 blk = get_nodes_block(n);
2814 /* skip a potential Pin */
2815 mem = skip_Pin(mem);
2816 turn_into_tuple(n, pn_Div_max);
2817 set_Tuple_pred(n, pn_Div_M, mem);
2818 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2819 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2820 set_Tuple_pred(n, pn_Div_res, value);
2823 } /* transform_node_Div */
2826 * Transform a Mod node.
2828 static ir_node *transform_node_Mod(ir_node *n) {
2829 ir_mode *mode = get_Mod_resmode(n);
2830 ir_node *a = get_Mod_left(n);
2831 ir_node *b = get_Mod_right(n);
2835 if (is_Const(b) && is_const_Phi(a)) {
2836 /* check for Div(Phi, Const) */
2837 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2839 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2843 else if (is_Const(a) && is_const_Phi(b)) {
2844 /* check for Div(Const, Phi) */
2845 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2847 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2851 else if (is_const_Phi(a) && is_const_Phi(b)) {
2852 /* check for Div(Phi, Phi) */
2853 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2855 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2862 if (tv != tarval_bad) {
2863 value = new_Const(get_tarval_mode(tv), tv);
2865 DBG_OPT_CSTEVAL(n, value);
2868 ir_node *a = get_Mod_left(n);
2869 ir_node *b = get_Mod_right(n);
2870 const ir_node *dummy;
2872 if (a == b && value_not_zero(a, &dummy)) {
2873 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2874 value = new_Const(mode, get_mode_null(mode));
2875 DBG_OPT_CSTEVAL(n, value);
2878 if (mode_is_signed(mode) && is_Const(b)) {
2879 tarval *tv = get_Const_tarval(b);
2881 if (tv == get_mode_minus_one(mode)) {
2883 value = new_Const(mode, get_mode_null(mode));
2884 DBG_OPT_CSTEVAL(n, value);
2888 /* Try architecture dependent optimization */
2889 value = arch_dep_replace_mod_by_const(n);
2897 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2898 mem = get_Mod_mem(n);
2899 blk = get_nodes_block(n);
2901 /* skip a potential Pin */
2902 mem = skip_Pin(mem);
2903 turn_into_tuple(n, pn_Mod_max);
2904 set_Tuple_pred(n, pn_Mod_M, mem);
2905 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2906 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2907 set_Tuple_pred(n, pn_Mod_res, value);
2910 } /* transform_node_Mod */
2913 * Transform a DivMod node.
2915 static ir_node *transform_node_DivMod(ir_node *n) {
2916 const ir_node *dummy;
2917 ir_node *a = get_DivMod_left(n);
2918 ir_node *b = get_DivMod_right(n);
2919 ir_mode *mode = get_DivMod_resmode(n);
2924 if (is_Const(b) && is_const_Phi(a)) {
2925 /* check for Div(Phi, Const) */
2926 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2927 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2929 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2930 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2934 else if (is_Const(a) && is_const_Phi(b)) {
2935 /* check for Div(Const, Phi) */
2936 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2937 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2939 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2940 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2944 else if (is_const_Phi(a) && is_const_Phi(b)) {
2945 /* check for Div(Phi, Phi) */
2946 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2947 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2949 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2950 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2957 if (tb != tarval_bad) {
2958 if (tb == get_mode_one(get_tarval_mode(tb))) {
2960 vb = new_Const(mode, get_mode_null(mode));
2961 DBG_OPT_CSTEVAL(n, vb);
2963 } else if (ta != tarval_bad) {
2964 tarval *resa, *resb;
2965 resa = tarval_div(ta, tb);
2966 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2967 Jmp for X result!? */
2968 resb = tarval_mod(ta, tb);
2969 if (resb == tarval_bad) return n; /* Causes exception! */
2970 va = new_Const(mode, resa);
2971 vb = new_Const(mode, resb);
2972 DBG_OPT_CSTEVAL(n, va);
2973 DBG_OPT_CSTEVAL(n, vb);
2975 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2976 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2977 vb = new_Const(mode, get_mode_null(mode));
2978 DBG_OPT_CSTEVAL(n, va);
2979 DBG_OPT_CSTEVAL(n, vb);
2981 } else { /* Try architecture dependent optimization */
2984 arch_dep_replace_divmod_by_const(&va, &vb, n);
2985 evaluated = va != NULL;
2987 } else if (a == b) {
2988 if (value_not_zero(a, &dummy)) {
2990 va = new_Const(mode, get_mode_one(mode));
2991 vb = new_Const(mode, get_mode_null(mode));
2992 DBG_OPT_CSTEVAL(n, va);
2993 DBG_OPT_CSTEVAL(n, vb);
2996 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2999 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3000 /* 0 / non-Const = 0 */
3005 if (evaluated) { /* replace by tuple */
3009 mem = get_DivMod_mem(n);
3010 /* skip a potential Pin */
3011 mem = skip_Pin(mem);
3013 blk = get_nodes_block(n);
3014 turn_into_tuple(n, pn_DivMod_max);
3015 set_Tuple_pred(n, pn_DivMod_M, mem);
3016 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3017 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3018 set_Tuple_pred(n, pn_DivMod_res_div, va);
3019 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3023 } /* transform_node_DivMod */
3026 * Optimize x / c to x * (1/c)
3028 static ir_node *transform_node_Quot(ir_node *n) {
3029 ir_mode *mode = get_Quot_resmode(n);
3032 if (get_mode_arithmetic(mode) == irma_ieee754) {
3033 ir_node *b = get_Quot_right(n);
3034 tarval *tv = value_of(b);
3036 if (tv != tarval_bad) {
3040 * Floating point constant folding might be disabled here to
3042 * However, as we check for exact result, doing it is safe.
3045 rem = tarval_enable_fp_ops(1);
3046 tv = tarval_quo(get_mode_one(mode), tv);
3047 (void)tarval_enable_fp_ops(rem);
3049 /* Do the transformation if the result is either exact or we are not
3050 using strict rules. */
3051 if (tv != tarval_bad &&
3052 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3053 ir_node *blk = get_nodes_block(n);
3054 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3055 ir_node *a = get_Quot_left(n);
3056 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3057 ir_node *mem = get_Quot_mem(n);
3059 /* skip a potential Pin */
3060 mem = skip_Pin(mem);
3061 turn_into_tuple(n, pn_Quot_max);
3062 set_Tuple_pred(n, pn_Quot_M, mem);
3063 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3064 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3065 set_Tuple_pred(n, pn_Quot_res, m);
3066 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3071 } /* transform_node_Quot */
3074 * Optimize Abs(x) into x if x is Confirmed >= 0
3075 * Optimize Abs(x) into -x if x is Confirmed <= 0
3076 * Optimize Abs(-x) int Abs(x)
3078 static ir_node *transform_node_Abs(ir_node *n) {
3079 ir_node *c, *oldn = n;
3080 ir_node *a = get_Abs_op(n);
3083 HANDLE_UNOP_PHI(tarval_abs, a, c);
3085 switch (classify_value_sign(a)) {
3086 case value_classified_negative:
3087 mode = get_irn_mode(n);
3090 * We can replace the Abs by -x here.
3091 * We even could add a new Confirm here
3092 * (if not twos complement)
3094 * Note that -x would create a new node, so we could
3095 * not run it in the equivalent_node() context.
3097 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3098 get_nodes_block(n), a, mode);
3100 DBG_OPT_CONFIRM(oldn, n);
3102 case value_classified_positive:
3103 /* n is positive, Abs is not needed */
3106 DBG_OPT_CONFIRM(oldn, n);
3112 /* Abs(-x) = Abs(x) */
3113 mode = get_irn_mode(n);
3114 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3115 get_nodes_block(n), get_Minus_op(a), mode);
3116 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3120 } /* transform_node_Abs */
3123 * Optimize -a CMP -b into b CMP a.
3124 * This works only for for modes where unary Minus
3126 * Note that two-complement integers can Overflow
3127 * so it will NOT work.
3129 * For == and != can be handled in Proj(Cmp)
3131 static ir_node *transform_node_Cmp(ir_node *n) {
3133 ir_node *left = get_Cmp_left(n);
3134 ir_node *right = get_Cmp_right(n);
3136 if (is_Minus(left) && is_Minus(right) &&
3137 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3138 ir_node *const new_left = get_Minus_op(right);
3139 ir_node *const new_right = get_Minus_op(left);
3140 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3141 get_nodes_block(n), new_left, new_right);
3142 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3145 } /* transform_node_Cmp */
3149 * Transform a Cond node.
3151 * Replace the Cond by a Jmp if it branches on a constant
3154 static ir_node *transform_node_Cond(ir_node *n) {
3157 ir_node *a = get_Cond_selector(n);
3158 tarval *ta = value_of(a);
3160 /* we need block info which is not available in floating irgs */
3161 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3164 if ((ta != tarval_bad) &&
3165 (get_irn_mode(a) == mode_b) &&
3166 (get_opt_unreachable_code())) {
3167 /* It's a boolean Cond, branching on a boolean constant.
3168 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3169 ir_node *blk = get_nodes_block(n);
3170 jmp = new_r_Jmp(current_ir_graph, blk);
3171 turn_into_tuple(n, pn_Cond_max);
3172 if (ta == tarval_b_true) {
3173 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3174 set_Tuple_pred(n, pn_Cond_true, jmp);
3176 set_Tuple_pred(n, pn_Cond_false, jmp);
3177 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3179 /* We might generate an endless loop, so keep it alive. */
3180 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3183 } /* transform_node_Cond */
3186 * Prototype of a recursive transform function
3187 * for bitwise distributive transformations.
