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
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
46 #include "opt_confirms.h"
47 #include "opt_polymorphy.h"
52 /* Make types visible to allow most efficient access */
56 * Returns the tarval of a Const node or tarval_bad for all other nodes.
58 static tarval *default_value_of(const ir_node *n) {
60 return get_Const_tarval(n); /* might return tarval_bad */
65 value_of_func value_of_ptr = default_value_of;
67 /* * Set a new value_of function. */
68 void set_value_of_func(value_of_func func) {
72 value_of_ptr = default_value_of;
76 * Return the value of a Constant.
78 static tarval *computed_value_Const(const ir_node *n) {
79 return get_Const_tarval(n);
80 } /* computed_value_Const */
83 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
85 static tarval *computed_value_SymConst(const ir_node *n) {
89 switch (get_SymConst_kind(n)) {
90 case symconst_type_size:
91 type = get_SymConst_type(n);
92 if (get_type_state(type) == layout_fixed)
93 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
95 case symconst_type_align:
96 type = get_SymConst_type(n);
97 if (get_type_state(type) == layout_fixed)
98 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
100 case symconst_ofs_ent:
101 ent = get_SymConst_entity(n);
102 type = get_entity_owner(ent);
103 if (get_type_state(type) == layout_fixed)
104 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
110 } /* computed_value_SymConst */
113 * Return the value of an Add.
115 static tarval *computed_value_Add(const ir_node *n) {
116 ir_node *a = get_Add_left(n);
117 ir_node *b = get_Add_right(n);
119 tarval *ta = value_of(a);
120 tarval *tb = value_of(b);
122 if ((ta != tarval_bad) && (tb != tarval_bad))
123 return tarval_add(ta, tb);
126 } /* computed_value_Add */
129 * Return the value of a Sub.
130 * Special case: a - a
132 static tarval *computed_value_Sub(const ir_node *n) {
133 ir_mode *mode = get_irn_mode(n);
134 ir_node *a = get_Sub_left(n);
135 ir_node *b = get_Sub_right(n);
140 if (! mode_is_float(mode)) {
143 return get_mode_null(mode);
149 if ((ta != tarval_bad) && (tb != tarval_bad))
150 return tarval_sub(ta, tb, mode);
153 } /* computed_value_Sub */
156 * Return the value of a Carry.
157 * Special : a op 0, 0 op b
159 static tarval *computed_value_Carry(const ir_node *n) {
160 ir_node *a = get_binop_left(n);
161 ir_node *b = get_binop_right(n);
162 ir_mode *m = get_irn_mode(n);
164 tarval *ta = value_of(a);
165 tarval *tb = value_of(b);
167 if ((ta != tarval_bad) && (tb != tarval_bad)) {
169 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
171 if (tarval_is_null(ta) || tarval_is_null(tb))
172 return get_mode_null(m);
175 } /* computed_value_Carry */
178 * Return the value of a Borrow.
181 static tarval *computed_value_Borrow(const ir_node *n) {
182 ir_node *a = get_binop_left(n);
183 ir_node *b = get_binop_right(n);
184 ir_mode *m = get_irn_mode(n);
186 tarval *ta = value_of(a);
187 tarval *tb = value_of(b);
189 if ((ta != tarval_bad) && (tb != tarval_bad)) {
190 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
191 } else if (tarval_is_null(ta)) {
192 return get_mode_null(m);
195 } /* computed_value_Borrow */
198 * Return the value of an unary Minus.
200 static tarval *computed_value_Minus(const ir_node *n) {
201 ir_node *a = get_Minus_op(n);
202 tarval *ta = value_of(a);
204 if (ta != tarval_bad)
205 return tarval_neg(ta);
208 } /* computed_value_Minus */
211 * Return the value of a Mul.
213 static tarval *computed_value_Mul(const ir_node *n) {
214 ir_node *a = get_Mul_left(n);
215 ir_node *b = get_Mul_right(n);
218 tarval *ta = value_of(a);
219 tarval *tb = value_of(b);
221 mode = get_irn_mode(n);
222 if (mode != get_irn_mode(a)) {
223 /* n * n = 2n bit multiplication */
224 ta = tarval_convert_to(ta, mode);
225 tb = tarval_convert_to(tb, mode);
228 if (ta != tarval_bad && tb != tarval_bad) {
229 return tarval_mul(ta, tb);
231 /* a * 0 != 0 if a == NaN or a == Inf */
232 if (!mode_is_float(mode)) {
233 /* a*0 = 0 or 0*b = 0 */
234 if (ta == get_mode_null(mode))
236 if (tb == get_mode_null(mode))
241 } /* computed_value_Mul */
244 * Return the value of an Abs.
246 static tarval *computed_value_Abs(const ir_node *n) {
247 ir_node *a = get_Abs_op(n);
248 tarval *ta = value_of(a);
250 if (ta != tarval_bad)
251 return tarval_abs(ta);
254 } /* computed_value_Abs */
257 * Return the value of an And.
258 * Special case: a & 0, 0 & b
260 static tarval *computed_value_And(const ir_node *n) {
261 ir_node *a = get_And_left(n);
262 ir_node *b = get_And_right(n);
264 tarval *ta = value_of(a);
265 tarval *tb = value_of(b);
267 if ((ta != tarval_bad) && (tb != tarval_bad)) {
268 return tarval_and (ta, tb);
270 if (tarval_is_null(ta)) return ta;
271 if (tarval_is_null(tb)) return tb;
274 } /* computed_value_And */
277 * Return the value of an Or.
278 * Special case: a | 1...1, 1...1 | b
280 static tarval *computed_value_Or(const ir_node *n) {
281 ir_node *a = get_Or_left(n);
282 ir_node *b = get_Or_right(n);
284 tarval *ta = value_of(a);
285 tarval *tb = value_of(b);
287 if ((ta != tarval_bad) && (tb != tarval_bad)) {
288 return tarval_or (ta, tb);
290 if (tarval_is_all_one(ta)) return ta;
291 if (tarval_is_all_one(tb)) return tb;
294 } /* computed_value_Or */
297 * Return the value of an Eor.
299 static tarval *computed_value_Eor(const ir_node *n) {
300 ir_node *a = get_Eor_left(n);
301 ir_node *b = get_Eor_right(n);
306 return get_mode_null(get_irn_mode(n));
311 if ((ta != tarval_bad) && (tb != tarval_bad)) {
312 return tarval_eor (ta, tb);
315 } /* computed_value_Eor */
318 * Return the value of a Not.
320 static tarval *computed_value_Not(const ir_node *n) {
321 ir_node *a = get_Not_op(n);
322 tarval *ta = value_of(a);
324 if (ta != tarval_bad)
325 return tarval_not(ta);
328 } /* computed_value_Not */
331 * Return the value of a Shl.
333 static tarval *computed_value_Shl(const ir_node *n) {
334 ir_node *a = get_Shl_left(n);
335 ir_node *b = get_Shl_right(n);
337 tarval *ta = value_of(a);
338 tarval *tb = value_of(b);
340 if ((ta != tarval_bad) && (tb != tarval_bad)) {
341 return tarval_shl (ta, tb);
344 } /* computed_value_Shl */
347 * Return the value of a Shr.
349 static tarval *computed_value_Shr(const ir_node *n) {
350 ir_node *a = get_Shr_left(n);
351 ir_node *b = get_Shr_right(n);
353 tarval *ta = value_of(a);
354 tarval *tb = value_of(b);
356 if ((ta != tarval_bad) && (tb != tarval_bad)) {
357 return tarval_shr (ta, tb);
360 } /* computed_value_Shr */
363 * Return the value of a Shrs.
365 static tarval *computed_value_Shrs(const ir_node *n) {
366 ir_node *a = get_Shrs_left(n);
367 ir_node *b = get_Shrs_right(n);
369 tarval *ta = value_of(a);
370 tarval *tb = value_of(b);
372 if ((ta != tarval_bad) && (tb != tarval_bad)) {
373 return tarval_shrs (ta, tb);
376 } /* computed_value_Shrs */
379 * Return the value of a Rotl.
381 static tarval *computed_value_Rotl(const ir_node *n) {
382 ir_node *a = get_Rotl_left(n);
383 ir_node *b = get_Rotl_right(n);
385 tarval *ta = value_of(a);
386 tarval *tb = value_of(b);
388 if ((ta != tarval_bad) && (tb != tarval_bad)) {
389 return tarval_rotl(ta, tb);
392 } /* computed_value_Rotl */
395 * Return the value of a Conv.
397 static tarval *computed_value_Conv(const ir_node *n) {
398 ir_node *a = get_Conv_op(n);
399 tarval *ta = value_of(a);
401 if (ta != tarval_bad)
402 return tarval_convert_to(ta, get_irn_mode(n));
405 } /* computed_value_Conv */
408 * Calculate the value of a Mux: can be evaluated, if the
409 * sel and the right input are known.
411 static tarval *computed_value_Mux(const ir_node *n) {
412 ir_node *sel = get_Mux_sel(n);
413 tarval *ts = value_of(sel);
415 if (ts == get_tarval_b_true()) {
416 ir_node *v = get_Mux_true(n);
419 else if (ts == get_tarval_b_false()) {
420 ir_node *v = get_Mux_false(n);
424 } /* computed_value_Mux */
427 * Calculate the value of a Confirm: can be evaluated,
428 * if it has the form Confirm(x, '=', Const).
430 static tarval *computed_value_Confirm(const ir_node *n) {
432 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
433 * Do NOT optimize them away (CondEval wants them), so wait until
434 * remove_confirm is activated.
436 if (get_opt_remove_confirm()) {
437 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
438 tarval *tv = value_of(get_Confirm_bound(n));
439 if (tv != tarval_bad)
443 return value_of(get_Confirm_value(n));
444 } /* computed_value_Confirm */
447 * Return the value of a Proj(Cmp).
449 * This performs a first step of unreachable code elimination.
450 * Proj can not be computed, but folding a Cmp above the Proj here is
451 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
453 * There are several case where we can evaluate a Cmp node, see later.
455 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
456 ir_node *a = get_Proj_pred(n);
457 ir_node *aa = get_Cmp_left(a);
458 ir_node *ab = get_Cmp_right(a);
459 long proj_nr = get_Proj_proj(n);
462 * BEWARE: a == a is NOT always True for floating Point values, as
463 * NaN != NaN is defined, so we must check this here.
466 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
469 /* This is a trick with the bits used for encoding the Cmp
470 Proj numbers, the following statement is not the same:
471 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
472 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
475 tarval *taa = value_of(aa);
476 tarval *tab = value_of(ab);
477 ir_mode *mode = get_irn_mode(aa);
480 * The predecessors of Cmp are target values. We can evaluate
483 if ((taa != tarval_bad) && (tab != tarval_bad)) {
484 /* strange checks... */
485 pn_Cmp flags = tarval_cmp(taa, tab);
486 if (flags != pn_Cmp_False) {
487 return new_tarval_from_long (proj_nr & flags, mode_b);
490 /* for integer values, we can check against MIN/MAX */
491 else if (mode_is_int(mode)) {
492 /* MIN <=/> x. This results in true/false. */
493 if (taa == get_mode_min(mode)) {
494 /* a compare with the MIN value */
495 if (proj_nr == pn_Cmp_Le)
496 return get_tarval_b_true();
497 else if (proj_nr == pn_Cmp_Gt)
498 return get_tarval_b_false();
500 /* x >=/< MIN. This results in true/false. */
502 if (tab == get_mode_min(mode)) {
503 /* a compare with the MIN value */
504 if (proj_nr == pn_Cmp_Ge)
505 return get_tarval_b_true();
506 else if (proj_nr == pn_Cmp_Lt)
507 return get_tarval_b_false();
509 /* MAX >=/< x. This results in true/false. */
510 else if (taa == get_mode_max(mode)) {
511 if (proj_nr == pn_Cmp_Ge)
512 return get_tarval_b_true();
513 else if (proj_nr == pn_Cmp_Lt)
514 return get_tarval_b_false();
516 /* x <=/> MAX. This results in true/false. */
517 else if (tab == get_mode_max(mode)) {
518 if (proj_nr == pn_Cmp_Le)
519 return get_tarval_b_true();
520 else if (proj_nr == pn_Cmp_Gt)
521 return get_tarval_b_false();
525 * The predecessors are Allocs or (void*)(0) constants. Allocs never
526 * return NULL, they raise an exception. Therefore we can predict
530 ir_node *aaa = skip_Proj(aa);
531 ir_node *aba = skip_Proj(ab);
533 if ( ( (/* aa is ProjP and aaa is Alloc */
535 && mode_is_reference(get_irn_mode(aa))
537 && ( (/* ab is NULL */
538 mode_is_reference(get_irn_mode(ab))
539 && tarval_is_null(tab))
540 || (/* ab is other Alloc */
542 && mode_is_reference(get_irn_mode(ab))
545 || (/* aa is NULL and aba is Alloc */
546 mode_is_reference(get_irn_mode(aa))
547 && tarval_is_null(taa)
549 && mode_is_reference(get_irn_mode(ab))
552 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
555 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
556 } /* computed_value_Proj_Cmp */
559 * Return the value of a floating point Quot.
561 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
562 tarval *ta = value_of(a);
563 tarval *tb = value_of(b);
565 /* cannot optimize 0 / b = 0 because of NaN */
566 if (ta != tarval_bad && tb != tarval_bad)
567 return tarval_quo(ta, tb);
569 } /* do_computed_value_Quot */
572 * Calculate the value of an integer Div of two nodes.
573 * Special case: 0 / b
575 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
576 tarval *ta = value_of(a);
578 const ir_node *dummy;
580 /* Compute c1 / c2 or 0 / a, a != 0 */
581 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
582 return ta; /* 0 / b == 0 */
584 if (ta != tarval_bad && tb != tarval_bad)
585 return tarval_div(ta, tb);
587 } /* do_computed_value_Div */
590 * Calculate the value of an integer Mod of two nodes.
591 * Special case: a % 1
593 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
594 tarval *ta = value_of(a);
595 tarval *tb = value_of(b);
597 /* Compute a % 1 or c1 % c2 */
598 if (tarval_is_one(tb))
599 return get_mode_null(get_irn_mode(a));
600 if (ta != tarval_bad && tb != tarval_bad)
601 return tarval_mod(ta, tb);
603 } /* do_computed_value_Mod */
606 * Return the value of a Proj(DivMod).
608 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
609 long proj_nr = get_Proj_proj(n);
611 /* compute either the Div or the Mod part */
612 if (proj_nr == pn_DivMod_res_div) {
613 const ir_node *a = get_Proj_pred(n);
614 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
615 } else if (proj_nr == pn_DivMod_res_mod) {
616 const ir_node *a = get_Proj_pred(n);
617 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
620 } /* computed_value_Proj_DivMod */
623 * Return the value of a Proj(Div).
625 static tarval *computed_value_Proj_Div(const ir_node *n) {
626 long proj_nr = get_Proj_proj(n);
628 if (proj_nr == pn_Div_res) {
629 const ir_node *a = get_Proj_pred(n);
630 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
633 } /* computed_value_Proj_Div */
636 * Return the value of a Proj(Mod).
638 static tarval *computed_value_Proj_Mod(const ir_node *n) {
639 long proj_nr = get_Proj_proj(n);
641 if (proj_nr == pn_Mod_res) {
642 const ir_node *a = get_Proj_pred(n);
643 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
646 } /* computed_value_Proj_Mod */
649 * Return the value of a Proj(Quot).
651 static tarval *computed_value_Proj_Quot(const ir_node *n) {
652 long proj_nr = get_Proj_proj(n);
654 if (proj_nr == pn_Quot_res) {
655 const ir_node *a = get_Proj_pred(n);
656 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
659 } /* computed_value_Proj_Quot */
662 * Return the value of a Proj.
664 static tarval *computed_value_Proj(const ir_node *proj) {
665 ir_node *n = get_Proj_pred(proj);
667 if (n->op->ops.computed_value_Proj != NULL)
668 return n->op->ops.computed_value_Proj(proj);
670 } /* computed_value_Proj */
673 * If the parameter n can be computed, return its value, else tarval_bad.
674 * Performs constant folding.
676 * @param n The node this should be evaluated
678 tarval *computed_value(const ir_node *n) {
679 if (n->op->ops.computed_value)
680 return n->op->ops.computed_value(n);
682 } /* computed_value */
685 * Set the default computed_value evaluator in an ir_op_ops.
687 * @param code the opcode for the default operation
688 * @param ops the operations initialized
693 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
697 ops->computed_value = computed_value_##a; \
699 #define CASE_PROJ(a) \
701 ops->computed_value_Proj = computed_value_Proj_##a; \
738 } /* firm_set_default_computed_value */
741 * Returns a equivalent block for another block.
742 * If the block has only one predecessor, this is
743 * the equivalent one. If the only predecessor of a block is
744 * the block itself, this is a dead block.
746 * If both predecessors of a block are the branches of a binary
747 * Cond, the equivalent block is Cond's block.
749 * If all predecessors of a block are bad or lies in a dead
750 * block, the current block is dead as well.
752 * Note, that blocks are NEVER turned into Bad's, instead
753 * the dead_block flag is set. So, never test for is_Bad(block),
754 * always use is_dead_Block(block).
756 static ir_node *equivalent_node_Block(ir_node *n)
761 /* don't optimize dead blocks */
762 if (is_Block_dead(n))
765 n_preds = get_Block_n_cfgpreds(n);
767 /* The Block constructor does not call optimize, but mature_immBlock()
768 calls the optimization. */
769 assert(get_Block_matured(n));
771 /* Straightening: a single entry Block following a single exit Block
772 can be merged, if it is not the Start block. */
773 /* !!! Beware, all Phi-nodes of n must have been optimized away.
774 This should be true, as the block is matured before optimize is called.
775 But what about Phi-cycles with the Phi0/Id that could not be resolved?
776 Remaining Phi nodes are just Ids. */
778 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
781 ir_node *predblock = get_nodes_block(pred);
782 if (predblock == oldn) {
783 /* Jmp jumps into the block it is in -- deal self cycle. */
784 n = set_Block_dead(n);
785 DBG_OPT_DEAD_BLOCK(oldn, n);
786 } else if (get_opt_control_flow_straightening()) {
788 DBG_OPT_STG(oldn, n);
790 } else if (is_Cond(pred)) {
791 ir_node *predblock = get_nodes_block(pred);
792 if (predblock == oldn) {
793 /* Jmp jumps into the block it is in -- deal self cycle. */
794 n = set_Block_dead(n);
795 DBG_OPT_DEAD_BLOCK(oldn, n);
798 } else if ((n_preds == 2) &&
799 (get_opt_control_flow_weak_simplification())) {
800 /* Test whether Cond jumps twice to this block
801 * The more general case which more than 2 predecessors is handles
802 * in optimize_cf(), we handle only this special case for speed here.
