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 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if (tarval_is_null(ta) || tarval_is_null(tb))
149 return get_mode_null(m);
152 } /* computed_value_Carry */
155 * Return the value of a Borrow.
158 static tarval *computed_value_Borrow(ir_node *n) {
159 ir_node *a = get_binop_left(n);
160 ir_node *b = get_binop_right(n);
161 ir_mode *m = get_irn_mode(n);
163 tarval *ta = value_of(a);
164 tarval *tb = value_of(b);
166 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
168 } else if (tarval_is_null(ta)) {
169 return get_mode_null(m);
172 } /* computed_value_Borrow */
175 * Return the value of an unary Minus.
177 static tarval *computed_value_Minus(ir_node *n) {
178 ir_node *a = get_Minus_op(n);
179 tarval *ta = value_of(a);
181 if (ta != tarval_bad)
182 return tarval_neg(ta);
185 } /* computed_value_Minus */
188 * Return the value of a Mul.
190 static tarval *computed_value_Mul(ir_node *n) {
191 ir_node *a = get_Mul_left(n);
192 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 mode = get_irn_mode(n);
199 if (mode != get_irn_mode(a)) {
200 /* n * n = 2n bit multiplication */
201 ta = tarval_convert_to(ta, mode);
202 tb = tarval_convert_to(tb, mode);
205 if (ta != tarval_bad && tb != tarval_bad) {
206 return tarval_mul(ta, tb);
208 /* a*0 = 0 or 0*b = 0 */
209 if (ta == get_mode_null(mode))
211 if (tb == get_mode_null(mode))
215 } /* computed_value_Mul */
218 * Return the value of a floating point Quot.
220 static tarval *computed_value_Quot(ir_node *n) {
221 ir_node *a = get_Quot_left(n);
222 ir_node *b = get_Quot_right(n);
224 tarval *ta = value_of(a);
225 tarval *tb = value_of(b);
227 /* This was missing in original implementation. Why? */
228 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
229 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
230 return tarval_quo(ta, tb);
233 } /* computed_value_Quot */
236 * Calculate the value of an integer Div of two nodes.
237 * Special case: 0 / b
239 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
240 tarval *ta = value_of(a);
241 tarval *tb = value_of(b);
243 /* Compute c1 / c2 or 0 / a, a != 0 */
244 if (ta != tarval_bad) {
245 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
246 return tarval_div(ta, tb);
247 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
251 } /* do_computed_value_Div */
254 * Return the value of an integer Div.
256 static tarval *computed_value_Div(ir_node *n) {
257 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
258 } /* computed_value_Div */
261 * Calculate the value of an integer Mod of two nodes.
262 * Special case: a % 1
264 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 /* Compute c1 % c2 or a % 1 */
269 if (tb != tarval_bad) {
270 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
271 return tarval_mod(ta, tb);
272 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
273 return get_mode_null(get_irn_mode(a));
276 } /* do_computed_value_Mod */
279 * Return the value of an integer Mod.
281 static tarval *computed_value_Mod(ir_node *n) {
282 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
283 } /* computed_value_Mod */
286 * Return the value of an Abs.
288 static tarval *computed_value_Abs(ir_node *n) {
289 ir_node *a = get_Abs_op(n);
290 tarval *ta = value_of(a);
292 if (ta != tarval_bad)
293 return tarval_abs(ta);
296 } /* computed_value_Abs */
299 * Return the value of an And.
300 * Special case: a & 0, 0 & b
302 static tarval *computed_value_And(ir_node *n) {
303 ir_node *a = get_And_left(n);
304 ir_node *b = get_And_right(n);
306 tarval *ta = value_of(a);
307 tarval *tb = value_of(b);
309 if ((ta != tarval_bad) && (tb != tarval_bad)) {
310 return tarval_and (ta, tb);
312 if (tarval_is_null(ta)) return ta;
313 if (tarval_is_null(tb)) return tb;
316 } /* computed_value_And */
319 * Return the value of an Or.
320 * Special case: a | 1...1, 1...1 | b
322 static tarval *computed_value_Or(ir_node *n) {
323 ir_node *a = get_Or_left(n);
324 ir_node *b = get_Or_right(n);
326 tarval *ta = value_of(a);
327 tarval *tb = value_of(b);
329 if ((ta != tarval_bad) && (tb != tarval_bad)) {
330 return tarval_or (ta, tb);
332 if (tarval_is_all_one(ta)) return ta;
333 if (tarval_is_all_one(tb)) return tb;
336 } /* computed_value_Or */
339 * Return the value of an Eor.
341 static tarval *computed_value_Eor(ir_node *n) {
342 ir_node *a = get_Eor_left(n);
343 ir_node *b = get_Eor_right(n);
348 return get_mode_null(get_irn_mode(n));
353 if ((ta != tarval_bad) && (tb != tarval_bad)) {
354 return tarval_eor (ta, tb);
357 } /* computed_value_Eor */
360 * Return the value of a Not.
362 static tarval *computed_value_Not(ir_node *n) {
363 ir_node *a = get_Not_op(n);
364 tarval *ta = value_of(a);
366 if (ta != tarval_bad)
367 return tarval_not(ta);
370 } /* computed_value_Not */
373 * Return the value of a Shl.
375 static tarval *computed_value_Shl(ir_node *n) {
376 ir_node *a = get_Shl_left(n);
377 ir_node *b = get_Shl_right(n);
379 tarval *ta = value_of(a);
380 tarval *tb = value_of(b);
382 if ((ta != tarval_bad) && (tb != tarval_bad)) {
383 return tarval_shl (ta, tb);
386 } /* computed_value_Shl */
389 * Return the value of a Shr.
391 static tarval *computed_value_Shr(ir_node *n) {
392 ir_node *a = get_Shr_left(n);
393 ir_node *b = get_Shr_right(n);
395 tarval *ta = value_of(a);
396 tarval *tb = value_of(b);
398 if ((ta != tarval_bad) && (tb != tarval_bad)) {
399 return tarval_shr (ta, tb);
402 } /* computed_value_Shr */
405 * Return the value of a Shrs.
407 static tarval *computed_value_Shrs(ir_node *n) {
408 ir_node *a = get_Shrs_left(n);
409 ir_node *b = get_Shrs_right(n);
411 tarval *ta = value_of(a);
412 tarval *tb = value_of(b);
414 if ((ta != tarval_bad) && (tb != tarval_bad)) {
415 return tarval_shrs (ta, tb);
418 } /* computed_value_Shrs */
421 * Return the value of a Rot.
423 static tarval *computed_value_Rot(ir_node *n) {
424 ir_node *a = get_Rot_left(n);
425 ir_node *b = get_Rot_right(n);
427 tarval *ta = value_of(a);
428 tarval *tb = value_of(b);
430 if ((ta != tarval_bad) && (tb != tarval_bad)) {
431 return tarval_rot (ta, tb);
434 } /* computed_value_Rot */
437 * Return the value of a Conv.
439 static tarval *computed_value_Conv(ir_node *n) {
440 ir_node *a = get_Conv_op(n);
441 tarval *ta = value_of(a);
443 if (ta != tarval_bad)
444 return tarval_convert_to(ta, get_irn_mode(n));
447 } /* computed_value_Conv */
450 * Return the value of a Proj(Cmp).
452 * This performs a first step of unreachable code elimination.
453 * Proj can not be computed, but folding a Cmp above the Proj here is
454 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
456 * There are several case where we can evaluate a Cmp node, see later.
458 static tarval *computed_value_Proj_Cmp(ir_node *n) {
459 ir_node *a = get_Proj_pred(n);
460 ir_node *aa = get_Cmp_left(a);
461 ir_node *ab = get_Cmp_right(a);
462 long proj_nr = get_Proj_proj(n);
465 * BEWARE: a == a is NOT always True for floating Point values, as
466 * NaN != NaN is defined, so we must check this here.
469 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
472 /* This is a trick with the bits used for encoding the Cmp
473 Proj numbers, the following statement is not the same:
474 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
475 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
478 tarval *taa = value_of(aa);
479 tarval *tab = value_of(ab);
480 ir_mode *mode = get_irn_mode(aa);
483 * The predecessors of Cmp are target values. We can evaluate
486 if ((taa != tarval_bad) && (tab != tarval_bad)) {
487 /* strange checks... */
488 pn_Cmp flags = tarval_cmp(taa, tab);
489 if (flags != pn_Cmp_False) {
490 return new_tarval_from_long (proj_nr & flags, mode_b);
493 /* for integer values, we can check against MIN/MAX */
494 else if (mode_is_int(mode)) {
495 /* MIN <=/> x. This results in true/false. */
496 if (taa == get_mode_min(mode)) {
497 /* a compare with the MIN value */
498 if (proj_nr == pn_Cmp_Le)
499 return get_tarval_b_true();
500 else if (proj_nr == pn_Cmp_Gt)
501 return get_tarval_b_false();
503 /* x >=/< MIN. This results in true/false. */
505 if (tab == get_mode_min(mode)) {
506 /* a compare with the MIN value */
507 if (proj_nr == pn_Cmp_Ge)
508 return get_tarval_b_true();
509 else if (proj_nr == pn_Cmp_Lt)
510 return get_tarval_b_false();
512 /* MAX >=/< x. This results in true/false. */
513 else if (taa == get_mode_max(mode)) {
514 if (proj_nr == pn_Cmp_Ge)
515 return get_tarval_b_true();
516 else if (proj_nr == pn_Cmp_Lt)
517 return get_tarval_b_false();
519 /* x <=/> MAX. This results in true/false. */
520 else if (tab == get_mode_max(mode)) {
521 if (proj_nr == pn_Cmp_Le)
522 return get_tarval_b_true();
523 else if (proj_nr == pn_Cmp_Gt)
524 return get_tarval_b_false();
528 * The predecessors are Allocs or (void*)(0) constants. Allocs never
529 * return NULL, they raise an exception. Therefore we can predict
533 ir_node *aaa = skip_Id(skip_Proj(aa));
534 ir_node *aba = skip_Id(skip_Proj(ab));
536 if ( ( (/* aa is ProjP and aaa is Alloc */
538 && mode_is_reference(get_irn_mode(aa))
540 && ( (/* ab is NULL */
542 && mode_is_reference(get_irn_mode(ab))
543 && is_Const_null(ab))
544 || (/* ab is other Alloc */
546 && mode_is_reference(get_irn_mode(ab))
549 || (/* aa is NULL and aba is Alloc */
551 && mode_is_reference(get_irn_mode(aa))
554 && mode_is_reference(get_irn_mode(ab))
557 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
560 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
561 } /* computed_value_Proj_Cmp */
564 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
565 * Proj(DivMod) and Proj(Quot).
567 static tarval *computed_value_Proj(ir_node *n) {
568 ir_node *a = get_Proj_pred(n);
571 switch (get_irn_opcode(a)) {
573 return computed_value_Proj_Cmp(n);
576 /* compute either the Div or the Mod part */
577 proj_nr = get_Proj_proj(n);
578 if (proj_nr == pn_DivMod_res_div)
579 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
580 else if (proj_nr == pn_DivMod_res_mod)
581 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
585 if (get_Proj_proj(n) == pn_Div_res)
586 return computed_value(a);
590 if (get_Proj_proj(n) == pn_Mod_res)
591 return computed_value(a);
595 if (get_Proj_proj(n) == pn_Quot_res)
596 return computed_value(a);
603 } /* computed_value_Proj */
606 * Calculate the value of a Mux: can be evaluated, if the
607 * sel and the right input are known.
609 static tarval *computed_value_Mux(ir_node *n) {
610 ir_node *sel = get_Mux_sel(n);
611 tarval *ts = value_of(sel);
613 if (ts == get_tarval_b_true()) {
614 ir_node *v = get_Mux_true(n);
617 else if (ts == get_tarval_b_false()) {
618 ir_node *v = get_Mux_false(n);
622 } /* computed_value_Mux */
625 * Calculate the value of a Psi: can be evaluated, if a condition is true
626 * and all previous conditions are false. If all conditions are false
627 * we evaluate to the default one.
629 static tarval *computed_value_Psi(ir_node *n) {
631 return computed_value_Mux(n);
633 } /* computed_value_Psi */
636 * Calculate the value of a Confirm: can be evaluated,
637 * if it has the form Confirm(x, '=', Const).
639 static tarval *computed_value_Confirm(ir_node *n) {
640 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
641 value_of(get_Confirm_bound(n)) : tarval_bad;
642 } /* computed_value_Confirm */
645 * If the parameter n can be computed, return its value, else tarval_bad.
646 * Performs constant folding.
648 * @param n The node this should be evaluated
650 tarval *computed_value(ir_node *n) {
651 if (n->op->ops.computed_value)
652 return n->op->ops.computed_value(n);
654 } /* computed_value */
657 * Set the default computed_value evaluator in an ir_op_ops.
659 * @param code the opcode for the default operation
660 * @param ops the operations initialized
665 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
669 ops->computed_value = computed_value_##a; \
704 } /* firm_set_default_computed_value */
707 * Returns a equivalent block for another block.
708 * If the block has only one predecessor, this is
709 * the equivalent one. If the only predecessor of a block is
710 * the block itself, this is a dead block.
712 * If both predecessors of a block are the branches of a binary
713 * Cond, the equivalent block is Cond's block.
715 * If all predecessors of a block are bad or lies in a dead
716 * block, the current block is dead as well.
718 * Note, that blocks are NEVER turned into Bad's, instead
719 * the dead_block flag is set. So, never test for is_Bad(block),
720 * always use is_dead_Block(block).
722 static ir_node *equivalent_node_Block(ir_node *n)
725 int n_preds = get_Block_n_cfgpreds(n);
727 /* The Block constructor does not call optimize, but mature_immBlock
728 calls the optimization. */
729 assert(get_Block_matured(n));
731 /* Straightening: a single entry Block following a single exit Block
732 can be merged, if it is not the Start block. */
733 /* !!! Beware, all Phi-nodes of n must have been optimized away.
734 This should be true, as the block is matured before optimize is called.
735 But what about Phi-cycles with the Phi0/Id that could not be resolved?
736 Remaining Phi nodes are just Ids. */
737 if (n_preds == 1 && is_Jmp(get_Block_cfgpred(n, 0))) {
738 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
739 if (predblock == oldn) {
740 /* Jmp jumps into the block it is in -- deal self cycle. */
741 n = set_Block_dead(n);
742 DBG_OPT_DEAD_BLOCK(oldn, n);
743 } else if (get_opt_control_flow_straightening()) {
745 DBG_OPT_STG(oldn, n);
747 } else if (n_preds == 1 && is_Cond(skip_Proj(get_Block_cfgpred(n, 0)))) {
748 ir_node *predblock = get_Block_cfgpred_block(n, 0);
749 if (predblock == oldn) {
750 /* Jmp jumps into the block it is in -- deal self cycle. */
751 n = set_Block_dead(n);
752 DBG_OPT_DEAD_BLOCK(oldn, n);
754 } else if ((n_preds == 2) &&
755 (get_opt_control_flow_weak_simplification())) {
756 /* Test whether Cond jumps twice to this block
757 * The more general case which more than 2 predecessors is handles
758 * in optimize_cf(), we handle only this special case for speed here.
760 ir_node *a = get_Block_cfgpred(n, 0);
761 ir_node *b = get_Block_cfgpred(n, 1);
765 (get_Proj_pred(a) == get_Proj_pred(b)) &&
766 is_Cond(get_Proj_pred(a)) &&
767 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
768 /* Also a single entry Block following a single exit Block. Phis have
769 twice the same operand and will be optimized away. */
770 n = get_nodes_block(get_Proj_pred(a));
771 DBG_OPT_IFSIM1(oldn, a, b, n);
773 } else if (get_opt_unreachable_code() &&
774 (n != get_irg_start_block(current_ir_graph)) &&
775 (n != get_irg_end_block(current_ir_graph)) ) {
778 /* If all inputs are dead, this block is dead too, except if it is
779 the start or end block. This is one step of unreachable code
781 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
782 ir_node *pred = get_Block_cfgpred(n, i);
785 if (is_Bad(pred)) continue;
786 pred_blk = get_nodes_block(skip_Proj(pred));
788 if (is_Block_dead(pred_blk)) continue;
791 /* really found a living input */
796 n = set_Block_dead(n);
797 DBG_OPT_DEAD_BLOCK(oldn, n);
802 } /* equivalent_node_Block */
805 * Returns a equivalent node for a Jmp, a Bad :-)
806 * Of course this only happens if the Block of the Jmp is dead.
808 static ir_node *equivalent_node_Jmp(ir_node *n) {
809 /* unreachable code elimination */
810 if (is_Block_dead(get_nodes_block(n)))
814 } /* equivalent_node_Jmp */
816 /** Raise is handled in the same way as Jmp. */
817 #define equivalent_node_Raise equivalent_node_Jmp
820 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
821 See transform_node_Proj_Cond(). */
824 * Optimize operations that are commutative and have neutral 0,
825 * so a op 0 = 0 op a = a.
