2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if (tarval_is_null(ta) || tarval_is_null(tb))
149 return get_mode_null(m);
152 } /* computed_value_Carry */
155 * Return the value of a Borrow.
158 static tarval *computed_value_Borrow(ir_node *n) {
159 ir_node *a = get_binop_left(n);
160 ir_node *b = get_binop_right(n);
161 ir_mode *m = get_irn_mode(n);
163 tarval *ta = value_of(a);
164 tarval *tb = value_of(b);
166 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
168 } else if (tarval_is_null(ta)) {
169 return get_mode_null(m);
172 } /* computed_value_Borrow */
175 * Return the value of an unary Minus.
177 static tarval *computed_value_Minus(ir_node *n) {
178 ir_node *a = get_Minus_op(n);
179 tarval *ta = value_of(a);
181 if (ta != tarval_bad)
182 return tarval_neg(ta);
185 } /* computed_value_Minus */
188 * Return the value of a Mul.
190 static tarval *computed_value_Mul(ir_node *n) {
191 ir_node *a = get_Mul_left(n);
192 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 mode = get_irn_mode(n);
199 if (mode != get_irn_mode(a)) {
200 /* n * n = 2n bit multiplication */
201 ta = tarval_convert_to(ta, mode);
202 tb = tarval_convert_to(tb, mode);
205 if (ta != tarval_bad && tb != tarval_bad) {
206 return tarval_mul(ta, tb);
208 /* a*0 = 0 or 0*b = 0 */
209 if (ta == get_mode_null(mode))
211 if (tb == get_mode_null(mode))
215 } /* computed_value_Mul */
218 * Return the value of a floating point Quot.
220 static tarval *computed_value_Quot(ir_node *n) {
221 ir_node *a = get_Quot_left(n);
222 ir_node *b = get_Quot_right(n);
224 tarval *ta = value_of(a);
225 tarval *tb = value_of(b);
227 /* This was missing in original implementation. Why? */
228 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
229 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
230 return tarval_quo(ta, tb);
233 } /* computed_value_Quot */
236 * Calculate the value of an integer Div of two nodes.
237 * Special case: 0 / b
239 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
240 tarval *ta = value_of(a);
241 tarval *tb = value_of(b);
243 /* Compute c1 / c2 or 0 / a, a != 0 */
244 if (ta != tarval_bad) {
245 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
246 return tarval_div(ta, tb);
247 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
251 } /* do_computed_value_Div */
254 * Return the value of an integer Div.
256 static tarval *computed_value_Div(ir_node *n) {
257 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
258 } /* computed_value_Div */
261 * Calculate the value of an integer Mod of two nodes.
262 * Special case: a % 1
264 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 /* Compute c1 % c2 or a % 1 */
269 if (tb != tarval_bad) {
270 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
271 return tarval_mod(ta, tb);
272 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
273 return get_mode_null(get_irn_mode(a));
276 } /* do_computed_value_Mod */
279 * Return the value of an integer Mod.
281 static tarval *computed_value_Mod(ir_node *n) {
282 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
283 } /* computed_value_Mod */
286 * Return the value of an Abs.
288 static tarval *computed_value_Abs(ir_node *n) {
289 ir_node *a = get_Abs_op(n);
290 tarval *ta = value_of(a);
292 if (ta != tarval_bad)
293 return tarval_abs(ta);
296 } /* computed_value_Abs */
299 * Return the value of an And.
300 * Special case: a & 0, 0 & b
302 static tarval *computed_value_And(ir_node *n) {
303 ir_node *a = get_And_left(n);
304 ir_node *b = get_And_right(n);
306 tarval *ta = value_of(a);
307 tarval *tb = value_of(b);
309 if ((ta != tarval_bad) && (tb != tarval_bad)) {
310 return tarval_and (ta, tb);
312 if (tarval_is_null(ta)) return ta;
313 if (tarval_is_null(tb)) return tb;
316 } /* computed_value_And */
319 * Return the value of an Or.
320 * Special case: a | 1...1, 1...1 | b
322 static tarval *computed_value_Or(ir_node *n) {
323 ir_node *a = get_Or_left(n);
324 ir_node *b = get_Or_right(n);
326 tarval *ta = value_of(a);
327 tarval *tb = value_of(b);
329 if ((ta != tarval_bad) && (tb != tarval_bad)) {
330 return tarval_or (ta, tb);
332 if (tarval_is_all_one(ta)) return ta;
333 if (tarval_is_all_one(tb)) return tb;
336 } /* computed_value_Or */
339 * Return the value of an Eor.
341 static tarval *computed_value_Eor(ir_node *n) {
342 ir_node *a = get_Eor_left(n);
343 ir_node *b = get_Eor_right(n);
348 return get_mode_null(get_irn_mode(n));
353 if ((ta != tarval_bad) && (tb != tarval_bad)) {
354 return tarval_eor (ta, tb);
357 } /* computed_value_Eor */
360 * Return the value of a Not.
362 static tarval *computed_value_Not(ir_node *n) {
363 ir_node *a = get_Not_op(n);
364 tarval *ta = value_of(a);
366 if (ta != tarval_bad)
367 return tarval_not(ta);
370 } /* computed_value_Not */
373 * Return the value of a Shl.
375 static tarval *computed_value_Shl(ir_node *n) {
376 ir_node *a = get_Shl_left(n);
377 ir_node *b = get_Shl_right(n);
379 tarval *ta = value_of(a);
380 tarval *tb = value_of(b);
382 if ((ta != tarval_bad) && (tb != tarval_bad)) {
383 return tarval_shl (ta, tb);
386 } /* computed_value_Shl */
389 * Return the value of a Shr.
391 static tarval *computed_value_Shr(ir_node *n) {
392 ir_node *a = get_Shr_left(n);
393 ir_node *b = get_Shr_right(n);
395 tarval *ta = value_of(a);
396 tarval *tb = value_of(b);
398 if ((ta != tarval_bad) && (tb != tarval_bad)) {
399 return tarval_shr (ta, tb);
402 } /* computed_value_Shr */
405 * Return the value of a Shrs.
407 static tarval *computed_value_Shrs(ir_node *n) {
408 ir_node *a = get_Shrs_left(n);
409 ir_node *b = get_Shrs_right(n);
411 tarval *ta = value_of(a);
412 tarval *tb = value_of(b);
414 if ((ta != tarval_bad) && (tb != tarval_bad)) {
415 return tarval_shrs (ta, tb);
418 } /* computed_value_Shrs */
421 * Return the value of a Rot.
423 static tarval *computed_value_Rot(ir_node *n) {
424 ir_node *a = get_Rot_left(n);
425 ir_node *b = get_Rot_right(n);
427 tarval *ta = value_of(a);
428 tarval *tb = value_of(b);
430 if ((ta != tarval_bad) && (tb != tarval_bad)) {
431 return tarval_rot (ta, tb);
434 } /* computed_value_Rot */
437 * Return the value of a Conv.
439 static tarval *computed_value_Conv(ir_node *n) {
440 ir_node *a = get_Conv_op(n);
441 tarval *ta = value_of(a);
443 if (ta != tarval_bad)
444 return tarval_convert_to(ta, get_irn_mode(n));
447 } /* computed_value_Conv */
450 * Return the value of a Proj(Cmp).
452 * This performs a first step of unreachable code elimination.
453 * Proj can not be computed, but folding a Cmp above the Proj here is
454 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
456 * There are several case where we can evaluate a Cmp node, see later.
458 static tarval *computed_value_Proj_Cmp(ir_node *n) {
459 ir_node *a = get_Proj_pred(n);
460 ir_node *aa = get_Cmp_left(a);
461 ir_node *ab = get_Cmp_right(a);
462 long proj_nr = get_Proj_proj(n);
465 * BEWARE: a == a is NOT always True for floating Point values, as
466 * NaN != NaN is defined, so we must check this here.
469 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
472 /* This is a trick with the bits used for encoding the Cmp
473 Proj numbers, the following statement is not the same:
474 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
475 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
478 tarval *taa = value_of(aa);
479 tarval *tab = value_of(ab);
480 ir_mode *mode = get_irn_mode(aa);
483 * The predecessors of Cmp are target values. We can evaluate
486 if ((taa != tarval_bad) && (tab != tarval_bad)) {
487 /* strange checks... */
488 pn_Cmp flags = tarval_cmp(taa, tab);
489 if (flags != pn_Cmp_False) {
490 return new_tarval_from_long (proj_nr & flags, mode_b);
493 /* for integer values, we can check against MIN/MAX */
494 else if (mode_is_int(mode)) {
495 /* MIN <=/> x. This results in true/false. */
496 if (taa == get_mode_min(mode)) {
497 /* a compare with the MIN value */
498 if (proj_nr == pn_Cmp_Le)
499 return get_tarval_b_true();
500 else if (proj_nr == pn_Cmp_Gt)
501 return get_tarval_b_false();
503 /* x >=/< MIN. This results in true/false. */
505 if (tab == get_mode_min(mode)) {
506 /* a compare with the MIN value */
507 if (proj_nr == pn_Cmp_Ge)
508 return get_tarval_b_true();
509 else if (proj_nr == pn_Cmp_Lt)
510 return get_tarval_b_false();
512 /* MAX >=/< x. This results in true/false. */
513 else if (taa == get_mode_max(mode)) {
514 if (proj_nr == pn_Cmp_Ge)
515 return get_tarval_b_true();
516 else if (proj_nr == pn_Cmp_Lt)
517 return get_tarval_b_false();
519 /* x <=/> MAX. This results in true/false. */
520 else if (tab == get_mode_max(mode)) {
521 if (proj_nr == pn_Cmp_Le)
522 return get_tarval_b_true();
523 else if (proj_nr == pn_Cmp_Gt)
524 return get_tarval_b_false();
528 * The predecessors are Allocs or (void*)(0) constants. Allocs never
529 * return NULL, they raise an exception. Therefore we can predict
533 ir_node *aaa = skip_Id(skip_Proj(aa));
534 ir_node *aba = skip_Id(skip_Proj(ab));
536 if ( ( (/* aa is ProjP and aaa is Alloc */
538 && mode_is_reference(get_irn_mode(aa))
540 && ( (/* ab is NULL */
542 && mode_is_reference(get_irn_mode(ab))
543 && is_Const_null(ab))
544 || (/* ab is other Alloc */
546 && mode_is_reference(get_irn_mode(ab))
549 || (/* aa is NULL and aba is Alloc */
551 && mode_is_reference(get_irn_mode(aa))
554 && mode_is_reference(get_irn_mode(ab))
557 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
560 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
561 } /* computed_value_Proj_Cmp */
564 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
565 * Proj(DivMod) and Proj(Quot).
567 static tarval *computed_value_Proj(ir_node *n) {
568 ir_node *a = get_Proj_pred(n);
571 switch (get_irn_opcode(a)) {
573 return computed_value_Proj_Cmp(n);
576 /* compute either the Div or the Mod part */
577 proj_nr = get_Proj_proj(n);
578 if (proj_nr == pn_DivMod_res_div)
579 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
580 else if (proj_nr == pn_DivMod_res_mod)
581 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
585 if (get_Proj_proj(n) == pn_Div_res)
586 return computed_value(a);
590 if (get_Proj_proj(n) == pn_Mod_res)
591 return computed_value(a);
595 if (get_Proj_proj(n) == pn_Quot_res)
596 return computed_value(a);
603 } /* computed_value_Proj */
606 * Calculate the value of a Mux: can be evaluated, if the
607 * sel and the right input are known.
609 static tarval *computed_value_Mux(ir_node *n) {
610 ir_node *sel = get_Mux_sel(n);
611 tarval *ts = value_of(sel);
613 if (ts == get_tarval_b_true()) {
614 ir_node *v = get_Mux_true(n);
617 else if (ts == get_tarval_b_false()) {
618 ir_node *v = get_Mux_false(n);
622 } /* computed_value_Mux */
625 * Calculate the value of a Psi: can be evaluated, if a condition is true
626 * and all previous conditions are false. If all conditions are false
627 * we evaluate to the default one.
629 static tarval *computed_value_Psi(ir_node *n) {
631 return computed_value_Mux(n);
633 } /* computed_value_Psi */
636 * Calculate the value of a Confirm: can be evaluated,
637 * if it has the form Confirm(x, '=', Const).
639 static tarval *computed_value_Confirm(ir_node *n) {
640 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
641 value_of(get_Confirm_bound(n)) : tarval_bad;
642 } /* computed_value_Confirm */
645 * If the parameter n can be computed, return its value, else tarval_bad.
646 * Performs constant folding.
648 * @param n The node this should be evaluated
650 tarval *computed_value(ir_node *n) {
651 if (n->op->ops.computed_value)
652 return n->op->ops.computed_value(n);
654 } /* computed_value */
657 * Set the default computed_value evaluator in an ir_op_ops.
659 * @param code the opcode for the default operation
660 * @param ops the operations initialized
665 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
669 ops->computed_value = computed_value_##a; \
704 } /* firm_set_default_computed_value */
707 * Returns a equivalent block for another block.
708 * If the block has only one predecessor, this is
709 * the equivalent one. If the only predecessor of a block is
710 * the block itself, this is a dead block.
712 * If both predecessors of a block are the branches of a binary
713 * Cond, the equivalent block is Cond's block.
715 * If all predecessors of a block are bad or lies in a dead
716 * block, the current block is dead as well.
718 * Note, that blocks are NEVER turned into Bad's, instead
719 * the dead_block flag is set. So, never test for is_Bad(block),
720 * always use is_dead_Block(block).
722 static ir_node *equivalent_node_Block(ir_node *n)
727 /* don't optimize dead blocks */
728 if (is_Block_dead(n))
731 n_preds = get_Block_n_cfgpreds(n);
733 /* The Block constructor does not call optimize, but mature_immBlock()
734 calls the optimization. */
735 assert(get_Block_matured(n));
737 /* Straightening: a single entry Block following a single exit Block
738 can be merged, if it is not the Start block. */
739 /* !!! Beware, all Phi-nodes of n must have been optimized away.
740 This should be true, as the block is matured before optimize is called.
741 But what about Phi-cycles with the Phi0/Id that could not be resolved?
742 Remaining Phi nodes are just Ids. */
743 if (n_preds == 1 && is_Jmp(get_Block_cfgpred(n, 0))) {
744 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
745 if (predblock == oldn) {
746 /* Jmp jumps into the block it is in -- deal self cycle. */
747 n = set_Block_dead(n);
748 DBG_OPT_DEAD_BLOCK(oldn, n);
749 } else if (get_opt_control_flow_straightening()) {
751 DBG_OPT_STG(oldn, n);
753 } else if (n_preds == 1 && is_Cond(skip_Proj(get_Block_cfgpred(n, 0)))) {
754 ir_node *predblock = get_Block_cfgpred_block(n, 0);
755 if (predblock == oldn) {
756 /* Jmp jumps into the block it is in -- deal self cycle. */
757 n = set_Block_dead(n);
758 DBG_OPT_DEAD_BLOCK(oldn, n);
760 } else if ((n_preds == 2) &&
761 (get_opt_control_flow_weak_simplification())) {
762 /* Test whether Cond jumps twice to this block
763 * The more general case which more than 2 predecessors is handles
764 * in optimize_cf(), we handle only this special case for speed here.
766 ir_node *a = get_Block_cfgpred(n, 0);
767 ir_node *b = get_Block_cfgpred(n, 1);
771 (get_Proj_pred(a) == get_Proj_pred(b)) &&
772 is_Cond(get_Proj_pred(a)) &&
773 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
774 /* Also a single entry Block following a single exit Block. Phis have
775 twice the same operand and will be optimized away. */
776 n = get_nodes_block(get_Proj_pred(a));
777 DBG_OPT_IFSIM1(oldn, a, b, n);
779 } else if (get_opt_unreachable_code() &&
780 (n != get_irg_start_block(current_ir_graph)) &&
781 (n != get_irg_end_block(current_ir_graph)) ) {
784 /* If all inputs are dead, this block is dead too, except if it is
785 the start or end block. This is one step of unreachable code
787 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
788 ir_node *pred = get_Block_cfgpred(n, i);
791 if (is_Bad(pred)) continue;
792 pred_blk = get_nodes_block(skip_Proj(pred));
794 if (is_Block_dead(pred_blk)) continue;
797 /* really found a living input */
802 n = set_Block_dead(n);
803 DBG_OPT_DEAD_BLOCK(oldn, n);
808 } /* equivalent_node_Block */
811 * Returns a equivalent node for a Jmp, a Bad :-)
812 * Of course this only happens if the Block of the Jmp is dead.
814 static ir_node *equivalent_node_Jmp(ir_node *n) {
815 /* unreachable code elimination */
816 if (is_Block_dead(get_nodes_block(n)))
820 } /* equivalent_node_Jmp */
822 /** Raise is handled in the same way as Jmp. */
823 #define equivalent_node_Raise equivalent_node_Jmp
826 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
827 See transform_node_Proj_Cond(). */
830 * Optimize operations that are commutative and have neutral 0,
831 * so a op 0 = 0 op a = a.
