2 * Copyright (C) 1995-2007 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) && mode_is_signed(get_irn_mode(a)))
182 return tarval_neg(ta);
185 } /* computed_value_Minus */
188 * Return the value of a Mul.
190 static tarval *computed_value_Mul(ir_node *n) {
191 ir_node *a = get_Mul_left(n);
192 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 mode = get_irn_mode(n);
199 if (mode != get_irn_mode(a)) {
200 /* n * n = 2n bit multiplication */
201 ta = tarval_convert_to(ta, mode);
202 tb = tarval_convert_to(tb, mode);
205 if (ta != tarval_bad && tb != tarval_bad) {
206 return tarval_mul(ta, tb);
208 /* a*0 = 0 or 0*b = 0 */
209 if (ta == get_mode_null(mode))
211 if (tb == get_mode_null(mode))
215 } /* computed_value_Mul */
218 * Return the value of a floating point Quot.
220 static tarval *computed_value_Quot(ir_node *n) {
221 ir_node *a = get_Quot_left(n);
222 ir_node *b = get_Quot_right(n);
224 tarval *ta = value_of(a);
225 tarval *tb = value_of(b);
227 /* This was missing in original implementation. Why? */
228 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
229 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
230 return tarval_quo(ta, tb);
233 } /* computed_value_Quot */
236 * Calculate the value of an integer Div of two nodes.
237 * Special case: 0 / b
239 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
240 tarval *ta = value_of(a);
241 tarval *tb = value_of(b);
243 /* Compute c1 / c2 or 0 / a, a != 0 */
244 if (ta != tarval_bad) {
245 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
246 return tarval_div(ta, tb);
247 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
251 } /* do_computed_value_Div */
254 * Return the value of an integer Div.
256 static tarval *computed_value_Div(ir_node *n) {
257 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
258 } /* computed_value_Div */
261 * Calculate the value of an integer Mod of two nodes.
262 * Special case: a % 1
264 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 /* Compute c1 % c2 or a % 1 */
269 if (tb != tarval_bad) {
270 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
271 return tarval_mod(ta, tb);
272 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
273 return get_mode_null(get_irn_mode(a));
276 } /* do_computed_value_Mod */
279 * Return the value of an integer Mod.
281 static tarval *computed_value_Mod(ir_node *n) {
282 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
283 } /* computed_value_Mod */
286 * Return the value of an Abs.
288 static tarval *computed_value_Abs(ir_node *n) {
289 ir_node *a = get_Abs_op(n);
290 tarval *ta = value_of(a);
292 if (ta != tarval_bad)
293 return tarval_abs(ta);
296 } /* computed_value_Abs */
299 * Return the value of an And.
300 * Special case: a & 0, 0 & b
302 static tarval *computed_value_And(ir_node *n) {
303 ir_node *a = get_And_left(n);
304 ir_node *b = get_And_right(n);
306 tarval *ta = value_of(a);
307 tarval *tb = value_of(b);
309 if ((ta != tarval_bad) && (tb != tarval_bad)) {
310 return tarval_and (ta, tb);
312 if (tarval_is_null(ta)) return ta;
313 if (tarval_is_null(tb)) return tb;
316 } /* computed_value_And */
319 * Return the value of an Or.
320 * Special case: a | 1...1, 1...1 | b
322 static tarval *computed_value_Or(ir_node *n) {
323 ir_node *a = get_Or_left(n);
324 ir_node *b = get_Or_right(n);
326 tarval *ta = value_of(a);
327 tarval *tb = value_of(b);
329 if ((ta != tarval_bad) && (tb != tarval_bad)) {
330 return tarval_or (ta, tb);
332 if (tarval_is_all_one(ta)) return ta;
333 if (tarval_is_all_one(tb)) return tb;
336 } /* computed_value_Or */
339 * Return the value of an Eor.
341 static tarval *computed_value_Eor(ir_node *n) {
342 ir_node *a = get_Eor_left(n);
343 ir_node *b = get_Eor_right(n);
348 return get_mode_null(get_irn_mode(n));
353 if ((ta != tarval_bad) && (tb != tarval_bad)) {
354 return tarval_eor (ta, tb);
357 } /* computed_value_Eor */
360 * Return the value of a Not.
362 static tarval *computed_value_Not(ir_node *n) {
363 ir_node *a = get_Not_op(n);
364 tarval *ta = value_of(a);
366 if (ta != tarval_bad)
367 return tarval_not(ta);
370 } /* computed_value_Not */
373 * Return the value of a Shl.
375 static tarval *computed_value_Shl(ir_node *n) {
376 ir_node *a = get_Shl_left(n);
377 ir_node *b = get_Shl_right(n);
379 tarval *ta = value_of(a);
380 tarval *tb = value_of(b);
382 if ((ta != tarval_bad) && (tb != tarval_bad)) {
383 return tarval_shl (ta, tb);
386 } /* computed_value_Shl */
389 * Return the value of a Shr.
391 static tarval *computed_value_Shr(ir_node *n) {
392 ir_node *a = get_Shr_left(n);
393 ir_node *b = get_Shr_right(n);
395 tarval *ta = value_of(a);
396 tarval *tb = value_of(b);
398 if ((ta != tarval_bad) && (tb != tarval_bad)) {
399 return tarval_shr (ta, tb);
402 } /* computed_value_Shr */
405 * Return the value of a Shrs.
407 static tarval *computed_value_Shrs(ir_node *n) {
408 ir_node *a = get_Shrs_left(n);
409 ir_node *b = get_Shrs_right(n);
411 tarval *ta = value_of(a);
412 tarval *tb = value_of(b);
414 if ((ta != tarval_bad) && (tb != tarval_bad)) {
415 return tarval_shrs (ta, tb);
418 } /* computed_value_Shrs */
421 * Return the value of a Rot.
423 static tarval *computed_value_Rot(ir_node *n) {
424 ir_node *a = get_Rot_left(n);
425 ir_node *b = get_Rot_right(n);
427 tarval *ta = value_of(a);
428 tarval *tb = value_of(b);
430 if ((ta != tarval_bad) && (tb != tarval_bad)) {
431 return tarval_rot (ta, tb);
434 } /* computed_value_Rot */
437 * Return the value of a Conv.
439 static tarval *computed_value_Conv(ir_node *n) {
440 ir_node *a = get_Conv_op(n);
441 tarval *ta = value_of(a);
443 if (ta != tarval_bad)
444 return tarval_convert_to(ta, get_irn_mode(n));
447 } /* computed_value_Conv */
450 * Return the value of a Proj(Cmp).
452 * This performs a first step of unreachable code elimination.
453 * Proj can not be computed, but folding a Cmp above the Proj here is
454 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
456 * There are several case where we can evaluate a Cmp node, see later.
458 static tarval *computed_value_Proj_Cmp(ir_node *n) {
459 ir_node *a = get_Proj_pred(n);
460 ir_node *aa = get_Cmp_left(a);
461 ir_node *ab = get_Cmp_right(a);
462 long proj_nr = get_Proj_proj(n);
465 * BEWARE: a == a is NOT always True for floating Point values, as
466 * NaN != NaN is defined, so we must check this here.
469 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
472 /* This is a trick with the bits used for encoding the Cmp
473 Proj numbers, the following statement is not the same:
474 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
475 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
478 tarval *taa = value_of(aa);
479 tarval *tab = value_of(ab);
480 ir_mode *mode = get_irn_mode(aa);
483 * The predecessors of Cmp are target values. We can evaluate
486 if ((taa != tarval_bad) && (tab != tarval_bad)) {
487 /* strange checks... */
488 pn_Cmp flags = tarval_cmp(taa, tab);
489 if (flags != pn_Cmp_False) {
490 return new_tarval_from_long (proj_nr & flags, mode_b);
493 /* for integer values, we can check against MIN/MAX */
494 else if (mode_is_int(mode)) {
495 /* MIN <=/> x. This results in true/false. */
496 if (taa == get_mode_min(mode)) {
497 /* a compare with the MIN value */
498 if (proj_nr == pn_Cmp_Le)
499 return get_tarval_b_true();
500 else if (proj_nr == pn_Cmp_Gt)
501 return get_tarval_b_false();
503 /* x >=/< MIN. This results in true/false. */
505 if (tab == get_mode_min(mode)) {
506 /* a compare with the MIN value */
507 if (proj_nr == pn_Cmp_Ge)
508 return get_tarval_b_true();
509 else if (proj_nr == pn_Cmp_Lt)
510 return get_tarval_b_false();
512 /* MAX >=/< x. This results in true/false. */
513 else if (taa == get_mode_max(mode)) {
514 if (proj_nr == pn_Cmp_Ge)
515 return get_tarval_b_true();
516 else if (proj_nr == pn_Cmp_Lt)
517 return get_tarval_b_false();
519 /* x <=/> MAX. This results in true/false. */
520 else if (tab == get_mode_max(mode)) {
521 if (proj_nr == pn_Cmp_Le)
522 return get_tarval_b_true();
523 else if (proj_nr == pn_Cmp_Gt)
524 return get_tarval_b_false();
528 * The predecessors are Allocs or (void*)(0) constants. Allocs never
529 * return NULL, they raise an exception. Therefore we can predict
533 ir_node *aaa = skip_Id(skip_Proj(aa));
534 ir_node *aba = skip_Id(skip_Proj(ab));
536 if ( ( (/* aa is ProjP and aaa is Alloc */
538 && mode_is_reference(get_irn_mode(aa))
540 && ( (/* ab is NULL */
542 && mode_is_reference(get_irn_mode(ab))
543 && is_Const_null(ab))
544 || (/* ab is other Alloc */
546 && mode_is_reference(get_irn_mode(ab))
549 || (/* aa is NULL and aba is Alloc */
551 && mode_is_reference(get_irn_mode(aa))
554 && mode_is_reference(get_irn_mode(ab))
557 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
560 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
561 } /* computed_value_Proj_Cmp */
564 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
565 * Proj(DivMod) and Proj(Quot).
567 static tarval *computed_value_Proj(ir_node *n) {
568 ir_node *a = get_Proj_pred(n);
571 switch (get_irn_opcode(a)) {
573 return computed_value_Proj_Cmp(n);
576 /* compute either the Div or the Mod part */
577 proj_nr = get_Proj_proj(n);
578 if (proj_nr == pn_DivMod_res_div)
579 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
580 else if (proj_nr == pn_DivMod_res_mod)
581 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
585 if (get_Proj_proj(n) == pn_Div_res)
586 return computed_value(a);
590 if (get_Proj_proj(n) == pn_Mod_res)
591 return computed_value(a);
595 if (get_Proj_proj(n) == pn_Quot_res)
596 return computed_value(a);
603 } /* computed_value_Proj */
606 * Calculate the value of a Mux: can be evaluated, if the
607 * sel and the right input are known.
609 static tarval *computed_value_Mux(ir_node *n) {
610 ir_node *sel = get_Mux_sel(n);
611 tarval *ts = value_of(sel);
613 if (ts == get_tarval_b_true()) {
614 ir_node *v = get_Mux_true(n);
617 else if (ts == get_tarval_b_false()) {
618 ir_node *v = get_Mux_false(n);
622 } /* computed_value_Mux */
625 * Calculate the value of a Psi: can be evaluated, if a condition is true
626 * and all previous conditions are false. If all conditions are false
627 * we evaluate to the default one.
629 static tarval *computed_value_Psi(ir_node *n) {
631 return computed_value_Mux(n);
633 } /* computed_value_Psi */
636 * Calculate the value of a Confirm: can be evaluated,
637 * if it has the form Confirm(x, '=', Const).
639 static tarval *computed_value_Confirm(ir_node *n) {
640 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
641 value_of(get_Confirm_bound(n)) : tarval_bad;
642 } /* computed_value_Confirm */
645 * If the parameter n can be computed, return its value, else tarval_bad.
646 * Performs constant folding.
648 * @param n The node this should be evaluated
650 tarval *computed_value(ir_node *n) {
651 if (n->op->ops.computed_value)
652 return n->op->ops.computed_value(n);
654 } /* computed_value */
657 * Set the default computed_value evaluator in an ir_op_ops.
659 * @param code the opcode for the default operation
660 * @param ops the operations initialized
665 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
669 ops->computed_value = computed_value_##a; \
704 } /* firm_set_default_computed_value */
707 * Returns a equivalent block for another block.
708 * If the block has only one predecessor, this is
709 * the equivalent one. If the only predecessor of a block is
710 * the block itself, this is a dead block.
712 * If both predecessors of a block are the branches of a binary
713 * Cond, the equivalent block is Cond's block.
715 * If all predecessors of a block are bad or lies in a dead
716 * block, the current block is dead as well.
718 * Note, that blocks are NEVER turned into Bad's, instead
719 * the dead_block flag is set. So, never test for is_Bad(block),
720 * always use is_dead_Block(block).
722 static ir_node *equivalent_node_Block(ir_node *n)
725 int n_preds = get_Block_n_cfgpreds(n);
727 /* The Block constructor does not call optimize, but mature_immBlock
728 calls the optimization. */
729 assert(get_Block_matured(n));
731 /* Straightening: a single entry Block following a single exit Block
732 can be merged, if it is not the Start block. */
733 /* !!! Beware, all Phi-nodes of n must have been optimized away.
734 This should be true, as the block is matured before optimize is called.
735 But what about Phi-cycles with the Phi0/Id that could not be resolved?
736 Remaining Phi nodes are just Ids. */
737 if (n_preds == 1 && is_Jmp(get_Block_cfgpred(n, 0))) {
738 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
739 if (predblock == oldn) {
740 /* Jmp jumps into the block it is in -- deal self cycle. */
741 n = set_Block_dead(n);
742 DBG_OPT_DEAD_BLOCK(oldn, n);
743 } else if (get_opt_control_flow_straightening()) {
745 DBG_OPT_STG(oldn, n);
747 } else if (n_preds == 1 && is_Cond(skip_Proj(get_Block_cfgpred(n, 0)))) {
748 ir_node *predblock = get_Block_cfgpred_block(n, 0);
749 if (predblock == oldn) {
750 /* Jmp jumps into the block it is in -- deal self cycle. */
751 n = set_Block_dead(n);
752 DBG_OPT_DEAD_BLOCK(oldn, n);
754 } else if ((n_preds == 2) &&
755 (get_opt_control_flow_weak_simplification())) {
756 /* Test whether Cond jumps twice to this block
757 * The more general case which more than 2 predecessors is handles
758 * in optimize_cf(), we handle only this special case for speed here.
760 ir_node *a = get_Block_cfgpred(n, 0);
761 ir_node *b = get_Block_cfgpred(n, 1);
765 (get_Proj_pred(a) == get_Proj_pred(b)) &&
766 is_Cond(get_Proj_pred(a)) &&
767 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
768 /* Also a single entry Block following a single exit Block. Phis have
769 twice the same operand and will be optimized away. */
770 n = get_nodes_block(get_Proj_pred(a));
771 DBG_OPT_IFSIM1(oldn, a, b, n);
773 } else if (get_opt_unreachable_code() &&
774 (n != get_irg_start_block(current_ir_graph)) &&
775 (n != get_irg_end_block(current_ir_graph)) ) {
778 /* If all inputs are dead, this block is dead too, except if it is
779 the start or end block. This is one step of unreachable code
781 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
782 ir_node *pred = get_Block_cfgpred(n, i);
785 if (is_Bad(pred)) continue;
786 pred_blk = get_nodes_block(skip_Proj(pred));
788 if (is_Block_dead(pred_blk)) continue;
791 /* really found a living input */
796 n = set_Block_dead(n);
797 DBG_OPT_DEAD_BLOCK(oldn, n);
802 } /* equivalent_node_Block */
805 * Returns a equivalent node for a Jmp, a Bad :-)
806 * Of course this only happens if the Block of the Jmp is dead.