827 static ir_node *equivalent_node_neutral_zero(ir_node *n)
831 ir_node *a = get_binop_left(n);
832 ir_node *b = get_binop_right(n);
837 /* After running compute_node there is only one constant predecessor.
838 Find this predecessors value and remember the other node: */
839 if ((tv = value_of(a)) != tarval_bad) {
841 } else if ((tv = value_of(b)) != tarval_bad) {
846 /* If this predecessors constant value is zero, the operation is
847 * unnecessary. Remove it.
849 * Beware: If n is a Add, the mode of on and n might be different
850 * which happens in this rare construction: NULL + 3.
851 * Then, a Conv would be needed which we cannot include here.
853 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
856 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
860 } /* equivalent_node_neutral_zero */
863 * Eor is commutative and has neutral 0.
865 #define equivalent_node_Eor equivalent_node_neutral_zero
868 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
870 * The second one looks strange, but this construct
871 * is used heavily in the LCC sources :-).
873 * Beware: The Mode of an Add may be different than the mode of its
874 * predecessors, so we could not return a predecessors in all cases.
876 static ir_node *equivalent_node_Add(ir_node *n) {
878 ir_node *left, *right;
879 ir_mode *mode = get_irn_mode(n);
881 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
882 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
885 n = equivalent_node_neutral_zero(n);
889 left = get_Add_left(n);
890 right = get_Add_right(n);
893 if (get_Sub_right(left) == right) {
896 n = get_Sub_left(left);
897 if (mode == get_irn_mode(n)) {
898 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
904 if (get_Sub_right(right) == left) {
907 n = get_Sub_left(right);
908 if (mode == get_irn_mode(n)) {
909 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
915 } /* equivalent_node_Add */
918 * optimize operations that are not commutative but have neutral 0 on left,
921 static ir_node *equivalent_node_left_zero(ir_node *n) {
924 ir_node *a = get_binop_left(n);
925 ir_node *b = get_binop_right(n);
927 if (is_Const(b) && is_Const_null(b)) {
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
933 } /* equivalent_node_left_zero */
935 #define equivalent_node_Shl equivalent_node_left_zero
936 #define equivalent_node_Shr equivalent_node_left_zero
937 #define equivalent_node_Shrs equivalent_node_left_zero
938 #define equivalent_node_Rot equivalent_node_left_zero
941 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
943 * The second one looks strange, but this construct
944 * is used heavily in the LCC sources :-).
946 * Beware: The Mode of a Sub may be different than the mode of its
947 * predecessors, so we could not return a predecessors in all cases.
949 static ir_node *equivalent_node_Sub(ir_node *n) {
952 ir_mode *mode = get_irn_mode(n);
954 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
955 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
958 b = get_Sub_right(n);
960 /* Beware: modes might be different */
961 if (is_Const(b) && is_Const_null(b)) {
962 ir_node *a = get_Sub_left(n);
963 if (mode == get_irn_mode(a)) {
966 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
970 } /* equivalent_node_Sub */
974 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
977 * -(-a) == a, but might overflow two times.
978 * We handle it anyway here but the better way would be a
979 * flag. This would be needed for Pascal for instance.
981 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
983 ir_node *pred = get_unop_op(n);
985 /* optimize symmetric unop */
986 if (get_irn_op(pred) == get_irn_op(n)) {
987 n = get_unop_op(pred);
988 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
991 } /* equivalent_node_idempotent_unop */
993 /** Optimize Not(Not(x)) == x. */
994 #define equivalent_node_Not equivalent_node_idempotent_unop
996 /** -(-x) == x ??? Is this possible or can --x raise an
997 out of bounds exception if min =! max? */
998 #define equivalent_node_Minus equivalent_node_idempotent_unop
1001 * Optimize a * 1 = 1 * a = a.
1003 static ir_node *equivalent_node_Mul(ir_node *n) {
1005 ir_node *a = get_Mul_left(n);
1007 /* we can handle here only the n * n = n bit cases */
1008 if (get_irn_mode(n) == get_irn_mode(a)) {
1009 ir_node *b = get_Mul_right(n);
1011 /* Mul is commutative and has again an other neutral element. */
1012 if (is_Const(a) && is_Const_one(a)) {
1014 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1015 } else if (is_Const(b) && is_Const_one(b)) {
1017 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1021 } /* equivalent_node_Mul */
1024 * Optimize a / 1 = a.
1026 static ir_node *equivalent_node_Div(ir_node *n) {
1027 ir_node *a = get_Div_left(n);
1028 ir_node *b = get_Div_right(n);
1030 /* Div is not commutative. */
1031 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1032 /* Turn Div into a tuple (mem, bad, a) */
1033 ir_node *mem = get_Div_mem(n);
1034 ir_node *blk = get_irn_n(n, -1);
1035 turn_into_tuple(n, pn_Div_max);
1036 set_Tuple_pred(n, pn_Div_M, mem);
1037 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1038 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1039 set_Tuple_pred(n, pn_Div_res, a);
1042 } /* equivalent_node_Div */
1045 * Optimize a / 1.0 = a.
1047 static ir_node *equivalent_node_Quot(ir_node *n) {
1048 ir_node *a = get_Quot_left(n);
1049 ir_node *b = get_Quot_right(n);
1051 /* Div is not commutative. */
1052 if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */
1053 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1054 ir_node *mem = get_Quot_mem(n);
1055 ir_node *blk = get_irn_n(n, -1);
1056 turn_into_tuple(n, pn_Quot_max);
1057 set_Tuple_pred(n, pn_Quot_M, mem);
1058 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1059 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1060 set_Tuple_pred(n, pn_Quot_res, a);
1063 } /* equivalent_node_Quot */
1066 * Optimize a / 1 = a.
1068 static ir_node *equivalent_node_DivMod(ir_node *n) {
1069 ir_node *b = get_DivMod_right(n);
1071 /* Div is not commutative. */
1072 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1073 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1074 ir_node *a = get_DivMod_left(n);
1075 ir_node *mem = get_Div_mem(n);
1076 ir_node *blk = get_irn_n(n, -1);
1077 ir_mode *mode = get_DivMod_resmode(n);
1079 turn_into_tuple(n, pn_DivMod_max);
1080 set_Tuple_pred(n, pn_DivMod_M, mem);
1081 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1082 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1083 set_Tuple_pred(n, pn_DivMod_res_div, a);
1084 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1087 } /* equivalent_node_DivMod */
1090 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1092 static ir_node *equivalent_node_Or(ir_node *n) {
1095 ir_node *a = get_Or_left(n);
1096 ir_node *b = get_Or_right(n);
1099 n = a; /* Or has it's own neutral element */
1100 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1101 } else if (is_Const(a) && is_Const_null(a)) {
1103 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1104 } else if (is_Const(b) && is_Const_null(b)) {
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1110 } /* equivalent_node_Or */
1113 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1115 static ir_node *equivalent_node_And(ir_node *n) {
1118 ir_node *a = get_And_left(n);
1119 ir_node *b = get_And_right(n);
1122 n = a; /* And has it's own neutral element */
1123 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1126 if (is_Const(a) && is_Const_all_one(a)) {
1128 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1131 if (is_Const(b) && is_Const_all_one(b)) {
1133 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1137 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1140 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1145 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1148 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1154 } /* equivalent_node_And */
1157 * Try to remove useless Conv's:
1159 static ir_node *equivalent_node_Conv(ir_node *n) {
1161 ir_node *a = get_Conv_op(n);
1163 ir_mode *n_mode = get_irn_mode(n);
1164 ir_mode *a_mode = get_irn_mode(a);
1166 if (n_mode == a_mode) { /* No Conv necessary */
1167 if (get_Conv_strict(n)) {
1168 /* special case: the predecessor might be a also a Conv */
1170 if (! get_Conv_strict(a)) {
1171 /* first one is not strict, kick it */
1172 set_Conv_op(n, get_Conv_op(a));
1175 /* else both are strict conv, second is superflous */
1176 } else if(is_Proj(a)) {
1177 ir_node *pred = get_Proj_pred(a);
1179 /* loads always return with the exact precision of n_mode */
1180 assert(get_Load_mode(pred) == n_mode);
1185 /* leave strict floating point Conv's */
1189 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1190 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1191 ir_node *b = get_Conv_op(a);
1192 ir_mode *b_mode = get_irn_mode(b);
1194 if (n_mode == b_mode) {
1195 if (n_mode == mode_b) {
1196 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1197 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1198 } else if (mode_is_int(n_mode)) {
1199 if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) {
1200 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1201 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1207 } /* equivalent_node_Conv */
1210 * A Cast may be removed if the type of the previous node
1211 * is already the type of the Cast.
1213 static ir_node *equivalent_node_Cast(ir_node *n) {
1215 ir_node *pred = get_Cast_op(n);
1217 if (get_irn_type(pred) == get_Cast_type(n)) {
1219 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1222 } /* equivalent_node_Cast */
1225 * Several optimizations:
1226 * - no Phi in start block.
1227 * - remove Id operators that are inputs to Phi
1228 * - fold Phi-nodes, iff they have only one predecessor except
1231 static ir_node *equivalent_node_Phi(ir_node *n) {
1236 ir_node *first_val = NULL; /* to shutup gcc */
1238 if (!get_opt_normalize()) return n;
1240 n_preds = get_Phi_n_preds(n);
1242 block = get_nodes_block(n);
1243 if ((is_Block_dead(block)) || /* Control dead */
1244 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1245 return new_Bad(); /* in the Start Block. */
1247 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1249 /* If the Block has a Bad pred, we also have one. */
1250 for (i = 0; i < n_preds; ++i)
1251 if (is_Bad(get_Block_cfgpred(block, i)))
1252 set_Phi_pred(n, i, new_Bad());
1254 /* Find first non-self-referencing input */
1255 for (i = 0; i < n_preds; ++i) {
1256 first_val = get_Phi_pred(n, i);
1257 if ( (first_val != n) /* not self pointer */
1259 && (! is_Bad(first_val))
1261 ) { /* value not dead */
1262 break; /* then found first value. */
1267 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1271 /* search for rest of inputs, determine if any of these
1272 are non-self-referencing */
1273 while (++i < n_preds) {
1274 ir_node *scnd_val = get_Phi_pred(n, i);
1275 if ( (scnd_val != n)
1276 && (scnd_val != first_val)
1278 && (! is_Bad(scnd_val))
1286 /* Fold, if no multiple distinct non-self-referencing inputs */
1288 DBG_OPT_PHI(oldn, n);
1291 } /* equivalent_node_Phi */
1294 * Several optimizations:
1295 * - no Sync in start block.
1296 * - fold Sync-nodes, iff they have only one predecessor except
1299 static ir_node *equivalent_node_Sync(ir_node *n) {
1303 ir_node *first_val = NULL; /* to shutup gcc */
1305 if (!get_opt_normalize()) return n;
1307 n_preds = get_Sync_n_preds(n);
1309 /* Find first non-self-referencing input */
1310 for (i = 0; i < n_preds; ++i) {
1311 first_val = get_Sync_pred(n, i);
1312 if ((first_val != n) /* not self pointer */ &&
1313 (! is_Bad(first_val))
1314 ) { /* value not dead */
1315 break; /* then found first value. */
1320 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1323 /* search the rest of inputs, determine if any of these
1324 are non-self-referencing */
1325 while (++i < n_preds) {
1326 ir_node *scnd_val = get_Sync_pred(n, i);
1327 if ((scnd_val != n) &&
1328 (scnd_val != first_val) &&
1329 (! is_Bad(scnd_val))
1335 /* Fold, if no multiple distinct non-self-referencing inputs */
1337 DBG_OPT_SYNC(oldn, n);
1340 } /* equivalent_node_Sync */
1343 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1344 * ProjX(Load) and ProjX(Store).
1346 static ir_node *equivalent_node_Proj(ir_node *proj) {
1347 ir_node *oldn = proj;
1348 ir_node *a = get_Proj_pred(proj);
1351 /* Remove the Tuple/Proj combination. */
1352 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1353 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1354 DBG_OPT_TUPLE(oldn, a, proj);
1356 /* This should not happen! */
1357 assert(! "found a Proj with higher number than Tuple predecessors");
1360 } else if (get_irn_mode(proj) == mode_X) {
1361 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1362 /* Remove dead control flow -- early gigo(). */
1364 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1366 /* get the Load address */
1367 ir_node *addr = get_Load_ptr(a);
1368 ir_node *blk = get_irn_n(a, -1);
1371 if (value_not_null(addr, &confirm)) {
1372 if (confirm == NULL) {
1373 /* this node may float if it did not depend on a Confirm */
1374 set_irn_pinned(a, op_pin_state_floats);
1376 if (get_Proj_proj(proj) == pn_Load_X_except) {
1377 DBG_OPT_EXC_REM(proj);
1380 return new_r_Jmp(current_ir_graph, blk);
1382 } else if (is_Store(a)) {
1383 /* get the load/store address */
1384 ir_node *addr = get_Store_ptr(a);
1385 ir_node *blk = get_irn_n(a, -1);
1388 if (value_not_null(addr, &confirm)) {
1389 if (confirm == NULL) {
1390 /* this node may float if it did not depend on a Confirm */
1391 set_irn_pinned(a, op_pin_state_floats);
1393 if (get_Proj_proj(proj) == pn_Store_X_except) {
1394 DBG_OPT_EXC_REM(proj);
1397 return new_r_Jmp(current_ir_graph, blk);
1404 } /* equivalent_node_Proj */
1409 static ir_node *equivalent_node_Id(ir_node *n) {
1414 } while (get_irn_op(n) == op_Id);
1416 DBG_OPT_ID(oldn, n);
1418 } /* equivalent_node_Id */
1423 static ir_node *equivalent_node_Mux(ir_node *n)
1425 ir_node *oldn = n, *sel = get_Mux_sel(n);
1426 tarval *ts = value_of(sel);
1428 /* Mux(true, f, t) == t */
1429 if (ts == tarval_b_true) {
1430 n = get_Mux_true(n);
1431 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1433 /* Mux(false, f, t) == f */
1434 else if (ts == tarval_b_false) {
1435 n = get_Mux_false(n);
1436 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1438 /* Mux(v, x, x) == x */
1439 else if (get_Mux_false(n) == get_Mux_true(n)) {
1440 n = get_Mux_true(n);
1441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1443 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1444 ir_node *cmp = get_Proj_pred(sel);
1445 long proj_nr = get_Proj_proj(sel);
1446 ir_node *f = get_Mux_false(n);
1447 ir_node *t = get_Mux_true(n);
1450 * Note further that these optimization work even for floating point
1451 * with NaN's because -NaN == NaN.
1452 * However, if +0 and -0 is handled differently, we cannot use the first one.
1455 ir_node *const cmp_l = get_Cmp_left(cmp);
1456 ir_node *const cmp_r = get_Cmp_right(cmp);
1460 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1461 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1463 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1470 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1471 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1480 * Note: normalization puts the constant on the right side,
1481 * so we check only one case.
1483 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1484 /* Mux(t CMP 0, X, t) */
1485 if (is_Minus(f) && get_Minus_op(f) == t) {
1486 /* Mux(t CMP 0, -t, t) */
1487 if (proj_nr == pn_Cmp_Eq) {
1488 /* Mux(t == 0, -t, t) ==> -t */
1490 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1491 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1492 /* Mux(t != 0, -t, t) ==> t */
1494 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1501 } /* equivalent_node_Mux */
1504 * Returns a equivalent node of a Psi: if a condition is true
1505 * and all previous conditions are false we know its value.
1506 * If all conditions are false its value is the default one.
1508 static ir_node *equivalent_node_Psi(ir_node *n) {
1510 return equivalent_node_Mux(n);
1512 } /* equivalent_node_Psi */
1515 * Optimize -a CMP -b into b CMP a.
1516 * This works only for for modes where unary Minus
1518 * Note that two-complement integers can Overflow
1519 * so it will NOT work.
1521 * For == and != can be handled in Proj(Cmp)
1523 static ir_node *equivalent_node_Cmp(ir_node *n) {
1524 ir_node *left = get_Cmp_left(n);
1525 ir_node *right = get_Cmp_right(n);
1527 if (is_Minus(left) && is_Minus(right) &&
1528 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1529 left = get_Minus_op(left);
1530 right = get_Minus_op(right);
1531 set_Cmp_left(n, right);
1532 set_Cmp_right(n, left);
1535 } /* equivalent_node_Cmp */
1538 * Remove Confirm nodes if setting is on.
1539 * Replace Confirms(x, '=', Constlike) by Constlike.