804 ir_node *a = get_Block_cfgpred(n, 0);
805 ir_node *b = get_Block_cfgpred(n, 1);
807 if (is_Proj(a) && is_Proj(b)) {
808 ir_node *cond = get_Proj_pred(a);
810 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
811 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
812 /* Also a single entry Block following a single exit Block. Phis have
813 twice the same operand and will be optimized away. */
814 n = get_nodes_block(cond);
815 DBG_OPT_IFSIM1(oldn, a, b, n);
818 } else if (get_opt_unreachable_code() &&
819 (n != get_irg_start_block(current_ir_graph)) &&
820 (n != get_irg_end_block(current_ir_graph)) ) {
823 /* If all inputs are dead, this block is dead too, except if it is
824 the start or end block. This is one step of unreachable code
826 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
827 ir_node *pred = get_Block_cfgpred(n, i);
830 if (is_Bad(pred)) continue;
831 pred_blk = get_nodes_block(skip_Proj(pred));
833 if (is_Block_dead(pred_blk)) continue;
836 /* really found a living input */
841 n = set_Block_dead(n);
842 DBG_OPT_DEAD_BLOCK(oldn, n);
847 } /* equivalent_node_Block */
850 * Returns a equivalent node for a Jmp, a Bad :-)
851 * Of course this only happens if the Block of the Jmp is dead.
853 static ir_node *equivalent_node_Jmp(ir_node *n) {
856 /* unreachable code elimination */
857 if (is_Block_dead(get_nodes_block(n))) {
858 n = get_irg_bad(current_ir_graph);
859 DBG_OPT_DEAD_BLOCK(oldn, n);
862 } /* equivalent_node_Jmp */
864 /** Raise is handled in the same way as Jmp. */
865 #define equivalent_node_Raise equivalent_node_Jmp
868 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
869 See transform_node_Proj_Cond(). */
872 * Optimize operations that are commutative and have neutral 0,
873 * so a op 0 = 0 op a = a.
875 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
878 ir_node *a = get_binop_left(n);
879 ir_node *b = get_binop_right(n);
884 /* After running compute_node there is only one constant predecessor.
885 Find this predecessors value and remember the other node: */
886 if ((tv = value_of(a)) != tarval_bad) {
888 } else if ((tv = value_of(b)) != tarval_bad) {
893 /* If this predecessors constant value is zero, the operation is
894 * unnecessary. Remove it.
896 * Beware: If n is a Add, the mode of on and n might be different
897 * which happens in this rare construction: NULL + 3.
898 * Then, a Conv would be needed which we cannot include here.
900 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
903 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
907 } /* equivalent_node_neutral_zero */
910 * Eor is commutative and has neutral 0.
912 static ir_node *equivalent_node_Eor(ir_node *n) {
917 n = equivalent_node_neutral_zero(n);
918 if (n != oldn) return n;
921 b = get_Eor_right(n);
924 ir_node *aa = get_Eor_left(a);
925 ir_node *ab = get_Eor_right(a);
928 /* (a ^ b) ^ a -> b */
930 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
932 } else if (ab == b) {
933 /* (a ^ b) ^ b -> a */
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
940 ir_node *ba = get_Eor_left(b);
941 ir_node *bb = get_Eor_right(b);
944 /* a ^ (a ^ b) -> b */
946 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
948 } else if (bb == a) {
949 /* a ^ (b ^ a) -> b */
951 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
959 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
961 * The second one looks strange, but this construct
962 * is used heavily in the LCC sources :-).
964 * Beware: The Mode of an Add may be different than the mode of its
965 * predecessors, so we could not return a predecessors in all cases.
967 static ir_node *equivalent_node_Add(ir_node *n) {
969 ir_node *left, *right;
970 ir_mode *mode = get_irn_mode(n);
972 n = equivalent_node_neutral_zero(n);
976 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
977 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
980 left = get_Add_left(n);
981 right = get_Add_right(n);
984 if (get_Sub_right(left) == right) {
987 n = get_Sub_left(left);
988 if (mode == get_irn_mode(n)) {
989 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
995 if (get_Sub_right(right) == left) {
998 n = get_Sub_left(right);
999 if (mode == get_irn_mode(n)) {
1000 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1006 } /* equivalent_node_Add */
1009 * optimize operations that are not commutative but have neutral 0 on left,
1012 static ir_node *equivalent_node_left_zero(ir_node *n) {
1015 ir_node *a = get_binop_left(n);
1016 ir_node *b = get_binop_right(n);
1017 tarval *tb = value_of(b);
1019 if (tarval_is_null(tb)) {
1022 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1025 } /* equivalent_node_left_zero */
1027 #define equivalent_node_Shl equivalent_node_left_zero
1028 #define equivalent_node_Shr equivalent_node_left_zero
1029 #define equivalent_node_Shrs equivalent_node_left_zero
1030 #define equivalent_node_Rotl equivalent_node_left_zero
1033 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1035 * The second one looks strange, but this construct
1036 * is used heavily in the LCC sources :-).
1038 * Beware: The Mode of a Sub may be different than the mode of its
1039 * predecessors, so we could not return a predecessors in all cases.
1041 static ir_node *equivalent_node_Sub(ir_node *n) {
1044 ir_mode *mode = get_irn_mode(n);
1047 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1048 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1051 b = get_Sub_right(n);
1054 /* Beware: modes might be different */
1055 if (tarval_is_null(tb)) {
1056 ir_node *a = get_Sub_left(n);
1057 if (mode == get_irn_mode(a)) {
1060 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1064 } /* equivalent_node_Sub */
1068 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1071 * -(-a) == a, but might overflow two times.
1072 * We handle it anyway here but the better way would be a
1073 * flag. This would be needed for Pascal for instance.
1075 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1077 ir_node *pred = get_unop_op(n);
1079 /* optimize symmetric unop */
1080 if (get_irn_op(pred) == get_irn_op(n)) {
1081 n = get_unop_op(pred);
1082 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1085 } /* equivalent_node_idempotent_unop */
1087 /** Optimize Not(Not(x)) == x. */
1088 #define equivalent_node_Not equivalent_node_idempotent_unop
1090 /** -(-x) == x ??? Is this possible or can --x raise an
1091 out of bounds exception if min =! max? */
1092 #define equivalent_node_Minus equivalent_node_idempotent_unop
1095 * Optimize a * 1 = 1 * a = a.
1097 static ir_node *equivalent_node_Mul(ir_node *n) {
1099 ir_node *a = get_Mul_left(n);
1101 /* we can handle here only the n * n = n bit cases */
1102 if (get_irn_mode(n) == get_irn_mode(a)) {
1103 ir_node *b = get_Mul_right(n);
1107 * Mul is commutative and has again an other neutral element.
1108 * Constants are place right, so check this case first.
1111 if (tarval_is_one(tv)) {
1113 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1116 if (tarval_is_one(tv)) {
1118 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1123 } /* equivalent_node_Mul */
1126 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1128 static ir_node *equivalent_node_Or(ir_node *n) {
1131 ir_node *a = get_Or_left(n);
1132 ir_node *b = get_Or_right(n);
1136 n = a; /* Or has it's own neutral element */
1137 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1140 /* constants are cormalized to right, check this site first */
1142 if (tarval_is_null(tv)) {
1144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1148 if (tarval_is_null(tv)) {
1150 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1155 } /* equivalent_node_Or */
1158 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1160 static ir_node *equivalent_node_And(ir_node *n) {
1163 ir_node *a = get_And_left(n);
1164 ir_node *b = get_And_right(n);
1168 n = a; /* And has it's own neutral element */
1169 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1172 /* constants are cormalized to right, check this site first */
1174 if (tarval_is_all_one(tv)) {
1176 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1180 if (tarval_is_all_one(tv)) {
1182 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1186 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1189 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1194 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1197 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1202 } /* equivalent_node_And */
1205 * Try to remove useless Conv's:
1207 static ir_node *equivalent_node_Conv(ir_node *n) {
1209 ir_node *a = get_Conv_op(n);
1211 ir_mode *n_mode = get_irn_mode(n);
1212 ir_mode *a_mode = get_irn_mode(a);
1215 if (n_mode == a_mode) { /* No Conv necessary */
1216 if (get_Conv_strict(n)) {
1220 /* neither Minus nor Abs nor Confirm change the precision,
1221 so we can "look-through" */
1225 p = get_Minus_op(p);
1226 } else if (is_Abs(p)) {
1229 } else if (is_Confirm(p)) {
1231 p = get_Confirm_value(p);
1238 /* special case: the predecessor is also a Conv */
1239 if (! get_Conv_strict(p)) {
1240 /* first one is not strict, kick it */
1242 a_mode = get_irn_mode(a);
1246 /* else both are strict conv, second is superfluous */
1249 ir_node *pred = get_Proj_pred(p);
1250 if (is_Load(pred)) {
1251 /* loads always return with the exact precision of n_mode */
1252 assert(get_Load_mode(pred) == n_mode);
1256 /* leave strict floating point Conv's */
1261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1262 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1263 ir_node *b = get_Conv_op(a);
1264 ir_mode *b_mode = get_irn_mode(b);
1266 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1267 /* both are strict conv */
1268 if (smaller_mode(a_mode, n_mode)) {
1269 /* both are strict, but the first is smaller, so
1270 the second cannot remove more precision, remove the
1272 set_Conv_strict(n, 0);
1275 if (n_mode == b_mode) {
1276 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1277 if (n_mode == mode_b) {
1278 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1279 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1281 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1282 if (values_in_mode(b_mode, a_mode)) {
1283 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1284 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1289 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1290 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1291 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1292 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1294 if (float_mantissa >= int_mantissa) {
1296 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1301 if (smaller_mode(b_mode, a_mode)) {
1302 if (get_Conv_strict(n))
1303 set_Conv_strict(b, 1);
1304 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1305 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1312 } /* equivalent_node_Conv */
1315 * A Cast may be removed if the type of the previous node
1316 * is already the type of the Cast.
1318 static ir_node *equivalent_node_Cast(ir_node *n) {
1320 ir_node *pred = get_Cast_op(n);
1322 if (get_irn_type(pred) == get_Cast_type(n)) {
1324 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1327 } /* equivalent_node_Cast */
1330 * - fold Phi-nodes, iff they have only one predecessor except
1333 static ir_node *equivalent_node_Phi(ir_node *n) {
1338 ir_node *first_val = NULL; /* to shutup gcc */
1340 if (!get_opt_normalize()) return n;
1342 n_preds = get_Phi_n_preds(n);
1344 block = get_nodes_block(n);
1345 if (is_Block_dead(block)) /* Control dead */
1346 return get_irg_bad(current_ir_graph);
1348 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1350 /* Find first non-self-referencing input */
1351 for (i = 0; i < n_preds; ++i) {
1352 first_val = get_Phi_pred(n, i);
1353 if ( (first_val != n) /* not self pointer */
1355 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1356 * predecessors. Then, Phi nodes in dead code might be removed, causing
1357 * nodes pointing to themself (Add's for instance).
1358 * This is really bad and causes endless recursions in several
1359 * code pathes, so we do NOT optimize such a code.
1360 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1361 * (and bad Phi predecessors), so live code is optimized later.
1363 && (! is_Bad(get_Block_cfgpred(block, i)))
1365 ) { /* value not dead */
1366 break; /* then found first value. */
1371 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1372 return get_irg_bad(current_ir_graph);
1375 /* search for rest of inputs, determine if any of these
1376 are non-self-referencing */
1377 while (++i < n_preds) {
1378 ir_node *scnd_val = get_Phi_pred(n, i);
1379 if ( (scnd_val != n)
1380 && (scnd_val != first_val)
1383 && (! is_Bad(get_Block_cfgpred(block, i)))
1391 /* Fold, if no multiple distinct non-self-referencing inputs */
1393 DBG_OPT_PHI(oldn, n);
1396 } /* equivalent_node_Phi */
1399 * Several optimizations:
1400 * - fold Sync-nodes, iff they have only one predecessor except
1403 static ir_node *equivalent_node_Sync(ir_node *n) {
1404 int arity = get_Sync_n_preds(n);
1407 for (i = 0; i < arity;) {
1408 ir_node *pred = get_Sync_pred(n, i);
1411 /* Remove Bad predecessors */
1418 /* Remove duplicate predecessors */
1424 if (get_Sync_pred(n, j) == pred) {
1432 if (arity == 0) return get_irg_bad(current_ir_graph);
1433 if (arity == 1) return get_Sync_pred(n, 0);
1435 } /* equivalent_node_Sync */
1438 * Optimize Proj(Tuple).
1440 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1441 ir_node *oldn = proj;
1442 ir_node *tuple = get_Proj_pred(proj);
1444 /* Remove the Tuple/Proj combination. */
1445 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1446 DBG_OPT_TUPLE(oldn, tuple, proj);
1449 } /* equivalent_node_Proj_Tuple */
1452 * Optimize a / 1 = a.
1454 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1455 ir_node *oldn = proj;
1456 ir_node *div = get_Proj_pred(proj);
1457 ir_node *b = get_Div_right(div);
1458 tarval *tb = value_of(b);
1460 /* Div is not commutative. */
1461 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1462 switch (get_Proj_proj(proj)) {
1464 proj = get_Div_mem(div);
1465 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1469 proj = get_Div_left(div);
1470 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1474 /* we cannot replace the exception Proj's here, this is done in
1475 transform_node_Proj_Div() */
1480 } /* equivalent_node_Proj_Div */
1483 * Optimize a / 1.0 = a.
1485 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1486 ir_node *oldn = proj;
1487 ir_node *quot = get_Proj_pred(proj);
1488 ir_node *b = get_Quot_right(quot);
1489 tarval *tb = value_of(b);
1491 /* Div is not commutative. */
1492 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1493 switch (get_Proj_proj(proj)) {
1495 proj = get_Quot_mem(quot);
1496 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1500 proj = get_Quot_left(quot);
1501 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1505 /* we cannot replace the exception Proj's here, this is done in
1506 transform_node_Proj_Quot() */
1511 } /* equivalent_node_Proj_Quot */
1514 * Optimize a / 1 = a.
1516 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1517 ir_node *oldn = proj;
1518 ir_node *divmod = get_Proj_pred(proj);
1519 ir_node *b = get_DivMod_right(divmod);
1520 tarval *tb = value_of(b);
1522 /* Div is not commutative. */
1523 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1524 switch (get_Proj_proj(proj)) {
1526 proj = get_DivMod_mem(divmod);
1527 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1530 case pn_DivMod_res_div:
1531 proj = get_DivMod_left(divmod);
1532 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1536 /* we cannot replace the exception Proj's here, this is done in
1537 transform_node_Proj_DivMod().
1538 Note further that the pn_DivMod_res_div case is handled in
1539 computed_value_Proj(). */
1544 } /* equivalent_node_Proj_DivMod */
1547 * Optimize CopyB(mem, x, x) into a Nop.
1549 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1550 ir_node *oldn = proj;
1551 ir_node *copyb = get_Proj_pred(proj);
1552 ir_node *a = get_CopyB_dst(copyb);
1553 ir_node *b = get_CopyB_src(copyb);
1556 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1557 switch (get_Proj_proj(proj)) {
1558 case pn_CopyB_M_regular:
1559 proj = get_CopyB_mem(copyb);
1560 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1563 case pn_CopyB_M_except:
1564 case pn_CopyB_X_except:
1565 DBG_OPT_EXC_REM(proj);
1566 proj = get_irg_bad(current_ir_graph);
1571 } /* equivalent_node_Proj_CopyB */
1574 * Optimize Bounds(idx, idx, upper) into idx.
1576 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1577 ir_node *oldn = proj;
1578 ir_node *bound = get_Proj_pred(proj);
1579 ir_node *idx = get_Bound_index(bound);
1580 ir_node *pred = skip_Proj(idx);
1583 if (idx == get_Bound_lower(bound))
1585 else if (is_Bound(pred)) {
1587 * idx was Bounds checked in the same MacroBlock previously,
1588 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1590 ir_node *lower = get_Bound_lower(bound);
1591 ir_node *upper = get_Bound_upper(bound);
1592 if (get_Bound_lower(pred) == lower &&
1593 get_Bound_upper(pred) == upper &&
1594 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1596 * One could expect that we simply return the previous
1597 * Bound here. However, this would be wrong, as we could
1598 * add an exception Proj to a new location then.
1599 * So, we must turn in into a tuple.
1605 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1606 switch (get_Proj_proj(proj)) {
1608 DBG_OPT_EXC_REM(proj);
1609 proj = get_Bound_mem(bound);
1611 case pn_Bound_X_except:
1612 DBG_OPT_EXC_REM(proj);
1613 proj = get_irg_bad(current_ir_graph);
1617 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1620 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1625 } /* equivalent_node_Proj_Bound */
1628 * Optimize an Exception Proj(Load) with a non-null address.
1630 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1631 if (get_opt_ldst_only_null_ptr_exceptions()) {
1632 if (get_irn_mode(proj) == mode_X) {
1633 ir_node *load = get_Proj_pred(proj);
1635 /* get the Load address */
1636 const ir_node *addr = get_Load_ptr(load);
1637 const ir_node *confirm;
1639 if (value_not_null(addr, &confirm)) {
1640 if (get_Proj_proj(proj) == pn_Load_X_except) {
1641 DBG_OPT_EXC_REM(proj);
1642 return get_irg_bad(current_ir_graph);
1648 } /* equivalent_node_Proj_Load */
1651 * Optimize an Exception Proj(Store) with a non-null address.
1653 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1654 if (get_opt_ldst_only_null_ptr_exceptions()) {
1655 if (get_irn_mode(proj) == mode_X) {
1656 ir_node *store = get_Proj_pred(proj);
1658 /* get the load/store address */
1659 const ir_node *addr = get_Store_ptr(store);
1660 const ir_node *confirm;
1662 if (value_not_null(addr, &confirm)) {
1663 if (get_Proj_proj(proj) == pn_Store_X_except) {
1664 DBG_OPT_EXC_REM(proj);
1665 return get_irg_bad(current_ir_graph);
1671 } /* equivalent_node_Proj_Store */
1674 * Does all optimizations on nodes that must be done on it's Proj's
1675 * because of creating new nodes.
