2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if (tarval_is_null(ta) || tarval_is_null(tb))
149 return get_mode_null(m);
152 } /* computed_value_Carry */
155 * Return the value of a Borrow.
158 static tarval *computed_value_Borrow(ir_node *n) {
159 ir_node *a = get_binop_left(n);
160 ir_node *b = get_binop_right(n);
161 ir_mode *m = get_irn_mode(n);
163 tarval *ta = value_of(a);
164 tarval *tb = value_of(b);
166 if ((ta != tarval_bad) && (tb != tarval_bad)) {
167 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
168 } else if (tarval_is_null(ta)) {
169 return get_mode_null(m);
172 } /* computed_value_Borrow */
175 * Return the value of an unary Minus.
177 static tarval *computed_value_Minus(ir_node *n) {
178 ir_node *a = get_Minus_op(n);
179 tarval *ta = value_of(a);
181 if (ta != tarval_bad)
182 return tarval_neg(ta);
185 } /* computed_value_Minus */
188 * Return the value of a Mul.
190 static tarval *computed_value_Mul(ir_node *n) {
191 ir_node *a = get_Mul_left(n);
192 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 mode = get_irn_mode(n);
199 if (mode != get_irn_mode(a)) {
200 /* n * n = 2n bit multiplication */
201 ta = tarval_convert_to(ta, mode);
202 tb = tarval_convert_to(tb, mode);
205 if (ta != tarval_bad && tb != tarval_bad) {
206 return tarval_mul(ta, tb);
208 /* a*0 = 0 or 0*b = 0 */
209 if (ta == get_mode_null(mode))
211 if (tb == get_mode_null(mode))
215 } /* computed_value_Mul */
218 * Return the value of a floating point Quot.
220 static tarval *computed_value_Quot(ir_node *n) {
221 ir_node *a = get_Quot_left(n);
222 ir_node *b = get_Quot_right(n);
224 tarval *ta = value_of(a);
225 tarval *tb = value_of(b);
227 /* This was missing in original implementation. Why? */
228 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
229 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
230 return tarval_quo(ta, tb);
233 } /* computed_value_Quot */
236 * Calculate the value of an integer Div of two nodes.
237 * Special case: 0 / b
239 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
240 tarval *ta = value_of(a);
241 tarval *tb = value_of(b);
243 /* Compute c1 / c2 or 0 / a, a != 0 */
244 if (ta != tarval_bad) {
245 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
246 return tarval_div(ta, tb);
247 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
251 } /* do_computed_value_Div */
254 * Return the value of an integer Div.
256 static tarval *computed_value_Div(ir_node *n) {
257 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
258 } /* computed_value_Div */
261 * Calculate the value of an integer Mod of two nodes.
262 * Special case: a % 1
264 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 /* Compute c1 % c2 or a % 1 */
269 if (tb != tarval_bad) {
270 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
271 return tarval_mod(ta, tb);
272 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
273 return get_mode_null(get_irn_mode(a));
276 } /* do_computed_value_Mod */
279 * Return the value of an integer Mod.
281 static tarval *computed_value_Mod(ir_node *n) {
282 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
283 } /* computed_value_Mod */
286 * Return the value of an Abs.
288 static tarval *computed_value_Abs(ir_node *n) {
289 ir_node *a = get_Abs_op(n);
290 tarval *ta = value_of(a);
292 if (ta != tarval_bad)
293 return tarval_abs(ta);
296 } /* computed_value_Abs */
299 * Return the value of an And.
300 * Special case: a & 0, 0 & b
302 static tarval *computed_value_And(ir_node *n) {
303 ir_node *a = get_And_left(n);
304 ir_node *b = get_And_right(n);
306 tarval *ta = value_of(a);
307 tarval *tb = value_of(b);
309 if ((ta != tarval_bad) && (tb != tarval_bad)) {
310 return tarval_and (ta, tb);
312 if (tarval_is_null(ta)) return ta;
313 if (tarval_is_null(tb)) return tb;
316 } /* computed_value_And */
319 * Return the value of an Or.
320 * Special case: a | 1...1, 1...1 | b
322 static tarval *computed_value_Or(ir_node *n) {
323 ir_node *a = get_Or_left(n);
324 ir_node *b = get_Or_right(n);
326 tarval *ta = value_of(a);
327 tarval *tb = value_of(b);
329 if ((ta != tarval_bad) && (tb != tarval_bad)) {
330 return tarval_or (ta, tb);
332 if (tarval_is_all_one(ta)) return ta;
333 if (tarval_is_all_one(tb)) return tb;
336 } /* computed_value_Or */
339 * Return the value of an Eor.
341 static tarval *computed_value_Eor(ir_node *n) {
342 ir_node *a = get_Eor_left(n);
343 ir_node *b = get_Eor_right(n);
348 return get_mode_null(get_irn_mode(n));
353 if ((ta != tarval_bad) && (tb != tarval_bad)) {
354 return tarval_eor (ta, tb);
357 } /* computed_value_Eor */
360 * Return the value of a Not.
362 static tarval *computed_value_Not(ir_node *n) {
363 ir_node *a = get_Not_op(n);
364 tarval *ta = value_of(a);
366 if (ta != tarval_bad)
367 return tarval_not(ta);
370 } /* computed_value_Not */
373 * Return the value of a Shl.
375 static tarval *computed_value_Shl(ir_node *n) {
376 ir_node *a = get_Shl_left(n);
377 ir_node *b = get_Shl_right(n);
379 tarval *ta = value_of(a);
380 tarval *tb = value_of(b);
382 if ((ta != tarval_bad) && (tb != tarval_bad)) {
383 return tarval_shl (ta, tb);
386 } /* computed_value_Shl */
389 * Return the value of a Shr.
391 static tarval *computed_value_Shr(ir_node *n) {
392 ir_node *a = get_Shr_left(n);
393 ir_node *b = get_Shr_right(n);
395 tarval *ta = value_of(a);
396 tarval *tb = value_of(b);
398 if ((ta != tarval_bad) && (tb != tarval_bad)) {
399 return tarval_shr (ta, tb);
402 } /* computed_value_Shr */
405 * Return the value of a Shrs.
407 static tarval *computed_value_Shrs(ir_node *n) {
408 ir_node *a = get_Shrs_left(n);
409 ir_node *b = get_Shrs_right(n);
411 tarval *ta = value_of(a);
412 tarval *tb = value_of(b);
414 if ((ta != tarval_bad) && (tb != tarval_bad)) {
415 return tarval_shrs (ta, tb);
418 } /* computed_value_Shrs */
421 * Return the value of a Rot.
423 static tarval *computed_value_Rot(ir_node *n) {
424 ir_node *a = get_Rot_left(n);
425 ir_node *b = get_Rot_right(n);
427 tarval *ta = value_of(a);
428 tarval *tb = value_of(b);
430 if ((ta != tarval_bad) && (tb != tarval_bad)) {
431 return tarval_rot (ta, tb);
434 } /* computed_value_Rot */
437 * Return the value of a Conv.
439 static tarval *computed_value_Conv(ir_node *n) {
440 ir_node *a = get_Conv_op(n);
441 tarval *ta = value_of(a);
443 if (ta != tarval_bad)
444 return tarval_convert_to(ta, get_irn_mode(n));
447 } /* computed_value_Conv */
450 * Return the value of a Proj(Cmp).
452 * This performs a first step of unreachable code elimination.
453 * Proj can not be computed, but folding a Cmp above the Proj here is
454 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
456 * There are several case where we can evaluate a Cmp node, see later.
458 static tarval *computed_value_Proj_Cmp(ir_node *n) {
459 ir_node *a = get_Proj_pred(n);
460 ir_node *aa = get_Cmp_left(a);
461 ir_node *ab = get_Cmp_right(a);
462 long proj_nr = get_Proj_proj(n);
465 * BEWARE: a == a is NOT always True for floating Point values, as
466 * NaN != NaN is defined, so we must check this here.
469 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
472 /* This is a trick with the bits used for encoding the Cmp
473 Proj numbers, the following statement is not the same:
474 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
475 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
478 tarval *taa = value_of(aa);
479 tarval *tab = value_of(ab);
480 ir_mode *mode = get_irn_mode(aa);
483 * The predecessors of Cmp are target values. We can evaluate
486 if ((taa != tarval_bad) && (tab != tarval_bad)) {
487 /* strange checks... */
488 pn_Cmp flags = tarval_cmp(taa, tab);
489 if (flags != pn_Cmp_False) {
490 return new_tarval_from_long (proj_nr & flags, mode_b);
493 /* for integer values, we can check against MIN/MAX */
494 else if (mode_is_int(mode)) {
495 /* MIN <=/> x. This results in true/false. */
496 if (taa == get_mode_min(mode)) {
497 /* a compare with the MIN value */
498 if (proj_nr == pn_Cmp_Le)
499 return get_tarval_b_true();
500 else if (proj_nr == pn_Cmp_Gt)
501 return get_tarval_b_false();
503 /* x >=/< MIN. This results in true/false. */
505 if (tab == get_mode_min(mode)) {
506 /* a compare with the MIN value */
507 if (proj_nr == pn_Cmp_Ge)
508 return get_tarval_b_true();
509 else if (proj_nr == pn_Cmp_Lt)
510 return get_tarval_b_false();
512 /* MAX >=/< x. This results in true/false. */
513 else if (taa == get_mode_max(mode)) {
514 if (proj_nr == pn_Cmp_Ge)
515 return get_tarval_b_true();
516 else if (proj_nr == pn_Cmp_Lt)
517 return get_tarval_b_false();
519 /* x <=/> MAX. This results in true/false. */
520 else if (tab == get_mode_max(mode)) {
521 if (proj_nr == pn_Cmp_Le)
522 return get_tarval_b_true();
523 else if (proj_nr == pn_Cmp_Gt)
524 return get_tarval_b_false();
528 * The predecessors are Allocs or (void*)(0) constants. Allocs never
529 * return NULL, they raise an exception. Therefore we can predict
533 ir_node *aaa = skip_Id(skip_Proj(aa));
534 ir_node *aba = skip_Id(skip_Proj(ab));
536 if ( ( (/* aa is ProjP and aaa is Alloc */
538 && mode_is_reference(get_irn_mode(aa))
540 && ( (/* ab is NULL */
542 && mode_is_reference(get_irn_mode(ab))
543 && is_Const_null(ab))
544 || (/* ab is other Alloc */
546 && mode_is_reference(get_irn_mode(ab))
549 || (/* aa is NULL and aba is Alloc */
551 && mode_is_reference(get_irn_mode(aa))
554 && mode_is_reference(get_irn_mode(ab))
557 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
560 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
561 } /* computed_value_Proj_Cmp */
564 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
565 * Proj(DivMod) and Proj(Quot).
567 static tarval *computed_value_Proj(ir_node *n) {
568 ir_node *a = get_Proj_pred(n);
571 switch (get_irn_opcode(a)) {
573 return computed_value_Proj_Cmp(n);
576 /* compute either the Div or the Mod part */
577 proj_nr = get_Proj_proj(n);
578 if (proj_nr == pn_DivMod_res_div)
579 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
580 else if (proj_nr == pn_DivMod_res_mod)
581 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
585 if (get_Proj_proj(n) == pn_Div_res)
586 return computed_value(a);
590 if (get_Proj_proj(n) == pn_Mod_res)
591 return computed_value(a);
595 if (get_Proj_proj(n) == pn_Quot_res)
596 return computed_value(a);
603 } /* computed_value_Proj */
606 * Calculate the value of a Mux: can be evaluated, if the
607 * sel and the right input are known.
609 static tarval *computed_value_Mux(ir_node *n) {
610 ir_node *sel = get_Mux_sel(n);
611 tarval *ts = value_of(sel);
613 if (ts == get_tarval_b_true()) {
614 ir_node *v = get_Mux_true(n);
617 else if (ts == get_tarval_b_false()) {
618 ir_node *v = get_Mux_false(n);
622 } /* computed_value_Mux */
625 * Calculate the value of a Psi: can be evaluated, if a condition is true
626 * and all previous conditions are false. If all conditions are false
627 * we evaluate to the default one.
629 static tarval *computed_value_Psi(ir_node *n) {
631 return computed_value_Mux(n);
633 } /* computed_value_Psi */
636 * Calculate the value of a Confirm: can be evaluated,
637 * if it has the form Confirm(x, '=', Const).
639 static tarval *computed_value_Confirm(ir_node *n) {
640 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
641 value_of(get_Confirm_bound(n)) : tarval_bad;
642 } /* computed_value_Confirm */
645 * If the parameter n can be computed, return its value, else tarval_bad.
646 * Performs constant folding.
648 * @param n The node this should be evaluated
650 tarval *computed_value(ir_node *n) {
651 if (n->op->ops.computed_value)
652 return n->op->ops.computed_value(n);
654 } /* computed_value */
657 * Set the default computed_value evaluator in an ir_op_ops.
659 * @param code the opcode for the default operation
660 * @param ops the operations initialized
665 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
669 ops->computed_value = computed_value_##a; \
704 } /* firm_set_default_computed_value */
707 * Returns a equivalent block for another block.
708 * If the block has only one predecessor, this is
709 * the equivalent one. If the only predecessor of a block is
710 * the block itself, this is a dead block.
712 * If both predecessors of a block are the branches of a binary
713 * Cond, the equivalent block is Cond's block.
715 * If all predecessors of a block are bad or lies in a dead
716 * block, the current block is dead as well.
718 * Note, that blocks are NEVER turned into Bad's, instead
719 * the dead_block flag is set. So, never test for is_Bad(block),
720 * always use is_dead_Block(block).
722 static ir_node *equivalent_node_Block(ir_node *n)
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 "self-inverse 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);
1163 ir_mode *n_mode = get_irn_mode(n);
1164 ir_mode *a_mode = get_irn_mode(a);
1166 if (n_mode == a_mode) { /* No Conv necessary */
1167 if (get_Conv_strict(n)) {
1168 /* special case: the predecessor might be a also a Conv */
1170 if (! get_Conv_strict(a)) {
1171 /* first one is not strict, kick it */
1172 set_Conv_op(n, get_Conv_op(a));
1175 /* else both are strict conv, second is superflous */
1176 } else if(is_Proj(a)) {
1177 ir_node *pred = get_Proj_pred(a);
1179 /* loads always return with the exact precision of n_mode */
1180 assert(get_Load_mode(pred) == n_mode);
1185 /* leave strict floating point Conv's */
1189 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1190 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1191 ir_node *b = get_Conv_op(a);
1192 ir_mode *b_mode = get_irn_mode(b);
1194 if (n_mode == b_mode) {
1195 if (n_mode == mode_b) {
1196 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1197 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1198 } else if (mode_is_int(n_mode)) {
1199 if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) {
1200 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1201 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1207 } /* equivalent_node_Conv */
1210 * A Cast may be removed if the type of the previous node
1211 * is already the type of the Cast.
1213 static ir_node *equivalent_node_Cast(ir_node *n) {
1215 ir_node *pred = get_Cast_op(n);
1217 if (get_irn_type(pred) == get_Cast_type(n)) {
1219 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1222 } /* equivalent_node_Cast */
1225 * Several optimizations:
1226 * - no Phi in start block.
