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 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
942 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
945 n = equivalent_node_neutral_zero(n);
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);
1226 if (n_mode == a_mode) { /* No Conv necessary */
1227 if (get_Conv_strict(n)) {
1228 /* special case: the predecessor might be a also a Conv */
1230 if (! get_Conv_strict(a)) {
1231 /* first one is not strict, kick it */
1232 set_Conv_op(n, get_Conv_op(a));
1235 /* else both are strict conv, second is superfluous */
1236 } else if(is_Proj(a)) {
1237 ir_node *pred = get_Proj_pred(a);
1239 /* loads always return with the exact precision of n_mode */
1240 assert(get_Load_mode(pred) == n_mode);
1245 /* leave strict floating point Conv's */
1249 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1250 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1251 ir_node *b = get_Conv_op(a);
1252 ir_mode *b_mode = get_irn_mode(b);
1254 if (n_mode == b_mode) {
1255 if (n_mode == mode_b) {
1256 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1257 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1258 } else if (mode_is_int(n_mode)) {
1259 if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) {
1260 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1261 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1267 } /* equivalent_node_Conv */
1270 * A Cast may be removed if the type of the previous node
1271 * is already the type of the Cast.
1273 static ir_node *equivalent_node_Cast(ir_node *n) {
1275 ir_node *pred = get_Cast_op(n);
1277 if (get_irn_type(pred) == get_Cast_type(n)) {
1279 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1282 } /* equivalent_node_Cast */
1285 * Several optimizations:
1286 * - no Phi in start block.
1287 * - remove Id operators that are inputs to Phi
1288 * - fold Phi-nodes, iff they have only one predecessor except
1291 static ir_node *equivalent_node_Phi(ir_node *n) {
1296 ir_node *first_val = NULL; /* to shutup gcc */
1298 if (!get_opt_normalize()) return n;
1300 n_preds = get_Phi_n_preds(n);
1302 block = get_nodes_block(n);
1303 if ((is_Block_dead(block)) || /* Control dead */
1304 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1305 return new_Bad(); /* in the Start Block. */
1307 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1309 /* If the Block has a Bad pred, we also have one. */
1310 for (i = 0; i < n_preds; ++i)
1311 if (is_Bad(get_Block_cfgpred(block, i)))
1312 set_Phi_pred(n, i, new_Bad());
1314 /* Find first non-self-referencing input */
1315 for (i = 0; i < n_preds; ++i) {
1316 first_val = get_Phi_pred(n, i);
1317 if ( (first_val != n) /* not self pointer */
1319 && (! is_Bad(first_val))
1321 ) { /* value not dead */
1322 break; /* then found first value. */
1327 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1331 /* search for rest of inputs, determine if any of these
1332 are non-self-referencing */
1333 while (++i < n_preds) {
1334 ir_node *scnd_val = get_Phi_pred(n, i);
1335 if ( (scnd_val != n)
1336 && (scnd_val != first_val)
1338 && (! is_Bad(scnd_val))
1346 /* Fold, if no multiple distinct non-self-referencing inputs */
1348 DBG_OPT_PHI(oldn, n);
1351 } /* equivalent_node_Phi */
1354 * Several optimizations:
1355 * - no Sync in start block.
1356 * - fold Sync-nodes, iff they have only one predecessor except
1359 static ir_node *equivalent_node_Sync(ir_node *n) {
1360 int arity = get_Sync_n_preds(n);
1363 for (i = 0; i < arity;) {
1364 ir_node *pred = get_Sync_pred(n, i);
1367 /* Remove Bad predecessors */
1374 /* Remove duplicate predecessors */
1380 if (get_Sync_pred(n, j) == pred) {
1388 if (arity == 0) return new_Bad();
1389 if (arity == 1) return get_Sync_pred(n, 0);
1391 } /* equivalent_node_Sync */
1394 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1395 * ProjX(Load) and ProjX(Store).
1397 static ir_node *equivalent_node_Proj(ir_node *proj) {
1398 ir_node *oldn = proj;
1399 ir_node *a = get_Proj_pred(proj);
1402 /* Remove the Tuple/Proj combination. */
1403 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1404 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1405 DBG_OPT_TUPLE(oldn, a, proj);
1407 /* This should not happen! */
1408 assert(! "found a Proj with higher number than Tuple predecessors");
1411 } else if (get_irn_mode(proj) == mode_X) {
1412 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1413 /* Remove dead control flow -- early gigo(). */
1415 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1417 /* get the Load address */
1418 ir_node *addr = get_Load_ptr(a);
1419 ir_node *blk = get_irn_n(a, -1);
1422 if (value_not_null(addr, &confirm)) {
1423 if (confirm == NULL) {
1424 /* this node may float if it did not depend on a Confirm */
1425 set_irn_pinned(a, op_pin_state_floats);
1427 if (get_Proj_proj(proj) == pn_Load_X_except) {
1428 DBG_OPT_EXC_REM(proj);
1431 return new_r_Jmp(current_ir_graph, blk);
1433 } else if (is_Store(a)) {
1434 /* get the load/store address */
1435 ir_node *addr = get_Store_ptr(a);
1436 ir_node *blk = get_irn_n(a, -1);
1439 if (value_not_null(addr, &confirm)) {
1440 if (confirm == NULL) {
1441 /* this node may float if it did not depend on a Confirm */
1442 set_irn_pinned(a, op_pin_state_floats);
1444 if (get_Proj_proj(proj) == pn_Store_X_except) {
1445 DBG_OPT_EXC_REM(proj);
1448 return new_r_Jmp(current_ir_graph, blk);
1455 } /* equivalent_node_Proj */
1460 static ir_node *equivalent_node_Id(ir_node *n) {
1465 } while (get_irn_op(n) == op_Id);
1467 DBG_OPT_ID(oldn, n);
1469 } /* equivalent_node_Id */
1474 static ir_node *equivalent_node_Mux(ir_node *n)
1476 ir_node *oldn = n, *sel = get_Mux_sel(n);
1477 tarval *ts = value_of(sel);
1479 /* Mux(true, f, t) == t */
1480 if (ts == tarval_b_true) {
1481 n = get_Mux_true(n);
1482 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1484 /* Mux(false, f, t) == f */
1485 else if (ts == tarval_b_false) {
1486 n = get_Mux_false(n);
1487 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1489 /* Mux(v, x, x) == x */
1490 else if (get_Mux_false(n) == get_Mux_true(n)) {
1491 n = get_Mux_true(n);
1492 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1494 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1495 ir_node *cmp = get_Proj_pred(sel);
1496 long proj_nr = get_Proj_proj(sel);
1497 ir_node *f = get_Mux_false(n);
1498 ir_node *t = get_Mux_true(n);
1501 * Note further that these optimization work even for floating point
1502 * with NaN's because -NaN == NaN.
1503 * However, if +0 and -0 is handled differently, we cannot use the first one.
1506 ir_node *const cmp_l = get_Cmp_left(cmp);
1507 ir_node *const cmp_r = get_Cmp_right(cmp);
1511 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1512 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1514 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1521 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1522 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1524 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1531 * Note: normalization puts the constant on the right side,
1532 * so we check only one case.
1534 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1535 /* Mux(t CMP 0, X, t) */
1536 if (is_Minus(f) && get_Minus_op(f) == t) {
1537 /* Mux(t CMP 0, -t, t) */
1538 if (proj_nr == pn_Cmp_Eq) {
1539 /* Mux(t == 0, -t, t) ==> -t */
1541 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1542 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1543 /* Mux(t != 0, -t, t) ==> t */
1545 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1552 } /* equivalent_node_Mux */
1555 * Returns a equivalent node of a Psi: if a condition is true
1556 * and all previous conditions are false we know its value.
1557 * If all conditions are false its value is the default one.
1559 static ir_node *equivalent_node_Psi(ir_node *n) {
1561 return equivalent_node_Mux(n);
1563 } /* equivalent_node_Psi */
1566 * Optimize -a CMP -b into b CMP a.
1567 * This works only for for modes where unary Minus
1569 * Note that two-complement integers can Overflow
1570 * so it will NOT work.
1572 * For == and != can be handled in Proj(Cmp)
1574 static ir_node *equivalent_node_Cmp(ir_node *n) {
1575 ir_node *left = get_Cmp_left(n);
1576 ir_node *right = get_Cmp_right(n);
1578 if (is_Minus(left) && is_Minus(right) &&
1579 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1580 left = get_Minus_op(left);
1581 right = get_Minus_op(right);
1582 set_Cmp_left(n, right);
1583 set_Cmp_right(n, left);
1586 } /* equivalent_node_Cmp */
1589 * Remove Confirm nodes if setting is on.
1590 * Replace Confirms(x, '=', Constlike) by Constlike.
1592 static ir_node *equivalent_node_Confirm(ir_node *n) {
1593 ir_node *pred = get_Confirm_value(n);
1594 pn_Cmp pnc = get_Confirm_cmp(n);
1596 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1598 * rare case: two identical Confirms one after another,
1599 * replace the second one with the first.
1603 if (get_opt_remove_confirm())
1604 return get_Confirm_value(n);
1609 * Optimize CopyB(mem, x, x) into a Nop.
1611 static ir_node *equivalent_node_CopyB(ir_node *n) {
1612 ir_node *a = get_CopyB_dst(n);
1613 ir_node *b = get_CopyB_src(n);
1616 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1617 ir_node *mem = get_CopyB_mem(n);
1618 ir_node *blk = get_nodes_block(n);
1619 turn_into_tuple(n, pn_CopyB_max);
1620 set_Tuple_pred(n, pn_CopyB_M, mem);
1621 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1622 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1623 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1626 } /* equivalent_node_CopyB */
1629 * Optimize Bounds(idx, idx, upper) into idx.
1631 static ir_node *equivalent_node_Bound(ir_node *n) {
1632 ir_node *idx = get_Bound_index(n);
1633 ir_node *pred = skip_Proj(idx);
1636 if (is_Bound(pred)) {
1638 * idx was Bounds checked in the same MacroBlock previously,
1639 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1641 ir_node *lower = get_Bound_lower(n);
1642 ir_node *upper = get_Bound_upper(n);
1643 if (get_Bound_lower(pred) == lower &&
1644 get_Bound_upper(pred) == upper &&
1645 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1647 * One could expect that we simply return the previous
1648 * Bound here. However, this would be wrong, as we could
1649 * add an exception Proj to a new location then.
1650 * So, we must turn in into a tuple.
1656 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1657 ir_node *mem = get_Bound_mem(n);
1658 ir_node *blk = get_nodes_block(n);
1659 turn_into_tuple(n, pn_Bound_max);
1660 set_Tuple_pred(n, pn_Bound_M, mem);
1661 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1662 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1663 set_Tuple_pred(n, pn_Bound_res, idx);
1666 } /* equivalent_node_Bound */
1669 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1670 * perform no actual computation, as, e.g., the Id nodes. It does not create
1671 * new nodes. It is therefore safe to free n if the node returned is not n.
1672 * If a node returns a Tuple we can not just skip it. If the size of the
1673 * in array fits, we transform n into a tuple (e.g., Div).
1675 ir_node *equivalent_node(ir_node *n) {
1676 if (n->op->ops.equivalent_node)
1677 return n->op->ops.equivalent_node(n);
1679 } /* equivalent_node */
1682 * Sets the default equivalent node operation for an ir_op_ops.
1684 * @param code the opcode for the default operation
1685 * @param ops the operations initialized
1690 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1694 ops->equivalent_node = equivalent_node_##a; \
1734 } /* firm_set_default_equivalent_node */
1737 * Returns non-zero if a node is a Phi node
1738 * with all predecessors constant.
1740 static int is_const_Phi(ir_node *n) {
1743 if (! is_Phi(n) || get_irn_arity(n) == 0)
1745 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1746 if (! is_Const(get_irn_n(n, i)))
1749 } /* is_const_Phi */
1752 * Apply an evaluator on a binop with a constant operators (and one Phi).
1754 * @param phi the Phi node
1755 * @param other the other operand
1756 * @param eval an evaluator function
1757 * @param mode the mode of the result, may be different from the mode of the Phi!
1758 * @param left if non-zero, other is the left operand, else the right
1760 * @return a new Phi node if the conversion was successful, NULL else
1762 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1767 int i, n = get_irn_arity(phi);
1769 NEW_ARR_A(void *, res, n);
1771 for (i = 0; i < n; ++i) {
1772 pred = get_irn_n(phi, i);
1773 tv = get_Const_tarval(pred);
1774 tv = eval(other, tv);
1776 if (tv == tarval_bad) {
1777 /* folding failed, bad */
1783 for (i = 0; i < n; ++i) {
1784 pred = get_irn_n(phi, i);
1785 tv = get_Const_tarval(pred);
1786 tv = eval(tv, other);
1788 if (tv == tarval_bad) {
1789 /* folding failed, bad */
1795 irg = current_ir_graph;
1796 for (i = 0; i < n; ++i) {
1797 pred = get_irn_n(phi, i);
1798 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1799 mode, res[i], get_Const_type(pred));
1801 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1802 } /* apply_binop_on_phi */
1805 * Apply an evaluator on a binop with two constant Phi.
1807 * @param a the left Phi node
1808 * @param b the right Phi node
1809 * @param eval an evaluator function
1810 * @param mode the mode of the result, may be different from the mode of the Phi!
1812 * @return a new Phi node if the conversion was successful, NULL else
1814 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1815 tarval *tv_l, *tv_r, *tv;
1821 if (get_nodes_block(a) != get_nodes_block(b))
1824 n = get_irn_arity(a);
1825 NEW_ARR_A(void *, res, n);
1827 for (i = 0; i < n; ++i) {
1828 pred = get_irn_n(a, i);
1829 tv_l = get_Const_tarval(pred);
1830 pred = get_irn_n(b, i);
1831 tv_r = get_Const_tarval(pred);
1832 tv = eval(tv_l, tv_r);
1834 if (tv == tarval_bad) {
1835 /* folding failed, bad */
1840 irg = current_ir_graph;
1841 for (i = 0; i < n; ++i) {
1842 pred = get_irn_n(a, i);
1843 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1845 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1846 } /* apply_binop_on_2_phis */
1849 * Apply an evaluator on a unop with a constant operator (a Phi).
