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 "idempotent 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) {
1235 ir_node *block = NULL; /* to shutup gcc */
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 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1244 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1245 if ((is_Block_dead(block)) || /* Control dead */
1246 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1247 return new_Bad(); /* in the Start Block. */
1249 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1251 /* If the Block has a Bad pred, we also have one. */
1252 for (i = 0; i < n_preds; ++i)
1253 if (is_Bad(get_Block_cfgpred(block, i)))
1254 set_Phi_pred(n, i, new_Bad());
1256 /* Find first non-self-referencing input */
1257 for (i = 0; i < n_preds; ++i) {
1258 first_val = get_Phi_pred(n, i);
1259 if ( (first_val != n) /* not self pointer */
1261 && (! is_Bad(first_val))
1263 ) { /* value not dead */
1264 break; /* then found first value. */
1269 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1273 /* search for rest of inputs, determine if any of these
1274 are non-self-referencing */
1275 while (++i < n_preds) {
1276 ir_node *scnd_val = get_Phi_pred(n, i);
1277 if ( (scnd_val != n)
1278 && (scnd_val != first_val)
1280 && (! is_Bad(scnd_val))
1288 /* Fold, if no multiple distinct non-self-referencing inputs */
1290 DBG_OPT_PHI(oldn, n);
1293 } /* equivalent_node_Phi */
1296 * Several optimizations:
1297 * - no Sync in start block.
1298 * - fold Sync-nodes, iff they have only one predecessor except
1301 static ir_node *equivalent_node_Sync(ir_node *n) {
1305 ir_node *first_val = NULL; /* to shutup gcc */
1307 if (!get_opt_normalize()) return n;
1309 n_preds = get_Sync_n_preds(n);
1311 /* Find first non-self-referencing input */
1312 for (i = 0; i < n_preds; ++i) {
1313 first_val = get_Sync_pred(n, i);
1314 if ((first_val != n) /* not self pointer */ &&
1315 (! is_Bad(first_val))
1316 ) { /* value not dead */
1317 break; /* then found first value. */
1322 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1325 /* search the rest of inputs, determine if any of these
1326 are non-self-referencing */
1327 while (++i < n_preds) {
1328 ir_node *scnd_val = get_Sync_pred(n, i);
1329 if ((scnd_val != n) &&
1330 (scnd_val != first_val) &&
1331 (! is_Bad(scnd_val))
1337 /* Fold, if no multiple distinct non-self-referencing inputs */
1339 DBG_OPT_SYNC(oldn, n);
1342 } /* equivalent_node_Sync */
1345 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1346 * ProjX(Load) and ProjX(Store).
1348 static ir_node *equivalent_node_Proj(ir_node *proj) {
1349 ir_node *oldn = proj;
1350 ir_node *a = get_Proj_pred(proj);
1353 /* Remove the Tuple/Proj combination. */
1354 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1355 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1356 DBG_OPT_TUPLE(oldn, a, proj);
1358 /* This should not happen! */
1359 assert(! "found a Proj with higher number than Tuple predecessors");
1362 } else if (get_irn_mode(proj) == mode_X) {
1363 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1364 /* Remove dead control flow -- early gigo(). */
1366 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1368 /* get the Load address */
1369 ir_node *addr = get_Load_ptr(a);
1370 ir_node *blk = get_irn_n(a, -1);
1373 if (value_not_null(addr, &confirm)) {
1374 if (confirm == NULL) {
1375 /* this node may float if it did not depend on a Confirm */
1376 set_irn_pinned(a, op_pin_state_floats);
1378 if (get_Proj_proj(proj) == pn_Load_X_except) {
1379 DBG_OPT_EXC_REM(proj);
1382 return new_r_Jmp(current_ir_graph, blk);
1384 } else if (is_Store(a)) {
1385 /* get the load/store address */
1386 ir_node *addr = get_Store_ptr(a);
1387 ir_node *blk = get_irn_n(a, -1);
1390 if (value_not_null(addr, &confirm)) {
1391 if (confirm == NULL) {
1392 /* this node may float if it did not depend on a Confirm */
1393 set_irn_pinned(a, op_pin_state_floats);
1395 if (get_Proj_proj(proj) == pn_Store_X_except) {
1396 DBG_OPT_EXC_REM(proj);
1399 return new_r_Jmp(current_ir_graph, blk);
1406 } /* equivalent_node_Proj */
1411 static ir_node *equivalent_node_Id(ir_node *n) {
1416 } while (get_irn_op(n) == op_Id);
1418 DBG_OPT_ID(oldn, n);
1420 } /* equivalent_node_Id */
1425 static ir_node *equivalent_node_Mux(ir_node *n)
1427 ir_node *oldn = n, *sel = get_Mux_sel(n);
1428 tarval *ts = value_of(sel);
1430 /* Mux(true, f, t) == t */
1431 if (ts == tarval_b_true) {
1432 n = get_Mux_true(n);
1433 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1435 /* Mux(false, f, t) == f */
1436 else if (ts == tarval_b_false) {
1437 n = get_Mux_false(n);
1438 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1440 /* Mux(v, x, x) == x */
1441 else if (get_Mux_false(n) == get_Mux_true(n)) {
1442 n = get_Mux_true(n);
1443 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1445 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1446 ir_node *cmp = get_Proj_pred(sel);
1447 long proj_nr = get_Proj_proj(sel);
1448 ir_node *b = get_Mux_false(n);
1449 ir_node *a = get_Mux_true(n);
1452 * Note: normalization puts the constant on the right site,
1453 * so we check only one case.
1455 * Note further that these optimization work even for floating point
1456 * with NaN's because -NaN == NaN.
1457 * However, if +0 and -0 is handled differently, we cannot use the first one.
1459 if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) {
1460 ir_node *cmp_r = get_Cmp_right(cmp);
1461 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
1462 /* Mux(a CMP 0, X, a) */
1463 if (is_Minus(b) && get_Minus_op(b) == a) {
1464 /* Mux(a CMP 0, -a, a) */
1465 if (proj_nr == pn_Cmp_Eq) {
1466 /* Mux(a == 0, -a, a) ==> -a */
1468 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1469 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1470 /* Mux(a != 0, -a, a) ==> a */
1472 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1474 } else if (is_Const(b) && is_Const_null(b)) {
1475 /* Mux(a CMP 0, 0, a) */
1476 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1477 /* Mux(a != 0, 0, a) ==> a */
1479 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1480 } else if (proj_nr == pn_Cmp_Eq) {
1481 /* Mux(a == 0, 0, a) ==> 0 */
1483 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1490 } /* equivalent_node_Mux */
1493 * Returns a equivalent node of a Psi: if a condition is true
1494 * and all previous conditions are false we know its value.
1495 * If all conditions are false its value is the default one.
1497 static ir_node *equivalent_node_Psi(ir_node *n) {
1499 return equivalent_node_Mux(n);
1501 } /* equivalent_node_Psi */
1504 * Optimize -a CMP -b into b CMP a.
1505 * This works only for for modes where unary Minus
1507 * Note that two-complement integers can Overflow
1508 * so it will NOT work.
1510 * For == and != can be handled in Proj(Cmp)
1512 static ir_node *equivalent_node_Cmp(ir_node *n) {
1513 ir_node *left = get_Cmp_left(n);
1514 ir_node *right = get_Cmp_right(n);
1516 if (is_Minus(left) && is_Minus(right) &&
1517 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1518 left = get_Minus_op(left);
1519 right = get_Minus_op(right);
1520 set_Cmp_left(n, right);
1521 set_Cmp_right(n, left);
1524 } /* equivalent_node_Cmp */
1527 * Remove Confirm nodes if setting is on.
1528 * Replace Confirms(x, '=', Constlike) by Constlike.
1530 static ir_node *equivalent_node_Confirm(ir_node *n) {
1531 ir_node *pred = get_Confirm_value(n);
1532 pn_Cmp pnc = get_Confirm_cmp(n);
1534 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1536 * rare case: two identical Confirms one after another,
1537 * replace the second one with the first.
1541 if (pnc == pn_Cmp_Eq) {
1542 ir_node *bound = get_Confirm_bound(n);
1545 * Optimize a rare case:
1546 * Confirm(x, '=', Constlike) ==> Constlike
1548 if (is_irn_constlike(bound)) {
1549 DBG_OPT_CONFIRM(n, bound);
1553 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1557 * Optimize CopyB(mem, x, x) into a Nop.
1559 static ir_node *equivalent_node_CopyB(ir_node *n) {
1560 ir_node *a = get_CopyB_dst(n);
1561 ir_node *b = get_CopyB_src(n);
1564 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1565 ir_node *mem = get_CopyB_mem(n);
1566 ir_node *blk = get_nodes_block(n);
1567 turn_into_tuple(n, pn_CopyB_max);
1568 set_Tuple_pred(n, pn_CopyB_M, mem);
1569 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1570 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1571 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1574 } /* equivalent_node_CopyB */
1577 * Optimize Bounds(idx, idx, upper) into idx.
1579 static ir_node *equivalent_node_Bound(ir_node *n) {
1580 ir_node *idx = get_Bound_index(n);
1581 ir_node *pred = skip_Proj(idx);
1584 if (is_Bound(pred)) {
1586 * idx was Bounds checked in the same MacroBlock previously,
1587 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1589 ir_node *lower = get_Bound_lower(n);
1590 ir_node *upper = get_Bound_upper(n);
1591 if (get_Bound_lower(pred) == lower &&
1592 get_Bound_upper(pred) == upper &&
1593 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1595 * One could expect that we simply return the previous
1596 * Bound here. However, this would be wrong, as we could
1597 * add an exception Proj to a new location then.
1598 * So, we must turn in into a tuple.
1604 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1605 ir_node *mem = get_Bound_mem(n);
1606 ir_node *blk = get_nodes_block(n);
1607 turn_into_tuple(n, pn_Bound_max);
1608 set_Tuple_pred(n, pn_Bound_M, mem);
1609 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1610 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1611 set_Tuple_pred(n, pn_Bound_res, idx);
1614 } /* equivalent_node_Bound */
1617 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1618 * perform no actual computation, as, e.g., the Id nodes. It does not create
1619 * new nodes. It is therefore safe to free n if the node returned is not n.
1620 * If a node returns a Tuple we can not just skip it. If the size of the
1621 * in array fits, we transform n into a tuple (e.g., Div).
1623 ir_node *equivalent_node(ir_node *n) {
1624 if (n->op->ops.equivalent_node)
1625 return n->op->ops.equivalent_node(n);
1627 } /* equivalent_node */
1630 * Sets the default equivalent node operation for an ir_op_ops.
1632 * @param code the opcode for the default operation
1633 * @param ops the operations initialized
1638 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1642 ops->equivalent_node = equivalent_node_##a; \
1682 } /* firm_set_default_equivalent_node */
1685 * Returns non-zero if a node is a Phi node
1686 * with all predecessors constant.
1688 static int is_const_Phi(ir_node *n) {
1691 if (! is_Phi(n) || get_irn_arity(n) == 0)
1693 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1694 if (! is_Const(get_irn_n(n, i)))
1697 } /* is_const_Phi */
1700 * Apply an evaluator on a binop with a constant operators (and one Phi).
1702 * @param phi the Phi node
1703 * @param other the other operand
1704 * @param eval an evaluator function
1705 * @param mode the mode of the result, may be different from the mode of the Phi!
1706 * @param left if non-zero, other is the left operand, else the right
1708 * @return a new Phi node if the conversion was successful, NULL else
1710 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1715 int i, n = get_irn_arity(phi);
1717 NEW_ARR_A(void *, res, n);
1719 for (i = 0; i < n; ++i) {
1720 pred = get_irn_n(phi, i);
1721 tv = get_Const_tarval(pred);
1722 tv = eval(other, tv);
1724 if (tv == tarval_bad) {
1725 /* folding failed, bad */
1731 for (i = 0; i < n; ++i) {
1732 pred = get_irn_n(phi, i);
1733 tv = get_Const_tarval(pred);
1734 tv = eval(tv, other);
1736 if (tv == tarval_bad) {
1737 /* folding failed, bad */
1743 irg = current_ir_graph;
1744 for (i = 0; i < n; ++i) {
1745 pred = get_irn_n(phi, i);
1746 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1747 mode, res[i], get_Const_type(pred));
1749 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1750 } /* apply_binop_on_phi */
1753 * Apply an evaluator on a binop with two constant Phi.
1755 * @param a the left Phi node
1756 * @param b the right Phi node
1757 * @param eval an evaluator function
1758 * @param mode the mode of the result, may be different from the mode of the Phi!
1760 * @return a new Phi node if the conversion was successful, NULL else
1762 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1763 tarval *tv_l, *tv_r, *tv;
1769 if (get_nodes_block(a) != get_nodes_block(b))
1772 n = get_irn_arity(a);
1773 NEW_ARR_A(void *, res, n);
1775 for (i = 0; i < n; ++i) {
1776 pred = get_irn_n(a, i);
1777 tv_l = get_Const_tarval(pred);
1778 pred = get_irn_n(b, i);
1779 tv_r = get_Const_tarval(pred);
1780 tv = eval(tv_l, tv_r);
1782 if (tv == tarval_bad) {
1783 /* folding failed, bad */
1788 irg = current_ir_graph;
1789 for (i = 0; i < n; ++i) {
1790 pred = get_irn_n(a, i);
1791 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1793 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1794 } /* apply_binop_on_2_phis */
1797 * Apply an evaluator on a unop with a constant operator (a Phi).
