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);
1281 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1284 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1285 ir_node *b = get_Conv_op(a);
1286 ir_mode *b_mode = get_irn_mode(b);
1288 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1289 /* both are strict conv */
1290 if (smaller_mode(a_mode, n_mode)) {
1291 /* both are strict, but the first is smaller, so
1292 the second cannot remove more precision, remove the
1294 set_Conv_strict(n, 0);
1297 if (n_mode == b_mode) {
1298 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1299 if (n_mode == mode_b) {
1300 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1301 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1303 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1304 if (values_in_mode(b_mode, a_mode)) {
1305 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1306 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1311 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1312 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1313 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1314 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1316 if (float_mantissa >= int_mantissa) {
1318 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1323 if (smaller_mode(b_mode, a_mode)) {
1324 if (get_Conv_strict(n))
1325 set_Conv_strict(b, 1);
1326 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1327 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1334 } /* equivalent_node_Conv */
1337 * A Cast may be removed if the type of the previous node
1338 * is already the type of the Cast.
1340 static ir_node *equivalent_node_Cast(ir_node *n) {
1342 ir_node *pred = get_Cast_op(n);
1344 if (get_irn_type(pred) == get_Cast_type(n)) {
1346 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1349 } /* equivalent_node_Cast */
1352 * - fold Phi-nodes, iff they have only one predecessor except
1355 static ir_node *equivalent_node_Phi(ir_node *n) {
1360 ir_node *first_val = NULL; /* to shutup gcc */
1362 if (!get_opt_normalize()) return n;
1364 n_preds = get_Phi_n_preds(n);
1366 block = get_nodes_block(n);
1367 if (is_Block_dead(block)) /* Control dead */
1368 return get_irg_bad(current_ir_graph);
1370 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1372 /* Find first non-self-referencing input */
1373 for (i = 0; i < n_preds; ++i) {
1374 first_val = get_Phi_pred(n, i);
1375 if ( (first_val != n) /* not self pointer */
1377 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1378 * predecessors. Then, Phi nodes in dead code might be removed, causing
1379 * nodes pointing to themself (Add's for instance).
1380 * This is really bad and causes endless recursions in several
1381 * code pathes, so we do NOT optimize such a code.
1382 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1383 * (and bad Phi predecessors), so live code is optimized later.
1385 && (! is_Bad(get_Block_cfgpred(block, i)))
1387 ) { /* value not dead */
1388 break; /* then found first value. */
1393 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1394 return get_irg_bad(current_ir_graph);
1397 /* search for rest of inputs, determine if any of these
1398 are non-self-referencing */
1399 while (++i < n_preds) {
1400 ir_node *scnd_val = get_Phi_pred(n, i);
1401 if ( (scnd_val != n)
1402 && (scnd_val != first_val)
1405 && (! is_Bad(get_Block_cfgpred(block, i)))
1413 /* Fold, if no multiple distinct non-self-referencing inputs */
1415 DBG_OPT_PHI(oldn, n);
1418 } /* equivalent_node_Phi */
1421 * Several optimizations:
1422 * - fold Sync-nodes, iff they have only one predecessor except
1425 static ir_node *equivalent_node_Sync(ir_node *n) {
1426 int arity = get_Sync_n_preds(n);
1429 for (i = 0; i < arity;) {
1430 ir_node *pred = get_Sync_pred(n, i);
1433 /* Remove Bad predecessors */
1440 /* Remove duplicate predecessors */
1446 if (get_Sync_pred(n, j) == pred) {
1454 if (arity == 0) return get_irg_bad(current_ir_graph);
1455 if (arity == 1) return get_Sync_pred(n, 0);
1457 } /* equivalent_node_Sync */
1460 * Optimize Proj(Tuple).
1462 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1463 ir_node *oldn = proj;
1464 ir_node *tuple = get_Proj_pred(proj);
1466 /* Remove the Tuple/Proj combination. */
1467 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1468 DBG_OPT_TUPLE(oldn, tuple, proj);
1471 } /* equivalent_node_Proj_Tuple */
1474 * Optimize a / 1 = a.
1476 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1477 ir_node *oldn = proj;
1478 ir_node *div = get_Proj_pred(proj);
1479 ir_node *b = get_Div_right(div);
1480 tarval *tb = value_of(b);
1482 /* Div is not commutative. */
1483 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1484 switch (get_Proj_proj(proj)) {
1486 proj = get_Div_mem(div);
1487 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1491 proj = get_Div_left(div);
1492 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1496 /* we cannot replace the exception Proj's here, this is done in
1497 transform_node_Proj_Div() */
1502 } /* equivalent_node_Proj_Div */
1505 * Optimize a / 1.0 = a.