1677 static ir_node *equivalent_node_Proj(ir_node *proj) {
1678 ir_node *n = get_Proj_pred(proj);
1680 if (get_irn_mode(proj) == mode_X) {
1681 if (is_Block_dead(get_nodes_block(n))) {
1682 /* Remove dead control flow -- early gigo(). */
1683 return get_irg_bad(current_ir_graph);
1686 if (n->op->ops.equivalent_node_Proj)
1687 return n->op->ops.equivalent_node_Proj(proj);
1689 } /* equivalent_node_Proj */
1694 static ir_node *equivalent_node_Id(ir_node *n) {
1701 DBG_OPT_ID(oldn, n);
1703 } /* equivalent_node_Id */
1708 static ir_node *equivalent_node_Mux(ir_node *n)
1710 ir_node *oldn = n, *sel = get_Mux_sel(n);
1711 tarval *ts = value_of(sel);
1713 /* Mux(true, f, t) == t */
1714 if (ts == tarval_b_true) {
1715 n = get_Mux_true(n);
1716 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1718 /* Mux(false, f, t) == f */
1719 else if (ts == tarval_b_false) {
1720 n = get_Mux_false(n);
1721 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1723 /* Mux(v, x, x) == x */
1724 else if (get_Mux_false(n) == get_Mux_true(n)) {
1725 n = get_Mux_true(n);
1726 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1728 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1729 ir_node *cmp = get_Proj_pred(sel);
1730 long proj_nr = get_Proj_proj(sel);
1731 ir_node *f = get_Mux_false(n);
1732 ir_node *t = get_Mux_true(n);
1735 * Note further that these optimization work even for floating point
1736 * with NaN's because -NaN == NaN.
1737 * However, if +0 and -0 is handled differently, we cannot use the first one.
1740 ir_node *const cmp_l = get_Cmp_left(cmp);
1741 ir_node *const cmp_r = get_Cmp_right(cmp);
1745 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1746 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1748 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1755 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1756 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1758 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1765 * Note: normalization puts the constant on the right side,
1766 * so we check only one case.
1768 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1769 /* Mux(t CMP 0, X, t) */
1770 if (is_Minus(f) && get_Minus_op(f) == t) {
1771 /* Mux(t CMP 0, -t, t) */
1772 if (proj_nr == pn_Cmp_Eq) {
1773 /* Mux(t == 0, -t, t) ==> -t */
1775 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1776 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1777 /* Mux(t != 0, -t, t) ==> t */
1779 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1786 } /* equivalent_node_Mux */
1789 * Remove Confirm nodes if setting is on.
1790 * Replace Confirms(x, '=', Constlike) by Constlike.
1792 static ir_node *equivalent_node_Confirm(ir_node *n) {
1793 ir_node *pred = get_Confirm_value(n);
1794 pn_Cmp pnc = get_Confirm_cmp(n);
1796 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1798 * rare case: two identical Confirms one after another,
1799 * replace the second one with the first.
1802 pred = get_Confirm_value(n);
1803 pnc = get_Confirm_cmp(n);
1805 if (get_opt_remove_confirm())
1806 return get_Confirm_value(n);
1811 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1812 * perform no actual computation, as, e.g., the Id nodes. It does not create
1813 * new nodes. It is therefore safe to free n if the node returned is not n.
1814 * If a node returns a Tuple we can not just skip it. If the size of the
1815 * in array fits, we transform n into a tuple (e.g., Div).
1817 ir_node *equivalent_node(ir_node *n) {
1818 if (n->op->ops.equivalent_node)
1819 return n->op->ops.equivalent_node(n);
1821 } /* equivalent_node */
1824 * Sets the default equivalent node operation for an ir_op_ops.
1826 * @param code the opcode for the default operation
1827 * @param ops the operations initialized
1832 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1836 ops->equivalent_node = equivalent_node_##a; \
1838 #define CASE_PROJ(a) \
1840 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1882 } /* firm_set_default_equivalent_node */
1885 * Returns non-zero if a node is a Phi node
1886 * with all predecessors constant.
1888 static int is_const_Phi(ir_node *n) {
1891 if (! is_Phi(n) || get_irn_arity(n) == 0)
1893 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1894 if (! is_Const(get_irn_n(n, i)))
1898 } /* is_const_Phi */
1900 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1901 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1904 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1906 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1907 if (eval == tarval_sub) {
1908 tarval_sub_type func = (tarval_sub_type)eval;
1910 return func(a, b, mode);
1912 tarval_binop_type func = (tarval_binop_type)eval;
1919 * Apply an evaluator on a binop with a constant operators (and one Phi).
1921 * @param phi the Phi node
1922 * @param other the other operand
1923 * @param eval an evaluator function
1924 * @param mode the mode of the result, may be different from the mode of the Phi!
1925 * @param left if non-zero, other is the left operand, else the right
1927 * @return a new Phi node if the conversion was successful, NULL else
1929 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1934 int i, n = get_irn_arity(phi);
1936 NEW_ARR_A(void *, res, n);
1938 for (i = 0; i < n; ++i) {
1939 pred = get_irn_n(phi, i);
1940 tv = get_Const_tarval(pred);
1941 tv = do_eval(eval, other, tv, mode);
1943 if (tv == tarval_bad) {
1944 /* folding failed, bad */
1950 for (i = 0; i < n; ++i) {
1951 pred = get_irn_n(phi, i);
1952 tv = get_Const_tarval(pred);
1953 tv = do_eval(eval, tv, other, mode);
1955 if (tv == tarval_bad) {
1956 /* folding failed, bad */
1962 irg = current_ir_graph;
1963 for (i = 0; i < n; ++i) {
1964 pred = get_irn_n(phi, i);
1965 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1966 mode, res[i], get_Const_type(pred));
1968 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1969 } /* apply_binop_on_phi */
1972 * Apply an evaluator on a binop with two constant Phi.
1974 * @param a the left Phi node
1975 * @param b the right Phi node
1976 * @param eval an evaluator function
1977 * @param mode the mode of the result, may be different from the mode of the Phi!
1979 * @return a new Phi node if the conversion was successful, NULL else
1981 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1982 tarval *tv_l, *tv_r, *tv;
1988 if (get_nodes_block(a) != get_nodes_block(b))
1991 n = get_irn_arity(a);
1992 NEW_ARR_A(void *, res, n);
1994 for (i = 0; i < n; ++i) {
1995 pred = get_irn_n(a, i);
1996 tv_l = get_Const_tarval(pred);
1997 pred = get_irn_n(b, i);
1998 tv_r = get_Const_tarval(pred);
1999 tv = do_eval(eval, tv_l, tv_r, mode);
2001 if (tv == tarval_bad) {
2002 /* folding failed, bad */
2007 irg = current_ir_graph;
2008 for (i = 0; i < n; ++i) {
2009 pred = get_irn_n(a, i);
2010 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
2012 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2013 } /* apply_binop_on_2_phis */
2016 * Apply an evaluator on a unop with a constant operator (a Phi).
2018 * @param phi the Phi node
2019 * @param eval an evaluator function
2021 * @return a new Phi node if the conversion was successful, NULL else
2023 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2029 int i, n = get_irn_arity(phi);
2031 NEW_ARR_A(void *, res, n);
2032 for (i = 0; i < n; ++i) {
2033 pred = get_irn_n(phi, i);
2034 tv = get_Const_tarval(pred);
2037 if (tv == tarval_bad) {
2038 /* folding failed, bad */
2043 mode = get_irn_mode(phi);
2044 irg = current_ir_graph;
2045 for (i = 0; i < n; ++i) {
2046 pred = get_irn_n(phi, i);
2047 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2048 mode, res[i], get_Const_type(pred));
2050 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2051 } /* apply_unop_on_phi */
2054 * Apply a conversion on a constant operator (a Phi).
2056 * @param phi the Phi node
2058 * @return a new Phi node if the conversion was successful, NULL else
2060 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2065 int i, n = get_irn_arity(phi);
2067 NEW_ARR_A(void *, res, n);
2068 for (i = 0; i < n; ++i) {
2069 pred = get_irn_n(phi, i);
2070 tv = get_Const_tarval(pred);
2071 tv = tarval_convert_to(tv, mode);
2073 if (tv == tarval_bad) {
2074 /* folding failed, bad */
2079 irg = current_ir_graph;
2080 for (i = 0; i < n; ++i) {
2081 pred = get_irn_n(phi, i);
2082 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2083 mode, res[i], get_Const_type(pred));
2085 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2086 } /* apply_conv_on_phi */
2089 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2090 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2091 * If possible, remove the Conv's.
2093 static ir_node *transform_node_AddSub(ir_node *n) {
2094 ir_mode *mode = get_irn_mode(n);
2096 if (mode_is_reference(mode)) {
2097 ir_node *left = get_binop_left(n);
2098 ir_node *right = get_binop_right(n);
2099 unsigned ref_bits = get_mode_size_bits(mode);
2101 if (is_Conv(left)) {
2102 ir_mode *lmode = get_irn_mode(left);
2103 unsigned bits = get_mode_size_bits(lmode);
2105 if (ref_bits == bits &&
2106 mode_is_int(lmode) &&
2107 get_mode_arithmetic(lmode) == irma_twos_complement) {
2108 ir_node *pre = get_Conv_op(left);
2109 ir_mode *pre_mode = get_irn_mode(pre);
2111 if (mode_is_int(pre_mode) &&
2112 get_mode_size_bits(pre_mode) == bits &&
2113 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2114 /* ok, this conv just changes to sign, moreover the calculation
2115 * is done with same number of bits as our address mode, so
2116 * we can ignore the conv as address calculation can be viewed
2117 * as either signed or unsigned
2119 set_binop_left(n, pre);
2124 if (is_Conv(right)) {
2125 ir_mode *rmode = get_irn_mode(right);
2126 unsigned bits = get_mode_size_bits(rmode);
2128 if (ref_bits == bits &&
2129 mode_is_int(rmode) &&
2130 get_mode_arithmetic(rmode) == irma_twos_complement) {
2131 ir_node *pre = get_Conv_op(right);
2132 ir_mode *pre_mode = get_irn_mode(pre);
2134 if (mode_is_int(pre_mode) &&
2135 get_mode_size_bits(pre_mode) == bits &&
2136 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2137 /* ok, this conv just changes to sign, moreover the calculation
2138 * is done with same number of bits as our address mode, so
2139 * we can ignore the conv as address calculation can be viewed
2140 * as either signed or unsigned
2142 set_binop_right(n, pre);
2147 /* let address arithmetic use unsigned modes */
2148 if (is_Const(right)) {
2149 ir_mode *rmode = get_irn_mode(right);
2151 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2152 /* convert a AddP(P, *s) into AddP(P, *u) */
2153 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2155 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2156 set_binop_right(n, pre);
2162 } /* transform_node_AddSub */
2164 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2166 if (is_Const(b) && is_const_Phi(a)) { \
2167 /* check for Op(Phi, Const) */ \
2168 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2170 else if (is_Const(a) && is_const_Phi(b)) { \
2171 /* check for Op(Const, Phi) */ \
2172 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2174 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2175 /* check for Op(Phi, Phi) */ \
2176 c = apply_binop_on_2_phis(a, b, eval, mode); \
2179 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2183 #define HANDLE_UNOP_PHI(eval, a, c) \
2185 if (is_const_Phi(a)) { \
2186 /* check for Op(Phi) */ \
2187 c = apply_unop_on_phi(a, eval); \
2189 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2195 * Do the AddSub optimization, then Transform
2196 * Constant folding on Phi
2197 * Add(a,a) -> Mul(a, 2)
2198 * Add(Mul(a, x), a) -> Mul(a, x+1)
2199 * if the mode is integer or float.
2200 * Transform Add(a,-b) into Sub(a,b).
2201 * Reassociation might fold this further.
2203 static ir_node *transform_node_Add(ir_node *n) {
2205 ir_node *a, *b, *c, *oldn = n;
2207 n = transform_node_AddSub(n);
2209 a = get_Add_left(n);
2210 b = get_Add_right(n);
2212 mode = get_irn_mode(n);
2214 if (mode_is_reference(mode)) {
2215 ir_mode *lmode = get_irn_mode(a);
2217 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2218 /* an Add(a, NULL) is a hidden Conv */
2219 dbg_info *dbg = get_irn_dbg_info(n);
2220 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2224 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2226 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2227 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2230 if (mode_is_num(mode)) {
2231 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2232 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2233 ir_node *block = get_nodes_block(n);
2236 get_irn_dbg_info(n),
2240 new_r_Const_long(current_ir_graph, block, mode, 2),
2242 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2247 get_irn_dbg_info(n),
2253 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2258 get_irn_dbg_info(n),
2264 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2267 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2268 /* Here we rely on constants be on the RIGHT side */
2270 ir_node *op = get_Not_op(a);
2272 if (is_Const(b) && is_Const_one(b)) {
2274 ir_node *blk = get_nodes_block(n);
2275 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2276 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2281 ir_node *blk = get_nodes_block(n);
2282 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2283 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2288 ir_node *op = get_Not_op(b);
2292 ir_node *blk = get_nodes_block(n);
2293 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2294 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2301 } /* transform_node_Add */
2304 * returns -cnst or NULL if impossible
2306 static ir_node *const_negate(ir_node *cnst) {
2307 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2308 dbg_info *dbgi = get_irn_dbg_info(cnst);
2309 ir_graph *irg = get_irn_irg(cnst);
2310 ir_node *block = get_nodes_block(cnst);
2311 ir_mode *mode = get_irn_mode(cnst);
2312 if (tv == tarval_bad) return NULL;
2313 return new_rd_Const(dbgi, irg, block, mode, tv);
2317 * Do the AddSub optimization, then Transform
2318 * Constant folding on Phi
2319 * Sub(0,a) -> Minus(a)
2320 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2321 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2322 * Sub(Add(a, x), x) -> a
2323 * Sub(x, Add(x, a)) -> -a
2324 * Sub(x, Const) -> Add(x, -Const)
2326 static ir_node *transform_node_Sub(ir_node *n) {
2331 n = transform_node_AddSub(n);
2333 a = get_Sub_left(n);
2334 b = get_Sub_right(n);
2336 mode = get_irn_mode(n);
2338 if (mode_is_int(mode)) {
2339 ir_mode *lmode = get_irn_mode(a);
2341 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2342 /* a Sub(a, NULL) is a hidden Conv */
2343 dbg_info *dbg = get_irn_dbg_info(n);
2344 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2345 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2349 if (mode == lmode &&
2350 get_mode_arithmetic(mode) == irma_twos_complement &&
2352 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2354 dbg_info *dbg = get_irn_dbg_info(n);
2355 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2356 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2362 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2364 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2365 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2368 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2369 /* a - C -> a + (-C) */
2370 ir_node *cnst = const_negate(b);
2372 ir_node *block = get_nodes_block(n);
2373 dbg_info *dbgi = get_irn_dbg_info(n);
2374 ir_graph *irg = get_irn_irg(n);
2376 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2377 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2382 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2383 ir_graph *irg = current_ir_graph;
2384 dbg_info *dbg = get_irn_dbg_info(n);
2385 ir_node *block = get_nodes_block(n);
2386 ir_node *left = get_Minus_op(a);
2387 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2389 n = new_rd_Minus(dbg, irg, block, add, mode);
2390 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2392 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2393 ir_graph *irg = current_ir_graph;
2394 dbg_info *dbg = get_irn_dbg_info(n);
2395 ir_node *block = get_nodes_block(n);
2396 ir_node *right = get_Minus_op(b);
2398 n = new_rd_Add(dbg, irg, block, a, right, mode);
2399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2401 } else if (is_Sub(b)) {
2402 /* a - (b - c) -> a + (c - b)
2403 * -> (a - b) + c iff (b - c) is a pointer */
2404 ir_graph *irg = current_ir_graph;
2405 dbg_info *s_dbg = get_irn_dbg_info(b);
2406 ir_node *s_block = get_nodes_block(b);
2407 ir_node *s_left = get_Sub_left(b);
2408 ir_node *s_right = get_Sub_right(b);
2409 ir_mode *s_mode = get_irn_mode(b);
2410 if (s_mode == mode_P) {
2411 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2412 dbg_info *a_dbg = get_irn_dbg_info(n);
2413 ir_node *a_block = get_nodes_block(n);
2416 s_right = new_r_Conv(irg, a_block, s_right, mode);
2417 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2419 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2420 dbg_info *a_dbg = get_irn_dbg_info(n);
2421 ir_node *a_block = get_nodes_block(n);
2423 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2425 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2427 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2428 ir_node *m_right = get_Mul_right(b);
2429 if (is_Const(m_right)) {
2430 ir_node *cnst2 = const_negate(m_right);
2431 if (cnst2 != NULL) {
2432 ir_graph *irg = current_ir_graph;
2433 dbg_info *m_dbg = get_irn_dbg_info(b);
2434 ir_node *m_block = get_nodes_block(b);
2435 ir_node *m_left = get_Mul_left(b);
2436 ir_mode *m_mode = get_irn_mode(b);
2437 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2438 dbg_info *a_dbg = get_irn_dbg_info(n);
2439 ir_node *a_block = get_nodes_block(n);
2441 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2448 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2449 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2451 get_irn_dbg_info(n),
2456 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2460 if (mode_wrap_around(mode)) {
2461 ir_node *left = get_Add_left(a);
2462 ir_node *right = get_Add_right(a);
2464 /* FIXME: Does the Conv's work only for two complement or generally? */
2466 if (mode != get_irn_mode(right)) {
2467 /* This Sub is an effective Cast */
2468 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2471 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2473 } else if (right == b) {
2474 if (mode != get_irn_mode(left)) {
2475 /* This Sub is an effective Cast */
2476 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2479 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2485 if (mode_wrap_around(mode)) {
2486 ir_node *left = get_Add_left(b);
2487 ir_node *right = get_Add_right(b);
2489 /* FIXME: Does the Conv's work only for two complement or generally? */
2491 ir_mode *r_mode = get_irn_mode(right);
2493 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2494 if (mode != r_mode) {
2495 /* This Sub is an effective Cast */
2496 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2498 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2500 } else if (right == a) {
2501 ir_mode *l_mode = get_irn_mode(left);
2503 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2504 if (mode != l_mode) {
2505 /* This Sub is an effective Cast */
2506 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2508 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2513 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2514 ir_mode *mode = get_irn_mode(a);
2516 if (mode == get_irn_mode(b)) {
2518 ir_node *op_a = get_Conv_op(a);
2519 ir_node *op_b = get_Conv_op(b);
2521 /* check if it's allowed to skip the conv */
2522 ma = get_irn_mode(op_a);
2523 mb = get_irn_mode(op_b);
2525 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2526 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2529 set_Sub_right(n, b);
2535 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2536 if (!is_reassoc_running() && is_Mul(a)) {
2537 ir_node *ma = get_Mul_left(a);
2538 ir_node *mb = get_Mul_right(a);
2541 ir_node *blk = get_nodes_block(n);
2543 get_irn_dbg_info(n),
2544 current_ir_graph, blk,
2547 get_irn_dbg_info(n),
2548 current_ir_graph, blk,
2550 new_r_Const_long(current_ir_graph, blk, mode, 1),
2553 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2555 } else if (mb == b) {
2556 ir_node *blk = get_nodes_block(n);
2558 get_irn_dbg_info(n),
2559 current_ir_graph, blk,
2562 get_irn_dbg_info(n),
2563 current_ir_graph, blk,
2565 new_r_Const_long(current_ir_graph, blk, mode, 1),
2568 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2572 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2573 ir_node *x = get_Sub_left(a);
2574 ir_node *y = get_Sub_right(a);
2575 ir_node *blk = get_nodes_block(n);
2576 ir_mode *m_b = get_irn_mode(b);
2577 ir_mode *m_y = get_irn_mode(y);
2581 /* Determine the right mode for the Add. */
2584 else if (mode_is_reference(m_b))
2586 else if (mode_is_reference(m_y))
2590 * Both modes are different but none is reference,
2591 * happens for instance in SubP(SubP(P, Iu), Is).