1799 * @param phi the Phi node
1800 * @param eval an evaluator function
1802 * @return a new Phi node if the conversion was successful, NULL else
1804 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1810 int i, n = get_irn_arity(phi);
1812 NEW_ARR_A(void *, res, n);
1813 for (i = 0; i < n; ++i) {
1814 pred = get_irn_n(phi, i);
1815 tv = get_Const_tarval(pred);
1818 if (tv == tarval_bad) {
1819 /* folding failed, bad */
1824 mode = get_irn_mode(phi);
1825 irg = current_ir_graph;
1826 for (i = 0; i < n; ++i) {
1827 pred = get_irn_n(phi, i);
1828 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1829 mode, res[i], get_Const_type(pred));
1831 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1832 } /* apply_unop_on_phi */
1835 * Apply a conversion on a constant operator (a Phi).
1837 * @param phi the Phi node
1839 * @return a new Phi node if the conversion was successful, NULL else
1841 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1846 int i, n = get_irn_arity(phi);
1848 NEW_ARR_A(void *, res, n);
1849 for (i = 0; i < n; ++i) {
1850 pred = get_irn_n(phi, i);
1851 tv = get_Const_tarval(pred);
1852 tv = tarval_convert_to(tv, mode);
1854 if (tv == tarval_bad) {
1855 /* folding failed, bad */
1860 irg = current_ir_graph;
1861 for (i = 0; i < n; ++i) {
1862 pred = get_irn_n(phi, i);
1863 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1864 mode, res[i], get_Const_type(pred));
1866 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1867 } /* apply_conv_on_phi */
1870 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1871 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1872 * If possible, remove the Conv's.
1874 static ir_node *transform_node_AddSub(ir_node *n) {
1875 ir_mode *mode = get_irn_mode(n);
1877 if (mode_is_reference(mode)) {
1878 ir_node *left = get_binop_left(n);
1879 ir_node *right = get_binop_right(n);
1880 unsigned ref_bits = get_mode_size_bits(mode);
1882 if (is_Conv(left)) {
1883 ir_mode *mode = get_irn_mode(left);
1884 unsigned bits = get_mode_size_bits(mode);
1886 if (ref_bits == bits &&
1887 mode_is_int(mode) &&
1888 get_mode_arithmetic(mode) == irma_twos_complement) {
1889 ir_node *pre = get_Conv_op(left);
1890 ir_mode *pre_mode = get_irn_mode(pre);
1892 if (mode_is_int(pre_mode) &&
1893 get_mode_size_bits(pre_mode) == bits &&
1894 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1895 /* ok, this conv just changes to sign, moreover the calculation
1896 * is done with same number of bits as our address mode, so
1897 * we can ignore the conv as address calculation can be viewed
1898 * as either signed or unsigned
1900 set_binop_left(n, pre);
1905 if (is_Conv(right)) {
1906 ir_mode *mode = get_irn_mode(right);
1907 unsigned bits = get_mode_size_bits(mode);
1909 if (ref_bits == bits &&
1910 mode_is_int(mode) &&
1911 get_mode_arithmetic(mode) == irma_twos_complement) {
1912 ir_node *pre = get_Conv_op(right);
1913 ir_mode *pre_mode = get_irn_mode(pre);
1915 if (mode_is_int(pre_mode) &&
1916 get_mode_size_bits(pre_mode) == bits &&
1917 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1918 /* ok, this conv just changes to sign, moreover the calculation
1919 * is done with same number of bits as our address mode, so
1920 * we can ignore the conv as address calculation can be viewed
1921 * as either signed or unsigned
1923 set_binop_right(n, pre);
1929 } /* transform_node_AddSub */
1931 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1933 if (is_Const(b) && is_const_Phi(a)) { \
1934 /* check for Op(Phi, Const) */ \
1935 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1937 else if (is_Const(a) && is_const_Phi(b)) { \
1938 /* check for Op(Const, Phi) */ \
1939 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1941 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1942 /* check for Op(Phi, Phi) */ \
1943 c = apply_binop_on_2_phis(a, b, eval, mode); \
1946 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1950 #define HANDLE_UNOP_PHI(eval, a, c) \
1952 if (is_const_Phi(a)) { \
1953 /* check for Op(Phi) */ \
1954 c = apply_unop_on_phi(a, eval); \
1956 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1962 * Do the AddSub optimization, then Transform
1963 * Constant folding on Phi
1964 * Add(a,a) -> Mul(a, 2)
1965 * Add(Mul(a, x), a) -> Mul(a, x+1)
1966 * if the mode is integer or float.
1967 * Transform Add(a,-b) into Sub(a,b).
1968 * Reassociation might fold this further.
1970 static ir_node *transform_node_Add(ir_node *n) {
1972 ir_node *a, *b, *c, *oldn = n;
1974 n = transform_node_AddSub(n);
1976 a = get_Add_left(n);
1977 b = get_Add_right(n);
1979 mode = get_irn_mode(n);
1980 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
1982 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1983 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1986 if (mode_is_num(mode)) {
1987 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
1988 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
1989 ir_node *block = get_irn_n(n, -1);
1992 get_irn_dbg_info(n),
1996 new_r_Const_long(current_ir_graph, block, mode, 2),
1998 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2003 get_irn_dbg_info(n),
2009 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2014 get_irn_dbg_info(n),
2020 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2023 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2024 /* Here we rely on constants be on the RIGHT side */
2026 ir_node *op = get_Not_op(a);
2028 if (is_Const(b) && is_Const_one(b)) {
2030 ir_node *blk = get_irn_n(n, -1);
2031 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2032 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2037 ir_node *blk = get_irn_n(n, -1);
2038 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2039 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2044 ir_node *op = get_Not_op(b);
2048 ir_node *blk = get_irn_n(n, -1);
2049 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2050 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2057 } /* transform_node_Add */
2060 * returns -cnst or NULL if impossible
2062 static ir_node *const_negate(ir_node *cnst) {
2063 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2064 dbg_info *dbgi = get_irn_dbg_info(cnst);
2065 ir_graph *irg = get_irn_irg(cnst);
2066 ir_node *block = get_nodes_block(cnst);
2067 ir_mode *mode = get_irn_mode(cnst);
2068 if (tv == tarval_bad) return NULL;
2069 return new_rd_Const(dbgi, irg, block, mode, tv);
2073 * Do the AddSub optimization, then Transform
2074 * Constant folding on Phi
2075 * Sub(0,a) -> Minus(a)
2076 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2077 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2078 * Sub(Add(a, x), x) -> a
2079 * Sub(x, Add(x, a)) -> -a
2080 * Sub(x, Const) -> Add(x, -Const)
2082 static ir_node *transform_node_Sub(ir_node *n) {
2087 n = transform_node_AddSub(n);
2089 a = get_Sub_left(n);
2090 b = get_Sub_right(n);
2092 mode = get_irn_mode(n);
2095 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2097 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2098 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2101 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2102 /* a - C -> a + (-C) */
2103 ir_node *cnst = const_negate(b);
2105 ir_node *block = get_nodes_block(n);
2106 dbg_info *dbgi = get_irn_dbg_info(n);
2107 ir_graph *irg = get_irn_irg(n);
2109 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2110 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2115 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2116 ir_graph *irg = current_ir_graph;
2117 dbg_info *dbg = get_irn_dbg_info(n);
2118 ir_node *block = get_nodes_block(n);
2119 ir_node *left = get_Minus_op(a);
2120 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2122 n = new_rd_Minus(dbg, irg, block, add, mode);
2123 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2125 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2126 ir_graph *irg = current_ir_graph;
2127 dbg_info *dbg = get_irn_dbg_info(n);
2128 ir_node *block = get_nodes_block(n);
2129 ir_node *right = get_Minus_op(b);
2131 n = new_rd_Add(dbg, irg, block, a, right, mode);
2132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2134 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2135 ir_graph *irg = current_ir_graph;
2136 dbg_info *s_dbg = get_irn_dbg_info(b);
2137 ir_node *s_block = get_nodes_block(b);
2138 ir_node *s_left = get_Sub_right(b);
2139 ir_node *s_right = get_Sub_left(b);
2140 ir_mode *s_mode = get_irn_mode(b);
2141 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2142 dbg_info *a_dbg = get_irn_dbg_info(n);
2143 ir_node *a_block = get_nodes_block(n);
2145 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2148 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2149 ir_node *m_right = get_Mul_right(b);
2150 if (is_Const(m_right)) {
2151 ir_node *cnst2 = const_negate(m_right);
2152 if (cnst2 != NULL) {
2153 ir_graph *irg = current_ir_graph;
2154 dbg_info *m_dbg = get_irn_dbg_info(b);
2155 ir_node *m_block = get_nodes_block(b);
2156 ir_node *m_left = get_Mul_left(b);
2157 ir_mode *m_mode = get_irn_mode(b);
2158 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2159 dbg_info *a_dbg = get_irn_dbg_info(n);
2160 ir_node *a_block = get_nodes_block(n);
2162 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2163 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2169 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2170 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2172 get_irn_dbg_info(n),
2177 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2181 if (mode_wrap_around(mode)) {
2182 ir_node *left = get_Add_left(a);
2183 ir_node *right = get_Add_right(a);
2185 /* FIXME: Does the Conv's work only for two complement or generally? */
2187 if (mode != get_irn_mode(right)) {
2188 /* This Sub is an effective Cast */
2189 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2192 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2194 } else if (right == b) {
2195 if (mode != get_irn_mode(left)) {
2196 /* This Sub is an effective Cast */
2197 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2200 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2206 if (mode_wrap_around(mode)) {
2207 ir_node *left = get_Add_left(b);
2208 ir_node *right = get_Add_right(b);
2210 /* FIXME: Does the Conv's work only for two complement or generally? */
2212 ir_mode *r_mode = get_irn_mode(right);
2214 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2215 if (mode != r_mode) {
2216 /* This Sub is an effective Cast */
2217 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2219 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2221 } else if (right == a) {
2222 ir_mode *l_mode = get_irn_mode(left);
2224 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2225 if (mode != l_mode) {
2226 /* This Sub is an effective Cast */
2227 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2229 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2234 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2235 ir_mode *mode = get_irn_mode(a);
2237 if (mode == get_irn_mode(b)) {
2239 ir_node *op_a = get_Conv_op(a);
2240 ir_node *op_b = get_Conv_op(b);
2242 /* check if it's allowed to skip the conv */
2243 ma = get_irn_mode(op_a);
2244 mb = get_irn_mode(op_b);
2246 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2247 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2250 set_Sub_right(n, b);
2256 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2257 if (!is_reassoc_running() && is_Mul(a)) {
2258 ir_node *ma = get_Mul_left(a);
2259 ir_node *mb = get_Mul_right(a);
2262 ir_node *blk = get_irn_n(n, -1);
2264 get_irn_dbg_info(n),
2265 current_ir_graph, blk,
2268 get_irn_dbg_info(n),
2269 current_ir_graph, blk,
2271 new_r_Const_long(current_ir_graph, blk, mode, 1),
2274 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2276 } else if (mb == b) {
2277 ir_node *blk = get_irn_n(n, -1);
2279 get_irn_dbg_info(n),
2280 current_ir_graph, blk,
2283 get_irn_dbg_info(n),
2284 current_ir_graph, blk,
2286 new_r_Const_long(current_ir_graph, blk, mode, 1),
2289 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2293 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2294 ir_node *x = get_Sub_left(a);
2295 ir_node *y = get_Sub_right(a);
2296 ir_node *blk = get_irn_n(n, -1);
2297 ir_mode *m_b = get_irn_mode(b);
2298 ir_mode *m_y = get_irn_mode(y);
2302 /* Determine the right mode for the Add. */
2305 else if (mode_is_reference(m_b))
2307 else if (mode_is_reference(m_y))
2311 * Both modes are different but none is reference,
2312 * happens for instance in SubP(SubP(P, Iu), Is).
2313 * We have two possibilities here: Cast or ignore.
2314 * Currently we ignore this case.
2319 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2321 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2322 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2326 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2327 if (is_Const(a) && is_Not(b)) {
2328 /* c - ~X = X + (c+1) */
2329 tarval *tv = get_Const_tarval(a);
2331 tv = tarval_add(tv, get_mode_one(mode));
2332 if (tv != tarval_bad) {
2333 ir_node *blk = get_irn_n(n, -1);
2334 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2335 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2336 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2342 } /* transform_node_Sub */
2345 * Several transformation done on n*n=2n bits mul.
2346 * These transformations must be done here because new nodes may be produced.
2348 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2350 ir_node *a = get_Mul_left(n);
2351 ir_node *b = get_Mul_right(n);
2352 tarval *ta = value_of(a);
2353 tarval *tb = value_of(b);
2354 ir_mode *smode = get_irn_mode(a);
2356 if (ta == get_mode_one(smode)) {
2357 /* (L)1 * (L)b = (L)b */
2358 ir_node *blk = get_irn_n(n, -1);
2359 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2360 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2363 else if (ta == get_mode_minus_one(smode)) {
2364 /* (L)-1 * (L)b = (L)b */
2365 ir_node *blk = get_irn_n(n, -1);
2366 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2367 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2368 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2371 if (tb == get_mode_one(smode)) {
2372 /* (L)a * (L)1 = (L)a */
2373 ir_node *blk = get_irn_n(a, -1);
2374 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2375 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2378 else if (tb == get_mode_minus_one(smode)) {
2379 /* (L)a * (L)-1 = (L)-a */
2380 ir_node *blk = get_irn_n(n, -1);
2381 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2382 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2383 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2390 * Transform Mul(a,-1) into -a.