833 static ir_node *equivalent_node_neutral_zero(ir_node *n)
837 ir_node *a = get_binop_left(n);
838 ir_node *b = get_binop_right(n);
843 /* After running compute_node there is only one constant predecessor.
844 Find this predecessors value and remember the other node: */
845 if ((tv = value_of(a)) != tarval_bad) {
847 } else if ((tv = value_of(b)) != tarval_bad) {
852 /* If this predecessors constant value is zero, the operation is
853 * unnecessary. Remove it.
855 * Beware: If n is a Add, the mode of on and n might be different
856 * which happens in this rare construction: NULL + 3.
857 * Then, a Conv would be needed which we cannot include here.
859 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
862 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
866 } /* equivalent_node_neutral_zero */
869 * Eor is commutative and has neutral 0.
871 static ir_node *equivalent_node_Eor(ir_node *n)
877 n = equivalent_node_neutral_zero(n);
878 if (n != oldn) return n;
881 b = get_Eor_right(n);
884 ir_node *aa = get_Eor_left(a);
885 ir_node *ab = get_Eor_right(a);
888 /* (a ^ b) ^ a -> b */
890 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
891 } else if (ab == b) {
892 /* (a ^ b) ^ b -> a */
894 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
898 ir_node *ba = get_Eor_left(b);
899 ir_node *bb = get_Eor_right(b);
902 /* a ^ (a ^ b) -> b */
904 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
905 } else if (bb == a) {
906 /* a ^ (b ^ a) -> b */
908 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
916 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
918 * The second one looks strange, but this construct
919 * is used heavily in the LCC sources :-).
921 * Beware: The Mode of an Add may be different than the mode of its
922 * predecessors, so we could not return a predecessors in all cases.
924 static ir_node *equivalent_node_Add(ir_node *n) {
926 ir_node *left, *right;
927 ir_mode *mode = get_irn_mode(n);
929 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
930 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
933 n = equivalent_node_neutral_zero(n);
937 left = get_Add_left(n);
938 right = get_Add_right(n);
941 if (get_Sub_right(left) == right) {
944 n = get_Sub_left(left);
945 if (mode == get_irn_mode(n)) {
946 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
952 if (get_Sub_right(right) == left) {
955 n = get_Sub_left(right);
956 if (mode == get_irn_mode(n)) {
957 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
963 } /* equivalent_node_Add */
966 * optimize operations that are not commutative but have neutral 0 on left,
969 static ir_node *equivalent_node_left_zero(ir_node *n) {
972 ir_node *a = get_binop_left(n);
973 ir_node *b = get_binop_right(n);
975 if (is_Const(b) && is_Const_null(b)) {
978 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
981 } /* equivalent_node_left_zero */
983 #define equivalent_node_Shl equivalent_node_left_zero
984 #define equivalent_node_Shr equivalent_node_left_zero
985 #define equivalent_node_Shrs equivalent_node_left_zero
986 #define equivalent_node_Rot equivalent_node_left_zero
989 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
991 * The second one looks strange, but this construct
992 * is used heavily in the LCC sources :-).
994 * Beware: The Mode of a Sub may be different than the mode of its
995 * predecessors, so we could not return a predecessors in all cases.
997 static ir_node *equivalent_node_Sub(ir_node *n) {
1000 ir_mode *mode = get_irn_mode(n);
1002 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1003 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1006 b = get_Sub_right(n);
1008 /* Beware: modes might be different */
1009 if (is_Const(b) && is_Const_null(b)) {
1010 ir_node *a = get_Sub_left(n);
1011 if (mode == get_irn_mode(a)) {
1014 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1018 } /* equivalent_node_Sub */
1022 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1025 * -(-a) == a, but might overflow two times.
1026 * We handle it anyway here but the better way would be a
1027 * flag. This would be needed for Pascal for instance.
1029 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1031 ir_node *pred = get_unop_op(n);
1033 /* optimize symmetric unop */
1034 if (get_irn_op(pred) == get_irn_op(n)) {
1035 n = get_unop_op(pred);
1036 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1039 } /* equivalent_node_idempotent_unop */
1041 /** Optimize Not(Not(x)) == x. */
1042 #define equivalent_node_Not equivalent_node_idempotent_unop
1044 /** -(-x) == x ??? Is this possible or can --x raise an
1045 out of bounds exception if min =! max? */
1046 #define equivalent_node_Minus equivalent_node_idempotent_unop
1049 * Optimize a * 1 = 1 * a = a.
1051 static ir_node *equivalent_node_Mul(ir_node *n) {
1053 ir_node *a = get_Mul_left(n);
1055 /* we can handle here only the n * n = n bit cases */
1056 if (get_irn_mode(n) == get_irn_mode(a)) {
1057 ir_node *b = get_Mul_right(n);
1059 /* Mul is commutative and has again an other neutral element. */
1060 if (is_Const(a) && is_Const_one(a)) {
1062 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1063 } else if (is_Const(b) && is_Const_one(b)) {
1065 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1069 } /* equivalent_node_Mul */
1072 * Optimize a / 1 = a.
1074 static ir_node *equivalent_node_Div(ir_node *n) {
1075 ir_node *a = get_Div_left(n);
1076 ir_node *b = get_Div_right(n);
1078 /* Div is not commutative. */
1079 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1080 /* Turn Div into a tuple (mem, bad, a) */
1081 ir_node *mem = get_Div_mem(n);
1082 ir_node *blk = get_irn_n(n, -1);
1083 turn_into_tuple(n, pn_Div_max);
1084 set_Tuple_pred(n, pn_Div_M, mem);
1085 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1086 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1087 set_Tuple_pred(n, pn_Div_res, a);
1090 } /* equivalent_node_Div */
1093 * Optimize a / 1.0 = a.
1095 static ir_node *equivalent_node_Quot(ir_node *n) {
1096 ir_node *a = get_Quot_left(n);
1097 ir_node *b = get_Quot_right(n);
1099 /* Div is not commutative. */
1100 if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */
1101 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1102 ir_node *mem = get_Quot_mem(n);
1103 ir_node *blk = get_irn_n(n, -1);
1104 turn_into_tuple(n, pn_Quot_max);
1105 set_Tuple_pred(n, pn_Quot_M, mem);
1106 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1107 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1108 set_Tuple_pred(n, pn_Quot_res, a);
1111 } /* equivalent_node_Quot */
1114 * Optimize a / 1 = a.
1116 static ir_node *equivalent_node_DivMod(ir_node *n) {
1117 ir_node *b = get_DivMod_right(n);
1119 /* Div is not commutative. */
1120 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1121 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1122 ir_node *a = get_DivMod_left(n);
1123 ir_node *mem = get_Div_mem(n);
1124 ir_node *blk = get_irn_n(n, -1);
1125 ir_mode *mode = get_DivMod_resmode(n);
1127 turn_into_tuple(n, pn_DivMod_max);
1128 set_Tuple_pred(n, pn_DivMod_M, mem);
1129 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1130 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1131 set_Tuple_pred(n, pn_DivMod_res_div, a);
1132 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1135 } /* equivalent_node_DivMod */
1138 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1140 static ir_node *equivalent_node_Or(ir_node *n) {
1143 ir_node *a = get_Or_left(n);
1144 ir_node *b = get_Or_right(n);
1147 n = a; /* Or has it's own neutral element */
1148 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1149 } else if (is_Const(a) && is_Const_null(a)) {
1151 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1152 } else if (is_Const(b) && is_Const_null(b)) {
1154 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1158 } /* equivalent_node_Or */
1161 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1163 static ir_node *equivalent_node_And(ir_node *n) {
1166 ir_node *a = get_And_left(n);
1167 ir_node *b = get_And_right(n);
1170 n = a; /* And has it's own neutral element */
1171 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1174 if (is_Const(a) && is_Const_all_one(a)) {
1176 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1179 if (is_Const(b) && is_Const_all_one(b)) {
1181 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1185 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1188 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1193 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1196 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1202 } /* equivalent_node_And */
1205 * Try to remove useless Conv's:
1207 static ir_node *equivalent_node_Conv(ir_node *n) {
1209 ir_node *a = get_Conv_op(n);
1211 ir_mode *n_mode = get_irn_mode(n);
1212 ir_mode *a_mode = get_irn_mode(a);
1214 if (n_mode == a_mode) { /* No Conv necessary */
1215 if (get_Conv_strict(n)) {
1216 /* special case: the predecessor might be a also a Conv */
1218 if (! get_Conv_strict(a)) {
1219 /* first one is not strict, kick it */
1220 set_Conv_op(n, get_Conv_op(a));
1223 /* else both are strict conv, second is superflous */
1224 } else if(is_Proj(a)) {
1225 ir_node *pred = get_Proj_pred(a);
1227 /* loads always return with the exact precision of n_mode */
1228 assert(get_Load_mode(pred) == n_mode);
1233 /* leave strict floating point Conv's */
1237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1238 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1239 ir_node *b = get_Conv_op(a);
1240 ir_mode *b_mode = get_irn_mode(b);
1242 if (n_mode == b_mode) {
1243 if (n_mode == mode_b) {
1244 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1245 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1246 } else if (mode_is_int(n_mode)) {
1247 if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) {
1248 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1249 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1255 } /* equivalent_node_Conv */
1258 * A Cast may be removed if the type of the previous node
1259 * is already the type of the Cast.
1261 static ir_node *equivalent_node_Cast(ir_node *n) {
1263 ir_node *pred = get_Cast_op(n);
1265 if (get_irn_type(pred) == get_Cast_type(n)) {
1267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1270 } /* equivalent_node_Cast */
1273 * Several optimizations:
1274 * - no Phi in start block.
1275 * - remove Id operators that are inputs to Phi
1276 * - fold Phi-nodes, iff they have only one predecessor except
1279 static ir_node *equivalent_node_Phi(ir_node *n) {
1284 ir_node *first_val = NULL; /* to shutup gcc */
1286 if (!get_opt_normalize()) return n;
1288 n_preds = get_Phi_n_preds(n);
1290 block = get_nodes_block(n);
1291 if ((is_Block_dead(block)) || /* Control dead */
1292 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1293 return new_Bad(); /* in the Start Block. */
1295 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1297 /* If the Block has a Bad pred, we also have one. */
1298 for (i = 0; i < n_preds; ++i)
1299 if (is_Bad(get_Block_cfgpred(block, i)))
1300 set_Phi_pred(n, i, new_Bad());
1302 /* Find first non-self-referencing input */
1303 for (i = 0; i < n_preds; ++i) {
1304 first_val = get_Phi_pred(n, i);
1305 if ( (first_val != n) /* not self pointer */
1307 && (! is_Bad(first_val))
1309 ) { /* value not dead */
1310 break; /* then found first value. */
1315 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1319 /* search for rest of inputs, determine if any of these
1320 are non-self-referencing */
1321 while (++i < n_preds) {
1322 ir_node *scnd_val = get_Phi_pred(n, i);
1323 if ( (scnd_val != n)
1324 && (scnd_val != first_val)
1326 && (! is_Bad(scnd_val))
1334 /* Fold, if no multiple distinct non-self-referencing inputs */
1336 DBG_OPT_PHI(oldn, n);
1339 } /* equivalent_node_Phi */
1342 * Several optimizations:
1343 * - no Sync in start block.
1344 * - fold Sync-nodes, iff they have only one predecessor except
1347 static ir_node *equivalent_node_Sync(ir_node *n) {
1348 int arity = get_Sync_n_preds(n);
1351 for (i = 0; i < arity;) {
1352 ir_node *pred = get_Sync_pred(n, i);
1355 /* Remove Bad predecessors */
1362 /* Remove duplicate predecessors */
1368 if (get_Sync_pred(n, j) == pred) {
1376 if (arity == 0) return new_Bad();
1377 if (arity == 1) return get_Sync_pred(n, 0);
1379 } /* equivalent_node_Sync */
1382 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1383 * ProjX(Load) and ProjX(Store).
1385 static ir_node *equivalent_node_Proj(ir_node *proj) {
1386 ir_node *oldn = proj;
1387 ir_node *a = get_Proj_pred(proj);
1390 /* Remove the Tuple/Proj combination. */
1391 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1392 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1393 DBG_OPT_TUPLE(oldn, a, proj);
1395 /* This should not happen! */
1396 assert(! "found a Proj with higher number than Tuple predecessors");
1399 } else if (get_irn_mode(proj) == mode_X) {
1400 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1401 /* Remove dead control flow -- early gigo(). */
1403 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1405 /* get the Load address */
1406 ir_node *addr = get_Load_ptr(a);
1407 ir_node *blk = get_irn_n(a, -1);
1410 if (value_not_null(addr, &confirm)) {
1411 if (confirm == NULL) {
1412 /* this node may float if it did not depend on a Confirm */
1413 set_irn_pinned(a, op_pin_state_floats);
1415 if (get_Proj_proj(proj) == pn_Load_X_except) {
1416 DBG_OPT_EXC_REM(proj);
1419 return new_r_Jmp(current_ir_graph, blk);
1421 } else if (is_Store(a)) {
1422 /* get the load/store address */
1423 ir_node *addr = get_Store_ptr(a);
1424 ir_node *blk = get_irn_n(a, -1);
1427 if (value_not_null(addr, &confirm)) {
1428 if (confirm == NULL) {
1429 /* this node may float if it did not depend on a Confirm */
1430 set_irn_pinned(a, op_pin_state_floats);
1432 if (get_Proj_proj(proj) == pn_Store_X_except) {
1433 DBG_OPT_EXC_REM(proj);
1436 return new_r_Jmp(current_ir_graph, blk);
1443 } /* equivalent_node_Proj */
1448 static ir_node *equivalent_node_Id(ir_node *n) {
1453 } while (get_irn_op(n) == op_Id);
1455 DBG_OPT_ID(oldn, n);
1457 } /* equivalent_node_Id */
1462 static ir_node *equivalent_node_Mux(ir_node *n)
1464 ir_node *oldn = n, *sel = get_Mux_sel(n);
1465 tarval *ts = value_of(sel);
1467 /* Mux(true, f, t) == t */
1468 if (ts == tarval_b_true) {
1469 n = get_Mux_true(n);
1470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1472 /* Mux(false, f, t) == f */
1473 else if (ts == tarval_b_false) {
1474 n = get_Mux_false(n);
1475 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1477 /* Mux(v, x, x) == x */
1478 else if (get_Mux_false(n) == get_Mux_true(n)) {
1479 n = get_Mux_true(n);
1480 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1482 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1483 ir_node *cmp = get_Proj_pred(sel);
1484 long proj_nr = get_Proj_proj(sel);
1485 ir_node *f = get_Mux_false(n);
1486 ir_node *t = get_Mux_true(n);
1489 * Note further that these optimization work even for floating point
1490 * with NaN's because -NaN == NaN.
1491 * However, if +0 and -0 is handled differently, we cannot use the first one.
1494 ir_node *const cmp_l = get_Cmp_left(cmp);
1495 ir_node *const cmp_r = get_Cmp_right(cmp);
1499 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1500 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1502 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1509 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1510 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1512 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1519 * Note: normalization puts the constant on the right side,
1520 * so we check only one case.
1522 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1523 /* Mux(t CMP 0, X, t) */
1524 if (is_Minus(f) && get_Minus_op(f) == t) {
1525 /* Mux(t CMP 0, -t, t) */
1526 if (proj_nr == pn_Cmp_Eq) {
1527 /* Mux(t == 0, -t, t) ==> -t */
1529 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1530 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1531 /* Mux(t != 0, -t, t) ==> t */
1533 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1540 } /* equivalent_node_Mux */
1543 * Returns a equivalent node of a Psi: if a condition is true
1544 * and all previous conditions are false we know its value.
1545 * If all conditions are false its value is the default one.
1547 static ir_node *equivalent_node_Psi(ir_node *n) {
1549 return equivalent_node_Mux(n);
1551 } /* equivalent_node_Psi */
1554 * Optimize -a CMP -b into b CMP a.
1555 * This works only for for modes where unary Minus
1557 * Note that two-complement integers can Overflow
1558 * so it will NOT work.
1560 * For == and != can be handled in Proj(Cmp)
1562 static ir_node *equivalent_node_Cmp(ir_node *n) {
1563 ir_node *left = get_Cmp_left(n);
1564 ir_node *right = get_Cmp_right(n);
1566 if (is_Minus(left) && is_Minus(right) &&
1567 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1568 left = get_Minus_op(left);
1569 right = get_Minus_op(right);
1570 set_Cmp_left(n, right);
1571 set_Cmp_right(n, left);
1574 } /* equivalent_node_Cmp */
1577 * Remove Confirm nodes if setting is on.
1578 * Replace Confirms(x, '=', Constlike) by Constlike.
