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);
892 } else if (ab == b) {
893 /* (a ^ b) ^ b -> a */
895 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
900 ir_node *ba = get_Eor_left(b);
901 ir_node *bb = get_Eor_right(b);
904 /* a ^ (a ^ b) -> b */
906 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
908 } else if (bb == a) {
909 /* a ^ (b ^ a) -> b */
911 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
920 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
922 * The second one looks strange, but this construct
923 * is used heavily in the LCC sources :-).
925 * Beware: The Mode of an Add may be different than the mode of its
926 * predecessors, so we could not return a predecessors in all cases.
928 static ir_node *equivalent_node_Add(ir_node *n) {
930 ir_node *left, *right;
931 ir_mode *mode = get_irn_mode(n);
933 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
934 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
937 n = equivalent_node_neutral_zero(n);
941 left = get_Add_left(n);
942 right = get_Add_right(n);
945 if (get_Sub_right(left) == right) {
948 n = get_Sub_left(left);
949 if (mode == get_irn_mode(n)) {
950 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
956 if (get_Sub_right(right) == left) {
959 n = get_Sub_left(right);
960 if (mode == get_irn_mode(n)) {
961 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
967 } /* equivalent_node_Add */
970 * optimize operations that are not commutative but have neutral 0 on left,
973 static ir_node *equivalent_node_left_zero(ir_node *n) {
976 ir_node *a = get_binop_left(n);
977 ir_node *b = get_binop_right(n);
979 if (is_Const(b) && is_Const_null(b)) {
982 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
985 } /* equivalent_node_left_zero */
987 #define equivalent_node_Shl equivalent_node_left_zero
988 #define equivalent_node_Shr equivalent_node_left_zero
989 #define equivalent_node_Shrs equivalent_node_left_zero
990 #define equivalent_node_Rot equivalent_node_left_zero
993 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
995 * The second one looks strange, but this construct
996 * is used heavily in the LCC sources :-).
998 * Beware: The Mode of a Sub may be different than the mode of its
999 * predecessors, so we could not return a predecessors in all cases.
1001 static ir_node *equivalent_node_Sub(ir_node *n) {
1004 ir_mode *mode = get_irn_mode(n);
1006 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1007 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1010 b = get_Sub_right(n);
1012 /* Beware: modes might be different */
1013 if (is_Const(b) && is_Const_null(b)) {
1014 ir_node *a = get_Sub_left(n);
1015 if (mode == get_irn_mode(a)) {
1018 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1022 } /* equivalent_node_Sub */
1026 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1029 * -(-a) == a, but might overflow two times.
1030 * We handle it anyway here but the better way would be a
1031 * flag. This would be needed for Pascal for instance.
1033 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1035 ir_node *pred = get_unop_op(n);
1037 /* optimize symmetric unop */
1038 if (get_irn_op(pred) == get_irn_op(n)) {
1039 n = get_unop_op(pred);
1040 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1043 } /* equivalent_node_idempotent_unop */
1045 /** Optimize Not(Not(x)) == x. */
1046 #define equivalent_node_Not equivalent_node_idempotent_unop
1048 /** -(-x) == x ??? Is this possible or can --x raise an
1049 out of bounds exception if min =! max? */
1050 #define equivalent_node_Minus equivalent_node_idempotent_unop
1053 * Optimize a * 1 = 1 * a = a.
1055 static ir_node *equivalent_node_Mul(ir_node *n) {
1057 ir_node *a = get_Mul_left(n);
1059 /* we can handle here only the n * n = n bit cases */
1060 if (get_irn_mode(n) == get_irn_mode(a)) {
1061 ir_node *b = get_Mul_right(n);
1063 /* Mul is commutative and has again an other neutral element. */
1064 if (is_Const(a) && is_Const_one(a)) {
1066 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1067 } else if (is_Const(b) && is_Const_one(b)) {
1069 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1073 } /* equivalent_node_Mul */
1076 * Optimize a / 1 = a.
1078 static ir_node *equivalent_node_Div(ir_node *n) {
1079 ir_node *a = get_Div_left(n);
1080 ir_node *b = get_Div_right(n);
1082 /* Div is not commutative. */
1083 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1084 /* Turn Div into a tuple (mem, bad, a) */
1085 ir_node *mem = get_Div_mem(n);
1086 ir_node *blk = get_irn_n(n, -1);
1087 turn_into_tuple(n, pn_Div_max);
1088 set_Tuple_pred(n, pn_Div_M, mem);
1089 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1090 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1091 set_Tuple_pred(n, pn_Div_res, a);
1094 } /* equivalent_node_Div */
1097 * Optimize a / 1.0 = a.
1099 static ir_node *equivalent_node_Quot(ir_node *n) {
1100 ir_node *a = get_Quot_left(n);
1101 ir_node *b = get_Quot_right(n);
1103 /* Div is not commutative. */
1104 if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */
1105 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1106 ir_node *mem = get_Quot_mem(n);
1107 ir_node *blk = get_irn_n(n, -1);
1108 turn_into_tuple(n, pn_Quot_max);
1109 set_Tuple_pred(n, pn_Quot_M, mem);
1110 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1111 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1112 set_Tuple_pred(n, pn_Quot_res, a);
1115 } /* equivalent_node_Quot */
1118 * Optimize a / 1 = a.
1120 static ir_node *equivalent_node_DivMod(ir_node *n) {
1121 ir_node *b = get_DivMod_right(n);
1123 /* Div is not commutative. */
1124 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1125 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1126 ir_node *a = get_DivMod_left(n);
1127 ir_node *mem = get_Div_mem(n);
1128 ir_node *blk = get_irn_n(n, -1);
1129 ir_mode *mode = get_DivMod_resmode(n);
1131 turn_into_tuple(n, pn_DivMod_max);
1132 set_Tuple_pred(n, pn_DivMod_M, mem);
1133 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1134 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1135 set_Tuple_pred(n, pn_DivMod_res_div, a);
1136 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1139 } /* equivalent_node_DivMod */
1142 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1144 static ir_node *equivalent_node_Or(ir_node *n) {
1147 ir_node *a = get_Or_left(n);
1148 ir_node *b = get_Or_right(n);
1151 n = a; /* Or has it's own neutral element */
1152 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1153 } else if (is_Const(a) && is_Const_null(a)) {
1155 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1156 } else if (is_Const(b) && is_Const_null(b)) {
1158 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1162 } /* equivalent_node_Or */
1165 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1167 static ir_node *equivalent_node_And(ir_node *n) {
1170 ir_node *a = get_And_left(n);
1171 ir_node *b = get_And_right(n);
1174 n = a; /* And has it's own neutral element */
1175 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1178 if (is_Const(a) && is_Const_all_one(a)) {
1180 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1183 if (is_Const(b) && is_Const_all_one(b)) {
1185 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1189 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1192 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1197 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1200 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1206 } /* equivalent_node_And */
1209 * Try to remove useless Conv's:
1211 static ir_node *equivalent_node_Conv(ir_node *n) {
1213 ir_node *a = get_Conv_op(n);
1215 ir_mode *n_mode = get_irn_mode(n);
1216 ir_mode *a_mode = get_irn_mode(a);
1218 if (n_mode == a_mode) { /* No Conv necessary */
1219 if (get_Conv_strict(n)) {
1220 /* special case: the predecessor might be a also a Conv */
1222 if (! get_Conv_strict(a)) {
1223 /* first one is not strict, kick it */
1224 set_Conv_op(n, get_Conv_op(a));
1227 /* else both are strict conv, second is superfluous */
1228 } else if(is_Proj(a)) {
1229 ir_node *pred = get_Proj_pred(a);
1231 /* loads always return with the exact precision of n_mode */
1232 assert(get_Load_mode(pred) == n_mode);
1237 /* leave strict floating point Conv's */
1241 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1242 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1243 ir_node *b = get_Conv_op(a);
1244 ir_mode *b_mode = get_irn_mode(b);
1246 if (n_mode == b_mode) {
1247 if (n_mode == mode_b) {
1248 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1249 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1250 } else if (mode_is_int(n_mode)) {
1251 if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) {
1252 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1253 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1259 } /* equivalent_node_Conv */
1262 * A Cast may be removed if the type of the previous node
1263 * is already the type of the Cast.
1265 static ir_node *equivalent_node_Cast(ir_node *n) {
1267 ir_node *pred = get_Cast_op(n);
1269 if (get_irn_type(pred) == get_Cast_type(n)) {
1271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1274 } /* equivalent_node_Cast */
1277 * Several optimizations:
1278 * - no Phi in start block.
1279 * - remove Id operators that are inputs to Phi
1280 * - fold Phi-nodes, iff they have only one predecessor except
1283 static ir_node *equivalent_node_Phi(ir_node *n) {
1288 ir_node *first_val = NULL; /* to shutup gcc */
1290 if (!get_opt_normalize()) return n;
1292 n_preds = get_Phi_n_preds(n);
1294 block = get_nodes_block(n);
1295 if ((is_Block_dead(block)) || /* Control dead */
1296 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1297 return new_Bad(); /* in the Start Block. */
1299 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1301 /* If the Block has a Bad pred, we also have one. */
1302 for (i = 0; i < n_preds; ++i)
1303 if (is_Bad(get_Block_cfgpred(block, i)))
1304 set_Phi_pred(n, i, new_Bad());
1306 /* Find first non-self-referencing input */
1307 for (i = 0; i < n_preds; ++i) {
1308 first_val = get_Phi_pred(n, i);
1309 if ( (first_val != n) /* not self pointer */
1311 && (! is_Bad(first_val))
1313 ) { /* value not dead */
1314 break; /* then found first value. */
1319 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1323 /* search for rest of inputs, determine if any of these
1324 are non-self-referencing */
1325 while (++i < n_preds) {
1326 ir_node *scnd_val = get_Phi_pred(n, i);
1327 if ( (scnd_val != n)
1328 && (scnd_val != first_val)
1330 && (! is_Bad(scnd_val))
1338 /* Fold, if no multiple distinct non-self-referencing inputs */
1340 DBG_OPT_PHI(oldn, n);
1343 } /* equivalent_node_Phi */
1346 * Several optimizations:
1347 * - no Sync in start block.
1348 * - fold Sync-nodes, iff they have only one predecessor except
1351 static ir_node *equivalent_node_Sync(ir_node *n) {
1352 int arity = get_Sync_n_preds(n);
1355 for (i = 0; i < arity;) {
1356 ir_node *pred = get_Sync_pred(n, i);
1359 /* Remove Bad predecessors */
1366 /* Remove duplicate predecessors */
1372 if (get_Sync_pred(n, j) == pred) {
1380 if (arity == 0) return new_Bad();
1381 if (arity == 1) return get_Sync_pred(n, 0);
1383 } /* equivalent_node_Sync */
1386 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1387 * ProjX(Load) and ProjX(Store).
1389 static ir_node *equivalent_node_Proj(ir_node *proj) {
1390 ir_node *oldn = proj;
1391 ir_node *a = get_Proj_pred(proj);
1394 /* Remove the Tuple/Proj combination. */
1395 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1396 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1397 DBG_OPT_TUPLE(oldn, a, proj);
1399 /* This should not happen! */
1400 assert(! "found a Proj with higher number than Tuple predecessors");
1403 } else if (get_irn_mode(proj) == mode_X) {
1404 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1405 /* Remove dead control flow -- early gigo(). */
1407 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1409 /* get the Load address */
1410 ir_node *addr = get_Load_ptr(a);
1411 ir_node *blk = get_irn_n(a, -1);
1414 if (value_not_null(addr, &confirm)) {
1415 if (confirm == NULL) {
1416 /* this node may float if it did not depend on a Confirm */
1417 set_irn_pinned(a, op_pin_state_floats);
1419 if (get_Proj_proj(proj) == pn_Load_X_except) {
1420 DBG_OPT_EXC_REM(proj);
1423 return new_r_Jmp(current_ir_graph, blk);
1425 } else if (is_Store(a)) {
1426 /* get the load/store address */
1427 ir_node *addr = get_Store_ptr(a);
1428 ir_node *blk = get_irn_n(a, -1);
1431 if (value_not_null(addr, &confirm)) {
1432 if (confirm == NULL) {
1433 /* this node may float if it did not depend on a Confirm */
1434 set_irn_pinned(a, op_pin_state_floats);
1436 if (get_Proj_proj(proj) == pn_Store_X_except) {
1437 DBG_OPT_EXC_REM(proj);
1440 return new_r_Jmp(current_ir_graph, blk);
1447 } /* equivalent_node_Proj */
1452 static ir_node *equivalent_node_Id(ir_node *n) {
1457 } while (get_irn_op(n) == op_Id);
1459 DBG_OPT_ID(oldn, n);
1461 } /* equivalent_node_Id */
1466 static ir_node *equivalent_node_Mux(ir_node *n)
1468 ir_node *oldn = n, *sel = get_Mux_sel(n);
1469 tarval *ts = value_of(sel);
1471 /* Mux(true, f, t) == t */
1472 if (ts == tarval_b_true) {
1473 n = get_Mux_true(n);
1474 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1476 /* Mux(false, f, t) == f */
1477 else if (ts == tarval_b_false) {
1478 n = get_Mux_false(n);
1479 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1481 /* Mux(v, x, x) == x */
1482 else if (get_Mux_false(n) == get_Mux_true(n)) {
1483 n = get_Mux_true(n);
1484 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1486 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1487 ir_node *cmp = get_Proj_pred(sel);
1488 long proj_nr = get_Proj_proj(sel);
1489 ir_node *f = get_Mux_false(n);
1490 ir_node *t = get_Mux_true(n);
1493 * Note further that these optimization work even for floating point
1494 * with NaN's because -NaN == NaN.
1495 * However, if +0 and -0 is handled differently, we cannot use the first one.
1498 ir_node *const cmp_l = get_Cmp_left(cmp);
1499 ir_node *const cmp_r = get_Cmp_right(cmp);
1503 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1504 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1506 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1513 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1514 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1516 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1523 * Note: normalization puts the constant on the right side,
1524 * so we check only one case.
1526 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1527 /* Mux(t CMP 0, X, t) */
1528 if (is_Minus(f) && get_Minus_op(f) == t) {
1529 /* Mux(t CMP 0, -t, t) */
1530 if (proj_nr == pn_Cmp_Eq) {
1531 /* Mux(t == 0, -t, t) ==> -t */
1533 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1534 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1535 /* Mux(t != 0, -t, t) ==> t */
1537 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1544 } /* equivalent_node_Mux */
1547 * Returns a equivalent node of a Psi: if a condition is true
1548 * and all previous conditions are false we know its value.
1549 * If all conditions are false its value is the default one.
1551 static ir_node *equivalent_node_Psi(ir_node *n) {
1553 return equivalent_node_Mux(n);
1555 } /* equivalent_node_Psi */
1558 * Optimize -a CMP -b into b CMP a.
1559 * This works only for for modes where unary Minus
1561 * Note that two-complement integers can Overflow
1562 * so it will NOT work.