2391 * Do constant evaluation of Phi nodes.
2392 * Do architecture dependent optimizations on Mul nodes
2394 static ir_node *transform_node_Mul(ir_node *n) {
2395 ir_node *c, *oldn = n;
2396 ir_mode *mode = get_irn_mode(n);
2397 ir_node *a = get_Mul_left(n);
2398 ir_node *b = get_Mul_right(n);
2400 if (is_Bad(a) || is_Bad(b))
2403 if (mode != get_irn_mode(a))
2404 return transform_node_Mul2n(n, mode);
2406 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2408 if (mode_is_signed(mode)) {
2411 if (value_of(a) == get_mode_minus_one(mode))
2413 else if (value_of(b) == get_mode_minus_one(mode))
2416 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2417 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2422 if (is_Const(b)) { /* (-a) * const -> a * -const */
2423 ir_node *cnst = const_negate(b);
2425 dbg_info *dbgi = get_irn_dbg_info(n);
2426 ir_node *block = get_nodes_block(n);
2427 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2428 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2431 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2432 dbg_info *dbgi = get_irn_dbg_info(n);
2433 ir_node *block = get_nodes_block(n);
2434 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2435 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2437 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2438 ir_node *sub_l = get_Sub_left(b);
2439 ir_node *sub_r = get_Sub_right(b);
2440 dbg_info *dbgi = get_irn_dbg_info(n);
2441 ir_graph *irg = current_ir_graph;
2442 ir_node *block = get_nodes_block(n);
2443 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2444 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2445 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2448 } else if (is_Minus(b)) {
2449 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2450 ir_node *sub_l = get_Sub_left(a);
2451 ir_node *sub_r = get_Sub_right(a);
2452 dbg_info *dbgi = get_irn_dbg_info(n);
2453 ir_graph *irg = current_ir_graph;
2454 ir_node *block = get_nodes_block(n);
2455 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2456 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2457 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2461 if (get_mode_arithmetic(mode) == irma_ieee754) {
2463 tarval *tv = get_Const_tarval(a);
2464 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2465 /* 2.0 * b = b + b */
2466 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2467 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2471 else if (is_Const(b)) {
2472 tarval *tv = get_Const_tarval(b);
2473 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2474 /* a * 2.0 = a + a */
2475 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2476 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2481 return arch_dep_replace_mul_with_shifts(n);
2482 } /* transform_node_Mul */
2485 * Transform a Div Node.
2487 static ir_node *transform_node_Div(ir_node *n) {
2488 ir_mode *mode = get_Div_resmode(n);
2489 ir_node *a = get_Div_left(n);
2490 ir_node *b = get_Div_right(n);
2494 if (is_Const(b) && is_const_Phi(a)) {
2495 /* check for Div(Phi, Const) */
2496 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2498 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2502 else if (is_Const(a) && is_const_Phi(b)) {
2503 /* check for Div(Const, Phi) */
2504 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2506 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2510 else if (is_const_Phi(a) && is_const_Phi(b)) {
2511 /* check for Div(Phi, Phi) */
2512 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2514 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2521 if (tv != tarval_bad) {
2522 value = new_Const(get_tarval_mode(tv), tv);
2524 DBG_OPT_CSTEVAL(n, value);
2527 ir_node *a = get_Div_left(n);
2528 ir_node *b = get_Div_right(n);
2531 if (a == b && value_not_zero(a, &dummy)) {
2532 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2533 value = new_Const(mode, get_mode_one(mode));
2534 DBG_OPT_CSTEVAL(n, value);
2537 if (mode_is_signed(mode) && is_Const(b)) {
2538 tarval *tv = get_Const_tarval(b);
2540 if (tv == get_mode_minus_one(mode)) {
2542 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2543 DBG_OPT_CSTEVAL(n, value);
2547 /* Try architecture dependent optimization */
2548 value = arch_dep_replace_div_by_const(n);
2556 /* Turn Div into a tuple (mem, jmp, bad, value) */
2557 mem = get_Div_mem(n);
2558 blk = get_irn_n(n, -1);
2560 /* skip a potential Pin */
2562 mem = get_Pin_op(mem);
2563 turn_into_tuple(n, pn_Div_max);
2564 set_Tuple_pred(n, pn_Div_M, mem);
2565 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2566 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2567 set_Tuple_pred(n, pn_Div_res, value);
2570 } /* transform_node_Div */
2573 * Transform a Mod node.
2575 static ir_node *transform_node_Mod(ir_node *n) {
2576 ir_mode *mode = get_Mod_resmode(n);
2577 ir_node *a = get_Mod_left(n);
2578 ir_node *b = get_Mod_right(n);
2582 if (is_Const(b) && is_const_Phi(a)) {
2583 /* check for Div(Phi, Const) */
2584 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2586 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2590 else if (is_Const(a) && is_const_Phi(b)) {
2591 /* check for Div(Const, Phi) */
2592 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2594 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2598 else if (is_const_Phi(a) && is_const_Phi(b)) {
2599 /* check for Div(Phi, Phi) */
2600 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2602 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2609 if (tv != tarval_bad) {
2610 value = new_Const(get_tarval_mode(tv), tv);
2612 DBG_OPT_CSTEVAL(n, value);
2615 ir_node *a = get_Mod_left(n);
2616 ir_node *b = get_Mod_right(n);
2619 if (a == b && value_not_zero(a, &dummy)) {
2620 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2621 value = new_Const(mode, get_mode_null(mode));
2622 DBG_OPT_CSTEVAL(n, value);
2625 if (mode_is_signed(mode) && is_Const(b)) {
2626 tarval *tv = get_Const_tarval(b);
2628 if (tv == get_mode_minus_one(mode)) {
2630 value = new_Const(mode, get_mode_null(mode));
2631 DBG_OPT_CSTEVAL(n, value);
2635 /* Try architecture dependent optimization */
2636 value = arch_dep_replace_mod_by_const(n);
2644 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2645 mem = get_Mod_mem(n);
2646 blk = get_irn_n(n, -1);
2648 /* skip a potential Pin */
2650 mem = get_Pin_op(mem);
2651 turn_into_tuple(n, pn_Mod_max);
2652 set_Tuple_pred(n, pn_Mod_M, mem);
2653 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2654 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2655 set_Tuple_pred(n, pn_Mod_res, value);
2658 } /* transform_node_Mod */
2661 * Transform a DivMod node.
2663 static ir_node *transform_node_DivMod(ir_node *n) {
2665 ir_node *a = get_DivMod_left(n);
2666 ir_node *b = get_DivMod_right(n);
2667 ir_mode *mode = get_DivMod_resmode(n);
2672 if (is_Const(b) && is_const_Phi(a)) {
2673 /* check for Div(Phi, Const) */
2674 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2675 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2677 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2678 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2682 else if (is_Const(a) && is_const_Phi(b)) {
2683 /* check for Div(Const, Phi) */
2684 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2685 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2687 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2688 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2692 else if (is_const_Phi(a) && is_const_Phi(b)) {
2693 /* check for Div(Phi, Phi) */
2694 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2695 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2697 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2698 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2705 if (tb != tarval_bad) {
2706 if (tb == get_mode_one(get_tarval_mode(tb))) {
2708 vb = new_Const(mode, get_mode_null(mode));
2709 DBG_OPT_CSTEVAL(n, vb);
2711 } else if (ta != tarval_bad) {
2712 tarval *resa, *resb;
2713 resa = tarval_div(ta, tb);
2714 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2715 Jmp for X result!? */
2716 resb = tarval_mod(ta, tb);
2717 if (resb == tarval_bad) return n; /* Causes exception! */
2718 va = new_Const(mode, resa);
2719 vb = new_Const(mode, resb);
2720 DBG_OPT_CSTEVAL(n, va);
2721 DBG_OPT_CSTEVAL(n, vb);
2723 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2724 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2725 vb = new_Const(mode, get_mode_null(mode));
2726 DBG_OPT_CSTEVAL(n, va);
2727 DBG_OPT_CSTEVAL(n, vb);
2729 } else { /* Try architecture dependent optimization */
2732 arch_dep_replace_divmod_by_const(&va, &vb, n);
2733 evaluated = va != NULL;
2735 } else if (a == b) {
2736 if (value_not_zero(a, &dummy)) {
2738 va = new_Const(mode, get_mode_one(mode));
2739 vb = new_Const(mode, get_mode_null(mode));
2740 DBG_OPT_CSTEVAL(n, va);
2741 DBG_OPT_CSTEVAL(n, vb);
2744 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2747 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2748 /* 0 / non-Const = 0 */
2753 if (evaluated) { /* replace by tuple */
2757 mem = get_DivMod_mem(n);
2758 /* skip a potential Pin */
2760 mem = get_Pin_op(mem);
2762 blk = get_irn_n(n, -1);
2763 turn_into_tuple(n, pn_DivMod_max);
2764 set_Tuple_pred(n, pn_DivMod_M, mem);
2765 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2766 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2767 set_Tuple_pred(n, pn_DivMod_res_div, va);
2768 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2772 } /* transform_node_DivMod */
2775 * Optimize x / c to x * (1/c)
2777 static ir_node *transform_node_Quot(ir_node *n) {
2778 ir_mode *mode = get_Quot_resmode(n);
2781 if (get_mode_arithmetic(mode) == irma_ieee754) {
2782 ir_node *b = get_Quot_right(n);
2785 tarval *tv = get_Const_tarval(b);
2787 tv = tarval_quo(get_mode_one(mode), tv);
2789 /* Do the transformation if the result is either exact or we are not
2790 using strict rules. */
2791 if (tv != tarval_bad &&
2792 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2793 ir_node *blk = get_irn_n(n, -1);
2794 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2795 ir_node *a = get_Quot_left(n);
2796 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2797 ir_node *mem = get_Quot_mem(n);
2799 /* skip a potential Pin */
2801 mem = get_Pin_op(mem);
2802 turn_into_tuple(n, pn_Quot_max);
2803 set_Tuple_pred(n, pn_Quot_M, mem);
2804 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2805 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2806 set_Tuple_pred(n, pn_Quot_res, m);
2807 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2812 } /* transform_node_Quot */
2815 * Optimize Abs(x) into x if x is Confirmed >= 0
2816 * Optimize Abs(x) into -x if x is Confirmed <= 0
2817 * Optimize Abs(-x) int Abs(x)
2819 static ir_node *transform_node_Abs(ir_node *n) {
2820 ir_node *c, *oldn = n;
2821 ir_node *a = get_Abs_op(n);
2824 HANDLE_UNOP_PHI(tarval_abs, a, c);
2826 switch (classify_value_sign(a)) {
2827 case value_classified_negative:
2828 mode = get_irn_mode(n);
2831 * We can replace the Abs by -x here.
2832 * We even could add a new Confirm here.
2834 * Note that -x would create a new node, so we could
2835 * not run it in the equivalent_node() context.
2837 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2838 get_nodes_block(n), a, mode);
2840 DBG_OPT_CONFIRM(oldn, n);
2842 case value_classified_positive:
2843 /* n is positive, Abs is not needed */
2846 DBG_OPT_CONFIRM(oldn, n);
2852 /* Abs(-x) = Abs(x) */
2853 mode = get_irn_mode(n);
2854 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2855 get_nodes_block(n), get_Minus_op(a), mode);
2856 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2860 } /* transform_node_Abs */
2863 * Transform a Cond node.
2865 * Replace the Cond by a Jmp if it branches on a constant
2868 static ir_node *transform_node_Cond(ir_node *n) {
2871 ir_node *a = get_Cond_selector(n);
2872 tarval *ta = value_of(a);
2874 /* we need block info which is not available in floating irgs */
2875 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2878 if ((ta != tarval_bad) &&
2879 (get_irn_mode(a) == mode_b) &&
2880 (get_opt_unreachable_code())) {
2881 /* It's a boolean Cond, branching on a boolean constant.
2882 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2883 ir_node *blk = get_nodes_block(n);
2884 jmp = new_r_Jmp(current_ir_graph, blk);
2885 turn_into_tuple(n, pn_Cond_max);
2886 if (ta == tarval_b_true) {
2887 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2888 set_Tuple_pred(n, pn_Cond_true, jmp);
2890 set_Tuple_pred(n, pn_Cond_false, jmp);
2891 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2893 /* We might generate an endless loop, so keep it alive. */
2894 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2897 } /* transform_node_Cond */
2900 * Prototype of a recursive transform function
2901 * for bitwise distributive transformations.