2592 * We have two possibilities here: Cast or ignore.
2593 * Currently we ignore this case.
2598 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2600 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2601 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2605 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2606 if (is_Const(a) && is_Not(b)) {
2607 /* c - ~X = X + (c+1) */
2608 tarval *tv = get_Const_tarval(a);
2610 tv = tarval_add(tv, get_mode_one(mode));
2611 if (tv != tarval_bad) {
2612 ir_node *blk = get_nodes_block(n);
2613 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2614 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2615 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2621 } /* transform_node_Sub */
2624 * Several transformation done on n*n=2n bits mul.
2625 * These transformations must be done here because new nodes may be produced.
2627 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2629 ir_node *a = get_Mul_left(n);
2630 ir_node *b = get_Mul_right(n);
2631 tarval *ta = value_of(a);
2632 tarval *tb = value_of(b);
2633 ir_mode *smode = get_irn_mode(a);
2635 if (ta == get_mode_one(smode)) {
2636 /* (L)1 * (L)b = (L)b */
2637 ir_node *blk = get_nodes_block(n);
2638 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2639 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2642 else if (ta == get_mode_minus_one(smode)) {
2643 /* (L)-1 * (L)b = (L)b */
2644 ir_node *blk = get_nodes_block(n);
2645 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2646 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2647 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2650 if (tb == get_mode_one(smode)) {
2651 /* (L)a * (L)1 = (L)a */
2652 ir_node *blk = get_irn_n(a, -1);
2653 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2654 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2657 else if (tb == get_mode_minus_one(smode)) {
2658 /* (L)a * (L)-1 = (L)-a */
2659 ir_node *blk = get_nodes_block(n);
2660 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2661 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2662 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2669 * Transform Mul(a,-1) into -a.
2670 * Do constant evaluation of Phi nodes.
2671 * Do architecture dependent optimizations on Mul nodes
2673 static ir_node *transform_node_Mul(ir_node *n) {
2674 ir_node *c, *oldn = n;
2675 ir_mode *mode = get_irn_mode(n);
2676 ir_node *a = get_Mul_left(n);
2677 ir_node *b = get_Mul_right(n);
2679 if (is_Bad(a) || is_Bad(b))
2682 if (mode != get_irn_mode(a))
2683 return transform_node_Mul2n(n, mode);
2685 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2687 if (mode_is_signed(mode)) {
2690 if (value_of(a) == get_mode_minus_one(mode))
2692 else if (value_of(b) == get_mode_minus_one(mode))
2695 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2696 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2701 if (is_Const(b)) { /* (-a) * const -> a * -const */
2702 ir_node *cnst = const_negate(b);
2704 dbg_info *dbgi = get_irn_dbg_info(n);
2705 ir_node *block = get_nodes_block(n);
2706 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2707 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2710 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2711 dbg_info *dbgi = get_irn_dbg_info(n);
2712 ir_node *block = get_nodes_block(n);
2713 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2714 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2716 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2717 ir_node *sub_l = get_Sub_left(b);
2718 ir_node *sub_r = get_Sub_right(b);
2719 dbg_info *dbgi = get_irn_dbg_info(n);
2720 ir_graph *irg = current_ir_graph;
2721 ir_node *block = get_nodes_block(n);
2722 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2723 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2724 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2727 } else if (is_Minus(b)) {
2728 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2729 ir_node *sub_l = get_Sub_left(a);
2730 ir_node *sub_r = get_Sub_right(a);
2731 dbg_info *dbgi = get_irn_dbg_info(n);
2732 ir_graph *irg = current_ir_graph;
2733 ir_node *block = get_nodes_block(n);
2734 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2735 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2736 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2740 if (get_mode_arithmetic(mode) == irma_ieee754) {
2742 tarval *tv = get_Const_tarval(a);
2743 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2744 && !tarval_is_negative(tv)) {
2745 /* 2.0 * b = b + b */
2746 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2747 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2751 else if (is_Const(b)) {
2752 tarval *tv = get_Const_tarval(b);
2753 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2754 && !tarval_is_negative(tv)) {
2755 /* a * 2.0 = a + a */
2756 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2757 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2762 return arch_dep_replace_mul_with_shifts(n);
2763 } /* transform_node_Mul */
2766 * Transform a Div Node.
2768 static ir_node *transform_node_Div(ir_node *n) {
2769 ir_mode *mode = get_Div_resmode(n);
2770 ir_node *a = get_Div_left(n);
2771 ir_node *b = get_Div_right(n);
2775 if (is_Const(b) && is_const_Phi(a)) {
2776 /* check for Div(Phi, Const) */
2777 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2779 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2783 else if (is_Const(a) && is_const_Phi(b)) {
2784 /* check for Div(Const, Phi) */
2785 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2787 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2791 else if (is_const_Phi(a) && is_const_Phi(b)) {
2792 /* check for Div(Phi, Phi) */
2793 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2795 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2802 if (tv != tarval_bad) {
2803 value = new_Const(get_tarval_mode(tv), tv);
2805 DBG_OPT_CSTEVAL(n, value);
2808 ir_node *a = get_Div_left(n);
2809 ir_node *b = get_Div_right(n);
2810 const ir_node *dummy;
2812 if (a == b && value_not_zero(a, &dummy)) {
2813 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2814 value = new_Const(mode, get_mode_one(mode));
2815 DBG_OPT_CSTEVAL(n, value);
2818 if (mode_is_signed(mode) && is_Const(b)) {
2819 tarval *tv = get_Const_tarval(b);
2821 if (tv == get_mode_minus_one(mode)) {
2823 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2824 DBG_OPT_CSTEVAL(n, value);
2828 /* Try architecture dependent optimization */
2829 value = arch_dep_replace_div_by_const(n);
2837 /* Turn Div into a tuple (mem, jmp, bad, value) */
2838 mem = get_Div_mem(n);
2839 blk = get_nodes_block(n);
2841 /* skip a potential Pin */
2842 mem = skip_Pin(mem);
2843 turn_into_tuple(n, pn_Div_max);
2844 set_Tuple_pred(n, pn_Div_M, mem);
2845 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2846 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2847 set_Tuple_pred(n, pn_Div_res, value);
2850 } /* transform_node_Div */
2853 * Transform a Mod node.
2855 static ir_node *transform_node_Mod(ir_node *n) {
2856 ir_mode *mode = get_Mod_resmode(n);
2857 ir_node *a = get_Mod_left(n);
2858 ir_node *b = get_Mod_right(n);
2862 if (is_Const(b) && is_const_Phi(a)) {
2863 /* check for Div(Phi, Const) */
2864 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2866 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2870 else if (is_Const(a) && is_const_Phi(b)) {
2871 /* check for Div(Const, Phi) */
2872 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2874 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2878 else if (is_const_Phi(a) && is_const_Phi(b)) {
2879 /* check for Div(Phi, Phi) */
2880 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2882 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2889 if (tv != tarval_bad) {
2890 value = new_Const(get_tarval_mode(tv), tv);
2892 DBG_OPT_CSTEVAL(n, value);
2895 ir_node *a = get_Mod_left(n);
2896 ir_node *b = get_Mod_right(n);
2897 const ir_node *dummy;
2899 if (a == b && value_not_zero(a, &dummy)) {
2900 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2901 value = new_Const(mode, get_mode_null(mode));
2902 DBG_OPT_CSTEVAL(n, value);
2905 if (mode_is_signed(mode) && is_Const(b)) {
2906 tarval *tv = get_Const_tarval(b);
2908 if (tv == get_mode_minus_one(mode)) {
2910 value = new_Const(mode, get_mode_null(mode));
2911 DBG_OPT_CSTEVAL(n, value);
2915 /* Try architecture dependent optimization */
2916 value = arch_dep_replace_mod_by_const(n);
2924 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2925 mem = get_Mod_mem(n);
2926 blk = get_nodes_block(n);
2928 /* skip a potential Pin */
2929 mem = skip_Pin(mem);
2930 turn_into_tuple(n, pn_Mod_max);
2931 set_Tuple_pred(n, pn_Mod_M, mem);
2932 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2933 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2934 set_Tuple_pred(n, pn_Mod_res, value);
2937 } /* transform_node_Mod */
2940 * Transform a DivMod node.
2942 static ir_node *transform_node_DivMod(ir_node *n) {
2943 const ir_node *dummy;
2944 ir_node *a = get_DivMod_left(n);
2945 ir_node *b = get_DivMod_right(n);
2946 ir_mode *mode = get_DivMod_resmode(n);
2951 if (is_Const(b) && is_const_Phi(a)) {
2952 /* check for Div(Phi, Const) */
2953 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2954 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2956 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2957 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2961 else if (is_Const(a) && is_const_Phi(b)) {
2962 /* check for Div(Const, Phi) */
2963 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2964 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2966 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2967 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2971 else if (is_const_Phi(a) && is_const_Phi(b)) {
2972 /* check for Div(Phi, Phi) */
2973 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2974 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2976 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2977 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2984 if (tb != tarval_bad) {
2985 if (tb == get_mode_one(get_tarval_mode(tb))) {
2987 vb = new_Const(mode, get_mode_null(mode));
2988 DBG_OPT_CSTEVAL(n, vb);
2990 } else if (ta != tarval_bad) {
2991 tarval *resa, *resb;
2992 resa = tarval_div(ta, tb);
2993 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2994 Jmp for X result!? */
2995 resb = tarval_mod(ta, tb);
2996 if (resb == tarval_bad) return n; /* Causes exception! */
2997 va = new_Const(mode, resa);
2998 vb = new_Const(mode, resb);
2999 DBG_OPT_CSTEVAL(n, va);
3000 DBG_OPT_CSTEVAL(n, vb);
3002 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3003 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
3004 vb = new_Const(mode, get_mode_null(mode));
3005 DBG_OPT_CSTEVAL(n, va);
3006 DBG_OPT_CSTEVAL(n, vb);
3008 } else { /* Try architecture dependent optimization */
3011 arch_dep_replace_divmod_by_const(&va, &vb, n);
3012 evaluated = va != NULL;
3014 } else if (a == b) {
3015 if (value_not_zero(a, &dummy)) {
3017 va = new_Const(mode, get_mode_one(mode));
3018 vb = new_Const(mode, get_mode_null(mode));
3019 DBG_OPT_CSTEVAL(n, va);
3020 DBG_OPT_CSTEVAL(n, vb);
3023 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3026 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3027 /* 0 / non-Const = 0 */
3032 if (evaluated) { /* replace by tuple */
3036 mem = get_DivMod_mem(n);
3037 /* skip a potential Pin */
3038 mem = skip_Pin(mem);
3040 blk = get_nodes_block(n);
3041 turn_into_tuple(n, pn_DivMod_max);
3042 set_Tuple_pred(n, pn_DivMod_M, mem);
3043 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3044 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3045 set_Tuple_pred(n, pn_DivMod_res_div, va);
3046 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3050 } /* transform_node_DivMod */
3053 * Optimize x / c to x * (1/c)
3055 static ir_node *transform_node_Quot(ir_node *n) {
3056 ir_mode *mode = get_Quot_resmode(n);
3059 if (get_mode_arithmetic(mode) == irma_ieee754) {
3060 ir_node *b = get_Quot_right(n);
3061 tarval *tv = value_of(b);
3063 if (tv != tarval_bad) {
3067 * Floating point constant folding might be disabled here to
3069 * However, as we check for exact result, doing it is safe.
3072 rem = tarval_enable_fp_ops(1);
3073 tv = tarval_quo(get_mode_one(mode), tv);
3074 (void)tarval_enable_fp_ops(rem);
3076 /* Do the transformation if the result is either exact or we are not
3077 using strict rules. */
3078 if (tv != tarval_bad &&
3079 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3080 ir_node *blk = get_nodes_block(n);
3081 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3082 ir_node *a = get_Quot_left(n);
3083 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3084 ir_node *mem = get_Quot_mem(n);
3086 /* skip a potential Pin */
3087 mem = skip_Pin(mem);
3088 turn_into_tuple(n, pn_Quot_max);
3089 set_Tuple_pred(n, pn_Quot_M, mem);
3090 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3091 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3092 set_Tuple_pred(n, pn_Quot_res, m);
3093 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3098 } /* transform_node_Quot */
3101 * Optimize Abs(x) into x if x is Confirmed >= 0
3102 * Optimize Abs(x) into -x if x is Confirmed <= 0
3103 * Optimize Abs(-x) int Abs(x)
3105 static ir_node *transform_node_Abs(ir_node *n) {
3106 ir_node *c, *oldn = n;
3107 ir_node *a = get_Abs_op(n);
3110 HANDLE_UNOP_PHI(tarval_abs, a, c);
3112 switch (classify_value_sign(a)) {
3113 case value_classified_negative:
3114 mode = get_irn_mode(n);
3117 * We can replace the Abs by -x here.
3118 * We even could add a new Confirm here
3119 * (if not twos complement)
3121 * Note that -x would create a new node, so we could
3122 * not run it in the equivalent_node() context.
3124 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3125 get_nodes_block(n), a, mode);
3127 DBG_OPT_CONFIRM(oldn, n);
3129 case value_classified_positive:
3130 /* n is positive, Abs is not needed */
3133 DBG_OPT_CONFIRM(oldn, n);
3139 /* Abs(-x) = Abs(x) */
3140 mode = get_irn_mode(n);
3141 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3142 get_nodes_block(n), get_Minus_op(a), mode);
3143 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3147 } /* transform_node_Abs */
3150 * Optimize -a CMP -b into b CMP a.
3151 * This works only for for modes where unary Minus
3153 * Note that two-complement integers can Overflow
3154 * so it will NOT work.
3156 * For == and != can be handled in Proj(Cmp)
3158 static ir_node *transform_node_Cmp(ir_node *n) {
3160 ir_node *left = get_Cmp_left(n);
3161 ir_node *right = get_Cmp_right(n);
3163 if (is_Minus(left) && is_Minus(right) &&
3164 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3165 ir_node *const new_left = get_Minus_op(right);
3166 ir_node *const new_right = get_Minus_op(left);
3167 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3168 get_nodes_block(n), new_left, new_right);
3169 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3172 } /* transform_node_Cmp */
3176 * Transform a Cond node.
3178 * Replace the Cond by a Jmp if it branches on a constant
3181 static ir_node *transform_node_Cond(ir_node *n) {
3184 ir_node *a = get_Cond_selector(n);
3185 tarval *ta = value_of(a);
3187 /* we need block info which is not available in floating irgs */
3188 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3191 if ((ta != tarval_bad) &&
3192 (get_irn_mode(a) == mode_b) &&
3193 (get_opt_unreachable_code())) {
3194 /* It's a boolean Cond, branching on a boolean constant.
3195 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3196 ir_node *blk = get_nodes_block(n);
3197 jmp = new_r_Jmp(current_ir_graph, blk);
3198 turn_into_tuple(n, pn_Cond_max);
3199 if (ta == tarval_b_true) {
3200 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3201 set_Tuple_pred(n, pn_Cond_true, jmp);
3203 set_Tuple_pred(n, pn_Cond_false, jmp);
3204 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3206 /* We might generate an endless loop, so keep it alive. */
3207 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3210 } /* transform_node_Cond */
3213 * Prototype of a recursive transform function
3214 * for bitwise distributive transformations.