1851 * @param phi the Phi node
1852 * @param eval an evaluator function
1854 * @return a new Phi node if the conversion was successful, NULL else
1856 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1862 int i, n = get_irn_arity(phi);
1864 NEW_ARR_A(void *, res, n);
1865 for (i = 0; i < n; ++i) {
1866 pred = get_irn_n(phi, i);
1867 tv = get_Const_tarval(pred);
1870 if (tv == tarval_bad) {
1871 /* folding failed, bad */
1876 mode = get_irn_mode(phi);
1877 irg = current_ir_graph;
1878 for (i = 0; i < n; ++i) {
1879 pred = get_irn_n(phi, i);
1880 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1881 mode, res[i], get_Const_type(pred));
1883 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1884 } /* apply_unop_on_phi */
1887 * Apply a conversion on a constant operator (a Phi).
1889 * @param phi the Phi node
1891 * @return a new Phi node if the conversion was successful, NULL else
1893 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1898 int i, n = get_irn_arity(phi);
1900 NEW_ARR_A(void *, res, n);
1901 for (i = 0; i < n; ++i) {
1902 pred = get_irn_n(phi, i);
1903 tv = get_Const_tarval(pred);
1904 tv = tarval_convert_to(tv, mode);
1906 if (tv == tarval_bad) {
1907 /* folding failed, bad */
1912 irg = current_ir_graph;
1913 for (i = 0; i < n; ++i) {
1914 pred = get_irn_n(phi, i);
1915 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1916 mode, res[i], get_Const_type(pred));
1918 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1919 } /* apply_conv_on_phi */
1922 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1923 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1924 * If possible, remove the Conv's.
1926 static ir_node *transform_node_AddSub(ir_node *n) {
1927 ir_mode *mode = get_irn_mode(n);
1929 if (mode_is_reference(mode)) {
1930 ir_node *left = get_binop_left(n);
1931 ir_node *right = get_binop_right(n);
1932 unsigned ref_bits = get_mode_size_bits(mode);
1934 if (is_Conv(left)) {
1935 ir_mode *lmode = get_irn_mode(left);
1936 unsigned bits = get_mode_size_bits(lmode);
1938 if (ref_bits == bits &&
1939 mode_is_int(lmode) &&
1940 get_mode_arithmetic(lmode) == irma_twos_complement) {
1941 ir_node *pre = get_Conv_op(left);
1942 ir_mode *pre_mode = get_irn_mode(pre);
1944 if (mode_is_int(pre_mode) &&
1945 get_mode_size_bits(pre_mode) == bits &&
1946 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1947 /* ok, this conv just changes to sign, moreover the calculation
1948 * is done with same number of bits as our address mode, so
1949 * we can ignore the conv as address calculation can be viewed
1950 * as either signed or unsigned
1952 set_binop_left(n, pre);
1957 if (is_Conv(right)) {
1958 ir_mode *rmode = get_irn_mode(right);
1959 unsigned bits = get_mode_size_bits(rmode);
1961 if (ref_bits == bits &&
1962 mode_is_int(rmode) &&
1963 get_mode_arithmetic(rmode) == irma_twos_complement) {
1964 ir_node *pre = get_Conv_op(right);
1965 ir_mode *pre_mode = get_irn_mode(pre);
1967 if (mode_is_int(pre_mode) &&
1968 get_mode_size_bits(pre_mode) == bits &&
1969 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1970 /* ok, this conv just changes to sign, moreover the calculation
1971 * is done with same number of bits as our address mode, so
1972 * we can ignore the conv as address calculation can be viewed
1973 * as either signed or unsigned
1975 set_binop_right(n, pre);
1980 /* let address arithmetic use unsigned modes */
1981 if (is_Const(right)) {
1982 ir_mode *rmode = get_irn_mode(right);
1984 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1985 /* convert a AddP(P, *s) into AddP(P, *u) */
1986 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1988 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
1989 set_binop_right(n, pre);
1994 } /* transform_node_AddSub */
1996 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1998 if (is_Const(b) && is_const_Phi(a)) { \
1999 /* check for Op(Phi, Const) */ \
2000 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2002 else if (is_Const(a) && is_const_Phi(b)) { \
2003 /* check for Op(Const, Phi) */ \
2004 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2006 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2007 /* check for Op(Phi, Phi) */ \
2008 c = apply_binop_on_2_phis(a, b, eval, mode); \
2011 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2015 #define HANDLE_UNOP_PHI(eval, a, c) \
2017 if (is_const_Phi(a)) { \
2018 /* check for Op(Phi) */ \
2019 c = apply_unop_on_phi(a, eval); \
2021 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2027 * Do the AddSub optimization, then Transform
2028 * Constant folding on Phi
2029 * Add(a,a) -> Mul(a, 2)
2030 * Add(Mul(a, x), a) -> Mul(a, x+1)
2031 * if the mode is integer or float.
2032 * Transform Add(a,-b) into Sub(a,b).
2033 * Reassociation might fold this further.
2035 static ir_node *transform_node_Add(ir_node *n) {
2037 ir_node *a, *b, *c, *oldn = n;
2039 n = transform_node_AddSub(n);
2041 a = get_Add_left(n);
2042 b = get_Add_right(n);
2044 mode = get_irn_mode(n);
2045 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2047 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2048 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2051 if (mode_is_num(mode)) {
2052 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2053 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2054 ir_node *block = get_irn_n(n, -1);
2057 get_irn_dbg_info(n),
2061 new_r_Const_long(current_ir_graph, block, mode, 2),
2063 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2068 get_irn_dbg_info(n),
2074 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2079 get_irn_dbg_info(n),
2085 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2088 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2089 /* Here we rely on constants be on the RIGHT side */
2091 ir_node *op = get_Not_op(a);
2093 if (is_Const(b) && is_Const_one(b)) {
2095 ir_node *blk = get_irn_n(n, -1);
2096 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2097 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2102 ir_node *blk = get_irn_n(n, -1);
2103 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2104 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2109 ir_node *op = get_Not_op(b);
2113 ir_node *blk = get_irn_n(n, -1);
2114 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2115 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2122 } /* transform_node_Add */
2125 * returns -cnst or NULL if impossible
2127 static ir_node *const_negate(ir_node *cnst) {
2128 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2129 dbg_info *dbgi = get_irn_dbg_info(cnst);
2130 ir_graph *irg = get_irn_irg(cnst);
2131 ir_node *block = get_nodes_block(cnst);
2132 ir_mode *mode = get_irn_mode(cnst);
2133 if (tv == tarval_bad) return NULL;
2134 return new_rd_Const(dbgi, irg, block, mode, tv);
2138 * Do the AddSub optimization, then Transform
2139 * Constant folding on Phi
2140 * Sub(0,a) -> Minus(a)
2141 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2142 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2143 * Sub(Add(a, x), x) -> a
2144 * Sub(x, Add(x, a)) -> -a
2145 * Sub(x, Const) -> Add(x, -Const)
2147 static ir_node *transform_node_Sub(ir_node *n) {
2152 n = transform_node_AddSub(n);
2154 a = get_Sub_left(n);
2155 b = get_Sub_right(n);
2157 mode = get_irn_mode(n);
2160 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2162 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2163 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2166 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2167 /* a - C -> a + (-C) */
2168 ir_node *cnst = const_negate(b);
2170 ir_node *block = get_nodes_block(n);
2171 dbg_info *dbgi = get_irn_dbg_info(n);
2172 ir_graph *irg = get_irn_irg(n);
2174 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2175 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2180 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2181 ir_graph *irg = current_ir_graph;
2182 dbg_info *dbg = get_irn_dbg_info(n);
2183 ir_node *block = get_nodes_block(n);
2184 ir_node *left = get_Minus_op(a);
2185 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2187 n = new_rd_Minus(dbg, irg, block, add, mode);
2188 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2190 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2191 ir_graph *irg = current_ir_graph;
2192 dbg_info *dbg = get_irn_dbg_info(n);
2193 ir_node *block = get_nodes_block(n);
2194 ir_node *right = get_Minus_op(b);
2196 n = new_rd_Add(dbg, irg, block, a, right, mode);
2197 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2199 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2200 ir_graph *irg = current_ir_graph;
2201 dbg_info *s_dbg = get_irn_dbg_info(b);
2202 ir_node *s_block = get_nodes_block(b);
2203 ir_node *s_left = get_Sub_right(b);
2204 ir_node *s_right = get_Sub_left(b);
2205 ir_mode *s_mode = get_irn_mode(b);
2206 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2207 dbg_info *a_dbg = get_irn_dbg_info(n);
2208 ir_node *a_block = get_nodes_block(n);
2210 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2211 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2213 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2214 ir_node *m_right = get_Mul_right(b);
2215 if (is_Const(m_right)) {
2216 ir_node *cnst2 = const_negate(m_right);
2217 if (cnst2 != NULL) {
2218 ir_graph *irg = current_ir_graph;
2219 dbg_info *m_dbg = get_irn_dbg_info(b);
2220 ir_node *m_block = get_nodes_block(b);
2221 ir_node *m_left = get_Mul_left(b);
2222 ir_mode *m_mode = get_irn_mode(b);
2223 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2224 dbg_info *a_dbg = get_irn_dbg_info(n);
2225 ir_node *a_block = get_nodes_block(n);
2227 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2228 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2234 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2235 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2237 get_irn_dbg_info(n),
2242 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2246 if (mode_wrap_around(mode)) {
2247 ir_node *left = get_Add_left(a);
2248 ir_node *right = get_Add_right(a);
2250 /* FIXME: Does the Conv's work only for two complement or generally? */
2252 if (mode != get_irn_mode(right)) {
2253 /* This Sub is an effective Cast */
2254 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2257 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2259 } else if (right == b) {
2260 if (mode != get_irn_mode(left)) {
2261 /* This Sub is an effective Cast */
2262 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2265 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2271 if (mode_wrap_around(mode)) {
2272 ir_node *left = get_Add_left(b);
2273 ir_node *right = get_Add_right(b);
2275 /* FIXME: Does the Conv's work only for two complement or generally? */
2277 ir_mode *r_mode = get_irn_mode(right);
2279 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2280 if (mode != r_mode) {
2281 /* This Sub is an effective Cast */
2282 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2284 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2286 } else if (right == a) {
2287 ir_mode *l_mode = get_irn_mode(left);
2289 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2290 if (mode != l_mode) {
2291 /* This Sub is an effective Cast */
2292 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2294 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2299 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2300 ir_mode *mode = get_irn_mode(a);
2302 if (mode == get_irn_mode(b)) {
2304 ir_node *op_a = get_Conv_op(a);
2305 ir_node *op_b = get_Conv_op(b);
2307 /* check if it's allowed to skip the conv */
2308 ma = get_irn_mode(op_a);
2309 mb = get_irn_mode(op_b);
2311 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2312 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2315 set_Sub_right(n, b);
2321 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2322 if (!is_reassoc_running() && is_Mul(a)) {
2323 ir_node *ma = get_Mul_left(a);
2324 ir_node *mb = get_Mul_right(a);
2327 ir_node *blk = get_irn_n(n, -1);
2329 get_irn_dbg_info(n),
2330 current_ir_graph, blk,
2333 get_irn_dbg_info(n),
2334 current_ir_graph, blk,
2336 new_r_Const_long(current_ir_graph, blk, mode, 1),
2339 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2341 } else if (mb == b) {
2342 ir_node *blk = get_irn_n(n, -1);
2344 get_irn_dbg_info(n),
2345 current_ir_graph, blk,
2348 get_irn_dbg_info(n),
2349 current_ir_graph, blk,
2351 new_r_Const_long(current_ir_graph, blk, mode, 1),
2354 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2358 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2359 ir_node *x = get_Sub_left(a);
2360 ir_node *y = get_Sub_right(a);
2361 ir_node *blk = get_irn_n(n, -1);
2362 ir_mode *m_b = get_irn_mode(b);
2363 ir_mode *m_y = get_irn_mode(y);
2367 /* Determine the right mode for the Add. */
2370 else if (mode_is_reference(m_b))
2372 else if (mode_is_reference(m_y))
2376 * Both modes are different but none is reference,
2377 * happens for instance in SubP(SubP(P, Iu), Is).
2378 * We have two possibilities here: Cast or ignore.
2379 * Currently we ignore this case.
2384 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2386 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2387 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2391 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2392 if (is_Const(a) && is_Not(b)) {
2393 /* c - ~X = X + (c+1) */
2394 tarval *tv = get_Const_tarval(a);
2396 tv = tarval_add(tv, get_mode_one(mode));
2397 if (tv != tarval_bad) {
2398 ir_node *blk = get_irn_n(n, -1);
2399 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2400 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2401 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2407 } /* transform_node_Sub */
2410 * Several transformation done on n*n=2n bits mul.
2411 * These transformations must be done here because new nodes may be produced.
2413 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2415 ir_node *a = get_Mul_left(n);
2416 ir_node *b = get_Mul_right(n);
2417 tarval *ta = value_of(a);
2418 tarval *tb = value_of(b);
2419 ir_mode *smode = get_irn_mode(a);
2421 if (ta == get_mode_one(smode)) {
2422 /* (L)1 * (L)b = (L)b */
2423 ir_node *blk = get_irn_n(n, -1);
2424 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2425 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2428 else if (ta == get_mode_minus_one(smode)) {
2429 /* (L)-1 * (L)b = (L)b */
2430 ir_node *blk = get_irn_n(n, -1);
2431 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2432 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2433 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2436 if (tb == get_mode_one(smode)) {
2437 /* (L)a * (L)1 = (L)a */
2438 ir_node *blk = get_irn_n(a, -1);
2439 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2440 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2443 else if (tb == get_mode_minus_one(smode)) {
2444 /* (L)a * (L)-1 = (L)-a */
2445 ir_node *blk = get_irn_n(n, -1);
2446 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2447 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2448 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2455 * Transform Mul(a,-1) into -a.