1541 static ir_node *equivalent_node_Confirm(ir_node *n) {
1542 ir_node *pred = get_Confirm_value(n);
1543 pn_Cmp pnc = get_Confirm_cmp(n);
1545 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1547 * rare case: two identical Confirms one after another,
1548 * replace the second one with the first.
1552 if (pnc == pn_Cmp_Eq) {
1553 ir_node *bound = get_Confirm_bound(n);
1556 * Optimize a rare case:
1557 * Confirm(x, '=', Constlike) ==> Constlike
1559 if (is_irn_constlike(bound)) {
1560 DBG_OPT_CONFIRM(n, bound);
1564 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1568 * Optimize CopyB(mem, x, x) into a Nop.
1570 static ir_node *equivalent_node_CopyB(ir_node *n) {
1571 ir_node *a = get_CopyB_dst(n);
1572 ir_node *b = get_CopyB_src(n);
1575 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1576 ir_node *mem = get_CopyB_mem(n);
1577 ir_node *blk = get_nodes_block(n);
1578 turn_into_tuple(n, pn_CopyB_max);
1579 set_Tuple_pred(n, pn_CopyB_M, mem);
1580 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1581 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1582 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1585 } /* equivalent_node_CopyB */
1588 * Optimize Bounds(idx, idx, upper) into idx.
1590 static ir_node *equivalent_node_Bound(ir_node *n) {
1591 ir_node *idx = get_Bound_index(n);
1592 ir_node *pred = skip_Proj(idx);
1595 if (is_Bound(pred)) {
1597 * idx was Bounds checked in the same MacroBlock previously,
1598 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1600 ir_node *lower = get_Bound_lower(n);
1601 ir_node *upper = get_Bound_upper(n);
1602 if (get_Bound_lower(pred) == lower &&
1603 get_Bound_upper(pred) == upper &&
1604 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1606 * One could expect that we simply return the previous
1607 * Bound here. However, this would be wrong, as we could
1608 * add an exception Proj to a new location then.
1609 * So, we must turn in into a tuple.
1615 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1616 ir_node *mem = get_Bound_mem(n);
1617 ir_node *blk = get_nodes_block(n);
1618 turn_into_tuple(n, pn_Bound_max);
1619 set_Tuple_pred(n, pn_Bound_M, mem);
1620 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1621 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1622 set_Tuple_pred(n, pn_Bound_res, idx);
1625 } /* equivalent_node_Bound */
1628 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1629 * perform no actual computation, as, e.g., the Id nodes. It does not create
1630 * new nodes. It is therefore safe to free n if the node returned is not n.
1631 * If a node returns a Tuple we can not just skip it. If the size of the
1632 * in array fits, we transform n into a tuple (e.g., Div).
1634 ir_node *equivalent_node(ir_node *n) {
1635 if (n->op->ops.equivalent_node)
1636 return n->op->ops.equivalent_node(n);
1638 } /* equivalent_node */
1641 * Sets the default equivalent node operation for an ir_op_ops.
1643 * @param code the opcode for the default operation
1644 * @param ops the operations initialized
1649 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1653 ops->equivalent_node = equivalent_node_##a; \
1693 } /* firm_set_default_equivalent_node */
1696 * Returns non-zero if a node is a Phi node
1697 * with all predecessors constant.
1699 static int is_const_Phi(ir_node *n) {
1702 if (! is_Phi(n) || get_irn_arity(n) == 0)
1704 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1705 if (! is_Const(get_irn_n(n, i)))
1708 } /* is_const_Phi */
1711 * Apply an evaluator on a binop with a constant operators (and one Phi).
1713 * @param phi the Phi node
1714 * @param other the other operand
1715 * @param eval an evaluator function
1716 * @param mode the mode of the result, may be different from the mode of the Phi!
1717 * @param left if non-zero, other is the left operand, else the right
1719 * @return a new Phi node if the conversion was successful, NULL else
1721 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1726 int i, n = get_irn_arity(phi);
1728 NEW_ARR_A(void *, res, n);
1730 for (i = 0; i < n; ++i) {
1731 pred = get_irn_n(phi, i);
1732 tv = get_Const_tarval(pred);
1733 tv = eval(other, tv);
1735 if (tv == tarval_bad) {
1736 /* folding failed, bad */
1742 for (i = 0; i < n; ++i) {
1743 pred = get_irn_n(phi, i);
1744 tv = get_Const_tarval(pred);
1745 tv = eval(tv, other);
1747 if (tv == tarval_bad) {
1748 /* folding failed, bad */
1754 irg = current_ir_graph;
1755 for (i = 0; i < n; ++i) {
1756 pred = get_irn_n(phi, i);
1757 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1758 mode, res[i], get_Const_type(pred));
1760 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1761 } /* apply_binop_on_phi */
1764 * Apply an evaluator on a binop with two constant Phi.
1766 * @param a the left Phi node
1767 * @param b the right Phi node
1768 * @param eval an evaluator function
1769 * @param mode the mode of the result, may be different from the mode of the Phi!
1771 * @return a new Phi node if the conversion was successful, NULL else
1773 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1774 tarval *tv_l, *tv_r, *tv;
1780 if (get_nodes_block(a) != get_nodes_block(b))
1783 n = get_irn_arity(a);
1784 NEW_ARR_A(void *, res, n);
1786 for (i = 0; i < n; ++i) {
1787 pred = get_irn_n(a, i);
1788 tv_l = get_Const_tarval(pred);
1789 pred = get_irn_n(b, i);
1790 tv_r = get_Const_tarval(pred);
1791 tv = eval(tv_l, tv_r);
1793 if (tv == tarval_bad) {
1794 /* folding failed, bad */
1799 irg = current_ir_graph;
1800 for (i = 0; i < n; ++i) {
1801 pred = get_irn_n(a, i);
1802 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1804 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1805 } /* apply_binop_on_2_phis */
1808 * Apply an evaluator on a unop with a constant operator (a Phi).
1810 * @param phi the Phi node
1811 * @param eval an evaluator function
1813 * @return a new Phi node if the conversion was successful, NULL else
1815 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1821 int i, n = get_irn_arity(phi);
1823 NEW_ARR_A(void *, res, n);
1824 for (i = 0; i < n; ++i) {
1825 pred = get_irn_n(phi, i);
1826 tv = get_Const_tarval(pred);
1829 if (tv == tarval_bad) {
1830 /* folding failed, bad */
1835 mode = get_irn_mode(phi);
1836 irg = current_ir_graph;
1837 for (i = 0; i < n; ++i) {
1838 pred = get_irn_n(phi, i);
1839 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1840 mode, res[i], get_Const_type(pred));
1842 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1843 } /* apply_unop_on_phi */
1846 * Apply a conversion on a constant operator (a Phi).
1848 * @param phi the Phi node
1850 * @return a new Phi node if the conversion was successful, NULL else
1852 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1857 int i, n = get_irn_arity(phi);
1859 NEW_ARR_A(void *, res, n);
1860 for (i = 0; i < n; ++i) {
1861 pred = get_irn_n(phi, i);
1862 tv = get_Const_tarval(pred);
1863 tv = tarval_convert_to(tv, mode);
1865 if (tv == tarval_bad) {
1866 /* folding failed, bad */
1871 irg = current_ir_graph;
1872 for (i = 0; i < n; ++i) {
1873 pred = get_irn_n(phi, i);
1874 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1875 mode, res[i], get_Const_type(pred));
1877 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1878 } /* apply_conv_on_phi */
1881 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1882 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1883 * If possible, remove the Conv's.
1885 static ir_node *transform_node_AddSub(ir_node *n) {
1886 ir_mode *mode = get_irn_mode(n);
1888 if (mode_is_reference(mode)) {
1889 ir_node *left = get_binop_left(n);
1890 ir_node *right = get_binop_right(n);
1891 unsigned ref_bits = get_mode_size_bits(mode);
1893 if (is_Conv(left)) {
1894 ir_mode *lmode = get_irn_mode(left);
1895 unsigned bits = get_mode_size_bits(lmode);
1897 if (ref_bits == bits &&
1898 mode_is_int(lmode) &&
1899 get_mode_arithmetic(lmode) == irma_twos_complement) {
1900 ir_node *pre = get_Conv_op(left);
1901 ir_mode *pre_mode = get_irn_mode(pre);
1903 if (mode_is_int(pre_mode) &&
1904 get_mode_size_bits(pre_mode) == bits &&
1905 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1906 /* ok, this conv just changes to sign, moreover the calculation
1907 * is done with same number of bits as our address mode, so
1908 * we can ignore the conv as address calculation can be viewed
1909 * as either signed or unsigned
1911 set_binop_left(n, pre);
1916 if (is_Conv(right)) {
1917 ir_mode *rmode = get_irn_mode(right);
1918 unsigned bits = get_mode_size_bits(rmode);
1920 if (ref_bits == bits &&
1921 mode_is_int(rmode) &&
1922 get_mode_arithmetic(rmode) == irma_twos_complement) {
1923 ir_node *pre = get_Conv_op(right);
1924 ir_mode *pre_mode = get_irn_mode(pre);
1926 if (mode_is_int(pre_mode) &&
1927 get_mode_size_bits(pre_mode) == bits &&
1928 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1929 /* ok, this conv just changes to sign, moreover the calculation
1930 * is done with same number of bits as our address mode, so
1931 * we can ignore the conv as address calculation can be viewed
1932 * as either signed or unsigned
1934 set_binop_right(n, pre);
1939 /* let address arithmetic use unsigned modes */
1940 if (is_Const(right)) {
1941 ir_mode *rmode = get_irn_mode(right);
1943 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1944 /* convert a AddP(P, *s) into AddP(P, *u) */
1945 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1947 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
1948 set_binop_right(n, pre);
1953 } /* transform_node_AddSub */
1955 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1957 if (is_Const(b) && is_const_Phi(a)) { \
1958 /* check for Op(Phi, Const) */ \
1959 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1961 else if (is_Const(a) && is_const_Phi(b)) { \
1962 /* check for Op(Const, Phi) */ \
1963 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1965 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1966 /* check for Op(Phi, Phi) */ \
1967 c = apply_binop_on_2_phis(a, b, eval, mode); \
1970 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1974 #define HANDLE_UNOP_PHI(eval, a, c) \
1976 if (is_const_Phi(a)) { \
1977 /* check for Op(Phi) */ \
1978 c = apply_unop_on_phi(a, eval); \
1980 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1986 * Do the AddSub optimization, then Transform
1987 * Constant folding on Phi
1988 * Add(a,a) -> Mul(a, 2)
1989 * Add(Mul(a, x), a) -> Mul(a, x+1)
1990 * if the mode is integer or float.
1991 * Transform Add(a,-b) into Sub(a,b).
1992 * Reassociation might fold this further.
1994 static ir_node *transform_node_Add(ir_node *n) {
1996 ir_node *a, *b, *c, *oldn = n;
1998 n = transform_node_AddSub(n);
2000 a = get_Add_left(n);
2001 b = get_Add_right(n);
2003 mode = get_irn_mode(n);
2004 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2006 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2007 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2010 if (mode_is_num(mode)) {
2011 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2012 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2013 ir_node *block = get_irn_n(n, -1);
2016 get_irn_dbg_info(n),
2020 new_r_Const_long(current_ir_graph, block, mode, 2),
2022 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2027 get_irn_dbg_info(n),
2033 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2038 get_irn_dbg_info(n),
2044 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2047 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2048 /* Here we rely on constants be on the RIGHT side */
2050 ir_node *op = get_Not_op(a);
2052 if (is_Const(b) && is_Const_one(b)) {
2054 ir_node *blk = get_irn_n(n, -1);
2055 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2056 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2061 ir_node *blk = get_irn_n(n, -1);
2062 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2063 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2068 ir_node *op = get_Not_op(b);
2072 ir_node *blk = get_irn_n(n, -1);
2073 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2074 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2081 } /* transform_node_Add */
2084 * returns -cnst or NULL if impossible
2086 static ir_node *const_negate(ir_node *cnst) {
2087 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2088 dbg_info *dbgi = get_irn_dbg_info(cnst);
2089 ir_graph *irg = get_irn_irg(cnst);
2090 ir_node *block = get_nodes_block(cnst);
2091 ir_mode *mode = get_irn_mode(cnst);
2092 if (tv == tarval_bad) return NULL;
2093 return new_rd_Const(dbgi, irg, block, mode, tv);
2097 * Do the AddSub optimization, then Transform
2098 * Constant folding on Phi
2099 * Sub(0,a) -> Minus(a)
2100 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2101 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2102 * Sub(Add(a, x), x) -> a
2103 * Sub(x, Add(x, a)) -> -a
2104 * Sub(x, Const) -> Add(x, -Const)
2106 static ir_node *transform_node_Sub(ir_node *n) {
2111 n = transform_node_AddSub(n);
2113 a = get_Sub_left(n);
2114 b = get_Sub_right(n);
2116 mode = get_irn_mode(n);
2119 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2121 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2122 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2125 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2126 /* a - C -> a + (-C) */
2127 ir_node *cnst = const_negate(b);
2129 ir_node *block = get_nodes_block(n);
2130 dbg_info *dbgi = get_irn_dbg_info(n);
2131 ir_graph *irg = get_irn_irg(n);
2133 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2134 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2139 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2140 ir_graph *irg = current_ir_graph;
2141 dbg_info *dbg = get_irn_dbg_info(n);
2142 ir_node *block = get_nodes_block(n);
2143 ir_node *left = get_Minus_op(a);
2144 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2146 n = new_rd_Minus(dbg, irg, block, add, mode);
2147 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2149 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2150 ir_graph *irg = current_ir_graph;
2151 dbg_info *dbg = get_irn_dbg_info(n);
2152 ir_node *block = get_nodes_block(n);
2153 ir_node *right = get_Minus_op(b);
2155 n = new_rd_Add(dbg, irg, block, a, right, mode);
2156 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2158 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2159 ir_graph *irg = current_ir_graph;
2160 dbg_info *s_dbg = get_irn_dbg_info(b);
2161 ir_node *s_block = get_nodes_block(b);
2162 ir_node *s_left = get_Sub_right(b);
2163 ir_node *s_right = get_Sub_left(b);
2164 ir_mode *s_mode = get_irn_mode(b);
2165 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2166 dbg_info *a_dbg = get_irn_dbg_info(n);
2167 ir_node *a_block = get_nodes_block(n);
2169 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2170 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2172 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2173 ir_node *m_right = get_Mul_right(b);
2174 if (is_Const(m_right)) {
2175 ir_node *cnst2 = const_negate(m_right);
2176 if (cnst2 != NULL) {
2177 ir_graph *irg = current_ir_graph;
2178 dbg_info *m_dbg = get_irn_dbg_info(b);
2179 ir_node *m_block = get_nodes_block(b);
2180 ir_node *m_left = get_Mul_left(b);
2181 ir_mode *m_mode = get_irn_mode(b);
2182 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2183 dbg_info *a_dbg = get_irn_dbg_info(n);
2184 ir_node *a_block = get_nodes_block(n);
2186 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2187 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2193 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2194 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2196 get_irn_dbg_info(n),
2201 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2205 if (mode_wrap_around(mode)) {
2206 ir_node *left = get_Add_left(a);
2207 ir_node *right = get_Add_right(a);
2209 /* FIXME: Does the Conv's work only for two complement or generally? */
2211 if (mode != get_irn_mode(right)) {
2212 /* This Sub is an effective Cast */
2213 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2216 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2218 } else if (right == b) {
2219 if (mode != get_irn_mode(left)) {
2220 /* This Sub is an effective Cast */
2221 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2224 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2230 if (mode_wrap_around(mode)) {
2231 ir_node *left = get_Add_left(b);
2232 ir_node *right = get_Add_right(b);
2234 /* FIXME: Does the Conv's work only for two complement or generally? */
2236 ir_mode *r_mode = get_irn_mode(right);
2238 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2239 if (mode != r_mode) {
2240 /* This Sub is an effective Cast */
2241 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2243 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2245 } else if (right == a) {
2246 ir_mode *l_mode = get_irn_mode(left);
2248 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2249 if (mode != l_mode) {
2250 /* This Sub is an effective Cast */
2251 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2253 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2258 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2259 ir_mode *mode = get_irn_mode(a);
2261 if (mode == get_irn_mode(b)) {
2263 ir_node *op_a = get_Conv_op(a);
2264 ir_node *op_b = get_Conv_op(b);
2266 /* check if it's allowed to skip the conv */
2267 ma = get_irn_mode(op_a);
2268 mb = get_irn_mode(op_b);
2270 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2271 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2274 set_Sub_right(n, b);
2280 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2281 if (!is_reassoc_running() && is_Mul(a)) {
2282 ir_node *ma = get_Mul_left(a);
2283 ir_node *mb = get_Mul_right(a);
2286 ir_node *blk = get_irn_n(n, -1);
2288 get_irn_dbg_info(n),
2289 current_ir_graph, blk,
2292 get_irn_dbg_info(n),
2293 current_ir_graph, blk,
2295 new_r_Const_long(current_ir_graph, blk, mode, 1),
2298 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2300 } else if (mb == b) {
2301 ir_node *blk = get_irn_n(n, -1);
2303 get_irn_dbg_info(n),
2304 current_ir_graph, blk,
2307 get_irn_dbg_info(n),
2308 current_ir_graph, blk,
2310 new_r_Const_long(current_ir_graph, blk, mode, 1),
2313 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2317 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2318 ir_node *x = get_Sub_left(a);
2319 ir_node *y = get_Sub_right(a);
2320 ir_node *blk = get_irn_n(n, -1);
2321 ir_mode *m_b = get_irn_mode(b);
2322 ir_mode *m_y = get_irn_mode(y);
2326 /* Determine the right mode for the Add. */
2329 else if (mode_is_reference(m_b))
2331 else if (mode_is_reference(m_y))
2335 * Both modes are different but none is reference,
2336 * happens for instance in SubP(SubP(P, Iu), Is).
