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) {
641 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
642 * Do NOT optimize them away (CondEval wants them), so wait until
643 * remove_confirm is activated.
645 if (get_opt_remove_confirm()) {
646 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
647 value_of(get_Confirm_bound(n)) : tarval_bad;
650 } /* computed_value_Confirm */
653 * If the parameter n can be computed, return its value, else tarval_bad.
654 * Performs constant folding.
656 * @param n The node this should be evaluated
658 tarval *computed_value(ir_node *n) {
659 if (n->op->ops.computed_value)
660 return n->op->ops.computed_value(n);
662 } /* computed_value */
665 * Set the default computed_value evaluator in an ir_op_ops.
667 * @param code the opcode for the default operation
668 * @param ops the operations initialized
673 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
677 ops->computed_value = computed_value_##a; \
712 } /* firm_set_default_computed_value */
715 * Returns a equivalent block for another block.
716 * If the block has only one predecessor, this is
717 * the equivalent one. If the only predecessor of a block is
718 * the block itself, this is a dead block.
720 * If both predecessors of a block are the branches of a binary
721 * Cond, the equivalent block is Cond's block.
723 * If all predecessors of a block are bad or lies in a dead
724 * block, the current block is dead as well.
726 * Note, that blocks are NEVER turned into Bad's, instead
727 * the dead_block flag is set. So, never test for is_Bad(block),
728 * always use is_dead_Block(block).
730 static ir_node *equivalent_node_Block(ir_node *n)
735 /* don't optimize dead blocks */
736 if (is_Block_dead(n))
739 n_preds = get_Block_n_cfgpreds(n);
741 /* The Block constructor does not call optimize, but mature_immBlock()
742 calls the optimization. */
743 assert(get_Block_matured(n));
745 /* Straightening: a single entry Block following a single exit Block
746 can be merged, if it is not the Start block. */
747 /* !!! Beware, all Phi-nodes of n must have been optimized away.
748 This should be true, as the block is matured before optimize is called.
749 But what about Phi-cycles with the Phi0/Id that could not be resolved?
750 Remaining Phi nodes are just Ids. */
751 if (n_preds == 1 && is_Jmp(get_Block_cfgpred(n, 0))) {
752 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
753 if (predblock == oldn) {
754 /* Jmp jumps into the block it is in -- deal self cycle. */
755 n = set_Block_dead(n);
756 DBG_OPT_DEAD_BLOCK(oldn, n);
757 } else if (get_opt_control_flow_straightening()) {
759 DBG_OPT_STG(oldn, n);
761 } else if (n_preds == 1 && is_Cond(skip_Proj(get_Block_cfgpred(n, 0)))) {
762 ir_node *predblock = get_Block_cfgpred_block(n, 0);
763 if (predblock == oldn) {
764 /* Jmp jumps into the block it is in -- deal self cycle. */
765 n = set_Block_dead(n);
766 DBG_OPT_DEAD_BLOCK(oldn, n);
768 } else if ((n_preds == 2) &&
769 (get_opt_control_flow_weak_simplification())) {
770 /* Test whether Cond jumps twice to this block
771 * The more general case which more than 2 predecessors is handles
772 * in optimize_cf(), we handle only this special case for speed here.
774 ir_node *a = get_Block_cfgpred(n, 0);
775 ir_node *b = get_Block_cfgpred(n, 1);
779 (get_Proj_pred(a) == get_Proj_pred(b)) &&
780 is_Cond(get_Proj_pred(a)) &&
781 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
782 /* Also a single entry Block following a single exit Block. Phis have
783 twice the same operand and will be optimized away. */
784 n = get_nodes_block(get_Proj_pred(a));
785 DBG_OPT_IFSIM1(oldn, a, b, n);
787 } else if (get_opt_unreachable_code() &&
788 (n != get_irg_start_block(current_ir_graph)) &&
789 (n != get_irg_end_block(current_ir_graph)) ) {
792 /* If all inputs are dead, this block is dead too, except if it is
793 the start or end block. This is one step of unreachable code
795 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
796 ir_node *pred = get_Block_cfgpred(n, i);
799 if (is_Bad(pred)) continue;
800 pred_blk = get_nodes_block(skip_Proj(pred));
802 if (is_Block_dead(pred_blk)) continue;
805 /* really found a living input */
810 n = set_Block_dead(n);
811 DBG_OPT_DEAD_BLOCK(oldn, n);
816 } /* equivalent_node_Block */
819 * Returns a equivalent node for a Jmp, a Bad :-)
820 * Of course this only happens if the Block of the Jmp is dead.
822 static ir_node *equivalent_node_Jmp(ir_node *n) {
823 /* unreachable code elimination */
824 if (is_Block_dead(get_nodes_block(n)))
828 } /* equivalent_node_Jmp */
830 /** Raise is handled in the same way as Jmp. */
831 #define equivalent_node_Raise equivalent_node_Jmp
834 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
835 See transform_node_Proj_Cond(). */
838 * Optimize operations that are commutative and have neutral 0,
839 * so a op 0 = 0 op a = a.
841 static ir_node *equivalent_node_neutral_zero(ir_node *n)
845 ir_node *a = get_binop_left(n);
846 ir_node *b = get_binop_right(n);
851 /* After running compute_node there is only one constant predecessor.
852 Find this predecessors value and remember the other node: */
853 if ((tv = value_of(a)) != tarval_bad) {
855 } else if ((tv = value_of(b)) != tarval_bad) {
860 /* If this predecessors constant value is zero, the operation is
861 * unnecessary. Remove it.
863 * Beware: If n is a Add, the mode of on and n might be different
864 * which happens in this rare construction: NULL + 3.
865 * Then, a Conv would be needed which we cannot include here.
867 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
870 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
874 } /* equivalent_node_neutral_zero */
877 * Eor is commutative and has neutral 0.
879 static ir_node *equivalent_node_Eor(ir_node *n)
885 n = equivalent_node_neutral_zero(n);
886 if (n != oldn) return n;
889 b = get_Eor_right(n);
892 ir_node *aa = get_Eor_left(a);
893 ir_node *ab = get_Eor_right(a);
896 /* (a ^ b) ^ a -> b */
898 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
900 } else if (ab == b) {
901 /* (a ^ b) ^ b -> a */
903 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
908 ir_node *ba = get_Eor_left(b);
909 ir_node *bb = get_Eor_right(b);
912 /* a ^ (a ^ b) -> b */
914 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
916 } else if (bb == a) {
917 /* a ^ (b ^ a) -> b */
919 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
928 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
930 * The second one looks strange, but this construct
931 * is used heavily in the LCC sources :-).
933 * Beware: The Mode of an Add may be different than the mode of its
934 * predecessors, so we could not return a predecessors in all cases.
936 static ir_node *equivalent_node_Add(ir_node *n) {
938 ir_node *left, *right;
939 ir_mode *mode = get_irn_mode(n);
941 n = equivalent_node_neutral_zero(n);
945 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
946 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
949 left = get_Add_left(n);
950 right = get_Add_right(n);
953 if (get_Sub_right(left) == right) {
956 n = get_Sub_left(left);
957 if (mode == get_irn_mode(n)) {
958 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
964 if (get_Sub_right(right) == left) {
967 n = get_Sub_left(right);
968 if (mode == get_irn_mode(n)) {
969 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
975 } /* equivalent_node_Add */
978 * optimize operations that are not commutative but have neutral 0 on left,
981 static ir_node *equivalent_node_left_zero(ir_node *n) {
984 ir_node *a = get_binop_left(n);
985 ir_node *b = get_binop_right(n);
987 if (is_Const(b) && is_Const_null(b)) {
990 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
993 } /* equivalent_node_left_zero */
995 #define equivalent_node_Shl equivalent_node_left_zero
996 #define equivalent_node_Shr equivalent_node_left_zero
997 #define equivalent_node_Shrs equivalent_node_left_zero
998 #define equivalent_node_Rot equivalent_node_left_zero
1001 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1003 * The second one looks strange, but this construct
1004 * is used heavily in the LCC sources :-).
1006 * Beware: The Mode of a Sub may be different than the mode of its
1007 * predecessors, so we could not return a predecessors in all cases.
1009 static ir_node *equivalent_node_Sub(ir_node *n) {
1012 ir_mode *mode = get_irn_mode(n);
1014 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1015 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1018 b = get_Sub_right(n);
1020 /* Beware: modes might be different */
1021 if (is_Const(b) && is_Const_null(b)) {
1022 ir_node *a = get_Sub_left(n);
1023 if (mode == get_irn_mode(a)) {
1026 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1030 } /* equivalent_node_Sub */
1034 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1037 * -(-a) == a, but might overflow two times.
1038 * We handle it anyway here but the better way would be a
1039 * flag. This would be needed for Pascal for instance.
1041 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1043 ir_node *pred = get_unop_op(n);
1045 /* optimize symmetric unop */
1046 if (get_irn_op(pred) == get_irn_op(n)) {
1047 n = get_unop_op(pred);
1048 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1051 } /* equivalent_node_idempotent_unop */
1053 /** Optimize Not(Not(x)) == x. */
1054 #define equivalent_node_Not equivalent_node_idempotent_unop
1056 /** -(-x) == x ??? Is this possible or can --x raise an
1057 out of bounds exception if min =! max? */
1058 #define equivalent_node_Minus equivalent_node_idempotent_unop
1061 * Optimize a * 1 = 1 * a = a.
1063 static ir_node *equivalent_node_Mul(ir_node *n) {
1065 ir_node *a = get_Mul_left(n);
1067 /* we can handle here only the n * n = n bit cases */
1068 if (get_irn_mode(n) == get_irn_mode(a)) {
1069 ir_node *b = get_Mul_right(n);
1071 /* Mul is commutative and has again an other neutral element. */
1072 if (is_Const(a) && is_Const_one(a)) {
1074 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1075 } else if (is_Const(b) && is_Const_one(b)) {
1077 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1081 } /* equivalent_node_Mul */
1084 * Optimize a / 1 = a.
1086 static ir_node *equivalent_node_Div(ir_node *n) {
1087 ir_node *a = get_Div_left(n);
1088 ir_node *b = get_Div_right(n);
1090 /* Div is not commutative. */
1091 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1092 /* Turn Div into a tuple (mem, bad, a) */
1093 ir_node *mem = get_Div_mem(n);
1094 ir_node *blk = get_irn_n(n, -1);
1095 turn_into_tuple(n, pn_Div_max);
1096 set_Tuple_pred(n, pn_Div_M, mem);
1097 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1098 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1099 set_Tuple_pred(n, pn_Div_res, a);
1102 } /* equivalent_node_Div */
1105 * Optimize a / 1.0 = a.
1107 static ir_node *equivalent_node_Quot(ir_node *n) {
1108 ir_node *a = get_Quot_left(n);
1109 ir_node *b = get_Quot_right(n);
1111 /* Div is not commutative. */
1112 if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */
1113 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1114 ir_node *mem = get_Quot_mem(n);
1115 ir_node *blk = get_irn_n(n, -1);
1116 turn_into_tuple(n, pn_Quot_max);
1117 set_Tuple_pred(n, pn_Quot_M, mem);
1118 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1119 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1120 set_Tuple_pred(n, pn_Quot_res, a);
1123 } /* equivalent_node_Quot */
1126 * Optimize a / 1 = a.
1128 static ir_node *equivalent_node_DivMod(ir_node *n) {
1129 ir_node *b = get_DivMod_right(n);
1131 /* Div is not commutative. */
1132 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1133 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1134 ir_node *a = get_DivMod_left(n);
1135 ir_node *mem = get_Div_mem(n);
1136 ir_node *blk = get_irn_n(n, -1);
1137 ir_mode *mode = get_DivMod_resmode(n);
1139 turn_into_tuple(n, pn_DivMod_max);
1140 set_Tuple_pred(n, pn_DivMod_M, mem);
1141 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1142 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1143 set_Tuple_pred(n, pn_DivMod_res_div, a);
1144 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1147 } /* equivalent_node_DivMod */
1150 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1152 static ir_node *equivalent_node_Or(ir_node *n) {
1155 ir_node *a = get_Or_left(n);
1156 ir_node *b = get_Or_right(n);
1159 n = a; /* Or has it's own neutral element */
1160 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1161 } else if (is_Const(a) && is_Const_null(a)) {
1163 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1164 } else if (is_Const(b) && is_Const_null(b)) {
1166 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1170 } /* equivalent_node_Or */
1173 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1175 static ir_node *equivalent_node_And(ir_node *n) {
1178 ir_node *a = get_And_left(n);
1179 ir_node *b = get_And_right(n);
1182 n = a; /* And has it's own neutral element */
1183 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1186 if (is_Const(a) && is_Const_all_one(a)) {
1188 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1191 if (is_Const(b) && is_Const_all_one(b)) {
1193 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1197 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1200 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1205 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1208 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1214 } /* equivalent_node_And */
1217 * Try to remove useless Conv's:
1219 static ir_node *equivalent_node_Conv(ir_node *n) {
1221 ir_node *a = get_Conv_op(n);
1223 ir_mode *n_mode = get_irn_mode(n);
1224 ir_mode *a_mode = get_irn_mode(a);
1227 if (n_mode == a_mode) { /* No Conv necessary */
1228 if (get_Conv_strict(n)) {
1229 /* special case: the predecessor might be a also a Conv */
1231 if (! get_Conv_strict(a)) {
1232 /* first one is not strict, kick it */
1234 a_mode = get_irn_mode(a);
1238 /* else both are strict conv, second is superfluous */
1239 } else if (is_Proj(a)) {
1240 ir_node *pred = get_Proj_pred(a);
1241 if (is_Load(pred)) {
1242 /* loads always return with the exact precision of n_mode */
1243 assert(get_Load_mode(pred) == n_mode);
1248 /* leave strict floating point Conv's */
1252 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1253 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1254 ir_node *b = get_Conv_op(a);
1255 ir_mode *b_mode = get_irn_mode(b);
1257 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1258 /* both are strict conv */
1259 if (smaller_mode(a_mode, n_mode)) {
1260 /* both are strict, but the first is smaller, so
1261 the second cannot remove more precision, remove the
1263 set_Conv_strict(n, 0);
1266 if (n_mode == b_mode) {
1267 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1268 if (n_mode == mode_b) {
1269 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1270 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1271 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1272 if (smaller_mode(b_mode, a_mode)) {
1273 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1274 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1279 if (get_Conv_strict(n))
1280 set_Conv_strict(b, 1);
1281 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1282 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1287 } /* equivalent_node_Conv */
1290 * A Cast may be removed if the type of the previous node
1291 * is already the type of the Cast.
1293 static ir_node *equivalent_node_Cast(ir_node *n) {
1295 ir_node *pred = get_Cast_op(n);
1297 if (get_irn_type(pred) == get_Cast_type(n)) {
1299 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1302 } /* equivalent_node_Cast */
1305 * Several optimizations:
1306 * - no Phi in start block.
1307 * - remove Id operators that are inputs to Phi
1308 * - fold Phi-nodes, iff they have only one predecessor except
1311 static ir_node *equivalent_node_Phi(ir_node *n) {
1316 ir_node *first_val = NULL; /* to shutup gcc */
1318 if (!get_opt_normalize()) return n;
1320 n_preds = get_Phi_n_preds(n);
1322 block = get_nodes_block(n);
1323 if ((is_Block_dead(block)) || /* Control dead */
1324 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1325 return new_Bad(); /* in the Start Block. */
1327 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1329 /* If the Block has a Bad pred, we also have one. */
1330 for (i = 0; i < n_preds; ++i)
1331 if (is_Bad(get_Block_cfgpred(block, i)))
1332 set_Phi_pred(n, i, new_Bad());
1334 /* Find first non-self-referencing input */
1335 for (i = 0; i < n_preds; ++i) {
1336 first_val = get_Phi_pred(n, i);
1337 if ( (first_val != n) /* not self pointer */
1339 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1340 * predecessors. Then, Phi nodes in dead code might be removed, causing
1341 * nodes pointing to themself (Add's for instance).