2903 typedef ir_node* (*recursive_transform)(ir_node *n);
2906 * makes use of distributive laws for and, or, eor
2907 * and(a OP c, b OP c) -> and(a, b) OP c
2908 * note, might return a different op than n
2910 static ir_node *transform_bitwise_distributive(ir_node *n,
2911 recursive_transform trans_func)
2914 ir_node *a = get_binop_left(n);
2915 ir_node *b = get_binop_right(n);
2916 ir_op *op = get_irn_op(a);
2917 ir_op *op_root = get_irn_op(n);
2919 if(op != get_irn_op(b))
2922 if (op == op_Conv) {
2923 ir_node *a_op = get_Conv_op(a);
2924 ir_node *b_op = get_Conv_op(b);
2925 ir_mode *a_mode = get_irn_mode(a_op);
2926 ir_mode *b_mode = get_irn_mode(b_op);
2927 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2928 ir_node *blk = get_irn_n(n, -1);
2931 set_binop_left(n, a_op);
2932 set_binop_right(n, b_op);
2933 set_irn_mode(n, a_mode);
2935 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2937 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2943 /* nothing to gain here */
2947 if (op == op_Shrs || op == op_Shr || op == op_Shl
2948 || op == op_And || op == op_Or || op == op_Eor) {
2949 ir_node *a_left = get_binop_left(a);
2950 ir_node *a_right = get_binop_right(a);
2951 ir_node *b_left = get_binop_left(b);
2952 ir_node *b_right = get_binop_right(b);
2954 ir_node *op1 = NULL;
2955 ir_node *op2 = NULL;
2957 if (is_op_commutative(op)) {
2958 if (a_left == b_left) {
2962 } else if(a_left == b_right) {
2966 } else if(a_right == b_left) {
2972 if(a_right == b_right) {
2979 /* (a sop c) & (b sop c) => (a & b) sop c */
2980 ir_node *blk = get_irn_n(n, -1);
2982 ir_node *new_n = exact_copy(n);
2983 set_binop_left(new_n, op1);
2984 set_binop_right(new_n, op2);
2985 new_n = trans_func(new_n);
2987 if(op_root == op_Eor && op == op_Or) {
2988 dbg_info *dbgi = get_irn_dbg_info(n);
2989 ir_graph *irg = current_ir_graph;
2990 ir_mode *mode = get_irn_mode(c);
2992 c = new_rd_Not(dbgi, irg, blk, c, mode);
2993 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
2996 set_nodes_block(n, blk);
2997 set_binop_left(n, new_n);
2998 set_binop_right(n, c);
2999 add_identities(current_ir_graph->value_table, n);
3002 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3013 static ir_node *transform_node_And(ir_node *n) {
3014 ir_node *c, *oldn = n;
3015 ir_node *a = get_And_left(n);
3016 ir_node *b = get_And_right(n);
3019 mode = get_irn_mode(n);
3020 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3022 /* we can evaluate 2 Projs of the same Cmp */
3023 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3024 ir_node *pred_a = get_Proj_pred(a);
3025 ir_node *pred_b = get_Proj_pred(b);
3026 if (pred_a == pred_b) {
3027 dbg_info *dbgi = get_irn_dbg_info(n);
3028 ir_node *block = get_nodes_block(pred_a);
3029 pn_Cmp pn_a = get_Proj_proj(a);
3030 pn_Cmp pn_b = get_Proj_proj(b);
3031 /* yes, we can simply calculate with pncs */
3032 pn_Cmp new_pnc = pn_a & pn_b;
3034 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3039 ir_node *op = get_Not_op(b);
3041 ir_node *ba = get_And_left(op);
3042 ir_node *bb = get_And_right(op);
3044 /* it's enough to test the following cases due to normalization! */
3045 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3046 /* (a|b) & ~(a&b) = a^b */
3047 ir_node *block = get_nodes_block(n);
3049 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3050 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3058 ir_node *op = get_Not_op(a);
3060 ir_node *aa = get_And_left(op);
3061 ir_node *ab = get_And_right(op);
3063 /* it's enough to test the following cases due to normalization! */
3064 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3065 /* (a|b) & ~(a&b) = a^b */
3066 ir_node *block = get_nodes_block(n);
3068 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3069 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3076 ir_node *al = get_Eor_left(a);
3077 ir_node *ar = get_Eor_right(a);
3080 /* (b ^ a) & b -> ~a & b */
3081 dbg_info *dbg = get_irn_dbg_info(n);
3082 ir_node *block = get_nodes_block(n);
3084 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3085 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3086 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3090 /* (a ^ b) & b -> ~a & b */
3091 dbg_info *dbg = get_irn_dbg_info(n);
3092 ir_node *block = get_nodes_block(n);
3094 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3095 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3096 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3101 ir_node *bl = get_Eor_left(b);
3102 ir_node *br = get_Eor_right(b);
3105 /* a & (a ^ b) -> a & ~b */
3106 dbg_info *dbg = get_irn_dbg_info(n);
3107 ir_node *block = get_nodes_block(n);
3109 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3110 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3111 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3115 /* a & (b ^ a) -> a & ~b */
3116 dbg_info *dbg = get_irn_dbg_info(n);
3117 ir_node *block = get_nodes_block(n);
3119 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3120 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3121 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3125 if (is_Not(a) && is_Not(b)) {
3126 /* ~a & ~b = ~(a|b) */
3127 ir_node *block = get_nodes_block(n);
3128 ir_mode *mode = get_irn_mode(n);
3132 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3133 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3134 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3138 n = transform_bitwise_distributive(n, transform_node_And);
3141 } /* transform_node_And */
3146 static ir_node *transform_node_Eor(ir_node *n) {
3147 ir_node *c, *oldn = n;
3148 ir_node *a = get_Eor_left(n);
3149 ir_node *b = get_Eor_right(n);
3150 ir_mode *mode = get_irn_mode(n);
3152 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3154 /* we can evaluate 2 Projs of the same Cmp */
3155 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3156 ir_node *pred_a = get_Proj_pred(a);
3157 ir_node *pred_b = get_Proj_pred(b);
3158 if(pred_a == pred_b) {
3159 dbg_info *dbgi = get_irn_dbg_info(n);
3160 ir_node *block = get_nodes_block(pred_a);
3161 pn_Cmp pn_a = get_Proj_proj(a);
3162 pn_Cmp pn_b = get_Proj_proj(b);
3163 /* yes, we can simply calculate with pncs */
3164 pn_Cmp new_pnc = pn_a ^ pn_b;
3166 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3173 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3174 mode, get_mode_null(mode));
3175 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3176 } else if (mode == mode_b &&
3178 is_Const(b) && is_Const_one(b) &&
3179 is_Cmp(get_Proj_pred(a))) {
3180 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3181 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3182 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3184 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3185 } else if (is_Const(b)) {
3186 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3187 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3188 ir_node *not_op = get_Not_op(a);
3189 dbg_info *dbg = get_irn_dbg_info(n);
3190 ir_graph *irg = current_ir_graph;
3191 ir_node *block = get_nodes_block(n);
3192 ir_mode *mode = get_irn_mode(n);
3193 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3195 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3196 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3197 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3200 n = transform_bitwise_distributive(n, transform_node_Eor);
3204 } /* transform_node_Eor */
3209 static ir_node *transform_node_Not(ir_node *n) {
3210 ir_node *c, *oldn = n;
3211 ir_node *a = get_Not_op(n);
3212 ir_mode *mode = get_irn_mode(n);
3214 HANDLE_UNOP_PHI(tarval_not,a,c);
3216 /* check for a boolean Not */
3217 if (mode == mode_b &&
3219 is_Cmp(get_Proj_pred(a))) {
3220 /* We negate a Cmp. The Cmp has the negated result anyways! */
3221 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3222 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3223 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3227 ir_node *eor_b = get_Eor_right(a);
3228 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3229 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3230 ir_node *eor_a = get_Eor_left(a);
3231 dbg_info *dbg = get_irn_dbg_info(n);
3232 ir_graph *irg = current_ir_graph;
3233 ir_node *block = get_nodes_block(n);
3234 ir_mode *mode = get_irn_mode(n);
3235 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3239 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3240 if (is_Minus(a)) { /* ~-x -> x + -1 */
3241 dbg_info *dbg = get_irn_dbg_info(n);
3242 ir_graph *irg = current_ir_graph;
3243 ir_node *block = get_nodes_block(n);
3244 ir_node *add_l = get_Minus_op(a);
3245 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3246 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3247 } else if (is_Add(a)) {
3248 ir_node *add_r = get_Add_right(a);
3249 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3250 /* ~(x + -1) = -x */
3251 ir_node *op = get_Add_left(a);
3252 ir_node *blk = get_irn_n(n, -1);
3253 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3254 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3259 } /* transform_node_Not */
3262 * Transform a Minus.
3266 * -(a >>u (size-1)) = a >>s (size-1)
3267 * -(a >>s (size-1)) = a >>u (size-1)
3268 * -(a * const) -> a * -const
3270 static ir_node *transform_node_Minus(ir_node *n) {
3271 ir_node *c, *oldn = n;
3272 ir_node *a = get_Minus_op(n);
3275 HANDLE_UNOP_PHI(tarval_neg,a,c);
3277 mode = get_irn_mode(a);
3278 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3279 /* the following rules are only to twos-complement */
3282 ir_node *op = get_Not_op(a);
3283 tarval *tv = get_mode_one(mode);
3284 ir_node *blk = get_irn_n(n, -1);
3285 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3286 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3287 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3291 ir_node *c = get_Shr_right(a);
3294 tarval *tv = get_Const_tarval(c);
3296 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3297 /* -(a >>u (size-1)) = a >>s (size-1) */
3298 ir_node *v = get_Shr_left(a);
3300 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3301 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3307 ir_node *c = get_Shrs_right(a);
3310 tarval *tv = get_Const_tarval(c);
3312 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3313 /* -(a >>s (size-1)) = a >>u (size-1) */
3314 ir_node *v = get_Shrs_left(a);
3316 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3317 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3324 /* - (a-b) = b - a */
3325 ir_node *la = get_Sub_left(a);
3326 ir_node *ra = get_Sub_right(a);
3327 ir_node *blk = get_irn_n(n, -1);
3329 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3330 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3334 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3335 ir_node *mul_l = get_Mul_left(a);
3336 ir_node *mul_r = get_Mul_right(a);
3337 if (is_Const(mul_r)) {
3338 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3339 if(tv != tarval_bad) {
3340 ir_node *cnst = new_Const(mode, tv);
3341 dbg_info *dbg = get_irn_dbg_info(a);
3342 ir_graph *irg = current_ir_graph;
3343 ir_node *block = get_nodes_block(a);
3344 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3345 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3352 } /* transform_node_Minus */
3355 * Transform a Cast_type(Const) into a new Const_type
3357 static ir_node *transform_node_Cast(ir_node *n) {
3359 ir_node *pred = get_Cast_op(n);
3360 ir_type *tp = get_irn_type(n);
3362 if (is_Const(pred) && get_Const_type(pred) != tp) {
3363 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3364 get_Const_tarval(pred), tp);
3365 DBG_OPT_CSTEVAL(oldn, n);
3366 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3367 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3368 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3369 DBG_OPT_CSTEVAL(oldn, n);
3373 } /* transform_node_Cast */
3376 * Transform a Proj(Div) with a non-zero value.
3377 * Removes the exceptions and routes the memory to the NoMem node.
3379 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3380 ir_node *div = get_Proj_pred(proj);
3381 ir_node *b = get_Div_right(div);
3382 ir_node *confirm, *res, *new_mem;
3385 if (value_not_zero(b, &confirm)) {
3386 /* div(x, y) && y != 0 */
3387 if (confirm == NULL) {
3388 /* we are sure we have a Const != 0 */
3389 new_mem = get_Div_mem(div);
3390 if (is_Pin(new_mem))
3391 new_mem = get_Pin_op(new_mem);
3392 set_Div_mem(div, new_mem);
3393 set_irn_pinned(div, op_pin_state_floats);
3396 proj_nr = get_Proj_proj(proj);
3398 case pn_Div_X_regular:
3399 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3401 case pn_Div_X_except:
3402 /* we found an exception handler, remove it */
3403 DBG_OPT_EXC_REM(proj);
3407 res = get_Div_mem(div);
3408 new_mem = get_irg_no_mem(current_ir_graph);
3411 /* This node can only float up to the Confirm block */
3412 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3414 set_irn_pinned(div, op_pin_state_floats);
3415 /* this is a Div without exception, we can remove the memory edge */
3416 set_Div_mem(div, new_mem);
3421 } /* transform_node_Proj_Div */
3424 * Transform a Proj(Mod) with a non-zero value.
3425 * Removes the exceptions and routes the memory to the NoMem node.
3427 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3428 ir_node *mod = get_Proj_pred(proj);
3429 ir_node *b = get_Mod_right(mod);
3430 ir_node *confirm, *res, *new_mem;
3433 if (value_not_zero(b, &confirm)) {
3434 /* mod(x, y) && y != 0 */
3435 proj_nr = get_Proj_proj(proj);
3437 if (confirm == NULL) {
3438 /* we are sure we have a Const != 0 */
3439 new_mem = get_Mod_mem(mod);
3440 if (is_Pin(new_mem))
3441 new_mem = get_Pin_op(new_mem);
3442 set_Mod_mem(mod, new_mem);
3443 set_irn_pinned(mod, op_pin_state_floats);
3448 case pn_Mod_X_regular:
3449 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3451 case pn_Mod_X_except:
3452 /* we found an exception handler, remove it */
3453 DBG_OPT_EXC_REM(proj);
3457 res = get_Mod_mem(mod);
3458 new_mem = get_irg_no_mem(current_ir_graph);
3461 /* This node can only float up to the Confirm block */
3462 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3464 /* this is a Mod without exception, we can remove the memory edge */
3465 set_Mod_mem(mod, new_mem);
3468 if (get_Mod_left(mod) == b) {
3469 /* a % a = 0 if a != 0 */
3470 ir_mode *mode = get_irn_mode(proj);
3471 ir_node *res = new_Const(mode, get_mode_null(mode));
3473 DBG_OPT_CSTEVAL(mod, res);
3479 } /* transform_node_Proj_Mod */
3482 * Transform a Proj(DivMod) with a non-zero value.
3483 * Removes the exceptions and routes the memory to the NoMem node.