1227 * - remove Id operators that are inputs to Phi
1228 * - fold Phi-nodes, iff they have only one predecessor except
1231 static ir_node *equivalent_node_Phi(ir_node *n) {
1236 ir_node *first_val = NULL; /* to shutup gcc */
1238 if (!get_opt_normalize()) return n;
1240 n_preds = get_Phi_n_preds(n);
1242 block = get_nodes_block(n);
1243 if ((is_Block_dead(block)) || /* Control dead */
1244 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1245 return new_Bad(); /* in the Start Block. */
1247 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1249 /* If the Block has a Bad pred, we also have one. */
1250 for (i = 0; i < n_preds; ++i)
1251 if (is_Bad(get_Block_cfgpred(block, i)))
1252 set_Phi_pred(n, i, new_Bad());
1254 /* Find first non-self-referencing input */
1255 for (i = 0; i < n_preds; ++i) {
1256 first_val = get_Phi_pred(n, i);
1257 if ( (first_val != n) /* not self pointer */
1259 && (! is_Bad(first_val))
1261 ) { /* value not dead */
1262 break; /* then found first value. */
1267 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1271 /* search for rest of inputs, determine if any of these
1272 are non-self-referencing */
1273 while (++i < n_preds) {
1274 ir_node *scnd_val = get_Phi_pred(n, i);
1275 if ( (scnd_val != n)
1276 && (scnd_val != first_val)
1278 && (! is_Bad(scnd_val))
1286 /* Fold, if no multiple distinct non-self-referencing inputs */
1288 DBG_OPT_PHI(oldn, n);
1291 } /* equivalent_node_Phi */
1294 * Several optimizations:
1295 * - no Sync in start block.
1296 * - fold Sync-nodes, iff they have only one predecessor except
1299 static ir_node *equivalent_node_Sync(ir_node *n) {
1300 int arity = get_Sync_n_preds(n);
1303 for (i = 0; i < arity;) {
1304 ir_node *pred = get_Sync_pred(n, i);
1307 /* Remove Bad predecessors */
1314 /* Remove duplicate predecessors */
1320 if (get_Sync_pred(n, j) == pred) {
1328 if (arity == 0) return new_Bad();
1329 if (arity == 1) return get_Sync_pred(n, 0);
1331 } /* equivalent_node_Sync */
1334 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1335 * ProjX(Load) and ProjX(Store).
1337 static ir_node *equivalent_node_Proj(ir_node *proj) {
1338 ir_node *oldn = proj;
1339 ir_node *a = get_Proj_pred(proj);
1342 /* Remove the Tuple/Proj combination. */
1343 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1344 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1345 DBG_OPT_TUPLE(oldn, a, proj);
1347 /* This should not happen! */
1348 assert(! "found a Proj with higher number than Tuple predecessors");
1351 } else if (get_irn_mode(proj) == mode_X) {
1352 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1353 /* Remove dead control flow -- early gigo(). */
1355 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1357 /* get the Load address */
1358 ir_node *addr = get_Load_ptr(a);
1359 ir_node *blk = get_irn_n(a, -1);
1362 if (value_not_null(addr, &confirm)) {
1363 if (confirm == NULL) {
1364 /* this node may float if it did not depend on a Confirm */
1365 set_irn_pinned(a, op_pin_state_floats);
1367 if (get_Proj_proj(proj) == pn_Load_X_except) {
1368 DBG_OPT_EXC_REM(proj);
1371 return new_r_Jmp(current_ir_graph, blk);
1373 } else if (is_Store(a)) {
1374 /* get the load/store address */
1375 ir_node *addr = get_Store_ptr(a);
1376 ir_node *blk = get_irn_n(a, -1);
1379 if (value_not_null(addr, &confirm)) {
1380 if (confirm == NULL) {
1381 /* this node may float if it did not depend on a Confirm */
1382 set_irn_pinned(a, op_pin_state_floats);
1384 if (get_Proj_proj(proj) == pn_Store_X_except) {
1385 DBG_OPT_EXC_REM(proj);
1388 return new_r_Jmp(current_ir_graph, blk);
1395 } /* equivalent_node_Proj */
1400 static ir_node *equivalent_node_Id(ir_node *n) {
1405 } while (get_irn_op(n) == op_Id);
1407 DBG_OPT_ID(oldn, n);
1409 } /* equivalent_node_Id */
1414 static ir_node *equivalent_node_Mux(ir_node *n)
1416 ir_node *oldn = n, *sel = get_Mux_sel(n);
1417 tarval *ts = value_of(sel);
1419 /* Mux(true, f, t) == t */
1420 if (ts == tarval_b_true) {
1421 n = get_Mux_true(n);
1422 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1424 /* Mux(false, f, t) == f */
1425 else if (ts == tarval_b_false) {
1426 n = get_Mux_false(n);
1427 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1429 /* Mux(v, x, x) == x */
1430 else if (get_Mux_false(n) == get_Mux_true(n)) {
1431 n = get_Mux_true(n);
1432 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1434 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1435 ir_node *cmp = get_Proj_pred(sel);
1436 long proj_nr = get_Proj_proj(sel);
1437 ir_node *f = get_Mux_false(n);
1438 ir_node *t = get_Mux_true(n);
1441 * Note further that these optimization work even for floating point
1442 * with NaN's because -NaN == NaN.
1443 * However, if +0 and -0 is handled differently, we cannot use the first one.
1446 ir_node *const cmp_l = get_Cmp_left(cmp);
1447 ir_node *const cmp_r = get_Cmp_right(cmp);
1451 if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
1452 (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
1454 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1461 if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
1462 (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
1464 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1471 * Note: normalization puts the constant on the right side,
1472 * so we check only one case.
1474 if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
1475 /* Mux(t CMP 0, X, t) */
1476 if (is_Minus(f) && get_Minus_op(f) == t) {
1477 /* Mux(t CMP 0, -t, t) */
1478 if (proj_nr == pn_Cmp_Eq) {
1479 /* Mux(t == 0, -t, t) ==> -t */
1481 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1482 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1483 /* Mux(t != 0, -t, t) ==> t */
1485 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1492 } /* equivalent_node_Mux */
1495 * Returns a equivalent node of a Psi: if a condition is true
1496 * and all previous conditions are false we know its value.
1497 * If all conditions are false its value is the default one.
1499 static ir_node *equivalent_node_Psi(ir_node *n) {
1501 return equivalent_node_Mux(n);
1503 } /* equivalent_node_Psi */
1506 * Optimize -a CMP -b into b CMP a.
1507 * This works only for for modes where unary Minus
1509 * Note that two-complement integers can Overflow
1510 * so it will NOT work.
1512 * For == and != can be handled in Proj(Cmp)
1514 static ir_node *equivalent_node_Cmp(ir_node *n) {
1515 ir_node *left = get_Cmp_left(n);
1516 ir_node *right = get_Cmp_right(n);
1518 if (is_Minus(left) && is_Minus(right) &&
1519 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1520 left = get_Minus_op(left);
1521 right = get_Minus_op(right);
1522 set_Cmp_left(n, right);
1523 set_Cmp_right(n, left);
1526 } /* equivalent_node_Cmp */
1529 * Remove Confirm nodes if setting is on.
1530 * Replace Confirms(x, '=', Constlike) by Constlike.
1532 static ir_node *equivalent_node_Confirm(ir_node *n) {
1533 ir_node *pred = get_Confirm_value(n);
1534 pn_Cmp pnc = get_Confirm_cmp(n);
1536 if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1538 * rare case: two identical Confirms one after another,
1539 * replace the second one with the first.
1543 if (pnc == pn_Cmp_Eq) {
1544 ir_node *bound = get_Confirm_bound(n);
1547 * Optimize a rare case:
1548 * Confirm(x, '=', Constlike) ==> Constlike
1550 if (is_irn_constlike(bound)) {
1551 DBG_OPT_CONFIRM(n, bound);
1555 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1559 * Optimize CopyB(mem, x, x) into a Nop.
1561 static ir_node *equivalent_node_CopyB(ir_node *n) {
1562 ir_node *a = get_CopyB_dst(n);
1563 ir_node *b = get_CopyB_src(n);
1566 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1567 ir_node *mem = get_CopyB_mem(n);
1568 ir_node *blk = get_nodes_block(n);
1569 turn_into_tuple(n, pn_CopyB_max);
1570 set_Tuple_pred(n, pn_CopyB_M, mem);
1571 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1572 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1573 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1576 } /* equivalent_node_CopyB */
1579 * Optimize Bounds(idx, idx, upper) into idx.
1581 static ir_node *equivalent_node_Bound(ir_node *n) {
1582 ir_node *idx = get_Bound_index(n);
1583 ir_node *pred = skip_Proj(idx);
1586 if (is_Bound(pred)) {
1588 * idx was Bounds checked in the same MacroBlock previously,
1589 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1591 ir_node *lower = get_Bound_lower(n);
1592 ir_node *upper = get_Bound_upper(n);
1593 if (get_Bound_lower(pred) == lower &&
1594 get_Bound_upper(pred) == upper &&
1595 get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) {
1597 * One could expect that we simply return the previous
1598 * Bound here. However, this would be wrong, as we could
1599 * add an exception Proj to a new location then.
1600 * So, we must turn in into a tuple.
1606 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1607 ir_node *mem = get_Bound_mem(n);
1608 ir_node *blk = get_nodes_block(n);
1609 turn_into_tuple(n, pn_Bound_max);
1610 set_Tuple_pred(n, pn_Bound_M, mem);
1611 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1612 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1613 set_Tuple_pred(n, pn_Bound_res, idx);
1616 } /* equivalent_node_Bound */
1619 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1620 * perform no actual computation, as, e.g., the Id nodes. It does not create
1621 * new nodes. It is therefore safe to free n if the node returned is not n.
1622 * If a node returns a Tuple we can not just skip it. If the size of the
1623 * in array fits, we transform n into a tuple (e.g., Div).
1625 ir_node *equivalent_node(ir_node *n) {
1626 if (n->op->ops.equivalent_node)
1627 return n->op->ops.equivalent_node(n);
1629 } /* equivalent_node */
1632 * Sets the default equivalent node operation for an ir_op_ops.
1634 * @param code the opcode for the default operation
1635 * @param ops the operations initialized
1640 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1644 ops->equivalent_node = equivalent_node_##a; \
1684 } /* firm_set_default_equivalent_node */
1687 * Returns non-zero if a node is a Phi node
1688 * with all predecessors constant.
1690 static int is_const_Phi(ir_node *n) {
1693 if (! is_Phi(n) || get_irn_arity(n) == 0)
1695 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1696 if (! is_Const(get_irn_n(n, i)))
1699 } /* is_const_Phi */
1702 * Apply an evaluator on a binop with a constant operators (and one Phi).
1704 * @param phi the Phi node
1705 * @param other the other operand
1706 * @param eval an evaluator function
1707 * @param mode the mode of the result, may be different from the mode of the Phi!
1708 * @param left if non-zero, other is the left operand, else the right
1710 * @return a new Phi node if the conversion was successful, NULL else
1712 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) {
1717 int i, n = get_irn_arity(phi);
1719 NEW_ARR_A(void *, res, n);
1721 for (i = 0; i < n; ++i) {
1722 pred = get_irn_n(phi, i);
1723 tv = get_Const_tarval(pred);
1724 tv = eval(other, tv);
1726 if (tv == tarval_bad) {
1727 /* folding failed, bad */
1733 for (i = 0; i < n; ++i) {
1734 pred = get_irn_n(phi, i);
1735 tv = get_Const_tarval(pred);
1736 tv = eval(tv, other);
1738 if (tv == tarval_bad) {
1739 /* folding failed, bad */
1745 irg = current_ir_graph;
1746 for (i = 0; i < n; ++i) {
1747 pred = get_irn_n(phi, i);
1748 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1749 mode, res[i], get_Const_type(pred));
1751 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1752 } /* apply_binop_on_phi */
1755 * Apply an evaluator on a binop with two constant Phi.
1757 * @param a the left Phi node
1758 * @param b the right Phi node
1759 * @param eval an evaluator function
1760 * @param mode the mode of the result, may be different from the mode of the Phi!
1762 * @return a new Phi node if the conversion was successful, NULL else
1764 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) {
1765 tarval *tv_l, *tv_r, *tv;
1771 if (get_nodes_block(a) != get_nodes_block(b))
1774 n = get_irn_arity(a);
1775 NEW_ARR_A(void *, res, n);
1777 for (i = 0; i < n; ++i) {
1778 pred = get_irn_n(a, i);
1779 tv_l = get_Const_tarval(pred);
1780 pred = get_irn_n(b, i);
1781 tv_r = get_Const_tarval(pred);
1782 tv = eval(tv_l, tv_r);
1784 if (tv == tarval_bad) {
1785 /* folding failed, bad */
1790 irg = current_ir_graph;
1791 for (i = 0; i < n; ++i) {
1792 pred = get_irn_n(a, i);
1793 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
1795 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
1796 } /* apply_binop_on_2_phis */
1799 * Apply an evaluator on a unop with a constant operator (a Phi).
1801 * @param phi the Phi node
1802 * @param eval an evaluator function
1804 * @return a new Phi node if the conversion was successful, NULL else
1806 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1812 int i, n = get_irn_arity(phi);
1814 NEW_ARR_A(void *, res, n);
1815 for (i = 0; i < n; ++i) {
1816 pred = get_irn_n(phi, i);
1817 tv = get_Const_tarval(pred);
1820 if (tv == tarval_bad) {
1821 /* folding failed, bad */
1826 mode = get_irn_mode(phi);
1827 irg = current_ir_graph;
1828 for (i = 0; i < n; ++i) {
1829 pred = get_irn_n(phi, i);
1830 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1831 mode, res[i], get_Const_type(pred));
1833 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1834 } /* apply_unop_on_phi */
1837 * Apply a conversion on a constant operator (a Phi).
1839 * @param phi the Phi node
1841 * @return a new Phi node if the conversion was successful, NULL else
1843 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
1848 int i, n = get_irn_arity(phi);
1850 NEW_ARR_A(void *, res, n);
1851 for (i = 0; i < n; ++i) {
1852 pred = get_irn_n(phi, i);
1853 tv = get_Const_tarval(pred);
1854 tv = tarval_convert_to(tv, mode);
1856 if (tv == tarval_bad) {
1857 /* folding failed, bad */
1862 irg = current_ir_graph;
1863 for (i = 0; i < n; ++i) {
1864 pred = get_irn_n(phi, i);
1865 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1866 mode, res[i], get_Const_type(pred));
1868 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1869 } /* apply_conv_on_phi */
1872 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1873 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1874 * If possible, remove the Conv's.