1342 * This is really bad and causes endless recursions in several
1343 * code pathes, so we do NOT optimize such a code.
1344 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1345 * (and bad Phi predecessors), so live code is optimized later.
1347 && (! is_Bad(first_val))
1349 ) { /* value not dead */
1350 break; /* then found first value. */
1355 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1359 /* search for rest of inputs, determine if any of these
1360 are non-self-referencing */
1361 while (++i < n_preds) {
1362 ir_node *scnd_val = get_Phi_pred(n, i);
1363 if ( (scnd_val != n)
1364 && (scnd_val != first_val)
1367 && (! is_Bad(scnd_val))
1375 /* Fold, if no multiple distinct non-self-referencing inputs */
1377 DBG_OPT_PHI(oldn, n);
1380 } /* equivalent_node_Phi */
1383 * Several optimizations:
1384 * - no Sync in start block.
1385 * - fold Sync-nodes, iff they have only one predecessor except
1388 static ir_node *equivalent_node_Sync(ir_node *n) {
1389 int arity = get_Sync_n_preds(n);
1392 for (i = 0; i < arity;) {
1393 ir_node *pred = get_Sync_pred(n, i);
1396 /* Remove Bad predecessors */
1403 /* Remove duplicate predecessors */
1409 if (get_Sync_pred(n, j) == pred) {
1417 if (arity == 0) return new_Bad();
1418 if (arity == 1) return get_Sync_pred(n, 0);
1420 } /* equivalent_node_Sync */
1423 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1424 * ProjX(Load) and ProjX(Store).
1426 static ir_node *equivalent_node_Proj(ir_node *proj) {
1427 ir_node *oldn = proj;
1428 ir_node *a = get_Proj_pred(proj);
1431 /* Remove the Tuple/Proj combination. */
1432 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1433 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1434 DBG_OPT_TUPLE(oldn, a, proj);
1436 /* This should not happen! */
1437 assert(! "found a Proj with higher number than Tuple predecessors");
1440 } else if (get_irn_mode(proj) == mode_X) {
1441 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1442 /* Remove dead control flow -- early gigo(). */
1444 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1446 /* get the Load address */
1447 ir_node *addr = get_Load_ptr(a);
1448 ir_node *blk = get_irn_n(a, -1);
1451 if (value_not_null(addr, &confirm)) {
1452 if (confirm == NULL) {
1453 /* this node may float if it did not depend on a Confirm */
1454 set_irn_pinned(a, op_pin_state_floats);
1456 if (get_Proj_proj(proj) == pn_Load_X_except) {
1457 DBG_OPT_EXC_REM(proj);
1460 return new_r_Jmp(current_ir_graph, blk);
1462 } else if (is_Store(a)) {
1463 /* get the load/store address */
1464 ir_node *addr = get_Store_ptr(a);
1465 ir_node *blk = get_irn_n(a, -1);
1468 if (value_not_null(addr, &confirm)) {
1469 if (confirm == NULL) {
1470 /* this node may float if it did not depend on a Confirm */
1471 set_irn_pinned(a, op_pin_state_floats);
1473 if (get_Proj_proj(proj) == pn_Store_X_except) {
1474 DBG_OPT_EXC_REM(proj);
1477 return new_r_Jmp(current_ir_graph, blk);
1484 } /* equivalent_node_Proj */
1489 static ir_node *equivalent_node_Id(ir_node *n) {
1494 } while (get_irn_op(n) == op_Id);
1496 DBG_OPT_ID(oldn, n);
1498 } /* equivalent_node_Id */
1503 static ir_node *equivalent_node_Mux(ir_node *n)
1505 ir_node *oldn = n, *sel = get_Mux_sel(n);
1506 tarval *ts = value_of(sel);
1508 /* Mux(true, f, t) == t */
1509 if (ts == tarval_b_true) {
1510 n = get_Mux_true(n);
1511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1513 /* Mux(false, f, t) == f */
1514 else if (ts == tarval_b_false) {
1515 n = get_Mux_false(n);
1516 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1518 /* Mux(v, x, x) == x */
1519 else if (get_Mux_false(n) == get_Mux_true(n)) {
1520 n = get_Mux_true(n);
1521 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1523 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1524 ir_node *cmp = get_Proj_pred(sel);
1525 long proj_nr = get_Proj_proj(sel);
1526 ir_node *f = get_Mux_false(n);
1527 ir_node *t = get_Mux_true(n);
1530 * Note further that these optimization work even for floating point
1531 * with NaN's because -NaN == NaN.
1532 * However, if +0 and -0 is handled differently, we cannot use the first one.
1535 ir_node *const cmp_l = get_Cmp_left(cmp);
1536 ir_node *const cmp_r = get_Cmp_right(cmp);
1540 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1541 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1543 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1550 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1551 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1553 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1560 * Note: normalization puts the constant on the right side,
1561 * so we check only one case.
1563 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1564 /* Mux(t CMP 0, X, t) */
1565 if (is_Minus(f) && get_Minus_op(f) == t) {
1566 /* Mux(t CMP 0, -t, t) */
1567 if (proj_nr == pn_Cmp_Eq) {
1568 /* Mux(t == 0, -t, t) ==> -t */
1570 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1571 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1572 /* Mux(t != 0, -t, t) ==> t */
1574 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1581 } /* equivalent_node_Mux */
1584 * Returns a equivalent node of a Psi: if a condition is true
1585 * and all previous conditions are false we know its value.
1586 * If all conditions are false its value is the default one.
1588 static ir_node *equivalent_node_Psi(ir_node *n) {
1590 return equivalent_node_Mux(n);
1592 } /* equivalent_node_Psi */
1595 * Optimize -a CMP -b into b CMP a.
1596 * This works only for for modes where unary Minus
1598 * Note that two-complement integers can Overflow
1599 * so it will NOT work.
1601 * For == and != can be handled in Proj(Cmp)
1603 static ir_node *equivalent_node_Cmp(ir_node *n) {
1604 ir_node *left = get_Cmp_left(n);
1605 ir_node *right = get_Cmp_right(n);
1607 if (is_Minus(left) && is_Minus(right) &&
1608 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1609 left = get_Minus_op(left);
1610 right = get_Minus_op(right);
1611 set_Cmp_left(n, right);
1612 set_Cmp_right(n, left);
1615 } /* equivalent_node_Cmp */
1618 * Remove Confirm nodes if setting is on.
1619 * Replace Confirms(x, '=', Constlike) by Constlike.
1621 static ir_node *equivalent_node_Confirm(ir_node *n) {
1622 ir_node *pred = get_Confirm_value(n);
1623 pn_Cmp pnc = get_Confirm_cmp(n);
1625 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1627 * rare case: two identical Confirms one after another,
1628 * replace the second one with the first.
1632 if (get_opt_remove_confirm())
1633 return get_Confirm_value(n);
1638 * Optimize CopyB(mem, x, x) into a Nop.
1640 static ir_node *equivalent_node_CopyB(ir_node *n) {
1641 ir_node *a = get_CopyB_dst(n);
1642 ir_node *b = get_CopyB_src(n);
1645 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1646 ir_node *mem = get_CopyB_mem(n);
1647 ir_node *blk = get_nodes_block(n);
1648 turn_into_tuple(n, pn_CopyB_max);
1649 set_Tuple_pred(n, pn_CopyB_M, mem);
1650 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1651 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1652 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1655 } /* equivalent_node_CopyB */
1658 * Optimize Bounds(idx, idx, upper) into idx.
1660 static ir_node *equivalent_node_Bound(ir_node *n) {
1661 ir_node *idx = get_Bound_index(n);
1662 ir_node *pred = skip_Proj(idx);
1665 if (is_Bound(pred)) {
1667 * idx was Bounds checked in the same MacroBlock previously,
1668 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1670 ir_node *lower = get_Bound_lower(n);
1671 ir_node *upper = get_Bound_upper(n);
1672 if (get_Bound_lower(pred) == lower &&
1673 get_Bound_upper(pred) == upper &&
1674 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1676 * One could expect that we simply return the previous
1677 * Bound here. However, this would be wrong, as we could
1678 * add an exception Proj to a new location then.
1679 * So, we must turn in into a tuple.
1685 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1686 ir_node *mem = get_Bound_mem(n);
1687 ir_node *blk = get_nodes_block(n);
1688 turn_into_tuple(n, pn_Bound_max);
1689 set_Tuple_pred(n, pn_Bound_M, mem);
1690 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1691 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1692 set_Tuple_pred(n, pn_Bound_res, idx);
1695 } /* equivalent_node_Bound */
1698 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1699 * perform no actual computation, as, e.g., the Id nodes. It does not create
1700 * new nodes. It is therefore safe to free n if the node returned is not n.
1701 * If a node returns a Tuple we can not just skip it. If the size of the
1702 * in array fits, we transform n into a tuple (e.g., Div).
1704 ir_node *equivalent_node(ir_node *n) {
1705 if (n->op->ops.equivalent_node)
1706 return n->op->ops.equivalent_node(n);
1708 } /* equivalent_node */
1711 * Sets the default equivalent node operation for an ir_op_ops.
1713 * @param code the opcode for the default operation
1714 * @param ops the operations initialized
1719 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1723 ops->equivalent_node = equivalent_node_##a; \
1763 } /* firm_set_default_equivalent_node */
1766 * Returns non-zero if a node is a Phi node
1767 * with all predecessors constant.
1769 static int is_const_Phi(ir_node *n) {
1772 if (! is_Phi(n) || get_irn_arity(n) == 0)
1774 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1775 if (! is_Const(get_irn_n(n, i)))
1778 } /* is_const_Phi */
1781 * Apply an evaluator on a binop with a constant operators (and one Phi).
1783 * @param phi the Phi node
1784 * @param other the other operand
1785 * @param eval an evaluator function
1786 * @param mode the mode of the result, may be different from the mode of the Phi!
1787 * @param left if non-zero, other is the left operand, else the right
1789 * @return a new Phi node if the conversion was successful, NULL else
1791 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1796 int i, n = get_irn_arity(phi);
1798 NEW_ARR_A(void *, res, n);
1800 for (i = 0; i < n; ++i) {
1801 pred = get_irn_n(phi, i);
1802 tv = get_Const_tarval(pred);
1803 tv = eval(other, tv);
1805 if (tv == tarval_bad) {
1806 /* folding failed, bad */
1812 for (i = 0; i < n; ++i) {
1813 pred = get_irn_n(phi, i);
1814 tv = get_Const_tarval(pred);
1815 tv = eval(tv, other);
1817 if (tv == tarval_bad) {
1818 /* folding failed, bad */
1824 irg = current_ir_graph;
1825 for (i = 0; i < n; ++i) {
1826 pred = get_irn_n(phi, i);
1827 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1828 mode, res[i], get_Const_type(pred));
1830 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1831 } /* apply_binop_on_phi */
1834 * Apply an evaluator on a binop with two constant Phi.
1836 * @param a the left Phi node
1837 * @param b the right Phi node
1838 * @param eval an evaluator function
1839 * @param mode the mode of the result, may be different from the mode of the Phi!
1841 * @return a new Phi node if the conversion was successful, NULL else
1843 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1844 tarval *tv_l, *tv_r, *tv;
1850 if (get_nodes_block(a) != get_nodes_block(b))
1853 n = get_irn_arity(a);
1854 NEW_ARR_A(void *, res, n);
1856 for (i = 0; i < n; ++i) {
1857 pred = get_irn_n(a, i);
1858 tv_l = get_Const_tarval(pred);
1859 pred = get_irn_n(b, i);
1860 tv_r = get_Const_tarval(pred);
1861 tv = eval(tv_l, tv_r);
1863 if (tv == tarval_bad) {
1864 /* folding failed, bad */
1869 irg = current_ir_graph;
1870 for (i = 0; i < n; ++i) {
1871 pred = get_irn_n(a, i);
1872 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1874 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1875 } /* apply_binop_on_2_phis */
1878 * Apply an evaluator on a unop with a constant operator (a Phi).
1880 * @param phi the Phi node
1881 * @param eval an evaluator function
1883 * @return a new Phi node if the conversion was successful, NULL else
1885 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1891 int i, n = get_irn_arity(phi);
1893 NEW_ARR_A(void *, res, n);
1894 for (i = 0; i < n; ++i) {
1895 pred = get_irn_n(phi, i);
1896 tv = get_Const_tarval(pred);
1899 if (tv == tarval_bad) {
1900 /* folding failed, bad */
1905 mode = get_irn_mode(phi);
1906 irg = current_ir_graph;
1907 for (i = 0; i < n; ++i) {
1908 pred = get_irn_n(phi, i);
1909 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1910 mode, res[i], get_Const_type(pred));
1912 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1913 } /* apply_unop_on_phi */
1916 * Apply a conversion on a constant operator (a Phi).
1918 * @param phi the Phi node
1920 * @return a new Phi node if the conversion was successful, NULL else
1922 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1927 int i, n = get_irn_arity(phi);
1929 NEW_ARR_A(void *, res, n);
1930 for (i = 0; i < n; ++i) {
1931 pred = get_irn_n(phi, i);
1932 tv = get_Const_tarval(pred);
1933 tv = tarval_convert_to(tv, mode);
1935 if (tv == tarval_bad) {
1936 /* folding failed, bad */
1941 irg = current_ir_graph;
1942 for (i = 0; i < n; ++i) {
1943 pred = get_irn_n(phi, i);
1944 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1945 mode, res[i], get_Const_type(pred));
1947 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1948 } /* apply_conv_on_phi */
1951 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1952 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1953 * If possible, remove the Conv's.
1955 static ir_node *transform_node_AddSub(ir_node *n) {
1956 ir_mode *mode = get_irn_mode(n);
1958 if (mode_is_reference(mode)) {
1959 ir_node *left = get_binop_left(n);
1960 ir_node *right = get_binop_right(n);
1961 unsigned ref_bits = get_mode_size_bits(mode);
1963 if (is_Conv(left)) {
1964 ir_mode *lmode = get_irn_mode(left);
1965 unsigned bits = get_mode_size_bits(lmode);
1967 if (ref_bits == bits &&
1968 mode_is_int(lmode) &&
1969 get_mode_arithmetic(lmode) == irma_twos_complement) {
1970 ir_node *pre = get_Conv_op(left);
1971 ir_mode *pre_mode = get_irn_mode(pre);
1973 if (mode_is_int(pre_mode) &&
1974 get_mode_size_bits(pre_mode) == bits &&
1975 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1976 /* ok, this conv just changes to sign, moreover the calculation
1977 * is done with same number of bits as our address mode, so
1978 * we can ignore the conv as address calculation can be viewed
1979 * as either signed or unsigned
1981 set_binop_left(n, pre);
1986 if (is_Conv(right)) {
1987 ir_mode *rmode = get_irn_mode(right);
1988 unsigned bits = get_mode_size_bits(rmode);
1990 if (ref_bits == bits &&
1991 mode_is_int(rmode) &&
1992 get_mode_arithmetic(rmode) == irma_twos_complement) {
1993 ir_node *pre = get_Conv_op(right);
1994 ir_mode *pre_mode = get_irn_mode(pre);
1996 if (mode_is_int(pre_mode) &&
1997 get_mode_size_bits(pre_mode) == bits &&
1998 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1999 /* ok, this conv just changes to sign, moreover the calculation
2000 * is done with same number of bits as our address mode, so
2001 * we can ignore the conv as address calculation can be viewed
2002 * as either signed or unsigned
2004 set_binop_right(n, pre);
2009 /* let address arithmetic use unsigned modes */
2010 if (is_Const(right)) {
2011 ir_mode *rmode = get_irn_mode(right);
2013 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2014 /* convert a AddP(P, *s) into AddP(P, *u) */
2015 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2017 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2018 set_binop_right(n, pre);
2024 } /* transform_node_AddSub */
2026 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2028 if (is_Const(b) && is_const_Phi(a)) { \
2029 /* check for Op(Phi, Const) */ \
2030 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2032 else if (is_Const(a) && is_const_Phi(b)) { \
2033 /* check for Op(Const, Phi) */ \
2034 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2036 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2037 /* check for Op(Phi, Phi) */ \
2038 c = apply_binop_on_2_phis(a, b, eval, mode); \
2041 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2045 #define HANDLE_UNOP_PHI(eval, a, c) \
2047 if (is_const_Phi(a)) { \
2048 /* check for Op(Phi) */ \
2049 c = apply_unop_on_phi(a, eval); \
2051 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2057 * Do the AddSub optimization, then Transform
2058 * Constant folding on Phi
2059 * Add(a,a) -> Mul(a, 2)
2060 * Add(Mul(a, x), a) -> Mul(a, x+1)
2061 * if the mode is integer or float.
