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);
1255 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1256 if (smaller_mode(b_mode, a_mode)) {
1257 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1258 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1262 if (mode_is_int(n_mode) && mode_is_float(a_mode)) {
1263 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1264 size_t int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1265 size_t float_mantissa;
1266 /* FIXME There is no way to get the mantissa size of a mode */
1267 switch (get_mode_size_bits(a_mode)) {
1268 case 32: float_mantissa = 23 + 1; break; // + 1 for implicit 1
1269 case 64: float_mantissa = 52 + 1; break;
1270 case 80: float_mantissa = 64; break;
1271 default: float_mantissa = 0; break;
1273 if (float_mantissa >= int_mantissa) {
1275 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1280 if (smaller_mode(b_mode, a_mode)) {
1281 if (get_Conv_strict(n))
1282 set_Conv_strict(b, 1);
1283 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1284 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1290 } /* equivalent_node_Conv */
1293 * A Cast may be removed if the type of the previous node
1294 * is already the type of the Cast.
1296 static ir_node *equivalent_node_Cast(ir_node *n) {
1298 ir_node *pred = get_Cast_op(n);
1300 if (get_irn_type(pred) == get_Cast_type(n)) {
1302 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1305 } /* equivalent_node_Cast */
1308 * - fold Phi-nodes, iff they have only one predecessor except
1311 static ir_node *equivalent_node_Phi(ir_node *n) {
1316 ir_node *first_val = NULL; /* to shutup gcc */
1318 if (!get_opt_normalize()) return n;
1320 n_preds = get_Phi_n_preds(n);
1322 block = get_nodes_block(n);
1323 if (is_Block_dead(block)) /* Control dead */
1324 return get_irg_bad(current_ir_graph);
1326 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1328 /* Find first non-self-referencing input */
1329 for (i = 0; i < n_preds; ++i) {
1330 first_val = get_Phi_pred(n, i);
1331 if ( (first_val != n) /* not self pointer */
1333 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1334 * predecessors. Then, Phi nodes in dead code might be removed, causing
1335 * nodes pointing to themself (Add's for instance).
1336 * This is really bad and causes endless recursions in several
1337 * code pathes, so we do NOT optimize such a code.
1338 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1339 * (and bad Phi predecessors), so live code is optimized later.
1341 && (! is_Bad(get_Block_cfgpred(block, i)))
1343 ) { /* value not dead */
1344 break; /* then found first value. */
1349 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1350 return get_irg_bad(current_ir_graph);
1353 /* search for rest of inputs, determine if any of these
1354 are non-self-referencing */
1355 while (++i < n_preds) {
1356 ir_node *scnd_val = get_Phi_pred(n, i);
1357 if ( (scnd_val != n)
1358 && (scnd_val != first_val)
1361 && (! is_Bad(get_Block_cfgpred(block, i)))
1369 /* Fold, if no multiple distinct non-self-referencing inputs */
1371 DBG_OPT_PHI(oldn, n);
1374 } /* equivalent_node_Phi */
1377 * Several optimizations:
1378 * - fold Sync-nodes, iff they have only one predecessor except
1381 static ir_node *equivalent_node_Sync(ir_node *n) {
1382 int arity = get_Sync_n_preds(n);
1385 for (i = 0; i < arity;) {
1386 ir_node *pred = get_Sync_pred(n, i);
1389 /* Remove Bad predecessors */
1396 /* Remove duplicate predecessors */
1402 if (get_Sync_pred(n, j) == pred) {
1410 if (arity == 0) return get_irg_bad(current_ir_graph);
1411 if (arity == 1) return get_Sync_pred(n, 0);
1413 } /* equivalent_node_Sync */
1416 * Optimize Proj(Tuple).
1418 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1419 ir_node *oldn = proj;
1420 ir_node *tuple = get_Proj_pred(proj);
1422 /* Remove the Tuple/Proj combination. */
1423 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1424 DBG_OPT_TUPLE(oldn, tuple, proj);
1427 } /* equivalent_node_Proj_Tuple */
1430 * Optimize a / 1 = a.
1432 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1433 ir_node *oldn = proj;
1434 ir_node *div = get_Proj_pred(proj);
1435 ir_node *b = get_Div_right(div);
1436 tarval *tb = value_of(b);
1438 /* Div is not commutative. */
1439 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1440 switch (get_Proj_proj(proj)) {
1442 proj = get_Div_mem(div);
1443 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1447 proj = get_Div_left(div);
1448 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1452 /* we cannot replace the exception Proj's here, this is done in
1453 transform_node_Proj_Div() */
1458 } /* equivalent_node_Proj_Div */
1461 * Optimize a / 1.0 = a.
1463 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1464 ir_node *oldn = proj;
1465 ir_node *quot = get_Proj_pred(proj);
1466 ir_node *b = get_Quot_right(quot);
1467 tarval *tb = value_of(b);
1469 /* Div is not commutative. */
1470 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1471 switch (get_Proj_proj(proj)) {
1473 proj = get_Quot_mem(quot);
1474 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1478 proj = get_Quot_left(quot);
1479 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1483 /* we cannot replace the exception Proj's here, this is done in
1484 transform_node_Proj_Quot() */
1489 } /* equivalent_node_Proj_Quot */
1492 * Optimize a / 1 = a.
1494 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1495 ir_node *oldn = proj;
1496 ir_node *divmod = get_Proj_pred(proj);
1497 ir_node *b = get_DivMod_right(divmod);
1498 tarval *tb = value_of(b);
1500 /* Div is not commutative. */
1501 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1502 switch (get_Proj_proj(proj)) {
1504 proj = get_DivMod_mem(divmod);
1505 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1508 case pn_DivMod_res_div:
1509 proj = get_DivMod_left(divmod);
1510 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1514 /* we cannot replace the exception Proj's here, this is done in
1515 transform_node_Proj_DivMod().
1516 Note further that the pn_DivMod_res_div case is handled in
1517 computed_value_Proj(). */
1522 } /* equivalent_node_Proj_DivMod */
1525 * Optimize CopyB(mem, x, x) into a Nop.
1527 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1528 ir_node *oldn = proj;
1529 ir_node *copyb = get_Proj_pred(proj);
1530 ir_node *a = get_CopyB_dst(copyb);
1531 ir_node *b = get_CopyB_src(copyb);
1534 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1535 switch (get_Proj_proj(proj)) {
1536 case pn_CopyB_M_regular:
1537 proj = get_CopyB_mem(copyb);
1538 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1541 case pn_CopyB_M_except:
1542 case pn_CopyB_X_except:
1543 DBG_OPT_EXC_REM(proj);
1544 proj = get_irg_bad(current_ir_graph);
1549 } /* equivalent_node_Proj_CopyB */
1552 * Optimize Bounds(idx, idx, upper) into idx.
1554 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1555 ir_node *oldn = proj;
1556 ir_node *bound = get_Proj_pred(proj);
1557 ir_node *idx = get_Bound_index(bound);
1558 ir_node *pred = skip_Proj(idx);
1561 if (idx == get_Bound_lower(bound))
1563 else if (is_Bound(pred)) {
1565 * idx was Bounds checked in the same MacroBlock previously,
1566 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1568 ir_node *lower = get_Bound_lower(bound);
1569 ir_node *upper = get_Bound_upper(bound);
1570 if (get_Bound_lower(pred) == lower &&
1571 get_Bound_upper(pred) == upper &&
1572 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1574 * One could expect that we simply return the previous
1575 * Bound here. However, this would be wrong, as we could
1576 * add an exception Proj to a new location then.
1577 * So, we must turn in into a tuple.
1583 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1584 switch (get_Proj_proj(proj)) {
1586 DBG_OPT_EXC_REM(proj);
1587 proj = get_Bound_mem(bound);
1589 case pn_Bound_X_except:
1590 DBG_OPT_EXC_REM(proj);
1591 proj = get_irg_bad(current_ir_graph);
1595 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1598 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1603 } /* equivalent_node_Proj_Bound */
1606 * Optimize an Exception Proj(Load) with a non-null address.
1608 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1609 if (get_opt_ldst_only_null_ptr_exceptions()) {
1610 if (get_irn_mode(proj) == mode_X) {
1611 ir_node *load = get_Proj_pred(proj);
1613 /* get the Load address */
1614 const ir_node *addr = get_Load_ptr(load);
1615 const ir_node *confirm;
1617 if (value_not_null(addr, &confirm)) {
1618 if (get_Proj_proj(proj) == pn_Load_X_except) {
1619 DBG_OPT_EXC_REM(proj);
1620 return get_irg_bad(current_ir_graph);
1626 } /* equivalent_node_Proj_Load */
1629 * Optimize an Exception Proj(Store) with a non-null address.
1631 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1632 if (get_opt_ldst_only_null_ptr_exceptions()) {
1633 if (get_irn_mode(proj) == mode_X) {
1634 ir_node *store = get_Proj_pred(proj);
1636 /* get the load/store address */
1637 const ir_node *addr = get_Store_ptr(store);
1638 const ir_node *confirm;
1640 if (value_not_null(addr, &confirm)) {
1641 if (get_Proj_proj(proj) == pn_Store_X_except) {
1642 DBG_OPT_EXC_REM(proj);
1643 return get_irg_bad(current_ir_graph);
1649 } /* equivalent_node_Proj_Store */
1652 * Does all optimizations on nodes that must be done on it's Proj's
1653 * because of creating new nodes.
1655 static ir_node *equivalent_node_Proj(ir_node *proj) {
1656 ir_node *n = get_Proj_pred(proj);
1658 if (get_irn_mode(proj) == mode_X) {
1659 if (is_Block_dead(get_nodes_block(n))) {
1660 /* Remove dead control flow -- early gigo(). */
1661 return get_irg_bad(current_ir_graph);
1664 if (n->op->ops.equivalent_node_Proj)
1665 return n->op->ops.equivalent_node_Proj(proj);
1667 } /* equivalent_node_Proj */
1672 static ir_node *equivalent_node_Id(ir_node *n) {
1679 DBG_OPT_ID(oldn, n);
1681 } /* equivalent_node_Id */
1686 static ir_node *equivalent_node_Mux(ir_node *n)
1688 ir_node *oldn = n, *sel = get_Mux_sel(n);
1689 tarval *ts = value_of(sel);
1691 /* Mux(true, f, t) == t */
1692 if (ts == tarval_b_true) {
1693 n = get_Mux_true(n);
1694 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1696 /* Mux(false, f, t) == f */
1697 else if (ts == tarval_b_false) {
1698 n = get_Mux_false(n);
1699 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1701 /* Mux(v, x, x) == x */
1702 else if (get_Mux_false(n) == get_Mux_true(n)) {
1703 n = get_Mux_true(n);
1704 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1706 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1707 ir_node *cmp = get_Proj_pred(sel);
1708 long proj_nr = get_Proj_proj(sel);
1709 ir_node *f = get_Mux_false(n);
1710 ir_node *t = get_Mux_true(n);
1713 * Note further that these optimization work even for floating point
1714 * with NaN's because -NaN == NaN.
1715 * However, if +0 and -0 is handled differently, we cannot use the first one.
1718 ir_node *const cmp_l = get_Cmp_left(cmp);
1719 ir_node *const cmp_r = get_Cmp_right(cmp);
1723 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1724 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1726 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1733 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1734 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1736 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1743 * Note: normalization puts the constant on the right side,
1744 * so we check only one case.
1746 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1747 /* Mux(t CMP 0, X, t) */
1748 if (is_Minus(f) && get_Minus_op(f) == t) {
1749 /* Mux(t CMP 0, -t, t) */
1750 if (proj_nr == pn_Cmp_Eq) {
1751 /* Mux(t == 0, -t, t) ==> -t */
1753 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1754 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1755 /* Mux(t != 0, -t, t) ==> t */
1757 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1764 } /* equivalent_node_Mux */
1767 * Remove Confirm nodes if setting is on.
1768 * Replace Confirms(x, '=', Constlike) by Constlike.
1770 static ir_node *equivalent_node_Confirm(ir_node *n) {
1771 ir_node *pred = get_Confirm_value(n);
1772 pn_Cmp pnc = get_Confirm_cmp(n);
1774 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1776 * rare case: two identical Confirms one after another,
1777 * replace the second one with the first.
1780 pred = get_Confirm_value(n);
1781 pnc = get_Confirm_cmp(n);
1783 if (get_opt_remove_confirm())
1784 return get_Confirm_value(n);
1789 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1790 * perform no actual computation, as, e.g., the Id nodes. It does not create
1791 * new nodes. It is therefore safe to free n if the node returned is not n.
1792 * If a node returns a Tuple we can not just skip it. If the size of the
1793 * in array fits, we transform n into a tuple (e.g., Div).
1795 ir_node *equivalent_node(ir_node *n) {
1796 if (n->op->ops.equivalent_node)
1797 return n->op->ops.equivalent_node(n);
1799 } /* equivalent_node */
1802 * Sets the default equivalent node operation for an ir_op_ops.
1804 * @param code the opcode for the default operation
1805 * @param ops the operations initialized
1810 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1814 ops->equivalent_node = equivalent_node_##a; \
1816 #define CASE_PROJ(a) \
1818 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1860 } /* firm_set_default_equivalent_node */
1863 * Returns non-zero if a node is a Phi node
1864 * with all predecessors constant.
1866 static int is_const_Phi(ir_node *n) {
1869 if (! is_Phi(n) || get_irn_arity(n) == 0)
1871 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1872 if (! is_Const(get_irn_n(n, i)))
1875 } /* is_const_Phi */
1877 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1878 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1881 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1883 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1884 if (eval == tarval_sub) {
1885 tarval_sub_type func = (tarval_sub_type)eval;
1887 return func(a, b, mode);
1889 tarval_binop_type func = (tarval_binop_type)eval;
1896 * Apply an evaluator on a binop with a constant operators (and one Phi).
1898 * @param phi the Phi node
1899 * @param other the other operand
1900 * @param eval an evaluator function
1901 * @param mode the mode of the result, may be different from the mode of the Phi!
1902 * @param left if non-zero, other is the left operand, else the right
1904 * @return a new Phi node if the conversion was successful, NULL else
1906 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1911 int i, n = get_irn_arity(phi);
1913 NEW_ARR_A(void *, res, n);
1915 for (i = 0; i < n; ++i) {
1916 pred = get_irn_n(phi, i);
1917 tv = get_Const_tarval(pred);
1918 tv = do_eval(eval, other, tv, mode);
1920 if (tv == tarval_bad) {
1921 /* folding failed, bad */
1927 for (i = 0; i < n; ++i) {
1928 pred = get_irn_n(phi, i);
1929 tv = get_Const_tarval(pred);
1930 tv = do_eval(eval, tv, other, mode);
1932 if (tv == tarval_bad) {
1933 /* folding failed, bad */
1939 irg = current_ir_graph;
1940 for (i = 0; i < n; ++i) {
1941 pred = get_irn_n(phi, i);
1942 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1943 mode, res[i], get_Const_type(pred));
1945 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1946 } /* apply_binop_on_phi */
1949 * Apply an evaluator on a binop with two constant Phi.
1951 * @param a the left Phi node
1952 * @param b the right Phi node
1953 * @param eval an evaluator function
1954 * @param mode the mode of the result, may be different from the mode of the Phi!
