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)) {
1219 /* neither Minus nor Abs nor Confirm change the precision,
1220 so we can "look-through" */
1223 p = get_Minus_op(p);
1224 } else if (is_Abs(p)) {
1226 } else if (is_Confirm(p)) {
1227 p = get_Confirm_value(p);
1233 if (is_Conv(p) && get_Conv_strict(p)) {
1234 /* we known already, that a_mode == n_mode, and neither
1235 Abs nor Minus change the mode, so the second Conv
1237 assert(get_irn_mode(p) == n_mode);
1239 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1243 ir_node *pred = get_Proj_pred(p);
1244 if (is_Load(pred)) {
1245 /* Loads always return with the exact precision of n_mode */
1246 assert(get_Load_mode(pred) == n_mode);
1248 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1251 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1252 pred = get_Proj_pred(pred);
1253 if (is_Start(pred)) {
1254 /* Arguments always return with the exact precision,
1255 as strictConv's are place before Call -- if the
1256 caller was compiled with the same setting.
1257 Otherwise, the semantics is probably still right. */
1258 assert(get_irn_mode(p) == n_mode);
1260 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1266 /* special case: the immediate predecessor is also a Conv */
1267 if (! get_Conv_strict(a)) {
1268 /* first one is not strict, kick it */
1270 a_mode = get_irn_mode(a);
1274 /* else both are strict conv, second is superfluous */
1276 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1279 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1280 ir_node *b = get_Conv_op(a);
1281 ir_mode *b_mode = get_irn_mode(b);
1283 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1284 /* both are strict conv */
1285 if (smaller_mode(a_mode, n_mode)) {
1286 /* both are strict, but the first is smaller, so
1287 the second cannot remove more precision, remove the
1289 set_Conv_strict(n, 0);
1292 if (n_mode == b_mode) {
1293 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1294 if (n_mode == mode_b) {
1295 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1296 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1298 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1299 if (values_in_mode(b_mode, a_mode)) {
1300 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1301 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1306 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1307 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1308 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1309 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1311 if (float_mantissa >= int_mantissa) {
1313 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1318 if (smaller_mode(b_mode, a_mode)) {
1319 if (get_Conv_strict(n))
1320 set_Conv_strict(b, 1);
1321 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1322 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1329 } /* equivalent_node_Conv */
1332 * A Cast may be removed if the type of the previous node
1333 * is already the type of the Cast.
1335 static ir_node *equivalent_node_Cast(ir_node *n) {
1337 ir_node *pred = get_Cast_op(n);
1339 if (get_irn_type(pred) == get_Cast_type(n)) {
1341 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1344 } /* equivalent_node_Cast */
1347 * - fold Phi-nodes, iff they have only one predecessor except
1350 static ir_node *equivalent_node_Phi(ir_node *n) {
1355 ir_node *first_val = NULL; /* to shutup gcc */
1357 if (!get_opt_normalize()) return n;
1359 n_preds = get_Phi_n_preds(n);
1361 block = get_nodes_block(n);
1362 if (is_Block_dead(block)) /* Control dead */
1363 return get_irg_bad(current_ir_graph);
1365 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1367 /* Find first non-self-referencing input */
1368 for (i = 0; i < n_preds; ++i) {
1369 first_val = get_Phi_pred(n, i);
1370 if ( (first_val != n) /* not self pointer */
1372 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1373 * predecessors. Then, Phi nodes in dead code might be removed, causing
1374 * nodes pointing to themself (Add's for instance).
1375 * This is really bad and causes endless recursions in several
1376 * code pathes, so we do NOT optimize such a code.
1377 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1378 * (and bad Phi predecessors), so live code is optimized later.
1380 && (! is_Bad(get_Block_cfgpred(block, i)))
1382 ) { /* value not dead */
1383 break; /* then found first value. */
1388 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1389 return get_irg_bad(current_ir_graph);
1392 /* search for rest of inputs, determine if any of these
1393 are non-self-referencing */
1394 while (++i < n_preds) {
1395 ir_node *scnd_val = get_Phi_pred(n, i);
1396 if ( (scnd_val != n)
1397 && (scnd_val != first_val)
1400 && (! is_Bad(get_Block_cfgpred(block, i)))
1408 /* Fold, if no multiple distinct non-self-referencing inputs */
1410 DBG_OPT_PHI(oldn, n);
1413 } /* equivalent_node_Phi */
1416 * Several optimizations:
1417 * - fold Sync-nodes, iff they have only one predecessor except
1420 static ir_node *equivalent_node_Sync(ir_node *n) {
1421 int arity = get_Sync_n_preds(n);
1424 for (i = 0; i < arity;) {
1425 ir_node *pred = get_Sync_pred(n, i);
1428 /* Remove Bad predecessors */
1435 /* Remove duplicate predecessors */
1441 if (get_Sync_pred(n, j) == pred) {
1449 if (arity == 0) return get_irg_bad(current_ir_graph);
1450 if (arity == 1) return get_Sync_pred(n, 0);
1452 } /* equivalent_node_Sync */
1455 * Optimize Proj(Tuple).
1457 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1458 ir_node *oldn = proj;
1459 ir_node *tuple = get_Proj_pred(proj);
1461 /* Remove the Tuple/Proj combination. */
1462 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1463 DBG_OPT_TUPLE(oldn, tuple, proj);
1466 } /* equivalent_node_Proj_Tuple */
1469 * Optimize a / 1 = a.
1471 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1472 ir_node *oldn = proj;
1473 ir_node *div = get_Proj_pred(proj);
1474 ir_node *b = get_Div_right(div);
1475 tarval *tb = value_of(b);
1477 /* Div is not commutative. */
1478 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1479 switch (get_Proj_proj(proj)) {
1481 proj = get_Div_mem(div);
1482 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1486 proj = get_Div_left(div);
1487 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1491 /* we cannot replace the exception Proj's here, this is done in
1492 transform_node_Proj_Div() */
1497 } /* equivalent_node_Proj_Div */
1500 * Optimize a / 1.0 = a.
1502 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1503 ir_node *oldn = proj;
1504 ir_node *quot = get_Proj_pred(proj);
1505 ir_node *b = get_Quot_right(quot);
1506 tarval *tb = value_of(b);
1508 /* Div is not commutative. */
1509 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1510 switch (get_Proj_proj(proj)) {
1512 proj = get_Quot_mem(quot);
1513 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1517 proj = get_Quot_left(quot);
1518 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1522 /* we cannot replace the exception Proj's here, this is done in
1523 transform_node_Proj_Quot() */
1528 } /* equivalent_node_Proj_Quot */
1531 * Optimize a / 1 = a.
1533 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1534 ir_node *oldn = proj;
1535 ir_node *divmod = get_Proj_pred(proj);
1536 ir_node *b = get_DivMod_right(divmod);
1537 tarval *tb = value_of(b);
1539 /* Div is not commutative. */
1540 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1541 switch (get_Proj_proj(proj)) {
1543 proj = get_DivMod_mem(divmod);
1544 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1547 case pn_DivMod_res_div:
1548 proj = get_DivMod_left(divmod);
1549 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1553 /* we cannot replace the exception Proj's here, this is done in
1554 transform_node_Proj_DivMod().
1555 Note further that the pn_DivMod_res_div case is handled in
1556 computed_value_Proj(). */
1561 } /* equivalent_node_Proj_DivMod */
1564 * Optimize CopyB(mem, x, x) into a Nop.
1566 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1567 ir_node *oldn = proj;
1568 ir_node *copyb = get_Proj_pred(proj);
1569 ir_node *a = get_CopyB_dst(copyb);
1570 ir_node *b = get_CopyB_src(copyb);
1573 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1574 switch (get_Proj_proj(proj)) {
1575 case pn_CopyB_M_regular:
1576 proj = get_CopyB_mem(copyb);
1577 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1580 case pn_CopyB_M_except:
1581 case pn_CopyB_X_except:
1582 DBG_OPT_EXC_REM(proj);
1583 proj = get_irg_bad(current_ir_graph);
1588 } /* equivalent_node_Proj_CopyB */
1591 * Optimize Bounds(idx, idx, upper) into idx.
1593 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1594 ir_node *oldn = proj;
1595 ir_node *bound = get_Proj_pred(proj);
1596 ir_node *idx = get_Bound_index(bound);
1597 ir_node *pred = skip_Proj(idx);
1600 if (idx == get_Bound_lower(bound))
1602 else if (is_Bound(pred)) {
1604 * idx was Bounds checked in the same MacroBlock previously,
1605 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1607 ir_node *lower = get_Bound_lower(bound);
1608 ir_node *upper = get_Bound_upper(bound);
1609 if (get_Bound_lower(pred) == lower &&
1610 get_Bound_upper(pred) == upper &&
1611 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1613 * One could expect that we simply return the previous
1614 * Bound here. However, this would be wrong, as we could
1615 * add an exception Proj to a new location then.
1616 * So, we must turn in into a tuple.
1622 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1623 switch (get_Proj_proj(proj)) {
1625 DBG_OPT_EXC_REM(proj);
1626 proj = get_Bound_mem(bound);
1628 case pn_Bound_X_except:
1629 DBG_OPT_EXC_REM(proj);
1630 proj = get_irg_bad(current_ir_graph);
1634 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1637 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1642 } /* equivalent_node_Proj_Bound */
1645 * Optimize an Exception Proj(Load) with a non-null address.
1647 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1648 if (get_opt_ldst_only_null_ptr_exceptions()) {
1649 if (get_irn_mode(proj) == mode_X) {
1650 ir_node *load = get_Proj_pred(proj);
1652 /* get the Load address */
1653 const ir_node *addr = get_Load_ptr(load);
1654 const ir_node *confirm;
1656 if (value_not_null(addr, &confirm)) {
1657 if (get_Proj_proj(proj) == pn_Load_X_except) {
1658 DBG_OPT_EXC_REM(proj);
1659 return get_irg_bad(current_ir_graph);
1665 } /* equivalent_node_Proj_Load */
1668 * Optimize an Exception Proj(Store) with a non-null address.
1670 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1671 if (get_opt_ldst_only_null_ptr_exceptions()) {
1672 if (get_irn_mode(proj) == mode_X) {
1673 ir_node *store = get_Proj_pred(proj);
1675 /* get the load/store address */
1676 const ir_node *addr = get_Store_ptr(store);
1677 const ir_node *confirm;
1679 if (value_not_null(addr, &confirm)) {
1680 if (get_Proj_proj(proj) == pn_Store_X_except) {
1681 DBG_OPT_EXC_REM(proj);
1682 return get_irg_bad(current_ir_graph);
1688 } /* equivalent_node_Proj_Store */
1691 * Does all optimizations on nodes that must be done on it's Proj's
1692 * because of creating new nodes.
1694 static ir_node *equivalent_node_Proj(ir_node *proj) {
1695 ir_node *n = get_Proj_pred(proj);
1697 if (get_irn_mode(proj) == mode_X) {
1698 if (is_Block_dead(get_nodes_block(n))) {
1699 /* Remove dead control flow -- early gigo(). */
1700 return get_irg_bad(current_ir_graph);
1703 if (n->op->ops.equivalent_node_Proj)
1704 return n->op->ops.equivalent_node_Proj(proj);
1706 } /* equivalent_node_Proj */
1711 static ir_node *equivalent_node_Id(ir_node *n) {
1718 DBG_OPT_ID(oldn, n);
1720 } /* equivalent_node_Id */
1725 static ir_node *equivalent_node_Mux(ir_node *n)
1727 ir_node *oldn = n, *sel = get_Mux_sel(n);
1728 tarval *ts = value_of(sel);
1730 /* Mux(true, f, t) == t */
1731 if (ts == tarval_b_true) {
1732 n = get_Mux_true(n);
1733 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1735 /* Mux(false, f, t) == f */
1736 else if (ts == tarval_b_false) {
1737 n = get_Mux_false(n);
1738 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1740 /* Mux(v, x, x) == x */
1741 else if (get_Mux_false(n) == get_Mux_true(n)) {
1742 n = get_Mux_true(n);
1743 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1745 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1746 ir_node *cmp = get_Proj_pred(sel);
1747 long proj_nr = get_Proj_proj(sel);
1748 ir_node *f = get_Mux_false(n);
1749 ir_node *t = get_Mux_true(n);
1752 * Note further that these optimization work even for floating point
1753 * with NaN's because -NaN == NaN.
1754 * However, if +0 and -0 is handled differently, we cannot use the first one.
1757 ir_node *const cmp_l = get_Cmp_left(cmp);
1758 ir_node *const cmp_r = get_Cmp_right(cmp);
1762 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1763 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1765 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1772 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1773 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1775 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1782 * Note: normalization puts the constant on the right side,
1783 * so we check only one case.
1785 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1786 /* Mux(t CMP 0, X, t) */
1787 if (is_Minus(f) && get_Minus_op(f) == t) {
1788 /* Mux(t CMP 0, -t, t) */
1789 if (proj_nr == pn_Cmp_Eq) {
1790 /* Mux(t == 0, -t, t) ==> -t */
1792 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1793 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1794 /* Mux(t != 0, -t, t) ==> t */
1796 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1803 } /* equivalent_node_Mux */
1806 * Remove Confirm nodes if setting is on.
1807 * Replace Confirms(x, '=', Constlike) by Constlike.
1809 static ir_node *equivalent_node_Confirm(ir_node *n) {
1810 ir_node *pred = get_Confirm_value(n);
1811 pn_Cmp pnc = get_Confirm_cmp(n);
1813 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1815 * rare case: two identical Confirms one after another,
1816 * replace the second one with the first.
1819 pred = get_Confirm_value(n);
1820 pnc = get_Confirm_cmp(n);
1822 if (get_opt_remove_confirm())
1823 return get_Confirm_value(n);
1828 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1829 * perform no actual computation, as, e.g., the Id nodes. It does not create
1830 * new nodes. It is therefore safe to free n if the node returned is not n.
1831 * If a node returns a Tuple we can not just skip it. If the size of the
1832 * in array fits, we transform n into a tuple (e.g., Div).
1834 ir_node *equivalent_node(ir_node *n) {
1835 if (n->op->ops.equivalent_node)
1836 return n->op->ops.equivalent_node(n);
1838 } /* equivalent_node */
1841 * Sets the default equivalent node operation for an ir_op_ops.
1843 * @param code the opcode for the default operation
1844 * @param ops the operations initialized
1849 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1853 ops->equivalent_node = equivalent_node_##a; \
1855 #define CASE_PROJ(a) \
1857 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1899 } /* firm_set_default_equivalent_node */
1902 * Returns non-zero if a node is a Phi node
1903 * with all predecessors constant.
1905 static int is_const_Phi(ir_node *n) {
1908 if (! is_Phi(n) || get_irn_arity(n) == 0)
1910 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1911 if (! is_Const(get_irn_n(n, i)))
1915 } /* is_const_Phi */
1917 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1918 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1921 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1923 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1924 if (eval == tarval_sub) {
1925 tarval_sub_type func = (tarval_sub_type)eval;
1927 return func(a, b, mode);
1929 tarval_binop_type func = (tarval_binop_type)eval;
1936 * Apply an evaluator on a binop with a constant operators (and one Phi).
1938 * @param phi the Phi node
1939 * @param other the other operand
1940 * @param eval an evaluator function
1941 * @param mode the mode of the result, may be different from the mode of the Phi!