1507 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1508 ir_node *oldn = proj;
1509 ir_node *quot = get_Proj_pred(proj);
1510 ir_node *b = get_Quot_right(quot);
1511 tarval *tb = value_of(b);
1513 /* Div is not commutative. */
1514 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1515 switch (get_Proj_proj(proj)) {
1517 proj = get_Quot_mem(quot);
1518 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1522 proj = get_Quot_left(quot);
1523 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1527 /* we cannot replace the exception Proj's here, this is done in
1528 transform_node_Proj_Quot() */
1533 } /* equivalent_node_Proj_Quot */
1536 * Optimize a / 1 = a.
1538 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1539 ir_node *oldn = proj;
1540 ir_node *divmod = get_Proj_pred(proj);
1541 ir_node *b = get_DivMod_right(divmod);
1542 tarval *tb = value_of(b);
1544 /* Div is not commutative. */
1545 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1546 switch (get_Proj_proj(proj)) {
1548 proj = get_DivMod_mem(divmod);
1549 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1552 case pn_DivMod_res_div:
1553 proj = get_DivMod_left(divmod);
1554 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1558 /* we cannot replace the exception Proj's here, this is done in
1559 transform_node_Proj_DivMod().
1560 Note further that the pn_DivMod_res_div case is handled in
1561 computed_value_Proj(). */
1566 } /* equivalent_node_Proj_DivMod */
1569 * Optimize CopyB(mem, x, x) into a Nop.
1571 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1572 ir_node *oldn = proj;
1573 ir_node *copyb = get_Proj_pred(proj);
1574 ir_node *a = get_CopyB_dst(copyb);
1575 ir_node *b = get_CopyB_src(copyb);
1578 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1579 switch (get_Proj_proj(proj)) {
1580 case pn_CopyB_M_regular:
1581 proj = get_CopyB_mem(copyb);
1582 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1585 case pn_CopyB_M_except:
1586 case pn_CopyB_X_except:
1587 DBG_OPT_EXC_REM(proj);
1588 proj = get_irg_bad(current_ir_graph);
1593 } /* equivalent_node_Proj_CopyB */
1596 * Optimize Bounds(idx, idx, upper) into idx.
1598 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1599 ir_node *oldn = proj;
1600 ir_node *bound = get_Proj_pred(proj);
1601 ir_node *idx = get_Bound_index(bound);
1602 ir_node *pred = skip_Proj(idx);
1605 if (idx == get_Bound_lower(bound))
1607 else if (is_Bound(pred)) {
1609 * idx was Bounds checked in the same MacroBlock previously,
1610 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1612 ir_node *lower = get_Bound_lower(bound);
1613 ir_node *upper = get_Bound_upper(bound);
1614 if (get_Bound_lower(pred) == lower &&
1615 get_Bound_upper(pred) == upper &&
1616 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1618 * One could expect that we simply return the previous
1619 * Bound here. However, this would be wrong, as we could
1620 * add an exception Proj to a new location then.
1621 * So, we must turn in into a tuple.
1627 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1628 switch (get_Proj_proj(proj)) {
1630 DBG_OPT_EXC_REM(proj);
1631 proj = get_Bound_mem(bound);
1633 case pn_Bound_X_except:
1634 DBG_OPT_EXC_REM(proj);
1635 proj = get_irg_bad(current_ir_graph);
1639 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1642 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1647 } /* equivalent_node_Proj_Bound */
1650 * Optimize an Exception Proj(Load) with a non-null address.
1652 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1653 if (get_opt_ldst_only_null_ptr_exceptions()) {
1654 if (get_irn_mode(proj) == mode_X) {
1655 ir_node *load = get_Proj_pred(proj);
1657 /* get the Load address */
1658 const ir_node *addr = get_Load_ptr(load);
1659 const ir_node *confirm;
1661 if (value_not_null(addr, &confirm)) {
1662 if (get_Proj_proj(proj) == pn_Load_X_except) {
1663 DBG_OPT_EXC_REM(proj);
1664 return get_irg_bad(current_ir_graph);
1670 } /* equivalent_node_Proj_Load */
1673 * Optimize an Exception Proj(Store) with a non-null address.
1675 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1676 if (get_opt_ldst_only_null_ptr_exceptions()) {
1677 if (get_irn_mode(proj) == mode_X) {
1678 ir_node *store = get_Proj_pred(proj);
1680 /* get the load/store address */
1681 const ir_node *addr = get_Store_ptr(store);
1682 const ir_node *confirm;
1684 if (value_not_null(addr, &confirm)) {
1685 if (get_Proj_proj(proj) == pn_Store_X_except) {
1686 DBG_OPT_EXC_REM(proj);
1687 return get_irg_bad(current_ir_graph);
1693 } /* equivalent_node_Proj_Store */
1696 * Does all optimizations on nodes that must be done on it's Proj's
1697 * because of creating new nodes.