3189 typedef ir_node* (*recursive_transform)(ir_node *n);
3192 * makes use of distributive laws for and, or, eor
3193 * and(a OP c, b OP c) -> and(a, b) OP c
3194 * note, might return a different op than n
3196 static ir_node *transform_bitwise_distributive(ir_node *n,
3197 recursive_transform trans_func)
3200 ir_node *a = get_binop_left(n);
3201 ir_node *b = get_binop_right(n);
3202 ir_op *op = get_irn_op(a);
3203 ir_op *op_root = get_irn_op(n);
3205 if(op != get_irn_op(b))
3208 if (op == op_Conv) {
3209 ir_node *a_op = get_Conv_op(a);
3210 ir_node *b_op = get_Conv_op(b);
3211 ir_mode *a_mode = get_irn_mode(a_op);
3212 ir_mode *b_mode = get_irn_mode(b_op);
3213 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3214 ir_node *blk = get_nodes_block(n);
3217 set_binop_left(n, a_op);
3218 set_binop_right(n, b_op);
3219 set_irn_mode(n, a_mode);
3221 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3223 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3229 /* nothing to gain here */
3233 if (op == op_Shrs || op == op_Shr || op == op_Shl
3234 || op == op_And || op == op_Or || op == op_Eor) {
3235 ir_node *a_left = get_binop_left(a);
3236 ir_node *a_right = get_binop_right(a);
3237 ir_node *b_left = get_binop_left(b);
3238 ir_node *b_right = get_binop_right(b);
3240 ir_node *op1 = NULL;
3241 ir_node *op2 = NULL;
3243 if (is_op_commutative(op)) {
3244 if (a_left == b_left) {
3248 } else if(a_left == b_right) {
3252 } else if(a_right == b_left) {
3258 if(a_right == b_right) {
3265 /* (a sop c) & (b sop c) => (a & b) sop c */
3266 ir_node *blk = get_nodes_block(n);
3268 ir_node *new_n = exact_copy(n);
3269 set_binop_left(new_n, op1);
3270 set_binop_right(new_n, op2);
3271 new_n = trans_func(new_n);
3273 if(op_root == op_Eor && op == op_Or) {
3274 dbg_info *dbgi = get_irn_dbg_info(n);
3275 ir_graph *irg = current_ir_graph;
3276 ir_mode *mode = get_irn_mode(c);
3278 c = new_rd_Not(dbgi, irg, blk, c, mode);
3279 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3282 set_nodes_block(n, blk);
3283 set_binop_left(n, new_n);
3284 set_binop_right(n, c);
3285 add_identities(current_ir_graph->value_table, n);
3288 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3299 static ir_node *transform_node_And(ir_node *n) {
3300 ir_node *c, *oldn = n;
3301 ir_node *a = get_And_left(n);
3302 ir_node *b = get_And_right(n);
3305 mode = get_irn_mode(n);
3306 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3308 /* we can evaluate 2 Projs of the same Cmp */
3309 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3310 ir_node *pred_a = get_Proj_pred(a);
3311 ir_node *pred_b = get_Proj_pred(b);
3312 if (pred_a == pred_b) {
3313 dbg_info *dbgi = get_irn_dbg_info(n);
3314 ir_node *block = get_nodes_block(pred_a);
3315 pn_Cmp pn_a = get_Proj_proj(a);
3316 pn_Cmp pn_b = get_Proj_proj(b);
3317 /* yes, we can simply calculate with pncs */
3318 pn_Cmp new_pnc = pn_a & pn_b;
3320 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3325 ir_node *op = get_Not_op(b);
3327 ir_node *ba = get_And_left(op);
3328 ir_node *bb = get_And_right(op);
3330 /* it's enough to test the following cases due to normalization! */
3331 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3332 /* (a|b) & ~(a&b) = a^b */
3333 ir_node *block = get_nodes_block(n);
3335 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3336 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3344 ir_node *op = get_Not_op(a);
3346 ir_node *aa = get_And_left(op);
3347 ir_node *ab = get_And_right(op);
3349 /* it's enough to test the following cases due to normalization! */
3350 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3351 /* (a|b) & ~(a&b) = a^b */
3352 ir_node *block = get_nodes_block(n);
3354 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3355 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3362 ir_node *al = get_Eor_left(a);
3363 ir_node *ar = get_Eor_right(a);
3366 /* (b ^ a) & b -> ~a & b */
3367 dbg_info *dbg = get_irn_dbg_info(n);
3368 ir_node *block = get_nodes_block(n);
3370 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3371 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3372 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3376 /* (a ^ b) & b -> ~a & b */
3377 dbg_info *dbg = get_irn_dbg_info(n);
3378 ir_node *block = get_nodes_block(n);
3380 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3381 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3382 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3387 ir_node *bl = get_Eor_left(b);
3388 ir_node *br = get_Eor_right(b);
3391 /* a & (a ^ b) -> a & ~b */
3392 dbg_info *dbg = get_irn_dbg_info(n);
3393 ir_node *block = get_nodes_block(n);
3395 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3396 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3397 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3401 /* a & (b ^ a) -> a & ~b */
3402 dbg_info *dbg = get_irn_dbg_info(n);
3403 ir_node *block = get_nodes_block(n);
3405 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3406 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3407 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3411 if (is_Not(a) && is_Not(b)) {
3412 /* ~a & ~b = ~(a|b) */
3413 ir_node *block = get_nodes_block(n);
3414 ir_mode *mode = get_irn_mode(n);
3418 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3419 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3420 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3424 n = transform_bitwise_distributive(n, transform_node_And);
3427 } /* transform_node_And */
3432 static ir_node *transform_node_Eor(ir_node *n) {
3433 ir_node *c, *oldn = n;
3434 ir_node *a = get_Eor_left(n);
3435 ir_node *b = get_Eor_right(n);
3436 ir_mode *mode = get_irn_mode(n);
3438 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3440 /* we can evaluate 2 Projs of the same Cmp */
3441 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3442 ir_node *pred_a = get_Proj_pred(a);
3443 ir_node *pred_b = get_Proj_pred(b);
3444 if(pred_a == pred_b) {
3445 dbg_info *dbgi = get_irn_dbg_info(n);
3446 ir_node *block = get_nodes_block(pred_a);
3447 pn_Cmp pn_a = get_Proj_proj(a);
3448 pn_Cmp pn_b = get_Proj_proj(b);
3449 /* yes, we can simply calculate with pncs */
3450 pn_Cmp new_pnc = pn_a ^ pn_b;
3452 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3459 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
3460 mode, get_mode_null(mode));
3461 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3462 } else if (mode == mode_b &&
3464 is_Const(b) && is_Const_one(b) &&
3465 is_Cmp(get_Proj_pred(a))) {
3466 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3467 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3468 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3471 } else if (is_Const(b)) {
3472 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3473 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3474 ir_node *not_op = get_Not_op(a);
3475 dbg_info *dbg = get_irn_dbg_info(n);
3476 ir_graph *irg = current_ir_graph;
3477 ir_node *block = get_nodes_block(n);
3478 ir_mode *mode = get_irn_mode(n);
3479 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3481 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3482 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3483 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3486 n = transform_bitwise_distributive(n, transform_node_Eor);
3490 } /* transform_node_Eor */
3495 static ir_node *transform_node_Not(ir_node *n) {
3496 ir_node *c, *oldn = n;
3497 ir_node *a = get_Not_op(n);
3498 ir_mode *mode = get_irn_mode(n);
3500 HANDLE_UNOP_PHI(tarval_not,a,c);
3502 /* check for a boolean Not */
3503 if (mode == mode_b &&
3505 is_Cmp(get_Proj_pred(a))) {
3506 /* We negate a Cmp. The Cmp has the negated result anyways! */
3507 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3508 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3509 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3513 ir_node *eor_b = get_Eor_right(a);
3514 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3515 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3516 ir_node *eor_a = get_Eor_left(a);
3517 dbg_info *dbg = get_irn_dbg_info(n);
3518 ir_graph *irg = current_ir_graph;
3519 ir_node *block = get_nodes_block(n);
3520 ir_mode *mode = get_irn_mode(n);
3521 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3525 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3526 if (is_Minus(a)) { /* ~-x -> x + -1 */
3527 dbg_info *dbg = get_irn_dbg_info(n);
3528 ir_graph *irg = current_ir_graph;
3529 ir_node *block = get_nodes_block(n);
3530 ir_node *add_l = get_Minus_op(a);
3531 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3532 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3533 } else if (is_Add(a)) {
3534 ir_node *add_r = get_Add_right(a);
3535 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3536 /* ~(x + -1) = -x */
3537 ir_node *op = get_Add_left(a);
3538 ir_node *blk = get_nodes_block(n);
3539 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3540 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3545 } /* transform_node_Not */
3548 * Transform a Minus.
3552 * -(a >>u (size-1)) = a >>s (size-1)
3553 * -(a >>s (size-1)) = a >>u (size-1)
3554 * -(a * const) -> a * -const
3556 static ir_node *transform_node_Minus(ir_node *n) {
3557 ir_node *c, *oldn = n;
3558 ir_node *a = get_Minus_op(n);
3561 HANDLE_UNOP_PHI(tarval_neg,a,c);
3563 mode = get_irn_mode(a);
3564 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3565 /* the following rules are only to twos-complement */
3568 ir_node *op = get_Not_op(a);
3569 tarval *tv = get_mode_one(mode);
3570 ir_node *blk = get_nodes_block(n);
3571 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3572 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3573 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3577 ir_node *c = get_Shr_right(a);
3580 tarval *tv = get_Const_tarval(c);
3582 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3583 /* -(a >>u (size-1)) = a >>s (size-1) */
3584 ir_node *v = get_Shr_left(a);
3586 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3587 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3593 ir_node *c = get_Shrs_right(a);
3596 tarval *tv = get_Const_tarval(c);
3598 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3599 /* -(a >>s (size-1)) = a >>u (size-1) */
3600 ir_node *v = get_Shrs_left(a);
3602 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3603 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3610 /* - (a-b) = b - a */
3611 ir_node *la = get_Sub_left(a);
3612 ir_node *ra = get_Sub_right(a);
3613 ir_node *blk = get_nodes_block(n);
3615 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3616 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3620 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3621 ir_node *mul_l = get_Mul_left(a);
3622 ir_node *mul_r = get_Mul_right(a);
3623 tarval *tv = value_of(mul_r);
3624 if (tv != tarval_bad) {
3625 tv = tarval_neg(tv);
3626 if (tv != tarval_bad) {
3627 ir_node *cnst = new_Const(mode, tv);
3628 dbg_info *dbg = get_irn_dbg_info(a);
3629 ir_graph *irg = current_ir_graph;
3630 ir_node *block = get_nodes_block(a);
3631 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3632 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3639 } /* transform_node_Minus */
3642 * Transform a Cast_type(Const) into a new Const_type
3644 static ir_node *transform_node_Cast(ir_node *n) {
3646 ir_node *pred = get_Cast_op(n);
3647 ir_type *tp = get_irn_type(n);
3649 if (is_Const(pred) && get_Const_type(pred) != tp) {
3650 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3651 get_Const_tarval(pred), tp);
3652 DBG_OPT_CSTEVAL(oldn, n);
3653 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3654 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3655 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3656 DBG_OPT_CSTEVAL(oldn, n);
3660 } /* transform_node_Cast */
3663 * Transform a Proj(Load) with a non-null address.
3665 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3666 if (get_opt_ldst_only_null_ptr_exceptions()) {
3667 if (get_irn_mode(proj) == mode_X) {
3668 ir_node *load = get_Proj_pred(proj);
3670 /* get the Load address */
3671 const ir_node *addr = get_Load_ptr(load);
3672 const ir_node *confirm;
3674 if (value_not_null(addr, &confirm)) {
3675 if (confirm == NULL) {
3676 /* this node may float if it did not depend on a Confirm */
3677 set_irn_pinned(load, op_pin_state_floats);
3679 if (get_Proj_proj(proj) == pn_Load_X_except) {
3680 DBG_OPT_EXC_REM(proj);
3681 return get_irg_bad(current_ir_graph);
3683 ir_node *blk = get_nodes_block(load);
3684 return new_r_Jmp(current_ir_graph, blk);
3690 } /* transform_node_Proj_Load */
3693 * Transform a Proj(Store) with a non-null address.