1564 * For == and != can be handled in Proj(Cmp)
1566 static ir_node *equivalent_node_Cmp(ir_node *n) {
1567 ir_node *left = get_Cmp_left(n);
1568 ir_node *right = get_Cmp_right(n);
1570 if (is_Minus(left) && is_Minus(right) &&
1571 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1572 left = get_Minus_op(left);
1573 right = get_Minus_op(right);
1574 set_Cmp_left(n, right);
1575 set_Cmp_right(n, left);
1578 } /* equivalent_node_Cmp */
1581 * Remove Confirm nodes if setting is on.
1582 * Replace Confirms(x, '=', Constlike) by Constlike.
1584 static ir_node *equivalent_node_Confirm(ir_node *n) {
1585 ir_node *pred = get_Confirm_value(n);
1586 pn_Cmp pnc = get_Confirm_cmp(n);
1588 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1590 * rare case: two identical Confirms one after another,
1591 * replace the second one with the first.
1595 if (pnc == pn_Cmp_Eq) {
1596 ir_node *bound = get_Confirm_bound(n);
1599 * Optimize a rare case:
1600 * Confirm(x, '=', Constlike) ==> Constlike
1602 if (is_irn_constlike(bound)) {
1603 DBG_OPT_CONFIRM(n, bound);
1607 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1611 * Optimize CopyB(mem, x, x) into a Nop.
1613 static ir_node *equivalent_node_CopyB(ir_node *n) {
1614 ir_node *a = get_CopyB_dst(n);
1615 ir_node *b = get_CopyB_src(n);
1618 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1619 ir_node *mem = get_CopyB_mem(n);
1620 ir_node *blk = get_nodes_block(n);
1621 turn_into_tuple(n, pn_CopyB_max);
1622 set_Tuple_pred(n, pn_CopyB_M, mem);
1623 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1624 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1625 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1628 } /* equivalent_node_CopyB */
1631 * Optimize Bounds(idx, idx, upper) into idx.
1633 static ir_node *equivalent_node_Bound(ir_node *n) {
1634 ir_node *idx = get_Bound_index(n);
1635 ir_node *pred = skip_Proj(idx);
1638 if (is_Bound(pred)) {
1640 * idx was Bounds checked in the same MacroBlock previously,
1641 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1643 ir_node *lower = get_Bound_lower(n);
1644 ir_node *upper = get_Bound_upper(n);
1645 if (get_Bound_lower(pred) == lower &&
1646 get_Bound_upper(pred) == upper &&
1647 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1649 * One could expect that we simply return the previous
1650 * Bound here. However, this would be wrong, as we could
1651 * add an exception Proj to a new location then.
1652 * So, we must turn in into a tuple.
1658 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1659 ir_node *mem = get_Bound_mem(n);
1660 ir_node *blk = get_nodes_block(n);
1661 turn_into_tuple(n, pn_Bound_max);
1662 set_Tuple_pred(n, pn_Bound_M, mem);
1663 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1664 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1665 set_Tuple_pred(n, pn_Bound_res, idx);
1668 } /* equivalent_node_Bound */
1671 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1672 * perform no actual computation, as, e.g., the Id nodes. It does not create
1673 * new nodes. It is therefore safe to free n if the node returned is not n.
1674 * If a node returns a Tuple we can not just skip it. If the size of the
1675 * in array fits, we transform n into a tuple (e.g., Div).
1677 ir_node *equivalent_node(ir_node *n) {
1678 if (n->op->ops.equivalent_node)
1679 return n->op->ops.equivalent_node(n);
1681 } /* equivalent_node */
1684 * Sets the default equivalent node operation for an ir_op_ops.
1686 * @param code the opcode for the default operation
1687 * @param ops the operations initialized
1692 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1696 ops->equivalent_node = equivalent_node_##a; \
1736 } /* firm_set_default_equivalent_node */
1739 * Returns non-zero if a node is a Phi node
1740 * with all predecessors constant.
1742 static int is_const_Phi(ir_node *n) {
1745 if (! is_Phi(n) || get_irn_arity(n) == 0)
1747 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1748 if (! is_Const(get_irn_n(n, i)))
1751 } /* is_const_Phi */
1754 * Apply an evaluator on a binop with a constant operators (and one Phi).
1756 * @param phi the Phi node
1757 * @param other the other operand
1758 * @param eval an evaluator function
1759 * @param mode the mode of the result, may be different from the mode of the Phi!
1760 * @param left if non-zero, other is the left operand, else the right
1762 * @return a new Phi node if the conversion was successful, NULL else
1764 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1769 int i, n = get_irn_arity(phi);
1771 NEW_ARR_A(void *, res, n);
1773 for (i = 0; i < n; ++i) {
1774 pred = get_irn_n(phi, i);
1775 tv = get_Const_tarval(pred);
1776 tv = eval(other, tv);
1778 if (tv == tarval_bad) {
1779 /* folding failed, bad */
1785 for (i = 0; i < n; ++i) {
1786 pred = get_irn_n(phi, i);
1787 tv = get_Const_tarval(pred);
1788 tv = eval(tv, other);
1790 if (tv == tarval_bad) {
1791 /* folding failed, bad */
1797 irg = current_ir_graph;
1798 for (i = 0; i < n; ++i) {
1799 pred = get_irn_n(phi, i);
1800 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1801 mode, res[i], get_Const_type(pred));
1803 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1804 } /* apply_binop_on_phi */
1807 * Apply an evaluator on a binop with two constant Phi.
1809 * @param a the left Phi node
1810 * @param b the right Phi node
1811 * @param eval an evaluator function
1812 * @param mode the mode of the result, may be different from the mode of the Phi!
1814 * @return a new Phi node if the conversion was successful, NULL else
1816 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1817 tarval *tv_l, *tv_r, *tv;
1823 if (get_nodes_block(a) != get_nodes_block(b))
1826 n = get_irn_arity(a);
1827 NEW_ARR_A(void *, res, n);
1829 for (i = 0; i < n; ++i) {
1830 pred = get_irn_n(a, i);
1831 tv_l = get_Const_tarval(pred);
1832 pred = get_irn_n(b, i);
1833 tv_r = get_Const_tarval(pred);
1834 tv = eval(tv_l, tv_r);
1836 if (tv == tarval_bad) {
1837 /* folding failed, bad */
1842 irg = current_ir_graph;
1843 for (i = 0; i < n; ++i) {
1844 pred = get_irn_n(a, i);
1845 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1847 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1848 } /* apply_binop_on_2_phis */
1851 * Apply an evaluator on a unop with a constant operator (a Phi).
1853 * @param phi the Phi node
1854 * @param eval an evaluator function
1856 * @return a new Phi node if the conversion was successful, NULL else
1858 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1864 int i, n = get_irn_arity(phi);
1866 NEW_ARR_A(void *, res, n);
1867 for (i = 0; i < n; ++i) {
1868 pred = get_irn_n(phi, i);
1869 tv = get_Const_tarval(pred);
1872 if (tv == tarval_bad) {
1873 /* folding failed, bad */
1878 mode = get_irn_mode(phi);
1879 irg = current_ir_graph;
1880 for (i = 0; i < n; ++i) {
1881 pred = get_irn_n(phi, i);
1882 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1883 mode, res[i], get_Const_type(pred));
1885 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1886 } /* apply_unop_on_phi */
1889 * Apply a conversion on a constant operator (a Phi).
1891 * @param phi the Phi node
1893 * @return a new Phi node if the conversion was successful, NULL else
1895 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1900 int i, n = get_irn_arity(phi);
1902 NEW_ARR_A(void *, res, n);
1903 for (i = 0; i < n; ++i) {
1904 pred = get_irn_n(phi, i);
1905 tv = get_Const_tarval(pred);
1906 tv = tarval_convert_to(tv, mode);
1908 if (tv == tarval_bad) {
1909 /* folding failed, bad */
1914 irg = current_ir_graph;
1915 for (i = 0; i < n; ++i) {
1916 pred = get_irn_n(phi, i);
1917 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1918 mode, res[i], get_Const_type(pred));
1920 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1921 } /* apply_conv_on_phi */
1924 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1925 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1926 * If possible, remove the Conv's.
1928 static ir_node *transform_node_AddSub(ir_node *n) {
1929 ir_mode *mode = get_irn_mode(n);
1931 if (mode_is_reference(mode)) {
1932 ir_node *left = get_binop_left(n);
1933 ir_node *right = get_binop_right(n);
1934 unsigned ref_bits = get_mode_size_bits(mode);
1936 if (is_Conv(left)) {
1937 ir_mode *lmode = get_irn_mode(left);
1938 unsigned bits = get_mode_size_bits(lmode);
1940 if (ref_bits == bits &&
1941 mode_is_int(lmode) &&
1942 get_mode_arithmetic(lmode) == irma_twos_complement) {
1943 ir_node *pre = get_Conv_op(left);
1944 ir_mode *pre_mode = get_irn_mode(pre);
1946 if (mode_is_int(pre_mode) &&
1947 get_mode_size_bits(pre_mode) == bits &&
1948 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1949 /* ok, this conv just changes to sign, moreover the calculation
1950 * is done with same number of bits as our address mode, so
1951 * we can ignore the conv as address calculation can be viewed
1952 * as either signed or unsigned
1954 set_binop_left(n, pre);
1959 if (is_Conv(right)) {
1960 ir_mode *rmode = get_irn_mode(right);
1961 unsigned bits = get_mode_size_bits(rmode);
1963 if (ref_bits == bits &&
1964 mode_is_int(rmode) &&
1965 get_mode_arithmetic(rmode) == irma_twos_complement) {
1966 ir_node *pre = get_Conv_op(right);
1967 ir_mode *pre_mode = get_irn_mode(pre);
1969 if (mode_is_int(pre_mode) &&
1970 get_mode_size_bits(pre_mode) == bits &&
1971 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1972 /* ok, this conv just changes to sign, moreover the calculation
1973 * is done with same number of bits as our address mode, so
1974 * we can ignore the conv as address calculation can be viewed
1975 * as either signed or unsigned
1977 set_binop_right(n, pre);
1982 /* let address arithmetic use unsigned modes */
1983 if (is_Const(right)) {
1984 ir_mode *rmode = get_irn_mode(right);
1986 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1987 /* convert a AddP(P, *s) into AddP(P, *u) */
1988 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1990 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
1991 set_binop_right(n, pre);
1996 } /* transform_node_AddSub */
1998 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2000 if (is_Const(b) && is_const_Phi(a)) { \
2001 /* check for Op(Phi, Const) */ \
2002 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2004 else if (is_Const(a) && is_const_Phi(b)) { \
2005 /* check for Op(Const, Phi) */ \
2006 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2008 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2009 /* check for Op(Phi, Phi) */ \
2010 c = apply_binop_on_2_phis(a, b, eval, mode); \
2013 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2017 #define HANDLE_UNOP_PHI(eval, a, c) \
2019 if (is_const_Phi(a)) { \
2020 /* check for Op(Phi) */ \
2021 c = apply_unop_on_phi(a, eval); \
2023 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2029 * Do the AddSub optimization, then Transform
2030 * Constant folding on Phi
2031 * Add(a,a) -> Mul(a, 2)
2032 * Add(Mul(a, x), a) -> Mul(a, x+1)
2033 * if the mode is integer or float.
2034 * Transform Add(a,-b) into Sub(a,b).
2035 * Reassociation might fold this further.