2062 * Transform Add(a,-b) into Sub(a,b).
2063 * Reassociation might fold this further.
2065 static ir_node *transform_node_Add(ir_node *n) {
2067 ir_node *a, *b, *c, *oldn = n;
2069 n = transform_node_AddSub(n);
2071 a = get_Add_left(n);
2072 b = get_Add_right(n);
2074 mode = get_irn_mode(n);
2076 if (mode_is_reference(mode)) {
2077 ir_mode *lmode = get_irn_mode(a);
2079 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2080 /* an Add(a, NULL) is a hidden Conv */
2081 dbg_info *dbg = get_irn_dbg_info(n);
2082 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2086 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2088 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2089 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2092 if (mode_is_num(mode)) {
2093 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2094 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2095 ir_node *block = get_nodes_block(n);
2098 get_irn_dbg_info(n),
2102 new_r_Const_long(current_ir_graph, block, mode, 2),
2104 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2109 get_irn_dbg_info(n),
2115 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2120 get_irn_dbg_info(n),
2126 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2129 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2130 /* Here we rely on constants be on the RIGHT side */
2132 ir_node *op = get_Not_op(a);
2134 if (is_Const(b) && is_Const_one(b)) {
2136 ir_node *blk = get_irn_n(n, -1);
2137 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2138 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2143 ir_node *blk = get_irn_n(n, -1);
2144 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2145 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2150 ir_node *op = get_Not_op(b);
2154 ir_node *blk = get_irn_n(n, -1);
2155 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2156 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2163 } /* transform_node_Add */
2166 * returns -cnst or NULL if impossible
2168 static ir_node *const_negate(ir_node *cnst) {
2169 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2170 dbg_info *dbgi = get_irn_dbg_info(cnst);
2171 ir_graph *irg = get_irn_irg(cnst);
2172 ir_node *block = get_nodes_block(cnst);
2173 ir_mode *mode = get_irn_mode(cnst);
2174 if (tv == tarval_bad) return NULL;
2175 return new_rd_Const(dbgi, irg, block, mode, tv);
2179 * Do the AddSub optimization, then Transform
2180 * Constant folding on Phi
2181 * Sub(0,a) -> Minus(a)
2182 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2183 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2184 * Sub(Add(a, x), x) -> a
2185 * Sub(x, Add(x, a)) -> -a
2186 * Sub(x, Const) -> Add(x, -Const)
2188 static ir_node *transform_node_Sub(ir_node *n) {
2193 n = transform_node_AddSub(n);
2195 a = get_Sub_left(n);
2196 b = get_Sub_right(n);
2198 mode = get_irn_mode(n);
2200 if (mode_is_int(mode)) {
2201 ir_mode *lmode = get_irn_mode(a);
2203 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2204 /* a Sub(a, NULL) is a hidden Conv */
2205 dbg_info *dbg = get_irn_dbg_info(n);
2206 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2211 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2213 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2214 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2217 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2218 /* a - C -> a + (-C) */
2219 ir_node *cnst = const_negate(b);
2221 ir_node *block = get_nodes_block(n);
2222 dbg_info *dbgi = get_irn_dbg_info(n);
2223 ir_graph *irg = get_irn_irg(n);
2225 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2226 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2231 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2232 ir_graph *irg = current_ir_graph;
2233 dbg_info *dbg = get_irn_dbg_info(n);
2234 ir_node *block = get_nodes_block(n);
2235 ir_node *left = get_Minus_op(a);
2236 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2238 n = new_rd_Minus(dbg, irg, block, add, mode);
2239 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2241 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2242 ir_graph *irg = current_ir_graph;
2243 dbg_info *dbg = get_irn_dbg_info(n);
2244 ir_node *block = get_nodes_block(n);
2245 ir_node *right = get_Minus_op(b);
2247 n = new_rd_Add(dbg, irg, block, a, right, mode);
2248 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2250 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2251 ir_graph *irg = current_ir_graph;
2252 dbg_info *s_dbg = get_irn_dbg_info(b);
2253 ir_node *s_block = get_nodes_block(b);
2254 ir_node *s_left = get_Sub_right(b);
2255 ir_node *s_right = get_Sub_left(b);
2256 ir_mode *s_mode = get_irn_mode(b);
2257 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2258 dbg_info *a_dbg = get_irn_dbg_info(n);
2259 ir_node *a_block = get_nodes_block(n);
2261 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2262 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2264 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2265 ir_node *m_right = get_Mul_right(b);
2266 if (is_Const(m_right)) {
2267 ir_node *cnst2 = const_negate(m_right);
2268 if (cnst2 != NULL) {
2269 ir_graph *irg = current_ir_graph;
2270 dbg_info *m_dbg = get_irn_dbg_info(b);
2271 ir_node *m_block = get_nodes_block(b);
2272 ir_node *m_left = get_Mul_left(b);
2273 ir_mode *m_mode = get_irn_mode(b);
2274 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2275 dbg_info *a_dbg = get_irn_dbg_info(n);
2276 ir_node *a_block = get_nodes_block(n);
2278 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2279 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2285 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2286 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2288 get_irn_dbg_info(n),
2293 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2297 if (mode_wrap_around(mode)) {
2298 ir_node *left = get_Add_left(a);
2299 ir_node *right = get_Add_right(a);
2301 /* FIXME: Does the Conv's work only for two complement or generally? */
2303 if (mode != get_irn_mode(right)) {
2304 /* This Sub is an effective Cast */
2305 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2308 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2310 } else if (right == b) {
2311 if (mode != get_irn_mode(left)) {
2312 /* This Sub is an effective Cast */
2313 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2316 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2322 if (mode_wrap_around(mode)) {
2323 ir_node *left = get_Add_left(b);
2324 ir_node *right = get_Add_right(b);
2326 /* FIXME: Does the Conv's work only for two complement or generally? */
2328 ir_mode *r_mode = get_irn_mode(right);
2330 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2331 if (mode != r_mode) {
2332 /* This Sub is an effective Cast */
2333 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2335 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2337 } else if (right == a) {
2338 ir_mode *l_mode = get_irn_mode(left);
2340 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2341 if (mode != l_mode) {
2342 /* This Sub is an effective Cast */
2343 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2345 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2350 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2351 ir_mode *mode = get_irn_mode(a);
2353 if (mode == get_irn_mode(b)) {
2355 ir_node *op_a = get_Conv_op(a);
2356 ir_node *op_b = get_Conv_op(b);
2358 /* check if it's allowed to skip the conv */
2359 ma = get_irn_mode(op_a);
2360 mb = get_irn_mode(op_b);
2362 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2363 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2366 set_Sub_right(n, b);
2372 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2373 if (!is_reassoc_running() && is_Mul(a)) {
2374 ir_node *ma = get_Mul_left(a);
2375 ir_node *mb = get_Mul_right(a);
2378 ir_node *blk = get_irn_n(n, -1);
2380 get_irn_dbg_info(n),
2381 current_ir_graph, blk,
2384 get_irn_dbg_info(n),
2385 current_ir_graph, blk,
2387 new_r_Const_long(current_ir_graph, blk, mode, 1),
2390 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2392 } else if (mb == b) {
2393 ir_node *blk = get_irn_n(n, -1);
2395 get_irn_dbg_info(n),
2396 current_ir_graph, blk,
2399 get_irn_dbg_info(n),
2400 current_ir_graph, blk,
2402 new_r_Const_long(current_ir_graph, blk, mode, 1),
2405 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2409 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2410 ir_node *x = get_Sub_left(a);
2411 ir_node *y = get_Sub_right(a);
2412 ir_node *blk = get_irn_n(n, -1);
2413 ir_mode *m_b = get_irn_mode(b);
2414 ir_mode *m_y = get_irn_mode(y);
2418 /* Determine the right mode for the Add. */
2421 else if (mode_is_reference(m_b))
2423 else if (mode_is_reference(m_y))
2427 * Both modes are different but none is reference,
2428 * happens for instance in SubP(SubP(P, Iu), Is).
2429 * We have two possibilities here: Cast or ignore.
2430 * Currently we ignore this case.
2435 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2437 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2438 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2442 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2443 if (is_Const(a) && is_Not(b)) {
2444 /* c - ~X = X + (c+1) */
2445 tarval *tv = get_Const_tarval(a);
2447 tv = tarval_add(tv, get_mode_one(mode));
2448 if (tv != tarval_bad) {
2449 ir_node *blk = get_irn_n(n, -1);
2450 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2451 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2452 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2458 } /* transform_node_Sub */
2461 * Several transformation done on n*n=2n bits mul.
2462 * These transformations must be done here because new nodes may be produced.
2464 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2466 ir_node *a = get_Mul_left(n);
2467 ir_node *b = get_Mul_right(n);
2468 tarval *ta = value_of(a);
2469 tarval *tb = value_of(b);
2470 ir_mode *smode = get_irn_mode(a);
2472 if (ta == get_mode_one(smode)) {
2473 /* (L)1 * (L)b = (L)b */
2474 ir_node *blk = get_irn_n(n, -1);
2475 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2476 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2479 else if (ta == get_mode_minus_one(smode)) {
2480 /* (L)-1 * (L)b = (L)b */
2481 ir_node *blk = get_irn_n(n, -1);
2482 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2483 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2484 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2487 if (tb == get_mode_one(smode)) {
2488 /* (L)a * (L)1 = (L)a */
2489 ir_node *blk = get_irn_n(a, -1);
2490 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2491 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2494 else if (tb == get_mode_minus_one(smode)) {
2495 /* (L)a * (L)-1 = (L)-a */
2496 ir_node *blk = get_irn_n(n, -1);
2497 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2498 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2499 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2506 * Transform Mul(a,-1) into -a.
2507 * Do constant evaluation of Phi nodes.
2508 * Do architecture dependent optimizations on Mul nodes
2510 static ir_node *transform_node_Mul(ir_node *n) {
2511 ir_node *c, *oldn = n;
2512 ir_mode *mode = get_irn_mode(n);
2513 ir_node *a = get_Mul_left(n);
2514 ir_node *b = get_Mul_right(n);
2516 if (is_Bad(a) || is_Bad(b))
2519 if (mode != get_irn_mode(a))
2520 return transform_node_Mul2n(n, mode);
2522 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2524 if (mode_is_signed(mode)) {
2527 if (value_of(a) == get_mode_minus_one(mode))
2529 else if (value_of(b) == get_mode_minus_one(mode))
2532 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2533 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2538 if (is_Const(b)) { /* (-a) * const -> a * -const */
2539 ir_node *cnst = const_negate(b);
2541 dbg_info *dbgi = get_irn_dbg_info(n);
2542 ir_node *block = get_nodes_block(n);
2543 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2544 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2547 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2548 dbg_info *dbgi = get_irn_dbg_info(n);
2549 ir_node *block = get_nodes_block(n);
2550 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2551 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2553 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2554 ir_node *sub_l = get_Sub_left(b);
2555 ir_node *sub_r = get_Sub_right(b);
2556 dbg_info *dbgi = get_irn_dbg_info(n);
2557 ir_graph *irg = current_ir_graph;
2558 ir_node *block = get_nodes_block(n);
2559 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2560 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2561 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2564 } else if (is_Minus(b)) {
2565 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2566 ir_node *sub_l = get_Sub_left(a);
2567 ir_node *sub_r = get_Sub_right(a);
2568 dbg_info *dbgi = get_irn_dbg_info(n);
2569 ir_graph *irg = current_ir_graph;
2570 ir_node *block = get_nodes_block(n);
2571 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2572 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2573 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2577 if (get_mode_arithmetic(mode) == irma_ieee754) {
2579 tarval *tv = get_Const_tarval(a);
2580 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2581 /* 2.0 * b = b + b */
2582 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2583 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2587 else if (is_Const(b)) {
2588 tarval *tv = get_Const_tarval(b);
2589 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2590 /* a * 2.0 = a + a */
2591 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2592 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2597 return arch_dep_replace_mul_with_shifts(n);
2598 } /* transform_node_Mul */
2601 * Transform a Div Node.
2603 static ir_node *transform_node_Div(ir_node *n) {
2604 ir_mode *mode = get_Div_resmode(n);
2605 ir_node *a = get_Div_left(n);
2606 ir_node *b = get_Div_right(n);
2610 if (is_Const(b) && is_const_Phi(a)) {
2611 /* check for Div(Phi, Const) */
2612 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2614 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2618 else if (is_Const(a) && is_const_Phi(b)) {
2619 /* check for Div(Const, Phi) */
2620 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2622 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2626 else if (is_const_Phi(a) && is_const_Phi(b)) {
2627 /* check for Div(Phi, Phi) */
2628 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2630 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2637 if (tv != tarval_bad) {
2638 value = new_Const(get_tarval_mode(tv), tv);
2640 DBG_OPT_CSTEVAL(n, value);
2643 ir_node *a = get_Div_left(n);
2644 ir_node *b = get_Div_right(n);
2647 if (a == b && value_not_zero(a, &dummy)) {
2648 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2649 value = new_Const(mode, get_mode_one(mode));
2650 DBG_OPT_CSTEVAL(n, value);
2653 if (mode_is_signed(mode) && is_Const(b)) {
2654 tarval *tv = get_Const_tarval(b);
2656 if (tv == get_mode_minus_one(mode)) {
2658 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2659 DBG_OPT_CSTEVAL(n, value);
2663 /* Try architecture dependent optimization */
2664 value = arch_dep_replace_div_by_const(n);
2672 /* Turn Div into a tuple (mem, jmp, bad, value) */
2673 mem = get_Div_mem(n);
2674 blk = get_irn_n(n, -1);
2676 /* skip a potential Pin */
2678 mem = get_Pin_op(mem);
2679 turn_into_tuple(n, pn_Div_max);
2680 set_Tuple_pred(n, pn_Div_M, mem);
2681 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2682 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2683 set_Tuple_pred(n, pn_Div_res, value);
2686 } /* transform_node_Div */
2689 * Transform a Mod node.