808 static ir_node *equivalent_node_Jmp(ir_node *n) {
809 /* unreachable code elimination */
810 if (is_Block_dead(get_nodes_block(n)))
814 } /* equivalent_node_Jmp */
816 /** Raise is handled in the same way as Jmp. */
817 #define equivalent_node_Raise equivalent_node_Jmp
820 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
821 See transform_node_Proj_Cond(). */
824 * Optimize operations that are commutative and have neutral 0,
825 * so a op 0 = 0 op a = a.
827 static ir_node *equivalent_node_neutral_zero(ir_node *n)
831 ir_node *a = get_binop_left(n);
832 ir_node *b = get_binop_right(n);
837 /* After running compute_node there is only one constant predecessor.
838 Find this predecessors value and remember the other node: */
839 if ((tv = value_of(a)) != tarval_bad) {
841 } else if ((tv = value_of(b)) != tarval_bad) {
846 /* If this predecessors constant value is zero, the operation is
847 * unnecessary. Remove it.
849 * Beware: If n is a Add, the mode of on and n might be different
850 * which happens in this rare construction: NULL + 3.
851 * Then, a Conv would be needed which we cannot include here.
853 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
856 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
860 } /* equivalent_node_neutral_zero */
863 * Eor is commutative and has neutral 0.
865 #define equivalent_node_Eor equivalent_node_neutral_zero
868 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
870 * The second one looks strange, but this construct
871 * is used heavily in the LCC sources :-).
873 * Beware: The Mode of an Add may be different than the mode of its
874 * predecessors, so we could not return a predecessors in all cases.
876 static ir_node *equivalent_node_Add(ir_node *n) {
878 ir_node *left, *right;
879 ir_mode *mode = get_irn_mode(n);
881 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
882 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
885 n = equivalent_node_neutral_zero(n);
889 left = get_Add_left(n);
890 right = get_Add_right(n);
893 if (get_Sub_right(left) == right) {
896 n = get_Sub_left(left);
897 if (mode == get_irn_mode(n)) {
898 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
904 if (get_Sub_right(right) == left) {
907 n = get_Sub_left(right);
908 if (mode == get_irn_mode(n)) {
909 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
915 } /* equivalent_node_Add */
918 * optimize operations that are not commutative but have neutral 0 on left,
921 static ir_node *equivalent_node_left_zero(ir_node *n) {
924 ir_node *a = get_binop_left(n);
925 ir_node *b = get_binop_right(n);
927 if (is_Const(b) && is_Const_null(b)) {
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
933 } /* equivalent_node_left_zero */
935 #define equivalent_node_Shl equivalent_node_left_zero
936 #define equivalent_node_Shr equivalent_node_left_zero
937 #define equivalent_node_Shrs equivalent_node_left_zero
938 #define equivalent_node_Rot equivalent_node_left_zero
941 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
943 * The second one looks strange, but this construct
944 * is used heavily in the LCC sources :-).
946 * Beware: The Mode of a Sub may be different than the mode of its
947 * predecessors, so we could not return a predecessors in all cases.
949 static ir_node *equivalent_node_Sub(ir_node *n) {
952 ir_mode *mode = get_irn_mode(n);
954 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
955 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
958 b = get_Sub_right(n);
960 /* Beware: modes might be different */
961 if (is_Const(b) && is_Const_null(b)) {
962 ir_node *a = get_Sub_left(n);
963 if (mode == get_irn_mode(a)) {
966 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
970 } /* equivalent_node_Sub */
974 * Optimize an "idempotent unary op", ie op(op(n)) = n.
977 * -(-a) == a, but might overflow two times.
978 * We handle it anyway here but the better way would be a
979 * flag. This would be needed for Pascal for instance.
981 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
983 ir_node *pred = get_unop_op(n);
985 /* optimize symmetric unop */
986 if (get_irn_op(pred) == get_irn_op(n)) {
987 n = get_unop_op(pred);
988 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
991 } /* equivalent_node_idempotent_unop */
993 /** Optimize Not(Not(x)) == x. */
994 #define equivalent_node_Not equivalent_node_idempotent_unop
996 /** -(-x) == x ??? Is this possible or can --x raise an
997 out of bounds exception if min =! max? */
998 #define equivalent_node_Minus equivalent_node_idempotent_unop
1001 * Optimize a * 1 = 1 * a = a.
1003 static ir_node *equivalent_node_Mul(ir_node *n) {
1005 ir_node *a = get_Mul_left(n);
1007 /* we can handle here only the n * n = n bit cases */
1008 if (get_irn_mode(n) == get_irn_mode(a)) {
1009 ir_node *b = get_Mul_right(n);
1011 /* Mul is commutative and has again an other neutral element. */
1012 if (is_Const(a) && is_Const_one(a)) {
1014 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1015 } else if (is_Const(b) && is_Const_one(b)) {
1017 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1021 } /* equivalent_node_Mul */
1024 * Optimize a / 1 = a.
1026 static ir_node *equivalent_node_Div(ir_node *n) {
1027 ir_node *a = get_Div_left(n);
1028 ir_node *b = get_Div_right(n);
1030 /* Div is not commutative. */
1031 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1032 /* Turn Div into a tuple (mem, bad, a) */
1033 ir_node *mem = get_Div_mem(n);
1034 ir_node *blk = get_irn_n(n, -1);
1035 turn_into_tuple(n, pn_Div_max);
1036 set_Tuple_pred(n, pn_Div_M, mem);
1037 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1038 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1039 set_Tuple_pred(n, pn_Div_res, a);
1042 } /* equivalent_node_Div */
1045 * Optimize a / 1.0 = a.
1047 static ir_node *equivalent_node_Quot(ir_node *n) {
1048 ir_node *a = get_Quot_left(n);
1049 ir_node *b = get_Quot_right(n);
1051 /* Div is not commutative. */
1052 if (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */
1053 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1054 ir_node *mem = get_Quot_mem(n);
1055 ir_node *blk = get_irn_n(n, -1);
1056 turn_into_tuple(n, pn_Quot_max);
1057 set_Tuple_pred(n, pn_Quot_M, mem);
1058 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1059 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1060 set_Tuple_pred(n, pn_Quot_res, a);
1063 } /* equivalent_node_Quot */
1066 * Optimize a / 1 = a.
1068 static ir_node *equivalent_node_DivMod(ir_node *n) {
1069 ir_node *b = get_DivMod_right(n);
1071 /* Div is not commutative. */
1072 if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */
1073 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1074 ir_node *a = get_DivMod_left(n);
1075 ir_node *mem = get_Div_mem(n);
1076 ir_node *blk = get_irn_n(n, -1);
1077 ir_mode *mode = get_DivMod_resmode(n);
1079 turn_into_tuple(n, pn_DivMod_max);
1080 set_Tuple_pred(n, pn_DivMod_M, mem);
1081 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1082 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1083 set_Tuple_pred(n, pn_DivMod_res_div, a);
1084 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1087 } /* equivalent_node_DivMod */
1090 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1092 static ir_node *equivalent_node_Or(ir_node *n) {
1095 ir_node *a = get_Or_left(n);
1096 ir_node *b = get_Or_right(n);
1099 n = a; /* Or has it's own neutral element */
1100 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1101 } else if (is_Const(a) && is_Const_null(a)) {
1103 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1104 } else if (is_Const(b) && is_Const_null(b)) {
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1110 } /* equivalent_node_Or */
1113 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1115 static ir_node *equivalent_node_And(ir_node *n) {
1118 ir_node *a = get_And_left(n);
1119 ir_node *b = get_And_right(n);
1122 n = a; /* And has it's own neutral element */
1123 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1126 if (is_Const(a) && is_Const_all_one(a)) {
1128 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1131 if (is_Const(b) && is_Const_all_one(b)) {
1133 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1137 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1140 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1145 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1148 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1154 } /* equivalent_node_And */
1157 * Try to remove useless Conv's:
1159 static ir_node *equivalent_node_Conv(ir_node *n) {
1161 ir_node *a = get_Conv_op(n);
1164 ir_mode *n_mode = get_irn_mode(n);
1165 ir_mode *a_mode = get_irn_mode(a);
1167 if (n_mode == a_mode) { /* No Conv necessary */
1168 if (get_Conv_strict(n)) {
1169 /* special case: the predecessor might be a also a Conv */
1171 if (! get_Conv_strict(a)) {
1172 /* first one is not strict, kick it */
1173 set_Conv_op(n, get_Conv_op(a));
1176 /* else both are strict conv, second is superflous */
1177 } else if(is_Proj(a)) {
1178 ir_node *pred = get_Proj_pred(a);
1180 /* loads always return with the exact precision of n_mode */
1181 assert(get_Load_mode(pred) == n_mode);
1186 /* leave strict floating point Conv's */
1190 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1191 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1195 n_mode = get_irn_mode(n);
1196 b_mode = get_irn_mode(b);
1198 if (n_mode == b_mode) {
1199 if (n_mode == mode_b) {
1200 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1201 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1202 } else if (mode_is_int(n_mode)) {
1203 if (smaller_mode(b_mode, a_mode)){
1204 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1205 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1211 } /* equivalent_node_Conv */
1214 * A Cast may be removed if the type of the previous node
1215 * is already the type of the Cast.
1217 static ir_node *equivalent_node_Cast(ir_node *n) {
1219 ir_node *pred = get_Cast_op(n);
1221 if (get_irn_type(pred) == get_Cast_type(n)) {
1223 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1226 } /* equivalent_node_Cast */
1229 * Several optimizations:
1230 * - no Phi in start block.
1231 * - remove Id operators that are inputs to Phi
1232 * - fold Phi-nodes, iff they have only one predecessor except
1235 static ir_node *equivalent_node_Phi(ir_node *n) {
1239 ir_node *block = NULL; /* to shutup gcc */
1240 ir_node *first_val = NULL; /* to shutup gcc */
1242 if (!get_opt_normalize()) return n;
1244 n_preds = get_Phi_n_preds(n);
1246 block = get_nodes_block(n);
1247 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1248 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1249 if ((is_Block_dead(block)) || /* Control dead */
1250 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1251 return new_Bad(); /* in the Start Block. */
1253 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1255 /* If the Block has a Bad pred, we also have one. */
1256 for (i = 0; i < n_preds; ++i)
1257 if (is_Bad(get_Block_cfgpred(block, i)))
1258 set_Phi_pred(n, i, new_Bad());
1260 /* Find first non-self-referencing input */
1261 for (i = 0; i < n_preds; ++i) {
1262 first_val = get_Phi_pred(n, i);
1263 if ( (first_val != n) /* not self pointer */
1265 && (! is_Bad(first_val))
1267 ) { /* value not dead */
1268 break; /* then found first value. */
1273 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1277 /* search for rest of inputs, determine if any of these
1278 are non-self-referencing */
1279 while (++i < n_preds) {
1280 ir_node *scnd_val = get_Phi_pred(n, i);
1281 if ( (scnd_val != n)
1282 && (scnd_val != first_val)
1284 && (! is_Bad(scnd_val))
1292 /* Fold, if no multiple distinct non-self-referencing inputs */
1294 DBG_OPT_PHI(oldn, n);
1297 } /* equivalent_node_Phi */
1300 * Several optimizations:
1301 * - no Sync in start block.
1302 * - fold Sync-nodes, iff they have only one predecessor except
1305 static ir_node *equivalent_node_Sync(ir_node *n) {
1309 ir_node *first_val = NULL; /* to shutup gcc */
1311 if (!get_opt_normalize()) return n;
1313 n_preds = get_Sync_n_preds(n);
1315 /* Find first non-self-referencing input */
1316 for (i = 0; i < n_preds; ++i) {
1317 first_val = get_Sync_pred(n, i);
1318 if ((first_val != n) /* not self pointer */ &&
1319 (! is_Bad(first_val))
1320 ) { /* value not dead */
1321 break; /* then found first value. */
1326 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1329 /* search the rest of inputs, determine if any of these
1330 are non-self-referencing */
1331 while (++i < n_preds) {
1332 ir_node *scnd_val = get_Sync_pred(n, i);
1333 if ((scnd_val != n) &&
1334 (scnd_val != first_val) &&
1335 (! is_Bad(scnd_val))
1341 /* Fold, if no multiple distinct non-self-referencing inputs */
1343 DBG_OPT_SYNC(oldn, n);
1346 } /* equivalent_node_Sync */
1349 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1350 * ProjX(Load) and ProjX(Store).
1352 static ir_node *equivalent_node_Proj(ir_node *proj) {
1353 ir_node *oldn = proj;
1354 ir_node *a = get_Proj_pred(proj);
1357 /* Remove the Tuple/Proj combination. */
1358 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1359 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1360 DBG_OPT_TUPLE(oldn, a, proj);
1362 /* This should not happen! */
1363 assert(! "found a Proj with higher number than Tuple predecessors");
1366 } else if (get_irn_mode(proj) == mode_X) {
1367 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1368 /* Remove dead control flow -- early gigo(). */
1370 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1372 /* get the Load address */
1373 ir_node *addr = get_Load_ptr(a);
1374 ir_node *blk = get_irn_n(a, -1);
1377 if (value_not_null(addr, &confirm)) {
1378 if (confirm == NULL) {
1379 /* this node may float if it did not depend on a Confirm */
1380 set_irn_pinned(a, op_pin_state_floats);
1382 if (get_Proj_proj(proj) == pn_Load_X_except) {
1383 DBG_OPT_EXC_REM(proj);
1386 return new_r_Jmp(current_ir_graph, blk);
1388 } else if (is_Store(a)) {
1389 /* get the load/store address */
1390 ir_node *addr = get_Store_ptr(a);
1391 ir_node *blk = get_irn_n(a, -1);
1394 if (value_not_null(addr, &confirm)) {
1395 if (confirm == NULL) {
1396 /* this node may float if it did not depend on a Confirm */
1397 set_irn_pinned(a, op_pin_state_floats);
1399 if (get_Proj_proj(proj) == pn_Store_X_except) {
1400 DBG_OPT_EXC_REM(proj);
1403 return new_r_Jmp(current_ir_graph, blk);
1410 } /* equivalent_node_Proj */
1415 static ir_node *equivalent_node_Id(ir_node *n) {
1420 } while (get_irn_op(n) == op_Id);
1422 DBG_OPT_ID(oldn, n);
1424 } /* equivalent_node_Id */
1429 static ir_node *equivalent_node_Mux(ir_node *n)
1431 ir_node *oldn = n, *sel = get_Mux_sel(n);
1432 tarval *ts = value_of(sel);
1434 /* Mux(true, f, t) == t */
1435 if (ts == tarval_b_true) {
1436 n = get_Mux_true(n);
1437 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1439 /* Mux(false, f, t) == f */
1440 else if (ts == tarval_b_false) {
1441 n = get_Mux_false(n);
1442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1444 /* Mux(v, x, x) == x */
1445 else if (get_Mux_false(n) == get_Mux_true(n)) {
1446 n = get_Mux_true(n);
1447 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1449 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1450 ir_node *cmp = get_Proj_pred(sel);
1451 long proj_nr = get_Proj_proj(sel);
1452 ir_node *b = get_Mux_false(n);
1453 ir_node *a = get_Mux_true(n);
1456 * Note: normalization puts the constant on the right site,
1457 * so we check only one case.
1459 * Note further that these optimization work even for floating point
1460 * with NaN's because -NaN == NaN.
1461 * However, if +0 and -0 is handled differently, we cannot use the first one.