1580 static ir_node *equivalent_node_Confirm(ir_node *n) {
1581 ir_node *pred = get_Confirm_value(n);
1582 pn_Cmp pnc = get_Confirm_cmp(n);
1584 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1586 * rare case: two identical Confirms one after another,
1587 * replace the second one with the first.
1591 if (pnc == pn_Cmp_Eq) {
1592 ir_node *bound = get_Confirm_bound(n);
1595 * Optimize a rare case:
1596 * Confirm(x, '=', Constlike) ==> Constlike
1598 if (is_irn_constlike(bound)) {
1599 DBG_OPT_CONFIRM(n, bound);
1603 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1607 * Optimize CopyB(mem, x, x) into a Nop.
1609 static ir_node *equivalent_node_CopyB(ir_node *n) {
1610 ir_node *a = get_CopyB_dst(n);
1611 ir_node *b = get_CopyB_src(n);
1614 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1615 ir_node *mem = get_CopyB_mem(n);
1616 ir_node *blk = get_nodes_block(n);
1617 turn_into_tuple(n, pn_CopyB_max);
1618 set_Tuple_pred(n, pn_CopyB_M, mem);
1619 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1620 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1621 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1624 } /* equivalent_node_CopyB */
1627 * Optimize Bounds(idx, idx, upper) into idx.
1629 static ir_node *equivalent_node_Bound(ir_node *n) {
1630 ir_node *idx = get_Bound_index(n);
1631 ir_node *pred = skip_Proj(idx);
1634 if (is_Bound(pred)) {
1636 * idx was Bounds checked in the same MacroBlock previously,
1637 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1639 ir_node *lower = get_Bound_lower(n);
1640 ir_node *upper = get_Bound_upper(n);
1641 if (get_Bound_lower(pred) == lower &&
1642 get_Bound_upper(pred) == upper &&
1643 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1645 * One could expect that we simply return the previous
1646 * Bound here. However, this would be wrong, as we could
1647 * add an exception Proj to a new location then.
1648 * So, we must turn in into a tuple.
1654 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1655 ir_node *mem = get_Bound_mem(n);
1656 ir_node *blk = get_nodes_block(n);
1657 turn_into_tuple(n, pn_Bound_max);
1658 set_Tuple_pred(n, pn_Bound_M, mem);
1659 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1660 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1661 set_Tuple_pred(n, pn_Bound_res, idx);
1664 } /* equivalent_node_Bound */
1667 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1668 * perform no actual computation, as, e.g., the Id nodes. It does not create
1669 * new nodes. It is therefore safe to free n if the node returned is not n.
1670 * If a node returns a Tuple we can not just skip it. If the size of the
1671 * in array fits, we transform n into a tuple (e.g., Div).
1673 ir_node *equivalent_node(ir_node *n) {
1674 if (n->op->ops.equivalent_node)
1675 return n->op->ops.equivalent_node(n);
1677 } /* equivalent_node */
1680 * Sets the default equivalent node operation for an ir_op_ops.
1682 * @param code the opcode for the default operation
1683 * @param ops the operations initialized
1688 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1692 ops->equivalent_node = equivalent_node_##a; \
1732 } /* firm_set_default_equivalent_node */
1735 * Returns non-zero if a node is a Phi node
1736 * with all predecessors constant.
1738 static int is_const_Phi(ir_node *n) {
1741 if (! is_Phi(n) || get_irn_arity(n) == 0)
1743 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1744 if (! is_Const(get_irn_n(n, i)))
1747 } /* is_const_Phi */
1750 * Apply an evaluator on a binop with a constant operators (and one Phi).
1752 * @param phi the Phi node
1753 * @param other the other operand
1754 * @param eval an evaluator function
1755 * @param mode the mode of the result, may be different from the mode of the Phi!
1756 * @param left if non-zero, other is the left operand, else the right
1758 * @return a new Phi node if the conversion was successful, NULL else
1760 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1765 int i, n = get_irn_arity(phi);
1767 NEW_ARR_A(void *, res, n);
1769 for (i = 0; i < n; ++i) {
1770 pred = get_irn_n(phi, i);
1771 tv = get_Const_tarval(pred);
1772 tv = eval(other, tv);
1774 if (tv == tarval_bad) {
1775 /* folding failed, bad */
1781 for (i = 0; i < n; ++i) {
1782 pred = get_irn_n(phi, i);
1783 tv = get_Const_tarval(pred);
1784 tv = eval(tv, other);
1786 if (tv == tarval_bad) {
1787 /* folding failed, bad */
1793 irg = current_ir_graph;
1794 for (i = 0; i < n; ++i) {
1795 pred = get_irn_n(phi, i);
1796 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1797 mode, res[i], get_Const_type(pred));
1799 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1800 } /* apply_binop_on_phi */
1803 * Apply an evaluator on a binop with two constant Phi.
1805 * @param a the left Phi node
1806 * @param b the right Phi node
1807 * @param eval an evaluator function
1808 * @param mode the mode of the result, may be different from the mode of the Phi!
1810 * @return a new Phi node if the conversion was successful, NULL else
1812 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1813 tarval *tv_l, *tv_r, *tv;
1819 if (get_nodes_block(a) != get_nodes_block(b))
1822 n = get_irn_arity(a);
1823 NEW_ARR_A(void *, res, n);
1825 for (i = 0; i < n; ++i) {
1826 pred = get_irn_n(a, i);
1827 tv_l = get_Const_tarval(pred);
1828 pred = get_irn_n(b, i);
1829 tv_r = get_Const_tarval(pred);
1830 tv = eval(tv_l, tv_r);
1832 if (tv == tarval_bad) {
1833 /* folding failed, bad */
1838 irg = current_ir_graph;
1839 for (i = 0; i < n; ++i) {
1840 pred = get_irn_n(a, i);
1841 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1843 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1844 } /* apply_binop_on_2_phis */
1847 * Apply an evaluator on a unop with a constant operator (a Phi).
1849 * @param phi the Phi node
1850 * @param eval an evaluator function
1852 * @return a new Phi node if the conversion was successful, NULL else
1854 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1860 int i, n = get_irn_arity(phi);
1862 NEW_ARR_A(void *, res, n);
1863 for (i = 0; i < n; ++i) {
1864 pred = get_irn_n(phi, i);
1865 tv = get_Const_tarval(pred);
1868 if (tv == tarval_bad) {
1869 /* folding failed, bad */
1874 mode = get_irn_mode(phi);
1875 irg = current_ir_graph;
1876 for (i = 0; i < n; ++i) {
1877 pred = get_irn_n(phi, i);
1878 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1879 mode, res[i], get_Const_type(pred));
1881 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1882 } /* apply_unop_on_phi */
1885 * Apply a conversion on a constant operator (a Phi).
1887 * @param phi the Phi node
1889 * @return a new Phi node if the conversion was successful, NULL else
1891 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1896 int i, n = get_irn_arity(phi);
1898 NEW_ARR_A(void *, res, n);
1899 for (i = 0; i < n; ++i) {
1900 pred = get_irn_n(phi, i);
1901 tv = get_Const_tarval(pred);
1902 tv = tarval_convert_to(tv, mode);
1904 if (tv == tarval_bad) {
1905 /* folding failed, bad */
1910 irg = current_ir_graph;
1911 for (i = 0; i < n; ++i) {
1912 pred = get_irn_n(phi, i);
1913 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1914 mode, res[i], get_Const_type(pred));
1916 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1917 } /* apply_conv_on_phi */
1920 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1921 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1922 * If possible, remove the Conv's.
1924 static ir_node *transform_node_AddSub(ir_node *n) {
1925 ir_mode *mode = get_irn_mode(n);
1927 if (mode_is_reference(mode)) {
1928 ir_node *left = get_binop_left(n);
1929 ir_node *right = get_binop_right(n);
1930 unsigned ref_bits = get_mode_size_bits(mode);
1932 if (is_Conv(left)) {
1933 ir_mode *lmode = get_irn_mode(left);
1934 unsigned bits = get_mode_size_bits(lmode);
1936 if (ref_bits == bits &&
1937 mode_is_int(lmode) &&
1938 get_mode_arithmetic(lmode) == irma_twos_complement) {
1939 ir_node *pre = get_Conv_op(left);
1940 ir_mode *pre_mode = get_irn_mode(pre);
1942 if (mode_is_int(pre_mode) &&
1943 get_mode_size_bits(pre_mode) == bits &&
1944 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1945 /* ok, this conv just changes to sign, moreover the calculation
1946 * is done with same number of bits as our address mode, so
1947 * we can ignore the conv as address calculation can be viewed
1948 * as either signed or unsigned
1950 set_binop_left(n, pre);
1955 if (is_Conv(right)) {
1956 ir_mode *rmode = get_irn_mode(right);
1957 unsigned bits = get_mode_size_bits(rmode);
1959 if (ref_bits == bits &&
1960 mode_is_int(rmode) &&
1961 get_mode_arithmetic(rmode) == irma_twos_complement) {
1962 ir_node *pre = get_Conv_op(right);
1963 ir_mode *pre_mode = get_irn_mode(pre);
1965 if (mode_is_int(pre_mode) &&
1966 get_mode_size_bits(pre_mode) == bits &&
1967 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1968 /* ok, this conv just changes to sign, moreover the calculation
1969 * is done with same number of bits as our address mode, so
1970 * we can ignore the conv as address calculation can be viewed
1971 * as either signed or unsigned
1973 set_binop_right(n, pre);
1978 /* let address arithmetic use unsigned modes */
1979 if (is_Const(right)) {
1980 ir_mode *rmode = get_irn_mode(right);
1982 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1983 /* convert a AddP(P, *s) into AddP(P, *u) */
1984 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1986 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
1987 set_binop_right(n, pre);
1992 } /* transform_node_AddSub */
1994 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1996 if (is_Const(b) && is_const_Phi(a)) { \
1997 /* check for Op(Phi, Const) */ \
1998 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2000 else if (is_Const(a) && is_const_Phi(b)) { \
2001 /* check for Op(Const, Phi) */ \
2002 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2004 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2005 /* check for Op(Phi, Phi) */ \
2006 c = apply_binop_on_2_phis(a, b, eval, mode); \
2009 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2013 #define HANDLE_UNOP_PHI(eval, a, c) \
2015 if (is_const_Phi(a)) { \
2016 /* check for Op(Phi) */ \
2017 c = apply_unop_on_phi(a, eval); \
2019 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2025 * Do the AddSub optimization, then Transform
2026 * Constant folding on Phi
2027 * Add(a,a) -> Mul(a, 2)
2028 * Add(Mul(a, x), a) -> Mul(a, x+1)
2029 * if the mode is integer or float.
2030 * Transform Add(a,-b) into Sub(a,b).
2031 * Reassociation might fold this further.
2033 static ir_node *transform_node_Add(ir_node *n) {
2035 ir_node *a, *b, *c, *oldn = n;
2037 n = transform_node_AddSub(n);
2039 a = get_Add_left(n);
2040 b = get_Add_right(n);
2042 mode = get_irn_mode(n);
2043 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2045 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2046 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2049 if (mode_is_num(mode)) {
2050 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2051 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2052 ir_node *block = get_irn_n(n, -1);
2055 get_irn_dbg_info(n),
2059 new_r_Const_long(current_ir_graph, block, mode, 2),
2061 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2066 get_irn_dbg_info(n),
2072 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2077 get_irn_dbg_info(n),
2083 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2086 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2087 /* Here we rely on constants be on the RIGHT side */
2089 ir_node *op = get_Not_op(a);
2091 if (is_Const(b) && is_Const_one(b)) {
2093 ir_node *blk = get_irn_n(n, -1);
2094 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2095 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2100 ir_node *blk = get_irn_n(n, -1);
2101 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2102 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2107 ir_node *op = get_Not_op(b);
2111 ir_node *blk = get_irn_n(n, -1);
2112 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2113 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2120 } /* transform_node_Add */
2123 * returns -cnst or NULL if impossible
2125 static ir_node *const_negate(ir_node *cnst) {
2126 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2127 dbg_info *dbgi = get_irn_dbg_info(cnst);
2128 ir_graph *irg = get_irn_irg(cnst);
2129 ir_node *block = get_nodes_block(cnst);
2130 ir_mode *mode = get_irn_mode(cnst);
2131 if (tv == tarval_bad) return NULL;
2132 return new_rd_Const(dbgi, irg, block, mode, tv);
2136 * Do the AddSub optimization, then Transform
2137 * Constant folding on Phi
2138 * Sub(0,a) -> Minus(a)
2139 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2140 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2141 * Sub(Add(a, x), x) -> a
2142 * Sub(x, Add(x, a)) -> -a
2143 * Sub(x, Const) -> Add(x, -Const)
2145 static ir_node *transform_node_Sub(ir_node *n) {
2150 n = transform_node_AddSub(n);
2152 a = get_Sub_left(n);
2153 b = get_Sub_right(n);
2155 mode = get_irn_mode(n);
2158 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2160 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2161 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2164 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2165 /* a - C -> a + (-C) */
2166 ir_node *cnst = const_negate(b);
2168 ir_node *block = get_nodes_block(n);
2169 dbg_info *dbgi = get_irn_dbg_info(n);
2170 ir_graph *irg = get_irn_irg(n);
2172 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2173 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2178 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2179 ir_graph *irg = current_ir_graph;
2180 dbg_info *dbg = get_irn_dbg_info(n);
2181 ir_node *block = get_nodes_block(n);
2182 ir_node *left = get_Minus_op(a);
2183 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2185 n = new_rd_Minus(dbg, irg, block, add, mode);
2186 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2188 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2189 ir_graph *irg = current_ir_graph;
2190 dbg_info *dbg = get_irn_dbg_info(n);
2191 ir_node *block = get_nodes_block(n);
2192 ir_node *right = get_Minus_op(b);
2194 n = new_rd_Add(dbg, irg, block, a, right, mode);
2195 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2197 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2198 ir_graph *irg = current_ir_graph;
2199 dbg_info *s_dbg = get_irn_dbg_info(b);
2200 ir_node *s_block = get_nodes_block(b);
2201 ir_node *s_left = get_Sub_right(b);
2202 ir_node *s_right = get_Sub_left(b);
2203 ir_mode *s_mode = get_irn_mode(b);
2204 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2205 dbg_info *a_dbg = get_irn_dbg_info(n);
2206 ir_node *a_block = get_nodes_block(n);
2208 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2209 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2211 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2212 ir_node *m_right = get_Mul_right(b);
2213 if (is_Const(m_right)) {
2214 ir_node *cnst2 = const_negate(m_right);
2215 if (cnst2 != NULL) {
2216 ir_graph *irg = current_ir_graph;
2217 dbg_info *m_dbg = get_irn_dbg_info(b);
2218 ir_node *m_block = get_nodes_block(b);
2219 ir_node *m_left = get_Mul_left(b);
2220 ir_mode *m_mode = get_irn_mode(b);
2221 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2222 dbg_info *a_dbg = get_irn_dbg_info(n);
2223 ir_node *a_block = get_nodes_block(n);
2225 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2226 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2232 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2233 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2235 get_irn_dbg_info(n),
2240 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2244 if (mode_wrap_around(mode)) {
2245 ir_node *left = get_Add_left(a);
2246 ir_node *right = get_Add_right(a);
2248 /* FIXME: Does the Conv's work only for two complement or generally? */
2250 if (mode != get_irn_mode(right)) {
2251 /* This Sub is an effective Cast */
2252 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2255 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2257 } else if (right == b) {
2258 if (mode != get_irn_mode(left)) {
2259 /* This Sub is an effective Cast */
2260 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2263 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2269 if (mode_wrap_around(mode)) {
2270 ir_node *left = get_Add_left(b);
2271 ir_node *right = get_Add_right(b);
2273 /* FIXME: Does the Conv's work only for two complement or generally? */
2275 ir_mode *r_mode = get_irn_mode(right);
2277 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2278 if (mode != r_mode) {
2279 /* This Sub is an effective Cast */
2280 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2282 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2284 } else if (right == a) {
2285 ir_mode *l_mode = get_irn_mode(left);
2287 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2288 if (mode != l_mode) {
2289 /* This Sub is an effective Cast */
2290 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2292 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2297 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2298 ir_mode *mode = get_irn_mode(a);
2300 if (mode == get_irn_mode(b)) {
2302 ir_node *op_a = get_Conv_op(a);
2303 ir_node *op_b = get_Conv_op(b);
2305 /* check if it's allowed to skip the conv */
2306 ma = get_irn_mode(op_a);
2307 mb = get_irn_mode(op_b);
2309 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2310 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2313 set_Sub_right(n, b);
2319 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2320 if (!is_reassoc_running() && is_Mul(a)) {
2321 ir_node *ma = get_Mul_left(a);
2322 ir_node *mb = get_Mul_right(a);
2325 ir_node *blk = get_irn_n(n, -1);
2327 get_irn_dbg_info(n),
2328 current_ir_graph, blk,
2331 get_irn_dbg_info(n),
2332 current_ir_graph, blk,
2334 new_r_Const_long(current_ir_graph, blk, mode, 1),
2337 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2339 } else if (mb == b) {
2340 ir_node *blk = get_irn_n(n, -1);
2342 get_irn_dbg_info(n),
2343 current_ir_graph, blk,
2346 get_irn_dbg_info(n),
2347 current_ir_graph, blk,
2349 new_r_Const_long(current_ir_graph, blk, mode, 1),
2352 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2356 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2357 ir_node *x = get_Sub_left(a);
2358 ir_node *y = get_Sub_right(a);
2359 ir_node *blk = get_irn_n(n, -1);
2360 ir_mode *m_b = get_irn_mode(b);
2361 ir_mode *m_y = get_irn_mode(y);
2365 /* Determine the right mode for the Add. */
2368 else if (mode_is_reference(m_b))
2370 else if (mode_is_reference(m_y))
2374 * Both modes are different but none is reference,
2375 * happens for instance in SubP(SubP(P, Iu), Is).