1956 * @return a new Phi node if the conversion was successful, NULL else
1958 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1959 tarval *tv_l, *tv_r, *tv;
1965 if (get_nodes_block(a) != get_nodes_block(b))
1968 n = get_irn_arity(a);
1969 NEW_ARR_A(void *, res, n);
1971 for (i = 0; i < n; ++i) {
1972 pred = get_irn_n(a, i);
1973 tv_l = get_Const_tarval(pred);
1974 pred = get_irn_n(b, i);
1975 tv_r = get_Const_tarval(pred);
1976 tv = do_eval(eval, tv_l, tv_r, mode);
1978 if (tv == tarval_bad) {
1979 /* folding failed, bad */
1984 irg = current_ir_graph;
1985 for (i = 0; i < n; ++i) {
1986 pred = get_irn_n(a, i);
1987 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1989 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1990 } /* apply_binop_on_2_phis */
1993 * Apply an evaluator on a unop with a constant operator (a Phi).
1995 * @param phi the Phi node
1996 * @param eval an evaluator function
1998 * @return a new Phi node if the conversion was successful, NULL else
2000 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2006 int i, n = get_irn_arity(phi);
2008 NEW_ARR_A(void *, res, n);
2009 for (i = 0; i < n; ++i) {
2010 pred = get_irn_n(phi, i);
2011 tv = get_Const_tarval(pred);
2014 if (tv == tarval_bad) {
2015 /* folding failed, bad */
2020 mode = get_irn_mode(phi);
2021 irg = current_ir_graph;
2022 for (i = 0; i < n; ++i) {
2023 pred = get_irn_n(phi, i);
2024 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2025 mode, res[i], get_Const_type(pred));
2027 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2028 } /* apply_unop_on_phi */
2031 * Apply a conversion on a constant operator (a Phi).
2033 * @param phi the Phi node
2035 * @return a new Phi node if the conversion was successful, NULL else
2037 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2042 int i, n = get_irn_arity(phi);
2044 NEW_ARR_A(void *, res, n);
2045 for (i = 0; i < n; ++i) {
2046 pred = get_irn_n(phi, i);
2047 tv = get_Const_tarval(pred);
2048 tv = tarval_convert_to(tv, mode);
2050 if (tv == tarval_bad) {
2051 /* folding failed, bad */
2056 irg = current_ir_graph;
2057 for (i = 0; i < n; ++i) {
2058 pred = get_irn_n(phi, i);
2059 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2060 mode, res[i], get_Const_type(pred));
2062 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2063 } /* apply_conv_on_phi */
2066 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2067 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2068 * If possible, remove the Conv's.
2070 static ir_node *transform_node_AddSub(ir_node *n) {
2071 ir_mode *mode = get_irn_mode(n);
2073 if (mode_is_reference(mode)) {
2074 ir_node *left = get_binop_left(n);
2075 ir_node *right = get_binop_right(n);
2076 unsigned ref_bits = get_mode_size_bits(mode);
2078 if (is_Conv(left)) {
2079 ir_mode *lmode = get_irn_mode(left);
2080 unsigned bits = get_mode_size_bits(lmode);
2082 if (ref_bits == bits &&
2083 mode_is_int(lmode) &&
2084 get_mode_arithmetic(lmode) == irma_twos_complement) {
2085 ir_node *pre = get_Conv_op(left);
2086 ir_mode *pre_mode = get_irn_mode(pre);
2088 if (mode_is_int(pre_mode) &&
2089 get_mode_size_bits(pre_mode) == bits &&
2090 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2091 /* ok, this conv just changes to sign, moreover the calculation
2092 * is done with same number of bits as our address mode, so
2093 * we can ignore the conv as address calculation can be viewed
2094 * as either signed or unsigned
2096 set_binop_left(n, pre);
2101 if (is_Conv(right)) {
2102 ir_mode *rmode = get_irn_mode(right);
2103 unsigned bits = get_mode_size_bits(rmode);
2105 if (ref_bits == bits &&
2106 mode_is_int(rmode) &&
2107 get_mode_arithmetic(rmode) == irma_twos_complement) {
2108 ir_node *pre = get_Conv_op(right);
2109 ir_mode *pre_mode = get_irn_mode(pre);
2111 if (mode_is_int(pre_mode) &&
2112 get_mode_size_bits(pre_mode) == bits &&
2113 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2114 /* ok, this conv just changes to sign, moreover the calculation
2115 * is done with same number of bits as our address mode, so
2116 * we can ignore the conv as address calculation can be viewed
2117 * as either signed or unsigned
2119 set_binop_right(n, pre);
2124 /* let address arithmetic use unsigned modes */
2125 if (is_Const(right)) {
2126 ir_mode *rmode = get_irn_mode(right);
2128 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2129 /* convert a AddP(P, *s) into AddP(P, *u) */
2130 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2132 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2133 set_binop_right(n, pre);
2139 } /* transform_node_AddSub */
2141 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2143 if (is_Const(b) && is_const_Phi(a)) { \
2144 /* check for Op(Phi, Const) */ \
2145 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2147 else if (is_Const(a) && is_const_Phi(b)) { \
2148 /* check for Op(Const, Phi) */ \
2149 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2151 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2152 /* check for Op(Phi, Phi) */ \
2153 c = apply_binop_on_2_phis(a, b, eval, mode); \
2156 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2160 #define HANDLE_UNOP_PHI(eval, a, c) \
2162 if (is_const_Phi(a)) { \
2163 /* check for Op(Phi) */ \
2164 c = apply_unop_on_phi(a, eval); \
2166 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2172 * Do the AddSub optimization, then Transform
2173 * Constant folding on Phi
2174 * Add(a,a) -> Mul(a, 2)
2175 * Add(Mul(a, x), a) -> Mul(a, x+1)
2176 * if the mode is integer or float.
2177 * Transform Add(a,-b) into Sub(a,b).
2178 * Reassociation might fold this further.
2180 static ir_node *transform_node_Add(ir_node *n) {
2182 ir_node *a, *b, *c, *oldn = n;
2184 n = transform_node_AddSub(n);
2186 a = get_Add_left(n);
2187 b = get_Add_right(n);
2189 mode = get_irn_mode(n);
2191 if (mode_is_reference(mode)) {
2192 ir_mode *lmode = get_irn_mode(a);
2194 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2195 /* an Add(a, NULL) is a hidden Conv */
2196 dbg_info *dbg = get_irn_dbg_info(n);
2197 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2201 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2203 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2204 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2207 if (mode_is_num(mode)) {
2208 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2209 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2210 ir_node *block = get_nodes_block(n);
2213 get_irn_dbg_info(n),
2217 new_r_Const_long(current_ir_graph, block, mode, 2),
2219 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2224 get_irn_dbg_info(n),
2230 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2235 get_irn_dbg_info(n),
2241 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2244 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2245 /* Here we rely on constants be on the RIGHT side */
2247 ir_node *op = get_Not_op(a);
2249 if (is_Const(b) && is_Const_one(b)) {
2251 ir_node *blk = get_nodes_block(n);
2252 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2253 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2258 ir_node *blk = get_nodes_block(n);
2259 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2260 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2265 ir_node *op = get_Not_op(b);
2269 ir_node *blk = get_nodes_block(n);
2270 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2278 } /* transform_node_Add */
2281 * returns -cnst or NULL if impossible
2283 static ir_node *const_negate(ir_node *cnst) {
2284 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2285 dbg_info *dbgi = get_irn_dbg_info(cnst);
2286 ir_graph *irg = get_irn_irg(cnst);
2287 ir_node *block = get_nodes_block(cnst);
2288 ir_mode *mode = get_irn_mode(cnst);
2289 if (tv == tarval_bad) return NULL;
2290 return new_rd_Const(dbgi, irg, block, mode, tv);
2294 * Do the AddSub optimization, then Transform
2295 * Constant folding on Phi
2296 * Sub(0,a) -> Minus(a)
2297 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2298 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2299 * Sub(Add(a, x), x) -> a
2300 * Sub(x, Add(x, a)) -> -a
2301 * Sub(x, Const) -> Add(x, -Const)
2303 static ir_node *transform_node_Sub(ir_node *n) {
2308 n = transform_node_AddSub(n);
2310 a = get_Sub_left(n);
2311 b = get_Sub_right(n);
2313 mode = get_irn_mode(n);
2315 if (mode_is_int(mode)) {
2316 ir_mode *lmode = get_irn_mode(a);
2318 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2319 /* a Sub(a, NULL) is a hidden Conv */
2320 dbg_info *dbg = get_irn_dbg_info(n);
2321 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2322 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2326 if (mode == lmode &&
2327 get_mode_arithmetic(mode) == irma_twos_complement &&
2329 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2331 dbg_info *dbg = get_irn_dbg_info(n);
2332 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2333 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2339 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2341 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2342 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2345 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2346 /* a - C -> a + (-C) */
2347 ir_node *cnst = const_negate(b);
2349 ir_node *block = get_nodes_block(n);
2350 dbg_info *dbgi = get_irn_dbg_info(n);
2351 ir_graph *irg = get_irn_irg(n);
2353 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2354 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2359 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2360 ir_graph *irg = current_ir_graph;
2361 dbg_info *dbg = get_irn_dbg_info(n);
2362 ir_node *block = get_nodes_block(n);
2363 ir_node *left = get_Minus_op(a);
2364 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2366 n = new_rd_Minus(dbg, irg, block, add, mode);
2367 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2369 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2370 ir_graph *irg = current_ir_graph;
2371 dbg_info *dbg = get_irn_dbg_info(n);
2372 ir_node *block = get_nodes_block(n);
2373 ir_node *right = get_Minus_op(b);
2375 n = new_rd_Add(dbg, irg, block, a, right, mode);
2376 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2378 } else if (is_Sub(b)) {
2379 /* a - (b - c) -> a + (c - b)
2380 * -> (a - b) + c iff (b - c) is a pointer */
2381 ir_graph *irg = current_ir_graph;
2382 dbg_info *s_dbg = get_irn_dbg_info(b);
2383 ir_node *s_block = get_nodes_block(b);
2384 ir_node *s_left = get_Sub_left(b);
2385 ir_node *s_right = get_Sub_right(b);
2386 ir_mode *s_mode = get_irn_mode(b);
2387 if (s_mode == mode_P) {
2388 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2389 dbg_info *a_dbg = get_irn_dbg_info(n);
2390 ir_node *a_block = get_nodes_block(n);
2393 s_right = new_r_Conv(irg, a_block, s_right, mode);
2394 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2396 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2397 dbg_info *a_dbg = get_irn_dbg_info(n);
2398 ir_node *a_block = get_nodes_block(n);
2400 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2402 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2404 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2405 ir_node *m_right = get_Mul_right(b);
2406 if (is_Const(m_right)) {
2407 ir_node *cnst2 = const_negate(m_right);
2408 if (cnst2 != NULL) {
2409 ir_graph *irg = current_ir_graph;
2410 dbg_info *m_dbg = get_irn_dbg_info(b);
2411 ir_node *m_block = get_nodes_block(b);
2412 ir_node *m_left = get_Mul_left(b);
2413 ir_mode *m_mode = get_irn_mode(b);
2414 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2415 dbg_info *a_dbg = get_irn_dbg_info(n);
2416 ir_node *a_block = get_nodes_block(n);
2418 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2419 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2425 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2426 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2428 get_irn_dbg_info(n),
2433 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2437 if (mode_wrap_around(mode)) {
2438 ir_node *left = get_Add_left(a);
2439 ir_node *right = get_Add_right(a);
2441 /* FIXME: Does the Conv's work only for two complement or generally? */
2443 if (mode != get_irn_mode(right)) {
2444 /* This Sub is an effective Cast */
2445 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2448 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2450 } else if (right == b) {
2451 if (mode != get_irn_mode(left)) {
2452 /* This Sub is an effective Cast */
2453 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2456 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2462 if (mode_wrap_around(mode)) {
2463 ir_node *left = get_Add_left(b);
2464 ir_node *right = get_Add_right(b);
2466 /* FIXME: Does the Conv's work only for two complement or generally? */
2468 ir_mode *r_mode = get_irn_mode(right);
2470 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2471 if (mode != r_mode) {
2472 /* This Sub is an effective Cast */
2473 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2475 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2477 } else if (right == a) {
2478 ir_mode *l_mode = get_irn_mode(left);
2480 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2481 if (mode != l_mode) {
2482 /* This Sub is an effective Cast */
2483 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2485 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2490 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2491 ir_mode *mode = get_irn_mode(a);
2493 if (mode == get_irn_mode(b)) {
2495 ir_node *op_a = get_Conv_op(a);
2496 ir_node *op_b = get_Conv_op(b);
2498 /* check if it's allowed to skip the conv */
2499 ma = get_irn_mode(op_a);
2500 mb = get_irn_mode(op_b);
2502 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2503 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2506 set_Sub_right(n, b);
2512 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2513 if (!is_reassoc_running() && is_Mul(a)) {
2514 ir_node *ma = get_Mul_left(a);
2515 ir_node *mb = get_Mul_right(a);
2518 ir_node *blk = get_nodes_block(n);
2520 get_irn_dbg_info(n),
2521 current_ir_graph, blk,
2524 get_irn_dbg_info(n),
2525 current_ir_graph, blk,
2527 new_r_Const_long(current_ir_graph, blk, mode, 1),
2530 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2532 } else if (mb == b) {
2533 ir_node *blk = get_nodes_block(n);
2535 get_irn_dbg_info(n),
2536 current_ir_graph, blk,
2539 get_irn_dbg_info(n),
2540 current_ir_graph, blk,
2542 new_r_Const_long(current_ir_graph, blk, mode, 1),
2545 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2549 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2550 ir_node *x = get_Sub_left(a);
2551 ir_node *y = get_Sub_right(a);
2552 ir_node *blk = get_nodes_block(n);
2553 ir_mode *m_b = get_irn_mode(b);
2554 ir_mode *m_y = get_irn_mode(y);
2558 /* Determine the right mode for the Add. */
2561 else if (mode_is_reference(m_b))
2563 else if (mode_is_reference(m_y))
2567 * Both modes are different but none is reference,
2568 * happens for instance in SubP(SubP(P, Iu), Is).
2569 * We have two possibilities here: Cast or ignore.
2570 * Currently we ignore this case.
2575 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2577 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2578 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2582 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2583 if (is_Const(a) && is_Not(b)) {
2584 /* c - ~X = X + (c+1) */
2585 tarval *tv = get_Const_tarval(a);
2587 tv = tarval_add(tv, get_mode_one(mode));
2588 if (tv != tarval_bad) {
2589 ir_node *blk = get_nodes_block(n);
2590 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2591 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2592 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2598 } /* transform_node_Sub */
2601 * Several transformation done on n*n=2n bits mul.
2602 * These transformations must be done here because new nodes may be produced.
2604 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2606 ir_node *a = get_Mul_left(n);
2607 ir_node *b = get_Mul_right(n);
2608 tarval *ta = value_of(a);
2609 tarval *tb = value_of(b);
2610 ir_mode *smode = get_irn_mode(a);
2612 if (ta == get_mode_one(smode)) {
2613 /* (L)1 * (L)b = (L)b */
2614 ir_node *blk = get_nodes_block(n);
2615 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2616 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2619 else if (ta == get_mode_minus_one(smode)) {
2620 /* (L)-1 * (L)b = (L)b */
2621 ir_node *blk = get_nodes_block(n);
2622 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2623 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2624 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2627 if (tb == get_mode_one(smode)) {
2628 /* (L)a * (L)1 = (L)a */
2629 ir_node *blk = get_irn_n(a, -1);
2630 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2631 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2634 else if (tb == get_mode_minus_one(smode)) {
2635 /* (L)a * (L)-1 = (L)-a */
2636 ir_node *blk = get_nodes_block(n);
2637 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2638 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2639 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2646 * Transform Mul(a,-1) into -a.