1463 if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) {
1464 ir_node *cmp_r = get_Cmp_right(cmp);
1465 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
1466 /* Mux(a CMP 0, X, a) */
1467 if (is_Minus(b) && get_Minus_op(b) == a) {
1468 /* Mux(a CMP 0, -a, a) */
1469 if (proj_nr == pn_Cmp_Eq) {
1470 /* Mux(a == 0, -a, a) ==> -a */
1472 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1473 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1474 /* Mux(a != 0, -a, a) ==> a */
1476 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1478 } else if (is_Const(b) && is_Const_null(b)) {
1479 /* Mux(a CMP 0, 0, a) */
1480 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1481 /* Mux(a != 0, 0, a) ==> a */
1483 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1484 } else if (proj_nr == pn_Cmp_Eq) {
1485 /* Mux(a == 0, 0, a) ==> 0 */
1487 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1494 } /* equivalent_node_Mux */
1497 * Returns a equivalent node of a Psi: if a condition is true
1498 * and all previous conditions are false we know its value.
1499 * If all conditions are false its value is the default one.
1501 static ir_node *equivalent_node_Psi(ir_node *n) {
1503 return equivalent_node_Mux(n);
1505 } /* equivalent_node_Psi */
1508 * Optimize -a CMP -b into b CMP a.
1509 * This works only for for modes where unary Minus
1511 * Note that two-complement integers can Overflow
1512 * so it will NOT work.
1514 * For == and != can be handled in Proj(Cmp)
1516 static ir_node *equivalent_node_Cmp(ir_node *n) {
1517 ir_node *left = get_Cmp_left(n);
1518 ir_node *right = get_Cmp_right(n);
1520 if (is_Minus(left) && is_Minus(right) &&
1521 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1522 left = get_Minus_op(left);
1523 right = get_Minus_op(right);
1524 set_Cmp_left(n, right);
1525 set_Cmp_right(n, left);
1528 } /* equivalent_node_Cmp */
1531 * Remove Confirm nodes if setting is on.
1532 * Replace Confirms(x, '=', Constlike) by Constlike.
1534 static ir_node *equivalent_node_Confirm(ir_node *n) {
1535 ir_node *pred = get_Confirm_value(n);
1536 pn_Cmp pnc = get_Confirm_cmp(n);
1538 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1540 * rare case: two identical Confirms one after another,
1541 * replace the second one with the first.
1545 if (pnc == pn_Cmp_Eq) {
1546 ir_node *bound = get_Confirm_bound(n);
1549 * Optimize a rare case:
1550 * Confirm(x, '=', Constlike) ==> Constlike
1552 if (is_irn_constlike(bound)) {
1553 DBG_OPT_CONFIRM(n, bound);
1557 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1561 * Optimize CopyB(mem, x, x) into a Nop.
1563 static ir_node *equivalent_node_CopyB(ir_node *n) {
1564 ir_node *a = get_CopyB_dst(n);
1565 ir_node *b = get_CopyB_src(n);
1568 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1569 ir_node *mem = get_CopyB_mem(n);
1570 ir_node *blk = get_nodes_block(n);
1571 turn_into_tuple(n, pn_CopyB_max);
1572 set_Tuple_pred(n, pn_CopyB_M, mem);
1573 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1574 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1575 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1578 } /* equivalent_node_CopyB */
1581 * Optimize Bounds(idx, idx, upper) into idx.
1583 static ir_node *equivalent_node_Bound(ir_node *n) {
1584 ir_node *idx = get_Bound_index(n);
1585 ir_node *lower = get_Bound_lower(n);
1588 /* By definition lower < upper, so if idx == lower -->
1589 lower <= idx && idx < upper */
1591 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1594 ir_node *pred = skip_Proj(idx);
1596 if (get_irn_op(pred) == op_Bound) {
1598 * idx was Bounds_check previously, it is still valid if
1599 * lower <= pred_lower && pred_upper <= upper.
1601 ir_node *upper = get_Bound_upper(n);
1602 if (get_Bound_lower(pred) == lower &&
1603 get_Bound_upper(pred) == upper) {
1605 * One could expect that we simply return the previous
1606 * Bound here. However, this would be wrong, as we could
1607 * add an exception Proj to a new location then.
1608 * So, we must turn in into a tuple.
1615 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1616 ir_node *mem = get_Bound_mem(n);
1617 ir_node *blk = get_nodes_block(n);
1618 turn_into_tuple(n, pn_Bound_max);
1619 set_Tuple_pred(n, pn_Bound_M, mem);
1620 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1621 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1622 set_Tuple_pred(n, pn_Bound_res, idx);
1625 } /* equivalent_node_Bound */
1628 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1629 * perform no actual computation, as, e.g., the Id nodes. It does not create
1630 * new nodes. It is therefore safe to free n if the node returned is not n.
1631 * If a node returns a Tuple we can not just skip it. If the size of the
1632 * in array fits, we transform n into a tuple (e.g., Div).
1634 ir_node *equivalent_node(ir_node *n) {
1635 if (n->op->ops.equivalent_node)
1636 return n->op->ops.equivalent_node(n);
1638 } /* equivalent_node */
1641 * Sets the default equivalent node operation for an ir_op_ops.
1643 * @param code the opcode for the default operation
1644 * @param ops the operations initialized
1649 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1653 ops->equivalent_node = equivalent_node_##a; \
1693 } /* firm_set_default_equivalent_node */
1696 * Returns non-zero if a node is a Phi node
1697 * with all predecessors constant.
1699 static int is_const_Phi(ir_node *n) {
1704 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1705 if (! is_Const(get_irn_n(n, i)))
1708 } /* is_const_Phi */
1711 * Apply an evaluator on a binop with a constant operators (and one Phi).
1713 * @param phi the Phi node
1714 * @param other the other operand
1715 * @param eval an evaluator function
1716 * @param left if non-zero, other is the left operand, else the right
1718 * @return a new Phi node if the conversion was successful, NULL else
1720 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1726 int i, n = get_irn_arity(phi);
1728 NEW_ARR_A(void *, res, n);
1730 for (i = 0; i < n; ++i) {
1731 pred = get_irn_n(phi, i);
1732 tv = get_Const_tarval(pred);
1733 tv = eval(other, tv);
1735 if (tv == tarval_bad) {
1736 /* folding failed, bad */
1742 for (i = 0; i < n; ++i) {
1743 pred = get_irn_n(phi, i);
1744 tv = get_Const_tarval(pred);
1745 tv = eval(tv, other);
1747 if (tv == tarval_bad) {
1748 /* folding failed, bad */
1754 mode = get_irn_mode(phi);
1755 irg = current_ir_graph;
1756 for (i = 0; i < n; ++i) {
1757 pred = get_irn_n(phi, i);
1758 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1759 mode, res[i], get_Const_type(pred));
1761 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1762 } /* apply_binop_on_phi */
1765 * Apply an evaluator on a unop with a constant operator (a Phi).
1767 * @param phi the Phi node
1768 * @param eval an evaluator function
1770 * @return a new Phi node if the conversion was successful, NULL else
1772 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1778 int i, n = get_irn_arity(phi);
1780 NEW_ARR_A(void *, res, n);
1781 for (i = 0; i < n; ++i) {
1782 pred = get_irn_n(phi, i);
1783 tv = get_Const_tarval(pred);
1786 if (tv == tarval_bad) {
1787 /* folding failed, bad */
1792 mode = get_irn_mode(phi);
1793 irg = current_ir_graph;
1794 for (i = 0; i < n; ++i) {
1795 pred = get_irn_n(phi, i);
1796 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1797 mode, res[i], get_Const_type(pred));
1799 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1800 } /* apply_unop_on_phi */
1803 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1804 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1805 * If possible, remove the Conv's.
1807 static ir_node *transform_node_AddSub(ir_node *n) {
1808 ir_mode *mode = get_irn_mode(n);
1810 if (mode_is_reference(mode)) {
1811 ir_node *left = get_binop_left(n);
1812 ir_node *right = get_binop_right(n);
1813 int ref_bits = get_mode_size_bits(mode);
1815 if (is_Conv(left)) {
1816 ir_mode *mode = get_irn_mode(left);
1817 int bits = get_mode_size_bits(mode);
1819 if (ref_bits == bits &&
1820 mode_is_int(mode) &&
1821 get_mode_arithmetic(mode) == irma_twos_complement) {
1822 ir_node *pre = get_Conv_op(left);
1823 ir_mode *pre_mode = get_irn_mode(pre);
1825 if (mode_is_int(pre_mode) &&
1826 get_mode_size_bits(pre_mode) == bits &&
1827 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1828 /* ok, this conv just changes to sign, moreover the calculation
1829 * is done with same number of bits as our address mode, so
1830 * we can ignore the conv as address calculation can be viewed
1831 * as either signed or unsigned
1833 set_binop_left(n, pre);
1838 if (is_Conv(right)) {
1839 ir_mode *mode = get_irn_mode(right);
1840 int bits = get_mode_size_bits(mode);
1842 if (ref_bits == bits &&
1843 mode_is_int(mode) &&
1844 get_mode_arithmetic(mode) == irma_twos_complement) {
1845 ir_node *pre = get_Conv_op(right);
1846 ir_mode *pre_mode = get_irn_mode(pre);
1848 if (mode_is_int(pre_mode) &&
1849 get_mode_size_bits(pre_mode) == bits &&
1850 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1851 /* ok, this conv just changes to sign, moreover the calculation
1852 * is done with same number of bits as our address mode, so
1853 * we can ignore the conv as address calculation can be viewed
1854 * as either signed or unsigned
1856 set_binop_right(n, pre);
1862 } /* transform_node_AddSub */
1864 #define HANDLE_BINOP_PHI(op,a,b,c) \
1866 if (is_Const(b) && is_const_Phi(a)) { \
1867 /* check for Op(Phi, Const) */ \
1868 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1870 else if (is_Const(a) && is_const_Phi(b)) { \
1871 /* check for Op(Const, Phi) */ \
1872 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1875 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1879 #define HANDLE_UNOP_PHI(op,a,c) \
1881 if (is_const_Phi(a)) { \
1882 /* check for Op(Phi) */ \
1883 c = apply_unop_on_phi(a, op); \
1885 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1891 * Do the AddSub optimization, then Transform
1892 * Constant folding on Phi
1893 * Add(a,a) -> Mul(a, 2)
1894 * Add(Mul(a, x), a) -> Mul(a, x+1)
1895 * if the mode is integer or float.
1896 * Transform Add(a,-b) into Sub(a,b).
1897 * Reassociation might fold this further.
1899 static ir_node *transform_node_Add(ir_node *n) {
1901 ir_node *a, *b, *c, *oldn = n;
1903 n = transform_node_AddSub(n);
1905 a = get_Add_left(n);
1906 b = get_Add_right(n);
1908 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1910 mode = get_irn_mode(n);
1912 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1913 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1916 if (mode_is_num(mode)) {
1917 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
1918 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
1919 ir_node *block = get_irn_n(n, -1);
1922 get_irn_dbg_info(n),
1926 new_r_Const_long(current_ir_graph, block, mode, 2),
1928 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1933 get_irn_dbg_info(n),
1939 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1944 get_irn_dbg_info(n),
1950 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1953 if (! is_reassoc_running()) {
1954 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1956 ir_node *ma = get_Mul_left(a);
1957 ir_node *mb = get_Mul_right(a);
1960 ir_node *blk = get_irn_n(n, -1);
1962 get_irn_dbg_info(n), current_ir_graph, blk,
1965 get_irn_dbg_info(n), current_ir_graph, blk,
1967 new_r_Const_long(current_ir_graph, blk, mode, 1),
1970 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1972 } else if (b == mb) {
1973 ir_node *blk = get_irn_n(n, -1);
1975 get_irn_dbg_info(n), current_ir_graph, blk,
1978 get_irn_dbg_info(n), current_ir_graph, blk,
1980 new_r_Const_long(current_ir_graph, blk, mode, 1),
1983 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1988 ir_node *ma = get_Mul_left(b);
1989 ir_node *mb = get_Mul_right(b);
1992 ir_node *blk = get_irn_n(n, -1);
1994 get_irn_dbg_info(n), current_ir_graph, blk,
1997 get_irn_dbg_info(n), current_ir_graph, blk,
1999 new_r_Const_long(current_ir_graph, blk, mode, 1),
2002 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2006 ir_node *blk = get_irn_n(n, -1);
2008 get_irn_dbg_info(n), current_ir_graph, blk,
2011 get_irn_dbg_info(n), current_ir_graph, blk,
2013 new_r_Const_long(current_ir_graph, blk, mode, 1),
2016 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
2021 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2022 /* Here we rely on constants be on the RIGHT side */
2024 ir_node *op = get_Not_op(a);
2026 if (is_Const(b) && is_Const_one(b)) {
2028 ir_node *blk = get_irn_n(n, -1);
2029 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2030 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2035 ir_node *blk = get_irn_n(n, -1);
2036 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2037 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2042 ir_node *op = get_Not_op(b);
2046 ir_node *blk = get_irn_n(n, -1);
2047 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2048 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2055 } /* transform_node_Add */
2058 * returns -cnst or NULL if impossible
2060 static ir_node *const_negate(ir_node *cnst) {
2061 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2062 dbg_info *dbgi = get_irn_dbg_info(cnst);
2063 ir_graph *irg = get_irn_irg(cnst);
2064 ir_node *block = get_nodes_block(cnst);
2065 ir_mode *mode = get_irn_mode(cnst);
2066 if (tv == tarval_bad) return NULL;
2067 return new_rd_Const(dbgi, irg, block, mode, tv);
2071 * Do the AddSub optimization, then Transform
2072 * Constant folding on Phi
2073 * Sub(0,a) -> Minus(a)
2074 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2075 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2076 * Sub(Add(a, x), x) -> a
2077 * Sub(x, Add(x, a)) -> -a
2078 * Sub(x, Const) -> Add(x, -Const)
2080 static ir_node *transform_node_Sub(ir_node *n) {
2085 n = transform_node_AddSub(n);
2087 a = get_Sub_left(n);
2088 b = get_Sub_right(n);
2090 mode = get_irn_mode(n);
2093 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2095 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2096 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2099 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2100 /* a - C -> a + (-C) */
2101 ir_node *cnst = const_negate(b);
2103 ir_node *block = get_nodes_block(n);
2104 dbg_info *dbgi = get_irn_dbg_info(n);
2105 ir_graph *irg = get_irn_irg(n);
2107 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2108 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2113 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2114 ir_graph *irg = current_ir_graph;
2115 dbg_info *dbg = get_irn_dbg_info(n);
2116 ir_node *block = get_nodes_block(n);
2117 ir_node *left = get_Minus_op(a);
2118 ir_mode *mode = get_irn_mode(n);
2119 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2121 n = new_rd_Minus(dbg, irg, block, add, mode);
2122 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2124 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2125 ir_graph *irg = current_ir_graph;
2126 dbg_info *dbg = get_irn_dbg_info(n);
2127 ir_node *block = get_nodes_block(n);
2128 ir_node *right = get_Minus_op(b);
2129 ir_mode *mode = get_irn_mode(n);
2131 n = new_rd_Add(dbg, irg, block, a, right, mode);
2132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2134 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2135 ir_graph *irg = current_ir_graph;
2136 dbg_info *s_dbg = get_irn_dbg_info(b);
2137 ir_node *s_block = get_nodes_block(b);
2138 ir_node *s_left = get_Sub_right(b);
2139 ir_node *s_right = get_Sub_left(b);
2140 ir_mode *s_mode = get_irn_mode(b);
2141 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2142 dbg_info *a_dbg = get_irn_dbg_info(n);
2143 ir_node *a_block = get_nodes_block(n);
2144 ir_mode *a_mode = get_irn_mode(n);
2146 n = new_rd_Add(a_dbg, irg, a_block, a, sub, a_mode);
2147 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2149 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2150 ir_node *m_right = get_Mul_right(b);
2151 if (is_Const(m_right)) {
2152 ir_node *cnst2 = const_negate(m_right);
2153 if (cnst2 != NULL) {
2154 ir_graph *irg = current_ir_graph;
2155 dbg_info *m_dbg = get_irn_dbg_info(b);
2156 ir_node *m_block = get_nodes_block(b);
2157 ir_node *m_left = get_Mul_left(b);
2158 ir_mode *m_mode = get_irn_mode(b);
2159 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2160 dbg_info *a_dbg = get_irn_dbg_info(n);
2161 ir_node *a_block = get_nodes_block(n);
2162 ir_mode *a_mode = get_irn_mode(n);
2164 n = new_rd_Add(a_dbg, irg, a_block, a, mul, a_mode);
2165 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2171 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2172 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2174 get_irn_dbg_info(n),
2179 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2183 if (mode_wrap_around(mode)) {
2184 ir_node *left = get_Add_left(a);
2185 ir_node *right = get_Add_right(a);
2187 /* FIXME: Does the Conv's work only for two complement or generally? */
2189 if (mode != get_irn_mode(right)) {
2190 /* This Sub is an effective Cast */
2191 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2194 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2196 } else if (right == b) {
2197 if (mode != get_irn_mode(left)) {
2198 /* This Sub is an effective Cast */
2199 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2202 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2208 if (mode_wrap_around(mode)) {
2209 ir_node *left = get_Add_left(b);
2210 ir_node *right = get_Add_right(b);
2212 /* FIXME: Does the Conv's work only for two complement or generally? */
2214 ir_mode *r_mode = get_irn_mode(right);
2216 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2217 if (mode != r_mode) {
2218 /* This Sub is an effective Cast */
2219 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2221 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2223 } else if (right == a) {
2224 ir_mode *l_mode = get_irn_mode(left);
2226 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2227 if (mode != l_mode) {
2228 /* This Sub is an effective Cast */
2229 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2231 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2236 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2237 ir_mode *mode = get_irn_mode(a);
2239 if (mode == get_irn_mode(b)) {
2245 /* check if it's allowed to skip the conv */
2246 ma = get_irn_mode(a);
2247 mb = get_irn_mode(b);
2249 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2250 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2252 set_Sub_right(n, b);
2258 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2259 if (!is_reassoc_running() && is_Mul(a)) {
2260 ir_node *ma = get_Mul_left(a);
2261 ir_node *mb = get_Mul_right(a);
2264 ir_node *blk = get_irn_n(n, -1);
2266 get_irn_dbg_info(n),
2267 current_ir_graph, blk,
2270 get_irn_dbg_info(n),
2271 current_ir_graph, blk,
2273 new_r_Const_long(current_ir_graph, blk, mode, 1),
2276 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2278 } else if (mb == b) {
2279 ir_node *blk = get_irn_n(n, -1);
2281 get_irn_dbg_info(n),
2282 current_ir_graph, blk,
2285 get_irn_dbg_info(n),
2286 current_ir_graph, blk,
2288 new_r_Const_long(current_ir_graph, blk, mode, 1),
2291 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2295 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2296 ir_node *x = get_Sub_left(a);
2297 ir_node *y = get_Sub_right(a);
2298 ir_node *blk = get_irn_n(n, -1);
2299 ir_mode *m_b = get_irn_mode(b);
2300 ir_mode *m_y = get_irn_mode(y);
2304 /* Determine the right mode for the Add. */
2307 else if (mode_is_reference(m_b))
2309 else if (mode_is_reference(m_y))
2313 * Both modes are different but none is reference,
2314 * happens for instance in SubP(SubP(P, Iu), Is).