2691 static ir_node *transform_node_Mod(ir_node *n) {
2692 ir_mode *mode = get_Mod_resmode(n);
2693 ir_node *a = get_Mod_left(n);
2694 ir_node *b = get_Mod_right(n);
2698 if (is_Const(b) && is_const_Phi(a)) {
2699 /* check for Div(Phi, Const) */
2700 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2702 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2706 else if (is_Const(a) && is_const_Phi(b)) {
2707 /* check for Div(Const, Phi) */
2708 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2710 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2714 else if (is_const_Phi(a) && is_const_Phi(b)) {
2715 /* check for Div(Phi, Phi) */
2716 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2718 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2725 if (tv != tarval_bad) {
2726 value = new_Const(get_tarval_mode(tv), tv);
2728 DBG_OPT_CSTEVAL(n, value);
2731 ir_node *a = get_Mod_left(n);
2732 ir_node *b = get_Mod_right(n);
2735 if (a == b && value_not_zero(a, &dummy)) {
2736 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2737 value = new_Const(mode, get_mode_null(mode));
2738 DBG_OPT_CSTEVAL(n, value);
2741 if (mode_is_signed(mode) && is_Const(b)) {
2742 tarval *tv = get_Const_tarval(b);
2744 if (tv == get_mode_minus_one(mode)) {
2746 value = new_Const(mode, get_mode_null(mode));
2747 DBG_OPT_CSTEVAL(n, value);
2751 /* Try architecture dependent optimization */
2752 value = arch_dep_replace_mod_by_const(n);
2760 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2761 mem = get_Mod_mem(n);
2762 blk = get_irn_n(n, -1);
2764 /* skip a potential Pin */
2766 mem = get_Pin_op(mem);
2767 turn_into_tuple(n, pn_Mod_max);
2768 set_Tuple_pred(n, pn_Mod_M, mem);
2769 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2770 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2771 set_Tuple_pred(n, pn_Mod_res, value);
2774 } /* transform_node_Mod */
2777 * Transform a DivMod node.
2779 static ir_node *transform_node_DivMod(ir_node *n) {
2781 ir_node *a = get_DivMod_left(n);
2782 ir_node *b = get_DivMod_right(n);
2783 ir_mode *mode = get_DivMod_resmode(n);
2788 if (is_Const(b) && is_const_Phi(a)) {
2789 /* check for Div(Phi, Const) */
2790 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2791 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2793 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2794 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2798 else if (is_Const(a) && is_const_Phi(b)) {
2799 /* check for Div(Const, Phi) */
2800 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2801 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2803 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2804 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2808 else if (is_const_Phi(a) && is_const_Phi(b)) {
2809 /* check for Div(Phi, Phi) */
2810 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2811 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2813 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2814 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2821 if (tb != tarval_bad) {
2822 if (tb == get_mode_one(get_tarval_mode(tb))) {
2824 vb = new_Const(mode, get_mode_null(mode));
2825 DBG_OPT_CSTEVAL(n, vb);
2827 } else if (ta != tarval_bad) {
2828 tarval *resa, *resb;
2829 resa = tarval_div(ta, tb);
2830 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2831 Jmp for X result!? */
2832 resb = tarval_mod(ta, tb);
2833 if (resb == tarval_bad) return n; /* Causes exception! */
2834 va = new_Const(mode, resa);
2835 vb = new_Const(mode, resb);
2836 DBG_OPT_CSTEVAL(n, va);
2837 DBG_OPT_CSTEVAL(n, vb);
2839 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2840 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2841 vb = new_Const(mode, get_mode_null(mode));
2842 DBG_OPT_CSTEVAL(n, va);
2843 DBG_OPT_CSTEVAL(n, vb);
2845 } else { /* Try architecture dependent optimization */
2848 arch_dep_replace_divmod_by_const(&va, &vb, n);
2849 evaluated = va != NULL;
2851 } else if (a == b) {
2852 if (value_not_zero(a, &dummy)) {
2854 va = new_Const(mode, get_mode_one(mode));
2855 vb = new_Const(mode, get_mode_null(mode));
2856 DBG_OPT_CSTEVAL(n, va);
2857 DBG_OPT_CSTEVAL(n, vb);
2860 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2863 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2864 /* 0 / non-Const = 0 */
2869 if (evaluated) { /* replace by tuple */
2873 mem = get_DivMod_mem(n);
2874 /* skip a potential Pin */
2876 mem = get_Pin_op(mem);
2878 blk = get_irn_n(n, -1);
2879 turn_into_tuple(n, pn_DivMod_max);
2880 set_Tuple_pred(n, pn_DivMod_M, mem);
2881 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2882 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2883 set_Tuple_pred(n, pn_DivMod_res_div, va);
2884 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2888 } /* transform_node_DivMod */
2891 * Optimize x / c to x * (1/c)
2893 static ir_node *transform_node_Quot(ir_node *n) {
2894 ir_mode *mode = get_Quot_resmode(n);
2897 if (get_mode_arithmetic(mode) == irma_ieee754) {
2898 ir_node *b = get_Quot_right(n);
2901 tarval *tv = get_Const_tarval(b);
2905 * Floating point constant folding might be disabled here to
2907 * However, as we check for exact result, doing it is safe.
2910 rem = tarval_enable_fp_ops(1);
2911 tv = tarval_quo(get_mode_one(mode), tv);
2912 (void)tarval_enable_fp_ops(rem);
2914 /* Do the transformation if the result is either exact or we are not
2915 using strict rules. */
2916 if (tv != tarval_bad &&
2917 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2918 ir_node *blk = get_irn_n(n, -1);
2919 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2920 ir_node *a = get_Quot_left(n);
2921 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2922 ir_node *mem = get_Quot_mem(n);
2924 /* skip a potential Pin */
2926 mem = get_Pin_op(mem);
2927 turn_into_tuple(n, pn_Quot_max);
2928 set_Tuple_pred(n, pn_Quot_M, mem);
2929 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2930 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2931 set_Tuple_pred(n, pn_Quot_res, m);
2932 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2937 } /* transform_node_Quot */
2940 * Optimize Abs(x) into x if x is Confirmed >= 0
2941 * Optimize Abs(x) into -x if x is Confirmed <= 0
2942 * Optimize Abs(-x) int Abs(x)
2944 static ir_node *transform_node_Abs(ir_node *n) {
2945 ir_node *c, *oldn = n;
2946 ir_node *a = get_Abs_op(n);
2949 HANDLE_UNOP_PHI(tarval_abs, a, c);
2951 switch (classify_value_sign(a)) {
2952 case value_classified_negative:
2953 mode = get_irn_mode(n);
2956 * We can replace the Abs by -x here.
2957 * We even could add a new Confirm here
2958 * (if not twos complement)
2960 * Note that -x would create a new node, so we could
2961 * not run it in the equivalent_node() context.
2963 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2964 get_nodes_block(n), a, mode);
2966 DBG_OPT_CONFIRM(oldn, n);
2968 case value_classified_positive:
2969 /* n is positive, Abs is not needed */
2972 DBG_OPT_CONFIRM(oldn, n);
2978 /* Abs(-x) = Abs(x) */
2979 mode = get_irn_mode(n);
2980 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2981 get_nodes_block(n), get_Minus_op(a), mode);
2982 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2986 } /* transform_node_Abs */
2989 * Transform a Cond node.
2991 * Replace the Cond by a Jmp if it branches on a constant
2994 static ir_node *transform_node_Cond(ir_node *n) {
2997 ir_node *a = get_Cond_selector(n);
2998 tarval *ta = value_of(a);
3000 /* we need block info which is not available in floating irgs */
3001 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3004 if ((ta != tarval_bad) &&
3005 (get_irn_mode(a) == mode_b) &&
3006 (get_opt_unreachable_code())) {
3007 /* It's a boolean Cond, branching on a boolean constant.
3008 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3009 ir_node *blk = get_nodes_block(n);
3010 jmp = new_r_Jmp(current_ir_graph, blk);
3011 turn_into_tuple(n, pn_Cond_max);
3012 if (ta == tarval_b_true) {
3013 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3014 set_Tuple_pred(n, pn_Cond_true, jmp);
3016 set_Tuple_pred(n, pn_Cond_false, jmp);
3017 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3019 /* We might generate an endless loop, so keep it alive. */
3020 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3023 } /* transform_node_Cond */
3026 * Prototype of a recursive transform function
3027 * for bitwise distributive transformations.
3029 typedef ir_node* (*recursive_transform)(ir_node *n);
3032 * makes use of distributive laws for and, or, eor
3033 * and(a OP c, b OP c) -> and(a, b) OP c
3034 * note, might return a different op than n
3036 static ir_node *transform_bitwise_distributive(ir_node *n,
3037 recursive_transform trans_func)
3040 ir_node *a = get_binop_left(n);
3041 ir_node *b = get_binop_right(n);
3042 ir_op *op = get_irn_op(a);
3043 ir_op *op_root = get_irn_op(n);
3045 if(op != get_irn_op(b))
3048 if (op == op_Conv) {
3049 ir_node *a_op = get_Conv_op(a);
3050 ir_node *b_op = get_Conv_op(b);
3051 ir_mode *a_mode = get_irn_mode(a_op);
3052 ir_mode *b_mode = get_irn_mode(b_op);
3053 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3054 ir_node *blk = get_irn_n(n, -1);
3057 set_binop_left(n, a_op);
3058 set_binop_right(n, b_op);
3059 set_irn_mode(n, a_mode);
3061 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3063 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3069 /* nothing to gain here */
3073 if (op == op_Shrs || op == op_Shr || op == op_Shl
3074 || op == op_And || op == op_Or || op == op_Eor) {
3075 ir_node *a_left = get_binop_left(a);
3076 ir_node *a_right = get_binop_right(a);
3077 ir_node *b_left = get_binop_left(b);
3078 ir_node *b_right = get_binop_right(b);
3080 ir_node *op1 = NULL;
3081 ir_node *op2 = NULL;
3083 if (is_op_commutative(op)) {
3084 if (a_left == b_left) {
3088 } else if(a_left == b_right) {
3092 } else if(a_right == b_left) {
3098 if(a_right == b_right) {
3105 /* (a sop c) & (b sop c) => (a & b) sop c */
3106 ir_node *blk = get_irn_n(n, -1);
3108 ir_node *new_n = exact_copy(n);
3109 set_binop_left(new_n, op1);
3110 set_binop_right(new_n, op2);
3111 new_n = trans_func(new_n);
3113 if(op_root == op_Eor && op == op_Or) {
3114 dbg_info *dbgi = get_irn_dbg_info(n);
3115 ir_graph *irg = current_ir_graph;
3116 ir_mode *mode = get_irn_mode(c);
3118 c = new_rd_Not(dbgi, irg, blk, c, mode);
3119 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3122 set_nodes_block(n, blk);
3123 set_binop_left(n, new_n);
3124 set_binop_right(n, c);
3125 add_identities(current_ir_graph->value_table, n);
3128 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3139 static ir_node *transform_node_And(ir_node *n) {
3140 ir_node *c, *oldn = n;
3141 ir_node *a = get_And_left(n);
3142 ir_node *b = get_And_right(n);
3145 mode = get_irn_mode(n);
3146 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3148 /* we can evaluate 2 Projs of the same Cmp */
3149 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3150 ir_node *pred_a = get_Proj_pred(a);
3151 ir_node *pred_b = get_Proj_pred(b);
3152 if (pred_a == pred_b) {
3153 dbg_info *dbgi = get_irn_dbg_info(n);
3154 ir_node *block = get_nodes_block(pred_a);
3155 pn_Cmp pn_a = get_Proj_proj(a);
3156 pn_Cmp pn_b = get_Proj_proj(b);
3157 /* yes, we can simply calculate with pncs */
3158 pn_Cmp new_pnc = pn_a & pn_b;
3160 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3165 ir_node *op = get_Not_op(b);
3167 ir_node *ba = get_And_left(op);
3168 ir_node *bb = get_And_right(op);
3170 /* it's enough to test the following cases due to normalization! */
3171 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3172 /* (a|b) & ~(a&b) = a^b */
3173 ir_node *block = get_nodes_block(n);
3175 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3176 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3184 ir_node *op = get_Not_op(a);
3186 ir_node *aa = get_And_left(op);
3187 ir_node *ab = get_And_right(op);
3189 /* it's enough to test the following cases due to normalization! */
3190 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3191 /* (a|b) & ~(a&b) = a^b */
3192 ir_node *block = get_nodes_block(n);
3194 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3195 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3202 ir_node *al = get_Eor_left(a);
3203 ir_node *ar = get_Eor_right(a);
3206 /* (b ^ a) & b -> ~a & b */
3207 dbg_info *dbg = get_irn_dbg_info(n);
3208 ir_node *block = get_nodes_block(n);
3210 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3211 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3212 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3216 /* (a ^ b) & b -> ~a & b */
3217 dbg_info *dbg = get_irn_dbg_info(n);
3218 ir_node *block = get_nodes_block(n);
3220 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3221 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3222 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3227 ir_node *bl = get_Eor_left(b);
3228 ir_node *br = get_Eor_right(b);
3231 /* a & (a ^ b) -> a & ~b */
3232 dbg_info *dbg = get_irn_dbg_info(n);
3233 ir_node *block = get_nodes_block(n);
3235 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3236 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3241 /* a & (b ^ a) -> a & ~b */
3242 dbg_info *dbg = get_irn_dbg_info(n);
3243 ir_node *block = get_nodes_block(n);
3245 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3246 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3247 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3251 if (is_Not(a) && is_Not(b)) {
3252 /* ~a & ~b = ~(a|b) */
3253 ir_node *block = get_nodes_block(n);
3254 ir_mode *mode = get_irn_mode(n);
3258 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3259 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3260 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3264 n = transform_bitwise_distributive(n, transform_node_And);
3267 } /* transform_node_And */
3272 static ir_node *transform_node_Eor(ir_node *n) {
3273 ir_node *c, *oldn = n;
3274 ir_node *a = get_Eor_left(n);
3275 ir_node *b = get_Eor_right(n);
3276 ir_mode *mode = get_irn_mode(n);
3278 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3280 /* we can evaluate 2 Projs of the same Cmp */
3281 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3282 ir_node *pred_a = get_Proj_pred(a);
3283 ir_node *pred_b = get_Proj_pred(b);
3284 if(pred_a == pred_b) {
3285 dbg_info *dbgi = get_irn_dbg_info(n);
3286 ir_node *block = get_nodes_block(pred_a);
3287 pn_Cmp pn_a = get_Proj_proj(a);
3288 pn_Cmp pn_b = get_Proj_proj(b);
3289 /* yes, we can simply calculate with pncs */
3290 pn_Cmp new_pnc = pn_a ^ pn_b;
3292 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3299 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3300 mode, get_mode_null(mode));
3301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3302 } else if (mode == mode_b &&
3304 is_Const(b) && is_Const_one(b) &&
3305 is_Cmp(get_Proj_pred(a))) {
3306 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3307 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3308 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3310 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3311 } else if (is_Const(b)) {
3312 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3313 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3314 ir_node *not_op = get_Not_op(a);
3315 dbg_info *dbg = get_irn_dbg_info(n);
3316 ir_graph *irg = current_ir_graph;
3317 ir_node *block = get_nodes_block(n);
3318 ir_mode *mode = get_irn_mode(n);
3319 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3321 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3322 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3323 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3326 n = transform_bitwise_distributive(n, transform_node_Eor);
3330 } /* transform_node_Eor */
3335 static ir_node *transform_node_Not(ir_node *n) {
3336 ir_node *c, *oldn = n;
3337 ir_node *a = get_Not_op(n);
3338 ir_mode *mode = get_irn_mode(n);
3340 HANDLE_UNOP_PHI(tarval_not,a,c);
3342 /* check for a boolean Not */
3343 if (mode == mode_b &&
3345 is_Cmp(get_Proj_pred(a))) {
3346 /* We negate a Cmp. The Cmp has the negated result anyways! */
3347 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3348 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3349 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3353 ir_node *eor_b = get_Eor_right(a);
3354 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3355 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3356 ir_node *eor_a = get_Eor_left(a);
3357 dbg_info *dbg = get_irn_dbg_info(n);
3358 ir_graph *irg = current_ir_graph;
3359 ir_node *block = get_nodes_block(n);
3360 ir_mode *mode = get_irn_mode(n);
3361 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3365 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3366 if (is_Minus(a)) { /* ~-x -> x + -1 */
3367 dbg_info *dbg = get_irn_dbg_info(n);
3368 ir_graph *irg = current_ir_graph;
3369 ir_node *block = get_nodes_block(n);
3370 ir_node *add_l = get_Minus_op(a);
3371 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3372 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3373 } else if (is_Add(a)) {
3374 ir_node *add_r = get_Add_right(a);
3375 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3376 /* ~(x + -1) = -x */
3377 ir_node *op = get_Add_left(a);
3378 ir_node *blk = get_irn_n(n, -1);
3379 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3380 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3385 } /* transform_node_Not */
3388 * Transform a Minus.