2337 * We have two possibilities here: Cast or ignore.
2338 * Currently we ignore this case.
2343 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2345 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2346 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2350 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2351 if (is_Const(a) && is_Not(b)) {
2352 /* c - ~X = X + (c+1) */
2353 tarval *tv = get_Const_tarval(a);
2355 tv = tarval_add(tv, get_mode_one(mode));
2356 if (tv != tarval_bad) {
2357 ir_node *blk = get_irn_n(n, -1);
2358 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2359 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2360 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2366 } /* transform_node_Sub */
2369 * Several transformation done on n*n=2n bits mul.
2370 * These transformations must be done here because new nodes may be produced.
2372 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2374 ir_node *a = get_Mul_left(n);
2375 ir_node *b = get_Mul_right(n);
2376 tarval *ta = value_of(a);
2377 tarval *tb = value_of(b);
2378 ir_mode *smode = get_irn_mode(a);
2380 if (ta == get_mode_one(smode)) {
2381 /* (L)1 * (L)b = (L)b */
2382 ir_node *blk = get_irn_n(n, -1);
2383 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2384 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2387 else if (ta == get_mode_minus_one(smode)) {
2388 /* (L)-1 * (L)b = (L)b */
2389 ir_node *blk = get_irn_n(n, -1);
2390 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2391 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2392 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2395 if (tb == get_mode_one(smode)) {
2396 /* (L)a * (L)1 = (L)a */
2397 ir_node *blk = get_irn_n(a, -1);
2398 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2399 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2402 else if (tb == get_mode_minus_one(smode)) {
2403 /* (L)a * (L)-1 = (L)-a */
2404 ir_node *blk = get_irn_n(n, -1);
2405 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2406 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2407 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2414 * Transform Mul(a,-1) into -a.
2415 * Do constant evaluation of Phi nodes.
2416 * Do architecture dependent optimizations on Mul nodes
2418 static ir_node *transform_node_Mul(ir_node *n) {
2419 ir_node *c, *oldn = n;
2420 ir_mode *mode = get_irn_mode(n);
2421 ir_node *a = get_Mul_left(n);
2422 ir_node *b = get_Mul_right(n);
2424 if (is_Bad(a) || is_Bad(b))
2427 if (mode != get_irn_mode(a))
2428 return transform_node_Mul2n(n, mode);
2430 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2432 if (mode_is_signed(mode)) {
2435 if (value_of(a) == get_mode_minus_one(mode))
2437 else if (value_of(b) == get_mode_minus_one(mode))
2440 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2441 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2446 if (is_Const(b)) { /* (-a) * const -> a * -const */
2447 ir_node *cnst = const_negate(b);
2449 dbg_info *dbgi = get_irn_dbg_info(n);
2450 ir_node *block = get_nodes_block(n);
2451 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2452 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2455 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2456 dbg_info *dbgi = get_irn_dbg_info(n);
2457 ir_node *block = get_nodes_block(n);
2458 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2459 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2461 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2462 ir_node *sub_l = get_Sub_left(b);
2463 ir_node *sub_r = get_Sub_right(b);
2464 dbg_info *dbgi = get_irn_dbg_info(n);
2465 ir_graph *irg = current_ir_graph;
2466 ir_node *block = get_nodes_block(n);
2467 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2468 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2469 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2472 } else if (is_Minus(b)) {
2473 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2474 ir_node *sub_l = get_Sub_left(a);
2475 ir_node *sub_r = get_Sub_right(a);
2476 dbg_info *dbgi = get_irn_dbg_info(n);
2477 ir_graph *irg = current_ir_graph;
2478 ir_node *block = get_nodes_block(n);
2479 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2480 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2481 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2485 if (get_mode_arithmetic(mode) == irma_ieee754) {
2487 tarval *tv = get_Const_tarval(a);
2488 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2489 /* 2.0 * b = b + b */
2490 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2491 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2495 else if (is_Const(b)) {
2496 tarval *tv = get_Const_tarval(b);
2497 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2498 /* a * 2.0 = a + a */
2499 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2500 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2505 return arch_dep_replace_mul_with_shifts(n);
2506 } /* transform_node_Mul */
2509 * Transform a Div Node.
2511 static ir_node *transform_node_Div(ir_node *n) {
2512 ir_mode *mode = get_Div_resmode(n);
2513 ir_node *a = get_Div_left(n);
2514 ir_node *b = get_Div_right(n);
2518 if (is_Const(b) && is_const_Phi(a)) {
2519 /* check for Div(Phi, Const) */
2520 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2522 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2526 else if (is_Const(a) && is_const_Phi(b)) {
2527 /* check for Div(Const, Phi) */
2528 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2530 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2534 else if (is_const_Phi(a) && is_const_Phi(b)) {
2535 /* check for Div(Phi, Phi) */
2536 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2538 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2545 if (tv != tarval_bad) {
2546 value = new_Const(get_tarval_mode(tv), tv);
2548 DBG_OPT_CSTEVAL(n, value);
2551 ir_node *a = get_Div_left(n);
2552 ir_node *b = get_Div_right(n);
2555 if (a == b && value_not_zero(a, &dummy)) {
2556 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2557 value = new_Const(mode, get_mode_one(mode));
2558 DBG_OPT_CSTEVAL(n, value);
2561 if (mode_is_signed(mode) && is_Const(b)) {
2562 tarval *tv = get_Const_tarval(b);
2564 if (tv == get_mode_minus_one(mode)) {
2566 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2567 DBG_OPT_CSTEVAL(n, value);
2571 /* Try architecture dependent optimization */
2572 value = arch_dep_replace_div_by_const(n);
2580 /* Turn Div into a tuple (mem, jmp, bad, value) */
2581 mem = get_Div_mem(n);
2582 blk = get_irn_n(n, -1);
2584 /* skip a potential Pin */
2586 mem = get_Pin_op(mem);
2587 turn_into_tuple(n, pn_Div_max);
2588 set_Tuple_pred(n, pn_Div_M, mem);
2589 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2590 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2591 set_Tuple_pred(n, pn_Div_res, value);
2594 } /* transform_node_Div */
2597 * Transform a Mod node.
2599 static ir_node *transform_node_Mod(ir_node *n) {
2600 ir_mode *mode = get_Mod_resmode(n);
2601 ir_node *a = get_Mod_left(n);
2602 ir_node *b = get_Mod_right(n);
2606 if (is_Const(b) && is_const_Phi(a)) {
2607 /* check for Div(Phi, Const) */
2608 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2610 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2614 else if (is_Const(a) && is_const_Phi(b)) {
2615 /* check for Div(Const, Phi) */
2616 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2618 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2622 else if (is_const_Phi(a) && is_const_Phi(b)) {
2623 /* check for Div(Phi, Phi) */
2624 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2626 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2633 if (tv != tarval_bad) {
2634 value = new_Const(get_tarval_mode(tv), tv);
2636 DBG_OPT_CSTEVAL(n, value);
2639 ir_node *a = get_Mod_left(n);
2640 ir_node *b = get_Mod_right(n);
2643 if (a == b && value_not_zero(a, &dummy)) {
2644 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2645 value = new_Const(mode, get_mode_null(mode));
2646 DBG_OPT_CSTEVAL(n, value);
2649 if (mode_is_signed(mode) && is_Const(b)) {
2650 tarval *tv = get_Const_tarval(b);
2652 if (tv == get_mode_minus_one(mode)) {
2654 value = new_Const(mode, get_mode_null(mode));
2655 DBG_OPT_CSTEVAL(n, value);
2659 /* Try architecture dependent optimization */
2660 value = arch_dep_replace_mod_by_const(n);
2668 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2669 mem = get_Mod_mem(n);
2670 blk = get_irn_n(n, -1);
2672 /* skip a potential Pin */
2674 mem = get_Pin_op(mem);
2675 turn_into_tuple(n, pn_Mod_max);
2676 set_Tuple_pred(n, pn_Mod_M, mem);
2677 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2678 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2679 set_Tuple_pred(n, pn_Mod_res, value);
2682 } /* transform_node_Mod */
2685 * Transform a DivMod node.
2687 static ir_node *transform_node_DivMod(ir_node *n) {
2689 ir_node *a = get_DivMod_left(n);
2690 ir_node *b = get_DivMod_right(n);
2691 ir_mode *mode = get_DivMod_resmode(n);
2696 if (is_Const(b) && is_const_Phi(a)) {
2697 /* check for Div(Phi, Const) */
2698 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2699 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2701 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2702 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2706 else if (is_Const(a) && is_const_Phi(b)) {
2707 /* check for Div(Const, Phi) */
2708 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2709 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2711 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2712 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2716 else if (is_const_Phi(a) && is_const_Phi(b)) {
2717 /* check for Div(Phi, Phi) */
2718 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2719 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2721 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2722 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2729 if (tb != tarval_bad) {
2730 if (tb == get_mode_one(get_tarval_mode(tb))) {
2732 vb = new_Const(mode, get_mode_null(mode));
2733 DBG_OPT_CSTEVAL(n, vb);
2735 } else if (ta != tarval_bad) {
2736 tarval *resa, *resb;
2737 resa = tarval_div(ta, tb);
2738 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2739 Jmp for X result!? */
2740 resb = tarval_mod(ta, tb);
2741 if (resb == tarval_bad) return n; /* Causes exception! */
2742 va = new_Const(mode, resa);
2743 vb = new_Const(mode, resb);
2744 DBG_OPT_CSTEVAL(n, va);
2745 DBG_OPT_CSTEVAL(n, vb);
2747 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2748 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2749 vb = new_Const(mode, get_mode_null(mode));
2750 DBG_OPT_CSTEVAL(n, va);
2751 DBG_OPT_CSTEVAL(n, vb);
2753 } else { /* Try architecture dependent optimization */
2756 arch_dep_replace_divmod_by_const(&va, &vb, n);
2757 evaluated = va != NULL;
2759 } else if (a == b) {
2760 if (value_not_zero(a, &dummy)) {
2762 va = new_Const(mode, get_mode_one(mode));
2763 vb = new_Const(mode, get_mode_null(mode));
2764 DBG_OPT_CSTEVAL(n, va);
2765 DBG_OPT_CSTEVAL(n, vb);
2768 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2771 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2772 /* 0 / non-Const = 0 */
2777 if (evaluated) { /* replace by tuple */
2781 mem = get_DivMod_mem(n);
2782 /* skip a potential Pin */
2784 mem = get_Pin_op(mem);
2786 blk = get_irn_n(n, -1);
2787 turn_into_tuple(n, pn_DivMod_max);
2788 set_Tuple_pred(n, pn_DivMod_M, mem);
2789 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2790 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2791 set_Tuple_pred(n, pn_DivMod_res_div, va);
2792 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2796 } /* transform_node_DivMod */
2799 * Optimize x / c to x * (1/c)
2801 static ir_node *transform_node_Quot(ir_node *n) {
2802 ir_mode *mode = get_Quot_resmode(n);
2805 if (get_mode_arithmetic(mode) == irma_ieee754) {
2806 ir_node *b = get_Quot_right(n);
2809 tarval *tv = get_Const_tarval(b);
2811 tv = tarval_quo(get_mode_one(mode), tv);
2813 /* Do the transformation if the result is either exact or we are not
2814 using strict rules. */
2815 if (tv != tarval_bad &&
2816 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2817 ir_node *blk = get_irn_n(n, -1);
2818 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2819 ir_node *a = get_Quot_left(n);
2820 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2821 ir_node *mem = get_Quot_mem(n);
2823 /* skip a potential Pin */
2825 mem = get_Pin_op(mem);
2826 turn_into_tuple(n, pn_Quot_max);
2827 set_Tuple_pred(n, pn_Quot_M, mem);
2828 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2829 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2830 set_Tuple_pred(n, pn_Quot_res, m);
2831 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2836 } /* transform_node_Quot */
2839 * Optimize Abs(x) into x if x is Confirmed >= 0
2840 * Optimize Abs(x) into -x if x is Confirmed <= 0
2841 * Optimize Abs(-x) int Abs(x)
2843 static ir_node *transform_node_Abs(ir_node *n) {
2844 ir_node *c, *oldn = n;
2845 ir_node *a = get_Abs_op(n);
2848 HANDLE_UNOP_PHI(tarval_abs, a, c);
2850 switch (classify_value_sign(a)) {
2851 case value_classified_negative:
2852 mode = get_irn_mode(n);
2855 * We can replace the Abs by -x here.
2856 * We even could add a new Confirm here
2857 * (if not twos complement)
2859 * Note that -x would create a new node, so we could
2860 * not run it in the equivalent_node() context.
2862 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2863 get_nodes_block(n), a, mode);
2865 DBG_OPT_CONFIRM(oldn, n);
2867 case value_classified_positive:
2868 /* n is positive, Abs is not needed */
2871 DBG_OPT_CONFIRM(oldn, n);
2877 /* Abs(-x) = Abs(x) */
2878 mode = get_irn_mode(n);
2879 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2880 get_nodes_block(n), get_Minus_op(a), mode);
2881 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2885 } /* transform_node_Abs */
2888 * Transform a Cond node.
2890 * Replace the Cond by a Jmp if it branches on a constant
2893 static ir_node *transform_node_Cond(ir_node *n) {
2896 ir_node *a = get_Cond_selector(n);
2897 tarval *ta = value_of(a);
2899 /* we need block info which is not available in floating irgs */
2900 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2903 if ((ta != tarval_bad) &&
2904 (get_irn_mode(a) == mode_b) &&
2905 (get_opt_unreachable_code())) {
2906 /* It's a boolean Cond, branching on a boolean constant.
2907 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2908 ir_node *blk = get_nodes_block(n);
2909 jmp = new_r_Jmp(current_ir_graph, blk);
2910 turn_into_tuple(n, pn_Cond_max);
2911 if (ta == tarval_b_true) {
2912 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2913 set_Tuple_pred(n, pn_Cond_true, jmp);
2915 set_Tuple_pred(n, pn_Cond_false, jmp);
2916 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2918 /* We might generate an endless loop, so keep it alive. */
2919 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2922 } /* transform_node_Cond */
2925 * Prototype of a recursive transform function
2926 * for bitwise distributive transformations.