2456 * Do constant evaluation of Phi nodes.
2457 * Do architecture dependent optimizations on Mul nodes
2459 static ir_node *transform_node_Mul(ir_node *n) {
2460 ir_node *c, *oldn = n;
2461 ir_mode *mode = get_irn_mode(n);
2462 ir_node *a = get_Mul_left(n);
2463 ir_node *b = get_Mul_right(n);
2465 if (is_Bad(a) || is_Bad(b))
2468 if (mode != get_irn_mode(a))
2469 return transform_node_Mul2n(n, mode);
2471 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2473 if (mode_is_signed(mode)) {
2476 if (value_of(a) == get_mode_minus_one(mode))
2478 else if (value_of(b) == get_mode_minus_one(mode))
2481 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2482 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2487 if (is_Const(b)) { /* (-a) * const -> a * -const */
2488 ir_node *cnst = const_negate(b);
2490 dbg_info *dbgi = get_irn_dbg_info(n);
2491 ir_node *block = get_nodes_block(n);
2492 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2493 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2496 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2497 dbg_info *dbgi = get_irn_dbg_info(n);
2498 ir_node *block = get_nodes_block(n);
2499 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2500 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2502 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2503 ir_node *sub_l = get_Sub_left(b);
2504 ir_node *sub_r = get_Sub_right(b);
2505 dbg_info *dbgi = get_irn_dbg_info(n);
2506 ir_graph *irg = current_ir_graph;
2507 ir_node *block = get_nodes_block(n);
2508 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2509 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2510 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2513 } else if (is_Minus(b)) {
2514 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2515 ir_node *sub_l = get_Sub_left(a);
2516 ir_node *sub_r = get_Sub_right(a);
2517 dbg_info *dbgi = get_irn_dbg_info(n);
2518 ir_graph *irg = current_ir_graph;
2519 ir_node *block = get_nodes_block(n);
2520 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2521 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2522 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2526 if (get_mode_arithmetic(mode) == irma_ieee754) {
2528 tarval *tv = get_Const_tarval(a);
2529 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2530 /* 2.0 * b = b + b */
2531 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2532 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2536 else if (is_Const(b)) {
2537 tarval *tv = get_Const_tarval(b);
2538 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2539 /* a * 2.0 = a + a */
2540 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2541 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2546 return arch_dep_replace_mul_with_shifts(n);
2547 } /* transform_node_Mul */
2550 * Transform a Div Node.
2552 static ir_node *transform_node_Div(ir_node *n) {
2553 ir_mode *mode = get_Div_resmode(n);
2554 ir_node *a = get_Div_left(n);
2555 ir_node *b = get_Div_right(n);
2559 if (is_Const(b) && is_const_Phi(a)) {
2560 /* check for Div(Phi, Const) */
2561 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2563 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2567 else if (is_Const(a) && is_const_Phi(b)) {
2568 /* check for Div(Const, Phi) */
2569 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2571 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2575 else if (is_const_Phi(a) && is_const_Phi(b)) {
2576 /* check for Div(Phi, Phi) */
2577 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2579 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2586 if (tv != tarval_bad) {
2587 value = new_Const(get_tarval_mode(tv), tv);
2589 DBG_OPT_CSTEVAL(n, value);
2592 ir_node *a = get_Div_left(n);
2593 ir_node *b = get_Div_right(n);
2596 if (a == b && value_not_zero(a, &dummy)) {
2597 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2598 value = new_Const(mode, get_mode_one(mode));
2599 DBG_OPT_CSTEVAL(n, value);
2602 if (mode_is_signed(mode) && is_Const(b)) {
2603 tarval *tv = get_Const_tarval(b);
2605 if (tv == get_mode_minus_one(mode)) {
2607 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2608 DBG_OPT_CSTEVAL(n, value);
2612 /* Try architecture dependent optimization */
2613 value = arch_dep_replace_div_by_const(n);
2621 /* Turn Div into a tuple (mem, jmp, bad, value) */
2622 mem = get_Div_mem(n);
2623 blk = get_irn_n(n, -1);
2625 /* skip a potential Pin */
2627 mem = get_Pin_op(mem);
2628 turn_into_tuple(n, pn_Div_max);
2629 set_Tuple_pred(n, pn_Div_M, mem);
2630 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2631 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2632 set_Tuple_pred(n, pn_Div_res, value);
2635 } /* transform_node_Div */
2638 * Transform a Mod node.
2640 static ir_node *transform_node_Mod(ir_node *n) {
2641 ir_mode *mode = get_Mod_resmode(n);
2642 ir_node *a = get_Mod_left(n);
2643 ir_node *b = get_Mod_right(n);
2647 if (is_Const(b) && is_const_Phi(a)) {
2648 /* check for Div(Phi, Const) */
2649 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2651 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2655 else if (is_Const(a) && is_const_Phi(b)) {
2656 /* check for Div(Const, Phi) */
2657 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2659 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2663 else if (is_const_Phi(a) && is_const_Phi(b)) {
2664 /* check for Div(Phi, Phi) */
2665 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2667 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2674 if (tv != tarval_bad) {
2675 value = new_Const(get_tarval_mode(tv), tv);
2677 DBG_OPT_CSTEVAL(n, value);
2680 ir_node *a = get_Mod_left(n);
2681 ir_node *b = get_Mod_right(n);
2684 if (a == b && value_not_zero(a, &dummy)) {
2685 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2686 value = new_Const(mode, get_mode_null(mode));
2687 DBG_OPT_CSTEVAL(n, value);
2690 if (mode_is_signed(mode) && is_Const(b)) {
2691 tarval *tv = get_Const_tarval(b);
2693 if (tv == get_mode_minus_one(mode)) {
2695 value = new_Const(mode, get_mode_null(mode));
2696 DBG_OPT_CSTEVAL(n, value);
2700 /* Try architecture dependent optimization */
2701 value = arch_dep_replace_mod_by_const(n);
2709 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2710 mem = get_Mod_mem(n);
2711 blk = get_irn_n(n, -1);
2713 /* skip a potential Pin */
2715 mem = get_Pin_op(mem);
2716 turn_into_tuple(n, pn_Mod_max);
2717 set_Tuple_pred(n, pn_Mod_M, mem);
2718 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2719 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2720 set_Tuple_pred(n, pn_Mod_res, value);
2723 } /* transform_node_Mod */
2726 * Transform a DivMod node.
2728 static ir_node *transform_node_DivMod(ir_node *n) {
2730 ir_node *a = get_DivMod_left(n);
2731 ir_node *b = get_DivMod_right(n);
2732 ir_mode *mode = get_DivMod_resmode(n);
2737 if (is_Const(b) && is_const_Phi(a)) {
2738 /* check for Div(Phi, Const) */
2739 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2740 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2742 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2743 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2747 else if (is_Const(a) && is_const_Phi(b)) {
2748 /* check for Div(Const, Phi) */
2749 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2750 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2752 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2753 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2757 else if (is_const_Phi(a) && is_const_Phi(b)) {
2758 /* check for Div(Phi, Phi) */
2759 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2760 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2762 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2763 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2770 if (tb != tarval_bad) {
2771 if (tb == get_mode_one(get_tarval_mode(tb))) {
2773 vb = new_Const(mode, get_mode_null(mode));
2774 DBG_OPT_CSTEVAL(n, vb);
2776 } else if (ta != tarval_bad) {
2777 tarval *resa, *resb;
2778 resa = tarval_div(ta, tb);
2779 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2780 Jmp for X result!? */
2781 resb = tarval_mod(ta, tb);
2782 if (resb == tarval_bad) return n; /* Causes exception! */
2783 va = new_Const(mode, resa);
2784 vb = new_Const(mode, resb);
2785 DBG_OPT_CSTEVAL(n, va);
2786 DBG_OPT_CSTEVAL(n, vb);
2788 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2789 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2790 vb = new_Const(mode, get_mode_null(mode));
2791 DBG_OPT_CSTEVAL(n, va);
2792 DBG_OPT_CSTEVAL(n, vb);
2794 } else { /* Try architecture dependent optimization */
2797 arch_dep_replace_divmod_by_const(&va, &vb, n);
2798 evaluated = va != NULL;
2800 } else if (a == b) {
2801 if (value_not_zero(a, &dummy)) {
2803 va = new_Const(mode, get_mode_one(mode));
2804 vb = new_Const(mode, get_mode_null(mode));
2805 DBG_OPT_CSTEVAL(n, va);
2806 DBG_OPT_CSTEVAL(n, vb);
2809 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2812 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2813 /* 0 / non-Const = 0 */
2818 if (evaluated) { /* replace by tuple */
2822 mem = get_DivMod_mem(n);
2823 /* skip a potential Pin */
2825 mem = get_Pin_op(mem);
2827 blk = get_irn_n(n, -1);
2828 turn_into_tuple(n, pn_DivMod_max);
2829 set_Tuple_pred(n, pn_DivMod_M, mem);
2830 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2831 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2832 set_Tuple_pred(n, pn_DivMod_res_div, va);
2833 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2837 } /* transform_node_DivMod */
2840 * Optimize x / c to x * (1/c)
2842 static ir_node *transform_node_Quot(ir_node *n) {
2843 ir_mode *mode = get_Quot_resmode(n);
2846 if (get_mode_arithmetic(mode) == irma_ieee754) {
2847 ir_node *b = get_Quot_right(n);
2850 tarval *tv = get_Const_tarval(b);
2854 * Floating point constant folding might be disabled here to
2856 * However, as we check for exact result, doing it is safe.
2859 rem = tarval_enable_fp_ops(1);
2860 tv = tarval_quo(get_mode_one(mode), tv);
2861 (void)tarval_enable_fp_ops(rem);
2863 /* Do the transformation if the result is either exact or we are not
2864 using strict rules. */
2865 if (tv != tarval_bad &&
2866 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2867 ir_node *blk = get_irn_n(n, -1);
2868 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2869 ir_node *a = get_Quot_left(n);
2870 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2871 ir_node *mem = get_Quot_mem(n);
2873 /* skip a potential Pin */
2875 mem = get_Pin_op(mem);
2876 turn_into_tuple(n, pn_Quot_max);
2877 set_Tuple_pred(n, pn_Quot_M, mem);
2878 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2879 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2880 set_Tuple_pred(n, pn_Quot_res, m);
2881 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2886 } /* transform_node_Quot */
2889 * Optimize Abs(x) into x if x is Confirmed >= 0
2890 * Optimize Abs(x) into -x if x is Confirmed <= 0
2891 * Optimize Abs(-x) int Abs(x)
2893 static ir_node *transform_node_Abs(ir_node *n) {
2894 ir_node *c, *oldn = n;
2895 ir_node *a = get_Abs_op(n);
2898 HANDLE_UNOP_PHI(tarval_abs, a, c);
2900 switch (classify_value_sign(a)) {
2901 case value_classified_negative:
2902 mode = get_irn_mode(n);
2905 * We can replace the Abs by -x here.
2906 * We even could add a new Confirm here
2907 * (if not twos complement)
2909 * Note that -x would create a new node, so we could
2910 * not run it in the equivalent_node() context.
2912 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2913 get_nodes_block(n), a, mode);
2915 DBG_OPT_CONFIRM(oldn, n);
2917 case value_classified_positive:
2918 /* n is positive, Abs is not needed */
2921 DBG_OPT_CONFIRM(oldn, n);
2927 /* Abs(-x) = Abs(x) */
2928 mode = get_irn_mode(n);
2929 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2930 get_nodes_block(n), get_Minus_op(a), mode);
2931 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2935 } /* transform_node_Abs */
2938 * Transform a Cond node.
2940 * Replace the Cond by a Jmp if it branches on a constant
2943 static ir_node *transform_node_Cond(ir_node *n) {
2946 ir_node *a = get_Cond_selector(n);
2947 tarval *ta = value_of(a);
2949 /* we need block info which is not available in floating irgs */
2950 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2953 if ((ta != tarval_bad) &&
2954 (get_irn_mode(a) == mode_b) &&
2955 (get_opt_unreachable_code())) {
2956 /* It's a boolean Cond, branching on a boolean constant.
2957 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2958 ir_node *blk = get_nodes_block(n);
2959 jmp = new_r_Jmp(current_ir_graph, blk);
2960 turn_into_tuple(n, pn_Cond_max);
2961 if (ta == tarval_b_true) {
2962 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2963 set_Tuple_pred(n, pn_Cond_true, jmp);
2965 set_Tuple_pred(n, pn_Cond_false, jmp);
2966 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2968 /* We might generate an endless loop, so keep it alive. */
2969 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2972 } /* transform_node_Cond */
2975 * Prototype of a recursive transform function
2976 * for bitwise distributive transformations.