2376 * We have two possibilities here: Cast or ignore.
2377 * Currently we ignore this case.
2382 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2384 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2385 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2389 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2390 if (is_Const(a) && is_Not(b)) {
2391 /* c - ~X = X + (c+1) */
2392 tarval *tv = get_Const_tarval(a);
2394 tv = tarval_add(tv, get_mode_one(mode));
2395 if (tv != tarval_bad) {
2396 ir_node *blk = get_irn_n(n, -1);
2397 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2398 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2405 } /* transform_node_Sub */
2408 * Several transformation done on n*n=2n bits mul.
2409 * These transformations must be done here because new nodes may be produced.
2411 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2413 ir_node *a = get_Mul_left(n);
2414 ir_node *b = get_Mul_right(n);
2415 tarval *ta = value_of(a);
2416 tarval *tb = value_of(b);
2417 ir_mode *smode = get_irn_mode(a);
2419 if (ta == get_mode_one(smode)) {
2420 /* (L)1 * (L)b = (L)b */
2421 ir_node *blk = get_irn_n(n, -1);
2422 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2423 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2426 else if (ta == get_mode_minus_one(smode)) {
2427 /* (L)-1 * (L)b = (L)b */
2428 ir_node *blk = get_irn_n(n, -1);
2429 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2430 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2431 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2434 if (tb == get_mode_one(smode)) {
2435 /* (L)a * (L)1 = (L)a */
2436 ir_node *blk = get_irn_n(a, -1);
2437 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2438 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2441 else if (tb == get_mode_minus_one(smode)) {
2442 /* (L)a * (L)-1 = (L)-a */
2443 ir_node *blk = get_irn_n(n, -1);
2444 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2445 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2446 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2453 * Transform Mul(a,-1) into -a.
2454 * Do constant evaluation of Phi nodes.
2455 * Do architecture dependent optimizations on Mul nodes
2457 static ir_node *transform_node_Mul(ir_node *n) {
2458 ir_node *c, *oldn = n;
2459 ir_mode *mode = get_irn_mode(n);
2460 ir_node *a = get_Mul_left(n);
2461 ir_node *b = get_Mul_right(n);
2463 if (is_Bad(a) || is_Bad(b))
2466 if (mode != get_irn_mode(a))
2467 return transform_node_Mul2n(n, mode);
2469 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2471 if (mode_is_signed(mode)) {
2474 if (value_of(a) == get_mode_minus_one(mode))
2476 else if (value_of(b) == get_mode_minus_one(mode))
2479 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2480 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2485 if (is_Const(b)) { /* (-a) * const -> a * -const */
2486 ir_node *cnst = const_negate(b);
2488 dbg_info *dbgi = get_irn_dbg_info(n);
2489 ir_node *block = get_nodes_block(n);
2490 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2491 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2494 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2495 dbg_info *dbgi = get_irn_dbg_info(n);
2496 ir_node *block = get_nodes_block(n);
2497 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2498 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2500 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2501 ir_node *sub_l = get_Sub_left(b);
2502 ir_node *sub_r = get_Sub_right(b);
2503 dbg_info *dbgi = get_irn_dbg_info(n);
2504 ir_graph *irg = current_ir_graph;
2505 ir_node *block = get_nodes_block(n);
2506 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2507 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2508 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2511 } else if (is_Minus(b)) {
2512 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2513 ir_node *sub_l = get_Sub_left(a);
2514 ir_node *sub_r = get_Sub_right(a);
2515 dbg_info *dbgi = get_irn_dbg_info(n);
2516 ir_graph *irg = current_ir_graph;
2517 ir_node *block = get_nodes_block(n);
2518 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2519 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2520 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2524 if (get_mode_arithmetic(mode) == irma_ieee754) {
2526 tarval *tv = get_Const_tarval(a);
2527 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2528 /* 2.0 * b = b + b */
2529 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2530 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2534 else if (is_Const(b)) {
2535 tarval *tv = get_Const_tarval(b);
2536 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2537 /* a * 2.0 = a + a */
2538 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2539 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2544 return arch_dep_replace_mul_with_shifts(n);
2545 } /* transform_node_Mul */
2548 * Transform a Div Node.
2550 static ir_node *transform_node_Div(ir_node *n) {
2551 ir_mode *mode = get_Div_resmode(n);
2552 ir_node *a = get_Div_left(n);
2553 ir_node *b = get_Div_right(n);
2557 if (is_Const(b) && is_const_Phi(a)) {
2558 /* check for Div(Phi, Const) */
2559 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2561 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2565 else if (is_Const(a) && is_const_Phi(b)) {
2566 /* check for Div(Const, Phi) */
2567 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2569 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2573 else if (is_const_Phi(a) && is_const_Phi(b)) {
2574 /* check for Div(Phi, Phi) */
2575 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2577 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2584 if (tv != tarval_bad) {
2585 value = new_Const(get_tarval_mode(tv), tv);
2587 DBG_OPT_CSTEVAL(n, value);
2590 ir_node *a = get_Div_left(n);
2591 ir_node *b = get_Div_right(n);
2594 if (a == b && value_not_zero(a, &dummy)) {
2595 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2596 value = new_Const(mode, get_mode_one(mode));
2597 DBG_OPT_CSTEVAL(n, value);
2600 if (mode_is_signed(mode) && is_Const(b)) {
2601 tarval *tv = get_Const_tarval(b);
2603 if (tv == get_mode_minus_one(mode)) {
2605 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2606 DBG_OPT_CSTEVAL(n, value);
2610 /* Try architecture dependent optimization */
2611 value = arch_dep_replace_div_by_const(n);
2619 /* Turn Div into a tuple (mem, jmp, bad, value) */
2620 mem = get_Div_mem(n);
2621 blk = get_irn_n(n, -1);
2623 /* skip a potential Pin */
2625 mem = get_Pin_op(mem);
2626 turn_into_tuple(n, pn_Div_max);
2627 set_Tuple_pred(n, pn_Div_M, mem);
2628 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2629 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2630 set_Tuple_pred(n, pn_Div_res, value);
2633 } /* transform_node_Div */
2636 * Transform a Mod node.
2638 static ir_node *transform_node_Mod(ir_node *n) {
2639 ir_mode *mode = get_Mod_resmode(n);
2640 ir_node *a = get_Mod_left(n);
2641 ir_node *b = get_Mod_right(n);
2645 if (is_Const(b) && is_const_Phi(a)) {
2646 /* check for Div(Phi, Const) */
2647 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2649 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2653 else if (is_Const(a) && is_const_Phi(b)) {
2654 /* check for Div(Const, Phi) */
2655 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2657 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2661 else if (is_const_Phi(a) && is_const_Phi(b)) {
2662 /* check for Div(Phi, Phi) */
2663 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2665 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2672 if (tv != tarval_bad) {
2673 value = new_Const(get_tarval_mode(tv), tv);
2675 DBG_OPT_CSTEVAL(n, value);
2678 ir_node *a = get_Mod_left(n);
2679 ir_node *b = get_Mod_right(n);
2682 if (a == b && value_not_zero(a, &dummy)) {
2683 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2684 value = new_Const(mode, get_mode_null(mode));
2685 DBG_OPT_CSTEVAL(n, value);
2688 if (mode_is_signed(mode) && is_Const(b)) {
2689 tarval *tv = get_Const_tarval(b);
2691 if (tv == get_mode_minus_one(mode)) {
2693 value = new_Const(mode, get_mode_null(mode));
2694 DBG_OPT_CSTEVAL(n, value);
2698 /* Try architecture dependent optimization */
2699 value = arch_dep_replace_mod_by_const(n);
2707 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2708 mem = get_Mod_mem(n);
2709 blk = get_irn_n(n, -1);
2711 /* skip a potential Pin */
2713 mem = get_Pin_op(mem);
2714 turn_into_tuple(n, pn_Mod_max);
2715 set_Tuple_pred(n, pn_Mod_M, mem);
2716 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2717 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2718 set_Tuple_pred(n, pn_Mod_res, value);
2721 } /* transform_node_Mod */
2724 * Transform a DivMod node.
2726 static ir_node *transform_node_DivMod(ir_node *n) {
2728 ir_node *a = get_DivMod_left(n);
2729 ir_node *b = get_DivMod_right(n);
2730 ir_mode *mode = get_DivMod_resmode(n);
2735 if (is_Const(b) && is_const_Phi(a)) {
2736 /* check for Div(Phi, Const) */
2737 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2738 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2740 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2741 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2745 else if (is_Const(a) && is_const_Phi(b)) {
2746 /* check for Div(Const, Phi) */
2747 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2748 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2750 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2751 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2755 else if (is_const_Phi(a) && is_const_Phi(b)) {
2756 /* check for Div(Phi, Phi) */
2757 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2758 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2760 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2761 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2768 if (tb != tarval_bad) {
2769 if (tb == get_mode_one(get_tarval_mode(tb))) {
2771 vb = new_Const(mode, get_mode_null(mode));
2772 DBG_OPT_CSTEVAL(n, vb);
2774 } else if (ta != tarval_bad) {
2775 tarval *resa, *resb;
2776 resa = tarval_div(ta, tb);
2777 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2778 Jmp for X result!? */
2779 resb = tarval_mod(ta, tb);
2780 if (resb == tarval_bad) return n; /* Causes exception! */
2781 va = new_Const(mode, resa);
2782 vb = new_Const(mode, resb);
2783 DBG_OPT_CSTEVAL(n, va);
2784 DBG_OPT_CSTEVAL(n, vb);
2786 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2787 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2788 vb = new_Const(mode, get_mode_null(mode));
2789 DBG_OPT_CSTEVAL(n, va);
2790 DBG_OPT_CSTEVAL(n, vb);
2792 } else { /* Try architecture dependent optimization */
2795 arch_dep_replace_divmod_by_const(&va, &vb, n);
2796 evaluated = va != NULL;
2798 } else if (a == b) {
2799 if (value_not_zero(a, &dummy)) {
2801 va = new_Const(mode, get_mode_one(mode));
2802 vb = new_Const(mode, get_mode_null(mode));
2803 DBG_OPT_CSTEVAL(n, va);
2804 DBG_OPT_CSTEVAL(n, vb);
2807 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2810 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2811 /* 0 / non-Const = 0 */
2816 if (evaluated) { /* replace by tuple */
2820 mem = get_DivMod_mem(n);
2821 /* skip a potential Pin */
2823 mem = get_Pin_op(mem);
2825 blk = get_irn_n(n, -1);
2826 turn_into_tuple(n, pn_DivMod_max);
2827 set_Tuple_pred(n, pn_DivMod_M, mem);
2828 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2829 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2830 set_Tuple_pred(n, pn_DivMod_res_div, va);
2831 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2835 } /* transform_node_DivMod */
2838 * Optimize x / c to x * (1/c)
2840 static ir_node *transform_node_Quot(ir_node *n) {
2841 ir_mode *mode = get_Quot_resmode(n);
2844 if (get_mode_arithmetic(mode) == irma_ieee754) {
2845 ir_node *b = get_Quot_right(n);
2848 tarval *tv = get_Const_tarval(b);
2852 * Floating point constant folding might be disabled here to
2854 * However, as we check for exact result, doing it is safe.
2857 rem = tarval_enable_fp_ops(1);
2858 tv = tarval_quo(get_mode_one(mode), tv);
2859 (void)tarval_enable_fp_ops(rem);
2861 /* Do the transformation if the result is either exact or we are not
2862 using strict rules. */
2863 if (tv != tarval_bad &&
2864 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2865 ir_node *blk = get_irn_n(n, -1);
2866 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2867 ir_node *a = get_Quot_left(n);
2868 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2869 ir_node *mem = get_Quot_mem(n);
2871 /* skip a potential Pin */
2873 mem = get_Pin_op(mem);
2874 turn_into_tuple(n, pn_Quot_max);
2875 set_Tuple_pred(n, pn_Quot_M, mem);
2876 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2877 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2878 set_Tuple_pred(n, pn_Quot_res, m);
2879 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2884 } /* transform_node_Quot */
2887 * Optimize Abs(x) into x if x is Confirmed >= 0
2888 * Optimize Abs(x) into -x if x is Confirmed <= 0
2889 * Optimize Abs(-x) int Abs(x)
2891 static ir_node *transform_node_Abs(ir_node *n) {
2892 ir_node *c, *oldn = n;
2893 ir_node *a = get_Abs_op(n);
2896 HANDLE_UNOP_PHI(tarval_abs, a, c);
2898 switch (classify_value_sign(a)) {
2899 case value_classified_negative:
2900 mode = get_irn_mode(n);
2903 * We can replace the Abs by -x here.
2904 * We even could add a new Confirm here
2905 * (if not twos complement)
2907 * Note that -x would create a new node, so we could
2908 * not run it in the equivalent_node() context.
2910 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2911 get_nodes_block(n), a, mode);
2913 DBG_OPT_CONFIRM(oldn, n);
2915 case value_classified_positive:
2916 /* n is positive, Abs is not needed */
2919 DBG_OPT_CONFIRM(oldn, n);
2925 /* Abs(-x) = Abs(x) */
2926 mode = get_irn_mode(n);
2927 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2928 get_nodes_block(n), get_Minus_op(a), mode);
2929 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2933 } /* transform_node_Abs */
2936 * Transform a Cond node.
2938 * Replace the Cond by a Jmp if it branches on a constant
2941 static ir_node *transform_node_Cond(ir_node *n) {
2944 ir_node *a = get_Cond_selector(n);
2945 tarval *ta = value_of(a);
2947 /* we need block info which is not available in floating irgs */
2948 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2951 if ((ta != tarval_bad) &&
2952 (get_irn_mode(a) == mode_b) &&
2953 (get_opt_unreachable_code())) {
2954 /* It's a boolean Cond, branching on a boolean constant.
2955 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2956 ir_node *blk = get_nodes_block(n);
2957 jmp = new_r_Jmp(current_ir_graph, blk);
2958 turn_into_tuple(n, pn_Cond_max);
2959 if (ta == tarval_b_true) {
2960 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2961 set_Tuple_pred(n, pn_Cond_true, jmp);
2963 set_Tuple_pred(n, pn_Cond_false, jmp);
2964 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2966 /* We might generate an endless loop, so keep it alive. */
2967 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2970 } /* transform_node_Cond */
2973 * Prototype of a recursive transform function
2974 * for bitwise distributive transformations.
