2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Returns the tarval of a Const node or tarval_bad for all other nodes.
59 static tarval *default_value_of(const ir_node *n) {
61 return get_Const_tarval(n); /* might return tarval_bad */
66 value_of_func value_of_ptr = default_value_of;
68 /* * Set a new value_of function. */
69 void set_value_of_func(value_of_func func) {
73 value_of_ptr = default_value_of;
77 * Return the value of a Constant.
79 static tarval *computed_value_Const(const ir_node *n) {
80 return get_Const_tarval(n);
81 } /* computed_value_Const */
84 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
86 static tarval *computed_value_SymConst(const ir_node *n) {
90 switch (get_SymConst_kind(n)) {
91 case symconst_type_size:
92 type = get_SymConst_type(n);
93 if (get_type_state(type) == layout_fixed)
94 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
96 case symconst_type_align:
97 type = get_SymConst_type(n);
98 if (get_type_state(type) == layout_fixed)
99 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
101 case symconst_ofs_ent:
102 ent = get_SymConst_entity(n);
103 type = get_entity_owner(ent);
104 if (get_type_state(type) == layout_fixed)
105 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
111 } /* computed_value_SymConst */
114 * Return the value of an Add.
116 static tarval *computed_value_Add(const ir_node *n) {
117 ir_node *a = get_Add_left(n);
118 ir_node *b = get_Add_right(n);
120 tarval *ta = value_of(a);
121 tarval *tb = value_of(b);
123 if ((ta != tarval_bad) && (tb != tarval_bad))
124 return tarval_add(ta, tb);
127 } /* computed_value_Add */
130 * Return the value of a Sub.
131 * Special case: a - a
133 static tarval *computed_value_Sub(const ir_node *n) {
134 ir_mode *mode = get_irn_mode(n);
135 ir_node *a = get_Sub_left(n);
136 ir_node *b = get_Sub_right(n);
141 if (a == b && !is_Bad(a))
142 return get_mode_null(mode);
147 if ((ta != tarval_bad) && (tb != tarval_bad))
148 return tarval_sub(ta, tb, mode);
151 } /* computed_value_Sub */
154 * Return the value of a Carry.
155 * Special : a op 0, 0 op b
157 static tarval *computed_value_Carry(const ir_node *n) {
158 ir_node *a = get_binop_left(n);
159 ir_node *b = get_binop_right(n);
160 ir_mode *m = get_irn_mode(n);
162 tarval *ta = value_of(a);
163 tarval *tb = value_of(b);
165 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
169 if (tarval_is_null(ta) || tarval_is_null(tb))
170 return get_mode_null(m);
173 } /* computed_value_Carry */
176 * Return the value of a Borrow.
179 static tarval *computed_value_Borrow(const ir_node *n) {
180 ir_node *a = get_binop_left(n);
181 ir_node *b = get_binop_right(n);
182 ir_mode *m = get_irn_mode(n);
184 tarval *ta = value_of(a);
185 tarval *tb = value_of(b);
187 if ((ta != tarval_bad) && (tb != tarval_bad)) {
188 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
189 } else if (tarval_is_null(ta)) {
190 return get_mode_null(m);
193 } /* computed_value_Borrow */
196 * Return the value of an unary Minus.
198 static tarval *computed_value_Minus(const ir_node *n) {
199 ir_node *a = get_Minus_op(n);
200 tarval *ta = value_of(a);
202 if (ta != tarval_bad)
203 return tarval_neg(ta);
206 } /* computed_value_Minus */
209 * Return the value of a Mul.
211 static tarval *computed_value_Mul(const ir_node *n) {
212 ir_node *a = get_Mul_left(n);
213 ir_node *b = get_Mul_right(n);
216 tarval *ta = value_of(a);
217 tarval *tb = value_of(b);
219 mode = get_irn_mode(n);
220 if (mode != get_irn_mode(a)) {
221 /* n * n = 2n bit multiplication */
222 ta = tarval_convert_to(ta, mode);
223 tb = tarval_convert_to(tb, mode);
226 if (ta != tarval_bad && tb != tarval_bad) {
227 return tarval_mul(ta, tb);
229 /* a*0 = 0 or 0*b = 0 */
230 if (ta == get_mode_null(mode))
232 if (tb == get_mode_null(mode))
236 } /* computed_value_Mul */
239 * Return the value of an Abs.
241 static tarval *computed_value_Abs(const ir_node *n) {
242 ir_node *a = get_Abs_op(n);
243 tarval *ta = value_of(a);
245 if (ta != tarval_bad)
246 return tarval_abs(ta);
249 } /* computed_value_Abs */
252 * Return the value of an And.
253 * Special case: a & 0, 0 & b
255 static tarval *computed_value_And(const ir_node *n) {
256 ir_node *a = get_And_left(n);
257 ir_node *b = get_And_right(n);
259 tarval *ta = value_of(a);
260 tarval *tb = value_of(b);
262 if ((ta != tarval_bad) && (tb != tarval_bad)) {
263 return tarval_and (ta, tb);
265 if (tarval_is_null(ta)) return ta;
266 if (tarval_is_null(tb)) return tb;
269 } /* computed_value_And */
272 * Return the value of an Or.
273 * Special case: a | 1...1, 1...1 | b
275 static tarval *computed_value_Or(const ir_node *n) {
276 ir_node *a = get_Or_left(n);
277 ir_node *b = get_Or_right(n);
279 tarval *ta = value_of(a);
280 tarval *tb = value_of(b);
282 if ((ta != tarval_bad) && (tb != tarval_bad)) {
283 return tarval_or (ta, tb);
285 if (tarval_is_all_one(ta)) return ta;
286 if (tarval_is_all_one(tb)) return tb;
289 } /* computed_value_Or */
292 * Return the value of an Eor.
294 static tarval *computed_value_Eor(const ir_node *n) {
295 ir_node *a = get_Eor_left(n);
296 ir_node *b = get_Eor_right(n);
301 return get_mode_null(get_irn_mode(n));
306 if ((ta != tarval_bad) && (tb != tarval_bad)) {
307 return tarval_eor (ta, tb);
310 } /* computed_value_Eor */
313 * Return the value of a Not.
315 static tarval *computed_value_Not(const ir_node *n) {
316 ir_node *a = get_Not_op(n);
317 tarval *ta = value_of(a);
319 if (ta != tarval_bad)
320 return tarval_not(ta);
323 } /* computed_value_Not */
326 * Return the value of a Shl.
328 static tarval *computed_value_Shl(const ir_node *n) {
329 ir_node *a = get_Shl_left(n);
330 ir_node *b = get_Shl_right(n);
332 tarval *ta = value_of(a);
333 tarval *tb = value_of(b);
335 if ((ta != tarval_bad) && (tb != tarval_bad)) {
336 return tarval_shl (ta, tb);
339 } /* computed_value_Shl */
342 * Return the value of a Shr.
344 static tarval *computed_value_Shr(const ir_node *n) {
345 ir_node *a = get_Shr_left(n);
346 ir_node *b = get_Shr_right(n);
348 tarval *ta = value_of(a);
349 tarval *tb = value_of(b);
351 if ((ta != tarval_bad) && (tb != tarval_bad)) {
352 return tarval_shr (ta, tb);
355 } /* computed_value_Shr */
358 * Return the value of a Shrs.
360 static tarval *computed_value_Shrs(const ir_node *n) {
361 ir_node *a = get_Shrs_left(n);
362 ir_node *b = get_Shrs_right(n);
364 tarval *ta = value_of(a);
365 tarval *tb = value_of(b);
367 if ((ta != tarval_bad) && (tb != tarval_bad)) {
368 return tarval_shrs (ta, tb);
371 } /* computed_value_Shrs */
374 * Return the value of a Rotl.
376 static tarval *computed_value_Rotl(const ir_node *n) {
377 ir_node *a = get_Rotl_left(n);
378 ir_node *b = get_Rotl_right(n);
380 tarval *ta = value_of(a);
381 tarval *tb = value_of(b);
383 if ((ta != tarval_bad) && (tb != tarval_bad)) {
384 return tarval_rotl(ta, tb);
387 } /* computed_value_Rotl */
390 * Return the value of a Conv.
392 static tarval *computed_value_Conv(const ir_node *n) {
393 ir_node *a = get_Conv_op(n);
394 tarval *ta = value_of(a);
396 if (ta != tarval_bad)
397 return tarval_convert_to(ta, get_irn_mode(n));
400 } /* computed_value_Conv */
403 * Calculate the value of a Mux: can be evaluated, if the
404 * sel and the right input are known.
406 static tarval *computed_value_Mux(const ir_node *n) {
407 ir_node *sel = get_Mux_sel(n);
408 tarval *ts = value_of(sel);
410 if (ts == get_tarval_b_true()) {
411 ir_node *v = get_Mux_true(n);
414 else if (ts == get_tarval_b_false()) {
415 ir_node *v = get_Mux_false(n);
419 } /* computed_value_Mux */
422 * Calculate the value of a Confirm: can be evaluated,
423 * if it has the form Confirm(x, '=', Const).
425 static tarval *computed_value_Confirm(const ir_node *n) {
427 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
428 * Do NOT optimize them away (CondEval wants them), so wait until
429 * remove_confirm is activated.
431 if (get_opt_remove_confirm()) {
432 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
433 tarval *tv = value_of(get_Confirm_bound(n));
434 if (tv != tarval_bad)
438 return value_of(get_Confirm_value(n));
439 } /* computed_value_Confirm */
442 * Return the value of a Proj(Cmp).
444 * This performs a first step of unreachable code elimination.
445 * Proj can not be computed, but folding a Cmp above the Proj here is
446 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
448 * There are several case where we can evaluate a Cmp node, see later.
450 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
451 ir_node *a = get_Proj_pred(n);
452 ir_node *aa = get_Cmp_left(a);
453 ir_node *ab = get_Cmp_right(a);
454 long proj_nr = get_Proj_proj(n);
457 * BEWARE: a == a is NOT always True for floating Point values, as
458 * NaN != NaN is defined, so we must check this here.
461 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
464 /* This is a trick with the bits used for encoding the Cmp
465 Proj numbers, the following statement is not the same:
466 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
467 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
470 tarval *taa = value_of(aa);
471 tarval *tab = value_of(ab);
472 ir_mode *mode = get_irn_mode(aa);
475 * The predecessors of Cmp are target values. We can evaluate
478 if ((taa != tarval_bad) && (tab != tarval_bad)) {
479 /* strange checks... */
480 pn_Cmp flags = tarval_cmp(taa, tab);
481 if (flags != pn_Cmp_False) {
482 return new_tarval_from_long (proj_nr & flags, mode_b);
485 /* for integer values, we can check against MIN/MAX */
486 else if (mode_is_int(mode)) {
487 /* MIN <=/> x. This results in true/false. */
488 if (taa == get_mode_min(mode)) {
489 /* a compare with the MIN value */
490 if (proj_nr == pn_Cmp_Le)
491 return get_tarval_b_true();
492 else if (proj_nr == pn_Cmp_Gt)
493 return get_tarval_b_false();
495 /* x >=/< MIN. This results in true/false. */
497 if (tab == get_mode_min(mode)) {
498 /* a compare with the MIN value */
499 if (proj_nr == pn_Cmp_Ge)
500 return get_tarval_b_true();
501 else if (proj_nr == pn_Cmp_Lt)
502 return get_tarval_b_false();
504 /* MAX >=/< x. This results in true/false. */
505 else if (taa == get_mode_max(mode)) {
506 if (proj_nr == pn_Cmp_Ge)
507 return get_tarval_b_true();
508 else if (proj_nr == pn_Cmp_Lt)
509 return get_tarval_b_false();
511 /* x <=/> MAX. This results in true/false. */
512 else if (tab == get_mode_max(mode)) {
513 if (proj_nr == pn_Cmp_Le)
514 return get_tarval_b_true();
515 else if (proj_nr == pn_Cmp_Gt)
516 return get_tarval_b_false();
520 * The predecessors are Allocs or (void*)(0) constants. Allocs never
521 * return NULL, they raise an exception. Therefore we can predict
525 ir_node *aaa = skip_Proj(aa);
526 ir_node *aba = skip_Proj(ab);
528 if ( ( (/* aa is ProjP and aaa is Alloc */
530 && mode_is_reference(get_irn_mode(aa))
532 && ( (/* ab is NULL */
533 mode_is_reference(get_irn_mode(ab))
534 && tarval_is_null(tab))
535 || (/* ab is other Alloc */
537 && mode_is_reference(get_irn_mode(ab))
540 || (/* aa is NULL and aba is Alloc */
541 mode_is_reference(get_irn_mode(aa))
542 && tarval_is_null(taa)
544 && mode_is_reference(get_irn_mode(ab))
547 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
550 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
551 } /* computed_value_Proj_Cmp */
554 * Return the value of a floating point Quot.
556 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
557 tarval *ta = value_of(a);
558 tarval *tb = value_of(b);
560 /* cannot optimize 0 / b = 0 because of NaN */
561 if (ta != tarval_bad && tb != tarval_bad)
562 return tarval_quo(ta, tb);
564 } /* do_computed_value_Quot */
567 * Calculate the value of an integer Div of two nodes.
568 * Special case: 0 / b
570 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
571 tarval *ta = value_of(a);
573 const ir_node *dummy;
575 /* Compute c1 / c2 or 0 / a, a != 0 */
576 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
577 return ta; /* 0 / b == 0 */
579 if (ta != tarval_bad && tb != tarval_bad)
580 return tarval_div(ta, tb);
582 } /* do_computed_value_Div */
585 * Calculate the value of an integer Mod of two nodes.
586 * Special case: a % 1
588 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
589 tarval *ta = value_of(a);
590 tarval *tb = value_of(b);
592 /* Compute a % 1 or c1 % c2 */
593 if (tarval_is_one(tb))
594 return get_mode_null(get_irn_mode(a));
595 if (ta != tarval_bad && tb != tarval_bad)
596 return tarval_mod(ta, tb);
598 } /* do_computed_value_Mod */
601 * Return the value of a Proj(DivMod).
603 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
604 long proj_nr = get_Proj_proj(n);
606 /* compute either the Div or the Mod part */
607 if (proj_nr == pn_DivMod_res_div) {
608 const ir_node *a = get_Proj_pred(n);
609 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
610 } else if (proj_nr == pn_DivMod_res_mod) {
611 const ir_node *a = get_Proj_pred(n);
612 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
615 } /* computed_value_Proj_DivMod */
618 * Return the value of a Proj(Div).
620 static tarval *computed_value_Proj_Div(const ir_node *n) {
621 long proj_nr = get_Proj_proj(n);
623 if (proj_nr == pn_Div_res) {
624 const ir_node *a = get_Proj_pred(n);
625 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
628 } /* computed_value_Proj_Div */
631 * Return the value of a Proj(Mod).
633 static tarval *computed_value_Proj_Mod(const ir_node *n) {
634 long proj_nr = get_Proj_proj(n);
636 if (proj_nr == pn_Mod_res) {
637 const ir_node *a = get_Proj_pred(n);
638 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
641 } /* computed_value_Proj_Mod */
644 * Return the value of a Proj(Quot).
646 static tarval *computed_value_Proj_Quot(const ir_node *n) {
647 long proj_nr = get_Proj_proj(n);
649 if (proj_nr == pn_Quot_res) {
650 const ir_node *a = get_Proj_pred(n);
651 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
654 } /* computed_value_Proj_Quot */
657 * Return the value of a Proj.
659 static tarval *computed_value_Proj(const ir_node *proj) {
660 ir_node *n = get_Proj_pred(proj);
662 if (n->op->ops.computed_value_Proj != NULL)
663 return n->op->ops.computed_value_Proj(proj);
665 } /* computed_value_Proj */
668 * If the parameter n can be computed, return its value, else tarval_bad.
669 * Performs constant folding.
671 * @param n The node this should be evaluated
673 tarval *computed_value(const ir_node *n) {
674 if (n->op->ops.computed_value)
675 return n->op->ops.computed_value(n);
677 } /* computed_value */
680 * Set the default computed_value evaluator in an ir_op_ops.
682 * @param code the opcode for the default operation
683 * @param ops the operations initialized
688 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
692 ops->computed_value = computed_value_##a; \
694 #define CASE_PROJ(a) \
696 ops->computed_value_Proj = computed_value_Proj_##a; \
733 } /* firm_set_default_computed_value */
736 * Returns a equivalent block for another block.
737 * If the block has only one predecessor, this is
738 * the equivalent one. If the only predecessor of a block is
739 * the block itself, this is a dead block.
741 * If both predecessors of a block are the branches of a binary
742 * Cond, the equivalent block is Cond's block.
744 * If all predecessors of a block are bad or lies in a dead
745 * block, the current block is dead as well.
747 * Note, that blocks are NEVER turned into Bad's, instead
748 * the dead_block flag is set. So, never test for is_Bad(block),
749 * always use is_dead_Block(block).
751 static ir_node *equivalent_node_Block(ir_node *n)
756 /* don't optimize dead blocks */
757 if (is_Block_dead(n))
760 n_preds = get_Block_n_cfgpreds(n);
762 /* The Block constructor does not call optimize, but mature_immBlock()
763 calls the optimization. */
764 assert(get_Block_matured(n));
766 /* Straightening: a single entry Block following a single exit Block
767 can be merged, if it is not the Start block. */
768 /* !!! Beware, all Phi-nodes of n must have been optimized away.
769 This should be true, as the block is matured before optimize is called.
770 But what about Phi-cycles with the Phi0/Id that could not be resolved?
771 Remaining Phi nodes are just Ids. */
773 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
776 ir_node *predblock = get_nodes_block(pred);
777 if (predblock == oldn) {
778 /* Jmp jumps into the block it is in -- deal self cycle. */
779 n = set_Block_dead(n);
780 DBG_OPT_DEAD_BLOCK(oldn, n);
781 } else if (get_opt_control_flow_straightening()) {
783 DBG_OPT_STG(oldn, n);
785 } else if (is_Cond(pred)) {
786 ir_node *predblock = get_nodes_block(pred);
787 if (predblock == oldn) {
788 /* Jmp jumps into the block it is in -- deal self cycle. */
789 n = set_Block_dead(n);
790 DBG_OPT_DEAD_BLOCK(oldn, n);
793 } else if ((n_preds == 2) &&
794 (get_opt_control_flow_weak_simplification())) {
795 /* Test whether Cond jumps twice to this block
796 * The more general case which more than 2 predecessors is handles
797 * in optimize_cf(), we handle only this special case for speed here.