2647 * Do constant evaluation of Phi nodes.
2648 * Do architecture dependent optimizations on Mul nodes
2650 static ir_node *transform_node_Mul(ir_node *n) {
2651 ir_node *c, *oldn = n;
2652 ir_mode *mode = get_irn_mode(n);
2653 ir_node *a = get_Mul_left(n);
2654 ir_node *b = get_Mul_right(n);
2656 if (is_Bad(a) || is_Bad(b))
2659 if (mode != get_irn_mode(a))
2660 return transform_node_Mul2n(n, mode);
2662 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2664 if (mode_is_signed(mode)) {
2667 if (value_of(a) == get_mode_minus_one(mode))
2669 else if (value_of(b) == get_mode_minus_one(mode))
2672 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2673 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2678 if (is_Const(b)) { /* (-a) * const -> a * -const */
2679 ir_node *cnst = const_negate(b);
2681 dbg_info *dbgi = get_irn_dbg_info(n);
2682 ir_node *block = get_nodes_block(n);
2683 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2684 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2687 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2688 dbg_info *dbgi = get_irn_dbg_info(n);
2689 ir_node *block = get_nodes_block(n);
2690 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2691 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2693 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2694 ir_node *sub_l = get_Sub_left(b);
2695 ir_node *sub_r = get_Sub_right(b);
2696 dbg_info *dbgi = get_irn_dbg_info(n);
2697 ir_graph *irg = current_ir_graph;
2698 ir_node *block = get_nodes_block(n);
2699 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2700 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2701 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2704 } else if (is_Minus(b)) {
2705 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2706 ir_node *sub_l = get_Sub_left(a);
2707 ir_node *sub_r = get_Sub_right(a);
2708 dbg_info *dbgi = get_irn_dbg_info(n);
2709 ir_graph *irg = current_ir_graph;
2710 ir_node *block = get_nodes_block(n);
2711 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2712 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2713 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2717 if (get_mode_arithmetic(mode) == irma_ieee754) {
2719 tarval *tv = get_Const_tarval(a);
2720 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2721 && !tarval_is_negative(tv)) {
2722 /* 2.0 * b = b + b */
2723 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2724 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2728 else if (is_Const(b)) {
2729 tarval *tv = get_Const_tarval(b);
2730 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2731 && !tarval_is_negative(tv)) {
2732 /* a * 2.0 = a + a */
2733 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2734 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2739 return arch_dep_replace_mul_with_shifts(n);
2740 } /* transform_node_Mul */
2743 * Transform a Div Node.
2745 static ir_node *transform_node_Div(ir_node *n) {
2746 ir_mode *mode = get_Div_resmode(n);
2747 ir_node *a = get_Div_left(n);
2748 ir_node *b = get_Div_right(n);
2752 if (is_Const(b) && is_const_Phi(a)) {
2753 /* check for Div(Phi, Const) */
2754 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2756 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2760 else if (is_Const(a) && is_const_Phi(b)) {
2761 /* check for Div(Const, Phi) */
2762 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2764 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2768 else if (is_const_Phi(a) && is_const_Phi(b)) {
2769 /* check for Div(Phi, Phi) */
2770 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2772 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2779 if (tv != tarval_bad) {
2780 value = new_Const(get_tarval_mode(tv), tv);
2782 DBG_OPT_CSTEVAL(n, value);
2785 ir_node *a = get_Div_left(n);
2786 ir_node *b = get_Div_right(n);
2787 const ir_node *dummy;
2789 if (a == b && value_not_zero(a, &dummy)) {
2790 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2791 value = new_Const(mode, get_mode_one(mode));
2792 DBG_OPT_CSTEVAL(n, value);
2795 if (mode_is_signed(mode) && is_Const(b)) {
2796 tarval *tv = get_Const_tarval(b);
2798 if (tv == get_mode_minus_one(mode)) {
2800 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2801 DBG_OPT_CSTEVAL(n, value);
2805 /* Try architecture dependent optimization */
2806 value = arch_dep_replace_div_by_const(n);
2814 /* Turn Div into a tuple (mem, jmp, bad, value) */
2815 mem = get_Div_mem(n);
2816 blk = get_nodes_block(n);
2818 /* skip a potential Pin */
2819 mem = skip_Pin(mem);
2820 turn_into_tuple(n, pn_Div_max);
2821 set_Tuple_pred(n, pn_Div_M, mem);
2822 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2823 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2824 set_Tuple_pred(n, pn_Div_res, value);
2827 } /* transform_node_Div */
2830 * Transform a Mod node.
2832 static ir_node *transform_node_Mod(ir_node *n) {
2833 ir_mode *mode = get_Mod_resmode(n);
2834 ir_node *a = get_Mod_left(n);
2835 ir_node *b = get_Mod_right(n);
2839 if (is_Const(b) && is_const_Phi(a)) {
2840 /* check for Div(Phi, Const) */
2841 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2843 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2847 else if (is_Const(a) && is_const_Phi(b)) {
2848 /* check for Div(Const, Phi) */
2849 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2851 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2855 else if (is_const_Phi(a) && is_const_Phi(b)) {
2856 /* check for Div(Phi, Phi) */
2857 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2859 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2866 if (tv != tarval_bad) {
2867 value = new_Const(get_tarval_mode(tv), tv);
2869 DBG_OPT_CSTEVAL(n, value);
2872 ir_node *a = get_Mod_left(n);
2873 ir_node *b = get_Mod_right(n);
2874 const ir_node *dummy;
2876 if (a == b && value_not_zero(a, &dummy)) {
2877 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2878 value = new_Const(mode, get_mode_null(mode));
2879 DBG_OPT_CSTEVAL(n, value);
2882 if (mode_is_signed(mode) && is_Const(b)) {
2883 tarval *tv = get_Const_tarval(b);
2885 if (tv == get_mode_minus_one(mode)) {
2887 value = new_Const(mode, get_mode_null(mode));
2888 DBG_OPT_CSTEVAL(n, value);
2892 /* Try architecture dependent optimization */
2893 value = arch_dep_replace_mod_by_const(n);
2901 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2902 mem = get_Mod_mem(n);
2903 blk = get_nodes_block(n);
2905 /* skip a potential Pin */
2906 mem = skip_Pin(mem);
2907 turn_into_tuple(n, pn_Mod_max);
2908 set_Tuple_pred(n, pn_Mod_M, mem);
2909 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2910 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2911 set_Tuple_pred(n, pn_Mod_res, value);
2914 } /* transform_node_Mod */
2917 * Transform a DivMod node.
2919 static ir_node *transform_node_DivMod(ir_node *n) {
2920 const ir_node *dummy;
2921 ir_node *a = get_DivMod_left(n);
2922 ir_node *b = get_DivMod_right(n);
2923 ir_mode *mode = get_DivMod_resmode(n);
2928 if (is_Const(b) && is_const_Phi(a)) {
2929 /* check for Div(Phi, Const) */
2930 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2931 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2933 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2934 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2938 else if (is_Const(a) && is_const_Phi(b)) {
2939 /* check for Div(Const, Phi) */
2940 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2941 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2943 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2944 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2948 else if (is_const_Phi(a) && is_const_Phi(b)) {
2949 /* check for Div(Phi, Phi) */
2950 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2951 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2953 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2954 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2961 if (tb != tarval_bad) {
2962 if (tb == get_mode_one(get_tarval_mode(tb))) {
2964 vb = new_Const(mode, get_mode_null(mode));
2965 DBG_OPT_CSTEVAL(n, vb);
2967 } else if (ta != tarval_bad) {
2968 tarval *resa, *resb;
2969 resa = tarval_div(ta, tb);
2970 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2971 Jmp for X result!? */
2972 resb = tarval_mod(ta, tb);
2973 if (resb == tarval_bad) return n; /* Causes exception! */
2974 va = new_Const(mode, resa);
2975 vb = new_Const(mode, resb);
2976 DBG_OPT_CSTEVAL(n, va);
2977 DBG_OPT_CSTEVAL(n, vb);
2979 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2980 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2981 vb = new_Const(mode, get_mode_null(mode));
2982 DBG_OPT_CSTEVAL(n, va);
2983 DBG_OPT_CSTEVAL(n, vb);
2985 } else { /* Try architecture dependent optimization */
2988 arch_dep_replace_divmod_by_const(&va, &vb, n);
2989 evaluated = va != NULL;
2991 } else if (a == b) {
2992 if (value_not_zero(a, &dummy)) {
2994 va = new_Const(mode, get_mode_one(mode));
2995 vb = new_Const(mode, get_mode_null(mode));
2996 DBG_OPT_CSTEVAL(n, va);
2997 DBG_OPT_CSTEVAL(n, vb);
3000 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3003 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3004 /* 0 / non-Const = 0 */
3009 if (evaluated) { /* replace by tuple */
3013 mem = get_DivMod_mem(n);
3014 /* skip a potential Pin */
3015 mem = skip_Pin(mem);
3017 blk = get_nodes_block(n);
3018 turn_into_tuple(n, pn_DivMod_max);
3019 set_Tuple_pred(n, pn_DivMod_M, mem);
3020 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3021 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3022 set_Tuple_pred(n, pn_DivMod_res_div, va);
3023 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3027 } /* transform_node_DivMod */
3030 * Optimize x / c to x * (1/c)
3032 static ir_node *transform_node_Quot(ir_node *n) {
3033 ir_mode *mode = get_Quot_resmode(n);
3036 if (get_mode_arithmetic(mode) == irma_ieee754) {
3037 ir_node *b = get_Quot_right(n);
3038 tarval *tv = value_of(b);
3040 if (tv != tarval_bad) {
3044 * Floating point constant folding might be disabled here to
3046 * However, as we check for exact result, doing it is safe.
3049 rem = tarval_enable_fp_ops(1);
3050 tv = tarval_quo(get_mode_one(mode), tv);
3051 (void)tarval_enable_fp_ops(rem);
3053 /* Do the transformation if the result is either exact or we are not
3054 using strict rules. */
3055 if (tv != tarval_bad &&
3056 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3057 ir_node *blk = get_nodes_block(n);
3058 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3059 ir_node *a = get_Quot_left(n);
3060 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3061 ir_node *mem = get_Quot_mem(n);
3063 /* skip a potential Pin */
3064 mem = skip_Pin(mem);
3065 turn_into_tuple(n, pn_Quot_max);
3066 set_Tuple_pred(n, pn_Quot_M, mem);
3067 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3068 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3069 set_Tuple_pred(n, pn_Quot_res, m);
3070 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3075 } /* transform_node_Quot */
3078 * Optimize Abs(x) into x if x is Confirmed >= 0
3079 * Optimize Abs(x) into -x if x is Confirmed <= 0
3080 * Optimize Abs(-x) int Abs(x)
3082 static ir_node *transform_node_Abs(ir_node *n) {
3083 ir_node *c, *oldn = n;
3084 ir_node *a = get_Abs_op(n);
3087 HANDLE_UNOP_PHI(tarval_abs, a, c);
3089 switch (classify_value_sign(a)) {
3090 case value_classified_negative:
3091 mode = get_irn_mode(n);
3094 * We can replace the Abs by -x here.
3095 * We even could add a new Confirm here
3096 * (if not twos complement)
3098 * Note that -x would create a new node, so we could
3099 * not run it in the equivalent_node() context.
3101 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3102 get_nodes_block(n), a, mode);
3104 DBG_OPT_CONFIRM(oldn, n);
3106 case value_classified_positive:
3107 /* n is positive, Abs is not needed */
3110 DBG_OPT_CONFIRM(oldn, n);
3116 /* Abs(-x) = Abs(x) */
3117 mode = get_irn_mode(n);
3118 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3119 get_nodes_block(n), get_Minus_op(a), mode);
3120 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3124 } /* transform_node_Abs */
3127 * Optimize -a CMP -b into b CMP a.
3128 * This works only for for modes where unary Minus
3130 * Note that two-complement integers can Overflow
3131 * so it will NOT work.
3133 * For == and != can be handled in Proj(Cmp)
3135 static ir_node *transform_node_Cmp(ir_node *n) {
3137 ir_node *left = get_Cmp_left(n);
3138 ir_node *right = get_Cmp_right(n);
3140 if (is_Minus(left) && is_Minus(right) &&
3141 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3142 ir_node *const new_left = get_Minus_op(right);
3143 ir_node *const new_right = get_Minus_op(left);
3144 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3145 get_nodes_block(n), new_left, new_right);
3146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3149 } /* transform_node_Cmp */
3153 * Transform a Cond node.
3155 * Replace the Cond by a Jmp if it branches on a constant
3158 static ir_node *transform_node_Cond(ir_node *n) {
3161 ir_node *a = get_Cond_selector(n);
3162 tarval *ta = value_of(a);
3164 /* we need block info which is not available in floating irgs */
3165 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3168 if ((ta != tarval_bad) &&
3169 (get_irn_mode(a) == mode_b) &&
3170 (get_opt_unreachable_code())) {
3171 /* It's a boolean Cond, branching on a boolean constant.
3172 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3173 ir_node *blk = get_nodes_block(n);
3174 jmp = new_r_Jmp(current_ir_graph, blk);
3175 turn_into_tuple(n, pn_Cond_max);
3176 if (ta == tarval_b_true) {
3177 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3178 set_Tuple_pred(n, pn_Cond_true, jmp);
3180 set_Tuple_pred(n, pn_Cond_false, jmp);
3181 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3183 /* We might generate an endless loop, so keep it alive. */
3184 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3187 } /* transform_node_Cond */
3190 * Prototype of a recursive transform function
3191 * for bitwise distributive transformations.