1699 static ir_node *equivalent_node_Proj(ir_node *proj) {
1700 ir_node *n = get_Proj_pred(proj);
1702 if (get_irn_mode(proj) == mode_X) {
1703 if (is_Block_dead(get_nodes_block(n))) {
1704 /* Remove dead control flow -- early gigo(). */
1705 return get_irg_bad(current_ir_graph);
1708 if (n->op->ops.equivalent_node_Proj)
1709 return n->op->ops.equivalent_node_Proj(proj);
1711 } /* equivalent_node_Proj */
1716 static ir_node *equivalent_node_Id(ir_node *n) {
1723 DBG_OPT_ID(oldn, n);
1725 } /* equivalent_node_Id */
1730 static ir_node *equivalent_node_Mux(ir_node *n)
1732 ir_node *oldn = n, *sel = get_Mux_sel(n);
1733 tarval *ts = value_of(sel);
1735 /* Mux(true, f, t) == t */
1736 if (ts == tarval_b_true) {
1737 n = get_Mux_true(n);
1738 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1740 /* Mux(false, f, t) == f */
1741 else if (ts == tarval_b_false) {
1742 n = get_Mux_false(n);
1743 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1745 /* Mux(v, x, x) == x */
1746 else if (get_Mux_false(n) == get_Mux_true(n)) {
1747 n = get_Mux_true(n);
1748 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1750 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1751 ir_node *cmp = get_Proj_pred(sel);
1752 long proj_nr = get_Proj_proj(sel);
1753 ir_node *f = get_Mux_false(n);
1754 ir_node *t = get_Mux_true(n);
1757 * Note further that these optimization work even for floating point
1758 * with NaN's because -NaN == NaN.
1759 * However, if +0 and -0 is handled differently, we cannot use the first one.
1762 ir_node *const cmp_l = get_Cmp_left(cmp);
1763 ir_node *const cmp_r = get_Cmp_right(cmp);
1767 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1768 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1770 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1777 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1778 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1780 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1787 * Note: normalization puts the constant on the right side,
1788 * so we check only one case.
1790 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1791 /* Mux(t CMP 0, X, t) */
1792 if (is_Minus(f) && get_Minus_op(f) == t) {
1793 /* Mux(t CMP 0, -t, t) */
1794 if (proj_nr == pn_Cmp_Eq) {
1795 /* Mux(t == 0, -t, t) ==> -t */
1797 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1798 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1799 /* Mux(t != 0, -t, t) ==> t */
1801 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1808 } /* equivalent_node_Mux */
1811 * Remove Confirm nodes if setting is on.
1812 * Replace Confirms(x, '=', Constlike) by Constlike.
1814 static ir_node *equivalent_node_Confirm(ir_node *n) {
1815 ir_node *pred = get_Confirm_value(n);
1816 pn_Cmp pnc = get_Confirm_cmp(n);
1818 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1820 * rare case: two identical Confirms one after another,
1821 * replace the second one with the first.
1824 pred = get_Confirm_value(n);
1825 pnc = get_Confirm_cmp(n);
1827 if (get_opt_remove_confirm())
1828 return get_Confirm_value(n);
1833 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1834 * perform no actual computation, as, e.g., the Id nodes. It does not create
1835 * new nodes. It is therefore safe to free n if the node returned is not n.
1836 * If a node returns a Tuple we can not just skip it. If the size of the
1837 * in array fits, we transform n into a tuple (e.g., Div).
1839 ir_node *equivalent_node(ir_node *n) {
1840 if (n->op->ops.equivalent_node)
1841 return n->op->ops.equivalent_node(n);
1843 } /* equivalent_node */
1846 * Sets the default equivalent node operation for an ir_op_ops.
1848 * @param code the opcode for the default operation
1849 * @param ops the operations initialized
1854 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1858 ops->equivalent_node = equivalent_node_##a; \
1860 #define CASE_PROJ(a) \
1862 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1904 } /* firm_set_default_equivalent_node */
1907 * Returns non-zero if a node is a Phi node
1908 * with all predecessors constant.