799 ir_node *a = get_Block_cfgpred(n, 0);
800 ir_node *b = get_Block_cfgpred(n, 1);
802 if (is_Proj(a) && is_Proj(b)) {
803 ir_node *cond = get_Proj_pred(a);
805 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
806 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
807 /* Also a single entry Block following a single exit Block. Phis have
808 twice the same operand and will be optimized away. */
809 n = get_nodes_block(cond);
810 DBG_OPT_IFSIM1(oldn, a, b, n);
813 } else if (get_opt_unreachable_code() &&
814 (n != get_irg_start_block(current_ir_graph)) &&
815 (n != get_irg_end_block(current_ir_graph)) ) {
818 /* If all inputs are dead, this block is dead too, except if it is
819 the start or end block. This is one step of unreachable code
821 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
822 ir_node *pred = get_Block_cfgpred(n, i);
825 if (is_Bad(pred)) continue;
826 pred_blk = get_nodes_block(skip_Proj(pred));
828 if (is_Block_dead(pred_blk)) continue;
831 /* really found a living input */
836 n = set_Block_dead(n);
837 DBG_OPT_DEAD_BLOCK(oldn, n);
842 } /* equivalent_node_Block */
845 * Returns a equivalent node for a Jmp, a Bad :-)
846 * Of course this only happens if the Block of the Jmp is dead.
848 static ir_node *equivalent_node_Jmp(ir_node *n) {
851 /* unreachable code elimination */
852 if (is_Block_dead(get_nodes_block(n))) {
853 n = get_irg_bad(current_ir_graph);
854 DBG_OPT_DEAD_BLOCK(oldn, n);
857 } /* equivalent_node_Jmp */
859 /** Raise is handled in the same way as Jmp. */
860 #define equivalent_node_Raise equivalent_node_Jmp
863 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
864 See transform_node_Proj_Cond(). */
867 * Optimize operations that are commutative and have neutral 0,
868 * so a op 0 = 0 op a = a.
870 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
873 ir_node *a = get_binop_left(n);
874 ir_node *b = get_binop_right(n);
879 /* After running compute_node there is only one constant predecessor.
880 Find this predecessors value and remember the other node: */
881 if ((tv = value_of(a)) != tarval_bad) {
883 } else if ((tv = value_of(b)) != tarval_bad) {
888 /* If this predecessors constant value is zero, the operation is
889 * unnecessary. Remove it.
891 * Beware: If n is a Add, the mode of on and n might be different
892 * which happens in this rare construction: NULL + 3.
893 * Then, a Conv would be needed which we cannot include here.
895 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
898 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
902 } /* equivalent_node_neutral_zero */
905 * Eor is commutative and has neutral 0.
907 static ir_node *equivalent_node_Eor(ir_node *n) {
912 n = equivalent_node_neutral_zero(n);
913 if (n != oldn) return n;
916 b = get_Eor_right(n);
919 ir_node *aa = get_Eor_left(a);
920 ir_node *ab = get_Eor_right(a);
923 /* (a ^ b) ^ a -> b */
925 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
927 } else if (ab == b) {
928 /* (a ^ b) ^ b -> a */
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
935 ir_node *ba = get_Eor_left(b);
936 ir_node *bb = get_Eor_right(b);
939 /* a ^ (a ^ b) -> b */
941 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
943 } else if (bb == a) {
944 /* a ^ (b ^ a) -> b */
946 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
954 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
956 * The second one looks strange, but this construct
957 * is used heavily in the LCC sources :-).
959 * Beware: The Mode of an Add may be different than the mode of its
960 * predecessors, so we could not return a predecessors in all cases.
962 static ir_node *equivalent_node_Add(ir_node *n) {
964 ir_node *left, *right;
965 ir_mode *mode = get_irn_mode(n);
967 n = equivalent_node_neutral_zero(n);
971 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
972 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
975 left = get_Add_left(n);
976 right = get_Add_right(n);
979 if (get_Sub_right(left) == right) {
982 n = get_Sub_left(left);
983 if (mode == get_irn_mode(n)) {
984 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
990 if (get_Sub_right(right) == left) {
993 n = get_Sub_left(right);
994 if (mode == get_irn_mode(n)) {
995 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1001 } /* equivalent_node_Add */
1004 * optimize operations that are not commutative but have neutral 0 on left,
1007 static ir_node *equivalent_node_left_zero(ir_node *n) {
1010 ir_node *a = get_binop_left(n);
1011 ir_node *b = get_binop_right(n);
1012 tarval *tb = value_of(b);
1014 if (tarval_is_null(tb)) {
1017 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1020 } /* equivalent_node_left_zero */
1022 #define equivalent_node_Shl equivalent_node_left_zero
1023 #define equivalent_node_Shr equivalent_node_left_zero
1024 #define equivalent_node_Shrs equivalent_node_left_zero
1025 #define equivalent_node_Rotl equivalent_node_left_zero
1028 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1030 * The second one looks strange, but this construct
1031 * is used heavily in the LCC sources :-).
1033 * Beware: The Mode of a Sub may be different than the mode of its
1034 * predecessors, so we could not return a predecessors in all cases.
1036 static ir_node *equivalent_node_Sub(ir_node *n) {
1039 ir_mode *mode = get_irn_mode(n);
1042 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1043 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1046 b = get_Sub_right(n);
1049 /* Beware: modes might be different */
1050 if (tarval_is_null(tb)) {
1051 ir_node *a = get_Sub_left(n);
1052 if (mode == get_irn_mode(a)) {
1055 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1059 } /* equivalent_node_Sub */
1063 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1066 * -(-a) == a, but might overflow two times.
1067 * We handle it anyway here but the better way would be a
1068 * flag. This would be needed for Pascal for instance.
1070 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1072 ir_node *pred = get_unop_op(n);
1074 /* optimize symmetric unop */
1075 if (get_irn_op(pred) == get_irn_op(n)) {
1076 n = get_unop_op(pred);
1077 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1080 } /* equivalent_node_idempotent_unop */
1082 /** Optimize Not(Not(x)) == x. */
1083 #define equivalent_node_Not equivalent_node_idempotent_unop
1085 /** -(-x) == x ??? Is this possible or can --x raise an
1086 out of bounds exception if min =! max? */
1087 #define equivalent_node_Minus equivalent_node_idempotent_unop
1090 * Optimize a * 1 = 1 * a = a.
1092 static ir_node *equivalent_node_Mul(ir_node *n) {
1094 ir_node *a = get_Mul_left(n);
1096 /* we can handle here only the n * n = n bit cases */
1097 if (get_irn_mode(n) == get_irn_mode(a)) {
1098 ir_node *b = get_Mul_right(n);
1102 * Mul is commutative and has again an other neutral element.
1103 * Constants are place right, so check this case first.
1106 if (tarval_is_one(tv)) {
1108 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1111 if (tarval_is_one(tv)) {
1113 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1118 } /* equivalent_node_Mul */
1121 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1123 static ir_node *equivalent_node_Or(ir_node *n) {
1126 ir_node *a = get_Or_left(n);
1127 ir_node *b = get_Or_right(n);
1131 n = a; /* Or has it's own neutral element */
1132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1135 /* constants are cormalized to right, check this site first */
1137 if (tarval_is_null(tv)) {
1139 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1143 if (tarval_is_null(tv)) {
1145 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1150 } /* equivalent_node_Or */
1153 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1155 static ir_node *equivalent_node_And(ir_node *n) {
1158 ir_node *a = get_And_left(n);
1159 ir_node *b = get_And_right(n);
1163 n = a; /* And has it's own neutral element */
1164 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1167 /* constants are cormalized to right, check this site first */
1169 if (tarval_is_all_one(tv)) {
1171 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1175 if (tarval_is_all_one(tv)) {
1177 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1181 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1184 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1189 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1192 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1197 } /* equivalent_node_And */
1200 * Try to remove useless Conv's:
1202 static ir_node *equivalent_node_Conv(ir_node *n) {
1204 ir_node *a = get_Conv_op(n);
1206 ir_mode *n_mode = get_irn_mode(n);
1207 ir_mode *a_mode = get_irn_mode(a);
1210 if (n_mode == a_mode) { /* No Conv necessary */
1211 if (get_Conv_strict(n)) {
1212 /* special case: the predecessor might be a also a Conv */
1214 if (! get_Conv_strict(a)) {
1215 /* first one is not strict, kick it */
1217 a_mode = get_irn_mode(a);
1221 /* else both are strict conv, second is superfluous */
1224 ir_node *pred = get_Proj_pred(a);
1225 if (is_Load(pred)) {
1226 /* loads always return with the exact precision of n_mode */
1227 assert(get_Load_mode(pred) == n_mode);
1231 /* leave strict floating point Conv's */
1236 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1237 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1238 ir_node *b = get_Conv_op(a);
1239 ir_mode *b_mode = get_irn_mode(b);
1241 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1242 /* both are strict conv */
1243 if (smaller_mode(a_mode, n_mode)) {
1244 /* both are strict, but the first is smaller, so
1245 the second cannot remove more precision, remove the
1247 set_Conv_strict(n, 0);
1250 if (n_mode == b_mode) {
1251 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1252 if (n_mode == mode_b) {
1253 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1254 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1256 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1257 if (values_in_mode(b_mode, a_mode)) {
1258 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1259 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1264 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1265 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1266 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1267 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1269 if (float_mantissa >= int_mantissa) {
1271 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1276 if (smaller_mode(b_mode, a_mode)) {
1277 if (get_Conv_strict(n))
1278 set_Conv_strict(b, 1);
1279 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1280 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1287 } /* equivalent_node_Conv */
1290 * A Cast may be removed if the type of the previous node
1291 * is already the type of the Cast.
1293 static ir_node *equivalent_node_Cast(ir_node *n) {
1295 ir_node *pred = get_Cast_op(n);
1297 if (get_irn_type(pred) == get_Cast_type(n)) {
1299 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1302 } /* equivalent_node_Cast */
1305 * - fold Phi-nodes, iff they have only one predecessor except
1308 static ir_node *equivalent_node_Phi(ir_node *n) {
1313 ir_node *first_val = NULL; /* to shutup gcc */
1315 if (!get_opt_normalize()) return n;
1317 n_preds = get_Phi_n_preds(n);
1319 block = get_nodes_block(n);
1320 if (is_Block_dead(block)) /* Control dead */
1321 return get_irg_bad(current_ir_graph);
1323 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1325 /* Find first non-self-referencing input */
1326 for (i = 0; i < n_preds; ++i) {
1327 first_val = get_Phi_pred(n, i);
1328 if ( (first_val != n) /* not self pointer */
1330 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1331 * predecessors. Then, Phi nodes in dead code might be removed, causing
1332 * nodes pointing to themself (Add's for instance).
1333 * This is really bad and causes endless recursions in several
1334 * code pathes, so we do NOT optimize such a code.
1335 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1336 * (and bad Phi predecessors), so live code is optimized later.
1338 && (! is_Bad(get_Block_cfgpred(block, i)))
1340 ) { /* value not dead */
1341 break; /* then found first value. */
1346 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1347 return get_irg_bad(current_ir_graph);
1350 /* search for rest of inputs, determine if any of these
1351 are non-self-referencing */
1352 while (++i < n_preds) {
1353 ir_node *scnd_val = get_Phi_pred(n, i);
1354 if ( (scnd_val != n)
1355 && (scnd_val != first_val)
1358 && (! is_Bad(get_Block_cfgpred(block, i)))
1366 /* Fold, if no multiple distinct non-self-referencing inputs */
1368 DBG_OPT_PHI(oldn, n);
1371 } /* equivalent_node_Phi */
1374 * Several optimizations:
1375 * - fold Sync-nodes, iff they have only one predecessor except
1378 static ir_node *equivalent_node_Sync(ir_node *n) {
1379 int arity = get_Sync_n_preds(n);
1382 for (i = 0; i < arity;) {
1383 ir_node *pred = get_Sync_pred(n, i);
1386 /* Remove Bad predecessors */
1393 /* Remove duplicate predecessors */
1399 if (get_Sync_pred(n, j) == pred) {
1407 if (arity == 0) return get_irg_bad(current_ir_graph);
1408 if (arity == 1) return get_Sync_pred(n, 0);
1410 } /* equivalent_node_Sync */
1413 * Optimize Proj(Tuple).
1415 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1416 ir_node *oldn = proj;
1417 ir_node *tuple = get_Proj_pred(proj);
1419 /* Remove the Tuple/Proj combination. */
1420 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1421 DBG_OPT_TUPLE(oldn, tuple, proj);
1424 } /* equivalent_node_Proj_Tuple */
1427 * Optimize a / 1 = a.
1429 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1430 ir_node *oldn = proj;
1431 ir_node *div = get_Proj_pred(proj);
1432 ir_node *b = get_Div_right(div);
1433 tarval *tb = value_of(b);
1435 /* Div is not commutative. */
1436 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1437 switch (get_Proj_proj(proj)) {
1439 proj = get_Div_mem(div);
1440 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1444 proj = get_Div_left(div);
1445 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1449 /* we cannot replace the exception Proj's here, this is done in
1450 transform_node_Proj_Div() */
1455 } /* equivalent_node_Proj_Div */
1458 * Optimize a / 1.0 = a.
1460 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1461 ir_node *oldn = proj;
1462 ir_node *quot = get_Proj_pred(proj);
1463 ir_node *b = get_Quot_right(quot);
1464 tarval *tb = value_of(b);
1466 /* Div is not commutative. */
1467 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1468 switch (get_Proj_proj(proj)) {
1470 proj = get_Quot_mem(quot);
1471 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1475 proj = get_Quot_left(quot);
1476 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1480 /* we cannot replace the exception Proj's here, this is done in
1481 transform_node_Proj_Quot() */
1486 } /* equivalent_node_Proj_Quot */
1489 * Optimize a / 1 = a.
1491 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1492 ir_node *oldn = proj;
1493 ir_node *divmod = get_Proj_pred(proj);
1494 ir_node *b = get_DivMod_right(divmod);
1495 tarval *tb = value_of(b);
1497 /* Div is not commutative. */
1498 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1499 switch (get_Proj_proj(proj)) {
1501 proj = get_DivMod_mem(divmod);
1502 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1505 case pn_DivMod_res_div:
1506 proj = get_DivMod_left(divmod);
1507 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1511 /* we cannot replace the exception Proj's here, this is done in
1512 transform_node_Proj_DivMod().
1513 Note further that the pn_DivMod_res_div case is handled in
1514 computed_value_Proj(). */
1519 } /* equivalent_node_Proj_DivMod */
1522 * Optimize CopyB(mem, x, x) into a Nop.
1524 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1525 ir_node *oldn = proj;
1526 ir_node *copyb = get_Proj_pred(proj);
1527 ir_node *a = get_CopyB_dst(copyb);
1528 ir_node *b = get_CopyB_src(copyb);
1531 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1532 switch (get_Proj_proj(proj)) {
1533 case pn_CopyB_M_regular:
1534 proj = get_CopyB_mem(copyb);
1535 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1538 case pn_CopyB_M_except:
1539 case pn_CopyB_X_except:
1540 DBG_OPT_EXC_REM(proj);
1541 proj = get_irg_bad(current_ir_graph);
1546 } /* equivalent_node_Proj_CopyB */
1549 * Optimize Bounds(idx, idx, upper) into idx.
1551 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1552 ir_node *oldn = proj;
1553 ir_node *bound = get_Proj_pred(proj);
1554 ir_node *idx = get_Bound_index(bound);
1555 ir_node *pred = skip_Proj(idx);
1558 if (idx == get_Bound_lower(bound))
1560 else if (is_Bound(pred)) {
1562 * idx was Bounds checked in the same MacroBlock previously,
1563 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1565 ir_node *lower = get_Bound_lower(bound);
1566 ir_node *upper = get_Bound_upper(bound);
1567 if (get_Bound_lower(pred) == lower &&
1568 get_Bound_upper(pred) == upper &&
1569 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1571 * One could expect that we simply return the previous
1572 * Bound here. However, this would be wrong, as we could
1573 * add an exception Proj to a new location then.
1574 * So, we must turn in into a tuple.
1580 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1581 switch (get_Proj_proj(proj)) {
1583 DBG_OPT_EXC_REM(proj);
1584 proj = get_Bound_mem(bound);
1586 case pn_Bound_X_except:
1587 DBG_OPT_EXC_REM(proj);
1588 proj = get_irg_bad(current_ir_graph);
1592 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1595 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1600 } /* equivalent_node_Proj_Bound */
1603 * Optimize an Exception Proj(Load) with a non-null address.
1605 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1606 if (get_opt_ldst_only_null_ptr_exceptions()) {
1607 if (get_irn_mode(proj) == mode_X) {
1608 ir_node *load = get_Proj_pred(proj);
1610 /* get the Load address */
1611 const ir_node *addr = get_Load_ptr(load);
1612 const ir_node *confirm;
1614 if (value_not_null(addr, &confirm)) {
1615 if (get_Proj_proj(proj) == pn_Load_X_except) {
1616 DBG_OPT_EXC_REM(proj);
1617 return get_irg_bad(current_ir_graph);
1623 } /* equivalent_node_Proj_Load */
1626 * Optimize an Exception Proj(Store) with a non-null address.
1628 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1629 if (get_opt_ldst_only_null_ptr_exceptions()) {
1630 if (get_irn_mode(proj) == mode_X) {
1631 ir_node *store = get_Proj_pred(proj);
1633 /* get the load/store address */
1634 const ir_node *addr = get_Store_ptr(store);
1635 const ir_node *confirm;
1637 if (value_not_null(addr, &confirm)) {
1638 if (get_Proj_proj(proj) == pn_Store_X_except) {
1639 DBG_OPT_EXC_REM(proj);
1640 return get_irg_bad(current_ir_graph);
1646 } /* equivalent_node_Proj_Store */
1649 * Does all optimizations on nodes that must be done on it's Proj's
1650 * because of creating new nodes.
1652 static ir_node *equivalent_node_Proj(ir_node *proj) {
1653 ir_node *n = get_Proj_pred(proj);
1655 if (get_irn_mode(proj) == mode_X) {
1656 if (is_Block_dead(get_nodes_block(n))) {
1657 /* Remove dead control flow -- early gigo(). */
1658 return get_irg_bad(current_ir_graph);
1661 if (n->op->ops.equivalent_node_Proj)
1662 return n->op->ops.equivalent_node_Proj(proj);
1664 } /* equivalent_node_Proj */
1669 static ir_node *equivalent_node_Id(ir_node *n) {
1676 DBG_OPT_ID(oldn, n);
1678 } /* equivalent_node_Id */
1683 static ir_node *equivalent_node_Mux(ir_node *n)
1685 ir_node *oldn = n, *sel = get_Mux_sel(n);
1686 tarval *ts = value_of(sel);
1688 /* Mux(true, f, t) == t */
1689 if (ts == tarval_b_true) {
1690 n = get_Mux_true(n);
1691 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1693 /* Mux(false, f, t) == f */
1694 else if (ts == tarval_b_false) {
1695 n = get_Mux_false(n);
1696 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1698 /* Mux(v, x, x) == x */
1699 else if (get_Mux_false(n) == get_Mux_true(n)) {
1700 n = get_Mux_true(n);
1701 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1703 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1704 ir_node *cmp = get_Proj_pred(sel);
1705 long proj_nr = get_Proj_proj(sel);
1706 ir_node *f = get_Mux_false(n);
1707 ir_node *t = get_Mux_true(n);
1710 * Note further that these optimization work even for floating point
1711 * with NaN's because -NaN == NaN.