1942 * @param left if non-zero, other is the left operand, else the right
1944 * @return a new Phi node if the conversion was successful, NULL else
1946 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1951 int i, n = get_irn_arity(phi);
1953 NEW_ARR_A(void *, res, n);
1955 for (i = 0; i < n; ++i) {
1956 pred = get_irn_n(phi, i);
1957 tv = get_Const_tarval(pred);
1958 tv = do_eval(eval, other, tv, mode);
1960 if (tv == tarval_bad) {
1961 /* folding failed, bad */
1967 for (i = 0; i < n; ++i) {
1968 pred = get_irn_n(phi, i);
1969 tv = get_Const_tarval(pred);
1970 tv = do_eval(eval, tv, other, mode);
1972 if (tv == tarval_bad) {
1973 /* folding failed, bad */
1979 irg = current_ir_graph;
1980 for (i = 0; i < n; ++i) {
1981 pred = get_irn_n(phi, i);
1982 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1983 mode, res[i], get_Const_type(pred));
1985 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1986 } /* apply_binop_on_phi */
1989 * Apply an evaluator on a binop with two constant Phi.
1991 * @param a the left Phi node
1992 * @param b the right Phi node
1993 * @param eval an evaluator function
1994 * @param mode the mode of the result, may be different from the mode of the Phi!
1996 * @return a new Phi node if the conversion was successful, NULL else
1998 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
1999 tarval *tv_l, *tv_r, *tv;
2005 if (get_nodes_block(a) != get_nodes_block(b))
2008 n = get_irn_arity(a);
2009 NEW_ARR_A(void *, res, n);
2011 for (i = 0; i < n; ++i) {
2012 pred = get_irn_n(a, i);
2013 tv_l = get_Const_tarval(pred);
2014 pred = get_irn_n(b, i);
2015 tv_r = get_Const_tarval(pred);
2016 tv = do_eval(eval, tv_l, tv_r, mode);
2018 if (tv == tarval_bad) {
2019 /* folding failed, bad */
2024 irg = current_ir_graph;
2025 for (i = 0; i < n; ++i) {
2026 pred = get_irn_n(a, i);
2027 res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred));
2029 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2030 } /* apply_binop_on_2_phis */
2033 * Apply an evaluator on a unop with a constant operator (a Phi).
2035 * @param phi the Phi node
2036 * @param eval an evaluator function
2038 * @return a new Phi node if the conversion was successful, NULL else
2040 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2046 int i, n = get_irn_arity(phi);
2048 NEW_ARR_A(void *, res, n);
2049 for (i = 0; i < n; ++i) {
2050 pred = get_irn_n(phi, i);
2051 tv = get_Const_tarval(pred);
2054 if (tv == tarval_bad) {
2055 /* folding failed, bad */
2060 mode = get_irn_mode(phi);
2061 irg = current_ir_graph;
2062 for (i = 0; i < n; ++i) {
2063 pred = get_irn_n(phi, i);
2064 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2065 mode, res[i], get_Const_type(pred));
2067 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2068 } /* apply_unop_on_phi */
2071 * Apply a conversion on a constant operator (a Phi).
2073 * @param phi the Phi node
2075 * @return a new Phi node if the conversion was successful, NULL else
2077 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2082 int i, n = get_irn_arity(phi);
2084 NEW_ARR_A(void *, res, n);
2085 for (i = 0; i < n; ++i) {
2086 pred = get_irn_n(phi, i);
2087 tv = get_Const_tarval(pred);
2088 tv = tarval_convert_to(tv, mode);
2090 if (tv == tarval_bad) {
2091 /* folding failed, bad */
2096 irg = current_ir_graph;
2097 for (i = 0; i < n; ++i) {
2098 pred = get_irn_n(phi, i);
2099 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
2100 mode, res[i], get_Const_type(pred));
2102 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2103 } /* apply_conv_on_phi */
2106 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2107 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2108 * If possible, remove the Conv's.
2110 static ir_node *transform_node_AddSub(ir_node *n) {
2111 ir_mode *mode = get_irn_mode(n);
2113 if (mode_is_reference(mode)) {
2114 ir_node *left = get_binop_left(n);
2115 ir_node *right = get_binop_right(n);
2116 unsigned ref_bits = get_mode_size_bits(mode);
2118 if (is_Conv(left)) {
2119 ir_mode *lmode = get_irn_mode(left);
2120 unsigned bits = get_mode_size_bits(lmode);
2122 if (ref_bits == bits &&
2123 mode_is_int(lmode) &&
2124 get_mode_arithmetic(lmode) == irma_twos_complement) {
2125 ir_node *pre = get_Conv_op(left);
2126 ir_mode *pre_mode = get_irn_mode(pre);
2128 if (mode_is_int(pre_mode) &&
2129 get_mode_size_bits(pre_mode) == bits &&
2130 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2131 /* ok, this conv just changes to sign, moreover the calculation
2132 * is done with same number of bits as our address mode, so
2133 * we can ignore the conv as address calculation can be viewed
2134 * as either signed or unsigned
2136 set_binop_left(n, pre);
2141 if (is_Conv(right)) {
2142 ir_mode *rmode = get_irn_mode(right);
2143 unsigned bits = get_mode_size_bits(rmode);
2145 if (ref_bits == bits &&
2146 mode_is_int(rmode) &&
2147 get_mode_arithmetic(rmode) == irma_twos_complement) {
2148 ir_node *pre = get_Conv_op(right);
2149 ir_mode *pre_mode = get_irn_mode(pre);
2151 if (mode_is_int(pre_mode) &&
2152 get_mode_size_bits(pre_mode) == bits &&
2153 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2154 /* ok, this conv just changes to sign, moreover the calculation
2155 * is done with same number of bits as our address mode, so
2156 * we can ignore the conv as address calculation can be viewed
2157 * as either signed or unsigned
2159 set_binop_right(n, pre);
2164 /* let address arithmetic use unsigned modes */
2165 if (is_Const(right)) {
2166 ir_mode *rmode = get_irn_mode(right);
2168 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2169 /* convert a AddP(P, *s) into AddP(P, *u) */
2170 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2172 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2173 set_binop_right(n, pre);
2179 } /* transform_node_AddSub */
2181 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2183 if (is_Const(b) && is_const_Phi(a)) { \
2184 /* check for Op(Phi, Const) */ \
2185 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2187 else if (is_Const(a) && is_const_Phi(b)) { \
2188 /* check for Op(Const, Phi) */ \
2189 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2191 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2192 /* check for Op(Phi, Phi) */ \
2193 c = apply_binop_on_2_phis(a, b, eval, mode); \
2196 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2200 #define HANDLE_UNOP_PHI(eval, a, c) \
2202 if (is_const_Phi(a)) { \
2203 /* check for Op(Phi) */ \
2204 c = apply_unop_on_phi(a, eval); \
2206 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2212 * Do the AddSub optimization, then Transform
2213 * Constant folding on Phi
2214 * Add(a,a) -> Mul(a, 2)
2215 * Add(Mul(a, x), a) -> Mul(a, x+1)
2216 * if the mode is integer or float.
2217 * Transform Add(a,-b) into Sub(a,b).
2218 * Reassociation might fold this further.
2220 static ir_node *transform_node_Add(ir_node *n) {
2222 ir_node *a, *b, *c, *oldn = n;
2224 n = transform_node_AddSub(n);
2226 a = get_Add_left(n);
2227 b = get_Add_right(n);
2229 mode = get_irn_mode(n);
2231 if (mode_is_reference(mode)) {
2232 ir_mode *lmode = get_irn_mode(a);
2234 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2235 /* an Add(a, NULL) is a hidden Conv */
2236 dbg_info *dbg = get_irn_dbg_info(n);
2237 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2241 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2243 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2244 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2247 if (mode_is_num(mode)) {
2248 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2249 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2250 ir_node *block = get_nodes_block(n);
2253 get_irn_dbg_info(n),
2257 new_r_Const_long(current_ir_graph, block, mode, 2),
2259 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2264 get_irn_dbg_info(n),
2270 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2275 get_irn_dbg_info(n),
2281 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2284 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2285 /* Here we rely on constants be on the RIGHT side */
2287 ir_node *op = get_Not_op(a);
2289 if (is_Const(b) && is_Const_one(b)) {
2291 ir_node *blk = get_nodes_block(n);
2292 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2293 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2298 ir_node *blk = get_nodes_block(n);
2299 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2300 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2305 ir_node *op = get_Not_op(b);
2309 ir_node *blk = get_nodes_block(n);
2310 n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode));
2311 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2318 } /* transform_node_Add */
2321 * returns -cnst or NULL if impossible
2323 static ir_node *const_negate(ir_node *cnst) {
2324 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2325 dbg_info *dbgi = get_irn_dbg_info(cnst);
2326 ir_graph *irg = get_irn_irg(cnst);
2327 ir_node *block = get_nodes_block(cnst);
2328 ir_mode *mode = get_irn_mode(cnst);
2329 if (tv == tarval_bad) return NULL;
2330 return new_rd_Const(dbgi, irg, block, mode, tv);
2334 * Do the AddSub optimization, then Transform
2335 * Constant folding on Phi
2336 * Sub(0,a) -> Minus(a)
2337 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2338 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2339 * Sub(Add(a, x), x) -> a
2340 * Sub(x, Add(x, a)) -> -a
2341 * Sub(x, Const) -> Add(x, -Const)
2343 static ir_node *transform_node_Sub(ir_node *n) {
2348 n = transform_node_AddSub(n);
2350 a = get_Sub_left(n);
2351 b = get_Sub_right(n);
2353 mode = get_irn_mode(n);
2355 if (mode_is_int(mode)) {
2356 ir_mode *lmode = get_irn_mode(a);
2358 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2359 /* a Sub(a, NULL) is a hidden Conv */
2360 dbg_info *dbg = get_irn_dbg_info(n);
2361 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2362 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2366 if (mode == lmode &&
2367 get_mode_arithmetic(mode) == irma_twos_complement &&
2369 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2371 dbg_info *dbg = get_irn_dbg_info(n);
2372 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2373 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2379 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2381 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2382 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2385 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2386 /* a - C -> a + (-C) */
2387 ir_node *cnst = const_negate(b);
2389 ir_node *block = get_nodes_block(n);
2390 dbg_info *dbgi = get_irn_dbg_info(n);
2391 ir_graph *irg = get_irn_irg(n);
2393 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2394 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2399 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2400 ir_graph *irg = current_ir_graph;
2401 dbg_info *dbg = get_irn_dbg_info(n);
2402 ir_node *block = get_nodes_block(n);
2403 ir_node *left = get_Minus_op(a);
2404 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2406 n = new_rd_Minus(dbg, irg, block, add, mode);
2407 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2409 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2410 ir_graph *irg = current_ir_graph;
2411 dbg_info *dbg = get_irn_dbg_info(n);
2412 ir_node *block = get_nodes_block(n);
2413 ir_node *right = get_Minus_op(b);
2415 n = new_rd_Add(dbg, irg, block, a, right, mode);
2416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2418 } else if (is_Sub(b)) {
2419 /* a - (b - c) -> a + (c - b)
2420 * -> (a - b) + c iff (b - c) is a pointer */
2421 ir_graph *irg = current_ir_graph;
2422 dbg_info *s_dbg = get_irn_dbg_info(b);
2423 ir_node *s_block = get_nodes_block(b);
2424 ir_node *s_left = get_Sub_left(b);
2425 ir_node *s_right = get_Sub_right(b);
2426 ir_mode *s_mode = get_irn_mode(b);
2427 if (s_mode == mode_P) {
2428 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2429 dbg_info *a_dbg = get_irn_dbg_info(n);
2430 ir_node *a_block = get_nodes_block(n);
2433 s_right = new_r_Conv(irg, a_block, s_right, mode);
2434 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2436 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2437 dbg_info *a_dbg = get_irn_dbg_info(n);
2438 ir_node *a_block = get_nodes_block(n);
2440 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2444 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2445 ir_node *m_right = get_Mul_right(b);
2446 if (is_Const(m_right)) {
2447 ir_node *cnst2 = const_negate(m_right);
2448 if (cnst2 != NULL) {
2449 ir_graph *irg = current_ir_graph;
2450 dbg_info *m_dbg = get_irn_dbg_info(b);
2451 ir_node *m_block = get_nodes_block(b);
2452 ir_node *m_left = get_Mul_left(b);
2453 ir_mode *m_mode = get_irn_mode(b);
2454 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2455 dbg_info *a_dbg = get_irn_dbg_info(n);
2456 ir_node *a_block = get_nodes_block(n);
2458 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2459 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2465 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2466 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2468 get_irn_dbg_info(n),
2473 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2477 if (mode_wrap_around(mode)) {
2478 ir_node *left = get_Add_left(a);
2479 ir_node *right = get_Add_right(a);
2481 /* FIXME: Does the Conv's work only for two complement or generally? */
2483 if (mode != get_irn_mode(right)) {
2484 /* This Sub is an effective Cast */
2485 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2488 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2490 } else if (right == b) {
2491 if (mode != get_irn_mode(left)) {
2492 /* This Sub is an effective Cast */
2493 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2496 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2502 if (mode_wrap_around(mode)) {
2503 ir_node *left = get_Add_left(b);
2504 ir_node *right = get_Add_right(b);
2506 /* FIXME: Does the Conv's work only for two complement or generally? */
2508 ir_mode *r_mode = get_irn_mode(right);
2510 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2511 if (mode != r_mode) {
2512 /* This Sub is an effective Cast */
2513 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2515 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2517 } else if (right == a) {
2518 ir_mode *l_mode = get_irn_mode(left);
2520 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2521 if (mode != l_mode) {
2522 /* This Sub is an effective Cast */
2523 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2525 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2530 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2531 ir_mode *mode = get_irn_mode(a);
2533 if (mode == get_irn_mode(b)) {
2535 ir_node *op_a = get_Conv_op(a);
2536 ir_node *op_b = get_Conv_op(b);
2538 /* check if it's allowed to skip the conv */
2539 ma = get_irn_mode(op_a);
2540 mb = get_irn_mode(op_b);
2542 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2543 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2546 set_Sub_right(n, b);
2552 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2553 if (!is_reassoc_running() && is_Mul(a)) {
2554 ir_node *ma = get_Mul_left(a);
2555 ir_node *mb = get_Mul_right(a);
2558 ir_node *blk = get_nodes_block(n);
2560 get_irn_dbg_info(n),
2561 current_ir_graph, blk,
2564 get_irn_dbg_info(n),
2565 current_ir_graph, blk,
2567 new_r_Const_long(current_ir_graph, blk, mode, 1),
2570 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2572 } else if (mb == b) {
2573 ir_node *blk = get_nodes_block(n);
2575 get_irn_dbg_info(n),
2576 current_ir_graph, blk,
2579 get_irn_dbg_info(n),
2580 current_ir_graph, blk,
2582 new_r_Const_long(current_ir_graph, blk, mode, 1),
2585 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2589 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2590 ir_node *x = get_Sub_left(a);
2591 ir_node *y = get_Sub_right(a);
2592 ir_node *blk = get_nodes_block(n);
2593 ir_mode *m_b = get_irn_mode(b);
2594 ir_mode *m_y = get_irn_mode(y);
2598 /* Determine the right mode for the Add. */
2601 else if (mode_is_reference(m_b))
2603 else if (mode_is_reference(m_y))
2607 * Both modes are different but none is reference,
2608 * happens for instance in SubP(SubP(P, Iu), Is).
2609 * We have two possibilities here: Cast or ignore.
2610 * Currently we ignore this case.
2615 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2617 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2618 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2622 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2623 if (is_Const(a) && is_Not(b)) {
2624 /* c - ~X = X + (c+1) */
2625 tarval *tv = get_Const_tarval(a);
2627 tv = tarval_add(tv, get_mode_one(mode));
2628 if (tv != tarval_bad) {
2629 ir_node *blk = get_nodes_block(n);
2630 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2631 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2632 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2638 } /* transform_node_Sub */
2641 * Several transformation done on n*n=2n bits mul.
2642 * These transformations must be done here because new nodes may be produced.