3193 typedef ir_node* (*recursive_transform)(ir_node *n);
3196 * makes use of distributive laws for and, or, eor
3197 * and(a OP c, b OP c) -> and(a, b) OP c
3198 * note, might return a different op than n
3200 static ir_node *transform_bitwise_distributive(ir_node *n,
3201 recursive_transform trans_func)
3204 ir_node *a = get_binop_left(n);
3205 ir_node *b = get_binop_right(n);
3206 ir_op *op = get_irn_op(a);
3207 ir_op *op_root = get_irn_op(n);
3209 if(op != get_irn_op(b))
3212 if (op == op_Conv) {
3213 ir_node *a_op = get_Conv_op(a);
3214 ir_node *b_op = get_Conv_op(b);
3215 ir_mode *a_mode = get_irn_mode(a_op);
3216 ir_mode *b_mode = get_irn_mode(b_op);
3217 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3218 ir_node *blk = get_nodes_block(n);
3221 set_binop_left(n, a_op);
3222 set_binop_right(n, b_op);
3223 set_irn_mode(n, a_mode);
3225 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3227 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3233 /* nothing to gain here */
3237 if (op == op_Shrs || op == op_Shr || op == op_Shl
3238 || op == op_And || op == op_Or || op == op_Eor) {
3239 ir_node *a_left = get_binop_left(a);
3240 ir_node *a_right = get_binop_right(a);
3241 ir_node *b_left = get_binop_left(b);
3242 ir_node *b_right = get_binop_right(b);
3244 ir_node *op1 = NULL;
3245 ir_node *op2 = NULL;
3247 if (is_op_commutative(op)) {
3248 if (a_left == b_left) {
3252 } else if(a_left == b_right) {
3256 } else if(a_right == b_left) {
3262 if(a_right == b_right) {
3269 /* (a sop c) & (b sop c) => (a & b) sop c */
3270 ir_node *blk = get_nodes_block(n);
3272 ir_node *new_n = exact_copy(n);
3273 set_binop_left(new_n, op1);
3274 set_binop_right(new_n, op2);
3275 new_n = trans_func(new_n);
3277 if(op_root == op_Eor && op == op_Or) {
3278 dbg_info *dbgi = get_irn_dbg_info(n);
3279 ir_graph *irg = current_ir_graph;
3280 ir_mode *mode = get_irn_mode(c);
3282 c = new_rd_Not(dbgi, irg, blk, c, mode);
3283 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3286 set_nodes_block(n, blk);
3287 set_binop_left(n, new_n);
3288 set_binop_right(n, c);
3289 add_identities(current_ir_graph->value_table, n);
3292 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3303 static ir_node *transform_node_And(ir_node *n) {
3304 ir_node *c, *oldn = n;
3305 ir_node *a = get_And_left(n);
3306 ir_node *b = get_And_right(n);
3309 mode = get_irn_mode(n);
3310 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3312 /* we can evaluate 2 Projs of the same Cmp */
3313 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3314 ir_node *pred_a = get_Proj_pred(a);
3315 ir_node *pred_b = get_Proj_pred(b);
3316 if (pred_a == pred_b) {
3317 dbg_info *dbgi = get_irn_dbg_info(n);
3318 ir_node *block = get_nodes_block(pred_a);
3319 pn_Cmp pn_a = get_Proj_proj(a);
3320 pn_Cmp pn_b = get_Proj_proj(b);
3321 /* yes, we can simply calculate with pncs */
3322 pn_Cmp new_pnc = pn_a & pn_b;
3324 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3329 ir_node *op = get_Not_op(b);
3331 ir_node *ba = get_And_left(op);
3332 ir_node *bb = get_And_right(op);
3334 /* it's enough to test the following cases due to normalization! */
3335 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3336 /* (a|b) & ~(a&b) = a^b */
3337 ir_node *block = get_nodes_block(n);
3339 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3340 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3348 ir_node *op = get_Not_op(a);
3350 ir_node *aa = get_And_left(op);
3351 ir_node *ab = get_And_right(op);
3353 /* it's enough to test the following cases due to normalization! */
3354 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3355 /* (a|b) & ~(a&b) = a^b */
3356 ir_node *block = get_nodes_block(n);
3358 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3359 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3366 ir_node *al = get_Eor_left(a);
3367 ir_node *ar = get_Eor_right(a);
3370 /* (b ^ a) & b -> ~a & b */
3371 dbg_info *dbg = get_irn_dbg_info(n);
3372 ir_node *block = get_nodes_block(n);
3374 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3375 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3376 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3380 /* (a ^ b) & b -> ~a & b */
3381 dbg_info *dbg = get_irn_dbg_info(n);
3382 ir_node *block = get_nodes_block(n);
3384 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3385 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3386 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3391 ir_node *bl = get_Eor_left(b);
3392 ir_node *br = get_Eor_right(b);
3395 /* a & (a ^ b) -> a & ~b */
3396 dbg_info *dbg = get_irn_dbg_info(n);
3397 ir_node *block = get_nodes_block(n);
3399 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3400 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3401 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3405 /* a & (b ^ a) -> a & ~b */
3406 dbg_info *dbg = get_irn_dbg_info(n);
3407 ir_node *block = get_nodes_block(n);
3409 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3410 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3411 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3415 if (is_Not(a) && is_Not(b)) {
3416 /* ~a & ~b = ~(a|b) */
3417 ir_node *block = get_nodes_block(n);
3418 ir_mode *mode = get_irn_mode(n);
3422 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3423 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3424 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3428 n = transform_bitwise_distributive(n, transform_node_And);
3431 } /* transform_node_And */
3436 static ir_node *transform_node_Eor(ir_node *n) {
3437 ir_node *c, *oldn = n;
3438 ir_node *a = get_Eor_left(n);
3439 ir_node *b = get_Eor_right(n);
3440 ir_mode *mode = get_irn_mode(n);
3442 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3444 /* we can evaluate 2 Projs of the same Cmp */
3445 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3446 ir_node *pred_a = get_Proj_pred(a);
3447 ir_node *pred_b = get_Proj_pred(b);
3448 if(pred_a == pred_b) {
3449 dbg_info *dbgi = get_irn_dbg_info(n);
3450 ir_node *block = get_nodes_block(pred_a);
3451 pn_Cmp pn_a = get_Proj_proj(a);
3452 pn_Cmp pn_b = get_Proj_proj(b);
3453 /* yes, we can simply calculate with pncs */
3454 pn_Cmp new_pnc = pn_a ^ pn_b;
3456 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3463 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
3464 mode, get_mode_null(mode));
3465 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3466 } else if (mode == mode_b &&
3468 is_Const(b) && is_Const_one(b) &&
3469 is_Cmp(get_Proj_pred(a))) {
3470 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3471 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3472 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3474 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3475 } else if (is_Const(b)) {
3476 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3477 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3478 ir_node *not_op = get_Not_op(a);
3479 dbg_info *dbg = get_irn_dbg_info(n);
3480 ir_graph *irg = current_ir_graph;
3481 ir_node *block = get_nodes_block(n);
3482 ir_mode *mode = get_irn_mode(n);
3483 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3485 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3486 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3487 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3490 n = transform_bitwise_distributive(n, transform_node_Eor);
3494 } /* transform_node_Eor */
3499 static ir_node *transform_node_Not(ir_node *n) {
3500 ir_node *c, *oldn = n;
3501 ir_node *a = get_Not_op(n);
3502 ir_mode *mode = get_irn_mode(n);
3504 HANDLE_UNOP_PHI(tarval_not,a,c);
3506 /* check for a boolean Not */
3507 if (mode == mode_b &&
3509 is_Cmp(get_Proj_pred(a))) {
3510 /* We negate a Cmp. The Cmp has the negated result anyways! */
3511 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3512 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3513 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3517 ir_node *eor_b = get_Eor_right(a);
3518 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3519 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3520 ir_node *eor_a = get_Eor_left(a);
3521 dbg_info *dbg = get_irn_dbg_info(n);
3522 ir_graph *irg = current_ir_graph;
3523 ir_node *block = get_nodes_block(n);
3524 ir_mode *mode = get_irn_mode(n);
3525 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3529 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3530 if (is_Minus(a)) { /* ~-x -> x + -1 */
3531 dbg_info *dbg = get_irn_dbg_info(n);
3532 ir_graph *irg = current_ir_graph;
3533 ir_node *block = get_nodes_block(n);
3534 ir_node *add_l = get_Minus_op(a);
3535 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3536 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3537 } else if (is_Add(a)) {
3538 ir_node *add_r = get_Add_right(a);
3539 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3540 /* ~(x + -1) = -x */
3541 ir_node *op = get_Add_left(a);
3542 ir_node *blk = get_nodes_block(n);
3543 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3544 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3549 } /* transform_node_Not */
3552 * Transform a Minus.
3556 * -(a >>u (size-1)) = a >>s (size-1)
3557 * -(a >>s (size-1)) = a >>u (size-1)
3558 * -(a * const) -> a * -const
3560 static ir_node *transform_node_Minus(ir_node *n) {
3561 ir_node *c, *oldn = n;
3562 ir_node *a = get_Minus_op(n);
3565 HANDLE_UNOP_PHI(tarval_neg,a,c);
3567 mode = get_irn_mode(a);
3568 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3569 /* the following rules are only to twos-complement */
3572 ir_node *op = get_Not_op(a);
3573 tarval *tv = get_mode_one(mode);
3574 ir_node *blk = get_nodes_block(n);
3575 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3576 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3577 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3581 ir_node *c = get_Shr_right(a);
3584 tarval *tv = get_Const_tarval(c);
3586 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3587 /* -(a >>u (size-1)) = a >>s (size-1) */
3588 ir_node *v = get_Shr_left(a);
3590 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3591 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3597 ir_node *c = get_Shrs_right(a);
3600 tarval *tv = get_Const_tarval(c);
3602 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3603 /* -(a >>s (size-1)) = a >>u (size-1) */
3604 ir_node *v = get_Shrs_left(a);
3606 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3607 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3614 /* - (a-b) = b - a */
3615 ir_node *la = get_Sub_left(a);
3616 ir_node *ra = get_Sub_right(a);
3617 ir_node *blk = get_nodes_block(n);
3619 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3620 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3624 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3625 ir_node *mul_l = get_Mul_left(a);
3626 ir_node *mul_r = get_Mul_right(a);
3627 tarval *tv = value_of(mul_r);
3628 if (tv != tarval_bad) {
3629 tv = tarval_neg(tv);
3630 if (tv != tarval_bad) {
3631 ir_node *cnst = new_Const(mode, tv);
3632 dbg_info *dbg = get_irn_dbg_info(a);
3633 ir_graph *irg = current_ir_graph;
3634 ir_node *block = get_nodes_block(a);
3635 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3636 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3643 } /* transform_node_Minus */
3646 * Transform a Cast_type(Const) into a new Const_type
3648 static ir_node *transform_node_Cast(ir_node *n) {
3650 ir_node *pred = get_Cast_op(n);
3651 ir_type *tp = get_irn_type(n);
3653 if (is_Const(pred) && get_Const_type(pred) != tp) {
3654 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3655 get_Const_tarval(pred), tp);
3656 DBG_OPT_CSTEVAL(oldn, n);
3657 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3658 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3659 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3660 DBG_OPT_CSTEVAL(oldn, n);
3664 } /* transform_node_Cast */
3667 * Transform a Proj(Load) with a non-null address.
3669 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3670 if (get_opt_ldst_only_null_ptr_exceptions()) {
3671 if (get_irn_mode(proj) == mode_X) {
3672 ir_node *load = get_Proj_pred(proj);
3674 /* get the Load address */
3675 const ir_node *addr = get_Load_ptr(load);
3676 const ir_node *confirm;
3678 if (value_not_null(addr, &confirm)) {
3679 if (confirm == NULL) {
3680 /* this node may float if it did not depend on a Confirm */
3681 set_irn_pinned(load, op_pin_state_floats);
3683 if (get_Proj_proj(proj) == pn_Load_X_except) {
3684 DBG_OPT_EXC_REM(proj);
3685 return get_irg_bad(current_ir_graph);
3687 ir_node *blk = get_nodes_block(load);
3688 return new_r_Jmp(current_ir_graph, blk);
3694 } /* transform_node_Proj_Load */
3697 * Transform a Proj(Store) with a non-null address.
3699 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3700 if (get_opt_ldst_only_null_ptr_exceptions()) {
3701 if (get_irn_mode(proj) == mode_X) {
3702 ir_node *store = get_Proj_pred(proj);
3704 /* get the load/store address */
3705 const ir_node *addr = get_Store_ptr(store);
3706 const ir_node *confirm;
3708 if (value_not_null(addr, &confirm)) {
3709 if (confirm == NULL) {
3710 /* this node may float if it did not depend on a Confirm */
3711 set_irn_pinned(store, op_pin_state_floats);
3713 if (get_Proj_proj(proj) == pn_Store_X_except) {
3714 DBG_OPT_EXC_REM(proj);
3715 return get_irg_bad(current_ir_graph);
3717 ir_node *blk = get_nodes_block(store);
3718 return new_r_Jmp(current_ir_graph, blk);
3724 } /* transform_node_Proj_Store */
3727 * Transform a Proj(Div) with a non-zero value.
3728 * Removes the exceptions and routes the memory to the NoMem node.
3730 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3731 ir_node *div = get_Proj_pred(proj);
3732 ir_node *b = get_Div_right(div);
3733 ir_node *res, *new_mem;
3734 const ir_node *confirm;
3737 if (value_not_zero(b, &confirm)) {
3738 /* div(x, y) && y != 0 */
3739 if (confirm == NULL) {
3740 /* we are sure we have a Const != 0 */
3741 new_mem = get_Div_mem(div);
3742 new_mem = skip_Pin(new_mem);
3743 set_Div_mem(div, new_mem);
3744 set_irn_pinned(div, op_pin_state_floats);
3747 proj_nr = get_Proj_proj(proj);
3749 case pn_Div_X_regular:
3750 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3752 case pn_Div_X_except:
3753 /* we found an exception handler, remove it */
3754 DBG_OPT_EXC_REM(proj);
3758 res = get_Div_mem(div);
3759 new_mem = get_irg_no_mem(current_ir_graph);
3762 /* This node can only float up to the Confirm block */
3763 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3765 set_irn_pinned(div, op_pin_state_floats);
3766 /* this is a Div without exception, we can remove the memory edge */
3767 set_Div_mem(div, new_mem);
3772 } /* transform_node_Proj_Div */
3775 * Transform a Proj(Mod) with a non-zero value.
3776 * Removes the exceptions and routes the memory to the NoMem node.
3778 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3779 ir_node *mod = get_Proj_pred(proj);
3780 ir_node *b = get_Mod_right(mod);
3781 ir_node *res, *new_mem;
3782 const ir_node *confirm;
3785 if (value_not_zero(b, &confirm)) {
3786 /* mod(x, y) && y != 0 */
3787 proj_nr = get_Proj_proj(proj);
3789 if (confirm == NULL) {
3790 /* we are sure we have a Const != 0 */
3791 new_mem = get_Mod_mem(mod);
3792 new_mem = skip_Pin(new_mem);
3793 set_Mod_mem(mod, new_mem);
3794 set_irn_pinned(mod, op_pin_state_floats);
3799 case pn_Mod_X_regular:
3800 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3802 case pn_Mod_X_except:
3803 /* we found an exception handler, remove it */
3804 DBG_OPT_EXC_REM(proj);
3808 res = get_Mod_mem(mod);
3809 new_mem = get_irg_no_mem(current_ir_graph);
3812 /* This node can only float up to the Confirm block */
3813 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3815 /* this is a Mod without exception, we can remove the memory edge */
3816 set_Mod_mem(mod, new_mem);
3819 if (get_Mod_left(mod) == b) {
3820 /* a % a = 0 if a != 0 */
3821 ir_mode *mode = get_irn_mode(proj);
3822 ir_node *res = new_Const(mode, get_mode_null(mode));
3824 DBG_OPT_CSTEVAL(mod, res);
3830 } /* transform_node_Proj_Mod */
3833 * Transform a Proj(DivMod) with a non-zero value.
3834 * Removes the exceptions and routes the memory to the NoMem node.