1712 * However, if +0 and -0 is handled differently, we cannot use the first one.
1715 ir_node *const cmp_l = get_Cmp_left(cmp);
1716 ir_node *const cmp_r = get_Cmp_right(cmp);
1720 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1721 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1723 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1730 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1731 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1733 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1740 * Note: normalization puts the constant on the right side,
1741 * so we check only one case.
1743 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1744 /* Mux(t CMP 0, X, t) */
1745 if (is_Minus(f) && get_Minus_op(f) == t) {
1746 /* Mux(t CMP 0, -t, t) */
1747 if (proj_nr == pn_Cmp_Eq) {
1748 /* Mux(t == 0, -t, t) ==> -t */
1750 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1751 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1752 /* Mux(t != 0, -t, t) ==> t */
1754 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1761 } /* equivalent_node_Mux */
1764 * Remove Confirm nodes if setting is on.
1765 * Replace Confirms(x, '=', Constlike) by Constlike.
1767 static ir_node *equivalent_node_Confirm(ir_node *n) {
1768 ir_node *pred = get_Confirm_value(n);
1769 pn_Cmp pnc = get_Confirm_cmp(n);
1771 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1773 * rare case: two identical Confirms one after another,
1774 * replace the second one with the first.
1777 pred = get_Confirm_value(n);
1778 pnc = get_Confirm_cmp(n);
1780 if (get_opt_remove_confirm())
1781 return get_Confirm_value(n);
1786 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1787 * perform no actual computation, as, e.g., the Id nodes. It does not create
1788 * new nodes. It is therefore safe to free n if the node returned is not n.
1789 * If a node returns a Tuple we can not just skip it. If the size of the
1790 * in array fits, we transform n into a tuple (e.g., Div).
1792 ir_node *equivalent_node(ir_node *n) {
1793 if (n->op->ops.equivalent_node)
1794 return n->op->ops.equivalent_node(n);
1796 } /* equivalent_node */
1799 * Sets the default equivalent node operation for an ir_op_ops.
1801 * @param code the opcode for the default operation
1802 * @param ops the operations initialized
1807 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1811 ops->equivalent_node = equivalent_node_##a; \
1813 #define CASE_PROJ(a) \
1815 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1857 } /* firm_set_default_equivalent_node */
1860 * Returns non-zero if a node is a Phi node
1861 * with all predecessors constant.
1863 static int is_const_Phi(ir_node *n) {
1866 if (! is_Phi(n) || get_irn_arity(n) == 0)
1868 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1869 if (! is_Const(get_irn_n(n, i)))
1873 } /* is_const_Phi */
1875 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1876 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1879 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1881 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1882 if (eval == tarval_sub) {
1883 tarval_sub_type func = (tarval_sub_type)eval;
1885 return func(a, b, mode);
1887 tarval_binop_type func = (tarval_binop_type)eval;
1894 * Apply an evaluator on a binop with a constant operators (and one Phi).
1896 * @param phi the Phi node
1897 * @param other the other operand
1898 * @param eval an evaluator function
1899 * @param mode the mode of the result, may be different from the mode of the Phi!
1900 * @param left if non-zero, other is the left operand, else the right
1902 * @return a new Phi node if the conversion was successful, NULL else
1904 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1909 int i, n = get_irn_arity(phi);
1911 NEW_ARR_A(void *, res, n);
1913 for (i = 0; i < n; ++i) {
1914 pred = get_irn_n(phi, i);
1915 tv = get_Const_tarval(pred);
1916 tv = do_eval(eval, other, tv, mode);
1918 if (tv == tarval_bad) {
1919 /* folding failed, bad */
1925 for (i = 0; i < n; ++i) {
1926 pred = get_irn_n(phi, i);
1927 tv = get_Const_tarval(pred);
1928 tv = do_eval(eval, tv, other, mode);
1930 if (tv == tarval_bad) {
1931 /* folding failed, bad */
1937 irg = current_ir_graph;
1938 for (i = 0; i < n; ++i) {
1939 pred = get_irn_n(phi, i);
1940 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1941 mode, res[i], get_Const_type(pred));
1943 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1944 } /* apply_binop_on_phi */
1947 * Apply an evaluator on a binop with two constant Phi.
1949 * @param a the left Phi node
1950 * @param b the right Phi node
1951 * @param eval an evaluator function
1952 * @param mode the mode of the result, may be different from the mode of the Phi!
1954 * @return a new Phi node if the conversion was successful, NULL else
1956 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1957 tarval *tv_l, *tv_r, *tv;
1963 if (get_nodes_block(a) != get_nodes_block(b))
1966 n = get_irn_arity(a);
1967 NEW_ARR_A(void *, res, n);
1969 for (i = 0; i < n; ++i) {
1970 pred = get_irn_n(a, i);
1971 tv_l = get_Const_tarval(pred);
1972 pred = get_irn_n(b, i);
1973 tv_r = get_Const_tarval(pred);
1974 tv = do_eval(eval, tv_l, tv_r, mode);
1976 if (tv == tarval_bad) {
1977 /* folding failed, bad */
1982 irg = current_ir_graph;
1983 for (i = 0; i < n; ++i) {
1984 pred = get_irn_n(a, i);
1985 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1987 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1988 } /* apply_binop_on_2_phis */
1991 * Apply an evaluator on a unop with a constant operator (a Phi).
1993 * @param phi the Phi node
1994 * @param eval an evaluator function
1996 * @return a new Phi node if the conversion was successful, NULL else
1998 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2004 int i, n = get_irn_arity(phi);
2006 NEW_ARR_A(void *, res, n);
2007 for (i = 0; i < n; ++i) {
2008 pred = get_irn_n(phi, i);
2009 tv = get_Const_tarval(pred);
2012 if (tv == tarval_bad) {
2013 /* folding failed, bad */
2018 mode = get_irn_mode(phi);
2019 irg = current_ir_graph;
2020 for (i = 0; i < n; ++i) {
2021 pred = get_irn_n(phi, i);
2022 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2023 mode, res[i], get_Const_type(pred));
2025 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2026 } /* apply_unop_on_phi */
2029 * Apply a conversion on a constant operator (a Phi).
2031 * @param phi the Phi node
2033 * @return a new Phi node if the conversion was successful, NULL else
2035 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2040 int i, n = get_irn_arity(phi);
2042 NEW_ARR_A(void *, res, n);
2043 for (i = 0; i < n; ++i) {
2044 pred = get_irn_n(phi, i);
2045 tv = get_Const_tarval(pred);
2046 tv = tarval_convert_to(tv, mode);
2048 if (tv == tarval_bad) {
2049 /* folding failed, bad */
2054 irg = current_ir_graph;
2055 for (i = 0; i < n; ++i) {
2056 pred = get_irn_n(phi, i);
2057 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2058 mode, res[i], get_Const_type(pred));
2060 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2061 } /* apply_conv_on_phi */
2064 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2065 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2066 * If possible, remove the Conv's.
2068 static ir_node *transform_node_AddSub(ir_node *n) {
2069 ir_mode *mode = get_irn_mode(n);
2071 if (mode_is_reference(mode)) {
2072 ir_node *left = get_binop_left(n);
2073 ir_node *right = get_binop_right(n);
2074 unsigned ref_bits = get_mode_size_bits(mode);
2076 if (is_Conv(left)) {
2077 ir_mode *lmode = get_irn_mode(left);
2078 unsigned bits = get_mode_size_bits(lmode);
2080 if (ref_bits == bits &&
2081 mode_is_int(lmode) &&
2082 get_mode_arithmetic(lmode) == irma_twos_complement) {
2083 ir_node *pre = get_Conv_op(left);
2084 ir_mode *pre_mode = get_irn_mode(pre);
2086 if (mode_is_int(pre_mode) &&
2087 get_mode_size_bits(pre_mode) == bits &&
2088 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2089 /* ok, this conv just changes to sign, moreover the calculation
2090 * is done with same number of bits as our address mode, so
2091 * we can ignore the conv as address calculation can be viewed
2092 * as either signed or unsigned
2094 set_binop_left(n, pre);
2099 if (is_Conv(right)) {
2100 ir_mode *rmode = get_irn_mode(right);
2101 unsigned bits = get_mode_size_bits(rmode);
2103 if (ref_bits == bits &&
2104 mode_is_int(rmode) &&
2105 get_mode_arithmetic(rmode) == irma_twos_complement) {
2106 ir_node *pre = get_Conv_op(right);
2107 ir_mode *pre_mode = get_irn_mode(pre);
2109 if (mode_is_int(pre_mode) &&
2110 get_mode_size_bits(pre_mode) == bits &&
2111 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2112 /* ok, this conv just changes to sign, moreover the calculation
2113 * is done with same number of bits as our address mode, so
2114 * we can ignore the conv as address calculation can be viewed
2115 * as either signed or unsigned
2117 set_binop_right(n, pre);
2122 /* let address arithmetic use unsigned modes */
2123 if (is_Const(right)) {
2124 ir_mode *rmode = get_irn_mode(right);
2126 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2127 /* convert a AddP(P, *s) into AddP(P, *u) */
2128 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2130 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2131 set_binop_right(n, pre);
2137 } /* transform_node_AddSub */
2139 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2141 if (is_Const(b) && is_const_Phi(a)) { \
2142 /* check for Op(Phi, Const) */ \
2143 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2145 else if (is_Const(a) && is_const_Phi(b)) { \
2146 /* check for Op(Const, Phi) */ \
2147 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2149 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2150 /* check for Op(Phi, Phi) */ \
2151 c = apply_binop_on_2_phis(a, b, eval, mode); \
2154 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2158 #define HANDLE_UNOP_PHI(eval, a, c) \
2160 if (is_const_Phi(a)) { \
2161 /* check for Op(Phi) */ \
2162 c = apply_unop_on_phi(a, eval); \
2164 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2170 * Do the AddSub optimization, then Transform
2171 * Constant folding on Phi
2172 * Add(a,a) -> Mul(a, 2)
2173 * Add(Mul(a, x), a) -> Mul(a, x+1)
2174 * if the mode is integer or float.
2175 * Transform Add(a,-b) into Sub(a,b).
2176 * Reassociation might fold this further.
2178 static ir_node *transform_node_Add(ir_node *n) {
2180 ir_node *a, *b, *c, *oldn = n;
2182 n = transform_node_AddSub(n);
2184 a = get_Add_left(n);
2185 b = get_Add_right(n);
2187 mode = get_irn_mode(n);
2189 if (mode_is_reference(mode)) {
2190 ir_mode *lmode = get_irn_mode(a);
2192 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2193 /* an Add(a, NULL) is a hidden Conv */
2194 dbg_info *dbg = get_irn_dbg_info(n);
2195 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2199 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2201 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2202 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2205 if (mode_is_num(mode)) {
2206 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2207 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2208 ir_node *block = get_nodes_block(n);
2211 get_irn_dbg_info(n),
2215 new_r_Const_long(current_ir_graph, block, mode, 2),
2217 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2222 get_irn_dbg_info(n),
2228 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2233 get_irn_dbg_info(n),
2239 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2242 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2243 /* Here we rely on constants be on the RIGHT side */
2245 ir_node *op = get_Not_op(a);
2247 if (is_Const(b) && is_Const_one(b)) {
2249 ir_node *blk = get_nodes_block(n);
2250 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2251 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2256 ir_node *blk = get_nodes_block(n);
2257 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2258 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2263 ir_node *op = get_Not_op(b);
2267 ir_node *blk = get_nodes_block(n);
2268 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2269 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2276 } /* transform_node_Add */
2279 * returns -cnst or NULL if impossible
2281 static ir_node *const_negate(ir_node *cnst) {
2282 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2283 dbg_info *dbgi = get_irn_dbg_info(cnst);
2284 ir_graph *irg = get_irn_irg(cnst);
2285 ir_node *block = get_nodes_block(cnst);
2286 ir_mode *mode = get_irn_mode(cnst);
2287 if (tv == tarval_bad) return NULL;
2288 return new_rd_Const(dbgi, irg, block, mode, tv);
2292 * Do the AddSub optimization, then Transform
2293 * Constant folding on Phi
2294 * Sub(0,a) -> Minus(a)
2295 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2296 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2297 * Sub(Add(a, x), x) -> a
2298 * Sub(x, Add(x, a)) -> -a
2299 * Sub(x, Const) -> Add(x, -Const)
2301 static ir_node *transform_node_Sub(ir_node *n) {
2306 n = transform_node_AddSub(n);
2308 a = get_Sub_left(n);
2309 b = get_Sub_right(n);
2311 mode = get_irn_mode(n);
2313 if (mode_is_int(mode)) {
2314 ir_mode *lmode = get_irn_mode(a);
2316 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2317 /* a Sub(a, NULL) is a hidden Conv */
2318 dbg_info *dbg = get_irn_dbg_info(n);
2319 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2320 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2324 if (mode == lmode &&
2325 get_mode_arithmetic(mode) == irma_twos_complement &&
2327 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2329 dbg_info *dbg = get_irn_dbg_info(n);
2330 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2331 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2337 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2339 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2340 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2343 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2344 /* a - C -> a + (-C) */
2345 ir_node *cnst = const_negate(b);
2347 ir_node *block = get_nodes_block(n);
2348 dbg_info *dbgi = get_irn_dbg_info(n);
2349 ir_graph *irg = get_irn_irg(n);
2351 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2352 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2357 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2358 ir_graph *irg = current_ir_graph;
2359 dbg_info *dbg = get_irn_dbg_info(n);
2360 ir_node *block = get_nodes_block(n);
2361 ir_node *left = get_Minus_op(a);
2362 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2364 n = new_rd_Minus(dbg, irg, block, add, mode);
2365 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2367 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2368 ir_graph *irg = current_ir_graph;
2369 dbg_info *dbg = get_irn_dbg_info(n);
2370 ir_node *block = get_nodes_block(n);
2371 ir_node *right = get_Minus_op(b);
2373 n = new_rd_Add(dbg, irg, block, a, right, mode);
2374 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2376 } else if (is_Sub(b)) {
2377 /* a - (b - c) -> a + (c - b)
2378 * -> (a - b) + c iff (b - c) is a pointer */
2379 ir_graph *irg = current_ir_graph;
2380 dbg_info *s_dbg = get_irn_dbg_info(b);
2381 ir_node *s_block = get_nodes_block(b);
2382 ir_node *s_left = get_Sub_left(b);
2383 ir_node *s_right = get_Sub_right(b);
2384 ir_mode *s_mode = get_irn_mode(b);
2385 if (s_mode == mode_P) {
2386 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2387 dbg_info *a_dbg = get_irn_dbg_info(n);
2388 ir_node *a_block = get_nodes_block(n);
2391 s_right = new_r_Conv(irg, a_block, s_right, mode);
2392 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2394 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2395 dbg_info *a_dbg = get_irn_dbg_info(n);
2396 ir_node *a_block = get_nodes_block(n);
2398 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2400 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2402 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2403 ir_node *m_right = get_Mul_right(b);
2404 if (is_Const(m_right)) {
2405 ir_node *cnst2 = const_negate(m_right);
2406 if (cnst2 != NULL) {
2407 ir_graph *irg = current_ir_graph;
2408 dbg_info *m_dbg = get_irn_dbg_info(b);
2409 ir_node *m_block = get_nodes_block(b);
2410 ir_node *m_left = get_Mul_left(b);
2411 ir_mode *m_mode = get_irn_mode(b);
2412 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2413 dbg_info *a_dbg = get_irn_dbg_info(n);
2414 ir_node *a_block = get_nodes_block(n);
2416 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2417 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2423 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2424 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2426 get_irn_dbg_info(n),
2431 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2435 if (mode_wrap_around(mode)) {
2436 ir_node *left = get_Add_left(a);
2437 ir_node *right = get_Add_right(a);
2439 /* FIXME: Does the Conv's work only for two complement or generally? */
2441 if (mode != get_irn_mode(right)) {
2442 /* This Sub is an effective Cast */
2443 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2446 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2448 } else if (right == b) {
2449 if (mode != get_irn_mode(left)) {
2450 /* This Sub is an effective Cast */
2451 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2454 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2460 if (mode_wrap_around(mode)) {
2461 ir_node *left = get_Add_left(b);
2462 ir_node *right = get_Add_right(b);
2464 /* FIXME: Does the Conv's work only for two complement or generally? */
2466 ir_mode *r_mode = get_irn_mode(right);
2468 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2469 if (mode != r_mode) {
2470 /* This Sub is an effective Cast */
2471 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2473 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2475 } else if (right == a) {
2476 ir_mode *l_mode = get_irn_mode(left);
2478 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2479 if (mode != l_mode) {
2480 /* This Sub is an effective Cast */
2481 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2483 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2488 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2489 ir_mode *mode = get_irn_mode(a);
2491 if (mode == get_irn_mode(b)) {
2493 ir_node *op_a = get_Conv_op(a);
2494 ir_node *op_b = get_Conv_op(b);
2496 /* check if it's allowed to skip the conv */
2497 ma = get_irn_mode(op_a);
2498 mb = get_irn_mode(op_b);
2500 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2501 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2504 set_Sub_right(n, b);
2510 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2511 if (!is_reassoc_running() && is_Mul(a)) {
2512 ir_node *ma = get_Mul_left(a);
2513 ir_node *mb = get_Mul_right(a);
2516 ir_node *blk = get_nodes_block(n);
2518 get_irn_dbg_info(n),
2519 current_ir_graph, blk,
2522 get_irn_dbg_info(n),
2523 current_ir_graph, blk,
2525 new_r_Const_long(current_ir_graph, blk, mode, 1),
2528 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2530 } else if (mb == b) {
2531 ir_node *blk = get_nodes_block(n);
2533 get_irn_dbg_info(n),
2534 current_ir_graph, blk,
2537 get_irn_dbg_info(n),
2538 current_ir_graph, blk,
2540 new_r_Const_long(current_ir_graph, blk, mode, 1),
2543 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2547 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2548 ir_node *x = get_Sub_left(a);
2549 ir_node *y = get_Sub_right(a);
2550 ir_node *blk = get_nodes_block(n);
2551 ir_mode *m_b = get_irn_mode(b);
2552 ir_mode *m_y = get_irn_mode(y);
2556 /* Determine the right mode for the Add. */
2559 else if (mode_is_reference(m_b))
2561 else if (mode_is_reference(m_y))
2565 * Both modes are different but none is reference,
2566 * happens for instance in SubP(SubP(P, Iu), Is).