2037 static ir_node *transform_node_Add(ir_node *n) {
2039 ir_node *a, *b, *c, *oldn = n;
2041 n = transform_node_AddSub(n);
2043 a = get_Add_left(n);
2044 b = get_Add_right(n);
2046 mode = get_irn_mode(n);
2047 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2049 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2050 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2053 if (mode_is_num(mode)) {
2054 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2055 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2056 ir_node *block = get_irn_n(n, -1);
2059 get_irn_dbg_info(n),
2063 new_r_Const_long(current_ir_graph, block, mode, 2),
2065 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2070 get_irn_dbg_info(n),
2076 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2081 get_irn_dbg_info(n),
2087 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2090 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2091 /* Here we rely on constants be on the RIGHT side */
2093 ir_node *op = get_Not_op(a);
2095 if (is_Const(b) && is_Const_one(b)) {
2097 ir_node *blk = get_irn_n(n, -1);
2098 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2099 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2104 ir_node *blk = get_irn_n(n, -1);
2105 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2106 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2111 ir_node *op = get_Not_op(b);
2115 ir_node *blk = get_irn_n(n, -1);
2116 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2117 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2124 } /* transform_node_Add */
2127 * returns -cnst or NULL if impossible
2129 static ir_node *const_negate(ir_node *cnst) {
2130 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2131 dbg_info *dbgi = get_irn_dbg_info(cnst);
2132 ir_graph *irg = get_irn_irg(cnst);
2133 ir_node *block = get_nodes_block(cnst);
2134 ir_mode *mode = get_irn_mode(cnst);
2135 if (tv == tarval_bad) return NULL;
2136 return new_rd_Const(dbgi, irg, block, mode, tv);
2140 * Do the AddSub optimization, then Transform
2141 * Constant folding on Phi
2142 * Sub(0,a) -> Minus(a)
2143 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2144 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2145 * Sub(Add(a, x), x) -> a
2146 * Sub(x, Add(x, a)) -> -a
2147 * Sub(x, Const) -> Add(x, -Const)
2149 static ir_node *transform_node_Sub(ir_node *n) {
2154 n = transform_node_AddSub(n);
2156 a = get_Sub_left(n);
2157 b = get_Sub_right(n);
2159 mode = get_irn_mode(n);
2162 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2164 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2165 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2168 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2169 /* a - C -> a + (-C) */
2170 ir_node *cnst = const_negate(b);
2172 ir_node *block = get_nodes_block(n);
2173 dbg_info *dbgi = get_irn_dbg_info(n);
2174 ir_graph *irg = get_irn_irg(n);
2176 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2177 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2182 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2183 ir_graph *irg = current_ir_graph;
2184 dbg_info *dbg = get_irn_dbg_info(n);
2185 ir_node *block = get_nodes_block(n);
2186 ir_node *left = get_Minus_op(a);
2187 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2189 n = new_rd_Minus(dbg, irg, block, add, mode);
2190 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2192 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2193 ir_graph *irg = current_ir_graph;
2194 dbg_info *dbg = get_irn_dbg_info(n);
2195 ir_node *block = get_nodes_block(n);
2196 ir_node *right = get_Minus_op(b);
2198 n = new_rd_Add(dbg, irg, block, a, right, mode);
2199 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2201 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2202 ir_graph *irg = current_ir_graph;
2203 dbg_info *s_dbg = get_irn_dbg_info(b);
2204 ir_node *s_block = get_nodes_block(b);
2205 ir_node *s_left = get_Sub_right(b);
2206 ir_node *s_right = get_Sub_left(b);
2207 ir_mode *s_mode = get_irn_mode(b);
2208 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2209 dbg_info *a_dbg = get_irn_dbg_info(n);
2210 ir_node *a_block = get_nodes_block(n);
2212 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2213 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2215 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2216 ir_node *m_right = get_Mul_right(b);
2217 if (is_Const(m_right)) {
2218 ir_node *cnst2 = const_negate(m_right);
2219 if (cnst2 != NULL) {
2220 ir_graph *irg = current_ir_graph;
2221 dbg_info *m_dbg = get_irn_dbg_info(b);
2222 ir_node *m_block = get_nodes_block(b);
2223 ir_node *m_left = get_Mul_left(b);
2224 ir_mode *m_mode = get_irn_mode(b);
2225 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2226 dbg_info *a_dbg = get_irn_dbg_info(n);
2227 ir_node *a_block = get_nodes_block(n);
2229 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2230 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2236 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2237 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2239 get_irn_dbg_info(n),
2244 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2248 if (mode_wrap_around(mode)) {
2249 ir_node *left = get_Add_left(a);
2250 ir_node *right = get_Add_right(a);
2252 /* FIXME: Does the Conv's work only for two complement or generally? */
2254 if (mode != get_irn_mode(right)) {
2255 /* This Sub is an effective Cast */
2256 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2259 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2261 } else if (right == b) {
2262 if (mode != get_irn_mode(left)) {
2263 /* This Sub is an effective Cast */
2264 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2267 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2273 if (mode_wrap_around(mode)) {
2274 ir_node *left = get_Add_left(b);
2275 ir_node *right = get_Add_right(b);
2277 /* FIXME: Does the Conv's work only for two complement or generally? */
2279 ir_mode *r_mode = get_irn_mode(right);
2281 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2282 if (mode != r_mode) {
2283 /* This Sub is an effective Cast */
2284 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2286 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2288 } else if (right == a) {
2289 ir_mode *l_mode = get_irn_mode(left);
2291 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2292 if (mode != l_mode) {
2293 /* This Sub is an effective Cast */
2294 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2296 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2301 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2302 ir_mode *mode = get_irn_mode(a);
2304 if (mode == get_irn_mode(b)) {
2306 ir_node *op_a = get_Conv_op(a);
2307 ir_node *op_b = get_Conv_op(b);
2309 /* check if it's allowed to skip the conv */
2310 ma = get_irn_mode(op_a);
2311 mb = get_irn_mode(op_b);
2313 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2314 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2317 set_Sub_right(n, b);
2323 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2324 if (!is_reassoc_running() && is_Mul(a)) {
2325 ir_node *ma = get_Mul_left(a);
2326 ir_node *mb = get_Mul_right(a);
2329 ir_node *blk = get_irn_n(n, -1);
2331 get_irn_dbg_info(n),
2332 current_ir_graph, blk,
2335 get_irn_dbg_info(n),
2336 current_ir_graph, blk,
2338 new_r_Const_long(current_ir_graph, blk, mode, 1),
2341 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2343 } else if (mb == b) {
2344 ir_node *blk = get_irn_n(n, -1);
2346 get_irn_dbg_info(n),
2347 current_ir_graph, blk,
2350 get_irn_dbg_info(n),
2351 current_ir_graph, blk,
2353 new_r_Const_long(current_ir_graph, blk, mode, 1),
2356 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2360 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2361 ir_node *x = get_Sub_left(a);
2362 ir_node *y = get_Sub_right(a);
2363 ir_node *blk = get_irn_n(n, -1);
2364 ir_mode *m_b = get_irn_mode(b);
2365 ir_mode *m_y = get_irn_mode(y);
2369 /* Determine the right mode for the Add. */
2372 else if (mode_is_reference(m_b))
2374 else if (mode_is_reference(m_y))
2378 * Both modes are different but none is reference,
2379 * happens for instance in SubP(SubP(P, Iu), Is).
2380 * We have two possibilities here: Cast or ignore.
2381 * Currently we ignore this case.
2386 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2388 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2389 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2393 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2394 if (is_Const(a) && is_Not(b)) {
2395 /* c - ~X = X + (c+1) */
2396 tarval *tv = get_Const_tarval(a);
2398 tv = tarval_add(tv, get_mode_one(mode));
2399 if (tv != tarval_bad) {
2400 ir_node *blk = get_irn_n(n, -1);
2401 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2402 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2403 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2409 } /* transform_node_Sub */
2412 * Several transformation done on n*n=2n bits mul.
2413 * These transformations must be done here because new nodes may be produced.
2415 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2417 ir_node *a = get_Mul_left(n);
2418 ir_node *b = get_Mul_right(n);
2419 tarval *ta = value_of(a);
2420 tarval *tb = value_of(b);
2421 ir_mode *smode = get_irn_mode(a);
2423 if (ta == get_mode_one(smode)) {
2424 /* (L)1 * (L)b = (L)b */
2425 ir_node *blk = get_irn_n(n, -1);
2426 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2427 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2430 else if (ta == get_mode_minus_one(smode)) {
2431 /* (L)-1 * (L)b = (L)b */
2432 ir_node *blk = get_irn_n(n, -1);
2433 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2434 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2435 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2438 if (tb == get_mode_one(smode)) {
2439 /* (L)a * (L)1 = (L)a */
2440 ir_node *blk = get_irn_n(a, -1);
2441 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2442 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2445 else if (tb == get_mode_minus_one(smode)) {
2446 /* (L)a * (L)-1 = (L)-a */
2447 ir_node *blk = get_irn_n(n, -1);
2448 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2449 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2450 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2457 * Transform Mul(a,-1) into -a.
2458 * Do constant evaluation of Phi nodes.
2459 * Do architecture dependent optimizations on Mul nodes
2461 static ir_node *transform_node_Mul(ir_node *n) {
2462 ir_node *c, *oldn = n;
2463 ir_mode *mode = get_irn_mode(n);
2464 ir_node *a = get_Mul_left(n);
2465 ir_node *b = get_Mul_right(n);
2467 if (is_Bad(a) || is_Bad(b))
2470 if (mode != get_irn_mode(a))
2471 return transform_node_Mul2n(n, mode);
2473 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2475 if (mode_is_signed(mode)) {
2478 if (value_of(a) == get_mode_minus_one(mode))
2480 else if (value_of(b) == get_mode_minus_one(mode))
2483 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2484 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2489 if (is_Const(b)) { /* (-a) * const -> a * -const */
2490 ir_node *cnst = const_negate(b);
2492 dbg_info *dbgi = get_irn_dbg_info(n);
2493 ir_node *block = get_nodes_block(n);
2494 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2495 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2498 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2499 dbg_info *dbgi = get_irn_dbg_info(n);
2500 ir_node *block = get_nodes_block(n);
2501 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2502 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2504 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2505 ir_node *sub_l = get_Sub_left(b);
2506 ir_node *sub_r = get_Sub_right(b);
2507 dbg_info *dbgi = get_irn_dbg_info(n);
2508 ir_graph *irg = current_ir_graph;
2509 ir_node *block = get_nodes_block(n);
2510 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2511 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2512 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2515 } else if (is_Minus(b)) {
2516 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2517 ir_node *sub_l = get_Sub_left(a);
2518 ir_node *sub_r = get_Sub_right(a);
2519 dbg_info *dbgi = get_irn_dbg_info(n);
2520 ir_graph *irg = current_ir_graph;
2521 ir_node *block = get_nodes_block(n);
2522 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2523 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2524 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2528 if (get_mode_arithmetic(mode) == irma_ieee754) {
2530 tarval *tv = get_Const_tarval(a);
2531 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2532 /* 2.0 * b = b + b */
2533 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2534 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2538 else if (is_Const(b)) {
2539 tarval *tv = get_Const_tarval(b);
2540 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2541 /* a * 2.0 = a + a */
2542 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2543 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2548 return arch_dep_replace_mul_with_shifts(n);
2549 } /* transform_node_Mul */
2552 * Transform a Div Node.
2554 static ir_node *transform_node_Div(ir_node *n) {
2555 ir_mode *mode = get_Div_resmode(n);
2556 ir_node *a = get_Div_left(n);
2557 ir_node *b = get_Div_right(n);
2561 if (is_Const(b) && is_const_Phi(a)) {
2562 /* check for Div(Phi, Const) */
2563 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2565 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2569 else if (is_Const(a) && is_const_Phi(b)) {
2570 /* check for Div(Const, Phi) */
2571 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2573 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2577 else if (is_const_Phi(a) && is_const_Phi(b)) {
2578 /* check for Div(Phi, Phi) */
2579 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2581 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2588 if (tv != tarval_bad) {
2589 value = new_Const(get_tarval_mode(tv), tv);
2591 DBG_OPT_CSTEVAL(n, value);
2594 ir_node *a = get_Div_left(n);
2595 ir_node *b = get_Div_right(n);
2598 if (a == b && value_not_zero(a, &dummy)) {
2599 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2600 value = new_Const(mode, get_mode_one(mode));
2601 DBG_OPT_CSTEVAL(n, value);
2604 if (mode_is_signed(mode) && is_Const(b)) {
2605 tarval *tv = get_Const_tarval(b);
2607 if (tv == get_mode_minus_one(mode)) {
2609 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2610 DBG_OPT_CSTEVAL(n, value);
2614 /* Try architecture dependent optimization */
2615 value = arch_dep_replace_div_by_const(n);
2623 /* Turn Div into a tuple (mem, jmp, bad, value) */
2624 mem = get_Div_mem(n);
2625 blk = get_irn_n(n, -1);
2627 /* skip a potential Pin */
2629 mem = get_Pin_op(mem);
2630 turn_into_tuple(n, pn_Div_max);
2631 set_Tuple_pred(n, pn_Div_M, mem);
2632 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2633 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2634 set_Tuple_pred(n, pn_Div_res, value);
2637 } /* transform_node_Div */
2640 * Transform a Mod node.
2642 static ir_node *transform_node_Mod(ir_node *n) {
2643 ir_mode *mode = get_Mod_resmode(n);
2644 ir_node *a = get_Mod_left(n);
2645 ir_node *b = get_Mod_right(n);
2649 if (is_Const(b) && is_const_Phi(a)) {
2650 /* check for Div(Phi, Const) */
2651 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2653 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2657 else if (is_Const(a) && is_const_Phi(b)) {
2658 /* check for Div(Const, Phi) */
2659 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2661 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2665 else if (is_const_Phi(a) && is_const_Phi(b)) {
2666 /* check for Div(Phi, Phi) */
2667 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2669 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2676 if (tv != tarval_bad) {
2677 value = new_Const(get_tarval_mode(tv), tv);
2679 DBG_OPT_CSTEVAL(n, value);
2682 ir_node *a = get_Mod_left(n);
2683 ir_node *b = get_Mod_right(n);
2686 if (a == b && value_not_zero(a, &dummy)) {
2687 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2688 value = new_Const(mode, get_mode_null(mode));
2689 DBG_OPT_CSTEVAL(n, value);
2692 if (mode_is_signed(mode) && is_Const(b)) {
2693 tarval *tv = get_Const_tarval(b);
2695 if (tv == get_mode_minus_one(mode)) {
2697 value = new_Const(mode, get_mode_null(mode));
2698 DBG_OPT_CSTEVAL(n, value);
2702 /* Try architecture dependent optimization */
2703 value = arch_dep_replace_mod_by_const(n);
2711 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2712 mem = get_Mod_mem(n);
2713 blk = get_irn_n(n, -1);
2715 /* skip a potential Pin */
2717 mem = get_Pin_op(mem);
2718 turn_into_tuple(n, pn_Mod_max);
2719 set_Tuple_pred(n, pn_Mod_M, mem);
2720 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2721 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2722 set_Tuple_pred(n, pn_Mod_res, value);
2725 } /* transform_node_Mod */
2728 * Transform a DivMod node.
2730 static ir_node *transform_node_DivMod(ir_node *n) {
2732 ir_node *a = get_DivMod_left(n);
2733 ir_node *b = get_DivMod_right(n);
2734 ir_mode *mode = get_DivMod_resmode(n);
2739 if (is_Const(b) && is_const_Phi(a)) {
2740 /* check for Div(Phi, Const) */
2741 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2742 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2744 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2745 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2749 else if (is_Const(a) && is_const_Phi(b)) {
2750 /* check for Div(Const, Phi) */
2751 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2752 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2754 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2755 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2759 else if (is_const_Phi(a) && is_const_Phi(b)) {
2760 /* check for Div(Phi, Phi) */
2761 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2762 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2764 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2765 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2772 if (tb != tarval_bad) {
2773 if (tb == get_mode_one(get_tarval_mode(tb))) {
2775 vb = new_Const(mode, get_mode_null(mode));
2776 DBG_OPT_CSTEVAL(n, vb);
2778 } else if (ta != tarval_bad) {
2779 tarval *resa, *resb;
2780 resa = tarval_div(ta, tb);
2781 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2782 Jmp for X result!? */
2783 resb = tarval_mod(ta, tb);
2784 if (resb == tarval_bad) return n; /* Causes exception! */
2785 va = new_Const(mode, resa);
2786 vb = new_Const(mode, resb);
2787 DBG_OPT_CSTEVAL(n, va);
2788 DBG_OPT_CSTEVAL(n, vb);
2790 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2791 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2792 vb = new_Const(mode, get_mode_null(mode));
2793 DBG_OPT_CSTEVAL(n, va);
2794 DBG_OPT_CSTEVAL(n, vb);
2796 } else { /* Try architecture dependent optimization */
2799 arch_dep_replace_divmod_by_const(&va, &vb, n);
2800 evaluated = va != NULL;
2802 } else if (a == b) {
2803 if (value_not_zero(a, &dummy)) {
2805 va = new_Const(mode, get_mode_one(mode));
2806 vb = new_Const(mode, get_mode_null(mode));
2807 DBG_OPT_CSTEVAL(n, va);
2808 DBG_OPT_CSTEVAL(n, vb);
2811 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2814 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2815 /* 0 / non-Const = 0 */
2820 if (evaluated) { /* replace by tuple */
2824 mem = get_DivMod_mem(n);
2825 /* skip a potential Pin */
2827 mem = get_Pin_op(mem);
2829 blk = get_irn_n(n, -1);
2830 turn_into_tuple(n, pn_DivMod_max);
2831 set_Tuple_pred(n, pn_DivMod_M, mem);
2832 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2833 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2834 set_Tuple_pred(n, pn_DivMod_res_div, va);
2835 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2839 } /* transform_node_DivMod */
2842 * Optimize x / c to x * (1/c)
2844 static ir_node *transform_node_Quot(ir_node *n) {
2845 ir_mode *mode = get_Quot_resmode(n);
2848 if (get_mode_arithmetic(mode) == irma_ieee754) {
2849 ir_node *b = get_Quot_right(n);
2852 tarval *tv = get_Const_tarval(b);
2856 * Floating point constant folding might be disabled here to
2858 * However, as we check for exact result, doing it is safe.