2644 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2646 ir_node *a = get_Mul_left(n);
2647 ir_node *b = get_Mul_right(n);
2648 tarval *ta = value_of(a);
2649 tarval *tb = value_of(b);
2650 ir_mode *smode = get_irn_mode(a);
2652 if (ta == get_mode_one(smode)) {
2653 /* (L)1 * (L)b = (L)b */
2654 ir_node *blk = get_nodes_block(n);
2655 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2656 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2659 else if (ta == get_mode_minus_one(smode)) {
2660 /* (L)-1 * (L)b = (L)b */
2661 ir_node *blk = get_nodes_block(n);
2662 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2663 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2664 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2667 if (tb == get_mode_one(smode)) {
2668 /* (L)a * (L)1 = (L)a */
2669 ir_node *blk = get_irn_n(a, -1);
2670 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2671 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2674 else if (tb == get_mode_minus_one(smode)) {
2675 /* (L)a * (L)-1 = (L)-a */
2676 ir_node *blk = get_nodes_block(n);
2677 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2678 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2679 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2686 * Transform Mul(a,-1) into -a.
2687 * Do constant evaluation of Phi nodes.
2688 * Do architecture dependent optimizations on Mul nodes
2690 static ir_node *transform_node_Mul(ir_node *n) {
2691 ir_node *c, *oldn = n;
2692 ir_mode *mode = get_irn_mode(n);
2693 ir_node *a = get_Mul_left(n);
2694 ir_node *b = get_Mul_right(n);
2696 if (is_Bad(a) || is_Bad(b))
2699 if (mode != get_irn_mode(a))
2700 return transform_node_Mul2n(n, mode);
2702 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2704 if (mode_is_signed(mode)) {
2707 if (value_of(a) == get_mode_minus_one(mode))
2709 else if (value_of(b) == get_mode_minus_one(mode))
2712 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2713 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2718 if (is_Const(b)) { /* (-a) * const -> a * -const */
2719 ir_node *cnst = const_negate(b);
2721 dbg_info *dbgi = get_irn_dbg_info(n);
2722 ir_node *block = get_nodes_block(n);
2723 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2724 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2727 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2728 dbg_info *dbgi = get_irn_dbg_info(n);
2729 ir_node *block = get_nodes_block(n);
2730 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2731 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2733 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2734 ir_node *sub_l = get_Sub_left(b);
2735 ir_node *sub_r = get_Sub_right(b);
2736 dbg_info *dbgi = get_irn_dbg_info(n);
2737 ir_graph *irg = current_ir_graph;
2738 ir_node *block = get_nodes_block(n);
2739 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2740 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2741 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2744 } else if (is_Minus(b)) {
2745 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2746 ir_node *sub_l = get_Sub_left(a);
2747 ir_node *sub_r = get_Sub_right(a);
2748 dbg_info *dbgi = get_irn_dbg_info(n);
2749 ir_graph *irg = current_ir_graph;
2750 ir_node *block = get_nodes_block(n);
2751 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2752 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2753 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2757 if (get_mode_arithmetic(mode) == irma_ieee754) {
2759 tarval *tv = get_Const_tarval(a);
2760 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2761 && !tarval_is_negative(tv)) {
2762 /* 2.0 * b = b + b */
2763 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2764 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2768 else if (is_Const(b)) {
2769 tarval *tv = get_Const_tarval(b);
2770 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2771 && !tarval_is_negative(tv)) {
2772 /* a * 2.0 = a + a */
2773 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2774 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2779 return arch_dep_replace_mul_with_shifts(n);
2780 } /* transform_node_Mul */
2783 * Transform a Div Node.
2785 static ir_node *transform_node_Div(ir_node *n) {
2786 ir_mode *mode = get_Div_resmode(n);
2787 ir_node *a = get_Div_left(n);
2788 ir_node *b = get_Div_right(n);
2792 if (is_Const(b) && is_const_Phi(a)) {
2793 /* check for Div(Phi, Const) */
2794 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2796 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2800 else if (is_Const(a) && is_const_Phi(b)) {
2801 /* check for Div(Const, Phi) */
2802 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2804 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2808 else if (is_const_Phi(a) && is_const_Phi(b)) {
2809 /* check for Div(Phi, Phi) */
2810 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2812 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2819 if (tv != tarval_bad) {
2820 value = new_Const(get_tarval_mode(tv), tv);
2822 DBG_OPT_CSTEVAL(n, value);
2825 ir_node *a = get_Div_left(n);
2826 ir_node *b = get_Div_right(n);
2827 const ir_node *dummy;
2829 if (a == b && value_not_zero(a, &dummy)) {
2830 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2831 value = new_Const(mode, get_mode_one(mode));
2832 DBG_OPT_CSTEVAL(n, value);
2835 if (mode_is_signed(mode) && is_Const(b)) {
2836 tarval *tv = get_Const_tarval(b);
2838 if (tv == get_mode_minus_one(mode)) {
2840 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2841 DBG_OPT_CSTEVAL(n, value);
2845 /* Try architecture dependent optimization */
2846 value = arch_dep_replace_div_by_const(n);
2854 /* Turn Div into a tuple (mem, jmp, bad, value) */
2855 mem = get_Div_mem(n);
2856 blk = get_nodes_block(n);
2858 /* skip a potential Pin */
2859 mem = skip_Pin(mem);
2860 turn_into_tuple(n, pn_Div_max);
2861 set_Tuple_pred(n, pn_Div_M, mem);
2862 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2863 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2864 set_Tuple_pred(n, pn_Div_res, value);
2867 } /* transform_node_Div */
2870 * Transform a Mod node.
2872 static ir_node *transform_node_Mod(ir_node *n) {
2873 ir_mode *mode = get_Mod_resmode(n);
2874 ir_node *a = get_Mod_left(n);
2875 ir_node *b = get_Mod_right(n);
2879 if (is_Const(b) && is_const_Phi(a)) {
2880 /* check for Div(Phi, Const) */
2881 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2883 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2887 else if (is_Const(a) && is_const_Phi(b)) {
2888 /* check for Div(Const, Phi) */
2889 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2891 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2895 else if (is_const_Phi(a) && is_const_Phi(b)) {
2896 /* check for Div(Phi, Phi) */
2897 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2899 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2906 if (tv != tarval_bad) {
2907 value = new_Const(get_tarval_mode(tv), tv);
2909 DBG_OPT_CSTEVAL(n, value);
2912 ir_node *a = get_Mod_left(n);
2913 ir_node *b = get_Mod_right(n);
2914 const ir_node *dummy;
2916 if (a == b && value_not_zero(a, &dummy)) {
2917 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2918 value = new_Const(mode, get_mode_null(mode));
2919 DBG_OPT_CSTEVAL(n, value);
2922 if (mode_is_signed(mode) && is_Const(b)) {
2923 tarval *tv = get_Const_tarval(b);
2925 if (tv == get_mode_minus_one(mode)) {
2927 value = new_Const(mode, get_mode_null(mode));
2928 DBG_OPT_CSTEVAL(n, value);
2932 /* Try architecture dependent optimization */
2933 value = arch_dep_replace_mod_by_const(n);
2941 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2942 mem = get_Mod_mem(n);
2943 blk = get_nodes_block(n);
2945 /* skip a potential Pin */
2946 mem = skip_Pin(mem);
2947 turn_into_tuple(n, pn_Mod_max);
2948 set_Tuple_pred(n, pn_Mod_M, mem);
2949 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2950 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2951 set_Tuple_pred(n, pn_Mod_res, value);
2954 } /* transform_node_Mod */
2957 * Transform a DivMod node.
2959 static ir_node *transform_node_DivMod(ir_node *n) {
2960 const ir_node *dummy;
2961 ir_node *a = get_DivMod_left(n);
2962 ir_node *b = get_DivMod_right(n);
2963 ir_mode *mode = get_DivMod_resmode(n);
2968 if (is_Const(b) && is_const_Phi(a)) {
2969 /* check for Div(Phi, Const) */
2970 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2971 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2973 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2974 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2978 else if (is_Const(a) && is_const_Phi(b)) {
2979 /* check for Div(Const, Phi) */
2980 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2981 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2983 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2984 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2988 else if (is_const_Phi(a) && is_const_Phi(b)) {
2989 /* check for Div(Phi, Phi) */
2990 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2991 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2993 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2994 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3001 if (tb != tarval_bad) {
3002 if (tb == get_mode_one(get_tarval_mode(tb))) {
3004 vb = new_Const(mode, get_mode_null(mode));
3005 DBG_OPT_CSTEVAL(n, vb);
3007 } else if (ta != tarval_bad) {
3008 tarval *resa, *resb;
3009 resa = tarval_div(ta, tb);
3010 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3011 Jmp for X result!? */
3012 resb = tarval_mod(ta, tb);
3013 if (resb == tarval_bad) return n; /* Causes exception! */
3014 va = new_Const(mode, resa);
3015 vb = new_Const(mode, resb);
3016 DBG_OPT_CSTEVAL(n, va);
3017 DBG_OPT_CSTEVAL(n, vb);
3019 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3020 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
3021 vb = new_Const(mode, get_mode_null(mode));
3022 DBG_OPT_CSTEVAL(n, va);
3023 DBG_OPT_CSTEVAL(n, vb);
3025 } else { /* Try architecture dependent optimization */
3028 arch_dep_replace_divmod_by_const(&va, &vb, n);
3029 evaluated = va != NULL;
3031 } else if (a == b) {
3032 if (value_not_zero(a, &dummy)) {
3034 va = new_Const(mode, get_mode_one(mode));
3035 vb = new_Const(mode, get_mode_null(mode));
3036 DBG_OPT_CSTEVAL(n, va);
3037 DBG_OPT_CSTEVAL(n, vb);
3040 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3043 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3044 /* 0 / non-Const = 0 */
3049 if (evaluated) { /* replace by tuple */
3053 mem = get_DivMod_mem(n);
3054 /* skip a potential Pin */
3055 mem = skip_Pin(mem);
3057 blk = get_nodes_block(n);
3058 turn_into_tuple(n, pn_DivMod_max);
3059 set_Tuple_pred(n, pn_DivMod_M, mem);
3060 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3061 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3062 set_Tuple_pred(n, pn_DivMod_res_div, va);
3063 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3067 } /* transform_node_DivMod */
3070 * Optimize x / c to x * (1/c)
3072 static ir_node *transform_node_Quot(ir_node *n) {
3073 ir_mode *mode = get_Quot_resmode(n);
3076 if (get_mode_arithmetic(mode) == irma_ieee754) {
3077 ir_node *b = get_Quot_right(n);
3078 tarval *tv = value_of(b);
3080 if (tv != tarval_bad) {
3084 * Floating point constant folding might be disabled here to
3086 * However, as we check for exact result, doing it is safe.
3089 rem = tarval_enable_fp_ops(1);
3090 tv = tarval_quo(get_mode_one(mode), tv);
3091 (void)tarval_enable_fp_ops(rem);
3093 /* Do the transformation if the result is either exact or we are not
3094 using strict rules. */
3095 if (tv != tarval_bad &&
3096 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3097 ir_node *blk = get_nodes_block(n);
3098 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3099 ir_node *a = get_Quot_left(n);
3100 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3101 ir_node *mem = get_Quot_mem(n);
3103 /* skip a potential Pin */
3104 mem = skip_Pin(mem);
3105 turn_into_tuple(n, pn_Quot_max);
3106 set_Tuple_pred(n, pn_Quot_M, mem);
3107 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3108 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3109 set_Tuple_pred(n, pn_Quot_res, m);
3110 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3115 } /* transform_node_Quot */
3118 * Optimize Abs(x) into x if x is Confirmed >= 0
3119 * Optimize Abs(x) into -x if x is Confirmed <= 0
3120 * Optimize Abs(-x) int Abs(x)
3122 static ir_node *transform_node_Abs(ir_node *n) {
3123 ir_node *c, *oldn = n;
3124 ir_node *a = get_Abs_op(n);
3127 HANDLE_UNOP_PHI(tarval_abs, a, c);
3129 switch (classify_value_sign(a)) {
3130 case value_classified_negative:
3131 mode = get_irn_mode(n);
3134 * We can replace the Abs by -x here.
3135 * We even could add a new Confirm here
3136 * (if not twos complement)
3138 * Note that -x would create a new node, so we could
3139 * not run it in the equivalent_node() context.
3141 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3142 get_nodes_block(n), a, mode);
3144 DBG_OPT_CONFIRM(oldn, n);
3146 case value_classified_positive:
3147 /* n is positive, Abs is not needed */
3150 DBG_OPT_CONFIRM(oldn, n);
3156 /* Abs(-x) = Abs(x) */
3157 mode = get_irn_mode(n);
3158 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3159 get_nodes_block(n), get_Minus_op(a), mode);
3160 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3164 } /* transform_node_Abs */
3167 * Optimize -a CMP -b into b CMP a.
3168 * This works only for for modes where unary Minus
3170 * Note that two-complement integers can Overflow
3171 * so it will NOT work.
3173 * For == and != can be handled in Proj(Cmp)
3175 static ir_node *transform_node_Cmp(ir_node *n) {
3177 ir_node *left = get_Cmp_left(n);
3178 ir_node *right = get_Cmp_right(n);
3180 if (is_Minus(left) && is_Minus(right) &&
3181 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3182 ir_node *const new_left = get_Minus_op(right);
3183 ir_node *const new_right = get_Minus_op(left);
3184 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3185 get_nodes_block(n), new_left, new_right);
3186 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3189 } /* transform_node_Cmp */
3193 * Transform a Cond node.
3195 * Replace the Cond by a Jmp if it branches on a constant
3198 static ir_node *transform_node_Cond(ir_node *n) {
3201 ir_node *a = get_Cond_selector(n);
3202 tarval *ta = value_of(a);
3204 /* we need block info which is not available in floating irgs */
3205 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3208 if ((ta != tarval_bad) &&
3209 (get_irn_mode(a) == mode_b) &&
3210 (get_opt_unreachable_code())) {
3211 /* It's a boolean Cond, branching on a boolean constant.
3212 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3213 ir_node *blk = get_nodes_block(n);
3214 jmp = new_r_Jmp(current_ir_graph, blk);
3215 turn_into_tuple(n, pn_Cond_max);
3216 if (ta == tarval_b_true) {
3217 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3218 set_Tuple_pred(n, pn_Cond_true, jmp);
3220 set_Tuple_pred(n, pn_Cond_false, jmp);
3221 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3223 /* We might generate an endless loop, so keep it alive. */
3224 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3227 } /* transform_node_Cond */
3230 * Prototype of a recursive transform function
3231 * for bitwise distributive transformations.