3695 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3696 if (get_opt_ldst_only_null_ptr_exceptions()) {
3697 if (get_irn_mode(proj) == mode_X) {
3698 ir_node *store = get_Proj_pred(proj);
3700 /* get the load/store address */
3701 const ir_node *addr = get_Store_ptr(store);
3702 const ir_node *confirm;
3704 if (value_not_null(addr, &confirm)) {
3705 if (confirm == NULL) {
3706 /* this node may float if it did not depend on a Confirm */
3707 set_irn_pinned(store, op_pin_state_floats);
3709 if (get_Proj_proj(proj) == pn_Store_X_except) {
3710 DBG_OPT_EXC_REM(proj);
3711 return get_irg_bad(current_ir_graph);
3713 ir_node *blk = get_nodes_block(store);
3714 return new_r_Jmp(current_ir_graph, blk);
3720 } /* transform_node_Proj_Store */
3723 * Transform a Proj(Div) with a non-zero value.
3724 * Removes the exceptions and routes the memory to the NoMem node.
3726 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3727 ir_node *div = get_Proj_pred(proj);
3728 ir_node *b = get_Div_right(div);
3729 ir_node *res, *new_mem;
3730 const ir_node *confirm;
3733 if (value_not_zero(b, &confirm)) {
3734 /* div(x, y) && y != 0 */
3735 if (confirm == NULL) {
3736 /* we are sure we have a Const != 0 */
3737 new_mem = get_Div_mem(div);
3738 new_mem = skip_Pin(new_mem);
3739 set_Div_mem(div, new_mem);
3740 set_irn_pinned(div, op_pin_state_floats);
3743 proj_nr = get_Proj_proj(proj);
3745 case pn_Div_X_regular:
3746 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3748 case pn_Div_X_except:
3749 /* we found an exception handler, remove it */
3750 DBG_OPT_EXC_REM(proj);
3754 res = get_Div_mem(div);
3755 new_mem = get_irg_no_mem(current_ir_graph);
3758 /* This node can only float up to the Confirm block */
3759 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3761 set_irn_pinned(div, op_pin_state_floats);
3762 /* this is a Div without exception, we can remove the memory edge */
3763 set_Div_mem(div, new_mem);
3768 } /* transform_node_Proj_Div */
3771 * Transform a Proj(Mod) with a non-zero value.
3772 * Removes the exceptions and routes the memory to the NoMem node.
3774 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3775 ir_node *mod = get_Proj_pred(proj);
3776 ir_node *b = get_Mod_right(mod);
3777 ir_node *res, *new_mem;
3778 const ir_node *confirm;
3781 if (value_not_zero(b, &confirm)) {
3782 /* mod(x, y) && y != 0 */
3783 proj_nr = get_Proj_proj(proj);
3785 if (confirm == NULL) {
3786 /* we are sure we have a Const != 0 */
3787 new_mem = get_Mod_mem(mod);
3788 new_mem = skip_Pin(new_mem);
3789 set_Mod_mem(mod, new_mem);
3790 set_irn_pinned(mod, op_pin_state_floats);
3795 case pn_Mod_X_regular:
3796 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3798 case pn_Mod_X_except:
3799 /* we found an exception handler, remove it */
3800 DBG_OPT_EXC_REM(proj);
3804 res = get_Mod_mem(mod);
3805 new_mem = get_irg_no_mem(current_ir_graph);
3808 /* This node can only float up to the Confirm block */
3809 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3811 /* this is a Mod without exception, we can remove the memory edge */
3812 set_Mod_mem(mod, new_mem);
3815 if (get_Mod_left(mod) == b) {
3816 /* a % a = 0 if a != 0 */
3817 ir_mode *mode = get_irn_mode(proj);
3818 ir_node *res = new_Const(mode, get_mode_null(mode));
3820 DBG_OPT_CSTEVAL(mod, res);
3826 } /* transform_node_Proj_Mod */
3829 * Transform a Proj(DivMod) with a non-zero value.
3830 * Removes the exceptions and routes the memory to the NoMem node.
3832 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3833 ir_node *divmod = get_Proj_pred(proj);
3834 ir_node *b = get_DivMod_right(divmod);
3835 ir_node *res, *new_mem;
3836 const ir_node *confirm;
3839 if (value_not_zero(b, &confirm)) {
3840 /* DivMod(x, y) && y != 0 */
3841 proj_nr = get_Proj_proj(proj);
3843 if (confirm == NULL) {
3844 /* we are sure we have a Const != 0 */
3845 new_mem = get_DivMod_mem(divmod);
3846 new_mem = skip_Pin(new_mem);
3847 set_DivMod_mem(divmod, new_mem);
3848 set_irn_pinned(divmod, op_pin_state_floats);
3853 case pn_DivMod_X_regular:
3854 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3856 case pn_DivMod_X_except:
3857 /* we found an exception handler, remove it */
3858 DBG_OPT_EXC_REM(proj);
3862 res = get_DivMod_mem(divmod);
3863 new_mem = get_irg_no_mem(current_ir_graph);
3866 /* This node can only float up to the Confirm block */
3867 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3869 /* this is a DivMod without exception, we can remove the memory edge */
3870 set_DivMod_mem(divmod, new_mem);
3873 case pn_DivMod_res_mod:
3874 if (get_DivMod_left(divmod) == b) {
3875 /* a % a = 0 if a != 0 */
3876 ir_mode *mode = get_irn_mode(proj);
3877 ir_node *res = new_Const(mode, get_mode_null(mode));
3879 DBG_OPT_CSTEVAL(divmod, res);
3885 } /* transform_node_Proj_DivMod */
3888 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3890 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3891 if (get_opt_unreachable_code()) {
3892 ir_node *n = get_Proj_pred(proj);
3893 ir_node *b = get_Cond_selector(n);
3895 if (mode_is_int(get_irn_mode(b))) {
3896 tarval *tb = value_of(b);
3898 if (tb != tarval_bad) {
3899 /* we have a constant switch */
3900 long num = get_Proj_proj(proj);
3902 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3903 if (get_tarval_long(tb) == num) {
3904 /* Do NOT create a jump here, or we will have 2 control flow ops
3905 * in a block. This case is optimized away in optimize_cf(). */
3908 /* this case will NEVER be taken, kill it */
3909 return get_irg_bad(current_ir_graph);
3916 } /* transform_node_Proj_Cond */
3919 * Create a 0 constant of given mode.
3921 static ir_node *create_zero_const(ir_mode *mode) {
3922 tarval *tv = get_mode_null(mode);
3923 ir_node *cnst = new_Const(mode, tv);
3928 /* the order of the values is important! */
3929 typedef enum const_class {
3935 static const_class classify_const(const ir_node* n)
3937 if (is_Const(n)) return const_const;
3938 if (is_irn_constlike(n)) return const_like;
3943 * Determines whether r is more constlike or has a larger index (in that order)
3946 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3948 const const_class l_order = classify_const(l);
3949 const const_class r_order = classify_const(r);
3951 l_order > r_order ||
3952 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3956 * Normalizes and optimizes Cmp nodes.
3958 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3959 ir_node *n = get_Proj_pred(proj);
3960 ir_node *left = get_Cmp_left(n);
3961 ir_node *right = get_Cmp_right(n);
3964 ir_mode *mode = NULL;
3965 long proj_nr = get_Proj_proj(proj);
3967 /* we can evaluate some cases directly */
3970 return new_Const(mode_b, get_tarval_b_false());
3972 return new_Const(mode_b, get_tarval_b_true());
3974 if (!mode_is_float(get_irn_mode(left)))
3975 return new_Const(mode_b, get_tarval_b_true());
3981 /* remove Casts of both sides */
3982 left = skip_Cast(left);
3983 right = skip_Cast(right);
3985 /* Remove unnecessary conversions */
3986 /* TODO handle constants */
3987 if (is_Conv(left) && is_Conv(right)) {
3988 ir_mode *mode = get_irn_mode(left);
3989 ir_node *op_left = get_Conv_op(left);
3990 ir_node *op_right = get_Conv_op(right);
3991 ir_mode *mode_left = get_irn_mode(op_left);
3992 ir_mode *mode_right = get_irn_mode(op_right);
3994 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3995 && mode_left != mode_b && mode_right != mode_b) {
3996 ir_graph *irg = current_ir_graph;
3997 ir_node *block = get_nodes_block(n);
3999 if (mode_left == mode_right) {
4003 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4004 } else if (smaller_mode(mode_left, mode_right)) {
4005 left = new_r_Conv(irg, block, op_left, mode_right);
4008 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4009 } else if (smaller_mode(mode_right, mode_left)) {
4011 right = new_r_Conv(irg, block, op_right, mode_left);
4013 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4018 /* remove operation on both sides if possible */
4019 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4021 * The following operations are NOT safe for floating point operations, for instance
4022 * 1.0 + inf == 2.0 + inf, =/=> x == y
4024 if (mode_is_int(get_irn_mode(left))) {
4025 unsigned lop = get_irn_opcode(left);
4027 if (lop == get_irn_opcode(right)) {
4028 ir_node *ll, *lr, *rl, *rr;
4030 /* same operation on both sides, try to remove */
4034 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4035 left = get_unop_op(left);
4036 right = get_unop_op(right);
4038 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4041 ll = get_Add_left(left);
4042 lr = get_Add_right(left);
4043 rl = get_Add_left(right);
4044 rr = get_Add_right(right);
4047 /* X + a CMP X + b ==> a CMP b */
4051 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4052 } else if (ll == rr) {
4053 /* X + a CMP b + X ==> a CMP b */
4057 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4058 } else if (lr == rl) {
4059 /* a + X CMP X + b ==> a CMP b */
4063 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4064 } else if (lr == rr) {
4065 /* a + X CMP b + X ==> a CMP b */
4069 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4073 ll = get_Sub_left(left);
4074 lr = get_Sub_right(left);
4075 rl = get_Sub_left(right);
4076 rr = get_Sub_right(right);
4079 /* X - a CMP X - b ==> a CMP b */
4083 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4084 } else if (lr == rr) {
4085 /* a - X CMP b - X ==> a CMP b */
4089 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4093 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4094 /* a ROTL X CMP b ROTL X ==> a CMP b */
4095 left = get_Rotl_left(left);
4096 right = get_Rotl_left(right);
4098 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4106 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4107 if (is_Add(left) || is_Sub(left)) {
4108 ir_node *ll = get_binop_left(left);
4109 ir_node *lr = get_binop_right(left);
4111 if (lr == right && is_Add(left)) {
4118 right = create_zero_const(get_irn_mode(left));
4120 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4123 if (is_Add(right) || is_Sub(right)) {
4124 ir_node *rl = get_binop_left(right);
4125 ir_node *rr = get_binop_right(right);
4127 if (rr == left && is_Add(right)) {
4134 right = create_zero_const(get_irn_mode(left));
4136 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4139 } /* mode_is_int(...) */
4140 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4142 /* replace mode_b compares with ands/ors */
4143 if (get_irn_mode(left) == mode_b) {
4144 ir_graph *irg = current_ir_graph;
4145 ir_node *block = get_nodes_block(n);
4149 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4150 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4151 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4152 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4153 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4154 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4155 default: bres = NULL;
4158 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4164 * First step: normalize the compare op
4165 * by placing the constant on the right side
4166 * or moving the lower address node to the left.
4168 if (!operands_are_normalized(left, right)) {
4174 proj_nr = get_inversed_pnc(proj_nr);
4179 * Second step: Try to reduce the magnitude
4180 * of a constant. This may help to generate better code
4181 * later and may help to normalize more compares.