3836 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3837 ir_node *divmod = get_Proj_pred(proj);
3838 ir_node *b = get_DivMod_right(divmod);
3839 ir_node *res, *new_mem;
3840 const ir_node *confirm;
3843 if (value_not_zero(b, &confirm)) {
3844 /* DivMod(x, y) && y != 0 */
3845 proj_nr = get_Proj_proj(proj);
3847 if (confirm == NULL) {
3848 /* we are sure we have a Const != 0 */
3849 new_mem = get_DivMod_mem(divmod);
3850 new_mem = skip_Pin(new_mem);
3851 set_DivMod_mem(divmod, new_mem);
3852 set_irn_pinned(divmod, op_pin_state_floats);
3857 case pn_DivMod_X_regular:
3858 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3860 case pn_DivMod_X_except:
3861 /* we found an exception handler, remove it */
3862 DBG_OPT_EXC_REM(proj);
3866 res = get_DivMod_mem(divmod);
3867 new_mem = get_irg_no_mem(current_ir_graph);
3870 /* This node can only float up to the Confirm block */
3871 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3873 /* this is a DivMod without exception, we can remove the memory edge */
3874 set_DivMod_mem(divmod, new_mem);
3877 case pn_DivMod_res_mod:
3878 if (get_DivMod_left(divmod) == b) {
3879 /* a % a = 0 if a != 0 */
3880 ir_mode *mode = get_irn_mode(proj);
3881 ir_node *res = new_Const(mode, get_mode_null(mode));
3883 DBG_OPT_CSTEVAL(divmod, res);
3889 } /* transform_node_Proj_DivMod */
3892 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3894 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3895 if (get_opt_unreachable_code()) {
3896 ir_node *n = get_Proj_pred(proj);
3897 ir_node *b = get_Cond_selector(n);
3899 if (mode_is_int(get_irn_mode(b))) {
3900 tarval *tb = value_of(b);
3902 if (tb != tarval_bad) {
3903 /* we have a constant switch */
3904 long num = get_Proj_proj(proj);
3906 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3907 if (get_tarval_long(tb) == num) {
3908 /* Do NOT create a jump here, or we will have 2 control flow ops
3909 * in a block. This case is optimized away in optimize_cf(). */
3912 /* this case will NEVER be taken, kill it */
3913 return get_irg_bad(current_ir_graph);
3920 } /* transform_node_Proj_Cond */
3923 * Create a 0 constant of given mode.
3925 static ir_node *create_zero_const(ir_mode *mode) {
3926 tarval *tv = get_mode_null(mode);
3927 ir_node *cnst = new_Const(mode, tv);
3932 /* the order of the values is important! */
3933 typedef enum const_class {
3939 static const_class classify_const(const ir_node* n)
3941 if (is_Const(n)) return const_const;
3942 if (is_irn_constlike(n)) return const_like;
3947 * Determines whether r is more constlike or has a larger index (in that order)
3950 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3952 const const_class l_order = classify_const(l);
3953 const const_class r_order = classify_const(r);
3955 l_order > r_order ||
3956 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3960 * Normalizes and optimizes Cmp nodes.
3962 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3963 ir_node *n = get_Proj_pred(proj);
3964 ir_node *left = get_Cmp_left(n);
3965 ir_node *right = get_Cmp_right(n);
3968 ir_mode *mode = NULL;
3969 long proj_nr = get_Proj_proj(proj);
3971 /* we can evaluate some cases directly */
3974 return new_Const(mode_b, get_tarval_b_false());
3976 return new_Const(mode_b, get_tarval_b_true());
3978 if (!mode_is_float(get_irn_mode(left)))
3979 return new_Const(mode_b, get_tarval_b_true());
3985 /* remove Casts of both sides */
3986 left = skip_Cast(left);
3987 right = skip_Cast(right);
3989 /* Remove unnecessary conversions */
3990 /* TODO handle constants */
3991 if (is_Conv(left) && is_Conv(right)) {
3992 ir_mode *mode = get_irn_mode(left);
3993 ir_node *op_left = get_Conv_op(left);
3994 ir_node *op_right = get_Conv_op(right);
3995 ir_mode *mode_left = get_irn_mode(op_left);
3996 ir_mode *mode_right = get_irn_mode(op_right);
3998 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3999 && mode_left != mode_b && mode_right != mode_b) {
4000 ir_graph *irg = current_ir_graph;
4001 ir_node *block = get_nodes_block(n);
4003 if (mode_left == mode_right) {
4007 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4008 } else if (smaller_mode(mode_left, mode_right)) {
4009 left = new_r_Conv(irg, block, op_left, mode_right);
4012 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4013 } else if (smaller_mode(mode_right, mode_left)) {
4015 right = new_r_Conv(irg, block, op_right, mode_left);
4017 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4022 /* remove operation on both sides if possible */
4023 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4025 * The following operations are NOT safe for floating point operations, for instance
4026 * 1.0 + inf == 2.0 + inf, =/=> x == y
4028 if (mode_is_int(get_irn_mode(left))) {
4029 unsigned lop = get_irn_opcode(left);
4031 if (lop == get_irn_opcode(right)) {
4032 ir_node *ll, *lr, *rl, *rr;
4034 /* same operation on both sides, try to remove */
4038 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4039 left = get_unop_op(left);
4040 right = get_unop_op(right);
4042 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4045 ll = get_Add_left(left);
4046 lr = get_Add_right(left);
4047 rl = get_Add_left(right);
4048 rr = get_Add_right(right);
4051 /* X + a CMP X + b ==> a CMP b */
4055 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4056 } else if (ll == rr) {
4057 /* X + a CMP b + X ==> a CMP b */
4061 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4062 } else if (lr == rl) {
4063 /* a + X CMP X + b ==> a CMP b */
4067 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4068 } else if (lr == rr) {
4069 /* a + X CMP b + X ==> a CMP b */
4073 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4077 ll = get_Sub_left(left);
4078 lr = get_Sub_right(left);
4079 rl = get_Sub_left(right);
4080 rr = get_Sub_right(right);
4083 /* X - a CMP X - b ==> a CMP b */
4087 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4088 } else if (lr == rr) {
4089 /* a - X CMP b - X ==> a CMP b */
4093 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4097 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4098 /* a ROTL X CMP b ROTL X ==> a CMP b */
4099 left = get_Rotl_left(left);
4100 right = get_Rotl_left(right);
4102 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4110 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4111 if (is_Add(left) || is_Sub(left)) {
4112 ir_node *ll = get_binop_left(left);
4113 ir_node *lr = get_binop_right(left);
4115 if (lr == right && is_Add(left)) {
4122 right = create_zero_const(get_irn_mode(left));
4124 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4127 if (is_Add(right) || is_Sub(right)) {
4128 ir_node *rl = get_binop_left(right);
4129 ir_node *rr = get_binop_right(right);
4131 if (rr == left && is_Add(right)) {
4138 right = create_zero_const(get_irn_mode(left));
4140 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4143 } /* mode_is_int(...) */
4144 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4146 /* replace mode_b compares with ands/ors */
4147 if (get_irn_mode(left) == mode_b) {
4148 ir_graph *irg = current_ir_graph;
4149 ir_node *block = get_nodes_block(n);
4153 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4154 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4155 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4156 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4157 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4158 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4159 default: bres = NULL;
4162 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4168 * First step: normalize the compare op
4169 * by placing the constant on the right side
4170 * or moving the lower address node to the left.
4172 if (!operands_are_normalized(left, right)) {
4178 proj_nr = get_inversed_pnc(proj_nr);
4183 * Second step: Try to reduce the magnitude
4184 * of a constant. This may help to generate better code
4185 * later and may help to normalize more compares.
4186 * Of course this is only possible for integer values.
4188 tv = value_of(right);
4189 if (tv != tarval_bad) {
4190 mode = get_irn_mode(right);
4192 /* TODO extend to arbitrary constants */
4193 if (is_Conv(left) && tarval_is_null(tv)) {
4194 ir_node *op = get_Conv_op(left);
4195 ir_mode *op_mode = get_irn_mode(op);
4198 * UpConv(x) REL 0 ==> x REL 0
4200 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4201 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4202 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4203 tv = get_mode_null(op_mode);
4207 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4211 if (tv != tarval_bad) {
4212 /* the following optimization is possible on modes without Overflow
4213 * on Unary Minus or on == and !=:
4214 * -a CMP c ==> a swap(CMP) -c
4216 * Beware: for two-complement Overflow may occur, so only == and != can
4217 * be optimized, see this:
4218 * -MININT < 0 =/=> MININT > 0 !!!
4220 if (is_Minus(left) &&
4221 (!mode_overflow_on_unary_Minus(mode) ||
4222 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4223 tv = tarval_neg(tv);
4225 if (tv != tarval_bad) {
4226 left = get_Minus_op(left);
4227 proj_nr = get_inversed_pnc(proj_nr);
4229 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4231 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4232 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4233 tv = tarval_not(tv);
4235 if (tv != tarval_bad) {
4236 left = get_Not_op(left);
4238 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4242 /* for integer modes, we have more */
4243 if (mode_is_int(mode)) {
4244 /* Ne includes Unordered which is not possible on integers.
4245 * However, frontends often use this wrong, so fix it here */
4246 if (proj_nr & pn_Cmp_Uo) {
4247 proj_nr &= ~pn_Cmp_Uo;
4248 set_Proj_proj(proj, proj_nr);
4251 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4252 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4253 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4254 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4256 if (tv != tarval_bad) {
4257 proj_nr ^= pn_Cmp_Eq;
4259 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4262 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4263 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4264 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4265 tv = tarval_add(tv, get_mode_one(mode));
4267 if (tv != tarval_bad) {
4268 proj_nr ^= pn_Cmp_Eq;
4270 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4274 /* the following reassociations work only for == and != */
4275 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4277 #if 0 /* Might be not that good in general */
4278 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4279 if (tarval_is_null(tv) && is_Sub(left)) {
4280 right = get_Sub_right(left);
4281 left = get_Sub_left(left);
4283 tv = value_of(right);
4285 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4289 if (tv != tarval_bad) {
4290 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4292 ir_node *c1 = get_Sub_right(left);
4293 tarval *tv2 = value_of(c1);
4295 if (tv2 != tarval_bad) {
4296 tv2 = tarval_add(tv, value_of(c1));
4298 if (tv2 != tarval_bad) {
4299 left = get_Sub_left(left);
4302 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4306 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4307 else if (is_Add(left)) {
4308 ir_node *a_l = get_Add_left(left);
4309 ir_node *a_r = get_Add_right(left);
4313 if (is_Const(a_l)) {
4315 tv2 = value_of(a_l);
4318 tv2 = value_of(a_r);
4321 if (tv2 != tarval_bad) {
4322 tv2 = tarval_sub(tv, tv2, NULL);
4324 if (tv2 != tarval_bad) {
4328 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4332 /* -a == c ==> a == -c, -a != c ==> a != -c */
4333 else if (is_Minus(left)) {
4334 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4336 if (tv2 != tarval_bad) {
4337 left = get_Minus_op(left);
4340 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4345 /* the following reassociations work only for <= */
4346 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4347 if (tv != tarval_bad) {
4348 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4349 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4355 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4356 switch (get_irn_opcode(left)) {
4360 c1 = get_And_right(left);
4363 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4364 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4366 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4368 /* TODO: move to constant evaluation */
4369 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4370 c1 = new_Const(mode_b, tv);
4371 DBG_OPT_CSTEVAL(proj, c1);
4375 if (tarval_is_single_bit(tv)) {
4377 * optimization for AND:
4379 * And(x, C) == C ==> And(x, C) != 0
4380 * And(x, C) != C ==> And(X, C) == 0
4382 * if C is a single Bit constant.
4385 /* check for Constant's match. We have check hare the tarvals,
4386 because our const might be changed */
4387 if (get_Const_tarval(c1) == tv) {
4388 /* fine: do the transformation */
4389 tv = get_mode_null(get_tarval_mode(tv));
4390 proj_nr ^= pn_Cmp_Leg;
4392 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4398 c1 = get_Or_right(left);
4399 if (is_Const(c1) && tarval_is_null(tv)) {
4401 * Or(x, C) == 0 && C != 0 ==> FALSE
4402 * Or(x, C) != 0 && C != 0 ==> TRUE
4404 if (! tarval_is_null(get_Const_tarval(c1))) {
4405 /* TODO: move to constant evaluation */
4406 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4407 c1 = new_Const(mode_b, tv);
4408 DBG_OPT_CSTEVAL(proj, c1);
4415 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4417 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4420 c1 = get_Shl_right(left);
4422 tarval *tv1 = get_Const_tarval(c1);
4423 ir_mode *mode = get_irn_mode(left);
4424 tarval *minus1 = get_mode_all_one(mode);
4425 tarval *amask = tarval_shr(minus1, tv1);
4426 tarval *cmask = tarval_shl(minus1, tv1);
4429 if (tarval_and(tv, cmask) != tv) {
4430 /* condition not met */
4431 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4432 c1 = new_Const(mode_b, tv);
4433 DBG_OPT_CSTEVAL(proj, c1);
4436 sl = get_Shl_left(left);
4437 blk = get_nodes_block(n);
4438 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4439 tv = tarval_shr(tv, tv1);
4441 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4446 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4448 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4451 c1 = get_Shr_right(left);
4453 tarval *tv1 = get_Const_tarval(c1);
4454 ir_mode *mode = get_irn_mode(left);
4455 tarval *minus1 = get_mode_all_one(mode);
4456 tarval *amask = tarval_shl(minus1, tv1);
4457 tarval *cmask = tarval_shr(minus1, tv1);
4460 if (tarval_and(tv, cmask) != tv) {
4461 /* condition not met */
4462 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4463 c1 = new_Const(mode_b, tv);
4464 DBG_OPT_CSTEVAL(proj, c1);
4467 sl = get_Shr_left(left);
4468 blk = get_nodes_block(n);
4469 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4470 tv = tarval_shl(tv, tv1);
4472 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4477 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4479 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4482 c1 = get_Shrs_right(left);
4484 tarval *tv1 = get_Const_tarval(c1);
4485 ir_mode *mode = get_irn_mode(left);
4486 tarval *minus1 = get_mode_all_one(mode);
4487 tarval *amask = tarval_shl(minus1, tv1);
4488 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4491 cond = tarval_sub(cond, tv1, NULL);
4492 cond = tarval_shrs(tv, cond);
4494 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4495 /* condition not met */
4496 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4497 c1 = new_Const(mode_b, tv);
4498 DBG_OPT_CSTEVAL(proj, c1);
4501 sl = get_Shrs_left(left);
4502 blk = get_nodes_block(n);
4503 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4504 tv = tarval_shl(tv, tv1);
4506 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4511 } /* tarval != bad */
4514 if (changed & 2) /* need a new Const */
4515 right = new_Const(mode, tv);
4517 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4518 ir_node *op = get_Proj_pred(left);
4520 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4521 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4522 ir_node *c = get_binop_right(op);
4525 tarval *tv = get_Const_tarval(c);
4527 if (tarval_is_single_bit(tv)) {
4528 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4529 ir_node *v = get_binop_left(op);
4530 ir_node *blk = get_irn_n(op, -1);
4531 ir_mode *mode = get_irn_mode(v);
4533 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4534 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4536 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4543 ir_node *block = get_nodes_block(n);
4545 /* create a new compare */
4546 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4547 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4551 } /* transform_node_Proj_Cmp */
4554 * Optimize CopyB(mem, x, x) into a Nop.
4556 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4557 ir_node *copyb = get_Proj_pred(proj);
4558 ir_node *a = get_CopyB_dst(copyb);
4559 ir_node *b = get_CopyB_src(copyb);
4562 switch (get_Proj_proj(proj)) {
4563 case pn_CopyB_X_regular:
4564 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4565 DBG_OPT_EXC_REM(proj);
4566 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4568 case pn_CopyB_M_except:
4569 case pn_CopyB_X_except:
4570 DBG_OPT_EXC_REM(proj);
4571 proj = get_irg_bad(current_ir_graph);
4578 } /* transform_node_Proj_CopyB */
4581 * Optimize Bounds(idx, idx, upper) into idx.