3216 typedef ir_node* (*recursive_transform)(ir_node *n);
3219 * makes use of distributive laws for and, or, eor
3220 * and(a OP c, b OP c) -> and(a, b) OP c
3221 * note, might return a different op than n
3223 static ir_node *transform_bitwise_distributive(ir_node *n,
3224 recursive_transform trans_func)
3227 ir_node *a = get_binop_left(n);
3228 ir_node *b = get_binop_right(n);
3229 ir_op *op = get_irn_op(a);
3230 ir_op *op_root = get_irn_op(n);
3232 if(op != get_irn_op(b))
3235 if (op == op_Conv) {
3236 ir_node *a_op = get_Conv_op(a);
3237 ir_node *b_op = get_Conv_op(b);
3238 ir_mode *a_mode = get_irn_mode(a_op);
3239 ir_mode *b_mode = get_irn_mode(b_op);
3240 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3241 ir_node *blk = get_nodes_block(n);
3244 set_binop_left(n, a_op);
3245 set_binop_right(n, b_op);
3246 set_irn_mode(n, a_mode);
3248 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3250 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3256 /* nothing to gain here */
3260 if (op == op_Shrs || op == op_Shr || op == op_Shl
3261 || op == op_And || op == op_Or || op == op_Eor) {
3262 ir_node *a_left = get_binop_left(a);
3263 ir_node *a_right = get_binop_right(a);
3264 ir_node *b_left = get_binop_left(b);
3265 ir_node *b_right = get_binop_right(b);
3267 ir_node *op1 = NULL;
3268 ir_node *op2 = NULL;
3270 if (is_op_commutative(op)) {
3271 if (a_left == b_left) {
3275 } else if(a_left == b_right) {
3279 } else if(a_right == b_left) {
3285 if(a_right == b_right) {
3292 /* (a sop c) & (b sop c) => (a & b) sop c */
3293 ir_node *blk = get_nodes_block(n);
3295 ir_node *new_n = exact_copy(n);
3296 set_binop_left(new_n, op1);
3297 set_binop_right(new_n, op2);
3298 new_n = trans_func(new_n);
3300 if(op_root == op_Eor && op == op_Or) {
3301 dbg_info *dbgi = get_irn_dbg_info(n);
3302 ir_graph *irg = current_ir_graph;
3303 ir_mode *mode = get_irn_mode(c);
3305 c = new_rd_Not(dbgi, irg, blk, c, mode);
3306 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3309 set_nodes_block(n, blk);
3310 set_binop_left(n, new_n);
3311 set_binop_right(n, c);
3312 add_identities(current_ir_graph->value_table, n);
3315 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3326 static ir_node *transform_node_And(ir_node *n) {
3327 ir_node *c, *oldn = n;
3328 ir_node *a = get_And_left(n);
3329 ir_node *b = get_And_right(n);
3332 mode = get_irn_mode(n);
3333 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3335 /* we can evaluate 2 Projs of the same Cmp */
3336 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3337 ir_node *pred_a = get_Proj_pred(a);
3338 ir_node *pred_b = get_Proj_pred(b);
3339 if (pred_a == pred_b) {
3340 dbg_info *dbgi = get_irn_dbg_info(n);
3341 ir_node *block = get_nodes_block(pred_a);
3342 pn_Cmp pn_a = get_Proj_proj(a);
3343 pn_Cmp pn_b = get_Proj_proj(b);
3344 /* yes, we can simply calculate with pncs */
3345 pn_Cmp new_pnc = pn_a & pn_b;
3347 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3352 ir_node *op = get_Not_op(b);
3354 ir_node *ba = get_And_left(op);
3355 ir_node *bb = get_And_right(op);
3357 /* it's enough to test the following cases due to normalization! */
3358 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3359 /* (a|b) & ~(a&b) = a^b */
3360 ir_node *block = get_nodes_block(n);
3362 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3363 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3371 ir_node *op = get_Not_op(a);
3373 ir_node *aa = get_And_left(op);
3374 ir_node *ab = get_And_right(op);
3376 /* it's enough to test the following cases due to normalization! */
3377 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3378 /* (a|b) & ~(a&b) = a^b */
3379 ir_node *block = get_nodes_block(n);
3381 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3382 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3389 ir_node *al = get_Eor_left(a);
3390 ir_node *ar = get_Eor_right(a);
3393 /* (b ^ a) & b -> ~a & b */
3394 dbg_info *dbg = get_irn_dbg_info(n);
3395 ir_node *block = get_nodes_block(n);
3397 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3398 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3399 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3403 /* (a ^ b) & b -> ~a & b */
3404 dbg_info *dbg = get_irn_dbg_info(n);
3405 ir_node *block = get_nodes_block(n);
3407 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3408 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3409 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3414 ir_node *bl = get_Eor_left(b);
3415 ir_node *br = get_Eor_right(b);
3418 /* a & (a ^ b) -> a & ~b */
3419 dbg_info *dbg = get_irn_dbg_info(n);
3420 ir_node *block = get_nodes_block(n);
3422 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3423 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3424 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3428 /* a & (b ^ a) -> a & ~b */
3429 dbg_info *dbg = get_irn_dbg_info(n);
3430 ir_node *block = get_nodes_block(n);
3432 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3433 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3434 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3438 if (is_Not(a) && is_Not(b)) {
3439 /* ~a & ~b = ~(a|b) */
3440 ir_node *block = get_nodes_block(n);
3441 ir_mode *mode = get_irn_mode(n);
3445 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3446 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3447 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3451 n = transform_bitwise_distributive(n, transform_node_And);
3454 } /* transform_node_And */
3459 static ir_node *transform_node_Eor(ir_node *n) {
3460 ir_node *c, *oldn = n;
3461 ir_node *a = get_Eor_left(n);
3462 ir_node *b = get_Eor_right(n);
3463 ir_mode *mode = get_irn_mode(n);
3465 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3467 /* we can evaluate 2 Projs of the same Cmp */
3468 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3469 ir_node *pred_a = get_Proj_pred(a);
3470 ir_node *pred_b = get_Proj_pred(b);
3471 if(pred_a == pred_b) {
3472 dbg_info *dbgi = get_irn_dbg_info(n);
3473 ir_node *block = get_nodes_block(pred_a);
3474 pn_Cmp pn_a = get_Proj_proj(a);
3475 pn_Cmp pn_b = get_Proj_proj(b);
3476 /* yes, we can simply calculate with pncs */
3477 pn_Cmp new_pnc = pn_a ^ pn_b;
3479 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3486 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
3487 mode, get_mode_null(mode));
3488 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3489 } else if (mode == mode_b &&
3491 is_Const(b) && is_Const_one(b) &&
3492 is_Cmp(get_Proj_pred(a))) {
3493 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3494 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3495 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3497 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3498 } else if (is_Const(b)) {
3499 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3500 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3501 ir_node *not_op = get_Not_op(a);
3502 dbg_info *dbg = get_irn_dbg_info(n);
3503 ir_graph *irg = current_ir_graph;
3504 ir_node *block = get_nodes_block(n);
3505 ir_mode *mode = get_irn_mode(n);
3506 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3508 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3509 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3510 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3513 n = transform_bitwise_distributive(n, transform_node_Eor);
3517 } /* transform_node_Eor */
3522 static ir_node *transform_node_Not(ir_node *n) {
3523 ir_node *c, *oldn = n;
3524 ir_node *a = get_Not_op(n);
3525 ir_mode *mode = get_irn_mode(n);
3527 HANDLE_UNOP_PHI(tarval_not,a,c);
3529 /* check for a boolean Not */
3530 if (mode == mode_b &&
3532 is_Cmp(get_Proj_pred(a))) {
3533 /* We negate a Cmp. The Cmp has the negated result anyways! */
3534 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3535 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3536 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3540 ir_node *eor_b = get_Eor_right(a);
3541 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3542 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3543 ir_node *eor_a = get_Eor_left(a);
3544 dbg_info *dbg = get_irn_dbg_info(n);
3545 ir_graph *irg = current_ir_graph;
3546 ir_node *block = get_nodes_block(n);
3547 ir_mode *mode = get_irn_mode(n);
3548 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3552 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3553 if (is_Minus(a)) { /* ~-x -> x + -1 */
3554 dbg_info *dbg = get_irn_dbg_info(n);
3555 ir_graph *irg = current_ir_graph;
3556 ir_node *block = get_nodes_block(n);
3557 ir_node *add_l = get_Minus_op(a);
3558 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3559 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3560 } else if (is_Add(a)) {
3561 ir_node *add_r = get_Add_right(a);
3562 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3563 /* ~(x + -1) = -x */
3564 ir_node *op = get_Add_left(a);
3565 ir_node *blk = get_nodes_block(n);
3566 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3567 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3572 } /* transform_node_Not */
3575 * Transform a Minus.
3579 * -(a >>u (size-1)) = a >>s (size-1)
3580 * -(a >>s (size-1)) = a >>u (size-1)
3581 * -(a * const) -> a * -const
3583 static ir_node *transform_node_Minus(ir_node *n) {
3584 ir_node *c, *oldn = n;
3585 ir_node *a = get_Minus_op(n);
3588 HANDLE_UNOP_PHI(tarval_neg,a,c);
3590 mode = get_irn_mode(a);
3591 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3592 /* the following rules are only to twos-complement */
3595 ir_node *op = get_Not_op(a);
3596 tarval *tv = get_mode_one(mode);
3597 ir_node *blk = get_nodes_block(n);
3598 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3599 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3600 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3604 ir_node *c = get_Shr_right(a);
3607 tarval *tv = get_Const_tarval(c);
3609 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3610 /* -(a >>u (size-1)) = a >>s (size-1) */
3611 ir_node *v = get_Shr_left(a);
3613 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3614 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3620 ir_node *c = get_Shrs_right(a);
3623 tarval *tv = get_Const_tarval(c);
3625 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3626 /* -(a >>s (size-1)) = a >>u (size-1) */
3627 ir_node *v = get_Shrs_left(a);
3629 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3630 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3637 /* - (a-b) = b - a */
3638 ir_node *la = get_Sub_left(a);
3639 ir_node *ra = get_Sub_right(a);
3640 ir_node *blk = get_nodes_block(n);
3642 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3643 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3647 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3648 ir_node *mul_l = get_Mul_left(a);
3649 ir_node *mul_r = get_Mul_right(a);
3650 tarval *tv = value_of(mul_r);
3651 if (tv != tarval_bad) {
3652 tv = tarval_neg(tv);
3653 if (tv != tarval_bad) {
3654 ir_node *cnst = new_Const(mode, tv);
3655 dbg_info *dbg = get_irn_dbg_info(a);
3656 ir_graph *irg = current_ir_graph;
3657 ir_node *block = get_nodes_block(a);
3658 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3659 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3666 } /* transform_node_Minus */
3669 * Transform a Cast_type(Const) into a new Const_type
3671 static ir_node *transform_node_Cast(ir_node *n) {
3673 ir_node *pred = get_Cast_op(n);
3674 ir_type *tp = get_irn_type(n);
3676 if (is_Const(pred) && get_Const_type(pred) != tp) {
3677 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3678 get_Const_tarval(pred), tp);
3679 DBG_OPT_CSTEVAL(oldn, n);
3680 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3681 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3682 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3683 DBG_OPT_CSTEVAL(oldn, n);
3687 } /* transform_node_Cast */
3690 * Transform a Proj(Load) with a non-null address.
3692 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3693 if (get_opt_ldst_only_null_ptr_exceptions()) {
3694 if (get_irn_mode(proj) == mode_X) {
3695 ir_node *load = get_Proj_pred(proj);
3697 /* get the Load address */
3698 const ir_node *addr = get_Load_ptr(load);
3699 const ir_node *confirm;
3701 if (value_not_null(addr, &confirm)) {
3702 if (confirm == NULL) {
3703 /* this node may float if it did not depend on a Confirm */
3704 set_irn_pinned(load, op_pin_state_floats);
3706 if (get_Proj_proj(proj) == pn_Load_X_except) {
3707 DBG_OPT_EXC_REM(proj);
3708 return get_irg_bad(current_ir_graph);
3710 ir_node *blk = get_nodes_block(load);
3711 return new_r_Jmp(current_ir_graph, blk);
3717 } /* transform_node_Proj_Load */
3720 * Transform a Proj(Store) with a non-null address.
3722 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3723 if (get_opt_ldst_only_null_ptr_exceptions()) {
3724 if (get_irn_mode(proj) == mode_X) {
3725 ir_node *store = get_Proj_pred(proj);
3727 /* get the load/store address */
3728 const ir_node *addr = get_Store_ptr(store);
3729 const ir_node *confirm;
3731 if (value_not_null(addr, &confirm)) {
3732 if (confirm == NULL) {
3733 /* this node may float if it did not depend on a Confirm */
3734 set_irn_pinned(store, op_pin_state_floats);
3736 if (get_Proj_proj(proj) == pn_Store_X_except) {
3737 DBG_OPT_EXC_REM(proj);
3738 return get_irg_bad(current_ir_graph);
3740 ir_node *blk = get_nodes_block(store);
3741 return new_r_Jmp(current_ir_graph, blk);
3747 } /* transform_node_Proj_Store */
3750 * Transform a Proj(Div) with a non-zero value.
3751 * Removes the exceptions and routes the memory to the NoMem node.
3753 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3754 ir_node *div = get_Proj_pred(proj);
3755 ir_node *b = get_Div_right(div);
3756 ir_node *res, *new_mem;
3757 const ir_node *confirm;
3760 if (value_not_zero(b, &confirm)) {
3761 /* div(x, y) && y != 0 */
3762 if (confirm == NULL) {
3763 /* we are sure we have a Const != 0 */
3764 new_mem = get_Div_mem(div);
3765 new_mem = skip_Pin(new_mem);
3766 set_Div_mem(div, new_mem);
3767 set_irn_pinned(div, op_pin_state_floats);
3770 proj_nr = get_Proj_proj(proj);
3772 case pn_Div_X_regular:
3773 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3775 case pn_Div_X_except:
3776 /* we found an exception handler, remove it */
3777 DBG_OPT_EXC_REM(proj);
3781 res = get_Div_mem(div);
3782 new_mem = get_irg_no_mem(current_ir_graph);
3785 /* This node can only float up to the Confirm block */
3786 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3788 set_irn_pinned(div, op_pin_state_floats);
3789 /* this is a Div without exception, we can remove the memory edge */
3790 set_Div_mem(div, new_mem);
3795 } /* transform_node_Proj_Div */
3798 * Transform a Proj(Mod) with a non-zero value.
3799 * Removes the exceptions and routes the memory to the NoMem node.
3801 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3802 ir_node *mod = get_Proj_pred(proj);
3803 ir_node *b = get_Mod_right(mod);
3804 ir_node *res, *new_mem;
3805 const ir_node *confirm;
3808 if (value_not_zero(b, &confirm)) {
3809 /* mod(x, y) && y != 0 */
3810 proj_nr = get_Proj_proj(proj);
3812 if (confirm == NULL) {
3813 /* we are sure we have a Const != 0 */
3814 new_mem = get_Mod_mem(mod);
3815 new_mem = skip_Pin(new_mem);
3816 set_Mod_mem(mod, new_mem);
3817 set_irn_pinned(mod, op_pin_state_floats);
3822 case pn_Mod_X_regular:
3823 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3825 case pn_Mod_X_except:
3826 /* we found an exception handler, remove it */
3827 DBG_OPT_EXC_REM(proj);
3831 res = get_Mod_mem(mod);
3832 new_mem = get_irg_no_mem(current_ir_graph);
3835 /* This node can only float up to the Confirm block */
3836 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3838 /* this is a Mod without exception, we can remove the memory edge */
3839 set_Mod_mem(mod, new_mem);
3842 if (get_Mod_left(mod) == b) {
3843 /* a % a = 0 if a != 0 */
3844 ir_mode *mode = get_irn_mode(proj);
3845 ir_node *res = new_Const(mode, get_mode_null(mode));
3847 DBG_OPT_CSTEVAL(mod, res);
3853 } /* transform_node_Proj_Mod */
3856 * Transform a Proj(DivMod) with a non-zero value.
3857 * Removes the exceptions and routes the memory to the NoMem node.
3859 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3860 ir_node *divmod = get_Proj_pred(proj);
3861 ir_node *b = get_DivMod_right(divmod);
3862 ir_node *res, *new_mem;
3863 const ir_node *confirm;
3866 if (value_not_zero(b, &confirm)) {
3867 /* DivMod(x, y) && y != 0 */
3868 proj_nr = get_Proj_proj(proj);
3870 if (confirm == NULL) {
3871 /* we are sure we have a Const != 0 */
3872 new_mem = get_DivMod_mem(divmod);
3873 new_mem = skip_Pin(new_mem);
3874 set_DivMod_mem(divmod, new_mem);
3875 set_irn_pinned(divmod, op_pin_state_floats);
3880 case pn_DivMod_X_regular:
3881 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3883 case pn_DivMod_X_except:
3884 /* we found an exception handler, remove it */
3885 DBG_OPT_EXC_REM(proj);
3889 res = get_DivMod_mem(divmod);
3890 new_mem = get_irg_no_mem(current_ir_graph);
3893 /* This node can only float up to the Confirm block */
3894 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3896 /* this is a DivMod without exception, we can remove the memory edge */
3897 set_DivMod_mem(divmod, new_mem);
3900 case pn_DivMod_res_mod:
3901 if (get_DivMod_left(divmod) == b) {
3902 /* a % a = 0 if a != 0 */
3903 ir_mode *mode = get_irn_mode(proj);
3904 ir_node *res = new_Const(mode, get_mode_null(mode));
3906 DBG_OPT_CSTEVAL(divmod, res);
3912 } /* transform_node_Proj_DivMod */
3915 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3917 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3918 if (get_opt_unreachable_code()) {
3919 ir_node *n = get_Proj_pred(proj);
3920 ir_node *b = get_Cond_selector(n);
3922 if (mode_is_int(get_irn_mode(b))) {
3923 tarval *tb = value_of(b);
3925 if (tb != tarval_bad) {
3926 /* we have a constant switch */
3927 long num = get_Proj_proj(proj);
3929 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3930 if (get_tarval_long(tb) == num) {
3931 /* Do NOT create a jump here, or we will have 2 control flow ops
3932 * in a block. This case is optimized away in optimize_cf(). */
3935 /* this case will NEVER be taken, kill it */
3936 return get_irg_bad(current_ir_graph);
3943 } /* transform_node_Proj_Cond */
3946 * Create a 0 constant of given mode.
3948 static ir_node *create_zero_const(ir_mode *mode) {
3949 tarval *tv = get_mode_null(mode);
3950 ir_node *cnst = new_Const(mode, tv);
3955 /* the order of the values is important! */
3956 typedef enum const_class {
3962 static const_class classify_const(const ir_node* n)
3964 if (is_Const(n)) return const_const;
3965 if (is_irn_constlike(n)) return const_like;
3970 * Determines whether r is more constlike or has a larger index (in that order)
3973 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3975 const const_class l_order = classify_const(l);
3976 const const_class r_order = classify_const(r);
3978 l_order > r_order ||
3979 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3983 * Normalizes and optimizes Cmp nodes.
3985 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3986 ir_node *n = get_Proj_pred(proj);
3987 ir_node *left = get_Cmp_left(n);
3988 ir_node *right = get_Cmp_right(n);
3991 ir_mode *mode = NULL;
3992 long proj_nr = get_Proj_proj(proj);
3994 /* we can evaluate some cases directly */
3997 return new_Const(mode_b, get_tarval_b_false());
3999 return new_Const(mode_b, get_tarval_b_true());
4001 if (!mode_is_float(get_irn_mode(left)))
4002 return new_Const(mode_b, get_tarval_b_true());
4008 /* remove Casts of both sides */
4009 left = skip_Cast(left);
4010 right = skip_Cast(right);
4012 /* Remove unnecessary conversions */
4013 /* TODO handle constants */
4014 if (is_Conv(left) && is_Conv(right)) {
4015 ir_mode *mode = get_irn_mode(left);
4016 ir_node *op_left = get_Conv_op(left);
4017 ir_node *op_right = get_Conv_op(right);
4018 ir_mode *mode_left = get_irn_mode(op_left);
4019 ir_mode *mode_right = get_irn_mode(op_right);
4021 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4022 && mode_left != mode_b && mode_right != mode_b) {
4023 ir_graph *irg = current_ir_graph;
4024 ir_node *block = get_nodes_block(n);
4026 if (mode_left == mode_right) {
4030 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4031 } else if (smaller_mode(mode_left, mode_right)) {
4032 left = new_r_Conv(irg, block, op_left, mode_right);
4035 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4036 } else if (smaller_mode(mode_right, mode_left)) {
4038 right = new_r_Conv(irg, block, op_right, mode_left);
4040 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4045 /* remove operation on both sides if possible */
4046 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4048 * The following operations are NOT safe for floating point operations, for instance
4049 * 1.0 + inf == 2.0 + inf, =/=> x == y
4051 if (mode_is_int(get_irn_mode(left))) {
4052 unsigned lop = get_irn_opcode(left);
4054 if (lop == get_irn_opcode(right)) {
4055 ir_node *ll, *lr, *rl, *rr;
4057 /* same operation on both sides, try to remove */
4061 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4062 left = get_unop_op(left);
4063 right = get_unop_op(right);
4065 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4068 ll = get_Add_left(left);
4069 lr = get_Add_right(left);
4070 rl = get_Add_left(right);
4071 rr = get_Add_right(right);
4074 /* X + a CMP X + b ==> a CMP b */
4078 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4079 } else if (ll == rr) {
4080 /* X + a CMP b + X ==> a CMP b */
4084 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4085 } else if (lr == rl) {
4086 /* a + X CMP X + b ==> a CMP b */
4090 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4091 } else if (lr == rr) {
4092 /* a + X CMP b + X ==> a CMP b */
4096 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4100 ll = get_Sub_left(left);
4101 lr = get_Sub_right(left);
4102 rl = get_Sub_left(right);
4103 rr = get_Sub_right(right);
4106 /* X - a CMP X - b ==> a CMP b */
4110 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4111 } else if (lr == rr) {
4112 /* a - X CMP b - X ==> a CMP b */
4116 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4120 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4121 /* a ROTL X CMP b ROTL X ==> a CMP b */
4122 left = get_Rotl_left(left);
4123 right = get_Rotl_left(right);
4125 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4133 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4134 if (is_Add(left) || is_Sub(left)) {
4135 ir_node *ll = get_binop_left(left);
4136 ir_node *lr = get_binop_right(left);
4138 if (lr == right && is_Add(left)) {
4145 right = create_zero_const(get_irn_mode(left));
4147 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4150 if (is_Add(right) || is_Sub(right)) {
4151 ir_node *rl = get_binop_left(right);
4152 ir_node *rr = get_binop_right(right);
4154 if (rr == left && is_Add(right)) {
4161 right = create_zero_const(get_irn_mode(left));
4163 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4166 } /* mode_is_int(...) */
4167 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4169 /* replace mode_b compares with ands/ors */
4170 if (get_irn_mode(left) == mode_b) {
4171 ir_graph *irg = current_ir_graph;
4172 ir_node *block = get_nodes_block(n);
4176 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4177 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4178 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4179 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4180 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4181 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4182 default: bres = NULL;
4185 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4191 * First step: normalize the compare op
4192 * by placing the constant on the right side
4193 * or moving the lower address node to the left.