2978 typedef ir_node* (*recursive_transform)(ir_node *n);
2981 * makes use of distributive laws for and, or, eor
2982 * and(a OP c, b OP c) -> and(a, b) OP c
2983 * note, might return a different op than n
2985 static ir_node *transform_bitwise_distributive(ir_node *n,
2986 recursive_transform trans_func)
2989 ir_node *a = get_binop_left(n);
2990 ir_node *b = get_binop_right(n);
2991 ir_op *op = get_irn_op(a);
2992 ir_op *op_root = get_irn_op(n);
2994 if(op != get_irn_op(b))
2997 if (op == op_Conv) {
2998 ir_node *a_op = get_Conv_op(a);
2999 ir_node *b_op = get_Conv_op(b);
3000 ir_mode *a_mode = get_irn_mode(a_op);
3001 ir_mode *b_mode = get_irn_mode(b_op);
3002 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3003 ir_node *blk = get_irn_n(n, -1);
3006 set_binop_left(n, a_op);
3007 set_binop_right(n, b_op);
3008 set_irn_mode(n, a_mode);
3010 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3012 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3018 /* nothing to gain here */
3022 if (op == op_Shrs || op == op_Shr || op == op_Shl
3023 || op == op_And || op == op_Or || op == op_Eor) {
3024 ir_node *a_left = get_binop_left(a);
3025 ir_node *a_right = get_binop_right(a);
3026 ir_node *b_left = get_binop_left(b);
3027 ir_node *b_right = get_binop_right(b);
3029 ir_node *op1 = NULL;
3030 ir_node *op2 = NULL;
3032 if (is_op_commutative(op)) {
3033 if (a_left == b_left) {
3037 } else if(a_left == b_right) {
3041 } else if(a_right == b_left) {
3047 if(a_right == b_right) {
3054 /* (a sop c) & (b sop c) => (a & b) sop c */
3055 ir_node *blk = get_irn_n(n, -1);
3057 ir_node *new_n = exact_copy(n);
3058 set_binop_left(new_n, op1);
3059 set_binop_right(new_n, op2);
3060 new_n = trans_func(new_n);
3062 if(op_root == op_Eor && op == op_Or) {
3063 dbg_info *dbgi = get_irn_dbg_info(n);
3064 ir_graph *irg = current_ir_graph;
3065 ir_mode *mode = get_irn_mode(c);
3067 c = new_rd_Not(dbgi, irg, blk, c, mode);
3068 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3071 set_nodes_block(n, blk);
3072 set_binop_left(n, new_n);
3073 set_binop_right(n, c);
3074 add_identities(current_ir_graph->value_table, n);
3077 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3088 static ir_node *transform_node_And(ir_node *n) {
3089 ir_node *c, *oldn = n;
3090 ir_node *a = get_And_left(n);
3091 ir_node *b = get_And_right(n);
3094 mode = get_irn_mode(n);
3095 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3097 /* we can evaluate 2 Projs of the same Cmp */
3098 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3099 ir_node *pred_a = get_Proj_pred(a);
3100 ir_node *pred_b = get_Proj_pred(b);
3101 if (pred_a == pred_b) {
3102 dbg_info *dbgi = get_irn_dbg_info(n);
3103 ir_node *block = get_nodes_block(pred_a);
3104 pn_Cmp pn_a = get_Proj_proj(a);
3105 pn_Cmp pn_b = get_Proj_proj(b);
3106 /* yes, we can simply calculate with pncs */
3107 pn_Cmp new_pnc = pn_a & pn_b;
3109 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3114 ir_node *op = get_Not_op(b);
3116 ir_node *ba = get_And_left(op);
3117 ir_node *bb = get_And_right(op);
3119 /* it's enough to test the following cases due to normalization! */
3120 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3121 /* (a|b) & ~(a&b) = a^b */
3122 ir_node *block = get_nodes_block(n);
3124 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3125 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3133 ir_node *op = get_Not_op(a);
3135 ir_node *aa = get_And_left(op);
3136 ir_node *ab = get_And_right(op);
3138 /* it's enough to test the following cases due to normalization! */
3139 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3140 /* (a|b) & ~(a&b) = a^b */
3141 ir_node *block = get_nodes_block(n);
3143 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3151 ir_node *al = get_Eor_left(a);
3152 ir_node *ar = get_Eor_right(a);
3155 /* (b ^ a) & b -> ~a & b */
3156 dbg_info *dbg = get_irn_dbg_info(n);
3157 ir_node *block = get_nodes_block(n);
3159 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3160 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3161 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3165 /* (a ^ b) & b -> ~a & b */
3166 dbg_info *dbg = get_irn_dbg_info(n);
3167 ir_node *block = get_nodes_block(n);
3169 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3170 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3171 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3176 ir_node *bl = get_Eor_left(b);
3177 ir_node *br = get_Eor_right(b);
3180 /* a & (a ^ b) -> a & ~b */
3181 dbg_info *dbg = get_irn_dbg_info(n);
3182 ir_node *block = get_nodes_block(n);
3184 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3185 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3186 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3190 /* a & (b ^ a) -> a & ~b */
3191 dbg_info *dbg = get_irn_dbg_info(n);
3192 ir_node *block = get_nodes_block(n);
3194 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3195 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3196 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3200 if (is_Not(a) && is_Not(b)) {
3201 /* ~a & ~b = ~(a|b) */
3202 ir_node *block = get_nodes_block(n);
3203 ir_mode *mode = get_irn_mode(n);
3207 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3208 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3209 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3213 n = transform_bitwise_distributive(n, transform_node_And);
3216 } /* transform_node_And */
3221 static ir_node *transform_node_Eor(ir_node *n) {
3222 ir_node *c, *oldn = n;
3223 ir_node *a = get_Eor_left(n);
3224 ir_node *b = get_Eor_right(n);
3225 ir_mode *mode = get_irn_mode(n);
3227 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3229 /* we can evaluate 2 Projs of the same Cmp */
3230 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3231 ir_node *pred_a = get_Proj_pred(a);
3232 ir_node *pred_b = get_Proj_pred(b);
3233 if(pred_a == pred_b) {
3234 dbg_info *dbgi = get_irn_dbg_info(n);
3235 ir_node *block = get_nodes_block(pred_a);
3236 pn_Cmp pn_a = get_Proj_proj(a);
3237 pn_Cmp pn_b = get_Proj_proj(b);
3238 /* yes, we can simply calculate with pncs */
3239 pn_Cmp new_pnc = pn_a ^ pn_b;
3241 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3248 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3249 mode, get_mode_null(mode));
3250 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3251 } else if (mode == mode_b &&
3253 is_Const(b) && is_Const_one(b) &&
3254 is_Cmp(get_Proj_pred(a))) {
3255 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3256 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3257 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3259 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3260 } else if (is_Const(b)) {
3261 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3262 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3263 ir_node *not_op = get_Not_op(a);
3264 dbg_info *dbg = get_irn_dbg_info(n);
3265 ir_graph *irg = current_ir_graph;
3266 ir_node *block = get_nodes_block(n);
3267 ir_mode *mode = get_irn_mode(n);
3268 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3270 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3271 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3272 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3275 n = transform_bitwise_distributive(n, transform_node_Eor);
3279 } /* transform_node_Eor */
3284 static ir_node *transform_node_Not(ir_node *n) {
3285 ir_node *c, *oldn = n;
3286 ir_node *a = get_Not_op(n);
3287 ir_mode *mode = get_irn_mode(n);
3289 HANDLE_UNOP_PHI(tarval_not,a,c);
3291 /* check for a boolean Not */
3292 if (mode == mode_b &&
3294 is_Cmp(get_Proj_pred(a))) {
3295 /* We negate a Cmp. The Cmp has the negated result anyways! */
3296 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3297 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3298 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3302 ir_node *eor_b = get_Eor_right(a);
3303 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3304 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3305 ir_node *eor_a = get_Eor_left(a);
3306 dbg_info *dbg = get_irn_dbg_info(n);
3307 ir_graph *irg = current_ir_graph;
3308 ir_node *block = get_nodes_block(n);
3309 ir_mode *mode = get_irn_mode(n);
3310 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3314 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3315 if (is_Minus(a)) { /* ~-x -> x + -1 */
3316 dbg_info *dbg = get_irn_dbg_info(n);
3317 ir_graph *irg = current_ir_graph;
3318 ir_node *block = get_nodes_block(n);
3319 ir_node *add_l = get_Minus_op(a);
3320 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3321 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3322 } else if (is_Add(a)) {
3323 ir_node *add_r = get_Add_right(a);
3324 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3325 /* ~(x + -1) = -x */
3326 ir_node *op = get_Add_left(a);
3327 ir_node *blk = get_irn_n(n, -1);
3328 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3329 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3334 } /* transform_node_Not */
3337 * Transform a Minus.
3341 * -(a >>u (size-1)) = a >>s (size-1)
3342 * -(a >>s (size-1)) = a >>u (size-1)
3343 * -(a * const) -> a * -const
3345 static ir_node *transform_node_Minus(ir_node *n) {
3346 ir_node *c, *oldn = n;
3347 ir_node *a = get_Minus_op(n);
3350 HANDLE_UNOP_PHI(tarval_neg,a,c);
3352 mode = get_irn_mode(a);
3353 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3354 /* the following rules are only to twos-complement */
3357 ir_node *op = get_Not_op(a);
3358 tarval *tv = get_mode_one(mode);
3359 ir_node *blk = get_irn_n(n, -1);
3360 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3361 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3362 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3366 ir_node *c = get_Shr_right(a);
3369 tarval *tv = get_Const_tarval(c);
3371 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3372 /* -(a >>u (size-1)) = a >>s (size-1) */
3373 ir_node *v = get_Shr_left(a);
3375 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3376 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3382 ir_node *c = get_Shrs_right(a);
3385 tarval *tv = get_Const_tarval(c);
3387 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3388 /* -(a >>s (size-1)) = a >>u (size-1) */
3389 ir_node *v = get_Shrs_left(a);
3391 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3392 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3399 /* - (a-b) = b - a */
3400 ir_node *la = get_Sub_left(a);
3401 ir_node *ra = get_Sub_right(a);
3402 ir_node *blk = get_irn_n(n, -1);
3404 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3405 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3409 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3410 ir_node *mul_l = get_Mul_left(a);
3411 ir_node *mul_r = get_Mul_right(a);
3412 if (is_Const(mul_r)) {
3413 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3414 if(tv != tarval_bad) {
3415 ir_node *cnst = new_Const(mode, tv);
3416 dbg_info *dbg = get_irn_dbg_info(a);
3417 ir_graph *irg = current_ir_graph;
3418 ir_node *block = get_nodes_block(a);
3419 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3420 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3427 } /* transform_node_Minus */
3430 * Transform a Cast_type(Const) into a new Const_type
3432 static ir_node *transform_node_Cast(ir_node *n) {
3434 ir_node *pred = get_Cast_op(n);
3435 ir_type *tp = get_irn_type(n);
3437 if (is_Const(pred) && get_Const_type(pred) != tp) {
3438 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3439 get_Const_tarval(pred), tp);
3440 DBG_OPT_CSTEVAL(oldn, n);
3441 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3442 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3443 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3444 DBG_OPT_CSTEVAL(oldn, n);
3448 } /* transform_node_Cast */
3451 * Transform a Proj(Div) with a non-zero value.
3452 * Removes the exceptions and routes the memory to the NoMem node.
3454 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3455 ir_node *div = get_Proj_pred(proj);
3456 ir_node *b = get_Div_right(div);
3457 ir_node *confirm, *res, *new_mem;
3460 if (value_not_zero(b, &confirm)) {
3461 /* div(x, y) && y != 0 */
3462 if (confirm == NULL) {
3463 /* we are sure we have a Const != 0 */
3464 new_mem = get_Div_mem(div);
3465 if (is_Pin(new_mem))
3466 new_mem = get_Pin_op(new_mem);
3467 set_Div_mem(div, new_mem);
3468 set_irn_pinned(div, op_pin_state_floats);
3471 proj_nr = get_Proj_proj(proj);
3473 case pn_Div_X_regular:
3474 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3476 case pn_Div_X_except:
3477 /* we found an exception handler, remove it */
3478 DBG_OPT_EXC_REM(proj);
3482 res = get_Div_mem(div);
3483 new_mem = get_irg_no_mem(current_ir_graph);
3486 /* This node can only float up to the Confirm block */
3487 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3489 set_irn_pinned(div, op_pin_state_floats);
3490 /* this is a Div without exception, we can remove the memory edge */
3491 set_Div_mem(div, new_mem);
3496 } /* transform_node_Proj_Div */
3499 * Transform a Proj(Mod) with a non-zero value.
3500 * Removes the exceptions and routes the memory to the NoMem node.
3502 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3503 ir_node *mod = get_Proj_pred(proj);
3504 ir_node *b = get_Mod_right(mod);
3505 ir_node *confirm, *res, *new_mem;
3508 if (value_not_zero(b, &confirm)) {
3509 /* mod(x, y) && y != 0 */
3510 proj_nr = get_Proj_proj(proj);
3512 if (confirm == NULL) {
3513 /* we are sure we have a Const != 0 */
3514 new_mem = get_Mod_mem(mod);
3515 if (is_Pin(new_mem))
3516 new_mem = get_Pin_op(new_mem);
3517 set_Mod_mem(mod, new_mem);
3518 set_irn_pinned(mod, op_pin_state_floats);
3523 case pn_Mod_X_regular:
3524 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3526 case pn_Mod_X_except:
3527 /* we found an exception handler, remove it */
3528 DBG_OPT_EXC_REM(proj);
3532 res = get_Mod_mem(mod);
3533 new_mem = get_irg_no_mem(current_ir_graph);
3536 /* This node can only float up to the Confirm block */
3537 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3539 /* this is a Mod without exception, we can remove the memory edge */
3540 set_Mod_mem(mod, new_mem);
3543 if (get_Mod_left(mod) == b) {
3544 /* a % a = 0 if a != 0 */
3545 ir_mode *mode = get_irn_mode(proj);
3546 ir_node *res = new_Const(mode, get_mode_null(mode));
3548 DBG_OPT_CSTEVAL(mod, res);
3554 } /* transform_node_Proj_Mod */
3557 * Transform a Proj(DivMod) with a non-zero value.
3558 * Removes the exceptions and routes the memory to the NoMem node.