4583 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4584 ir_node *oldn = proj;
4585 ir_node *bound = get_Proj_pred(proj);
4586 ir_node *idx = get_Bound_index(bound);
4587 ir_node *pred = skip_Proj(idx);
4590 if (idx == get_Bound_lower(bound))
4592 else if (is_Bound(pred)) {
4594 * idx was Bounds checked in the same MacroBlock previously,
4595 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4597 ir_node *lower = get_Bound_lower(bound);
4598 ir_node *upper = get_Bound_upper(bound);
4599 if (get_Bound_lower(pred) == lower &&
4600 get_Bound_upper(pred) == upper &&
4601 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4603 * One could expect that we simply return the previous
4604 * Bound here. However, this would be wrong, as we could
4605 * add an exception Proj to a new location then.
4606 * So, we must turn in into a tuple.
4612 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4613 switch (get_Proj_proj(proj)) {
4615 DBG_OPT_EXC_REM(proj);
4616 proj = get_Bound_mem(bound);
4618 case pn_Bound_X_except:
4619 DBG_OPT_EXC_REM(proj);
4620 proj = get_irg_bad(current_ir_graph);
4624 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4626 case pn_Bound_X_regular:
4627 DBG_OPT_EXC_REM(proj);
4628 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4635 } /* transform_node_Proj_Bound */
4638 * Does all optimizations on nodes that must be done on it's Proj's
4639 * because of creating new nodes.
4641 static ir_node *transform_node_Proj(ir_node *proj) {
4642 ir_node *n = get_Proj_pred(proj);
4644 if (n->op->ops.transform_node_Proj)
4645 return n->op->ops.transform_node_Proj(proj);
4647 } /* transform_node_Proj */
4650 * Move Confirms down through Phi nodes.
4652 static ir_node *transform_node_Phi(ir_node *phi) {
4654 ir_mode *mode = get_irn_mode(phi);
4656 if (mode_is_reference(mode)) {
4657 n = get_irn_arity(phi);
4659 /* Beware of Phi0 */
4661 ir_node *pred = get_irn_n(phi, 0);
4662 ir_node *bound, *new_Phi, *block, **in;
4665 if (! is_Confirm(pred))
4668 bound = get_Confirm_bound(pred);
4669 pnc = get_Confirm_cmp(pred);
4671 NEW_ARR_A(ir_node *, in, n);
4672 in[0] = get_Confirm_value(pred);
4674 for (i = 1; i < n; ++i) {
4675 pred = get_irn_n(phi, i);
4677 if (! is_Confirm(pred) ||
4678 get_Confirm_bound(pred) != bound ||
4679 get_Confirm_cmp(pred) != pnc)
4681 in[i] = get_Confirm_value(pred);
4683 /* move the Confirm nodes "behind" the Phi */
4684 block = get_irn_n(phi, -1);
4685 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4686 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4690 } /* transform_node_Phi */
4693 * Returns the operands of a commutative bin-op, if one operand is
4694 * a const, it is returned as the second one.
4696 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4697 ir_node *op_a = get_binop_left(binop);
4698 ir_node *op_b = get_binop_right(binop);
4700 assert(is_op_commutative(get_irn_op(binop)));
4702 if (is_Const(op_a)) {
4709 } /* get_comm_Binop_Ops */
4712 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4713 * Such pattern may arise in bitfield stores.
4715 * value c4 value c4 & c2
4716 * AND c3 AND c1 | c3
4723 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4726 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4729 ir_node *and_l, *c3;
4730 ir_node *value, *c4;
4731 ir_node *new_and, *new_const, *block;
4732 ir_mode *mode = get_irn_mode(or);
4734 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4737 get_comm_Binop_Ops(or, &and, &c1);
4738 if (!is_Const(c1) || !is_And(and))
4741 get_comm_Binop_Ops(and, &or_l, &c2);
4745 tv1 = get_Const_tarval(c1);
4746 tv2 = get_Const_tarval(c2);
4748 tv = tarval_or(tv1, tv2);
4749 if (tarval_is_all_one(tv)) {
4750 /* the AND does NOT clear a bit with isn't set by the OR */
4751 set_Or_left(or, or_l);
4752 set_Or_right(or, c1);
4754 /* check for more */
4761 get_comm_Binop_Ops(or_l, &and_l, &c3);
4762 if (!is_Const(c3) || !is_And(and_l))
4765 get_comm_Binop_Ops(and_l, &value, &c4);
4769 /* ok, found the pattern, check for conditions */
4770 assert(mode == get_irn_mode(and));
4771 assert(mode == get_irn_mode(or_l));
4772 assert(mode == get_irn_mode(and_l));
4774 tv3 = get_Const_tarval(c3);
4775 tv4 = get_Const_tarval(c4);
4777 tv = tarval_or(tv4, tv2);
4778 if (!tarval_is_all_one(tv)) {
4779 /* have at least one 0 at the same bit position */
4783 n_tv4 = tarval_not(tv4);
4784 if (tv3 != tarval_and(tv3, n_tv4)) {
4785 /* bit in the or_mask is outside the and_mask */
4789 n_tv2 = tarval_not(tv2);
4790 if (tv1 != tarval_and(tv1, n_tv2)) {
4791 /* bit in the or_mask is outside the and_mask */
4795 /* ok, all conditions met */
4796 block = get_irn_n(or, -1);
4798 new_and = new_r_And(current_ir_graph, block,
4799 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4801 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4803 set_Or_left(or, new_and);
4804 set_Or_right(or, new_const);
4806 /* check for more */
4808 } /* transform_node_Or_bf_store */
4811 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4813 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4814 ir_mode *mode = get_irn_mode(or);
4815 ir_node *shl, *shr, *block;
4816 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4819 if (! mode_is_int(mode))
4822 shl = get_binop_left(or);
4823 shr = get_binop_right(or);
4832 } else if (!is_Shl(shl)) {
4834 } else if (!is_Shr(shr)) {
4837 x = get_Shl_left(shl);
4838 if (x != get_Shr_left(shr))
4841 c1 = get_Shl_right(shl);
4842 c2 = get_Shr_right(shr);
4843 if (is_Const(c1) && is_Const(c2)) {
4844 tv1 = get_Const_tarval(c1);
4845 if (! tarval_is_long(tv1))
4848 tv2 = get_Const_tarval(c2);
4849 if (! tarval_is_long(tv2))
4852 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4853 != (int) get_mode_size_bits(mode))
4856 /* yet, condition met */
4857 block = get_nodes_block(or);
4859 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4861 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4868 rotval = sub; /* a Rot right is not supported, so use a rot left */
4869 } else if (is_Sub(c2)) {
4875 if (get_Sub_right(sub) != v)
4878 c1 = get_Sub_left(sub);
4882 tv1 = get_Const_tarval(c1);
4883 if (! tarval_is_long(tv1))
4886 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4889 /* yet, condition met */
4890 block = get_nodes_block(or);
4892 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4894 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4896 } /* transform_node_Or_Rotl */
4901 static ir_node *transform_node_Or(ir_node *n) {
4902 ir_node *c, *oldn = n;
4903 ir_node *a = get_Or_left(n);
4904 ir_node *b = get_Or_right(n);
4907 if (is_Not(a) && is_Not(b)) {
4908 /* ~a | ~b = ~(a&b) */
4909 ir_node *block = get_nodes_block(n);
4911 mode = get_irn_mode(n);
4914 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4915 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4916 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4920 /* we can evaluate 2 Projs of the same Cmp */
4921 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4922 ir_node *pred_a = get_Proj_pred(a);
4923 ir_node *pred_b = get_Proj_pred(b);
4924 if (pred_a == pred_b) {
4925 dbg_info *dbgi = get_irn_dbg_info(n);
4926 ir_node *block = get_nodes_block(pred_a);
4927 pn_Cmp pn_a = get_Proj_proj(a);
4928 pn_Cmp pn_b = get_Proj_proj(b);
4929 /* yes, we can simply calculate with pncs */
4930 pn_Cmp new_pnc = pn_a | pn_b;
4932 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4937 mode = get_irn_mode(n);
4938 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4940 n = transform_node_Or_bf_store(n);
4941 n = transform_node_Or_Rotl(n);
4945 n = transform_bitwise_distributive(n, transform_node_Or);
4948 } /* transform_node_Or */
4952 static ir_node *transform_node(ir_node *n);
4955 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4957 * Should be moved to reassociation?
4959 static ir_node *transform_node_shift(ir_node *n) {
4960 ir_node *left, *right;
4962 tarval *tv1, *tv2, *res;
4963 ir_node *in[2], *irn, *block;
4965 left = get_binop_left(n);
4967 /* different operations */
4968 if (get_irn_op(left) != get_irn_op(n))
4971 right = get_binop_right(n);
4972 tv1 = value_of(right);
4973 if (tv1 == tarval_bad)
4976 tv2 = value_of(get_binop_right(left));
4977 if (tv2 == tarval_bad)
4980 res = tarval_add(tv1, tv2);
4981 mode = get_irn_mode(n);
4983 /* beware: a simple replacement works only, if res < modulo shift */
4985 int modulo_shf = get_mode_modulo_shift(mode);
4986 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4987 if (modulo_shf > 0) {
4988 tarval *modulo = new_tarval_from_long(modulo_shf,
4989 get_tarval_mode(res));
4991 /* shifting too much */
4992 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4994 ir_graph *irg = get_irn_irg(n);
4995 ir_node *block = get_nodes_block(n);
4996 dbg_info *dbgi = get_irn_dbg_info(n);
4997 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
4998 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5002 return new_Const(mode, get_mode_null(mode));
5006 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5009 /* ok, we can replace it */
5010 block = get_nodes_block(n);
5012 in[0] = get_binop_left(left);
5013 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5015 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5017 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5019 return transform_node(irn);
5020 } /* transform_node_shift */
5023 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5025 * - and, or, xor instead of &
5026 * - Shl, Shr, Shrs, rotl instead of >>
5027 * (with a special case for Or/Xor + Shrs)
5029 static ir_node *transform_node_bitop_shift(ir_node *n) {
5031 ir_node *right = get_binop_right(n);
5032 ir_mode *mode = get_irn_mode(n);
5033 ir_node *bitop_left;
5034 ir_node *bitop_right;
5046 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5048 if (!is_Const(right))
5051 left = get_binop_left(n);
5052 op_left = get_irn_op(left);
5053 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5056 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5057 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5058 /* TODO: test if sign bit is affectes */
5062 bitop_right = get_binop_right(left);
5063 if (!is_Const(bitop_right))
5066 bitop_left = get_binop_left(left);
5068 irg = get_irn_irg(n);
5069 block = get_nodes_block(n);
5070 dbgi = get_irn_dbg_info(n);
5071 tv1 = get_Const_tarval(bitop_right);
5072 tv2 = get_Const_tarval(right);
5074 assert(get_tarval_mode(tv1) == mode);
5077 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5078 tv_shift = tarval_shl(tv1, tv2);
5079 } else if(is_Shr(n)) {
5080 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5081 tv_shift = tarval_shr(tv1, tv2);
5082 } else if(is_Shrs(n)) {
5083 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5084 tv_shift = tarval_shrs(tv1, tv2);
5087 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5088 tv_shift = tarval_rotl(tv1, tv2);
5091 assert(get_tarval_mode(tv_shift) == mode);
5092 new_const = new_Const(mode, tv_shift);
5094 if (op_left == op_And) {
5095 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5096 } else if(op_left == op_Or) {
5097 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5099 assert(op_left == op_Eor);
5100 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5108 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5110 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5111 * (also with x >>s c1 when c1>=c2)
5113 static ir_node *transform_node_shl_shr(ir_node *n) {
5115 ir_node *right = get_binop_right(n);
5131 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5133 if (!is_Const(right))
5136 left = get_binop_left(n);
5137 mode = get_irn_mode(n);
5138 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5139 ir_node *shr_right = get_binop_right(left);
5141 if (!is_Const(shr_right))
5144 x = get_binop_left(left);
5145 tv_shr = get_Const_tarval(shr_right);
5146 tv_shl = get_Const_tarval(right);
5148 if (is_Shrs(left)) {
5149 /* shrs variant only allowed if c1 >= c2 */
5150 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5153 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5156 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5158 tv_mask = tarval_shl(tv_mask, tv_shl);
5159 } else if(is_Shr(n) && is_Shl(left)) {
5160 ir_node *shl_right = get_Shl_right(left);
5162 if (!is_Const(shl_right))
5165 x = get_Shl_left(left);
5166 tv_shr = get_Const_tarval(right);
5167 tv_shl = get_Const_tarval(shl_right);
5169 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5170 tv_mask = tarval_shr(tv_mask, tv_shr);
5175 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5176 assert(tv_mask != tarval_bad);
5177 assert(get_tarval_mode(tv_mask) == mode);
5179 irg = get_irn_irg(n);
5180 block = get_nodes_block(n);
5181 dbgi = get_irn_dbg_info(n);
5183 pnc = tarval_cmp(tv_shl, tv_shr);
5184 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5185 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5186 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5188 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5190 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5193 assert(pnc == pn_Cmp_Gt);
5194 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5195 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5196 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5199 new_const = new_Const(mode, tv_mask);
5200 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5208 static ir_node *transform_node_Shr(ir_node *n) {
5209 ir_node *c, *oldn = n;
5210 ir_node *left = get_Shr_left(n);
5211 ir_node *right = get_Shr_right(n);
5212 ir_mode *mode = get_irn_mode(n);
5214 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5215 n = transform_node_shift(n);
5218 n = transform_node_shl_shr(n);
5220 n = transform_node_bitop_shift(n);
5223 } /* transform_node_Shr */
5228 static ir_node *transform_node_Shrs(ir_node *n) {
5229 ir_node *c, *oldn = n;
5230 ir_node *a = get_Shrs_left(n);
5231 ir_node *b = get_Shrs_right(n);
5232 ir_mode *mode = get_irn_mode(n);
5234 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5235 n = transform_node_shift(n);
5238 n = transform_node_bitop_shift(n);
5241 } /* transform_node_Shrs */
5246 static ir_node *transform_node_Shl(ir_node *n) {
5247 ir_node *c, *oldn = n;
5248 ir_node *a = get_Shl_left(n);
5249 ir_node *b = get_Shl_right(n);
5250 ir_mode *mode = get_irn_mode(n);
5252 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5253 n = transform_node_shift(n);
5256 n = transform_node_shl_shr(n);
5258 n = transform_node_bitop_shift(n);
5261 } /* transform_node_Shl */
5266 static ir_node *transform_node_Rotl(ir_node *n) {
5267 ir_node *c, *oldn = n;
5268 ir_node *a = get_Rotl_left(n);
5269 ir_node *b = get_Rotl_right(n);
5270 ir_mode *mode = get_irn_mode(n);
5272 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5273 n = transform_node_shift(n);
5276 n = transform_node_bitop_shift(n);
5279 } /* transform_node_Rotl */
5284 static ir_node *transform_node_Conv(ir_node *n) {
5285 ir_node *c, *oldn = n;
5286 ir_node *a = get_Conv_op(n);
5288 if (is_const_Phi(a)) {
5289 c = apply_conv_on_phi(a, get_irn_mode(n));
5291 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5296 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5297 ir_mode *mode = get_irn_mode(n);
5298 return new_r_Unknown(current_ir_graph, mode);
5302 } /* transform_node_Conv */
5305 * Remove dead blocks and nodes in dead blocks
5306 * in keep alive list. We do not generate a new End node.