4182 * Of course this is only possible for integer values.
4184 tv = value_of(right);
4185 if (tv != tarval_bad) {
4186 mode = get_irn_mode(right);
4188 /* TODO extend to arbitrary constants */
4189 if (is_Conv(left) && tarval_is_null(tv)) {
4190 ir_node *op = get_Conv_op(left);
4191 ir_mode *op_mode = get_irn_mode(op);
4194 * UpConv(x) REL 0 ==> x REL 0
4196 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4197 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4198 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4199 tv = get_mode_null(op_mode);
4203 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4207 if (tv != tarval_bad) {
4208 /* the following optimization is possible on modes without Overflow
4209 * on Unary Minus or on == and !=:
4210 * -a CMP c ==> a swap(CMP) -c
4212 * Beware: for two-complement Overflow may occur, so only == and != can
4213 * be optimized, see this:
4214 * -MININT < 0 =/=> MININT > 0 !!!
4216 if (is_Minus(left) &&
4217 (!mode_overflow_on_unary_Minus(mode) ||
4218 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4219 tv = tarval_neg(tv);
4221 if (tv != tarval_bad) {
4222 left = get_Minus_op(left);
4223 proj_nr = get_inversed_pnc(proj_nr);
4225 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4227 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4228 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4229 tv = tarval_not(tv);
4231 if (tv != tarval_bad) {
4232 left = get_Not_op(left);
4234 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4238 /* for integer modes, we have more */
4239 if (mode_is_int(mode)) {
4240 /* Ne includes Unordered which is not possible on integers.
4241 * However, frontends often use this wrong, so fix it here */
4242 if (proj_nr & pn_Cmp_Uo) {
4243 proj_nr &= ~pn_Cmp_Uo;
4244 set_Proj_proj(proj, proj_nr);
4247 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4248 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4249 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4250 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4252 if (tv != tarval_bad) {
4253 proj_nr ^= pn_Cmp_Eq;
4255 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4258 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4259 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4260 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4261 tv = tarval_add(tv, get_mode_one(mode));
4263 if (tv != tarval_bad) {
4264 proj_nr ^= pn_Cmp_Eq;
4266 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4270 /* the following reassociations work only for == and != */
4271 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4273 #if 0 /* Might be not that good in general */
4274 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4275 if (tarval_is_null(tv) && is_Sub(left)) {
4276 right = get_Sub_right(left);
4277 left = get_Sub_left(left);
4279 tv = value_of(right);
4281 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4285 if (tv != tarval_bad) {
4286 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4288 ir_node *c1 = get_Sub_right(left);
4289 tarval *tv2 = value_of(c1);
4291 if (tv2 != tarval_bad) {
4292 tv2 = tarval_add(tv, value_of(c1));
4294 if (tv2 != tarval_bad) {
4295 left = get_Sub_left(left);
4298 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4302 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4303 else if (is_Add(left)) {
4304 ir_node *a_l = get_Add_left(left);
4305 ir_node *a_r = get_Add_right(left);
4309 if (is_Const(a_l)) {
4311 tv2 = value_of(a_l);
4314 tv2 = value_of(a_r);
4317 if (tv2 != tarval_bad) {
4318 tv2 = tarval_sub(tv, tv2, NULL);
4320 if (tv2 != tarval_bad) {
4324 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4328 /* -a == c ==> a == -c, -a != c ==> a != -c */
4329 else if (is_Minus(left)) {
4330 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4332 if (tv2 != tarval_bad) {
4333 left = get_Minus_op(left);
4336 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4341 /* the following reassociations work only for <= */
4342 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4343 if (tv != tarval_bad) {
4344 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4345 if (is_Abs(left)) { // TODO something is missing here
4351 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4352 switch (get_irn_opcode(left)) {
4356 c1 = get_And_right(left);
4359 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4360 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4362 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4364 /* TODO: move to constant evaluation */
4365 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4366 c1 = new_Const(mode_b, tv);
4367 DBG_OPT_CSTEVAL(proj, c1);
4371 if (tarval_is_single_bit(tv)) {
4373 * optimization for AND:
4375 * And(x, C) == C ==> And(x, C) != 0
4376 * And(x, C) != C ==> And(X, C) == 0
4378 * if C is a single Bit constant.
4381 /* check for Constant's match. We have check hare the tarvals,
4382 because our const might be changed */
4383 if (get_Const_tarval(c1) == tv) {
4384 /* fine: do the transformation */
4385 tv = get_mode_null(get_tarval_mode(tv));
4386 proj_nr ^= pn_Cmp_Leg;
4388 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4394 c1 = get_Or_right(left);
4395 if (is_Const(c1) && tarval_is_null(tv)) {
4397 * Or(x, C) == 0 && C != 0 ==> FALSE
4398 * Or(x, C) != 0 && C != 0 ==> TRUE
4400 if (! tarval_is_null(get_Const_tarval(c1))) {
4401 /* TODO: move to constant evaluation */
4402 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4403 c1 = new_Const(mode_b, tv);
4404 DBG_OPT_CSTEVAL(proj, c1);
4411 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4413 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4416 c1 = get_Shl_right(left);
4418 tarval *tv1 = get_Const_tarval(c1);
4419 ir_mode *mode = get_irn_mode(left);
4420 tarval *minus1 = get_mode_all_one(mode);
4421 tarval *amask = tarval_shr(minus1, tv1);
4422 tarval *cmask = tarval_shl(minus1, tv1);
4425 if (tarval_and(tv, cmask) != tv) {
4426 /* condition not met */
4427 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4428 c1 = new_Const(mode_b, tv);
4429 DBG_OPT_CSTEVAL(proj, c1);
4432 sl = get_Shl_left(left);
4433 blk = get_nodes_block(n);
4434 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4435 tv = tarval_shr(tv, tv1);
4437 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4442 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4444 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4447 c1 = get_Shr_right(left);
4449 tarval *tv1 = get_Const_tarval(c1);
4450 ir_mode *mode = get_irn_mode(left);
4451 tarval *minus1 = get_mode_all_one(mode);
4452 tarval *amask = tarval_shl(minus1, tv1);
4453 tarval *cmask = tarval_shr(minus1, tv1);
4456 if (tarval_and(tv, cmask) != tv) {
4457 /* condition not met */
4458 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4459 c1 = new_Const(mode_b, tv);
4460 DBG_OPT_CSTEVAL(proj, c1);
4463 sl = get_Shr_left(left);
4464 blk = get_nodes_block(n);
4465 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4466 tv = tarval_shl(tv, tv1);
4468 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4473 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4475 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4478 c1 = get_Shrs_right(left);
4480 tarval *tv1 = get_Const_tarval(c1);
4481 ir_mode *mode = get_irn_mode(left);
4482 tarval *minus1 = get_mode_all_one(mode);
4483 tarval *amask = tarval_shl(minus1, tv1);
4484 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4487 cond = tarval_sub(cond, tv1, NULL);
4488 cond = tarval_shrs(tv, cond);
4490 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4491 /* condition not met */
4492 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4493 c1 = new_Const(mode_b, tv);
4494 DBG_OPT_CSTEVAL(proj, c1);
4497 sl = get_Shrs_left(left);
4498 blk = get_nodes_block(n);
4499 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4500 tv = tarval_shl(tv, tv1);
4502 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4507 } /* tarval != bad */
4510 if (changed & 2) /* need a new Const */
4511 right = new_Const(mode, tv);
4513 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4514 ir_node *op = get_Proj_pred(left);
4516 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4517 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4518 ir_node *c = get_binop_right(op);
4521 tarval *tv = get_Const_tarval(c);
4523 if (tarval_is_single_bit(tv)) {
4524 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4525 ir_node *v = get_binop_left(op);
4526 ir_node *blk = get_irn_n(op, -1);
4527 ir_mode *mode = get_irn_mode(v);
4529 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4530 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4532 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4539 ir_node *block = get_nodes_block(n);
4541 /* create a new compare */
4542 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4543 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4547 } /* transform_node_Proj_Cmp */
4550 * Optimize CopyB(mem, x, x) into a Nop.
4552 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4553 ir_node *copyb = get_Proj_pred(proj);
4554 ir_node *a = get_CopyB_dst(copyb);
4555 ir_node *b = get_CopyB_src(copyb);
4558 switch (get_Proj_proj(proj)) {
4559 case pn_CopyB_X_regular:
4560 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4561 DBG_OPT_EXC_REM(proj);
4562 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4564 case pn_CopyB_M_except:
4565 case pn_CopyB_X_except:
4566 DBG_OPT_EXC_REM(proj);
4567 proj = get_irg_bad(current_ir_graph);
4574 } /* transform_node_Proj_CopyB */
4577 * Optimize Bounds(idx, idx, upper) into idx.
4579 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4580 ir_node *oldn = proj;
4581 ir_node *bound = get_Proj_pred(proj);
4582 ir_node *idx = get_Bound_index(bound);
4583 ir_node *pred = skip_Proj(idx);
4586 if (idx == get_Bound_lower(bound))
4588 else if (is_Bound(pred)) {
4590 * idx was Bounds checked in the same MacroBlock previously,
4591 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4593 ir_node *lower = get_Bound_lower(bound);
4594 ir_node *upper = get_Bound_upper(bound);
4595 if (get_Bound_lower(pred) == lower &&
4596 get_Bound_upper(pred) == upper &&
4597 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4599 * One could expect that we simply return the previous
4600 * Bound here. However, this would be wrong, as we could
4601 * add an exception Proj to a new location then.
4602 * So, we must turn in into a tuple.
4608 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4609 switch (get_Proj_proj(proj)) {
4611 DBG_OPT_EXC_REM(proj);
4612 proj = get_Bound_mem(bound);
4614 case pn_Bound_X_except:
4615 DBG_OPT_EXC_REM(proj);
4616 proj = get_irg_bad(current_ir_graph);
4620 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4622 case pn_Bound_X_regular:
4623 DBG_OPT_EXC_REM(proj);
4624 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4631 } /* transform_node_Proj_Bound */
4634 * Does all optimizations on nodes that must be done on it's Proj's
4635 * because of creating new nodes.
4637 static ir_node *transform_node_Proj(ir_node *proj) {
4638 ir_node *n = get_Proj_pred(proj);
4640 if (n->op->ops.transform_node_Proj)
4641 return n->op->ops.transform_node_Proj(proj);
4643 } /* transform_node_Proj */
4646 * Move Confirms down through Phi nodes.
4648 static ir_node *transform_node_Phi(ir_node *phi) {
4650 ir_mode *mode = get_irn_mode(phi);
4652 if (mode_is_reference(mode)) {
4653 n = get_irn_arity(phi);
4655 /* Beware of Phi0 */
4657 ir_node *pred = get_irn_n(phi, 0);
4658 ir_node *bound, *new_Phi, *block, **in;
4661 if (! is_Confirm(pred))
4664 bound = get_Confirm_bound(pred);
4665 pnc = get_Confirm_cmp(pred);
4667 NEW_ARR_A(ir_node *, in, n);
4668 in[0] = get_Confirm_value(pred);
4670 for (i = 1; i < n; ++i) {
4671 pred = get_irn_n(phi, i);
4673 if (! is_Confirm(pred) ||
4674 get_Confirm_bound(pred) != bound ||
4675 get_Confirm_cmp(pred) != pnc)
4677 in[i] = get_Confirm_value(pred);
4679 /* move the Confirm nodes "behind" the Phi */
4680 block = get_irn_n(phi, -1);
4681 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4682 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4686 } /* transform_node_Phi */
4689 * Returns the operands of a commutative bin-op, if one operand is
4690 * a const, it is returned as the second one.