1876 static ir_node *transform_node_AddSub(ir_node *n) {
1877 ir_mode *mode = get_irn_mode(n);
1879 if (mode_is_reference(mode)) {
1880 ir_node *left = get_binop_left(n);
1881 ir_node *right = get_binop_right(n);
1882 unsigned ref_bits = get_mode_size_bits(mode);
1884 if (is_Conv(left)) {
1885 ir_mode *lmode = get_irn_mode(left);
1886 unsigned bits = get_mode_size_bits(lmode);
1888 if (ref_bits == bits &&
1889 mode_is_int(lmode) &&
1890 get_mode_arithmetic(lmode) == irma_twos_complement) {
1891 ir_node *pre = get_Conv_op(left);
1892 ir_mode *pre_mode = get_irn_mode(pre);
1894 if (mode_is_int(pre_mode) &&
1895 get_mode_size_bits(pre_mode) == bits &&
1896 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1897 /* ok, this conv just changes to sign, moreover the calculation
1898 * is done with same number of bits as our address mode, so
1899 * we can ignore the conv as address calculation can be viewed
1900 * as either signed or unsigned
1902 set_binop_left(n, pre);
1907 if (is_Conv(right)) {
1908 ir_mode *rmode = get_irn_mode(right);
1909 unsigned bits = get_mode_size_bits(rmode);
1911 if (ref_bits == bits &&
1912 mode_is_int(rmode) &&
1913 get_mode_arithmetic(rmode) == irma_twos_complement) {
1914 ir_node *pre = get_Conv_op(right);
1915 ir_mode *pre_mode = get_irn_mode(pre);
1917 if (mode_is_int(pre_mode) &&
1918 get_mode_size_bits(pre_mode) == bits &&
1919 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1920 /* ok, this conv just changes to sign, moreover the calculation
1921 * is done with same number of bits as our address mode, so
1922 * we can ignore the conv as address calculation can be viewed
1923 * as either signed or unsigned
1925 set_binop_right(n, pre);
1930 /* let address arithmetic use unsigned modes */
1931 if (is_Const(right)) {
1932 ir_mode *rmode = get_irn_mode(right);
1934 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
1935 /* convert a AddP(P, *s) into AddP(P, *u) */
1936 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
1938 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
1939 set_binop_right(n, pre);
1944 } /* transform_node_AddSub */
1946 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
1948 if (is_Const(b) && is_const_Phi(a)) { \
1949 /* check for Op(Phi, Const) */ \
1950 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
1952 else if (is_Const(a) && is_const_Phi(b)) { \
1953 /* check for Op(Const, Phi) */ \
1954 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
1956 else if (is_const_Phi(a) && is_const_Phi(b)) { \
1957 /* check for Op(Phi, Phi) */ \
1958 c = apply_binop_on_2_phis(a, b, eval, mode); \
1961 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1965 #define HANDLE_UNOP_PHI(eval, a, c) \
1967 if (is_const_Phi(a)) { \
1968 /* check for Op(Phi) */ \
1969 c = apply_unop_on_phi(a, eval); \
1971 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1977 * Do the AddSub optimization, then Transform
1978 * Constant folding on Phi
1979 * Add(a,a) -> Mul(a, 2)
1980 * Add(Mul(a, x), a) -> Mul(a, x+1)
1981 * if the mode is integer or float.
1982 * Transform Add(a,-b) into Sub(a,b).
1983 * Reassociation might fold this further.
1985 static ir_node *transform_node_Add(ir_node *n) {
1987 ir_node *a, *b, *c, *oldn = n;
1989 n = transform_node_AddSub(n);
1991 a = get_Add_left(n);
1992 b = get_Add_right(n);
1994 mode = get_irn_mode(n);
1995 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
1997 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1998 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2001 if (mode_is_num(mode)) {
2002 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2003 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2004 ir_node *block = get_irn_n(n, -1);
2007 get_irn_dbg_info(n),
2011 new_r_Const_long(current_ir_graph, block, mode, 2),
2013 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2018 get_irn_dbg_info(n),
2024 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2029 get_irn_dbg_info(n),
2035 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2038 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2039 /* Here we rely on constants be on the RIGHT side */
2041 ir_node *op = get_Not_op(a);
2043 if (is_Const(b) && is_Const_one(b)) {
2045 ir_node *blk = get_irn_n(n, -1);
2046 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2047 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2052 ir_node *blk = get_irn_n(n, -1);
2053 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2054 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2059 ir_node *op = get_Not_op(b);
2063 ir_node *blk = get_irn_n(n, -1);
2064 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2065 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2072 } /* transform_node_Add */
2075 * returns -cnst or NULL if impossible
2077 static ir_node *const_negate(ir_node *cnst) {
2078 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2079 dbg_info *dbgi = get_irn_dbg_info(cnst);
2080 ir_graph *irg = get_irn_irg(cnst);
2081 ir_node *block = get_nodes_block(cnst);
2082 ir_mode *mode = get_irn_mode(cnst);
2083 if (tv == tarval_bad) return NULL;
2084 return new_rd_Const(dbgi, irg, block, mode, tv);
2088 * Do the AddSub optimization, then Transform
2089 * Constant folding on Phi
2090 * Sub(0,a) -> Minus(a)
2091 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2092 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2093 * Sub(Add(a, x), x) -> a
2094 * Sub(x, Add(x, a)) -> -a
2095 * Sub(x, Const) -> Add(x, -Const)
2097 static ir_node *transform_node_Sub(ir_node *n) {
2102 n = transform_node_AddSub(n);
2104 a = get_Sub_left(n);
2105 b = get_Sub_right(n);
2107 mode = get_irn_mode(n);
2110 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2112 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2113 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2116 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2117 /* a - C -> a + (-C) */
2118 ir_node *cnst = const_negate(b);
2120 ir_node *block = get_nodes_block(n);
2121 dbg_info *dbgi = get_irn_dbg_info(n);
2122 ir_graph *irg = get_irn_irg(n);
2124 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2125 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2130 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2131 ir_graph *irg = current_ir_graph;
2132 dbg_info *dbg = get_irn_dbg_info(n);
2133 ir_node *block = get_nodes_block(n);
2134 ir_node *left = get_Minus_op(a);
2135 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2137 n = new_rd_Minus(dbg, irg, block, add, mode);
2138 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2140 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2141 ir_graph *irg = current_ir_graph;
2142 dbg_info *dbg = get_irn_dbg_info(n);
2143 ir_node *block = get_nodes_block(n);
2144 ir_node *right = get_Minus_op(b);
2146 n = new_rd_Add(dbg, irg, block, a, right, mode);
2147 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2149 } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */
2150 ir_graph *irg = current_ir_graph;
2151 dbg_info *s_dbg = get_irn_dbg_info(b);
2152 ir_node *s_block = get_nodes_block(b);
2153 ir_node *s_left = get_Sub_right(b);
2154 ir_node *s_right = get_Sub_left(b);
2155 ir_mode *s_mode = get_irn_mode(b);
2156 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode);
2157 dbg_info *a_dbg = get_irn_dbg_info(n);
2158 ir_node *a_block = get_nodes_block(n);
2160 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2161 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2163 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2164 ir_node *m_right = get_Mul_right(b);
2165 if (is_Const(m_right)) {
2166 ir_node *cnst2 = const_negate(m_right);
2167 if (cnst2 != NULL) {
2168 ir_graph *irg = current_ir_graph;
2169 dbg_info *m_dbg = get_irn_dbg_info(b);
2170 ir_node *m_block = get_nodes_block(b);
2171 ir_node *m_left = get_Mul_left(b);
2172 ir_mode *m_mode = get_irn_mode(b);
2173 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2174 dbg_info *a_dbg = get_irn_dbg_info(n);
2175 ir_node *a_block = get_nodes_block(n);
2177 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2178 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2184 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2185 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2187 get_irn_dbg_info(n),
2192 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2196 if (mode_wrap_around(mode)) {
2197 ir_node *left = get_Add_left(a);
2198 ir_node *right = get_Add_right(a);
2200 /* FIXME: Does the Conv's work only for two complement or generally? */
2202 if (mode != get_irn_mode(right)) {
2203 /* This Sub is an effective Cast */
2204 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2207 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2209 } else if (right == b) {
2210 if (mode != get_irn_mode(left)) {
2211 /* This Sub is an effective Cast */
2212 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2215 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2221 if (mode_wrap_around(mode)) {
2222 ir_node *left = get_Add_left(b);
2223 ir_node *right = get_Add_right(b);
2225 /* FIXME: Does the Conv's work only for two complement or generally? */
2227 ir_mode *r_mode = get_irn_mode(right);
2229 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode);
2230 if (mode != r_mode) {
2231 /* This Sub is an effective Cast */
2232 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2234 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2236 } else if (right == a) {
2237 ir_mode *l_mode = get_irn_mode(left);
2239 n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode);
2240 if (mode != l_mode) {
2241 /* This Sub is an effective Cast */
2242 n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode);
2244 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2249 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2250 ir_mode *mode = get_irn_mode(a);
2252 if (mode == get_irn_mode(b)) {
2254 ir_node *op_a = get_Conv_op(a);
2255 ir_node *op_b = get_Conv_op(b);
2257 /* check if it's allowed to skip the conv */
2258 ma = get_irn_mode(op_a);
2259 mb = get_irn_mode(op_b);
2261 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2262 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2265 set_Sub_right(n, b);
2271 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2272 if (!is_reassoc_running() && is_Mul(a)) {
2273 ir_node *ma = get_Mul_left(a);
2274 ir_node *mb = get_Mul_right(a);
2277 ir_node *blk = get_irn_n(n, -1);
2279 get_irn_dbg_info(n),
2280 current_ir_graph, blk,
2283 get_irn_dbg_info(n),
2284 current_ir_graph, blk,
2286 new_r_Const_long(current_ir_graph, blk, mode, 1),
2289 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2291 } else if (mb == b) {
2292 ir_node *blk = get_irn_n(n, -1);
2294 get_irn_dbg_info(n),
2295 current_ir_graph, blk,
2298 get_irn_dbg_info(n),
2299 current_ir_graph, blk,
2301 new_r_Const_long(current_ir_graph, blk, mode, 1),
2304 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2308 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2309 ir_node *x = get_Sub_left(a);
2310 ir_node *y = get_Sub_right(a);
2311 ir_node *blk = get_irn_n(n, -1);
2312 ir_mode *m_b = get_irn_mode(b);
2313 ir_mode *m_y = get_irn_mode(y);
2317 /* Determine the right mode for the Add. */
2320 else if (mode_is_reference(m_b))
2322 else if (mode_is_reference(m_y))
2326 * Both modes are different but none is reference,
2327 * happens for instance in SubP(SubP(P, Iu), Is).
2328 * We have two possibilities here: Cast or ignore.
2329 * Currently we ignore this case.
2334 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2336 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2337 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2341 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2342 if (is_Const(a) && is_Not(b)) {
2343 /* c - ~X = X + (c+1) */
2344 tarval *tv = get_Const_tarval(a);
2346 tv = tarval_add(tv, get_mode_one(mode));
2347 if (tv != tarval_bad) {
2348 ir_node *blk = get_irn_n(n, -1);
2349 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2350 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2351 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2357 } /* transform_node_Sub */
2360 * Several transformation done on n*n=2n bits mul.
2361 * These transformations must be done here because new nodes may be produced.
2363 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2365 ir_node *a = get_Mul_left(n);
2366 ir_node *b = get_Mul_right(n);
2367 tarval *ta = value_of(a);
2368 tarval *tb = value_of(b);
2369 ir_mode *smode = get_irn_mode(a);
2371 if (ta == get_mode_one(smode)) {
2372 /* (L)1 * (L)b = (L)b */
2373 ir_node *blk = get_irn_n(n, -1);
2374 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2375 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2378 else if (ta == get_mode_minus_one(smode)) {
2379 /* (L)-1 * (L)b = (L)b */
2380 ir_node *blk = get_irn_n(n, -1);
2381 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2382 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2383 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2386 if (tb == get_mode_one(smode)) {
2387 /* (L)a * (L)1 = (L)a */
2388 ir_node *blk = get_irn_n(a, -1);
2389 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2390 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2393 else if (tb == get_mode_minus_one(smode)) {
2394 /* (L)a * (L)-1 = (L)-a */
2395 ir_node *blk = get_irn_n(n, -1);
2396 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2397 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2398 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2405 * Transform Mul(a,-1) into -a.
2406 * Do constant evaluation of Phi nodes.
2407 * Do architecture dependent optimizations on Mul nodes
2409 static ir_node *transform_node_Mul(ir_node *n) {
2410 ir_node *c, *oldn = n;
2411 ir_mode *mode = get_irn_mode(n);
2412 ir_node *a = get_Mul_left(n);
2413 ir_node *b = get_Mul_right(n);
2415 if (is_Bad(a) || is_Bad(b))
2418 if (mode != get_irn_mode(a))
2419 return transform_node_Mul2n(n, mode);
2421 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2423 if (mode_is_signed(mode)) {
2426 if (value_of(a) == get_mode_minus_one(mode))
2428 else if (value_of(b) == get_mode_minus_one(mode))
2431 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2432 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2437 if (is_Const(b)) { /* (-a) * const -> a * -const */
2438 ir_node *cnst = const_negate(b);
2440 dbg_info *dbgi = get_irn_dbg_info(n);
2441 ir_node *block = get_nodes_block(n);
2442 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2443 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2446 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2447 dbg_info *dbgi = get_irn_dbg_info(n);
2448 ir_node *block = get_nodes_block(n);
2449 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2450 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2452 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2453 ir_node *sub_l = get_Sub_left(b);
2454 ir_node *sub_r = get_Sub_right(b);
2455 dbg_info *dbgi = get_irn_dbg_info(n);
2456 ir_graph *irg = current_ir_graph;
2457 ir_node *block = get_nodes_block(n);
2458 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2459 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2460 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2463 } else if (is_Minus(b)) {
2464 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2465 ir_node *sub_l = get_Sub_left(a);
2466 ir_node *sub_r = get_Sub_right(a);
2467 dbg_info *dbgi = get_irn_dbg_info(n);
2468 ir_graph *irg = current_ir_graph;
2469 ir_node *block = get_nodes_block(n);
2470 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2471 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2472 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2476 if (get_mode_arithmetic(mode) == irma_ieee754) {
2478 tarval *tv = get_Const_tarval(a);
2479 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2480 /* 2.0 * b = b + b */
2481 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2482 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2486 else if (is_Const(b)) {
2487 tarval *tv = get_Const_tarval(b);
2488 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) {
2489 /* a * 2.0 = a + a */
2490 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2491 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2496 return arch_dep_replace_mul_with_shifts(n);
2497 } /* transform_node_Mul */
2500 * Transform a Div Node.
2502 static ir_node *transform_node_Div(ir_node *n) {
2503 ir_mode *mode = get_Div_resmode(n);
2504 ir_node *a = get_Div_left(n);
2505 ir_node *b = get_Div_right(n);
2509 if (is_Const(b) && is_const_Phi(a)) {
2510 /* check for Div(Phi, Const) */
2511 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2513 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2517 else if (is_Const(a) && is_const_Phi(b)) {
2518 /* check for Div(Const, Phi) */
2519 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2521 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2525 else if (is_const_Phi(a) && is_const_Phi(b)) {
2526 /* check for Div(Phi, Phi) */
2527 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2529 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2536 if (tv != tarval_bad) {
2537 value = new_Const(get_tarval_mode(tv), tv);
2539 DBG_OPT_CSTEVAL(n, value);
2542 ir_node *a = get_Div_left(n);
2543 ir_node *b = get_Div_right(n);
2546 if (a == b && value_not_zero(a, &dummy)) {
2547 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2548 value = new_Const(mode, get_mode_one(mode));
2549 DBG_OPT_CSTEVAL(n, value);
2552 if (mode_is_signed(mode) && is_Const(b)) {
2553 tarval *tv = get_Const_tarval(b);
2555 if (tv == get_mode_minus_one(mode)) {
2557 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2558 DBG_OPT_CSTEVAL(n, value);
2562 /* Try architecture dependent optimization */
2563 value = arch_dep_replace_div_by_const(n);
2571 /* Turn Div into a tuple (mem, jmp, bad, value) */
2572 mem = get_Div_mem(n);
2573 blk = get_irn_n(n, -1);
2575 /* skip a potential Pin */
2577 mem = get_Pin_op(mem);
2578 turn_into_tuple(n, pn_Div_max);
2579 set_Tuple_pred(n, pn_Div_M, mem);
2580 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2581 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2582 set_Tuple_pred(n, pn_Div_res, value);
2585 } /* transform_node_Div */
2588 * Transform a Mod node.