3392 * -(a >>u (size-1)) = a >>s (size-1)
3393 * -(a >>s (size-1)) = a >>u (size-1)
3394 * -(a * const) -> a * -const
3396 static ir_node *transform_node_Minus(ir_node *n) {
3397 ir_node *c, *oldn = n;
3398 ir_node *a = get_Minus_op(n);
3401 HANDLE_UNOP_PHI(tarval_neg,a,c);
3403 mode = get_irn_mode(a);
3404 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3405 /* the following rules are only to twos-complement */
3408 ir_node *op = get_Not_op(a);
3409 tarval *tv = get_mode_one(mode);
3410 ir_node *blk = get_irn_n(n, -1);
3411 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3412 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3413 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3417 ir_node *c = get_Shr_right(a);
3420 tarval *tv = get_Const_tarval(c);
3422 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3423 /* -(a >>u (size-1)) = a >>s (size-1) */
3424 ir_node *v = get_Shr_left(a);
3426 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3427 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3433 ir_node *c = get_Shrs_right(a);
3436 tarval *tv = get_Const_tarval(c);
3438 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3439 /* -(a >>s (size-1)) = a >>u (size-1) */
3440 ir_node *v = get_Shrs_left(a);
3442 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3443 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3450 /* - (a-b) = b - a */
3451 ir_node *la = get_Sub_left(a);
3452 ir_node *ra = get_Sub_right(a);
3453 ir_node *blk = get_irn_n(n, -1);
3455 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3456 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3460 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3461 ir_node *mul_l = get_Mul_left(a);
3462 ir_node *mul_r = get_Mul_right(a);
3463 if (is_Const(mul_r)) {
3464 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3465 if(tv != tarval_bad) {
3466 ir_node *cnst = new_Const(mode, tv);
3467 dbg_info *dbg = get_irn_dbg_info(a);
3468 ir_graph *irg = current_ir_graph;
3469 ir_node *block = get_nodes_block(a);
3470 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3471 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3478 } /* transform_node_Minus */
3481 * Transform a Cast_type(Const) into a new Const_type
3483 static ir_node *transform_node_Cast(ir_node *n) {
3485 ir_node *pred = get_Cast_op(n);
3486 ir_type *tp = get_irn_type(n);
3488 if (is_Const(pred) && get_Const_type(pred) != tp) {
3489 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3490 get_Const_tarval(pred), tp);
3491 DBG_OPT_CSTEVAL(oldn, n);
3492 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3493 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3494 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3495 DBG_OPT_CSTEVAL(oldn, n);
3499 } /* transform_node_Cast */
3502 * Transform a Proj(Div) with a non-zero value.
3503 * Removes the exceptions and routes the memory to the NoMem node.
3505 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3506 ir_node *div = get_Proj_pred(proj);
3507 ir_node *b = get_Div_right(div);
3508 ir_node *confirm, *res, *new_mem;
3511 if (value_not_zero(b, &confirm)) {
3512 /* div(x, y) && y != 0 */
3513 if (confirm == NULL) {
3514 /* we are sure we have a Const != 0 */
3515 new_mem = get_Div_mem(div);
3516 if (is_Pin(new_mem))
3517 new_mem = get_Pin_op(new_mem);
3518 set_Div_mem(div, new_mem);
3519 set_irn_pinned(div, op_pin_state_floats);
3522 proj_nr = get_Proj_proj(proj);
3524 case pn_Div_X_regular:
3525 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3527 case pn_Div_X_except:
3528 /* we found an exception handler, remove it */
3529 DBG_OPT_EXC_REM(proj);
3533 res = get_Div_mem(div);
3534 new_mem = get_irg_no_mem(current_ir_graph);
3537 /* This node can only float up to the Confirm block */
3538 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3540 set_irn_pinned(div, op_pin_state_floats);
3541 /* this is a Div without exception, we can remove the memory edge */
3542 set_Div_mem(div, new_mem);
3547 } /* transform_node_Proj_Div */
3550 * Transform a Proj(Mod) with a non-zero value.
3551 * Removes the exceptions and routes the memory to the NoMem node.
3553 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3554 ir_node *mod = get_Proj_pred(proj);
3555 ir_node *b = get_Mod_right(mod);
3556 ir_node *confirm, *res, *new_mem;
3559 if (value_not_zero(b, &confirm)) {
3560 /* mod(x, y) && y != 0 */
3561 proj_nr = get_Proj_proj(proj);
3563 if (confirm == NULL) {
3564 /* we are sure we have a Const != 0 */
3565 new_mem = get_Mod_mem(mod);
3566 if (is_Pin(new_mem))
3567 new_mem = get_Pin_op(new_mem);
3568 set_Mod_mem(mod, new_mem);
3569 set_irn_pinned(mod, op_pin_state_floats);
3574 case pn_Mod_X_regular:
3575 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3577 case pn_Mod_X_except:
3578 /* we found an exception handler, remove it */
3579 DBG_OPT_EXC_REM(proj);
3583 res = get_Mod_mem(mod);
3584 new_mem = get_irg_no_mem(current_ir_graph);
3587 /* This node can only float up to the Confirm block */
3588 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3590 /* this is a Mod without exception, we can remove the memory edge */
3591 set_Mod_mem(mod, new_mem);
3594 if (get_Mod_left(mod) == b) {
3595 /* a % a = 0 if a != 0 */
3596 ir_mode *mode = get_irn_mode(proj);
3597 ir_node *res = new_Const(mode, get_mode_null(mode));
3599 DBG_OPT_CSTEVAL(mod, res);
3605 } /* transform_node_Proj_Mod */
3608 * Transform a Proj(DivMod) with a non-zero value.
3609 * Removes the exceptions and routes the memory to the NoMem node.
3611 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3612 ir_node *divmod = get_Proj_pred(proj);
3613 ir_node *b = get_DivMod_right(divmod);
3614 ir_node *confirm, *res, *new_mem;
3617 if (value_not_zero(b, &confirm)) {
3618 /* DivMod(x, y) && y != 0 */
3619 proj_nr = get_Proj_proj(proj);
3621 if (confirm == NULL) {
3622 /* we are sure we have a Const != 0 */
3623 new_mem = get_DivMod_mem(divmod);
3624 if (is_Pin(new_mem))
3625 new_mem = get_Pin_op(new_mem);
3626 set_DivMod_mem(divmod, new_mem);
3627 set_irn_pinned(divmod, op_pin_state_floats);
3632 case pn_DivMod_X_regular:
3633 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3635 case pn_DivMod_X_except:
3636 /* we found an exception handler, remove it */
3637 DBG_OPT_EXC_REM(proj);
3641 res = get_DivMod_mem(divmod);
3642 new_mem = get_irg_no_mem(current_ir_graph);
3645 /* This node can only float up to the Confirm block */
3646 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3648 /* this is a DivMod without exception, we can remove the memory edge */
3649 set_DivMod_mem(divmod, new_mem);
3652 case pn_DivMod_res_mod:
3653 if (get_DivMod_left(divmod) == b) {
3654 /* a % a = 0 if a != 0 */
3655 ir_mode *mode = get_irn_mode(proj);
3656 ir_node *res = new_Const(mode, get_mode_null(mode));
3658 DBG_OPT_CSTEVAL(divmod, res);
3664 } /* transform_node_Proj_DivMod */
3667 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3669 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3670 if (get_opt_unreachable_code()) {
3671 ir_node *n = get_Proj_pred(proj);
3672 ir_node *b = get_Cond_selector(n);
3674 if (mode_is_int(get_irn_mode(b))) {
3675 tarval *tb = value_of(b);
3677 if (tb != tarval_bad) {
3678 /* we have a constant switch */
3679 long num = get_Proj_proj(proj);
3681 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3682 if (get_tarval_long(tb) == num) {
3683 /* Do NOT create a jump here, or we will have 2 control flow ops
3684 * in a block. This case is optimized away in optimize_cf(). */
3687 /* this case will NEVER be taken, kill it */
3695 } /* transform_node_Proj_Cond */
3698 * Create a 0 constant of given mode.
3700 static ir_node *create_zero_const(ir_mode *mode) {
3701 tarval *tv = get_mode_null(mode);
3702 ir_node *cnst = new_Const(mode, tv);
3707 /* the order of the values is important! */
3708 typedef enum const_class {
3714 static const_class classify_const(const ir_node* n)
3716 if (is_Const(n)) return const_const;
3717 if (is_irn_constlike(n)) return const_like;
3722 * Determines whether r is more constlike or has a larger index (in that order)
3725 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3727 const const_class l_order = classify_const(l);
3728 const const_class r_order = classify_const(r);
3730 l_order > r_order ||
3731 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3735 * Normalizes and optimizes Cmp nodes.
3737 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3738 ir_node *n = get_Proj_pred(proj);
3739 ir_node *left = get_Cmp_left(n);
3740 ir_node *right = get_Cmp_right(n);
3743 ir_mode *mode = NULL;
3744 long proj_nr = get_Proj_proj(proj);
3746 /* we can evaluate some cases directly */
3749 return new_Const(mode_b, get_tarval_b_false());
3751 return new_Const(mode_b, get_tarval_b_true());
3753 if (!mode_is_float(get_irn_mode(left)))
3754 return new_Const(mode_b, get_tarval_b_true());
3760 /* remove Casts of both sides */
3761 left = skip_Cast(left);
3762 right = skip_Cast(right);
3764 /* Remove unnecessary conversions */
3765 /* TODO handle constants */
3766 if (is_Conv(left) && is_Conv(right)) {
3767 ir_mode *mode = get_irn_mode(left);
3768 ir_node *op_left = get_Conv_op(left);
3769 ir_node *op_right = get_Conv_op(right);
3770 ir_mode *mode_left = get_irn_mode(op_left);
3771 ir_mode *mode_right = get_irn_mode(op_right);
3773 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3774 && mode_left != mode_b && mode_right != mode_b) {
3775 ir_graph *irg = current_ir_graph;
3776 ir_node *block = get_nodes_block(n);
3778 if (mode_left == mode_right) {
3782 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3783 } else if (smaller_mode(mode_left, mode_right)) {
3784 left = new_r_Conv(irg, block, op_left, mode_right);
3787 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3788 } else if (smaller_mode(mode_right, mode_left)) {
3790 right = new_r_Conv(irg, block, op_right, mode_left);
3792 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3797 /* remove operation on both sides if possible */
3798 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3800 * The following operations are NOT safe for floating point operations, for instance
3801 * 1.0 + inf == 2.0 + inf, =/=> x == y
3803 if (mode_is_int(get_irn_mode(left))) {
3804 unsigned lop = get_irn_opcode(left);
3806 if (lop == get_irn_opcode(right)) {
3807 ir_node *ll, *lr, *rl, *rr;
3809 /* same operation on both sides, try to remove */
3813 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3814 left = get_unop_op(left);
3815 right = get_unop_op(right);
3817 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3820 ll = get_Add_left(left);
3821 lr = get_Add_right(left);
3822 rl = get_Add_left(right);
3823 rr = get_Add_right(right);
3826 /* X + a CMP X + b ==> a CMP b */
3830 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3831 } else if (ll == rr) {
3832 /* X + a CMP b + X ==> a CMP b */
3836 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3837 } else if (lr == rl) {
3838 /* a + X CMP X + b ==> a CMP b */
3842 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3843 } else if (lr == rr) {
3844 /* a + X CMP b + X ==> a CMP b */
3848 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3852 ll = get_Sub_left(left);
3853 lr = get_Sub_right(left);
3854 rl = get_Sub_left(right);
3855 rr = get_Sub_right(right);
3858 /* X - a CMP X - b ==> a CMP b */
3862 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3863 } else if (lr == rr) {
3864 /* a - X CMP b - X ==> a CMP b */
3868 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3872 if (get_Rot_right(left) == get_Rot_right(right)) {
3873 /* a ROT X CMP b ROT X ==> a CMP b */
3874 left = get_Rot_left(left);
3875 right = get_Rot_left(right);
3877 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3885 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3886 if (is_Add(left) || is_Sub(left)) {
3887 ir_node *ll = get_binop_left(left);
3888 ir_node *lr = get_binop_right(left);
3890 if (lr == right && is_Add(left)) {
3897 right = create_zero_const(get_irn_mode(left));
3899 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3902 if (is_Add(right) || is_Sub(right)) {
3903 ir_node *rl = get_binop_left(right);
3904 ir_node *rr = get_binop_right(right);
3906 if (rr == left && is_Add(right)) {
3913 right = create_zero_const(get_irn_mode(left));
3915 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3918 } /* mode_is_int(...) */
3919 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3921 /* replace mode_b compares with ands/ors */
3922 if (get_irn_mode(left) == mode_b) {
3923 ir_graph *irg = current_ir_graph;
3924 ir_node *block = get_nodes_block(n);
3928 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3929 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3930 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3931 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3932 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3933 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3934 default: bres = NULL;
3937 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3943 * First step: normalize the compare op
3944 * by placing the constant on the right side
3945 * or moving the lower address node to the left.
3947 if (!operands_are_normalized(left, right)) {
3953 proj_nr = get_inversed_pnc(proj_nr);
3958 * Second step: Try to reduce the magnitude
3959 * of a constant. This may help to generate better code
3960 * later and may help to normalize more compares.
3961 * Of course this is only possible for integer values.
3963 if (is_Const(right)) {
3964 mode = get_irn_mode(right);
3965 tv = get_Const_tarval(right);
3967 /* TODO extend to arbitrary constants */
3968 if (is_Conv(left) && tarval_is_null(tv)) {
3969 ir_node *op = get_Conv_op(left);
3970 ir_mode *op_mode = get_irn_mode(op);
3973 * UpConv(x) REL 0 ==> x REL 0
3975 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3976 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3977 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3978 tv = get_mode_null(op_mode);
3982 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3986 if (tv != tarval_bad) {
3987 /* the following optimization is possible on modes without Overflow
3988 * on Unary Minus or on == and !=:
3989 * -a CMP c ==> a swap(CMP) -c
3991 * Beware: for two-complement Overflow may occur, so only == and != can
3992 * be optimized, see this:
3993 * -MININT < 0 =/=> MININT > 0 !!!
3995 if (is_Minus(left) &&
3996 (!mode_overflow_on_unary_Minus(mode) ||
3997 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3998 tv = tarval_neg(tv);
4000 if (tv != tarval_bad) {
4001 left = get_Minus_op(left);
4002 proj_nr = get_inversed_pnc(proj_nr);
4004 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4006 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4007 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4008 tv = tarval_not(tv);
4010 if (tv != tarval_bad) {
4011 left = get_Not_op(left);
4013 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4017 /* for integer modes, we have more */
4018 if (mode_is_int(mode)) {
4019 /* Ne includes Unordered which is not possible on integers.