2861 rem = tarval_enable_fp_ops(1);
2862 tv = tarval_quo(get_mode_one(mode), tv);
2863 (void)tarval_enable_fp_ops(rem);
2865 /* Do the transformation if the result is either exact or we are not
2866 using strict rules. */
2867 if (tv != tarval_bad &&
2868 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2869 ir_node *blk = get_irn_n(n, -1);
2870 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2871 ir_node *a = get_Quot_left(n);
2872 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2873 ir_node *mem = get_Quot_mem(n);
2875 /* skip a potential Pin */
2877 mem = get_Pin_op(mem);
2878 turn_into_tuple(n, pn_Quot_max);
2879 set_Tuple_pred(n, pn_Quot_M, mem);
2880 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2881 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2882 set_Tuple_pred(n, pn_Quot_res, m);
2883 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2888 } /* transform_node_Quot */
2891 * Optimize Abs(x) into x if x is Confirmed >= 0
2892 * Optimize Abs(x) into -x if x is Confirmed <= 0
2893 * Optimize Abs(-x) int Abs(x)
2895 static ir_node *transform_node_Abs(ir_node *n) {
2896 ir_node *c, *oldn = n;
2897 ir_node *a = get_Abs_op(n);
2900 HANDLE_UNOP_PHI(tarval_abs, a, c);
2902 switch (classify_value_sign(a)) {
2903 case value_classified_negative:
2904 mode = get_irn_mode(n);
2907 * We can replace the Abs by -x here.
2908 * We even could add a new Confirm here
2909 * (if not twos complement)
2911 * Note that -x would create a new node, so we could
2912 * not run it in the equivalent_node() context.
2914 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2915 get_nodes_block(n), a, mode);
2917 DBG_OPT_CONFIRM(oldn, n);
2919 case value_classified_positive:
2920 /* n is positive, Abs is not needed */
2923 DBG_OPT_CONFIRM(oldn, n);
2929 /* Abs(-x) = Abs(x) */
2930 mode = get_irn_mode(n);
2931 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2932 get_nodes_block(n), get_Minus_op(a), mode);
2933 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2937 } /* transform_node_Abs */
2940 * Transform a Cond node.
2942 * Replace the Cond by a Jmp if it branches on a constant
2945 static ir_node *transform_node_Cond(ir_node *n) {
2948 ir_node *a = get_Cond_selector(n);
2949 tarval *ta = value_of(a);
2951 /* we need block info which is not available in floating irgs */
2952 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2955 if ((ta != tarval_bad) &&
2956 (get_irn_mode(a) == mode_b) &&
2957 (get_opt_unreachable_code())) {
2958 /* It's a boolean Cond, branching on a boolean constant.
2959 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2960 ir_node *blk = get_nodes_block(n);
2961 jmp = new_r_Jmp(current_ir_graph, blk);
2962 turn_into_tuple(n, pn_Cond_max);
2963 if (ta == tarval_b_true) {
2964 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2965 set_Tuple_pred(n, pn_Cond_true, jmp);
2967 set_Tuple_pred(n, pn_Cond_false, jmp);
2968 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2970 /* We might generate an endless loop, so keep it alive. */
2971 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2974 } /* transform_node_Cond */
2977 * Prototype of a recursive transform function
2978 * for bitwise distributive transformations.
2980 typedef ir_node* (*recursive_transform)(ir_node *n);
2983 * makes use of distributive laws for and, or, eor
2984 * and(a OP c, b OP c) -> and(a, b) OP c
2985 * note, might return a different op than n
2987 static ir_node *transform_bitwise_distributive(ir_node *n,
2988 recursive_transform trans_func)
2991 ir_node *a = get_binop_left(n);
2992 ir_node *b = get_binop_right(n);
2993 ir_op *op = get_irn_op(a);
2994 ir_op *op_root = get_irn_op(n);
2996 if(op != get_irn_op(b))
2999 if (op == op_Conv) {
3000 ir_node *a_op = get_Conv_op(a);
3001 ir_node *b_op = get_Conv_op(b);
3002 ir_mode *a_mode = get_irn_mode(a_op);
3003 ir_mode *b_mode = get_irn_mode(b_op);
3004 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3005 ir_node *blk = get_irn_n(n, -1);
3008 set_binop_left(n, a_op);
3009 set_binop_right(n, b_op);
3010 set_irn_mode(n, a_mode);
3012 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3014 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3020 /* nothing to gain here */
3024 if (op == op_Shrs || op == op_Shr || op == op_Shl
3025 || op == op_And || op == op_Or || op == op_Eor) {
3026 ir_node *a_left = get_binop_left(a);
3027 ir_node *a_right = get_binop_right(a);
3028 ir_node *b_left = get_binop_left(b);
3029 ir_node *b_right = get_binop_right(b);
3031 ir_node *op1 = NULL;
3032 ir_node *op2 = NULL;
3034 if (is_op_commutative(op)) {
3035 if (a_left == b_left) {
3039 } else if(a_left == b_right) {
3043 } else if(a_right == b_left) {
3049 if(a_right == b_right) {
3056 /* (a sop c) & (b sop c) => (a & b) sop c */
3057 ir_node *blk = get_irn_n(n, -1);
3059 ir_node *new_n = exact_copy(n);
3060 set_binop_left(new_n, op1);
3061 set_binop_right(new_n, op2);
3062 new_n = trans_func(new_n);
3064 if(op_root == op_Eor && op == op_Or) {
3065 dbg_info *dbgi = get_irn_dbg_info(n);
3066 ir_graph *irg = current_ir_graph;
3067 ir_mode *mode = get_irn_mode(c);
3069 c = new_rd_Not(dbgi, irg, blk, c, mode);
3070 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3073 set_nodes_block(n, blk);
3074 set_binop_left(n, new_n);
3075 set_binop_right(n, c);
3076 add_identities(current_ir_graph->value_table, n);
3079 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3090 static ir_node *transform_node_And(ir_node *n) {
3091 ir_node *c, *oldn = n;
3092 ir_node *a = get_And_left(n);
3093 ir_node *b = get_And_right(n);
3096 mode = get_irn_mode(n);
3097 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3099 /* we can evaluate 2 Projs of the same Cmp */
3100 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3101 ir_node *pred_a = get_Proj_pred(a);
3102 ir_node *pred_b = get_Proj_pred(b);
3103 if (pred_a == pred_b) {
3104 dbg_info *dbgi = get_irn_dbg_info(n);
3105 ir_node *block = get_nodes_block(pred_a);
3106 pn_Cmp pn_a = get_Proj_proj(a);
3107 pn_Cmp pn_b = get_Proj_proj(b);
3108 /* yes, we can simply calculate with pncs */
3109 pn_Cmp new_pnc = pn_a & pn_b;
3111 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3116 ir_node *op = get_Not_op(b);
3118 ir_node *ba = get_And_left(op);
3119 ir_node *bb = get_And_right(op);
3121 /* it's enough to test the following cases due to normalization! */
3122 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3123 /* (a|b) & ~(a&b) = a^b */
3124 ir_node *block = get_nodes_block(n);
3126 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3127 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3135 ir_node *op = get_Not_op(a);
3137 ir_node *aa = get_And_left(op);
3138 ir_node *ab = get_And_right(op);
3140 /* it's enough to test the following cases due to normalization! */
3141 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3142 /* (a|b) & ~(a&b) = a^b */
3143 ir_node *block = get_nodes_block(n);
3145 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3146 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3153 ir_node *al = get_Eor_left(a);
3154 ir_node *ar = get_Eor_right(a);
3157 /* (b ^ a) & b -> ~a & b */
3158 dbg_info *dbg = get_irn_dbg_info(n);
3159 ir_node *block = get_nodes_block(n);
3161 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3162 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3163 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3167 /* (a ^ b) & b -> ~a & b */
3168 dbg_info *dbg = get_irn_dbg_info(n);
3169 ir_node *block = get_nodes_block(n);
3171 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3172 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3173 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3178 ir_node *bl = get_Eor_left(b);
3179 ir_node *br = get_Eor_right(b);
3182 /* a & (a ^ b) -> a & ~b */
3183 dbg_info *dbg = get_irn_dbg_info(n);
3184 ir_node *block = get_nodes_block(n);
3186 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3187 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3188 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3192 /* a & (b ^ a) -> a & ~b */
3193 dbg_info *dbg = get_irn_dbg_info(n);
3194 ir_node *block = get_nodes_block(n);
3196 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3197 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3198 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3202 if (is_Not(a) && is_Not(b)) {
3203 /* ~a & ~b = ~(a|b) */
3204 ir_node *block = get_nodes_block(n);
3205 ir_mode *mode = get_irn_mode(n);
3209 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3210 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3211 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3215 n = transform_bitwise_distributive(n, transform_node_And);
3218 } /* transform_node_And */
3223 static ir_node *transform_node_Eor(ir_node *n) {
3224 ir_node *c, *oldn = n;
3225 ir_node *a = get_Eor_left(n);
3226 ir_node *b = get_Eor_right(n);
3227 ir_mode *mode = get_irn_mode(n);
3229 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3231 /* we can evaluate 2 Projs of the same Cmp */
3232 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3233 ir_node *pred_a = get_Proj_pred(a);
3234 ir_node *pred_b = get_Proj_pred(b);
3235 if(pred_a == pred_b) {
3236 dbg_info *dbgi = get_irn_dbg_info(n);
3237 ir_node *block = get_nodes_block(pred_a);
3238 pn_Cmp pn_a = get_Proj_proj(a);
3239 pn_Cmp pn_b = get_Proj_proj(b);
3240 /* yes, we can simply calculate with pncs */
3241 pn_Cmp new_pnc = pn_a ^ pn_b;
3243 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3250 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3251 mode, get_mode_null(mode));
3252 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3253 } else if (mode == mode_b &&
3255 is_Const(b) && is_Const_one(b) &&
3256 is_Cmp(get_Proj_pred(a))) {
3257 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3258 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3259 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3262 } else if (is_Const(b)) {
3263 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3264 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3265 ir_node *not_op = get_Not_op(a);
3266 dbg_info *dbg = get_irn_dbg_info(n);
3267 ir_graph *irg = current_ir_graph;
3268 ir_node *block = get_nodes_block(n);
3269 ir_mode *mode = get_irn_mode(n);
3270 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3272 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3273 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3274 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3277 n = transform_bitwise_distributive(n, transform_node_Eor);
3281 } /* transform_node_Eor */
3286 static ir_node *transform_node_Not(ir_node *n) {
3287 ir_node *c, *oldn = n;
3288 ir_node *a = get_Not_op(n);
3289 ir_mode *mode = get_irn_mode(n);
3291 HANDLE_UNOP_PHI(tarval_not,a,c);
3293 /* check for a boolean Not */
3294 if (mode == mode_b &&
3296 is_Cmp(get_Proj_pred(a))) {
3297 /* We negate a Cmp. The Cmp has the negated result anyways! */
3298 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3299 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3300 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3304 ir_node *eor_b = get_Eor_right(a);
3305 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3306 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3307 ir_node *eor_a = get_Eor_left(a);
3308 dbg_info *dbg = get_irn_dbg_info(n);
3309 ir_graph *irg = current_ir_graph;
3310 ir_node *block = get_nodes_block(n);
3311 ir_mode *mode = get_irn_mode(n);
3312 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3316 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3317 if (is_Minus(a)) { /* ~-x -> x + -1 */
3318 dbg_info *dbg = get_irn_dbg_info(n);
3319 ir_graph *irg = current_ir_graph;
3320 ir_node *block = get_nodes_block(n);
3321 ir_node *add_l = get_Minus_op(a);
3322 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3323 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3324 } else if (is_Add(a)) {
3325 ir_node *add_r = get_Add_right(a);
3326 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3327 /* ~(x + -1) = -x */
3328 ir_node *op = get_Add_left(a);
3329 ir_node *blk = get_irn_n(n, -1);
3330 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3331 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3336 } /* transform_node_Not */
3339 * Transform a Minus.
3343 * -(a >>u (size-1)) = a >>s (size-1)
3344 * -(a >>s (size-1)) = a >>u (size-1)
3345 * -(a * const) -> a * -const
3347 static ir_node *transform_node_Minus(ir_node *n) {
3348 ir_node *c, *oldn = n;
3349 ir_node *a = get_Minus_op(n);
3352 HANDLE_UNOP_PHI(tarval_neg,a,c);
3354 mode = get_irn_mode(a);
3355 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3356 /* the following rules are only to twos-complement */
3359 ir_node *op = get_Not_op(a);
3360 tarval *tv = get_mode_one(mode);
3361 ir_node *blk = get_irn_n(n, -1);
3362 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3363 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3364 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3368 ir_node *c = get_Shr_right(a);
3371 tarval *tv = get_Const_tarval(c);
3373 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3374 /* -(a >>u (size-1)) = a >>s (size-1) */
3375 ir_node *v = get_Shr_left(a);
3377 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3378 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3384 ir_node *c = get_Shrs_right(a);
3387 tarval *tv = get_Const_tarval(c);
3389 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3390 /* -(a >>s (size-1)) = a >>u (size-1) */
3391 ir_node *v = get_Shrs_left(a);
3393 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3394 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3401 /* - (a-b) = b - a */
3402 ir_node *la = get_Sub_left(a);
3403 ir_node *ra = get_Sub_right(a);
3404 ir_node *blk = get_irn_n(n, -1);
3406 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3407 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3411 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3412 ir_node *mul_l = get_Mul_left(a);
3413 ir_node *mul_r = get_Mul_right(a);
3414 if (is_Const(mul_r)) {
3415 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3416 if(tv != tarval_bad) {
3417 ir_node *cnst = new_Const(mode, tv);
3418 dbg_info *dbg = get_irn_dbg_info(a);
3419 ir_graph *irg = current_ir_graph;
3420 ir_node *block = get_nodes_block(a);
3421 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3422 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3429 } /* transform_node_Minus */
3432 * Transform a Cast_type(Const) into a new Const_type
3434 static ir_node *transform_node_Cast(ir_node *n) {
3436 ir_node *pred = get_Cast_op(n);
3437 ir_type *tp = get_irn_type(n);
3439 if (is_Const(pred) && get_Const_type(pred) != tp) {
3440 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3441 get_Const_tarval(pred), tp);
3442 DBG_OPT_CSTEVAL(oldn, n);
3443 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3444 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3445 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3446 DBG_OPT_CSTEVAL(oldn, n);
3450 } /* transform_node_Cast */
3453 * Transform a Proj(Div) with a non-zero value.
3454 * Removes the exceptions and routes the memory to the NoMem node.