3233 typedef ir_node* (*recursive_transform)(ir_node *n);
3236 * makes use of distributive laws for and, or, eor
3237 * and(a OP c, b OP c) -> and(a, b) OP c
3238 * note, might return a different op than n
3240 static ir_node *transform_bitwise_distributive(ir_node *n,
3241 recursive_transform trans_func)
3244 ir_node *a = get_binop_left(n);
3245 ir_node *b = get_binop_right(n);
3246 ir_op *op = get_irn_op(a);
3247 ir_op *op_root = get_irn_op(n);
3249 if(op != get_irn_op(b))
3252 if (op == op_Conv) {
3253 ir_node *a_op = get_Conv_op(a);
3254 ir_node *b_op = get_Conv_op(b);
3255 ir_mode *a_mode = get_irn_mode(a_op);
3256 ir_mode *b_mode = get_irn_mode(b_op);
3257 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3258 ir_node *blk = get_nodes_block(n);
3261 set_binop_left(n, a_op);
3262 set_binop_right(n, b_op);
3263 set_irn_mode(n, a_mode);
3265 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3267 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3273 /* nothing to gain here */
3277 if (op == op_Shrs || op == op_Shr || op == op_Shl
3278 || op == op_And || op == op_Or || op == op_Eor) {
3279 ir_node *a_left = get_binop_left(a);
3280 ir_node *a_right = get_binop_right(a);
3281 ir_node *b_left = get_binop_left(b);
3282 ir_node *b_right = get_binop_right(b);
3284 ir_node *op1 = NULL;
3285 ir_node *op2 = NULL;
3287 if (is_op_commutative(op)) {
3288 if (a_left == b_left) {
3292 } else if(a_left == b_right) {
3296 } else if(a_right == b_left) {
3302 if(a_right == b_right) {
3309 /* (a sop c) & (b sop c) => (a & b) sop c */
3310 ir_node *blk = get_nodes_block(n);
3312 ir_node *new_n = exact_copy(n);
3313 set_binop_left(new_n, op1);
3314 set_binop_right(new_n, op2);
3315 new_n = trans_func(new_n);
3317 if(op_root == op_Eor && op == op_Or) {
3318 dbg_info *dbgi = get_irn_dbg_info(n);
3319 ir_graph *irg = current_ir_graph;
3320 ir_mode *mode = get_irn_mode(c);
3322 c = new_rd_Not(dbgi, irg, blk, c, mode);
3323 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3326 set_nodes_block(n, blk);
3327 set_binop_left(n, new_n);
3328 set_binop_right(n, c);
3329 add_identities(current_ir_graph->value_table, n);
3332 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3343 static ir_node *transform_node_And(ir_node *n) {
3344 ir_node *c, *oldn = n;
3345 ir_node *a = get_And_left(n);
3346 ir_node *b = get_And_right(n);
3349 mode = get_irn_mode(n);
3350 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3352 /* we can evaluate 2 Projs of the same Cmp */
3353 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3354 ir_node *pred_a = get_Proj_pred(a);
3355 ir_node *pred_b = get_Proj_pred(b);
3356 if (pred_a == pred_b) {
3357 dbg_info *dbgi = get_irn_dbg_info(n);
3358 ir_node *block = get_nodes_block(pred_a);
3359 pn_Cmp pn_a = get_Proj_proj(a);
3360 pn_Cmp pn_b = get_Proj_proj(b);
3361 /* yes, we can simply calculate with pncs */
3362 pn_Cmp new_pnc = pn_a & pn_b;
3364 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3369 ir_node *op = get_Not_op(b);
3371 ir_node *ba = get_And_left(op);
3372 ir_node *bb = get_And_right(op);
3374 /* it's enough to test the following cases due to normalization! */
3375 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3376 /* (a|b) & ~(a&b) = a^b */
3377 ir_node *block = get_nodes_block(n);
3379 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3380 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3388 ir_node *op = get_Not_op(a);
3390 ir_node *aa = get_And_left(op);
3391 ir_node *ab = get_And_right(op);
3393 /* it's enough to test the following cases due to normalization! */
3394 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3395 /* (a|b) & ~(a&b) = a^b */
3396 ir_node *block = get_nodes_block(n);
3398 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3399 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3406 ir_node *al = get_Eor_left(a);
3407 ir_node *ar = get_Eor_right(a);
3410 /* (b ^ a) & b -> ~a & b */
3411 dbg_info *dbg = get_irn_dbg_info(n);
3412 ir_node *block = get_nodes_block(n);
3414 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3415 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3420 /* (a ^ b) & b -> ~a & b */
3421 dbg_info *dbg = get_irn_dbg_info(n);
3422 ir_node *block = get_nodes_block(n);
3424 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3425 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3431 ir_node *bl = get_Eor_left(b);
3432 ir_node *br = get_Eor_right(b);
3435 /* a & (a ^ b) -> a & ~b */
3436 dbg_info *dbg = get_irn_dbg_info(n);
3437 ir_node *block = get_nodes_block(n);
3439 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3440 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3445 /* a & (b ^ a) -> a & ~b */
3446 dbg_info *dbg = get_irn_dbg_info(n);
3447 ir_node *block = get_nodes_block(n);
3449 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3450 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3451 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3455 if (is_Not(a) && is_Not(b)) {
3456 /* ~a & ~b = ~(a|b) */
3457 ir_node *block = get_nodes_block(n);
3458 ir_mode *mode = get_irn_mode(n);
3462 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3463 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3464 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3468 n = transform_bitwise_distributive(n, transform_node_And);
3471 } /* transform_node_And */
3476 static ir_node *transform_node_Eor(ir_node *n) {
3477 ir_node *c, *oldn = n;
3478 ir_node *a = get_Eor_left(n);
3479 ir_node *b = get_Eor_right(n);
3480 ir_mode *mode = get_irn_mode(n);
3482 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3484 /* we can evaluate 2 Projs of the same Cmp */
3485 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3486 ir_node *pred_a = get_Proj_pred(a);
3487 ir_node *pred_b = get_Proj_pred(b);
3488 if(pred_a == pred_b) {
3489 dbg_info *dbgi = get_irn_dbg_info(n);
3490 ir_node *block = get_nodes_block(pred_a);
3491 pn_Cmp pn_a = get_Proj_proj(a);
3492 pn_Cmp pn_b = get_Proj_proj(b);
3493 /* yes, we can simply calculate with pncs */
3494 pn_Cmp new_pnc = pn_a ^ pn_b;
3496 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3503 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n),
3504 mode, get_mode_null(mode));
3505 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3506 } else if (mode == mode_b &&
3508 is_Const(b) && is_Const_one(b) &&
3509 is_Cmp(get_Proj_pred(a))) {
3510 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3511 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3512 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3514 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3515 } else if (is_Const(b)) {
3516 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3517 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3518 ir_node *not_op = get_Not_op(a);
3519 dbg_info *dbg = get_irn_dbg_info(n);
3520 ir_graph *irg = current_ir_graph;
3521 ir_node *block = get_nodes_block(n);
3522 ir_mode *mode = get_irn_mode(n);
3523 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3525 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3526 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3527 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3530 n = transform_bitwise_distributive(n, transform_node_Eor);
3534 } /* transform_node_Eor */
3539 static ir_node *transform_node_Not(ir_node *n) {
3540 ir_node *c, *oldn = n;
3541 ir_node *a = get_Not_op(n);
3542 ir_mode *mode = get_irn_mode(n);
3544 HANDLE_UNOP_PHI(tarval_not,a,c);
3546 /* check for a boolean Not */
3547 if (mode == mode_b &&
3549 is_Cmp(get_Proj_pred(a))) {
3550 /* We negate a Cmp. The Cmp has the negated result anyways! */
3551 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3552 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3553 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3557 ir_node *eor_b = get_Eor_right(a);
3558 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3559 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3560 ir_node *eor_a = get_Eor_left(a);
3561 dbg_info *dbg = get_irn_dbg_info(n);
3562 ir_graph *irg = current_ir_graph;
3563 ir_node *block = get_nodes_block(n);
3564 ir_mode *mode = get_irn_mode(n);
3565 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3569 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3570 if (is_Minus(a)) { /* ~-x -> x + -1 */
3571 dbg_info *dbg = get_irn_dbg_info(n);
3572 ir_graph *irg = current_ir_graph;
3573 ir_node *block = get_nodes_block(n);
3574 ir_node *add_l = get_Minus_op(a);
3575 ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode));
3576 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3577 } else if (is_Add(a)) {
3578 ir_node *add_r = get_Add_right(a);
3579 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3580 /* ~(x + -1) = -x */
3581 ir_node *op = get_Add_left(a);
3582 ir_node *blk = get_nodes_block(n);
3583 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3584 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3589 } /* transform_node_Not */
3592 * Transform a Minus.
3596 * -(a >>u (size-1)) = a >>s (size-1)
3597 * -(a >>s (size-1)) = a >>u (size-1)
3598 * -(a * const) -> a * -const
3600 static ir_node *transform_node_Minus(ir_node *n) {
3601 ir_node *c, *oldn = n;
3602 ir_node *a = get_Minus_op(n);
3605 HANDLE_UNOP_PHI(tarval_neg,a,c);
3607 mode = get_irn_mode(a);
3608 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3609 /* the following rules are only to twos-complement */
3612 ir_node *op = get_Not_op(a);
3613 tarval *tv = get_mode_one(mode);
3614 ir_node *blk = get_nodes_block(n);
3615 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
3616 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3617 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3621 ir_node *c = get_Shr_right(a);
3624 tarval *tv = get_Const_tarval(c);
3626 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3627 /* -(a >>u (size-1)) = a >>s (size-1) */
3628 ir_node *v = get_Shr_left(a);
3630 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3631 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3637 ir_node *c = get_Shrs_right(a);
3640 tarval *tv = get_Const_tarval(c);
3642 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3643 /* -(a >>s (size-1)) = a >>u (size-1) */
3644 ir_node *v = get_Shrs_left(a);
3646 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3647 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3654 /* - (a-b) = b - a */
3655 ir_node *la = get_Sub_left(a);
3656 ir_node *ra = get_Sub_right(a);
3657 ir_node *blk = get_nodes_block(n);
3659 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3660 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3664 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3665 ir_node *mul_l = get_Mul_left(a);
3666 ir_node *mul_r = get_Mul_right(a);
3667 tarval *tv = value_of(mul_r);
3668 if (tv != tarval_bad) {
3669 tv = tarval_neg(tv);
3670 if (tv != tarval_bad) {
3671 ir_node *cnst = new_Const(mode, tv);
3672 dbg_info *dbg = get_irn_dbg_info(a);
3673 ir_graph *irg = current_ir_graph;
3674 ir_node *block = get_nodes_block(a);
3675 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3676 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3683 } /* transform_node_Minus */
3686 * Transform a Cast_type(Const) into a new Const_type
3688 static ir_node *transform_node_Cast(ir_node *n) {
3690 ir_node *pred = get_Cast_op(n);
3691 ir_type *tp = get_irn_type(n);
3693 if (is_Const(pred) && get_Const_type(pred) != tp) {
3694 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3695 get_Const_tarval(pred), tp);
3696 DBG_OPT_CSTEVAL(oldn, n);
3697 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3698 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3699 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3700 DBG_OPT_CSTEVAL(oldn, n);
3704 } /* transform_node_Cast */
3707 * Transform a Proj(Load) with a non-null address.
3709 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3710 if (get_opt_ldst_only_null_ptr_exceptions()) {
3711 if (get_irn_mode(proj) == mode_X) {
3712 ir_node *load = get_Proj_pred(proj);
3714 /* get the Load address */
3715 const ir_node *addr = get_Load_ptr(load);
3716 const ir_node *confirm;
3718 if (value_not_null(addr, &confirm)) {
3719 if (confirm == NULL) {
3720 /* this node may float if it did not depend on a Confirm */
3721 set_irn_pinned(load, op_pin_state_floats);
3723 if (get_Proj_proj(proj) == pn_Load_X_except) {
3724 DBG_OPT_EXC_REM(proj);
3725 return get_irg_bad(current_ir_graph);
3727 ir_node *blk = get_nodes_block(load);
3728 return new_r_Jmp(current_ir_graph, blk);
3734 } /* transform_node_Proj_Load */
3737 * Transform a Proj(Store) with a non-null address.
3739 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3740 if (get_opt_ldst_only_null_ptr_exceptions()) {
3741 if (get_irn_mode(proj) == mode_X) {
3742 ir_node *store = get_Proj_pred(proj);
3744 /* get the load/store address */
3745 const ir_node *addr = get_Store_ptr(store);
3746 const ir_node *confirm;
3748 if (value_not_null(addr, &confirm)) {
3749 if (confirm == NULL) {
3750 /* this node may float if it did not depend on a Confirm */
3751 set_irn_pinned(store, op_pin_state_floats);
3753 if (get_Proj_proj(proj) == pn_Store_X_except) {
3754 DBG_OPT_EXC_REM(proj);
3755 return get_irg_bad(current_ir_graph);
3757 ir_node *blk = get_nodes_block(store);
3758 return new_r_Jmp(current_ir_graph, blk);
3764 } /* transform_node_Proj_Store */
3767 * Transform a Proj(Div) with a non-zero value.
3768 * Removes the exceptions and routes the memory to the NoMem node.
3770 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3771 ir_node *div = get_Proj_pred(proj);
3772 ir_node *b = get_Div_right(div);
3773 ir_node *res, *new_mem;
3774 const ir_node *confirm;
3777 if (value_not_zero(b, &confirm)) {
3778 /* div(x, y) && y != 0 */
3779 if (confirm == NULL) {
3780 /* we are sure we have a Const != 0 */
3781 new_mem = get_Div_mem(div);
3782 new_mem = skip_Pin(new_mem);
3783 set_Div_mem(div, new_mem);
3784 set_irn_pinned(div, op_pin_state_floats);
3787 proj_nr = get_Proj_proj(proj);
3789 case pn_Div_X_regular:
3790 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3792 case pn_Div_X_except:
3793 /* we found an exception handler, remove it */
3794 DBG_OPT_EXC_REM(proj);
3798 res = get_Div_mem(div);
3799 new_mem = get_irg_no_mem(current_ir_graph);
3802 /* This node can only float up to the Confirm block */
3803 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3805 set_irn_pinned(div, op_pin_state_floats);
3806 /* this is a Div without exception, we can remove the memory edge */
3807 set_Div_mem(div, new_mem);
3812 } /* transform_node_Proj_Div */
3815 * Transform a Proj(Mod) with a non-zero value.
3816 * Removes the exceptions and routes the memory to the NoMem node.
3818 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3819 ir_node *mod = get_Proj_pred(proj);
3820 ir_node *b = get_Mod_right(mod);
3821 ir_node *res, *new_mem;
3822 const ir_node *confirm;
3825 if (value_not_zero(b, &confirm)) {
3826 /* mod(x, y) && y != 0 */
3827 proj_nr = get_Proj_proj(proj);
3829 if (confirm == NULL) {
3830 /* we are sure we have a Const != 0 */
3831 new_mem = get_Mod_mem(mod);
3832 new_mem = skip_Pin(new_mem);
3833 set_Mod_mem(mod, new_mem);
3834 set_irn_pinned(mod, op_pin_state_floats);
3839 case pn_Mod_X_regular:
3840 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3842 case pn_Mod_X_except:
3843 /* we found an exception handler, remove it */
3844 DBG_OPT_EXC_REM(proj);
3848 res = get_Mod_mem(mod);
3849 new_mem = get_irg_no_mem(current_ir_graph);
3852 /* This node can only float up to the Confirm block */
3853 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3855 /* this is a Mod without exception, we can remove the memory edge */
3856 set_Mod_mem(mod, new_mem);
3859 if (get_Mod_left(mod) == b) {
3860 /* a % a = 0 if a != 0 */
3861 ir_mode *mode = get_irn_mode(proj);
3862 ir_node *res = new_Const(mode, get_mode_null(mode));
3864 DBG_OPT_CSTEVAL(mod, res);
3870 } /* transform_node_Proj_Mod */
3873 * Transform a Proj(DivMod) with a non-zero value.
3874 * Removes the exceptions and routes the memory to the NoMem node.