2315 * We have two possibilities here: Cast or ignore.
2316 * Currently we ignore this case.
2321 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2323 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2324 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2328 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2329 if (is_Const(a) && is_Not(b)) {
2330 /* c - ~X = X + (c+1) */
2331 tarval *tv = get_Const_tarval(a);
2333 tv = tarval_add(tv, get_mode_one(mode));
2334 if (tv != tarval_bad) {
2335 ir_node *blk = get_irn_n(n, -1);
2336 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2337 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2338 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2344 } /* transform_node_Sub */
2347 * Several transformation done on n*n=2n bits mul.
2348 * These transformations must be done here because new nodes may be produced.
2350 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2352 ir_node *a = get_Mul_left(n);
2353 ir_node *b = get_Mul_right(n);
2354 tarval *ta = value_of(a);
2355 tarval *tb = value_of(b);
2356 ir_mode *smode = get_irn_mode(a);
2358 if (ta == get_mode_one(smode)) {
2359 ir_node *blk = get_irn_n(n, -1);
2360 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2361 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2364 else if (ta == get_mode_minus_one(smode)) {
2365 ir_node *blk = get_irn_n(n, -1);
2366 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2367 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2368 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2371 if (tb == get_mode_one(smode)) {
2372 ir_node *blk = get_irn_n(a, -1);
2373 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2374 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2377 else if (tb == get_mode_minus_one(smode)) {
2378 ir_node *blk = get_irn_n(n, -1);
2379 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2380 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2381 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2388 * Transform Mul(a,-1) into -a.
2389 * Do constant evaluation of Phi nodes.
2390 * Do architecture dependent optimizations on Mul nodes
2392 static ir_node *transform_node_Mul(ir_node *n) {
2393 ir_node *c, *oldn = n;
2394 ir_mode *mode = get_irn_mode(n);
2395 ir_node *a = get_Mul_left(n);
2396 ir_node *b = get_Mul_right(n);
2398 if (is_Bad(a) || is_Bad(b))
2401 if (mode != get_irn_mode(a))
2402 return transform_node_Mul2n(n, mode);
2404 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2406 if (mode_is_signed(mode)) {
2409 if (value_of(a) == get_mode_minus_one(mode))
2411 else if (value_of(b) == get_mode_minus_one(mode))
2414 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2415 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2420 if (is_Const(b)) { /* (-a) * const -> a * -const */
2421 ir_node *cnst = const_negate(b);
2423 dbg_info *dbgi = get_irn_dbg_info(n);
2424 ir_node *block = get_nodes_block(n);
2425 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2426 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2429 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2430 dbg_info *dbgi = get_irn_dbg_info(n);
2431 ir_node *block = get_nodes_block(n);
2432 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2433 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2435 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2436 ir_node *sub_l = get_Sub_left(b);
2437 ir_node *sub_r = get_Sub_right(b);
2438 dbg_info *dbgi = get_irn_dbg_info(n);
2439 ir_graph *irg = current_ir_graph;
2440 ir_node *block = get_nodes_block(n);
2441 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2442 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2443 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2446 } else if (is_Minus(b)) {
2447 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2448 ir_node *sub_l = get_Sub_left(a);
2449 ir_node *sub_r = get_Sub_right(a);
2450 dbg_info *dbgi = get_irn_dbg_info(n);
2451 ir_graph *irg = current_ir_graph;
2452 ir_node *block = get_nodes_block(n);
2453 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2454 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2455 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2459 if (get_mode_arithmetic(mode) == irma_ieee754) {
2461 tarval *tv = get_Const_tarval(a);
2462 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2463 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), b, b, mode);
2464 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2468 else if (is_Const(b)) {
2469 tarval *tv = get_Const_tarval(b);
2470 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2471 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, a, mode);
2472 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2477 return arch_dep_replace_mul_with_shifts(n);
2478 } /* transform_node_Mul */
2481 * Transform a Div Node.
2483 static ir_node *transform_node_Div(ir_node *n) {
2484 tarval *tv = value_of(n);
2485 ir_mode *mode = get_Div_resmode(n);
2488 if (tv != tarval_bad) {
2489 value = new_Const(get_tarval_mode(tv), tv);
2491 DBG_OPT_CSTEVAL(n, value);
2494 ir_node *a = get_Div_left(n);
2495 ir_node *b = get_Div_right(n);
2498 if (a == b && value_not_zero(a, &dummy)) {
2499 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2500 value = new_Const(mode, get_mode_one(mode));
2501 DBG_OPT_CSTEVAL(n, value);
2504 if (mode_is_signed(mode) && is_Const(b)) {
2505 tarval *tv = get_Const_tarval(b);
2507 if (tv == get_mode_minus_one(mode)) {
2509 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2510 DBG_OPT_CSTEVAL(n, value);
2514 /* Try architecture dependent optimization */
2515 value = arch_dep_replace_div_by_const(n);
2523 /* Turn Div into a tuple (mem, jmp, bad, value) */
2524 mem = get_Div_mem(n);
2525 blk = get_irn_n(n, -1);
2527 /* skip a potential Pin */
2529 mem = get_Pin_op(mem);
2530 turn_into_tuple(n, pn_Div_max);
2531 set_Tuple_pred(n, pn_Div_M, mem);
2532 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2533 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2534 set_Tuple_pred(n, pn_Div_res, value);
2537 } /* transform_node_Div */
2540 * Transform a Mod node.
2542 static ir_node *transform_node_Mod(ir_node *n) {
2543 tarval *tv = value_of(n);
2544 ir_mode *mode = get_Mod_resmode(n);
2547 if (tv != tarval_bad) {
2548 value = new_Const(get_tarval_mode(tv), tv);
2550 DBG_OPT_CSTEVAL(n, value);
2553 ir_node *a = get_Mod_left(n);
2554 ir_node *b = get_Mod_right(n);
2557 if (a == b && value_not_zero(a, &dummy)) {
2558 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2559 value = new_Const(mode, get_mode_null(mode));
2560 DBG_OPT_CSTEVAL(n, value);
2563 if (mode_is_signed(mode) && is_Const(b)) {
2564 tarval *tv = get_Const_tarval(b);
2566 if (tv == get_mode_minus_one(mode)) {
2568 value = new_Const(mode, get_mode_null(mode));
2569 DBG_OPT_CSTEVAL(n, value);
2573 /* Try architecture dependent optimization */
2574 value = arch_dep_replace_mod_by_const(n);
2582 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2583 mem = get_Mod_mem(n);
2584 blk = get_irn_n(n, -1);
2586 /* skip a potential Pin */
2588 mem = get_Pin_op(mem);
2589 turn_into_tuple(n, pn_Mod_max);
2590 set_Tuple_pred(n, pn_Mod_M, mem);
2591 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2592 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2593 set_Tuple_pred(n, pn_Mod_res, value);
2596 } /* transform_node_Mod */
2599 * Transform a DivMod node.
2601 static ir_node *transform_node_DivMod(ir_node *n) {
2603 ir_node *a = get_DivMod_left(n);
2604 ir_node *b = get_DivMod_right(n);
2605 ir_mode *mode = get_DivMod_resmode(n);
2606 tarval *ta = value_of(a);
2607 tarval *tb = value_of(b);
2610 if (tb != tarval_bad) {
2611 if (tb == get_mode_one(get_tarval_mode(tb))) {
2612 b = new_Const(mode, get_mode_null(mode));
2613 DBG_OPT_CSTEVAL(n, b);
2615 } else if (ta != tarval_bad) {
2616 tarval *resa, *resb;
2617 resa = tarval_div(ta, tb);
2618 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2619 Jmp for X result!? */
2620 resb = tarval_mod(ta, tb);
2621 if (resb == tarval_bad) return n; /* Causes exception! */
2622 a = new_Const(mode, resa);
2623 b = new_Const(mode, resb);
2624 DBG_OPT_CSTEVAL(n, a);
2625 DBG_OPT_CSTEVAL(n, b);
2627 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2628 a = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2629 b = new_Const(mode, get_mode_null(mode));
2630 DBG_OPT_CSTEVAL(n, a);
2631 DBG_OPT_CSTEVAL(n, b);
2633 } else { /* Try architecture dependent optimization */
2634 arch_dep_replace_divmod_by_const(&a, &b, n);
2635 evaluated = a != NULL;
2637 } else if (a == b) {
2638 if (value_not_zero(a, &dummy)) {
2640 a = new_Const(mode, get_mode_one(mode));
2641 b = new_Const(mode, get_mode_null(mode));
2642 DBG_OPT_CSTEVAL(n, a);
2643 DBG_OPT_CSTEVAL(n, b);
2646 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2649 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2650 /* 0 / non-Const = 0 */
2655 if (evaluated) { /* replace by tuple */
2659 mem = get_DivMod_mem(n);
2660 /* skip a potential Pin */
2662 mem = get_Pin_op(mem);
2664 blk = get_irn_n(n, -1);
2665 turn_into_tuple(n, pn_DivMod_max);
2666 set_Tuple_pred(n, pn_DivMod_M, mem);
2667 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2668 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2669 set_Tuple_pred(n, pn_DivMod_res_div, a);
2670 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2674 } /* transform_node_DivMod */
2677 * Optimize x / c to x * (1/c)
2679 static ir_node *transform_node_Quot(ir_node *n) {
2680 ir_mode *mode = get_Quot_resmode(n);
2683 if (get_mode_arithmetic(mode) == irma_ieee754) {
2684 ir_node *b = get_Quot_right(n);
2687 tarval *tv = get_Const_tarval(b);
2689 tv = tarval_quo(get_mode_one(mode), tv);
2691 /* Do the transformation if the result is either exact or we are not
2692 using strict rules. */
2693 if (tv != tarval_bad &&
2694 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2695 ir_node *blk = get_irn_n(n, -1);
2696 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2697 ir_node *a = get_Quot_left(n);
2698 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2699 ir_node *mem = get_Quot_mem(n);
2701 /* skip a potential Pin */
2703 mem = get_Pin_op(mem);
2704 turn_into_tuple(n, pn_Quot_max);
2705 set_Tuple_pred(n, pn_Quot_M, mem);
2706 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2707 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2708 set_Tuple_pred(n, pn_Quot_res, m);
2709 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2714 } /* transform_node_Quot */
2717 * Optimize Abs(x) into x if x is Confirmed >= 0
2718 * Optimize Abs(x) into -x if x is Confirmed <= 0
2720 static ir_node *transform_node_Abs(ir_node *n) {
2722 ir_node *a = get_Abs_op(n);
2723 value_classify_sign sign = classify_value_sign(a);
2725 if (sign == value_classified_negative) {
2726 ir_mode *mode = get_irn_mode(n);
2729 * We can replace the Abs by -x here.
2730 * We even could add a new Confirm here.
2732 * Note that -x would create a new node, so we could
2733 * not run it in the equivalent_node() context.
2735 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2736 get_irn_n(n, -1), a, mode);
2738 DBG_OPT_CONFIRM(oldn, n);
2739 } else if (sign == value_classified_positive) {
2740 /* n is positive, Abs is not needed */
2743 DBG_OPT_CONFIRM(oldn, n);
2747 } /* transform_node_Abs */
2750 * Transform a Cond node.
2752 * Replace the Cond by a Jmp if it branches on a constant
2755 static ir_node *transform_node_Cond(ir_node *n) {
2758 ir_node *a = get_Cond_selector(n);
2759 tarval *ta = value_of(a);
2761 /* we need block info which is not available in floating irgs */
2762 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2765 if ((ta != tarval_bad) &&
2766 (get_irn_mode(a) == mode_b) &&
2767 (get_opt_unreachable_code())) {
2768 /* It's a boolean Cond, branching on a boolean constant.