1910 static int is_const_Phi(ir_node *n) {
1913 if (! is_Phi(n) || get_irn_arity(n) == 0)
1915 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1916 if (! is_Const(get_irn_n(n, i)))
1920 } /* is_const_Phi */
1922 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1923 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1926 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1928 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1929 if (eval == tarval_sub) {
1930 tarval_sub_type func = (tarval_sub_type)eval;
1932 return func(a, b, mode);
1934 tarval_binop_type func = (tarval_binop_type)eval;
1941 * Apply an evaluator on a binop with a constant operators (and one Phi).
1943 * @param phi the Phi node
1944 * @param other the other operand
1945 * @param eval an evaluator function
1946 * @param mode the mode of the result, may be different from the mode of the Phi!
1947 * @param left if non-zero, other is the left operand, else the right
1949 * @return a new Phi node if the conversion was successful, NULL else
1951 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1956 int i, n = get_irn_arity(phi);
1958 NEW_ARR_A(void *, res, n);
1960 for (i = 0; i < n; ++i) {
1961 pred = get_irn_n(phi, i);
1962 tv = get_Const_tarval(pred);
1963 tv = do_eval(eval, other, tv, mode);
1965 if (tv == tarval_bad) {
1966 /* folding failed, bad */
1972 for (i = 0; i < n; ++i) {
1973 pred = get_irn_n(phi, i);
1974 tv = get_Const_tarval(pred);
1975 tv = do_eval(eval, tv, other, mode);
1977 if (tv == tarval_bad) {
1978 /* folding failed, bad */
1984 irg = current_ir_graph;
1985 for (i = 0; i < n; ++i) {
1986 pred = get_irn_n(phi, i);
1987 res[i] = new_r_Const_type(irg, mode, res[i], get_Const_type(pred));
1989 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1990 } /* apply_binop_on_phi */
1993 * Apply an evaluator on a binop with two constant Phi.
1995 * @param a the left Phi node
1996 * @param b the right Phi node
1997 * @param eval an evaluator function
1998 * @param mode the mode of the result, may be different from the mode of the Phi!
2000 * @return a new Phi node if the conversion was successful, NULL else
2002 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
2003 tarval *tv_l, *tv_r, *tv;
2009 if (get_nodes_block(a) != get_nodes_block(b))
2012 n = get_irn_arity(a);
2013 NEW_ARR_A(void *, res, n);
2015 for (i = 0; i < n; ++i) {
2016 pred = get_irn_n(a, i);
2017 tv_l = get_Const_tarval(pred);
2018 pred = get_irn_n(b, i);
2019 tv_r = get_Const_tarval(pred);
2020 tv = do_eval(eval, tv_l, tv_r, mode);
2022 if (tv == tarval_bad) {
2023 /* folding failed, bad */
2028 irg = current_ir_graph;
2029 for (i = 0; i < n; ++i) {
2030 pred = get_irn_n(a, i);
2031 res[i] = new_r_Const_type(irg, mode, res[i], get_Const_type(pred));
2033 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2034 } /* apply_binop_on_2_phis */
2037 * Apply an evaluator on a unop with a constant operator (a Phi).
2039 * @param phi the Phi node
2040 * @param eval an evaluator function
2042 * @return a new Phi node if the conversion was successful, NULL else
2044 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2050 int i, n = get_irn_arity(phi);
2052 NEW_ARR_A(void *, res, n);
2053 for (i = 0; i < n; ++i) {
2054 pred = get_irn_n(phi, i);
2055 tv = get_Const_tarval(pred);
2058 if (tv == tarval_bad) {
2059 /* folding failed, bad */
2064 mode = get_irn_mode(phi);
2065 irg = current_ir_graph;
2066 for (i = 0; i < n; ++i) {
2067 pred = get_irn_n(phi, i);
2068 res[i] = new_r_Const_type(irg, mode, res[i], get_Const_type(pred));
2070 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2071 } /* apply_unop_on_phi */
2074 * Apply a conversion on a constant operator (a Phi).
2076 * @param phi the Phi node
2078 * @return a new Phi node if the conversion was successful, NULL else
2080 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2085 int i, n = get_irn_arity(phi);
2087 NEW_ARR_A(void *, res, n);
2088 for (i = 0; i < n; ++i) {
2089 pred = get_irn_n(phi, i);
2090 tv = get_Const_tarval(pred);
2091 tv = tarval_convert_to(tv, mode);
2093 if (tv == tarval_bad) {
2094 /* folding failed, bad */
2099 irg = current_ir_graph;
2100 for (i = 0; i < n; ++i) {
2101 pred = get_irn_n(phi, i);
2102 res[i] = new_r_Const_type(irg, mode, res[i], get_Const_type(pred));
2104 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2105 } /* apply_conv_on_phi */
2108 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2109 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2110 * If possible, remove the Conv's.