2928 typedef ir_node* (*recursive_transform)(ir_node *n);
2931 * makes use of distributive laws for and, or, eor
2932 * and(a OP c, b OP c) -> and(a, b) OP c
2933 * note, might return a different op than n
2935 static ir_node *transform_bitwise_distributive(ir_node *n,
2936 recursive_transform trans_func)
2939 ir_node *a = get_binop_left(n);
2940 ir_node *b = get_binop_right(n);
2941 ir_op *op = get_irn_op(a);
2942 ir_op *op_root = get_irn_op(n);
2944 if(op != get_irn_op(b))
2947 if (op == op_Conv) {
2948 ir_node *a_op = get_Conv_op(a);
2949 ir_node *b_op = get_Conv_op(b);
2950 ir_mode *a_mode = get_irn_mode(a_op);
2951 ir_mode *b_mode = get_irn_mode(b_op);
2952 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2953 ir_node *blk = get_irn_n(n, -1);
2956 set_binop_left(n, a_op);
2957 set_binop_right(n, b_op);
2958 set_irn_mode(n, a_mode);
2960 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2962 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2968 /* nothing to gain here */
2972 if (op == op_Shrs || op == op_Shr || op == op_Shl
2973 || op == op_And || op == op_Or || op == op_Eor) {
2974 ir_node *a_left = get_binop_left(a);
2975 ir_node *a_right = get_binop_right(a);
2976 ir_node *b_left = get_binop_left(b);
2977 ir_node *b_right = get_binop_right(b);
2979 ir_node *op1 = NULL;
2980 ir_node *op2 = NULL;
2982 if (is_op_commutative(op)) {
2983 if (a_left == b_left) {
2987 } else if(a_left == b_right) {
2991 } else if(a_right == b_left) {
2997 if(a_right == b_right) {
3004 /* (a sop c) & (b sop c) => (a & b) sop c */
3005 ir_node *blk = get_irn_n(n, -1);
3007 ir_node *new_n = exact_copy(n);
3008 set_binop_left(new_n, op1);
3009 set_binop_right(new_n, op2);
3010 new_n = trans_func(new_n);
3012 if(op_root == op_Eor && op == op_Or) {
3013 dbg_info *dbgi = get_irn_dbg_info(n);
3014 ir_graph *irg = current_ir_graph;
3015 ir_mode *mode = get_irn_mode(c);
3017 c = new_rd_Not(dbgi, irg, blk, c, mode);
3018 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3021 set_nodes_block(n, blk);
3022 set_binop_left(n, new_n);
3023 set_binop_right(n, c);
3024 add_identities(current_ir_graph->value_table, n);
3027 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3038 static ir_node *transform_node_And(ir_node *n) {
3039 ir_node *c, *oldn = n;
3040 ir_node *a = get_And_left(n);
3041 ir_node *b = get_And_right(n);
3044 mode = get_irn_mode(n);
3045 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3047 /* we can evaluate 2 Projs of the same Cmp */
3048 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3049 ir_node *pred_a = get_Proj_pred(a);
3050 ir_node *pred_b = get_Proj_pred(b);
3051 if (pred_a == pred_b) {
3052 dbg_info *dbgi = get_irn_dbg_info(n);
3053 ir_node *block = get_nodes_block(pred_a);
3054 pn_Cmp pn_a = get_Proj_proj(a);
3055 pn_Cmp pn_b = get_Proj_proj(b);
3056 /* yes, we can simply calculate with pncs */
3057 pn_Cmp new_pnc = pn_a & pn_b;
3059 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3064 ir_node *op = get_Not_op(b);
3066 ir_node *ba = get_And_left(op);
3067 ir_node *bb = get_And_right(op);
3069 /* it's enough to test the following cases due to normalization! */
3070 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3071 /* (a|b) & ~(a&b) = a^b */
3072 ir_node *block = get_nodes_block(n);
3074 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3075 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3083 ir_node *op = get_Not_op(a);
3085 ir_node *aa = get_And_left(op);
3086 ir_node *ab = get_And_right(op);
3088 /* it's enough to test the following cases due to normalization! */
3089 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3090 /* (a|b) & ~(a&b) = a^b */
3091 ir_node *block = get_nodes_block(n);
3093 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3094 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3101 ir_node *al = get_Eor_left(a);
3102 ir_node *ar = get_Eor_right(a);
3105 /* (b ^ a) & b -> ~a & b */
3106 dbg_info *dbg = get_irn_dbg_info(n);
3107 ir_node *block = get_nodes_block(n);
3109 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3110 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3111 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3115 /* (a ^ b) & b -> ~a & b */
3116 dbg_info *dbg = get_irn_dbg_info(n);
3117 ir_node *block = get_nodes_block(n);
3119 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3120 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3121 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3126 ir_node *bl = get_Eor_left(b);
3127 ir_node *br = get_Eor_right(b);
3130 /* a & (a ^ b) -> a & ~b */
3131 dbg_info *dbg = get_irn_dbg_info(n);
3132 ir_node *block = get_nodes_block(n);
3134 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3135 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3140 /* a & (b ^ a) -> a & ~b */
3141 dbg_info *dbg = get_irn_dbg_info(n);
3142 ir_node *block = get_nodes_block(n);
3144 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3145 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3150 if (is_Not(a) && is_Not(b)) {
3151 /* ~a & ~b = ~(a|b) */
3152 ir_node *block = get_nodes_block(n);
3153 ir_mode *mode = get_irn_mode(n);
3157 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3158 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3159 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3163 n = transform_bitwise_distributive(n, transform_node_And);
3166 } /* transform_node_And */
3171 static ir_node *transform_node_Eor(ir_node *n) {
3172 ir_node *c, *oldn = n;
3173 ir_node *a = get_Eor_left(n);
3174 ir_node *b = get_Eor_right(n);
3175 ir_mode *mode = get_irn_mode(n);
3177 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3179 /* we can evaluate 2 Projs of the same Cmp */
3180 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3181 ir_node *pred_a = get_Proj_pred(a);
3182 ir_node *pred_b = get_Proj_pred(b);
3183 if(pred_a == pred_b) {
3184 dbg_info *dbgi = get_irn_dbg_info(n);
3185 ir_node *block = get_nodes_block(pred_a);
3186 pn_Cmp pn_a = get_Proj_proj(a);
3187 pn_Cmp pn_b = get_Proj_proj(b);
3188 /* yes, we can simply calculate with pncs */
3189 pn_Cmp new_pnc = pn_a ^ pn_b;
3191 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3198 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3199 mode, get_mode_null(mode));
3200 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3201 } else if (mode == mode_b &&
3203 is_Const(b) && is_Const_one(b) &&
3204 is_Cmp(get_Proj_pred(a))) {
3205 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3206 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3207 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3209 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3210 } else if (is_Const(b)) {
3211 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3212 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3213 ir_node *not_op = get_Not_op(a);
3214 dbg_info *dbg = get_irn_dbg_info(n);
3215 ir_graph *irg = current_ir_graph;
3216 ir_node *block = get_nodes_block(n);
3217 ir_mode *mode = get_irn_mode(n);
3218 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3220 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3221 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3222 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3225 n = transform_bitwise_distributive(n, transform_node_Eor);
3229 } /* transform_node_Eor */
3234 static ir_node *transform_node_Not(ir_node *n) {
3235 ir_node *c, *oldn = n;
3236 ir_node *a = get_Not_op(n);
3237 ir_mode *mode = get_irn_mode(n);
3239 HANDLE_UNOP_PHI(tarval_not,a,c);
3241 /* check for a boolean Not */
3242 if (mode == mode_b &&
3244 is_Cmp(get_Proj_pred(a))) {
3245 /* We negate a Cmp. The Cmp has the negated result anyways! */
3246 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3247 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3248 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3252 ir_node *eor_b = get_Eor_right(a);
3253 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3254 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3255 ir_node *eor_a = get_Eor_left(a);
3256 dbg_info *dbg = get_irn_dbg_info(n);
3257 ir_graph *irg = current_ir_graph;
3258 ir_node *block = get_nodes_block(n);
3259 ir_mode *mode = get_irn_mode(n);
3260 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3264 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3265 if (is_Minus(a)) { /* ~-x -> x + -1 */
3266 dbg_info *dbg = get_irn_dbg_info(n);
3267 ir_graph *irg = current_ir_graph;
3268 ir_node *block = get_nodes_block(n);
3269 ir_node *add_l = get_Minus_op(a);
3270 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3271 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3272 } else if (is_Add(a)) {
3273 ir_node *add_r = get_Add_right(a);
3274 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3275 /* ~(x + -1) = -x */
3276 ir_node *op = get_Add_left(a);
3277 ir_node *blk = get_irn_n(n, -1);
3278 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3279 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3284 } /* transform_node_Not */
3287 * Transform a Minus.
3291 * -(a >>u (size-1)) = a >>s (size-1)
3292 * -(a >>s (size-1)) = a >>u (size-1)
3293 * -(a * const) -> a * -const
3295 static ir_node *transform_node_Minus(ir_node *n) {
3296 ir_node *c, *oldn = n;
3297 ir_node *a = get_Minus_op(n);
3300 HANDLE_UNOP_PHI(tarval_neg,a,c);
3302 mode = get_irn_mode(a);
3303 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3304 /* the following rules are only to twos-complement */
3307 ir_node *op = get_Not_op(a);
3308 tarval *tv = get_mode_one(mode);
3309 ir_node *blk = get_irn_n(n, -1);
3310 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3311 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3312 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3316 ir_node *c = get_Shr_right(a);
3319 tarval *tv = get_Const_tarval(c);
3321 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3322 /* -(a >>u (size-1)) = a >>s (size-1) */
3323 ir_node *v = get_Shr_left(a);
3325 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3326 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3332 ir_node *c = get_Shrs_right(a);
3335 tarval *tv = get_Const_tarval(c);
3337 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3338 /* -(a >>s (size-1)) = a >>u (size-1) */
3339 ir_node *v = get_Shrs_left(a);
3341 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3342 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3349 /* - (a-b) = b - a */
3350 ir_node *la = get_Sub_left(a);
3351 ir_node *ra = get_Sub_right(a);
3352 ir_node *blk = get_irn_n(n, -1);
3354 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3355 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3359 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3360 ir_node *mul_l = get_Mul_left(a);
3361 ir_node *mul_r = get_Mul_right(a);
3362 if (is_Const(mul_r)) {
3363 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3364 if(tv != tarval_bad) {
3365 ir_node *cnst = new_Const(mode, tv);
3366 dbg_info *dbg = get_irn_dbg_info(a);
3367 ir_graph *irg = current_ir_graph;
3368 ir_node *block = get_nodes_block(a);
3369 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3370 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3377 } /* transform_node_Minus */
3380 * Transform a Cast_type(Const) into a new Const_type
3382 static ir_node *transform_node_Cast(ir_node *n) {
3384 ir_node *pred = get_Cast_op(n);
3385 ir_type *tp = get_irn_type(n);
3387 if (is_Const(pred) && get_Const_type(pred) != tp) {
3388 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3389 get_Const_tarval(pred), tp);
3390 DBG_OPT_CSTEVAL(oldn, n);
3391 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3392 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3393 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3394 DBG_OPT_CSTEVAL(oldn, n);
3398 } /* transform_node_Cast */
3401 * Transform a Proj(Div) with a non-zero value.
3402 * Removes the exceptions and routes the memory to the NoMem node.
3404 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3405 ir_node *div = get_Proj_pred(proj);
3406 ir_node *b = get_Div_right(div);
3407 ir_node *confirm, *res, *new_mem;
3410 if (value_not_zero(b, &confirm)) {
3411 /* div(x, y) && y != 0 */
3412 if (confirm == NULL) {
3413 /* we are sure we have a Const != 0 */
3414 new_mem = get_Div_mem(div);
3415 if (is_Pin(new_mem))
3416 new_mem = get_Pin_op(new_mem);
3417 set_Div_mem(div, new_mem);
3418 set_irn_pinned(div, op_pin_state_floats);
3421 proj_nr = get_Proj_proj(proj);
3423 case pn_Div_X_regular:
3424 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3426 case pn_Div_X_except:
3427 /* we found an exception handler, remove it */
3428 DBG_OPT_EXC_REM(proj);
3432 res = get_Div_mem(div);
3433 new_mem = get_irg_no_mem(current_ir_graph);
3436 /* This node can only float up to the Confirm block */
3437 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3439 set_irn_pinned(div, op_pin_state_floats);
3440 /* this is a Div without exception, we can remove the memory edge */
3441 set_Div_mem(div, new_mem);
3446 } /* transform_node_Proj_Div */
3449 * Transform a Proj(Mod) with a non-zero value.
3450 * Removes the exceptions and routes the memory to the NoMem node.
3452 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3453 ir_node *mod = get_Proj_pred(proj);
3454 ir_node *b = get_Mod_right(mod);
3455 ir_node *confirm, *res, *new_mem;
3458 if (value_not_zero(b, &confirm)) {
3459 /* mod(x, y) && y != 0 */
3460 proj_nr = get_Proj_proj(proj);
3462 if (confirm == NULL) {
3463 /* we are sure we have a Const != 0 */
3464 new_mem = get_Mod_mem(mod);
3465 if (is_Pin(new_mem))
3466 new_mem = get_Pin_op(new_mem);
3467 set_Mod_mem(mod, new_mem);
3468 set_irn_pinned(mod, op_pin_state_floats);
3473 case pn_Mod_X_regular:
3474 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3476 case pn_Mod_X_except:
3477 /* we found an exception handler, remove it */
3478 DBG_OPT_EXC_REM(proj);
3482 res = get_Mod_mem(mod);
3483 new_mem = get_irg_no_mem(current_ir_graph);
3486 /* This node can only float up to the Confirm block */
3487 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3489 /* this is a Mod without exception, we can remove the memory edge */
3490 set_Mod_mem(mod, new_mem);
3493 if (get_Mod_left(mod) == b) {
3494 /* a % a = 0 if a != 0 */
3495 ir_mode *mode = get_irn_mode(proj);
3496 ir_node *res = new_Const(mode, get_mode_null(mode));
3498 DBG_OPT_CSTEVAL(mod, res);
3504 } /* transform_node_Proj_Mod */
3507 * Transform a Proj(DivMod) with a non-zero value.
3508 * Removes the exceptions and routes the memory to the NoMem node.
3510 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3511 ir_node *divmod = get_Proj_pred(proj);
3512 ir_node *b = get_DivMod_right(divmod);
3513 ir_node *confirm, *res, *new_mem;
3516 if (value_not_zero(b, &confirm)) {
3517 /* DivMod(x, y) && y != 0 */
3518 proj_nr = get_Proj_proj(proj);
3520 if (confirm == NULL) {
3521 /* we are sure we have a Const != 0 */
3522 new_mem = get_DivMod_mem(divmod);
3523 if (is_Pin(new_mem))
3524 new_mem = get_Pin_op(new_mem);
3525 set_DivMod_mem(divmod, new_mem);
3526 set_irn_pinned(divmod, op_pin_state_floats);
3531 case pn_DivMod_X_regular:
3532 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3534 case pn_DivMod_X_except:
3535 /* we found an exception handler, remove it */
3536 DBG_OPT_EXC_REM(proj);
3540 res = get_DivMod_mem(divmod);
3541 new_mem = get_irg_no_mem(current_ir_graph);
3544 /* This node can only float up to the Confirm block */
3545 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3547 /* this is a DivMod without exception, we can remove the memory edge */
3548 set_DivMod_mem(divmod, new_mem);
3551 case pn_DivMod_res_mod:
3552 if (get_DivMod_left(divmod) == b) {
3553 /* a % a = 0 if a != 0 */
3554 ir_mode *mode = get_irn_mode(proj);
3555 ir_node *res = new_Const(mode, get_mode_null(mode));
3557 DBG_OPT_CSTEVAL(divmod, res);
3563 } /* transform_node_Proj_DivMod */
3566 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3568 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3569 if (get_opt_unreachable_code()) {
3570 ir_node *n = get_Proj_pred(proj);
3571 ir_node *b = get_Cond_selector(n);
3573 if (mode_is_int(get_irn_mode(b))) {
3574 tarval *tb = value_of(b);
3576 if (tb != tarval_bad) {
3577 /* we have a constant switch */
3578 long num = get_Proj_proj(proj);
3580 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3581 if (get_tarval_long(tb) == num) {
3582 /* Do NOT create a jump here, or we will have 2 control flow ops
3583 * in a block. This case is optimized away in optimize_cf(). */
3586 /* this case will NEVER be taken, kill it */
3594 } /* transform_node_Proj_Cond */
3597 * Create a 0 constant of given mode.
3599 static ir_node *create_zero_const(ir_mode *mode) {
3600 tarval *tv = get_mode_null(mode);
3601 ir_node *cnst = new_Const(mode, tv);
3606 /* the order of the values is important! */
3607 typedef enum const_class {
3613 static const_class classify_const(const ir_node* n)
3615 if (is_Const(n)) return const_const;
3616 if (is_irn_constlike(n)) return const_like;
3621 * Determines whether r is more constlike or has a larger index (in that order)
3624 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3626 const const_class l_order = classify_const(l);
3627 const const_class r_order = classify_const(r);
3629 l_order > r_order ||
3630 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3634 * Normalizes and optimizes Cmp nodes.