2769 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2770 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2771 turn_into_tuple(n, pn_Cond_max);
2772 if (ta == tarval_b_true) {
2773 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2774 set_Tuple_pred(n, pn_Cond_true, jmp);
2776 set_Tuple_pred(n, pn_Cond_false, jmp);
2777 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2779 /* We might generate an endless loop, so keep it alive. */
2780 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2783 } /* transform_node_Cond */
2785 typedef ir_node* (*recursive_transform) (ir_node *n);
2788 * makes use of distributive laws for and, or, eor
2789 * and(a OP c, b OP c) -> and(a, b) OP c
2790 * note, might return a different op than n
2792 static ir_node *transform_bitwise_distributive(ir_node *n,
2793 recursive_transform trans_func)
2796 ir_node *a = get_binop_left(n);
2797 ir_node *b = get_binop_right(n);
2798 ir_op *op = get_irn_op(a);
2799 ir_op *op_root = get_irn_op(n);
2801 if(op != get_irn_op(b))
2804 if (op == op_Conv) {
2805 ir_node *a_op = get_Conv_op(a);
2806 ir_node *b_op = get_Conv_op(b);
2807 ir_mode *a_mode = get_irn_mode(a_op);
2808 ir_mode *b_mode = get_irn_mode(b_op);
2809 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2810 ir_node *blk = get_irn_n(n, -1);
2813 set_binop_left(n, a_op);
2814 set_binop_right(n, b_op);
2815 set_irn_mode(n, a_mode);
2817 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2819 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2825 /* nothing to gain here */
2829 if (op == op_Shrs || op == op_Shr || op == op_Shl
2830 || op == op_And || op == op_Or || op == op_Eor) {
2831 ir_node *a_left = get_binop_left(a);
2832 ir_node *a_right = get_binop_right(a);
2833 ir_node *b_left = get_binop_left(b);
2834 ir_node *b_right = get_binop_right(b);
2838 if (is_op_commutative(op)) {
2839 if (a_left == b_left) {
2843 } else if(a_left == b_right) {
2847 } else if(a_right == b_left) {
2853 if(a_right == b_right) {
2860 /* (a sop c) & (b sop c) => (a & b) sop c */
2861 ir_node *blk = get_irn_n(n, -1);
2863 ir_node *new_n = exact_copy(n);
2864 set_binop_left(new_n, op1);
2865 set_binop_right(new_n, op2);
2866 new_n = trans_func(new_n);
2868 if(op_root == op_Eor && op == op_Or) {
2869 dbg_info *dbgi = get_irn_dbg_info(n);
2870 ir_graph *irg = current_ir_graph;
2871 ir_mode *mode = get_irn_mode(c);
2873 c = new_rd_Not(dbgi, irg, blk, c, mode);
2874 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
2877 set_irn_n(n, -1, blk);
2878 set_binop_left(n, new_n);
2879 set_binop_right(n, c);
2882 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2893 static ir_node *transform_node_And(ir_node *n) {
2894 ir_node *c, *oldn = n;
2895 ir_node *a = get_And_left(n);
2896 ir_node *b = get_And_right(n);
2899 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2901 mode = get_irn_mode(n);
2903 /* we can evaluate 2 Projs of the same Cmp */
2904 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
2905 ir_node *pred_a = get_Proj_pred(a);
2906 ir_node *pred_b = get_Proj_pred(b);
2907 if (pred_a == pred_b) {
2908 dbg_info *dbgi = get_irn_dbg_info(n);
2909 ir_node *block = get_nodes_block(pred_a);
2910 pn_Cmp pn_a = get_Proj_proj(a);
2911 pn_Cmp pn_b = get_Proj_proj(b);
2912 /* yes, we can simply calculate with pncs */
2913 pn_Cmp new_pnc = pn_a & pn_b;
2915 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
2920 ir_node *op = get_Not_op(b);
2922 ir_node *ba = get_And_left(op);
2923 ir_node *bb = get_And_right(op);
2925 /* it's enough to test the following cases due to normalization! */
2926 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
2927 /* (a|b) & ~(a&b) = a^b */
2928 ir_node *block = get_nodes_block(n);
2930 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
2931 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
2939 ir_node *op = get_Not_op(a);
2941 ir_node *aa = get_And_left(op);
2942 ir_node *ab = get_And_right(op);
2944 /* it's enough to test the following cases due to normalization! */
2945 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
2946 /* (a|b) & ~(a&b) = a^b */
2947 ir_node *block = get_nodes_block(n);
2949 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
2950 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
2957 ir_node *al = get_Eor_left(a);
2958 ir_node *ar = get_Eor_right(a);
2961 /* (b ^ a) & b -> ~a & b */
2962 dbg_info *dbg = get_irn_dbg_info(n);
2963 ir_node *block = get_nodes_block(n);
2965 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
2966 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
2967 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2971 /* (a ^ b) & b -> ~a & b */
2972 dbg_info *dbg = get_irn_dbg_info(n);
2973 ir_node *block = get_nodes_block(n);
2975 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
2976 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
2977 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2982 ir_node *bl = get_Eor_left(b);
2983 ir_node *br = get_Eor_right(b);
2986 /* a & (a ^ b) -> a & ~b */
2987 dbg_info *dbg = get_irn_dbg_info(n);
2988 ir_node *block = get_nodes_block(n);
2990 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
2991 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
2992 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2996 /* a & (b ^ a) -> a & ~b */
2997 dbg_info *dbg = get_irn_dbg_info(n);
2998 ir_node *block = get_nodes_block(n);
3000 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3001 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3002 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3006 if (is_Not(a) && is_Not(b)) {
3007 /* ~a & ~b = ~(a|b) */
3008 ir_node *block = get_nodes_block(n);
3009 ir_mode *mode = get_irn_mode(n);
3013 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3014 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3015 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3019 n = transform_bitwise_distributive(n, transform_node_And);
3022 } /* transform_node_And */
3027 static ir_node *transform_node_Eor(ir_node *n) {
3028 ir_node *c, *oldn = n;
3029 ir_node *a = get_Eor_left(n);
3030 ir_node *b = get_Eor_right(n);
3031 ir_mode *mode = get_irn_mode(n);
3033 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
3035 /* we can evaluate 2 Projs of the same Cmp */
3036 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3037 ir_node *pred_a = get_Proj_pred(a);
3038 ir_node *pred_b = get_Proj_pred(b);
3039 if(pred_a == pred_b) {
3040 dbg_info *dbgi = get_irn_dbg_info(n);
3041 ir_node *block = get_nodes_block(pred_a);
3042 pn_Cmp pn_a = get_Proj_proj(a);
3043 pn_Cmp pn_b = get_Proj_proj(b);
3044 /* yes, we can simply calculate with pncs */
3045 pn_Cmp new_pnc = pn_a ^ pn_b;
3047 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3054 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3055 mode, get_mode_null(mode));
3056 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3057 } else if (mode == mode_b &&
3059 is_Const(b) && is_Const_one(b) &&
3060 is_Cmp(get_Proj_pred(a))) {
3061 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3062 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3063 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3065 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3066 } else if (is_Const(b)) {
3067 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3068 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3069 ir_node *not_op = get_Not_op(a);
3070 dbg_info *dbg = get_irn_dbg_info(n);
3071 ir_graph *irg = current_ir_graph;
3072 ir_node *block = get_nodes_block(n);
3073 ir_mode *mode = get_irn_mode(n);
3074 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3076 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3077 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3078 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3081 n = transform_bitwise_distributive(n, transform_node_Eor);
3085 } /* transform_node_Eor */
3090 static ir_node *transform_node_Not(ir_node *n) {
3091 ir_node *c, *oldn = n;
3092 ir_node *a = get_Not_op(n);
3093 ir_mode *mode = get_irn_mode(n);
3095 HANDLE_UNOP_PHI(tarval_not,a,c);
3097 /* check for a boolean Not */
3098 if (mode == mode_b &&
3100 is_Cmp(get_Proj_pred(a))) {
3101 /* We negate a Cmp. The Cmp has the negated result anyways! */
3102 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3103 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3104 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3108 ir_node *eor_b = get_Eor_right(a);
3109 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3110 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3111 ir_node *eor_a = get_Eor_left(a);
3112 dbg_info *dbg = get_irn_dbg_info(n);
3113 ir_graph *irg = current_ir_graph;
3114 ir_node *block = get_nodes_block(n);
3115 ir_mode *mode = get_irn_mode(n);
3116 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3120 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3121 if (is_Minus(a)) { /* ~-x -> x + -1 */
3122 dbg_info *dbg = get_irn_dbg_info(n);
3123 ir_graph *irg = current_ir_graph;
3124 ir_node *block = get_nodes_block(n);
3125 ir_node *add_l = get_Minus_op(a);
3126 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3127 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3128 } else if (is_Add(a)) {
3129 ir_node *add_r = get_Add_right(a);
3130 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3131 /* ~(x + -1) = -x */
3132 ir_node *op = get_Add_left(a);
3133 ir_node *blk = get_irn_n(n, -1);
3134 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3135 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3140 } /* transform_node_Not */
3143 * Transform a Minus.
3148 static ir_node *transform_node_Minus(ir_node *n) {
3149 ir_node *c, *oldn = n;
3150 ir_node *a = get_Minus_op(n);
3153 HANDLE_UNOP_PHI(tarval_neg,a,c);
3155 mode = get_irn_mode(a);
3156 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3157 /* the following rules are only to twos-complement */
3160 ir_node *op = get_Not_op(a);
3161 tarval *tv = get_mode_one(mode);
3162 ir_node *blk = get_irn_n(n, -1);
3163 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3164 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3165 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3169 ir_node *c = get_Shr_right(a);
3172 tarval *tv = get_Const_tarval(c);
3174 if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) {
3175 /* -(a >>u (size-1)) = a >>s (size-1) */
3176 ir_node *v = get_Shr_left(a);
3178 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3179 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3185 ir_node *c = get_Shrs_right(a);
3188 tarval *tv = get_Const_tarval(c);
3190 if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) {
3191 /* -(a >>s (size-1)) = a >>u (size-1) */
3192 ir_node *v = get_Shrs_left(a);
3194 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3195 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3202 /* - (a-b) = b - a */
3203 ir_node *la = get_Sub_left(a);
3204 ir_node *ra = get_Sub_right(a);
3205 ir_node *blk = get_irn_n(n, -1);
3207 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3208 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3212 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3213 ir_node *mul_l = get_Mul_left(a);
3214 ir_node *mul_r = get_Mul_right(a);
3215 if (is_Const(mul_r)) {
3216 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3217 ir_node *cnst = new_Const(mode, tv);
3218 dbg_info *dbg = get_irn_dbg_info(a);
3219 ir_graph *irg = current_ir_graph;
3220 ir_node *block = get_nodes_block(a);
3221 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3227 } /* transform_node_Minus */
3230 * Transform a Cast_type(Const) into a new Const_type
3232 static ir_node *transform_node_Cast(ir_node *n) {
3234 ir_node *pred = get_Cast_op(n);
3235 ir_type *tp = get_irn_type(n);
3237 if (is_Const(pred) && get_Const_type(pred) != tp) {
3238 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3239 get_Const_tarval(pred), tp);
3240 DBG_OPT_CSTEVAL(oldn, n);
3241 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3242 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
3243 get_SymConst_kind(pred), tp);
3244 DBG_OPT_CSTEVAL(oldn, n);
3248 } /* transform_node_Cast */
3251 * Transform a Proj(Div) with a non-zero value.
3252 * Removes the exceptions and routes the memory to the NoMem node.
3254 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3255 ir_node *div = get_Proj_pred(proj);
3256 ir_node *b = get_Div_right(div);
3257 ir_node *confirm, *res, *new_mem;
3260 if (value_not_zero(b, &confirm)) {
3261 /* div(x, y) && y != 0 */
3262 if (confirm == NULL) {
3263 /* we are sure we have a Const != 0 */
3264 new_mem = get_Div_mem(div);
3265 if (is_Pin(new_mem))
3266 new_mem = get_Pin_op(new_mem);
3267 set_Div_mem(div, new_mem);
3268 set_irn_pinned(div, op_pin_state_floats);
3271 proj_nr = get_Proj_proj(proj);
3273 case pn_Div_X_regular:
3274 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3276 case pn_Div_X_except:
3277 /* we found an exception handler, remove it */
3278 DBG_OPT_EXC_REM(proj);
3282 res = get_Div_mem(div);
3283 new_mem = get_irg_no_mem(current_ir_graph);
3286 /* This node can only float up to the Confirm block */
3287 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3289 set_irn_pinned(div, op_pin_state_floats);
3290 /* this is a Div without exception, we can remove the memory edge */
3291 set_Div_mem(div, new_mem);
3296 } /* transform_node_Proj_Div */
3299 * Transform a Proj(Mod) with a non-zero value.
3300 * Removes the exceptions and routes the memory to the NoMem node.
3302 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3303 ir_node *mod = get_Proj_pred(proj);
3304 ir_node *b = get_Mod_right(mod);
3305 ir_node *confirm, *res, *new_mem;
3308 if (value_not_zero(b, &confirm)) {
3309 /* mod(x, y) && y != 0 */
3310 proj_nr = get_Proj_proj(proj);
3312 if (confirm == NULL) {
3313 /* we are sure we have a Const != 0 */
3314 new_mem = get_Mod_mem(mod);
3315 if (is_Pin(new_mem))
3316 new_mem = get_Pin_op(new_mem);
3317 set_Mod_mem(mod, new_mem);
3318 set_irn_pinned(mod, op_pin_state_floats);
3323 case pn_Mod_X_regular:
3324 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3326 case pn_Mod_X_except:
3327 /* we found an exception handler, remove it */
3328 DBG_OPT_EXC_REM(proj);
3332 res = get_Mod_mem(mod);
3333 new_mem = get_irg_no_mem(current_ir_graph);
3336 /* This node can only float up to the Confirm block */
3337 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3339 /* this is a Mod without exception, we can remove the memory edge */
3340 set_Mod_mem(mod, new_mem);
3343 if (get_Mod_left(mod) == b) {
3344 /* a % a = 0 if a != 0 */
3345 ir_mode *mode = get_irn_mode(proj);
3346 ir_node *res = new_Const(mode, get_mode_null(mode));
3348 DBG_OPT_CSTEVAL(mod, res);
3354 } /* transform_node_Proj_Mod */
3357 * Transform a Proj(DivMod) with a non-zero value.
3358 * Removes the exceptions and routes the memory to the NoMem node.
3360 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3361 ir_node *divmod = get_Proj_pred(proj);
3362 ir_node *b = get_DivMod_right(divmod);
3363 ir_node *confirm, *res, *new_mem;
3366 if (value_not_zero(b, &confirm)) {
3367 /* DivMod(x, y) && y != 0 */
3368 proj_nr = get_Proj_proj(proj);
3370 if (confirm == NULL) {
3371 /* we are sure we have a Const != 0 */
3372 new_mem = get_DivMod_mem(divmod);
3373 if (is_Pin(new_mem))
3374 new_mem = get_Pin_op(new_mem);
3375 set_DivMod_mem(divmod, new_mem);
3376 set_irn_pinned(divmod, op_pin_state_floats);
3381 case pn_DivMod_X_regular:
3382 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3384 case pn_DivMod_X_except:
3385 /* we found an exception handler, remove it */
3386 DBG_OPT_EXC_REM(proj);
3390 res = get_DivMod_mem(divmod);
3391 new_mem = get_irg_no_mem(current_ir_graph);
3394 /* This node can only float up to the Confirm block */
3395 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3397 /* this is a DivMod without exception, we can remove the memory edge */
3398 set_DivMod_mem(divmod, new_mem);
3401 case pn_DivMod_res_mod:
3402 if (get_DivMod_left(divmod) == b) {
3403 /* a % a = 0 if a != 0 */
3404 ir_mode *mode = get_irn_mode(proj);
3405 ir_node *res = new_Const(mode, get_mode_null(mode));
3407 DBG_OPT_CSTEVAL(divmod, res);
3413 } /* transform_node_Proj_DivMod */
3416 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3418 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3419 if (get_opt_unreachable_code()) {
3420 ir_node *n = get_Proj_pred(proj);
3421 ir_node *b = get_Cond_selector(n);
3423 if (mode_is_int(get_irn_mode(b))) {
3424 tarval *tb = value_of(b);
3426 if (tb != tarval_bad) {
3427 /* we have a constant switch */
3428 long num = get_Proj_proj(proj);
3430 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3431 if (get_tarval_long(tb) == num) {
3432 /* Do NOT create a jump here, or we will have 2 control flow ops
3433 * in a block. This case is optimized away in optimize_cf(). */
3436 /* this case will NEVER be taken, kill it */
3444 } /* transform_node_Proj_Cond */
3447 * Normalizes and optimizes Cmp nodes.