2112 static ir_node *transform_node_AddSub(ir_node *n) {
2113 ir_mode *mode = get_irn_mode(n);
2115 if (mode_is_reference(mode)) {
2116 ir_node *left = get_binop_left(n);
2117 ir_node *right = get_binop_right(n);
2118 unsigned ref_bits = get_mode_size_bits(mode);
2120 if (is_Conv(left)) {
2121 ir_mode *lmode = get_irn_mode(left);
2122 unsigned bits = get_mode_size_bits(lmode);
2124 if (ref_bits == bits &&
2125 mode_is_int(lmode) &&
2126 get_mode_arithmetic(lmode) == irma_twos_complement) {
2127 ir_node *pre = get_Conv_op(left);
2128 ir_mode *pre_mode = get_irn_mode(pre);
2130 if (mode_is_int(pre_mode) &&
2131 get_mode_size_bits(pre_mode) == bits &&
2132 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2133 /* ok, this conv just changes to sign, moreover the calculation
2134 * is done with same number of bits as our address mode, so
2135 * we can ignore the conv as address calculation can be viewed
2136 * as either signed or unsigned
2138 set_binop_left(n, pre);
2143 if (is_Conv(right)) {
2144 ir_mode *rmode = get_irn_mode(right);
2145 unsigned bits = get_mode_size_bits(rmode);
2147 if (ref_bits == bits &&
2148 mode_is_int(rmode) &&
2149 get_mode_arithmetic(rmode) == irma_twos_complement) {
2150 ir_node *pre = get_Conv_op(right);
2151 ir_mode *pre_mode = get_irn_mode(pre);
2153 if (mode_is_int(pre_mode) &&
2154 get_mode_size_bits(pre_mode) == bits &&
2155 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2156 /* ok, this conv just changes to sign, moreover the calculation
2157 * is done with same number of bits as our address mode, so
2158 * we can ignore the conv as address calculation can be viewed
2159 * as either signed or unsigned
2161 set_binop_right(n, pre);
2166 /* let address arithmetic use unsigned modes */
2167 if (is_Const(right)) {
2168 ir_mode *rmode = get_irn_mode(right);
2170 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2171 /* convert a AddP(P, *s) into AddP(P, *u) */
2172 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2174 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2175 set_binop_right(n, pre);
2181 } /* transform_node_AddSub */
2183 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2185 if (is_Const(b) && is_const_Phi(a)) { \
2186 /* check for Op(Phi, Const) */ \
2187 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2189 else if (is_Const(a) && is_const_Phi(b)) { \
2190 /* check for Op(Const, Phi) */ \
2191 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2193 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2194 /* check for Op(Phi, Phi) */ \
2195 c = apply_binop_on_2_phis(a, b, eval, mode); \
2198 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2202 #define HANDLE_UNOP_PHI(eval, a, c) \
2204 if (is_const_Phi(a)) { \
2205 /* check for Op(Phi) */ \
2206 c = apply_unop_on_phi(a, eval); \
2208 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2214 * Do the AddSub optimization, then Transform
2215 * Constant folding on Phi
2216 * Add(a,a) -> Mul(a, 2)
2217 * Add(Mul(a, x), a) -> Mul(a, x+1)
2218 * if the mode is integer or float.
2219 * Transform Add(a,-b) into Sub(a,b).
2220 * Reassociation might fold this further.