3636 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3637 ir_node *n = get_Proj_pred(proj);
3638 ir_node *left = get_Cmp_left(n);
3639 ir_node *right = get_Cmp_right(n);
3642 ir_mode *mode = NULL;
3643 long proj_nr = get_Proj_proj(proj);
3645 /* we can evaluate some cases directly */
3648 return new_Const(mode_b, get_tarval_b_false());
3650 return new_Const(mode_b, get_tarval_b_true());
3652 if (!mode_is_float(get_irn_mode(left)))
3653 return new_Const(mode_b, get_tarval_b_true());
3661 left = get_Cast_op(left);
3663 right = get_Cast_op(right);
3665 /* Remove unnecessary conversions */
3666 /* TODO handle constants */
3667 if (is_Conv(left) && is_Conv(right)) {
3668 ir_mode *mode = get_irn_mode(left);
3669 ir_node *op_left = get_Conv_op(left);
3670 ir_node *op_right = get_Conv_op(right);
3671 ir_mode *mode_left = get_irn_mode(op_left);
3672 ir_mode *mode_right = get_irn_mode(op_right);
3674 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3675 && mode_left != mode_b && mode_right != mode_b) {
3676 ir_graph *irg = current_ir_graph;
3677 ir_node *block = get_nodes_block(n);
3679 if (mode_left == mode_right) {
3683 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3684 } else if (smaller_mode(mode_left, mode_right)) {
3685 left = new_r_Conv(irg, block, op_left, mode_right);
3688 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3689 } else if (smaller_mode(mode_right, mode_left)) {
3691 right = new_r_Conv(irg, block, op_right, mode_left);
3693 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3698 /* remove operation of both sides if possible */
3699 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3701 * The following operations are NOT safe for floating point operations, for instance
3702 * 1.0 + inf == 2.0 + inf, =/=> x == y
3704 if (mode_is_int(get_irn_mode(left))) {
3705 unsigned lop = get_irn_opcode(left);
3707 if (lop == get_irn_opcode(right)) {
3708 ir_node *ll, *lr, *rl, *rr;
3710 /* same operation on both sides, try to remove */
3714 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3715 left = get_unop_op(left);
3716 right = get_unop_op(right);
3718 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3721 ll = get_Add_left(left);
3722 lr = get_Add_right(left);
3723 rl = get_Add_left(right);
3724 rr = get_Add_right(right);
3727 /* X + a CMP X + b ==> a CMP b */
3731 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3732 } else if (ll == rr) {
3733 /* X + a CMP b + X ==> a CMP b */
3737 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3738 } else if (lr == rl) {
3739 /* a + X CMP X + b ==> a CMP b */
3743 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3744 } else if (lr == rr) {
3745 /* a + X CMP b + X ==> a CMP b */
3749 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3753 ll = get_Sub_left(left);
3754 lr = get_Sub_right(left);
3755 rl = get_Sub_left(right);
3756 rr = get_Sub_right(right);
3759 /* X - a CMP X - b ==> a CMP b */
3763 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3764 } else if (lr == rr) {
3765 /* a - X CMP b - X ==> a CMP b */
3769 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3773 if (get_Rot_right(left) == get_Rot_right(right)) {
3774 /* a ROT X CMP b ROT X ==> a CMP b */
3775 left = get_Rot_left(left);
3776 right = get_Rot_left(right);
3778 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3786 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3787 if (is_Add(left) || is_Sub(left)) {
3788 ir_node *ll = get_binop_left(left);
3789 ir_node *lr = get_binop_right(left);
3791 if (lr == right && is_Add(left)) {
3798 right = create_zero_const(get_irn_mode(left));
3800 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3803 if (is_Add(right) || is_Sub(right)) {
3804 ir_node *rl = get_binop_left(right);
3805 ir_node *rr = get_binop_right(right);
3807 if (rr == left && is_Add(right)) {
3814 right = create_zero_const(get_irn_mode(left));
3816 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3819 } /* mode_is_int(...) */
3820 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3822 /* replace mode_b compares with ands/ors */
3823 if (get_irn_mode(left) == mode_b) {
3824 ir_graph *irg = current_ir_graph;
3825 ir_node *block = get_nodes_block(n);
3829 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3830 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3831 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3832 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3833 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3834 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3835 default: bres = NULL;
3838 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3844 * First step: normalize the compare op
3845 * by placing the constant on the right side
3846 * or moving the lower address node to the left.
3848 if (!operands_are_normalized(left, right)) {
3854 proj_nr = get_inversed_pnc(proj_nr);
3859 * Second step: Try to reduce the magnitude
3860 * of a constant. This may help to generate better code
3861 * later and may help to normalize more compares.
3862 * Of course this is only possible for integer values.
3864 if (is_Const(right)) {
3865 mode = get_irn_mode(right);
3866 tv = get_Const_tarval(right);
3868 /* TODO extend to arbitrary constants */
3869 if (is_Conv(left) && tarval_is_null(tv)) {
3870 ir_node *op = get_Conv_op(left);
3871 ir_mode *op_mode = get_irn_mode(op);
3874 * UpConv(x) REL 0 ==> x REL 0
3876 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3877 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3878 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3879 tv = get_mode_null(op_mode);
3883 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3887 if (tv != tarval_bad) {
3888 /* the following optimization is possible on modes without Overflow
3889 * on Unary Minus or on == and !=:
3890 * -a CMP c ==> a swap(CMP) -c
3892 * Beware: for two-complement Overflow may occur, so only == and != can
3893 * be optimized, see this:
3894 * -MININT < 0 =/=> MININT > 0 !!!
3896 if (is_Minus(left) &&
3897 (!mode_overflow_on_unary_Minus(mode) ||
3898 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3899 tv = tarval_neg(tv);
3901 if (tv != tarval_bad) {
3902 left = get_Minus_op(left);
3903 proj_nr = get_inversed_pnc(proj_nr);
3905 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3907 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3908 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3909 tv = tarval_not(tv);
3911 if (tv != tarval_bad) {
3912 left = get_Not_op(left);
3914 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3918 /* for integer modes, we have more */
3919 if (mode_is_int(mode)) {
3920 /* Ne includes Unordered which is not possible on integers.
3921 * However, frontends often use this wrong, so fix it here */
3922 if (proj_nr & pn_Cmp_Uo) {
3923 proj_nr &= ~pn_Cmp_Uo;
3924 set_Proj_proj(proj, proj_nr);
3927 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3928 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3929 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3930 tv = tarval_sub(tv, get_mode_one(mode));
3932 if (tv != tarval_bad) {
3933 proj_nr ^= pn_Cmp_Eq;
3935 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3938 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3939 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3940 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3941 tv = tarval_add(tv, get_mode_one(mode));
3943 if (tv != tarval_bad) {
3944 proj_nr ^= pn_Cmp_Eq;
3946 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3950 /* the following reassociations work only for == and != */
3951 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3953 #if 0 /* Might be not that good in general */
3954 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
3955 if (tarval_is_null(tv) && is_Sub(left)) {
3956 right = get_Sub_right(left);
3957 left = get_Sub_left(left);
3959 tv = value_of(right);
3961 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3965 if (tv != tarval_bad) {
3966 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
3968 ir_node *c1 = get_Sub_right(left);
3969 tarval *tv2 = value_of(c1);
3971 if (tv2 != tarval_bad) {
3972 tv2 = tarval_add(tv, value_of(c1));
3974 if (tv2 != tarval_bad) {
3975 left = get_Sub_left(left);
3978 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3982 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
3983 else if (is_Add(left)) {
3984 ir_node *a_l = get_Add_left(left);
3985 ir_node *a_r = get_Add_right(left);
3989 if (is_Const(a_l)) {
3991 tv2 = value_of(a_l);
3994 tv2 = value_of(a_r);
3997 if (tv2 != tarval_bad) {
3998 tv2 = tarval_sub(tv, tv2);
4000 if (tv2 != tarval_bad) {
4004 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4008 /* -a == c ==> a == -c, -a != c ==> a != -c */
4009 else if (is_Minus(left)) {
4010 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4012 if (tv2 != tarval_bad) {
4013 left = get_Minus_op(left);
4016 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4021 /* the following reassociations work only for <= */
4022 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4023 if (tv != tarval_bad) {
4024 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4025 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4031 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4032 switch (get_irn_opcode(left)) {
4036 c1 = get_And_right(left);
4039 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4040 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4042 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4044 /* TODO: move to constant evaluation */
4045 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4046 c1 = new_Const(mode_b, tv);
4047 DBG_OPT_CSTEVAL(proj, c1);
4051 if (tarval_is_single_bit(tv)) {
4053 * optimization for AND:
4055 * And(x, C) == C ==> And(x, C) != 0
4056 * And(x, C) != C ==> And(X, C) == 0
4058 * if C is a single Bit constant.
4061 /* check for Constant's match. We have check hare the tarvals,
4062 because our const might be changed */
4063 if (get_Const_tarval(c1) == tv) {
4064 /* fine: do the transformation */
4065 tv = get_mode_null(get_tarval_mode(tv));
4066 proj_nr ^= pn_Cmp_Leg;
4068 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4074 c1 = get_Or_right(left);
4075 if (is_Const(c1) && tarval_is_null(tv)) {
4077 * Or(x, C) == 0 && C != 0 ==> FALSE
4078 * Or(x, C) != 0 && C != 0 ==> TRUE
4080 if (! tarval_is_null(get_Const_tarval(c1))) {
4081 /* TODO: move to constant evaluation */
4082 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4083 c1 = new_Const(mode_b, tv);
4084 DBG_OPT_CSTEVAL(proj, c1);
4091 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4093 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4096 c1 = get_Shl_right(left);
4098 tarval *tv1 = get_Const_tarval(c1);
4099 ir_mode *mode = get_irn_mode(left);
4100 tarval *minus1 = get_mode_all_one(mode);
4101 tarval *amask = tarval_shr(minus1, tv1);
4102 tarval *cmask = tarval_shl(minus1, tv1);
4105 if (tarval_and(tv, cmask) != tv) {
4106 /* condition not met */
4107 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4108 c1 = new_Const(mode_b, tv);
4109 DBG_OPT_CSTEVAL(proj, c1);
4112 sl = get_Shl_left(left);
4113 blk = get_nodes_block(n);
4114 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4115 tv = tarval_shr(tv, tv1);
4117 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4122 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4124 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4127 c1 = get_Shr_right(left);
4129 tarval *tv1 = get_Const_tarval(c1);
4130 ir_mode *mode = get_irn_mode(left);
4131 tarval *minus1 = get_mode_all_one(mode);
4132 tarval *amask = tarval_shl(minus1, tv1);
4133 tarval *cmask = tarval_shr(minus1, tv1);
4136 if (tarval_and(tv, cmask) != tv) {
4137 /* condition not met */
4138 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4139 c1 = new_Const(mode_b, tv);
4140 DBG_OPT_CSTEVAL(proj, c1);
4143 sl = get_Shr_left(left);
4144 blk = get_nodes_block(n);
4145 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4146 tv = tarval_shl(tv, tv1);
4148 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4153 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4155 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4158 c1 = get_Shrs_right(left);
4160 tarval *tv1 = get_Const_tarval(c1);
4161 ir_mode *mode = get_irn_mode(left);
4162 tarval *minus1 = get_mode_all_one(mode);
4163 tarval *amask = tarval_shl(minus1, tv1);
4164 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4167 cond = tarval_sub(cond, tv1);
4168 cond = tarval_shrs(tv, cond);
4170 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4171 /* condition not met */
4172 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4173 c1 = new_Const(mode_b, tv);
4174 DBG_OPT_CSTEVAL(proj, c1);
4177 sl = get_Shrs_left(left);
4178 blk = get_nodes_block(n);
4179 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4180 tv = tarval_shl(tv, tv1);
4182 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4187 } /* tarval != bad */
4190 if (changed & 2) /* need a new Const */
4191 right = new_Const(mode, tv);
4193 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4194 ir_node *op = get_Proj_pred(left);
4196 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4197 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4198 ir_node *c = get_binop_right(op);
4201 tarval *tv = get_Const_tarval(c);
4203 if (tarval_is_single_bit(tv)) {
4204 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4205 ir_node *v = get_binop_left(op);
4206 ir_node *blk = get_irn_n(op, -1);
4207 ir_mode *mode = get_irn_mode(v);
4209 tv = tarval_sub(tv, get_mode_one(mode));
4210 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4212 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4219 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4221 /* create a new compare */
4222 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4223 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4227 } /* transform_node_Proj_Cmp */
4230 * Does all optimizations on nodes that must be done on it's Proj's
4231 * because of creating new nodes.
4233 static ir_node *transform_node_Proj(ir_node *proj) {
4234 ir_node *n = get_Proj_pred(proj);
4236 switch (get_irn_opcode(n)) {
4238 return transform_node_Proj_Div(proj);
4241 return transform_node_Proj_Mod(proj);
4244 return transform_node_Proj_DivMod(proj);
4247 return transform_node_Proj_Cond(proj);
4250 return transform_node_Proj_Cmp(proj);
4253 /* should not happen, but if it does will be optimized away */
4254 return equivalent_node_Proj(proj);
4260 } /* transform_node_Proj */
4263 * Move Confirms down through Phi nodes.
4265 static ir_node *transform_node_Phi(ir_node *phi) {
4267 ir_mode *mode = get_irn_mode(phi);
4269 if (mode_is_reference(mode)) {
4270 n = get_irn_arity(phi);
4272 /* Beware of Phi0 */
4274 ir_node *pred = get_irn_n(phi, 0);
4275 ir_node *bound, *new_Phi, *block, **in;
4278 if (! is_Confirm(pred))
4281 bound = get_Confirm_bound(pred);
4282 pnc = get_Confirm_cmp(pred);
4284 NEW_ARR_A(ir_node *, in, n);
4285 in[0] = get_Confirm_value(pred);
4287 for (i = 1; i < n; ++i) {
4288 pred = get_irn_n(phi, i);
4290 if (! is_Confirm(pred) ||
4291 get_Confirm_bound(pred) != bound ||
4292 get_Confirm_cmp(pred) != pnc)
4294 in[i] = get_Confirm_value(pred);
4296 /* move the Confirm nodes "behind" the Phi */
4297 block = get_irn_n(phi, -1);
4298 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4299 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4303 } /* transform_node_Phi */
4306 * Returns the operands of a commutative bin-op, if one operand is
4307 * a const, it is returned as the second one.
4309 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4310 ir_node *op_a = get_binop_left(binop);
4311 ir_node *op_b = get_binop_right(binop);
4313 assert(is_op_commutative(get_irn_op(binop)));
4315 if (is_Const(op_a)) {
4322 } /* get_comm_Binop_Ops */
4325 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4326 * Such pattern may arise in bitfield stores.
4328 * value c4 value c4 & c2
4329 * AND c3 AND c1 | c3
4336 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4339 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4342 ir_node *and_l, *c3;
4343 ir_node *value, *c4;
4344 ir_node *new_and, *new_const, *block;
4345 ir_mode *mode = get_irn_mode(or);
4347 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4350 get_comm_Binop_Ops(or, &and, &c1);
4351 if (!is_Const(c1) || !is_And(and))
4354 get_comm_Binop_Ops(and, &or_l, &c2);
4358 tv1 = get_Const_tarval(c1);
4359 tv2 = get_Const_tarval(c2);
4361 tv = tarval_or(tv1, tv2);
4362 if (tarval_is_all_one(tv)) {
4363 /* the AND does NOT clear a bit with isn't set by the OR */
4364 set_Or_left(or, or_l);
4365 set_Or_right(or, c1);
4367 /* check for more */
4374 get_comm_Binop_Ops(or_l, &and_l, &c3);
4375 if (!is_Const(c3) || !is_And(and_l))
4378 get_comm_Binop_Ops(and_l, &value, &c4);
4382 /* ok, found the pattern, check for conditions */
4383 assert(mode == get_irn_mode(and));
4384 assert(mode == get_irn_mode(or_l));
4385 assert(mode == get_irn_mode(and_l));
4387 tv3 = get_Const_tarval(c3);
4388 tv4 = get_Const_tarval(c4);
4390 tv = tarval_or(tv4, tv2);
4391 if (!tarval_is_all_one(tv)) {
4392 /* have at least one 0 at the same bit position */
4396 n_tv4 = tarval_not(tv4);
4397 if (tv3 != tarval_and(tv3, n_tv4)) {
4398 /* bit in the or_mask is outside the and_mask */
4402 n_tv2 = tarval_not(tv2);
4403 if (tv1 != tarval_and(tv1, n_tv2)) {
4404 /* bit in the or_mask is outside the and_mask */
4408 /* ok, all conditions met */
4409 block = get_irn_n(or, -1);
4411 new_and = new_r_And(current_ir_graph, block,
4412 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4414 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4416 set_Or_left(or, new_and);
4417 set_Or_right(or, new_const);
4419 /* check for more */
4421 } /* transform_node_Or_bf_store */
4424 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4426 static ir_node *transform_node_Or_Rot(ir_node *or) {
4427 ir_mode *mode = get_irn_mode(or);
4428 ir_node *shl, *shr, *block;
4429 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4432 if (! mode_is_int(mode))
4435 shl = get_binop_left(or);
4436 shr = get_binop_right(or);
4445 } else if (!is_Shl(shl)) {
4447 } else if (!is_Shr(shr)) {
4450 x = get_Shl_left(shl);
4451 if (x != get_Shr_left(shr))
4454 c1 = get_Shl_right(shl);
4455 c2 = get_Shr_right(shr);
4456 if (is_Const(c1) && is_Const(c2)) {
4457 tv1 = get_Const_tarval(c1);
4458 if (! tarval_is_long(tv1))
4461 tv2 = get_Const_tarval(c2);
4462 if (! tarval_is_long(tv2))
4465 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4466 != (int) get_mode_size_bits(mode))
4469 /* yet, condition met */
4470 block = get_irn_n(or, -1);
4472 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4474 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4476 } else if (is_Sub(c1)) {
4480 if (get_Sub_right(sub) != v)
4483 c1 = get_Sub_left(sub);
4487 tv1 = get_Const_tarval(c1);
4488 if (! tarval_is_long(tv1))
4491 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4494 /* yet, condition met */
4495 block = get_nodes_block(or);
4497 /* a Rot right is not supported, so use a rot left */
4498 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4500 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4502 } else if (is_Sub(c2)) {
4506 c1 = get_Sub_left(sub);
4510 tv1 = get_Const_tarval(c1);
4511 if (! tarval_is_long(tv1))
4514 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4517 /* yet, condition met */
4518 block = get_irn_n(or, -1);
4521 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4523 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4528 } /* transform_node_Or_Rot */
4533 static ir_node *transform_node_Or(ir_node *n) {
4534 ir_node *c, *oldn = n;
4535 ir_node *a = get_Or_left(n);
4536 ir_node *b = get_Or_right(n);
4539 if (is_Not(a) && is_Not(b)) {
4540 /* ~a | ~b = ~(a&b) */
4541 ir_node *block = get_nodes_block(n);
4543 mode = get_irn_mode(n);
4546 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4547 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4548 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4552 /* we can evaluate 2 Projs of the same Cmp */
4553 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4554 ir_node *pred_a = get_Proj_pred(a);
4555 ir_node *pred_b = get_Proj_pred(b);
4556 if (pred_a == pred_b) {
4557 dbg_info *dbgi = get_irn_dbg_info(n);
4558 ir_node *block = get_nodes_block(pred_a);
4559 pn_Cmp pn_a = get_Proj_proj(a);
4560 pn_Cmp pn_b = get_Proj_proj(b);
4561 /* yes, we can simply calculate with pncs */
4562 pn_Cmp new_pnc = pn_a | pn_b;
4564 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4569 mode = get_irn_mode(n);
4570 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4572 n = transform_node_Or_bf_store(n);
4573 n = transform_node_Or_Rot(n);
4577 n = transform_bitwise_distributive(n, transform_node_Or);
4580 } /* transform_node_Or */
4584 static ir_node *transform_node(ir_node *n);
4587 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4589 * Should be moved to reassociation?