4692 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4693 ir_node *op_a = get_binop_left(binop);
4694 ir_node *op_b = get_binop_right(binop);
4696 assert(is_op_commutative(get_irn_op(binop)));
4698 if (is_Const(op_a)) {
4705 } /* get_comm_Binop_Ops */
4708 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4709 * Such pattern may arise in bitfield stores.
4711 * value c4 value c4 & c2
4712 * AND c3 AND c1 | c3
4719 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4722 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4725 ir_node *and_l, *c3;
4726 ir_node *value, *c4;
4727 ir_node *new_and, *new_const, *block;
4728 ir_mode *mode = get_irn_mode(or);
4730 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4733 get_comm_Binop_Ops(or, &and, &c1);
4734 if (!is_Const(c1) || !is_And(and))
4737 get_comm_Binop_Ops(and, &or_l, &c2);
4741 tv1 = get_Const_tarval(c1);
4742 tv2 = get_Const_tarval(c2);
4744 tv = tarval_or(tv1, tv2);
4745 if (tarval_is_all_one(tv)) {
4746 /* the AND does NOT clear a bit with isn't set by the OR */
4747 set_Or_left(or, or_l);
4748 set_Or_right(or, c1);
4750 /* check for more */
4757 get_comm_Binop_Ops(or_l, &and_l, &c3);
4758 if (!is_Const(c3) || !is_And(and_l))
4761 get_comm_Binop_Ops(and_l, &value, &c4);
4765 /* ok, found the pattern, check for conditions */
4766 assert(mode == get_irn_mode(and));
4767 assert(mode == get_irn_mode(or_l));
4768 assert(mode == get_irn_mode(and_l));
4770 tv3 = get_Const_tarval(c3);
4771 tv4 = get_Const_tarval(c4);
4773 tv = tarval_or(tv4, tv2);
4774 if (!tarval_is_all_one(tv)) {
4775 /* have at least one 0 at the same bit position */
4779 n_tv4 = tarval_not(tv4);
4780 if (tv3 != tarval_and(tv3, n_tv4)) {
4781 /* bit in the or_mask is outside the and_mask */
4785 n_tv2 = tarval_not(tv2);
4786 if (tv1 != tarval_and(tv1, n_tv2)) {
4787 /* bit in the or_mask is outside the and_mask */
4791 /* ok, all conditions met */
4792 block = get_irn_n(or, -1);
4794 new_and = new_r_And(current_ir_graph, block,
4795 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4797 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4799 set_Or_left(or, new_and);
4800 set_Or_right(or, new_const);
4802 /* check for more */
4804 } /* transform_node_Or_bf_store */
4807 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4809 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4810 ir_mode *mode = get_irn_mode(or);
4811 ir_node *shl, *shr, *block;
4812 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4815 if (! mode_is_int(mode))
4818 shl = get_binop_left(or);
4819 shr = get_binop_right(or);
4828 } else if (!is_Shl(shl)) {
4830 } else if (!is_Shr(shr)) {
4833 x = get_Shl_left(shl);
4834 if (x != get_Shr_left(shr))
4837 c1 = get_Shl_right(shl);
4838 c2 = get_Shr_right(shr);
4839 if (is_Const(c1) && is_Const(c2)) {
4840 tv1 = get_Const_tarval(c1);
4841 if (! tarval_is_long(tv1))
4844 tv2 = get_Const_tarval(c2);
4845 if (! tarval_is_long(tv2))
4848 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4849 != (int) get_mode_size_bits(mode))
4852 /* yet, condition met */
4853 block = get_nodes_block(or);
4855 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4857 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4864 rotval = sub; /* a Rot right is not supported, so use a rot left */
4865 } else if (is_Sub(c2)) {
4871 if (get_Sub_right(sub) != v)
4874 c1 = get_Sub_left(sub);
4878 tv1 = get_Const_tarval(c1);
4879 if (! tarval_is_long(tv1))
4882 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4885 /* yet, condition met */
4886 block = get_nodes_block(or);
4888 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4890 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4892 } /* transform_node_Or_Rotl */
4897 static ir_node *transform_node_Or(ir_node *n) {
4898 ir_node *c, *oldn = n;
4899 ir_node *a = get_Or_left(n);
4900 ir_node *b = get_Or_right(n);
4903 if (is_Not(a) && is_Not(b)) {
4904 /* ~a | ~b = ~(a&b) */
4905 ir_node *block = get_nodes_block(n);
4907 mode = get_irn_mode(n);
4910 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4911 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4912 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4916 /* we can evaluate 2 Projs of the same Cmp */
4917 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4918 ir_node *pred_a = get_Proj_pred(a);
4919 ir_node *pred_b = get_Proj_pred(b);
4920 if (pred_a == pred_b) {
4921 dbg_info *dbgi = get_irn_dbg_info(n);
4922 ir_node *block = get_nodes_block(pred_a);
4923 pn_Cmp pn_a = get_Proj_proj(a);
4924 pn_Cmp pn_b = get_Proj_proj(b);
4925 /* yes, we can simply calculate with pncs */
4926 pn_Cmp new_pnc = pn_a | pn_b;
4928 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4933 mode = get_irn_mode(n);
4934 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4936 n = transform_node_Or_bf_store(n);
4937 n = transform_node_Or_Rotl(n);
4941 n = transform_bitwise_distributive(n, transform_node_Or);
4944 } /* transform_node_Or */
4948 static ir_node *transform_node(ir_node *n);
4951 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4953 * Should be moved to reassociation?
4955 static ir_node *transform_node_shift(ir_node *n) {
4956 ir_node *left, *right;
4958 tarval *tv1, *tv2, *res;
4959 ir_node *in[2], *irn, *block;
4961 left = get_binop_left(n);
4963 /* different operations */
4964 if (get_irn_op(left) != get_irn_op(n))
4967 right = get_binop_right(n);
4968 tv1 = value_of(right);
4969 if (tv1 == tarval_bad)
4972 tv2 = value_of(get_binop_right(left));
4973 if (tv2 == tarval_bad)
4976 res = tarval_add(tv1, tv2);
4977 mode = get_irn_mode(n);
4979 /* beware: a simple replacement works only, if res < modulo shift */
4981 int modulo_shf = get_mode_modulo_shift(mode);
4982 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4983 if (modulo_shf > 0) {
4984 tarval *modulo = new_tarval_from_long(modulo_shf,
4985 get_tarval_mode(res));
4987 /* shifting too much */
4988 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4990 ir_graph *irg = get_irn_irg(n);
4991 ir_node *block = get_nodes_block(n);
4992 dbg_info *dbgi = get_irn_dbg_info(n);
4993 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
4994 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
4998 return new_Const(mode, get_mode_null(mode));
5002 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5005 /* ok, we can replace it */
5006 block = get_nodes_block(n);
5008 in[0] = get_binop_left(left);
5009 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5011 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5013 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5015 return transform_node(irn);
5016 } /* transform_node_shift */
5019 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5021 * - and, or, xor instead of &
5022 * - Shl, Shr, Shrs, rotl instead of >>
5023 * (with a special case for Or/Xor + Shrs)
5025 static ir_node *transform_node_bitop_shift(ir_node *n) {
5027 ir_node *right = get_binop_right(n);
5028 ir_mode *mode = get_irn_mode(n);
5029 ir_node *bitop_left;
5030 ir_node *bitop_right;
5042 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5044 if (!is_Const(right))
5047 left = get_binop_left(n);
5048 op_left = get_irn_op(left);
5049 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5052 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5053 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5054 /* TODO: test if sign bit is affectes */
5058 bitop_right = get_binop_right(left);
5059 if (!is_Const(bitop_right))
5062 bitop_left = get_binop_left(left);
5064 irg = get_irn_irg(n);
5065 block = get_nodes_block(n);
5066 dbgi = get_irn_dbg_info(n);
5067 tv1 = get_Const_tarval(bitop_right);
5068 tv2 = get_Const_tarval(right);
5070 assert(get_tarval_mode(tv1) == mode);
5073 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5074 tv_shift = tarval_shl(tv1, tv2);
5075 } else if(is_Shr(n)) {
5076 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5077 tv_shift = tarval_shr(tv1, tv2);
5078 } else if(is_Shrs(n)) {
5079 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5080 tv_shift = tarval_shrs(tv1, tv2);
5083 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5084 tv_shift = tarval_rotl(tv1, tv2);
5087 assert(get_tarval_mode(tv_shift) == mode);
5088 new_const = new_Const(mode, tv_shift);
5090 if (op_left == op_And) {
5091 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5092 } else if(op_left == op_Or) {
5093 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5095 assert(op_left == op_Eor);
5096 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5104 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5106 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5107 * (also with x >>s c1 when c1>=c2)
5109 static ir_node *transform_node_shl_shr(ir_node *n) {
5111 ir_node *right = get_binop_right(n);
5127 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5129 if (!is_Const(right))
5132 left = get_binop_left(n);
5133 mode = get_irn_mode(n);
5134 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5135 ir_node *shr_right = get_binop_right(left);
5137 if (!is_Const(shr_right))
5140 x = get_binop_left(left);
5141 tv_shr = get_Const_tarval(shr_right);
5142 tv_shl = get_Const_tarval(right);
5144 if (is_Shrs(left)) {
5145 /* shrs variant only allowed if c1 >= c2 */
5146 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5149 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5152 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5154 tv_mask = tarval_shl(tv_mask, tv_shl);
5155 } else if(is_Shr(n) && is_Shl(left)) {
5156 ir_node *shl_right = get_Shl_right(left);
5158 if (!is_Const(shl_right))
5161 x = get_Shl_left(left);
5162 tv_shr = get_Const_tarval(right);
5163 tv_shl = get_Const_tarval(shl_right);
5165 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5166 tv_mask = tarval_shr(tv_mask, tv_shr);
5171 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5172 assert(tv_mask != tarval_bad);
5173 assert(get_tarval_mode(tv_mask) == mode);
5175 irg = get_irn_irg(n);
5176 block = get_nodes_block(n);
5177 dbgi = get_irn_dbg_info(n);
5179 pnc = tarval_cmp(tv_shl, tv_shr);
5180 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5181 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5182 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5184 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5186 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5189 assert(pnc == pn_Cmp_Gt);
5190 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5191 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5192 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5195 new_const = new_Const(mode, tv_mask);
5196 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5204 static ir_node *transform_node_Shr(ir_node *n) {
5205 ir_node *c, *oldn = n;
5206 ir_node *left = get_Shr_left(n);
5207 ir_node *right = get_Shr_right(n);
5208 ir_mode *mode = get_irn_mode(n);
5210 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5211 n = transform_node_shift(n);
5214 n = transform_node_shl_shr(n);
5216 n = transform_node_bitop_shift(n);
5219 } /* transform_node_Shr */
5224 static ir_node *transform_node_Shrs(ir_node *n) {
5225 ir_node *c, *oldn = n;
5226 ir_node *a = get_Shrs_left(n);
5227 ir_node *b = get_Shrs_right(n);
5228 ir_mode *mode = get_irn_mode(n);
5230 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5231 n = transform_node_shift(n);
5234 n = transform_node_bitop_shift(n);
5237 } /* transform_node_Shrs */
5242 static ir_node *transform_node_Shl(ir_node *n) {
5243 ir_node *c, *oldn = n;
5244 ir_node *a = get_Shl_left(n);
5245 ir_node *b = get_Shl_right(n);
5246 ir_mode *mode = get_irn_mode(n);
5248 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5249 n = transform_node_shift(n);
5252 n = transform_node_shl_shr(n);
5254 n = transform_node_bitop_shift(n);
5257 } /* transform_node_Shl */
5262 static ir_node *transform_node_Rotl(ir_node *n) {
5263 ir_node *c, *oldn = n;
5264 ir_node *a = get_Rotl_left(n);
5265 ir_node *b = get_Rotl_right(n);
5266 ir_mode *mode = get_irn_mode(n);
5268 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5269 n = transform_node_shift(n);
5272 n = transform_node_bitop_shift(n);
5275 } /* transform_node_Rotl */
5280 static ir_node *transform_node_Conv(ir_node *n) {
5281 ir_node *c, *oldn = n;
5282 ir_mode *mode = get_irn_mode(n);
5283 ir_node *a = get_Conv_op(n);
5285 if (mode != mode_b && is_const_Phi(a)) {
5286 /* Do NOT optimize mode_b Conv's, this leads to remaining
5287 * Phib nodes later, because the conv_b_lower operation
5288 * is instantly reverted, when it tries to insert a Convb.