3560 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3561 ir_node *divmod = get_Proj_pred(proj);
3562 ir_node *b = get_DivMod_right(divmod);
3563 ir_node *confirm, *res, *new_mem;
3566 if (value_not_zero(b, &confirm)) {
3567 /* DivMod(x, y) && y != 0 */
3568 proj_nr = get_Proj_proj(proj);
3570 if (confirm == NULL) {
3571 /* we are sure we have a Const != 0 */
3572 new_mem = get_DivMod_mem(divmod);
3573 if (is_Pin(new_mem))
3574 new_mem = get_Pin_op(new_mem);
3575 set_DivMod_mem(divmod, new_mem);
3576 set_irn_pinned(divmod, op_pin_state_floats);
3581 case pn_DivMod_X_regular:
3582 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3584 case pn_DivMod_X_except:
3585 /* we found an exception handler, remove it */
3586 DBG_OPT_EXC_REM(proj);
3590 res = get_DivMod_mem(divmod);
3591 new_mem = get_irg_no_mem(current_ir_graph);
3594 /* This node can only float up to the Confirm block */
3595 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3597 /* this is a DivMod without exception, we can remove the memory edge */
3598 set_DivMod_mem(divmod, new_mem);
3601 case pn_DivMod_res_mod:
3602 if (get_DivMod_left(divmod) == b) {
3603 /* a % a = 0 if a != 0 */
3604 ir_mode *mode = get_irn_mode(proj);
3605 ir_node *res = new_Const(mode, get_mode_null(mode));
3607 DBG_OPT_CSTEVAL(divmod, res);
3613 } /* transform_node_Proj_DivMod */
3616 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3618 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3619 if (get_opt_unreachable_code()) {
3620 ir_node *n = get_Proj_pred(proj);
3621 ir_node *b = get_Cond_selector(n);
3623 if (mode_is_int(get_irn_mode(b))) {
3624 tarval *tb = value_of(b);
3626 if (tb != tarval_bad) {
3627 /* we have a constant switch */
3628 long num = get_Proj_proj(proj);
3630 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3631 if (get_tarval_long(tb) == num) {
3632 /* Do NOT create a jump here, or we will have 2 control flow ops
3633 * in a block. This case is optimized away in optimize_cf(). */
3636 /* this case will NEVER be taken, kill it */
3644 } /* transform_node_Proj_Cond */
3647 * Create a 0 constant of given mode.
3649 static ir_node *create_zero_const(ir_mode *mode) {
3650 tarval *tv = get_mode_null(mode);
3651 ir_node *cnst = new_Const(mode, tv);
3656 /* the order of the values is important! */
3657 typedef enum const_class {
3663 static const_class classify_const(const ir_node* n)
3665 if (is_Const(n)) return const_const;
3666 if (is_irn_constlike(n)) return const_like;
3671 * Determines whether r is more constlike or has a larger index (in that order)
3674 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3676 const const_class l_order = classify_const(l);
3677 const const_class r_order = classify_const(r);
3679 l_order > r_order ||
3680 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3684 * Normalizes and optimizes Cmp nodes.
3686 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3687 ir_node *n = get_Proj_pred(proj);
3688 ir_node *left = get_Cmp_left(n);
3689 ir_node *right = get_Cmp_right(n);
3692 ir_mode *mode = NULL;
3693 long proj_nr = get_Proj_proj(proj);
3695 /* we can evaluate some cases directly */
3698 return new_Const(mode_b, get_tarval_b_false());
3700 return new_Const(mode_b, get_tarval_b_true());
3702 if (!mode_is_float(get_irn_mode(left)))
3703 return new_Const(mode_b, get_tarval_b_true());
3709 /* remove Casts of both sides */
3710 left = skip_Cast(left);
3711 right = skip_Cast(right);
3713 /* Remove unnecessary conversions */
3714 /* TODO handle constants */
3715 if (is_Conv(left) && is_Conv(right)) {
3716 ir_mode *mode = get_irn_mode(left);
3717 ir_node *op_left = get_Conv_op(left);
3718 ir_node *op_right = get_Conv_op(right);
3719 ir_mode *mode_left = get_irn_mode(op_left);
3720 ir_mode *mode_right = get_irn_mode(op_right);
3722 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3723 && mode_left != mode_b && mode_right != mode_b) {
3724 ir_graph *irg = current_ir_graph;
3725 ir_node *block = get_nodes_block(n);
3727 if (mode_left == mode_right) {
3731 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3732 } else if (smaller_mode(mode_left, mode_right)) {
3733 left = new_r_Conv(irg, block, op_left, mode_right);
3736 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3737 } else if (smaller_mode(mode_right, mode_left)) {
3739 right = new_r_Conv(irg, block, op_right, mode_left);
3741 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3746 /* remove operation on both sides if possible */
3747 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3749 * The following operations are NOT safe for floating point operations, for instance
3750 * 1.0 + inf == 2.0 + inf, =/=> x == y
3752 if (mode_is_int(get_irn_mode(left))) {
3753 unsigned lop = get_irn_opcode(left);
3755 if (lop == get_irn_opcode(right)) {
3756 ir_node *ll, *lr, *rl, *rr;
3758 /* same operation on both sides, try to remove */
3762 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3763 left = get_unop_op(left);
3764 right = get_unop_op(right);
3766 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3769 ll = get_Add_left(left);
3770 lr = get_Add_right(left);
3771 rl = get_Add_left(right);
3772 rr = get_Add_right(right);
3775 /* X + a CMP X + b ==> a CMP b */
3779 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3780 } else if (ll == rr) {
3781 /* X + a CMP b + X ==> a CMP b */
3785 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3786 } else if (lr == rl) {
3787 /* a + X CMP X + b ==> a CMP b */
3791 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3792 } else if (lr == rr) {
3793 /* a + X CMP b + X ==> a CMP b */
3797 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3801 ll = get_Sub_left(left);
3802 lr = get_Sub_right(left);
3803 rl = get_Sub_left(right);
3804 rr = get_Sub_right(right);
3807 /* X - a CMP X - b ==> a CMP b */
3811 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3812 } else if (lr == rr) {
3813 /* a - X CMP b - X ==> a CMP b */
3817 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3821 if (get_Rot_right(left) == get_Rot_right(right)) {
3822 /* a ROT X CMP b ROT X ==> a CMP b */
3823 left = get_Rot_left(left);
3824 right = get_Rot_left(right);
3826 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3834 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3835 if (is_Add(left) || is_Sub(left)) {
3836 ir_node *ll = get_binop_left(left);
3837 ir_node *lr = get_binop_right(left);
3839 if (lr == right && is_Add(left)) {
3846 right = create_zero_const(get_irn_mode(left));
3848 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3851 if (is_Add(right) || is_Sub(right)) {
3852 ir_node *rl = get_binop_left(right);
3853 ir_node *rr = get_binop_right(right);
3855 if (rr == left && is_Add(right)) {
3862 right = create_zero_const(get_irn_mode(left));
3864 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3867 } /* mode_is_int(...) */
3868 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3870 /* replace mode_b compares with ands/ors */
3871 if (get_irn_mode(left) == mode_b) {
3872 ir_graph *irg = current_ir_graph;
3873 ir_node *block = get_nodes_block(n);
3877 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3878 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3879 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3880 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3881 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3882 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3883 default: bres = NULL;
3886 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3892 * First step: normalize the compare op
3893 * by placing the constant on the right side
3894 * or moving the lower address node to the left.
3896 if (!operands_are_normalized(left, right)) {
3902 proj_nr = get_inversed_pnc(proj_nr);
3907 * Second step: Try to reduce the magnitude
3908 * of a constant. This may help to generate better code
3909 * later and may help to normalize more compares.
3910 * Of course this is only possible for integer values.
3912 if (is_Const(right)) {
3913 mode = get_irn_mode(right);
3914 tv = get_Const_tarval(right);
3916 /* TODO extend to arbitrary constants */
3917 if (is_Conv(left) && tarval_is_null(tv)) {
3918 ir_node *op = get_Conv_op(left);
3919 ir_mode *op_mode = get_irn_mode(op);
3922 * UpConv(x) REL 0 ==> x REL 0
3924 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3925 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3926 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3927 tv = get_mode_null(op_mode);
3931 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3935 if (tv != tarval_bad) {
3936 /* the following optimization is possible on modes without Overflow
3937 * on Unary Minus or on == and !=:
3938 * -a CMP c ==> a swap(CMP) -c
3940 * Beware: for two-complement Overflow may occur, so only == and != can
3941 * be optimized, see this:
3942 * -MININT < 0 =/=> MININT > 0 !!!
3944 if (is_Minus(left) &&
3945 (!mode_overflow_on_unary_Minus(mode) ||
3946 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3947 tv = tarval_neg(tv);
3949 if (tv != tarval_bad) {
3950 left = get_Minus_op(left);
3951 proj_nr = get_inversed_pnc(proj_nr);
3953 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3955 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3956 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3957 tv = tarval_not(tv);
3959 if (tv != tarval_bad) {
3960 left = get_Not_op(left);
3962 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3966 /* for integer modes, we have more */
3967 if (mode_is_int(mode)) {
3968 /* Ne includes Unordered which is not possible on integers.
3969 * However, frontends often use this wrong, so fix it here */
3970 if (proj_nr & pn_Cmp_Uo) {
3971 proj_nr &= ~pn_Cmp_Uo;
3972 set_Proj_proj(proj, proj_nr);
3975 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3976 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3977 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3978 tv = tarval_sub(tv, get_mode_one(mode));
3980 if (tv != tarval_bad) {
3981 proj_nr ^= pn_Cmp_Eq;
3983 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3986 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3987 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3988 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3989 tv = tarval_add(tv, get_mode_one(mode));
3991 if (tv != tarval_bad) {
3992 proj_nr ^= pn_Cmp_Eq;
3994 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3998 /* the following reassociations work only for == and != */
3999 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4001 #if 0 /* Might be not that good in general */
4002 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4003 if (tarval_is_null(tv) && is_Sub(left)) {
4004 right = get_Sub_right(left);
4005 left = get_Sub_left(left);
4007 tv = value_of(right);
4009 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4013 if (tv != tarval_bad) {
4014 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4016 ir_node *c1 = get_Sub_right(left);
4017 tarval *tv2 = value_of(c1);
4019 if (tv2 != tarval_bad) {
4020 tv2 = tarval_add(tv, value_of(c1));
4022 if (tv2 != tarval_bad) {
4023 left = get_Sub_left(left);
4026 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4030 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4031 else if (is_Add(left)) {
4032 ir_node *a_l = get_Add_left(left);
4033 ir_node *a_r = get_Add_right(left);
4037 if (is_Const(a_l)) {
4039 tv2 = value_of(a_l);
4042 tv2 = value_of(a_r);
4045 if (tv2 != tarval_bad) {
4046 tv2 = tarval_sub(tv, tv2);
4048 if (tv2 != tarval_bad) {
4052 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4056 /* -a == c ==> a == -c, -a != c ==> a != -c */
4057 else if (is_Minus(left)) {
4058 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4060 if (tv2 != tarval_bad) {
4061 left = get_Minus_op(left);
4064 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4069 /* the following reassociations work only for <= */
4070 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4071 if (tv != tarval_bad) {
4072 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4073 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4079 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4080 switch (get_irn_opcode(left)) {
4084 c1 = get_And_right(left);
4087 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4088 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4090 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4092 /* TODO: move to constant evaluation */
4093 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4094 c1 = new_Const(mode_b, tv);
4095 DBG_OPT_CSTEVAL(proj, c1);
4099 if (tarval_is_single_bit(tv)) {
4101 * optimization for AND:
4103 * And(x, C) == C ==> And(x, C) != 0
4104 * And(x, C) != C ==> And(X, C) == 0
4106 * if C is a single Bit constant.
4109 /* check for Constant's match. We have check hare the tarvals,
4110 because our const might be changed */
4111 if (get_Const_tarval(c1) == tv) {
4112 /* fine: do the transformation */
4113 tv = get_mode_null(get_tarval_mode(tv));
4114 proj_nr ^= pn_Cmp_Leg;
4116 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4122 c1 = get_Or_right(left);
4123 if (is_Const(c1) && tarval_is_null(tv)) {
4125 * Or(x, C) == 0 && C != 0 ==> FALSE
4126 * Or(x, C) != 0 && C != 0 ==> TRUE
4128 if (! tarval_is_null(get_Const_tarval(c1))) {
4129 /* TODO: move to constant evaluation */
4130 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4131 c1 = new_Const(mode_b, tv);
4132 DBG_OPT_CSTEVAL(proj, c1);
4139 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4141 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4144 c1 = get_Shl_right(left);
4146 tarval *tv1 = get_Const_tarval(c1);
4147 ir_mode *mode = get_irn_mode(left);
4148 tarval *minus1 = get_mode_all_one(mode);
4149 tarval *amask = tarval_shr(minus1, tv1);
4150 tarval *cmask = tarval_shl(minus1, tv1);
4153 if (tarval_and(tv, cmask) != tv) {
4154 /* condition not met */
4155 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4156 c1 = new_Const(mode_b, tv);
4157 DBG_OPT_CSTEVAL(proj, c1);
4160 sl = get_Shl_left(left);
4161 blk = get_nodes_block(n);
4162 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4163 tv = tarval_shr(tv, tv1);
4165 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4170 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4172 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4175 c1 = get_Shr_right(left);
4177 tarval *tv1 = get_Const_tarval(c1);
4178 ir_mode *mode = get_irn_mode(left);
4179 tarval *minus1 = get_mode_all_one(mode);
4180 tarval *amask = tarval_shl(minus1, tv1);
4181 tarval *cmask = tarval_shr(minus1, tv1);
4184 if (tarval_and(tv, cmask) != tv) {
4185 /* condition not met */
4186 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4187 c1 = new_Const(mode_b, tv);
4188 DBG_OPT_CSTEVAL(proj, c1);
4191 sl = get_Shr_left(left);
4192 blk = get_nodes_block(n);
4193 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4194 tv = tarval_shl(tv, tv1);
4196 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4201 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4203 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4206 c1 = get_Shrs_right(left);
4208 tarval *tv1 = get_Const_tarval(c1);
4209 ir_mode *mode = get_irn_mode(left);
4210 tarval *minus1 = get_mode_all_one(mode);
4211 tarval *amask = tarval_shl(minus1, tv1);
4212 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4215 cond = tarval_sub(cond, tv1);
4216 cond = tarval_shrs(tv, cond);
4218 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4219 /* condition not met */
4220 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4221 c1 = new_Const(mode_b, tv);
4222 DBG_OPT_CSTEVAL(proj, c1);
4225 sl = get_Shrs_left(left);
4226 blk = get_nodes_block(n);
4227 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4228 tv = tarval_shl(tv, tv1);
4230 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4235 } /* tarval != bad */
4238 if (changed & 2) /* need a new Const */
4239 right = new_Const(mode, tv);
4241 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4242 ir_node *op = get_Proj_pred(left);
4244 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4245 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4246 ir_node *c = get_binop_right(op);
4249 tarval *tv = get_Const_tarval(c);
4251 if (tarval_is_single_bit(tv)) {
4252 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4253 ir_node *v = get_binop_left(op);
4254 ir_node *blk = get_irn_n(op, -1);
4255 ir_mode *mode = get_irn_mode(v);
4257 tv = tarval_sub(tv, get_mode_one(mode));
4258 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4260 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4267 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4269 /* create a new compare */
4270 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4271 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4275 } /* transform_node_Proj_Cmp */
4278 * Does all optimizations on nodes that must be done on it's Proj's
4279 * because of creating new nodes.