2976 typedef ir_node* (*recursive_transform)(ir_node *n);
2979 * makes use of distributive laws for and, or, eor
2980 * and(a OP c, b OP c) -> and(a, b) OP c
2981 * note, might return a different op than n
2983 static ir_node *transform_bitwise_distributive(ir_node *n,
2984 recursive_transform trans_func)
2987 ir_node *a = get_binop_left(n);
2988 ir_node *b = get_binop_right(n);
2989 ir_op *op = get_irn_op(a);
2990 ir_op *op_root = get_irn_op(n);
2992 if(op != get_irn_op(b))
2995 if (op == op_Conv) {
2996 ir_node *a_op = get_Conv_op(a);
2997 ir_node *b_op = get_Conv_op(b);
2998 ir_mode *a_mode = get_irn_mode(a_op);
2999 ir_mode *b_mode = get_irn_mode(b_op);
3000 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3001 ir_node *blk = get_irn_n(n, -1);
3004 set_binop_left(n, a_op);
3005 set_binop_right(n, b_op);
3006 set_irn_mode(n, a_mode);
3008 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3010 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3016 /* nothing to gain here */
3020 if (op == op_Shrs || op == op_Shr || op == op_Shl
3021 || op == op_And || op == op_Or || op == op_Eor) {
3022 ir_node *a_left = get_binop_left(a);
3023 ir_node *a_right = get_binop_right(a);
3024 ir_node *b_left = get_binop_left(b);
3025 ir_node *b_right = get_binop_right(b);
3027 ir_node *op1 = NULL;
3028 ir_node *op2 = NULL;
3030 if (is_op_commutative(op)) {
3031 if (a_left == b_left) {
3035 } else if(a_left == b_right) {
3039 } else if(a_right == b_left) {
3045 if(a_right == b_right) {
3052 /* (a sop c) & (b sop c) => (a & b) sop c */
3053 ir_node *blk = get_irn_n(n, -1);
3055 ir_node *new_n = exact_copy(n);
3056 set_binop_left(new_n, op1);
3057 set_binop_right(new_n, op2);
3058 new_n = trans_func(new_n);
3060 if(op_root == op_Eor && op == op_Or) {
3061 dbg_info *dbgi = get_irn_dbg_info(n);
3062 ir_graph *irg = current_ir_graph;
3063 ir_mode *mode = get_irn_mode(c);
3065 c = new_rd_Not(dbgi, irg, blk, c, mode);
3066 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3069 set_nodes_block(n, blk);
3070 set_binop_left(n, new_n);
3071 set_binop_right(n, c);
3072 add_identities(current_ir_graph->value_table, n);
3075 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3086 static ir_node *transform_node_And(ir_node *n) {
3087 ir_node *c, *oldn = n;
3088 ir_node *a = get_And_left(n);
3089 ir_node *b = get_And_right(n);
3092 mode = get_irn_mode(n);
3093 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3095 /* we can evaluate 2 Projs of the same Cmp */
3096 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3097 ir_node *pred_a = get_Proj_pred(a);
3098 ir_node *pred_b = get_Proj_pred(b);
3099 if (pred_a == pred_b) {
3100 dbg_info *dbgi = get_irn_dbg_info(n);
3101 ir_node *block = get_nodes_block(pred_a);
3102 pn_Cmp pn_a = get_Proj_proj(a);
3103 pn_Cmp pn_b = get_Proj_proj(b);
3104 /* yes, we can simply calculate with pncs */
3105 pn_Cmp new_pnc = pn_a & pn_b;
3107 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3112 ir_node *op = get_Not_op(b);
3114 ir_node *ba = get_And_left(op);
3115 ir_node *bb = get_And_right(op);
3117 /* it's enough to test the following cases due to normalization! */
3118 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3119 /* (a|b) & ~(a&b) = a^b */
3120 ir_node *block = get_nodes_block(n);
3122 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3123 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3131 ir_node *op = get_Not_op(a);
3133 ir_node *aa = get_And_left(op);
3134 ir_node *ab = get_And_right(op);
3136 /* it's enough to test the following cases due to normalization! */
3137 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3138 /* (a|b) & ~(a&b) = a^b */
3139 ir_node *block = get_nodes_block(n);
3141 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3142 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3149 ir_node *al = get_Eor_left(a);
3150 ir_node *ar = get_Eor_right(a);
3153 /* (b ^ a) & b -> ~a & b */
3154 dbg_info *dbg = get_irn_dbg_info(n);
3155 ir_node *block = get_nodes_block(n);
3157 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3158 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3159 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3163 /* (a ^ b) & b -> ~a & b */
3164 dbg_info *dbg = get_irn_dbg_info(n);
3165 ir_node *block = get_nodes_block(n);
3167 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3168 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3169 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3174 ir_node *bl = get_Eor_left(b);
3175 ir_node *br = get_Eor_right(b);
3178 /* a & (a ^ b) -> a & ~b */
3179 dbg_info *dbg = get_irn_dbg_info(n);
3180 ir_node *block = get_nodes_block(n);
3182 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3183 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3184 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3188 /* a & (b ^ a) -> a & ~b */
3189 dbg_info *dbg = get_irn_dbg_info(n);
3190 ir_node *block = get_nodes_block(n);
3192 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3193 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3194 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3198 if (is_Not(a) && is_Not(b)) {
3199 /* ~a & ~b = ~(a|b) */
3200 ir_node *block = get_nodes_block(n);
3201 ir_mode *mode = get_irn_mode(n);
3205 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3206 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3207 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3211 n = transform_bitwise_distributive(n, transform_node_And);
3214 } /* transform_node_And */
3219 static ir_node *transform_node_Eor(ir_node *n) {
3220 ir_node *c, *oldn = n;
3221 ir_node *a = get_Eor_left(n);
3222 ir_node *b = get_Eor_right(n);
3223 ir_mode *mode = get_irn_mode(n);
3225 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3227 /* we can evaluate 2 Projs of the same Cmp */
3228 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3229 ir_node *pred_a = get_Proj_pred(a);
3230 ir_node *pred_b = get_Proj_pred(b);
3231 if(pred_a == pred_b) {
3232 dbg_info *dbgi = get_irn_dbg_info(n);
3233 ir_node *block = get_nodes_block(pred_a);
3234 pn_Cmp pn_a = get_Proj_proj(a);
3235 pn_Cmp pn_b = get_Proj_proj(b);
3236 /* yes, we can simply calculate with pncs */
3237 pn_Cmp new_pnc = pn_a ^ pn_b;
3239 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3246 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3247 mode, get_mode_null(mode));
3248 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3249 } else if (mode == mode_b &&
3251 is_Const(b) && is_Const_one(b) &&
3252 is_Cmp(get_Proj_pred(a))) {
3253 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3254 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3255 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3257 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3258 } else if (is_Const(b)) {
3259 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3260 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3261 ir_node *not_op = get_Not_op(a);
3262 dbg_info *dbg = get_irn_dbg_info(n);
3263 ir_graph *irg = current_ir_graph;
3264 ir_node *block = get_nodes_block(n);
3265 ir_mode *mode = get_irn_mode(n);
3266 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3268 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3269 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3270 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3273 n = transform_bitwise_distributive(n, transform_node_Eor);
3277 } /* transform_node_Eor */
3282 static ir_node *transform_node_Not(ir_node *n) {
3283 ir_node *c, *oldn = n;
3284 ir_node *a = get_Not_op(n);
3285 ir_mode *mode = get_irn_mode(n);
3287 HANDLE_UNOP_PHI(tarval_not,a,c);
3289 /* check for a boolean Not */
3290 if (mode == mode_b &&
3292 is_Cmp(get_Proj_pred(a))) {
3293 /* We negate a Cmp. The Cmp has the negated result anyways! */
3294 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3295 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3296 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3300 ir_node *eor_b = get_Eor_right(a);
3301 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3302 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3303 ir_node *eor_a = get_Eor_left(a);
3304 dbg_info *dbg = get_irn_dbg_info(n);
3305 ir_graph *irg = current_ir_graph;
3306 ir_node *block = get_nodes_block(n);
3307 ir_mode *mode = get_irn_mode(n);
3308 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3312 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3313 if (is_Minus(a)) { /* ~-x -> x + -1 */
3314 dbg_info *dbg = get_irn_dbg_info(n);
3315 ir_graph *irg = current_ir_graph;
3316 ir_node *block = get_nodes_block(n);
3317 ir_node *add_l = get_Minus_op(a);
3318 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3319 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3320 } else if (is_Add(a)) {
3321 ir_node *add_r = get_Add_right(a);
3322 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3323 /* ~(x + -1) = -x */
3324 ir_node *op = get_Add_left(a);
3325 ir_node *blk = get_irn_n(n, -1);
3326 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3327 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3332 } /* transform_node_Not */
3335 * Transform a Minus.
3339 * -(a >>u (size-1)) = a >>s (size-1)
3340 * -(a >>s (size-1)) = a >>u (size-1)
3341 * -(a * const) -> a * -const
3343 static ir_node *transform_node_Minus(ir_node *n) {
3344 ir_node *c, *oldn = n;
3345 ir_node *a = get_Minus_op(n);
3348 HANDLE_UNOP_PHI(tarval_neg,a,c);
3350 mode = get_irn_mode(a);
3351 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3352 /* the following rules are only to twos-complement */
3355 ir_node *op = get_Not_op(a);
3356 tarval *tv = get_mode_one(mode);
3357 ir_node *blk = get_irn_n(n, -1);
3358 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3359 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3360 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3364 ir_node *c = get_Shr_right(a);
3367 tarval *tv = get_Const_tarval(c);
3369 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3370 /* -(a >>u (size-1)) = a >>s (size-1) */
3371 ir_node *v = get_Shr_left(a);
3373 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3374 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3380 ir_node *c = get_Shrs_right(a);
3383 tarval *tv = get_Const_tarval(c);
3385 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3386 /* -(a >>s (size-1)) = a >>u (size-1) */
3387 ir_node *v = get_Shrs_left(a);
3389 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3390 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3397 /* - (a-b) = b - a */
3398 ir_node *la = get_Sub_left(a);
3399 ir_node *ra = get_Sub_right(a);
3400 ir_node *blk = get_irn_n(n, -1);
3402 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3403 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3407 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3408 ir_node *mul_l = get_Mul_left(a);
3409 ir_node *mul_r = get_Mul_right(a);
3410 if (is_Const(mul_r)) {
3411 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3412 if(tv != tarval_bad) {
3413 ir_node *cnst = new_Const(mode, tv);
3414 dbg_info *dbg = get_irn_dbg_info(a);
3415 ir_graph *irg = current_ir_graph;
3416 ir_node *block = get_nodes_block(a);
3417 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3418 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3425 } /* transform_node_Minus */
3428 * Transform a Cast_type(Const) into a new Const_type
3430 static ir_node *transform_node_Cast(ir_node *n) {
3432 ir_node *pred = get_Cast_op(n);
3433 ir_type *tp = get_irn_type(n);
3435 if (is_Const(pred) && get_Const_type(pred) != tp) {
3436 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3437 get_Const_tarval(pred), tp);
3438 DBG_OPT_CSTEVAL(oldn, n);
3439 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3440 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3441 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3442 DBG_OPT_CSTEVAL(oldn, n);
3446 } /* transform_node_Cast */
3449 * Transform a Proj(Div) with a non-zero value.
3450 * Removes the exceptions and routes the memory to the NoMem node.
3452 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3453 ir_node *div = get_Proj_pred(proj);
3454 ir_node *b = get_Div_right(div);
3455 ir_node *confirm, *res, *new_mem;
3458 if (value_not_zero(b, &confirm)) {
3459 /* div(x, y) && y != 0 */
3460 if (confirm == NULL) {
3461 /* we are sure we have a Const != 0 */
3462 new_mem = get_Div_mem(div);
3463 if (is_Pin(new_mem))
3464 new_mem = get_Pin_op(new_mem);
3465 set_Div_mem(div, new_mem);
3466 set_irn_pinned(div, op_pin_state_floats);
3469 proj_nr = get_Proj_proj(proj);
3471 case pn_Div_X_regular:
3472 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3474 case pn_Div_X_except:
3475 /* we found an exception handler, remove it */
3476 DBG_OPT_EXC_REM(proj);
3480 res = get_Div_mem(div);
3481 new_mem = get_irg_no_mem(current_ir_graph);
3484 /* This node can only float up to the Confirm block */
3485 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3487 set_irn_pinned(div, op_pin_state_floats);
3488 /* this is a Div without exception, we can remove the memory edge */
3489 set_Div_mem(div, new_mem);
3494 } /* transform_node_Proj_Div */
3497 * Transform a Proj(Mod) with a non-zero value.
3498 * Removes the exceptions and routes the memory to the NoMem node.
3500 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3501 ir_node *mod = get_Proj_pred(proj);
3502 ir_node *b = get_Mod_right(mod);
3503 ir_node *confirm, *res, *new_mem;
3506 if (value_not_zero(b, &confirm)) {
3507 /* mod(x, y) && y != 0 */
3508 proj_nr = get_Proj_proj(proj);
3510 if (confirm == NULL) {
3511 /* we are sure we have a Const != 0 */
3512 new_mem = get_Mod_mem(mod);
3513 if (is_Pin(new_mem))
3514 new_mem = get_Pin_op(new_mem);
3515 set_Mod_mem(mod, new_mem);
3516 set_irn_pinned(mod, op_pin_state_floats);
3521 case pn_Mod_X_regular:
3522 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3524 case pn_Mod_X_except:
3525 /* we found an exception handler, remove it */
3526 DBG_OPT_EXC_REM(proj);
3530 res = get_Mod_mem(mod);
3531 new_mem = get_irg_no_mem(current_ir_graph);
3534 /* This node can only float up to the Confirm block */
3535 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3537 /* this is a Mod without exception, we can remove the memory edge */
3538 set_Mod_mem(mod, new_mem);
3541 if (get_Mod_left(mod) == b) {
3542 /* a % a = 0 if a != 0 */
3543 ir_mode *mode = get_irn_mode(proj);
3544 ir_node *res = new_Const(mode, get_mode_null(mode));
3546 DBG_OPT_CSTEVAL(mod, res);
3552 } /* transform_node_Proj_Mod */
3555 * Transform a Proj(DivMod) with a non-zero value.
3556 * Removes the exceptions and routes the memory to the NoMem node.
3558 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3559 ir_node *divmod = get_Proj_pred(proj);
3560 ir_node *b = get_DivMod_right(divmod);
3561 ir_node *confirm, *res, *new_mem;
3564 if (value_not_zero(b, &confirm)) {
3565 /* DivMod(x, y) && y != 0 */
3566 proj_nr = get_Proj_proj(proj);
3568 if (confirm == NULL) {
3569 /* we are sure we have a Const != 0 */
3570 new_mem = get_DivMod_mem(divmod);
3571 if (is_Pin(new_mem))
3572 new_mem = get_Pin_op(new_mem);
3573 set_DivMod_mem(divmod, new_mem);
3574 set_irn_pinned(divmod, op_pin_state_floats);
3579 case pn_DivMod_X_regular:
3580 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3582 case pn_DivMod_X_except:
3583 /* we found an exception handler, remove it */
3584 DBG_OPT_EXC_REM(proj);
3588 res = get_DivMod_mem(divmod);
3589 new_mem = get_irg_no_mem(current_ir_graph);
3592 /* This node can only float up to the Confirm block */
3593 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3595 /* this is a DivMod without exception, we can remove the memory edge */
3596 set_DivMod_mem(divmod, new_mem);
3599 case pn_DivMod_res_mod:
3600 if (get_DivMod_left(divmod) == b) {
3601 /* a % a = 0 if a != 0 */
3602 ir_mode *mode = get_irn_mode(proj);
3603 ir_node *res = new_Const(mode, get_mode_null(mode));
3605 DBG_OPT_CSTEVAL(divmod, res);
3611 } /* transform_node_Proj_DivMod */
3614 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3616 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3617 if (get_opt_unreachable_code()) {
3618 ir_node *n = get_Proj_pred(proj);
3619 ir_node *b = get_Cond_selector(n);
3621 if (mode_is_int(get_irn_mode(b))) {
3622 tarval *tb = value_of(b);
3624 if (tb != tarval_bad) {
3625 /* we have a constant switch */
3626 long num = get_Proj_proj(proj);
3628 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3629 if (get_tarval_long(tb) == num) {
3630 /* Do NOT create a jump here, or we will have 2 control flow ops
3631 * in a block. This case is optimized away in optimize_cf(). */
3634 /* this case will NEVER be taken, kill it */
3642 } /* transform_node_Proj_Cond */
3645 * Create a 0 constant of given mode.
3647 static ir_node *create_zero_const(ir_mode *mode) {
3648 tarval *tv = get_mode_null(mode);
3649 ir_node *cnst = new_Const(mode, tv);
3654 /* the order of the values is important! */
3655 typedef enum const_class {
3661 static const_class classify_const(const ir_node* n)
3663 if (is_Const(n)) return const_const;
3664 if (is_irn_constlike(n)) return const_like;
3669 * Determines whether r is more constlike or has a larger index (in that order)
3672 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3674 const const_class l_order = classify_const(l);
3675 const const_class r_order = classify_const(r);
3677 l_order > r_order ||
3678 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3682 * Normalizes and optimizes Cmp nodes.