5290 c = apply_conv_on_phi(a, mode);
5292 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5297 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5298 return new_r_Unknown(current_ir_graph, mode);
5301 if (mode_is_reference(mode) &&
5302 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5304 ir_node *l = get_Add_left(a);
5305 ir_node *r = get_Add_right(a);
5306 dbg_info *dbgi = get_irn_dbg_info(a);
5307 ir_node *block = get_nodes_block(n);
5309 ir_node *lop = get_Conv_op(l);
5310 if(get_irn_mode(lop) == mode) {
5311 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5312 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5317 ir_node *rop = get_Conv_op(r);
5318 if(get_irn_mode(rop) == mode) {
5319 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5320 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5327 } /* transform_node_Conv */
5330 * Remove dead blocks and nodes in dead blocks
5331 * in keep alive list. We do not generate a new End node.
5333 static ir_node *transform_node_End(ir_node *n) {
5334 int i, j, n_keepalives = get_End_n_keepalives(n);
5337 NEW_ARR_A(ir_node *, in, n_keepalives);
5339 for (i = j = 0; i < n_keepalives; ++i) {
5340 ir_node *ka = get_End_keepalive(n, i);
5342 if (! is_Block_dead(ka)) {
5346 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5348 } else if (is_Bad(ka)) {
5349 /* no need to keep Bad */
5354 if (j != n_keepalives)
5355 set_End_keepalives(n, j, in);
5357 } /* transform_node_End */
5359 /** returns 1 if a == -b */
5360 static int is_negated_value(ir_node *a, ir_node *b) {
5361 if (is_Minus(a) && get_Minus_op(a) == b)
5363 if (is_Minus(b) && get_Minus_op(b) == a)
5365 if (is_Sub(a) && is_Sub(b)) {
5366 ir_node *a_left = get_Sub_left(a);
5367 ir_node *a_right = get_Sub_right(a);
5368 ir_node *b_left = get_Sub_left(b);
5369 ir_node *b_right = get_Sub_right(b);
5371 if (a_left == b_right && a_right == b_left)
5379 * Optimize a Mux into some simpler cases.
5381 static ir_node *transform_node_Mux(ir_node *n) {
5382 ir_node *oldn = n, *sel = get_Mux_sel(n);
5383 ir_mode *mode = get_irn_mode(n);
5384 ir_node *t = get_Mux_true(n);
5385 ir_node *f = get_Mux_false(n);
5386 ir_graph *irg = current_ir_graph;
5388 /* first normalization step: move a possible zero to the false case */
5390 ir_node *cmp = get_Proj_pred(sel);
5393 if (is_Const(t) && is_Const_null(t)) {
5396 /* Mux(x, 0, y) => Mux(x, y, 0) */
5397 pn_Cmp pnc = get_Proj_proj(sel);
5398 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5399 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5400 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5408 /* note: after normalization, false can only happen on default */
5409 if (mode == mode_b) {
5410 dbg_info *dbg = get_irn_dbg_info(n);
5411 ir_node *block = get_nodes_block(n);
5412 ir_graph *irg = current_ir_graph;
5415 tarval *tv_t = get_Const_tarval(t);
5416 if (tv_t == tarval_b_true) {
5418 /* Muxb(sel, true, false) = sel */
5419 assert(get_Const_tarval(f) == tarval_b_false);
5420 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5423 /* Muxb(sel, true, x) = Or(sel, x) */
5424 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5425 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5429 } else if (is_Const(f)) {
5430 tarval *tv_f = get_Const_tarval(f);
5431 if (tv_f == tarval_b_true) {
5432 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5433 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5434 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5435 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5438 /* Muxb(sel, x, false) = And(sel, x) */
5439 assert(tv_f == tarval_b_false);
5440 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5447 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5448 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5449 tarval *a = get_Const_tarval(t);
5450 tarval *b = get_Const_tarval(f);
5451 tarval *null = get_tarval_null(mode);
5454 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5455 diff = tarval_sub(a, b, NULL);
5458 diff = tarval_sub(b, a, NULL);
5462 if (diff == get_tarval_one(mode) && min != null) {
5463 dbg_info *dbg = get_irn_dbg_info(n);
5464 ir_node *block = get_nodes_block(n);
5465 ir_graph *irg = current_ir_graph;
5466 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5467 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5468 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5469 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5475 ir_node *cmp = get_Proj_pred(sel);
5476 long pn = get_Proj_proj(sel);
5479 * Note: normalization puts the constant on the right side,
5480 * so we check only one case.
5482 * Note further that these optimization work even for floating point
5483 * with NaN's because -NaN == NaN.
5484 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5488 ir_node *cmp_r = get_Cmp_right(cmp);
5489 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5490 ir_node *block = get_nodes_block(n);
5491 ir_node *cmp_l = get_Cmp_left(cmp);
5493 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5496 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5497 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5499 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5500 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5502 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5504 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5505 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5507 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5508 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5510 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5512 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5517 if (mode_is_int(mode)) {
5519 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5520 /* Mux((a & b) != 0, c, 0) */
5521 ir_node *and_r = get_And_right(cmp_l);
5524 if (and_r == t && f == cmp_r) {
5525 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5526 if (pn == pn_Cmp_Lg) {
5527 /* Mux((a & 2^C) != 0, 2^C, 0) */
5529 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5531 /* Mux((a & 2^C) == 0, 2^C, 0) */
5532 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5533 block, cmp_l, t, mode);
5534 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5539 if (is_Shl(and_r)) {
5540 ir_node *shl_l = get_Shl_left(and_r);
5541 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5542 if (and_r == t && f == cmp_r) {
5543 if (pn == pn_Cmp_Lg) {
5544 /* (a & (1 << n)) != 0, (1 << n), 0) */
5546 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5548 /* (a & (1 << n)) == 0, (1 << n), 0) */
5549 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5550 block, cmp_l, t, mode);
5551 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5557 and_l = get_And_left(cmp_l);
5558 if (is_Shl(and_l)) {
5559 ir_node *shl_l = get_Shl_left(and_l);
5560 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5561 if (and_l == t && f == cmp_r) {
5562 if (pn == pn_Cmp_Lg) {
5563 /* ((1 << n) & a) != 0, (1 << n), 0) */
5565 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5567 /* ((1 << n) & a) == 0, (1 << n), 0) */
5568 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5569 block, cmp_l, t, mode);
5570 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5581 return arch_transform_node_Mux(n);
5582 } /* transform_node_Mux */
5585 * optimize Sync nodes that have other syncs as input we simply add the inputs
5586 * of the other sync to our own inputs
5588 static ir_node *transform_node_Sync(ir_node *n) {
5589 int arity = get_Sync_n_preds(n);
5592 for (i = 0; i < arity;) {
5593 ir_node *pred = get_Sync_pred(n, i);
5597 if (!is_Sync(pred)) {
5605 pred_arity = get_Sync_n_preds(pred);
5606 for (j = 0; j < pred_arity; ++j) {
5607 ir_node *pred_pred = get_Sync_pred(pred, j);
5612 add_irn_n(n, pred_pred);
5616 if (get_Sync_pred(n, k) == pred_pred) break;
5621 /* rehash the sync node */
5622 add_identities(current_ir_graph->value_table, n);
5628 * Tries several [inplace] [optimizing] transformations and returns an
5629 * equivalent node. The difference to equivalent_node() is that these
5630 * transformations _do_ generate new nodes, and thus the old node must
5631 * not be freed even if the equivalent node isn't the old one.
5633 static ir_node *transform_node(ir_node *n) {
5637 * Transform_node is the only "optimizing transformation" that might
5638 * return a node with a different opcode. We iterate HERE until fixpoint
5639 * to get the final result.
5643 if (n->op->ops.transform_node)
5644 n = n->op->ops.transform_node(n);
5645 } while (oldn != n);
5648 } /* transform_node */
5651 * Sets the default transform node operation for an ir_op_ops.