2567 * We have two possibilities here: Cast or ignore.
2568 * Currently we ignore this case.
2573 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2575 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2576 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2580 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2581 if (is_Const(a) && is_Not(b)) {
2582 /* c - ~X = X + (c+1) */
2583 tarval *tv = get_Const_tarval(a);
2585 tv = tarval_add(tv, get_mode_one(mode));
2586 if (tv != tarval_bad) {
2587 ir_node *blk = get_nodes_block(n);
2588 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2589 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2590 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2596 } /* transform_node_Sub */
2599 * Several transformation done on n*n=2n bits mul.
2600 * These transformations must be done here because new nodes may be produced.
2602 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2604 ir_node *a = get_Mul_left(n);
2605 ir_node *b = get_Mul_right(n);
2606 tarval *ta = value_of(a);
2607 tarval *tb = value_of(b);
2608 ir_mode *smode = get_irn_mode(a);
2610 if (ta == get_mode_one(smode)) {
2611 /* (L)1 * (L)b = (L)b */
2612 ir_node *blk = get_nodes_block(n);
2613 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2614 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2617 else if (ta == get_mode_minus_one(smode)) {
2618 /* (L)-1 * (L)b = (L)b */
2619 ir_node *blk = get_nodes_block(n);
2620 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2621 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2622 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2625 if (tb == get_mode_one(smode)) {
2626 /* (L)a * (L)1 = (L)a */
2627 ir_node *blk = get_irn_n(a, -1);
2628 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2629 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2632 else if (tb == get_mode_minus_one(smode)) {
2633 /* (L)a * (L)-1 = (L)-a */
2634 ir_node *blk = get_nodes_block(n);
2635 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2636 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2637 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2644 * Transform Mul(a,-1) into -a.
2645 * Do constant evaluation of Phi nodes.
2646 * Do architecture dependent optimizations on Mul nodes
2648 static ir_node *transform_node_Mul(ir_node *n) {
2649 ir_node *c, *oldn = n;
2650 ir_mode *mode = get_irn_mode(n);
2651 ir_node *a = get_Mul_left(n);
2652 ir_node *b = get_Mul_right(n);
2654 if (is_Bad(a) || is_Bad(b))
2657 if (mode != get_irn_mode(a))
2658 return transform_node_Mul2n(n, mode);
2660 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2662 if (mode_is_signed(mode)) {
2665 if (value_of(a) == get_mode_minus_one(mode))
2667 else if (value_of(b) == get_mode_minus_one(mode))
2670 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2671 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2676 if (is_Const(b)) { /* (-a) * const -> a * -const */
2677 ir_node *cnst = const_negate(b);
2679 dbg_info *dbgi = get_irn_dbg_info(n);
2680 ir_node *block = get_nodes_block(n);
2681 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2682 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2685 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2686 dbg_info *dbgi = get_irn_dbg_info(n);
2687 ir_node *block = get_nodes_block(n);
2688 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2689 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2691 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2692 ir_node *sub_l = get_Sub_left(b);
2693 ir_node *sub_r = get_Sub_right(b);
2694 dbg_info *dbgi = get_irn_dbg_info(n);
2695 ir_graph *irg = current_ir_graph;
2696 ir_node *block = get_nodes_block(n);
2697 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2698 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2699 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2702 } else if (is_Minus(b)) {
2703 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2704 ir_node *sub_l = get_Sub_left(a);
2705 ir_node *sub_r = get_Sub_right(a);
2706 dbg_info *dbgi = get_irn_dbg_info(n);
2707 ir_graph *irg = current_ir_graph;
2708 ir_node *block = get_nodes_block(n);
2709 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2710 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2711 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2715 if (get_mode_arithmetic(mode) == irma_ieee754) {
2717 tarval *tv = get_Const_tarval(a);
2718 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2719 && !tarval_is_negative(tv)) {
2720 /* 2.0 * b = b + b */
2721 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2722 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2726 else if (is_Const(b)) {
2727 tarval *tv = get_Const_tarval(b);
2728 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2729 && !tarval_is_negative(tv)) {
2730 /* a * 2.0 = a + a */
2731 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2732 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2737 return arch_dep_replace_mul_with_shifts(n);
2738 } /* transform_node_Mul */
2741 * Transform a Div Node.
2743 static ir_node *transform_node_Div(ir_node *n) {
2744 ir_mode *mode = get_Div_resmode(n);
2745 ir_node *a = get_Div_left(n);
2746 ir_node *b = get_Div_right(n);
2750 if (is_Const(b) && is_const_Phi(a)) {
2751 /* check for Div(Phi, Const) */
2752 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2754 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2758 else if (is_Const(a) && is_const_Phi(b)) {
2759 /* check for Div(Const, Phi) */
2760 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2762 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2766 else if (is_const_Phi(a) && is_const_Phi(b)) {
2767 /* check for Div(Phi, Phi) */
2768 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2770 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2777 if (tv != tarval_bad) {
2778 value = new_Const(get_tarval_mode(tv), tv);
2780 DBG_OPT_CSTEVAL(n, value);
2783 ir_node *a = get_Div_left(n);
2784 ir_node *b = get_Div_right(n);
2785 const ir_node *dummy;
2787 if (a == b && value_not_zero(a, &dummy)) {
2788 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2789 value = new_Const(mode, get_mode_one(mode));
2790 DBG_OPT_CSTEVAL(n, value);
2793 if (mode_is_signed(mode) && is_Const(b)) {
2794 tarval *tv = get_Const_tarval(b);
2796 if (tv == get_mode_minus_one(mode)) {
2798 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2799 DBG_OPT_CSTEVAL(n, value);
2803 /* Try architecture dependent optimization */
2804 value = arch_dep_replace_div_by_const(n);
2812 /* Turn Div into a tuple (mem, jmp, bad, value) */
2813 mem = get_Div_mem(n);
2814 blk = get_nodes_block(n);
2816 /* skip a potential Pin */
2817 mem = skip_Pin(mem);
2818 turn_into_tuple(n, pn_Div_max);
2819 set_Tuple_pred(n, pn_Div_M, mem);
2820 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2821 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2822 set_Tuple_pred(n, pn_Div_res, value);
2825 } /* transform_node_Div */
2828 * Transform a Mod node.
2830 static ir_node *transform_node_Mod(ir_node *n) {
2831 ir_mode *mode = get_Mod_resmode(n);
2832 ir_node *a = get_Mod_left(n);
2833 ir_node *b = get_Mod_right(n);
2837 if (is_Const(b) && is_const_Phi(a)) {
2838 /* check for Div(Phi, Const) */
2839 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2841 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2845 else if (is_Const(a) && is_const_Phi(b)) {
2846 /* check for Div(Const, Phi) */
2847 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2849 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2853 else if (is_const_Phi(a) && is_const_Phi(b)) {
2854 /* check for Div(Phi, Phi) */
2855 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2857 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2864 if (tv != tarval_bad) {
2865 value = new_Const(get_tarval_mode(tv), tv);
2867 DBG_OPT_CSTEVAL(n, value);
2870 ir_node *a = get_Mod_left(n);
2871 ir_node *b = get_Mod_right(n);
2872 const ir_node *dummy;
2874 if (a == b && value_not_zero(a, &dummy)) {
2875 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2876 value = new_Const(mode, get_mode_null(mode));
2877 DBG_OPT_CSTEVAL(n, value);
2880 if (mode_is_signed(mode) && is_Const(b)) {
2881 tarval *tv = get_Const_tarval(b);
2883 if (tv == get_mode_minus_one(mode)) {
2885 value = new_Const(mode, get_mode_null(mode));
2886 DBG_OPT_CSTEVAL(n, value);
2890 /* Try architecture dependent optimization */
2891 value = arch_dep_replace_mod_by_const(n);
2899 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2900 mem = get_Mod_mem(n);
2901 blk = get_nodes_block(n);
2903 /* skip a potential Pin */
2904 mem = skip_Pin(mem);
2905 turn_into_tuple(n, pn_Mod_max);
2906 set_Tuple_pred(n, pn_Mod_M, mem);
2907 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2908 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2909 set_Tuple_pred(n, pn_Mod_res, value);
2912 } /* transform_node_Mod */
2915 * Transform a DivMod node.
2917 static ir_node *transform_node_DivMod(ir_node *n) {
2918 const ir_node *dummy;
2919 ir_node *a = get_DivMod_left(n);
2920 ir_node *b = get_DivMod_right(n);
2921 ir_mode *mode = get_DivMod_resmode(n);
2926 if (is_Const(b) && is_const_Phi(a)) {
2927 /* check for Div(Phi, Const) */
2928 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2929 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2931 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2932 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2936 else if (is_Const(a) && is_const_Phi(b)) {
2937 /* check for Div(Const, Phi) */
2938 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2939 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2941 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2942 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2946 else if (is_const_Phi(a) && is_const_Phi(b)) {
2947 /* check for Div(Phi, Phi) */
2948 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2949 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2951 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2952 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2959 if (tb != tarval_bad) {
2960 if (tb == get_mode_one(get_tarval_mode(tb))) {
2962 vb = new_Const(mode, get_mode_null(mode));
2963 DBG_OPT_CSTEVAL(n, vb);
2965 } else if (ta != tarval_bad) {
2966 tarval *resa, *resb;
2967 resa = tarval_div(ta, tb);
2968 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2969 Jmp for X result!? */
2970 resb = tarval_mod(ta, tb);
2971 if (resb == tarval_bad) return n; /* Causes exception! */
2972 va = new_Const(mode, resa);
2973 vb = new_Const(mode, resb);
2974 DBG_OPT_CSTEVAL(n, va);
2975 DBG_OPT_CSTEVAL(n, vb);
2977 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2978 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2979 vb = new_Const(mode, get_mode_null(mode));
2980 DBG_OPT_CSTEVAL(n, va);
2981 DBG_OPT_CSTEVAL(n, vb);
2983 } else { /* Try architecture dependent optimization */
2986 arch_dep_replace_divmod_by_const(&va, &vb, n);
2987 evaluated = va != NULL;
2989 } else if (a == b) {
2990 if (value_not_zero(a, &dummy)) {
2992 va = new_Const(mode, get_mode_one(mode));
2993 vb = new_Const(mode, get_mode_null(mode));
2994 DBG_OPT_CSTEVAL(n, va);
2995 DBG_OPT_CSTEVAL(n, vb);
2998 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3001 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3002 /* 0 / non-Const = 0 */
3007 if (evaluated) { /* replace by tuple */
3011 mem = get_DivMod_mem(n);
3012 /* skip a potential Pin */
3013 mem = skip_Pin(mem);
3015 blk = get_nodes_block(n);
3016 turn_into_tuple(n, pn_DivMod_max);
3017 set_Tuple_pred(n, pn_DivMod_M, mem);
3018 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3019 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3020 set_Tuple_pred(n, pn_DivMod_res_div, va);
3021 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3025 } /* transform_node_DivMod */
3028 * Optimize x / c to x * (1/c)
3030 static ir_node *transform_node_Quot(ir_node *n) {
3031 ir_mode *mode = get_Quot_resmode(n);
3034 if (get_mode_arithmetic(mode) == irma_ieee754) {
3035 ir_node *b = get_Quot_right(n);
3036 tarval *tv = value_of(b);
3038 if (tv != tarval_bad) {
3042 * Floating point constant folding might be disabled here to
3044 * However, as we check for exact result, doing it is safe.
3047 rem = tarval_enable_fp_ops(1);
3048 tv = tarval_quo(get_mode_one(mode), tv);
3049 (void)tarval_enable_fp_ops(rem);
3051 /* Do the transformation if the result is either exact or we are not
3052 using strict rules. */
3053 if (tv != tarval_bad &&
3054 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3055 ir_node *blk = get_nodes_block(n);
3056 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3057 ir_node *a = get_Quot_left(n);
3058 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3059 ir_node *mem = get_Quot_mem(n);
3061 /* skip a potential Pin */
3062 mem = skip_Pin(mem);
3063 turn_into_tuple(n, pn_Quot_max);
3064 set_Tuple_pred(n, pn_Quot_M, mem);
3065 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3066 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3067 set_Tuple_pred(n, pn_Quot_res, m);
3068 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3073 } /* transform_node_Quot */
3076 * Optimize Abs(x) into x if x is Confirmed >= 0
3077 * Optimize Abs(x) into -x if x is Confirmed <= 0
3078 * Optimize Abs(-x) int Abs(x)
3080 static ir_node *transform_node_Abs(ir_node *n) {
3081 ir_node *c, *oldn = n;
3082 ir_node *a = get_Abs_op(n);
3085 HANDLE_UNOP_PHI(tarval_abs, a, c);
3087 switch (classify_value_sign(a)) {
3088 case value_classified_negative:
3089 mode = get_irn_mode(n);
3092 * We can replace the Abs by -x here.
3093 * We even could add a new Confirm here
3094 * (if not twos complement)
3096 * Note that -x would create a new node, so we could
3097 * not run it in the equivalent_node() context.
3099 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3100 get_nodes_block(n), a, mode);
3102 DBG_OPT_CONFIRM(oldn, n);
3104 case value_classified_positive:
3105 /* n is positive, Abs is not needed */
3108 DBG_OPT_CONFIRM(oldn, n);
3114 /* Abs(-x) = Abs(x) */
3115 mode = get_irn_mode(n);
3116 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3117 get_nodes_block(n), get_Minus_op(a), mode);
3118 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3122 } /* transform_node_Abs */
3125 * Optimize -a CMP -b into b CMP a.
3126 * This works only for for modes where unary Minus
3128 * Note that two-complement integers can Overflow
3129 * so it will NOT work.
3131 * For == and != can be handled in Proj(Cmp)
3133 static ir_node *transform_node_Cmp(ir_node *n) {
3135 ir_node *left = get_Cmp_left(n);
3136 ir_node *right = get_Cmp_right(n);
3138 if (is_Minus(left) && is_Minus(right) &&
3139 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3140 ir_node *const new_left = get_Minus_op(right);
3141 ir_node *const new_right = get_Minus_op(left);
3142 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3143 get_nodes_block(n), new_left, new_right);
3144 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3147 } /* transform_node_Cmp */
3151 * Transform a Cond node.
3153 * Replace the Cond by a Jmp if it branches on a constant
3156 static ir_node *transform_node_Cond(ir_node *n) {
3159 ir_node *a = get_Cond_selector(n);
3160 tarval *ta = value_of(a);
3162 /* we need block info which is not available in floating irgs */
3163 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3166 if ((ta != tarval_bad) &&
3167 (get_irn_mode(a) == mode_b) &&
3168 (get_opt_unreachable_code())) {
3169 /* It's a boolean Cond, branching on a boolean constant.
3170 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3171 ir_node *blk = get_nodes_block(n);
3172 jmp = new_r_Jmp(current_ir_graph, blk);
3173 turn_into_tuple(n, pn_Cond_max);
3174 if (ta == tarval_b_true) {
3175 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3176 set_Tuple_pred(n, pn_Cond_true, jmp);
3178 set_Tuple_pred(n, pn_Cond_false, jmp);
3179 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3181 /* We might generate an endless loop, so keep it alive. */
3182 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3185 } /* transform_node_Cond */
3188 * Prototype of a recursive transform function
3189 * for bitwise distributive transformations.