3485 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3486 ir_node *divmod = get_Proj_pred(proj);
3487 ir_node *b = get_DivMod_right(divmod);
3488 ir_node *confirm, *res, *new_mem;
3491 if (value_not_zero(b, &confirm)) {
3492 /* DivMod(x, y) && y != 0 */
3493 proj_nr = get_Proj_proj(proj);
3495 if (confirm == NULL) {
3496 /* we are sure we have a Const != 0 */
3497 new_mem = get_DivMod_mem(divmod);
3498 if (is_Pin(new_mem))
3499 new_mem = get_Pin_op(new_mem);
3500 set_DivMod_mem(divmod, new_mem);
3501 set_irn_pinned(divmod, op_pin_state_floats);
3506 case pn_DivMod_X_regular:
3507 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3509 case pn_DivMod_X_except:
3510 /* we found an exception handler, remove it */
3511 DBG_OPT_EXC_REM(proj);
3515 res = get_DivMod_mem(divmod);
3516 new_mem = get_irg_no_mem(current_ir_graph);
3519 /* This node can only float up to the Confirm block */
3520 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3522 /* this is a DivMod without exception, we can remove the memory edge */
3523 set_DivMod_mem(divmod, new_mem);
3526 case pn_DivMod_res_mod:
3527 if (get_DivMod_left(divmod) == b) {
3528 /* a % a = 0 if a != 0 */
3529 ir_mode *mode = get_irn_mode(proj);
3530 ir_node *res = new_Const(mode, get_mode_null(mode));
3532 DBG_OPT_CSTEVAL(divmod, res);
3538 } /* transform_node_Proj_DivMod */
3541 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3543 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3544 if (get_opt_unreachable_code()) {
3545 ir_node *n = get_Proj_pred(proj);
3546 ir_node *b = get_Cond_selector(n);
3548 if (mode_is_int(get_irn_mode(b))) {
3549 tarval *tb = value_of(b);
3551 if (tb != tarval_bad) {
3552 /* we have a constant switch */
3553 long num = get_Proj_proj(proj);
3555 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3556 if (get_tarval_long(tb) == num) {
3557 /* Do NOT create a jump here, or we will have 2 control flow ops
3558 * in a block. This case is optimized away in optimize_cf(). */
3561 /* this case will NEVER be taken, kill it */
3569 } /* transform_node_Proj_Cond */
3572 * Create a 0 constant of given mode.
3574 static ir_node *create_zero_const(ir_mode *mode) {
3575 tarval *tv = get_mode_null(mode);
3576 ir_node *cnst = new_Const(mode, tv);
3581 /* the order of the values is important! */
3582 typedef enum const_class {
3588 static const_class classify_const(const ir_node* n)
3590 if (is_Const(n)) return const_const;
3591 if (is_irn_constlike(n)) return const_like;
3596 * Determines whether r is more constlike or has a larger index (in that order)
3599 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3601 const const_class l_order = classify_const(l);
3602 const const_class r_order = classify_const(r);
3604 l_order > r_order ||
3605 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3609 * Normalizes and optimizes Cmp nodes.
3611 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3612 ir_node *n = get_Proj_pred(proj);
3613 ir_node *left = get_Cmp_left(n);
3614 ir_node *right = get_Cmp_right(n);
3617 ir_mode *mode = NULL;
3618 long proj_nr = get_Proj_proj(proj);
3620 /* we can evaluate some cases directly */
3623 return new_Const(mode_b, get_tarval_b_false());
3625 return new_Const(mode_b, get_tarval_b_true());
3627 if (!mode_is_float(get_irn_mode(left)))
3628 return new_Const(mode_b, get_tarval_b_true());
3636 left = get_Cast_op(left);
3638 right = get_Cast_op(right);
3640 /* Remove unnecessary conversions */
3641 /* TODO handle constants */
3642 if (is_Conv(left) && is_Conv(right)) {
3643 ir_mode *mode = get_irn_mode(left);
3644 ir_node *op_left = get_Conv_op(left);
3645 ir_node *op_right = get_Conv_op(right);
3646 ir_mode *mode_left = get_irn_mode(op_left);
3647 ir_mode *mode_right = get_irn_mode(op_right);
3649 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3650 && mode_left != mode_b && mode_right != mode_b) {
3651 ir_graph *irg = current_ir_graph;
3652 ir_node *block = get_nodes_block(n);
3654 if (mode_left == mode_right) {
3658 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3659 } else if (smaller_mode(mode_left, mode_right)) {
3660 left = new_r_Conv(irg, block, op_left, mode_right);
3663 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3664 } else if (smaller_mode(mode_right, mode_left)) {
3666 right = new_r_Conv(irg, block, op_right, mode_left);
3668 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3673 /* remove operation of both sides if possible */
3674 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3676 * The following operations are NOT safe for floating point operations, for instance
3677 * 1.0 + inf == 2.0 + inf, =/=> x == y
3679 if (mode_is_int(get_irn_mode(left))) {
3680 unsigned lop = get_irn_opcode(left);
3682 if (lop == get_irn_opcode(right)) {
3683 ir_node *ll, *lr, *rl, *rr;
3685 /* same operation on both sides, try to remove */
3689 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3690 left = get_unop_op(left);
3691 right = get_unop_op(right);
3693 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3696 ll = get_Add_left(left);
3697 lr = get_Add_right(left);
3698 rl = get_Add_left(right);
3699 rr = get_Add_right(right);
3702 /* X + a CMP X + b ==> a CMP b */
3706 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3707 } else if (ll == rr) {
3708 /* X + a CMP b + X ==> a CMP b */
3712 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3713 } else if (lr == rl) {
3714 /* a + X CMP X + b ==> a CMP b */
3718 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3719 } else if (lr == rr) {
3720 /* a + X CMP b + X ==> a CMP b */
3724 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3728 ll = get_Sub_left(left);
3729 lr = get_Sub_right(left);
3730 rl = get_Sub_left(right);
3731 rr = get_Sub_right(right);
3734 /* X - a CMP X - b ==> a CMP b */
3738 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3739 } else if (lr == rr) {
3740 /* a - X CMP b - X ==> a CMP b */
3744 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3748 if (get_Rot_right(left) == get_Rot_right(right)) {
3749 /* a ROT X CMP b ROT X ==> a CMP b */
3750 left = get_Rot_left(left);
3751 right = get_Rot_left(right);
3753 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3761 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3762 if (is_Add(left) || is_Sub(left)) {
3763 ir_node *ll = get_binop_left(left);
3764 ir_node *lr = get_binop_right(left);
3766 if (lr == right && is_Add(left)) {
3773 right = create_zero_const(get_irn_mode(left));
3775 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3778 if (is_Add(right) || is_Sub(right)) {
3779 ir_node *rl = get_binop_left(right);
3780 ir_node *rr = get_binop_right(right);
3782 if (rr == left && is_Add(right)) {
3789 right = create_zero_const(get_irn_mode(left));
3791 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3794 } /* mode_is_int(...) */
3795 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3797 /* replace mode_b compares with ands/ors */
3798 if (get_irn_mode(left) == mode_b) {
3799 ir_graph *irg = current_ir_graph;
3800 ir_node *block = get_nodes_block(n);
3804 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3805 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3806 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3807 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3808 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3809 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3810 default: bres = NULL;
3813 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3819 * First step: normalize the compare op
3820 * by placing the constant on the right side
3821 * or moving the lower address node to the left.
3823 if (!operands_are_normalized(left, right)) {
3829 proj_nr = get_inversed_pnc(proj_nr);
3834 * Second step: Try to reduce the magnitude
3835 * of a constant. This may help to generate better code
3836 * later and may help to normalize more compares.
3837 * Of course this is only possible for integer values.
3839 if (is_Const(right)) {
3840 mode = get_irn_mode(right);
3841 tv = get_Const_tarval(right);
3843 /* TODO extend to arbitrary constants */
3844 if (is_Conv(left) && tarval_is_null(tv)) {
3845 ir_node *op = get_Conv_op(left);
3846 ir_mode *op_mode = get_irn_mode(op);
3849 * UpConv(x) REL 0 ==> x REL 0
3851 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3852 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3853 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3854 tv = get_mode_null(op_mode);
3858 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3862 if (tv != tarval_bad) {
3863 /* the following optimization is possible on modes without Overflow
3864 * on Unary Minus or on == and !=:
3865 * -a CMP c ==> a swap(CMP) -c
3867 * Beware: for two-complement Overflow may occur, so only == and != can
3868 * be optimized, see this:
3869 * -MININT < 0 =/=> MININT > 0 !!!
3871 if (is_Minus(left) &&
3872 (!mode_overflow_on_unary_Minus(mode) ||
3873 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3874 tv = tarval_neg(tv);
3876 if (tv != tarval_bad) {
3877 left = get_Minus_op(left);
3878 proj_nr = get_inversed_pnc(proj_nr);
3880 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3882 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3883 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3884 tv = tarval_not(tv);
3886 if (tv != tarval_bad) {
3887 left = get_Not_op(left);
3889 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3893 /* for integer modes, we have more */
3894 if (mode_is_int(mode)) {
3895 /* Ne includes Unordered which is not possible on integers.
3896 * However, frontends often use this wrong, so fix it here */
3897 if (proj_nr & pn_Cmp_Uo) {
3898 proj_nr &= ~pn_Cmp_Uo;
3899 set_Proj_proj(proj, proj_nr);
3902 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3903 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3904 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3905 tv = tarval_sub(tv, get_mode_one(mode));
3907 if (tv != tarval_bad) {
3908 proj_nr ^= pn_Cmp_Eq;
3910 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3913 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3914 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3915 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3916 tv = tarval_add(tv, get_mode_one(mode));
3918 if (tv != tarval_bad) {
3919 proj_nr ^= pn_Cmp_Eq;
3921 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3925 /* the following reassociations work only for == and != */
3926 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3928 #if 0 /* Might be not that good in general */
3929 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
3930 if (tarval_is_null(tv) && is_Sub(left)) {
3931 right = get_Sub_right(left);
3932 left = get_Sub_left(left);
3934 tv = value_of(right);
3936 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3940 if (tv != tarval_bad) {
3941 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
3943 ir_node *c1 = get_Sub_right(left);
3944 tarval *tv2 = value_of(c1);
3946 if (tv2 != tarval_bad) {
3947 tv2 = tarval_add(tv, value_of(c1));
3949 if (tv2 != tarval_bad) {
3950 left = get_Sub_left(left);
3953 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3957 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
3958 else if (is_Add(left)) {
3959 ir_node *a_l = get_Add_left(left);
3960 ir_node *a_r = get_Add_right(left);
3964 if (is_Const(a_l)) {
3966 tv2 = value_of(a_l);
3969 tv2 = value_of(a_r);
3972 if (tv2 != tarval_bad) {
3973 tv2 = tarval_sub(tv, tv2);
3975 if (tv2 != tarval_bad) {
3979 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3983 /* -a == c ==> a == -c, -a != c ==> a != -c */
3984 else if (is_Minus(left)) {
3985 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
3987 if (tv2 != tarval_bad) {
3988 left = get_Minus_op(left);
3991 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3996 /* the following reassociations work only for <= */
3997 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3998 if (tv != tarval_bad) {
3999 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4000 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4006 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4007 switch (get_irn_opcode(left)) {
4011 c1 = get_And_right(left);
4014 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4015 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4017 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4019 /* TODO: move to constant evaluation */
4020 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4021 c1 = new_Const(mode_b, tv);
4022 DBG_OPT_CSTEVAL(proj, c1);
4026 if (tarval_is_single_bit(tv)) {
4028 * optimization for AND:
4030 * And(x, C) == C ==> And(x, C) != 0
4031 * And(x, C) != C ==> And(X, C) == 0
4033 * if C is a single Bit constant.
4036 /* check for Constant's match. We have check hare the tarvals,
4037 because our const might be changed */
4038 if (get_Const_tarval(c1) == tv) {
4039 /* fine: do the transformation */
4040 tv = get_mode_null(get_tarval_mode(tv));
4041 proj_nr ^= pn_Cmp_Leg;
4043 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4049 c1 = get_Or_right(left);
4050 if (is_Const(c1) && tarval_is_null(tv)) {
4052 * Or(x, C) == 0 && C != 0 ==> FALSE
4053 * Or(x, C) != 0 && C != 0 ==> TRUE
4055 if (! tarval_is_null(get_Const_tarval(c1))) {
4056 /* TODO: move to constant evaluation */
4057 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4058 c1 = new_Const(mode_b, tv);
4059 DBG_OPT_CSTEVAL(proj, c1);
4066 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4068 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4071 c1 = get_Shl_right(left);
4073 tarval *tv1 = get_Const_tarval(c1);
4074 ir_mode *mode = get_irn_mode(left);
4075 tarval *minus1 = get_mode_all_one(mode);
4076 tarval *amask = tarval_shr(minus1, tv1);
4077 tarval *cmask = tarval_shl(minus1, tv1);
4080 if (tarval_and(tv, cmask) != tv) {
4081 /* condition not met */
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);
4087 sl = get_Shl_left(left);
4088 blk = get_nodes_block(n);
4089 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4090 tv = tarval_shr(tv, tv1);
4092 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4097 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4099 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4102 c1 = get_Shr_right(left);
4104 tarval *tv1 = get_Const_tarval(c1);
4105 ir_mode *mode = get_irn_mode(left);
4106 tarval *minus1 = get_mode_all_one(mode);
4107 tarval *amask = tarval_shl(minus1, tv1);
4108 tarval *cmask = tarval_shr(minus1, tv1);
4111 if (tarval_and(tv, cmask) != tv) {
4112 /* condition not met */
4113 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4114 c1 = new_Const(mode_b, tv);
4115 DBG_OPT_CSTEVAL(proj, c1);
4118 sl = get_Shr_left(left);
4119 blk = get_nodes_block(n);
4120 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4121 tv = tarval_shl(tv, tv1);
4123 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4128 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4130 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4133 c1 = get_Shrs_right(left);
4135 tarval *tv1 = get_Const_tarval(c1);
4136 ir_mode *mode = get_irn_mode(left);
4137 tarval *minus1 = get_mode_all_one(mode);
4138 tarval *amask = tarval_shl(minus1, tv1);
4139 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4142 cond = tarval_sub(cond, tv1);
4143 cond = tarval_shrs(tv, cond);
4145 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4146 /* condition not met */
4147 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4148 c1 = new_Const(mode_b, tv);
4149 DBG_OPT_CSTEVAL(proj, c1);
4152 sl = get_Shrs_left(left);
4153 blk = get_nodes_block(n);
4154 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4155 tv = tarval_shl(tv, tv1);
4157 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4162 } /* tarval != bad */
4165 if (changed & 2) /* need a new Const */
4166 right = new_Const(mode, tv);
4168 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4169 ir_node *op = get_Proj_pred(left);
4171 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4172 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4173 ir_node *c = get_binop_right(op);
4176 tarval *tv = get_Const_tarval(c);
4178 if (tarval_is_single_bit(tv)) {
4179 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4180 ir_node *v = get_binop_left(op);
4181 ir_node *blk = get_irn_n(op, -1);
4182 ir_mode *mode = get_irn_mode(v);
4184 tv = tarval_sub(tv, get_mode_one(mode));
4185 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4187 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4194 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4196 /* create a new compare */
4197 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4198 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4202 } /* transform_node_Proj_Cmp */
4205 * Does all optimizations on nodes that must be done on it's Proj's
4206 * because of creating new nodes.