2222 static ir_node *transform_node_Add(ir_node *n) {
2224 ir_node *a, *b, *c, *oldn = n;
2226 n = transform_node_AddSub(n);
2228 a = get_Add_left(n);
2229 b = get_Add_right(n);
2231 mode = get_irn_mode(n);
2233 if (mode_is_reference(mode)) {
2234 ir_mode *lmode = get_irn_mode(a);
2236 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2237 /* an Add(a, NULL) is a hidden Conv */
2238 dbg_info *dbg = get_irn_dbg_info(n);
2239 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2243 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2245 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2246 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2249 if (mode_is_num(mode)) {
2250 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2251 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2252 ir_node *block = get_nodes_block(n);
2255 get_irn_dbg_info(n),
2259 new_r_Const_long(current_ir_graph, mode, 2),
2261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2266 get_irn_dbg_info(n),
2272 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2277 get_irn_dbg_info(n),
2283 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2286 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2287 /* Here we rely on constants be on the RIGHT side */
2289 ir_node *op = get_Not_op(a);
2291 if (is_Const(b) && is_Const_one(b)) {
2293 ir_node *blk = get_nodes_block(n);
2294 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2295 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2300 n = new_r_Const(current_ir_graph, mode, get_mode_minus_one(mode));
2301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2306 ir_node *op = get_Not_op(b);
2310 n = new_r_Const(current_ir_graph, 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_mode *mode = get_irn_mode(cnst);
2328 if (tv == tarval_bad) return NULL;
2329 return new_rd_Const(dbgi, irg, mode, tv);
2333 * Do the AddSub optimization, then Transform
2334 * Constant folding on Phi
2335 * Sub(0,a) -> Minus(a)
2336 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2337 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2338 * Sub(Add(a, x), x) -> a
2339 * Sub(x, Add(x, a)) -> -a
2340 * Sub(x, Const) -> Add(x, -Const)
2342 static ir_node *transform_node_Sub(ir_node *n) {
2347 n = transform_node_AddSub(n);
2349 a = get_Sub_left(n);
2350 b = get_Sub_right(n);
2352 mode = get_irn_mode(n);
2354 if (mode_is_int(mode)) {
2355 ir_mode *lmode = get_irn_mode(a);
2357 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2358 /* a Sub(a, NULL) is a hidden Conv */
2359 dbg_info *dbg = get_irn_dbg_info(n);
2360 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2361 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2365 if (mode == lmode &&
2366 get_mode_arithmetic(mode) == irma_twos_complement &&
2368 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2370 dbg_info *dbg = get_irn_dbg_info(n);
2371 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2372 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2378 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2380 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2381 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2384 if (is_Const(b) && get_irn_mode(b) != mode_P) {
2385 /* a - C -> a + (-C) */
2386 ir_node *cnst = const_negate(b);
2388 ir_node *block = get_nodes_block(n);
2389 dbg_info *dbgi = get_irn_dbg_info(n);
2390 ir_graph *irg = get_irn_irg(n);
2392 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2398 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2399 ir_graph *irg = current_ir_graph;
2400 dbg_info *dbg = get_irn_dbg_info(n);
2401 ir_node *block = get_nodes_block(n);
2402 ir_node *left = get_Minus_op(a);
2403 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2405 n = new_rd_Minus(dbg, irg, block, add, mode);
2406 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2408 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2409 ir_graph *irg = current_ir_graph;
2410 dbg_info *dbg = get_irn_dbg_info(n);
2411 ir_node *block = get_nodes_block(n);
2412 ir_node *right = get_Minus_op(b);
2414 n = new_rd_Add(dbg, irg, block, a, right, mode);
2415 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2417 } else if (is_Sub(b)) {
2418 /* a - (b - c) -> a + (c - b)
2419 * -> (a - b) + c iff (b - c) is a pointer */
2420 ir_graph *irg = current_ir_graph;
2421 dbg_info *s_dbg = get_irn_dbg_info(b);
2422 ir_node *s_block = get_nodes_block(b);
2423 ir_node *s_left = get_Sub_left(b);
2424 ir_node *s_right = get_Sub_right(b);
2425 ir_mode *s_mode = get_irn_mode(b);
2426 if (s_mode == mode_P) {
2427 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2428 dbg_info *a_dbg = get_irn_dbg_info(n);
2429 ir_node *a_block = get_nodes_block(n);
2432 s_right = new_r_Conv(irg, a_block, s_right, mode);
2433 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2435 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2436 dbg_info *a_dbg = get_irn_dbg_info(n);
2437 ir_node *a_block = get_nodes_block(n);
2439 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2443 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2444 ir_node *m_right = get_Mul_right(b);
2445 if (is_Const(m_right)) {
2446 ir_node *cnst2 = const_negate(m_right);
2447 if (cnst2 != NULL) {
2448 ir_graph *irg = current_ir_graph;
2449 dbg_info *m_dbg = get_irn_dbg_info(b);
2450 ir_node *m_block = get_nodes_block(b);
2451 ir_node *m_left = get_Mul_left(b);
2452 ir_mode *m_mode = get_irn_mode(b);
2453 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2454 dbg_info *a_dbg = get_irn_dbg_info(n);
2455 ir_node *a_block = get_nodes_block(n);