3684 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3685 ir_node *n = get_Proj_pred(proj);
3686 ir_node *left = get_Cmp_left(n);
3687 ir_node *right = get_Cmp_right(n);
3690 ir_mode *mode = NULL;
3691 long proj_nr = get_Proj_proj(proj);
3693 /* we can evaluate some cases directly */
3696 return new_Const(mode_b, get_tarval_b_false());
3698 return new_Const(mode_b, get_tarval_b_true());
3700 if (!mode_is_float(get_irn_mode(left)))
3701 return new_Const(mode_b, get_tarval_b_true());
3707 /* remove Casts of both sides */
3708 left = skip_Cast(left);
3709 right = skip_Cast(right);
3711 /* Remove unnecessary conversions */
3712 /* TODO handle constants */
3713 if (is_Conv(left) && is_Conv(right)) {
3714 ir_mode *mode = get_irn_mode(left);
3715 ir_node *op_left = get_Conv_op(left);
3716 ir_node *op_right = get_Conv_op(right);
3717 ir_mode *mode_left = get_irn_mode(op_left);
3718 ir_mode *mode_right = get_irn_mode(op_right);
3720 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3721 && mode_left != mode_b && mode_right != mode_b) {
3722 ir_graph *irg = current_ir_graph;
3723 ir_node *block = get_nodes_block(n);
3725 if (mode_left == mode_right) {
3729 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3730 } else if (smaller_mode(mode_left, mode_right)) {
3731 left = new_r_Conv(irg, block, op_left, mode_right);
3734 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3735 } else if (smaller_mode(mode_right, mode_left)) {
3737 right = new_r_Conv(irg, block, op_right, mode_left);
3739 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3744 /* remove operation on both sides if possible */
3745 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3747 * The following operations are NOT safe for floating point operations, for instance
3748 * 1.0 + inf == 2.0 + inf, =/=> x == y
3750 if (mode_is_int(get_irn_mode(left))) {
3751 unsigned lop = get_irn_opcode(left);
3753 if (lop == get_irn_opcode(right)) {
3754 ir_node *ll, *lr, *rl, *rr;
3756 /* same operation on both sides, try to remove */
3760 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3761 left = get_unop_op(left);
3762 right = get_unop_op(right);
3764 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3767 ll = get_Add_left(left);
3768 lr = get_Add_right(left);
3769 rl = get_Add_left(right);
3770 rr = get_Add_right(right);
3773 /* X + a CMP X + b ==> a CMP b */
3777 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3778 } else if (ll == rr) {
3779 /* X + a CMP b + X ==> a CMP b */
3783 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3784 } else if (lr == rl) {
3785 /* a + X CMP X + b ==> a CMP b */
3789 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3790 } else if (lr == rr) {
3791 /* a + X CMP b + X ==> a CMP b */
3795 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3799 ll = get_Sub_left(left);
3800 lr = get_Sub_right(left);
3801 rl = get_Sub_left(right);
3802 rr = get_Sub_right(right);
3805 /* X - a CMP X - b ==> a CMP b */
3809 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3810 } else if (lr == rr) {
3811 /* a - X CMP b - X ==> a CMP b */
3815 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3819 if (get_Rot_right(left) == get_Rot_right(right)) {
3820 /* a ROT X CMP b ROT X ==> a CMP b */
3821 left = get_Rot_left(left);
3822 right = get_Rot_left(right);
3824 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3832 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3833 if (is_Add(left) || is_Sub(left)) {
3834 ir_node *ll = get_binop_left(left);
3835 ir_node *lr = get_binop_right(left);
3837 if (lr == right && is_Add(left)) {
3844 right = create_zero_const(get_irn_mode(left));
3846 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3849 if (is_Add(right) || is_Sub(right)) {
3850 ir_node *rl = get_binop_left(right);
3851 ir_node *rr = get_binop_right(right);
3853 if (rr == left && is_Add(right)) {
3860 right = create_zero_const(get_irn_mode(left));
3862 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3865 } /* mode_is_int(...) */
3866 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3868 /* replace mode_b compares with ands/ors */
3869 if (get_irn_mode(left) == mode_b) {
3870 ir_graph *irg = current_ir_graph;
3871 ir_node *block = get_nodes_block(n);
3875 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3876 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3877 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3878 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3879 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3880 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3881 default: bres = NULL;
3884 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3890 * First step: normalize the compare op
3891 * by placing the constant on the right side
3892 * or moving the lower address node to the left.
3894 if (!operands_are_normalized(left, right)) {
3900 proj_nr = get_inversed_pnc(proj_nr);
3905 * Second step: Try to reduce the magnitude
3906 * of a constant. This may help to generate better code
3907 * later and may help to normalize more compares.
3908 * Of course this is only possible for integer values.
3910 if (is_Const(right)) {
3911 mode = get_irn_mode(right);
3912 tv = get_Const_tarval(right);
3914 /* TODO extend to arbitrary constants */
3915 if (is_Conv(left) && tarval_is_null(tv)) {
3916 ir_node *op = get_Conv_op(left);
3917 ir_mode *op_mode = get_irn_mode(op);
3920 * UpConv(x) REL 0 ==> x REL 0
3922 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3923 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3924 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3925 tv = get_mode_null(op_mode);
3929 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3933 if (tv != tarval_bad) {
3934 /* the following optimization is possible on modes without Overflow
3935 * on Unary Minus or on == and !=:
3936 * -a CMP c ==> a swap(CMP) -c
3938 * Beware: for two-complement Overflow may occur, so only == and != can
3939 * be optimized, see this:
3940 * -MININT < 0 =/=> MININT > 0 !!!
3942 if (is_Minus(left) &&
3943 (!mode_overflow_on_unary_Minus(mode) ||
3944 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3945 tv = tarval_neg(tv);
3947 if (tv != tarval_bad) {
3948 left = get_Minus_op(left);
3949 proj_nr = get_inversed_pnc(proj_nr);
3951 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3953 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3954 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3955 tv = tarval_not(tv);
3957 if (tv != tarval_bad) {
3958 left = get_Not_op(left);
3960 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3964 /* for integer modes, we have more */
3965 if (mode_is_int(mode)) {
3966 /* Ne includes Unordered which is not possible on integers.
3967 * However, frontends often use this wrong, so fix it here */
3968 if (proj_nr & pn_Cmp_Uo) {
3969 proj_nr &= ~pn_Cmp_Uo;
3970 set_Proj_proj(proj, proj_nr);
3973 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3974 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3975 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3976 tv = tarval_sub(tv, get_mode_one(mode));
3978 if (tv != tarval_bad) {
3979 proj_nr ^= pn_Cmp_Eq;
3981 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3984 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3985 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3986 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3987 tv = tarval_add(tv, get_mode_one(mode));
3989 if (tv != tarval_bad) {
3990 proj_nr ^= pn_Cmp_Eq;
3992 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3996 /* the following reassociations work only for == and != */
3997 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3999 #if 0 /* Might be not that good in general */
4000 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4001 if (tarval_is_null(tv) && is_Sub(left)) {
4002 right = get_Sub_right(left);
4003 left = get_Sub_left(left);
4005 tv = value_of(right);
4007 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4011 if (tv != tarval_bad) {
4012 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4014 ir_node *c1 = get_Sub_right(left);
4015 tarval *tv2 = value_of(c1);
4017 if (tv2 != tarval_bad) {
4018 tv2 = tarval_add(tv, value_of(c1));
4020 if (tv2 != tarval_bad) {
4021 left = get_Sub_left(left);
4024 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4028 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4029 else if (is_Add(left)) {
4030 ir_node *a_l = get_Add_left(left);
4031 ir_node *a_r = get_Add_right(left);
4035 if (is_Const(a_l)) {
4037 tv2 = value_of(a_l);
4040 tv2 = value_of(a_r);
4043 if (tv2 != tarval_bad) {
4044 tv2 = tarval_sub(tv, tv2);
4046 if (tv2 != tarval_bad) {
4050 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4054 /* -a == c ==> a == -c, -a != c ==> a != -c */
4055 else if (is_Minus(left)) {
4056 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4058 if (tv2 != tarval_bad) {
4059 left = get_Minus_op(left);
4062 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4067 /* the following reassociations work only for <= */
4068 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4069 if (tv != tarval_bad) {
4070 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4071 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4077 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4078 switch (get_irn_opcode(left)) {
4082 c1 = get_And_right(left);
4085 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4086 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4088 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4090 /* TODO: move to constant evaluation */
4091 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4092 c1 = new_Const(mode_b, tv);
4093 DBG_OPT_CSTEVAL(proj, c1);
4097 if (tarval_is_single_bit(tv)) {
4099 * optimization for AND:
4101 * And(x, C) == C ==> And(x, C) != 0
4102 * And(x, C) != C ==> And(X, C) == 0
4104 * if C is a single Bit constant.
4107 /* check for Constant's match. We have check hare the tarvals,
4108 because our const might be changed */
4109 if (get_Const_tarval(c1) == tv) {
4110 /* fine: do the transformation */
4111 tv = get_mode_null(get_tarval_mode(tv));
4112 proj_nr ^= pn_Cmp_Leg;
4114 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4120 c1 = get_Or_right(left);
4121 if (is_Const(c1) && tarval_is_null(tv)) {
4123 * Or(x, C) == 0 && C != 0 ==> FALSE
4124 * Or(x, C) != 0 && C != 0 ==> TRUE
4126 if (! tarval_is_null(get_Const_tarval(c1))) {
4127 /* TODO: move to constant evaluation */
4128 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4129 c1 = new_Const(mode_b, tv);
4130 DBG_OPT_CSTEVAL(proj, c1);
4137 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4139 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4142 c1 = get_Shl_right(left);
4144 tarval *tv1 = get_Const_tarval(c1);
4145 ir_mode *mode = get_irn_mode(left);
4146 tarval *minus1 = get_mode_all_one(mode);
4147 tarval *amask = tarval_shr(minus1, tv1);
4148 tarval *cmask = tarval_shl(minus1, tv1);
4151 if (tarval_and(tv, cmask) != tv) {
4152 /* condition not met */
4153 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4154 c1 = new_Const(mode_b, tv);
4155 DBG_OPT_CSTEVAL(proj, c1);
4158 sl = get_Shl_left(left);
4159 blk = get_nodes_block(n);
4160 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4161 tv = tarval_shr(tv, tv1);
4163 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4168 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4170 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4173 c1 = get_Shr_right(left);
4175 tarval *tv1 = get_Const_tarval(c1);
4176 ir_mode *mode = get_irn_mode(left);
4177 tarval *minus1 = get_mode_all_one(mode);
4178 tarval *amask = tarval_shl(minus1, tv1);
4179 tarval *cmask = tarval_shr(minus1, tv1);
4182 if (tarval_and(tv, cmask) != tv) {
4183 /* condition not met */
4184 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4185 c1 = new_Const(mode_b, tv);
4186 DBG_OPT_CSTEVAL(proj, c1);
4189 sl = get_Shr_left(left);
4190 blk = get_nodes_block(n);
4191 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4192 tv = tarval_shl(tv, tv1);
4194 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4199 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4201 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4204 c1 = get_Shrs_right(left);
4206 tarval *tv1 = get_Const_tarval(c1);
4207 ir_mode *mode = get_irn_mode(left);
4208 tarval *minus1 = get_mode_all_one(mode);
4209 tarval *amask = tarval_shl(minus1, tv1);
4210 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4213 cond = tarval_sub(cond, tv1);
4214 cond = tarval_shrs(tv, cond);
4216 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4217 /* condition not met */
4218 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4219 c1 = new_Const(mode_b, tv);
4220 DBG_OPT_CSTEVAL(proj, c1);
4223 sl = get_Shrs_left(left);
4224 blk = get_nodes_block(n);
4225 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4226 tv = tarval_shl(tv, tv1);
4228 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4233 } /* tarval != bad */
4236 if (changed & 2) /* need a new Const */
4237 right = new_Const(mode, tv);
4239 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4240 ir_node *op = get_Proj_pred(left);
4242 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4243 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4244 ir_node *c = get_binop_right(op);
4247 tarval *tv = get_Const_tarval(c);
4249 if (tarval_is_single_bit(tv)) {
4250 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4251 ir_node *v = get_binop_left(op);
4252 ir_node *blk = get_irn_n(op, -1);
4253 ir_mode *mode = get_irn_mode(v);
4255 tv = tarval_sub(tv, get_mode_one(mode));
4256 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4258 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4265 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4267 /* create a new compare */
4268 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4269 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4273 } /* transform_node_Proj_Cmp */
4276 * Does all optimizations on nodes that must be done on it's Proj's
4277 * because of creating new nodes.
4279 static ir_node *transform_node_Proj(ir_node *proj) {
4280 ir_node *n = get_Proj_pred(proj);
4282 switch (get_irn_opcode(n)) {
4284 return transform_node_Proj_Div(proj);
4287 return transform_node_Proj_Mod(proj);
4290 return transform_node_Proj_DivMod(proj);
4293 return transform_node_Proj_Cond(proj);
4296 return transform_node_Proj_Cmp(proj);
4299 /* should not happen, but if it does will be optimized away */
4300 return equivalent_node_Proj(proj);
4306 } /* transform_node_Proj */
4309 * Move Confirms down through Phi nodes.
4311 static ir_node *transform_node_Phi(ir_node *phi) {
4313 ir_mode *mode = get_irn_mode(phi);
4315 if (mode_is_reference(mode)) {
4316 n = get_irn_arity(phi);
4318 /* Beware of Phi0 */
4320 ir_node *pred = get_irn_n(phi, 0);
4321 ir_node *bound, *new_Phi, *block, **in;
4324 if (! is_Confirm(pred))
4327 bound = get_Confirm_bound(pred);
4328 pnc = get_Confirm_cmp(pred);
4330 NEW_ARR_A(ir_node *, in, n);
4331 in[0] = get_Confirm_value(pred);
4333 for (i = 1; i < n; ++i) {
4334 pred = get_irn_n(phi, i);
4336 if (! is_Confirm(pred) ||
4337 get_Confirm_bound(pred) != bound ||
4338 get_Confirm_cmp(pred) != pnc)
4340 in[i] = get_Confirm_value(pred);
4342 /* move the Confirm nodes "behind" the Phi */
4343 block = get_irn_n(phi, -1);
4344 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4345 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4349 } /* transform_node_Phi */
4352 * Returns the operands of a commutative bin-op, if one operand is
4353 * a const, it is returned as the second one.
4355 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4356 ir_node *op_a = get_binop_left(binop);
4357 ir_node *op_b = get_binop_right(binop);
4359 assert(is_op_commutative(get_irn_op(binop)));
4361 if (is_Const(op_a)) {
4368 } /* get_comm_Binop_Ops */
4371 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4372 * Such pattern may arise in bitfield stores.
4374 * value c4 value c4 & c2
4375 * AND c3 AND c1 | c3
4382 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4385 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4388 ir_node *and_l, *c3;
4389 ir_node *value, *c4;
4390 ir_node *new_and, *new_const, *block;
4391 ir_mode *mode = get_irn_mode(or);
4393 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4396 get_comm_Binop_Ops(or, &and, &c1);
4397 if (!is_Const(c1) || !is_And(and))
4400 get_comm_Binop_Ops(and, &or_l, &c2);
4404 tv1 = get_Const_tarval(c1);
4405 tv2 = get_Const_tarval(c2);
4407 tv = tarval_or(tv1, tv2);
4408 if (tarval_is_all_one(tv)) {
4409 /* the AND does NOT clear a bit with isn't set by the OR */
4410 set_Or_left(or, or_l);
4411 set_Or_right(or, c1);
4413 /* check for more */
4420 get_comm_Binop_Ops(or_l, &and_l, &c3);
4421 if (!is_Const(c3) || !is_And(and_l))
4424 get_comm_Binop_Ops(and_l, &value, &c4);
4428 /* ok, found the pattern, check for conditions */
4429 assert(mode == get_irn_mode(and));
4430 assert(mode == get_irn_mode(or_l));
4431 assert(mode == get_irn_mode(and_l));
4433 tv3 = get_Const_tarval(c3);
4434 tv4 = get_Const_tarval(c4);
4436 tv = tarval_or(tv4, tv2);
4437 if (!tarval_is_all_one(tv)) {
4438 /* have at least one 0 at the same bit position */
4442 n_tv4 = tarval_not(tv4);
4443 if (tv3 != tarval_and(tv3, n_tv4)) {
4444 /* bit in the or_mask is outside the and_mask */
4448 n_tv2 = tarval_not(tv2);
4449 if (tv1 != tarval_and(tv1, n_tv2)) {
4450 /* bit in the or_mask is outside the and_mask */
4454 /* ok, all conditions met */
4455 block = get_irn_n(or, -1);
4457 new_and = new_r_And(current_ir_graph, block,
4458 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4460 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4462 set_Or_left(or, new_and);
4463 set_Or_right(or, new_const);
4465 /* check for more */
4467 } /* transform_node_Or_bf_store */
4470 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4472 static ir_node *transform_node_Or_Rot(ir_node *or) {
4473 ir_mode *mode = get_irn_mode(or);
4474 ir_node *shl, *shr, *block;
4475 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4478 if (! mode_is_int(mode))
4481 shl = get_binop_left(or);
4482 shr = get_binop_right(or);
4491 } else if (!is_Shl(shl)) {
4493 } else if (!is_Shr(shr)) {
4496 x = get_Shl_left(shl);
4497 if (x != get_Shr_left(shr))
4500 c1 = get_Shl_right(shl);
4501 c2 = get_Shr_right(shr);
4502 if (is_Const(c1) && is_Const(c2)) {
4503 tv1 = get_Const_tarval(c1);
4504 if (! tarval_is_long(tv1))
4507 tv2 = get_Const_tarval(c2);
4508 if (! tarval_is_long(tv2))
4511 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4512 != (int) get_mode_size_bits(mode))
4515 /* yet, condition met */
4516 block = get_irn_n(or, -1);
4518 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4520 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4522 } else if (is_Sub(c1)) {
4526 if (get_Sub_right(sub) != v)
4529 c1 = get_Sub_left(sub);
4533 tv1 = get_Const_tarval(c1);
4534 if (! tarval_is_long(tv1))
4537 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4540 /* yet, condition met */
4541 block = get_nodes_block(or);
4543 /* a Rot right is not supported, so use a rot left */
4544 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4546 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4548 } else if (is_Sub(c2)) {
4552 c1 = get_Sub_left(sub);
4556 tv1 = get_Const_tarval(c1);
4557 if (! tarval_is_long(tv1))
4560 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4563 /* yet, condition met */
4564 block = get_irn_n(or, -1);
4567 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4569 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4574 } /* transform_node_Or_Rot */
4579 static ir_node *transform_node_Or(ir_node *n) {
4580 ir_node *c, *oldn = n;
4581 ir_node *a = get_Or_left(n);
4582 ir_node *b = get_Or_right(n);
4585 if (is_Not(a) && is_Not(b)) {
4586 /* ~a | ~b = ~(a&b) */
4587 ir_node *block = get_nodes_block(n);
4589 mode = get_irn_mode(n);
4592 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4593 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4594 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4598 /* we can evaluate 2 Projs of the same Cmp */
4599 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4600 ir_node *pred_a = get_Proj_pred(a);
4601 ir_node *pred_b = get_Proj_pred(b);
4602 if (pred_a == pred_b) {
4603 dbg_info *dbgi = get_irn_dbg_info(n);
4604 ir_node *block = get_nodes_block(pred_a);
4605 pn_Cmp pn_a = get_Proj_proj(a);
4606 pn_Cmp pn_b = get_Proj_proj(b);
4607 /* yes, we can simply calculate with pncs */
4608 pn_Cmp new_pnc = pn_a | pn_b;
4610 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4615 mode = get_irn_mode(n);
4616 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4618 n = transform_node_Or_bf_store(n);
4619 n = transform_node_Or_Rot(n);
4623 n = transform_bitwise_distributive(n, transform_node_Or);
4626 } /* transform_node_Or */
4630 static ir_node *transform_node(ir_node *n);
4633 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4635 * Should be moved to reassociation?