5308 static ir_node *transform_node_End(ir_node *n) {
5309 int i, j, n_keepalives = get_End_n_keepalives(n);
5312 NEW_ARR_A(ir_node *, in, n_keepalives);
5314 for (i = j = 0; i < n_keepalives; ++i) {
5315 ir_node *ka = get_End_keepalive(n, i);
5317 if (! is_Block_dead(ka)) {
5321 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5324 /* FIXME: beabi need to keep a Proj(M) */
5325 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
5328 if (j != n_keepalives)
5329 set_End_keepalives(n, j, in);
5331 } /* transform_node_End */
5333 /** returns 1 if a == -b */
5334 static int is_negated_value(ir_node *a, ir_node *b) {
5335 if (is_Minus(a) && get_Minus_op(a) == b)
5337 if (is_Minus(b) && get_Minus_op(b) == a)
5339 if (is_Sub(a) && is_Sub(b)) {
5340 ir_node *a_left = get_Sub_left(a);
5341 ir_node *a_right = get_Sub_right(a);
5342 ir_node *b_left = get_Sub_left(b);
5343 ir_node *b_right = get_Sub_right(b);
5345 if (a_left == b_right && a_right == b_left)
5353 * Optimize a Mux into some simpler cases.
5355 static ir_node *transform_node_Mux(ir_node *n) {
5356 ir_node *oldn = n, *sel = get_Mux_sel(n);
5357 ir_mode *mode = get_irn_mode(n);
5358 ir_node *t = get_Mux_true(n);
5359 ir_node *f = get_Mux_false(n);
5360 ir_graph *irg = current_ir_graph;
5362 /* first normalization step: move a possible zero to the false case */
5364 ir_node *cmp = get_Proj_pred(sel);
5367 if (is_Const(t) && is_Const_null(t)) {
5370 /* Mux(x, 0, y) => Mux(x, y, 0) */
5371 pn_Cmp pnc = get_Proj_proj(sel);
5372 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5373 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5374 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5382 /* note: after normalization, false can only happen on default */
5383 if (mode == mode_b) {
5384 dbg_info *dbg = get_irn_dbg_info(n);
5385 ir_node *block = get_nodes_block(n);
5386 ir_graph *irg = current_ir_graph;
5389 tarval *tv_t = get_Const_tarval(t);
5390 if (tv_t == tarval_b_true) {
5392 /* Muxb(sel, true, false) = sel */
5393 assert(get_Const_tarval(f) == tarval_b_false);
5394 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5397 /* Muxb(sel, true, x) = Or(sel, x) */
5398 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5403 } else if (is_Const(f)) {
5404 tarval *tv_f = get_Const_tarval(f);
5405 if (tv_f == tarval_b_true) {
5406 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5407 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5408 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5409 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5412 /* Muxb(sel, x, false) = And(sel, x) */
5413 assert(tv_f == tarval_b_false);
5414 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5415 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5421 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5422 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5423 tarval *a = get_Const_tarval(t);
5424 tarval *b = get_Const_tarval(f);
5425 tarval *null = get_tarval_null(mode);
5428 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5429 diff = tarval_sub(a, b, NULL);
5432 diff = tarval_sub(b, a, NULL);
5436 if (diff == get_tarval_one(mode) && min != null) {
5437 dbg_info *dbg = get_irn_dbg_info(n);
5438 ir_node *block = get_nodes_block(n);
5439 ir_graph *irg = current_ir_graph;
5440 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5441 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5442 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5443 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5449 ir_node *cmp = get_Proj_pred(sel);
5450 long pn = get_Proj_proj(sel);
5453 * Note: normalization puts the constant on the right side,
5454 * so we check only one case.
5456 * Note further that these optimization work even for floating point
5457 * with NaN's because -NaN == NaN.
5458 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5462 ir_node *cmp_r = get_Cmp_right(cmp);
5463 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5464 ir_node *block = get_nodes_block(n);
5465 ir_node *cmp_l = get_Cmp_left(cmp);
5467 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5470 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5471 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5473 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5474 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5476 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5478 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5479 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5481 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5482 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5484 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5486 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5491 if (mode_is_int(mode)) {
5493 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5494 /* Mux((a & b) != 0, c, 0) */
5495 ir_node *and_r = get_And_right(cmp_l);
5498 if (and_r == t && f == cmp_r) {
5499 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5500 if (pn == pn_Cmp_Lg) {
5501 /* Mux((a & 2^C) != 0, 2^C, 0) */
5503 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5505 /* Mux((a & 2^C) == 0, 2^C, 0) */
5506 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5507 block, cmp_l, t, mode);
5508 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5513 if (is_Shl(and_r)) {
5514 ir_node *shl_l = get_Shl_left(and_r);
5515 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5516 if (and_r == t && f == cmp_r) {
5517 if (pn == pn_Cmp_Lg) {
5518 /* (a & (1 << n)) != 0, (1 << n), 0) */
5520 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5522 /* (a & (1 << n)) == 0, (1 << n), 0) */
5523 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5524 block, cmp_l, t, mode);
5525 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5531 and_l = get_And_left(cmp_l);
5532 if (is_Shl(and_l)) {
5533 ir_node *shl_l = get_Shl_left(and_l);
5534 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5535 if (and_l == t && f == cmp_r) {
5536 if (pn == pn_Cmp_Lg) {
5537 /* ((1 << n) & a) != 0, (1 << n), 0) */
5539 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5541 /* ((1 << n) & a) == 0, (1 << n), 0) */
5542 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5543 block, cmp_l, t, mode);
5544 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5555 return arch_transform_node_Mux(n);
5556 } /* transform_node_Mux */
5559 * optimize Sync nodes that have other syncs as input we simply add the inputs
5560 * of the other sync to our own inputs
5562 static ir_node *transform_node_Sync(ir_node *n) {
5563 int arity = get_Sync_n_preds(n);
5566 for (i = 0; i < arity;) {
5567 ir_node *pred = get_Sync_pred(n, i);
5571 if (!is_Sync(pred)) {
5579 pred_arity = get_Sync_n_preds(pred);
5580 for (j = 0; j < pred_arity; ++j) {
5581 ir_node *pred_pred = get_Sync_pred(pred, j);
5586 add_irn_n(n, pred_pred);
5590 if (get_Sync_pred(n, k) == pred_pred) break;
5595 /* rehash the sync node */
5596 add_identities(current_ir_graph->value_table, n);
5602 * Tries several [inplace] [optimizing] transformations and returns an
5603 * equivalent node. The difference to equivalent_node() is that these
5604 * transformations _do_ generate new nodes, and thus the old node must
5605 * not be freed even if the equivalent node isn't the old one.
5607 static ir_node *transform_node(ir_node *n) {
5611 * Transform_node is the only "optimizing transformation" that might
5612 * return a node with a different opcode. We iterate HERE until fixpoint
5613 * to get the final result.
5617 if (n->op->ops.transform_node)
5618 n = n->op->ops.transform_node(n);
5619 } while (oldn != n);
5622 } /* transform_node */
5625 * Sets the default transform node operation for an ir_op_ops.
5627 * @param code the opcode for the default operation
5628 * @param ops the operations initialized
5633 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5637 ops->transform_node = transform_node_##a; \
5639 #define CASE_PROJ(a) \
5641 ops->transform_node_Proj = transform_node_Proj_##a; \
5643 #define CASE_PROJ_EX(a) \
5645 ops->transform_node = transform_node_##a; \
5646 ops->transform_node_Proj = transform_node_Proj_##a; \
5655 CASE_PROJ_EX(DivMod);
5689 } /* firm_set_default_transform_node */
5692 /* **************** Common Subexpression Elimination **************** */
5694 /** The size of the hash table used, should estimate the number of nodes
5696 #define N_IR_NODES 512
5698 /** Compares the attributes of two Const nodes. */
5699 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5700 return (get_Const_tarval(a) != get_Const_tarval(b))
5701 || (get_Const_type(a) != get_Const_type(b));
5702 } /* node_cmp_attr_Const */
5704 /** Compares the attributes of two Proj nodes. */
5705 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5706 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5707 } /* node_cmp_attr_Proj */
5709 /** Compares the attributes of two Filter nodes. */
5710 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5711 return get_Filter_proj(a) != get_Filter_proj(b);
5712 } /* node_cmp_attr_Filter */
5714 /** Compares the attributes of two Alloc nodes. */
5715 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5716 const alloc_attr *pa = get_irn_alloc_attr(a);
5717 const alloc_attr *pb = get_irn_alloc_attr(b);
5718 return (pa->where != pb->where) || (pa->type != pb->type);
5719 } /* node_cmp_attr_Alloc */
5721 /** Compares the attributes of two Free nodes. */
5722 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5723 const free_attr *pa = get_irn_free_attr(a);
5724 const free_attr *pb = get_irn_free_attr(b);
5725 return (pa->where != pb->where) || (pa->type != pb->type);
5726 } /* node_cmp_attr_Free */
5728 /** Compares the attributes of two SymConst nodes. */
5729 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5730 const symconst_attr *pa = get_irn_symconst_attr(a);
5731 const symconst_attr *pb = get_irn_symconst_attr(b);
5732 return (pa->kind != pb->kind)
5733 || (pa->sym.type_p != pb->sym.type_p)
5734 || (pa->tp != pb->tp);
5735 } /* node_cmp_attr_SymConst */
5737 /** Compares the attributes of two Call nodes. */
5738 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5739 return get_irn_call_attr(a) != get_irn_call_attr(b);
5740 } /* node_cmp_attr_Call */
5742 /** Compares the attributes of two Sel nodes. */
5743 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5744 const ir_entity *a_ent = get_Sel_entity(a);
5745 const ir_entity *b_ent = get_Sel_entity(b);
5748 (a_ent->kind != b_ent->kind) ||
5749 (a_ent->name != b_ent->name) ||
5750 (a_ent->owner != b_ent->owner) ||
5751 (a_ent->ld_name != b_ent->ld_name) ||
5752 (a_ent->type != b_ent->type);
5754 /* Matze: inlining of functions can produce 2 entities with same type,
5756 return a_ent != b_ent;
5757 } /* node_cmp_attr_Sel */
5759 /** Compares the attributes of two Phi nodes. */
5760 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5761 /* we can only enter this function if both nodes have the same number of inputs,
5762 hence it is enough to check if one of them is a Phi0 */
5764 /* check the Phi0 pos attribute */
5765 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5768 } /* node_cmp_attr_Phi */
5770 /** Compares the attributes of two Conv nodes. */
5771 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5772 return get_Conv_strict(a) != get_Conv_strict(b);
5773 } /* node_cmp_attr_Conv */
5775 /** Compares the attributes of two Cast nodes. */
5776 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5777 return get_Cast_type(a) != get_Cast_type(b);
5778 } /* node_cmp_attr_Cast */
5780 /** Compares the attributes of two Load nodes. */
5781 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5782 if (get_Load_volatility(a) == volatility_is_volatile ||
5783 get_Load_volatility(b) == volatility_is_volatile)
5784 /* NEVER do CSE on volatile Loads */
5786 /* do not CSE Loads with different alignment. Be conservative. */
5787 if (get_Load_align(a) != get_Load_align(b))
5790 return get_Load_mode(a) != get_Load_mode(b);
5791 } /* node_cmp_attr_Load */
5793 /** Compares the attributes of two Store nodes. */
5794 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5795 /* do not CSE Stores with different alignment. Be conservative. */
5796 if (get_Store_align(a) != get_Store_align(b))
5799 /* NEVER do CSE on volatile Stores */
5800 return (get_Store_volatility(a) == volatility_is_volatile ||
5801 get_Store_volatility(b) == volatility_is_volatile);
5802 } /* node_cmp_attr_Store */
5804 /** Compares two exception attributes */
5805 static int node_cmp_exception(ir_node *a, ir_node *b) {
5806 const except_attr *ea = get_irn_except_attr(a);
5807 const except_attr *eb = get_irn_except_attr(b);
5809 return ea->pin_state != eb->pin_state;
5812 #define node_cmp_attr_Bound node_cmp_exception
5814 /** Compares the attributes of two Div nodes. */
5815 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5816 const divmod_attr *ma = get_irn_divmod_attr(a);
5817 const divmod_attr *mb = get_irn_divmod_attr(b);
5818 return ma->exc.pin_state != mb->exc.pin_state ||
5819 ma->res_mode != mb->res_mode ||
5820 ma->no_remainder != mb->no_remainder;
5821 } /* node_cmp_attr_Div */
5823 /** Compares the attributes of two DivMod nodes. */
5824 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5825 const divmod_attr *ma = get_irn_divmod_attr(a);
5826 const divmod_attr *mb = get_irn_divmod_attr(b);
5827 return ma->exc.pin_state != mb->exc.pin_state ||
5828 ma->res_mode != mb->res_mode;
5829 } /* node_cmp_attr_DivMod */
5831 /** Compares the attributes of two Mod nodes. */
5832 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5833 const divmod_attr *ma = get_irn_divmod_attr(a);
5834 const divmod_attr *mb = get_irn_divmod_attr(b);
5835 return ma->exc.pin_state != mb->exc.pin_state ||
5836 ma->res_mode != mb->res_mode;
5837 } /* node_cmp_attr_Mod */
5839 /** Compares the attributes of two Quot nodes. */
5840 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5841 const divmod_attr *ma = get_irn_divmod_attr(a);
5842 const divmod_attr *mb = get_irn_divmod_attr(b);
5843 return ma->exc.pin_state != mb->exc.pin_state ||
5844 ma->res_mode != mb->res_mode;
5845 } /* node_cmp_attr_Quot */
5847 /** Compares the attributes of two Confirm nodes. */
5848 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5849 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5850 } /* node_cmp_attr_Confirm */
5852 /** Compares the attributes of two ASM nodes. */
5853 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5855 const ir_asm_constraint *ca;
5856 const ir_asm_constraint *cb;
5859 if (get_ASM_text(a) != get_ASM_text(b))
5862 /* Should we really check the constraints here? Should be better, but is strange. */
5863 n = get_ASM_n_input_constraints(a);
5864 if (n != get_ASM_n_input_constraints(b))
5867 ca = get_ASM_input_constraints(a);
5868 cb = get_ASM_input_constraints(b);
5869 for (i = 0; i < n; ++i) {
5870 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5874 n = get_ASM_n_output_constraints(a);
5875 if (n != get_ASM_n_output_constraints(b))
5878 ca = get_ASM_output_constraints(a);
5879 cb = get_ASM_output_constraints(b);
5880 for (i = 0; i < n; ++i) {
5881 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5885 n = get_ASM_n_clobbers(a);
5886 if (n != get_ASM_n_clobbers(b))
5889 cla = get_ASM_clobbers(a);
5890 clb = get_ASM_clobbers(b);
5891 for (i = 0; i < n; ++i) {
5892 if (cla[i] != clb[i])
5896 } /* node_cmp_attr_ASM */
5899 * Set the default node attribute compare operation for an ir_op_ops.