4208 static ir_node *transform_node_Proj(ir_node *proj) {
4209 ir_node *n = get_Proj_pred(proj);
4211 switch (get_irn_opcode(n)) {
4213 return transform_node_Proj_Div(proj);
4216 return transform_node_Proj_Mod(proj);
4219 return transform_node_Proj_DivMod(proj);
4222 return transform_node_Proj_Cond(proj);
4225 return transform_node_Proj_Cmp(proj);
4228 /* should not happen, but if it does will be optimized away */
4229 return equivalent_node_Proj(proj);
4235 } /* transform_node_Proj */
4238 * Move Confirms down through Phi nodes.
4240 static ir_node *transform_node_Phi(ir_node *phi) {
4242 ir_mode *mode = get_irn_mode(phi);
4244 if (mode_is_reference(mode)) {
4245 n = get_irn_arity(phi);
4247 /* Beware of Phi0 */
4249 ir_node *pred = get_irn_n(phi, 0);
4250 ir_node *bound, *new_Phi, *block, **in;
4253 if (! is_Confirm(pred))
4256 bound = get_Confirm_bound(pred);
4257 pnc = get_Confirm_cmp(pred);
4259 NEW_ARR_A(ir_node *, in, n);
4260 in[0] = get_Confirm_value(pred);
4262 for (i = 1; i < n; ++i) {
4263 pred = get_irn_n(phi, i);
4265 if (! is_Confirm(pred) ||
4266 get_Confirm_bound(pred) != bound ||
4267 get_Confirm_cmp(pred) != pnc)
4269 in[i] = get_Confirm_value(pred);
4271 /* move the Confirm nodes "behind" the Phi */
4272 block = get_irn_n(phi, -1);
4273 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4274 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4278 } /* transform_node_Phi */
4281 * Returns the operands of a commutative bin-op, if one operand is
4282 * a const, it is returned as the second one.
4284 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4285 ir_node *op_a = get_binop_left(binop);
4286 ir_node *op_b = get_binop_right(binop);
4288 assert(is_op_commutative(get_irn_op(binop)));
4290 if (is_Const(op_a)) {
4297 } /* get_comm_Binop_Ops */
4300 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4301 * Such pattern may arise in bitfield stores.
4303 * value c4 value c4 & c2
4304 * AND c3 AND c1 | c3
4311 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4314 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4317 ir_node *and_l, *c3;
4318 ir_node *value, *c4;
4319 ir_node *new_and, *new_const, *block;
4320 ir_mode *mode = get_irn_mode(or);
4322 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4325 get_comm_Binop_Ops(or, &and, &c1);
4326 if (!is_Const(c1) || !is_And(and))
4329 get_comm_Binop_Ops(and, &or_l, &c2);
4333 tv1 = get_Const_tarval(c1);
4334 tv2 = get_Const_tarval(c2);
4336 tv = tarval_or(tv1, tv2);
4337 if (tarval_is_all_one(tv)) {
4338 /* the AND does NOT clear a bit with isn't set by the OR */
4339 set_Or_left(or, or_l);
4340 set_Or_right(or, c1);
4342 /* check for more */
4349 get_comm_Binop_Ops(or_l, &and_l, &c3);
4350 if (!is_Const(c3) || !is_And(and_l))
4353 get_comm_Binop_Ops(and_l, &value, &c4);
4357 /* ok, found the pattern, check for conditions */
4358 assert(mode == get_irn_mode(and));
4359 assert(mode == get_irn_mode(or_l));
4360 assert(mode == get_irn_mode(and_l));
4362 tv3 = get_Const_tarval(c3);
4363 tv4 = get_Const_tarval(c4);
4365 tv = tarval_or(tv4, tv2);
4366 if (!tarval_is_all_one(tv)) {
4367 /* have at least one 0 at the same bit position */
4371 n_tv4 = tarval_not(tv4);
4372 if (tv3 != tarval_and(tv3, n_tv4)) {
4373 /* bit in the or_mask is outside the and_mask */
4377 n_tv2 = tarval_not(tv2);
4378 if (tv1 != tarval_and(tv1, n_tv2)) {
4379 /* bit in the or_mask is outside the and_mask */
4383 /* ok, all conditions met */
4384 block = get_irn_n(or, -1);
4386 new_and = new_r_And(current_ir_graph, block,
4387 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4389 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4391 set_Or_left(or, new_and);
4392 set_Or_right(or, new_const);
4394 /* check for more */
4396 } /* transform_node_Or_bf_store */
4399 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4401 static ir_node *transform_node_Or_Rot(ir_node *or) {
4402 ir_mode *mode = get_irn_mode(or);
4403 ir_node *shl, *shr, *block;
4404 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4407 if (! mode_is_int(mode))
4410 shl = get_binop_left(or);
4411 shr = get_binop_right(or);
4420 } else if (!is_Shl(shl)) {
4422 } else if (!is_Shr(shr)) {
4425 x = get_Shl_left(shl);
4426 if (x != get_Shr_left(shr))
4429 c1 = get_Shl_right(shl);
4430 c2 = get_Shr_right(shr);
4431 if (is_Const(c1) && is_Const(c2)) {
4432 tv1 = get_Const_tarval(c1);
4433 if (! tarval_is_long(tv1))
4436 tv2 = get_Const_tarval(c2);
4437 if (! tarval_is_long(tv2))
4440 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4441 != (int) get_mode_size_bits(mode))
4444 /* yet, condition met */
4445 block = get_irn_n(or, -1);
4447 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4449 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4451 } else if (is_Sub(c1)) {
4455 if (get_Sub_right(sub) != v)
4458 c1 = get_Sub_left(sub);
4462 tv1 = get_Const_tarval(c1);
4463 if (! tarval_is_long(tv1))
4466 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4469 /* yet, condition met */
4470 block = get_nodes_block(or);
4472 /* a Rot right is not supported, so use a rot left */
4473 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4475 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4477 } else if (is_Sub(c2)) {
4481 c1 = get_Sub_left(sub);
4485 tv1 = get_Const_tarval(c1);
4486 if (! tarval_is_long(tv1))
4489 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4492 /* yet, condition met */
4493 block = get_irn_n(or, -1);
4496 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4498 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4503 } /* transform_node_Or_Rot */
4508 static ir_node *transform_node_Or(ir_node *n) {
4509 ir_node *c, *oldn = n;
4510 ir_node *a = get_Or_left(n);
4511 ir_node *b = get_Or_right(n);
4514 if (is_Not(a) && is_Not(b)) {
4515 /* ~a | ~b = ~(a&b) */
4516 ir_node *block = get_nodes_block(n);
4518 mode = get_irn_mode(n);
4521 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4522 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4523 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4527 /* we can evaluate 2 Projs of the same Cmp */
4528 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4529 ir_node *pred_a = get_Proj_pred(a);
4530 ir_node *pred_b = get_Proj_pred(b);
4531 if (pred_a == pred_b) {
4532 dbg_info *dbgi = get_irn_dbg_info(n);
4533 ir_node *block = get_nodes_block(pred_a);
4534 pn_Cmp pn_a = get_Proj_proj(a);
4535 pn_Cmp pn_b = get_Proj_proj(b);
4536 /* yes, we can simply calculate with pncs */
4537 pn_Cmp new_pnc = pn_a | pn_b;
4539 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4544 mode = get_irn_mode(n);
4545 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4547 n = transform_node_Or_bf_store(n);
4548 n = transform_node_Or_Rot(n);
4552 n = transform_bitwise_distributive(n, transform_node_Or);
4555 } /* transform_node_Or */
4559 static ir_node *transform_node(ir_node *n);
4562 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4564 * Should be moved to reassociation?
4566 static ir_node *transform_node_shift(ir_node *n) {
4567 ir_node *left, *right;
4568 tarval *tv1, *tv2, *res;
4570 int modulo_shf, flag;
4572 left = get_binop_left(n);
4574 /* different operations */
4575 if (get_irn_op(left) != get_irn_op(n))
4578 right = get_binop_right(n);
4579 tv1 = value_of(right);
4580 if (tv1 == tarval_bad)
4583 tv2 = value_of(get_binop_right(left));
4584 if (tv2 == tarval_bad)
4587 res = tarval_add(tv1, tv2);
4589 /* beware: a simple replacement works only, if res < modulo shift */
4590 mode = get_irn_mode(n);
4594 modulo_shf = get_mode_modulo_shift(mode);
4595 if (modulo_shf > 0) {
4596 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4598 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4604 /* ok, we can replace it */
4605 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4607 in[0] = get_binop_left(left);
4608 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4610 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4612 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4614 return transform_node(irn);
4617 } /* transform_node_shift */
4622 static ir_node *transform_node_Shr(ir_node *n) {
4623 ir_node *c, *oldn = n;
4624 ir_node *a = get_Shr_left(n);
4625 ir_node *b = get_Shr_right(n);
4626 ir_mode *mode = get_irn_mode(n);
4628 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4629 return transform_node_shift(n);
4630 } /* transform_node_Shr */
4635 static ir_node *transform_node_Shrs(ir_node *n) {
4636 ir_node *c, *oldn = n;
4637 ir_node *a = get_Shrs_left(n);
4638 ir_node *b = get_Shrs_right(n);
4639 ir_mode *mode = get_irn_mode(n);
4641 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4642 return transform_node_shift(n);
4643 } /* transform_node_Shrs */
4648 static ir_node *transform_node_Shl(ir_node *n) {
4649 ir_node *c, *oldn = n;
4650 ir_node *a = get_Shl_left(n);
4651 ir_node *b = get_Shl_right(n);
4652 ir_mode *mode = get_irn_mode(n);
4654 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4655 return transform_node_shift(n);
4656 } /* transform_node_Shl */
4661 static ir_node *transform_node_Rot(ir_node *n) {
4662 ir_node *c, *oldn = n;
4663 ir_node *a = get_Rot_left(n);
4664 ir_node *b = get_Rot_right(n);
4665 ir_mode *mode = get_irn_mode(n);
4667 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4668 return transform_node_shift(n);
4669 } /* transform_node_Rot */
4674 static ir_node *transform_node_Conv(ir_node *n) {
4675 ir_node *c, *oldn = n;
4676 ir_node *a = get_Conv_op(n);
4678 if (is_const_Phi(a)) {
4679 c = apply_conv_on_phi(a, get_irn_mode(n));
4681 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4686 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4687 ir_mode *mode = get_irn_mode(n);
4688 return new_r_Unknown(current_ir_graph, mode);
4692 } /* transform_node_Conv */
4695 * Remove dead blocks and nodes in dead blocks
4696 * in keep alive list. We do not generate a new End node.
4698 static ir_node *transform_node_End(ir_node *n) {
4699 int i, j, n_keepalives = get_End_n_keepalives(n);
4702 NEW_ARR_A(ir_node *, in, n_keepalives);
4704 for (i = j = 0; i < n_keepalives; ++i) {
4705 ir_node *ka = get_End_keepalive(n, i);
4707 if (! is_Block_dead(ka)) {
4711 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4714 /* FIXME: beabi need to keep a Proj(M) */
4715 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4718 if (j != n_keepalives)
4719 set_End_keepalives(n, j, in);
4721 } /* transform_node_End */
4723 /** returns 1 if a == -b */
4724 static int is_negated_value(ir_node *a, ir_node *b) {
4725 if(is_Minus(a) && get_Minus_op(a) == b)
4727 if(is_Minus(b) && get_Minus_op(b) == a)
4729 if(is_Sub(a) && is_Sub(b)) {
4730 ir_node *a_left = get_Sub_left(a);
4731 ir_node *a_right = get_Sub_right(a);
4732 ir_node *b_left = get_Sub_left(b);
4733 ir_node *b_right = get_Sub_right(b);
4735 if(a_left == b_right && a_right == b_left)
4743 * Optimize a Mux into some simpler cases.