3191 typedef ir_node* (*recursive_transform)(ir_node *n);
3194 * makes use of distributive laws for and, or, eor
3195 * and(a OP c, b OP c) -> and(a, b) OP c
3196 * note, might return a different op than n
3198 static ir_node *transform_bitwise_distributive(ir_node *n,
3199 recursive_transform trans_func)
3202 ir_node *a = get_binop_left(n);
3203 ir_node *b = get_binop_right(n);
3204 ir_op *op = get_irn_op(a);
3205 ir_op *op_root = get_irn_op(n);
3207 if(op != get_irn_op(b))
3210 if (op == op_Conv) {
3211 ir_node *a_op = get_Conv_op(a);
3212 ir_node *b_op = get_Conv_op(b);
3213 ir_mode *a_mode = get_irn_mode(a_op);
3214 ir_mode *b_mode = get_irn_mode(b_op);
3215 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3216 ir_node *blk = get_nodes_block(n);
3219 set_binop_left(n, a_op);
3220 set_binop_right(n, b_op);
3221 set_irn_mode(n, a_mode);
3223 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3225 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3231 /* nothing to gain here */
3235 if (op == op_Shrs || op == op_Shr || op == op_Shl
3236 || op == op_And || op == op_Or || op == op_Eor) {
3237 ir_node *a_left = get_binop_left(a);
3238 ir_node *a_right = get_binop_right(a);
3239 ir_node *b_left = get_binop_left(b);
3240 ir_node *b_right = get_binop_right(b);
3242 ir_node *op1 = NULL;
3243 ir_node *op2 = NULL;
3245 if (is_op_commutative(op)) {
3246 if (a_left == b_left) {
3250 } else if(a_left == b_right) {
3254 } else if(a_right == b_left) {
3260 if(a_right == b_right) {
3267 /* (a sop c) & (b sop c) => (a & b) sop c */
3268 ir_node *blk = get_nodes_block(n);
3270 ir_node *new_n = exact_copy(n);
3271 set_binop_left(new_n, op1);
3272 set_binop_right(new_n, op2);
3273 new_n = trans_func(new_n);
3275 if(op_root == op_Eor && op == op_Or) {
3276 dbg_info *dbgi = get_irn_dbg_info(n);
3277 ir_graph *irg = current_ir_graph;
3278 ir_mode *mode = get_irn_mode(c);
3280 c = new_rd_Not(dbgi, irg, blk, c, mode);
3281 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3284 set_nodes_block(n, blk);
3285 set_binop_left(n, new_n);
3286 set_binop_right(n, c);
3287 add_identities(current_ir_graph->value_table, n);
3290 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3301 static ir_node *transform_node_And(ir_node *n) {
3302 ir_node *c, *oldn = n;
3303 ir_node *a = get_And_left(n);
3304 ir_node *b = get_And_right(n);
3307 mode = get_irn_mode(n);
3308 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3310 /* we can evaluate 2 Projs of the same Cmp */
3311 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3312 ir_node *pred_a = get_Proj_pred(a);
3313 ir_node *pred_b = get_Proj_pred(b);
3314 if (pred_a == pred_b) {
3315 dbg_info *dbgi = get_irn_dbg_info(n);
3316 ir_node *block = get_nodes_block(pred_a);
3317 pn_Cmp pn_a = get_Proj_proj(a);
3318 pn_Cmp pn_b = get_Proj_proj(b);
3319 /* yes, we can simply calculate with pncs */
3320 pn_Cmp new_pnc = pn_a & pn_b;
3322 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3327 ir_node *op = get_Not_op(b);
3329 ir_node *ba = get_And_left(op);
3330 ir_node *bb = get_And_right(op);
3332 /* it's enough to test the following cases due to normalization! */
3333 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3334 /* (a|b) & ~(a&b) = a^b */
3335 ir_node *block = get_nodes_block(n);
3337 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3338 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3346 ir_node *op = get_Not_op(a);
3348 ir_node *aa = get_And_left(op);
3349 ir_node *ab = get_And_right(op);
3351 /* it's enough to test the following cases due to normalization! */
3352 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3353 /* (a|b) & ~(a&b) = a^b */
3354 ir_node *block = get_nodes_block(n);
3356 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3357 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3364 ir_node *al = get_Eor_left(a);
3365 ir_node *ar = get_Eor_right(a);
3368 /* (b ^ a) & b -> ~a & b */
3369 dbg_info *dbg = get_irn_dbg_info(n);
3370 ir_node *block = get_nodes_block(n);
3372 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3373 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3374 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3378 /* (a ^ b) & b -> ~a & b */
3379 dbg_info *dbg = get_irn_dbg_info(n);
3380 ir_node *block = get_nodes_block(n);
3382 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3383 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3384 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3389 ir_node *bl = get_Eor_left(b);
3390 ir_node *br = get_Eor_right(b);
3393 /* a & (a ^ b) -> a & ~b */
3394 dbg_info *dbg = get_irn_dbg_info(n);
3395 ir_node *block = get_nodes_block(n);
3397 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3398 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3403 /* a & (b ^ a) -> a & ~b */
3404 dbg_info *dbg = get_irn_dbg_info(n);
3405 ir_node *block = get_nodes_block(n);
3407 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3408 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3409 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3413 if (is_Not(a) && is_Not(b)) {
3414 /* ~a & ~b = ~(a|b) */
3415 ir_node *block = get_nodes_block(n);
3416 ir_mode *mode = get_irn_mode(n);
3420 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3421 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3422 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3426 n = transform_bitwise_distributive(n, transform_node_And);
3429 } /* transform_node_And */
3434 static ir_node *transform_node_Eor(ir_node *n) {
3435 ir_node *c, *oldn = n;
3436 ir_node *a = get_Eor_left(n);
3437 ir_node *b = get_Eor_right(n);
3438 ir_mode *mode = get_irn_mode(n);
3440 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3442 /* we can evaluate 2 Projs of the same Cmp */
3443 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3444 ir_node *pred_a = get_Proj_pred(a);
3445 ir_node *pred_b = get_Proj_pred(b);
3446 if(pred_a == pred_b) {
3447 dbg_info *dbgi = get_irn_dbg_info(n);
3448 ir_node *block = get_nodes_block(pred_a);
3449 pn_Cmp pn_a = get_Proj_proj(a);
3450 pn_Cmp pn_b = get_Proj_proj(b);
3451 /* yes, we can simply calculate with pncs */
3452 pn_Cmp new_pnc = pn_a ^ pn_b;
3454 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3461 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
3462 mode, get_mode_null(mode));
3463 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3464 } else if (mode == mode_b &&
3466 is_Const(b) && is_Const_one(b) &&
3467 is_Cmp(get_Proj_pred(a))) {
3468 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3469 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3470 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3472 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3473 } else if (is_Const(b)) {
3474 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3475 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3476 ir_node *not_op = get_Not_op(a);
3477 dbg_info *dbg = get_irn_dbg_info(n);
3478 ir_graph *irg = current_ir_graph;
3479 ir_node *block = get_nodes_block(n);
3480 ir_mode *mode = get_irn_mode(n);
3481 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3483 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3484 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3485 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3488 n = transform_bitwise_distributive(n, transform_node_Eor);
3492 } /* transform_node_Eor */
3497 static ir_node *transform_node_Not(ir_node *n) {
3498 ir_node *c, *oldn = n;
3499 ir_node *a = get_Not_op(n);
3500 ir_mode *mode = get_irn_mode(n);
3502 HANDLE_UNOP_PHI(tarval_not,a,c);
3504 /* check for a boolean Not */
3505 if (mode == mode_b &&
3507 is_Cmp(get_Proj_pred(a))) {
3508 /* We negate a Cmp. The Cmp has the negated result anyways! */
3509 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3510 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3515 ir_node *eor_b = get_Eor_right(a);
3516 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3517 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3518 ir_node *eor_a = get_Eor_left(a);
3519 dbg_info *dbg = get_irn_dbg_info(n);
3520 ir_graph *irg = current_ir_graph;
3521 ir_node *block = get_nodes_block(n);
3522 ir_mode *mode = get_irn_mode(n);
3523 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3527 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3528 if (is_Minus(a)) { /* ~-x -> x + -1 */
3529 dbg_info *dbg = get_irn_dbg_info(n);
3530 ir_graph *irg = current_ir_graph;
3531 ir_node *block = get_nodes_block(n);
3532 ir_node *add_l = get_Minus_op(a);
3533 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3534 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3535 } else if (is_Add(a)) {
3536 ir_node *add_r = get_Add_right(a);
3537 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3538 /* ~(x + -1) = -x */
3539 ir_node *op = get_Add_left(a);
3540 ir_node *blk = get_nodes_block(n);
3541 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3542 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3547 } /* transform_node_Not */
3550 * Transform a Minus.
3554 * -(a >>u (size-1)) = a >>s (size-1)
3555 * -(a >>s (size-1)) = a >>u (size-1)
3556 * -(a * const) -> a * -const
3558 static ir_node *transform_node_Minus(ir_node *n) {
3559 ir_node *c, *oldn = n;
3560 ir_node *a = get_Minus_op(n);
3563 HANDLE_UNOP_PHI(tarval_neg,a,c);
3565 mode = get_irn_mode(a);
3566 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3567 /* the following rules are only to twos-complement */
3570 ir_node *op = get_Not_op(a);
3571 tarval *tv = get_mode_one(mode);
3572 ir_node *blk = get_nodes_block(n);
3573 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3574 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3575 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3579 ir_node *c = get_Shr_right(a);
3582 tarval *tv = get_Const_tarval(c);
3584 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3585 /* -(a >>u (size-1)) = a >>s (size-1) */
3586 ir_node *v = get_Shr_left(a);
3588 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3589 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3595 ir_node *c = get_Shrs_right(a);
3598 tarval *tv = get_Const_tarval(c);
3600 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3601 /* -(a >>s (size-1)) = a >>u (size-1) */
3602 ir_node *v = get_Shrs_left(a);
3604 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3605 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3612 /* - (a-b) = b - a */
3613 ir_node *la = get_Sub_left(a);
3614 ir_node *ra = get_Sub_right(a);
3615 ir_node *blk = get_nodes_block(n);
3617 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3618 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3622 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3623 ir_node *mul_l = get_Mul_left(a);
3624 ir_node *mul_r = get_Mul_right(a);
3625 tarval *tv = value_of(mul_r);
3626 if (tv != tarval_bad) {
3627 tv = tarval_neg(tv);
3628 if (tv != tarval_bad) {
3629 ir_node *cnst = new_Const(mode, tv);
3630 dbg_info *dbg = get_irn_dbg_info(a);
3631 ir_graph *irg = current_ir_graph;
3632 ir_node *block = get_nodes_block(a);
3633 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3634 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3641 } /* transform_node_Minus */
3644 * Transform a Cast_type(Const) into a new Const_type
3646 static ir_node *transform_node_Cast(ir_node *n) {
3648 ir_node *pred = get_Cast_op(n);
3649 ir_type *tp = get_irn_type(n);
3651 if (is_Const(pred) && get_Const_type(pred) != tp) {
3652 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3653 get_Const_tarval(pred), tp);
3654 DBG_OPT_CSTEVAL(oldn, n);
3655 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3656 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3657 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3658 DBG_OPT_CSTEVAL(oldn, n);
3662 } /* transform_node_Cast */
3665 * Transform a Proj(Load) with a non-null address.
3667 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3668 if (get_opt_ldst_only_null_ptr_exceptions()) {
3669 if (get_irn_mode(proj) == mode_X) {
3670 ir_node *load = get_Proj_pred(proj);
3672 /* get the Load address */
3673 const ir_node *addr = get_Load_ptr(load);
3674 const ir_node *confirm;
3676 if (value_not_null(addr, &confirm)) {
3677 if (confirm == NULL) {
3678 /* this node may float if it did not depend on a Confirm */
3679 set_irn_pinned(load, op_pin_state_floats);
3681 if (get_Proj_proj(proj) == pn_Load_X_except) {
3682 DBG_OPT_EXC_REM(proj);
3683 return get_irg_bad(current_ir_graph);
3685 ir_node *blk = get_nodes_block(load);
3686 return new_r_Jmp(current_ir_graph, blk);
3692 } /* transform_node_Proj_Load */
3695 * Transform a Proj(Store) with a non-null address.
3697 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3698 if (get_opt_ldst_only_null_ptr_exceptions()) {
3699 if (get_irn_mode(proj) == mode_X) {
3700 ir_node *store = get_Proj_pred(proj);
3702 /* get the load/store address */
3703 const ir_node *addr = get_Store_ptr(store);
3704 const ir_node *confirm;
3706 if (value_not_null(addr, &confirm)) {
3707 if (confirm == NULL) {
3708 /* this node may float if it did not depend on a Confirm */
3709 set_irn_pinned(store, op_pin_state_floats);
3711 if (get_Proj_proj(proj) == pn_Store_X_except) {
3712 DBG_OPT_EXC_REM(proj);
3713 return get_irg_bad(current_ir_graph);
3715 ir_node *blk = get_nodes_block(store);
3716 return new_r_Jmp(current_ir_graph, blk);
3722 } /* transform_node_Proj_Store */
3725 * Transform a Proj(Div) with a non-zero value.
3726 * Removes the exceptions and routes the memory to the NoMem node.
3728 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3729 ir_node *div = get_Proj_pred(proj);
3730 ir_node *b = get_Div_right(div);
3731 ir_node *res, *new_mem;
3732 const ir_node *confirm;
3735 if (value_not_zero(b, &confirm)) {
3736 /* div(x, y) && y != 0 */
3737 if (confirm == NULL) {
3738 /* we are sure we have a Const != 0 */
3739 new_mem = get_Div_mem(div);
3740 new_mem = skip_Pin(new_mem);
3741 set_Div_mem(div, new_mem);
3742 set_irn_pinned(div, op_pin_state_floats);
3745 proj_nr = get_Proj_proj(proj);
3747 case pn_Div_X_regular:
3748 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3750 case pn_Div_X_except:
3751 /* we found an exception handler, remove it */
3752 DBG_OPT_EXC_REM(proj);
3756 res = get_Div_mem(div);
3757 new_mem = get_irg_no_mem(current_ir_graph);
3760 /* This node can only float up to the Confirm block */
3761 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3763 set_irn_pinned(div, op_pin_state_floats);
3764 /* this is a Div without exception, we can remove the memory edge */
3765 set_Div_mem(div, new_mem);
3770 } /* transform_node_Proj_Div */
3773 * Transform a Proj(Mod) with a non-zero value.
3774 * Removes the exceptions and routes the memory to the NoMem node.
3776 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3777 ir_node *mod = get_Proj_pred(proj);
3778 ir_node *b = get_Mod_right(mod);
3779 ir_node *res, *new_mem;
3780 const ir_node *confirm;
3783 if (value_not_zero(b, &confirm)) {
3784 /* mod(x, y) && y != 0 */
3785 proj_nr = get_Proj_proj(proj);
3787 if (confirm == NULL) {
3788 /* we are sure we have a Const != 0 */
3789 new_mem = get_Mod_mem(mod);
3790 new_mem = skip_Pin(new_mem);
3791 set_Mod_mem(mod, new_mem);
3792 set_irn_pinned(mod, op_pin_state_floats);
3797 case pn_Mod_X_regular:
3798 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3800 case pn_Mod_X_except:
3801 /* we found an exception handler, remove it */
3802 DBG_OPT_EXC_REM(proj);
3806 res = get_Mod_mem(mod);
3807 new_mem = get_irg_no_mem(current_ir_graph);
3810 /* This node can only float up to the Confirm block */
3811 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3813 /* this is a Mod without exception, we can remove the memory edge */
3814 set_Mod_mem(mod, new_mem);
3817 if (get_Mod_left(mod) == b) {
3818 /* a % a = 0 if a != 0 */
3819 ir_mode *mode = get_irn_mode(proj);
3820 ir_node *res = new_Const(mode, get_mode_null(mode));
3822 DBG_OPT_CSTEVAL(mod, res);
3828 } /* transform_node_Proj_Mod */
3831 * Transform a Proj(DivMod) with a non-zero value.
3832 * Removes the exceptions and routes the memory to the NoMem node.
3834 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3835 ir_node *divmod = get_Proj_pred(proj);
3836 ir_node *b = get_DivMod_right(divmod);
3837 ir_node *res, *new_mem;
3838 const ir_node *confirm;
3841 if (value_not_zero(b, &confirm)) {
3842 /* DivMod(x, y) && y != 0 */
3843 proj_nr = get_Proj_proj(proj);
3845 if (confirm == NULL) {
3846 /* we are sure we have a Const != 0 */
3847 new_mem = get_DivMod_mem(divmod);
3848 new_mem = skip_Pin(new_mem);
3849 set_DivMod_mem(divmod, new_mem);
3850 set_irn_pinned(divmod, op_pin_state_floats);
3855 case pn_DivMod_X_regular:
3856 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3858 case pn_DivMod_X_except:
3859 /* we found an exception handler, remove it */
3860 DBG_OPT_EXC_REM(proj);
3864 res = get_DivMod_mem(divmod);
3865 new_mem = get_irg_no_mem(current_ir_graph);
3868 /* This node can only float up to the Confirm block */
3869 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3871 /* this is a DivMod without exception, we can remove the memory edge */
3872 set_DivMod_mem(divmod, new_mem);
3875 case pn_DivMod_res_mod:
3876 if (get_DivMod_left(divmod) == b) {
3877 /* a % a = 0 if a != 0 */
3878 ir_mode *mode = get_irn_mode(proj);
3879 ir_node *res = new_Const(mode, get_mode_null(mode));
3881 DBG_OPT_CSTEVAL(divmod, res);
3887 } /* transform_node_Proj_DivMod */
3890 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3892 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3893 if (get_opt_unreachable_code()) {
3894 ir_node *n = get_Proj_pred(proj);
3895 ir_node *b = get_Cond_selector(n);
3897 if (mode_is_int(get_irn_mode(b))) {
3898 tarval *tb = value_of(b);
3900 if (tb != tarval_bad) {
3901 /* we have a constant switch */
3902 long num = get_Proj_proj(proj);
3904 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3905 if (get_tarval_long(tb) == num) {
3906 /* Do NOT create a jump here, or we will have 2 control flow ops
3907 * in a block. This case is optimized away in optimize_cf(). */
3910 /* this case will NEVER be taken, kill it */
3911 return get_irg_bad(current_ir_graph);
3918 } /* transform_node_Proj_Cond */
3921 * Create a 0 constant of given mode.
3923 static ir_node *create_zero_const(ir_mode *mode) {
3924 tarval *tv = get_mode_null(mode);
3925 ir_node *cnst = new_Const(mode, tv);
3930 /* the order of the values is important! */
3931 typedef enum const_class {
3937 static const_class classify_const(const ir_node* n)
3939 if (is_Const(n)) return const_const;
3940 if (is_irn_constlike(n)) return const_like;
3945 * Determines whether r is more constlike or has a larger index (in that order)
3948 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3950 const const_class l_order = classify_const(l);
3951 const const_class r_order = classify_const(r);
3953 l_order > r_order ||
3954 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3958 * Normalizes and optimizes Cmp nodes.
3960 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3961 ir_node *n = get_Proj_pred(proj);
3962 ir_node *left = get_Cmp_left(n);
3963 ir_node *right = get_Cmp_right(n);
3966 ir_mode *mode = NULL;
3967 long proj_nr = get_Proj_proj(proj);
3969 /* we can evaluate some cases directly */
3972 return new_Const(mode_b, get_tarval_b_false());
3974 return new_Const(mode_b, get_tarval_b_true());
3976 if (!mode_is_float(get_irn_mode(left)))
3977 return new_Const(mode_b, get_tarval_b_true());
3983 /* remove Casts of both sides */
3984 left = skip_Cast(left);
3985 right = skip_Cast(right);
3987 /* Remove unnecessary conversions */
3988 /* TODO handle constants */
3989 if (is_Conv(left) && is_Conv(right)) {
3990 ir_mode *mode = get_irn_mode(left);
3991 ir_node *op_left = get_Conv_op(left);
3992 ir_node *op_right = get_Conv_op(right);
3993 ir_mode *mode_left = get_irn_mode(op_left);
3994 ir_mode *mode_right = get_irn_mode(op_right);
3996 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3997 && mode_left != mode_b && mode_right != mode_b) {
3998 ir_graph *irg = current_ir_graph;
3999 ir_node *block = get_nodes_block(n);
4001 if (mode_left == mode_right) {
4005 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4006 } else if (smaller_mode(mode_left, mode_right)) {
4007 left = new_r_Conv(irg, block, op_left, mode_right);
4010 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4011 } else if (smaller_mode(mode_right, mode_left)) {
4013 right = new_r_Conv(irg, block, op_right, mode_left);
4015 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4020 /* remove operation on both sides if possible */
4021 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4023 * The following operations are NOT safe for floating point operations, for instance
4024 * 1.0 + inf == 2.0 + inf, =/=> x == y
4026 if (mode_is_int(get_irn_mode(left))) {
4027 unsigned lop = get_irn_opcode(left);
4029 if (lop == get_irn_opcode(right)) {
4030 ir_node *ll, *lr, *rl, *rr;
4032 /* same operation on both sides, try to remove */
4036 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4037 left = get_unop_op(left);
4038 right = get_unop_op(right);
4040 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4043 ll = get_Add_left(left);
4044 lr = get_Add_right(left);
4045 rl = get_Add_left(right);
4046 rr = get_Add_right(right);
4049 /* X + a CMP X + b ==> a CMP b */
4053 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4054 } else if (ll == rr) {
4055 /* X + a CMP b + X ==> a CMP b */
4059 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4060 } else if (lr == rl) {
4061 /* a + X CMP X + b ==> a CMP b */
4065 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4066 } else if (lr == rr) {
4067 /* a + X CMP b + X ==> a CMP b */
4071 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4075 ll = get_Sub_left(left);
4076 lr = get_Sub_right(left);
4077 rl = get_Sub_left(right);
4078 rr = get_Sub_right(right);
4081 /* X - a CMP X - b ==> a CMP b */
4085 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4086 } else if (lr == rr) {
4087 /* a - X CMP b - X ==> a CMP b */
4091 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4095 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4096 /* a ROTL X CMP b ROTL X ==> a CMP b */
4097 left = get_Rotl_left(left);
4098 right = get_Rotl_left(right);
4100 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4108 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4109 if (is_Add(left) || is_Sub(left)) {
4110 ir_node *ll = get_binop_left(left);
4111 ir_node *lr = get_binop_right(left);
4113 if (lr == right && is_Add(left)) {
4120 right = create_zero_const(get_irn_mode(left));
4122 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4125 if (is_Add(right) || is_Sub(right)) {
4126 ir_node *rl = get_binop_left(right);
4127 ir_node *rr = get_binop_right(right);
4129 if (rr == left && is_Add(right)) {
4136 right = create_zero_const(get_irn_mode(left));
4138 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4141 } /* mode_is_int(...) */
4142 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4144 /* replace mode_b compares with ands/ors */
4145 if (get_irn_mode(left) == mode_b) {
4146 ir_graph *irg = current_ir_graph;
4147 ir_node *block = get_nodes_block(n);
4151 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4152 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4153 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4154 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4155 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4156 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4157 default: bres = NULL;
4160 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4166 * First step: normalize the compare op
4167 * by placing the constant on the right side
4168 * or moving the lower address node to the left.