3876 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3877 ir_node *divmod = get_Proj_pred(proj);
3878 ir_node *b = get_DivMod_right(divmod);
3879 ir_node *res, *new_mem;
3880 const ir_node *confirm;
3883 if (value_not_zero(b, &confirm)) {
3884 /* DivMod(x, y) && y != 0 */
3885 proj_nr = get_Proj_proj(proj);
3887 if (confirm == NULL) {
3888 /* we are sure we have a Const != 0 */
3889 new_mem = get_DivMod_mem(divmod);
3890 new_mem = skip_Pin(new_mem);
3891 set_DivMod_mem(divmod, new_mem);
3892 set_irn_pinned(divmod, op_pin_state_floats);
3897 case pn_DivMod_X_regular:
3898 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3900 case pn_DivMod_X_except:
3901 /* we found an exception handler, remove it */
3902 DBG_OPT_EXC_REM(proj);
3906 res = get_DivMod_mem(divmod);
3907 new_mem = get_irg_no_mem(current_ir_graph);
3910 /* This node can only float up to the Confirm block */
3911 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3913 /* this is a DivMod without exception, we can remove the memory edge */
3914 set_DivMod_mem(divmod, new_mem);
3917 case pn_DivMod_res_mod:
3918 if (get_DivMod_left(divmod) == b) {
3919 /* a % a = 0 if a != 0 */
3920 ir_mode *mode = get_irn_mode(proj);
3921 ir_node *res = new_Const(mode, get_mode_null(mode));
3923 DBG_OPT_CSTEVAL(divmod, res);
3929 } /* transform_node_Proj_DivMod */
3932 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3934 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3935 if (get_opt_unreachable_code()) {
3936 ir_node *n = get_Proj_pred(proj);
3937 ir_node *b = get_Cond_selector(n);
3939 if (mode_is_int(get_irn_mode(b))) {
3940 tarval *tb = value_of(b);
3942 if (tb != tarval_bad) {
3943 /* we have a constant switch */
3944 long num = get_Proj_proj(proj);
3946 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3947 if (get_tarval_long(tb) == num) {
3948 /* Do NOT create a jump here, or we will have 2 control flow ops
3949 * in a block. This case is optimized away in optimize_cf(). */
3952 /* this case will NEVER be taken, kill it */
3953 return get_irg_bad(current_ir_graph);
3960 } /* transform_node_Proj_Cond */
3963 * Create a 0 constant of given mode.
3965 static ir_node *create_zero_const(ir_mode *mode) {
3966 tarval *tv = get_mode_null(mode);
3967 ir_node *cnst = new_Const(mode, tv);
3972 /* the order of the values is important! */
3973 typedef enum const_class {
3979 static const_class classify_const(const ir_node* n)
3981 if (is_Const(n)) return const_const;
3982 if (is_irn_constlike(n)) return const_like;
3987 * Determines whether r is more constlike or has a larger index (in that order)
3990 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3992 const const_class l_order = classify_const(l);
3993 const const_class r_order = classify_const(r);
3995 l_order > r_order ||
3996 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4000 * Normalizes and optimizes Cmp nodes.
4002 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4003 ir_node *n = get_Proj_pred(proj);
4004 ir_node *left = get_Cmp_left(n);
4005 ir_node *right = get_Cmp_right(n);
4008 ir_mode *mode = NULL;
4009 long proj_nr = get_Proj_proj(proj);
4011 /* we can evaluate some cases directly */
4014 return new_Const(mode_b, get_tarval_b_false());
4016 return new_Const(mode_b, get_tarval_b_true());
4018 if (!mode_is_float(get_irn_mode(left)))
4019 return new_Const(mode_b, get_tarval_b_true());
4025 /* remove Casts of both sides */
4026 left = skip_Cast(left);
4027 right = skip_Cast(right);
4029 /* Remove unnecessary conversions */
4030 /* TODO handle constants */
4031 if (is_Conv(left) && is_Conv(right)) {
4032 ir_mode *mode = get_irn_mode(left);
4033 ir_node *op_left = get_Conv_op(left);
4034 ir_node *op_right = get_Conv_op(right);
4035 ir_mode *mode_left = get_irn_mode(op_left);
4036 ir_mode *mode_right = get_irn_mode(op_right);
4038 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4039 && mode_left != mode_b && mode_right != mode_b) {
4040 ir_graph *irg = current_ir_graph;
4041 ir_node *block = get_nodes_block(n);
4043 if (mode_left == mode_right) {
4047 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4048 } else if (smaller_mode(mode_left, mode_right)) {
4049 left = new_r_Conv(irg, block, op_left, mode_right);
4052 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4053 } else if (smaller_mode(mode_right, mode_left)) {
4055 right = new_r_Conv(irg, block, op_right, mode_left);
4057 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4062 /* remove operation on both sides if possible */
4063 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4065 * The following operations are NOT safe for floating point operations, for instance
4066 * 1.0 + inf == 2.0 + inf, =/=> x == y
4068 if (mode_is_int(get_irn_mode(left))) {
4069 unsigned lop = get_irn_opcode(left);
4071 if (lop == get_irn_opcode(right)) {
4072 ir_node *ll, *lr, *rl, *rr;
4074 /* same operation on both sides, try to remove */
4078 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4079 left = get_unop_op(left);
4080 right = get_unop_op(right);
4082 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4085 ll = get_Add_left(left);
4086 lr = get_Add_right(left);
4087 rl = get_Add_left(right);
4088 rr = get_Add_right(right);
4091 /* X + a CMP X + b ==> a CMP b */
4095 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4096 } else if (ll == rr) {
4097 /* X + a CMP b + X ==> a CMP b */
4101 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4102 } else if (lr == rl) {
4103 /* a + X CMP X + b ==> a CMP b */
4107 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4108 } else if (lr == rr) {
4109 /* a + X CMP b + X ==> a CMP b */
4113 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4117 ll = get_Sub_left(left);
4118 lr = get_Sub_right(left);
4119 rl = get_Sub_left(right);
4120 rr = get_Sub_right(right);
4123 /* X - a CMP X - b ==> a CMP b */
4127 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4128 } else if (lr == rr) {
4129 /* a - X CMP b - X ==> a CMP b */
4133 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4137 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4138 /* a ROTL X CMP b ROTL X ==> a CMP b */
4139 left = get_Rotl_left(left);
4140 right = get_Rotl_left(right);
4142 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4150 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4151 if (is_Add(left) || is_Sub(left)) {
4152 ir_node *ll = get_binop_left(left);
4153 ir_node *lr = get_binop_right(left);
4155 if (lr == right && is_Add(left)) {
4162 right = create_zero_const(get_irn_mode(left));
4164 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4167 if (is_Add(right) || is_Sub(right)) {
4168 ir_node *rl = get_binop_left(right);
4169 ir_node *rr = get_binop_right(right);
4171 if (rr == left && is_Add(right)) {
4178 right = create_zero_const(get_irn_mode(left));
4180 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4183 if (is_And(left) && is_Const(right)) {
4184 ir_node *ll = get_binop_left(left);
4185 ir_node *lr = get_binop_right(left);
4186 if (is_Shr(ll) && is_Const(lr)) {
4187 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4188 ir_node *block = get_nodes_block(n);
4189 ir_mode *mode = get_irn_mode(left);
4191 ir_node *llr = get_Shr_right(ll);
4192 if (is_Const(llr)) {
4193 ir_graph *irg = current_ir_graph;
4194 dbg_info *dbg = get_irn_dbg_info(left);
4196 tarval *c1 = get_Const_tarval(llr);
4197 tarval *c2 = get_Const_tarval(lr);
4198 tarval *c3 = get_Const_tarval(right);
4199 tarval *mask = tarval_shl(c2, c1);
4200 tarval *value = tarval_shl(c3, c1);
4202 left = new_rd_And(dbg, irg, block, get_Shr_left(ll), new_Const(mode, mask), mode);
4203 right = new_Const(mode, value);
4208 } /* mode_is_int(...) */
4209 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4211 /* replace mode_b compares with ands/ors */
4212 if (get_irn_mode(left) == mode_b) {
4213 ir_graph *irg = current_ir_graph;
4214 ir_node *block = get_nodes_block(n);
4218 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4219 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4220 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4221 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4222 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4223 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4224 default: bres = NULL;
4227 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4233 * First step: normalize the compare op
4234 * by placing the constant on the right side
4235 * or moving the lower address node to the left.
4237 if (!operands_are_normalized(left, right)) {
4243 proj_nr = get_inversed_pnc(proj_nr);
4248 * Second step: Try to reduce the magnitude
4249 * of a constant. This may help to generate better code
4250 * later and may help to normalize more compares.
4251 * Of course this is only possible for integer values.
4253 tv = value_of(right);
4254 if (tv != tarval_bad) {
4255 mode = get_irn_mode(right);
4257 /* TODO extend to arbitrary constants */
4258 if (is_Conv(left) && tarval_is_null(tv)) {
4259 ir_node *op = get_Conv_op(left);
4260 ir_mode *op_mode = get_irn_mode(op);
4263 * UpConv(x) REL 0 ==> x REL 0
4265 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4266 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4267 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4268 tv = get_mode_null(op_mode);
4272 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4276 if (tv != tarval_bad) {
4277 /* the following optimization is possible on modes without Overflow
4278 * on Unary Minus or on == and !=:
4279 * -a CMP c ==> a swap(CMP) -c
4281 * Beware: for two-complement Overflow may occur, so only == and != can
4282 * be optimized, see this:
4283 * -MININT < 0 =/=> MININT > 0 !!!
4285 if (is_Minus(left) &&
4286 (!mode_overflow_on_unary_Minus(mode) ||
4287 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4288 tv = tarval_neg(tv);
4290 if (tv != tarval_bad) {
4291 left = get_Minus_op(left);
4292 proj_nr = get_inversed_pnc(proj_nr);
4294 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4296 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4297 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4298 tv = tarval_not(tv);
4300 if (tv != tarval_bad) {
4301 left = get_Not_op(left);
4303 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4307 /* for integer modes, we have more */
4308 if (mode_is_int(mode)) {
4309 /* Ne includes Unordered which is not possible on integers.
4310 * However, frontends often use this wrong, so fix it here */
4311 if (proj_nr & pn_Cmp_Uo) {
4312 proj_nr &= ~pn_Cmp_Uo;
4313 set_Proj_proj(proj, proj_nr);
4316 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4317 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4318 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4319 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4321 if (tv != tarval_bad) {
4322 proj_nr ^= pn_Cmp_Eq;
4324 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4327 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4328 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4329 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4330 tv = tarval_add(tv, get_mode_one(mode));
4332 if (tv != tarval_bad) {
4333 proj_nr ^= pn_Cmp_Eq;
4335 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4339 /* the following reassociations work only for == and != */
4340 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4342 #if 0 /* Might be not that good in general */
4343 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4344 if (tarval_is_null(tv) && is_Sub(left)) {
4345 right = get_Sub_right(left);
4346 left = get_Sub_left(left);
4348 tv = value_of(right);
4350 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4354 if (tv != tarval_bad) {
4355 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4357 ir_node *c1 = get_Sub_right(left);
4358 tarval *tv2 = value_of(c1);
4360 if (tv2 != tarval_bad) {
4361 tv2 = tarval_add(tv, value_of(c1));
4363 if (tv2 != tarval_bad) {
4364 left = get_Sub_left(left);
4367 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4371 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4372 else if (is_Add(left)) {
4373 ir_node *a_l = get_Add_left(left);
4374 ir_node *a_r = get_Add_right(left);
4378 if (is_Const(a_l)) {
4380 tv2 = value_of(a_l);
4383 tv2 = value_of(a_r);
4386 if (tv2 != tarval_bad) {
4387 tv2 = tarval_sub(tv, tv2, NULL);
4389 if (tv2 != tarval_bad) {
4393 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4397 /* -a == c ==> a == -c, -a != c ==> a != -c */
4398 else if (is_Minus(left)) {
4399 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4401 if (tv2 != tarval_bad) {
4402 left = get_Minus_op(left);
4405 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4410 /* the following reassociations work only for <= */
4411 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4412 if (tv != tarval_bad) {
4413 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4414 if (is_Abs(left)) { // TODO something is missing here
4420 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4421 switch (get_irn_opcode(left)) {
4425 c1 = get_And_right(left);
4428 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4429 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4431 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4433 /* TODO: move to constant evaluation */
4434 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4435 c1 = new_Const(mode_b, tv);
4436 DBG_OPT_CSTEVAL(proj, c1);
4440 if (tarval_is_single_bit(tv)) {
4442 * optimization for AND:
4444 * And(x, C) == C ==> And(x, C) != 0
4445 * And(x, C) != C ==> And(X, C) == 0
4447 * if C is a single Bit constant.
4450 /* check for Constant's match. We have check hare the tarvals,
4451 because our const might be changed */
4452 if (get_Const_tarval(c1) == tv) {
4453 /* fine: do the transformation */
4454 tv = get_mode_null(get_tarval_mode(tv));
4455 proj_nr ^= pn_Cmp_Leg;
4457 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4463 c1 = get_Or_right(left);
4464 if (is_Const(c1) && tarval_is_null(tv)) {
4466 * Or(x, C) == 0 && C != 0 ==> FALSE
4467 * Or(x, C) != 0 && C != 0 ==> TRUE
4469 if (! tarval_is_null(get_Const_tarval(c1))) {
4470 /* TODO: move to constant evaluation */
4471 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4472 c1 = new_Const(mode_b, tv);
4473 DBG_OPT_CSTEVAL(proj, c1);
4480 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4482 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4485 c1 = get_Shl_right(left);
4487 tarval *tv1 = get_Const_tarval(c1);
4488 ir_mode *mode = get_irn_mode(left);
4489 tarval *minus1 = get_mode_all_one(mode);
4490 tarval *amask = tarval_shr(minus1, tv1);
4491 tarval *cmask = tarval_shl(minus1, tv1);
4494 if (tarval_and(tv, cmask) != tv) {
4495 /* condition not met */
4496 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4497 c1 = new_Const(mode_b, tv);
4498 DBG_OPT_CSTEVAL(proj, c1);
4501 sl = get_Shl_left(left);
4502 blk = get_nodes_block(n);
4503 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4504 tv = tarval_shr(tv, tv1);
4506 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4511 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4513 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4516 c1 = get_Shr_right(left);
4518 tarval *tv1 = get_Const_tarval(c1);
4519 ir_mode *mode = get_irn_mode(left);
4520 tarval *minus1 = get_mode_all_one(mode);
4521 tarval *amask = tarval_shl(minus1, tv1);
4522 tarval *cmask = tarval_shr(minus1, tv1);
4525 if (tarval_and(tv, cmask) != tv) {
4526 /* condition not met */
4527 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4528 c1 = new_Const(mode_b, tv);
4529 DBG_OPT_CSTEVAL(proj, c1);
4532 sl = get_Shr_left(left);
4533 blk = get_nodes_block(n);
4534 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4535 tv = tarval_shl(tv, tv1);
4537 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4542 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4544 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4547 c1 = get_Shrs_right(left);
4549 tarval *tv1 = get_Const_tarval(c1);
4550 ir_mode *mode = get_irn_mode(left);
4551 tarval *minus1 = get_mode_all_one(mode);
4552 tarval *amask = tarval_shl(minus1, tv1);
4553 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4556 cond = tarval_sub(cond, tv1, NULL);
4557 cond = tarval_shrs(tv, cond);
4559 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4560 /* condition not met */
4561 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4562 c1 = new_Const(mode_b, tv);
4563 DBG_OPT_CSTEVAL(proj, c1);
4566 sl = get_Shrs_left(left);
4567 blk = get_nodes_block(n);
4568 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4569 tv = tarval_shl(tv, tv1);
4571 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4576 } /* tarval != bad */
4579 if (changed & 2) /* need a new Const */
4580 right = new_Const(mode, tv);
4582 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4583 ir_node *op = get_Proj_pred(left);
4585 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4586 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4587 ir_node *c = get_binop_right(op);
4590 tarval *tv = get_Const_tarval(c);
4592 if (tarval_is_single_bit(tv)) {
4593 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4594 ir_node *v = get_binop_left(op);
4595 ir_node *blk = get_irn_n(op, -1);
4596 ir_mode *mode = get_irn_mode(v);
4598 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4599 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4601 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4608 ir_node *block = get_nodes_block(n);
4610 /* create a new compare */
4611 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4612 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4616 } /* transform_node_Proj_Cmp */
4619 * Optimize CopyB(mem, x, x) into a Nop.