2590 static ir_node *transform_node_Mod(ir_node *n) {
2591 ir_mode *mode = get_Mod_resmode(n);
2592 ir_node *a = get_Mod_left(n);
2593 ir_node *b = get_Mod_right(n);
2597 if (is_Const(b) && is_const_Phi(a)) {
2598 /* check for Div(Phi, Const) */
2599 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2601 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2605 else if (is_Const(a) && is_const_Phi(b)) {
2606 /* check for Div(Const, Phi) */
2607 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2609 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2613 else if (is_const_Phi(a) && is_const_Phi(b)) {
2614 /* check for Div(Phi, Phi) */
2615 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2617 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2624 if (tv != tarval_bad) {
2625 value = new_Const(get_tarval_mode(tv), tv);
2627 DBG_OPT_CSTEVAL(n, value);
2630 ir_node *a = get_Mod_left(n);
2631 ir_node *b = get_Mod_right(n);
2634 if (a == b && value_not_zero(a, &dummy)) {
2635 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2636 value = new_Const(mode, get_mode_null(mode));
2637 DBG_OPT_CSTEVAL(n, value);
2640 if (mode_is_signed(mode) && is_Const(b)) {
2641 tarval *tv = get_Const_tarval(b);
2643 if (tv == get_mode_minus_one(mode)) {
2645 value = new_Const(mode, get_mode_null(mode));
2646 DBG_OPT_CSTEVAL(n, value);
2650 /* Try architecture dependent optimization */
2651 value = arch_dep_replace_mod_by_const(n);
2659 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2660 mem = get_Mod_mem(n);
2661 blk = get_irn_n(n, -1);
2663 /* skip a potential Pin */
2665 mem = get_Pin_op(mem);
2666 turn_into_tuple(n, pn_Mod_max);
2667 set_Tuple_pred(n, pn_Mod_M, mem);
2668 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2669 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2670 set_Tuple_pred(n, pn_Mod_res, value);
2673 } /* transform_node_Mod */
2676 * Transform a DivMod node.
2678 static ir_node *transform_node_DivMod(ir_node *n) {
2680 ir_node *a = get_DivMod_left(n);
2681 ir_node *b = get_DivMod_right(n);
2682 ir_mode *mode = get_DivMod_resmode(n);
2687 if (is_Const(b) && is_const_Phi(a)) {
2688 /* check for Div(Phi, Const) */
2689 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2690 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2692 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2693 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2697 else if (is_Const(a) && is_const_Phi(b)) {
2698 /* check for Div(Const, Phi) */
2699 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2700 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2702 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2703 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2707 else if (is_const_Phi(a) && is_const_Phi(b)) {
2708 /* check for Div(Phi, Phi) */
2709 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2710 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2712 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2713 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2720 if (tb != tarval_bad) {
2721 if (tb == get_mode_one(get_tarval_mode(tb))) {
2723 vb = new_Const(mode, get_mode_null(mode));
2724 DBG_OPT_CSTEVAL(n, vb);
2726 } else if (ta != tarval_bad) {
2727 tarval *resa, *resb;
2728 resa = tarval_div(ta, tb);
2729 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2730 Jmp for X result!? */
2731 resb = tarval_mod(ta, tb);
2732 if (resb == tarval_bad) return n; /* Causes exception! */
2733 va = new_Const(mode, resa);
2734 vb = new_Const(mode, resb);
2735 DBG_OPT_CSTEVAL(n, va);
2736 DBG_OPT_CSTEVAL(n, vb);
2738 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
2739 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode);
2740 vb = new_Const(mode, get_mode_null(mode));
2741 DBG_OPT_CSTEVAL(n, va);
2742 DBG_OPT_CSTEVAL(n, vb);
2744 } else { /* Try architecture dependent optimization */
2747 arch_dep_replace_divmod_by_const(&va, &vb, n);
2748 evaluated = va != NULL;
2750 } else if (a == b) {
2751 if (value_not_zero(a, &dummy)) {
2753 va = new_Const(mode, get_mode_one(mode));
2754 vb = new_Const(mode, get_mode_null(mode));
2755 DBG_OPT_CSTEVAL(n, va);
2756 DBG_OPT_CSTEVAL(n, vb);
2759 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2762 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
2763 /* 0 / non-Const = 0 */
2768 if (evaluated) { /* replace by tuple */
2772 mem = get_DivMod_mem(n);
2773 /* skip a potential Pin */
2775 mem = get_Pin_op(mem);
2777 blk = get_irn_n(n, -1);
2778 turn_into_tuple(n, pn_DivMod_max);
2779 set_Tuple_pred(n, pn_DivMod_M, mem);
2780 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2781 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2782 set_Tuple_pred(n, pn_DivMod_res_div, va);
2783 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
2787 } /* transform_node_DivMod */
2790 * Optimize x / c to x * (1/c)
2792 static ir_node *transform_node_Quot(ir_node *n) {
2793 ir_mode *mode = get_Quot_resmode(n);
2796 if (get_mode_arithmetic(mode) == irma_ieee754) {
2797 ir_node *b = get_Quot_right(n);
2800 tarval *tv = get_Const_tarval(b);
2802 tv = tarval_quo(get_mode_one(mode), tv);
2804 /* Do the transformation if the result is either exact or we are not
2805 using strict rules. */
2806 if (tv != tarval_bad &&
2807 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
2808 ir_node *blk = get_irn_n(n, -1);
2809 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2810 ir_node *a = get_Quot_left(n);
2811 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
2812 ir_node *mem = get_Quot_mem(n);
2814 /* skip a potential Pin */
2816 mem = get_Pin_op(mem);
2817 turn_into_tuple(n, pn_Quot_max);
2818 set_Tuple_pred(n, pn_Quot_M, mem);
2819 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
2820 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
2821 set_Tuple_pred(n, pn_Quot_res, m);
2822 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
2827 } /* transform_node_Quot */
2830 * Optimize Abs(x) into x if x is Confirmed >= 0
2831 * Optimize Abs(x) into -x if x is Confirmed <= 0
2832 * Optimize Abs(-x) int Abs(x)
2834 static ir_node *transform_node_Abs(ir_node *n) {
2835 ir_node *c, *oldn = n;
2836 ir_node *a = get_Abs_op(n);
2839 HANDLE_UNOP_PHI(tarval_abs, a, c);
2841 switch (classify_value_sign(a)) {
2842 case value_classified_negative:
2843 mode = get_irn_mode(n);
2846 * We can replace the Abs by -x here.
2847 * We even could add a new Confirm here
2848 * (if not twos complement)
2850 * Note that -x would create a new node, so we could
2851 * not run it in the equivalent_node() context.
2853 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2854 get_nodes_block(n), a, mode);
2856 DBG_OPT_CONFIRM(oldn, n);
2858 case value_classified_positive:
2859 /* n is positive, Abs is not needed */
2862 DBG_OPT_CONFIRM(oldn, n);
2868 /* Abs(-x) = Abs(x) */
2869 mode = get_irn_mode(n);
2870 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
2871 get_nodes_block(n), get_Minus_op(a), mode);
2872 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
2876 } /* transform_node_Abs */
2879 * Transform a Cond node.
2881 * Replace the Cond by a Jmp if it branches on a constant
2884 static ir_node *transform_node_Cond(ir_node *n) {
2887 ir_node *a = get_Cond_selector(n);
2888 tarval *ta = value_of(a);
2890 /* we need block info which is not available in floating irgs */
2891 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2894 if ((ta != tarval_bad) &&
2895 (get_irn_mode(a) == mode_b) &&
2896 (get_opt_unreachable_code())) {
2897 /* It's a boolean Cond, branching on a boolean constant.
2898 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2899 ir_node *blk = get_nodes_block(n);
2900 jmp = new_r_Jmp(current_ir_graph, blk);
2901 turn_into_tuple(n, pn_Cond_max);
2902 if (ta == tarval_b_true) {
2903 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2904 set_Tuple_pred(n, pn_Cond_true, jmp);
2906 set_Tuple_pred(n, pn_Cond_false, jmp);
2907 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2909 /* We might generate an endless loop, so keep it alive. */
2910 add_End_keepalive(get_irg_end(current_ir_graph), blk);
2913 } /* transform_node_Cond */
2916 * Prototype of a recursive transform function
2917 * for bitwise distributive transformations.
2919 typedef ir_node* (*recursive_transform)(ir_node *n);
2922 * makes use of distributive laws for and, or, eor
2923 * and(a OP c, b OP c) -> and(a, b) OP c
2924 * note, might return a different op than n
2926 static ir_node *transform_bitwise_distributive(ir_node *n,
2927 recursive_transform trans_func)
2930 ir_node *a = get_binop_left(n);
2931 ir_node *b = get_binop_right(n);
2932 ir_op *op = get_irn_op(a);
2933 ir_op *op_root = get_irn_op(n);
2935 if(op != get_irn_op(b))
2938 if (op == op_Conv) {
2939 ir_node *a_op = get_Conv_op(a);
2940 ir_node *b_op = get_Conv_op(b);
2941 ir_mode *a_mode = get_irn_mode(a_op);
2942 ir_mode *b_mode = get_irn_mode(b_op);
2943 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2944 ir_node *blk = get_irn_n(n, -1);
2947 set_binop_left(n, a_op);
2948 set_binop_right(n, b_op);
2949 set_irn_mode(n, a_mode);
2951 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2953 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2959 /* nothing to gain here */
2963 if (op == op_Shrs || op == op_Shr || op == op_Shl
2964 || op == op_And || op == op_Or || op == op_Eor) {
2965 ir_node *a_left = get_binop_left(a);
2966 ir_node *a_right = get_binop_right(a);
2967 ir_node *b_left = get_binop_left(b);
2968 ir_node *b_right = get_binop_right(b);
2970 ir_node *op1 = NULL;
2971 ir_node *op2 = NULL;
2973 if (is_op_commutative(op)) {
2974 if (a_left == b_left) {
2978 } else if(a_left == b_right) {
2982 } else if(a_right == b_left) {
2988 if(a_right == b_right) {
2995 /* (a sop c) & (b sop c) => (a & b) sop c */
2996 ir_node *blk = get_irn_n(n, -1);
2998 ir_node *new_n = exact_copy(n);
2999 set_binop_left(new_n, op1);
3000 set_binop_right(new_n, op2);
3001 new_n = trans_func(new_n);
3003 if(op_root == op_Eor && op == op_Or) {
3004 dbg_info *dbgi = get_irn_dbg_info(n);
3005 ir_graph *irg = current_ir_graph;
3006 ir_mode *mode = get_irn_mode(c);
3008 c = new_rd_Not(dbgi, irg, blk, c, mode);
3009 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3012 set_nodes_block(n, blk);
3013 set_binop_left(n, new_n);
3014 set_binop_right(n, c);
3015 add_identities(current_ir_graph->value_table, n);
3018 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3029 static ir_node *transform_node_And(ir_node *n) {
3030 ir_node *c, *oldn = n;
3031 ir_node *a = get_And_left(n);
3032 ir_node *b = get_And_right(n);
3035 mode = get_irn_mode(n);
3036 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3038 /* we can evaluate 2 Projs of the same Cmp */
3039 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3040 ir_node *pred_a = get_Proj_pred(a);
3041 ir_node *pred_b = get_Proj_pred(b);
3042 if (pred_a == pred_b) {
3043 dbg_info *dbgi = get_irn_dbg_info(n);
3044 ir_node *block = get_nodes_block(pred_a);
3045 pn_Cmp pn_a = get_Proj_proj(a);
3046 pn_Cmp pn_b = get_Proj_proj(b);
3047 /* yes, we can simply calculate with pncs */
3048 pn_Cmp new_pnc = pn_a & pn_b;
3050 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3055 ir_node *op = get_Not_op(b);
3057 ir_node *ba = get_And_left(op);
3058 ir_node *bb = get_And_right(op);
3060 /* it's enough to test the following cases due to normalization! */
3061 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3062 /* (a|b) & ~(a&b) = a^b */
3063 ir_node *block = get_nodes_block(n);
3065 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3066 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3074 ir_node *op = get_Not_op(a);
3076 ir_node *aa = get_And_left(op);
3077 ir_node *ab = get_And_right(op);
3079 /* it's enough to test the following cases due to normalization! */
3080 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3081 /* (a|b) & ~(a&b) = a^b */
3082 ir_node *block = get_nodes_block(n);
3084 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3085 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3092 ir_node *al = get_Eor_left(a);
3093 ir_node *ar = get_Eor_right(a);
3096 /* (b ^ a) & b -> ~a & b */
3097 dbg_info *dbg = get_irn_dbg_info(n);
3098 ir_node *block = get_nodes_block(n);
3100 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3101 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3102 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3106 /* (a ^ b) & b -> ~a & b */
3107 dbg_info *dbg = get_irn_dbg_info(n);
3108 ir_node *block = get_nodes_block(n);
3110 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3111 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3112 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3117 ir_node *bl = get_Eor_left(b);
3118 ir_node *br = get_Eor_right(b);
3121 /* a & (a ^ b) -> a & ~b */
3122 dbg_info *dbg = get_irn_dbg_info(n);
3123 ir_node *block = get_nodes_block(n);
3125 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3126 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3127 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3131 /* a & (b ^ a) -> a & ~b */
3132 dbg_info *dbg = get_irn_dbg_info(n);
3133 ir_node *block = get_nodes_block(n);
3135 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3136 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3137 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3141 if (is_Not(a) && is_Not(b)) {
3142 /* ~a & ~b = ~(a|b) */
3143 ir_node *block = get_nodes_block(n);
3144 ir_mode *mode = get_irn_mode(n);
3148 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3149 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3150 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3154 n = transform_bitwise_distributive(n, transform_node_And);
3157 } /* transform_node_And */
3162 static ir_node *transform_node_Eor(ir_node *n) {
3163 ir_node *c, *oldn = n;
3164 ir_node *a = get_Eor_left(n);
3165 ir_node *b = get_Eor_right(n);
3166 ir_mode *mode = get_irn_mode(n);
3168 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3170 /* we can evaluate 2 Projs of the same Cmp */
3171 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3172 ir_node *pred_a = get_Proj_pred(a);
3173 ir_node *pred_b = get_Proj_pred(b);
3174 if(pred_a == pred_b) {
3175 dbg_info *dbgi = get_irn_dbg_info(n);
3176 ir_node *block = get_nodes_block(pred_a);
3177 pn_Cmp pn_a = get_Proj_proj(a);
3178 pn_Cmp pn_b = get_Proj_proj(b);
3179 /* yes, we can simply calculate with pncs */
3180 pn_Cmp new_pnc = pn_a ^ pn_b;
3182 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3189 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
3190 mode, get_mode_null(mode));
3191 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3192 } else if (mode == mode_b &&
3194 is_Const(b) && is_Const_one(b) &&
3195 is_Cmp(get_Proj_pred(a))) {
3196 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3197 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3198 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3200 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3201 } else if (is_Const(b)) {
3202 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3203 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3204 ir_node *not_op = get_Not_op(a);
3205 dbg_info *dbg = get_irn_dbg_info(n);
3206 ir_graph *irg = current_ir_graph;
3207 ir_node *block = get_nodes_block(n);
3208 ir_mode *mode = get_irn_mode(n);
3209 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3211 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3212 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3213 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3216 n = transform_bitwise_distributive(n, transform_node_Eor);
3220 } /* transform_node_Eor */
3225 static ir_node *transform_node_Not(ir_node *n) {
3226 ir_node *c, *oldn = n;
3227 ir_node *a = get_Not_op(n);
3228 ir_mode *mode = get_irn_mode(n);
3230 HANDLE_UNOP_PHI(tarval_not,a,c);
3232 /* check for a boolean Not */
3233 if (mode == mode_b &&
3235 is_Cmp(get_Proj_pred(a))) {
3236 /* We negate a Cmp. The Cmp has the negated result anyways! */
3237 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
3238 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3239 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3243 ir_node *eor_b = get_Eor_right(a);
3244 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3245 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3246 ir_node *eor_a = get_Eor_left(a);
3247 dbg_info *dbg = get_irn_dbg_info(n);
3248 ir_graph *irg = current_ir_graph;
3249 ir_node *block = get_nodes_block(n);
3250 ir_mode *mode = get_irn_mode(n);
3251 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3255 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3256 if (is_Minus(a)) { /* ~-x -> x + -1 */
3257 dbg_info *dbg = get_irn_dbg_info(n);
3258 ir_graph *irg = current_ir_graph;
3259 ir_node *block = get_nodes_block(n);
3260 ir_node *add_l = get_Minus_op(a);
3261 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3262 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3263 } else if (is_Add(a)) {
3264 ir_node *add_r = get_Add_right(a);
3265 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3266 /* ~(x + -1) = -x */
3267 ir_node *op = get_Add_left(a);
3268 ir_node *blk = get_irn_n(n, -1);
3269 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3270 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3275 } /* transform_node_Not */
3278 * Transform a Minus.