4637 static ir_node *transform_node_shift(ir_node *n) {
4638 ir_node *left, *right;
4639 tarval *tv1, *tv2, *res;
4641 int modulo_shf, flag;
4643 left = get_binop_left(n);
4645 /* different operations */
4646 if (get_irn_op(left) != get_irn_op(n))
4649 right = get_binop_right(n);
4650 tv1 = value_of(right);
4651 if (tv1 == tarval_bad)
4654 tv2 = value_of(get_binop_right(left));
4655 if (tv2 == tarval_bad)
4658 res = tarval_add(tv1, tv2);
4660 /* beware: a simple replacement works only, if res < modulo shift */
4661 mode = get_irn_mode(n);
4665 modulo_shf = get_mode_modulo_shift(mode);
4666 if (modulo_shf > 0) {
4667 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4669 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4675 /* ok, we can replace it */
4676 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4678 in[0] = get_binop_left(left);
4679 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4681 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4683 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4685 return transform_node(irn);
4688 } /* transform_node_shift */
4693 static ir_node *transform_node_Shr(ir_node *n) {
4694 ir_node *c, *oldn = n;
4695 ir_node *a = get_Shr_left(n);
4696 ir_node *b = get_Shr_right(n);
4697 ir_mode *mode = get_irn_mode(n);
4699 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4700 return transform_node_shift(n);
4701 } /* transform_node_Shr */
4706 static ir_node *transform_node_Shrs(ir_node *n) {
4707 ir_node *c, *oldn = n;
4708 ir_node *a = get_Shrs_left(n);
4709 ir_node *b = get_Shrs_right(n);
4710 ir_mode *mode = get_irn_mode(n);
4712 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4713 return transform_node_shift(n);
4714 } /* transform_node_Shrs */
4719 static ir_node *transform_node_Shl(ir_node *n) {
4720 ir_node *c, *oldn = n;
4721 ir_node *a = get_Shl_left(n);
4722 ir_node *b = get_Shl_right(n);
4723 ir_mode *mode = get_irn_mode(n);
4725 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4726 return transform_node_shift(n);
4727 } /* transform_node_Shl */
4732 static ir_node *transform_node_Rot(ir_node *n) {
4733 ir_node *c, *oldn = n;
4734 ir_node *a = get_Rot_left(n);
4735 ir_node *b = get_Rot_right(n);
4736 ir_mode *mode = get_irn_mode(n);
4738 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4739 return transform_node_shift(n);
4740 } /* transform_node_Rot */
4745 static ir_node *transform_node_Conv(ir_node *n) {
4746 ir_node *c, *oldn = n;
4747 ir_node *a = get_Conv_op(n);
4749 if (is_const_Phi(a)) {
4750 c = apply_conv_on_phi(a, get_irn_mode(n));
4752 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4757 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4758 ir_mode *mode = get_irn_mode(n);
4759 return new_r_Unknown(current_ir_graph, mode);
4763 } /* transform_node_Conv */
4766 * Remove dead blocks and nodes in dead blocks
4767 * in keep alive list. We do not generate a new End node.
4769 static ir_node *transform_node_End(ir_node *n) {
4770 int i, j, n_keepalives = get_End_n_keepalives(n);
4773 NEW_ARR_A(ir_node *, in, n_keepalives);
4775 for (i = j = 0; i < n_keepalives; ++i) {
4776 ir_node *ka = get_End_keepalive(n, i);
4778 if (! is_Block_dead(ka)) {
4782 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4785 /* FIXME: beabi need to keep a Proj(M) */
4786 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4789 if (j != n_keepalives)
4790 set_End_keepalives(n, j, in);
4792 } /* transform_node_End */
4794 /** returns 1 if a == -b */
4795 static int is_negated_value(ir_node *a, ir_node *b) {
4796 if (is_Minus(a) && get_Minus_op(a) == b)
4798 if (is_Minus(b) && get_Minus_op(b) == a)
4800 if (is_Sub(a) && is_Sub(b)) {
4801 ir_node *a_left = get_Sub_left(a);
4802 ir_node *a_right = get_Sub_right(a);
4803 ir_node *b_left = get_Sub_left(b);
4804 ir_node *b_right = get_Sub_right(b);
4806 if (a_left == b_right && a_right == b_left)
4814 * Optimize a Mux into some simpler cases.
4816 static ir_node *transform_node_Mux(ir_node *n) {
4817 ir_node *oldn = n, *sel = get_Mux_sel(n);
4818 ir_mode *mode = get_irn_mode(n);
4820 if (mode == mode_b) {
4821 ir_node *t = get_Mux_true(n);
4822 ir_node *f = get_Mux_false(n);
4823 dbg_info *dbg = get_irn_dbg_info(n);
4824 ir_node *block = get_irn_n(n, -1);
4825 ir_graph *irg = current_ir_graph;
4828 tarval *tv_t = get_Const_tarval(t);
4829 if (tv_t == tarval_b_true) {
4831 /* Muxb(sel, true, false) = sel */
4832 assert(get_Const_tarval(f) == tarval_b_false);
4833 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
4836 /* Muxb(sel, true, x) = Or(sel, x) */
4837 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
4838 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
4842 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4843 assert(tv_t == tarval_b_false);
4845 /* Muxb(sel, false, true) = Not(sel) */
4846 assert(get_Const_tarval(f) == tarval_b_true);
4847 DBG_OPT_ALGSIM0(oldn, not_sel, FS_OPT_MUX_NOT_BOOL);
4850 /* Muxb(sel, false, x) = And(Not(sel), x) */
4851 n = new_rd_And(dbg, irg, block, not_sel, f, mode_b);
4852 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ANDNOT_BOOL);
4856 } else if (is_Const(f)) {
4857 tarval *tv_f = get_Const_tarval(f);
4858 if (tv_f == tarval_b_true) {
4859 /* Muxb(sel, x, true) = Or(Not(sel), x) */
4860 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4861 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
4862 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4865 /* Muxb(sel, x, false) = And(sel, x) */
4866 assert(tv_f == tarval_b_false);
4867 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
4868 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
4874 if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
4875 ir_node *cmp = get_Proj_pred(sel);
4876 long pn = get_Proj_proj(sel);
4877 ir_node *f = get_Mux_false(n);
4878 ir_node *t = get_Mux_true(n);
4881 * Note: normalization puts the constant on the right side,
4882 * so we check only one case.
4884 * Note further that these optimization work even for floating point
4885 * with NaN's because -NaN == NaN.
4886 * However, if +0 and -0 is handled differently, we cannot use the first
4890 ir_node *cmp_r = get_Cmp_right(cmp);
4891 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4892 ir_node *block = get_irn_n(n, -1);
4894 if (is_negated_value(f, t)) {
4895 ir_node *cmp_left = get_Cmp_left(cmp);
4897 /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
4898 if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4899 || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4901 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4903 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4905 /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
4906 } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4907 || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4909 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4911 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4913 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4920 return arch_transform_node_Mux(n);
4921 } /* transform_node_Mux */
4924 * Optimize a Psi into some simpler cases.
4926 static ir_node *transform_node_Psi(ir_node *n) {
4928 return transform_node_Mux(n);
4931 } /* transform_node_Psi */
4934 * optimize sync nodes that have other syncs as input we simply add the inputs
4935 * of the other sync to our own inputs
4937 static ir_node *transform_node_Sync(ir_node *n) {
4938 int arity = get_Sync_n_preds(n);
4941 for (i = 0; i < arity;) {
4942 ir_node *pred = get_Sync_pred(n, i);
4946 if (!is_Sync(pred)) {
4954 pred_arity = get_Sync_n_preds(pred);
4955 for (j = 0; j < pred_arity; ++j) {
4956 ir_node *pred_pred = get_Sync_pred(pred, j);
4961 add_irn_n(n, pred_pred);
4965 if (get_Sync_pred(n, k) == pred_pred) break;
4970 /* rehash the sync node */
4971 add_identities(current_ir_graph->value_table, n);
4977 * Tries several [inplace] [optimizing] transformations and returns an
4978 * equivalent node. The difference to equivalent_node() is that these
4979 * transformations _do_ generate new nodes, and thus the old node must
4980 * not be freed even if the equivalent node isn't the old one.
4982 static ir_node *transform_node(ir_node *n) {
4986 * Transform_node is the only "optimizing transformation" that might
4987 * return a node with a different opcode. We iterate HERE until fixpoint
4988 * to get the final result.
4992 if (n->op->ops.transform_node)
4993 n = n->op->ops.transform_node(n);
4994 } while (oldn != n);
4997 } /* transform_node */
5000 * Sets the default transform node operation for an ir_op_ops.
5002 * @param code the opcode for the default operation
5003 * @param ops the operations initialized
5008 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5012 ops->transform_node = transform_node_##a; \
5049 } /* firm_set_default_transform_node */
5052 /* **************** Common Subexpression Elimination **************** */
5054 /** The size of the hash table used, should estimate the number of nodes
5056 #define N_IR_NODES 512
5058 /** Compares the attributes of two Const nodes. */
5059 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5060 return (get_Const_tarval(a) != get_Const_tarval(b))
5061 || (get_Const_type(a) != get_Const_type(b));
5062 } /* node_cmp_attr_Const */
5064 /** Compares the attributes of two Proj nodes. */
5065 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5066 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5067 } /* node_cmp_attr_Proj */
5069 /** Compares the attributes of two Filter nodes. */
5070 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5071 return get_Filter_proj(a) != get_Filter_proj(b);
5072 } /* node_cmp_attr_Filter */
5074 /** Compares the attributes of two Alloc nodes. */
5075 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5076 const alloc_attr *pa = get_irn_alloc_attr(a);
5077 const alloc_attr *pb = get_irn_alloc_attr(b);
5078 return (pa->where != pb->where) || (pa->type != pb->type);
5079 } /* node_cmp_attr_Alloc */
5081 /** Compares the attributes of two Free nodes. */
5082 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5083 const free_attr *pa = get_irn_free_attr(a);
5084 const free_attr *pb = get_irn_free_attr(b);
5085 return (pa->where != pb->where) || (pa->type != pb->type);
5086 } /* node_cmp_attr_Free */
5088 /** Compares the attributes of two SymConst nodes. */
5089 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5090 const symconst_attr *pa = get_irn_symconst_attr(a);
5091 const symconst_attr *pb = get_irn_symconst_attr(b);
5092 return (pa->kind != pb->kind)
5093 || (pa->sym.type_p != pb->sym.type_p)
5094 || (pa->tp != pb->tp);
5095 } /* node_cmp_attr_SymConst */
5097 /** Compares the attributes of two Call nodes. */
5098 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5099 return get_irn_call_attr(a) != get_irn_call_attr(b);
5100 } /* node_cmp_attr_Call */
5102 /** Compares the attributes of two Sel nodes. */
5103 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5104 const ir_entity *a_ent = get_Sel_entity(a);
5105 const ir_entity *b_ent = get_Sel_entity(b);
5107 (a_ent->kind != b_ent->kind) ||
5108 (a_ent->name != b_ent->name) ||
5109 (a_ent->owner != b_ent->owner) ||
5110 (a_ent->ld_name != b_ent->ld_name) ||
5111 (a_ent->type != b_ent->type);
5112 } /* node_cmp_attr_Sel */
5114 /** Compares the attributes of two Phi nodes. */
5115 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5116 /* we can only enter this function if both nodes have the same number of inputs,
5117 hence it is enough to check if one of them is a Phi0 */
5119 /* check the Phi0 pos attribute */
5120 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5123 } /* node_cmp_attr_Phi */
5125 /** Compares the attributes of two Conv nodes. */
5126 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5127 return get_Conv_strict(a) != get_Conv_strict(b);
5128 } /* node_cmp_attr_Conv */
5130 /** Compares the attributes of two Cast nodes. */
5131 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5132 return get_Cast_type(a) != get_Cast_type(b);
5133 } /* node_cmp_attr_Cast */
5135 /** Compares the attributes of two Load nodes. */
5136 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5137 if (get_Load_volatility(a) == volatility_is_volatile ||
5138 get_Load_volatility(b) == volatility_is_volatile)
5139 /* NEVER do CSE on volatile Loads */
5141 /* do not CSE Loads with different alignment. Be conservative. */
5142 if (get_Load_align(a) != get_Load_align(b))
5145 return get_Load_mode(a) != get_Load_mode(b);
5146 } /* node_cmp_attr_Load */
5148 /** Compares the attributes of two Store nodes. */
5149 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5150 /* do not CSE Stores with different alignment. Be conservative. */
5151 if (get_Store_align(a) != get_Store_align(b))
5154 /* NEVER do CSE on volatile Stores */
5155 return (get_Store_volatility(a) == volatility_is_volatile ||
5156 get_Store_volatility(b) == volatility_is_volatile);
5157 } /* node_cmp_attr_Store */
5159 /** Compares two exception attributes */
5160 static int node_cmp_exception(ir_node *a, ir_node *b) {
5161 const except_attr *ea = get_irn_except_attr(a);
5162 const except_attr *eb = get_irn_except_attr(b);
5164 return ea->pin_state != eb->pin_state;
5167 #define node_cmp_attr_Bound node_cmp_exception
5169 /** Compares the attributes of two Div nodes. */
5170 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5171 const divmod_attr *ma = get_irn_divmod_attr(a);
5172 const divmod_attr *mb = get_irn_divmod_attr(b);
5173 return ma->exc.pin_state != mb->exc.pin_state ||
5174 ma->res_mode != mb->res_mode ||
5175 ma->no_remainder != mb->no_remainder;
5176 } /* node_cmp_attr_Div */
5178 /** Compares the attributes of two DivMod nodes. */
5179 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5180 const divmod_attr *ma = get_irn_divmod_attr(a);
5181 const divmod_attr *mb = get_irn_divmod_attr(b);
5182 return ma->exc.pin_state != mb->exc.pin_state ||
5183 ma->res_mode != mb->res_mode;
5184 } /* node_cmp_attr_DivMod */
5186 /** Compares the attributes of two Mod nodes. */
5187 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5188 const divmod_attr *ma = get_irn_divmod_attr(a);
5189 const divmod_attr *mb = get_irn_divmod_attr(b);
5190 return ma->exc.pin_state != mb->exc.pin_state ||
5191 ma->res_mode != mb->res_mode;
5192 } /* node_cmp_attr_Mod */
5194 /** Compares the attributes of two Quot nodes. */
5195 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5196 const divmod_attr *ma = get_irn_divmod_attr(a);
5197 const divmod_attr *mb = get_irn_divmod_attr(b);
5198 return ma->exc.pin_state != mb->exc.pin_state ||
5199 ma->res_mode != mb->res_mode;
5200 } /* node_cmp_attr_Quot */
5202 /** Compares the attributes of two Confirm nodes. */
5203 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5204 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5205 } /* node_cmp_attr_Confirm */
5207 /** Compares the attributes of two ASM nodes. */
5208 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5210 const ir_asm_constraint *ca;
5211 const ir_asm_constraint *cb;
5214 if (get_ASM_text(a) != get_ASM_text(b))
5217 /* Should we really check the constraints here? Should be better, but is strange. */
5218 n = get_ASM_n_input_constraints(a);
5219 if (n != get_ASM_n_input_constraints(b))
5222 ca = get_ASM_input_constraints(a);
5223 cb = get_ASM_input_constraints(b);
5224 for (i = 0; i < n; ++i) {
5225 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5229 n = get_ASM_n_output_constraints(a);
5230 if (n != get_ASM_n_output_constraints(b))
5233 ca = get_ASM_output_constraints(a);
5234 cb = get_ASM_output_constraints(b);
5235 for (i = 0; i < n; ++i) {
5236 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5240 n = get_ASM_n_clobbers(a);
5241 if (n != get_ASM_n_clobbers(b))
5244 cla = get_ASM_clobbers(a);
5245 clb = get_ASM_clobbers(b);
5246 for (i = 0; i < n; ++i) {
5247 if (cla[i] != clb[i])
5251 } /* node_cmp_attr_ASM */
5254 * Set the default node attribute compare operation for an ir_op_ops.