4020 * However, frontends often use this wrong, so fix it here */
4021 if (proj_nr & pn_Cmp_Uo) {
4022 proj_nr &= ~pn_Cmp_Uo;
4023 set_Proj_proj(proj, proj_nr);
4026 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4027 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4028 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4029 tv = tarval_sub(tv, get_mode_one(mode));
4031 if (tv != tarval_bad) {
4032 proj_nr ^= pn_Cmp_Eq;
4034 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4037 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4038 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4039 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4040 tv = tarval_add(tv, get_mode_one(mode));
4042 if (tv != tarval_bad) {
4043 proj_nr ^= pn_Cmp_Eq;
4045 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4049 /* the following reassociations work only for == and != */
4050 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4052 #if 0 /* Might be not that good in general */
4053 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4054 if (tarval_is_null(tv) && is_Sub(left)) {
4055 right = get_Sub_right(left);
4056 left = get_Sub_left(left);
4058 tv = value_of(right);
4060 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4064 if (tv != tarval_bad) {
4065 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4067 ir_node *c1 = get_Sub_right(left);
4068 tarval *tv2 = value_of(c1);
4070 if (tv2 != tarval_bad) {
4071 tv2 = tarval_add(tv, value_of(c1));
4073 if (tv2 != tarval_bad) {
4074 left = get_Sub_left(left);
4077 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4081 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4082 else if (is_Add(left)) {
4083 ir_node *a_l = get_Add_left(left);
4084 ir_node *a_r = get_Add_right(left);
4088 if (is_Const(a_l)) {
4090 tv2 = value_of(a_l);
4093 tv2 = value_of(a_r);
4096 if (tv2 != tarval_bad) {
4097 tv2 = tarval_sub(tv, tv2);
4099 if (tv2 != tarval_bad) {
4103 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4107 /* -a == c ==> a == -c, -a != c ==> a != -c */
4108 else if (is_Minus(left)) {
4109 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4111 if (tv2 != tarval_bad) {
4112 left = get_Minus_op(left);
4115 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4120 /* the following reassociations work only for <= */
4121 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4122 if (tv != tarval_bad) {
4123 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4124 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4130 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4131 switch (get_irn_opcode(left)) {
4135 c1 = get_And_right(left);
4138 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4139 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4141 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4143 /* TODO: move to constant evaluation */
4144 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4145 c1 = new_Const(mode_b, tv);
4146 DBG_OPT_CSTEVAL(proj, c1);
4150 if (tarval_is_single_bit(tv)) {
4152 * optimization for AND:
4154 * And(x, C) == C ==> And(x, C) != 0
4155 * And(x, C) != C ==> And(X, C) == 0
4157 * if C is a single Bit constant.
4160 /* check for Constant's match. We have check hare the tarvals,
4161 because our const might be changed */
4162 if (get_Const_tarval(c1) == tv) {
4163 /* fine: do the transformation */
4164 tv = get_mode_null(get_tarval_mode(tv));
4165 proj_nr ^= pn_Cmp_Leg;
4167 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4173 c1 = get_Or_right(left);
4174 if (is_Const(c1) && tarval_is_null(tv)) {
4176 * Or(x, C) == 0 && C != 0 ==> FALSE
4177 * Or(x, C) != 0 && C != 0 ==> TRUE
4179 if (! tarval_is_null(get_Const_tarval(c1))) {
4180 /* TODO: move to constant evaluation */
4181 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4182 c1 = new_Const(mode_b, tv);
4183 DBG_OPT_CSTEVAL(proj, c1);
4190 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4192 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4195 c1 = get_Shl_right(left);
4197 tarval *tv1 = get_Const_tarval(c1);
4198 ir_mode *mode = get_irn_mode(left);
4199 tarval *minus1 = get_mode_all_one(mode);
4200 tarval *amask = tarval_shr(minus1, tv1);
4201 tarval *cmask = tarval_shl(minus1, tv1);
4204 if (tarval_and(tv, cmask) != tv) {
4205 /* condition not met */
4206 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4207 c1 = new_Const(mode_b, tv);
4208 DBG_OPT_CSTEVAL(proj, c1);
4211 sl = get_Shl_left(left);
4212 blk = get_nodes_block(n);
4213 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4214 tv = tarval_shr(tv, tv1);
4216 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4221 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4223 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4226 c1 = get_Shr_right(left);
4228 tarval *tv1 = get_Const_tarval(c1);
4229 ir_mode *mode = get_irn_mode(left);
4230 tarval *minus1 = get_mode_all_one(mode);
4231 tarval *amask = tarval_shl(minus1, tv1);
4232 tarval *cmask = tarval_shr(minus1, tv1);
4235 if (tarval_and(tv, cmask) != tv) {
4236 /* condition not met */
4237 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4238 c1 = new_Const(mode_b, tv);
4239 DBG_OPT_CSTEVAL(proj, c1);
4242 sl = get_Shr_left(left);
4243 blk = get_nodes_block(n);
4244 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4245 tv = tarval_shl(tv, tv1);
4247 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4252 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4254 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4257 c1 = get_Shrs_right(left);
4259 tarval *tv1 = get_Const_tarval(c1);
4260 ir_mode *mode = get_irn_mode(left);
4261 tarval *minus1 = get_mode_all_one(mode);
4262 tarval *amask = tarval_shl(minus1, tv1);
4263 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4266 cond = tarval_sub(cond, tv1);
4267 cond = tarval_shrs(tv, cond);
4269 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4270 /* condition not met */
4271 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4272 c1 = new_Const(mode_b, tv);
4273 DBG_OPT_CSTEVAL(proj, c1);
4276 sl = get_Shrs_left(left);
4277 blk = get_nodes_block(n);
4278 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4279 tv = tarval_shl(tv, tv1);
4281 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4286 } /* tarval != bad */
4289 if (changed & 2) /* need a new Const */
4290 right = new_Const(mode, tv);
4292 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4293 ir_node *op = get_Proj_pred(left);
4295 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4296 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4297 ir_node *c = get_binop_right(op);
4300 tarval *tv = get_Const_tarval(c);
4302 if (tarval_is_single_bit(tv)) {
4303 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4304 ir_node *v = get_binop_left(op);
4305 ir_node *blk = get_irn_n(op, -1);
4306 ir_mode *mode = get_irn_mode(v);
4308 tv = tarval_sub(tv, get_mode_one(mode));
4309 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4311 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4318 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4320 /* create a new compare */
4321 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4322 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4326 } /* transform_node_Proj_Cmp */
4329 * Does all optimizations on nodes that must be done on it's Proj's
4330 * because of creating new nodes.
4332 static ir_node *transform_node_Proj(ir_node *proj) {
4333 ir_node *n = get_Proj_pred(proj);
4335 switch (get_irn_opcode(n)) {
4337 return transform_node_Proj_Div(proj);
4340 return transform_node_Proj_Mod(proj);
4343 return transform_node_Proj_DivMod(proj);
4346 return transform_node_Proj_Cond(proj);
4349 return transform_node_Proj_Cmp(proj);
4352 /* should not happen, but if it does will be optimized away */
4353 return equivalent_node_Proj(proj);
4359 } /* transform_node_Proj */
4362 * Move Confirms down through Phi nodes.
4364 static ir_node *transform_node_Phi(ir_node *phi) {
4366 ir_mode *mode = get_irn_mode(phi);
4368 if (mode_is_reference(mode)) {
4369 n = get_irn_arity(phi);
4371 /* Beware of Phi0 */
4373 ir_node *pred = get_irn_n(phi, 0);
4374 ir_node *bound, *new_Phi, *block, **in;
4377 if (! is_Confirm(pred))
4380 bound = get_Confirm_bound(pred);
4381 pnc = get_Confirm_cmp(pred);
4383 NEW_ARR_A(ir_node *, in, n);
4384 in[0] = get_Confirm_value(pred);
4386 for (i = 1; i < n; ++i) {
4387 pred = get_irn_n(phi, i);
4389 if (! is_Confirm(pred) ||
4390 get_Confirm_bound(pred) != bound ||
4391 get_Confirm_cmp(pred) != pnc)
4393 in[i] = get_Confirm_value(pred);
4395 /* move the Confirm nodes "behind" the Phi */
4396 block = get_irn_n(phi, -1);
4397 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4398 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4402 } /* transform_node_Phi */
4405 * Returns the operands of a commutative bin-op, if one operand is
4406 * a const, it is returned as the second one.
4408 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4409 ir_node *op_a = get_binop_left(binop);
4410 ir_node *op_b = get_binop_right(binop);
4412 assert(is_op_commutative(get_irn_op(binop)));
4414 if (is_Const(op_a)) {
4421 } /* get_comm_Binop_Ops */
4424 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4425 * Such pattern may arise in bitfield stores.
4427 * value c4 value c4 & c2
4428 * AND c3 AND c1 | c3
4435 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4438 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4441 ir_node *and_l, *c3;
4442 ir_node *value, *c4;
4443 ir_node *new_and, *new_const, *block;
4444 ir_mode *mode = get_irn_mode(or);
4446 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4449 get_comm_Binop_Ops(or, &and, &c1);
4450 if (!is_Const(c1) || !is_And(and))
4453 get_comm_Binop_Ops(and, &or_l, &c2);
4457 tv1 = get_Const_tarval(c1);
4458 tv2 = get_Const_tarval(c2);
4460 tv = tarval_or(tv1, tv2);
4461 if (tarval_is_all_one(tv)) {
4462 /* the AND does NOT clear a bit with isn't set by the OR */
4463 set_Or_left(or, or_l);
4464 set_Or_right(or, c1);
4466 /* check for more */
4473 get_comm_Binop_Ops(or_l, &and_l, &c3);
4474 if (!is_Const(c3) || !is_And(and_l))
4477 get_comm_Binop_Ops(and_l, &value, &c4);
4481 /* ok, found the pattern, check for conditions */
4482 assert(mode == get_irn_mode(and));
4483 assert(mode == get_irn_mode(or_l));
4484 assert(mode == get_irn_mode(and_l));
4486 tv3 = get_Const_tarval(c3);
4487 tv4 = get_Const_tarval(c4);
4489 tv = tarval_or(tv4, tv2);
4490 if (!tarval_is_all_one(tv)) {
4491 /* have at least one 0 at the same bit position */
4495 n_tv4 = tarval_not(tv4);
4496 if (tv3 != tarval_and(tv3, n_tv4)) {
4497 /* bit in the or_mask is outside the and_mask */
4501 n_tv2 = tarval_not(tv2);
4502 if (tv1 != tarval_and(tv1, n_tv2)) {
4503 /* bit in the or_mask is outside the and_mask */
4507 /* ok, all conditions met */
4508 block = get_irn_n(or, -1);
4510 new_and = new_r_And(current_ir_graph, block,
4511 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4513 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4515 set_Or_left(or, new_and);
4516 set_Or_right(or, new_const);
4518 /* check for more */
4520 } /* transform_node_Or_bf_store */
4523 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4525 static ir_node *transform_node_Or_Rot(ir_node *or) {
4526 ir_mode *mode = get_irn_mode(or);
4527 ir_node *shl, *shr, *block;
4528 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4531 if (! mode_is_int(mode))
4534 shl = get_binop_left(or);
4535 shr = get_binop_right(or);
4544 } else if (!is_Shl(shl)) {
4546 } else if (!is_Shr(shr)) {
4549 x = get_Shl_left(shl);
4550 if (x != get_Shr_left(shr))
4553 c1 = get_Shl_right(shl);
4554 c2 = get_Shr_right(shr);
4555 if (is_Const(c1) && is_Const(c2)) {
4556 tv1 = get_Const_tarval(c1);
4557 if (! tarval_is_long(tv1))
4560 tv2 = get_Const_tarval(c2);
4561 if (! tarval_is_long(tv2))
4564 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4565 != (int) get_mode_size_bits(mode))
4568 /* yet, condition met */
4569 block = get_irn_n(or, -1);
4571 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4573 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4575 } else if (is_Sub(c1)) {
4579 if (get_Sub_right(sub) != v)
4582 c1 = get_Sub_left(sub);
4586 tv1 = get_Const_tarval(c1);
4587 if (! tarval_is_long(tv1))
4590 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4593 /* yet, condition met */
4594 block = get_nodes_block(or);
4596 /* a Rot right is not supported, so use a rot left */
4597 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4599 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4601 } else if (is_Sub(c2)) {
4605 c1 = get_Sub_left(sub);
4609 tv1 = get_Const_tarval(c1);
4610 if (! tarval_is_long(tv1))
4613 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4616 /* yet, condition met */
4617 block = get_irn_n(or, -1);
4620 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4622 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4627 } /* transform_node_Or_Rot */
4632 static ir_node *transform_node_Or(ir_node *n) {
4633 ir_node *c, *oldn = n;
4634 ir_node *a = get_Or_left(n);
4635 ir_node *b = get_Or_right(n);
4638 if (is_Not(a) && is_Not(b)) {
4639 /* ~a | ~b = ~(a&b) */
4640 ir_node *block = get_nodes_block(n);
4642 mode = get_irn_mode(n);
4645 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4646 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4647 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4651 /* we can evaluate 2 Projs of the same Cmp */
4652 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4653 ir_node *pred_a = get_Proj_pred(a);
4654 ir_node *pred_b = get_Proj_pred(b);
4655 if (pred_a == pred_b) {
4656 dbg_info *dbgi = get_irn_dbg_info(n);
4657 ir_node *block = get_nodes_block(pred_a);
4658 pn_Cmp pn_a = get_Proj_proj(a);
4659 pn_Cmp pn_b = get_Proj_proj(b);
4660 /* yes, we can simply calculate with pncs */
4661 pn_Cmp new_pnc = pn_a | pn_b;
4663 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4668 mode = get_irn_mode(n);
4669 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4671 n = transform_node_Or_bf_store(n);
4672 n = transform_node_Or_Rot(n);
4676 n = transform_bitwise_distributive(n, transform_node_Or);
4679 } /* transform_node_Or */
4683 static ir_node *transform_node(ir_node *n);
4686 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4688 * Should be moved to reassociation?
4690 static ir_node *transform_node_shift(ir_node *n) {
4691 ir_node *left, *right;
4692 tarval *tv1, *tv2, *res;
4694 int modulo_shf, flag;
4696 left = get_binop_left(n);
4698 /* different operations */
4699 if (get_irn_op(left) != get_irn_op(n))
4702 right = get_binop_right(n);
4703 tv1 = value_of(right);
4704 if (tv1 == tarval_bad)
4707 tv2 = value_of(get_binop_right(left));
4708 if (tv2 == tarval_bad)
4711 res = tarval_add(tv1, tv2);
4713 /* beware: a simple replacement works only, if res < modulo shift */
4714 mode = get_irn_mode(n);
4718 modulo_shf = get_mode_modulo_shift(mode);
4719 if (modulo_shf > 0) {
4720 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4722 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4728 /* ok, we can replace it */
4729 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4731 in[0] = get_binop_left(left);
4732 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4734 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4736 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4738 return transform_node(irn);
4741 } /* transform_node_shift */
4746 static ir_node *transform_node_Shr(ir_node *n) {
4747 ir_node *c, *oldn = n;
4748 ir_node *a = get_Shr_left(n);
4749 ir_node *b = get_Shr_right(n);
4750 ir_mode *mode = get_irn_mode(n);
4752 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4753 return transform_node_shift(n);
4754 } /* transform_node_Shr */
4759 static ir_node *transform_node_Shrs(ir_node *n) {
4760 ir_node *c, *oldn = n;
4761 ir_node *a = get_Shrs_left(n);
4762 ir_node *b = get_Shrs_right(n);
4763 ir_mode *mode = get_irn_mode(n);
4765 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4766 return transform_node_shift(n);
4767 } /* transform_node_Shrs */
4772 static ir_node *transform_node_Shl(ir_node *n) {
4773 ir_node *c, *oldn = n;
4774 ir_node *a = get_Shl_left(n);
4775 ir_node *b = get_Shl_right(n);
4776 ir_mode *mode = get_irn_mode(n);
4778 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4779 return transform_node_shift(n);
4780 } /* transform_node_Shl */
4785 static ir_node *transform_node_Rot(ir_node *n) {
4786 ir_node *c, *oldn = n;
4787 ir_node *a = get_Rot_left(n);
4788 ir_node *b = get_Rot_right(n);
4789 ir_mode *mode = get_irn_mode(n);
4791 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4792 return transform_node_shift(n);
4793 } /* transform_node_Rot */
4798 static ir_node *transform_node_Conv(ir_node *n) {
4799 ir_node *c, *oldn = n;
4800 ir_node *a = get_Conv_op(n);
4802 if (is_const_Phi(a)) {
4803 c = apply_conv_on_phi(a, get_irn_mode(n));
4805 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4810 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4811 ir_mode *mode = get_irn_mode(n);
4812 return new_r_Unknown(current_ir_graph, mode);
4816 } /* transform_node_Conv */
4819 * Remove dead blocks and nodes in dead blocks
4820 * in keep alive list. We do not generate a new End node.