3282 * -(a >>u (size-1)) = a >>s (size-1)
3283 * -(a >>s (size-1)) = a >>u (size-1)
3284 * -(a * const) -> a * -const
3286 static ir_node *transform_node_Minus(ir_node *n) {
3287 ir_node *c, *oldn = n;
3288 ir_node *a = get_Minus_op(n);
3291 HANDLE_UNOP_PHI(tarval_neg,a,c);
3293 mode = get_irn_mode(a);
3294 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3295 /* the following rules are only to twos-complement */
3298 ir_node *op = get_Not_op(a);
3299 tarval *tv = get_mode_one(mode);
3300 ir_node *blk = get_irn_n(n, -1);
3301 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3302 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3303 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3307 ir_node *c = get_Shr_right(a);
3310 tarval *tv = get_Const_tarval(c);
3312 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3313 /* -(a >>u (size-1)) = a >>s (size-1) */
3314 ir_node *v = get_Shr_left(a);
3316 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3317 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3323 ir_node *c = get_Shrs_right(a);
3326 tarval *tv = get_Const_tarval(c);
3328 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3329 /* -(a >>s (size-1)) = a >>u (size-1) */
3330 ir_node *v = get_Shrs_left(a);
3332 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode);
3333 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3340 /* - (a-b) = b - a */
3341 ir_node *la = get_Sub_left(a);
3342 ir_node *ra = get_Sub_right(a);
3343 ir_node *blk = get_irn_n(n, -1);
3345 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3346 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3350 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3351 ir_node *mul_l = get_Mul_left(a);
3352 ir_node *mul_r = get_Mul_right(a);
3353 if (is_Const(mul_r)) {
3354 tarval *tv = tarval_neg(get_Const_tarval(mul_r));
3355 if(tv != tarval_bad) {
3356 ir_node *cnst = new_Const(mode, tv);
3357 dbg_info *dbg = get_irn_dbg_info(a);
3358 ir_graph *irg = current_ir_graph;
3359 ir_node *block = get_nodes_block(a);
3360 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3361 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3368 } /* transform_node_Minus */
3371 * Transform a Cast_type(Const) into a new Const_type
3373 static ir_node *transform_node_Cast(ir_node *n) {
3375 ir_node *pred = get_Cast_op(n);
3376 ir_type *tp = get_irn_type(n);
3378 if (is_Const(pred) && get_Const_type(pred) != tp) {
3379 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3380 get_Const_tarval(pred), tp);
3381 DBG_OPT_CSTEVAL(oldn, n);
3382 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3383 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3384 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3385 DBG_OPT_CSTEVAL(oldn, n);
3389 } /* transform_node_Cast */
3392 * Transform a Proj(Div) with a non-zero value.
3393 * Removes the exceptions and routes the memory to the NoMem node.
3395 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3396 ir_node *div = get_Proj_pred(proj);
3397 ir_node *b = get_Div_right(div);
3398 ir_node *confirm, *res, *new_mem;
3401 if (value_not_zero(b, &confirm)) {
3402 /* div(x, y) && y != 0 */
3403 if (confirm == NULL) {
3404 /* we are sure we have a Const != 0 */
3405 new_mem = get_Div_mem(div);
3406 if (is_Pin(new_mem))
3407 new_mem = get_Pin_op(new_mem);
3408 set_Div_mem(div, new_mem);
3409 set_irn_pinned(div, op_pin_state_floats);
3412 proj_nr = get_Proj_proj(proj);
3414 case pn_Div_X_regular:
3415 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3417 case pn_Div_X_except:
3418 /* we found an exception handler, remove it */
3419 DBG_OPT_EXC_REM(proj);
3423 res = get_Div_mem(div);
3424 new_mem = get_irg_no_mem(current_ir_graph);
3427 /* This node can only float up to the Confirm block */
3428 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3430 set_irn_pinned(div, op_pin_state_floats);
3431 /* this is a Div without exception, we can remove the memory edge */
3432 set_Div_mem(div, new_mem);
3437 } /* transform_node_Proj_Div */
3440 * Transform a Proj(Mod) with a non-zero value.
3441 * Removes the exceptions and routes the memory to the NoMem node.
3443 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3444 ir_node *mod = get_Proj_pred(proj);
3445 ir_node *b = get_Mod_right(mod);
3446 ir_node *confirm, *res, *new_mem;
3449 if (value_not_zero(b, &confirm)) {
3450 /* mod(x, y) && y != 0 */
3451 proj_nr = get_Proj_proj(proj);
3453 if (confirm == NULL) {
3454 /* we are sure we have a Const != 0 */
3455 new_mem = get_Mod_mem(mod);
3456 if (is_Pin(new_mem))
3457 new_mem = get_Pin_op(new_mem);
3458 set_Mod_mem(mod, new_mem);
3459 set_irn_pinned(mod, op_pin_state_floats);
3464 case pn_Mod_X_regular:
3465 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3467 case pn_Mod_X_except:
3468 /* we found an exception handler, remove it */
3469 DBG_OPT_EXC_REM(proj);
3473 res = get_Mod_mem(mod);
3474 new_mem = get_irg_no_mem(current_ir_graph);
3477 /* This node can only float up to the Confirm block */
3478 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3480 /* this is a Mod without exception, we can remove the memory edge */
3481 set_Mod_mem(mod, new_mem);
3484 if (get_Mod_left(mod) == b) {
3485 /* a % a = 0 if a != 0 */
3486 ir_mode *mode = get_irn_mode(proj);
3487 ir_node *res = new_Const(mode, get_mode_null(mode));
3489 DBG_OPT_CSTEVAL(mod, res);
3495 } /* transform_node_Proj_Mod */
3498 * Transform a Proj(DivMod) with a non-zero value.
3499 * Removes the exceptions and routes the memory to the NoMem node.
3501 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3502 ir_node *divmod = get_Proj_pred(proj);
3503 ir_node *b = get_DivMod_right(divmod);
3504 ir_node *confirm, *res, *new_mem;
3507 if (value_not_zero(b, &confirm)) {
3508 /* DivMod(x, y) && y != 0 */
3509 proj_nr = get_Proj_proj(proj);
3511 if (confirm == NULL) {
3512 /* we are sure we have a Const != 0 */
3513 new_mem = get_DivMod_mem(divmod);
3514 if (is_Pin(new_mem))
3515 new_mem = get_Pin_op(new_mem);
3516 set_DivMod_mem(divmod, new_mem);
3517 set_irn_pinned(divmod, op_pin_state_floats);
3522 case pn_DivMod_X_regular:
3523 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3525 case pn_DivMod_X_except:
3526 /* we found an exception handler, remove it */
3527 DBG_OPT_EXC_REM(proj);
3531 res = get_DivMod_mem(divmod);
3532 new_mem = get_irg_no_mem(current_ir_graph);
3535 /* This node can only float up to the Confirm block */
3536 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3538 /* this is a DivMod without exception, we can remove the memory edge */
3539 set_DivMod_mem(divmod, new_mem);
3542 case pn_DivMod_res_mod:
3543 if (get_DivMod_left(divmod) == b) {
3544 /* a % a = 0 if a != 0 */
3545 ir_mode *mode = get_irn_mode(proj);
3546 ir_node *res = new_Const(mode, get_mode_null(mode));
3548 DBG_OPT_CSTEVAL(divmod, res);
3554 } /* transform_node_Proj_DivMod */
3557 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3559 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3560 if (get_opt_unreachable_code()) {
3561 ir_node *n = get_Proj_pred(proj);
3562 ir_node *b = get_Cond_selector(n);
3564 if (mode_is_int(get_irn_mode(b))) {
3565 tarval *tb = value_of(b);
3567 if (tb != tarval_bad) {
3568 /* we have a constant switch */
3569 long num = get_Proj_proj(proj);
3571 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3572 if (get_tarval_long(tb) == num) {
3573 /* Do NOT create a jump here, or we will have 2 control flow ops
3574 * in a block. This case is optimized away in optimize_cf(). */
3577 /* this case will NEVER be taken, kill it */
3585 } /* transform_node_Proj_Cond */
3588 * Create a 0 constant of given mode.
3590 static ir_node *create_zero_const(ir_mode *mode) {
3591 tarval *tv = get_mode_null(mode);
3592 ir_node *cnst = new_Const(mode, tv);
3597 /* the order of the values is important! */
3598 typedef enum const_class {
3604 static const_class classify_const(const ir_node* n)
3606 if (is_Const(n)) return const_const;
3607 if (is_irn_constlike(n)) return const_like;
3612 * Determines whether r is more constlike or has a larger index (in that order)
3615 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3617 const const_class l_order = classify_const(l);
3618 const const_class r_order = classify_const(r);
3620 l_order > r_order ||
3621 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3625 * Normalizes and optimizes Cmp nodes.
3627 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3628 ir_node *n = get_Proj_pred(proj);
3629 ir_node *left = get_Cmp_left(n);
3630 ir_node *right = get_Cmp_right(n);
3633 ir_mode *mode = NULL;
3634 long proj_nr = get_Proj_proj(proj);
3636 /* we can evaluate some cases directly */
3639 return new_Const(mode_b, get_tarval_b_false());
3641 return new_Const(mode_b, get_tarval_b_true());
3643 if (!mode_is_float(get_irn_mode(left)))
3644 return new_Const(mode_b, get_tarval_b_true());
3650 /* remove Casts of both sides */
3651 left = skip_Cast(left);
3652 right = skip_Cast(right);
3654 /* Remove unnecessary conversions */
3655 /* TODO handle constants */
3656 if (is_Conv(left) && is_Conv(right)) {
3657 ir_mode *mode = get_irn_mode(left);
3658 ir_node *op_left = get_Conv_op(left);
3659 ir_node *op_right = get_Conv_op(right);
3660 ir_mode *mode_left = get_irn_mode(op_left);
3661 ir_mode *mode_right = get_irn_mode(op_right);
3663 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3664 && mode_left != mode_b && mode_right != mode_b) {
3665 ir_graph *irg = current_ir_graph;
3666 ir_node *block = get_nodes_block(n);
3668 if (mode_left == mode_right) {
3672 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3673 } else if (smaller_mode(mode_left, mode_right)) {
3674 left = new_r_Conv(irg, block, op_left, mode_right);
3677 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3678 } else if (smaller_mode(mode_right, mode_left)) {
3680 right = new_r_Conv(irg, block, op_right, mode_left);
3682 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3687 /* remove operation on both sides if possible */
3688 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3690 * The following operations are NOT safe for floating point operations, for instance
3691 * 1.0 + inf == 2.0 + inf, =/=> x == y
3693 if (mode_is_int(get_irn_mode(left))) {
3694 unsigned lop = get_irn_opcode(left);
3696 if (lop == get_irn_opcode(right)) {
3697 ir_node *ll, *lr, *rl, *rr;
3699 /* same operation on both sides, try to remove */
3703 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3704 left = get_unop_op(left);
3705 right = get_unop_op(right);
3707 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3710 ll = get_Add_left(left);
3711 lr = get_Add_right(left);
3712 rl = get_Add_left(right);
3713 rr = get_Add_right(right);
3716 /* X + a CMP X + b ==> a CMP b */
3720 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3721 } else if (ll == rr) {
3722 /* X + a CMP b + X ==> a CMP b */
3726 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3727 } else if (lr == rl) {
3728 /* a + X CMP X + b ==> a CMP b */
3732 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3733 } else if (lr == rr) {
3734 /* a + X CMP b + X ==> a CMP b */
3738 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3742 ll = get_Sub_left(left);
3743 lr = get_Sub_right(left);
3744 rl = get_Sub_left(right);
3745 rr = get_Sub_right(right);
3748 /* X - a CMP X - b ==> a CMP b */
3752 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3753 } else if (lr == rr) {
3754 /* a - X CMP b - X ==> a CMP b */
3758 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3762 if (get_Rot_right(left) == get_Rot_right(right)) {
3763 /* a ROT X CMP b ROT X ==> a CMP b */
3764 left = get_Rot_left(left);
3765 right = get_Rot_left(right);
3767 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3775 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3776 if (is_Add(left) || is_Sub(left)) {
3777 ir_node *ll = get_binop_left(left);
3778 ir_node *lr = get_binop_right(left);
3780 if (lr == right && is_Add(left)) {
3787 right = create_zero_const(get_irn_mode(left));
3789 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3792 if (is_Add(right) || is_Sub(right)) {
3793 ir_node *rl = get_binop_left(right);
3794 ir_node *rr = get_binop_right(right);
3796 if (rr == left && is_Add(right)) {
3803 right = create_zero_const(get_irn_mode(left));
3805 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3808 } /* mode_is_int(...) */
3809 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
3811 /* replace mode_b compares with ands/ors */
3812 if (get_irn_mode(left) == mode_b) {
3813 ir_graph *irg = current_ir_graph;
3814 ir_node *block = get_nodes_block(n);
3818 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3819 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
3820 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3821 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
3822 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
3823 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
3824 default: bres = NULL;
3827 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
3833 * First step: normalize the compare op
3834 * by placing the constant on the right side
3835 * or moving the lower address node to the left.
3837 if (!operands_are_normalized(left, right)) {
3843 proj_nr = get_inversed_pnc(proj_nr);
3848 * Second step: Try to reduce the magnitude
3849 * of a constant. This may help to generate better code
3850 * later and may help to normalize more compares.
3851 * Of course this is only possible for integer values.
3853 if (is_Const(right)) {
3854 mode = get_irn_mode(right);
3855 tv = get_Const_tarval(right);
3857 /* TODO extend to arbitrary constants */
3858 if (is_Conv(left) && tarval_is_null(tv)) {
3859 ir_node *op = get_Conv_op(left);
3860 ir_mode *op_mode = get_irn_mode(op);
3863 * UpConv(x) REL 0 ==> x REL 0
3865 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
3866 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
3867 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
3868 tv = get_mode_null(op_mode);
3872 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3876 if (tv != tarval_bad) {
3877 /* the following optimization is possible on modes without Overflow
3878 * on Unary Minus or on == and !=:
3879 * -a CMP c ==> a swap(CMP) -c
3881 * Beware: for two-complement Overflow may occur, so only == and != can
3882 * be optimized, see this:
3883 * -MININT < 0 =/=> MININT > 0 !!!