4591 static ir_node *transform_node_shift(ir_node *n) {
4592 ir_node *left, *right;
4593 tarval *tv1, *tv2, *res;
4595 int modulo_shf, flag;
4597 left = get_binop_left(n);
4599 /* different operations */
4600 if (get_irn_op(left) != get_irn_op(n))
4603 right = get_binop_right(n);
4604 tv1 = value_of(right);
4605 if (tv1 == tarval_bad)
4608 tv2 = value_of(get_binop_right(left));
4609 if (tv2 == tarval_bad)
4612 res = tarval_add(tv1, tv2);
4614 /* beware: a simple replacement works only, if res < modulo shift */
4615 mode = get_irn_mode(n);
4619 modulo_shf = get_mode_modulo_shift(mode);
4620 if (modulo_shf > 0) {
4621 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4623 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4629 /* ok, we can replace it */
4630 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4632 in[0] = get_binop_left(left);
4633 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4635 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4637 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4639 return transform_node(irn);
4642 } /* transform_node_shift */
4647 static ir_node *transform_node_Shr(ir_node *n) {
4648 ir_node *c, *oldn = n;
4649 ir_node *a = get_Shr_left(n);
4650 ir_node *b = get_Shr_right(n);
4651 ir_mode *mode = get_irn_mode(n);
4653 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4654 return transform_node_shift(n);
4655 } /* transform_node_Shr */
4660 static ir_node *transform_node_Shrs(ir_node *n) {
4661 ir_node *c, *oldn = n;
4662 ir_node *a = get_Shrs_left(n);
4663 ir_node *b = get_Shrs_right(n);
4664 ir_mode *mode = get_irn_mode(n);
4666 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4667 return transform_node_shift(n);
4668 } /* transform_node_Shrs */
4673 static ir_node *transform_node_Shl(ir_node *n) {
4674 ir_node *c, *oldn = n;
4675 ir_node *a = get_Shl_left(n);
4676 ir_node *b = get_Shl_right(n);
4677 ir_mode *mode = get_irn_mode(n);
4679 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4680 return transform_node_shift(n);
4681 } /* transform_node_Shl */
4686 static ir_node *transform_node_Rot(ir_node *n) {
4687 ir_node *c, *oldn = n;
4688 ir_node *a = get_Rot_left(n);
4689 ir_node *b = get_Rot_right(n);
4690 ir_mode *mode = get_irn_mode(n);
4692 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4693 return transform_node_shift(n);
4694 } /* transform_node_Rot */
4699 static ir_node *transform_node_Conv(ir_node *n) {
4700 ir_node *c, *oldn = n;
4701 ir_node *a = get_Conv_op(n);
4703 if (is_const_Phi(a)) {
4704 c = apply_conv_on_phi(a, get_irn_mode(n));
4706 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4711 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4712 ir_mode *mode = get_irn_mode(n);
4713 return new_r_Unknown(current_ir_graph, mode);
4717 } /* transform_node_Conv */
4720 * Remove dead blocks and nodes in dead blocks
4721 * in keep alive list. We do not generate a new End node.
4723 static ir_node *transform_node_End(ir_node *n) {
4724 int i, j, n_keepalives = get_End_n_keepalives(n);
4727 NEW_ARR_A(ir_node *, in, n_keepalives);
4729 for (i = j = 0; i < n_keepalives; ++i) {
4730 ir_node *ka = get_End_keepalive(n, i);
4732 if (! is_Block_dead(ka)) {
4736 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4739 /* FIXME: beabi need to keep a Proj(M) */
4740 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4743 if (j != n_keepalives)
4744 set_End_keepalives(n, j, in);
4746 } /* transform_node_End */
4748 /** returns 1 if a == -b */
4749 static int is_negated_value(ir_node *a, ir_node *b) {
4750 if(is_Minus(a) && get_Minus_op(a) == b)
4752 if(is_Minus(b) && get_Minus_op(b) == a)
4754 if(is_Sub(a) && is_Sub(b)) {
4755 ir_node *a_left = get_Sub_left(a);
4756 ir_node *a_right = get_Sub_right(a);
4757 ir_node *b_left = get_Sub_left(b);
4758 ir_node *b_right = get_Sub_right(b);
4760 if(a_left == b_right && a_right == b_left)
4768 * Optimize a Mux into some simpler cases.
4770 static ir_node *transform_node_Mux(ir_node *n) {
4771 ir_node *oldn = n, *sel = get_Mux_sel(n);
4772 ir_mode *mode = get_irn_mode(n);
4774 if (mode == mode_b) {
4775 ir_node *t = get_Mux_true(n);
4776 ir_node *f = get_Mux_false(n);
4777 dbg_info *dbg = get_irn_dbg_info(n);
4778 ir_node *block = get_irn_n(n, -1);
4779 ir_graph *irg = current_ir_graph;
4782 tarval *tv_t = get_Const_tarval(t);
4783 if (tv_t == tarval_b_true) {
4785 assert(get_Const_tarval(f) == tarval_b_false);
4788 return new_rd_Or(dbg, irg, block, sel, f, mode_b);
4791 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4792 assert(tv_t == tarval_b_false);
4794 assert(get_Const_tarval(f) == tarval_b_true);
4797 return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
4800 } else if (is_Const(f)) {
4801 tarval *tv_f = get_Const_tarval(f);
4802 if (tv_f == tarval_b_true) {
4803 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4804 return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4806 assert(tv_f == tarval_b_false);
4807 return new_rd_And(dbg, irg, block, sel, t, mode_b);
4812 if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
4813 ir_node *cmp = get_Proj_pred(sel);
4814 long pn = get_Proj_proj(sel);
4815 ir_node *f = get_Mux_false(n);
4816 ir_node *t = get_Mux_true(n);
4819 * Note: normalization puts the constant on the right side,
4820 * so we check only one case.
4822 * Note further that these optimization work even for floating point
4823 * with NaN's because -NaN == NaN.
4824 * However, if +0 and -0 is handled differently, we cannot use the first
4828 ir_node *cmp_r = get_Cmp_right(cmp);
4829 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4830 ir_node *block = get_irn_n(n, -1);
4832 if(is_negated_value(f, t)) {
4833 ir_node *cmp_left = get_Cmp_left(cmp);
4835 /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
4836 if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4837 || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4839 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4841 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4843 /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
4844 } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4845 || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4847 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4849 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4851 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4858 return arch_transform_node_Mux(n);
4859 } /* transform_node_Mux */
4862 * Optimize a Psi into some simpler cases.
4864 static ir_node *transform_node_Psi(ir_node *n) {
4866 return transform_node_Mux(n);
4869 } /* transform_node_Psi */
4872 * optimize sync nodes that have other syncs as input we simply add the inputs
4873 * of the other sync to our own inputs
4875 static ir_node *transform_node_Sync(ir_node *n) {
4878 arity = get_irn_arity(n);
4879 for(i = 0; i < get_irn_arity(n); /*empty*/) {
4881 ir_node *in = get_irn_n(n, i);
4887 /* set sync input 0 instead of the sync */
4888 set_irn_n(n, i, get_irn_n(in, 0));
4889 /* so we check this input again for syncs */
4891 /* append all other inputs of the sync to our sync */
4892 arity2 = get_irn_arity(in);
4893 for(i2 = 1; i2 < arity2; ++i2) {
4894 ir_node *in_in = get_irn_n(in, i2);
4895 add_irn_n(n, in_in);
4896 /* increase arity so we also check the new inputs for syncs */
4901 /* rehash the sync node */
4902 add_identities(current_ir_graph->value_table, n);
4908 * Tries several [inplace] [optimizing] transformations and returns an
4909 * equivalent node. The difference to equivalent_node() is that these
4910 * transformations _do_ generate new nodes, and thus the old node must
4911 * not be freed even if the equivalent node isn't the old one.
4913 static ir_node *transform_node(ir_node *n) {
4917 * Transform_node is the only "optimizing transformation" that might
4918 * return a node with a different opcode. We iterate HERE until fixpoint
4919 * to get the final result.
4923 if (n->op->ops.transform_node)
4924 n = n->op->ops.transform_node(n);
4925 } while (oldn != n);
4928 } /* transform_node */
4931 * Sets the default transform node operation for an ir_op_ops.
4933 * @param code the opcode for the default operation
4934 * @param ops the operations initialized
4939 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
4943 ops->transform_node = transform_node_##a; \
4980 } /* firm_set_default_transform_node */
4983 /* **************** Common Subexpression Elimination **************** */
4985 /** The size of the hash table used, should estimate the number of nodes
4987 #define N_IR_NODES 512
4989 /** Compares the attributes of two Const nodes. */
4990 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
4991 return (get_Const_tarval(a) != get_Const_tarval(b))
4992 || (get_Const_type(a) != get_Const_type(b));
4993 } /* node_cmp_attr_Const */
4995 /** Compares the attributes of two Proj nodes. */
4996 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
4997 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
4998 } /* node_cmp_attr_Proj */
5000 /** Compares the attributes of two Filter nodes. */
5001 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5002 return get_Filter_proj(a) != get_Filter_proj(b);
5003 } /* node_cmp_attr_Filter */
5005 /** Compares the attributes of two Alloc nodes. */
5006 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5007 const alloc_attr *pa = get_irn_alloc_attr(a);
5008 const alloc_attr *pb = get_irn_alloc_attr(b);
5009 return (pa->where != pb->where) || (pa->type != pb->type);
5010 } /* node_cmp_attr_Alloc */
5012 /** Compares the attributes of two Free nodes. */
5013 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5014 const free_attr *pa = get_irn_free_attr(a);
5015 const free_attr *pb = get_irn_free_attr(b);
5016 return (pa->where != pb->where) || (pa->type != pb->type);
5017 } /* node_cmp_attr_Free */
5019 /** Compares the attributes of two SymConst nodes. */
5020 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5021 const symconst_attr *pa = get_irn_symconst_attr(a);
5022 const symconst_attr *pb = get_irn_symconst_attr(b);
5023 return (pa->num != pb->num)
5024 || (pa->sym.type_p != pb->sym.type_p)
5025 || (pa->tp != pb->tp);
5026 } /* node_cmp_attr_SymConst */
5028 /** Compares the attributes of two Call nodes. */
5029 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5030 return get_irn_call_attr(a) != get_irn_call_attr(b);
5031 } /* node_cmp_attr_Call */
5033 /** Compares the attributes of two Sel nodes. */
5034 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5035 const ir_entity *a_ent = get_Sel_entity(a);
5036 const ir_entity *b_ent = get_Sel_entity(b);
5038 (a_ent->kind != b_ent->kind) ||
5039 (a_ent->name != b_ent->name) ||
5040 (a_ent->owner != b_ent->owner) ||
5041 (a_ent->ld_name != b_ent->ld_name) ||
5042 (a_ent->type != b_ent->type);
5043 } /* node_cmp_attr_Sel */
5045 /** Compares the attributes of two Phi nodes. */
5046 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5047 /* we can only enter this function if both nodes have the same number of inputs,
5048 hence it is enough to check if one of them is a Phi0 */
5050 /* check the Phi0 pos attribute */
5051 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5054 } /* node_cmp_attr_Phi */
5056 /** Compares the attributes of two Conv nodes. */
5057 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5058 return get_Conv_strict(a) != get_Conv_strict(b);
5059 } /* node_cmp_attr_Conv */
5061 /** Compares the attributes of two Cast nodes. */
5062 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5063 return get_Cast_type(a) != get_Cast_type(b);
5064 } /* node_cmp_attr_Cast */
5066 /** Compares the attributes of two Load nodes. */
5067 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5068 if (get_Load_volatility(a) == volatility_is_volatile ||
5069 get_Load_volatility(b) == volatility_is_volatile)
5070 /* NEVER do CSE on volatile Loads */
5072 /* do not CSE Loads with different alignment. Be conservative. */
5073 if (get_Load_align(a) != get_Load_align(b))
5076 return get_Load_mode(a) != get_Load_mode(b);
5077 } /* node_cmp_attr_Load */
5079 /** Compares the attributes of two Store nodes. */
5080 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5081 /* do not CSE Stores with different alignment. Be conservative. */
5082 if (get_Store_align(a) != get_Store_align(b))
5085 /* NEVER do CSE on volatile Stores */
5086 return (get_Store_volatility(a) == volatility_is_volatile ||
5087 get_Store_volatility(b) == volatility_is_volatile);
5088 } /* node_cmp_attr_Store */
5090 /** Compares two exception attributes */
5091 static int node_cmp_exception(ir_node *a, ir_node *b) {
5092 const except_attr *ea = get_irn_except_attr(a);
5093 const except_attr *eb = get_irn_except_attr(b);
5095 return ea->pin_state != eb->pin_state;
5098 #define node_cmp_attr_Bound node_cmp_exception
5100 /** Compares the attributes of two Div nodes. */
5101 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5102 const divmod_attr *ma = get_irn_divmod_attr(a);
5103 const divmod_attr *mb = get_irn_divmod_attr(b);
5104 return ma->exc.pin_state != mb->exc.pin_state ||
5105 ma->res_mode != mb->res_mode ||
5106 ma->no_remainder != mb->no_remainder;
5107 } /* node_cmp_attr_Div */
5109 /** Compares the attributes of two DivMod nodes. */
5110 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5111 const divmod_attr *ma = get_irn_divmod_attr(a);
5112 const divmod_attr *mb = get_irn_divmod_attr(b);
5113 return ma->exc.pin_state != mb->exc.pin_state ||
5114 ma->res_mode != mb->res_mode;
5115 } /* node_cmp_attr_DivMod */
5117 /** Compares the attributes of two Mod nodes. */
5118 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5119 const divmod_attr *ma = get_irn_divmod_attr(a);
5120 const divmod_attr *mb = get_irn_divmod_attr(b);
5121 return ma->exc.pin_state != mb->exc.pin_state ||
5122 ma->res_mode != mb->res_mode;
5123 } /* node_cmp_attr_Mod */
5125 /** Compares the attributes of two Quot nodes. */
5126 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5127 const divmod_attr *ma = get_irn_divmod_attr(a);
5128 const divmod_attr *mb = get_irn_divmod_attr(b);
5129 return ma->exc.pin_state != mb->exc.pin_state ||
5130 ma->res_mode != mb->res_mode;
5131 } /* node_cmp_attr_Quot */
5133 /** Compares the attributes of two Confirm nodes. */
5134 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5135 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5136 } /* node_cmp_attr_Confirm */
5138 /** Compares the attributes of two ASM nodes. */
5139 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5141 const ir_asm_constraint *ca;
5142 const ir_asm_constraint *cb;
5145 if (get_ASM_text(a) != get_ASM_text(b))
5148 /* Should we really check the constraints here? Should be better, but is strange. */
5149 n = get_ASM_n_input_constraints(a);
5150 if (n != get_ASM_n_input_constraints(b))
5153 ca = get_ASM_input_constraints(a);
5154 cb = get_ASM_input_constraints(b);
5155 for (i = 0; i < n; ++i) {
5156 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5160 n = get_ASM_n_output_constraints(a);
5161 if (n != get_ASM_n_output_constraints(b))
5164 ca = get_ASM_output_constraints(a);
5165 cb = get_ASM_output_constraints(b);
5166 for (i = 0; i < n; ++i) {
5167 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5171 n = get_ASM_n_clobbers(a);
5172 if (n != get_ASM_n_clobbers(b))
5175 cla = get_ASM_clobbers(a);
5176 clb = get_ASM_clobbers(b);
5177 for (i = 0; i < n; ++i) {
5178 if (cla[i] != clb[i])
5182 } /* node_cmp_attr_ASM */
5185 * Set the default node attribute compare operation for an ir_op_ops.