5653 * @param code the opcode for the default operation
5654 * @param ops the operations initialized
5659 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5663 ops->transform_node = transform_node_##a; \
5665 #define CASE_PROJ(a) \
5667 ops->transform_node_Proj = transform_node_Proj_##a; \
5669 #define CASE_PROJ_EX(a) \
5671 ops->transform_node = transform_node_##a; \
5672 ops->transform_node_Proj = transform_node_Proj_##a; \
5681 CASE_PROJ_EX(DivMod);
5715 } /* firm_set_default_transform_node */
5718 /* **************** Common Subexpression Elimination **************** */
5720 /** The size of the hash table used, should estimate the number of nodes
5722 #define N_IR_NODES 512
5724 /** Compares the attributes of two Const nodes. */
5725 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5726 return (get_Const_tarval(a) != get_Const_tarval(b))
5727 || (get_Const_type(a) != get_Const_type(b));
5728 } /* node_cmp_attr_Const */
5730 /** Compares the attributes of two Proj nodes. */
5731 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5732 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5733 } /* node_cmp_attr_Proj */
5735 /** Compares the attributes of two Filter nodes. */
5736 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5737 return get_Filter_proj(a) != get_Filter_proj(b);
5738 } /* node_cmp_attr_Filter */
5740 /** Compares the attributes of two Alloc nodes. */
5741 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5742 const alloc_attr *pa = get_irn_alloc_attr(a);
5743 const alloc_attr *pb = get_irn_alloc_attr(b);
5744 return (pa->where != pb->where) || (pa->type != pb->type);
5745 } /* node_cmp_attr_Alloc */
5747 /** Compares the attributes of two Free nodes. */
5748 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5749 const free_attr *pa = get_irn_free_attr(a);
5750 const free_attr *pb = get_irn_free_attr(b);
5751 return (pa->where != pb->where) || (pa->type != pb->type);
5752 } /* node_cmp_attr_Free */
5754 /** Compares the attributes of two SymConst nodes. */
5755 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5756 const symconst_attr *pa = get_irn_symconst_attr(a);
5757 const symconst_attr *pb = get_irn_symconst_attr(b);
5758 return (pa->kind != pb->kind)
5759 || (pa->sym.type_p != pb->sym.type_p)
5760 || (pa->tp != pb->tp);
5761 } /* node_cmp_attr_SymConst */
5763 /** Compares the attributes of two Call nodes. */
5764 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5765 return get_irn_call_attr(a) != get_irn_call_attr(b);
5766 } /* node_cmp_attr_Call */
5768 /** Compares the attributes of two Sel nodes. */
5769 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5770 const ir_entity *a_ent = get_Sel_entity(a);
5771 const ir_entity *b_ent = get_Sel_entity(b);
5774 (a_ent->kind != b_ent->kind) ||
5775 (a_ent->name != b_ent->name) ||
5776 (a_ent->owner != b_ent->owner) ||
5777 (a_ent->ld_name != b_ent->ld_name) ||
5778 (a_ent->type != b_ent->type);
5780 /* Matze: inlining of functions can produce 2 entities with same type,
5782 return a_ent != b_ent;
5783 } /* node_cmp_attr_Sel */
5785 /** Compares the attributes of two Phi nodes. */
5786 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5787 /* we can only enter this function if both nodes have the same number of inputs,
5788 hence it is enough to check if one of them is a Phi0 */
5790 /* check the Phi0 pos attribute */
5791 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5794 } /* node_cmp_attr_Phi */
5796 /** Compares the attributes of two Conv nodes. */
5797 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5798 return get_Conv_strict(a) != get_Conv_strict(b);
5799 } /* node_cmp_attr_Conv */
5801 /** Compares the attributes of two Cast nodes. */
5802 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5803 return get_Cast_type(a) != get_Cast_type(b);
5804 } /* node_cmp_attr_Cast */
5806 /** Compares the attributes of two Load nodes. */
5807 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5808 if (get_Load_volatility(a) == volatility_is_volatile ||
5809 get_Load_volatility(b) == volatility_is_volatile)
5810 /* NEVER do CSE on volatile Loads */
5812 /* do not CSE Loads with different alignment. Be conservative. */
5813 if (get_Load_align(a) != get_Load_align(b))
5816 return get_Load_mode(a) != get_Load_mode(b);
5817 } /* node_cmp_attr_Load */
5819 /** Compares the attributes of two Store nodes. */
5820 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5821 /* do not CSE Stores with different alignment. Be conservative. */
5822 if (get_Store_align(a) != get_Store_align(b))
5825 /* NEVER do CSE on volatile Stores */
5826 return (get_Store_volatility(a) == volatility_is_volatile ||
5827 get_Store_volatility(b) == volatility_is_volatile);
5828 } /* node_cmp_attr_Store */
5830 /** Compares two exception attributes */
5831 static int node_cmp_exception(ir_node *a, ir_node *b) {
5832 const except_attr *ea = get_irn_except_attr(a);
5833 const except_attr *eb = get_irn_except_attr(b);
5835 return ea->pin_state != eb->pin_state;
5838 #define node_cmp_attr_Bound node_cmp_exception
5840 /** Compares the attributes of two Div nodes. */
5841 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5842 const divmod_attr *ma = get_irn_divmod_attr(a);
5843 const divmod_attr *mb = get_irn_divmod_attr(b);
5844 return ma->exc.pin_state != mb->exc.pin_state ||
5845 ma->res_mode != mb->res_mode ||
5846 ma->no_remainder != mb->no_remainder;
5847 } /* node_cmp_attr_Div */
5849 /** Compares the attributes of two DivMod nodes. */
5850 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5851 const divmod_attr *ma = get_irn_divmod_attr(a);
5852 const divmod_attr *mb = get_irn_divmod_attr(b);
5853 return ma->exc.pin_state != mb->exc.pin_state ||
5854 ma->res_mode != mb->res_mode;
5855 } /* node_cmp_attr_DivMod */
5857 /** Compares the attributes of two Mod nodes. */
5858 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5859 const divmod_attr *ma = get_irn_divmod_attr(a);
5860 const divmod_attr *mb = get_irn_divmod_attr(b);
5861 return ma->exc.pin_state != mb->exc.pin_state ||
5862 ma->res_mode != mb->res_mode;
5863 } /* node_cmp_attr_Mod */
5865 /** Compares the attributes of two Quot nodes. */
5866 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5867 const divmod_attr *ma = get_irn_divmod_attr(a);
5868 const divmod_attr *mb = get_irn_divmod_attr(b);
5869 return ma->exc.pin_state != mb->exc.pin_state ||
5870 ma->res_mode != mb->res_mode;
5871 } /* node_cmp_attr_Quot */
5873 /** Compares the attributes of two Confirm nodes. */
5874 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5875 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5876 } /* node_cmp_attr_Confirm */
5878 /** Compares the attributes of two ASM nodes. */
5879 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5881 const ir_asm_constraint *ca;
5882 const ir_asm_constraint *cb;
5885 if (get_ASM_text(a) != get_ASM_text(b))
5888 /* Should we really check the constraints here? Should be better, but is strange. */
5889 n = get_ASM_n_input_constraints(a);
5890 if (n != get_ASM_n_input_constraints(b))
5893 ca = get_ASM_input_constraints(a);
5894 cb = get_ASM_input_constraints(b);
5895 for (i = 0; i < n; ++i) {
5896 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5900 n = get_ASM_n_output_constraints(a);
5901 if (n != get_ASM_n_output_constraints(b))
5904 ca = get_ASM_output_constraints(a);
5905 cb = get_ASM_output_constraints(b);
5906 for (i = 0; i < n; ++i) {
5907 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5911 n = get_ASM_n_clobbers(a);
5912 if (n != get_ASM_n_clobbers(b))
5915 cla = get_ASM_clobbers(a);
5916 clb = get_ASM_clobbers(b);
5917 for (i = 0; i < n; ++i) {
5918 if (cla[i] != clb[i])
5922 } /* node_cmp_attr_ASM */
5925 * Set the default node attribute compare operation for an ir_op_ops.
5927 * @param code the opcode for the default operation
5928 * @param ops the operations initialized
5933 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5937 ops->node_cmp_attr = node_cmp_attr_##a; \
5968 } /* firm_set_default_node_cmp_attr */
5971 * Compare function for two nodes in the value table. Gets two
5972 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5974 int identities_cmp(const void *elt, const void *key) {
5975 ir_node *a = (ir_node *)elt;
5976 ir_node *b = (ir_node *)key;
5979 if (a == b) return 0;
5981 if ((get_irn_op(a) != get_irn_op(b)) ||
5982 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5984 /* compare if a's in and b's in are of equal length */
5985 irn_arity_a = get_irn_intra_arity(a);
5986 if (irn_arity_a != get_irn_intra_arity(b))
5989 if (get_irn_pinned(a) == op_pin_state_pinned) {
5990 /* for pinned nodes, the block inputs must be equal */
5991 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5993 } else if (! get_opt_global_cse()) {
5994 /* for block-local CSE both nodes must be in the same MacroBlock */
5995 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5999 /* compare a->in[0..ins] with b->in[0..ins] */
6000 for (i = 0; i < irn_arity_a; i++)
6001 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6005 * here, we already now that the nodes are identical except their
6008 if (a->op->ops.node_cmp_attr)
6009 return a->op->ops.node_cmp_attr(a, b);
6012 } /* identities_cmp */
6015 * Calculate a hash value of a node.
6017 * @param node The IR-node
6019 unsigned ir_node_hash(const ir_node *node) {
6020 return node->op->ops.hash(node);
6021 } /* ir_node_hash */
6024 pset *new_identities(void) {
6025 return new_pset(identities_cmp, N_IR_NODES);
6026 } /* new_identities */
6028 void del_identities(pset *value_table) {
6029 del_pset(value_table);
6030 } /* del_identities */
6033 * Normalize a node by putting constants (and operands with larger
6034 * node index) on the right (operator side).
6036 * @param n The node to normalize
6038 static void normalize_node(ir_node *n) {
6039 if (is_op_commutative(get_irn_op(n))) {
6040 ir_node *l = get_binop_left(n);
6041 ir_node *r = get_binop_right(n);
6043 /* For commutative operators perform a OP b == b OP a but keep
6044 * constants on the RIGHT side. This helps greatly in some
6045 * optimizations. Moreover we use the idx number to make the form
6047 if (!operands_are_normalized(l, r)) {
6048 set_binop_left(n, r);
6049 set_binop_right(n, l);
6052 } /* normalize_node */
6055 * Update the nodes after a match in the value table. If both nodes have
6056 * the same MacroBlock but different Blocks, we must ensure that the node
6057 * with the dominating Block (the node that is near to the MacroBlock header
6058 * is stored in the table.
6059 * Because a MacroBlock has only one "non-exception" flow, we don't need
6060 * dominance info here: We known, that one block must dominate the other and
6061 * following the only block input will allow to find it.
6063 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6064 ir_node *known_blk, *new_block, *block, *mbh;
6066 if (get_opt_global_cse()) {
6067 /* Block inputs are meaning less */
6070 known_blk = get_irn_n(known_irn, -1);
6071 new_block = get_irn_n(new_ir_node, -1);
6072 if (known_blk == new_block) {
6073 /* already in the same block */
6077 * We expect the typical case when we built the graph. In that case, the
6078 * known_irn is already the upper one, so checking this should be faster.
6081 mbh = get_Block_MacroBlock(new_block);
6083 if (block == known_blk) {
6084 /* ok, we have found it: known_block dominates new_block as expected */
6089 * We have reached the MacroBlock header NOT founding
6090 * the known_block. new_block must dominate known_block.
6093 set_irn_n(known_irn, -1, new_block);
6096 assert(get_Block_n_cfgpreds(block) == 1);
6097 block = get_Block_cfgpred_block(block, 0);
6099 } /* update_value_table */
6102 * Return the canonical node computing the same value as n.
6103 * Looks up the node in a hash table, enters it in the table
6104 * if it isn't there yet.
6106 * @param value_table the HashSet containing all nodes in the
6108 * @param n the node to look up
6110 * @return a node that computes the same value as n or n if no such
6111 * node could be found
6113 ir_node *identify_remember(pset *value_table, ir_node *n) {
6116 if (!value_table) return n;
6119 /* lookup or insert in hash table with given hash key. */
6120 o = pset_insert(value_table, n, ir_node_hash(n));
6123 update_known_irn(o, n);
6127 } /* identify_remember */
6130 * During construction we set the op_pin_state_pinned flag in the graph right when the
6131 * optimization is performed. The flag turning on procedure global cse could
6132 * be changed between two allocations. This way we are safe.