4195 if (!operands_are_normalized(left, right)) {
4201 proj_nr = get_inversed_pnc(proj_nr);
4206 * Second step: Try to reduce the magnitude
4207 * of a constant. This may help to generate better code
4208 * later and may help to normalize more compares.
4209 * Of course this is only possible for integer values.
4211 tv = value_of(right);
4212 if (tv != tarval_bad) {
4213 mode = get_irn_mode(right);
4215 /* TODO extend to arbitrary constants */
4216 if (is_Conv(left) && tarval_is_null(tv)) {
4217 ir_node *op = get_Conv_op(left);
4218 ir_mode *op_mode = get_irn_mode(op);
4221 * UpConv(x) REL 0 ==> x REL 0
4223 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4224 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4225 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4226 tv = get_mode_null(op_mode);
4230 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4234 if (tv != tarval_bad) {
4235 /* the following optimization is possible on modes without Overflow
4236 * on Unary Minus or on == and !=:
4237 * -a CMP c ==> a swap(CMP) -c
4239 * Beware: for two-complement Overflow may occur, so only == and != can
4240 * be optimized, see this:
4241 * -MININT < 0 =/=> MININT > 0 !!!
4243 if (is_Minus(left) &&
4244 (!mode_overflow_on_unary_Minus(mode) ||
4245 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4246 tv = tarval_neg(tv);
4248 if (tv != tarval_bad) {
4249 left = get_Minus_op(left);
4250 proj_nr = get_inversed_pnc(proj_nr);
4252 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4254 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4255 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4256 tv = tarval_not(tv);
4258 if (tv != tarval_bad) {
4259 left = get_Not_op(left);
4261 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4265 /* for integer modes, we have more */
4266 if (mode_is_int(mode)) {
4267 /* Ne includes Unordered which is not possible on integers.
4268 * However, frontends often use this wrong, so fix it here */
4269 if (proj_nr & pn_Cmp_Uo) {
4270 proj_nr &= ~pn_Cmp_Uo;
4271 set_Proj_proj(proj, proj_nr);
4274 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4275 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4276 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4277 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4279 if (tv != tarval_bad) {
4280 proj_nr ^= pn_Cmp_Eq;
4282 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4285 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4286 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4287 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4288 tv = tarval_add(tv, get_mode_one(mode));
4290 if (tv != tarval_bad) {
4291 proj_nr ^= pn_Cmp_Eq;
4293 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4297 /* the following reassociations work only for == and != */
4298 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4300 #if 0 /* Might be not that good in general */
4301 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4302 if (tarval_is_null(tv) && is_Sub(left)) {
4303 right = get_Sub_right(left);
4304 left = get_Sub_left(left);
4306 tv = value_of(right);
4308 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4312 if (tv != tarval_bad) {
4313 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4315 ir_node *c1 = get_Sub_right(left);
4316 tarval *tv2 = value_of(c1);
4318 if (tv2 != tarval_bad) {
4319 tv2 = tarval_add(tv, value_of(c1));
4321 if (tv2 != tarval_bad) {
4322 left = get_Sub_left(left);
4325 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4329 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4330 else if (is_Add(left)) {
4331 ir_node *a_l = get_Add_left(left);
4332 ir_node *a_r = get_Add_right(left);
4336 if (is_Const(a_l)) {
4338 tv2 = value_of(a_l);
4341 tv2 = value_of(a_r);
4344 if (tv2 != tarval_bad) {
4345 tv2 = tarval_sub(tv, tv2, NULL);
4347 if (tv2 != tarval_bad) {
4351 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4355 /* -a == c ==> a == -c, -a != c ==> a != -c */
4356 else if (is_Minus(left)) {
4357 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4359 if (tv2 != tarval_bad) {
4360 left = get_Minus_op(left);
4363 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4368 /* the following reassociations work only for <= */
4369 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4370 if (tv != tarval_bad) {
4371 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4372 if (is_Abs(left)) { // TODO something is missing here
4378 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4379 switch (get_irn_opcode(left)) {
4383 c1 = get_And_right(left);
4386 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4387 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4389 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4391 /* TODO: move to constant evaluation */
4392 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4393 c1 = new_Const(mode_b, tv);
4394 DBG_OPT_CSTEVAL(proj, c1);
4398 if (tarval_is_single_bit(tv)) {
4400 * optimization for AND:
4402 * And(x, C) == C ==> And(x, C) != 0
4403 * And(x, C) != C ==> And(X, C) == 0
4405 * if C is a single Bit constant.
4408 /* check for Constant's match. We have check hare the tarvals,
4409 because our const might be changed */
4410 if (get_Const_tarval(c1) == tv) {
4411 /* fine: do the transformation */
4412 tv = get_mode_null(get_tarval_mode(tv));
4413 proj_nr ^= pn_Cmp_Leg;
4415 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4421 c1 = get_Or_right(left);
4422 if (is_Const(c1) && tarval_is_null(tv)) {
4424 * Or(x, C) == 0 && C != 0 ==> FALSE
4425 * Or(x, C) != 0 && C != 0 ==> TRUE
4427 if (! tarval_is_null(get_Const_tarval(c1))) {
4428 /* TODO: move to constant evaluation */
4429 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4430 c1 = new_Const(mode_b, tv);
4431 DBG_OPT_CSTEVAL(proj, c1);
4438 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4440 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4443 c1 = get_Shl_right(left);
4445 tarval *tv1 = get_Const_tarval(c1);
4446 ir_mode *mode = get_irn_mode(left);
4447 tarval *minus1 = get_mode_all_one(mode);
4448 tarval *amask = tarval_shr(minus1, tv1);
4449 tarval *cmask = tarval_shl(minus1, tv1);
4452 if (tarval_and(tv, cmask) != tv) {
4453 /* condition not met */
4454 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4455 c1 = new_Const(mode_b, tv);
4456 DBG_OPT_CSTEVAL(proj, c1);
4459 sl = get_Shl_left(left);
4460 blk = get_nodes_block(n);
4461 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4462 tv = tarval_shr(tv, tv1);
4464 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4469 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4471 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4474 c1 = get_Shr_right(left);
4476 tarval *tv1 = get_Const_tarval(c1);
4477 ir_mode *mode = get_irn_mode(left);
4478 tarval *minus1 = get_mode_all_one(mode);
4479 tarval *amask = tarval_shl(minus1, tv1);
4480 tarval *cmask = tarval_shr(minus1, tv1);
4483 if (tarval_and(tv, cmask) != tv) {
4484 /* condition not met */
4485 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4486 c1 = new_Const(mode_b, tv);
4487 DBG_OPT_CSTEVAL(proj, c1);
4490 sl = get_Shr_left(left);
4491 blk = get_nodes_block(n);
4492 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4493 tv = tarval_shl(tv, tv1);
4495 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4500 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4502 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4505 c1 = get_Shrs_right(left);
4507 tarval *tv1 = get_Const_tarval(c1);
4508 ir_mode *mode = get_irn_mode(left);
4509 tarval *minus1 = get_mode_all_one(mode);
4510 tarval *amask = tarval_shl(minus1, tv1);
4511 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4514 cond = tarval_sub(cond, tv1, NULL);
4515 cond = tarval_shrs(tv, cond);
4517 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4518 /* condition not met */
4519 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4520 c1 = new_Const(mode_b, tv);
4521 DBG_OPT_CSTEVAL(proj, c1);
4524 sl = get_Shrs_left(left);
4525 blk = get_nodes_block(n);
4526 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4527 tv = tarval_shl(tv, tv1);
4529 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4534 } /* tarval != bad */
4537 if (changed & 2) /* need a new Const */
4538 right = new_Const(mode, tv);
4540 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4541 ir_node *op = get_Proj_pred(left);
4543 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4544 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4545 ir_node *c = get_binop_right(op);
4548 tarval *tv = get_Const_tarval(c);
4550 if (tarval_is_single_bit(tv)) {
4551 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4552 ir_node *v = get_binop_left(op);
4553 ir_node *blk = get_irn_n(op, -1);
4554 ir_mode *mode = get_irn_mode(v);
4556 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4557 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4559 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4566 ir_node *block = get_nodes_block(n);
4568 /* create a new compare */
4569 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4570 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4574 } /* transform_node_Proj_Cmp */
4577 * Optimize CopyB(mem, x, x) into a Nop.
4579 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4580 ir_node *copyb = get_Proj_pred(proj);
4581 ir_node *a = get_CopyB_dst(copyb);
4582 ir_node *b = get_CopyB_src(copyb);
4585 switch (get_Proj_proj(proj)) {
4586 case pn_CopyB_X_regular:
4587 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4588 DBG_OPT_EXC_REM(proj);
4589 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4591 case pn_CopyB_M_except:
4592 case pn_CopyB_X_except:
4593 DBG_OPT_EXC_REM(proj);
4594 proj = get_irg_bad(current_ir_graph);
4601 } /* transform_node_Proj_CopyB */
4604 * Optimize Bounds(idx, idx, upper) into idx.
4606 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4607 ir_node *oldn = proj;
4608 ir_node *bound = get_Proj_pred(proj);
4609 ir_node *idx = get_Bound_index(bound);
4610 ir_node *pred = skip_Proj(idx);
4613 if (idx == get_Bound_lower(bound))
4615 else if (is_Bound(pred)) {
4617 * idx was Bounds checked in the same MacroBlock previously,
4618 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4620 ir_node *lower = get_Bound_lower(bound);
4621 ir_node *upper = get_Bound_upper(bound);
4622 if (get_Bound_lower(pred) == lower &&
4623 get_Bound_upper(pred) == upper &&
4624 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4626 * One could expect that we simply return the previous
4627 * Bound here. However, this would be wrong, as we could
4628 * add an exception Proj to a new location then.
4629 * So, we must turn in into a tuple.
4635 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4636 switch (get_Proj_proj(proj)) {
4638 DBG_OPT_EXC_REM(proj);
4639 proj = get_Bound_mem(bound);
4641 case pn_Bound_X_except:
4642 DBG_OPT_EXC_REM(proj);
4643 proj = get_irg_bad(current_ir_graph);
4647 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4649 case pn_Bound_X_regular:
4650 DBG_OPT_EXC_REM(proj);
4651 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4658 } /* transform_node_Proj_Bound */
4661 * Does all optimizations on nodes that must be done on it's Proj's
4662 * because of creating new nodes.
4664 static ir_node *transform_node_Proj(ir_node *proj) {
4665 ir_node *n = get_Proj_pred(proj);
4667 if (n->op->ops.transform_node_Proj)
4668 return n->op->ops.transform_node_Proj(proj);
4670 } /* transform_node_Proj */
4673 * Move Confirms down through Phi nodes.
4675 static ir_node *transform_node_Phi(ir_node *phi) {
4677 ir_mode *mode = get_irn_mode(phi);
4679 if (mode_is_reference(mode)) {
4680 n = get_irn_arity(phi);
4682 /* Beware of Phi0 */
4684 ir_node *pred = get_irn_n(phi, 0);
4685 ir_node *bound, *new_Phi, *block, **in;
4688 if (! is_Confirm(pred))
4691 bound = get_Confirm_bound(pred);
4692 pnc = get_Confirm_cmp(pred);
4694 NEW_ARR_A(ir_node *, in, n);
4695 in[0] = get_Confirm_value(pred);
4697 for (i = 1; i < n; ++i) {
4698 pred = get_irn_n(phi, i);
4700 if (! is_Confirm(pred) ||
4701 get_Confirm_bound(pred) != bound ||
4702 get_Confirm_cmp(pred) != pnc)
4704 in[i] = get_Confirm_value(pred);
4706 /* move the Confirm nodes "behind" the Phi */
4707 block = get_irn_n(phi, -1);
4708 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4709 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4713 } /* transform_node_Phi */
4716 * Returns the operands of a commutative bin-op, if one operand is
4717 * a const, it is returned as the second one.
4719 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4720 ir_node *op_a = get_binop_left(binop);
4721 ir_node *op_b = get_binop_right(binop);
4723 assert(is_op_commutative(get_irn_op(binop)));
4725 if (is_Const(op_a)) {
4732 } /* get_comm_Binop_Ops */
4735 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4736 * Such pattern may arise in bitfield stores.
4738 * value c4 value c4 & c2
4739 * AND c3 AND c1 | c3
4746 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4749 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4752 ir_node *and_l, *c3;
4753 ir_node *value, *c4;
4754 ir_node *new_and, *new_const, *block;
4755 ir_mode *mode = get_irn_mode(or);
4757 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4760 get_comm_Binop_Ops(or, &and, &c1);
4761 if (!is_Const(c1) || !is_And(and))
4764 get_comm_Binop_Ops(and, &or_l, &c2);
4768 tv1 = get_Const_tarval(c1);
4769 tv2 = get_Const_tarval(c2);
4771 tv = tarval_or(tv1, tv2);
4772 if (tarval_is_all_one(tv)) {
4773 /* the AND does NOT clear a bit with isn't set by the OR */
4774 set_Or_left(or, or_l);
4775 set_Or_right(or, c1);
4777 /* check for more */
4784 get_comm_Binop_Ops(or_l, &and_l, &c3);
4785 if (!is_Const(c3) || !is_And(and_l))
4788 get_comm_Binop_Ops(and_l, &value, &c4);
4792 /* ok, found the pattern, check for conditions */
4793 assert(mode == get_irn_mode(and));
4794 assert(mode == get_irn_mode(or_l));
4795 assert(mode == get_irn_mode(and_l));
4797 tv3 = get_Const_tarval(c3);
4798 tv4 = get_Const_tarval(c4);
4800 tv = tarval_or(tv4, tv2);
4801 if (!tarval_is_all_one(tv)) {
4802 /* have at least one 0 at the same bit position */
4806 n_tv4 = tarval_not(tv4);
4807 if (tv3 != tarval_and(tv3, n_tv4)) {
4808 /* bit in the or_mask is outside the and_mask */
4812 n_tv2 = tarval_not(tv2);
4813 if (tv1 != tarval_and(tv1, n_tv2)) {
4814 /* bit in the or_mask is outside the and_mask */
4818 /* ok, all conditions met */
4819 block = get_irn_n(or, -1);
4821 new_and = new_r_And(current_ir_graph, block,
4822 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4824 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4826 set_Or_left(or, new_and);
4827 set_Or_right(or, new_const);
4829 /* check for more */
4831 } /* transform_node_Or_bf_store */
4834 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4836 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4837 ir_mode *mode = get_irn_mode(or);
4838 ir_node *shl, *shr, *block;
4839 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4842 if (! mode_is_int(mode))
4845 shl = get_binop_left(or);
4846 shr = get_binop_right(or);
4855 } else if (!is_Shl(shl)) {
4857 } else if (!is_Shr(shr)) {
4860 x = get_Shl_left(shl);
4861 if (x != get_Shr_left(shr))
4864 c1 = get_Shl_right(shl);
4865 c2 = get_Shr_right(shr);
4866 if (is_Const(c1) && is_Const(c2)) {
4867 tv1 = get_Const_tarval(c1);
4868 if (! tarval_is_long(tv1))
4871 tv2 = get_Const_tarval(c2);
4872 if (! tarval_is_long(tv2))
4875 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4876 != (int) get_mode_size_bits(mode))
4879 /* yet, condition met */
4880 block = get_nodes_block(or);
4882 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4884 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4891 rotval = sub; /* a Rot right is not supported, so use a rot left */
4892 } else if (is_Sub(c2)) {
4898 if (get_Sub_right(sub) != v)
4901 c1 = get_Sub_left(sub);
4905 tv1 = get_Const_tarval(c1);
4906 if (! tarval_is_long(tv1))
4909 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4912 /* yet, condition met */
4913 block = get_nodes_block(or);
4915 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4917 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4919 } /* transform_node_Or_Rotl */
4924 static ir_node *transform_node_Or(ir_node *n) {
4925 ir_node *c, *oldn = n;
4926 ir_node *a = get_Or_left(n);
4927 ir_node *b = get_Or_right(n);
4930 if (is_Not(a) && is_Not(b)) {
4931 /* ~a | ~b = ~(a&b) */
4932 ir_node *block = get_nodes_block(n);
4934 mode = get_irn_mode(n);
4937 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4938 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4939 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4943 /* we can evaluate 2 Projs of the same Cmp */
4944 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4945 ir_node *pred_a = get_Proj_pred(a);
4946 ir_node *pred_b = get_Proj_pred(b);
4947 if (pred_a == pred_b) {
4948 dbg_info *dbgi = get_irn_dbg_info(n);
4949 ir_node *block = get_nodes_block(pred_a);
4950 pn_Cmp pn_a = get_Proj_proj(a);
4951 pn_Cmp pn_b = get_Proj_proj(b);
4952 /* yes, we can simply calculate with pncs */
4953 pn_Cmp new_pnc = pn_a | pn_b;
4955 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4960 mode = get_irn_mode(n);
4961 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
4963 n = transform_node_Or_bf_store(n);
4964 n = transform_node_Or_Rotl(n);
4968 n = transform_bitwise_distributive(n, transform_node_Or);
4971 } /* transform_node_Or */
4975 static ir_node *transform_node(ir_node *n);
4978 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4980 * Should be moved to reassociation?