3885 if (is_Minus(left) &&
3886 (!mode_overflow_on_unary_Minus(mode) ||
3887 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
3888 tv = tarval_neg(tv);
3890 if (tv != tarval_bad) {
3891 left = get_Minus_op(left);
3892 proj_nr = get_inversed_pnc(proj_nr);
3894 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3896 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
3897 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
3898 tv = tarval_not(tv);
3900 if (tv != tarval_bad) {
3901 left = get_Not_op(left);
3903 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3907 /* for integer modes, we have more */
3908 if (mode_is_int(mode)) {
3909 /* Ne includes Unordered which is not possible on integers.
3910 * However, frontends often use this wrong, so fix it here */
3911 if (proj_nr & pn_Cmp_Uo) {
3912 proj_nr &= ~pn_Cmp_Uo;
3913 set_Proj_proj(proj, proj_nr);
3916 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
3917 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
3918 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
3919 tv = tarval_sub(tv, get_mode_one(mode));
3921 if (tv != tarval_bad) {
3922 proj_nr ^= pn_Cmp_Eq;
3924 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3927 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
3928 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
3929 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
3930 tv = tarval_add(tv, get_mode_one(mode));
3932 if (tv != tarval_bad) {
3933 proj_nr ^= pn_Cmp_Eq;
3935 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
3939 /* the following reassociations work only for == and != */
3940 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3942 #if 0 /* Might be not that good in general */
3943 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
3944 if (tarval_is_null(tv) && is_Sub(left)) {
3945 right = get_Sub_right(left);
3946 left = get_Sub_left(left);
3948 tv = value_of(right);
3950 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3954 if (tv != tarval_bad) {
3955 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
3957 ir_node *c1 = get_Sub_right(left);
3958 tarval *tv2 = value_of(c1);
3960 if (tv2 != tarval_bad) {
3961 tv2 = tarval_add(tv, value_of(c1));
3963 if (tv2 != tarval_bad) {
3964 left = get_Sub_left(left);
3967 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3971 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
3972 else if (is_Add(left)) {
3973 ir_node *a_l = get_Add_left(left);
3974 ir_node *a_r = get_Add_right(left);
3978 if (is_Const(a_l)) {
3980 tv2 = value_of(a_l);
3983 tv2 = value_of(a_r);
3986 if (tv2 != tarval_bad) {
3987 tv2 = tarval_sub(tv, tv2);
3989 if (tv2 != tarval_bad) {
3993 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
3997 /* -a == c ==> a == -c, -a != c ==> a != -c */
3998 else if (is_Minus(left)) {
3999 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
4001 if (tv2 != tarval_bad) {
4002 left = get_Minus_op(left);
4005 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4010 /* the following reassociations work only for <= */
4011 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4012 if (tv != tarval_bad) {
4013 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4014 if (get_irn_op(left) == op_Abs) { // TODO something is missing here
4020 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4021 switch (get_irn_opcode(left)) {
4025 c1 = get_And_right(left);
4028 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4029 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4031 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4033 /* TODO: move to constant evaluation */
4034 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4035 c1 = new_Const(mode_b, tv);
4036 DBG_OPT_CSTEVAL(proj, c1);
4040 if (tarval_is_single_bit(tv)) {
4042 * optimization for AND:
4044 * And(x, C) == C ==> And(x, C) != 0
4045 * And(x, C) != C ==> And(X, C) == 0
4047 * if C is a single Bit constant.
4050 /* check for Constant's match. We have check hare the tarvals,
4051 because our const might be changed */
4052 if (get_Const_tarval(c1) == tv) {
4053 /* fine: do the transformation */
4054 tv = get_mode_null(get_tarval_mode(tv));
4055 proj_nr ^= pn_Cmp_Leg;
4057 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4063 c1 = get_Or_right(left);
4064 if (is_Const(c1) && tarval_is_null(tv)) {
4066 * Or(x, C) == 0 && C != 0 ==> FALSE
4067 * Or(x, C) != 0 && C != 0 ==> TRUE
4069 if (! tarval_is_null(get_Const_tarval(c1))) {
4070 /* TODO: move to constant evaluation */
4071 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4072 c1 = new_Const(mode_b, tv);
4073 DBG_OPT_CSTEVAL(proj, c1);
4080 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4082 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4085 c1 = get_Shl_right(left);
4087 tarval *tv1 = get_Const_tarval(c1);
4088 ir_mode *mode = get_irn_mode(left);
4089 tarval *minus1 = get_mode_all_one(mode);
4090 tarval *amask = tarval_shr(minus1, tv1);
4091 tarval *cmask = tarval_shl(minus1, tv1);
4094 if (tarval_and(tv, cmask) != tv) {
4095 /* condition not met */
4096 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4097 c1 = new_Const(mode_b, tv);
4098 DBG_OPT_CSTEVAL(proj, c1);
4101 sl = get_Shl_left(left);
4102 blk = get_nodes_block(n);
4103 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4104 tv = tarval_shr(tv, tv1);
4106 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4111 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4113 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4116 c1 = get_Shr_right(left);
4118 tarval *tv1 = get_Const_tarval(c1);
4119 ir_mode *mode = get_irn_mode(left);
4120 tarval *minus1 = get_mode_all_one(mode);
4121 tarval *amask = tarval_shl(minus1, tv1);
4122 tarval *cmask = tarval_shr(minus1, tv1);
4125 if (tarval_and(tv, cmask) != tv) {
4126 /* condition not met */
4127 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4128 c1 = new_Const(mode_b, tv);
4129 DBG_OPT_CSTEVAL(proj, c1);
4132 sl = get_Shr_left(left);
4133 blk = get_nodes_block(n);
4134 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4135 tv = tarval_shl(tv, tv1);
4137 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4142 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4144 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4147 c1 = get_Shrs_right(left);
4149 tarval *tv1 = get_Const_tarval(c1);
4150 ir_mode *mode = get_irn_mode(left);
4151 tarval *minus1 = get_mode_all_one(mode);
4152 tarval *amask = tarval_shl(minus1, tv1);
4153 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4156 cond = tarval_sub(cond, tv1);
4157 cond = tarval_shrs(tv, cond);
4159 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4160 /* condition not met */
4161 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4162 c1 = new_Const(mode_b, tv);
4163 DBG_OPT_CSTEVAL(proj, c1);
4166 sl = get_Shrs_left(left);
4167 blk = get_nodes_block(n);
4168 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4169 tv = tarval_shl(tv, tv1);
4171 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4176 } /* tarval != bad */
4179 if (changed & 2) /* need a new Const */
4180 right = new_Const(mode, tv);
4182 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4183 ir_node *op = get_Proj_pred(left);
4185 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4186 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4187 ir_node *c = get_binop_right(op);
4190 tarval *tv = get_Const_tarval(c);
4192 if (tarval_is_single_bit(tv)) {
4193 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4194 ir_node *v = get_binop_left(op);
4195 ir_node *blk = get_irn_n(op, -1);
4196 ir_mode *mode = get_irn_mode(v);
4198 tv = tarval_sub(tv, get_mode_one(mode));
4199 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4201 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4208 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
4210 /* create a new compare */
4211 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4212 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4216 } /* transform_node_Proj_Cmp */
4219 * Does all optimizations on nodes that must be done on it's Proj's
4220 * because of creating new nodes.
4222 static ir_node *transform_node_Proj(ir_node *proj) {
4223 ir_node *n = get_Proj_pred(proj);
4225 switch (get_irn_opcode(n)) {
4227 return transform_node_Proj_Div(proj);
4230 return transform_node_Proj_Mod(proj);
4233 return transform_node_Proj_DivMod(proj);
4236 return transform_node_Proj_Cond(proj);
4239 return transform_node_Proj_Cmp(proj);
4242 /* should not happen, but if it does will be optimized away */
4243 return equivalent_node_Proj(proj);
4249 } /* transform_node_Proj */
4252 * Move Confirms down through Phi nodes.
4254 static ir_node *transform_node_Phi(ir_node *phi) {
4256 ir_mode *mode = get_irn_mode(phi);
4258 if (mode_is_reference(mode)) {
4259 n = get_irn_arity(phi);
4261 /* Beware of Phi0 */
4263 ir_node *pred = get_irn_n(phi, 0);
4264 ir_node *bound, *new_Phi, *block, **in;
4267 if (! is_Confirm(pred))
4270 bound = get_Confirm_bound(pred);
4271 pnc = get_Confirm_cmp(pred);
4273 NEW_ARR_A(ir_node *, in, n);
4274 in[0] = get_Confirm_value(pred);
4276 for (i = 1; i < n; ++i) {
4277 pred = get_irn_n(phi, i);
4279 if (! is_Confirm(pred) ||
4280 get_Confirm_bound(pred) != bound ||
4281 get_Confirm_cmp(pred) != pnc)
4283 in[i] = get_Confirm_value(pred);
4285 /* move the Confirm nodes "behind" the Phi */
4286 block = get_irn_n(phi, -1);
4287 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4288 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4292 } /* transform_node_Phi */
4295 * Returns the operands of a commutative bin-op, if one operand is
4296 * a const, it is returned as the second one.
4298 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4299 ir_node *op_a = get_binop_left(binop);
4300 ir_node *op_b = get_binop_right(binop);
4302 assert(is_op_commutative(get_irn_op(binop)));
4304 if (is_Const(op_a)) {
4311 } /* get_comm_Binop_Ops */
4314 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4315 * Such pattern may arise in bitfield stores.
4317 * value c4 value c4 & c2
4318 * AND c3 AND c1 | c3
4325 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4328 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4331 ir_node *and_l, *c3;
4332 ir_node *value, *c4;
4333 ir_node *new_and, *new_const, *block;
4334 ir_mode *mode = get_irn_mode(or);
4336 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4339 get_comm_Binop_Ops(or, &and, &c1);
4340 if (!is_Const(c1) || !is_And(and))
4343 get_comm_Binop_Ops(and, &or_l, &c2);
4347 tv1 = get_Const_tarval(c1);
4348 tv2 = get_Const_tarval(c2);
4350 tv = tarval_or(tv1, tv2);
4351 if (tarval_is_all_one(tv)) {
4352 /* the AND does NOT clear a bit with isn't set by the OR */
4353 set_Or_left(or, or_l);
4354 set_Or_right(or, c1);
4356 /* check for more */
4363 get_comm_Binop_Ops(or_l, &and_l, &c3);
4364 if (!is_Const(c3) || !is_And(and_l))
4367 get_comm_Binop_Ops(and_l, &value, &c4);
4371 /* ok, found the pattern, check for conditions */
4372 assert(mode == get_irn_mode(and));
4373 assert(mode == get_irn_mode(or_l));
4374 assert(mode == get_irn_mode(and_l));
4376 tv3 = get_Const_tarval(c3);
4377 tv4 = get_Const_tarval(c4);
4379 tv = tarval_or(tv4, tv2);
4380 if (!tarval_is_all_one(tv)) {
4381 /* have at least one 0 at the same bit position */
4385 n_tv4 = tarval_not(tv4);
4386 if (tv3 != tarval_and(tv3, n_tv4)) {
4387 /* bit in the or_mask is outside the and_mask */
4391 n_tv2 = tarval_not(tv2);
4392 if (tv1 != tarval_and(tv1, n_tv2)) {
4393 /* bit in the or_mask is outside the and_mask */
4397 /* ok, all conditions met */
4398 block = get_irn_n(or, -1);
4400 new_and = new_r_And(current_ir_graph, block,
4401 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4403 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4405 set_Or_left(or, new_and);
4406 set_Or_right(or, new_const);
4408 /* check for more */
4410 } /* transform_node_Or_bf_store */
4413 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
4415 static ir_node *transform_node_Or_Rot(ir_node *or) {
4416 ir_mode *mode = get_irn_mode(or);
4417 ir_node *shl, *shr, *block;
4418 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
4421 if (! mode_is_int(mode))
4424 shl = get_binop_left(or);
4425 shr = get_binop_right(or);
4434 } else if (!is_Shl(shl)) {
4436 } else if (!is_Shr(shr)) {
4439 x = get_Shl_left(shl);
4440 if (x != get_Shr_left(shr))
4443 c1 = get_Shl_right(shl);
4444 c2 = get_Shr_right(shr);
4445 if (is_Const(c1) && is_Const(c2)) {
4446 tv1 = get_Const_tarval(c1);
4447 if (! tarval_is_long(tv1))
4450 tv2 = get_Const_tarval(c2);
4451 if (! tarval_is_long(tv2))
4454 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4455 != (int) get_mode_size_bits(mode))
4458 /* yet, condition met */
4459 block = get_irn_n(or, -1);
4461 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
4463 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
4465 } else if (is_Sub(c1)) {
4469 if (get_Sub_right(sub) != v)
4472 c1 = get_Sub_left(sub);
4476 tv1 = get_Const_tarval(c1);
4477 if (! tarval_is_long(tv1))
4480 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4483 /* yet, condition met */
4484 block = get_nodes_block(or);
4486 /* a Rot right is not supported, so use a rot left */
4487 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
4489 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4491 } else if (is_Sub(c2)) {
4495 c1 = get_Sub_left(sub);
4499 tv1 = get_Const_tarval(c1);
4500 if (! tarval_is_long(tv1))
4503 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4506 /* yet, condition met */
4507 block = get_irn_n(or, -1);
4510 n = new_r_Rot(current_ir_graph, block, x, v, mode);
4512 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
4517 } /* transform_node_Or_Rot */
4522 static ir_node *transform_node_Or(ir_node *n) {
4523 ir_node *c, *oldn = n;
4524 ir_node *a = get_Or_left(n);
4525 ir_node *b = get_Or_right(n);
4528 if (is_Not(a) && is_Not(b)) {
4529 /* ~a | ~b = ~(a&b) */
4530 ir_node *block = get_nodes_block(n);
4532 mode = get_irn_mode(n);
4535 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4536 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4537 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4541 /* we can evaluate 2 Projs of the same Cmp */
4542 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4543 ir_node *pred_a = get_Proj_pred(a);
4544 ir_node *pred_b = get_Proj_pred(b);
4545 if (pred_a == pred_b) {
4546 dbg_info *dbgi = get_irn_dbg_info(n);
4547 ir_node *block = get_nodes_block(pred_a);
4548 pn_Cmp pn_a = get_Proj_proj(a);
4549 pn_Cmp pn_b = get_Proj_proj(b);
4550 /* yes, we can simply calculate with pncs */
4551 pn_Cmp new_pnc = pn_a | pn_b;
4553 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4558 mode = get_irn_mode(n);
4559 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4561 n = transform_node_Or_bf_store(n);
4562 n = transform_node_Or_Rot(n);
4566 n = transform_bitwise_distributive(n, transform_node_Or);
4569 } /* transform_node_Or */
4573 static ir_node *transform_node(ir_node *n);
4576 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot.
4578 * Should be moved to reassociation?