3449 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3450 ir_node *n = get_Proj_pred(proj);
3451 ir_node *left = get_Cmp_left(n);
3452 ir_node *right = get_Cmp_right(n);
3456 ir_mode *mode = NULL;
3457 long proj_nr = get_Proj_proj(proj);
3459 /* we can evaluate this direct */
3462 return new_Const(mode_b, get_tarval_b_false());
3464 return new_Const(mode_b, get_tarval_b_true());
3466 if(!mode_is_float(get_irn_mode(left)))
3467 return new_Const(mode_b, get_tarval_b_true());
3475 left = get_Cast_op(left);
3477 right = get_Cast_op(right);
3479 /* Remove unnecessary conversions */
3480 /* TODO handle constants */
3481 if (is_Conv(left) && is_Conv(right)) {
3482 ir_mode *mode = get_irn_mode(left);
3483 ir_node *op_left = get_Conv_op(left);
3484 ir_node *op_right = get_Conv_op(right);
3485 ir_mode *mode_left = get_irn_mode(op_left);
3486 ir_mode *mode_right = get_irn_mode(op_right);
3488 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)) {
3489 ir_graph *irg = current_ir_graph;
3490 ir_node *block = get_nodes_block(n);
3492 if (mode_left == mode_right) {
3496 } else if (smaller_mode(mode_left, mode_right)) {
3497 left = new_r_Conv(irg, block, op_left, mode_right);
3500 } else if (smaller_mode(mode_right, mode_left)) {
3502 right = new_r_Conv(irg, block, op_right, mode_left);
3508 /* TODO extend to arbitrary constants */
3509 if (is_Conv(left) && is_Const(right) && is_Const_null(right)) {
3510 ir_mode* mode = get_irn_mode(left);
3511 ir_node* op = get_Conv_op(left);
3512 ir_mode* op_mode = get_irn_mode(op);
3514 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3515 (mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3516 ir_node *null = new_Const(op_mode, get_mode_null(op_mode));
3517 set_Cmp_left( n, op);
3518 set_Cmp_right(n, null);
3523 /* remove operation of both sides if possible */
3524 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3525 ir_opcode lop = get_irn_opcode(left);
3527 if (lop == get_irn_opcode(right)) {
3528 ir_node *ll, *lr, *rl, *rr;
3530 /* same operation on both sides, try to remove */
3534 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3535 left = get_unop_op(left);
3536 right = get_unop_op(right);
3540 ll = get_Add_left(left);
3541 lr = get_Add_right(left);
3542 rl = get_Add_left(right);
3543 rr = get_Add_right(right);
3546 /* X + a CMP X + b ==> a CMP b */
3550 } else if (ll == rr) {
3551 /* X + a CMP b + X ==> a CMP b */
3555 } else if (lr == rl) {
3556 /* a + X CMP X + b ==> a CMP b */
3560 } else if (lr == rr) {
3561 /* a + X CMP b + X ==> a CMP b */
3568 ll = get_Sub_left(left);
3569 lr = get_Sub_right(left);
3570 rl = get_Sub_left(right);
3571 rr = get_Sub_right(right);
3574 /* X - a CMP X - b ==> a CMP b */
3578 } else if (lr == rr) {
3579 /* a - X CMP b - X ==> a CMP b */
3586 if (get_Rot_right(left) == get_Rot_right(right)) {
3587 /* a ROT X CMP b ROT X */
3588 left = get_Rot_left(left);
3589 right = get_Rot_left(right);
3599 if (get_irn_mode(left) == mode_b) {
3600 ir_graph *irg = current_ir_graph;
3601 ir_node *block = get_nodes_block(n);
3604 case pn_Cmp_Le: return new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b);
3605 case pn_Cmp_Lt: return new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b);
3606 case pn_Cmp_Ge: return new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b);
3607 case pn_Cmp_Gt: return new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b);
3608 case pn_Cmp_Lg: return new_r_Eor(irg, block, left, right, mode_b);
3609 case pn_Cmp_Eq: return new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b);
3613 if (!get_opt_reassociation())
3617 * First step: normalize the compare op
3618 * by placing the constant on the right side
3619 * or moving the lower address node to the left.
3620 * We ignore the case that both are constants
3621 * this case should be optimized away.
3623 if (is_Const(right)) {
3625 } else if (is_Const(left)) {
3630 proj_nr = get_inversed_pnc(proj_nr);
3632 } else if (get_irn_idx(left) > get_irn_idx(right)) {
3638 proj_nr = get_inversed_pnc(proj_nr);
3643 * Second step: Try to reduce the magnitude
3644 * of a constant. This may help to generate better code
3645 * later and may help to normalize more compares.
3646 * Of course this is only possible for integer values.
3649 mode = get_irn_mode(c);
3650 tv = get_Const_tarval(c);
3652 if (tv != tarval_bad) {
3653 /* the following optimization is possible on modes without Overflow
3654 * on Unary Minus or on == and !=:
3655 * -a CMP c ==> a swap(CMP) -c
3657 * Beware: for two-complement Overflow may occur, so only == and != can
3658 * be optimized, see this:
3659 * -MININT < 0 =/=> MININT > 0 !!!
3661 if (is_Minus(left) &&
3662 (!mode_overflow_on_unary_Minus(mode) ||
3663 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3664 tv = tarval_neg(tv);
3666 if (tv != tarval_bad) {
3667 left = get_Minus_op(left);
3668 proj_nr = get_inversed_pnc(proj_nr);
3671 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3672 tv = tarval_not(tv);
3674 if (tv != tarval_bad) {
3675 left = get_Not_op(left);
3680 /* for integer modes, we have more */
3681 if (mode_is_int(mode)) {
3682 /* Ne includes Unordered which is not possible on integers.
3683 * However, frontends often use this wrong, so fix it here */
3684 if (proj_nr & pn_Cmp_Uo) {
3685 proj_nr &= ~pn_Cmp_Uo;
3686 set_Proj_proj(proj, proj_nr);
3689 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3690 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3691 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3692 tv = tarval_sub(tv, get_mode_one(mode));
3694 if (tv != tarval_bad) {
3695 proj_nr ^= pn_Cmp_Eq;
3699 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3700 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3701 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3702 tv = tarval_add(tv, get_mode_one(mode));
3704 if (tv != tarval_bad) {
3705 proj_nr ^= pn_Cmp_Eq;
3710 /* the following reassociations work only for == and != */
3711 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3713 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
3714 if (tarval_is_null(tv) && is_Sub(left)) {
3715 right = get_Sub_right(left);
3716 left = get_Sub_left(left);
3718 tv = value_of(right);
3722 if (tv != tarval_bad) {
3723 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
3725 ir_node *c1 = get_Sub_right(left);
3726 tarval *tv2 = value_of(c1);
3728 if (tv2 != tarval_bad) {
3729 tv2 = tarval_add(tv, value_of(c1));
3731 if (tv2 != tarval_bad) {
3732 left = get_Sub_left(left);
3738 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
3739 else if (is_Add(left)) {
3740 ir_node *a_l = get_Add_left(left);
3741 ir_node *a_r = get_Add_right(left);
3745 if (is_Const(a_l)) {
3747 tv2 = value_of(a_l);
3750 tv2 = value_of(a_r);
3753 if (tv2 != tarval_bad) {
3754 tv2 = tarval_sub(tv, tv2);
3756 if (tv2 != tarval_bad) {
3763 /* -a == c ==> a == -c, -a != c ==> a != -c */
3764 else if (is_Minus(left)) {
3765 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
3767 if (tv2 != tarval_bad) {
3768 left = get_Minus_op(left);
3775 /* the following reassociations work only for <= */
3776 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3777 if (tv != tarval_bad) {
3778 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
3779 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
3785 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3786 switch (get_irn_opcode(left)) {
3791 * optimization for AND:
3793 * And(x, C) == C ==> And(x, C) != 0
3794 * And(x, C) != C ==> And(X, C) == 0
3796 * if C is a single Bit constant.
3798 if (tarval_is_single_bit(tv)) {
3799 /* check for Constant's match. We have check hare the tarvals,
3800 because our const might be changed */
3801 ir_node *ra = get_And_right(left);
3802 /* beware, tv might be != c here */
3803 if (is_Const(ra) && get_Const_tarval(ra) == tv) {
3804 /* fine: do the transformation */
3805 tv = get_mode_null(get_tarval_mode(tv));
3806 proj_nr ^= pn_Cmp_Leg;
3813 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
3815 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
3818 c1 = get_Shl_right(left);
3820 tarval *tv1 = get_Const_tarval(c1);
3821 ir_mode *mode = get_irn_mode(left);
3822 tarval *minus1 = get_mode_all_one(mode);
3823 tarval *amask = tarval_shr(minus1, tv1);
3824 tarval *cmask = tarval_shl(minus1, tv1);
3827 if (tarval_and(tv, cmask) != tv) {
3828 /* condition not met */
3829 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
3830 return new_Const(mode_b, tv);
3832 sl = get_Shl_left(left);
3833 blk = get_nodes_block(n);
3834 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
3835 tv = tarval_shr(tv, tv1);
3841 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
3843 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
3846 c1 = get_Shr_right(left);
3848 tarval *tv1 = get_Const_tarval(c1);
3849 ir_mode *mode = get_irn_mode(left);
3850 tarval *minus1 = get_mode_all_one(mode);
3851 tarval *amask = tarval_shl(minus1, tv1);
3852 tarval *cmask = tarval_shr(minus1, tv1);
3855 if (tarval_and(tv, cmask) != tv) {
3856 /* condition not met */
3857 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
3858 return new_Const(mode_b, tv);
3860 sl = get_Shr_left(left);
3861 blk = get_nodes_block(n);
3862 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
3863 tv = tarval_shl(tv, tv1);
3869 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
3871 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
3874 c1 = get_Shrs_right(left);
3876 tarval *tv1 = get_Const_tarval(c1);
3877 ir_mode *mode = get_irn_mode(left);
3878 tarval *minus1 = get_mode_all_one(mode);
3879 tarval *amask = tarval_shl(minus1, tv1);
3880 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
3883 cond = tarval_sub(cond, tv1);
3884 cond = tarval_shrs(tv, cond);
3886 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
3887 /* condition not met */
3888 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
3889 return new_Const(mode_b, tv);
3891 sl = get_Shrs_left(left);
3892 blk = get_nodes_block(n);
3893 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
3894 tv = tarval_shl(tv, tv1);
3900 } /* tarval != bad */
3903 if (changed & 2) /* need a new Const */
3904 right = new_Const(mode, tv);
3906 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
3907 ir_node *op = get_Proj_pred(left);
3909 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
3910 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
3911 ir_node *c = get_binop_right(op);
3914 tarval *tv = get_Const_tarval(c);
3916 if (tarval_is_single_bit(tv)) {
3917 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
3918 ir_node *v = get_binop_left(op);
3919 ir_node *blk = get_irn_n(op, -1);
3920 ir_mode *mode = get_irn_mode(v);
3922 tv = tarval_sub(tv, get_mode_one(mode));
3923 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
3931 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
3933 /* create a new compare */
3934 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
3936 set_Proj_pred(proj, n);
3937 set_Proj_proj(proj, proj_nr);
3941 } /* transform_node_Proj_Cmp */
3944 * Does all optimizations on nodes that must be done on it's Proj's
3945 * because of creating new nodes.
3947 static ir_node *transform_node_Proj(ir_node *proj) {
3948 ir_node *n = get_Proj_pred(proj);
3950 switch (get_irn_opcode(n)) {
3952 return transform_node_Proj_Div(proj);
3955 return transform_node_Proj_Mod(proj);
3958 return transform_node_Proj_DivMod(proj);
3961 return transform_node_Proj_Cond(proj);
3964 return transform_node_Proj_Cmp(proj);
3967 /* should not happen, but if it does will be optimized away */
3968 return equivalent_node_Proj(proj);
3974 } /* transform_node_Proj */
3977 * Move Confirms down through Phi nodes.
3979 static ir_node *transform_node_Phi(ir_node *phi) {
3981 ir_mode *mode = get_irn_mode(phi);
3983 if (mode_is_reference(mode)) {
3984 n = get_irn_arity(phi);
3986 /* Beware of Phi0 */
3988 ir_node *pred = get_irn_n(phi, 0);
3989 ir_node *bound, *new_Phi, *block, **in;
3992 if (! is_Confirm(pred))
3995 bound = get_Confirm_bound(pred);
3996 pnc = get_Confirm_cmp(pred);
3998 NEW_ARR_A(ir_node *, in, n);
3999 in[0] = get_Confirm_value(pred);
4001 for (i = 1; i < n; ++i) {
4002 pred = get_irn_n(phi, i);
4004 if (! is_Confirm(pred) ||
4005 get_Confirm_bound(pred) != bound ||
4006 get_Confirm_cmp(pred) != pnc)
4008 in[i] = get_Confirm_value(pred);
4010 /* move the Confirm nodes "behind" the Phi */
4011 block = get_irn_n(phi, -1);
4012 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4013 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4017 } /* transform_node_Phi */
4020 * Returns the operands of a commutative bin-op, if one operand is
4021 * a const, it is returned as the second one.
4023 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4024 ir_node *op_a = get_binop_left(binop);
4025 ir_node *op_b = get_binop_right(binop);
4027 assert(is_op_commutative(get_irn_op(binop)));
4029 if (is_Const(op_a)) {
4036 } /* get_comm_Binop_Ops */
4039 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4040 * Such pattern may arise in bitfield stores.