4281 static ir_node *transform_node_Proj(ir_node *proj) {
4282 ir_node *n = get_Proj_pred(proj);
4284 switch (get_irn_opcode(n)) {
4286 return transform_node_Proj_Div(proj);
4289 return transform_node_Proj_Mod(proj);
4292 return transform_node_Proj_DivMod(proj);
4295 return transform_node_Proj_Cond(proj);
4298 return transform_node_Proj_Cmp(proj);
4301 /* should not happen, but if it does will be optimized away */
4302 return equivalent_node_Proj(proj);
4308 } /* transform_node_Proj */
4311 * Move Confirms down through Phi nodes.
4313 static ir_node *transform_node_Phi(ir_node *phi) {
4315 ir_mode *mode = get_irn_mode(phi);
4317 if (mode_is_reference(mode)) {
4318 n = get_irn_arity(phi);
4320 /* Beware of Phi0 */
4322 ir_node *pred = get_irn_n(phi, 0);
4323 ir_node *bound, *new_Phi, *block, **in;
4326 if (! is_Confirm(pred))
4329 bound = get_Confirm_bound(pred);
4330 pnc = get_Confirm_cmp(pred);
4332 NEW_ARR_A(ir_node *, in, n);
4333 in[0] = get_Confirm_value(pred);
4335 for (i = 1; i < n; ++i) {
4336 pred = get_irn_n(phi, i);
4338 if (! is_Confirm(pred) ||
4339 get_Confirm_bound(pred) != bound ||
4340 get_Confirm_cmp(pred) != pnc)
4342 in[i] = get_Confirm_value(pred);
4344 /* move the Confirm nodes "behind" the Phi */
4345 block = get_irn_n(phi, -1);
4346 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4347 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4351 } /* transform_node_Phi */
4354 * Returns the operands of a commutative bin-op, if one operand is
4355 * a const, it is returned as the second one.
4357 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4358 ir_node *op_a = get_binop_left(binop);
4359 ir_node *op_b = get_binop_right(binop);
4361 assert(is_op_commutative(get_irn_op(binop)));
4363 if (is_Const(op_a)) {
4370 } /* get_comm_Binop_Ops */
4373 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4374 * Such pattern may arise in bitfield stores.
4376 * value c4 value c4 & c2
4377 * AND c3 AND c1 | c3
4384 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4387 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4390 ir_node *and_l, *c3;
4391 ir_node *value, *c4;
4392 ir_node *new_and, *new_const, *block;
4393 ir_mode *mode = get_irn_mode(or);
4395 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4398 get_comm_Binop_Ops(or, &and, &c1);
4399 if (!is_Const(c1) || !is_And(and))
4402 get_comm_Binop_Ops(and, &or_l, &c2);
4406 tv1 = get_Const_tarval(c1);
4407 tv2 = get_Const_tarval(c2);
4409 tv = tarval_or(tv1, tv2);
4410 if (tarval_is_all_one(tv)) {
4411 /* the AND does NOT clear a bit with isn't set by the OR */
4412 set_Or_left(or, or_l);
4413 set_Or_right(or, c1);
4415 /* check for more */
4422 get_comm_Binop_Ops(or_l, &and_l, &c3);
4423 if (!is_Const(c3) || !is_And(and_l))
4426 get_comm_Binop_Ops(and_l, &value, &c4);
4430 /* ok, found the pattern, check for conditions */
4431 assert(mode == get_irn_mode(and));
4432 assert(mode == get_irn_mode(or_l));
4433 assert(mode == get_irn_mode(and_l));
4435 tv3 = get_Const_tarval(c3);
4436 tv4 = get_Const_tarval(c4);
4438 tv = tarval_or(tv4, tv2);
4439 if (!tarval_is_all_one(tv)) {
4440 /* have at least one 0 at the same bit position */
4444 n_tv4 = tarval_not(tv4);
4445 if (tv3 != tarval_and(tv3, n_tv4)) {
4446 /* bit in the or_mask is outside the and_mask */
4450 n_tv2 = tarval_not(tv2);
4451 if (tv1 != tarval_and(tv1, n_tv2)) {
4452 /* bit in the or_mask is outside the and_mask */
4456 /* ok, all conditions met */
4457 block = get_irn_n(or, -1);
4459 new_and = new_r_And(current_ir_graph, block,
4460 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4462 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4464 set_Or_left(or, new_and);
4465 set_Or_right(or, new_const);
4467 /* check for more */
4469 } /* transform_node_Or_bf_store */
4472 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4474 static ir_node *transform_node_Or_Rot(ir_node *or) {
4475 ir_mode *mode = get_irn_mode(or);
4476 ir_node *shl, *shr, *block;
4477 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4480 if (! mode_is_int(mode))
4483 shl = get_binop_left(or);
4484 shr = get_binop_right(or);
4493 } else if (!is_Shl(shl)) {
4495 } else if (!is_Shr(shr)) {
4498 x = get_Shl_left(shl);
4499 if (x != get_Shr_left(shr))
4502 c1 = get_Shl_right(shl);
4503 c2 = get_Shr_right(shr);
4504 if (is_Const(c1) && is_Const(c2)) {
4505 tv1 = get_Const_tarval(c1);
4506 if (! tarval_is_long(tv1))
4509 tv2 = get_Const_tarval(c2);
4510 if (! tarval_is_long(tv2))
4513 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4514 != (int) get_mode_size_bits(mode))
4517 /* yet, condition met */
4518 block = get_irn_n(or, -1);
4520 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4522 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4524 } else if (is_Sub(c1)) {
4528 if (get_Sub_right(sub) != v)
4531 c1 = get_Sub_left(sub);
4535 tv1 = get_Const_tarval(c1);
4536 if (! tarval_is_long(tv1))
4539 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4542 /* yet, condition met */
4543 block = get_nodes_block(or);
4545 /* a Rot right is not supported, so use a rot left */
4546 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4548 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4550 } else if (is_Sub(c2)) {
4554 c1 = get_Sub_left(sub);
4558 tv1 = get_Const_tarval(c1);
4559 if (! tarval_is_long(tv1))
4562 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4565 /* yet, condition met */
4566 block = get_irn_n(or, -1);
4569 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4571 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4576 } /* transform_node_Or_Rot */
4581 static ir_node *transform_node_Or(ir_node *n) {
4582 ir_node *c, *oldn = n;
4583 ir_node *a = get_Or_left(n);
4584 ir_node *b = get_Or_right(n);
4587 if (is_Not(a) && is_Not(b)) {
4588 /* ~a | ~b = ~(a&b) */
4589 ir_node *block = get_nodes_block(n);
4591 mode = get_irn_mode(n);
4594 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4595 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4596 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4600 /* we can evaluate 2 Projs of the same Cmp */
4601 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4602 ir_node *pred_a = get_Proj_pred(a);
4603 ir_node *pred_b = get_Proj_pred(b);
4604 if (pred_a == pred_b) {
4605 dbg_info *dbgi = get_irn_dbg_info(n);
4606 ir_node *block = get_nodes_block(pred_a);
4607 pn_Cmp pn_a = get_Proj_proj(a);
4608 pn_Cmp pn_b = get_Proj_proj(b);
4609 /* yes, we can simply calculate with pncs */
4610 pn_Cmp new_pnc = pn_a | pn_b;
4612 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4617 mode = get_irn_mode(n);
4618 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4620 n = transform_node_Or_bf_store(n);
4621 n = transform_node_Or_Rot(n);
4625 n = transform_bitwise_distributive(n, transform_node_Or);
4628 } /* transform_node_Or */
4632 static ir_node *transform_node(ir_node *n);
4635 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4637 * Should be moved to reassociation?
4639 static ir_node *transform_node_shift(ir_node *n) {
4640 ir_node *left, *right;
4641 tarval *tv1, *tv2, *res;
4643 int modulo_shf, flag;
4645 left = get_binop_left(n);
4647 /* different operations */
4648 if (get_irn_op(left) != get_irn_op(n))
4651 right = get_binop_right(n);
4652 tv1 = value_of(right);
4653 if (tv1 == tarval_bad)
4656 tv2 = value_of(get_binop_right(left));
4657 if (tv2 == tarval_bad)
4660 res = tarval_add(tv1, tv2);
4662 /* beware: a simple replacement works only, if res < modulo shift */
4663 mode = get_irn_mode(n);
4667 modulo_shf = get_mode_modulo_shift(mode);
4668 if (modulo_shf > 0) {
4669 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4671 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4677 /* ok, we can replace it */
4678 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4680 in[0] = get_binop_left(left);
4681 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4683 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4685 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4687 return transform_node(irn);
4690 } /* transform_node_shift */
4695 static ir_node *transform_node_Shr(ir_node *n) {
4696 ir_node *c, *oldn = n;
4697 ir_node *a = get_Shr_left(n);
4698 ir_node *b = get_Shr_right(n);
4699 ir_mode *mode = get_irn_mode(n);
4701 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4702 return transform_node_shift(n);
4703 } /* transform_node_Shr */
4708 static ir_node *transform_node_Shrs(ir_node *n) {
4709 ir_node *c, *oldn = n;
4710 ir_node *a = get_Shrs_left(n);
4711 ir_node *b = get_Shrs_right(n);
4712 ir_mode *mode = get_irn_mode(n);
4714 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4715 return transform_node_shift(n);
4716 } /* transform_node_Shrs */
4721 static ir_node *transform_node_Shl(ir_node *n) {
4722 ir_node *c, *oldn = n;
4723 ir_node *a = get_Shl_left(n);
4724 ir_node *b = get_Shl_right(n);
4725 ir_mode *mode = get_irn_mode(n);
4727 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4728 return transform_node_shift(n);
4729 } /* transform_node_Shl */
4734 static ir_node *transform_node_Rot(ir_node *n) {
4735 ir_node *c, *oldn = n;
4736 ir_node *a = get_Rot_left(n);
4737 ir_node *b = get_Rot_right(n);
4738 ir_mode *mode = get_irn_mode(n);
4740 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4741 return transform_node_shift(n);
4742 } /* transform_node_Rot */
4747 static ir_node *transform_node_Conv(ir_node *n) {
4748 ir_node *c, *oldn = n;
4749 ir_node *a = get_Conv_op(n);
4751 if (is_const_Phi(a)) {
4752 c = apply_conv_on_phi(a, get_irn_mode(n));
4754 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4759 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4760 ir_mode *mode = get_irn_mode(n);
4761 return new_r_Unknown(current_ir_graph, mode);
4765 } /* transform_node_Conv */
4768 * Remove dead blocks and nodes in dead blocks
4769 * in keep alive list. We do not generate a new End node.
4771 static ir_node *transform_node_End(ir_node *n) {
4772 int i, j, n_keepalives = get_End_n_keepalives(n);
4775 NEW_ARR_A(ir_node *, in, n_keepalives);
4777 for (i = j = 0; i < n_keepalives; ++i) {
4778 ir_node *ka = get_End_keepalive(n, i);
4780 if (! is_Block_dead(ka)) {
4784 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4787 /* FIXME: beabi need to keep a Proj(M) */
4788 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4791 if (j != n_keepalives)
4792 set_End_keepalives(n, j, in);
4794 } /* transform_node_End */
4796 /** returns 1 if a == -b */
4797 static int is_negated_value(ir_node *a, ir_node *b) {
4798 if (is_Minus(a) && get_Minus_op(a) == b)
4800 if (is_Minus(b) && get_Minus_op(b) == a)
4802 if (is_Sub(a) && is_Sub(b)) {
4803 ir_node *a_left = get_Sub_left(a);
4804 ir_node *a_right = get_Sub_right(a);
4805 ir_node *b_left = get_Sub_left(b);
4806 ir_node *b_right = get_Sub_right(b);
4808 if (a_left == b_right && a_right == b_left)
4816 * Optimize a Mux into some simpler cases.
4818 static ir_node *transform_node_Mux(ir_node *n) {
4819 ir_node *oldn = n, *sel = get_Mux_sel(n);
4820 ir_mode *mode = get_irn_mode(n);
4821 ir_node *t = get_Mux_true(n);
4822 ir_node *f = get_Mux_false(n);
4823 ir_graph *irg = current_ir_graph;
4824 ir_node *conds[1], *vals[2];
4826 /* first normalization step: move a possible zero to the false case */
4828 ir_node *cmp = get_Proj_pred(sel);
4831 if (is_Const(t) && is_Const_null(t)) {
4832 /* Psi(x, 0, y) => Psi(x, y, 0) */
4833 pn_Cmp pnc = get_Proj_proj(sel);
4834 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
4835 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
4839 n = new_rd_Psi(get_irn_dbg_info(n), irg, get_nodes_block(n), 1, conds, vals, mode);
4846 /* note: after normalization, false can only happen on default */
4847 if (mode == mode_b) {
4848 dbg_info *dbg = get_irn_dbg_info(n);
4849 ir_node *block = get_nodes_block(n);
4850 ir_graph *irg = current_ir_graph;
4853 tarval *tv_t = get_Const_tarval(t);
4854 if (tv_t == tarval_b_true) {
4856 /* Muxb(sel, true, false) = sel */
4857 assert(get_Const_tarval(f) == tarval_b_false);
4858 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
4861 /* Muxb(sel, true, x) = Or(sel, x) */
4862 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
4863 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
4867 } else if (is_Const(f)) {
4868 tarval *tv_f = get_Const_tarval(f);
4869 if (tv_f == tarval_b_true) {
4870 /* Muxb(sel, x, true) = Or(Not(sel), x) */
4871 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4872 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
4873 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4876 /* Muxb(sel, x, false) = And(sel, x) */
4877 assert(tv_f == tarval_b_false);
4878 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
4879 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
4885 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
4886 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
4887 tarval *a = get_Const_tarval(t);
4888 tarval *b = get_Const_tarval(f);
4889 tarval *null = get_tarval_null(mode);
4892 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
4893 diff = tarval_sub(a, b);
4896 diff = tarval_sub(b, a);
4900 if (diff == get_tarval_one(mode) && min != null) {
4901 dbg_info *dbg = get_irn_dbg_info(n);
4902 ir_node *block = get_nodes_block(n);
4903 ir_graph *irg = current_ir_graph;
4907 vals[0] = new_Const(mode, tarval_sub(a, min));
4908 vals[1] = new_Const(mode, tarval_sub(b, min));
4909 n = new_rd_Psi(dbg, irg, block, 1, conds, vals, mode);
4910 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
4916 ir_node *cmp = get_Proj_pred(sel);
4917 long pn = get_Proj_proj(sel);
4920 * Note: normalization puts the constant on the right side,
4921 * so we check only one case.