4982 static ir_node *transform_node_shift(ir_node *n) {
4983 ir_node *left, *right;
4985 tarval *tv1, *tv2, *res;
4986 ir_node *in[2], *irn, *block;
4988 left = get_binop_left(n);
4990 /* different operations */
4991 if (get_irn_op(left) != get_irn_op(n))
4994 right = get_binop_right(n);
4995 tv1 = value_of(right);
4996 if (tv1 == tarval_bad)
4999 tv2 = value_of(get_binop_right(left));
5000 if (tv2 == tarval_bad)
5003 res = tarval_add(tv1, tv2);
5004 mode = get_irn_mode(n);
5006 /* beware: a simple replacement works only, if res < modulo shift */
5008 int modulo_shf = get_mode_modulo_shift(mode);
5009 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5010 if (modulo_shf > 0) {
5011 tarval *modulo = new_tarval_from_long(modulo_shf,
5012 get_tarval_mode(res));
5014 /* shifting too much */
5015 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5017 ir_graph *irg = get_irn_irg(n);
5018 ir_node *block = get_nodes_block(n);
5019 dbg_info *dbgi = get_irn_dbg_info(n);
5020 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
5021 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5025 return new_Const(mode, get_mode_null(mode));
5029 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5032 /* ok, we can replace it */
5033 block = get_nodes_block(n);
5035 in[0] = get_binop_left(left);
5036 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5038 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5040 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5042 return transform_node(irn);
5043 } /* transform_node_shift */
5046 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5048 * - and, or, xor instead of &
5049 * - Shl, Shr, Shrs, rotl instead of >>
5050 * (with a special case for Or/Xor + Shrs)
5052 static ir_node *transform_node_bitop_shift(ir_node *n) {
5054 ir_node *right = get_binop_right(n);
5055 ir_mode *mode = get_irn_mode(n);
5056 ir_node *bitop_left;
5057 ir_node *bitop_right;
5069 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5071 if (!is_Const(right))
5074 left = get_binop_left(n);
5075 op_left = get_irn_op(left);
5076 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5079 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5080 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5081 /* TODO: test if sign bit is affectes */
5085 bitop_right = get_binop_right(left);
5086 if (!is_Const(bitop_right))
5089 bitop_left = get_binop_left(left);
5091 irg = get_irn_irg(n);
5092 block = get_nodes_block(n);
5093 dbgi = get_irn_dbg_info(n);
5094 tv1 = get_Const_tarval(bitop_right);
5095 tv2 = get_Const_tarval(right);
5097 assert(get_tarval_mode(tv1) == mode);
5100 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5101 tv_shift = tarval_shl(tv1, tv2);
5102 } else if(is_Shr(n)) {
5103 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5104 tv_shift = tarval_shr(tv1, tv2);
5105 } else if(is_Shrs(n)) {
5106 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5107 tv_shift = tarval_shrs(tv1, tv2);
5110 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5111 tv_shift = tarval_rotl(tv1, tv2);
5114 assert(get_tarval_mode(tv_shift) == mode);
5115 new_const = new_Const(mode, tv_shift);
5117 if (op_left == op_And) {
5118 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5119 } else if(op_left == op_Or) {
5120 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5122 assert(op_left == op_Eor);
5123 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5131 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5133 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5134 * (also with x >>s c1 when c1>=c2)
5136 static ir_node *transform_node_shl_shr(ir_node *n) {
5138 ir_node *right = get_binop_right(n);
5154 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5156 if (!is_Const(right))
5159 left = get_binop_left(n);
5160 mode = get_irn_mode(n);
5161 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5162 ir_node *shr_right = get_binop_right(left);
5164 if (!is_Const(shr_right))
5167 x = get_binop_left(left);
5168 tv_shr = get_Const_tarval(shr_right);
5169 tv_shl = get_Const_tarval(right);
5171 if (is_Shrs(left)) {
5172 /* shrs variant only allowed if c1 >= c2 */
5173 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5176 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5179 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5181 tv_mask = tarval_shl(tv_mask, tv_shl);
5182 } else if(is_Shr(n) && is_Shl(left)) {
5183 ir_node *shl_right = get_Shl_right(left);
5185 if (!is_Const(shl_right))
5188 x = get_Shl_left(left);
5189 tv_shr = get_Const_tarval(right);
5190 tv_shl = get_Const_tarval(shl_right);
5192 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5193 tv_mask = tarval_shr(tv_mask, tv_shr);
5198 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5199 assert(tv_mask != tarval_bad);
5200 assert(get_tarval_mode(tv_mask) == mode);
5202 irg = get_irn_irg(n);
5203 block = get_nodes_block(n);
5204 dbgi = get_irn_dbg_info(n);
5206 pnc = tarval_cmp(tv_shl, tv_shr);
5207 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5208 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5209 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5211 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5213 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5216 assert(pnc == pn_Cmp_Gt);
5217 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5218 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5219 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5222 new_const = new_Const(mode, tv_mask);
5223 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5231 static ir_node *transform_node_Shr(ir_node *n) {
5232 ir_node *c, *oldn = n;
5233 ir_node *left = get_Shr_left(n);
5234 ir_node *right = get_Shr_right(n);
5235 ir_mode *mode = get_irn_mode(n);
5237 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5238 n = transform_node_shift(n);
5241 n = transform_node_shl_shr(n);
5243 n = transform_node_bitop_shift(n);
5246 } /* transform_node_Shr */
5251 static ir_node *transform_node_Shrs(ir_node *n) {
5252 ir_node *c, *oldn = n;
5253 ir_node *a = get_Shrs_left(n);
5254 ir_node *b = get_Shrs_right(n);
5255 ir_mode *mode = get_irn_mode(n);
5257 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5258 n = transform_node_shift(n);
5261 n = transform_node_bitop_shift(n);
5264 } /* transform_node_Shrs */
5269 static ir_node *transform_node_Shl(ir_node *n) {
5270 ir_node *c, *oldn = n;
5271 ir_node *a = get_Shl_left(n);
5272 ir_node *b = get_Shl_right(n);
5273 ir_mode *mode = get_irn_mode(n);
5275 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5276 n = transform_node_shift(n);
5279 n = transform_node_shl_shr(n);
5281 n = transform_node_bitop_shift(n);
5284 } /* transform_node_Shl */
5289 static ir_node *transform_node_Rotl(ir_node *n) {
5290 ir_node *c, *oldn = n;
5291 ir_node *a = get_Rotl_left(n);
5292 ir_node *b = get_Rotl_right(n);
5293 ir_mode *mode = get_irn_mode(n);
5295 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5296 n = transform_node_shift(n);
5299 n = transform_node_bitop_shift(n);
5302 } /* transform_node_Rotl */
5307 static ir_node *transform_node_Conv(ir_node *n) {
5308 ir_node *c, *oldn = n;
5309 ir_mode *mode = get_irn_mode(n);
5310 ir_node *a = get_Conv_op(n);
5312 if (mode != mode_b && is_const_Phi(a)) {
5313 /* Do NOT optimize mode_b Conv's, this leads to remaining
5314 * Phib nodes later, because the conv_b_lower operation
5315 * is instantly reverted, when it tries to insert a Convb.
5317 c = apply_conv_on_phi(a, mode);
5319 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5324 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5325 return new_r_Unknown(current_ir_graph, mode);
5328 if (mode_is_reference(mode) &&
5329 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5331 ir_node *l = get_Add_left(a);
5332 ir_node *r = get_Add_right(a);
5333 dbg_info *dbgi = get_irn_dbg_info(a);
5334 ir_node *block = get_nodes_block(n);
5336 ir_node *lop = get_Conv_op(l);
5337 if(get_irn_mode(lop) == mode) {
5338 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5339 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5344 ir_node *rop = get_Conv_op(r);
5345 if(get_irn_mode(rop) == mode) {
5346 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5347 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5354 } /* transform_node_Conv */
5357 * Remove dead blocks and nodes in dead blocks
5358 * in keep alive list. We do not generate a new End node.
5360 static ir_node *transform_node_End(ir_node *n) {
5361 int i, j, n_keepalives = get_End_n_keepalives(n);
5364 NEW_ARR_A(ir_node *, in, n_keepalives);
5366 for (i = j = 0; i < n_keepalives; ++i) {
5367 ir_node *ka = get_End_keepalive(n, i);
5369 if (! is_Block_dead(ka)) {
5373 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5375 } else if (is_Bad(ka)) {
5376 /* no need to keep Bad */
5381 if (j != n_keepalives)
5382 set_End_keepalives(n, j, in);
5384 } /* transform_node_End */
5386 /** returns 1 if a == -b */
5387 static int is_negated_value(ir_node *a, ir_node *b) {
5388 if (is_Minus(a) && get_Minus_op(a) == b)
5390 if (is_Minus(b) && get_Minus_op(b) == a)
5392 if (is_Sub(a) && is_Sub(b)) {
5393 ir_node *a_left = get_Sub_left(a);
5394 ir_node *a_right = get_Sub_right(a);
5395 ir_node *b_left = get_Sub_left(b);
5396 ir_node *b_right = get_Sub_right(b);
5398 if (a_left == b_right && a_right == b_left)
5406 * Optimize a Mux into some simpler cases.
5408 static ir_node *transform_node_Mux(ir_node *n) {
5409 ir_node *oldn = n, *sel = get_Mux_sel(n);
5410 ir_mode *mode = get_irn_mode(n);
5411 ir_node *t = get_Mux_true(n);
5412 ir_node *f = get_Mux_false(n);
5413 ir_graph *irg = current_ir_graph;
5415 /* first normalization step: move a possible zero to the false case */
5417 ir_node *cmp = get_Proj_pred(sel);
5420 if (is_Const(t) && is_Const_null(t)) {
5423 /* Mux(x, 0, y) => Mux(x, y, 0) */
5424 pn_Cmp pnc = get_Proj_proj(sel);
5425 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5426 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5427 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5435 /* note: after normalization, false can only happen on default */
5436 if (mode == mode_b) {
5437 dbg_info *dbg = get_irn_dbg_info(n);
5438 ir_node *block = get_nodes_block(n);
5439 ir_graph *irg = current_ir_graph;
5442 tarval *tv_t = get_Const_tarval(t);
5443 if (tv_t == tarval_b_true) {
5445 /* Muxb(sel, true, false) = sel */
5446 assert(get_Const_tarval(f) == tarval_b_false);
5447 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5450 /* Muxb(sel, true, x) = Or(sel, x) */
5451 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5452 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5456 } else if (is_Const(f)) {
5457 tarval *tv_f = get_Const_tarval(f);
5458 if (tv_f == tarval_b_true) {
5459 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5460 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5461 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5462 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5465 /* Muxb(sel, x, false) = And(sel, x) */
5466 assert(tv_f == tarval_b_false);
5467 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5468 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5474 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5475 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5476 tarval *a = get_Const_tarval(t);
5477 tarval *b = get_Const_tarval(f);
5478 tarval *null = get_tarval_null(mode);
5481 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5482 diff = tarval_sub(a, b, NULL);
5485 diff = tarval_sub(b, a, NULL);
5489 if (diff == get_tarval_one(mode) && min != null) {
5490 dbg_info *dbg = get_irn_dbg_info(n);
5491 ir_node *block = get_nodes_block(n);
5492 ir_graph *irg = current_ir_graph;
5493 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5494 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5495 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5496 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5502 ir_node *cmp = get_Proj_pred(sel);
5503 long pn = get_Proj_proj(sel);
5506 * Note: normalization puts the constant on the right side,
5507 * so we check only one case.
5509 * Note further that these optimization work even for floating point
5510 * with NaN's because -NaN == NaN.
5511 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5515 ir_node *cmp_r = get_Cmp_right(cmp);
5516 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5517 ir_node *block = get_nodes_block(n);
5518 ir_node *cmp_l = get_Cmp_left(cmp);
5520 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5523 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5524 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5526 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5527 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5529 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5531 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5532 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5534 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5535 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5537 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5539 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5544 if (mode_is_int(mode)) {
5546 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5547 /* Mux((a & b) != 0, c, 0) */
5548 ir_node *and_r = get_And_right(cmp_l);
5551 if (and_r == t && f == cmp_r) {
5552 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5553 if (pn == pn_Cmp_Lg) {
5554 /* Mux((a & 2^C) != 0, 2^C, 0) */
5556 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5558 /* Mux((a & 2^C) == 0, 2^C, 0) */
5559 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5560 block, cmp_l, t, mode);
5561 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5566 if (is_Shl(and_r)) {
5567 ir_node *shl_l = get_Shl_left(and_r);
5568 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5569 if (and_r == t && f == cmp_r) {
5570 if (pn == pn_Cmp_Lg) {
5571 /* (a & (1 << n)) != 0, (1 << n), 0) */
5573 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5575 /* (a & (1 << n)) == 0, (1 << n), 0) */
5576 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5577 block, cmp_l, t, mode);
5578 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5584 and_l = get_And_left(cmp_l);
5585 if (is_Shl(and_l)) {
5586 ir_node *shl_l = get_Shl_left(and_l);
5587 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5588 if (and_l == t && f == cmp_r) {
5589 if (pn == pn_Cmp_Lg) {
5590 /* ((1 << n) & a) != 0, (1 << n), 0) */
5592 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5594 /* ((1 << n) & a) == 0, (1 << n), 0) */
5595 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5596 block, cmp_l, t, mode);
5597 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5608 return arch_transform_node_Mux(n);
5609 } /* transform_node_Mux */
5612 * optimize Sync nodes that have other syncs as input we simply add the inputs
5613 * of the other sync to our own inputs
5615 static ir_node *transform_node_Sync(ir_node *n) {
5616 int arity = get_Sync_n_preds(n);
5619 for (i = 0; i < arity;) {
5620 ir_node *pred = get_Sync_pred(n, i);
5624 if (!is_Sync(pred)) {
5632 pred_arity = get_Sync_n_preds(pred);
5633 for (j = 0; j < pred_arity; ++j) {
5634 ir_node *pred_pred = get_Sync_pred(pred, j);
5639 add_irn_n(n, pred_pred);
5643 if (get_Sync_pred(n, k) == pred_pred) break;
5648 /* rehash the sync node */
5649 add_identities(current_ir_graph->value_table, n);
5655 * Tries several [inplace] [optimizing] transformations and returns an
5656 * equivalent node. The difference to equivalent_node() is that these
5657 * transformations _do_ generate new nodes, and thus the old node must
5658 * not be freed even if the equivalent node isn't the old one.
5660 static ir_node *transform_node(ir_node *n) {
5664 * Transform_node is the only "optimizing transformation" that might
5665 * return a node with a different opcode. We iterate HERE until fixpoint
5666 * to get the final result.
5670 if (n->op->ops.transform_node)
5671 n = n->op->ops.transform_node(n);
5672 } while (oldn != n);
5675 } /* transform_node */
5678 * Sets the default transform node operation for an ir_op_ops.
5680 * @param code the opcode for the default operation
5681 * @param ops the operations initialized
5686 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5690 ops->transform_node = transform_node_##a; \
5692 #define CASE_PROJ(a) \
5694 ops->transform_node_Proj = transform_node_Proj_##a; \
5696 #define CASE_PROJ_EX(a) \
5698 ops->transform_node = transform_node_##a; \
5699 ops->transform_node_Proj = transform_node_Proj_##a; \
5708 CASE_PROJ_EX(DivMod);
5742 } /* firm_set_default_transform_node */
5745 /* **************** Common Subexpression Elimination **************** */
5747 /** The size of the hash table used, should estimate the number of nodes
5749 #define N_IR_NODES 512
5751 /** Compares the attributes of two Const nodes. */
5752 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5753 return (get_Const_tarval(a) != get_Const_tarval(b))
5754 || (get_Const_type(a) != get_Const_type(b));
5755 } /* node_cmp_attr_Const */
5757 /** Compares the attributes of two Proj nodes. */
5758 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5759 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5760 } /* node_cmp_attr_Proj */
5762 /** Compares the attributes of two Filter nodes. */
5763 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5764 return get_Filter_proj(a) != get_Filter_proj(b);
5765 } /* node_cmp_attr_Filter */
5767 /** Compares the attributes of two Alloc nodes. */
5768 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5769 const alloc_attr *pa = get_irn_alloc_attr(a);
5770 const alloc_attr *pb = get_irn_alloc_attr(b);
5771 return (pa->where != pb->where) || (pa->type != pb->type);
5772 } /* node_cmp_attr_Alloc */
5774 /** Compares the attributes of two Free nodes. */
5775 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5776 const free_attr *pa = get_irn_free_attr(a);
5777 const free_attr *pb = get_irn_free_attr(b);
5778 return (pa->where != pb->where) || (pa->type != pb->type);
5779 } /* node_cmp_attr_Free */
5781 /** Compares the attributes of two SymConst nodes. */
5782 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5783 const symconst_attr *pa = get_irn_symconst_attr(a);
5784 const symconst_attr *pb = get_irn_symconst_attr(b);
5785 return (pa->kind != pb->kind)
5786 || (pa->sym.type_p != pb->sym.type_p)
5787 || (pa->tp != pb->tp);
5788 } /* node_cmp_attr_SymConst */
5790 /** Compares the attributes of two Call nodes. */
5791 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5792 return get_irn_call_attr(a) != get_irn_call_attr(b);
5793 } /* node_cmp_attr_Call */
5795 /** Compares the attributes of two Sel nodes. */
5796 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5797 const ir_entity *a_ent = get_Sel_entity(a);
5798 const ir_entity *b_ent = get_Sel_entity(b);
5801 (a_ent->kind != b_ent->kind) ||
5802 (a_ent->name != b_ent->name) ||
5803 (a_ent->owner != b_ent->owner) ||
5804 (a_ent->ld_name != b_ent->ld_name) ||
5805 (a_ent->type != b_ent->type);
5807 /* Matze: inlining of functions can produce 2 entities with same type,
5809 return a_ent != b_ent;
5810 } /* node_cmp_attr_Sel */
5812 /** Compares the attributes of two Phi nodes. */
5813 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5814 /* we can only enter this function if both nodes have the same number of inputs,
5815 hence it is enough to check if one of them is a Phi0 */
5817 /* check the Phi0 pos attribute */
5818 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5821 } /* node_cmp_attr_Phi */
5823 /** Compares the attributes of two Conv nodes. */
5824 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5825 return get_Conv_strict(a) != get_Conv_strict(b);
5826 } /* node_cmp_attr_Conv */
5828 /** Compares the attributes of two Cast nodes. */
5829 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5830 return get_Cast_type(a) != get_Cast_type(b);
5831 } /* node_cmp_attr_Cast */
5833 /** Compares the attributes of two Load nodes. */
5834 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5835 if (get_Load_volatility(a) == volatility_is_volatile ||
5836 get_Load_volatility(b) == volatility_is_volatile)
5837 /* NEVER do CSE on volatile Loads */
5839 /* do not CSE Loads with different alignment. Be conservative. */
5840 if (get_Load_align(a) != get_Load_align(b))
5843 return get_Load_mode(a) != get_Load_mode(b);
5844 } /* node_cmp_attr_Load */
5846 /** Compares the attributes of two Store nodes. */
5847 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5848 /* do not CSE Stores with different alignment. Be conservative. */
5849 if (get_Store_align(a) != get_Store_align(b))
5852 /* NEVER do CSE on volatile Stores */
5853 return (get_Store_volatility(a) == volatility_is_volatile ||
5854 get_Store_volatility(b) == volatility_is_volatile);
5855 } /* node_cmp_attr_Store */
5857 /** Compares two exception attributes */
5858 static int node_cmp_exception(ir_node *a, ir_node *b) {
5859 const except_attr *ea = get_irn_except_attr(a);
5860 const except_attr *eb = get_irn_except_attr(b);
5862 return ea->pin_state != eb->pin_state;
5865 #define node_cmp_attr_Bound node_cmp_exception
5867 /** Compares the attributes of two Div nodes. */
5868 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5869 const divmod_attr *ma = get_irn_divmod_attr(a);
5870 const divmod_attr *mb = get_irn_divmod_attr(b);
5871 return ma->exc.pin_state != mb->exc.pin_state ||
5872 ma->res_mode != mb->res_mode ||
5873 ma->no_remainder != mb->no_remainder;
5874 } /* node_cmp_attr_Div */
5876 /** Compares the attributes of two DivMod nodes. */
5877 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5878 const divmod_attr *ma = get_irn_divmod_attr(a);
5879 const divmod_attr *mb = get_irn_divmod_attr(b);
5880 return ma->exc.pin_state != mb->exc.pin_state ||
5881 ma->res_mode != mb->res_mode;
5882 } /* node_cmp_attr_DivMod */
5884 /** Compares the attributes of two Mod nodes. */
5885 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5886 const divmod_attr *ma = get_irn_divmod_attr(a);
5887 const divmod_attr *mb = get_irn_divmod_attr(b);
5888 return ma->exc.pin_state != mb->exc.pin_state ||
5889 ma->res_mode != mb->res_mode;
5890 } /* node_cmp_attr_Mod */
5892 /** Compares the attributes of two Quot nodes. */
5893 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5894 const divmod_attr *ma = get_irn_divmod_attr(a);
5895 const divmod_attr *mb = get_irn_divmod_attr(b);
5896 return ma->exc.pin_state != mb->exc.pin_state ||
5897 ma->res_mode != mb->res_mode;
5898 } /* node_cmp_attr_Quot */
5900 /** Compares the attributes of two Confirm nodes. */
5901 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5902 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5903 } /* node_cmp_attr_Confirm */
5905 /** Compares the attributes of two ASM nodes. */
5906 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5908 const ir_asm_constraint *ca;
5909 const ir_asm_constraint *cb;
5912 if (get_ASM_text(a) != get_ASM_text(b))
5915 /* Should we really check the constraints here? Should be better, but is strange. */
5916 n = get_ASM_n_input_constraints(a);
5917 if (n != get_ASM_n_input_constraints(b))
5920 ca = get_ASM_input_constraints(a);
5921 cb = get_ASM_input_constraints(b);
5922 for (i = 0; i < n; ++i) {
5923 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5927 n = get_ASM_n_output_constraints(a);
5928 if (n != get_ASM_n_output_constraints(b))
5931 ca = get_ASM_output_constraints(a);
5932 cb = get_ASM_output_constraints(b);
5933 for (i = 0; i < n; ++i) {
5934 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5938 n = get_ASM_n_clobbers(a);
5939 if (n != get_ASM_n_clobbers(b))
5942 cla = get_ASM_clobbers(a);
5943 clb = get_ASM_clobbers(b);
5944 for (i = 0; i < n; ++i) {
5945 if (cla[i] != clb[i])
5949 } /* node_cmp_attr_ASM */
5952 * Set the default node attribute compare operation for an ir_op_ops.