4745 static ir_node *transform_node_Mux(ir_node *n) {
4746 ir_node *oldn = n, *sel = get_Mux_sel(n);
4747 ir_mode *mode = get_irn_mode(n);
4749 if (mode == mode_b) {
4750 ir_node *t = get_Mux_true(n);
4751 ir_node *f = get_Mux_false(n);
4752 dbg_info *dbg = get_irn_dbg_info(n);
4753 ir_node *block = get_irn_n(n, -1);
4754 ir_graph *irg = current_ir_graph;
4757 tarval *tv_t = get_Const_tarval(t);
4758 if (tv_t == tarval_b_true) {
4760 assert(get_Const_tarval(f) == tarval_b_false);
4763 return new_rd_Or(dbg, irg, block, sel, f, mode_b);
4766 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4767 assert(tv_t == tarval_b_false);
4769 assert(get_Const_tarval(f) == tarval_b_true);
4772 return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
4775 } else if (is_Const(f)) {
4776 tarval *tv_f = get_Const_tarval(f);
4777 if (tv_f == tarval_b_true) {
4778 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4779 return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4781 assert(tv_f == tarval_b_false);
4782 return new_rd_And(dbg, irg, block, sel, t, mode_b);
4787 if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
4788 ir_node *cmp = get_Proj_pred(sel);
4789 long pn = get_Proj_proj(sel);
4790 ir_node *f = get_Mux_false(n);
4791 ir_node *t = get_Mux_true(n);
4794 * Note: normalization puts the constant on the right side,
4795 * so we check only one case.
4797 * Note further that these optimization work even for floating point
4798 * with NaN's because -NaN == NaN.
4799 * However, if +0 and -0 is handled differently, we cannot use the first
4803 ir_node *cmp_r = get_Cmp_right(cmp);
4804 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4805 ir_node *block = get_irn_n(n, -1);
4807 if(is_negated_value(f, t)) {
4808 ir_node *cmp_left = get_Cmp_left(cmp);
4810 /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
4811 if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4812 || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4814 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4816 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4818 /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
4819 } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4820 || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4822 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4824 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4826 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4833 return arch_transform_node_Mux(n);
4834 } /* transform_node_Mux */
4837 * Optimize a Psi into some simpler cases.
4839 static ir_node *transform_node_Psi(ir_node *n) {
4841 return transform_node_Mux(n);
4844 } /* transform_node_Psi */
4847 * optimize sync nodes that have other syncs as input we simply add the inputs
4848 * of the other sync to our own inputs
4850 static ir_node *transform_node_Sync(ir_node *n) {
4853 arity = get_irn_arity(n);
4854 for(i = 0; i < get_irn_arity(n); /*empty*/) {
4856 ir_node *in = get_irn_n(n, i);
4862 /* set sync input 0 instead of the sync */
4863 set_irn_n(n, i, get_irn_n(in, 0));
4864 /* so we check this input again for syncs */
4866 /* append all other inputs of the sync to our sync */
4867 arity2 = get_irn_arity(in);
4868 for(i2 = 1; i2 < arity2; ++i2) {
4869 ir_node *in_in = get_irn_n(in, i2);
4870 add_irn_n(n, in_in);
4871 /* increase arity so we also check the new inputs for syncs */
4876 /* rehash the sync node */
4877 add_identities(current_ir_graph->value_table, n);
4883 * Tries several [inplace] [optimizing] transformations and returns an
4884 * equivalent node. The difference to equivalent_node() is that these
4885 * transformations _do_ generate new nodes, and thus the old node must
4886 * not be freed even if the equivalent node isn't the old one.
4888 static ir_node *transform_node(ir_node *n) {
4892 * Transform_node is the only "optimizing transformation" that might
4893 * return a node with a different opcode. We iterate HERE until fixpoint
4894 * to get the final result.
4898 if (n->op->ops.transform_node)
4899 n = n->op->ops.transform_node(n);
4900 } while (oldn != n);
4903 } /* transform_node */
4906 * Sets the default transform node operation for an ir_op_ops.
4908 * @param code the opcode for the default operation
4909 * @param ops the operations initialized
4914 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
4918 ops->transform_node = transform_node_##a; \
4955 } /* firm_set_default_transform_node */
4958 /* **************** Common Subexpression Elimination **************** */
4960 /** The size of the hash table used, should estimate the number of nodes
4962 #define N_IR_NODES 512
4964 /** Compares the attributes of two Const nodes. */
4965 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
4966 return (get_Const_tarval(a) != get_Const_tarval(b))
4967 || (get_Const_type(a) != get_Const_type(b));
4968 } /* node_cmp_attr_Const */
4970 /** Compares the attributes of two Proj nodes. */
4971 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
4972 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
4973 } /* node_cmp_attr_Proj */
4975 /** Compares the attributes of two Filter nodes. */
4976 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
4977 return get_Filter_proj(a) != get_Filter_proj(b);
4978 } /* node_cmp_attr_Filter */
4980 /** Compares the attributes of two Alloc nodes. */
4981 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
4982 const alloc_attr *pa = get_irn_alloc_attr(a);
4983 const alloc_attr *pb = get_irn_alloc_attr(b);
4984 return (pa->where != pb->where) || (pa->type != pb->type);
4985 } /* node_cmp_attr_Alloc */
4987 /** Compares the attributes of two Free nodes. */
4988 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
4989 const free_attr *pa = get_irn_free_attr(a);
4990 const free_attr *pb = get_irn_free_attr(b);
4991 return (pa->where != pb->where) || (pa->type != pb->type);
4992 } /* node_cmp_attr_Free */
4994 /** Compares the attributes of two SymConst nodes. */
4995 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
4996 const symconst_attr *pa = get_irn_symconst_attr(a);
4997 const symconst_attr *pb = get_irn_symconst_attr(b);
4998 return (pa->num != pb->num)
4999 || (pa->sym.type_p != pb->sym.type_p)
5000 || (pa->tp != pb->tp);
5001 } /* node_cmp_attr_SymConst */
5003 /** Compares the attributes of two Call nodes. */
5004 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5005 return get_irn_call_attr(a) != get_irn_call_attr(b);
5006 } /* node_cmp_attr_Call */
5008 /** Compares the attributes of two Sel nodes. */
5009 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5010 const ir_entity *a_ent = get_Sel_entity(a);
5011 const ir_entity *b_ent = get_Sel_entity(b);
5013 (a_ent->kind != b_ent->kind) ||
5014 (a_ent->name != b_ent->name) ||
5015 (a_ent->owner != b_ent->owner) ||
5016 (a_ent->ld_name != b_ent->ld_name) ||
5017 (a_ent->type != b_ent->type);
5018 } /* node_cmp_attr_Sel */
5020 /** Compares the attributes of two Phi nodes. */
5021 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5022 /* we can only enter this function if both nodes have the same number of inputs,
5023 hence it is enough to check if one of them is a Phi0 */
5025 /* check the Phi0 pos attribute */
5026 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5029 } /* node_cmp_attr_Phi */
5031 /** Compares the attributes of two Conv nodes. */
5032 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5033 return get_Conv_strict(a) != get_Conv_strict(b);
5034 } /* node_cmp_attr_Conv */
5036 /** Compares the attributes of two Cast nodes. */
5037 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5038 return get_Cast_type(a) != get_Cast_type(b);
5039 } /* node_cmp_attr_Cast */
5041 /** Compares the attributes of two Load nodes. */
5042 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5043 if (get_Load_volatility(a) == volatility_is_volatile ||
5044 get_Load_volatility(b) == volatility_is_volatile)
5045 /* NEVER do CSE on volatile Loads */
5047 /* do not CSE Loads with different alignment. Be conservative. */
5048 if (get_Load_align(a) != get_Load_align(b))
5051 return get_Load_mode(a) != get_Load_mode(b);
5052 } /* node_cmp_attr_Load */
5054 /** Compares the attributes of two Store nodes. */
5055 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5056 /* do not CSE Stores with different alignment. Be conservative. */
5057 if (get_Store_align(a) != get_Store_align(b))
5060 /* NEVER do CSE on volatile Stores */
5061 return (get_Store_volatility(a) == volatility_is_volatile ||
5062 get_Store_volatility(b) == volatility_is_volatile);
5063 } /* node_cmp_attr_Store */
5065 /** Compares two exception attributes */
5066 static int node_cmp_exception(ir_node *a, ir_node *b) {
5067 const except_attr *ea = get_irn_except_attr(a);
5068 const except_attr *eb = get_irn_except_attr(b);
5070 return ea->pin_state != eb->pin_state;
5073 #define node_cmp_attr_Bound node_cmp_exception
5075 /** Compares the attributes of two Div nodes. */
5076 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5077 const divmod_attr *ma = get_irn_divmod_attr(a);
5078 const divmod_attr *mb = get_irn_divmod_attr(b);
5079 return ma->exc.pin_state != mb->exc.pin_state ||
5080 ma->res_mode != mb->res_mode ||
5081 ma->no_remainder != mb->no_remainder;
5082 } /* node_cmp_attr_Div */
5084 /** Compares the attributes of two DivMod nodes. */
5085 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5086 const divmod_attr *ma = get_irn_divmod_attr(a);
5087 const divmod_attr *mb = get_irn_divmod_attr(b);
5088 return ma->exc.pin_state != mb->exc.pin_state ||
5089 ma->res_mode != mb->res_mode;
5090 } /* node_cmp_attr_DivMod */
5092 /** Compares the attributes of two Mod nodes. */
5093 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5094 const divmod_attr *ma = get_irn_divmod_attr(a);
5095 const divmod_attr *mb = get_irn_divmod_attr(b);
5096 return ma->exc.pin_state != mb->exc.pin_state ||
5097 ma->res_mode != mb->res_mode;
5098 } /* node_cmp_attr_Mod */
5100 /** Compares the attributes of two Quot nodes. */
5101 static int node_cmp_attr_Quot(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 } /* node_cmp_attr_Quot */
5108 /** Compares the attributes of two Confirm nodes. */
5109 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5110 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5111 } /* node_cmp_attr_Confirm */
5113 /** Compares the attributes of two ASM nodes. */
5114 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5116 const ir_asm_constraint *ca;
5117 const ir_asm_constraint *cb;
5120 if (get_ASM_text(a) != get_ASM_text(b))
5123 /* Should we really check the constraints here? Should be better, but is strange. */
5124 n = get_ASM_n_input_constraints(a);
5125 if (n != get_ASM_n_input_constraints(b))
5128 ca = get_ASM_input_constraints(a);
5129 cb = get_ASM_input_constraints(b);
5130 for (i = 0; i < n; ++i) {
5131 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5135 n = get_ASM_n_output_constraints(a);
5136 if (n != get_ASM_n_output_constraints(b))
5139 ca = get_ASM_output_constraints(a);
5140 cb = get_ASM_output_constraints(b);
5141 for (i = 0; i < n; ++i) {
5142 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5146 n = get_ASM_n_clobbers(a);
5147 if (n != get_ASM_n_clobbers(b))
5150 cla = get_ASM_clobbers(a);
5151 clb = get_ASM_clobbers(b);
5152 for (i = 0; i < n; ++i) {
5153 if (cla[i] != clb[i])
5157 } /* node_cmp_attr_ASM */
5160 * Set the default node attribute compare operation for an ir_op_ops.
5162 * @param code the opcode for the default operation
5163 * @param ops the operations initialized
5168 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5172 ops->node_cmp_attr = node_cmp_attr_##a; \
5203 } /* firm_set_default_node_cmp_attr */
5206 * Compare function for two nodes in the value table. Gets two
5207 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5209 int identities_cmp(const void *elt, const void *key) {
5210 ir_node *a = (ir_node *)elt;
5211 ir_node *b = (ir_node *)key;
5214 if (a == b) return 0;
5216 if ((get_irn_op(a) != get_irn_op(b)) ||
5217 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5219 /* compare if a's in and b's in are of equal length */
5220 irn_arity_a = get_irn_intra_arity(a);
5221 if (irn_arity_a != get_irn_intra_arity(b))
5224 if (get_irn_pinned(a) == op_pin_state_pinned) {
5225 /* for pinned nodes, the block inputs must be equal */
5226 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5228 } else if (! get_opt_global_cse()) {
5229 /* for block-local CSE both nodes must be in the same MacroBlock */
5230 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5234 /* compare a->in[0..ins] with b->in[0..ins] */
5235 for (i = 0; i < irn_arity_a; i++)
5236 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5240 * here, we already now that the nodes are identical except their
5243 if (a->op->ops.node_cmp_attr)
5244 return a->op->ops.node_cmp_attr(a, b);
5247 } /* identities_cmp */
5250 * Calculate a hash value of a node.
5252 unsigned ir_node_hash(ir_node *node) {
5256 if (node->op == op_Const) {
5257 /* special value for const, as they only differ in their tarval. */
5258 h = HASH_PTR(node->attr.con.tv);
5259 h = 9*h + HASH_PTR(get_irn_mode(node));
5260 } else if (node->op == op_SymConst) {
5261 /* special value for const, as they only differ in their symbol. */
5262 h = HASH_PTR(node->attr.symc.sym.type_p);
5263 h = 9*h + HASH_PTR(get_irn_mode(node));
5266 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5267 h = irn_arity = get_irn_intra_arity(node);
5269 /* consider all in nodes... except the block if not a control flow. */
5270 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5271 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5275 h = 9*h + HASH_PTR(get_irn_mode(node));
5277 h = 9*h + HASH_PTR(get_irn_op(node));
5281 } /* ir_node_hash */
5283 pset *new_identities(void) {
5284 return new_pset(identities_cmp, N_IR_NODES);
5285 } /* new_identities */
5287 void del_identities(pset *value_table) {
5288 del_pset(value_table);
5289 } /* del_identities */
5292 * Normalize a node by putting constants (and operands with larger
5293 * node index) on the right (operator side).