2457 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2458 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2464 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2465 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2467 get_irn_dbg_info(n),
2472 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2476 if (mode_wrap_around(mode)) {
2477 ir_node *left = get_Add_left(a);
2478 ir_node *right = get_Add_right(a);
2480 /* FIXME: Does the Conv's work only for two complement or generally? */
2482 if (mode != get_irn_mode(right)) {
2483 /* This Sub is an effective Cast */
2484 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2487 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2489 } else if (right == b) {
2490 if (mode != get_irn_mode(left)) {
2491 /* This Sub is an effective Cast */
2492 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2495 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2501 if (mode_wrap_around(mode)) {
2502 ir_node *left = get_Add_left(b);
2503 ir_node *right = get_Add_right(b);
2505 /* FIXME: Does the Conv's work only for two complement or generally? */
2507 ir_mode *r_mode = get_irn_mode(right);
2509 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2510 if (mode != r_mode) {
2511 /* This Sub is an effective Cast */
2512 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2514 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2516 } else if (right == a) {
2517 ir_mode *l_mode = get_irn_mode(left);
2519 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2520 if (mode != l_mode) {
2521 /* This Sub is an effective Cast */
2522 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2524 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2529 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2530 ir_mode *mode = get_irn_mode(a);
2532 if (mode == get_irn_mode(b)) {
2534 ir_node *op_a = get_Conv_op(a);
2535 ir_node *op_b = get_Conv_op(b);
2537 /* check if it's allowed to skip the conv */
2538 ma = get_irn_mode(op_a);
2539 mb = get_irn_mode(op_b);
2541 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2542 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2545 set_Sub_right(n, b);
2551 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2552 if (!is_reassoc_running() && is_Mul(a)) {
2553 ir_node *ma = get_Mul_left(a);
2554 ir_node *mb = get_Mul_right(a);
2557 ir_node *blk = get_nodes_block(n);
2559 get_irn_dbg_info(n),
2560 current_ir_graph, blk,
2563 get_irn_dbg_info(n),
2564 current_ir_graph, blk,
2566 new_r_Const_long(current_ir_graph, mode, 1),
2569 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2571 } else if (mb == b) {
2572 ir_node *blk = get_nodes_block(n);
2574 get_irn_dbg_info(n),
2575 current_ir_graph, blk,
2578 get_irn_dbg_info(n),
2579 current_ir_graph, blk,
2581 new_r_Const_long(current_ir_graph, mode, 1),
2584 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2588 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2589 ir_node *x = get_Sub_left(a);
2590 ir_node *y = get_Sub_right(a);
2591 ir_node *blk = get_nodes_block(n);
2592 ir_mode *m_b = get_irn_mode(b);
2593 ir_mode *m_y = get_irn_mode(y);
2597 /* Determine the right mode for the Add. */
2600 else if (mode_is_reference(m_b))
2602 else if (mode_is_reference(m_y))
2606 * Both modes are different but none is reference,
2607 * happens for instance in SubP(SubP(P, Iu), Is).
2608 * We have two possibilities here: Cast or ignore.
2609 * Currently we ignore this case.
2614 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2616 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2617 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2621 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2622 if (is_Const(a) && is_Not(b)) {
2623 /* c - ~X = X + (c+1) */
2624 tarval *tv = get_Const_tarval(a);
2626 tv = tarval_add(tv, get_mode_one(mode));
2627 if (tv != tarval_bad) {
2628 ir_node *blk = get_nodes_block(n);
2629 ir_node *c = new_r_Const(current_ir_graph, mode, tv);
2630 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2631 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2637 } /* transform_node_Sub */
2640 * Several transformation done on n*n=2n bits mul.
2641 * These transformations must be done here because new nodes may be produced.
2643 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2645 ir_node *a = get_Mul_left(n);
2646 ir_node *b = get_Mul_right(n);
2647 tarval *ta = value_of(a);
2648 tarval *tb = value_of(b);
2649 ir_mode *smode = get_irn_mode(a);
2651 if (ta == get_mode_one(smode)) {
2652 /* (L)1 * (L)b = (L)b */
2653 ir_node *blk = get_nodes_block(n);
2654 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2655 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2658 else if (ta == get_mode_minus_one(smode)) {
2659 /* (L)-1 * (L)b = (L)b */
2660 ir_node *blk = get_nodes_block(n);
2661 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2662 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2663 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2666 if (tb == get_mode_one(smode)) {
2667 /* (L)a * (L)1 = (L)a */
2668 ir_node *blk = get_irn_n(a, -1);
2669 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2670 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2673 else if (tb == get_mode_minus_one(smode)) {
2674 /* (L)a * (L)-1 = (L)-a */
2675 ir_node *blk = get_nodes_block(n);
2676 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2677 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2678 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2685 * Transform Mul(a,-1) into -a.
2686 * Do constant evaluation of Phi nodes.