4621 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4622 ir_node *copyb = get_Proj_pred(proj);
4623 ir_node *a = get_CopyB_dst(copyb);
4624 ir_node *b = get_CopyB_src(copyb);
4627 switch (get_Proj_proj(proj)) {
4628 case pn_CopyB_X_regular:
4629 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4630 DBG_OPT_EXC_REM(proj);
4631 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4633 case pn_CopyB_M_except:
4634 case pn_CopyB_X_except:
4635 DBG_OPT_EXC_REM(proj);
4636 proj = get_irg_bad(current_ir_graph);
4643 } /* transform_node_Proj_CopyB */
4646 * Optimize Bounds(idx, idx, upper) into idx.
4648 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4649 ir_node *oldn = proj;
4650 ir_node *bound = get_Proj_pred(proj);
4651 ir_node *idx = get_Bound_index(bound);
4652 ir_node *pred = skip_Proj(idx);
4655 if (idx == get_Bound_lower(bound))
4657 else if (is_Bound(pred)) {
4659 * idx was Bounds checked in the same MacroBlock previously,
4660 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4662 ir_node *lower = get_Bound_lower(bound);
4663 ir_node *upper = get_Bound_upper(bound);
4664 if (get_Bound_lower(pred) == lower &&
4665 get_Bound_upper(pred) == upper &&
4666 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4668 * One could expect that we simply return the previous
4669 * Bound here. However, this would be wrong, as we could
4670 * add an exception Proj to a new location then.
4671 * So, we must turn in into a tuple.
4677 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4678 switch (get_Proj_proj(proj)) {
4680 DBG_OPT_EXC_REM(proj);
4681 proj = get_Bound_mem(bound);
4683 case pn_Bound_X_except:
4684 DBG_OPT_EXC_REM(proj);
4685 proj = get_irg_bad(current_ir_graph);
4689 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4691 case pn_Bound_X_regular:
4692 DBG_OPT_EXC_REM(proj);
4693 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4700 } /* transform_node_Proj_Bound */
4703 * Does all optimizations on nodes that must be done on it's Proj's
4704 * because of creating new nodes.
4706 static ir_node *transform_node_Proj(ir_node *proj) {
4707 ir_node *n = get_Proj_pred(proj);
4709 if (n->op->ops.transform_node_Proj)
4710 return n->op->ops.transform_node_Proj(proj);
4712 } /* transform_node_Proj */
4715 * Move Confirms down through Phi nodes.
4717 static ir_node *transform_node_Phi(ir_node *phi) {
4719 ir_mode *mode = get_irn_mode(phi);
4721 if (mode_is_reference(mode)) {
4722 n = get_irn_arity(phi);
4724 /* Beware of Phi0 */
4726 ir_node *pred = get_irn_n(phi, 0);
4727 ir_node *bound, *new_Phi, *block, **in;
4730 if (! is_Confirm(pred))
4733 bound = get_Confirm_bound(pred);
4734 pnc = get_Confirm_cmp(pred);
4736 NEW_ARR_A(ir_node *, in, n);
4737 in[0] = get_Confirm_value(pred);
4739 for (i = 1; i < n; ++i) {
4740 pred = get_irn_n(phi, i);
4742 if (! is_Confirm(pred) ||
4743 get_Confirm_bound(pred) != bound ||
4744 get_Confirm_cmp(pred) != pnc)
4746 in[i] = get_Confirm_value(pred);
4748 /* move the Confirm nodes "behind" the Phi */
4749 block = get_irn_n(phi, -1);
4750 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4751 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4755 } /* transform_node_Phi */
4758 * Returns the operands of a commutative bin-op, if one operand is
4759 * a const, it is returned as the second one.
4761 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4762 ir_node *op_a = get_binop_left(binop);
4763 ir_node *op_b = get_binop_right(binop);
4765 assert(is_op_commutative(get_irn_op(binop)));
4767 if (is_Const(op_a)) {
4774 } /* get_comm_Binop_Ops */
4777 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4778 * Such pattern may arise in bitfield stores.
4780 * value c4 value c4 & c2
4781 * AND c3 AND c1 | c3
4788 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4791 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4794 ir_node *and_l, *c3;
4795 ir_node *value, *c4;
4796 ir_node *new_and, *new_const, *block;
4797 ir_mode *mode = get_irn_mode(or);
4799 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4802 get_comm_Binop_Ops(or, &and, &c1);
4803 if (!is_Const(c1) || !is_And(and))
4806 get_comm_Binop_Ops(and, &or_l, &c2);
4810 tv1 = get_Const_tarval(c1);
4811 tv2 = get_Const_tarval(c2);
4813 tv = tarval_or(tv1, tv2);
4814 if (tarval_is_all_one(tv)) {
4815 /* the AND does NOT clear a bit with isn't set by the OR */
4816 set_Or_left(or, or_l);
4817 set_Or_right(or, c1);
4819 /* check for more */
4826 get_comm_Binop_Ops(or_l, &and_l, &c3);
4827 if (!is_Const(c3) || !is_And(and_l))
4830 get_comm_Binop_Ops(and_l, &value, &c4);
4834 /* ok, found the pattern, check for conditions */
4835 assert(mode == get_irn_mode(and));
4836 assert(mode == get_irn_mode(or_l));
4837 assert(mode == get_irn_mode(and_l));
4839 tv3 = get_Const_tarval(c3);
4840 tv4 = get_Const_tarval(c4);
4842 tv = tarval_or(tv4, tv2);
4843 if (!tarval_is_all_one(tv)) {
4844 /* have at least one 0 at the same bit position */
4848 n_tv4 = tarval_not(tv4);
4849 if (tv3 != tarval_and(tv3, n_tv4)) {
4850 /* bit in the or_mask is outside the and_mask */
4854 n_tv2 = tarval_not(tv2);
4855 if (tv1 != tarval_and(tv1, n_tv2)) {
4856 /* bit in the or_mask is outside the and_mask */
4860 /* ok, all conditions met */
4861 block = get_irn_n(or, -1);
4863 new_and = new_r_And(current_ir_graph, block,
4864 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
4866 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
4868 set_Or_left(or, new_and);
4869 set_Or_right(or, new_const);
4871 /* check for more */
4873 } /* transform_node_Or_bf_store */
4876 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4878 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4879 ir_mode *mode = get_irn_mode(or);
4880 ir_node *shl, *shr, *block;
4881 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4884 if (! mode_is_int(mode))
4887 shl = get_binop_left(or);
4888 shr = get_binop_right(or);
4897 } else if (!is_Shl(shl)) {
4899 } else if (!is_Shr(shr)) {
4902 x = get_Shl_left(shl);
4903 if (x != get_Shr_left(shr))
4906 c1 = get_Shl_right(shl);
4907 c2 = get_Shr_right(shr);
4908 if (is_Const(c1) && is_Const(c2)) {
4909 tv1 = get_Const_tarval(c1);
4910 if (! tarval_is_long(tv1))
4913 tv2 = get_Const_tarval(c2);
4914 if (! tarval_is_long(tv2))
4917 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4918 != (int) get_mode_size_bits(mode))
4921 /* yet, condition met */
4922 block = get_nodes_block(or);
4924 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4926 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4933 rotval = sub; /* a Rot right is not supported, so use a rot left */
4934 } else if (is_Sub(c2)) {
4940 if (get_Sub_right(sub) != v)
4943 c1 = get_Sub_left(sub);
4947 tv1 = get_Const_tarval(c1);
4948 if (! tarval_is_long(tv1))
4951 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4954 /* yet, condition met */
4955 block = get_nodes_block(or);
4957 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4959 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4961 } /* transform_node_Or_Rotl */
4966 static ir_node *transform_node_Or(ir_node *n) {
4967 ir_node *c, *oldn = n;
4968 ir_node *a = get_Or_left(n);
4969 ir_node *b = get_Or_right(n);
4972 if (is_Not(a) && is_Not(b)) {
4973 /* ~a | ~b = ~(a&b) */
4974 ir_node *block = get_nodes_block(n);
4976 mode = get_irn_mode(n);
4979 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4980 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4981 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4985 /* we can evaluate 2 Projs of the same Cmp */
4986 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4987 ir_node *pred_a = get_Proj_pred(a);
4988 ir_node *pred_b = get_Proj_pred(b);
4989 if (pred_a == pred_b) {
4990 dbg_info *dbgi = get_irn_dbg_info(n);
4991 ir_node *block = get_nodes_block(pred_a);
4992 pn_Cmp pn_a = get_Proj_proj(a);
4993 pn_Cmp pn_b = get_Proj_proj(b);
4994 /* yes, we can simply calculate with pncs */
4995 pn_Cmp new_pnc = pn_a | pn_b;
4997 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
5002 mode = get_irn_mode(n);
5003 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
5005 n = transform_node_Or_bf_store(n);
5006 n = transform_node_Or_Rotl(n);
5010 n = transform_bitwise_distributive(n, transform_node_Or);
5013 } /* transform_node_Or */
5017 static ir_node *transform_node(ir_node *n);
5020 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5022 * Should be moved to reassociation?
5024 static ir_node *transform_node_shift(ir_node *n) {
5025 ir_node *left, *right;
5027 tarval *tv1, *tv2, *res;
5028 ir_node *in[2], *irn, *block;
5030 left = get_binop_left(n);
5032 /* different operations */
5033 if (get_irn_op(left) != get_irn_op(n))
5036 right = get_binop_right(n);
5037 tv1 = value_of(right);
5038 if (tv1 == tarval_bad)
5041 tv2 = value_of(get_binop_right(left));
5042 if (tv2 == tarval_bad)
5045 res = tarval_add(tv1, tv2);
5046 mode = get_irn_mode(n);
5048 /* beware: a simple replacement works only, if res < modulo shift */
5050 int modulo_shf = get_mode_modulo_shift(mode);
5051 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5052 if (modulo_shf > 0) {
5053 tarval *modulo = new_tarval_from_long(modulo_shf,
5054 get_tarval_mode(res));
5056 /* shifting too much */
5057 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5059 ir_graph *irg = get_irn_irg(n);
5060 ir_node *block = get_nodes_block(n);
5061 dbg_info *dbgi = get_irn_dbg_info(n);
5062 ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu));
5063 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5067 return new_Const(mode, get_mode_null(mode));
5071 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5074 /* ok, we can replace it */
5075 block = get_nodes_block(n);
5077 in[0] = get_binop_left(left);
5078 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
5080 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5082 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5084 return transform_node(irn);
5085 } /* transform_node_shift */
5088 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5090 * - and, or, xor instead of &
5091 * - Shl, Shr, Shrs, rotl instead of >>
5092 * (with a special case for Or/Xor + Shrs)
5094 static ir_node *transform_node_bitop_shift(ir_node *n) {
5096 ir_node *right = get_binop_right(n);
5097 ir_mode *mode = get_irn_mode(n);
5098 ir_node *bitop_left;
5099 ir_node *bitop_right;
5111 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5113 if (!is_Const(right))
5116 left = get_binop_left(n);
5117 op_left = get_irn_op(left);
5118 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5121 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5122 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5123 /* TODO: test if sign bit is affectes */
5127 bitop_right = get_binop_right(left);
5128 if (!is_Const(bitop_right))
5131 bitop_left = get_binop_left(left);
5133 irg = get_irn_irg(n);
5134 block = get_nodes_block(n);
5135 dbgi = get_irn_dbg_info(n);
5136 tv1 = get_Const_tarval(bitop_right);
5137 tv2 = get_Const_tarval(right);
5139 assert(get_tarval_mode(tv1) == mode);
5142 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5143 tv_shift = tarval_shl(tv1, tv2);
5144 } else if(is_Shr(n)) {
5145 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5146 tv_shift = tarval_shr(tv1, tv2);
5147 } else if(is_Shrs(n)) {
5148 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5149 tv_shift = tarval_shrs(tv1, tv2);
5152 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5153 tv_shift = tarval_rotl(tv1, tv2);
5156 assert(get_tarval_mode(tv_shift) == mode);
5157 new_const = new_Const(mode, tv_shift);
5159 if (op_left == op_And) {
5160 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5161 } else if(op_left == op_Or) {
5162 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5164 assert(op_left == op_Eor);
5165 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5173 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5175 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5176 * (also with x >>s c1 when c1>=c2)
5178 static ir_node *transform_node_shl_shr(ir_node *n) {
5180 ir_node *right = get_binop_right(n);
5196 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5198 if (!is_Const(right))
5201 left = get_binop_left(n);
5202 mode = get_irn_mode(n);
5203 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5204 ir_node *shr_right = get_binop_right(left);
5206 if (!is_Const(shr_right))
5209 x = get_binop_left(left);
5210 tv_shr = get_Const_tarval(shr_right);
5211 tv_shl = get_Const_tarval(right);
5213 if (is_Shrs(left)) {
5214 /* shrs variant only allowed if c1 >= c2 */
5215 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5218 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5221 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5223 tv_mask = tarval_shl(tv_mask, tv_shl);
5224 } else if(is_Shr(n) && is_Shl(left)) {
5225 ir_node *shl_right = get_Shl_right(left);
5227 if (!is_Const(shl_right))
5230 x = get_Shl_left(left);
5231 tv_shr = get_Const_tarval(right);
5232 tv_shl = get_Const_tarval(shl_right);
5234 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5235 tv_mask = tarval_shr(tv_mask, tv_shr);
5240 assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr));
5241 assert(tv_mask != tarval_bad);
5242 assert(get_tarval_mode(tv_mask) == mode);
5244 irg = get_irn_irg(n);
5245 block = get_nodes_block(n);
5246 dbgi = get_irn_dbg_info(n);
5248 pnc = tarval_cmp(tv_shl, tv_shr);
5249 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5250 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5251 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5253 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5255 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5258 assert(pnc == pn_Cmp_Gt);
5259 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5260 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5261 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5264 new_const = new_Const(mode, tv_mask);
5265 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5273 static ir_node *transform_node_Shr(ir_node *n) {
5274 ir_node *c, *oldn = n;
5275 ir_node *left = get_Shr_left(n);
5276 ir_node *right = get_Shr_right(n);
5277 ir_mode *mode = get_irn_mode(n);
5279 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5280 n = transform_node_shift(n);
5283 n = transform_node_shl_shr(n);
5285 n = transform_node_bitop_shift(n);
5288 } /* transform_node_Shr */
5293 static ir_node *transform_node_Shrs(ir_node *n) {
5294 ir_node *c, *oldn = n;
5295 ir_node *a = get_Shrs_left(n);
5296 ir_node *b = get_Shrs_right(n);
5297 ir_mode *mode = get_irn_mode(n);
5299 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5300 n = transform_node_shift(n);
5303 n = transform_node_bitop_shift(n);
5306 } /* transform_node_Shrs */
5311 static ir_node *transform_node_Shl(ir_node *n) {
5312 ir_node *c, *oldn = n;
5313 ir_node *a = get_Shl_left(n);
5314 ir_node *b = get_Shl_right(n);
5315 ir_mode *mode = get_irn_mode(n);
5317 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5318 n = transform_node_shift(n);
5321 n = transform_node_shl_shr(n);
5323 n = transform_node_bitop_shift(n);
5326 } /* transform_node_Shl */
5331 static ir_node *transform_node_Rotl(ir_node *n) {
5332 ir_node *c, *oldn = n;
5333 ir_node *a = get_Rotl_left(n);
5334 ir_node *b = get_Rotl_right(n);
5335 ir_mode *mode = get_irn_mode(n);
5337 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5338 n = transform_node_shift(n);
5341 n = transform_node_bitop_shift(n);
5344 } /* transform_node_Rotl */
5349 static ir_node *transform_node_Conv(ir_node *n) {
5350 ir_node *c, *oldn = n;
5351 ir_mode *mode = get_irn_mode(n);
5352 ir_node *a = get_Conv_op(n);
5354 if (mode != mode_b && is_const_Phi(a)) {
5355 /* Do NOT optimize mode_b Conv's, this leads to remaining
5356 * Phib nodes later, because the conv_b_lower operation
5357 * is instantly reverted, when it tries to insert a Convb.