4580 static ir_node *transform_node_shift(ir_node *n) {
4581 ir_node *left, *right;
4582 tarval *tv1, *tv2, *res;
4584 int modulo_shf, flag;
4586 left = get_binop_left(n);
4588 /* different operations */
4589 if (get_irn_op(left) != get_irn_op(n))
4592 right = get_binop_right(n);
4593 tv1 = value_of(right);
4594 if (tv1 == tarval_bad)
4597 tv2 = value_of(get_binop_right(left));
4598 if (tv2 == tarval_bad)
4601 res = tarval_add(tv1, tv2);
4603 /* beware: a simple replacement works only, if res < modulo shift */
4604 mode = get_irn_mode(n);
4608 modulo_shf = get_mode_modulo_shift(mode);
4609 if (modulo_shf > 0) {
4610 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
4612 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
4618 /* ok, we can replace it */
4619 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
4621 in[0] = get_binop_left(left);
4622 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
4624 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
4626 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4628 return transform_node(irn);
4631 } /* transform_node_shift */
4636 static ir_node *transform_node_Shr(ir_node *n) {
4637 ir_node *c, *oldn = n;
4638 ir_node *a = get_Shr_left(n);
4639 ir_node *b = get_Shr_right(n);
4640 ir_mode *mode = get_irn_mode(n);
4642 HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode);
4643 return transform_node_shift(n);
4644 } /* transform_node_Shr */
4649 static ir_node *transform_node_Shrs(ir_node *n) {
4650 ir_node *c, *oldn = n;
4651 ir_node *a = get_Shrs_left(n);
4652 ir_node *b = get_Shrs_right(n);
4653 ir_mode *mode = get_irn_mode(n);
4655 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
4656 return transform_node_shift(n);
4657 } /* transform_node_Shrs */
4662 static ir_node *transform_node_Shl(ir_node *n) {
4663 ir_node *c, *oldn = n;
4664 ir_node *a = get_Shl_left(n);
4665 ir_node *b = get_Shl_right(n);
4666 ir_mode *mode = get_irn_mode(n);
4668 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
4669 return transform_node_shift(n);
4670 } /* transform_node_Shl */
4675 static ir_node *transform_node_Rot(ir_node *n) {
4676 ir_node *c, *oldn = n;
4677 ir_node *a = get_Rot_left(n);
4678 ir_node *b = get_Rot_right(n);
4679 ir_mode *mode = get_irn_mode(n);
4681 HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode);
4682 return transform_node_shift(n);
4683 } /* transform_node_Rot */
4688 static ir_node *transform_node_Conv(ir_node *n) {
4689 ir_node *c, *oldn = n;
4690 ir_node *a = get_Conv_op(n);
4692 if (is_const_Phi(a)) {
4693 c = apply_conv_on_phi(a, get_irn_mode(n));
4695 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
4700 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
4701 ir_mode *mode = get_irn_mode(n);
4702 return new_r_Unknown(current_ir_graph, mode);
4706 } /* transform_node_Conv */
4709 * Remove dead blocks and nodes in dead blocks
4710 * in keep alive list. We do not generate a new End node.
4712 static ir_node *transform_node_End(ir_node *n) {
4713 int i, j, n_keepalives = get_End_n_keepalives(n);
4716 NEW_ARR_A(ir_node *, in, n_keepalives);
4718 for (i = j = 0; i < n_keepalives; ++i) {
4719 ir_node *ka = get_End_keepalive(n, i);
4721 if (! is_Block_dead(ka)) {
4725 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
4728 /* FIXME: beabi need to keep a Proj(M) */
4729 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
4732 if (j != n_keepalives)
4733 set_End_keepalives(n, j, in);
4735 } /* transform_node_End */
4737 /** returns 1 if a == -b */
4738 static int is_negated_value(ir_node *a, ir_node *b) {
4739 if(is_Minus(a) && get_Minus_op(a) == b)
4741 if(is_Minus(b) && get_Minus_op(b) == a)
4743 if(is_Sub(a) && is_Sub(b)) {
4744 ir_node *a_left = get_Sub_left(a);
4745 ir_node *a_right = get_Sub_right(a);
4746 ir_node *b_left = get_Sub_left(b);
4747 ir_node *b_right = get_Sub_right(b);
4749 if(a_left == b_right && a_right == b_left)
4757 * Optimize a Mux into some simpler cases.
4759 static ir_node *transform_node_Mux(ir_node *n) {
4760 ir_node *oldn = n, *sel = get_Mux_sel(n);
4761 ir_mode *mode = get_irn_mode(n);
4763 if (mode == mode_b) {
4764 ir_node *t = get_Mux_true(n);
4765 ir_node *f = get_Mux_false(n);
4766 dbg_info *dbg = get_irn_dbg_info(n);
4767 ir_node *block = get_irn_n(n, -1);
4768 ir_graph *irg = current_ir_graph;
4771 tarval *tv_t = get_Const_tarval(t);
4772 if (tv_t == tarval_b_true) {
4774 assert(get_Const_tarval(f) == tarval_b_false);
4777 return new_rd_Or(dbg, irg, block, sel, f, mode_b);
4780 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4781 assert(tv_t == tarval_b_false);
4783 assert(get_Const_tarval(f) == tarval_b_true);
4786 return new_rd_And(dbg, irg, block, not_sel, f, mode_b);
4789 } else if (is_Const(f)) {
4790 tarval *tv_f = get_Const_tarval(f);
4791 if (tv_f == tarval_b_true) {
4792 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
4793 return new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
4795 assert(tv_f == tarval_b_false);
4796 return new_rd_And(dbg, irg, block, sel, t, mode_b);
4801 if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) {
4802 ir_node *cmp = get_Proj_pred(sel);
4803 long pn = get_Proj_proj(sel);
4804 ir_node *f = get_Mux_false(n);
4805 ir_node *t = get_Mux_true(n);
4808 * Note: normalization puts the constant on the right side,
4809 * so we check only one case.
4811 * Note further that these optimization work even for floating point
4812 * with NaN's because -NaN == NaN.
4813 * However, if +0 and -0 is handled differently, we cannot use the first
4817 ir_node *cmp_r = get_Cmp_right(cmp);
4818 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
4819 ir_node *block = get_irn_n(n, -1);
4821 if(is_negated_value(f, t)) {
4822 ir_node *cmp_left = get_Cmp_left(cmp);
4824 /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */
4825 if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
4826 || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
4828 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4830 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4832 /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */
4833 } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
4834 || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
4836 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
4838 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
4840 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
4847 return arch_transform_node_Mux(n);
4848 } /* transform_node_Mux */
4851 * Optimize a Psi into some simpler cases.
4853 static ir_node *transform_node_Psi(ir_node *n) {
4855 return transform_node_Mux(n);
4858 } /* transform_node_Psi */
4861 * optimize sync nodes that have other syncs as input we simply add the inputs
4862 * of the other sync to our own inputs
4864 static ir_node *transform_node_Sync(ir_node *n) {
4865 int arity = get_Sync_n_preds(n);
4868 for (i = 0; i < arity;) {
4869 ir_node *pred = get_Sync_pred(n, i);
4873 if (!is_Sync(pred)) {
4881 pred_arity = get_Sync_n_preds(pred);
4882 for (j = 0; j < pred_arity; ++j) {
4883 ir_node *pred_pred = get_Sync_pred(pred, j);
4888 add_irn_n(n, pred_pred);
4892 if (get_Sync_pred(n, k) == pred_pred) break;
4897 /* rehash the sync node */
4898 add_identities(current_ir_graph->value_table, n);
4904 * Tries several [inplace] [optimizing] transformations and returns an
4905 * equivalent node. The difference to equivalent_node() is that these
4906 * transformations _do_ generate new nodes, and thus the old node must
4907 * not be freed even if the equivalent node isn't the old one.
4909 static ir_node *transform_node(ir_node *n) {
4913 * Transform_node is the only "optimizing transformation" that might
4914 * return a node with a different opcode. We iterate HERE until fixpoint
4915 * to get the final result.
4919 if (n->op->ops.transform_node)
4920 n = n->op->ops.transform_node(n);
4921 } while (oldn != n);
4924 } /* transform_node */
4927 * Sets the default transform node operation for an ir_op_ops.
4929 * @param code the opcode for the default operation
4930 * @param ops the operations initialized
4935 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
4939 ops->transform_node = transform_node_##a; \
4976 } /* firm_set_default_transform_node */
4979 /* **************** Common Subexpression Elimination **************** */
4981 /** The size of the hash table used, should estimate the number of nodes
4983 #define N_IR_NODES 512
4985 /** Compares the attributes of two Const nodes. */
4986 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
4987 return (get_Const_tarval(a) != get_Const_tarval(b))
4988 || (get_Const_type(a) != get_Const_type(b));
4989 } /* node_cmp_attr_Const */
4991 /** Compares the attributes of two Proj nodes. */
4992 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
4993 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
4994 } /* node_cmp_attr_Proj */
4996 /** Compares the attributes of two Filter nodes. */
4997 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
4998 return get_Filter_proj(a) != get_Filter_proj(b);
4999 } /* node_cmp_attr_Filter */
5001 /** Compares the attributes of two Alloc nodes. */
5002 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5003 const alloc_attr *pa = get_irn_alloc_attr(a);
5004 const alloc_attr *pb = get_irn_alloc_attr(b);
5005 return (pa->where != pb->where) || (pa->type != pb->type);
5006 } /* node_cmp_attr_Alloc */
5008 /** Compares the attributes of two Free nodes. */
5009 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5010 const free_attr *pa = get_irn_free_attr(a);
5011 const free_attr *pb = get_irn_free_attr(b);
5012 return (pa->where != pb->where) || (pa->type != pb->type);
5013 } /* node_cmp_attr_Free */
5015 /** Compares the attributes of two SymConst nodes. */
5016 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5017 const symconst_attr *pa = get_irn_symconst_attr(a);
5018 const symconst_attr *pb = get_irn_symconst_attr(b);
5019 return (pa->num != pb->num)
5020 || (pa->sym.type_p != pb->sym.type_p)
5021 || (pa->tp != pb->tp);
5022 } /* node_cmp_attr_SymConst */
5024 /** Compares the attributes of two Call nodes. */
5025 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5026 return get_irn_call_attr(a) != get_irn_call_attr(b);
5027 } /* node_cmp_attr_Call */
5029 /** Compares the attributes of two Sel nodes. */
5030 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5031 const ir_entity *a_ent = get_Sel_entity(a);
5032 const ir_entity *b_ent = get_Sel_entity(b);
5034 (a_ent->kind != b_ent->kind) ||
5035 (a_ent->name != b_ent->name) ||
5036 (a_ent->owner != b_ent->owner) ||
5037 (a_ent->ld_name != b_ent->ld_name) ||
5038 (a_ent->type != b_ent->type);
5039 } /* node_cmp_attr_Sel */
5041 /** Compares the attributes of two Phi nodes. */
5042 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5043 /* we can only enter this function if both nodes have the same number of inputs,
5044 hence it is enough to check if one of them is a Phi0 */
5046 /* check the Phi0 pos attribute */
5047 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5050 } /* node_cmp_attr_Phi */
5052 /** Compares the attributes of two Conv nodes. */
5053 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5054 return get_Conv_strict(a) != get_Conv_strict(b);
5055 } /* node_cmp_attr_Conv */
5057 /** Compares the attributes of two Cast nodes. */
5058 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5059 return get_Cast_type(a) != get_Cast_type(b);
5060 } /* node_cmp_attr_Cast */
5062 /** Compares the attributes of two Load nodes. */
5063 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5064 if (get_Load_volatility(a) == volatility_is_volatile ||
5065 get_Load_volatility(b) == volatility_is_volatile)
5066 /* NEVER do CSE on volatile Loads */
5068 /* do not CSE Loads with different alignment. Be conservative. */
5069 if (get_Load_align(a) != get_Load_align(b))
5072 return get_Load_mode(a) != get_Load_mode(b);
5073 } /* node_cmp_attr_Load */
5075 /** Compares the attributes of two Store nodes. */
5076 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5077 /* do not CSE Stores with different alignment. Be conservative. */
5078 if (get_Store_align(a) != get_Store_align(b))
5081 /* NEVER do CSE on volatile Stores */
5082 return (get_Store_volatility(a) == volatility_is_volatile ||
5083 get_Store_volatility(b) == volatility_is_volatile);
5084 } /* node_cmp_attr_Store */
5086 /** Compares two exception attributes */
5087 static int node_cmp_exception(ir_node *a, ir_node *b) {
5088 const except_attr *ea = get_irn_except_attr(a);
5089 const except_attr *eb = get_irn_except_attr(b);
5091 return ea->pin_state != eb->pin_state;
5094 #define node_cmp_attr_Bound node_cmp_exception
5096 /** Compares the attributes of two Div nodes. */
5097 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5098 const divmod_attr *ma = get_irn_divmod_attr(a);
5099 const divmod_attr *mb = get_irn_divmod_attr(b);
5100 return ma->exc.pin_state != mb->exc.pin_state ||
5101 ma->res_mode != mb->res_mode ||
5102 ma->no_remainder != mb->no_remainder;
5103 } /* node_cmp_attr_Div */
5105 /** Compares the attributes of two DivMod nodes. */
5106 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5107 const divmod_attr *ma = get_irn_divmod_attr(a);
5108 const divmod_attr *mb = get_irn_divmod_attr(b);
5109 return ma->exc.pin_state != mb->exc.pin_state ||
5110 ma->res_mode != mb->res_mode;
5111 } /* node_cmp_attr_DivMod */
5113 /** Compares the attributes of two Mod nodes. */
5114 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5115 const divmod_attr *ma = get_irn_divmod_attr(a);
5116 const divmod_attr *mb = get_irn_divmod_attr(b);
5117 return ma->exc.pin_state != mb->exc.pin_state ||
5118 ma->res_mode != mb->res_mode;
5119 } /* node_cmp_attr_Mod */
5121 /** Compares the attributes of two Quot nodes. */
5122 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5123 const divmod_attr *ma = get_irn_divmod_attr(a);
5124 const divmod_attr *mb = get_irn_divmod_attr(b);
5125 return ma->exc.pin_state != mb->exc.pin_state ||
5126 ma->res_mode != mb->res_mode;
5127 } /* node_cmp_attr_Quot */
5129 /** Compares the attributes of two Confirm nodes. */
5130 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5131 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5132 } /* node_cmp_attr_Confirm */
5134 /** Compares the attributes of two ASM nodes. */
5135 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5137 const ir_asm_constraint *ca;
5138 const ir_asm_constraint *cb;
5141 if (get_ASM_text(a) != get_ASM_text(b))
5144 /* Should we really check the constraints here? Should be better, but is strange. */
5145 n = get_ASM_n_input_constraints(a);
5146 if (n != get_ASM_n_input_constraints(b))
5149 ca = get_ASM_input_constraints(a);
5150 cb = get_ASM_input_constraints(b);
5151 for (i = 0; i < n; ++i) {
5152 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5156 n = get_ASM_n_output_constraints(a);
5157 if (n != get_ASM_n_output_constraints(b))
5160 ca = get_ASM_output_constraints(a);
5161 cb = get_ASM_output_constraints(b);
5162 for (i = 0; i < n; ++i) {
5163 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5167 n = get_ASM_n_clobbers(a);
5168 if (n != get_ASM_n_clobbers(b))
5171 cla = get_ASM_clobbers(a);
5172 clb = get_ASM_clobbers(b);
5173 for (i = 0; i < n; ++i) {
5174 if (cla[i] != clb[i])
5178 } /* node_cmp_attr_ASM */
5181 * Set the default node attribute compare operation for an ir_op_ops.