3456 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3457 ir_node *div = get_Proj_pred(proj);
3458 ir_node *b = get_Div_right(div);
3459 ir_node *confirm, *res, *new_mem;
3462 if (value_not_zero(b, &confirm)) {
3463 /* div(x, y) && y != 0 */
3464 if (confirm == NULL) {
3465 /* we are sure we have a Const != 0 */
3466 new_mem = get_Div_mem(div);
3467 if (is_Pin(new_mem))
3468 new_mem = get_Pin_op(new_mem);
3469 set_Div_mem(div, new_mem);
3470 set_irn_pinned(div, op_pin_state_floats);
3473 proj_nr = get_Proj_proj(proj);
3475 case pn_Div_X_regular:
3476 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3478 case pn_Div_X_except:
3479 /* we found an exception handler, remove it */
3480 DBG_OPT_EXC_REM(proj);
3484 res = get_Div_mem(div);
3485 new_mem = get_irg_no_mem(current_ir_graph);
3488 /* This node can only float up to the Confirm block */
3489 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3491 set_irn_pinned(div, op_pin_state_floats);
3492 /* this is a Div without exception, we can remove the memory edge */
3493 set_Div_mem(div, new_mem);
3498 } /* transform_node_Proj_Div */
3501 * Transform a Proj(Mod) with a non-zero value.
3502 * Removes the exceptions and routes the memory to the NoMem node.
3504 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3505 ir_node *mod = get_Proj_pred(proj);
3506 ir_node *b = get_Mod_right(mod);
3507 ir_node *confirm, *res, *new_mem;
3510 if (value_not_zero(b, &confirm)) {
3511 /* mod(x, y) && y != 0 */
3512 proj_nr = get_Proj_proj(proj);
3514 if (confirm == NULL) {
3515 /* we are sure we have a Const != 0 */
3516 new_mem = get_Mod_mem(mod);
3517 if (is_Pin(new_mem))
3518 new_mem = get_Pin_op(new_mem);
3519 set_Mod_mem(mod, new_mem);
3520 set_irn_pinned(mod, op_pin_state_floats);
3525 case pn_Mod_X_regular:
3526 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3528 case pn_Mod_X_except:
3529 /* we found an exception handler, remove it */
3530 DBG_OPT_EXC_REM(proj);
3534 res = get_Mod_mem(mod);
3535 new_mem = get_irg_no_mem(current_ir_graph);
3538 /* This node can only float up to the Confirm block */
3539 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3541 /* this is a Mod without exception, we can remove the memory edge */
3542 set_Mod_mem(mod, new_mem);
3545 if (get_Mod_left(mod) == b) {
3546 /* a % a = 0 if a != 0 */
3547 ir_mode *mode = get_irn_mode(proj);
3548 ir_node *res = new_Const(mode, get_mode_null(mode));
3550 DBG_OPT_CSTEVAL(mod, res);
3556 } /* transform_node_Proj_Mod */
3559 * Transform a Proj(DivMod) with a non-zero value.
3560 * Removes the exceptions and routes the memory to the NoMem node.
3562 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3563 ir_node *divmod = get_Proj_pred(proj);
3564 ir_node *b = get_DivMod_right(divmod);
3565 ir_node *confirm, *res, *new_mem;
3568 if (value_not_zero(b, &confirm)) {
3569 /* DivMod(x, y) && y != 0 */
3570 proj_nr = get_Proj_proj(proj);
3572 if (confirm == NULL) {
3573 /* we are sure we have a Const != 0 */
3574 new_mem = get_DivMod_mem(divmod);
3575 if (is_Pin(new_mem))
3576 new_mem = get_Pin_op(new_mem);
3577 set_DivMod_mem(divmod, new_mem);
3578 set_irn_pinned(divmod, op_pin_state_floats);
3583 case pn_DivMod_X_regular:
3584 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3586 case pn_DivMod_X_except:
3587 /* we found an exception handler, remove it */
3588 DBG_OPT_EXC_REM(proj);
3592 res = get_DivMod_mem(divmod);
3593 new_mem = get_irg_no_mem(current_ir_graph);
3596 /* This node can only float up to the Confirm block */
3597 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3599 /* this is a DivMod without exception, we can remove the memory edge */
3600 set_DivMod_mem(divmod, new_mem);
3603 case pn_DivMod_res_mod:
3604 if (get_DivMod_left(divmod) == b) {
3605 /* a % a = 0 if a != 0 */
3606 ir_mode *mode = get_irn_mode(proj);
3607 ir_node *res = new_Const(mode, get_mode_null(mode));
3609 DBG_OPT_CSTEVAL(divmod, res);
3615 } /* transform_node_Proj_DivMod */
3618 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3620 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3621 if (get_opt_unreachable_code()) {
3622 ir_node *n = get_Proj_pred(proj);
3623 ir_node *b = get_Cond_selector(n);
3625 if (mode_is_int(get_irn_mode(b))) {
3626 tarval *tb = value_of(b);
3628 if (tb != tarval_bad) {
3629 /* we have a constant switch */
3630 long num = get_Proj_proj(proj);
3632 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3633 if (get_tarval_long(tb) == num) {
3634 /* Do NOT create a jump here, or we will have 2 control flow ops
3635 * in a block. This case is optimized away in optimize_cf(). */
3638 /* this case will NEVER be taken, kill it */
3646 } /* transform_node_Proj_Cond */
3649 * Create a 0 constant of given mode.
3651 static ir_node *create_zero_const(ir_mode *mode) {
3652 tarval *tv = get_mode_null(mode);
3653 ir_node *cnst = new_Const(mode, tv);
3658 /* the order of the values is important! */
3659 typedef enum const_class {
3665 static const_class classify_const(const ir_node* n)
3667 if (is_Const(n)) return const_const;
3668 if (is_irn_constlike(n)) return const_like;
3673 * Determines whether r is more constlike or has a larger index (in that order)
3676 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3678 const const_class l_order = classify_const(l);
3679 const const_class r_order = classify_const(r);
3681 l_order > r_order ||
3682 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3686 * Normalizes and optimizes Cmp nodes.
3688 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3689 ir_node *n = get_Proj_pred(proj);
3690 ir_node *left = get_Cmp_left(n);
3691 ir_node *right = get_Cmp_right(n);
3694 ir_mode *mode = NULL;
3695 long proj_nr = get_Proj_proj(proj);
3697 /* we can evaluate some cases directly */
3700 return new_Const(mode_b, get_tarval_b_false());
3702 return new_Const(mode_b, get_tarval_b_true());
3704 if (!mode_is_float(get_irn_mode(left)))
3705 return new_Const(mode_b, get_tarval_b_true());
3711 /* remove Casts of both sides */
3712 left = skip_Cast(left);
3713 right = skip_Cast(right);
3715 /* Remove unnecessary conversions */
3716 /* TODO handle constants */
3717 if (is_Conv(left) && is_Conv(right)) {
3718 ir_mode *mode = get_irn_mode(left);
3719 ir_node *op_left = get_Conv_op(left);
3720 ir_node *op_right = get_Conv_op(right);
3721 ir_mode *mode_left = get_irn_mode(op_left);
3722 ir_mode *mode_right = get_irn_mode(op_right);
3724 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3725 && mode_left != mode_b && mode_right != mode_b) {
3726 ir_graph *irg = current_ir_graph;
3727 ir_node *block = get_nodes_block(n);
3729 if (mode_left == mode_right) {
3733 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3734 } else if (smaller_mode(mode_left, mode_right)) {
3735 left = new_r_Conv(irg, block, op_left, mode_right);
3738 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3739 } else if (smaller_mode(mode_right, mode_left)) {
3741 right = new_r_Conv(irg, block, op_right, mode_left);
3743 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3748 /* remove operation on both sides if possible */
3749 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3751 * The following operations are NOT safe for floating point operations, for instance
3752 * 1.0 + inf == 2.0 + inf, =/=> x == y
3754 if (mode_is_int(get_irn_mode(left))) {
3755 unsigned lop = get_irn_opcode(left);
3757 if (lop == get_irn_opcode(right)) {
3758 ir_node *ll, *lr, *rl, *rr;
3760 /* same operation on both sides, try to remove */
3764 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3765 left = get_unop_op(left);
3766 right = get_unop_op(right);
3768 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3771 ll = get_Add_left(left);
3772 lr = get_Add_right(left);
3773 rl = get_Add_left(right);
3774 rr = get_Add_right(right);
3777 /* X + a CMP X + b ==> a CMP b */
3781 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3782 } else if (ll == rr) {
3783 /* X + a CMP b + X ==> a CMP b */
3787 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3788 } else if (lr == rl) {
3789 /* a + X CMP X + b ==> a CMP b */
3793 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3794 } else if (lr == rr) {
3795 /* a + X CMP b + X ==> a CMP b */
3799 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3803 ll = get_Sub_left(left);
3804 lr = get_Sub_right(left);
3805 rl = get_Sub_left(right);
3806 rr = get_Sub_right(right);
3809 /* X - a CMP X - b ==> a CMP b */
3813 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3814 } else if (lr == rr) {
3815 /* a - X CMP b - X ==> a CMP b */
3819 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3823 if (get_Rot_right(left) == get_Rot_right(right)) {
3824 /* a ROT X CMP b ROT X ==> a CMP b */
3825 left = get_Rot_left(left);
3826 right = get_Rot_left(right);
3828 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3836 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3837 if (is_Add(left) || is_Sub(left)) {
3838 ir_node *ll = get_binop_left(left);
3839 ir_node *lr = get_binop_right(left);
3841 if (lr == right && is_Add(left)) {
3848 right = create_zero_const(get_irn_mode(left));
3850 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3853 if (is_Add(right) || is_Sub(right)) {
3854 ir_node *rl = get_binop_left(right);
3855 ir_node *rr = get_binop_right(right);
3857 if (rr == left && is_Add(right)) {
3864 right = create_zero_const(get_irn_mode(left));
3866 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3869 } /* mode_is_int(...) */
3870 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3872 /* replace mode_b compares with ands/ors */
3873 if (get_irn_mode(left) == mode_b) {
3874 ir_graph *irg = current_ir_graph;
3875 ir_node *block = get_nodes_block(n);
3879 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3880 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3881 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3882 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3883 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3884 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3885 default: bres = NULL;
3888 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3894 * First step: normalize the compare op
3895 * by placing the constant on the right side
3896 * or moving the lower address node to the left.
3898 if (!operands_are_normalized(left, right)) {
3904 proj_nr = get_inversed_pnc(proj_nr);
3909 * Second step: Try to reduce the magnitude
3910 * of a constant. This may help to generate better code
3911 * later and may help to normalize more compares.
3912 * Of course this is only possible for integer values.
3914 if (is_Const(right)) {
3915 mode = get_irn_mode(right);
3916 tv = get_Const_tarval(right);
3918 /* TODO extend to arbitrary constants */
3919 if (is_Conv(left) && tarval_is_null(tv)) {
3920 ir_node *op = get_Conv_op(left);
3921 ir_mode *op_mode = get_irn_mode(op);
3924 * UpConv(x) REL 0 ==> x REL 0
3926 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3927 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3928 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3929 tv = get_mode_null(op_mode);
3933 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3937 if (tv != tarval_bad) {
3938 /* the following optimization is possible on modes without Overflow
3939 * on Unary Minus or on == and !=:
3940 * -a CMP c ==> a swap(CMP) -c
3942 * Beware: for two-complement Overflow may occur, so only == and != can
3943 * be optimized, see this:
3944 * -MININT < 0 =/=> MININT > 0 !!!
3946 if (is_Minus(left) &&
3947 (!mode_overflow_on_unary_Minus(mode) ||
3948 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3949 tv = tarval_neg(tv);
3951 if (tv != tarval_bad) {
3952 left = get_Minus_op(left);
3953 proj_nr = get_inversed_pnc(proj_nr);
3955 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3957 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3958 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3959 tv = tarval_not(tv);
3961 if (tv != tarval_bad) {
3962 left = get_Not_op(left);
3964 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3968 /* for integer modes, we have more */
3969 if (mode_is_int(mode)) {
3970 /* Ne includes Unordered which is not possible on integers.
3971 * However, frontends often use this wrong, so fix it here */
3972 if (proj_nr & pn_Cmp_Uo) {
3973 proj_nr &= ~pn_Cmp_Uo;
3974 set_Proj_proj(proj, proj_nr);
3977 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3978 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3979 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3980 tv = tarval_sub(tv, get_mode_one(mode));
3982 if (tv != tarval_bad) {
3983 proj_nr ^= pn_Cmp_Eq;
3985 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3988 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3989 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3990 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3991 tv = tarval_add(tv, get_mode_one(mode));
3993 if (tv != tarval_bad) {
3994 proj_nr ^= pn_Cmp_Eq;
3996 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4000 /* the following reassociations work only for == and != */
4001 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4003 #if 0 /* Might be not that good in general */
4004 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4005 if (tarval_is_null(tv) && is_Sub(left)) {
4006 right = get_Sub_right(left);
4007 left = get_Sub_left(left);
4009 tv = value_of(right);
4011 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4015 if (tv != tarval_bad) {
4016 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4018 ir_node *c1 = get_Sub_right(left);
4019 tarval *tv2 = value_of(c1);
4021 if (tv2 != tarval_bad) {
4022 tv2 = tarval_add(tv, value_of(c1));
4024 if (tv2 != tarval_bad) {
4025 left = get_Sub_left(left);
4028 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4032 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4033 else if (is_Add(left)) {
4034 ir_node *a_l = get_Add_left(left);
4035 ir_node *a_r = get_Add_right(left);
4039 if (is_Const(a_l)) {
4041 tv2 = value_of(a_l);
4044 tv2 = value_of(a_r);
4047 if (tv2 != tarval_bad) {
4048 tv2 = tarval_sub(tv, tv2);
4050 if (tv2 != tarval_bad) {
4054 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4058 /* -a == c ==> a == -c, -a != c ==> a != -c */
4059 else if (is_Minus(left)) {
4060 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4062 if (tv2 != tarval_bad) {
4063 left = get_Minus_op(left);
4066 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4071 /* the following reassociations work only for <= */
4072 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4073 if (tv != tarval_bad) {
4074 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4075 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4081 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4082 switch (get_irn_opcode(left)) {
4086 c1 = get_And_right(left);
4089 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4090 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4092 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4094 /* TODO: move to constant evaluation */
4095 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4096 c1 = new_Const(mode_b, tv);
4097 DBG_OPT_CSTEVAL(proj, c1);
4101 if (tarval_is_single_bit(tv)) {
4103 * optimization for AND:
4105 * And(x, C) == C ==> And(x, C) != 0
4106 * And(x, C) != C ==> And(X, C) == 0
4108 * if C is a single Bit constant.