5359 c = apply_conv_on_phi(a, mode);
5361 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5366 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5367 return new_r_Unknown(current_ir_graph, mode);
5370 if (mode_is_reference(mode) &&
5371 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5373 ir_node *l = get_Add_left(a);
5374 ir_node *r = get_Add_right(a);
5375 dbg_info *dbgi = get_irn_dbg_info(a);
5376 ir_node *block = get_nodes_block(n);
5378 ir_node *lop = get_Conv_op(l);
5379 if(get_irn_mode(lop) == mode) {
5380 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5381 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5386 ir_node *rop = get_Conv_op(r);
5387 if(get_irn_mode(rop) == mode) {
5388 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5389 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5396 } /* transform_node_Conv */
5399 * Remove dead blocks and nodes in dead blocks
5400 * in keep alive list. We do not generate a new End node.
5402 static ir_node *transform_node_End(ir_node *n) {
5403 int i, j, n_keepalives = get_End_n_keepalives(n);
5406 NEW_ARR_A(ir_node *, in, n_keepalives);
5408 for (i = j = 0; i < n_keepalives; ++i) {
5409 ir_node *ka = get_End_keepalive(n, i);
5411 if (! is_Block_dead(ka)) {
5415 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5417 } else if (is_Bad(ka)) {
5418 /* no need to keep Bad */
5423 if (j != n_keepalives)
5424 set_End_keepalives(n, j, in);
5426 } /* transform_node_End */
5428 /** returns 1 if a == -b */
5429 static int is_negated_value(ir_node *a, ir_node *b) {
5430 if (is_Minus(a) && get_Minus_op(a) == b)
5432 if (is_Minus(b) && get_Minus_op(b) == a)
5434 if (is_Sub(a) && is_Sub(b)) {
5435 ir_node *a_left = get_Sub_left(a);
5436 ir_node *a_right = get_Sub_right(a);
5437 ir_node *b_left = get_Sub_left(b);
5438 ir_node *b_right = get_Sub_right(b);
5440 if (a_left == b_right && a_right == b_left)
5448 * Optimize a Mux into some simpler cases.
5450 static ir_node *transform_node_Mux(ir_node *n) {
5451 ir_node *oldn = n, *sel = get_Mux_sel(n);
5452 ir_mode *mode = get_irn_mode(n);
5453 ir_node *t = get_Mux_true(n);
5454 ir_node *f = get_Mux_false(n);
5455 ir_graph *irg = current_ir_graph;
5457 /* first normalization step: move a possible zero to the false case */
5459 ir_node *cmp = get_Proj_pred(sel);
5462 if (is_Const(t) && is_Const_null(t)) {
5465 /* Mux(x, 0, y) => Mux(x, y, 0) */
5466 pn_Cmp pnc = get_Proj_proj(sel);
5467 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5468 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5469 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5477 /* note: after normalization, false can only happen on default */
5478 if (mode == mode_b) {
5479 dbg_info *dbg = get_irn_dbg_info(n);
5480 ir_node *block = get_nodes_block(n);
5481 ir_graph *irg = current_ir_graph;
5484 tarval *tv_t = get_Const_tarval(t);
5485 if (tv_t == tarval_b_true) {
5487 /* Muxb(sel, true, false) = sel */
5488 assert(get_Const_tarval(f) == tarval_b_false);
5489 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5492 /* Muxb(sel, true, x) = Or(sel, x) */
5493 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5494 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5498 } else if (is_Const(f)) {
5499 tarval *tv_f = get_Const_tarval(f);
5500 if (tv_f == tarval_b_true) {
5501 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5502 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5503 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5504 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5507 /* Muxb(sel, x, false) = And(sel, x) */
5508 assert(tv_f == tarval_b_false);
5509 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5510 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5516 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5517 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5518 tarval *a = get_Const_tarval(t);
5519 tarval *b = get_Const_tarval(f);
5520 tarval *null = get_tarval_null(mode);
5523 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5524 diff = tarval_sub(a, b, NULL);
5527 diff = tarval_sub(b, a, NULL);
5531 if (diff == get_tarval_one(mode) && min != null) {
5532 dbg_info *dbg = get_irn_dbg_info(n);
5533 ir_node *block = get_nodes_block(n);
5534 ir_graph *irg = current_ir_graph;
5535 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5536 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5537 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5538 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5544 ir_node *cmp = get_Proj_pred(sel);
5545 long pn = get_Proj_proj(sel);
5548 * Note: normalization puts the constant on the right side,
5549 * so we check only one case.
5551 * Note further that these optimization work even for floating point
5552 * with NaN's because -NaN == NaN.
5553 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5557 ir_node *cmp_r = get_Cmp_right(cmp);
5558 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5559 ir_node *block = get_nodes_block(n);
5560 ir_node *cmp_l = get_Cmp_left(cmp);
5562 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5565 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5566 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5568 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5569 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5571 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5573 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5574 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5576 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5577 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5579 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5581 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5586 if (mode_is_int(mode)) {
5588 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5589 /* Mux((a & b) != 0, c, 0) */
5590 ir_node *and_r = get_And_right(cmp_l);
5593 if (and_r == t && f == cmp_r) {
5594 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5595 if (pn == pn_Cmp_Lg) {
5596 /* Mux((a & 2^C) != 0, 2^C, 0) */
5598 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5600 /* Mux((a & 2^C) == 0, 2^C, 0) */
5601 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5602 block, cmp_l, t, mode);
5603 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5608 if (is_Shl(and_r)) {
5609 ir_node *shl_l = get_Shl_left(and_r);
5610 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5611 if (and_r == t && f == cmp_r) {
5612 if (pn == pn_Cmp_Lg) {
5613 /* (a & (1 << n)) != 0, (1 << n), 0) */
5615 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5617 /* (a & (1 << n)) == 0, (1 << n), 0) */
5618 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5619 block, cmp_l, t, mode);
5620 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5626 and_l = get_And_left(cmp_l);
5627 if (is_Shl(and_l)) {
5628 ir_node *shl_l = get_Shl_left(and_l);
5629 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5630 if (and_l == t && f == cmp_r) {
5631 if (pn == pn_Cmp_Lg) {
5632 /* ((1 << n) & a) != 0, (1 << n), 0) */
5634 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5636 /* ((1 << n) & a) == 0, (1 << n), 0) */
5637 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5638 block, cmp_l, t, mode);
5639 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5650 return arch_transform_node_Mux(n);
5651 } /* transform_node_Mux */
5654 * optimize Sync nodes that have other syncs as input we simply add the inputs
5655 * of the other sync to our own inputs
5657 static ir_node *transform_node_Sync(ir_node *n) {
5658 int arity = get_Sync_n_preds(n);
5661 for (i = 0; i < arity;) {
5662 ir_node *pred = get_Sync_pred(n, i);
5666 if (!is_Sync(pred)) {
5674 pred_arity = get_Sync_n_preds(pred);
5675 for (j = 0; j < pred_arity; ++j) {
5676 ir_node *pred_pred = get_Sync_pred(pred, j);
5681 add_irn_n(n, pred_pred);
5685 if (get_Sync_pred(n, k) == pred_pred) break;
5690 /* rehash the sync node */
5691 add_identities(current_ir_graph->value_table, n);
5697 * Tries several [inplace] [optimizing] transformations and returns an
5698 * equivalent node. The difference to equivalent_node() is that these
5699 * transformations _do_ generate new nodes, and thus the old node must
5700 * not be freed even if the equivalent node isn't the old one.
5702 static ir_node *transform_node(ir_node *n) {
5706 * Transform_node is the only "optimizing transformation" that might
5707 * return a node with a different opcode. We iterate HERE until fixpoint
5708 * to get the final result.
5712 if (n->op->ops.transform_node)
5713 n = n->op->ops.transform_node(n);
5714 } while (oldn != n);
5717 } /* transform_node */
5720 * Sets the default transform node operation for an ir_op_ops.
5722 * @param code the opcode for the default operation
5723 * @param ops the operations initialized
5728 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5732 ops->transform_node = transform_node_##a; \
5734 #define CASE_PROJ(a) \
5736 ops->transform_node_Proj = transform_node_Proj_##a; \
5738 #define CASE_PROJ_EX(a) \
5740 ops->transform_node = transform_node_##a; \
5741 ops->transform_node_Proj = transform_node_Proj_##a; \
5750 CASE_PROJ_EX(DivMod);
5784 } /* firm_set_default_transform_node */
5787 /* **************** Common Subexpression Elimination **************** */
5789 /** The size of the hash table used, should estimate the number of nodes
5791 #define N_IR_NODES 512
5793 /** Compares the attributes of two Const nodes. */
5794 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5795 return (get_Const_tarval(a) != get_Const_tarval(b))
5796 || (get_Const_type(a) != get_Const_type(b));
5797 } /* node_cmp_attr_Const */
5799 /** Compares the attributes of two Proj nodes. */
5800 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5801 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5802 } /* node_cmp_attr_Proj */
5804 /** Compares the attributes of two Filter nodes. */
5805 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5806 return get_Filter_proj(a) != get_Filter_proj(b);
5807 } /* node_cmp_attr_Filter */
5809 /** Compares the attributes of two Alloc nodes. */
5810 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5811 const alloc_attr *pa = get_irn_alloc_attr(a);
5812 const alloc_attr *pb = get_irn_alloc_attr(b);
5813 return (pa->where != pb->where) || (pa->type != pb->type);
5814 } /* node_cmp_attr_Alloc */
5816 /** Compares the attributes of two Free nodes. */
5817 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5818 const free_attr *pa = get_irn_free_attr(a);
5819 const free_attr *pb = get_irn_free_attr(b);
5820 return (pa->where != pb->where) || (pa->type != pb->type);
5821 } /* node_cmp_attr_Free */
5823 /** Compares the attributes of two SymConst nodes. */
5824 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5825 const symconst_attr *pa = get_irn_symconst_attr(a);
5826 const symconst_attr *pb = get_irn_symconst_attr(b);
5827 return (pa->kind != pb->kind)
5828 || (pa->sym.type_p != pb->sym.type_p)
5829 || (pa->tp != pb->tp);
5830 } /* node_cmp_attr_SymConst */
5832 /** Compares the attributes of two Call nodes. */
5833 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5834 return get_irn_call_attr(a) != get_irn_call_attr(b);
5835 } /* node_cmp_attr_Call */
5837 /** Compares the attributes of two Sel nodes. */
5838 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5839 const ir_entity *a_ent = get_Sel_entity(a);
5840 const ir_entity *b_ent = get_Sel_entity(b);
5843 (a_ent->kind != b_ent->kind) ||
5844 (a_ent->name != b_ent->name) ||
5845 (a_ent->owner != b_ent->owner) ||
5846 (a_ent->ld_name != b_ent->ld_name) ||
5847 (a_ent->type != b_ent->type);
5849 /* Matze: inlining of functions can produce 2 entities with same type,
5851 return a_ent != b_ent;
5852 } /* node_cmp_attr_Sel */
5854 /** Compares the attributes of two Phi nodes. */
5855 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5856 /* we can only enter this function if both nodes have the same number of inputs,
5857 hence it is enough to check if one of them is a Phi0 */
5859 /* check the Phi0 pos attribute */
5860 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5863 } /* node_cmp_attr_Phi */
5865 /** Compares the attributes of two Conv nodes. */
5866 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5867 return get_Conv_strict(a) != get_Conv_strict(b);
5868 } /* node_cmp_attr_Conv */
5870 /** Compares the attributes of two Cast nodes. */
5871 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5872 return get_Cast_type(a) != get_Cast_type(b);
5873 } /* node_cmp_attr_Cast */
5875 /** Compares the attributes of two Load nodes. */
5876 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5877 if (get_Load_volatility(a) == volatility_is_volatile ||
5878 get_Load_volatility(b) == volatility_is_volatile)
5879 /* NEVER do CSE on volatile Loads */
5881 /* do not CSE Loads with different alignment. Be conservative. */
5882 if (get_Load_align(a) != get_Load_align(b))
5885 return get_Load_mode(a) != get_Load_mode(b);
5886 } /* node_cmp_attr_Load */
5888 /** Compares the attributes of two Store nodes. */
5889 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5890 /* do not CSE Stores with different alignment. Be conservative. */
5891 if (get_Store_align(a) != get_Store_align(b))
5894 /* NEVER do CSE on volatile Stores */
5895 return (get_Store_volatility(a) == volatility_is_volatile ||
5896 get_Store_volatility(b) == volatility_is_volatile);
5897 } /* node_cmp_attr_Store */
5899 /** Compares two exception attributes */
5900 static int node_cmp_exception(ir_node *a, ir_node *b) {
5901 const except_attr *ea = get_irn_except_attr(a);
5902 const except_attr *eb = get_irn_except_attr(b);
5904 return ea->pin_state != eb->pin_state;
5907 #define node_cmp_attr_Bound node_cmp_exception
5909 /** Compares the attributes of two Div nodes. */
5910 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5911 const divmod_attr *ma = get_irn_divmod_attr(a);
5912 const divmod_attr *mb = get_irn_divmod_attr(b);
5913 return ma->exc.pin_state != mb->exc.pin_state ||
5914 ma->res_mode != mb->res_mode ||
5915 ma->no_remainder != mb->no_remainder;
5916 } /* node_cmp_attr_Div */
5918 /** Compares the attributes of two DivMod nodes. */
5919 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5920 const divmod_attr *ma = get_irn_divmod_attr(a);
5921 const divmod_attr *mb = get_irn_divmod_attr(b);
5922 return ma->exc.pin_state != mb->exc.pin_state ||
5923 ma->res_mode != mb->res_mode;
5924 } /* node_cmp_attr_DivMod */
5926 /** Compares the attributes of two Mod nodes. */
5927 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5928 const divmod_attr *ma = get_irn_divmod_attr(a);
5929 const divmod_attr *mb = get_irn_divmod_attr(b);
5930 return ma->exc.pin_state != mb->exc.pin_state ||
5931 ma->res_mode != mb->res_mode;
5932 } /* node_cmp_attr_Mod */
5934 /** Compares the attributes of two Quot nodes. */
5935 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5936 const divmod_attr *ma = get_irn_divmod_attr(a);
5937 const divmod_attr *mb = get_irn_divmod_attr(b);
5938 return ma->exc.pin_state != mb->exc.pin_state ||
5939 ma->res_mode != mb->res_mode;
5940 } /* node_cmp_attr_Quot */
5942 /** Compares the attributes of two Confirm nodes. */
5943 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5944 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5945 } /* node_cmp_attr_Confirm */
5947 /** Compares the attributes of two ASM nodes. */
5948 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5950 const ir_asm_constraint *ca;
5951 const ir_asm_constraint *cb;
5954 if (get_ASM_text(a) != get_ASM_text(b))
5957 /* Should we really check the constraints here? Should be better, but is strange. */
5958 n = get_ASM_n_input_constraints(a);
5959 if (n != get_ASM_n_input_constraints(b))
5962 ca = get_ASM_input_constraints(a);
5963 cb = get_ASM_input_constraints(b);
5964 for (i = 0; i < n; ++i) {
5965 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5969 n = get_ASM_n_output_constraints(a);
5970 if (n != get_ASM_n_output_constraints(b))
5973 ca = get_ASM_output_constraints(a);
5974 cb = get_ASM_output_constraints(b);
5975 for (i = 0; i < n; ++i) {
5976 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5980 n = get_ASM_n_clobbers(a);
5981 if (n != get_ASM_n_clobbers(b))
5984 cla = get_ASM_clobbers(a);
5985 clb = get_ASM_clobbers(b);
5986 for (i = 0; i < n; ++i) {
5987 if (cla[i] != clb[i])
5991 } /* node_cmp_attr_ASM */
5994 * Set the default node attribute compare operation for an ir_op_ops.