4170 if (!operands_are_normalized(left, right)) {
4176 proj_nr = get_inversed_pnc(proj_nr);
4181 * Second step: Try to reduce the magnitude
4182 * of a constant. This may help to generate better code
4183 * later and may help to normalize more compares.
4184 * Of course this is only possible for integer values.
4186 tv = value_of(right);
4187 if (tv != tarval_bad) {
4188 mode = get_irn_mode(right);
4190 /* TODO extend to arbitrary constants */
4191 if (is_Conv(left) && tarval_is_null(tv)) {
4192 ir_node *op = get_Conv_op(left);
4193 ir_mode *op_mode = get_irn_mode(op);
4196 * UpConv(x) REL 0 ==> x REL 0
4198 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4199 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4200 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4201 tv = get_mode_null(op_mode);
4205 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4209 if (tv != tarval_bad) {
4210 /* the following optimization is possible on modes without Overflow
4211 * on Unary Minus or on == and !=:
4212 * -a CMP c ==> a swap(CMP) -c
4214 * Beware: for two-complement Overflow may occur, so only == and != can
4215 * be optimized, see this:
4216 * -MININT < 0 =/=> MININT > 0 !!!
4218 if (is_Minus(left) &&
4219 (!mode_overflow_on_unary_Minus(mode) ||
4220 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4221 tv = tarval_neg(tv);
4223 if (tv != tarval_bad) {
4224 left = get_Minus_op(left);
4225 proj_nr = get_inversed_pnc(proj_nr);
4227 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4229 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4230 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4231 tv = tarval_not(tv);
4233 if (tv != tarval_bad) {
4234 left = get_Not_op(left);
4236 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4240 /* for integer modes, we have more */
4241 if (mode_is_int(mode)) {
4242 /* Ne includes Unordered which is not possible on integers.
4243 * However, frontends often use this wrong, so fix it here */
4244 if (proj_nr & pn_Cmp_Uo) {
4245 proj_nr &= ~pn_Cmp_Uo;
4246 set_Proj_proj(proj, proj_nr);
4249 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4250 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4251 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4252 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4254 if (tv != tarval_bad) {
4255 proj_nr ^= pn_Cmp_Eq;
4257 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4260 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4261 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4262 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4263 tv = tarval_add(tv, get_mode_one(mode));
4265 if (tv != tarval_bad) {
4266 proj_nr ^= pn_Cmp_Eq;
4268 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4272 /* the following reassociations work only for == and != */
4273 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4275 #if 0 /* Might be not that good in general */
4276 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4277 if (tarval_is_null(tv) && is_Sub(left)) {
4278 right = get_Sub_right(left);
4279 left = get_Sub_left(left);
4281 tv = value_of(right);
4283 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4287 if (tv != tarval_bad) {
4288 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4290 ir_node *c1 = get_Sub_right(left);
4291 tarval *tv2 = value_of(c1);
4293 if (tv2 != tarval_bad) {
4294 tv2 = tarval_add(tv, value_of(c1));
4296 if (tv2 != tarval_bad) {
4297 left = get_Sub_left(left);
4300 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4304 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4305 else if (is_Add(left)) {
4306 ir_node *a_l = get_Add_left(left);
4307 ir_node *a_r = get_Add_right(left);
4311 if (is_Const(a_l)) {
4313 tv2 = value_of(a_l);
4316 tv2 = value_of(a_r);
4319 if (tv2 != tarval_bad) {
4320 tv2 = tarval_sub(tv, tv2, NULL);
4322 if (tv2 != tarval_bad) {
4326 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4330 /* -a == c ==> a == -c, -a != c ==> a != -c */
4331 else if (is_Minus(left)) {
4332 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4334 if (tv2 != tarval_bad) {
4335 left = get_Minus_op(left);
4338 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4343 /* the following reassociations work only for <= */
4344 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4345 if (tv != tarval_bad) {
4346 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4347 if (is_Abs(left)) { // TODO something is missing here
4353 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4354 switch (get_irn_opcode(left)) {
4358 c1 = get_And_right(left);
4361 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4362 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4364 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4366 /* TODO: move to constant evaluation */
4367 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4368 c1 = new_Const(mode_b, tv);
4369 DBG_OPT_CSTEVAL(proj, c1);
4373 if (tarval_is_single_bit(tv)) {
4375 * optimization for AND:
4377 * And(x, C) == C ==> And(x, C) != 0
4378 * And(x, C) != C ==> And(X, C) == 0
4380 * if C is a single Bit constant.
4383 /* check for Constant's match. We have check hare the tarvals,
4384 because our const might be changed */
4385 if (get_Const_tarval(c1) == tv) {
4386 /* fine: do the transformation */
4387 tv = get_mode_null(get_tarval_mode(tv));
4388 proj_nr ^= pn_Cmp_Leg;
4390 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4396 c1 = get_Or_right(left);
4397 if (is_Const(c1) && tarval_is_null(tv)) {
4399 * Or(x, C) == 0 && C != 0 ==> FALSE
4400 * Or(x, C) != 0 && C != 0 ==> TRUE
4402 if (! tarval_is_null(get_Const_tarval(c1))) {
4403 /* TODO: move to constant evaluation */
4404 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4405 c1 = new_Const(mode_b, tv);
4406 DBG_OPT_CSTEVAL(proj, c1);
4413 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4415 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4418 c1 = get_Shl_right(left);
4420 tarval *tv1 = get_Const_tarval(c1);
4421 ir_mode *mode = get_irn_mode(left);
4422 tarval *minus1 = get_mode_all_one(mode);
4423 tarval *amask = tarval_shr(minus1, tv1);
4424 tarval *cmask = tarval_shl(minus1, tv1);
4427 if (tarval_and(tv, cmask) != tv) {
4428 /* condition not met */
4429 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4430 c1 = new_Const(mode_b, tv);
4431 DBG_OPT_CSTEVAL(proj, c1);
4434 sl = get_Shl_left(left);
4435 blk = get_nodes_block(n);
4436 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4437 tv = tarval_shr(tv, tv1);
4439 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4444 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4446 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4449 c1 = get_Shr_right(left);
4451 tarval *tv1 = get_Const_tarval(c1);
4452 ir_mode *mode = get_irn_mode(left);
4453 tarval *minus1 = get_mode_all_one(mode);
4454 tarval *amask = tarval_shl(minus1, tv1);
4455 tarval *cmask = tarval_shr(minus1, tv1);
4458 if (tarval_and(tv, cmask) != tv) {
4459 /* condition not met */
4460 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4461 c1 = new_Const(mode_b, tv);
4462 DBG_OPT_CSTEVAL(proj, c1);
4465 sl = get_Shr_left(left);
4466 blk = get_nodes_block(n);
4467 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4468 tv = tarval_shl(tv, tv1);
4470 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4475 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4477 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4480 c1 = get_Shrs_right(left);
4482 tarval *tv1 = get_Const_tarval(c1);
4483 ir_mode *mode = get_irn_mode(left);
4484 tarval *minus1 = get_mode_all_one(mode);
4485 tarval *amask = tarval_shl(minus1, tv1);
4486 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4489 cond = tarval_sub(cond, tv1, NULL);
4490 cond = tarval_shrs(tv, cond);
4492 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4493 /* condition not met */
4494 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4495 c1 = new_Const(mode_b, tv);
4496 DBG_OPT_CSTEVAL(proj, c1);
4499 sl = get_Shrs_left(left);
4500 blk = get_nodes_block(n);
4501 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4502 tv = tarval_shl(tv, tv1);
4504 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4509 } /* tarval != bad */
4512 if (changed & 2) /* need a new Const */
4513 right = new_Const(mode, tv);
4515 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4516 ir_node *op = get_Proj_pred(left);
4518 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4519 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4520 ir_node *c = get_binop_right(op);
4523 tarval *tv = get_Const_tarval(c);
4525 if (tarval_is_single_bit(tv)) {
4526 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4527 ir_node *v = get_binop_left(op);
4528 ir_node *blk = get_irn_n(op, -1);
4529 ir_mode *mode = get_irn_mode(v);
4531 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4532 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4534 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4541 ir_node *block = get_nodes_block(n);
4543 /* create a new compare */
4544 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4545 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4549 } /* transform_node_Proj_Cmp */
4552 * Optimize CopyB(mem, x, x) into a Nop.
4554 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4555 ir_node *copyb = get_Proj_pred(proj);
4556 ir_node *a = get_CopyB_dst(copyb);
4557 ir_node *b = get_CopyB_src(copyb);
4560 switch (get_Proj_proj(proj)) {
4561 case pn_CopyB_X_regular:
4562 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4563 DBG_OPT_EXC_REM(proj);
4564 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4566 case pn_CopyB_M_except:
4567 case pn_CopyB_X_except:
4568 DBG_OPT_EXC_REM(proj);
4569 proj = get_irg_bad(current_ir_graph);
4576 } /* transform_node_Proj_CopyB */
4579 * Optimize Bounds(idx, idx, upper) into idx.
4581 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4582 ir_node *oldn = proj;
4583 ir_node *bound = get_Proj_pred(proj);
4584 ir_node *idx = get_Bound_index(bound);
4585 ir_node *pred = skip_Proj(idx);
4588 if (idx == get_Bound_lower(bound))
4590 else if (is_Bound(pred)) {
4592 * idx was Bounds checked in the same MacroBlock previously,
4593 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4595 ir_node *lower = get_Bound_lower(bound);
4596 ir_node *upper = get_Bound_upper(bound);
4597 if (get_Bound_lower(pred) == lower &&
4598 get_Bound_upper(pred) == upper &&
4599 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4601 * One could expect that we simply return the previous
4602 * Bound here. However, this would be wrong, as we could
4603 * add an exception Proj to a new location then.
4604 * So, we must turn in into a tuple.
4610 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4611 switch (get_Proj_proj(proj)) {
4613 DBG_OPT_EXC_REM(proj);
4614 proj = get_Bound_mem(bound);
4616 case pn_Bound_X_except:
4617 DBG_OPT_EXC_REM(proj);
4618 proj = get_irg_bad(current_ir_graph);
4622 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4624 case pn_Bound_X_regular:
4625 DBG_OPT_EXC_REM(proj);
4626 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4633 } /* transform_node_Proj_Bound */
4636 * Does all optimizations on nodes that must be done on it's Proj's
4637 * because of creating new nodes.
4639 static ir_node *transform_node_Proj(ir_node *proj) {
4640 ir_node *n = get_Proj_pred(proj);
4642 if (n->op->ops.transform_node_Proj)
4643 return n->op->ops.transform_node_Proj(proj);
4645 } /* transform_node_Proj */
4648 * Move Confirms down through Phi nodes.
4650 static ir_node *transform_node_Phi(ir_node *phi) {
4652 ir_mode *mode = get_irn_mode(phi);
4654 if (mode_is_reference(mode)) {
4655 n = get_irn_arity(phi);
4657 /* Beware of Phi0 */
4659 ir_node *pred = get_irn_n(phi, 0);
4660 ir_node *bound, *new_Phi, *block, **in;
4663 if (! is_Confirm(pred))
4666 bound = get_Confirm_bound(pred);
4667 pnc = get_Confirm_cmp(pred);
4669 NEW_ARR_A(ir_node *, in, n);
4670 in[0] = get_Confirm_value(pred);
4672 for (i = 1; i < n; ++i) {
4673 pred = get_irn_n(phi, i);
4675 if (! is_Confirm(pred) ||
4676 get_Confirm_bound(pred) != bound ||
4677 get_Confirm_cmp(pred) != pnc)
4679 in[i] = get_Confirm_value(pred);
4681 /* move the Confirm nodes "behind" the Phi */
4682 block = get_irn_n(phi, -1);
4683 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4684 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4688 } /* transform_node_Phi */
4691 * Returns the operands of a commutative bin-op, if one operand is
4692 * a const, it is returned as the second one.
4694 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4695 ir_node *op_a = get_binop_left(binop);
4696 ir_node *op_b = get_binop_right(binop);
4698 assert(is_op_commutative(get_irn_op(binop)));
4700 if (is_Const(op_a)) {
4707 } /* get_comm_Binop_Ops */
4710 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4711 * Such pattern may arise in bitfield stores.
4713 * value c4 value c4 & c2
4714 * AND c3 AND c1 | c3
4721 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4724 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4727 ir_node *and_l, *c3;
4728 ir_node *value, *c4;
4729 ir_node *new_and, *new_const, *block;
4730 ir_mode *mode = get_irn_mode(or);
4732 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4735 get_comm_Binop_Ops(or, &and, &c1);
4736 if (!is_Const(c1) || !is_And(and))
4739 get_comm_Binop_Ops(and, &or_l, &c2);
4743 tv1 = get_Const_tarval(c1);
4744 tv2 = get_Const_tarval(c2);
4746 tv = tarval_or(tv1, tv2);
4747 if (tarval_is_all_one(tv)) {
4748 /* the AND does NOT clear a bit with isn't set by the OR */
4749 set_Or_left(or, or_l);
4750 set_Or_right(or, c1);
4752 /* check for more */
4759 get_comm_Binop_Ops(or_l, &and_l, &c3);
4760 if (!is_Const(c3) || !is_And(and_l))
4763 get_comm_Binop_Ops(and_l, &value, &c4);
4767 /* ok, found the pattern, check for conditions */
4768 assert(mode == get_irn_mode(and));
4769 assert(mode == get_irn_mode(or_l));
4770 assert(mode == get_irn_mode(and_l));
4772 tv3 = get_Const_tarval(c3);
4773 tv4 = get_Const_tarval(c4);
4775 tv = tarval_or(tv4, tv2);
4776 if (!tarval_is_all_one(tv)) {
4777 /* have at least one 0 at the same bit position */
4781 n_tv4 = tarval_not(tv4);
4782 if (tv3 != tarval_and(tv3, n_tv4)) {
4783 /* bit in the or_mask is outside the and_mask */
4787 n_tv2 = tarval_not(tv2);
4788 if (tv1 != tarval_and(tv1, n_tv2)) {
4789 /* bit in the or_mask is outside the and_mask */
4793 /* ok, all conditions met */
4794 block = get_irn_n(or, -1);
4796 new_and = new_r_And(current_ir_graph, block,
4797 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4799 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4801 set_Or_left(or, new_and);
4802 set_Or_right(or, new_const);
4804 /* check for more */
4806 } /* transform_node_Or_bf_store */
4809 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4811 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4812 ir_mode *mode = get_irn_mode(or);
4813 ir_node *shl, *shr, *block;
4814 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4817 if (! mode_is_int(mode))
4820 shl = get_binop_left(or);
4821 shr = get_binop_right(or);
4830 } else if (!is_Shl(shl)) {
4832 } else if (!is_Shr(shr)) {
4835 x = get_Shl_left(shl);
4836 if (x != get_Shr_left(shr))
4839 c1 = get_Shl_right(shl);
4840 c2 = get_Shr_right(shr);
4841 if (is_Const(c1) && is_Const(c2)) {
4842 tv1 = get_Const_tarval(c1);
4843 if (! tarval_is_long(tv1))
4846 tv2 = get_Const_tarval(c2);
4847 if (! tarval_is_long(tv2))
4850 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4851 != (int) get_mode_size_bits(mode))
4854 /* yet, condition met */
4855 block = get_nodes_block(or);
4857 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4859 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4866 rotval = sub; /* a Rot right is not supported, so use a rot left */
4867 } else if (is_Sub(c2)) {
4873 if (get_Sub_right(sub) != v)
4876 c1 = get_Sub_left(sub);
4880 tv1 = get_Const_tarval(c1);
4881 if (! tarval_is_long(tv1))
4884 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4887 /* yet, condition met */
4888 block = get_nodes_block(or);
4890 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4892 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4894 } /* transform_node_Or_Rotl */
4899 static ir_node *transform_node_Or(ir_node *n) {
4900 ir_node *c, *oldn = n;
4901 ir_node *a = get_Or_left(n);
4902 ir_node *b = get_Or_right(n);
4905 if (is_Not(a) && is_Not(b)) {
4906 /* ~a | ~b = ~(a&b) */
4907 ir_node *block = get_nodes_block(n);
4909 mode = get_irn_mode(n);
4912 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4913 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4914 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4918 /* we can evaluate 2 Projs of the same Cmp */
4919 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4920 ir_node *pred_a = get_Proj_pred(a);
4921 ir_node *pred_b = get_Proj_pred(b);
4922 if (pred_a == pred_b) {
4923 dbg_info *dbgi = get_irn_dbg_info(n);
4924 ir_node *block = get_nodes_block(pred_a);
4925 pn_Cmp pn_a = get_Proj_proj(a);
4926 pn_Cmp pn_b = get_Proj_proj(b);
4927 /* yes, we can simply calculate with pncs */
4928 pn_Cmp new_pnc = pn_a | pn_b;
4930 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4935 mode = get_irn_mode(n);
4936 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4938 n = transform_node_Or_bf_store(n);
4939 n = transform_node_Or_Rotl(n);
4943 n = transform_bitwise_distributive(n, transform_node_Or);
4946 } /* transform_node_Or */
4950 static ir_node *transform_node(ir_node *n);
4953 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4955 * Should be moved to reassociation?