4822 static ir_node *transform_node_End(ir_node *n) {
4823 int i, j, n_keepalives = get_End_n_keepalives(n);
4826 NEW_ARR_A(ir_node *, in, n_keepalives);
4828 for (i = j = 0; i < n_keepalives; ++i) {
4829 ir_node *ka = get_End_keepalive(n, i);
4831 if (! is_Block_dead(ka)) {
4835 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4838 /* FIXME: beabi need to keep a Proj(M) */
4839 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4842 if (j != n_keepalives)
4843 set_End_keepalives(n, j, in);
4845 } /* transform_node_End */
4847 /** returns 1 if a == -b */
4848 static int is_negated_value(ir_node *a, ir_node *b) {
4849 if (is_Minus(a) && get_Minus_op(a) == b)
4851 if (is_Minus(b) && get_Minus_op(b) == a)
4853 if (is_Sub(a) && is_Sub(b)) {
4854 ir_node *a_left = get_Sub_left(a);
4855 ir_node *a_right = get_Sub_right(a);
4856 ir_node *b_left = get_Sub_left(b);
4857 ir_node *b_right = get_Sub_right(b);
4859 if (a_left == b_right && a_right == b_left)
4867 * Optimize a Mux into some simpler cases.
4869 static ir_node *transform_node_Mux(ir_node *n) {
4870 ir_node *oldn = n, *sel = get_Mux_sel(n);
4871 ir_mode *mode = get_irn_mode(n);
4872 ir_node *t = get_Mux_true(n);
4873 ir_node *f = get_Mux_false(n);
4874 ir_graph *irg = current_ir_graph;
4875 ir_node *conds[1], *vals[2];
4877 /* first normalization step: move a possible zero to the false case */
4879 ir_node *cmp = get_Proj_pred(sel);
4882 if (is_Const(t) && is_Const_null(t)) {
4883 /* Psi(x, 0, y) => Psi(x, y, 0) */
4884 pn_Cmp pnc = get_Proj_proj(sel);
4885 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
4886 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
4890 n = new_rd_Psi(get_irn_dbg_info(n), irg, get_nodes_block(n), 1, conds, vals, mode);
4897 /* note: after normalization, false can only happen on default */
4898 if (mode == mode_b) {
4899 dbg_info *dbg = get_irn_dbg_info(n);
4900 ir_node *block = get_nodes_block(n);
4901 ir_graph *irg = current_ir_graph;
4904 tarval *tv_t = get_Const_tarval(t);
4905 if (tv_t == tarval_b_true) {
4907 /* Muxb(sel, true, false) = sel */
4908 assert(get_Const_tarval(f) == tarval_b_false);
4909 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
4912 /* Muxb(sel, true, x) = Or(sel, x) */
4913 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
4914 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
4918 } else if (is_Const(f)) {
4919 tarval *tv_f = get_Const_tarval(f);
4920 if (tv_f == tarval_b_true) {
4921 /* Muxb(sel, x, true) = Or(Not(sel), x) */
4922 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4923 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
4924 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4927 /* Muxb(sel, x, false) = And(sel, x) */
4928 assert(tv_f == tarval_b_false);
4929 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
4930 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
4936 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
4937 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
4938 tarval *a = get_Const_tarval(t);
4939 tarval *b = get_Const_tarval(f);
4940 tarval *null = get_tarval_null(mode);
4943 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
4944 diff = tarval_sub(a, b);
4947 diff = tarval_sub(b, a);
4951 if (diff == get_tarval_one(mode) && min != null) {
4952 dbg_info *dbg = get_irn_dbg_info(n);
4953 ir_node *block = get_nodes_block(n);
4954 ir_graph *irg = current_ir_graph;
4958 vals[0] = new_Const(mode, tarval_sub(a, min));
4959 vals[1] = new_Const(mode, tarval_sub(b, min));
4960 n = new_rd_Psi(dbg, irg, block, 1, conds, vals, mode);
4961 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
4967 ir_node *cmp = get_Proj_pred(sel);
4968 long pn = get_Proj_proj(sel);
4971 * Note: normalization puts the constant on the right side,
4972 * so we check only one case.
4974 * Note further that these optimization work even for floating point
4975 * with NaN's because -NaN == NaN.
4976 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
4980 ir_node *cmp_r = get_Cmp_right(cmp);
4981 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4982 ir_node *block = get_nodes_block(n);
4983 ir_node *cmp_l = get_Cmp_left(cmp);
4985 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
4988 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4989 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4991 /* Psi(a >/>= 0, a, -a) = Psi(a </<= 0, -a, a) ==> Abs(a) */
4992 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4994 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4996 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4997 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4999 /* Psi(a </<= 0, a, -a) = Psi(a >/>= 0, -a, a) ==> -Abs(a) */
5000 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5002 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5004 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5009 if (mode_is_int(mode)) {
5011 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5012 /* Psi((a & b) != 0, c, 0) */
5013 ir_node *and_r = get_And_right(cmp_l);
5016 if (and_r == t && f == cmp_r) {
5017 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5018 if (pn == pn_Cmp_Lg) {
5019 /* Psi((a & 2^C) != 0, 2^C, 0) */
5022 /* Psi((a & 2^C) == 0, 2^C, 0) */
5023 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5024 block, cmp_l, t, mode);
5029 if (is_Shl(and_r)) {
5030 ir_node *shl_l = get_Shl_left(and_r);
5031 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5032 if (and_r == t && f == cmp_r) {
5033 if (pn == pn_Cmp_Lg) {
5034 /* (a & (1 << n)) != 0, (1 << n), 0) */
5037 /* (a & (1 << n)) == 0, (1 << n), 0) */
5038 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5039 block, cmp_l, t, mode);
5045 and_l = get_And_left(cmp_l);
5046 if (is_Shl(and_l)) {
5047 ir_node *shl_l = get_Shl_left(and_l);
5048 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5049 if (and_l == t && f == cmp_r) {
5050 if (pn == pn_Cmp_Lg) {
5051 /* ((1 << n) & a) != 0, (1 << n), 0) */
5054 /* ((1 << n) & a) == 0, (1 << n), 0) */
5055 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5056 block, cmp_l, t, mode);
5067 return arch_transform_node_Mux(n);
5068 } /* transform_node_Mux */
5071 * Optimize a Psi into some simpler cases.
5073 static ir_node *transform_node_Psi(ir_node *n) {
5075 return transform_node_Mux(n);
5078 } /* transform_node_Psi */
5081 * optimize sync nodes that have other syncs as input we simply add the inputs
5082 * of the other sync to our own inputs
5084 static ir_node *transform_node_Sync(ir_node *n) {
5085 int arity = get_Sync_n_preds(n);
5088 for (i = 0; i < arity;) {
5089 ir_node *pred = get_Sync_pred(n, i);
5093 if (!is_Sync(pred)) {
5101 pred_arity = get_Sync_n_preds(pred);
5102 for (j = 0; j < pred_arity; ++j) {
5103 ir_node *pred_pred = get_Sync_pred(pred, j);
5108 add_irn_n(n, pred_pred);
5112 if (get_Sync_pred(n, k) == pred_pred) break;
5117 /* rehash the sync node */
5118 add_identities(current_ir_graph->value_table, n);
5124 * Tries several [inplace] [optimizing] transformations and returns an
5125 * equivalent node. The difference to equivalent_node() is that these
5126 * transformations _do_ generate new nodes, and thus the old node must
5127 * not be freed even if the equivalent node isn't the old one.
5129 static ir_node *transform_node(ir_node *n) {
5133 * Transform_node is the only "optimizing transformation" that might
5134 * return a node with a different opcode. We iterate HERE until fixpoint
5135 * to get the final result.
5139 if (n->op->ops.transform_node)
5140 n = n->op->ops.transform_node(n);
5141 } while (oldn != n);
5144 } /* transform_node */
5147 * Sets the default transform node operation for an ir_op_ops.
5149 * @param code the opcode for the default operation
5150 * @param ops the operations initialized
5155 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5159 ops->transform_node = transform_node_##a; \
5196 } /* firm_set_default_transform_node */
5199 /* **************** Common Subexpression Elimination **************** */
5201 /** The size of the hash table used, should estimate the number of nodes
5203 #define N_IR_NODES 512
5205 /** Compares the attributes of two Const nodes. */
5206 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5207 return (get_Const_tarval(a) != get_Const_tarval(b))
5208 || (get_Const_type(a) != get_Const_type(b));
5209 } /* node_cmp_attr_Const */
5211 /** Compares the attributes of two Proj nodes. */
5212 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5213 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5214 } /* node_cmp_attr_Proj */
5216 /** Compares the attributes of two Filter nodes. */
5217 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5218 return get_Filter_proj(a) != get_Filter_proj(b);
5219 } /* node_cmp_attr_Filter */
5221 /** Compares the attributes of two Alloc nodes. */
5222 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5223 const alloc_attr *pa = get_irn_alloc_attr(a);
5224 const alloc_attr *pb = get_irn_alloc_attr(b);
5225 return (pa->where != pb->where) || (pa->type != pb->type);
5226 } /* node_cmp_attr_Alloc */
5228 /** Compares the attributes of two Free nodes. */
5229 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5230 const free_attr *pa = get_irn_free_attr(a);
5231 const free_attr *pb = get_irn_free_attr(b);
5232 return (pa->where != pb->where) || (pa->type != pb->type);
5233 } /* node_cmp_attr_Free */
5235 /** Compares the attributes of two SymConst nodes. */
5236 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5237 const symconst_attr *pa = get_irn_symconst_attr(a);
5238 const symconst_attr *pb = get_irn_symconst_attr(b);
5239 return (pa->kind != pb->kind)
5240 || (pa->sym.type_p != pb->sym.type_p)
5241 || (pa->tp != pb->tp);
5242 } /* node_cmp_attr_SymConst */
5244 /** Compares the attributes of two Call nodes. */
5245 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5246 return get_irn_call_attr(a) != get_irn_call_attr(b);
5247 } /* node_cmp_attr_Call */
5249 /** Compares the attributes of two Sel nodes. */
5250 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5251 const ir_entity *a_ent = get_Sel_entity(a);
5252 const ir_entity *b_ent = get_Sel_entity(b);
5254 (a_ent->kind != b_ent->kind) ||
5255 (a_ent->name != b_ent->name) ||
5256 (a_ent->owner != b_ent->owner) ||
5257 (a_ent->ld_name != b_ent->ld_name) ||
5258 (a_ent->type != b_ent->type);
5259 } /* node_cmp_attr_Sel */
5261 /** Compares the attributes of two Phi nodes. */
5262 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5263 /* we can only enter this function if both nodes have the same number of inputs,
5264 hence it is enough to check if one of them is a Phi0 */
5266 /* check the Phi0 pos attribute */
5267 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5270 } /* node_cmp_attr_Phi */
5272 /** Compares the attributes of two Conv nodes. */
5273 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5274 return get_Conv_strict(a) != get_Conv_strict(b);
5275 } /* node_cmp_attr_Conv */
5277 /** Compares the attributes of two Cast nodes. */
5278 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5279 return get_Cast_type(a) != get_Cast_type(b);
5280 } /* node_cmp_attr_Cast */
5282 /** Compares the attributes of two Load nodes. */
5283 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5284 if (get_Load_volatility(a) == volatility_is_volatile ||
5285 get_Load_volatility(b) == volatility_is_volatile)
5286 /* NEVER do CSE on volatile Loads */
5288 /* do not CSE Loads with different alignment. Be conservative. */
5289 if (get_Load_align(a) != get_Load_align(b))
5292 return get_Load_mode(a) != get_Load_mode(b);
5293 } /* node_cmp_attr_Load */
5295 /** Compares the attributes of two Store nodes. */
5296 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5297 /* do not CSE Stores with different alignment. Be conservative. */
5298 if (get_Store_align(a) != get_Store_align(b))
5301 /* NEVER do CSE on volatile Stores */
5302 return (get_Store_volatility(a) == volatility_is_volatile ||
5303 get_Store_volatility(b) == volatility_is_volatile);
5304 } /* node_cmp_attr_Store */
5306 /** Compares two exception attributes */
5307 static int node_cmp_exception(ir_node *a, ir_node *b) {
5308 const except_attr *ea = get_irn_except_attr(a);
5309 const except_attr *eb = get_irn_except_attr(b);
5311 return ea->pin_state != eb->pin_state;
5314 #define node_cmp_attr_Bound node_cmp_exception
5316 /** Compares the attributes of two Div nodes. */
5317 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5318 const divmod_attr *ma = get_irn_divmod_attr(a);
5319 const divmod_attr *mb = get_irn_divmod_attr(b);
5320 return ma->exc.pin_state != mb->exc.pin_state ||
5321 ma->res_mode != mb->res_mode ||
5322 ma->no_remainder != mb->no_remainder;
5323 } /* node_cmp_attr_Div */
5325 /** Compares the attributes of two DivMod nodes. */
5326 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5327 const divmod_attr *ma = get_irn_divmod_attr(a);
5328 const divmod_attr *mb = get_irn_divmod_attr(b);
5329 return ma->exc.pin_state != mb->exc.pin_state ||
5330 ma->res_mode != mb->res_mode;
5331 } /* node_cmp_attr_DivMod */
5333 /** Compares the attributes of two Mod nodes. */
5334 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5335 const divmod_attr *ma = get_irn_divmod_attr(a);
5336 const divmod_attr *mb = get_irn_divmod_attr(b);
5337 return ma->exc.pin_state != mb->exc.pin_state ||
5338 ma->res_mode != mb->res_mode;
5339 } /* node_cmp_attr_Mod */
5341 /** Compares the attributes of two Quot nodes. */
5342 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5343 const divmod_attr *ma = get_irn_divmod_attr(a);
5344 const divmod_attr *mb = get_irn_divmod_attr(b);
5345 return ma->exc.pin_state != mb->exc.pin_state ||
5346 ma->res_mode != mb->res_mode;
5347 } /* node_cmp_attr_Quot */
5349 /** Compares the attributes of two Confirm nodes. */
5350 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5351 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5352 } /* node_cmp_attr_Confirm */
5354 /** Compares the attributes of two ASM nodes. */
5355 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5357 const ir_asm_constraint *ca;
5358 const ir_asm_constraint *cb;
5361 if (get_ASM_text(a) != get_ASM_text(b))
5364 /* Should we really check the constraints here? Should be better, but is strange. */
5365 n = get_ASM_n_input_constraints(a);
5366 if (n != get_ASM_n_input_constraints(b))
5369 ca = get_ASM_input_constraints(a);
5370 cb = get_ASM_input_constraints(b);
5371 for (i = 0; i < n; ++i) {
5372 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5376 n = get_ASM_n_output_constraints(a);
5377 if (n != get_ASM_n_output_constraints(b))
5380 ca = get_ASM_output_constraints(a);
5381 cb = get_ASM_output_constraints(b);
5382 for (i = 0; i < n; ++i) {
5383 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5387 n = get_ASM_n_clobbers(a);
5388 if (n != get_ASM_n_clobbers(b))
5391 cla = get_ASM_clobbers(a);
5392 clb = get_ASM_clobbers(b);
5393 for (i = 0; i < n; ++i) {
5394 if (cla[i] != clb[i])
5398 } /* node_cmp_attr_ASM */
5401 * Set the default node attribute compare operation for an ir_op_ops.