4923 * Note further that these optimization work even for floating point
4924 * with NaN's because -NaN == NaN.
4925 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
4929 ir_node *cmp_r = get_Cmp_right(cmp);
4930 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4931 ir_node *block = get_nodes_block(n);
4932 ir_node *cmp_l = get_Cmp_left(cmp);
4934 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
4937 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4938 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4940 /* Psi(a >/>= 0, a, -a) = Psi(a </<= 0, -a, a) ==> Abs(a) */
4941 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4943 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4945 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4946 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4948 /* Psi(a </<= 0, a, -a) = Psi(a >/>= 0, -a, a) ==> -Abs(a) */
4949 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4951 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4953 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4958 if (mode_is_int(mode)) {
4960 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
4961 /* Psi((a & b) != 0, c, 0) */
4962 ir_node *and_r = get_And_right(cmp_l);
4965 if (and_r == t && f == cmp_r) {
4966 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
4967 if (pn == pn_Cmp_Lg) {
4968 /* Psi((a & 2^C) != 0, 2^C, 0) */
4971 /* Psi((a & 2^C) == 0, 2^C, 0) */
4972 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
4973 block, cmp_l, t, mode);
4978 if (is_Shl(and_r)) {
4979 ir_node *shl_l = get_Shl_left(and_r);
4980 if (is_Const(shl_l) && is_Const_one(shl_l)) {
4981 if (and_r == t && f == cmp_r) {
4982 if (pn == pn_Cmp_Lg) {
4983 /* (a & (1 << n)) != 0, (1 << n), 0) */
4986 /* (a & (1 << n)) == 0, (1 << n), 0) */
4987 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
4988 block, cmp_l, t, mode);
4994 and_l = get_And_left(cmp_l);
4995 if (is_Shl(and_l)) {
4996 ir_node *shl_l = get_Shl_left(and_l);
4997 if (is_Const(shl_l) && is_Const_one(shl_l)) {
4998 if (and_l == t && f == cmp_r) {
4999 if (pn == pn_Cmp_Lg) {
5000 /* ((1 << n) & a) != 0, (1 << n), 0) */
5003 /* ((1 << n) & a) == 0, (1 << n), 0) */
5004 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5005 block, cmp_l, t, mode);
5016 return arch_transform_node_Mux(n);
5017 } /* transform_node_Mux */
5020 * Optimize a Psi into some simpler cases.
5022 static ir_node *transform_node_Psi(ir_node *n) {
5024 return transform_node_Mux(n);
5027 } /* transform_node_Psi */
5030 * optimize sync nodes that have other syncs as input we simply add the inputs
5031 * of the other sync to our own inputs
5033 static ir_node *transform_node_Sync(ir_node *n) {
5034 int arity = get_Sync_n_preds(n);
5037 for (i = 0; i < arity;) {
5038 ir_node *pred = get_Sync_pred(n, i);
5042 if (!is_Sync(pred)) {
5050 pred_arity = get_Sync_n_preds(pred);
5051 for (j = 0; j < pred_arity; ++j) {
5052 ir_node *pred_pred = get_Sync_pred(pred, j);
5057 add_irn_n(n, pred_pred);
5061 if (get_Sync_pred(n, k) == pred_pred) break;
5066 /* rehash the sync node */
5067 add_identities(current_ir_graph->value_table, n);
5073 * Tries several [inplace] [optimizing] transformations and returns an
5074 * equivalent node. The difference to equivalent_node() is that these
5075 * transformations _do_ generate new nodes, and thus the old node must
5076 * not be freed even if the equivalent node isn't the old one.
5078 static ir_node *transform_node(ir_node *n) {
5082 * Transform_node is the only "optimizing transformation" that might
5083 * return a node with a different opcode. We iterate HERE until fixpoint
5084 * to get the final result.
5088 if (n->op->ops.transform_node)
5089 n = n->op->ops.transform_node(n);
5090 } while (oldn != n);
5093 } /* transform_node */
5096 * Sets the default transform node operation for an ir_op_ops.
5098 * @param code the opcode for the default operation
5099 * @param ops the operations initialized
5104 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5108 ops->transform_node = transform_node_##a; \
5145 } /* firm_set_default_transform_node */
5148 /* **************** Common Subexpression Elimination **************** */
5150 /** The size of the hash table used, should estimate the number of nodes
5152 #define N_IR_NODES 512
5154 /** Compares the attributes of two Const nodes. */
5155 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5156 return (get_Const_tarval(a) != get_Const_tarval(b))
5157 || (get_Const_type(a) != get_Const_type(b));
5158 } /* node_cmp_attr_Const */
5160 /** Compares the attributes of two Proj nodes. */
5161 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5162 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5163 } /* node_cmp_attr_Proj */
5165 /** Compares the attributes of two Filter nodes. */
5166 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5167 return get_Filter_proj(a) != get_Filter_proj(b);
5168 } /* node_cmp_attr_Filter */
5170 /** Compares the attributes of two Alloc nodes. */
5171 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5172 const alloc_attr *pa = get_irn_alloc_attr(a);
5173 const alloc_attr *pb = get_irn_alloc_attr(b);
5174 return (pa->where != pb->where) || (pa->type != pb->type);
5175 } /* node_cmp_attr_Alloc */
5177 /** Compares the attributes of two Free nodes. */
5178 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5179 const free_attr *pa = get_irn_free_attr(a);
5180 const free_attr *pb = get_irn_free_attr(b);
5181 return (pa->where != pb->where) || (pa->type != pb->type);
5182 } /* node_cmp_attr_Free */
5184 /** Compares the attributes of two SymConst nodes. */
5185 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5186 const symconst_attr *pa = get_irn_symconst_attr(a);
5187 const symconst_attr *pb = get_irn_symconst_attr(b);
5188 return (pa->kind != pb->kind)
5189 || (pa->sym.type_p != pb->sym.type_p)
5190 || (pa->tp != pb->tp);
5191 } /* node_cmp_attr_SymConst */
5193 /** Compares the attributes of two Call nodes. */
5194 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5195 return get_irn_call_attr(a) != get_irn_call_attr(b);
5196 } /* node_cmp_attr_Call */
5198 /** Compares the attributes of two Sel nodes. */
5199 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5200 const ir_entity *a_ent = get_Sel_entity(a);
5201 const ir_entity *b_ent = get_Sel_entity(b);
5203 (a_ent->kind != b_ent->kind) ||
5204 (a_ent->name != b_ent->name) ||
5205 (a_ent->owner != b_ent->owner) ||
5206 (a_ent->ld_name != b_ent->ld_name) ||
5207 (a_ent->type != b_ent->type);
5208 } /* node_cmp_attr_Sel */
5210 /** Compares the attributes of two Phi nodes. */
5211 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5212 /* we can only enter this function if both nodes have the same number of inputs,
5213 hence it is enough to check if one of them is a Phi0 */
5215 /* check the Phi0 pos attribute */
5216 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5219 } /* node_cmp_attr_Phi */
5221 /** Compares the attributes of two Conv nodes. */
5222 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5223 return get_Conv_strict(a) != get_Conv_strict(b);
5224 } /* node_cmp_attr_Conv */
5226 /** Compares the attributes of two Cast nodes. */
5227 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5228 return get_Cast_type(a) != get_Cast_type(b);
5229 } /* node_cmp_attr_Cast */
5231 /** Compares the attributes of two Load nodes. */
5232 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5233 if (get_Load_volatility(a) == volatility_is_volatile ||
5234 get_Load_volatility(b) == volatility_is_volatile)
5235 /* NEVER do CSE on volatile Loads */
5237 /* do not CSE Loads with different alignment. Be conservative. */
5238 if (get_Load_align(a) != get_Load_align(b))
5241 return get_Load_mode(a) != get_Load_mode(b);
5242 } /* node_cmp_attr_Load */
5244 /** Compares the attributes of two Store nodes. */
5245 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5246 /* do not CSE Stores with different alignment. Be conservative. */
5247 if (get_Store_align(a) != get_Store_align(b))
5250 /* NEVER do CSE on volatile Stores */
5251 return (get_Store_volatility(a) == volatility_is_volatile ||
5252 get_Store_volatility(b) == volatility_is_volatile);
5253 } /* node_cmp_attr_Store */
5255 /** Compares two exception attributes */
5256 static int node_cmp_exception(ir_node *a, ir_node *b) {
5257 const except_attr *ea = get_irn_except_attr(a);
5258 const except_attr *eb = get_irn_except_attr(b);
5260 return ea->pin_state != eb->pin_state;
5263 #define node_cmp_attr_Bound node_cmp_exception
5265 /** Compares the attributes of two Div nodes. */
5266 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5267 const divmod_attr *ma = get_irn_divmod_attr(a);
5268 const divmod_attr *mb = get_irn_divmod_attr(b);
5269 return ma->exc.pin_state != mb->exc.pin_state ||
5270 ma->res_mode != mb->res_mode ||
5271 ma->no_remainder != mb->no_remainder;
5272 } /* node_cmp_attr_Div */
5274 /** Compares the attributes of two DivMod nodes. */
5275 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5276 const divmod_attr *ma = get_irn_divmod_attr(a);
5277 const divmod_attr *mb = get_irn_divmod_attr(b);
5278 return ma->exc.pin_state != mb->exc.pin_state ||
5279 ma->res_mode != mb->res_mode;
5280 } /* node_cmp_attr_DivMod */
5282 /** Compares the attributes of two Mod nodes. */
5283 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5284 const divmod_attr *ma = get_irn_divmod_attr(a);
5285 const divmod_attr *mb = get_irn_divmod_attr(b);
5286 return ma->exc.pin_state != mb->exc.pin_state ||
5287 ma->res_mode != mb->res_mode;
5288 } /* node_cmp_attr_Mod */
5290 /** Compares the attributes of two Quot nodes. */
5291 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5292 const divmod_attr *ma = get_irn_divmod_attr(a);
5293 const divmod_attr *mb = get_irn_divmod_attr(b);
5294 return ma->exc.pin_state != mb->exc.pin_state ||
5295 ma->res_mode != mb->res_mode;
5296 } /* node_cmp_attr_Quot */
5298 /** Compares the attributes of two Confirm nodes. */
5299 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5300 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5301 } /* node_cmp_attr_Confirm */
5303 /** Compares the attributes of two ASM nodes. */
5304 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5306 const ir_asm_constraint *ca;
5307 const ir_asm_constraint *cb;
5310 if (get_ASM_text(a) != get_ASM_text(b))
5313 /* Should we really check the constraints here? Should be better, but is strange. */
5314 n = get_ASM_n_input_constraints(a);
5315 if (n != get_ASM_n_input_constraints(b))
5318 ca = get_ASM_input_constraints(a);
5319 cb = get_ASM_input_constraints(b);
5320 for (i = 0; i < n; ++i) {
5321 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5325 n = get_ASM_n_output_constraints(a);
5326 if (n != get_ASM_n_output_constraints(b))
5329 ca = get_ASM_output_constraints(a);
5330 cb = get_ASM_output_constraints(b);
5331 for (i = 0; i < n; ++i) {
5332 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5336 n = get_ASM_n_clobbers(a);
5337 if (n != get_ASM_n_clobbers(b))
5340 cla = get_ASM_clobbers(a);
5341 clb = get_ASM_clobbers(b);
5342 for (i = 0; i < n; ++i) {
5343 if (cla[i] != clb[i])
5347 } /* node_cmp_attr_ASM */
5350 * Set the default node attribute compare operation for an ir_op_ops.
5352 * @param code the opcode for the default operation
5353 * @param ops the operations initialized
5358 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5362 ops->node_cmp_attr = node_cmp_attr_##a; \
5393 } /* firm_set_default_node_cmp_attr */
5396 * Compare function for two nodes in the value table. Gets two
5397 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5399 int identities_cmp(const void *elt, const void *key) {
5400 ir_node *a = (ir_node *)elt;
5401 ir_node *b = (ir_node *)key;
5404 if (a == b) return 0;
5406 if ((get_irn_op(a) != get_irn_op(b)) ||
5407 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5409 /* compare if a's in and b's in are of equal length */
5410 irn_arity_a = get_irn_intra_arity(a);
5411 if (irn_arity_a != get_irn_intra_arity(b))
5414 if (get_irn_pinned(a) == op_pin_state_pinned) {
5415 /* for pinned nodes, the block inputs must be equal */
5416 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5418 } else if (! get_opt_global_cse()) {
5419 /* for block-local CSE both nodes must be in the same MacroBlock */
5420 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5424 /* compare a->in[0..ins] with b->in[0..ins] */
5425 for (i = 0; i < irn_arity_a; i++)
5426 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5430 * here, we already now that the nodes are identical except their
5433 if (a->op->ops.node_cmp_attr)
5434 return a->op->ops.node_cmp_attr(a, b);
5437 } /* identities_cmp */
5440 * Calculate a hash value of a node.