5183 * @param code the opcode for the default operation
5184 * @param ops the operations initialized
5189 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5193 ops->node_cmp_attr = node_cmp_attr_##a; \
5224 } /* firm_set_default_node_cmp_attr */
5227 * Compare function for two nodes in the value table. Gets two
5228 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
5230 int identities_cmp(const void *elt, const void *key) {
5231 ir_node *a = (ir_node *)elt;
5232 ir_node *b = (ir_node *)key;
5235 if (a == b) return 0;
5237 if ((get_irn_op(a) != get_irn_op(b)) ||
5238 (get_irn_mode(a) != get_irn_mode(b))) return 1;
5240 /* compare if a's in and b's in are of equal length */
5241 irn_arity_a = get_irn_intra_arity(a);
5242 if (irn_arity_a != get_irn_intra_arity(b))
5245 if (get_irn_pinned(a) == op_pin_state_pinned) {
5246 /* for pinned nodes, the block inputs must be equal */
5247 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
5249 } else if (! get_opt_global_cse()) {
5250 /* for block-local CSE both nodes must be in the same MacroBlock */
5251 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
5255 /* compare a->in[0..ins] with b->in[0..ins] */
5256 for (i = 0; i < irn_arity_a; i++)
5257 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
5261 * here, we already now that the nodes are identical except their
5264 if (a->op->ops.node_cmp_attr)
5265 return a->op->ops.node_cmp_attr(a, b);
5268 } /* identities_cmp */
5271 * Calculate a hash value of a node.
5273 unsigned ir_node_hash(ir_node *node) {
5277 if (node->op == op_Const) {
5278 /* special value for const, as they only differ in their tarval. */
5279 h = HASH_PTR(node->attr.con.tv);
5280 h = 9*h + HASH_PTR(get_irn_mode(node));
5281 } else if (node->op == op_SymConst) {
5282 /* special value for const, as they only differ in their symbol. */
5283 h = HASH_PTR(node->attr.symc.sym.type_p);
5284 h = 9*h + HASH_PTR(get_irn_mode(node));
5287 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
5288 h = irn_arity = get_irn_intra_arity(node);
5290 /* consider all in nodes... except the block if not a control flow. */
5291 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
5292 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
5296 h = 9*h + HASH_PTR(get_irn_mode(node));
5298 h = 9*h + HASH_PTR(get_irn_op(node));
5302 } /* ir_node_hash */
5304 pset *new_identities(void) {
5305 return new_pset(identities_cmp, N_IR_NODES);
5306 } /* new_identities */
5308 void del_identities(pset *value_table) {
5309 del_pset(value_table);
5310 } /* del_identities */
5313 * Normalize a node by putting constants (and operands with larger
5314 * node index) on the right (operator side).
5316 * @param n The node to normalize
5318 static void normalize_node(ir_node *n) {
5319 if (is_op_commutative(get_irn_op(n))) {
5320 ir_node *l = get_binop_left(n);
5321 ir_node *r = get_binop_right(n);
5323 /* For commutative operators perform a OP b == b OP a but keep
5324 * constants on the RIGHT side. This helps greatly in some
5325 * optimizations. Moreover we use the idx number to make the form
5327 if (!operands_are_normalized(l, r)) {
5328 set_binop_left(n, r);
5329 set_binop_right(n, l);
5332 } /* normalize_node */
5335 * Update the nodes after a match in the value table. If both nodes have
5336 * the same MacroBlock but different Blocks, we must ensure that the node
5337 * with the dominating Block (the node that is near to the MacroBlock header
5338 * is stored in the table.
5339 * Because a MacroBlock has only one "non-exception" flow, we don't need
5340 * dominance info here: We known, that one block must dominate the other and
5341 * following the only block input will allow to find it.
5343 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
5344 ir_node *known_blk, *new_block, *block, *mbh;
5346 if (get_opt_global_cse()) {
5347 /* Block inputs are meaning less */
5350 known_blk = get_irn_n(known_irn, -1);
5351 new_block = get_irn_n(new_ir_node, -1);
5352 if (known_blk == new_block) {
5353 /* already in the same block */
5357 * We expect the typical case when we built the graph. In that case, the
5358 * known_irn is already the upper one, so checking this should be faster.
5361 mbh = get_Block_MacroBlock(new_block);
5363 if (block == known_blk) {
5364 /* ok, we have found it: known_block dominates new_block as expected */
5369 * We have reached the MacroBlock header NOT founding
5370 * the known_block. new_block must dominate known_block.
5373 set_irn_n(known_irn, -1, new_block);
5376 assert(get_Block_n_cfgpreds(block) == 1);
5377 block = get_Block_cfgpred_block(block, 0);
5379 } /* update_value_table */
5382 * Return the canonical node computing the same value as n.
5384 * @param value_table The value table
5385 * @param n The node to lookup
5387 * Looks up the node in a hash table.
5389 * For Const nodes this is performed in the constructor, too. Const
5390 * nodes are extremely time critical because of their frequent use in
5391 * constant string arrays.
5393 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
5396 if (!value_table) return n;
5400 o = pset_find(value_table, n, ir_node_hash(n));
5404 update_known_irn(o, n);
5411 * During construction we set the op_pin_state_pinned flag in the graph right when the
5412 * optimization is performed. The flag turning on procedure global cse could
5413 * be changed between two allocations. This way we are safe.
5415 * @param value_table The value table
5416 * @param n The node to lookup
5418 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
5421 n = identify(value_table, n);
5422 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
5423 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5425 } /* identify_cons */
5428 * Return the canonical node computing the same value as n.
5429 * Looks up the node in a hash table, enters it in the table
5430 * if it isn't there yet.
5432 * @param value_table the HashSet containing all nodes in the
5434 * @param n the node to look up
5436 * @return a node that computes the same value as n or n if no such
5437 * node could be found
5439 ir_node *identify_remember(pset *value_table, ir_node *n) {
5442 if (!value_table) return n;
5445 /* lookup or insert in hash table with given hash key. */
5446 o = pset_insert(value_table, n, ir_node_hash(n));
5449 update_known_irn(o, n);
5454 } /* identify_remember */
5456 /* Add a node to the identities value table. */
5457 void add_identities(pset *value_table, ir_node *node) {
5458 if (get_opt_cse() && is_no_Block(node))
5459 identify_remember(value_table, node);
5460 } /* add_identities */
5462 /* Visit each node in the value table of a graph. */
5463 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
5465 ir_graph *rem = current_ir_graph;
5467 current_ir_graph = irg;
5468 foreach_pset(irg->value_table, node)
5470 current_ir_graph = rem;
5471 } /* visit_all_identities */
5474 * Garbage in, garbage out. If a node has a dead input, i.e., the
5475 * Bad node is input to the node, return the Bad node.
5477 static ir_node *gigo(ir_node *node) {
5479 ir_op *op = get_irn_op(node);
5481 /* remove garbage blocks by looking at control flow that leaves the block
5482 and replacing the control flow by Bad. */
5483 if (get_irn_mode(node) == mode_X) {
5484 ir_node *block = get_nodes_block(skip_Proj(node));
5486 /* Don't optimize nodes in immature blocks. */
5487 if (!get_Block_matured(block))
5489 /* Don't optimize End, may have Bads. */
5490 if (op == op_End) return node;
5492 if (is_Block(block)) {
5493 if (is_Block_dead(block)) {
5494 /* control flow from dead block is dead */
5498 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
5499 if (!is_Bad(get_irn_n(block, i)))
5503 ir_graph *irg = get_irn_irg(block);
5504 /* the start block is never dead */
5505 if (block != get_irg_start_block(irg)
5506 && block != get_irg_end_block(irg)) {
5508 * Do NOT kill control flow without setting
5509 * the block to dead of bad things can happen:
5510 * We get a Block that is not reachable be irg_block_walk()
5511 * but can be found by irg_walk()!
5513 set_Block_dead(block);
5520 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
5521 blocks predecessors is dead. */
5522 if (op != op_Block && op != op_Phi && op != op_Tuple) {
5523 irn_arity = get_irn_arity(node);
5526 * Beware: we can only read the block of a non-floating node.
5528 if (is_irn_pinned_in_irg(node) &&
5529 is_Block_dead(get_nodes_block(skip_Proj(node))))
5532 for (i = 0; i < irn_arity; i++) {
5533 ir_node *pred = get_irn_n(node, i);
5538 /* Propagating Unknowns here seems to be a bad idea, because
5539 sometimes we need a node as a input and did not want that
5541 However, it might be useful to move this into a later phase
5542 (if you think that optimizing such code is useful). */
5543 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
5544 return new_Unknown(get_irn_mode(node));
5549 /* With this code we violate the agreement that local_optimize
5550 only leaves Bads in Block, Phi and Tuple nodes. */
5551 /* If Block has only Bads as predecessors it's garbage. */
5552 /* If Phi has only Bads as predecessors it's garbage. */
5553 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
5554 irn_arity = get_irn_arity(node);
5555 for (i = 0; i < irn_arity; i++) {
5556 if (!is_Bad(get_irn_n(node, i))) break;
5558 if (i == irn_arity) node = new_Bad();
5565 * These optimizations deallocate nodes from the obstack.
5566 * It can only be called if it is guaranteed that no other nodes
5567 * reference this one, i.e., right after construction of a node.
5569 * @param n The node to optimize
5571 * current_ir_graph must be set to the graph of the node!
5573 ir_node *optimize_node(ir_node *n) {
5576 ir_opcode iro = get_irn_opcode(n);
5578 /* Always optimize Phi nodes: part of the construction. */
5579 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
5581 /* constant expression evaluation / constant folding */
5582 if (get_opt_constant_folding()) {
5583 /* neither constants nor Tuple values can be evaluated */
5584 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
5585 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5586 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5587 /* try to evaluate */
5588 tv = computed_value(n);
5589 if (tv != tarval_bad) {
5591 ir_type *old_tp = get_irn_type(n);
5592 int i, arity = get_irn_arity(n);
5596 * Try to recover the type of the new expression.
5598 for (i = 0; i < arity && !old_tp; ++i)
5599 old_tp = get_irn_type(get_irn_n(n, i));
5602 * we MUST copy the node here temporary, because it's still needed
5603 * for DBG_OPT_CSTEVAL
5605 node_size = offsetof(ir_node, attr) + n->op->attr_size;
5606 oldn = alloca(node_size);
5608 memcpy(oldn, n, node_size);
5609 CLONE_ARR_A(ir_node *, oldn->in, n->in);
5611 /* ARG, copy the in array, we need it for statistics */
5612 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
5614 /* note the inplace edges module */
5615 edges_node_deleted(n, current_ir_graph);
5617 /* evaluation was successful -- replace the node. */
5618 irg_kill_node(current_ir_graph, n);
5619 nw = new_Const(get_tarval_mode(tv), tv);
5621 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5622 set_Const_type(nw, old_tp);
5623 DBG_OPT_CSTEVAL(oldn, nw);
5624 tarval_enable_fp_ops(old_fp_mode);
5627 tarval_enable_fp_ops(old_fp_mode);
5631 /* remove unnecessary nodes */
5632 if (get_opt_constant_folding() ||
5633 (iro == iro_Phi) || /* always optimize these nodes. */
5635 (iro == iro_Proj) ||
5636 (iro == iro_Block) ) /* Flags tested local. */
5637 n = equivalent_node(n);
5639 /* Common Subexpression Elimination.
5641 * Checks whether n is already available.
5642 * The block input is used to distinguish different subexpressions. Right
5643 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
5644 * subexpressions within a block.
5647 n = identify_cons(current_ir_graph->value_table, n);
5650 edges_node_deleted(oldn, current_ir_graph);
5652 /* We found an existing, better node, so we can deallocate the old node. */
5653 irg_kill_node(current_ir_graph, oldn);
5657 /* Some more constant expression evaluation that does not allow to
5659 iro = get_irn_opcode(n);
5660 if (get_opt_constant_folding() ||
5661 (iro == iro_Cond) ||
5662 (iro == iro_Proj)) /* Flags tested local. */
5663 n = transform_node(n);
5665 /* Remove nodes with dead (Bad) input.
5666 Run always for transformation induced Bads. */
5669 /* Now we have a legal, useful node. Enter it in hash table for CSE */
5670 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
5671 n = identify_remember(current_ir_graph->value_table, n);
5675 } /* optimize_node */
5679 * These optimizations never deallocate nodes (in place). This can cause dead
5680 * nodes lying on the obstack. Remove these by a dead node elimination,
5681 * i.e., a copying garbage collection.
5683 ir_node *optimize_in_place_2(ir_node *n) {
5686 ir_opcode iro = get_irn_opcode(n);
5688 if (!get_opt_optimize() && !is_Phi(n)) return n;
5690 /* constant expression evaluation / constant folding */
5691 if (get_opt_constant_folding()) {
5692 /* neither constants nor Tuple values can be evaluated */
5693 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
5694 unsigned fp_model = get_irg_fp_model(current_ir_graph);
5695 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
5696 /* try to evaluate */
5697 tv = computed_value(n);
5698 if (tv != tarval_bad) {
5699 /* evaluation was successful -- replace the node. */
5700 ir_type *old_tp = get_irn_type(n);
5701 int i, arity = get_irn_arity(n);
5704 * Try to recover the type of the new expression.
5706 for (i = 0; i < arity && !old_tp; ++i)
5707 old_tp = get_irn_type(get_irn_n(n, i));
5709 n = new_Const(get_tarval_mode(tv), tv);
5711 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
5712 set_Const_type(n, old_tp);
5714 DBG_OPT_CSTEVAL(oldn, n);
5715 tarval_enable_fp_ops(old_fp_mode);
5718 tarval_enable_fp_ops(old_fp_mode);
5722 /* remove unnecessary nodes */
5723 if (get_opt_constant_folding() ||
5724 (iro == iro_Phi) || /* always optimize these nodes. */
5725 (iro == iro_Id) || /* ... */
5726 (iro == iro_Proj) || /* ... */
5727 (iro == iro_Block) ) /* Flags tested local. */
5728 n = equivalent_node(n);
5730 /** common subexpression elimination **/
5731 /* Checks whether n is already available. */
5732 /* The block input is used to distinguish different subexpressions. Right
5733 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
5734 subexpressions within a block. */
5735 if (get_opt_cse()) {
5736 n = identify(current_ir_graph->value_table, n);
5739 /* Some more constant expression evaluation. */
5740 iro = get_irn_opcode(n);
5741 if (get_opt_constant_folding() ||
5742 (iro == iro_Cond) ||
5743 (iro == iro_Proj)) /* Flags tested local. */
5744 n = transform_node(n);
5746 /* Remove nodes with dead (Bad) input.
5747 Run always for transformation induced Bads. */
5750 /* Now we can verify the node, as it has no dead inputs any more. */
5753 /* Now we have a legal, useful node. Enter it in hash table for cse.
5754 Blocks should be unique anyways. (Except the successor of start:
5755 is cse with the start block!) */
5756 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
5757 n = identify_remember(current_ir_graph->value_table, n);
5760 } /* optimize_in_place_2 */
5763 * Wrapper for external use, set proper status bits after optimization.
5765 ir_node *optimize_in_place(ir_node *n) {
5766 /* Handle graph state */
5767 assert(get_irg_phase_state(current_ir_graph) != phase_building);
5769 if (get_opt_global_cse())
5770 set_irg_pinned(current_ir_graph, op_pin_state_floats);
5771 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
5772 set_irg_outs_inconsistent(current_ir_graph);
5774 /* FIXME: Maybe we could also test whether optimizing the node can
5775 change the control graph. */
5776 set_irg_doms_inconsistent(current_ir_graph);
5777 return optimize_in_place_2(n);
5778 } /* optimize_in_place */
5781 * Sets the default operation for an ir_ops.
5783 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
5784 ops = firm_set_default_computed_value(code, ops);
5785 ops = firm_set_default_equivalent_node(code, ops);
5786 ops = firm_set_default_transform_node(code, ops);
5787 ops = firm_set_default_node_cmp_attr(code, ops);
5788 ops = firm_set_default_get_type(code, ops);
5789 ops = firm_set_default_get_type_attr(code, ops);
5790 ops = firm_set_default_get_entity_attr(code, ops);
5793 } /* firm_set_default_operations */