5996 * @param code the opcode for the default operation
5997 * @param ops the operations initialized
6002 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6006 ops->node_cmp_attr = node_cmp_attr_##a; \
6037 } /* firm_set_default_node_cmp_attr */
6040 * Compare function for two nodes in the value table. Gets two
6041 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6043 int identities_cmp(const void *elt, const void *key) {
6044 ir_node *a = (ir_node *)elt;
6045 ir_node *b = (ir_node *)key;
6048 if (a == b) return 0;
6050 if ((get_irn_op(a) != get_irn_op(b)) ||
6051 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6053 /* compare if a's in and b's in are of equal length */
6054 irn_arity_a = get_irn_intra_arity(a);
6055 if (irn_arity_a != get_irn_intra_arity(b))
6058 if (get_irn_pinned(a) == op_pin_state_pinned) {
6059 /* for pinned nodes, the block inputs must be equal */
6060 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6062 } else if (! get_opt_global_cse()) {
6063 /* for block-local CSE both nodes must be in the same MacroBlock */
6064 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6068 /* compare a->in[0..ins] with b->in[0..ins] */
6069 for (i = 0; i < irn_arity_a; i++)
6070 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6074 * here, we already now that the nodes are identical except their
6077 if (a->op->ops.node_cmp_attr)
6078 return a->op->ops.node_cmp_attr(a, b);
6081 } /* identities_cmp */
6084 * Calculate a hash value of a node.
6086 * @param node The IR-node
6088 unsigned ir_node_hash(const ir_node *node) {
6089 return node->op->ops.hash(node);
6090 } /* ir_node_hash */
6093 pset *new_identities(void) {
6094 return new_pset(identities_cmp, N_IR_NODES);
6095 } /* new_identities */
6097 void del_identities(pset *value_table) {
6098 del_pset(value_table);
6099 } /* del_identities */
6101 /* Normalize a node by putting constants (and operands with larger
6102 * node index) on the right (operator side). */
6103 void ir_normalize_node(ir_node *n) {
6104 if (is_op_commutative(get_irn_op(n))) {
6105 ir_node *l = get_binop_left(n);
6106 ir_node *r = get_binop_right(n);
6108 /* For commutative operators perform a OP b == b OP a but keep
6109 * constants on the RIGHT side. This helps greatly in some
6110 * optimizations. Moreover we use the idx number to make the form
6112 if (!operands_are_normalized(l, r)) {
6113 set_binop_left(n, r);
6114 set_binop_right(n, l);
6118 } /* ir_normalize_node */
6121 * Update the nodes after a match in the value table. If both nodes have
6122 * the same MacroBlock but different Blocks, we must ensure that the node
6123 * with the dominating Block (the node that is near to the MacroBlock header
6124 * is stored in the table.
6125 * Because a MacroBlock has only one "non-exception" flow, we don't need
6126 * dominance info here: We known, that one block must dominate the other and
6127 * following the only block input will allow to find it.
6129 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6130 ir_node *known_blk, *new_block, *block, *mbh;
6132 if (get_opt_global_cse()) {
6133 /* Block inputs are meaning less */
6136 known_blk = get_irn_n(known_irn, -1);
6137 new_block = get_irn_n(new_ir_node, -1);
6138 if (known_blk == new_block) {
6139 /* already in the same block */
6143 * We expect the typical case when we built the graph. In that case, the
6144 * known_irn is already the upper one, so checking this should be faster.
6147 mbh = get_Block_MacroBlock(new_block);
6149 if (block == known_blk) {
6150 /* ok, we have found it: known_block dominates new_block as expected */
6155 * We have reached the MacroBlock header NOT founding
6156 * the known_block. new_block must dominate known_block.
6159 set_irn_n(known_irn, -1, new_block);
6162 assert(get_Block_n_cfgpreds(block) == 1);
6163 block = get_Block_cfgpred_block(block, 0);
6165 } /* update_value_table */
6168 * Return the canonical node computing the same value as n.
6169 * Looks up the node in a hash table, enters it in the table
6170 * if it isn't there yet.
6172 * @param value_table the HashSet containing all nodes in the
6174 * @param n the node to look up
6176 * @return a node that computes the same value as n or n if no such
6177 * node could be found
6179 ir_node *identify_remember(pset *value_table, ir_node *n) {
6182 if (!value_table) return n;
6184 ir_normalize_node(n);
6185 /* lookup or insert in hash table with given hash key. */
6186 o = pset_insert(value_table, n, ir_node_hash(n));
6189 update_known_irn(o, n);
6193 } /* identify_remember */
6196 * During construction we set the op_pin_state_pinned flag in the graph right when the
6197 * optimization is performed. The flag turning on procedure global cse could
6198 * be changed between two allocations. This way we are safe.
6200 * @param value_table The value table
6201 * @param n The node to lookup
6203 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6206 n = identify_remember(value_table, n);
6207 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6208 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6210 } /* identify_cons */
6212 /* Add a node to the identities value table. */
6213 void add_identities(pset *value_table, ir_node *node) {
6214 if (get_opt_cse() && is_no_Block(node))
6215 identify_remember(value_table, node);
6216 } /* add_identities */
6218 /* Visit each node in the value table of a graph. */
6219 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6221 ir_graph *rem = current_ir_graph;
6223 current_ir_graph = irg;
6224 foreach_pset(irg->value_table, node)
6226 current_ir_graph = rem;
6227 } /* visit_all_identities */
6230 * Garbage in, garbage out. If a node has a dead input, i.e., the
6231 * Bad node is input to the node, return the Bad node.
6233 static ir_node *gigo(ir_node *node) {
6235 ir_op *op = get_irn_op(node);
6237 /* remove garbage blocks by looking at control flow that leaves the block
6238 and replacing the control flow by Bad. */
6239 if (get_irn_mode(node) == mode_X) {
6240 ir_node *block = get_nodes_block(skip_Proj(node));
6242 /* Don't optimize nodes in immature blocks. */
6243 if (!get_Block_matured(block))
6245 /* Don't optimize End, may have Bads. */
6246 if (op == op_End) return node;
6248 if (is_Block(block)) {
6249 if (is_Block_dead(block)) {
6250 /* control flow from dead block is dead */
6254 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6255 if (!is_Bad(get_irn_n(block, i)))
6259 ir_graph *irg = get_irn_irg(block);
6260 /* the start block is never dead */
6261 if (block != get_irg_start_block(irg)
6262 && block != get_irg_end_block(irg)) {
6264 * Do NOT kill control flow without setting
6265 * the block to dead of bad things can happen:
6266 * We get a Block that is not reachable be irg_block_walk()
6267 * but can be found by irg_walk()!
6269 set_Block_dead(block);
6276 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6277 blocks predecessors is dead. */
6278 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6279 irn_arity = get_irn_arity(node);
6282 * Beware: we can only read the block of a non-floating node.
6284 if (is_irn_pinned_in_irg(node) &&
6285 is_Block_dead(get_nodes_block(skip_Proj(node))))
6288 for (i = 0; i < irn_arity; i++) {
6289 ir_node *pred = get_irn_n(node, i);
6294 /* Propagating Unknowns here seems to be a bad idea, because
6295 sometimes we need a node as a input and did not want that
6297 However, it might be useful to move this into a later phase
6298 (if you think that optimizing such code is useful). */
6299 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6300 return new_Unknown(get_irn_mode(node));
6305 /* With this code we violate the agreement that local_optimize
6306 only leaves Bads in Block, Phi and Tuple nodes. */
6307 /* If Block has only Bads as predecessors it's garbage. */
6308 /* If Phi has only Bads as predecessors it's garbage. */
6309 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6310 irn_arity = get_irn_arity(node);
6311 for (i = 0; i < irn_arity; i++) {
6312 if (!is_Bad(get_irn_n(node, i))) break;
6314 if (i == irn_arity) node = new_Bad();
6321 * These optimizations deallocate nodes from the obstack.
6322 * It can only be called if it is guaranteed that no other nodes
6323 * reference this one, i.e., right after construction of a node.
6325 * @param n The node to optimize
6327 * current_ir_graph must be set to the graph of the node!
6329 ir_node *optimize_node(ir_node *n) {
6332 ir_opcode iro = get_irn_opcode(n);
6334 /* Always optimize Phi nodes: part of the construction. */
6335 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6337 /* constant expression evaluation / constant folding */
6338 if (get_opt_constant_folding()) {
6339 /* neither constants nor Tuple values can be evaluated */
6340 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6341 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6342 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6343 /* try to evaluate */
6344 tv = computed_value(n);
6345 if (tv != tarval_bad) {
6347 ir_type *old_tp = get_irn_type(n);
6348 int i, arity = get_irn_arity(n);
6352 * Try to recover the type of the new expression.
6354 for (i = 0; i < arity && !old_tp; ++i)
6355 old_tp = get_irn_type(get_irn_n(n, i));
6358 * we MUST copy the node here temporary, because it's still needed
6359 * for DBG_OPT_CSTEVAL
6361 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6362 oldn = alloca(node_size);
6364 memcpy(oldn, n, node_size);
6365 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6367 /* ARG, copy the in array, we need it for statistics */
6368 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6370 /* note the inplace edges module */
6371 edges_node_deleted(n, current_ir_graph);
6373 /* evaluation was successful -- replace the node. */
6374 irg_kill_node(current_ir_graph, n);
6375 nw = new_Const(get_tarval_mode(tv), tv);
6377 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6378 set_Const_type(nw, old_tp);
6379 DBG_OPT_CSTEVAL(oldn, nw);
6380 tarval_enable_fp_ops(old_fp_mode);
6383 tarval_enable_fp_ops(old_fp_mode);
6387 /* remove unnecessary nodes */
6388 if (get_opt_algebraic_simplification() ||
6389 (iro == iro_Phi) || /* always optimize these nodes. */
6391 (iro == iro_Proj) ||
6392 (iro == iro_Block) ) /* Flags tested local. */
6393 n = equivalent_node(n);
6395 /* Common Subexpression Elimination.
6397 * Checks whether n is already available.
6398 * The block input is used to distinguish different subexpressions. Right
6399 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6400 * subexpressions within a block.
6403 n = identify_cons(current_ir_graph->value_table, n);
6406 edges_node_deleted(oldn, current_ir_graph);
6408 /* We found an existing, better node, so we can deallocate the old node. */
6409 irg_kill_node(current_ir_graph, oldn);
6413 /* Some more constant expression evaluation that does not allow to
6415 iro = get_irn_opcode(n);
6416 if (get_opt_algebraic_simplification() ||
6417 (iro == iro_Cond) ||
6418 (iro == iro_Proj)) /* Flags tested local. */
6419 n = transform_node(n);
6421 /* Remove nodes with dead (Bad) input.
6422 Run always for transformation induced Bads. */
6425 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6426 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6428 n = identify_remember(current_ir_graph->value_table, o);
6434 } /* optimize_node */
6438 * These optimizations never deallocate nodes (in place). This can cause dead
6439 * nodes lying on the obstack. Remove these by a dead node elimination,
6440 * i.e., a copying garbage collection.
6442 ir_node *optimize_in_place_2(ir_node *n) {
6445 ir_opcode iro = get_irn_opcode(n);
6447 if (!get_opt_optimize() && !is_Phi(n)) return n;
6449 /* constant expression evaluation / constant folding */
6450 if (get_opt_constant_folding()) {
6451 /* neither constants nor Tuple values can be evaluated */
6452 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6453 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6454 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6455 /* try to evaluate */
6456 tv = computed_value(n);
6457 if (tv != tarval_bad) {
6458 /* evaluation was successful -- replace the node. */
6459 ir_type *old_tp = get_irn_type(n);
6460 int i, arity = get_irn_arity(n);
6463 * Try to recover the type of the new expression.
6465 for (i = 0; i < arity && !old_tp; ++i)
6466 old_tp = get_irn_type(get_irn_n(n, i));
6468 n = new_Const(get_tarval_mode(tv), tv);
6470 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6471 set_Const_type(n, old_tp);
6473 DBG_OPT_CSTEVAL(oldn, n);
6474 tarval_enable_fp_ops(old_fp_mode);
6477 tarval_enable_fp_ops(old_fp_mode);
6481 /* remove unnecessary nodes */
6482 if (get_opt_constant_folding() ||
6483 (iro == iro_Phi) || /* always optimize these nodes. */
6484 (iro == iro_Id) || /* ... */
6485 (iro == iro_Proj) || /* ... */
6486 (iro == iro_Block) ) /* Flags tested local. */
6487 n = equivalent_node(n);
6489 /** common subexpression elimination **/
6490 /* Checks whether n is already available. */
6491 /* The block input is used to distinguish different subexpressions. Right
6492 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6493 subexpressions within a block. */
6494 if (get_opt_cse()) {
6496 n = identify_remember(current_ir_graph->value_table, o);
6501 /* Some more constant expression evaluation. */
6502 iro = get_irn_opcode(n);
6503 if (get_opt_constant_folding() ||
6504 (iro == iro_Cond) ||
6505 (iro == iro_Proj)) /* Flags tested local. */
6506 n = transform_node(n);
6508 /* Remove nodes with dead (Bad) input.
6509 Run always for transformation induced Bads. */
6512 /* Now we can verify the node, as it has no dead inputs any more. */
6515 /* Now we have a legal, useful node. Enter it in hash table for cse.
6516 Blocks should be unique anyways. (Except the successor of start:
6517 is cse with the start block!) */
6518 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6520 n = identify_remember(current_ir_graph->value_table, o);
6526 } /* optimize_in_place_2 */
6529 * Wrapper for external use, set proper status bits after optimization.
6531 ir_node *optimize_in_place(ir_node *n) {
6532 /* Handle graph state */
6533 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6535 if (get_opt_global_cse())
6536 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6537 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6538 set_irg_outs_inconsistent(current_ir_graph);
6540 /* FIXME: Maybe we could also test whether optimizing the node can
6541 change the control graph. */
6542 set_irg_doms_inconsistent(current_ir_graph);
6543 return optimize_in_place_2(n);
6544 } /* optimize_in_place */
6547 * Calculate a hash value of a Const node.
6549 static unsigned hash_Const(const ir_node *node) {
6552 /* special value for const, as they only differ in their tarval. */
6553 h = HASH_PTR(node->attr.con.tv);
6554 h = 9*h + HASH_PTR(get_irn_mode(node));
6560 * Calculate a hash value of a SymConst node.
6562 static unsigned hash_SymConst(const ir_node *node) {
6565 /* special value for const, as they only differ in their symbol. */
6566 h = HASH_PTR(node->attr.symc.sym.type_p);
6567 h = 9*h + HASH_PTR(get_irn_mode(node));
6570 } /* hash_SymConst */
6573 * Set the default hash operation in an ir_op_ops.
6575 * @param code the opcode for the default operation
6576 * @param ops the operations initialized
6581 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6585 ops->hash = hash_##a; \
6588 /* hash function already set */
6589 if (ops->hash != NULL)
6596 /* use input/mode default hash if no function was given */
6597 ops->hash = firm_default_hash;
6605 * Sets the default operation for an ir_ops.
6607 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6608 ops = firm_set_default_hash(code, ops);
6609 ops = firm_set_default_computed_value(code, ops);
6610 ops = firm_set_default_equivalent_node(code, ops);
6611 ops = firm_set_default_transform_node(code, ops);
6612 ops = firm_set_default_node_cmp_attr(code, ops);
6613 ops = firm_set_default_get_type(code, ops);
6614 ops = firm_set_default_get_type_attr(code, ops);
6615 ops = firm_set_default_get_entity_attr(code, ops);
6618 } /* firm_set_default_operations */