5403 * @param code the opcode for the default operation
5404 * @param ops the operations initialized
5409 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5413 ops->node_cmp_attr = node_cmp_attr_##a; \
5444 } /* firm_set_default_node_cmp_attr */
5447 * Compare function for two nodes in the value table. Gets two
5448 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5450 int identities_cmp(const void *elt, const void *key) {
5451 ir_node *a = (ir_node *)elt;
5452 ir_node *b = (ir_node *)key;
5455 if (a == b) return 0;
5457 if ((get_irn_op(a) != get_irn_op(b)) ||
5458 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5460 /* compare if a's in and b's in are of equal length */
5461 irn_arity_a = get_irn_intra_arity(a);
5462 if (irn_arity_a != get_irn_intra_arity(b))
5465 if (get_irn_pinned(a) == op_pin_state_pinned) {
5466 /* for pinned nodes, the block inputs must be equal */
5467 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5469 } else if (! get_opt_global_cse()) {
5470 /* for block-local CSE both nodes must be in the same MacroBlock */
5471 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5475 /* compare a->in[0..ins] with b->in[0..ins] */
5476 for (i = 0; i < irn_arity_a; i++)
5477 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5481 * here, we already now that the nodes are identical except their
5484 if (a->op->ops.node_cmp_attr)
5485 return a->op->ops.node_cmp_attr(a, b);
5488 } /* identities_cmp */
5491 * Calculate a hash value of a node.
5493 unsigned ir_node_hash(ir_node *node) {
5497 if (node->op == op_Const) {
5498 /* special value for const, as they only differ in their tarval. */
5499 h = HASH_PTR(node->attr.con.tv);
5500 h = 9*h + HASH_PTR(get_irn_mode(node));
5501 } else if (node->op == op_SymConst) {
5502 /* special value for const, as they only differ in their symbol. */
5503 h = HASH_PTR(node->attr.symc.sym.type_p);
5504 h = 9*h + HASH_PTR(get_irn_mode(node));
5507 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5508 h = irn_arity = get_irn_intra_arity(node);
5510 /* consider all in nodes... except the block if not a control flow. */
5511 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5512 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5516 h = 9*h + HASH_PTR(get_irn_mode(node));
5518 h = 9*h + HASH_PTR(get_irn_op(node));
5522 } /* ir_node_hash */
5524 pset *new_identities(void) {
5525 return new_pset(identities_cmp, N_IR_NODES);
5526 } /* new_identities */
5528 void del_identities(pset *value_table) {
5529 del_pset(value_table);
5530 } /* del_identities */
5533 * Normalize a node by putting constants (and operands with larger
5534 * node index) on the right (operator side).
5536 * @param n The node to normalize
5538 static void normalize_node(ir_node *n) {
5539 if (is_op_commutative(get_irn_op(n))) {
5540 ir_node *l = get_binop_left(n);
5541 ir_node *r = get_binop_right(n);
5543 /* For commutative operators perform a OP b == b OP a but keep
5544 * constants on the RIGHT side. This helps greatly in some
5545 * optimizations. Moreover we use the idx number to make the form
5547 if (!operands_are_normalized(l, r)) {
5548 set_binop_left(n, r);
5549 set_binop_right(n, l);
5552 } /* normalize_node */
5555 * Update the nodes after a match in the value table. If both nodes have
5556 * the same MacroBlock but different Blocks, we must ensure that the node
5557 * with the dominating Block (the node that is near to the MacroBlock header
5558 * is stored in the table.
5559 * Because a MacroBlock has only one "non-exception" flow, we don't need
5560 * dominance info here: We known, that one block must dominate the other and
5561 * following the only block input will allow to find it.
5563 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5564 ir_node *known_blk, *new_block, *block, *mbh;
5566 if (get_opt_global_cse()) {
5567 /* Block inputs are meaning less */
5570 known_blk = get_irn_n(known_irn, -1);
5571 new_block = get_irn_n(new_ir_node, -1);
5572 if (known_blk == new_block) {
5573 /* already in the same block */
5577 * We expect the typical case when we built the graph. In that case, the
5578 * known_irn is already the upper one, so checking this should be faster.
5581 mbh = get_Block_MacroBlock(new_block);
5583 if (block == known_blk) {
5584 /* ok, we have found it: known_block dominates new_block as expected */
5589 * We have reached the MacroBlock header NOT founding
5590 * the known_block. new_block must dominate known_block.
5593 set_irn_n(known_irn, -1, new_block);
5596 assert(get_Block_n_cfgpreds(block) == 1);
5597 block = get_Block_cfgpred_block(block, 0);
5599 } /* update_value_table */
5602 * Return the canonical node computing the same value as n.
5604 * @param value_table The value table
5605 * @param n The node to lookup
5607 * Looks up the node in a hash table.
5609 * For Const nodes this is performed in the constructor, too. Const
5610 * nodes are extremely time critical because of their frequent use in
5611 * constant string arrays.
5613 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5616 if (!value_table) return n;
5620 o = pset_find(value_table, n, ir_node_hash(n));
5624 update_known_irn(o, n);
5631 * During construction we set the op_pin_state_pinned flag in the graph right when the
5632 * optimization is performed. The flag turning on procedure global cse could
5633 * be changed between two allocations. This way we are safe.
5635 * @param value_table The value table
5636 * @param n The node to lookup
5638 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5641 n = identify(value_table, n);
5642 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5643 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5645 } /* identify_cons */
5648 * Return the canonical node computing the same value as n.
5649 * Looks up the node in a hash table, enters it in the table
5650 * if it isn't there yet.
5652 * @param value_table the HashSet containing all nodes in the
5654 * @param n the node to look up
5656 * @return a node that computes the same value as n or n if no such
5657 * node could be found
5659 ir_node *identify_remember(pset *value_table, ir_node *n) {
5662 if (!value_table) return n;
5665 /* lookup or insert in hash table with given hash key. */
5666 o = pset_insert(value_table, n, ir_node_hash(n));
5669 update_known_irn(o, n);
5674 } /* identify_remember */
5676 /* Add a node to the identities value table. */
5677 void add_identities(pset *value_table, ir_node *node) {
5678 if (get_opt_cse() && is_no_Block(node))
5679 identify_remember(value_table, node);
5680 } /* add_identities */
5682 /* Visit each node in the value table of a graph. */
5683 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5685 ir_graph *rem = current_ir_graph;
5687 current_ir_graph = irg;
5688 foreach_pset(irg->value_table, node)
5690 current_ir_graph = rem;
5691 } /* visit_all_identities */
5694 * Garbage in, garbage out. If a node has a dead input, i.e., the
5695 * Bad node is input to the node, return the Bad node.
5697 static ir_node *gigo(ir_node *node) {
5699 ir_op *op = get_irn_op(node);
5701 /* remove garbage blocks by looking at control flow that leaves the block
5702 and replacing the control flow by Bad. */
5703 if (get_irn_mode(node) == mode_X) {
5704 ir_node *block = get_nodes_block(skip_Proj(node));
5706 /* Don't optimize nodes in immature blocks. */
5707 if (!get_Block_matured(block))
5709 /* Don't optimize End, may have Bads. */
5710 if (op == op_End) return node;
5712 if (is_Block(block)) {
5713 if (is_Block_dead(block)) {
5714 /* control flow from dead block is dead */
5718 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5719 if (!is_Bad(get_irn_n(block, i)))
5723 ir_graph *irg = get_irn_irg(block);
5724 /* the start block is never dead */
5725 if (block != get_irg_start_block(irg)
5726 && block != get_irg_end_block(irg)) {
5728 * Do NOT kill control flow without setting
5729 * the block to dead of bad things can happen:
5730 * We get a Block that is not reachable be irg_block_walk()
5731 * but can be found by irg_walk()!
5733 set_Block_dead(block);
5740 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5741 blocks predecessors is dead. */
5742 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5743 irn_arity = get_irn_arity(node);
5746 * Beware: we can only read the block of a non-floating node.
5748 if (is_irn_pinned_in_irg(node) &&
5749 is_Block_dead(get_nodes_block(skip_Proj(node))))
5752 for (i = 0; i < irn_arity; i++) {
5753 ir_node *pred = get_irn_n(node, i);
5758 /* Propagating Unknowns here seems to be a bad idea, because
5759 sometimes we need a node as a input and did not want that
5761 However, it might be useful to move this into a later phase
5762 (if you think that optimizing such code is useful). */
5763 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5764 return new_Unknown(get_irn_mode(node));
5769 /* With this code we violate the agreement that local_optimize
5770 only leaves Bads in Block, Phi and Tuple nodes. */
5771 /* If Block has only Bads as predecessors it's garbage. */
5772 /* If Phi has only Bads as predecessors it's garbage. */
5773 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5774 irn_arity = get_irn_arity(node);
5775 for (i = 0; i < irn_arity; i++) {
5776 if (!is_Bad(get_irn_n(node, i))) break;
5778 if (i == irn_arity) node = new_Bad();
5785 * These optimizations deallocate nodes from the obstack.
5786 * It can only be called if it is guaranteed that no other nodes
5787 * reference this one, i.e., right after construction of a node.
5789 * @param n The node to optimize
5791 * current_ir_graph must be set to the graph of the node!
5793 ir_node *optimize_node(ir_node *n) {
5796 ir_opcode iro = get_irn_opcode(n);
5798 /* Always optimize Phi nodes: part of the construction. */
5799 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5801 /* constant expression evaluation / constant folding */
5802 if (get_opt_constant_folding()) {
5803 /* neither constants nor Tuple values can be evaluated */
5804 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5805 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5806 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5807 /* try to evaluate */
5808 tv = computed_value(n);
5809 if (tv != tarval_bad) {
5811 ir_type *old_tp = get_irn_type(n);
5812 int i, arity = get_irn_arity(n);
5816 * Try to recover the type of the new expression.
5818 for (i = 0; i < arity && !old_tp; ++i)
5819 old_tp = get_irn_type(get_irn_n(n, i));
5822 * we MUST copy the node here temporary, because it's still needed
5823 * for DBG_OPT_CSTEVAL
5825 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5826 oldn = alloca(node_size);
5828 memcpy(oldn, n, node_size);
5829 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5831 /* ARG, copy the in array, we need it for statistics */
5832 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5834 /* note the inplace edges module */
5835 edges_node_deleted(n, current_ir_graph);
5837 /* evaluation was successful -- replace the node. */
5838 irg_kill_node(current_ir_graph, n);
5839 nw = new_Const(get_tarval_mode(tv), tv);
5841 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5842 set_Const_type(nw, old_tp);
5843 DBG_OPT_CSTEVAL(oldn, nw);
5844 tarval_enable_fp_ops(old_fp_mode);
5847 tarval_enable_fp_ops(old_fp_mode);
5851 /* remove unnecessary nodes */
5852 if (get_opt_constant_folding() ||
5853 (iro == iro_Phi) || /* always optimize these nodes. */
5855 (iro == iro_Proj) ||
5856 (iro == iro_Block) ) /* Flags tested local. */
5857 n = equivalent_node(n);
5859 /* Common Subexpression Elimination.
5861 * Checks whether n is already available.
5862 * The block input is used to distinguish different subexpressions. Right
5863 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5864 * subexpressions within a block.
5867 n = identify_cons(current_ir_graph->value_table, n);
5870 edges_node_deleted(oldn, current_ir_graph);
5872 /* We found an existing, better node, so we can deallocate the old node. */
5873 irg_kill_node(current_ir_graph, oldn);
5877 /* Some more constant expression evaluation that does not allow to
5879 iro = get_irn_opcode(n);
5880 if (get_opt_constant_folding() ||
5881 (iro == iro_Cond) ||
5882 (iro == iro_Proj)) /* Flags tested local. */
5883 n = transform_node(n);
5885 /* Remove nodes with dead (Bad) input.
5886 Run always for transformation induced Bads. */
5889 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5890 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5891 n = identify_remember(current_ir_graph->value_table, n);
5895 } /* optimize_node */
5899 * These optimizations never deallocate nodes (in place). This can cause dead
5900 * nodes lying on the obstack. Remove these by a dead node elimination,
5901 * i.e., a copying garbage collection.
5903 ir_node *optimize_in_place_2(ir_node *n) {
5906 ir_opcode iro = get_irn_opcode(n);
5908 if (!get_opt_optimize() && !is_Phi(n)) return n;
5910 /* constant expression evaluation / constant folding */
5911 if (get_opt_constant_folding()) {
5912 /* neither constants nor Tuple values can be evaluated */
5913 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5914 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5915 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5916 /* try to evaluate */
5917 tv = computed_value(n);
5918 if (tv != tarval_bad) {
5919 /* evaluation was successful -- replace the node. */
5920 ir_type *old_tp = get_irn_type(n);
5921 int i, arity = get_irn_arity(n);
5924 * Try to recover the type of the new expression.
5926 for (i = 0; i < arity && !old_tp; ++i)
5927 old_tp = get_irn_type(get_irn_n(n, i));
5929 n = new_Const(get_tarval_mode(tv), tv);
5931 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5932 set_Const_type(n, old_tp);
5934 DBG_OPT_CSTEVAL(oldn, n);
5935 tarval_enable_fp_ops(old_fp_mode);
5938 tarval_enable_fp_ops(old_fp_mode);
5942 /* remove unnecessary nodes */
5943 if (get_opt_constant_folding() ||
5944 (iro == iro_Phi) || /* always optimize these nodes. */
5945 (iro == iro_Id) || /* ... */
5946 (iro == iro_Proj) || /* ... */
5947 (iro == iro_Block) ) /* Flags tested local. */
5948 n = equivalent_node(n);
5950 /** common subexpression elimination **/
5951 /* Checks whether n is already available. */
5952 /* The block input is used to distinguish different subexpressions. Right
5953 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5954 subexpressions within a block. */
5955 if (get_opt_cse()) {
5956 n = identify(current_ir_graph->value_table, n);
5959 /* Some more constant expression evaluation. */
5960 iro = get_irn_opcode(n);
5961 if (get_opt_constant_folding() ||
5962 (iro == iro_Cond) ||
5963 (iro == iro_Proj)) /* Flags tested local. */
5964 n = transform_node(n);
5966 /* Remove nodes with dead (Bad) input.
5967 Run always for transformation induced Bads. */
5970 /* Now we can verify the node, as it has no dead inputs any more. */
5973 /* Now we have a legal, useful node. Enter it in hash table for cse.
5974 Blocks should be unique anyways. (Except the successor of start:
5975 is cse with the start block!) */
5976 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5977 n = identify_remember(current_ir_graph->value_table, n);
5980 } /* optimize_in_place_2 */
5983 * Wrapper for external use, set proper status bits after optimization.
5985 ir_node *optimize_in_place(ir_node *n) {
5986 /* Handle graph state */
5987 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5989 if (get_opt_global_cse())
5990 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5991 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5992 set_irg_outs_inconsistent(current_ir_graph);
5994 /* FIXME: Maybe we could also test whether optimizing the node can
5995 change the control graph. */
5996 set_irg_doms_inconsistent(current_ir_graph);
5997 return optimize_in_place_2(n);
5998 } /* optimize_in_place */
6001 * Sets the default operation for an ir_ops.
6003 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6004 ops = firm_set_default_computed_value(code, ops);
6005 ops = firm_set_default_equivalent_node(code, ops);
6006 ops = firm_set_default_transform_node(code, ops);
6007 ops = firm_set_default_node_cmp_attr(code, ops);
6008 ops = firm_set_default_get_type(code, ops);
6009 ops = firm_set_default_get_type_attr(code, ops);
6010 ops = firm_set_default_get_entity_attr(code, ops);
6013 } /* firm_set_default_operations */