4042 * value c4 value c4 & c2
4043 * AND c3 AND c1 | c3
4050 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4053 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4056 ir_node *and_l, *c3;
4057 ir_node *value, *c4;
4058 ir_node *new_and, *new_const, *block;
4059 ir_mode *mode = get_irn_mode(or);
4061 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4064 get_comm_Binop_Ops(or, &and, &c1);
4065 if (!is_Const(c1) || !is_And(and))
4068 get_comm_Binop_Ops(and, &or_l, &c2);
4072 tv1 = get_Const_tarval(c1);
4073 tv2 = get_Const_tarval(c2);
4075 tv = tarval_or(tv1, tv2);
4076 if (tarval_is_all_one(tv)) {
4077 /* the AND does NOT clear a bit with isn't set by the OR */
4078 set_Or_left(or, or_l);
4079 set_Or_right(or, c1);
4081 /* check for more */
4088 get_comm_Binop_Ops(or_l, &and_l, &c3);
4089 if (!is_Const(c3) || !is_And(and_l))
4092 get_comm_Binop_Ops(and_l, &value, &c4);
4096 /* ok, found the pattern, check for conditions */
4097 assert(mode == get_irn_mode(and));
4098 assert(mode == get_irn_mode(or_l));
4099 assert(mode == get_irn_mode(and_l));
4101 tv3 = get_Const_tarval(c3);
4102 tv4 = get_Const_tarval(c4);
4104 tv = tarval_or(tv4, tv2);
4105 if (!tarval_is_all_one(tv)) {
4106 /* have at least one 0 at the same bit position */
4110 n_tv4 = tarval_not(tv4);
4111 if (tv3 != tarval_and(tv3, n_tv4)) {
4112 /* bit in the or_mask is outside the and_mask */
4116 n_tv2 = tarval_not(tv2);
4117 if (tv1 != tarval_and(tv1, n_tv2)) {
4118 /* bit in the or_mask is outside the and_mask */
4122 /* ok, all conditions met */
4123 block = get_irn_n(or, -1);
4125 new_and = new_r_And(current_ir_graph, block,
4126 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4128 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4130 set_Or_left(or, new_and);
4131 set_Or_right(or, new_const);
4133 /* check for more */
4135 } /* transform_node_Or_bf_store */
4138 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4140 static ir_node *transform_node_Or_Rot(ir_node *or) {
4141 ir_mode *mode = get_irn_mode(or);
4142 ir_node *shl, *shr, *block;
4143 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4146 if (! mode_is_int(mode))
4149 shl = get_binop_left(or);
4150 shr = get_binop_right(or);
4159 } else if (!is_Shl(shl)) {
4161 } else if (!is_Shr(shr)) {
4164 x = get_Shl_left(shl);
4165 if (x != get_Shr_left(shr))
4168 c1 = get_Shl_right(shl);
4169 c2 = get_Shr_right(shr);
4170 if (is_Const(c1) && is_Const(c2)) {
4171 tv1 = get_Const_tarval(c1);
4172 if (! tarval_is_long(tv1))
4175 tv2 = get_Const_tarval(c2);
4176 if (! tarval_is_long(tv2))
4179 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4180 != get_mode_size_bits(mode))
4183 /* yet, condition met */
4184 block = get_irn_n(or, -1);
4186 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4188 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4190 } else if (is_Sub(c1)) {
4194 if (get_Sub_right(sub) != v)
4197 c1 = get_Sub_left(sub);
4201 tv1 = get_Const_tarval(c1);
4202 if (! tarval_is_long(tv1))
4205 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
4208 /* yet, condition met */
4209 block = get_nodes_block(or);
4211 /* a Rot right is not supported, so use a rot left */
4212 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4214 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4216 } else if (is_Sub(c2)) {
4220 c1 = get_Sub_left(sub);
4224 tv1 = get_Const_tarval(c1);
4225 if (! tarval_is_long(tv1))
4228 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
4231 /* yet, condition met */
4232 block = get_irn_n(or, -1);
4235 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4237 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4242 } /* transform_node_Or_Rot */
4247 static ir_node *transform_node_Or(ir_node *n) {
4248 ir_node *c, *oldn = n;
4249 ir_node *a = get_Or_left(n);
4250 ir_node *b = get_Or_right(n);
4252 if (is_Not(a) && is_Not(b)) {
4253 /* ~a | ~b = ~(a&b) */
4254 ir_node *block = get_nodes_block(n);
4255 ir_mode *mode = get_irn_mode(n);
4259 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4260 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4265 /* we can evaluate 2 Projs of the same Cmp */
4266 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4267 ir_node *pred_a = get_Proj_pred(a);
4268 ir_node *pred_b = get_Proj_pred(b);
4269 if (pred_a == pred_b) {
4270 dbg_info *dbgi = get_irn_dbg_info(n);
4271 ir_node *block = get_nodes_block(pred_a);
4272 pn_Cmp pn_a = get_Proj_proj(a);
4273 pn_Cmp pn_b = get_Proj_proj(b);
4274 /* yes, we can simply calculate with pncs */
4275 pn_Cmp new_pnc = pn_a | pn_b;
4277 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4282 HANDLE_BINOP_PHI(tarval_or, a,b,c);
4284 n = transform_node_Or_bf_store(n);
4285 n = transform_node_Or_Rot(n);
4289 n = transform_bitwise_distributive(n, transform_node_Or);
4292 } /* transform_node_Or */
4296 static ir_node *transform_node(ir_node *n);
4299 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
4301 * Should be moved to reassociation?
4303 static ir_node *transform_node_shift(ir_node *n) {
4304 ir_node *left, *right;
4305 tarval *tv1, *tv2, *res;
4307 int modulo_shf, flag;
4309 left = get_binop_left(n);
4311 /* different operations */
4312 if (get_irn_op(left) != get_irn_op(n))
4315 right = get_binop_right(n);
4316 tv1 = value_of(right);
4317 if (tv1 == tarval_bad)
4320 tv2 = value_of(get_binop_right(left));
4321 if (tv2 == tarval_bad)
4324 res = tarval_add(tv1, tv2);
4326 /* beware: a simple replacement works only, if res < modulo shift */
4327 mode = get_irn_mode(n);
4331 modulo_shf = get_mode_modulo_shift(mode);
4332 if (modulo_shf > 0) {
4333 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4335 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4341 /* ok, we can replace it */
4342 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4344 in[0] = get_binop_left(left);
4345 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4347 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4349 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4351 return transform_node(irn);
4354 } /* transform_node_shift */
4359 static ir_node *transform_node_Shr(ir_node *n) {
4360 ir_node *c, *oldn = n;
4361 ir_node *a = get_Shr_left(n);
4362 ir_node *b = get_Shr_right(n);
4364 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
4365 return transform_node_shift(n);
4366 } /* transform_node_Shr */
4371 static ir_node *transform_node_Shrs(ir_node *n) {
4372 ir_node *c, *oldn = n;
4373 ir_node *a = get_Shrs_left(n);
4374 ir_node *b = get_Shrs_right(n);
4376 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
4377 return transform_node_shift(n);
4378 } /* transform_node_Shrs */
4383 static ir_node *transform_node_Shl(ir_node *n) {
4384 ir_node *c, *oldn = n;
4385 ir_node *a = get_Shl_left(n);
4386 ir_node *b = get_Shl_right(n);
4388 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
4389 return transform_node_shift(n);
4390 } /* transform_node_Shl */
4393 * Remove dead blocks and nodes in dead blocks
4394 * in keep alive list. We do not generate a new End node.
4396 static ir_node *transform_node_End(ir_node *n) {
4397 int i, j, n_keepalives = get_End_n_keepalives(n);
4400 NEW_ARR_A(ir_node *, in, n_keepalives);
4402 for (i = j = 0; i < n_keepalives; ++i) {
4403 ir_node *ka = get_End_keepalive(n, i);
4405 if (! is_Block_dead(ka)) {
4409 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4412 /* FIXME: beabi need to keep a Proj(M) */
4413 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4416 if (j != n_keepalives)
4417 set_End_keepalives(n, j, in);
4419 } /* transform_node_End */
4421 /** returns 1 if a == -b */
4422 static int is_negated_value(ir_node *a, ir_node *b) {
4423 if(is_Minus(a) && get_Minus_op(a) == b)
4425 if(is_Minus(b) && get_Minus_op(b) == a)
4427 if(is_Sub(a) && is_Sub(b)) {
4428 ir_node *a_left = get_Sub_left(a);
4429 ir_node *a_right = get_Sub_right(a);
4430 ir_node *b_left = get_Sub_left(b);
4431 ir_node *b_right = get_Sub_right(b);
4433 if(a_left == b_right && a_right == b_left)
4441 * Optimize a Mux into some simpler cases.
4443 static ir_node *transform_node_Mux(ir_node *n) {
4444 ir_node *oldn = n, *sel = get_Mux_sel(n);
4445 ir_mode *mode = get_irn_mode(n);
4447 if (mode == mode_b) {
4448 ir_node *t = get_Mux_true(n);
4449 ir_node *f = get_Mux_false(n);
4450 dbg_info *dbg = get_irn_dbg_info(n);
4451 ir_node *block = get_irn_n(n, -1);
4452 ir_graph *irg = current_ir_graph;
4455 tarval *tv_t = get_Const_tarval(t);
4456 if (tv_t == tarval_b_true) {
4458 assert(get_Const_tarval(f) == tarval_b_false);
4461 return new_rd_Or(dbg, irg, block, sel, f, mode_b);
4464 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4465 assert(tv_t == tarval_b_false);
4467 assert(get_Const_tarval(f) == tarval_b_true);
4470 return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
4473 } else if (is_Const(f)) {
4474 tarval *tv_f = get_Const_tarval(f);
4475 if (tv_f == tarval_b_true) {
4476 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4477 return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4479 assert(tv_f == tarval_b_false);
4480 return new_rd_And(dbg, irg, block, sel, t, mode_b);
4485 if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
4486 ir_node *cmp = get_Proj_pred(sel);
4487 long pn = get_Proj_proj(sel);
4488 ir_node *f = get_Mux_false(n);
4489 ir_node *t = get_Mux_true(n);
4492 * Note: normalization puts the constant on the right side,
4493 * so we check only one case.
4495 * Note further that these optimization work even for floating point
4496 * with NaN's because -NaN == NaN.
4497 * However, if +0 and -0 is handled differently, we cannot use the first
4501 ir_node *cmp_r = get_Cmp_right(cmp);
4502 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4503 ir_node *block = get_irn_n(n, -1);
4505 if(is_negated_value(f, t)) {
4506 ir_node *cmp_left = get_Cmp_left(cmp);
4508 /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
4509 if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4510 || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4512 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4514 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4516 /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
4517 } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4518 || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4520 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4522 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4524 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4532 return arch_transform_node_Mux(n);
4533 } /* transform_node_Mux */
4536 * Optimize a Psi into some simpler cases.
4538 static ir_node *transform_node_Psi(ir_node *n) {
4540 return transform_node_Mux(n);
4543 } /* transform_node_Psi */
4546 * optimize sync nodes that have other syncs as input we simply add the inputs
4547 * of the other sync to our own inputs
4549 static ir_node *transform_node_Sync(ir_node *n) {
4552 arity = get_irn_arity(n);
4553 for(i = 0; i < get_irn_arity(n); /*empty*/) {
4555 ir_node *in = get_irn_n(n, i);
4561 /* set sync input 0 instead of the sync */
4562 set_irn_n(n, i, get_irn_n(in, 0));
4563 /* so we check this input again for syncs */
4565 /* append all other inputs of the sync to our sync */
4566 arity2 = get_irn_arity(in);
4567 for(i2 = 1; i2 < arity2; ++i2) {
4568 ir_node *in_in = get_irn_n(in, i2);
4569 add_irn_n(n, in_in);
4570 /* increase arity so we also check the new inputs for syncs */
4575 /* rehash the sync node */
4576 add_identities(current_ir_graph->value_table, n);
4582 * Tries several [inplace] [optimizing] transformations and returns an
4583 * equivalent node. The difference to equivalent_node() is that these
4584 * transformations _do_ generate new nodes, and thus the old node must
4585 * not be freed even if the equivalent node isn't the old one.
4587 static ir_node *transform_node(ir_node *n) {
4591 * Transform_node is the only "optimizing transformation" that might
4592 * return a node with a different opcode. We iterate HERE until fixpoint
4593 * to get the final result.
4597 if (n->op->ops.transform_node)
4598 n = n->op->ops.transform_node(n);
4599 } while (oldn != n);
4602 } /* transform_node */
4605 * Sets the default transform node operation for an ir_op_ops.
4607 * @param code the opcode for the default operation
4608 * @param ops the operations initialized
4613 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
4617 ops->transform_node = transform_node_##a; \
4652 } /* firm_set_default_transform_node */
4655 /* **************** Common Subexpression Elimination **************** */
4657 /** The size of the hash table used, should estimate the number of nodes
4659 #define N_IR_NODES 512
4661 /** Compares the attributes of two Const nodes. */
4662 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
4663 return (get_Const_tarval(a) != get_Const_tarval(b))
4664 || (get_Const_type(a) != get_Const_type(b));
4665 } /* node_cmp_attr_Const */
4667 /** Compares the attributes of two Proj nodes. */
4668 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
4669 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
4670 } /* node_cmp_attr_Proj */
4672 /** Compares the attributes of two Filter nodes. */
4673 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
4674 return get_Filter_proj(a) != get_Filter_proj(b);
4675 } /* node_cmp_attr_Filter */
4677 /** Compares the attributes of two Alloc nodes. */
4678 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
4679 const alloc_attr *pa = get_irn_alloc_attr(a);
4680 const alloc_attr *pb = get_irn_alloc_attr(b);
4681 return (pa->where != pb->where) || (pa->type != pb->type);
4682 } /* node_cmp_attr_Alloc */
4684 /** Compares the attributes of two Free nodes. */
4685 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
4686 const free_attr *pa = get_irn_free_attr(a);
4687 const free_attr *pb = get_irn_free_attr(b);
4688 return (pa->where != pb->where) || (pa->type != pb->type);
4689 } /* node_cmp_attr_Free */
4691 /** Compares the attributes of two SymConst nodes. */
4692 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
4693 const symconst_attr *pa = get_irn_symconst_attr(a);
4694 const symconst_attr *pb = get_irn_symconst_attr(b);
4695 return (pa->num != pb->num)
4696 || (pa->sym.type_p != pb->sym.type_p)
4697 || (pa->tp != pb->tp);
4698 } /* node_cmp_attr_SymConst */
4700 /** Compares the attributes of two Call nodes. */
4701 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
4702 return (get_irn_call_attr(a) != get_irn_call_attr(b));
4703 } /* node_cmp_attr_Call */
4705 /** Compares the attributes of two Sel nodes. */
4706 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
4707 const ir_entity *a_ent = get_Sel_entity(a);
4708 const ir_entity *b_ent = get_Sel_entity(b);
4710 (a_ent->kind != b_ent->kind) ||
4711 (a_ent->name != b_ent->name) ||
4712 (a_ent->owner != b_ent->owner) ||
4713 (a_ent->ld_name != b_ent->ld_name) ||
4714 (a_ent->type != b_ent->type);
4715 } /* node_cmp_attr_Sel */
4717 /** Compares the attributes of two Phi nodes. */
4718 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
4719 /* we can only enter this function if both nodes have the same number of inputs,
4720 hence it is enough to check if one of them is a Phi0 */
4722 /* check the Phi0 attribute */
4723 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
4726 } /* node_cmp_attr_Phi */
4728 /** Compares the attributes of two Conv nodes. */
4729 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
4730 return get_Conv_strict(a) != get_Conv_strict(b);
4731 } /* node_cmp_attr_Conv */
4733 /** Compares the attributes of two Cast nodes. */
4734 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
4735 return get_Cast_type(a) != get_Cast_type(b);
4736 } /* node_cmp_attr_Cast */
4738 /** Compares the attributes of two Load nodes. */
4739 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
4740 if (get_Load_volatility(a) == volatility_is_volatile ||
4741 get_Load_volatility(b) == volatility_is_volatile)
4742 /* NEVER do CSE on volatile Loads */
4744 /* do not CSE Loads with different alignment. Be conservative. */
4745 if (get_Load_align(a) != get_Load_align(b))
4748 return get_Load_mode(a) != get_Load_mode(b);
4749 } /* node_cmp_attr_Load */
4751 /** Compares the attributes of two Store nodes. */
4752 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
4753 /* do not CSE Stores with different alignment. Be conservative. */
4754 if (get_Store_align(a) != get_Store_align(b))
4757 /* NEVER do CSE on volatile Stores */
4758 return (get_Store_volatility(a) == volatility_is_volatile ||
4759 get_Store_volatility(b) == volatility_is_volatile);
4760 } /* node_cmp_attr_Store */
4762 /** Compares the attributes of two Confirm nodes. */
4763 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
4764 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
4765 } /* node_cmp_attr_Confirm */
4767 /** Compares the attributes of two ASM nodes. */
4768 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
4770 const ir_asm_constraint *ca;
4771 const ir_asm_constraint *cb;
4774 if (get_ASM_text(a) != get_ASM_text(b))
4777 /* Should we really check the constraints here? Should be better, but is strange. */
4778 n = get_ASM_n_input_constraints(a);
4779 if (n != get_ASM_n_input_constraints(b))
4782 ca = get_ASM_input_constraints(a);
4783 cb = get_ASM_input_constraints(b);
4784 for (i = 0; i < n; ++i) {
4785 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
4789 n = get_ASM_n_output_constraints(a);
4790 if (n != get_ASM_n_output_constraints(b))
4793 ca = get_ASM_output_constraints(a);
4794 cb = get_ASM_output_constraints(b);
4795 for (i = 0; i < n; ++i) {
4796 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
4800 n = get_ASM_n_clobbers(a);
4801 if (n != get_ASM_n_clobbers(b))
4804 cla = get_ASM_clobbers(a);
4805 clb = get_ASM_clobbers(b);
4806 for (i = 0; i < n; ++i) {
4807 if (cla[i] != clb[i])
4811 } /* node_cmp_attr_ASM */
4814 * Set the default node attribute compare operation for an ir_op_ops.