4111 /* check for Constant's match. We have check hare the tarvals,
4112 because our const might be changed */
4113 if (get_Const_tarval(c1) == tv) {
4114 /* fine: do the transformation */
4115 tv = get_mode_null(get_tarval_mode(tv));
4116 proj_nr ^= pn_Cmp_Leg;
4118 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4124 c1 = get_Or_right(left);
4125 if (is_Const(c1) && tarval_is_null(tv)) {
4127 * Or(x, C) == 0 && C != 0 ==> FALSE
4128 * Or(x, C) != 0 && C != 0 ==> TRUE
4130 if (! tarval_is_null(get_Const_tarval(c1))) {
4131 /* TODO: move to constant evaluation */
4132 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4133 c1 = new_Const(mode_b, tv);
4134 DBG_OPT_CSTEVAL(proj, c1);
4141 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4143 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4146 c1 = get_Shl_right(left);
4148 tarval *tv1 = get_Const_tarval(c1);
4149 ir_mode *mode = get_irn_mode(left);
4150 tarval *minus1 = get_mode_all_one(mode);
4151 tarval *amask = tarval_shr(minus1, tv1);
4152 tarval *cmask = tarval_shl(minus1, tv1);
4155 if (tarval_and(tv, cmask) != tv) {
4156 /* condition not met */
4157 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4158 c1 = new_Const(mode_b, tv);
4159 DBG_OPT_CSTEVAL(proj, c1);
4162 sl = get_Shl_left(left);
4163 blk = get_nodes_block(n);
4164 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4165 tv = tarval_shr(tv, tv1);
4167 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4172 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4174 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4177 c1 = get_Shr_right(left);
4179 tarval *tv1 = get_Const_tarval(c1);
4180 ir_mode *mode = get_irn_mode(left);
4181 tarval *minus1 = get_mode_all_one(mode);
4182 tarval *amask = tarval_shl(minus1, tv1);
4183 tarval *cmask = tarval_shr(minus1, tv1);
4186 if (tarval_and(tv, cmask) != tv) {
4187 /* condition not met */
4188 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4189 c1 = new_Const(mode_b, tv);
4190 DBG_OPT_CSTEVAL(proj, c1);
4193 sl = get_Shr_left(left);
4194 blk = get_nodes_block(n);
4195 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4196 tv = tarval_shl(tv, tv1);
4198 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4203 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4205 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4208 c1 = get_Shrs_right(left);
4210 tarval *tv1 = get_Const_tarval(c1);
4211 ir_mode *mode = get_irn_mode(left);
4212 tarval *minus1 = get_mode_all_one(mode);
4213 tarval *amask = tarval_shl(minus1, tv1);
4214 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4217 cond = tarval_sub(cond, tv1);
4218 cond = tarval_shrs(tv, cond);
4220 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4221 /* condition not met */
4222 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4223 c1 = new_Const(mode_b, tv);
4224 DBG_OPT_CSTEVAL(proj, c1);
4227 sl = get_Shrs_left(left);
4228 blk = get_nodes_block(n);
4229 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4230 tv = tarval_shl(tv, tv1);
4232 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4237 } /* tarval != bad */
4240 if (changed & 2) /* need a new Const */
4241 right = new_Const(mode, tv);
4243 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4244 ir_node *op = get_Proj_pred(left);
4246 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4247 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4248 ir_node *c = get_binop_right(op);
4251 tarval *tv = get_Const_tarval(c);
4253 if (tarval_is_single_bit(tv)) {
4254 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4255 ir_node *v = get_binop_left(op);
4256 ir_node *blk = get_irn_n(op, -1);
4257 ir_mode *mode = get_irn_mode(v);
4259 tv = tarval_sub(tv, get_mode_one(mode));
4260 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4262 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4269 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4271 /* create a new compare */
4272 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4273 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4277 } /* transform_node_Proj_Cmp */
4280 * Does all optimizations on nodes that must be done on it's Proj's
4281 * because of creating new nodes.
4283 static ir_node *transform_node_Proj(ir_node *proj) {
4284 ir_node *n = get_Proj_pred(proj);
4286 switch (get_irn_opcode(n)) {
4288 return transform_node_Proj_Div(proj);
4291 return transform_node_Proj_Mod(proj);
4294 return transform_node_Proj_DivMod(proj);
4297 return transform_node_Proj_Cond(proj);
4300 return transform_node_Proj_Cmp(proj);
4303 /* should not happen, but if it does will be optimized away */
4304 return equivalent_node_Proj(proj);
4310 } /* transform_node_Proj */
4313 * Move Confirms down through Phi nodes.
4315 static ir_node *transform_node_Phi(ir_node *phi) {
4317 ir_mode *mode = get_irn_mode(phi);
4319 if (mode_is_reference(mode)) {
4320 n = get_irn_arity(phi);
4322 /* Beware of Phi0 */
4324 ir_node *pred = get_irn_n(phi, 0);
4325 ir_node *bound, *new_Phi, *block, **in;
4328 if (! is_Confirm(pred))
4331 bound = get_Confirm_bound(pred);
4332 pnc = get_Confirm_cmp(pred);
4334 NEW_ARR_A(ir_node *, in, n);
4335 in[0] = get_Confirm_value(pred);
4337 for (i = 1; i < n; ++i) {
4338 pred = get_irn_n(phi, i);
4340 if (! is_Confirm(pred) ||
4341 get_Confirm_bound(pred) != bound ||
4342 get_Confirm_cmp(pred) != pnc)
4344 in[i] = get_Confirm_value(pred);
4346 /* move the Confirm nodes "behind" the Phi */
4347 block = get_irn_n(phi, -1);
4348 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4349 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4353 } /* transform_node_Phi */
4356 * Returns the operands of a commutative bin-op, if one operand is
4357 * a const, it is returned as the second one.
4359 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4360 ir_node *op_a = get_binop_left(binop);
4361 ir_node *op_b = get_binop_right(binop);
4363 assert(is_op_commutative(get_irn_op(binop)));
4365 if (is_Const(op_a)) {
4372 } /* get_comm_Binop_Ops */
4375 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4376 * Such pattern may arise in bitfield stores.
4378 * value c4 value c4 & c2
4379 * AND c3 AND c1 | c3
4386 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4389 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4392 ir_node *and_l, *c3;
4393 ir_node *value, *c4;
4394 ir_node *new_and, *new_const, *block;
4395 ir_mode *mode = get_irn_mode(or);
4397 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4400 get_comm_Binop_Ops(or, &and, &c1);
4401 if (!is_Const(c1) || !is_And(and))
4404 get_comm_Binop_Ops(and, &or_l, &c2);
4408 tv1 = get_Const_tarval(c1);
4409 tv2 = get_Const_tarval(c2);
4411 tv = tarval_or(tv1, tv2);
4412 if (tarval_is_all_one(tv)) {
4413 /* the AND does NOT clear a bit with isn't set by the OR */
4414 set_Or_left(or, or_l);
4415 set_Or_right(or, c1);
4417 /* check for more */
4424 get_comm_Binop_Ops(or_l, &and_l, &c3);
4425 if (!is_Const(c3) || !is_And(and_l))
4428 get_comm_Binop_Ops(and_l, &value, &c4);
4432 /* ok, found the pattern, check for conditions */
4433 assert(mode == get_irn_mode(and));
4434 assert(mode == get_irn_mode(or_l));
4435 assert(mode == get_irn_mode(and_l));
4437 tv3 = get_Const_tarval(c3);
4438 tv4 = get_Const_tarval(c4);
4440 tv = tarval_or(tv4, tv2);
4441 if (!tarval_is_all_one(tv)) {
4442 /* have at least one 0 at the same bit position */
4446 n_tv4 = tarval_not(tv4);
4447 if (tv3 != tarval_and(tv3, n_tv4)) {
4448 /* bit in the or_mask is outside the and_mask */
4452 n_tv2 = tarval_not(tv2);
4453 if (tv1 != tarval_and(tv1, n_tv2)) {
4454 /* bit in the or_mask is outside the and_mask */
4458 /* ok, all conditions met */
4459 block = get_irn_n(or, -1);
4461 new_and = new_r_And(current_ir_graph, block,
4462 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4464 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4466 set_Or_left(or, new_and);
4467 set_Or_right(or, new_const);
4469 /* check for more */
4471 } /* transform_node_Or_bf_store */
4474 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4476 static ir_node *transform_node_Or_Rot(ir_node *or) {
4477 ir_mode *mode = get_irn_mode(or);
4478 ir_node *shl, *shr, *block;
4479 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4482 if (! mode_is_int(mode))
4485 shl = get_binop_left(or);
4486 shr = get_binop_right(or);
4495 } else if (!is_Shl(shl)) {
4497 } else if (!is_Shr(shr)) {
4500 x = get_Shl_left(shl);
4501 if (x != get_Shr_left(shr))
4504 c1 = get_Shl_right(shl);
4505 c2 = get_Shr_right(shr);
4506 if (is_Const(c1) && is_Const(c2)) {
4507 tv1 = get_Const_tarval(c1);
4508 if (! tarval_is_long(tv1))
4511 tv2 = get_Const_tarval(c2);
4512 if (! tarval_is_long(tv2))
4515 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4516 != (int) get_mode_size_bits(mode))
4519 /* yet, condition met */
4520 block = get_irn_n(or, -1);
4522 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4524 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4526 } else if (is_Sub(c1)) {
4530 if (get_Sub_right(sub) != v)
4533 c1 = get_Sub_left(sub);
4537 tv1 = get_Const_tarval(c1);
4538 if (! tarval_is_long(tv1))
4541 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4544 /* yet, condition met */
4545 block = get_nodes_block(or);
4547 /* a Rot right is not supported, so use a rot left */
4548 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4550 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4552 } else if (is_Sub(c2)) {
4556 c1 = get_Sub_left(sub);
4560 tv1 = get_Const_tarval(c1);
4561 if (! tarval_is_long(tv1))
4564 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4567 /* yet, condition met */
4568 block = get_irn_n(or, -1);
4571 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4573 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4578 } /* transform_node_Or_Rot */
4583 static ir_node *transform_node_Or(ir_node *n) {
4584 ir_node *c, *oldn = n;
4585 ir_node *a = get_Or_left(n);
4586 ir_node *b = get_Or_right(n);
4589 if (is_Not(a) && is_Not(b)) {
4590 /* ~a | ~b = ~(a&b) */
4591 ir_node *block = get_nodes_block(n);
4593 mode = get_irn_mode(n);
4596 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4597 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4598 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4602 /* we can evaluate 2 Projs of the same Cmp */
4603 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4604 ir_node *pred_a = get_Proj_pred(a);
4605 ir_node *pred_b = get_Proj_pred(b);
4606 if (pred_a == pred_b) {
4607 dbg_info *dbgi = get_irn_dbg_info(n);
4608 ir_node *block = get_nodes_block(pred_a);
4609 pn_Cmp pn_a = get_Proj_proj(a);
4610 pn_Cmp pn_b = get_Proj_proj(b);
4611 /* yes, we can simply calculate with pncs */
4612 pn_Cmp new_pnc = pn_a | pn_b;
4614 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4619 mode = get_irn_mode(n);
4620 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4622 n = transform_node_Or_bf_store(n);
4623 n = transform_node_Or_Rot(n);
4627 n = transform_bitwise_distributive(n, transform_node_Or);
4630 } /* transform_node_Or */
4634 static ir_node *transform_node(ir_node *n);
4637 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4639 * Should be moved to reassociation?
4641 static ir_node *transform_node_shift(ir_node *n) {
4642 ir_node *left, *right;
4643 tarval *tv1, *tv2, *res;
4645 int modulo_shf, flag;
4647 left = get_binop_left(n);
4649 /* different operations */
4650 if (get_irn_op(left) != get_irn_op(n))
4653 right = get_binop_right(n);
4654 tv1 = value_of(right);
4655 if (tv1 == tarval_bad)
4658 tv2 = value_of(get_binop_right(left));
4659 if (tv2 == tarval_bad)
4662 res = tarval_add(tv1, tv2);
4664 /* beware: a simple replacement works only, if res < modulo shift */
4665 mode = get_irn_mode(n);
4669 modulo_shf = get_mode_modulo_shift(mode);
4670 if (modulo_shf > 0) {
4671 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4673 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4679 /* ok, we can replace it */
4680 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4682 in[0] = get_binop_left(left);
4683 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4685 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4687 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4689 return transform_node(irn);
4692 } /* transform_node_shift */
4697 static ir_node *transform_node_Shr(ir_node *n) {
4698 ir_node *c, *oldn = n;
4699 ir_node *a = get_Shr_left(n);
4700 ir_node *b = get_Shr_right(n);
4701 ir_mode *mode = get_irn_mode(n);
4703 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4704 return transform_node_shift(n);
4705 } /* transform_node_Shr */
4710 static ir_node *transform_node_Shrs(ir_node *n) {
4711 ir_node *c, *oldn = n;
4712 ir_node *a = get_Shrs_left(n);
4713 ir_node *b = get_Shrs_right(n);
4714 ir_mode *mode = get_irn_mode(n);
4716 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4717 return transform_node_shift(n);
4718 } /* transform_node_Shrs */
4723 static ir_node *transform_node_Shl(ir_node *n) {
4724 ir_node *c, *oldn = n;
4725 ir_node *a = get_Shl_left(n);
4726 ir_node *b = get_Shl_right(n);
4727 ir_mode *mode = get_irn_mode(n);
4729 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4730 return transform_node_shift(n);
4731 } /* transform_node_Shl */
4736 static ir_node *transform_node_Rot(ir_node *n) {
4737 ir_node *c, *oldn = n;
4738 ir_node *a = get_Rot_left(n);
4739 ir_node *b = get_Rot_right(n);
4740 ir_mode *mode = get_irn_mode(n);
4742 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4743 return transform_node_shift(n);
4744 } /* transform_node_Rot */
4749 static ir_node *transform_node_Conv(ir_node *n) {
4750 ir_node *c, *oldn = n;
4751 ir_node *a = get_Conv_op(n);
4753 if (is_const_Phi(a)) {
4754 c = apply_conv_on_phi(a, get_irn_mode(n));
4756 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4761 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4762 ir_mode *mode = get_irn_mode(n);
4763 return new_r_Unknown(current_ir_graph, mode);
4767 } /* transform_node_Conv */
4770 * Remove dead blocks and nodes in dead blocks
4771 * in keep alive list. We do not generate a new End node.
4773 static ir_node *transform_node_End(ir_node *n) {
4774 int i, j, n_keepalives = get_End_n_keepalives(n);
4777 NEW_ARR_A(ir_node *, in, n_keepalives);
4779 for (i = j = 0; i < n_keepalives; ++i) {
4780 ir_node *ka = get_End_keepalive(n, i);
4782 if (! is_Block_dead(ka)) {
4786 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4789 /* FIXME: beabi need to keep a Proj(M) */
4790 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4793 if (j != n_keepalives)
4794 set_End_keepalives(n, j, in);
4796 } /* transform_node_End */
4798 /** returns 1 if a == -b */
4799 static int is_negated_value(ir_node *a, ir_node *b) {
4800 if (is_Minus(a) && get_Minus_op(a) == b)
4802 if (is_Minus(b) && get_Minus_op(b) == a)
4804 if (is_Sub(a) && is_Sub(b)) {
4805 ir_node *a_left = get_Sub_left(a);
4806 ir_node *a_right = get_Sub_right(a);
4807 ir_node *b_left = get_Sub_left(b);
4808 ir_node *b_right = get_Sub_right(b);
4810 if (a_left == b_right && a_right == b_left)
4818 * Optimize a Mux into some simpler cases.