6134 * @param value_table The value table
6135 * @param n The node to lookup
6137 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6140 n = identify_remember(value_table, n);
6141 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6142 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6144 } /* identify_cons */
6146 /* Add a node to the identities value table. */
6147 void add_identities(pset *value_table, ir_node *node) {
6148 if (get_opt_cse() && is_no_Block(node))
6149 identify_remember(value_table, node);
6150 } /* add_identities */
6152 /* Visit each node in the value table of a graph. */
6153 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6155 ir_graph *rem = current_ir_graph;
6157 current_ir_graph = irg;
6158 foreach_pset(irg->value_table, node)
6160 current_ir_graph = rem;
6161 } /* visit_all_identities */
6164 * Garbage in, garbage out. If a node has a dead input, i.e., the
6165 * Bad node is input to the node, return the Bad node.
6167 static ir_node *gigo(ir_node *node) {
6169 ir_op *op = get_irn_op(node);
6171 /* remove garbage blocks by looking at control flow that leaves the block
6172 and replacing the control flow by Bad. */
6173 if (get_irn_mode(node) == mode_X) {
6174 ir_node *block = get_nodes_block(skip_Proj(node));
6176 /* Don't optimize nodes in immature blocks. */
6177 if (!get_Block_matured(block))
6179 /* Don't optimize End, may have Bads. */
6180 if (op == op_End) return node;
6182 if (is_Block(block)) {
6183 if (is_Block_dead(block)) {
6184 /* control flow from dead block is dead */
6188 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6189 if (!is_Bad(get_irn_n(block, i)))
6193 ir_graph *irg = get_irn_irg(block);
6194 /* the start block is never dead */
6195 if (block != get_irg_start_block(irg)
6196 && block != get_irg_end_block(irg)) {
6198 * Do NOT kill control flow without setting
6199 * the block to dead of bad things can happen:
6200 * We get a Block that is not reachable be irg_block_walk()
6201 * but can be found by irg_walk()!
6203 set_Block_dead(block);
6210 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6211 blocks predecessors is dead. */
6212 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6213 irn_arity = get_irn_arity(node);
6216 * Beware: we can only read the block of a non-floating node.
6218 if (is_irn_pinned_in_irg(node) &&
6219 is_Block_dead(get_nodes_block(skip_Proj(node))))
6222 for (i = 0; i < irn_arity; i++) {
6223 ir_node *pred = get_irn_n(node, i);
6228 /* Propagating Unknowns here seems to be a bad idea, because
6229 sometimes we need a node as a input and did not want that
6231 However, it might be useful to move this into a later phase
6232 (if you think that optimizing such code is useful). */
6233 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6234 return new_Unknown(get_irn_mode(node));
6239 /* With this code we violate the agreement that local_optimize
6240 only leaves Bads in Block, Phi and Tuple nodes. */
6241 /* If Block has only Bads as predecessors it's garbage. */
6242 /* If Phi has only Bads as predecessors it's garbage. */
6243 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6244 irn_arity = get_irn_arity(node);
6245 for (i = 0; i < irn_arity; i++) {
6246 if (!is_Bad(get_irn_n(node, i))) break;
6248 if (i == irn_arity) node = new_Bad();
6255 * These optimizations deallocate nodes from the obstack.
6256 * It can only be called if it is guaranteed that no other nodes
6257 * reference this one, i.e., right after construction of a node.
6259 * @param n The node to optimize
6261 * current_ir_graph must be set to the graph of the node!
6263 ir_node *optimize_node(ir_node *n) {
6266 ir_opcode iro = get_irn_opcode(n);
6268 /* Always optimize Phi nodes: part of the construction. */
6269 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6271 /* constant expression evaluation / constant folding */
6272 if (get_opt_constant_folding()) {
6273 /* neither constants nor Tuple values can be evaluated */
6274 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6275 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6276 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6277 /* try to evaluate */
6278 tv = computed_value(n);
6279 if (tv != tarval_bad) {
6281 ir_type *old_tp = get_irn_type(n);
6282 int i, arity = get_irn_arity(n);
6286 * Try to recover the type of the new expression.
6288 for (i = 0; i < arity && !old_tp; ++i)
6289 old_tp = get_irn_type(get_irn_n(n, i));
6292 * we MUST copy the node here temporary, because it's still needed
6293 * for DBG_OPT_CSTEVAL
6295 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6296 oldn = alloca(node_size);
6298 memcpy(oldn, n, node_size);
6299 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6301 /* ARG, copy the in array, we need it for statistics */
6302 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6304 /* note the inplace edges module */
6305 edges_node_deleted(n, current_ir_graph);
6307 /* evaluation was successful -- replace the node. */
6308 irg_kill_node(current_ir_graph, n);
6309 nw = new_Const(get_tarval_mode(tv), tv);
6311 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6312 set_Const_type(nw, old_tp);
6313 DBG_OPT_CSTEVAL(oldn, nw);
6314 tarval_enable_fp_ops(old_fp_mode);
6317 tarval_enable_fp_ops(old_fp_mode);
6321 /* remove unnecessary nodes */
6322 if (get_opt_algebraic_simplification() ||
6323 (iro == iro_Phi) || /* always optimize these nodes. */
6325 (iro == iro_Proj) ||
6326 (iro == iro_Block) ) /* Flags tested local. */
6327 n = equivalent_node(n);
6329 /* Common Subexpression Elimination.
6331 * Checks whether n is already available.
6332 * The block input is used to distinguish different subexpressions. Right
6333 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6334 * subexpressions within a block.
6337 n = identify_cons(current_ir_graph->value_table, n);
6340 edges_node_deleted(oldn, current_ir_graph);
6342 /* We found an existing, better node, so we can deallocate the old node. */
6343 irg_kill_node(current_ir_graph, oldn);
6347 /* Some more constant expression evaluation that does not allow to
6349 iro = get_irn_opcode(n);
6350 if (get_opt_algebraic_simplification() ||
6351 (iro == iro_Cond) ||
6352 (iro == iro_Proj)) /* Flags tested local. */
6353 n = transform_node(n);
6355 /* Remove nodes with dead (Bad) input.
6356 Run always for transformation induced Bads. */
6359 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6360 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6362 n = identify_remember(current_ir_graph->value_table, o);
6368 } /* optimize_node */
6372 * These optimizations never deallocate nodes (in place). This can cause dead
6373 * nodes lying on the obstack. Remove these by a dead node elimination,
6374 * i.e., a copying garbage collection.
6376 ir_node *optimize_in_place_2(ir_node *n) {
6379 ir_opcode iro = get_irn_opcode(n);
6381 if (!get_opt_optimize() && !is_Phi(n)) return n;
6383 /* constant expression evaluation / constant folding */
6384 if (get_opt_constant_folding()) {
6385 /* neither constants nor Tuple values can be evaluated */
6386 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6387 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6388 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6389 /* try to evaluate */
6390 tv = computed_value(n);
6391 if (tv != tarval_bad) {
6392 /* evaluation was successful -- replace the node. */
6393 ir_type *old_tp = get_irn_type(n);
6394 int i, arity = get_irn_arity(n);
6397 * Try to recover the type of the new expression.
6399 for (i = 0; i < arity && !old_tp; ++i)
6400 old_tp = get_irn_type(get_irn_n(n, i));
6402 n = new_Const(get_tarval_mode(tv), tv);
6404 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6405 set_Const_type(n, old_tp);
6407 DBG_OPT_CSTEVAL(oldn, n);
6408 tarval_enable_fp_ops(old_fp_mode);
6411 tarval_enable_fp_ops(old_fp_mode);
6415 /* remove unnecessary nodes */
6416 if (get_opt_constant_folding() ||
6417 (iro == iro_Phi) || /* always optimize these nodes. */
6418 (iro == iro_Id) || /* ... */
6419 (iro == iro_Proj) || /* ... */
6420 (iro == iro_Block) ) /* Flags tested local. */
6421 n = equivalent_node(n);
6423 /** common subexpression elimination **/
6424 /* Checks whether n is already available. */
6425 /* The block input is used to distinguish different subexpressions. Right
6426 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6427 subexpressions within a block. */
6428 if (get_opt_cse()) {
6430 n = identify_remember(current_ir_graph->value_table, o);
6435 /* Some more constant expression evaluation. */
6436 iro = get_irn_opcode(n);
6437 if (get_opt_constant_folding() ||
6438 (iro == iro_Cond) ||
6439 (iro == iro_Proj)) /* Flags tested local. */
6440 n = transform_node(n);
6442 /* Remove nodes with dead (Bad) input.
6443 Run always for transformation induced Bads. */
6446 /* Now we can verify the node, as it has no dead inputs any more. */
6449 /* Now we have a legal, useful node. Enter it in hash table for cse.
6450 Blocks should be unique anyways. (Except the successor of start:
6451 is cse with the start block!) */
6452 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6454 n = identify_remember(current_ir_graph->value_table, o);
6460 } /* optimize_in_place_2 */
6463 * Wrapper for external use, set proper status bits after optimization.
6465 ir_node *optimize_in_place(ir_node *n) {
6466 /* Handle graph state */
6467 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6469 if (get_opt_global_cse())
6470 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6471 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6472 set_irg_outs_inconsistent(current_ir_graph);
6474 /* FIXME: Maybe we could also test whether optimizing the node can
6475 change the control graph. */
6476 set_irg_doms_inconsistent(current_ir_graph);
6477 return optimize_in_place_2(n);
6478 } /* optimize_in_place */
6481 * Calculate a hash value of a Const node.
6483 static unsigned hash_Const(const ir_node *node) {
6486 /* special value for const, as they only differ in their tarval. */
6487 h = HASH_PTR(node->attr.con.tv);
6488 h = 9*h + HASH_PTR(get_irn_mode(node));
6494 * Calculate a hash value of a SymConst node.
6496 static unsigned hash_SymConst(const ir_node *node) {
6499 /* special value for const, as they only differ in their symbol. */
6500 h = HASH_PTR(node->attr.symc.sym.type_p);
6501 h = 9*h + HASH_PTR(get_irn_mode(node));
6504 } /* hash_SymConst */
6507 * Set the default hash operation in an ir_op_ops.
6509 * @param code the opcode for the default operation
6510 * @param ops the operations initialized
6515 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6519 ops->hash = hash_##a; \
6522 /* hash function already set */
6523 if (ops->hash != NULL)
6530 /* use input/mode default hash if no function was given */
6531 ops->hash = firm_default_hash;
6539 * Sets the default operation for an ir_ops.
6541 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6542 ops = firm_set_default_hash(code, ops);
6543 ops = firm_set_default_computed_value(code, ops);
6544 ops = firm_set_default_equivalent_node(code, ops);
6545 ops = firm_set_default_transform_node(code, ops);
6546 ops = firm_set_default_node_cmp_attr(code, ops);
6547 ops = firm_set_default_get_type(code, ops);
6548 ops = firm_set_default_get_type_attr(code, ops);
6549 ops = firm_set_default_get_entity_attr(code, ops);
6552 } /* firm_set_default_operations */