5954 * @param code the opcode for the default operation
5955 * @param ops the operations initialized
5960 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
5964 ops->node_cmp_attr = node_cmp_attr_##a; \
5995 } /* firm_set_default_node_cmp_attr */
5998 * Compare function for two nodes in the value table. Gets two
5999 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6001 int identities_cmp(const void *elt, const void *key) {
6002 ir_node *a = (ir_node *)elt;
6003 ir_node *b = (ir_node *)key;
6006 if (a == b) return 0;
6008 if ((get_irn_op(a) != get_irn_op(b)) ||
6009 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6011 /* compare if a's in and b's in are of equal length */
6012 irn_arity_a = get_irn_intra_arity(a);
6013 if (irn_arity_a != get_irn_intra_arity(b))
6016 if (get_irn_pinned(a) == op_pin_state_pinned) {
6017 /* for pinned nodes, the block inputs must be equal */
6018 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6020 } else if (! get_opt_global_cse()) {
6021 /* for block-local CSE both nodes must be in the same MacroBlock */
6022 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6026 /* compare a->in[0..ins] with b->in[0..ins] */
6027 for (i = 0; i < irn_arity_a; i++)
6028 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6032 * here, we already now that the nodes are identical except their
6035 if (a->op->ops.node_cmp_attr)
6036 return a->op->ops.node_cmp_attr(a, b);
6039 } /* identities_cmp */
6042 * Calculate a hash value of a node.
6044 * @param node The IR-node
6046 unsigned ir_node_hash(const ir_node *node) {
6047 return node->op->ops.hash(node);
6048 } /* ir_node_hash */
6051 pset *new_identities(void) {
6052 return new_pset(identities_cmp, N_IR_NODES);
6053 } /* new_identities */
6055 void del_identities(pset *value_table) {
6056 del_pset(value_table);
6057 } /* del_identities */
6059 /* Normalize a node by putting constants (and operands with larger
6060 * node index) on the right (operator side). */
6061 void ir_normalize_node(ir_node *n) {
6062 if (is_op_commutative(get_irn_op(n))) {
6063 ir_node *l = get_binop_left(n);
6064 ir_node *r = get_binop_right(n);
6066 /* For commutative operators perform a OP b == b OP a but keep
6067 * constants on the RIGHT side. This helps greatly in some
6068 * optimizations. Moreover we use the idx number to make the form
6070 if (!operands_are_normalized(l, r)) {
6071 set_binop_left(n, r);
6072 set_binop_right(n, l);
6076 } /* ir_normalize_node */
6079 * Update the nodes after a match in the value table. If both nodes have
6080 * the same MacroBlock but different Blocks, we must ensure that the node
6081 * with the dominating Block (the node that is near to the MacroBlock header
6082 * is stored in the table.
6083 * Because a MacroBlock has only one "non-exception" flow, we don't need
6084 * dominance info here: We known, that one block must dominate the other and
6085 * following the only block input will allow to find it.
6087 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6088 ir_node *known_blk, *new_block, *block, *mbh;
6090 if (get_opt_global_cse()) {
6091 /* Block inputs are meaning less */
6094 known_blk = get_irn_n(known_irn, -1);
6095 new_block = get_irn_n(new_ir_node, -1);
6096 if (known_blk == new_block) {
6097 /* already in the same block */
6101 * We expect the typical case when we built the graph. In that case, the
6102 * known_irn is already the upper one, so checking this should be faster.
6105 mbh = get_Block_MacroBlock(new_block);
6107 if (block == known_blk) {
6108 /* ok, we have found it: known_block dominates new_block as expected */
6113 * We have reached the MacroBlock header NOT founding
6114 * the known_block. new_block must dominate known_block.
6117 set_irn_n(known_irn, -1, new_block);
6120 assert(get_Block_n_cfgpreds(block) == 1);
6121 block = get_Block_cfgpred_block(block, 0);
6123 } /* update_value_table */
6126 * Return the canonical node computing the same value as n.
6127 * Looks up the node in a hash table, enters it in the table
6128 * if it isn't there yet.
6130 * @param value_table the HashSet containing all nodes in the
6132 * @param n the node to look up
6134 * @return a node that computes the same value as n or n if no such
6135 * node could be found
6137 ir_node *identify_remember(pset *value_table, ir_node *n) {
6140 if (!value_table) return n;
6142 ir_normalize_node(n);
6143 /* lookup or insert in hash table with given hash key. */
6144 o = pset_insert(value_table, n, ir_node_hash(n));
6147 update_known_irn(o, n);
6151 } /* identify_remember */
6154 * During construction we set the op_pin_state_pinned flag in the graph right when the
6155 * optimization is performed. The flag turning on procedure global cse could
6156 * be changed between two allocations. This way we are safe.
6158 * @param value_table The value table
6159 * @param n The node to lookup
6161 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6164 n = identify_remember(value_table, n);
6165 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6166 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6168 } /* identify_cons */
6170 /* Add a node to the identities value table. */
6171 void add_identities(pset *value_table, ir_node *node) {
6172 if (get_opt_cse() && is_no_Block(node))
6173 identify_remember(value_table, node);
6174 } /* add_identities */
6176 /* Visit each node in the value table of a graph. */
6177 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6179 ir_graph *rem = current_ir_graph;
6181 current_ir_graph = irg;
6182 foreach_pset(irg->value_table, node)
6184 current_ir_graph = rem;
6185 } /* visit_all_identities */
6188 * Garbage in, garbage out. If a node has a dead input, i.e., the
6189 * Bad node is input to the node, return the Bad node.
6191 static ir_node *gigo(ir_node *node) {
6193 ir_op *op = get_irn_op(node);
6195 /* remove garbage blocks by looking at control flow that leaves the block
6196 and replacing the control flow by Bad. */
6197 if (get_irn_mode(node) == mode_X) {
6198 ir_node *block = get_nodes_block(skip_Proj(node));
6200 /* Don't optimize nodes in immature blocks. */
6201 if (!get_Block_matured(block))
6203 /* Don't optimize End, may have Bads. */
6204 if (op == op_End) return node;
6206 if (is_Block(block)) {
6207 if (is_Block_dead(block)) {
6208 /* control flow from dead block is dead */
6212 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6213 if (!is_Bad(get_irn_n(block, i)))
6217 ir_graph *irg = get_irn_irg(block);
6218 /* the start block is never dead */
6219 if (block != get_irg_start_block(irg)
6220 && block != get_irg_end_block(irg)) {
6222 * Do NOT kill control flow without setting
6223 * the block to dead of bad things can happen:
6224 * We get a Block that is not reachable be irg_block_walk()
6225 * but can be found by irg_walk()!
6227 set_Block_dead(block);
6234 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6235 blocks predecessors is dead. */
6236 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6237 irn_arity = get_irn_arity(node);
6240 * Beware: we can only read the block of a non-floating node.
6242 if (is_irn_pinned_in_irg(node) &&
6243 is_Block_dead(get_nodes_block(skip_Proj(node))))
6246 for (i = 0; i < irn_arity; i++) {
6247 ir_node *pred = get_irn_n(node, i);
6252 /* Propagating Unknowns here seems to be a bad idea, because
6253 sometimes we need a node as a input and did not want that
6255 However, it might be useful to move this into a later phase
6256 (if you think that optimizing such code is useful). */
6257 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6258 return new_Unknown(get_irn_mode(node));
6263 /* With this code we violate the agreement that local_optimize
6264 only leaves Bads in Block, Phi and Tuple nodes. */
6265 /* If Block has only Bads as predecessors it's garbage. */
6266 /* If Phi has only Bads as predecessors it's garbage. */
6267 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6268 irn_arity = get_irn_arity(node);
6269 for (i = 0; i < irn_arity; i++) {
6270 if (!is_Bad(get_irn_n(node, i))) break;
6272 if (i == irn_arity) node = new_Bad();
6279 * These optimizations deallocate nodes from the obstack.
6280 * It can only be called if it is guaranteed that no other nodes
6281 * reference this one, i.e., right after construction of a node.
6283 * @param n The node to optimize
6285 * current_ir_graph must be set to the graph of the node!
6287 ir_node *optimize_node(ir_node *n) {
6290 ir_opcode iro = get_irn_opcode(n);
6292 /* Always optimize Phi nodes: part of the construction. */
6293 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6295 /* constant expression evaluation / constant folding */
6296 if (get_opt_constant_folding()) {
6297 /* neither constants nor Tuple values can be evaluated */
6298 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6299 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6300 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6301 /* try to evaluate */
6302 tv = computed_value(n);
6303 if (tv != tarval_bad) {
6305 ir_type *old_tp = get_irn_type(n);
6306 int i, arity = get_irn_arity(n);
6310 * Try to recover the type of the new expression.
6312 for (i = 0; i < arity && !old_tp; ++i)
6313 old_tp = get_irn_type(get_irn_n(n, i));
6316 * we MUST copy the node here temporary, because it's still needed
6317 * for DBG_OPT_CSTEVAL
6319 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6320 oldn = alloca(node_size);
6322 memcpy(oldn, n, node_size);
6323 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6325 /* ARG, copy the in array, we need it for statistics */
6326 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6328 /* note the inplace edges module */
6329 edges_node_deleted(n, current_ir_graph);
6331 /* evaluation was successful -- replace the node. */
6332 irg_kill_node(current_ir_graph, n);
6333 nw = new_Const(get_tarval_mode(tv), tv);
6335 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6336 set_Const_type(nw, old_tp);
6337 DBG_OPT_CSTEVAL(oldn, nw);
6338 tarval_enable_fp_ops(old_fp_mode);
6341 tarval_enable_fp_ops(old_fp_mode);
6345 /* remove unnecessary nodes */
6346 if (get_opt_algebraic_simplification() ||
6347 (iro == iro_Phi) || /* always optimize these nodes. */
6349 (iro == iro_Proj) ||
6350 (iro == iro_Block) ) /* Flags tested local. */
6351 n = equivalent_node(n);
6353 /* Common Subexpression Elimination.
6355 * Checks whether n is already available.
6356 * The block input is used to distinguish different subexpressions. Right
6357 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6358 * subexpressions within a block.
6361 n = identify_cons(current_ir_graph->value_table, n);
6364 edges_node_deleted(oldn, current_ir_graph);
6366 /* We found an existing, better node, so we can deallocate the old node. */
6367 irg_kill_node(current_ir_graph, oldn);
6371 /* Some more constant expression evaluation that does not allow to
6373 iro = get_irn_opcode(n);
6374 if (get_opt_algebraic_simplification() ||
6375 (iro == iro_Cond) ||
6376 (iro == iro_Proj)) /* Flags tested local. */
6377 n = transform_node(n);
6379 /* Remove nodes with dead (Bad) input.
6380 Run always for transformation induced Bads. */
6383 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6384 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6386 n = identify_remember(current_ir_graph->value_table, o);
6392 } /* optimize_node */
6396 * These optimizations never deallocate nodes (in place). This can cause dead
6397 * nodes lying on the obstack. Remove these by a dead node elimination,
6398 * i.e., a copying garbage collection.
6400 ir_node *optimize_in_place_2(ir_node *n) {
6403 ir_opcode iro = get_irn_opcode(n);
6405 if (!get_opt_optimize() && !is_Phi(n)) return n;
6407 /* constant expression evaluation / constant folding */
6408 if (get_opt_constant_folding()) {
6409 /* neither constants nor Tuple values can be evaluated */
6410 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6411 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6412 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6413 /* try to evaluate */
6414 tv = computed_value(n);
6415 if (tv != tarval_bad) {
6416 /* evaluation was successful -- replace the node. */
6417 ir_type *old_tp = get_irn_type(n);
6418 int i, arity = get_irn_arity(n);
6421 * Try to recover the type of the new expression.
6423 for (i = 0; i < arity && !old_tp; ++i)
6424 old_tp = get_irn_type(get_irn_n(n, i));
6426 n = new_Const(get_tarval_mode(tv), tv);
6428 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6429 set_Const_type(n, old_tp);
6431 DBG_OPT_CSTEVAL(oldn, n);
6432 tarval_enable_fp_ops(old_fp_mode);
6435 tarval_enable_fp_ops(old_fp_mode);
6439 /* remove unnecessary nodes */
6440 if (get_opt_constant_folding() ||
6441 (iro == iro_Phi) || /* always optimize these nodes. */
6442 (iro == iro_Id) || /* ... */
6443 (iro == iro_Proj) || /* ... */
6444 (iro == iro_Block) ) /* Flags tested local. */
6445 n = equivalent_node(n);
6447 /** common subexpression elimination **/
6448 /* Checks whether n is already available. */
6449 /* The block input is used to distinguish different subexpressions. Right
6450 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6451 subexpressions within a block. */
6452 if (get_opt_cse()) {
6454 n = identify_remember(current_ir_graph->value_table, o);
6459 /* Some more constant expression evaluation. */
6460 iro = get_irn_opcode(n);
6461 if (get_opt_constant_folding() ||
6462 (iro == iro_Cond) ||
6463 (iro == iro_Proj)) /* Flags tested local. */
6464 n = transform_node(n);
6466 /* Remove nodes with dead (Bad) input.
6467 Run always for transformation induced Bads. */
6470 /* Now we can verify the node, as it has no dead inputs any more. */
6473 /* Now we have a legal, useful node. Enter it in hash table for cse.
6474 Blocks should be unique anyways. (Except the successor of start:
6475 is cse with the start block!) */
6476 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6478 n = identify_remember(current_ir_graph->value_table, o);
6484 } /* optimize_in_place_2 */
6487 * Wrapper for external use, set proper status bits after optimization.
6489 ir_node *optimize_in_place(ir_node *n) {
6490 /* Handle graph state */
6491 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6493 if (get_opt_global_cse())
6494 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6495 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6496 set_irg_outs_inconsistent(current_ir_graph);
6498 /* FIXME: Maybe we could also test whether optimizing the node can
6499 change the control graph. */
6500 set_irg_doms_inconsistent(current_ir_graph);
6501 return optimize_in_place_2(n);
6502 } /* optimize_in_place */
6505 * Calculate a hash value of a Const node.
6507 static unsigned hash_Const(const ir_node *node) {
6510 /* special value for const, as they only differ in their tarval. */
6511 h = HASH_PTR(node->attr.con.tv);
6512 h = 9*h + HASH_PTR(get_irn_mode(node));
6518 * Calculate a hash value of a SymConst node.
6520 static unsigned hash_SymConst(const ir_node *node) {
6523 /* special value for const, as they only differ in their symbol. */
6524 h = HASH_PTR(node->attr.symc.sym.type_p);
6525 h = 9*h + HASH_PTR(get_irn_mode(node));
6528 } /* hash_SymConst */
6531 * Set the default hash operation in an ir_op_ops.
6533 * @param code the opcode for the default operation
6534 * @param ops the operations initialized
6539 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6543 ops->hash = hash_##a; \
6546 /* hash function already set */
6547 if (ops->hash != NULL)
6554 /* use input/mode default hash if no function was given */
6555 ops->hash = firm_default_hash;
6563 * Sets the default operation for an ir_ops.
6565 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6566 ops = firm_set_default_hash(code, ops);
6567 ops = firm_set_default_computed_value(code, ops);
6568 ops = firm_set_default_equivalent_node(code, ops);
6569 ops = firm_set_default_transform_node(code, ops);
6570 ops = firm_set_default_node_cmp_attr(code, ops);
6571 ops = firm_set_default_get_type(code, ops);
6572 ops = firm_set_default_get_type_attr(code, ops);
6573 ops = firm_set_default_get_entity_attr(code, ops);
6576 } /* firm_set_default_operations */