4957 static ir_node *transform_node_shift(ir_node *n) {
4958 ir_node *left, *right;
4960 tarval *tv1, *tv2, *res;
4961 ir_node *in[2], *irn, *block;
4963 left = get_binop_left(n);
4965 /* different operations */
4966 if (get_irn_op(left) != get_irn_op(n))
4969 right = get_binop_right(n);
4970 tv1 = value_of(right);
4971 if (tv1 == tarval_bad)
4974 tv2 = value_of(get_binop_right(left));
4975 if (tv2 == tarval_bad)
4978 res = tarval_add(tv1, tv2);
4979 mode = get_irn_mode(n);
4981 /* beware: a simple replacement works only, if res < modulo shift */
4983 int modulo_shf = get_mode_modulo_shift(mode);
4984 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4985 if (modulo_shf > 0) {
4986 tarval *modulo = new_tarval_from_long(modulo_shf,
4987 get_tarval_mode(res));
4989 /* shifting too much */
4990 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4992 ir_graph *irg = get_irn_irg(n);
4993 ir_node *block = get_nodes_block(n);
4994 dbg_info *dbgi = get_irn_dbg_info(n);
4995 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
4996 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5000 return new_Const(mode, get_mode_null(mode));
5004 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5007 /* ok, we can replace it */
5008 block = get_nodes_block(n);
5010 in[0] = get_binop_left(left);
5011 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5013 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5015 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5017 return transform_node(irn);
5018 } /* transform_node_shift */
5021 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5023 * - and, or, xor instead of &
5024 * - Shl, Shr, Shrs, rotl instead of >>
5025 * (with a special case for Or/Xor + Shrs)
5027 static ir_node *transform_node_bitop_shift(ir_node *n) {
5029 ir_node *right = get_binop_right(n);
5030 ir_mode *mode = get_irn_mode(n);
5031 ir_node *bitop_left;
5032 ir_node *bitop_right;
5044 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5046 if (!is_Const(right))
5049 left = get_binop_left(n);
5050 op_left = get_irn_op(left);
5051 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5054 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5055 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5056 /* TODO: test if sign bit is affectes */
5060 bitop_right = get_binop_right(left);
5061 if (!is_Const(bitop_right))
5064 bitop_left = get_binop_left(left);
5066 irg = get_irn_irg(n);
5067 block = get_nodes_block(n);
5068 dbgi = get_irn_dbg_info(n);
5069 tv1 = get_Const_tarval(bitop_right);
5070 tv2 = get_Const_tarval(right);
5072 assert(get_tarval_mode(tv1) == mode);
5075 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5076 tv_shift = tarval_shl(tv1, tv2);
5077 } else if(is_Shr(n)) {
5078 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5079 tv_shift = tarval_shr(tv1, tv2);
5080 } else if(is_Shrs(n)) {
5081 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5082 tv_shift = tarval_shrs(tv1, tv2);
5085 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5086 tv_shift = tarval_rotl(tv1, tv2);
5089 assert(get_tarval_mode(tv_shift) == mode);
5090 new_const = new_Const(mode, tv_shift);
5092 if (op_left == op_And) {
5093 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5094 } else if(op_left == op_Or) {
5095 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5097 assert(op_left == op_Eor);
5098 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5106 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5108 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5109 * (also with x >>s c1 when c1>=c2)
5111 static ir_node *transform_node_shl_shr(ir_node *n) {
5113 ir_node *right = get_binop_right(n);
5129 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5131 if (!is_Const(right))
5134 left = get_binop_left(n);
5135 mode = get_irn_mode(n);
5136 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5137 ir_node *shr_right = get_binop_right(left);
5139 if (!is_Const(shr_right))
5142 x = get_binop_left(left);
5143 tv_shr = get_Const_tarval(shr_right);
5144 tv_shl = get_Const_tarval(right);
5146 if (is_Shrs(left)) {
5147 /* shrs variant only allowed if c1 >= c2 */
5148 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5151 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5154 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5156 tv_mask = tarval_shl(tv_mask, tv_shl);
5157 } else if(is_Shr(n) && is_Shl(left)) {
5158 ir_node *shl_right = get_Shl_right(left);
5160 if (!is_Const(shl_right))
5163 x = get_Shl_left(left);
5164 tv_shr = get_Const_tarval(right);
5165 tv_shl = get_Const_tarval(shl_right);
5167 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5168 tv_mask = tarval_shr(tv_mask, tv_shr);
5173 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5174 assert(tv_mask != tarval_bad);
5175 assert(get_tarval_mode(tv_mask) == mode);
5177 irg = get_irn_irg(n);
5178 block = get_nodes_block(n);
5179 dbgi = get_irn_dbg_info(n);
5181 pnc = tarval_cmp(tv_shl, tv_shr);
5182 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5183 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5184 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5186 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5188 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5191 assert(pnc == pn_Cmp_Gt);
5192 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5193 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5194 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5197 new_const = new_Const(mode, tv_mask);
5198 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5206 static ir_node *transform_node_Shr(ir_node *n) {
5207 ir_node *c, *oldn = n;
5208 ir_node *left = get_Shr_left(n);
5209 ir_node *right = get_Shr_right(n);
5210 ir_mode *mode = get_irn_mode(n);
5212 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5213 n = transform_node_shift(n);
5216 n = transform_node_shl_shr(n);
5218 n = transform_node_bitop_shift(n);
5221 } /* transform_node_Shr */
5226 static ir_node *transform_node_Shrs(ir_node *n) {
5227 ir_node *c, *oldn = n;
5228 ir_node *a = get_Shrs_left(n);
5229 ir_node *b = get_Shrs_right(n);
5230 ir_mode *mode = get_irn_mode(n);
5232 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5233 n = transform_node_shift(n);
5236 n = transform_node_bitop_shift(n);
5239 } /* transform_node_Shrs */
5244 static ir_node *transform_node_Shl(ir_node *n) {
5245 ir_node *c, *oldn = n;
5246 ir_node *a = get_Shl_left(n);
5247 ir_node *b = get_Shl_right(n);
5248 ir_mode *mode = get_irn_mode(n);
5250 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5251 n = transform_node_shift(n);
5254 n = transform_node_shl_shr(n);
5256 n = transform_node_bitop_shift(n);
5259 } /* transform_node_Shl */
5264 static ir_node *transform_node_Rotl(ir_node *n) {
5265 ir_node *c, *oldn = n;
5266 ir_node *a = get_Rotl_left(n);
5267 ir_node *b = get_Rotl_right(n);
5268 ir_mode *mode = get_irn_mode(n);
5270 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5271 n = transform_node_shift(n);
5274 n = transform_node_bitop_shift(n);
5277 } /* transform_node_Rotl */
5282 static ir_node *transform_node_Conv(ir_node *n) {
5283 ir_node *c, *oldn = n;
5284 ir_mode *mode = get_irn_mode(n);
5285 ir_node *a = get_Conv_op(n);
5287 if (mode != mode_b && is_const_Phi(a)) {
5288 /* Do NOT optimize mode_b Conv's, this leads to remaining
5289 * Phib nodes later, because the conv_b_lower operation
5290 * is instantly reverted, when it tries to insert a Convb.
5292 c = apply_conv_on_phi(a, mode);
5294 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5299 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5300 return new_r_Unknown(current_ir_graph, mode);
5303 if (mode_is_reference(mode) &&
5304 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5306 ir_node *l = get_Add_left(a);
5307 ir_node *r = get_Add_right(a);
5308 dbg_info *dbgi = get_irn_dbg_info(a);
5309 ir_node *block = get_nodes_block(n);
5311 ir_node *lop = get_Conv_op(l);
5312 if(get_irn_mode(lop) == mode) {
5313 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5314 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5319 ir_node *rop = get_Conv_op(r);
5320 if(get_irn_mode(rop) == mode) {
5321 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5322 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5329 } /* transform_node_Conv */
5332 * Remove dead blocks and nodes in dead blocks
5333 * in keep alive list. We do not generate a new End node.
5335 static ir_node *transform_node_End(ir_node *n) {
5336 int i, j, n_keepalives = get_End_n_keepalives(n);
5339 NEW_ARR_A(ir_node *, in, n_keepalives);
5341 for (i = j = 0; i < n_keepalives; ++i) {
5342 ir_node *ka = get_End_keepalive(n, i);
5344 if (! is_Block_dead(ka)) {
5348 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5350 } else if (is_Bad(ka)) {
5351 /* no need to keep Bad */
5356 if (j != n_keepalives)
5357 set_End_keepalives(n, j, in);
5359 } /* transform_node_End */
5361 /** returns 1 if a == -b */
5362 static int is_negated_value(ir_node *a, ir_node *b) {
5363 if (is_Minus(a) && get_Minus_op(a) == b)
5365 if (is_Minus(b) && get_Minus_op(b) == a)
5367 if (is_Sub(a) && is_Sub(b)) {
5368 ir_node *a_left = get_Sub_left(a);
5369 ir_node *a_right = get_Sub_right(a);
5370 ir_node *b_left = get_Sub_left(b);
5371 ir_node *b_right = get_Sub_right(b);
5373 if (a_left == b_right && a_right == b_left)
5381 * Optimize a Mux into some simpler cases.
5383 static ir_node *transform_node_Mux(ir_node *n) {
5384 ir_node *oldn = n, *sel = get_Mux_sel(n);
5385 ir_mode *mode = get_irn_mode(n);
5386 ir_node *t = get_Mux_true(n);
5387 ir_node *f = get_Mux_false(n);
5388 ir_graph *irg = current_ir_graph;
5390 /* first normalization step: move a possible zero to the false case */
5392 ir_node *cmp = get_Proj_pred(sel);
5395 if (is_Const(t) && is_Const_null(t)) {
5398 /* Mux(x, 0, y) => Mux(x, y, 0) */
5399 pn_Cmp pnc = get_Proj_proj(sel);
5400 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5401 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5402 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5410 /* note: after normalization, false can only happen on default */
5411 if (mode == mode_b) {
5412 dbg_info *dbg = get_irn_dbg_info(n);
5413 ir_node *block = get_nodes_block(n);
5414 ir_graph *irg = current_ir_graph;
5417 tarval *tv_t = get_Const_tarval(t);
5418 if (tv_t == tarval_b_true) {
5420 /* Muxb(sel, true, false) = sel */
5421 assert(get_Const_tarval(f) == tarval_b_false);
5422 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5425 /* Muxb(sel, true, x) = Or(sel, x) */
5426 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5427 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5431 } else if (is_Const(f)) {
5432 tarval *tv_f = get_Const_tarval(f);
5433 if (tv_f == tarval_b_true) {
5434 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5435 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5436 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5437 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5440 /* Muxb(sel, x, false) = And(sel, x) */
5441 assert(tv_f == tarval_b_false);
5442 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5443 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5449 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5450 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5451 tarval *a = get_Const_tarval(t);
5452 tarval *b = get_Const_tarval(f);
5453 tarval *null = get_tarval_null(mode);
5456 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5457 diff = tarval_sub(a, b, NULL);
5460 diff = tarval_sub(b, a, NULL);
5464 if (diff == get_tarval_one(mode) && min != null) {
5465 dbg_info *dbg = get_irn_dbg_info(n);
5466 ir_node *block = get_nodes_block(n);
5467 ir_graph *irg = current_ir_graph;
5468 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5469 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5470 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5471 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5477 ir_node *cmp = get_Proj_pred(sel);
5478 long pn = get_Proj_proj(sel);
5481 * Note: normalization puts the constant on the right side,
5482 * so we check only one case.
5484 * Note further that these optimization work even for floating point
5485 * with NaN's because -NaN == NaN.
5486 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5490 ir_node *cmp_r = get_Cmp_right(cmp);
5491 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5492 ir_node *block = get_nodes_block(n);
5493 ir_node *cmp_l = get_Cmp_left(cmp);
5495 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5498 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5499 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5501 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5502 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5504 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5506 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5507 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5509 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5510 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5512 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5514 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5519 if (mode_is_int(mode)) {
5521 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5522 /* Mux((a & b) != 0, c, 0) */
5523 ir_node *and_r = get_And_right(cmp_l);
5526 if (and_r == t && f == cmp_r) {
5527 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5528 if (pn == pn_Cmp_Lg) {
5529 /* Mux((a & 2^C) != 0, 2^C, 0) */
5531 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5533 /* Mux((a & 2^C) == 0, 2^C, 0) */
5534 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5535 block, cmp_l, t, mode);
5536 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5541 if (is_Shl(and_r)) {
5542 ir_node *shl_l = get_Shl_left(and_r);
5543 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5544 if (and_r == t && f == cmp_r) {
5545 if (pn == pn_Cmp_Lg) {
5546 /* (a & (1 << n)) != 0, (1 << n), 0) */
5548 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5550 /* (a & (1 << n)) == 0, (1 << n), 0) */
5551 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5552 block, cmp_l, t, mode);
5553 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5559 and_l = get_And_left(cmp_l);
5560 if (is_Shl(and_l)) {
5561 ir_node *shl_l = get_Shl_left(and_l);
5562 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5563 if (and_l == t && f == cmp_r) {
5564 if (pn == pn_Cmp_Lg) {
5565 /* ((1 << n) & a) != 0, (1 << n), 0) */
5567 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5569 /* ((1 << n) & a) == 0, (1 << n), 0) */
5570 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5571 block, cmp_l, t, mode);
5572 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5583 return arch_transform_node_Mux(n);
5584 } /* transform_node_Mux */
5587 * optimize Sync nodes that have other syncs as input we simply add the inputs
5588 * of the other sync to our own inputs
5590 static ir_node *transform_node_Sync(ir_node *n) {
5591 int arity = get_Sync_n_preds(n);
5594 for (i = 0; i < arity;) {
5595 ir_node *pred = get_Sync_pred(n, i);
5599 if (!is_Sync(pred)) {
5607 pred_arity = get_Sync_n_preds(pred);
5608 for (j = 0; j < pred_arity; ++j) {
5609 ir_node *pred_pred = get_Sync_pred(pred, j);
5614 add_irn_n(n, pred_pred);
5618 if (get_Sync_pred(n, k) == pred_pred) break;
5623 /* rehash the sync node */
5624 add_identities(current_ir_graph->value_table, n);
5630 * Tries several [inplace] [optimizing] transformations and returns an
5631 * equivalent node. The difference to equivalent_node() is that these
5632 * transformations _do_ generate new nodes, and thus the old node must
5633 * not be freed even if the equivalent node isn't the old one.
5635 static ir_node *transform_node(ir_node *n) {
5639 * Transform_node is the only "optimizing transformation" that might
5640 * return a node with a different opcode. We iterate HERE until fixpoint
5641 * to get the final result.
5645 if (n->op->ops.transform_node)
5646 n = n->op->ops.transform_node(n);
5647 } while (oldn != n);
5650 } /* transform_node */
5653 * Sets the default transform node operation for an ir_op_ops.
5655 * @param code the opcode for the default operation
5656 * @param ops the operations initialized
5661 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5665 ops->transform_node = transform_node_##a; \
5667 #define CASE_PROJ(a) \
5669 ops->transform_node_Proj = transform_node_Proj_##a; \
5671 #define CASE_PROJ_EX(a) \
5673 ops->transform_node = transform_node_##a; \
5674 ops->transform_node_Proj = transform_node_Proj_##a; \
5683 CASE_PROJ_EX(DivMod);
5717 } /* firm_set_default_transform_node */
5720 /* **************** Common Subexpression Elimination **************** */
5722 /** The size of the hash table used, should estimate the number of nodes
5724 #define N_IR_NODES 512
5726 /** Compares the attributes of two Const nodes. */
5727 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5728 return (get_Const_tarval(a) != get_Const_tarval(b))
5729 || (get_Const_type(a) != get_Const_type(b));
5730 } /* node_cmp_attr_Const */
5732 /** Compares the attributes of two Proj nodes. */
5733 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5734 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5735 } /* node_cmp_attr_Proj */
5737 /** Compares the attributes of two Filter nodes. */
5738 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5739 return get_Filter_proj(a) != get_Filter_proj(b);
5740 } /* node_cmp_attr_Filter */
5742 /** Compares the attributes of two Alloc nodes. */
5743 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5744 const alloc_attr *pa = get_irn_alloc_attr(a);
5745 const alloc_attr *pb = get_irn_alloc_attr(b);
5746 return (pa->where != pb->where) || (pa->type != pb->type);
5747 } /* node_cmp_attr_Alloc */
5749 /** Compares the attributes of two Free nodes. */
5750 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5751 const free_attr *pa = get_irn_free_attr(a);
5752 const free_attr *pb = get_irn_free_attr(b);
5753 return (pa->where != pb->where) || (pa->type != pb->type);
5754 } /* node_cmp_attr_Free */
5756 /** Compares the attributes of two SymConst nodes. */
5757 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5758 const symconst_attr *pa = get_irn_symconst_attr(a);
5759 const symconst_attr *pb = get_irn_symconst_attr(b);
5760 return (pa->kind != pb->kind)
5761 || (pa->sym.type_p != pb->sym.type_p)
5762 || (pa->tp != pb->tp);
5763 } /* node_cmp_attr_SymConst */
5765 /** Compares the attributes of two Call nodes. */
5766 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5767 return get_irn_call_attr(a) != get_irn_call_attr(b);
5768 } /* node_cmp_attr_Call */
5770 /** Compares the attributes of two Sel nodes. */
5771 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5772 const ir_entity *a_ent = get_Sel_entity(a);
5773 const ir_entity *b_ent = get_Sel_entity(b);
5776 (a_ent->kind != b_ent->kind) ||
5777 (a_ent->name != b_ent->name) ||
5778 (a_ent->owner != b_ent->owner) ||
5779 (a_ent->ld_name != b_ent->ld_name) ||
5780 (a_ent->type != b_ent->type);
5782 /* Matze: inlining of functions can produce 2 entities with same type,
5784 return a_ent != b_ent;
5785 } /* node_cmp_attr_Sel */
5787 /** Compares the attributes of two Phi nodes. */
5788 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5789 /* we can only enter this function if both nodes have the same number of inputs,
5790 hence it is enough to check if one of them is a Phi0 */
5792 /* check the Phi0 pos attribute */
5793 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5796 } /* node_cmp_attr_Phi */
5798 /** Compares the attributes of two Conv nodes. */
5799 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5800 return get_Conv_strict(a) != get_Conv_strict(b);
5801 } /* node_cmp_attr_Conv */
5803 /** Compares the attributes of two Cast nodes. */
5804 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5805 return get_Cast_type(a) != get_Cast_type(b);
5806 } /* node_cmp_attr_Cast */
5808 /** Compares the attributes of two Load nodes. */
5809 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5810 if (get_Load_volatility(a) == volatility_is_volatile ||
5811 get_Load_volatility(b) == volatility_is_volatile)
5812 /* NEVER do CSE on volatile Loads */
5814 /* do not CSE Loads with different alignment. Be conservative. */
5815 if (get_Load_align(a) != get_Load_align(b))
5818 return get_Load_mode(a) != get_Load_mode(b);
5819 } /* node_cmp_attr_Load */
5821 /** Compares the attributes of two Store nodes. */
5822 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5823 /* do not CSE Stores with different alignment. Be conservative. */
5824 if (get_Store_align(a) != get_Store_align(b))
5827 /* NEVER do CSE on volatile Stores */
5828 return (get_Store_volatility(a) == volatility_is_volatile ||
5829 get_Store_volatility(b) == volatility_is_volatile);
5830 } /* node_cmp_attr_Store */
5832 /** Compares two exception attributes */
5833 static int node_cmp_exception(ir_node *a, ir_node *b) {
5834 const except_attr *ea = get_irn_except_attr(a);
5835 const except_attr *eb = get_irn_except_attr(b);
5837 return ea->pin_state != eb->pin_state;
5840 #define node_cmp_attr_Bound node_cmp_exception
5842 /** Compares the attributes of two Div nodes. */
5843 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5844 const divmod_attr *ma = get_irn_divmod_attr(a);
5845 const divmod_attr *mb = get_irn_divmod_attr(b);
5846 return ma->exc.pin_state != mb->exc.pin_state ||
5847 ma->res_mode != mb->res_mode ||
5848 ma->no_remainder != mb->no_remainder;
5849 } /* node_cmp_attr_Div */
5851 /** Compares the attributes of two DivMod nodes. */
5852 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5853 const divmod_attr *ma = get_irn_divmod_attr(a);
5854 const divmod_attr *mb = get_irn_divmod_attr(b);
5855 return ma->exc.pin_state != mb->exc.pin_state ||
5856 ma->res_mode != mb->res_mode;
5857 } /* node_cmp_attr_DivMod */
5859 /** Compares the attributes of two Mod nodes. */
5860 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5861 const divmod_attr *ma = get_irn_divmod_attr(a);
5862 const divmod_attr *mb = get_irn_divmod_attr(b);
5863 return ma->exc.pin_state != mb->exc.pin_state ||
5864 ma->res_mode != mb->res_mode;
5865 } /* node_cmp_attr_Mod */
5867 /** Compares the attributes of two Quot nodes. */
5868 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5869 const divmod_attr *ma = get_irn_divmod_attr(a);
5870 const divmod_attr *mb = get_irn_divmod_attr(b);
5871 return ma->exc.pin_state != mb->exc.pin_state ||
5872 ma->res_mode != mb->res_mode;
5873 } /* node_cmp_attr_Quot */
5875 /** Compares the attributes of two Confirm nodes. */
5876 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5877 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5878 } /* node_cmp_attr_Confirm */
5880 /** Compares the attributes of two ASM nodes. */
5881 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5883 const ir_asm_constraint *ca;
5884 const ir_asm_constraint *cb;
5887 if (get_ASM_text(a) != get_ASM_text(b))
5890 /* Should we really check the constraints here? Should be better, but is strange. */
5891 n = get_ASM_n_input_constraints(a);
5892 if (n != get_ASM_n_input_constraints(b))
5895 ca = get_ASM_input_constraints(a);
5896 cb = get_ASM_input_constraints(b);
5897 for (i = 0; i < n; ++i) {
5898 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5902 n = get_ASM_n_output_constraints(a);
5903 if (n != get_ASM_n_output_constraints(b))
5906 ca = get_ASM_output_constraints(a);
5907 cb = get_ASM_output_constraints(b);
5908 for (i = 0; i < n; ++i) {
5909 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5913 n = get_ASM_n_clobbers(a);
5914 if (n != get_ASM_n_clobbers(b))
5917 cla = get_ASM_clobbers(a);
5918 clb = get_ASM_clobbers(b);
5919 for (i = 0; i < n; ++i) {
5920 if (cla[i] != clb[i])
5924 } /* node_cmp_attr_ASM */
5927 * Set the default node attribute compare operation for an ir_op_ops.