5901 * @param code the opcode for the default operation
5902 * @param ops the operations initialized
5907 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5911 ops->node_cmp_attr = node_cmp_attr_##a; \
5942 } /* firm_set_default_node_cmp_attr */
5945 * Compare function for two nodes in the value table. Gets two
5946 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5948 int identities_cmp(const void *elt, const void *key) {
5949 ir_node *a = (ir_node *)elt;
5950 ir_node *b = (ir_node *)key;
5953 if (a == b) return 0;
5955 if ((get_irn_op(a) != get_irn_op(b)) ||
5956 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5958 /* compare if a's in and b's in are of equal length */
5959 irn_arity_a = get_irn_intra_arity(a);
5960 if (irn_arity_a != get_irn_intra_arity(b))
5963 if (get_irn_pinned(a) == op_pin_state_pinned) {
5964 /* for pinned nodes, the block inputs must be equal */
5965 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5967 } else if (! get_opt_global_cse()) {
5968 /* for block-local CSE both nodes must be in the same MacroBlock */
5969 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5973 /* compare a->in[0..ins] with b->in[0..ins] */
5974 for (i = 0; i < irn_arity_a; i++)
5975 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5979 * here, we already now that the nodes are identical except their
5982 if (a->op->ops.node_cmp_attr)
5983 return a->op->ops.node_cmp_attr(a, b);
5986 } /* identities_cmp */
5989 * Calculate a hash value of a node.
5991 * @param node The IR-node
5993 unsigned ir_node_hash(const ir_node *node) {
5994 return node->op->ops.hash(node);
5995 } /* ir_node_hash */
5998 pset *new_identities(void) {
5999 return new_pset(identities_cmp, N_IR_NODES);
6000 } /* new_identities */
6002 void del_identities(pset *value_table) {
6003 del_pset(value_table);
6004 } /* del_identities */
6007 * Normalize a node by putting constants (and operands with larger
6008 * node index) on the right (operator side).
6010 * @param n The node to normalize
6012 static void normalize_node(ir_node *n) {
6013 if (is_op_commutative(get_irn_op(n))) {
6014 ir_node *l = get_binop_left(n);
6015 ir_node *r = get_binop_right(n);
6017 /* For commutative operators perform a OP b == b OP a but keep
6018 * constants on the RIGHT side. This helps greatly in some
6019 * optimizations. Moreover we use the idx number to make the form
6021 if (!operands_are_normalized(l, r)) {
6022 set_binop_left(n, r);
6023 set_binop_right(n, l);
6026 } /* normalize_node */
6029 * Update the nodes after a match in the value table. If both nodes have
6030 * the same MacroBlock but different Blocks, we must ensure that the node
6031 * with the dominating Block (the node that is near to the MacroBlock header
6032 * is stored in the table.
6033 * Because a MacroBlock has only one "non-exception" flow, we don't need
6034 * dominance info here: We known, that one block must dominate the other and
6035 * following the only block input will allow to find it.
6037 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6038 ir_node *known_blk, *new_block, *block, *mbh;
6040 if (get_opt_global_cse()) {
6041 /* Block inputs are meaning less */
6044 known_blk = get_irn_n(known_irn, -1);
6045 new_block = get_irn_n(new_ir_node, -1);
6046 if (known_blk == new_block) {
6047 /* already in the same block */
6051 * We expect the typical case when we built the graph. In that case, the
6052 * known_irn is already the upper one, so checking this should be faster.
6055 mbh = get_Block_MacroBlock(new_block);
6057 if (block == known_blk) {
6058 /* ok, we have found it: known_block dominates new_block as expected */
6063 * We have reached the MacroBlock header NOT founding
6064 * the known_block. new_block must dominate known_block.
6067 set_irn_n(known_irn, -1, new_block);
6070 assert(get_Block_n_cfgpreds(block) == 1);
6071 block = get_Block_cfgpred_block(block, 0);
6073 } /* update_value_table */
6076 * Return the canonical node computing the same value as n.
6077 * Looks up the node in a hash table, enters it in the table
6078 * if it isn't there yet.
6080 * @param value_table the HashSet containing all nodes in the
6082 * @param n the node to look up
6084 * @return a node that computes the same value as n or n if no such
6085 * node could be found
6087 ir_node *identify_remember(pset *value_table, ir_node *n) {
6090 if (!value_table) return n;
6093 /* lookup or insert in hash table with given hash key. */
6094 o = pset_insert(value_table, n, ir_node_hash(n));
6097 update_known_irn(o, n);
6101 } /* identify_remember */
6104 * During construction we set the op_pin_state_pinned flag in the graph right when the
6105 * optimization is performed. The flag turning on procedure global cse could
6106 * be changed between two allocations. This way we are safe.
6108 * @param value_table The value table
6109 * @param n The node to lookup
6111 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
6114 n = identify_remember(value_table, n);
6115 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6116 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6118 } /* identify_cons */
6120 /* Add a node to the identities value table. */
6121 void add_identities(pset *value_table, ir_node *node) {
6122 if (get_opt_cse() && is_no_Block(node))
6123 identify_remember(value_table, node);
6124 } /* add_identities */
6126 /* Visit each node in the value table of a graph. */
6127 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6129 ir_graph *rem = current_ir_graph;
6131 current_ir_graph = irg;
6132 foreach_pset(irg->value_table, node)
6134 current_ir_graph = rem;
6135 } /* visit_all_identities */
6138 * Garbage in, garbage out. If a node has a dead input, i.e., the
6139 * Bad node is input to the node, return the Bad node.
6141 static ir_node *gigo(ir_node *node) {
6143 ir_op *op = get_irn_op(node);
6145 /* remove garbage blocks by looking at control flow that leaves the block
6146 and replacing the control flow by Bad. */
6147 if (get_irn_mode(node) == mode_X) {
6148 ir_node *block = get_nodes_block(skip_Proj(node));
6150 /* Don't optimize nodes in immature blocks. */
6151 if (!get_Block_matured(block))
6153 /* Don't optimize End, may have Bads. */
6154 if (op == op_End) return node;
6156 if (is_Block(block)) {
6157 if (is_Block_dead(block)) {
6158 /* control flow from dead block is dead */
6162 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6163 if (!is_Bad(get_irn_n(block, i)))
6167 ir_graph *irg = get_irn_irg(block);
6168 /* the start block is never dead */
6169 if (block != get_irg_start_block(irg)
6170 && block != get_irg_end_block(irg)) {
6172 * Do NOT kill control flow without setting
6173 * the block to dead of bad things can happen:
6174 * We get a Block that is not reachable be irg_block_walk()
6175 * but can be found by irg_walk()!
6177 set_Block_dead(block);
6184 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6185 blocks predecessors is dead. */
6186 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6187 irn_arity = get_irn_arity(node);
6190 * Beware: we can only read the block of a non-floating node.
6192 if (is_irn_pinned_in_irg(node) &&
6193 is_Block_dead(get_nodes_block(skip_Proj(node))))
6196 for (i = 0; i < irn_arity; i++) {
6197 ir_node *pred = get_irn_n(node, i);
6202 /* Propagating Unknowns here seems to be a bad idea, because
6203 sometimes we need a node as a input and did not want that
6205 However, it might be useful to move this into a later phase
6206 (if you think that optimizing such code is useful). */
6207 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6208 return new_Unknown(get_irn_mode(node));
6213 /* With this code we violate the agreement that local_optimize
6214 only leaves Bads in Block, Phi and Tuple nodes. */
6215 /* If Block has only Bads as predecessors it's garbage. */
6216 /* If Phi has only Bads as predecessors it's garbage. */
6217 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6218 irn_arity = get_irn_arity(node);
6219 for (i = 0; i < irn_arity; i++) {
6220 if (!is_Bad(get_irn_n(node, i))) break;
6222 if (i == irn_arity) node = new_Bad();
6229 * These optimizations deallocate nodes from the obstack.
6230 * It can only be called if it is guaranteed that no other nodes
6231 * reference this one, i.e., right after construction of a node.
6233 * @param n The node to optimize
6235 * current_ir_graph must be set to the graph of the node!
6237 ir_node *optimize_node(ir_node *n) {
6240 ir_opcode iro = get_irn_opcode(n);
6242 /* Always optimize Phi nodes: part of the construction. */
6243 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6245 /* constant expression evaluation / constant folding */
6246 if (get_opt_constant_folding()) {
6247 /* neither constants nor Tuple values can be evaluated */
6248 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6249 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6250 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6251 /* try to evaluate */
6252 tv = computed_value(n);
6253 if (tv != tarval_bad) {
6255 ir_type *old_tp = get_irn_type(n);
6256 int i, arity = get_irn_arity(n);
6260 * Try to recover the type of the new expression.
6262 for (i = 0; i < arity && !old_tp; ++i)
6263 old_tp = get_irn_type(get_irn_n(n, i));
6266 * we MUST copy the node here temporary, because it's still needed
6267 * for DBG_OPT_CSTEVAL
6269 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6270 oldn = alloca(node_size);
6272 memcpy(oldn, n, node_size);
6273 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6275 /* ARG, copy the in array, we need it for statistics */
6276 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6278 /* note the inplace edges module */
6279 edges_node_deleted(n, current_ir_graph);
6281 /* evaluation was successful -- replace the node. */
6282 irg_kill_node(current_ir_graph, n);
6283 nw = new_Const(get_tarval_mode(tv), tv);
6285 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6286 set_Const_type(nw, old_tp);
6287 DBG_OPT_CSTEVAL(oldn, nw);
6288 tarval_enable_fp_ops(old_fp_mode);
6291 tarval_enable_fp_ops(old_fp_mode);
6295 /* remove unnecessary nodes */
6296 if (get_opt_algebraic_simplification() ||
6297 (iro == iro_Phi) || /* always optimize these nodes. */
6299 (iro == iro_Proj) ||
6300 (iro == iro_Block) ) /* Flags tested local. */
6301 n = equivalent_node(n);
6303 /* Common Subexpression Elimination.
6305 * Checks whether n is already available.
6306 * The block input is used to distinguish different subexpressions. Right
6307 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6308 * subexpressions within a block.
6311 n = identify_cons(current_ir_graph->value_table, n);
6314 edges_node_deleted(oldn, current_ir_graph);
6316 /* We found an existing, better node, so we can deallocate the old node. */
6317 irg_kill_node(current_ir_graph, oldn);
6321 /* Some more constant expression evaluation that does not allow to
6323 iro = get_irn_opcode(n);
6324 if (get_opt_algebraic_simplification() ||
6325 (iro == iro_Cond) ||
6326 (iro == iro_Proj)) /* Flags tested local. */
6327 n = transform_node(n);
6329 /* Remove nodes with dead (Bad) input.
6330 Run always for transformation induced Bads. */
6333 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6334 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6336 n = identify_remember(current_ir_graph->value_table, o);
6342 } /* optimize_node */
6346 * These optimizations never deallocate nodes (in place). This can cause dead
6347 * nodes lying on the obstack. Remove these by a dead node elimination,
6348 * i.e., a copying garbage collection.
6350 ir_node *optimize_in_place_2(ir_node *n) {
6353 ir_opcode iro = get_irn_opcode(n);
6355 if (!get_opt_optimize() && !is_Phi(n)) return n;
6357 /* constant expression evaluation / constant folding */
6358 if (get_opt_constant_folding()) {
6359 /* neither constants nor Tuple values can be evaluated */
6360 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6361 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6362 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6363 /* try to evaluate */
6364 tv = computed_value(n);
6365 if (tv != tarval_bad) {
6366 /* evaluation was successful -- replace the node. */
6367 ir_type *old_tp = get_irn_type(n);
6368 int i, arity = get_irn_arity(n);
6371 * Try to recover the type of the new expression.
6373 for (i = 0; i < arity && !old_tp; ++i)
6374 old_tp = get_irn_type(get_irn_n(n, i));
6376 n = new_Const(get_tarval_mode(tv), tv);
6378 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6379 set_Const_type(n, old_tp);
6381 DBG_OPT_CSTEVAL(oldn, n);
6382 tarval_enable_fp_ops(old_fp_mode);
6385 tarval_enable_fp_ops(old_fp_mode);
6389 /* remove unnecessary nodes */
6390 if (get_opt_constant_folding() ||
6391 (iro == iro_Phi) || /* always optimize these nodes. */
6392 (iro == iro_Id) || /* ... */
6393 (iro == iro_Proj) || /* ... */
6394 (iro == iro_Block) ) /* Flags tested local. */
6395 n = equivalent_node(n);
6397 /** common subexpression elimination **/
6398 /* Checks whether n is already available. */
6399 /* The block input is used to distinguish different subexpressions. Right
6400 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6401 subexpressions within a block. */
6402 if (get_opt_cse()) {
6404 n = identify_remember(current_ir_graph->value_table, o);
6409 /* Some more constant expression evaluation. */
6410 iro = get_irn_opcode(n);
6411 if (get_opt_constant_folding() ||
6412 (iro == iro_Cond) ||
6413 (iro == iro_Proj)) /* Flags tested local. */
6414 n = transform_node(n);
6416 /* Remove nodes with dead (Bad) input.
6417 Run always for transformation induced Bads. */
6420 /* Now we can verify the node, as it has no dead inputs any more. */
6423 /* Now we have a legal, useful node. Enter it in hash table for cse.
6424 Blocks should be unique anyways. (Except the successor of start:
6425 is cse with the start block!) */
6426 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6428 n = identify_remember(current_ir_graph->value_table, o);
6434 } /* optimize_in_place_2 */
6437 * Wrapper for external use, set proper status bits after optimization.
6439 ir_node *optimize_in_place(ir_node *n) {
6440 /* Handle graph state */
6441 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6443 if (get_opt_global_cse())
6444 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6445 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6446 set_irg_outs_inconsistent(current_ir_graph);
6448 /* FIXME: Maybe we could also test whether optimizing the node can
6449 change the control graph. */
6450 set_irg_doms_inconsistent(current_ir_graph);
6451 return optimize_in_place_2(n);
6452 } /* optimize_in_place */
6455 * Calculate a hash value of a Const node.
6457 static unsigned hash_Const(const ir_node *node) {
6460 /* special value for const, as they only differ in their tarval. */
6461 h = HASH_PTR(node->attr.con.tv);
6462 h = 9*h + HASH_PTR(get_irn_mode(node));
6468 * Calculate a hash value of a SymConst node.
6470 static unsigned hash_SymConst(const ir_node *node) {
6473 /* special value for const, as they only differ in their symbol. */
6474 h = HASH_PTR(node->attr.symc.sym.type_p);
6475 h = 9*h + HASH_PTR(get_irn_mode(node));
6478 } /* hash_SymConst */
6481 * Set the default hash operation in an ir_op_ops.
6483 * @param code the opcode for the default operation
6484 * @param ops the operations initialized
6489 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6493 ops->hash = hash_##a; \
6496 /* hash function already set */
6497 if (ops->hash != NULL)
6504 /* use input/mode default hash if no function was given */
6505 ops->hash = firm_default_hash;
6513 * Sets the default operation for an ir_ops.
6515 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6516 ops = firm_set_default_hash(code, ops);
6517 ops = firm_set_default_computed_value(code, ops);
6518 ops = firm_set_default_equivalent_node(code, ops);
6519 ops = firm_set_default_transform_node(code, ops);
6520 ops = firm_set_default_node_cmp_attr(code, ops);
6521 ops = firm_set_default_get_type(code, ops);
6522 ops = firm_set_default_get_type_attr(code, ops);
6523 ops = firm_set_default_get_entity_attr(code, ops);
6526 } /* firm_set_default_operations */