5442 unsigned ir_node_hash(ir_node *node) {
5446 if (node->op == op_Const) {
5447 /* special value for const, as they only differ in their tarval. */
5448 h = HASH_PTR(node->attr.con.tv);
5449 h = 9*h + HASH_PTR(get_irn_mode(node));
5450 } else if (node->op == op_SymConst) {
5451 /* special value for const, as they only differ in their symbol. */
5452 h = HASH_PTR(node->attr.symc.sym.type_p);
5453 h = 9*h + HASH_PTR(get_irn_mode(node));
5456 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5457 h = irn_arity = get_irn_intra_arity(node);
5459 /* consider all in nodes... except the block if not a control flow. */
5460 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5461 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5465 h = 9*h + HASH_PTR(get_irn_mode(node));
5467 h = 9*h + HASH_PTR(get_irn_op(node));
5471 } /* ir_node_hash */
5473 pset *new_identities(void) {
5474 return new_pset(identities_cmp, N_IR_NODES);
5475 } /* new_identities */
5477 void del_identities(pset *value_table) {
5478 del_pset(value_table);
5479 } /* del_identities */
5482 * Normalize a node by putting constants (and operands with larger
5483 * node index) on the right (operator side).
5485 * @param n The node to normalize
5487 static void normalize_node(ir_node *n) {
5488 if (is_op_commutative(get_irn_op(n))) {
5489 ir_node *l = get_binop_left(n);
5490 ir_node *r = get_binop_right(n);
5492 /* For commutative operators perform a OP b == b OP a but keep
5493 * constants on the RIGHT side. This helps greatly in some
5494 * optimizations. Moreover we use the idx number to make the form
5496 if (!operands_are_normalized(l, r)) {
5497 set_binop_left(n, r);
5498 set_binop_right(n, l);
5501 } /* normalize_node */
5504 * Update the nodes after a match in the value table. If both nodes have
5505 * the same MacroBlock but different Blocks, we must ensure that the node
5506 * with the dominating Block (the node that is near to the MacroBlock header
5507 * is stored in the table.
5508 * Because a MacroBlock has only one "non-exception" flow, we don't need
5509 * dominance info here: We known, that one block must dominate the other and
5510 * following the only block input will allow to find it.
5512 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5513 ir_node *known_blk, *new_block, *block, *mbh;
5515 if (get_opt_global_cse()) {
5516 /* Block inputs are meaning less */
5519 known_blk = get_irn_n(known_irn, -1);
5520 new_block = get_irn_n(new_ir_node, -1);
5521 if (known_blk == new_block) {
5522 /* already in the same block */
5526 * We expect the typical case when we built the graph. In that case, the
5527 * known_irn is already the upper one, so checking this should be faster.
5530 mbh = get_Block_MacroBlock(new_block);
5532 if (block == known_blk) {
5533 /* ok, we have found it: known_block dominates new_block as expected */
5538 * We have reached the MacroBlock header NOT founding
5539 * the known_block. new_block must dominate known_block.
5542 set_irn_n(known_irn, -1, new_block);
5545 assert(get_Block_n_cfgpreds(block) == 1);
5546 block = get_Block_cfgpred_block(block, 0);
5548 } /* update_value_table */
5551 * Return the canonical node computing the same value as n.
5553 * @param value_table The value table
5554 * @param n The node to lookup
5556 * Looks up the node in a hash table.
5558 * For Const nodes this is performed in the constructor, too. Const
5559 * nodes are extremely time critical because of their frequent use in
5560 * constant string arrays.
5562 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5565 if (!value_table) return n;
5569 o = pset_find(value_table, n, ir_node_hash(n));
5573 update_known_irn(o, n);
5580 * During construction we set the op_pin_state_pinned flag in the graph right when the
5581 * optimization is performed. The flag turning on procedure global cse could
5582 * be changed between two allocations. This way we are safe.
5584 * @param value_table The value table
5585 * @param n The node to lookup
5587 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5590 n = identify(value_table, n);
5591 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5592 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5594 } /* identify_cons */
5597 * Return the canonical node computing the same value as n.
5598 * Looks up the node in a hash table, enters it in the table
5599 * if it isn't there yet.
5601 * @param value_table the HashSet containing all nodes in the
5603 * @param n the node to look up
5605 * @return a node that computes the same value as n or n if no such
5606 * node could be found
5608 ir_node *identify_remember(pset *value_table, ir_node *n) {
5611 if (!value_table) return n;
5614 /* lookup or insert in hash table with given hash key. */
5615 o = pset_insert(value_table, n, ir_node_hash(n));
5618 update_known_irn(o, n);
5623 } /* identify_remember */
5625 /* Add a node to the identities value table. */
5626 void add_identities(pset *value_table, ir_node *node) {
5627 if (get_opt_cse() && is_no_Block(node))
5628 identify_remember(value_table, node);
5629 } /* add_identities */
5631 /* Visit each node in the value table of a graph. */
5632 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5634 ir_graph *rem = current_ir_graph;
5636 current_ir_graph = irg;
5637 foreach_pset(irg->value_table, node)
5639 current_ir_graph = rem;
5640 } /* visit_all_identities */
5643 * Garbage in, garbage out. If a node has a dead input, i.e., the
5644 * Bad node is input to the node, return the Bad node.
5646 static ir_node *gigo(ir_node *node) {
5648 ir_op *op = get_irn_op(node);
5650 /* remove garbage blocks by looking at control flow that leaves the block
5651 and replacing the control flow by Bad. */
5652 if (get_irn_mode(node) == mode_X) {
5653 ir_node *block = get_nodes_block(skip_Proj(node));
5655 /* Don't optimize nodes in immature blocks. */
5656 if (!get_Block_matured(block))
5658 /* Don't optimize End, may have Bads. */
5659 if (op == op_End) return node;
5661 if (is_Block(block)) {
5662 if (is_Block_dead(block)) {
5663 /* control flow from dead block is dead */
5667 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5668 if (!is_Bad(get_irn_n(block, i)))
5672 ir_graph *irg = get_irn_irg(block);
5673 /* the start block is never dead */
5674 if (block != get_irg_start_block(irg)
5675 && block != get_irg_end_block(irg)) {
5677 * Do NOT kill control flow without setting
5678 * the block to dead of bad things can happen:
5679 * We get a Block that is not reachable be irg_block_walk()
5680 * but can be found by irg_walk()!
5682 set_Block_dead(block);
5689 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5690 blocks predecessors is dead. */
5691 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5692 irn_arity = get_irn_arity(node);
5695 * Beware: we can only read the block of a non-floating node.
5697 if (is_irn_pinned_in_irg(node) &&
5698 is_Block_dead(get_nodes_block(skip_Proj(node))))
5701 for (i = 0; i < irn_arity; i++) {
5702 ir_node *pred = get_irn_n(node, i);
5707 /* Propagating Unknowns here seems to be a bad idea, because
5708 sometimes we need a node as a input and did not want that
5710 However, it might be useful to move this into a later phase
5711 (if you think that optimizing such code is useful). */
5712 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5713 return new_Unknown(get_irn_mode(node));
5718 /* With this code we violate the agreement that local_optimize
5719 only leaves Bads in Block, Phi and Tuple nodes. */
5720 /* If Block has only Bads as predecessors it's garbage. */
5721 /* If Phi has only Bads as predecessors it's garbage. */
5722 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5723 irn_arity = get_irn_arity(node);
5724 for (i = 0; i < irn_arity; i++) {
5725 if (!is_Bad(get_irn_n(node, i))) break;
5727 if (i == irn_arity) node = new_Bad();
5734 * These optimizations deallocate nodes from the obstack.
5735 * It can only be called if it is guaranteed that no other nodes
5736 * reference this one, i.e., right after construction of a node.
5738 * @param n The node to optimize
5740 * current_ir_graph must be set to the graph of the node!
5742 ir_node *optimize_node(ir_node *n) {
5745 ir_opcode iro = get_irn_opcode(n);
5747 /* Always optimize Phi nodes: part of the construction. */
5748 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5750 /* constant expression evaluation / constant folding */
5751 if (get_opt_constant_folding()) {
5752 /* neither constants nor Tuple values can be evaluated */
5753 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5754 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5755 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5756 /* try to evaluate */
5757 tv = computed_value(n);
5758 if (tv != tarval_bad) {
5760 ir_type *old_tp = get_irn_type(n);
5761 int i, arity = get_irn_arity(n);
5765 * Try to recover the type of the new expression.
5767 for (i = 0; i < arity && !old_tp; ++i)
5768 old_tp = get_irn_type(get_irn_n(n, i));
5771 * we MUST copy the node here temporary, because it's still needed
5772 * for DBG_OPT_CSTEVAL
5774 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5775 oldn = alloca(node_size);
5777 memcpy(oldn, n, node_size);
5778 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5780 /* ARG, copy the in array, we need it for statistics */
5781 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5783 /* note the inplace edges module */
5784 edges_node_deleted(n, current_ir_graph);
5786 /* evaluation was successful -- replace the node. */
5787 irg_kill_node(current_ir_graph, n);
5788 nw = new_Const(get_tarval_mode(tv), tv);
5790 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5791 set_Const_type(nw, old_tp);
5792 DBG_OPT_CSTEVAL(oldn, nw);
5793 tarval_enable_fp_ops(old_fp_mode);
5796 tarval_enable_fp_ops(old_fp_mode);
5800 /* remove unnecessary nodes */
5801 if (get_opt_constant_folding() ||
5802 (iro == iro_Phi) || /* always optimize these nodes. */
5804 (iro == iro_Proj) ||
5805 (iro == iro_Block) ) /* Flags tested local. */
5806 n = equivalent_node(n);
5808 /* Common Subexpression Elimination.
5810 * Checks whether n is already available.
5811 * The block input is used to distinguish different subexpressions. Right
5812 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5813 * subexpressions within a block.
5816 n = identify_cons(current_ir_graph->value_table, n);
5819 edges_node_deleted(oldn, current_ir_graph);
5821 /* We found an existing, better node, so we can deallocate the old node. */
5822 irg_kill_node(current_ir_graph, oldn);
5826 /* Some more constant expression evaluation that does not allow to
5828 iro = get_irn_opcode(n);
5829 if (get_opt_constant_folding() ||
5830 (iro == iro_Cond) ||
5831 (iro == iro_Proj)) /* Flags tested local. */
5832 n = transform_node(n);
5834 /* Remove nodes with dead (Bad) input.
5835 Run always for transformation induced Bads. */
5838 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5839 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5840 n = identify_remember(current_ir_graph->value_table, n);
5844 } /* optimize_node */
5848 * These optimizations never deallocate nodes (in place). This can cause dead
5849 * nodes lying on the obstack. Remove these by a dead node elimination,
5850 * i.e., a copying garbage collection.
5852 ir_node *optimize_in_place_2(ir_node *n) {
5855 ir_opcode iro = get_irn_opcode(n);
5857 if (!get_opt_optimize() && !is_Phi(n)) return n;
5859 /* constant expression evaluation / constant folding */
5860 if (get_opt_constant_folding()) {
5861 /* neither constants nor Tuple values can be evaluated */
5862 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5863 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5864 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5865 /* try to evaluate */
5866 tv = computed_value(n);
5867 if (tv != tarval_bad) {
5868 /* evaluation was successful -- replace the node. */
5869 ir_type *old_tp = get_irn_type(n);
5870 int i, arity = get_irn_arity(n);
5873 * Try to recover the type of the new expression.
5875 for (i = 0; i < arity && !old_tp; ++i)
5876 old_tp = get_irn_type(get_irn_n(n, i));
5878 n = new_Const(get_tarval_mode(tv), tv);
5880 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5881 set_Const_type(n, old_tp);
5883 DBG_OPT_CSTEVAL(oldn, n);
5884 tarval_enable_fp_ops(old_fp_mode);
5887 tarval_enable_fp_ops(old_fp_mode);
5891 /* remove unnecessary nodes */
5892 if (get_opt_constant_folding() ||
5893 (iro == iro_Phi) || /* always optimize these nodes. */
5894 (iro == iro_Id) || /* ... */
5895 (iro == iro_Proj) || /* ... */
5896 (iro == iro_Block) ) /* Flags tested local. */
5897 n = equivalent_node(n);
5899 /** common subexpression elimination **/
5900 /* Checks whether n is already available. */
5901 /* The block input is used to distinguish different subexpressions. Right
5902 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5903 subexpressions within a block. */
5904 if (get_opt_cse()) {
5905 n = identify(current_ir_graph->value_table, n);
5908 /* Some more constant expression evaluation. */
5909 iro = get_irn_opcode(n);
5910 if (get_opt_constant_folding() ||
5911 (iro == iro_Cond) ||
5912 (iro == iro_Proj)) /* Flags tested local. */
5913 n = transform_node(n);
5915 /* Remove nodes with dead (Bad) input.
5916 Run always for transformation induced Bads. */
5919 /* Now we can verify the node, as it has no dead inputs any more. */
5922 /* Now we have a legal, useful node. Enter it in hash table for cse.
5923 Blocks should be unique anyways. (Except the successor of start:
5924 is cse with the start block!) */
5925 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5926 n = identify_remember(current_ir_graph->value_table, n);
5929 } /* optimize_in_place_2 */
5932 * Wrapper for external use, set proper status bits after optimization.
5934 ir_node *optimize_in_place(ir_node *n) {
5935 /* Handle graph state */
5936 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5938 if (get_opt_global_cse())
5939 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5940 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5941 set_irg_outs_inconsistent(current_ir_graph);
5943 /* FIXME: Maybe we could also test whether optimizing the node can
5944 change the control graph. */
5945 set_irg_doms_inconsistent(current_ir_graph);
5946 return optimize_in_place_2(n);
5947 } /* optimize_in_place */
5950 * Sets the default operation for an ir_ops.
5952 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5953 ops = firm_set_default_computed_value(code, ops);
5954 ops = firm_set_default_equivalent_node(code, ops);
5955 ops = firm_set_default_transform_node(code, ops);
5956 ops = firm_set_default_node_cmp_attr(code, ops);
5957 ops = firm_set_default_get_type(code, ops);
5958 ops = firm_set_default_get_type_attr(code, ops);
5959 ops = firm_set_default_get_entity_attr(code, ops);
5962 } /* firm_set_default_operations */