4820 static ir_node *transform_node_Mux(ir_node *n) {
4821 ir_node *oldn = n, *sel = get_Mux_sel(n);
4822 ir_mode *mode = get_irn_mode(n);
4823 ir_node *t = get_Mux_true(n);
4824 ir_node *f = get_Mux_false(n);
4825 ir_graph *irg = current_ir_graph;
4826 ir_node *conds[1], *vals[2];
4828 /* first normalization step: move a possible zero to the false case */
4830 ir_node *cmp = get_Proj_pred(sel);
4833 if (is_Const(t) && is_Const_null(t)) {
4834 /* Psi(x, 0, y) => Psi(x, y, 0) */
4835 pn_Cmp pnc = get_Proj_proj(sel);
4836 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
4837 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
4841 n = new_rd_Psi(get_irn_dbg_info(n), irg, get_nodes_block(n), 1, conds, vals, mode);
4848 /* note: after normalization, false can only happen on default */
4849 if (mode == mode_b) {
4850 dbg_info *dbg = get_irn_dbg_info(n);
4851 ir_node *block = get_nodes_block(n);
4852 ir_graph *irg = current_ir_graph;
4855 tarval *tv_t = get_Const_tarval(t);
4856 if (tv_t == tarval_b_true) {
4858 /* Muxb(sel, true, false) = sel */
4859 assert(get_Const_tarval(f) == tarval_b_false);
4860 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
4863 /* Muxb(sel, true, x) = Or(sel, x) */
4864 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
4865 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
4869 } else if (is_Const(f)) {
4870 tarval *tv_f = get_Const_tarval(f);
4871 if (tv_f == tarval_b_true) {
4872 /* Muxb(sel, x, true) = Or(Not(sel), x) */
4873 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4874 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
4875 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4878 /* Muxb(sel, x, false) = And(sel, x) */
4879 assert(tv_f == tarval_b_false);
4880 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
4881 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
4887 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
4888 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
4889 tarval *a = get_Const_tarval(t);
4890 tarval *b = get_Const_tarval(f);
4891 tarval *null = get_tarval_null(mode);
4894 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
4895 diff = tarval_sub(a, b);
4898 diff = tarval_sub(b, a);
4902 if (diff == get_tarval_one(mode) && min != null) {
4903 dbg_info *dbg = get_irn_dbg_info(n);
4904 ir_node *block = get_nodes_block(n);
4905 ir_graph *irg = current_ir_graph;
4909 vals[0] = new_Const(mode, tarval_sub(a, min));
4910 vals[1] = new_Const(mode, tarval_sub(b, min));
4911 n = new_rd_Psi(dbg, irg, block, 1, conds, vals, mode);
4912 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
4918 ir_node *cmp = get_Proj_pred(sel);
4919 long pn = get_Proj_proj(sel);
4922 * Note: normalization puts the constant on the right side,
4923 * so we check only one case.
4925 * Note further that these optimization work even for floating point
4926 * with NaN's because -NaN == NaN.
4927 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
4931 ir_node *cmp_r = get_Cmp_right(cmp);
4932 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4933 ir_node *block = get_nodes_block(n);
4934 ir_node *cmp_l = get_Cmp_left(cmp);
4936 if (mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
4939 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4940 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4942 /* Psi(a >/>= 0, a, -a) = Psi(a </<= 0, -a, a) ==> Abs(a) */
4943 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4945 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4947 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4948 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4950 /* Psi(a </<= 0, a, -a) = Psi(a >/>= 0, -a, a) ==> -Abs(a) */
4951 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4953 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4955 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4960 if (mode_is_int(mode)) {
4962 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
4963 /* Psi((a & b) != 0, c, 0) */
4964 ir_node *and_r = get_And_right(cmp_l);
4967 if (and_r == t && f == cmp_r) {
4968 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
4969 if (pn == pn_Cmp_Lg) {
4970 /* Psi((a & 2^C) != 0, 2^C, 0) */
4973 /* Psi((a & 2^C) == 0, 2^C, 0) */
4974 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
4975 block, cmp_l, t, mode);
4980 if (is_Shl(and_r)) {
4981 ir_node *shl_l = get_Shl_left(and_r);
4982 if (is_Const(shl_l) && is_Const_one(shl_l)) {
4983 if (and_r == t && f == cmp_r) {
4984 if (pn == pn_Cmp_Lg) {
4985 /* (a & (1 << n)) != 0, (1 << n), 0) */
4988 /* (a & (1 << n)) == 0, (1 << n), 0) */
4989 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
4990 block, cmp_l, t, mode);
4996 and_l = get_And_left(cmp_l);
4997 if (is_Shl(and_l)) {
4998 ir_node *shl_l = get_Shl_left(and_l);
4999 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5000 if (and_l == t && f == cmp_r) {
5001 if (pn == pn_Cmp_Lg) {
5002 /* ((1 << n) & a) != 0, (1 << n), 0) */
5005 /* ((1 << n) & a) == 0, (1 << n), 0) */
5006 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5007 block, cmp_l, t, mode);
5018 return arch_transform_node_Mux(n);
5019 } /* transform_node_Mux */
5022 * Optimize a Psi into some simpler cases.
5024 static ir_node *transform_node_Psi(ir_node *n) {
5026 return transform_node_Mux(n);
5029 } /* transform_node_Psi */
5032 * optimize sync nodes that have other syncs as input we simply add the inputs
5033 * of the other sync to our own inputs
5035 static ir_node *transform_node_Sync(ir_node *n) {
5036 int arity = get_Sync_n_preds(n);
5039 for (i = 0; i < arity;) {
5040 ir_node *pred = get_Sync_pred(n, i);
5044 if (!is_Sync(pred)) {
5052 pred_arity = get_Sync_n_preds(pred);
5053 for (j = 0; j < pred_arity; ++j) {
5054 ir_node *pred_pred = get_Sync_pred(pred, j);
5059 add_irn_n(n, pred_pred);
5063 if (get_Sync_pred(n, k) == pred_pred) break;
5068 /* rehash the sync node */
5069 add_identities(current_ir_graph->value_table, n);
5075 * Tries several [inplace] [optimizing] transformations and returns an
5076 * equivalent node. The difference to equivalent_node() is that these
5077 * transformations _do_ generate new nodes, and thus the old node must
5078 * not be freed even if the equivalent node isn't the old one.
5080 static ir_node *transform_node(ir_node *n) {
5084 * Transform_node is the only "optimizing transformation" that might
5085 * return a node with a different opcode. We iterate HERE until fixpoint
5086 * to get the final result.
5090 if (n->op->ops.transform_node)
5091 n = n->op->ops.transform_node(n);
5092 } while (oldn != n);
5095 } /* transform_node */
5098 * Sets the default transform node operation for an ir_op_ops.
5100 * @param code the opcode for the default operation
5101 * @param ops the operations initialized
5106 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5110 ops->transform_node = transform_node_##a; \
5147 } /* firm_set_default_transform_node */
5150 /* **************** Common Subexpression Elimination **************** */
5152 /** The size of the hash table used, should estimate the number of nodes
5154 #define N_IR_NODES 512
5156 /** Compares the attributes of two Const nodes. */
5157 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5158 return (get_Const_tarval(a) != get_Const_tarval(b))
5159 || (get_Const_type(a) != get_Const_type(b));
5160 } /* node_cmp_attr_Const */
5162 /** Compares the attributes of two Proj nodes. */
5163 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5164 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5165 } /* node_cmp_attr_Proj */
5167 /** Compares the attributes of two Filter nodes. */
5168 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5169 return get_Filter_proj(a) != get_Filter_proj(b);
5170 } /* node_cmp_attr_Filter */
5172 /** Compares the attributes of two Alloc nodes. */
5173 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5174 const alloc_attr *pa = get_irn_alloc_attr(a);
5175 const alloc_attr *pb = get_irn_alloc_attr(b);
5176 return (pa->where != pb->where) || (pa->type != pb->type);
5177 } /* node_cmp_attr_Alloc */
5179 /** Compares the attributes of two Free nodes. */
5180 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5181 const free_attr *pa = get_irn_free_attr(a);
5182 const free_attr *pb = get_irn_free_attr(b);
5183 return (pa->where != pb->where) || (pa->type != pb->type);
5184 } /* node_cmp_attr_Free */
5186 /** Compares the attributes of two SymConst nodes. */
5187 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5188 const symconst_attr *pa = get_irn_symconst_attr(a);
5189 const symconst_attr *pb = get_irn_symconst_attr(b);
5190 return (pa->kind != pb->kind)
5191 || (pa->sym.type_p != pb->sym.type_p)
5192 || (pa->tp != pb->tp);
5193 } /* node_cmp_attr_SymConst */
5195 /** Compares the attributes of two Call nodes. */
5196 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5197 return get_irn_call_attr(a) != get_irn_call_attr(b);
5198 } /* node_cmp_attr_Call */
5200 /** Compares the attributes of two Sel nodes. */
5201 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5202 const ir_entity *a_ent = get_Sel_entity(a);
5203 const ir_entity *b_ent = get_Sel_entity(b);
5205 (a_ent->kind != b_ent->kind) ||
5206 (a_ent->name != b_ent->name) ||
5207 (a_ent->owner != b_ent->owner) ||
5208 (a_ent->ld_name != b_ent->ld_name) ||
5209 (a_ent->type != b_ent->type);
5210 } /* node_cmp_attr_Sel */
5212 /** Compares the attributes of two Phi nodes. */
5213 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5214 /* we can only enter this function if both nodes have the same number of inputs,
5215 hence it is enough to check if one of them is a Phi0 */
5217 /* check the Phi0 pos attribute */
5218 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5221 } /* node_cmp_attr_Phi */
5223 /** Compares the attributes of two Conv nodes. */
5224 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5225 return get_Conv_strict(a) != get_Conv_strict(b);
5226 } /* node_cmp_attr_Conv */
5228 /** Compares the attributes of two Cast nodes. */
5229 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5230 return get_Cast_type(a) != get_Cast_type(b);
5231 } /* node_cmp_attr_Cast */
5233 /** Compares the attributes of two Load nodes. */
5234 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5235 if (get_Load_volatility(a) == volatility_is_volatile ||
5236 get_Load_volatility(b) == volatility_is_volatile)
5237 /* NEVER do CSE on volatile Loads */
5239 /* do not CSE Loads with different alignment. Be conservative. */
5240 if (get_Load_align(a) != get_Load_align(b))
5243 return get_Load_mode(a) != get_Load_mode(b);
5244 } /* node_cmp_attr_Load */
5246 /** Compares the attributes of two Store nodes. */
5247 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5248 /* do not CSE Stores with different alignment. Be conservative. */
5249 if (get_Store_align(a) != get_Store_align(b))
5252 /* NEVER do CSE on volatile Stores */
5253 return (get_Store_volatility(a) == volatility_is_volatile ||
5254 get_Store_volatility(b) == volatility_is_volatile);
5255 } /* node_cmp_attr_Store */
5257 /** Compares two exception attributes */
5258 static int node_cmp_exception(ir_node *a, ir_node *b) {
5259 const except_attr *ea = get_irn_except_attr(a);
5260 const except_attr *eb = get_irn_except_attr(b);
5262 return ea->pin_state != eb->pin_state;
5265 #define node_cmp_attr_Bound node_cmp_exception
5267 /** Compares the attributes of two Div nodes. */
5268 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5269 const divmod_attr *ma = get_irn_divmod_attr(a);
5270 const divmod_attr *mb = get_irn_divmod_attr(b);
5271 return ma->exc.pin_state != mb->exc.pin_state ||
5272 ma->res_mode != mb->res_mode ||
5273 ma->no_remainder != mb->no_remainder;
5274 } /* node_cmp_attr_Div */
5276 /** Compares the attributes of two DivMod nodes. */
5277 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5278 const divmod_attr *ma = get_irn_divmod_attr(a);
5279 const divmod_attr *mb = get_irn_divmod_attr(b);
5280 return ma->exc.pin_state != mb->exc.pin_state ||
5281 ma->res_mode != mb->res_mode;
5282 } /* node_cmp_attr_DivMod */
5284 /** Compares the attributes of two Mod nodes. */
5285 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5286 const divmod_attr *ma = get_irn_divmod_attr(a);
5287 const divmod_attr *mb = get_irn_divmod_attr(b);
5288 return ma->exc.pin_state != mb->exc.pin_state ||
5289 ma->res_mode != mb->res_mode;
5290 } /* node_cmp_attr_Mod */
5292 /** Compares the attributes of two Quot nodes. */
5293 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5294 const divmod_attr *ma = get_irn_divmod_attr(a);
5295 const divmod_attr *mb = get_irn_divmod_attr(b);
5296 return ma->exc.pin_state != mb->exc.pin_state ||
5297 ma->res_mode != mb->res_mode;
5298 } /* node_cmp_attr_Quot */
5300 /** Compares the attributes of two Confirm nodes. */
5301 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5302 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5303 } /* node_cmp_attr_Confirm */
5305 /** Compares the attributes of two ASM nodes. */
5306 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5308 const ir_asm_constraint *ca;
5309 const ir_asm_constraint *cb;
5312 if (get_ASM_text(a) != get_ASM_text(b))
5315 /* Should we really check the constraints here? Should be better, but is strange. */
5316 n = get_ASM_n_input_constraints(a);
5317 if (n != get_ASM_n_input_constraints(b))
5320 ca = get_ASM_input_constraints(a);
5321 cb = get_ASM_input_constraints(b);
5322 for (i = 0; i < n; ++i) {
5323 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5327 n = get_ASM_n_output_constraints(a);
5328 if (n != get_ASM_n_output_constraints(b))
5331 ca = get_ASM_output_constraints(a);
5332 cb = get_ASM_output_constraints(b);
5333 for (i = 0; i < n; ++i) {
5334 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5338 n = get_ASM_n_clobbers(a);
5339 if (n != get_ASM_n_clobbers(b))
5342 cla = get_ASM_clobbers(a);
5343 clb = get_ASM_clobbers(b);
5344 for (i = 0; i < n; ++i) {
5345 if (cla[i] != clb[i])
5349 } /* node_cmp_attr_ASM */
5352 * Set the default node attribute compare operation for an ir_op_ops.
5354 * @param code the opcode for the default operation
5355 * @param ops the operations initialized
5360 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5364 ops->node_cmp_attr = node_cmp_attr_##a; \
5395 } /* firm_set_default_node_cmp_attr */
5398 * Compare function for two nodes in the value table. Gets two
5399 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5401 int identities_cmp(const void *elt, const void *key) {
5402 ir_node *a = (ir_node *)elt;
5403 ir_node *b = (ir_node *)key;
5406 if (a == b) return 0;
5408 if ((get_irn_op(a) != get_irn_op(b)) ||
5409 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5411 /* compare if a's in and b's in are of equal length */
5412 irn_arity_a = get_irn_intra_arity(a);
5413 if (irn_arity_a != get_irn_intra_arity(b))
5416 if (get_irn_pinned(a) == op_pin_state_pinned) {
5417 /* for pinned nodes, the block inputs must be equal */
5418 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5420 } else if (! get_opt_global_cse()) {
5421 /* for block-local CSE both nodes must be in the same MacroBlock */
5422 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5426 /* compare a->in[0..ins] with b->in[0..ins] */
5427 for (i = 0; i < irn_arity_a; i++)
5428 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5432 * here, we already now that the nodes are identical except their
5435 if (a->op->ops.node_cmp_attr)
5436 return a->op->ops.node_cmp_attr(a, b);
5439 } /* identities_cmp */
5442 * Calculate a hash value of a node.