5929 * @param code the opcode for the default operation
5930 * @param ops the operations initialized
5935 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5939 ops->node_cmp_attr = node_cmp_attr_##a; \
5970 } /* firm_set_default_node_cmp_attr */
5973 * Compare function for two nodes in the value table. Gets two
5974 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5976 int identities_cmp(const void *elt, const void *key) {
5977 ir_node *a = (ir_node *)elt;
5978 ir_node *b = (ir_node *)key;
5981 if (a == b) return 0;
5983 if ((get_irn_op(a) != get_irn_op(b)) ||
5984 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5986 /* compare if a's in and b's in are of equal length */
5987 irn_arity_a = get_irn_intra_arity(a);
5988 if (irn_arity_a != get_irn_intra_arity(b))
5991 if (get_irn_pinned(a) == op_pin_state_pinned) {
5992 /* for pinned nodes, the block inputs must be equal */
5993 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5995 } else if (! get_opt_global_cse()) {
5996 /* for block-local CSE both nodes must be in the same MacroBlock */
5997 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6001 /* compare a->in[0..ins] with b->in[0..ins] */
6002 for (i = 0; i < irn_arity_a; i++)
6003 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6007 * here, we already now that the nodes are identical except their
6010 if (a->op->ops.node_cmp_attr)
6011 return a->op->ops.node_cmp_attr(a, b);
6014 } /* identities_cmp */
6017 * Calculate a hash value of a node.
6019 * @param node The IR-node
6021 unsigned ir_node_hash(const ir_node *node) {
6022 return node->op->ops.hash(node);
6023 } /* ir_node_hash */
6026 pset *new_identities(void) {
6027 return new_pset(identities_cmp, N_IR_NODES);
6028 } /* new_identities */
6030 void del_identities(pset *value_table) {
6031 del_pset(value_table);
6032 } /* del_identities */
6035 * Normalize a node by putting constants (and operands with larger
6036 * node index) on the right (operator side).
6038 * @param n The node to normalize
6040 static void normalize_node(ir_node *n) {
6041 if (is_op_commutative(get_irn_op(n))) {
6042 ir_node *l = get_binop_left(n);
6043 ir_node *r = get_binop_right(n);
6045 /* For commutative operators perform a OP b == b OP a but keep
6046 * constants on the RIGHT side. This helps greatly in some
6047 * optimizations. Moreover we use the idx number to make the form
6049 if (!operands_are_normalized(l, r)) {
6050 set_binop_left(n, r);
6051 set_binop_right(n, l);
6054 } /* normalize_node */
6057 * Update the nodes after a match in the value table. If both nodes have
6058 * the same MacroBlock but different Blocks, we must ensure that the node
6059 * with the dominating Block (the node that is near to the MacroBlock header
6060 * is stored in the table.
6061 * Because a MacroBlock has only one "non-exception" flow, we don't need
6062 * dominance info here: We known, that one block must dominate the other and
6063 * following the only block input will allow to find it.
6065 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6066 ir_node *known_blk, *new_block, *block, *mbh;
6068 if (get_opt_global_cse()) {
6069 /* Block inputs are meaning less */
6072 known_blk = get_irn_n(known_irn, -1);
6073 new_block = get_irn_n(new_ir_node, -1);
6074 if (known_blk == new_block) {
6075 /* already in the same block */
6079 * We expect the typical case when we built the graph. In that case, the
6080 * known_irn is already the upper one, so checking this should be faster.
6083 mbh = get_Block_MacroBlock(new_block);
6085 if (block == known_blk) {
6086 /* ok, we have found it: known_block dominates new_block as expected */
6091 * We have reached the MacroBlock header NOT founding
6092 * the known_block. new_block must dominate known_block.
6095 set_irn_n(known_irn, -1, new_block);
6098 assert(get_Block_n_cfgpreds(block) == 1);
6099 block = get_Block_cfgpred_block(block, 0);
6101 } /* update_value_table */
6104 * Return the canonical node computing the same value as n.
6105 * Looks up the node in a hash table, enters it in the table
6106 * if it isn't there yet.
6108 * @param value_table the HashSet containing all nodes in the
6110 * @param n the node to look up
6112 * @return a node that computes the same value as n or n if no such
6113 * node could be found
6115 ir_node *identify_remember(pset *value_table, ir_node *n) {
6118 if (!value_table) return n;
6121 /* lookup or insert in hash table with given hash key. */
6122 o = pset_insert(value_table, n, ir_node_hash(n));
6125 update_known_irn(o, n);
6129 } /* identify_remember */
6132 * During construction we set the op_pin_state_pinned flag in the graph right when the
6133 * optimization is performed. The flag turning on procedure global cse could
6134 * be changed between two allocations. This way we are safe.
6136 * @param value_table The value table
6137 * @param n The node to lookup
6139 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
6142 n = identify_remember(value_table, n);
6143 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6144 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6146 } /* identify_cons */
6148 /* Add a node to the identities value table. */
6149 void add_identities(pset *value_table, ir_node *node) {
6150 if (get_opt_cse() && is_no_Block(node))
6151 identify_remember(value_table, node);
6152 } /* add_identities */
6154 /* Visit each node in the value table of a graph. */
6155 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6157 ir_graph *rem = current_ir_graph;
6159 current_ir_graph = irg;
6160 foreach_pset(irg->value_table, node)
6162 current_ir_graph = rem;
6163 } /* visit_all_identities */
6166 * Garbage in, garbage out. If a node has a dead input, i.e., the
6167 * Bad node is input to the node, return the Bad node.
6169 static ir_node *gigo(ir_node *node) {
6171 ir_op *op = get_irn_op(node);
6173 /* remove garbage blocks by looking at control flow that leaves the block
6174 and replacing the control flow by Bad. */
6175 if (get_irn_mode(node) == mode_X) {
6176 ir_node *block = get_nodes_block(skip_Proj(node));
6178 /* Don't optimize nodes in immature blocks. */
6179 if (!get_Block_matured(block))
6181 /* Don't optimize End, may have Bads. */
6182 if (op == op_End) return node;
6184 if (is_Block(block)) {
6185 if (is_Block_dead(block)) {
6186 /* control flow from dead block is dead */
6190 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6191 if (!is_Bad(get_irn_n(block, i)))
6195 ir_graph *irg = get_irn_irg(block);
6196 /* the start block is never dead */
6197 if (block != get_irg_start_block(irg)
6198 && block != get_irg_end_block(irg)) {
6200 * Do NOT kill control flow without setting
6201 * the block to dead of bad things can happen:
6202 * We get a Block that is not reachable be irg_block_walk()
6203 * but can be found by irg_walk()!
6205 set_Block_dead(block);
6212 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6213 blocks predecessors is dead. */
6214 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6215 irn_arity = get_irn_arity(node);
6218 * Beware: we can only read the block of a non-floating node.
6220 if (is_irn_pinned_in_irg(node) &&
6221 is_Block_dead(get_nodes_block(skip_Proj(node))))
6224 for (i = 0; i < irn_arity; i++) {
6225 ir_node *pred = get_irn_n(node, i);
6230 /* Propagating Unknowns here seems to be a bad idea, because
6231 sometimes we need a node as a input and did not want that
6233 However, it might be useful to move this into a later phase
6234 (if you think that optimizing such code is useful). */
6235 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6236 return new_Unknown(get_irn_mode(node));
6241 /* With this code we violate the agreement that local_optimize
6242 only leaves Bads in Block, Phi and Tuple nodes. */
6243 /* If Block has only Bads as predecessors it's garbage. */
6244 /* If Phi has only Bads as predecessors it's garbage. */
6245 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6246 irn_arity = get_irn_arity(node);
6247 for (i = 0; i < irn_arity; i++) {
6248 if (!is_Bad(get_irn_n(node, i))) break;
6250 if (i == irn_arity) node = new_Bad();
6257 * These optimizations deallocate nodes from the obstack.
6258 * It can only be called if it is guaranteed that no other nodes
6259 * reference this one, i.e., right after construction of a node.
6261 * @param n The node to optimize
6263 * current_ir_graph must be set to the graph of the node!
6265 ir_node *optimize_node(ir_node *n) {
6268 ir_opcode iro = get_irn_opcode(n);
6270 /* Always optimize Phi nodes: part of the construction. */
6271 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6273 /* constant expression evaluation / constant folding */
6274 if (get_opt_constant_folding()) {
6275 /* neither constants nor Tuple values can be evaluated */
6276 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6277 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6278 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6279 /* try to evaluate */
6280 tv = computed_value(n);
6281 if (tv != tarval_bad) {
6283 ir_type *old_tp = get_irn_type(n);
6284 int i, arity = get_irn_arity(n);
6288 * Try to recover the type of the new expression.
6290 for (i = 0; i < arity && !old_tp; ++i)
6291 old_tp = get_irn_type(get_irn_n(n, i));
6294 * we MUST copy the node here temporary, because it's still needed
6295 * for DBG_OPT_CSTEVAL
6297 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6298 oldn = alloca(node_size);
6300 memcpy(oldn, n, node_size);
6301 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6303 /* ARG, copy the in array, we need it for statistics */
6304 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6306 /* note the inplace edges module */
6307 edges_node_deleted(n, current_ir_graph);
6309 /* evaluation was successful -- replace the node. */
6310 irg_kill_node(current_ir_graph, n);
6311 nw = new_Const(get_tarval_mode(tv), tv);
6313 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6314 set_Const_type(nw, old_tp);
6315 DBG_OPT_CSTEVAL(oldn, nw);
6316 tarval_enable_fp_ops(old_fp_mode);
6319 tarval_enable_fp_ops(old_fp_mode);
6323 /* remove unnecessary nodes */
6324 if (get_opt_algebraic_simplification() ||
6325 (iro == iro_Phi) || /* always optimize these nodes. */
6327 (iro == iro_Proj) ||
6328 (iro == iro_Block) ) /* Flags tested local. */
6329 n = equivalent_node(n);
6331 /* Common Subexpression Elimination.
6333 * Checks whether n is already available.
6334 * The block input is used to distinguish different subexpressions. Right
6335 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6336 * subexpressions within a block.
6339 n = identify_cons(current_ir_graph->value_table, n);
6342 edges_node_deleted(oldn, current_ir_graph);
6344 /* We found an existing, better node, so we can deallocate the old node. */
6345 irg_kill_node(current_ir_graph, oldn);
6349 /* Some more constant expression evaluation that does not allow to
6351 iro = get_irn_opcode(n);
6352 if (get_opt_algebraic_simplification() ||
6353 (iro == iro_Cond) ||
6354 (iro == iro_Proj)) /* Flags tested local. */
6355 n = transform_node(n);
6357 /* Remove nodes with dead (Bad) input.
6358 Run always for transformation induced Bads. */
6361 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6362 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6364 n = identify_remember(current_ir_graph->value_table, o);
6370 } /* optimize_node */
6374 * These optimizations never deallocate nodes (in place). This can cause dead
6375 * nodes lying on the obstack. Remove these by a dead node elimination,
6376 * i.e., a copying garbage collection.
6378 ir_node *optimize_in_place_2(ir_node *n) {
6381 ir_opcode iro = get_irn_opcode(n);
6383 if (!get_opt_optimize() && !is_Phi(n)) return n;
6385 /* constant expression evaluation / constant folding */
6386 if (get_opt_constant_folding()) {
6387 /* neither constants nor Tuple values can be evaluated */
6388 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6389 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6390 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6391 /* try to evaluate */
6392 tv = computed_value(n);
6393 if (tv != tarval_bad) {
6394 /* evaluation was successful -- replace the node. */
6395 ir_type *old_tp = get_irn_type(n);
6396 int i, arity = get_irn_arity(n);
6399 * Try to recover the type of the new expression.
6401 for (i = 0; i < arity && !old_tp; ++i)
6402 old_tp = get_irn_type(get_irn_n(n, i));
6404 n = new_Const(get_tarval_mode(tv), tv);
6406 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6407 set_Const_type(n, old_tp);
6409 DBG_OPT_CSTEVAL(oldn, n);
6410 tarval_enable_fp_ops(old_fp_mode);
6413 tarval_enable_fp_ops(old_fp_mode);
6417 /* remove unnecessary nodes */
6418 if (get_opt_constant_folding() ||
6419 (iro == iro_Phi) || /* always optimize these nodes. */
6420 (iro == iro_Id) || /* ... */
6421 (iro == iro_Proj) || /* ... */
6422 (iro == iro_Block) ) /* Flags tested local. */
6423 n = equivalent_node(n);
6425 /** common subexpression elimination **/
6426 /* Checks whether n is already available. */
6427 /* The block input is used to distinguish different subexpressions. Right
6428 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6429 subexpressions within a block. */
6430 if (get_opt_cse()) {
6432 n = identify_remember(current_ir_graph->value_table, o);
6437 /* Some more constant expression evaluation. */
6438 iro = get_irn_opcode(n);
6439 if (get_opt_constant_folding() ||
6440 (iro == iro_Cond) ||
6441 (iro == iro_Proj)) /* Flags tested local. */
6442 n = transform_node(n);
6444 /* Remove nodes with dead (Bad) input.
6445 Run always for transformation induced Bads. */
6448 /* Now we can verify the node, as it has no dead inputs any more. */
6451 /* Now we have a legal, useful node. Enter it in hash table for cse.
6452 Blocks should be unique anyways. (Except the successor of start:
6453 is cse with the start block!) */
6454 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6456 n = identify_remember(current_ir_graph->value_table, o);
6462 } /* optimize_in_place_2 */
6465 * Wrapper for external use, set proper status bits after optimization.
6467 ir_node *optimize_in_place(ir_node *n) {
6468 /* Handle graph state */
6469 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6471 if (get_opt_global_cse())
6472 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6473 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6474 set_irg_outs_inconsistent(current_ir_graph);
6476 /* FIXME: Maybe we could also test whether optimizing the node can
6477 change the control graph. */
6478 set_irg_doms_inconsistent(current_ir_graph);
6479 return optimize_in_place_2(n);
6480 } /* optimize_in_place */
6483 * Calculate a hash value of a Const node.
6485 static unsigned hash_Const(const ir_node *node) {
6488 /* special value for const, as they only differ in their tarval. */
6489 h = HASH_PTR(node->attr.con.tv);
6490 h = 9*h + HASH_PTR(get_irn_mode(node));
6496 * Calculate a hash value of a SymConst node.
6498 static unsigned hash_SymConst(const ir_node *node) {
6501 /* special value for const, as they only differ in their symbol. */
6502 h = HASH_PTR(node->attr.symc.sym.type_p);
6503 h = 9*h + HASH_PTR(get_irn_mode(node));
6506 } /* hash_SymConst */
6509 * Set the default hash operation in an ir_op_ops.
6511 * @param code the opcode for the default operation
6512 * @param ops the operations initialized
6517 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6521 ops->hash = hash_##a; \
6524 /* hash function already set */
6525 if (ops->hash != NULL)
6532 /* use input/mode default hash if no function was given */
6533 ops->hash = firm_default_hash;
6541 * Sets the default operation for an ir_ops.
6543 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6544 ops = firm_set_default_hash(code, ops);
6545 ops = firm_set_default_computed_value(code, ops);
6546 ops = firm_set_default_equivalent_node(code, ops);
6547 ops = firm_set_default_transform_node(code, ops);
6548 ops = firm_set_default_node_cmp_attr(code, ops);
6549 ops = firm_set_default_get_type(code, ops);
6550 ops = firm_set_default_get_type_attr(code, ops);
6551 ops = firm_set_default_get_entity_attr(code, ops);
6554 } /* firm_set_default_operations */