5295 * @param n The node to normalize
5297 static void normalize_node(ir_node *n) {
5298 if (is_op_commutative(get_irn_op(n))) {
5299 ir_node *l = get_binop_left(n);
5300 ir_node *r = get_binop_right(n);
5302 /* For commutative operators perform a OP b == b OP a but keep
5303 * constants on the RIGHT side. This helps greatly in some
5304 * optimizations. Moreover we use the idx number to make the form
5306 if (!operands_are_normalized(l, r)) {
5307 set_binop_left(n, r);
5308 set_binop_right(n, l);
5311 } /* normalize_node */
5314 * Update the nodes after a match in the value table. If both nodes have
5315 * the same MacroBlock but different Blocks, we must ensure that the node
5316 * with the dominating Block (the node that is near to the MacroBlock header
5317 * is stored in the table.
5318 * Because a MacroBlock has only one "non-exception" flow, we don't need
5319 * dominance info here: We known, that one block must dominate the other and
5320 * following the only block input will allow to find it.
5322 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5323 ir_node *known_blk, *new_block, *block, *mbh;
5325 if (get_opt_global_cse()) {
5326 /* Block inputs are meaning less */
5329 known_blk = get_irn_n(known_irn, -1);
5330 new_block = get_irn_n(new_ir_node, -1);
5331 if (known_blk == new_block) {
5332 /* already in the same block */
5336 * We expect the typical case when we built the graph. In that case, the
5337 * known_irn is already the upper one, so checking this should be faster.
5340 mbh = get_Block_MacroBlock(new_block);
5342 if (block == known_blk) {
5343 /* ok, we have found it: known_block dominates new_block as expected */
5348 * We have reached the MacroBlock header NOT founding
5349 * the known_block. new_block must dominate known_block.
5352 set_irn_n(known_irn, -1, new_block);
5355 assert(get_Block_n_cfgpreds(block) == 1);
5356 block = get_Block_cfgpred_block(block, 0);
5358 } /* update_value_table */
5361 * Return the canonical node computing the same value as n.
5363 * @param value_table The value table
5364 * @param n The node to lookup
5366 * Looks up the node in a hash table.
5368 * For Const nodes this is performed in the constructor, too. Const
5369 * nodes are extremely time critical because of their frequent use in
5370 * constant string arrays.
5372 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5375 if (!value_table) return n;
5379 o = pset_find(value_table, n, ir_node_hash(n));
5383 update_known_irn(o, n);
5390 * During construction we set the op_pin_state_pinned flag in the graph right when the
5391 * optimization is performed. The flag turning on procedure global cse could
5392 * be changed between two allocations. This way we are safe.
5394 * @param value_table The value table
5395 * @param n The node to lookup
5397 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5400 n = identify(value_table, n);
5401 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5402 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5404 } /* identify_cons */
5407 * Return the canonical node computing the same value as n.
5408 * Looks up the node in a hash table, enters it in the table
5409 * if it isn't there yet.
5411 * @param value_table the HashSet containing all nodes in the
5413 * @param n the node to look up
5415 * @return a node that computes the same value as n or n if no such
5416 * node could be found
5418 ir_node *identify_remember(pset *value_table, ir_node *n) {
5421 if (!value_table) return n;
5424 /* lookup or insert in hash table with given hash key. */
5425 o = pset_insert(value_table, n, ir_node_hash(n));
5428 update_known_irn(o, n);
5433 } /* identify_remember */
5435 /* Add a node to the identities value table. */
5436 void add_identities(pset *value_table, ir_node *node) {
5437 if (get_opt_cse() && is_no_Block(node))
5438 identify_remember(value_table, node);
5439 } /* add_identities */
5441 /* Visit each node in the value table of a graph. */
5442 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5444 ir_graph *rem = current_ir_graph;
5446 current_ir_graph = irg;
5447 foreach_pset(irg->value_table, node)
5449 current_ir_graph = rem;
5450 } /* visit_all_identities */
5453 * Garbage in, garbage out. If a node has a dead input, i.e., the
5454 * Bad node is input to the node, return the Bad node.
5456 static ir_node *gigo(ir_node *node) {
5458 ir_op *op = get_irn_op(node);
5460 /* remove garbage blocks by looking at control flow that leaves the block
5461 and replacing the control flow by Bad. */
5462 if (get_irn_mode(node) == mode_X) {
5463 ir_node *block = get_nodes_block(skip_Proj(node));
5465 /* Don't optimize nodes in immature blocks. */
5466 if (!get_Block_matured(block))
5468 /* Don't optimize End, may have Bads. */
5469 if (op == op_End) return node;
5471 if (is_Block(block)) {
5472 if (is_Block_dead(block)) {
5473 /* control flow from dead block is dead */
5477 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5478 if (!is_Bad(get_irn_n(block, i)))
5482 ir_graph *irg = get_irn_irg(block);
5483 /* the start block is never dead */
5484 if (block != get_irg_start_block(irg)
5485 && block != get_irg_end_block(irg)) {
5487 * Do NOT kill control flow without setting
5488 * the block to dead of bad things can happen:
5489 * We get a Block that is not reachable be irg_block_walk()
5490 * but can be found by irg_walk()!
5492 set_Block_dead(block);
5499 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5500 blocks predecessors is dead. */
5501 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5502 irn_arity = get_irn_arity(node);
5505 * Beware: we can only read the block of a non-floating node.
5507 if (is_irn_pinned_in_irg(node) &&
5508 is_Block_dead(get_nodes_block(skip_Proj(node))))
5511 for (i = 0; i < irn_arity; i++) {
5512 ir_node *pred = get_irn_n(node, i);
5517 /* Propagating Unknowns here seems to be a bad idea, because
5518 sometimes we need a node as a input and did not want that
5520 However, it might be useful to move this into a later phase
5521 (if you think that optimizing such code is useful). */
5522 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5523 return new_Unknown(get_irn_mode(node));
5528 /* With this code we violate the agreement that local_optimize
5529 only leaves Bads in Block, Phi and Tuple nodes. */
5530 /* If Block has only Bads as predecessors it's garbage. */
5531 /* If Phi has only Bads as predecessors it's garbage. */
5532 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5533 irn_arity = get_irn_arity(node);
5534 for (i = 0; i < irn_arity; i++) {
5535 if (!is_Bad(get_irn_n(node, i))) break;
5537 if (i == irn_arity) node = new_Bad();
5544 * These optimizations deallocate nodes from the obstack.
5545 * It can only be called if it is guaranteed that no other nodes
5546 * reference this one, i.e., right after construction of a node.
5548 * @param n The node to optimize
5550 * current_ir_graph must be set to the graph of the node!
5552 ir_node *optimize_node(ir_node *n) {
5555 ir_opcode iro = get_irn_opcode(n);
5557 /* Always optimize Phi nodes: part of the construction. */
5558 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5560 /* constant expression evaluation / constant folding */
5561 if (get_opt_constant_folding()) {
5562 /* neither constants nor Tuple values can be evaluated */
5563 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5564 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5565 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5566 /* try to evaluate */
5567 tv = computed_value(n);
5568 if (tv != tarval_bad) {
5570 ir_type *old_tp = get_irn_type(n);
5571 int i, arity = get_irn_arity(n);
5575 * Try to recover the type of the new expression.
5577 for (i = 0; i < arity && !old_tp; ++i)
5578 old_tp = get_irn_type(get_irn_n(n, i));
5581 * we MUST copy the node here temporary, because it's still needed
5582 * for DBG_OPT_CSTEVAL
5584 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5585 oldn = alloca(node_size);
5587 memcpy(oldn, n, node_size);
5588 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5590 /* ARG, copy the in array, we need it for statistics */
5591 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5593 /* note the inplace edges module */
5594 edges_node_deleted(n, current_ir_graph);
5596 /* evaluation was successful -- replace the node. */
5597 irg_kill_node(current_ir_graph, n);
5598 nw = new_Const(get_tarval_mode(tv), tv);
5600 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5601 set_Const_type(nw, old_tp);
5602 DBG_OPT_CSTEVAL(oldn, nw);
5603 tarval_enable_fp_ops(old_fp_mode);
5606 tarval_enable_fp_ops(old_fp_mode);
5610 /* remove unnecessary nodes */
5611 if (get_opt_constant_folding() ||
5612 (iro == iro_Phi) || /* always optimize these nodes. */
5614 (iro == iro_Proj) ||
5615 (iro == iro_Block) ) /* Flags tested local. */
5616 n = equivalent_node(n);
5618 /* Common Subexpression Elimination.
5620 * Checks whether n is already available.
5621 * The block input is used to distinguish different subexpressions. Right
5622 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5623 * subexpressions within a block.
5626 n = identify_cons(current_ir_graph->value_table, n);
5629 edges_node_deleted(oldn, current_ir_graph);
5631 /* We found an existing, better node, so we can deallocate the old node. */
5632 irg_kill_node(current_ir_graph, oldn);
5636 /* Some more constant expression evaluation that does not allow to
5638 iro = get_irn_opcode(n);
5639 if (get_opt_constant_folding() ||
5640 (iro == iro_Cond) ||
5641 (iro == iro_Proj)) /* Flags tested local. */
5642 n = transform_node(n);
5644 /* Remove nodes with dead (Bad) input.
5645 Run always for transformation induced Bads. */
5648 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5649 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5650 n = identify_remember(current_ir_graph->value_table, n);
5654 } /* optimize_node */
5658 * These optimizations never deallocate nodes (in place). This can cause dead
5659 * nodes lying on the obstack. Remove these by a dead node elimination,
5660 * i.e., a copying garbage collection.
5662 ir_node *optimize_in_place_2(ir_node *n) {
5665 ir_opcode iro = get_irn_opcode(n);
5667 if (!get_opt_optimize() && !is_Phi(n)) return n;
5669 /* constant expression evaluation / constant folding */
5670 if (get_opt_constant_folding()) {
5671 /* neither constants nor Tuple values can be evaluated */
5672 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5673 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5674 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5675 /* try to evaluate */
5676 tv = computed_value(n);
5677 if (tv != tarval_bad) {
5678 /* evaluation was successful -- replace the node. */
5679 ir_type *old_tp = get_irn_type(n);
5680 int i, arity = get_irn_arity(n);
5683 * Try to recover the type of the new expression.
5685 for (i = 0; i < arity && !old_tp; ++i)
5686 old_tp = get_irn_type(get_irn_n(n, i));
5688 n = new_Const(get_tarval_mode(tv), tv);
5690 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5691 set_Const_type(n, old_tp);
5693 DBG_OPT_CSTEVAL(oldn, n);
5694 tarval_enable_fp_ops(old_fp_mode);
5697 tarval_enable_fp_ops(old_fp_mode);
5701 /* remove unnecessary nodes */
5702 if (get_opt_constant_folding() ||
5703 (iro == iro_Phi) || /* always optimize these nodes. */
5704 (iro == iro_Id) || /* ... */
5705 (iro == iro_Proj) || /* ... */
5706 (iro == iro_Block) ) /* Flags tested local. */
5707 n = equivalent_node(n);
5709 /** common subexpression elimination **/
5710 /* Checks whether n is already available. */
5711 /* The block input is used to distinguish different subexpressions. Right
5712 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5713 subexpressions within a block. */
5714 if (get_opt_cse()) {
5715 n = identify(current_ir_graph->value_table, n);
5718 /* Some more constant expression evaluation. */
5719 iro = get_irn_opcode(n);
5720 if (get_opt_constant_folding() ||
5721 (iro == iro_Cond) ||
5722 (iro == iro_Proj)) /* Flags tested local. */
5723 n = transform_node(n);
5725 /* Remove nodes with dead (Bad) input.
5726 Run always for transformation induced Bads. */
5729 /* Now we can verify the node, as it has no dead inputs any more. */
5732 /* Now we have a legal, useful node. Enter it in hash table for cse.
5733 Blocks should be unique anyways. (Except the successor of start:
5734 is cse with the start block!) */
5735 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5736 n = identify_remember(current_ir_graph->value_table, n);
5739 } /* optimize_in_place_2 */
5742 * Wrapper for external use, set proper status bits after optimization.
5744 ir_node *optimize_in_place(ir_node *n) {
5745 /* Handle graph state */
5746 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5748 if (get_opt_global_cse())
5749 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5750 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5751 set_irg_outs_inconsistent(current_ir_graph);
5753 /* FIXME: Maybe we could also test whether optimizing the node can
5754 change the control graph. */
5755 set_irg_doms_inconsistent(current_ir_graph);
5756 return optimize_in_place_2(n);
5757 } /* optimize_in_place */
5760 * Sets the default operation for an ir_ops.
5762 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5763 ops = firm_set_default_computed_value(code, ops);
5764 ops = firm_set_default_equivalent_node(code, ops);
5765 ops = firm_set_default_transform_node(code, ops);
5766 ops = firm_set_default_node_cmp_attr(code, ops);
5767 ops = firm_set_default_get_type(code, ops);
5768 ops = firm_set_default_get_type_attr(code, ops);
5769 ops = firm_set_default_get_entity_attr(code, ops);
5772 } /* firm_set_default_operations */