5444 unsigned ir_node_hash(ir_node *node) {
5448 if (node->op == op_Const) {
5449 /* special value for const, as they only differ in their tarval. */
5450 h = HASH_PTR(node->attr.con.tv);
5451 h = 9*h + HASH_PTR(get_irn_mode(node));
5452 } else if (node->op == op_SymConst) {
5453 /* special value for const, as they only differ in their symbol. */
5454 h = HASH_PTR(node->attr.symc.sym.type_p);
5455 h = 9*h + HASH_PTR(get_irn_mode(node));
5458 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5459 h = irn_arity = get_irn_intra_arity(node);
5461 /* consider all in nodes... except the block if not a control flow. */
5462 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5463 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5467 h = 9*h + HASH_PTR(get_irn_mode(node));
5469 h = 9*h + HASH_PTR(get_irn_op(node));
5473 } /* ir_node_hash */
5475 pset *new_identities(void) {
5476 return new_pset(identities_cmp, N_IR_NODES);
5477 } /* new_identities */
5479 void del_identities(pset *value_table) {
5480 del_pset(value_table);
5481 } /* del_identities */
5484 * Normalize a node by putting constants (and operands with larger
5485 * node index) on the right (operator side).
5487 * @param n The node to normalize
5489 static void normalize_node(ir_node *n) {
5490 if (is_op_commutative(get_irn_op(n))) {
5491 ir_node *l = get_binop_left(n);
5492 ir_node *r = get_binop_right(n);
5494 /* For commutative operators perform a OP b == b OP a but keep
5495 * constants on the RIGHT side. This helps greatly in some
5496 * optimizations. Moreover we use the idx number to make the form
5498 if (!operands_are_normalized(l, r)) {
5499 set_binop_left(n, r);
5500 set_binop_right(n, l);
5503 } /* normalize_node */
5506 * Update the nodes after a match in the value table. If both nodes have
5507 * the same MacroBlock but different Blocks, we must ensure that the node
5508 * with the dominating Block (the node that is near to the MacroBlock header
5509 * is stored in the table.
5510 * Because a MacroBlock has only one "non-exception" flow, we don't need
5511 * dominance info here: We known, that one block must dominate the other and
5512 * following the only block input will allow to find it.
5514 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5515 ir_node *known_blk, *new_block, *block, *mbh;
5517 if (get_opt_global_cse()) {
5518 /* Block inputs are meaning less */
5521 known_blk = get_irn_n(known_irn, -1);
5522 new_block = get_irn_n(new_ir_node, -1);
5523 if (known_blk == new_block) {
5524 /* already in the same block */
5528 * We expect the typical case when we built the graph. In that case, the
5529 * known_irn is already the upper one, so checking this should be faster.
5532 mbh = get_Block_MacroBlock(new_block);
5534 if (block == known_blk) {
5535 /* ok, we have found it: known_block dominates new_block as expected */
5540 * We have reached the MacroBlock header NOT founding
5541 * the known_block. new_block must dominate known_block.
5544 set_irn_n(known_irn, -1, new_block);
5547 assert(get_Block_n_cfgpreds(block) == 1);
5548 block = get_Block_cfgpred_block(block, 0);
5550 } /* update_value_table */
5553 * Return the canonical node computing the same value as n.
5555 * @param value_table The value table
5556 * @param n The node to lookup
5558 * Looks up the node in a hash table.
5560 * For Const nodes this is performed in the constructor, too. Const
5561 * nodes are extremely time critical because of their frequent use in
5562 * constant string arrays.
5564 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5567 if (!value_table) return n;
5571 o = pset_find(value_table, n, ir_node_hash(n));
5575 update_known_irn(o, n);
5582 * During construction we set the op_pin_state_pinned flag in the graph right when the
5583 * optimization is performed. The flag turning on procedure global cse could
5584 * be changed between two allocations. This way we are safe.
5586 * @param value_table The value table
5587 * @param n The node to lookup
5589 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5592 n = identify(value_table, n);
5593 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5594 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5596 } /* identify_cons */
5599 * Return the canonical node computing the same value as n.
5600 * Looks up the node in a hash table, enters it in the table
5601 * if it isn't there yet.
5603 * @param value_table the HashSet containing all nodes in the
5605 * @param n the node to look up
5607 * @return a node that computes the same value as n or n if no such
5608 * node could be found
5610 ir_node *identify_remember(pset *value_table, ir_node *n) {
5613 if (!value_table) return n;
5616 /* lookup or insert in hash table with given hash key. */
5617 o = pset_insert(value_table, n, ir_node_hash(n));
5620 update_known_irn(o, n);
5625 } /* identify_remember */
5627 /* Add a node to the identities value table. */
5628 void add_identities(pset *value_table, ir_node *node) {
5629 if (get_opt_cse() && is_no_Block(node))
5630 identify_remember(value_table, node);
5631 } /* add_identities */
5633 /* Visit each node in the value table of a graph. */
5634 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5636 ir_graph *rem = current_ir_graph;
5638 current_ir_graph = irg;
5639 foreach_pset(irg->value_table, node)
5641 current_ir_graph = rem;
5642 } /* visit_all_identities */
5645 * Garbage in, garbage out. If a node has a dead input, i.e., the
5646 * Bad node is input to the node, return the Bad node.
5648 static ir_node *gigo(ir_node *node) {
5650 ir_op *op = get_irn_op(node);
5652 /* remove garbage blocks by looking at control flow that leaves the block
5653 and replacing the control flow by Bad. */
5654 if (get_irn_mode(node) == mode_X) {
5655 ir_node *block = get_nodes_block(skip_Proj(node));
5657 /* Don't optimize nodes in immature blocks. */
5658 if (!get_Block_matured(block))
5660 /* Don't optimize End, may have Bads. */
5661 if (op == op_End) return node;
5663 if (is_Block(block)) {
5664 if (is_Block_dead(block)) {
5665 /* control flow from dead block is dead */
5669 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5670 if (!is_Bad(get_irn_n(block, i)))
5674 ir_graph *irg = get_irn_irg(block);
5675 /* the start block is never dead */
5676 if (block != get_irg_start_block(irg)
5677 && block != get_irg_end_block(irg)) {
5679 * Do NOT kill control flow without setting
5680 * the block to dead of bad things can happen:
5681 * We get a Block that is not reachable be irg_block_walk()
5682 * but can be found by irg_walk()!
5684 set_Block_dead(block);
5691 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5692 blocks predecessors is dead. */
5693 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5694 irn_arity = get_irn_arity(node);
5697 * Beware: we can only read the block of a non-floating node.
5699 if (is_irn_pinned_in_irg(node) &&
5700 is_Block_dead(get_nodes_block(skip_Proj(node))))
5703 for (i = 0; i < irn_arity; i++) {
5704 ir_node *pred = get_irn_n(node, i);
5709 /* Propagating Unknowns here seems to be a bad idea, because
5710 sometimes we need a node as a input and did not want that
5712 However, it might be useful to move this into a later phase
5713 (if you think that optimizing such code is useful). */
5714 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5715 return new_Unknown(get_irn_mode(node));
5720 /* With this code we violate the agreement that local_optimize
5721 only leaves Bads in Block, Phi and Tuple nodes. */
5722 /* If Block has only Bads as predecessors it's garbage. */
5723 /* If Phi has only Bads as predecessors it's garbage. */
5724 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5725 irn_arity = get_irn_arity(node);
5726 for (i = 0; i < irn_arity; i++) {
5727 if (!is_Bad(get_irn_n(node, i))) break;
5729 if (i == irn_arity) node = new_Bad();
5736 * These optimizations deallocate nodes from the obstack.
5737 * It can only be called if it is guaranteed that no other nodes
5738 * reference this one, i.e., right after construction of a node.
5740 * @param n The node to optimize
5742 * current_ir_graph must be set to the graph of the node!
5744 ir_node *optimize_node(ir_node *n) {
5747 ir_opcode iro = get_irn_opcode(n);
5749 /* Always optimize Phi nodes: part of the construction. */
5750 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5752 /* constant expression evaluation / constant folding */
5753 if (get_opt_constant_folding()) {
5754 /* neither constants nor Tuple values can be evaluated */
5755 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5756 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5757 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5758 /* try to evaluate */
5759 tv = computed_value(n);
5760 if (tv != tarval_bad) {
5762 ir_type *old_tp = get_irn_type(n);
5763 int i, arity = get_irn_arity(n);
5767 * Try to recover the type of the new expression.
5769 for (i = 0; i < arity && !old_tp; ++i)
5770 old_tp = get_irn_type(get_irn_n(n, i));
5773 * we MUST copy the node here temporary, because it's still needed
5774 * for DBG_OPT_CSTEVAL
5776 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5777 oldn = alloca(node_size);
5779 memcpy(oldn, n, node_size);
5780 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5782 /* ARG, copy the in array, we need it for statistics */
5783 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5785 /* note the inplace edges module */
5786 edges_node_deleted(n, current_ir_graph);
5788 /* evaluation was successful -- replace the node. */
5789 irg_kill_node(current_ir_graph, n);
5790 nw = new_Const(get_tarval_mode(tv), tv);
5792 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5793 set_Const_type(nw, old_tp);
5794 DBG_OPT_CSTEVAL(oldn, nw);
5795 tarval_enable_fp_ops(old_fp_mode);
5798 tarval_enable_fp_ops(old_fp_mode);
5802 /* remove unnecessary nodes */
5803 if (get_opt_constant_folding() ||
5804 (iro == iro_Phi) || /* always optimize these nodes. */
5806 (iro == iro_Proj) ||
5807 (iro == iro_Block) ) /* Flags tested local. */
5808 n = equivalent_node(n);
5810 /* Common Subexpression Elimination.
5812 * Checks whether n is already available.
5813 * The block input is used to distinguish different subexpressions. Right
5814 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5815 * subexpressions within a block.
5818 n = identify_cons(current_ir_graph->value_table, n);
5821 edges_node_deleted(oldn, current_ir_graph);
5823 /* We found an existing, better node, so we can deallocate the old node. */
5824 irg_kill_node(current_ir_graph, oldn);
5828 /* Some more constant expression evaluation that does not allow to
5830 iro = get_irn_opcode(n);
5831 if (get_opt_constant_folding() ||
5832 (iro == iro_Cond) ||
5833 (iro == iro_Proj)) /* Flags tested local. */
5834 n = transform_node(n);
5836 /* Remove nodes with dead (Bad) input.
5837 Run always for transformation induced Bads. */
5840 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5841 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5842 n = identify_remember(current_ir_graph->value_table, n);
5846 } /* optimize_node */
5850 * These optimizations never deallocate nodes (in place). This can cause dead
5851 * nodes lying on the obstack. Remove these by a dead node elimination,
5852 * i.e., a copying garbage collection.
5854 ir_node *optimize_in_place_2(ir_node *n) {
5857 ir_opcode iro = get_irn_opcode(n);
5859 if (!get_opt_optimize() && !is_Phi(n)) return n;
5861 /* constant expression evaluation / constant folding */
5862 if (get_opt_constant_folding()) {
5863 /* neither constants nor Tuple values can be evaluated */
5864 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5865 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5866 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5867 /* try to evaluate */
5868 tv = computed_value(n);
5869 if (tv != tarval_bad) {
5870 /* evaluation was successful -- replace the node. */
5871 ir_type *old_tp = get_irn_type(n);
5872 int i, arity = get_irn_arity(n);
5875 * Try to recover the type of the new expression.
5877 for (i = 0; i < arity && !old_tp; ++i)
5878 old_tp = get_irn_type(get_irn_n(n, i));
5880 n = new_Const(get_tarval_mode(tv), tv);
5882 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5883 set_Const_type(n, old_tp);
5885 DBG_OPT_CSTEVAL(oldn, n);
5886 tarval_enable_fp_ops(old_fp_mode);
5889 tarval_enable_fp_ops(old_fp_mode);
5893 /* remove unnecessary nodes */
5894 if (get_opt_constant_folding() ||
5895 (iro == iro_Phi) || /* always optimize these nodes. */
5896 (iro == iro_Id) || /* ... */
5897 (iro == iro_Proj) || /* ... */
5898 (iro == iro_Block) ) /* Flags tested local. */
5899 n = equivalent_node(n);
5901 /** common subexpression elimination **/
5902 /* Checks whether n is already available. */
5903 /* The block input is used to distinguish different subexpressions. Right
5904 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5905 subexpressions within a block. */
5906 if (get_opt_cse()) {
5907 n = identify(current_ir_graph->value_table, n);
5910 /* Some more constant expression evaluation. */
5911 iro = get_irn_opcode(n);
5912 if (get_opt_constant_folding() ||
5913 (iro == iro_Cond) ||
5914 (iro == iro_Proj)) /* Flags tested local. */
5915 n = transform_node(n);
5917 /* Remove nodes with dead (Bad) input.
5918 Run always for transformation induced Bads. */
5921 /* Now we can verify the node, as it has no dead inputs any more. */
5924 /* Now we have a legal, useful node. Enter it in hash table for cse.
5925 Blocks should be unique anyways. (Except the successor of start:
5926 is cse with the start block!) */
5927 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5928 n = identify_remember(current_ir_graph->value_table, n);
5931 } /* optimize_in_place_2 */
5934 * Wrapper for external use, set proper status bits after optimization.
5936 ir_node *optimize_in_place(ir_node *n) {
5937 /* Handle graph state */
5938 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5940 if (get_opt_global_cse())
5941 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5942 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5943 set_irg_outs_inconsistent(current_ir_graph);
5945 /* FIXME: Maybe we could also test whether optimizing the node can
5946 change the control graph. */
5947 set_irg_doms_inconsistent(current_ir_graph);
5948 return optimize_in_place_2(n);
5949 } /* optimize_in_place */
5952 * Sets the default operation for an ir_ops.
5954 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5955 ops = firm_set_default_computed_value(code, ops);
5956 ops = firm_set_default_equivalent_node(code, ops);
5957 ops = firm_set_default_transform_node(code, ops);
5958 ops = firm_set_default_node_cmp_attr(code, ops);
5959 ops = firm_set_default_get_type(code, ops);
5960 ops = firm_set_default_get_type_attr(code, ops);
5961 ops = firm_set_default_get_entity_attr(code, ops);
5964 } /* firm_set_default_operations */