5256 * @param code the opcode for the default operation
5257 * @param ops the operations initialized
5262 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5266 ops->node_cmp_attr = node_cmp_attr_##a; \
5297 } /* firm_set_default_node_cmp_attr */
5300 * Compare function for two nodes in the value table. Gets two
5301 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5303 int identities_cmp(const void *elt, const void *key) {
5304 ir_node *a = (ir_node *)elt;
5305 ir_node *b = (ir_node *)key;
5308 if (a == b) return 0;
5310 if ((get_irn_op(a) != get_irn_op(b)) ||
5311 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5313 /* compare if a's in and b's in are of equal length */
5314 irn_arity_a = get_irn_intra_arity(a);
5315 if (irn_arity_a != get_irn_intra_arity(b))
5318 if (get_irn_pinned(a) == op_pin_state_pinned) {
5319 /* for pinned nodes, the block inputs must be equal */
5320 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5322 } else if (! get_opt_global_cse()) {
5323 /* for block-local CSE both nodes must be in the same MacroBlock */
5324 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5328 /* compare a->in[0..ins] with b->in[0..ins] */
5329 for (i = 0; i < irn_arity_a; i++)
5330 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5334 * here, we already now that the nodes are identical except their
5337 if (a->op->ops.node_cmp_attr)
5338 return a->op->ops.node_cmp_attr(a, b);
5341 } /* identities_cmp */
5344 * Calculate a hash value of a node.
5346 unsigned ir_node_hash(ir_node *node) {
5350 if (node->op == op_Const) {
5351 /* special value for const, as they only differ in their tarval. */
5352 h = HASH_PTR(node->attr.con.tv);
5353 h = 9*h + HASH_PTR(get_irn_mode(node));
5354 } else if (node->op == op_SymConst) {
5355 /* special value for const, as they only differ in their symbol. */
5356 h = HASH_PTR(node->attr.symc.sym.type_p);
5357 h = 9*h + HASH_PTR(get_irn_mode(node));
5360 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5361 h = irn_arity = get_irn_intra_arity(node);
5363 /* consider all in nodes... except the block if not a control flow. */
5364 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5365 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5369 h = 9*h + HASH_PTR(get_irn_mode(node));
5371 h = 9*h + HASH_PTR(get_irn_op(node));
5375 } /* ir_node_hash */
5377 pset *new_identities(void) {
5378 return new_pset(identities_cmp, N_IR_NODES);
5379 } /* new_identities */
5381 void del_identities(pset *value_table) {
5382 del_pset(value_table);
5383 } /* del_identities */
5386 * Normalize a node by putting constants (and operands with larger
5387 * node index) on the right (operator side).
5389 * @param n The node to normalize
5391 static void normalize_node(ir_node *n) {
5392 if (is_op_commutative(get_irn_op(n))) {
5393 ir_node *l = get_binop_left(n);
5394 ir_node *r = get_binop_right(n);
5396 /* For commutative operators perform a OP b == b OP a but keep
5397 * constants on the RIGHT side. This helps greatly in some
5398 * optimizations. Moreover we use the idx number to make the form
5400 if (!operands_are_normalized(l, r)) {
5401 set_binop_left(n, r);
5402 set_binop_right(n, l);
5405 } /* normalize_node */
5408 * Update the nodes after a match in the value table. If both nodes have
5409 * the same MacroBlock but different Blocks, we must ensure that the node
5410 * with the dominating Block (the node that is near to the MacroBlock header
5411 * is stored in the table.
5412 * Because a MacroBlock has only one "non-exception" flow, we don't need
5413 * dominance info here: We known, that one block must dominate the other and
5414 * following the only block input will allow to find it.
5416 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5417 ir_node *known_blk, *new_block, *block, *mbh;
5419 if (get_opt_global_cse()) {
5420 /* Block inputs are meaning less */
5423 known_blk = get_irn_n(known_irn, -1);
5424 new_block = get_irn_n(new_ir_node, -1);
5425 if (known_blk == new_block) {
5426 /* already in the same block */
5430 * We expect the typical case when we built the graph. In that case, the
5431 * known_irn is already the upper one, so checking this should be faster.
5434 mbh = get_Block_MacroBlock(new_block);
5436 if (block == known_blk) {
5437 /* ok, we have found it: known_block dominates new_block as expected */
5442 * We have reached the MacroBlock header NOT founding
5443 * the known_block. new_block must dominate known_block.
5446 set_irn_n(known_irn, -1, new_block);
5449 assert(get_Block_n_cfgpreds(block) == 1);
5450 block = get_Block_cfgpred_block(block, 0);
5452 } /* update_value_table */
5455 * Return the canonical node computing the same value as n.
5457 * @param value_table The value table
5458 * @param n The node to lookup
5460 * Looks up the node in a hash table.
5462 * For Const nodes this is performed in the constructor, too. Const
5463 * nodes are extremely time critical because of their frequent use in
5464 * constant string arrays.
5466 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5469 if (!value_table) return n;
5473 o = pset_find(value_table, n, ir_node_hash(n));
5477 update_known_irn(o, n);
5484 * During construction we set the op_pin_state_pinned flag in the graph right when the
5485 * optimization is performed. The flag turning on procedure global cse could
5486 * be changed between two allocations. This way we are safe.
5488 * @param value_table The value table
5489 * @param n The node to lookup
5491 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5494 n = identify(value_table, n);
5495 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5496 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5498 } /* identify_cons */
5501 * Return the canonical node computing the same value as n.
5502 * Looks up the node in a hash table, enters it in the table
5503 * if it isn't there yet.
5505 * @param value_table the HashSet containing all nodes in the
5507 * @param n the node to look up
5509 * @return a node that computes the same value as n or n if no such
5510 * node could be found
5512 ir_node *identify_remember(pset *value_table, ir_node *n) {
5515 if (!value_table) return n;
5518 /* lookup or insert in hash table with given hash key. */
5519 o = pset_insert(value_table, n, ir_node_hash(n));
5522 update_known_irn(o, n);
5527 } /* identify_remember */
5529 /* Add a node to the identities value table. */
5530 void add_identities(pset *value_table, ir_node *node) {
5531 if (get_opt_cse() && is_no_Block(node))
5532 identify_remember(value_table, node);
5533 } /* add_identities */
5535 /* Visit each node in the value table of a graph. */
5536 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5538 ir_graph *rem = current_ir_graph;
5540 current_ir_graph = irg;
5541 foreach_pset(irg->value_table, node)
5543 current_ir_graph = rem;
5544 } /* visit_all_identities */
5547 * Garbage in, garbage out. If a node has a dead input, i.e., the
5548 * Bad node is input to the node, return the Bad node.
5550 static ir_node *gigo(ir_node *node) {
5552 ir_op *op = get_irn_op(node);
5554 /* remove garbage blocks by looking at control flow that leaves the block
5555 and replacing the control flow by Bad. */
5556 if (get_irn_mode(node) == mode_X) {
5557 ir_node *block = get_nodes_block(skip_Proj(node));
5559 /* Don't optimize nodes in immature blocks. */
5560 if (!get_Block_matured(block))
5562 /* Don't optimize End, may have Bads. */
5563 if (op == op_End) return node;
5565 if (is_Block(block)) {
5566 if (is_Block_dead(block)) {
5567 /* control flow from dead block is dead */
5571 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5572 if (!is_Bad(get_irn_n(block, i)))
5576 ir_graph *irg = get_irn_irg(block);
5577 /* the start block is never dead */
5578 if (block != get_irg_start_block(irg)
5579 && block != get_irg_end_block(irg)) {
5581 * Do NOT kill control flow without setting
5582 * the block to dead of bad things can happen:
5583 * We get a Block that is not reachable be irg_block_walk()
5584 * but can be found by irg_walk()!
5586 set_Block_dead(block);
5593 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5594 blocks predecessors is dead. */
5595 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5596 irn_arity = get_irn_arity(node);
5599 * Beware: we can only read the block of a non-floating node.
5601 if (is_irn_pinned_in_irg(node) &&
5602 is_Block_dead(get_nodes_block(skip_Proj(node))))
5605 for (i = 0; i < irn_arity; i++) {
5606 ir_node *pred = get_irn_n(node, i);
5611 /* Propagating Unknowns here seems to be a bad idea, because
5612 sometimes we need a node as a input and did not want that
5614 However, it might be useful to move this into a later phase
5615 (if you think that optimizing such code is useful). */
5616 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5617 return new_Unknown(get_irn_mode(node));
5622 /* With this code we violate the agreement that local_optimize
5623 only leaves Bads in Block, Phi and Tuple nodes. */
5624 /* If Block has only Bads as predecessors it's garbage. */
5625 /* If Phi has only Bads as predecessors it's garbage. */
5626 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5627 irn_arity = get_irn_arity(node);
5628 for (i = 0; i < irn_arity; i++) {
5629 if (!is_Bad(get_irn_n(node, i))) break;
5631 if (i == irn_arity) node = new_Bad();
5638 * These optimizations deallocate nodes from the obstack.
5639 * It can only be called if it is guaranteed that no other nodes
5640 * reference this one, i.e., right after construction of a node.
5642 * @param n The node to optimize
5644 * current_ir_graph must be set to the graph of the node!
5646 ir_node *optimize_node(ir_node *n) {
5649 ir_opcode iro = get_irn_opcode(n);
5651 /* Always optimize Phi nodes: part of the construction. */
5652 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5654 /* constant expression evaluation / constant folding */
5655 if (get_opt_constant_folding()) {
5656 /* neither constants nor Tuple values can be evaluated */
5657 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5658 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5659 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5660 /* try to evaluate */
5661 tv = computed_value(n);
5662 if (tv != tarval_bad) {
5664 ir_type *old_tp = get_irn_type(n);
5665 int i, arity = get_irn_arity(n);
5669 * Try to recover the type of the new expression.
5671 for (i = 0; i < arity && !old_tp; ++i)
5672 old_tp = get_irn_type(get_irn_n(n, i));
5675 * we MUST copy the node here temporary, because it's still needed
5676 * for DBG_OPT_CSTEVAL
5678 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5679 oldn = alloca(node_size);
5681 memcpy(oldn, n, node_size);
5682 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5684 /* ARG, copy the in array, we need it for statistics */
5685 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5687 /* note the inplace edges module */
5688 edges_node_deleted(n, current_ir_graph);
5690 /* evaluation was successful -- replace the node. */
5691 irg_kill_node(current_ir_graph, n);
5692 nw = new_Const(get_tarval_mode(tv), tv);
5694 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5695 set_Const_type(nw, old_tp);
5696 DBG_OPT_CSTEVAL(oldn, nw);
5697 tarval_enable_fp_ops(old_fp_mode);
5700 tarval_enable_fp_ops(old_fp_mode);
5704 /* remove unnecessary nodes */
5705 if (get_opt_constant_folding() ||
5706 (iro == iro_Phi) || /* always optimize these nodes. */
5708 (iro == iro_Proj) ||
5709 (iro == iro_Block) ) /* Flags tested local. */
5710 n = equivalent_node(n);
5712 /* Common Subexpression Elimination.
5714 * Checks whether n is already available.
5715 * The block input is used to distinguish different subexpressions. Right
5716 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5717 * subexpressions within a block.
5720 n = identify_cons(current_ir_graph->value_table, n);
5723 edges_node_deleted(oldn, current_ir_graph);
5725 /* We found an existing, better node, so we can deallocate the old node. */
5726 irg_kill_node(current_ir_graph, oldn);
5730 /* Some more constant expression evaluation that does not allow to
5732 iro = get_irn_opcode(n);
5733 if (get_opt_constant_folding() ||
5734 (iro == iro_Cond) ||
5735 (iro == iro_Proj)) /* Flags tested local. */
5736 n = transform_node(n);
5738 /* Remove nodes with dead (Bad) input.
5739 Run always for transformation induced Bads. */
5742 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5743 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5744 n = identify_remember(current_ir_graph->value_table, n);
5748 } /* optimize_node */
5752 * These optimizations never deallocate nodes (in place). This can cause dead
5753 * nodes lying on the obstack. Remove these by a dead node elimination,
5754 * i.e., a copying garbage collection.
5756 ir_node *optimize_in_place_2(ir_node *n) {
5759 ir_opcode iro = get_irn_opcode(n);
5761 if (!get_opt_optimize() && !is_Phi(n)) return n;
5763 /* constant expression evaluation / constant folding */
5764 if (get_opt_constant_folding()) {
5765 /* neither constants nor Tuple values can be evaluated */
5766 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5767 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5768 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5769 /* try to evaluate */
5770 tv = computed_value(n);
5771 if (tv != tarval_bad) {
5772 /* evaluation was successful -- replace the node. */
5773 ir_type *old_tp = get_irn_type(n);
5774 int i, arity = get_irn_arity(n);
5777 * Try to recover the type of the new expression.
5779 for (i = 0; i < arity && !old_tp; ++i)
5780 old_tp = get_irn_type(get_irn_n(n, i));
5782 n = new_Const(get_tarval_mode(tv), tv);
5784 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5785 set_Const_type(n, old_tp);
5787 DBG_OPT_CSTEVAL(oldn, n);
5788 tarval_enable_fp_ops(old_fp_mode);
5791 tarval_enable_fp_ops(old_fp_mode);
5795 /* remove unnecessary nodes */
5796 if (get_opt_constant_folding() ||
5797 (iro == iro_Phi) || /* always optimize these nodes. */
5798 (iro == iro_Id) || /* ... */
5799 (iro == iro_Proj) || /* ... */
5800 (iro == iro_Block) ) /* Flags tested local. */
5801 n = equivalent_node(n);
5803 /** common subexpression elimination **/
5804 /* Checks whether n is already available. */
5805 /* The block input is used to distinguish different subexpressions. Right
5806 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5807 subexpressions within a block. */
5808 if (get_opt_cse()) {
5809 n = identify(current_ir_graph->value_table, n);
5812 /* Some more constant expression evaluation. */
5813 iro = get_irn_opcode(n);
5814 if (get_opt_constant_folding() ||
5815 (iro == iro_Cond) ||
5816 (iro == iro_Proj)) /* Flags tested local. */
5817 n = transform_node(n);
5819 /* Remove nodes with dead (Bad) input.
5820 Run always for transformation induced Bads. */
5823 /* Now we can verify the node, as it has no dead inputs any more. */
5826 /* Now we have a legal, useful node. Enter it in hash table for cse.
5827 Blocks should be unique anyways. (Except the successor of start:
5828 is cse with the start block!) */
5829 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5830 n = identify_remember(current_ir_graph->value_table, n);
5833 } /* optimize_in_place_2 */
5836 * Wrapper for external use, set proper status bits after optimization.
5838 ir_node *optimize_in_place(ir_node *n) {
5839 /* Handle graph state */
5840 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5842 if (get_opt_global_cse())
5843 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5844 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5845 set_irg_outs_inconsistent(current_ir_graph);
5847 /* FIXME: Maybe we could also test whether optimizing the node can
5848 change the control graph. */
5849 set_irg_doms_inconsistent(current_ir_graph);
5850 return optimize_in_place_2(n);
5851 } /* optimize_in_place */
5854 * Sets the default operation for an ir_ops.
5856 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5857 ops = firm_set_default_computed_value(code, ops);
5858 ops = firm_set_default_equivalent_node(code, ops);
5859 ops = firm_set_default_transform_node(code, ops);
5860 ops = firm_set_default_node_cmp_attr(code, ops);
5861 ops = firm_set_default_get_type(code, ops);
5862 ops = firm_set_default_get_type_attr(code, ops);
5863 ops = firm_set_default_get_entity_attr(code, ops);
5866 } /* firm_set_default_operations */