4816 * @param code the opcode for the default operation
4817 * @param ops the operations initialized
4822 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
4826 ops->node_cmp_attr = node_cmp_attr_##a; \
4851 } /* firm_set_default_node_cmp_attr */
4854 * Compare function for two nodes in the hash table. Gets two
4855 * nodes as parameters. Returns 0 if the nodes are a cse.
4857 int identities_cmp(const void *elt, const void *key) {
4864 if (a == b) return 0;
4866 if ((get_irn_op(a) != get_irn_op(b)) ||
4867 (get_irn_mode(a) != get_irn_mode(b))) return 1;
4869 /* compare if a's in and b's in are of equal length */
4870 irn_arity_a = get_irn_intra_arity (a);
4871 if (irn_arity_a != get_irn_intra_arity(b))
4874 /* for block-local cse and op_pin_state_pinned nodes: */
4875 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
4876 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
4880 /* compare a->in[0..ins] with b->in[0..ins] */
4881 for (i = 0; i < irn_arity_a; i++)
4882 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
4886 * here, we already now that the nodes are identical except their
4889 if (a->op->ops.node_cmp_attr)
4890 return a->op->ops.node_cmp_attr(a, b);
4893 } /* identities_cmp */
4896 * Calculate a hash value of a node.
4898 unsigned ir_node_hash(ir_node *node) {
4902 if (node->op == op_Const) {
4903 /* special value for const, as they only differ in their tarval. */
4904 h = HASH_PTR(node->attr.con.tv);
4905 h = 9*h + HASH_PTR(get_irn_mode(node));
4906 } else if (node->op == op_SymConst) {
4907 /* special value for const, as they only differ in their symbol. */
4908 h = HASH_PTR(node->attr.symc.sym.type_p);
4909 h = 9*h + HASH_PTR(get_irn_mode(node));
4912 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
4913 h = irn_arity = get_irn_intra_arity(node);
4915 /* consider all in nodes... except the block if not a control flow. */
4916 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
4917 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
4921 h = 9*h + HASH_PTR(get_irn_mode(node));
4923 h = 9*h + HASH_PTR(get_irn_op(node));
4927 } /* ir_node_hash */
4929 pset *new_identities(void) {
4930 return new_pset(identities_cmp, N_IR_NODES);
4931 } /* new_identities */
4933 void del_identities(pset *value_table) {
4934 del_pset(value_table);
4935 } /* del_identities */
4938 * Normalize a node by putting constants (and operands with smaller
4939 * node index) on the right
4941 * @param n The node to normalize
4943 static void normalize_node(ir_node *n) {
4944 if (get_opt_reassociation()) {
4945 if (is_op_commutative(get_irn_op(n))) {
4946 ir_node *l = get_binop_left(n);
4947 ir_node *r = get_binop_right(n);
4948 int l_idx = get_irn_idx(l);
4949 int r_idx = get_irn_idx(r);
4951 /* For commutative operators perform a OP b == b OP a but keep
4952 constants on the RIGHT side. This helps greatly in some optimizations.
4953 Moreover we use the idx number to make the form deterministic. */
4954 if (is_irn_constlike(l))
4956 if (is_irn_constlike(r))
4958 if (l_idx < r_idx) {
4959 set_binop_left(n, r);
4960 set_binop_right(n, l);
4964 } /* normalize_node */
4967 * Return the canonical node computing the same value as n.
4969 * @param value_table The value table
4970 * @param n The node to lookup
4972 * Looks up the node in a hash table.
4974 * For Const nodes this is performed in the constructor, too. Const
4975 * nodes are extremely time critical because of their frequent use in
4976 * constant string arrays.
4978 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
4981 if (!value_table) return n;
4985 o = pset_find(value_table, n, ir_node_hash(n));
4994 * During construction we set the op_pin_state_pinned flag in the graph right when the
4995 * optimization is performed. The flag turning on procedure global cse could
4996 * be changed between two allocations. This way we are safe.
4998 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5001 n = identify(value_table, n);
5002 if (get_irn_n(old, -1) != get_irn_n(n, -1))
5003 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5005 } /* identify_cons */
5008 * Return the canonical node computing the same value as n.
5009 * Looks up the node in a hash table, enters it in the table
5010 * if it isn't there yet.
5012 ir_node *identify_remember(pset *value_table, ir_node *n) {
5015 if (!value_table) return n;
5018 /* lookup or insert in hash table with given hash key. */
5019 o = pset_insert(value_table, n, ir_node_hash(n));
5026 } /* identify_remember */
5028 /* Add a node to the identities value table. */
5029 void add_identities(pset *value_table, ir_node *node) {
5030 if (get_opt_cse() && is_no_Block(node))
5031 identify_remember(value_table, node);
5032 } /* add_identities */
5034 /* Visit each node in the value table of a graph. */
5035 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5037 ir_graph *rem = current_ir_graph;
5039 current_ir_graph = irg;
5040 foreach_pset(irg->value_table, node)
5042 current_ir_graph = rem;
5043 } /* visit_all_identities */
5046 * Garbage in, garbage out. If a node has a dead input, i.e., the
5047 * Bad node is input to the node, return the Bad node.
5049 static ir_node *gigo(ir_node *node) {
5051 ir_op *op = get_irn_op(node);
5053 /* remove garbage blocks by looking at control flow that leaves the block
5054 and replacing the control flow by Bad. */
5055 if (get_irn_mode(node) == mode_X) {
5056 ir_node *block = get_nodes_block(skip_Proj(node));
5058 /* Don't optimize nodes in immature blocks. */
5059 if (!get_Block_matured(block)) return node;
5060 /* Don't optimize End, may have Bads. */
5061 if (op == op_End) return node;
5063 if (is_Block(block)) {
5064 irn_arity = get_irn_arity(block);
5065 for (i = 0; i < irn_arity; i++) {
5066 if (!is_Bad(get_irn_n(block, i)))
5069 if (i == irn_arity) {
5070 ir_graph *irg = get_irn_irg(block);
5071 /* the start block is never dead */
5072 if (block != get_irg_start_block(irg)
5073 && block != get_irg_end_block(irg))
5079 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5080 blocks predecessors is dead. */
5081 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5082 irn_arity = get_irn_arity(node);
5085 * Beware: we can only read the block of a non-floating node.
5087 if (is_irn_pinned_in_irg(node) &&
5088 is_Block_dead(get_nodes_block(node)))
5091 for (i = 0; i < irn_arity; i++) {
5092 ir_node *pred = get_irn_n(node, i);
5097 /* Propagating Unknowns here seems to be a bad idea, because
5098 sometimes we need a node as a input and did not want that
5100 However, it might be useful to move this into a later phase
5101 (if you think that optimizing such code is useful). */
5102 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5103 return new_Unknown(get_irn_mode(node));
5108 /* With this code we violate the agreement that local_optimize
5109 only leaves Bads in Block, Phi and Tuple nodes. */
5110 /* If Block has only Bads as predecessors it's garbage. */
5111 /* If Phi has only Bads as predecessors it's garbage. */
5112 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5113 irn_arity = get_irn_arity(node);
5114 for (i = 0; i < irn_arity; i++) {
5115 if (!is_Bad(get_irn_n(node, i))) break;
5117 if (i == irn_arity) node = new_Bad();
5124 * These optimizations deallocate nodes from the obstack.
5125 * It can only be called if it is guaranteed that no other nodes
5126 * reference this one, i.e., right after construction of a node.
5128 * @param n The node to optimize
5130 * current_ir_graph must be set to the graph of the node!
5132 ir_node *optimize_node(ir_node *n) {
5135 ir_opcode iro = get_irn_opcode(n);
5137 /* Always optimize Phi nodes: part of the construction. */
5138 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5140 /* constant expression evaluation / constant folding */
5141 if (get_opt_constant_folding()) {
5142 /* neither constants nor Tuple values can be evaluated */
5143 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5144 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5145 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5146 /* try to evaluate */
5147 tv = computed_value(n);
5148 if (tv != tarval_bad) {
5150 ir_type *old_tp = get_irn_type(n);
5151 int i, arity = get_irn_arity(n);
5155 * Try to recover the type of the new expression.
5157 for (i = 0; i < arity && !old_tp; ++i)
5158 old_tp = get_irn_type(get_irn_n(n, i));
5161 * we MUST copy the node here temporary, because it's still needed
5162 * for DBG_OPT_CSTEVAL
5164 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5165 oldn = alloca(node_size);
5167 memcpy(oldn, n, node_size);
5168 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5170 /* ARG, copy the in array, we need it for statistics */
5171 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5173 /* note the inplace edges module */
5174 edges_node_deleted(n, current_ir_graph);
5176 /* evaluation was successful -- replace the node. */
5177 irg_kill_node(current_ir_graph, n);
5178 nw = new_Const(get_tarval_mode(tv), tv);
5180 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5181 set_Const_type(nw, old_tp);
5182 DBG_OPT_CSTEVAL(oldn, nw);
5183 tarval_enable_fp_ops(old_fp_mode);
5186 tarval_enable_fp_ops(old_fp_mode);
5190 /* remove unnecessary nodes */
5191 if (get_opt_constant_folding() ||
5192 (iro == iro_Phi) || /* always optimize these nodes. */
5194 (iro == iro_Proj) ||
5195 (iro == iro_Block) ) /* Flags tested local. */
5196 n = equivalent_node(n);
5198 /* Common Subexpression Elimination.
5200 * Checks whether n is already available.
5201 * The block input is used to distinguish different subexpressions. Right
5202 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5203 * subexpressions within a block.
5206 n = identify_cons(current_ir_graph->value_table, n);
5209 edges_node_deleted(oldn, current_ir_graph);
5211 /* We found an existing, better node, so we can deallocate the old node. */
5212 irg_kill_node(current_ir_graph, oldn);
5216 /* Some more constant expression evaluation that does not allow to
5218 iro = get_irn_opcode(n);
5219 if (get_opt_constant_folding() ||
5220 (iro == iro_Cond) ||
5221 (iro == iro_Proj)) /* Flags tested local. */
5222 n = transform_node(n);
5224 /* Remove nodes with dead (Bad) input.
5225 Run always for transformation induced Bads. */
5228 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5229 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5230 n = identify_remember(current_ir_graph->value_table, n);
5234 } /* optimize_node */
5238 * These optimizations never deallocate nodes (in place). This can cause dead
5239 * nodes lying on the obstack. Remove these by a dead node elimination,
5240 * i.e., a copying garbage collection.
5242 ir_node *optimize_in_place_2(ir_node *n) {
5245 ir_opcode iro = get_irn_opcode(n);
5247 if (!get_opt_optimize() && !is_Phi(n)) return n;
5249 /* constant expression evaluation / constant folding */
5250 if (get_opt_constant_folding()) {
5251 /* neither constants nor Tuple values can be evaluated */
5252 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5253 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5254 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5255 /* try to evaluate */
5256 tv = computed_value(n);
5257 if (tv != tarval_bad) {
5258 /* evaluation was successful -- replace the node. */
5259 ir_type *old_tp = get_irn_type(n);
5260 int i, arity = get_irn_arity(n);
5263 * Try to recover the type of the new expression.
5265 for (i = 0; i < arity && !old_tp; ++i)
5266 old_tp = get_irn_type(get_irn_n(n, i));
5268 n = new_Const(get_tarval_mode(tv), tv);
5270 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5271 set_Const_type(n, old_tp);
5273 DBG_OPT_CSTEVAL(oldn, n);
5274 tarval_enable_fp_ops(old_fp_mode);
5277 tarval_enable_fp_ops(old_fp_mode);
5281 /* remove unnecessary nodes */
5282 if (get_opt_constant_folding() ||
5283 (iro == iro_Phi) || /* always optimize these nodes. */
5284 (iro == iro_Id) || /* ... */
5285 (iro == iro_Proj) || /* ... */
5286 (iro == iro_Block) ) /* Flags tested local. */
5287 n = equivalent_node(n);
5289 /** common subexpression elimination **/
5290 /* Checks whether n is already available. */
5291 /* The block input is used to distinguish different subexpressions. Right
5292 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5293 subexpressions within a block. */
5294 if (get_opt_cse()) {
5295 n = identify(current_ir_graph->value_table, n);
5298 /* Some more constant expression evaluation. */
5299 iro = get_irn_opcode(n);
5300 if (get_opt_constant_folding() ||
5301 (iro == iro_Cond) ||
5302 (iro == iro_Proj)) /* Flags tested local. */
5303 n = transform_node(n);
5305 /* Remove nodes with dead (Bad) input.
5306 Run always for transformation induced Bads. */
5309 /* Now we can verify the node, as it has no dead inputs any more. */
5312 /* Now we have a legal, useful node. Enter it in hash table for cse.
5313 Blocks should be unique anyways. (Except the successor of start:
5314 is cse with the start block!) */
5315 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5316 n = identify_remember(current_ir_graph->value_table, n);
5319 } /* optimize_in_place_2 */
5322 * Wrapper for external use, set proper status bits after optimization.
5324 ir_node *optimize_in_place(ir_node *n) {
5325 /* Handle graph state */
5326 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5328 if (get_opt_global_cse())
5329 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5330 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5331 set_irg_outs_inconsistent(current_ir_graph);
5333 /* FIXME: Maybe we could also test whether optimizing the node can
5334 change the control graph. */
5335 set_irg_doms_inconsistent(current_ir_graph);
5336 return optimize_in_place_2(n);
5337 } /* optimize_in_place */
5340 * Sets the default operation for an ir_ops.
5342 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5343 ops = firm_set_default_computed_value(code, ops);
5344 ops = firm_set_default_equivalent_node(code, ops);
5345 ops = firm_set_default_transform_node(code, ops);
5346 ops = firm_set_default_node_cmp_attr(code, ops);
5347 ops = firm_set_default_get_type(code, ops);
5348 ops = firm_set_default_get_type_attr(code, ops);
5349 ops = firm_set_default_get_entity_attr(code, ops);
5352 } /* firm_set_default_operations */