5187 * @param code the opcode for the default operation
5188 * @param ops the operations initialized
5193 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5197 ops->node_cmp_attr = node_cmp_attr_##a; \
5228 } /* firm_set_default_node_cmp_attr */
5231 * Compare function for two nodes in the value table. Gets two
5232 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5234 int identities_cmp(const void *elt, const void *key) {
5235 ir_node *a = (ir_node *)elt;
5236 ir_node *b = (ir_node *)key;
5239 if (a == b) return 0;
5241 if ((get_irn_op(a) != get_irn_op(b)) ||
5242 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5244 /* compare if a's in and b's in are of equal length */
5245 irn_arity_a = get_irn_intra_arity(a);
5246 if (irn_arity_a != get_irn_intra_arity(b))
5249 if (get_irn_pinned(a) == op_pin_state_pinned) {
5250 /* for pinned nodes, the block inputs must be equal */
5251 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5253 } else if (! get_opt_global_cse()) {
5254 /* for block-local CSE both nodes must be in the same MacroBlock */
5255 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5259 /* compare a->in[0..ins] with b->in[0..ins] */
5260 for (i = 0; i < irn_arity_a; i++)
5261 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5265 * here, we already now that the nodes are identical except their
5268 if (a->op->ops.node_cmp_attr)
5269 return a->op->ops.node_cmp_attr(a, b);
5272 } /* identities_cmp */
5275 * Calculate a hash value of a node.
5277 unsigned ir_node_hash(ir_node *node) {
5281 if (node->op == op_Const) {
5282 /* special value for const, as they only differ in their tarval. */
5283 h = HASH_PTR(node->attr.con.tv);
5284 h = 9*h + HASH_PTR(get_irn_mode(node));
5285 } else if (node->op == op_SymConst) {
5286 /* special value for const, as they only differ in their symbol. */
5287 h = HASH_PTR(node->attr.symc.sym.type_p);
5288 h = 9*h + HASH_PTR(get_irn_mode(node));
5291 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5292 h = irn_arity = get_irn_intra_arity(node);
5294 /* consider all in nodes... except the block if not a control flow. */
5295 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5296 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5300 h = 9*h + HASH_PTR(get_irn_mode(node));
5302 h = 9*h + HASH_PTR(get_irn_op(node));
5306 } /* ir_node_hash */
5308 pset *new_identities(void) {
5309 return new_pset(identities_cmp, N_IR_NODES);
5310 } /* new_identities */
5312 void del_identities(pset *value_table) {
5313 del_pset(value_table);
5314 } /* del_identities */
5317 * Normalize a node by putting constants (and operands with larger
5318 * node index) on the right (operator side).
5320 * @param n The node to normalize
5322 static void normalize_node(ir_node *n) {
5323 if (is_op_commutative(get_irn_op(n))) {
5324 ir_node *l = get_binop_left(n);
5325 ir_node *r = get_binop_right(n);
5327 /* For commutative operators perform a OP b == b OP a but keep
5328 * constants on the RIGHT side. This helps greatly in some
5329 * optimizations. Moreover we use the idx number to make the form
5331 if (!operands_are_normalized(l, r)) {
5332 set_binop_left(n, r);
5333 set_binop_right(n, l);
5336 } /* normalize_node */
5339 * Update the nodes after a match in the value table. If both nodes have
5340 * the same MacroBlock but different Blocks, we must ensure that the node
5341 * with the dominating Block (the node that is near to the MacroBlock header
5342 * is stored in the table.
5343 * Because a MacroBlock has only one "non-exception" flow, we don't need
5344 * dominance info here: We known, that one block must dominate the other and
5345 * following the only block input will allow to find it.
5347 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5348 ir_node *known_blk, *new_block, *block, *mbh;
5350 if (get_opt_global_cse()) {
5351 /* Block inputs are meaning less */
5354 known_blk = get_irn_n(known_irn, -1);
5355 new_block = get_irn_n(new_ir_node, -1);
5356 if (known_blk == new_block) {
5357 /* already in the same block */
5361 * We expect the typical case when we built the graph. In that case, the
5362 * known_irn is already the upper one, so checking this should be faster.
5365 mbh = get_Block_MacroBlock(new_block);
5367 if (block == known_blk) {
5368 /* ok, we have found it: known_block dominates new_block as expected */
5373 * We have reached the MacroBlock header NOT founding
5374 * the known_block. new_block must dominate known_block.
5377 set_irn_n(known_irn, -1, new_block);
5380 assert(get_Block_n_cfgpreds(block) == 1);
5381 block = get_Block_cfgpred_block(block, 0);
5383 } /* update_value_table */
5386 * Return the canonical node computing the same value as n.
5388 * @param value_table The value table
5389 * @param n The node to lookup
5391 * Looks up the node in a hash table.
5393 * For Const nodes this is performed in the constructor, too. Const
5394 * nodes are extremely time critical because of their frequent use in
5395 * constant string arrays.
5397 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5400 if (!value_table) return n;
5404 o = pset_find(value_table, n, ir_node_hash(n));
5408 update_known_irn(o, n);
5415 * During construction we set the op_pin_state_pinned flag in the graph right when the
5416 * optimization is performed. The flag turning on procedure global cse could
5417 * be changed between two allocations. This way we are safe.
5419 * @param value_table The value table
5420 * @param n The node to lookup
5422 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5425 n = identify(value_table, n);
5426 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5427 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5429 } /* identify_cons */
5432 * Return the canonical node computing the same value as n.
5433 * Looks up the node in a hash table, enters it in the table
5434 * if it isn't there yet.
5436 * @param value_table the HashSet containing all nodes in the
5438 * @param n the node to look up
5440 * @return a node that computes the same value as n or n if no such
5441 * node could be found
5443 ir_node *identify_remember(pset *value_table, ir_node *n) {
5446 if (!value_table) return n;
5449 /* lookup or insert in hash table with given hash key. */
5450 o = pset_insert(value_table, n, ir_node_hash(n));
5453 update_known_irn(o, n);
5458 } /* identify_remember */
5460 /* Add a node to the identities value table. */
5461 void add_identities(pset *value_table, ir_node *node) {
5462 if (get_opt_cse() && is_no_Block(node))
5463 identify_remember(value_table, node);
5464 } /* add_identities */
5466 /* Visit each node in the value table of a graph. */
5467 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5469 ir_graph *rem = current_ir_graph;
5471 current_ir_graph = irg;
5472 foreach_pset(irg->value_table, node)
5474 current_ir_graph = rem;
5475 } /* visit_all_identities */
5478 * Garbage in, garbage out. If a node has a dead input, i.e., the
5479 * Bad node is input to the node, return the Bad node.
5481 static ir_node *gigo(ir_node *node) {
5483 ir_op *op = get_irn_op(node);
5485 /* remove garbage blocks by looking at control flow that leaves the block
5486 and replacing the control flow by Bad. */
5487 if (get_irn_mode(node) == mode_X) {
5488 ir_node *block = get_nodes_block(skip_Proj(node));
5490 /* Don't optimize nodes in immature blocks. */
5491 if (!get_Block_matured(block))
5493 /* Don't optimize End, may have Bads. */
5494 if (op == op_End) return node;
5496 if (is_Block(block)) {
5497 if (is_Block_dead(block)) {
5498 /* control flow from dead block is dead */
5502 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5503 if (!is_Bad(get_irn_n(block, i)))
5507 ir_graph *irg = get_irn_irg(block);
5508 /* the start block is never dead */
5509 if (block != get_irg_start_block(irg)
5510 && block != get_irg_end_block(irg)) {
5512 * Do NOT kill control flow without setting
5513 * the block to dead of bad things can happen:
5514 * We get a Block that is not reachable be irg_block_walk()
5515 * but can be found by irg_walk()!
5517 set_Block_dead(block);
5524 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5525 blocks predecessors is dead. */
5526 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5527 irn_arity = get_irn_arity(node);
5530 * Beware: we can only read the block of a non-floating node.
5532 if (is_irn_pinned_in_irg(node) &&
5533 is_Block_dead(get_nodes_block(skip_Proj(node))))
5536 for (i = 0; i < irn_arity; i++) {
5537 ir_node *pred = get_irn_n(node, i);
5542 /* Propagating Unknowns here seems to be a bad idea, because
5543 sometimes we need a node as a input and did not want that
5545 However, it might be useful to move this into a later phase
5546 (if you think that optimizing such code is useful). */
5547 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5548 return new_Unknown(get_irn_mode(node));
5553 /* With this code we violate the agreement that local_optimize
5554 only leaves Bads in Block, Phi and Tuple nodes. */
5555 /* If Block has only Bads as predecessors it's garbage. */
5556 /* If Phi has only Bads as predecessors it's garbage. */
5557 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5558 irn_arity = get_irn_arity(node);
5559 for (i = 0; i < irn_arity; i++) {
5560 if (!is_Bad(get_irn_n(node, i))) break;
5562 if (i == irn_arity) node = new_Bad();
5569 * These optimizations deallocate nodes from the obstack.
5570 * It can only be called if it is guaranteed that no other nodes
5571 * reference this one, i.e., right after construction of a node.
5573 * @param n The node to optimize
5575 * current_ir_graph must be set to the graph of the node!
5577 ir_node *optimize_node(ir_node *n) {
5580 ir_opcode iro = get_irn_opcode(n);
5582 /* Always optimize Phi nodes: part of the construction. */
5583 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5585 /* constant expression evaluation / constant folding */
5586 if (get_opt_constant_folding()) {
5587 /* neither constants nor Tuple values can be evaluated */
5588 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5589 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5590 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5591 /* try to evaluate */
5592 tv = computed_value(n);
5593 if (tv != tarval_bad) {
5595 ir_type *old_tp = get_irn_type(n);
5596 int i, arity = get_irn_arity(n);
5600 * Try to recover the type of the new expression.
5602 for (i = 0; i < arity && !old_tp; ++i)
5603 old_tp = get_irn_type(get_irn_n(n, i));
5606 * we MUST copy the node here temporary, because it's still needed
5607 * for DBG_OPT_CSTEVAL
5609 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5610 oldn = alloca(node_size);
5612 memcpy(oldn, n, node_size);
5613 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5615 /* ARG, copy the in array, we need it for statistics */
5616 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5618 /* note the inplace edges module */
5619 edges_node_deleted(n, current_ir_graph);
5621 /* evaluation was successful -- replace the node. */
5622 irg_kill_node(current_ir_graph, n);
5623 nw = new_Const(get_tarval_mode(tv), tv);
5625 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5626 set_Const_type(nw, old_tp);
5627 DBG_OPT_CSTEVAL(oldn, nw);
5628 tarval_enable_fp_ops(old_fp_mode);
5631 tarval_enable_fp_ops(old_fp_mode);
5635 /* remove unnecessary nodes */
5636 if (get_opt_constant_folding() ||
5637 (iro == iro_Phi) || /* always optimize these nodes. */
5639 (iro == iro_Proj) ||
5640 (iro == iro_Block) ) /* Flags tested local. */
5641 n = equivalent_node(n);
5643 /* Common Subexpression Elimination.
5645 * Checks whether n is already available.
5646 * The block input is used to distinguish different subexpressions. Right
5647 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5648 * subexpressions within a block.
5651 n = identify_cons(current_ir_graph->value_table, n);
5654 edges_node_deleted(oldn, current_ir_graph);
5656 /* We found an existing, better node, so we can deallocate the old node. */
5657 irg_kill_node(current_ir_graph, oldn);
5661 /* Some more constant expression evaluation that does not allow to
5663 iro = get_irn_opcode(n);
5664 if (get_opt_constant_folding() ||
5665 (iro == iro_Cond) ||
5666 (iro == iro_Proj)) /* Flags tested local. */
5667 n = transform_node(n);
5669 /* Remove nodes with dead (Bad) input.
5670 Run always for transformation induced Bads. */
5673 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5674 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5675 n = identify_remember(current_ir_graph->value_table, n);
5679 } /* optimize_node */
5683 * These optimizations never deallocate nodes (in place). This can cause dead
5684 * nodes lying on the obstack. Remove these by a dead node elimination,
5685 * i.e., a copying garbage collection.
5687 ir_node *optimize_in_place_2(ir_node *n) {
5690 ir_opcode iro = get_irn_opcode(n);
5692 if (!get_opt_optimize() && !is_Phi(n)) return n;
5694 /* constant expression evaluation / constant folding */
5695 if (get_opt_constant_folding()) {
5696 /* neither constants nor Tuple values can be evaluated */
5697 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5698 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5699 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5700 /* try to evaluate */
5701 tv = computed_value(n);
5702 if (tv != tarval_bad) {
5703 /* evaluation was successful -- replace the node. */
5704 ir_type *old_tp = get_irn_type(n);
5705 int i, arity = get_irn_arity(n);
5708 * Try to recover the type of the new expression.
5710 for (i = 0; i < arity && !old_tp; ++i)
5711 old_tp = get_irn_type(get_irn_n(n, i));
5713 n = new_Const(get_tarval_mode(tv), tv);
5715 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5716 set_Const_type(n, old_tp);
5718 DBG_OPT_CSTEVAL(oldn, n);
5719 tarval_enable_fp_ops(old_fp_mode);
5722 tarval_enable_fp_ops(old_fp_mode);
5726 /* remove unnecessary nodes */
5727 if (get_opt_constant_folding() ||
5728 (iro == iro_Phi) || /* always optimize these nodes. */
5729 (iro == iro_Id) || /* ... */
5730 (iro == iro_Proj) || /* ... */
5731 (iro == iro_Block) ) /* Flags tested local. */
5732 n = equivalent_node(n);
5734 /** common subexpression elimination **/
5735 /* Checks whether n is already available. */
5736 /* The block input is used to distinguish different subexpressions. Right
5737 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5738 subexpressions within a block. */
5739 if (get_opt_cse()) {
5740 n = identify(current_ir_graph->value_table, n);
5743 /* Some more constant expression evaluation. */
5744 iro = get_irn_opcode(n);
5745 if (get_opt_constant_folding() ||
5746 (iro == iro_Cond) ||
5747 (iro == iro_Proj)) /* Flags tested local. */
5748 n = transform_node(n);
5750 /* Remove nodes with dead (Bad) input.
5751 Run always for transformation induced Bads. */
5754 /* Now we can verify the node, as it has no dead inputs any more. */
5757 /* Now we have a legal, useful node. Enter it in hash table for cse.
5758 Blocks should be unique anyways. (Except the successor of start:
5759 is cse with the start block!) */
5760 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5761 n = identify_remember(current_ir_graph->value_table, n);
5764 } /* optimize_in_place_2 */
5767 * Wrapper for external use, set proper status bits after optimization.
5769 ir_node *optimize_in_place(ir_node *n) {
5770 /* Handle graph state */
5771 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5773 if (get_opt_global_cse())
5774 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5775 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5776 set_irg_outs_inconsistent(current_ir_graph);
5778 /* FIXME: Maybe we could also test whether optimizing the node can
5779 change the control graph. */
5780 set_irg_doms_inconsistent(current_ir_graph);
5781 return optimize_in_place_2(n);
5782 } /* optimize_in_place */
5785 * Sets the default operation for an ir_ops.
5787 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5788 ops = firm_set_default_computed_value(code, ops);
5789 ops = firm_set_default_equivalent_node(code, ops);
5790 ops = firm_set_default_transform_node(code, ops);
5791 ops = firm_set_default_node_cmp_attr(code, ops);
5792 ops = firm_set_default_get_type(code, ops);
5793 ops = firm_set_default_get_type_attr(code, ops);
5794 ops = firm_set_default_get_entity_attr(code, ops);
5797 } /* firm_set_default_operations */