2687 * Do architecture dependent optimizations on Mul nodes
2689 static ir_node *transform_node_Mul(ir_node *n) {
2690 ir_node *c, *oldn = n;
2691 ir_mode *mode = get_irn_mode(n);
2692 ir_node *a = get_Mul_left(n);
2693 ir_node *b = get_Mul_right(n);
2695 if (is_Bad(a) || is_Bad(b))
2698 if (mode != get_irn_mode(a))
2699 return transform_node_Mul2n(n, mode);
2701 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2703 if (mode_is_signed(mode)) {
2706 if (value_of(a) == get_mode_minus_one(mode))
2708 else if (value_of(b) == get_mode_minus_one(mode))
2711 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2712 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2717 if (is_Const(b)) { /* (-a) * const -> a * -const */
2718 ir_node *cnst = const_negate(b);
2720 dbg_info *dbgi = get_irn_dbg_info(n);
2721 ir_node *block = get_nodes_block(n);
2722 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2723 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2726 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2727 dbg_info *dbgi = get_irn_dbg_info(n);
2728 ir_node *block = get_nodes_block(n);
2729 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2730 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2732 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2733 ir_node *sub_l = get_Sub_left(b);
2734 ir_node *sub_r = get_Sub_right(b);
2735 dbg_info *dbgi = get_irn_dbg_info(n);
2736 ir_graph *irg = current_ir_graph;
2737 ir_node *block = get_nodes_block(n);
2738 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2739 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2740 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2743 } else if (is_Minus(b)) {
2744 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2745 ir_node *sub_l = get_Sub_left(a);
2746 ir_node *sub_r = get_Sub_right(a);
2747 dbg_info *dbgi = get_irn_dbg_info(n);
2748 ir_graph *irg = current_ir_graph;
2749 ir_node *block = get_nodes_block(n);
2750 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2751 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2752 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2756 if (get_mode_arithmetic(mode) == irma_ieee754) {
2758 tarval *tv = get_Const_tarval(a);
2759 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2760 && !tarval_is_negative(tv)) {
2761 /* 2.0 * b = b + b */
2762 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2763 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2767 else if (is_Const(b)) {
2768 tarval *tv = get_Const_tarval(b);
2769 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2770 && !tarval_is_negative(tv)) {
2771 /* a * 2.0 = a + a */
2772 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2773 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2778 return arch_dep_replace_mul_with_shifts(n);
2779 } /* transform_node_Mul */
2782 * Transform a Div Node.
2784 static ir_node *transform_node_Div(ir_node *n) {
2785 ir_mode *mode = get_Div_resmode(n);
2786 ir_node *a = get_Div_left(n);
2787 ir_node *b = get_Div_right(n);
2791 if (is_Const(b) && is_const_Phi(a)) {
2792 /* check for Div(Phi, Const) */
2793 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2795 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2799 else if (is_Const(a) && is_const_Phi(b)) {
2800 /* check for Div(Const, Phi) */
2801 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2803 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2807 else if (is_const_Phi(a) && is_const_Phi(b)) {
2808 /* check for Div(Phi, Phi) */
2809 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2811 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2818 if (tv != tarval_bad) {
2819 value = new_Const(get_tarval_mode(tv), tv);
2821 DBG_OPT_CSTEVAL(n, value);
2824 ir_node *a = get_Div_left(n);
2825 ir_node *b = get_Div_right(n);
2826 const ir_node *dummy;
2828 if (a == b && value_not_zero(a, &dummy)) {
2829 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2830 value = new_Const(mode, get_mode_one(mode));
2831 DBG_OPT_CSTEVAL(n, value);
2834 if (mode_is_signed(mode) && is_Const(b)) {
2835 tarval *tv = get_Const_tarval(b);
2837 if (tv == get_mode_minus_one(mode)) {
2839 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2840 DBG_OPT_CSTEVAL(n, value);
2844 /* Try architecture dependent optimization */
2845 value = arch_dep_replace_div_by_const(n);
2853 /* Turn Div into a tuple (mem, jmp, bad, value) */
2854 mem = get_Div_mem(n);
2855 blk = get_nodes_block(n);
2857 /* skip a potential Pin */
2858 mem = skip_Pin(mem);
2859 turn_into_tuple(n, pn_Div_max);
2860 set_Tuple_pred(n, pn_Div_M, mem);
2861 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2862 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2863 set_Tuple_pred(n, pn_Div_res, value);
2866 } /* transform_node_Div */
2869 * Transform a Mod node.
2871 static ir_node *transform_node_Mod(ir_node *n) {
2872 ir_mode *mode = get_Mod_resmode(n);
2873 ir_node *a = get_Mod_left(n);
2874 ir_node *b = get_Mod_right(n);
2878 if (is_Const(b) && is_const_Phi(a)) {
2879 /* check for Div(Phi, Const) */
2880 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2882 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2886 else if (is_Const(a) && is_const_Phi(b)) {
2887 /* check for Div(Const, Phi) */
2888 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2890 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2894 else if (is_const_Phi(a) && is_const_Phi(b)) {
2895 /* check for Div(Phi, Phi) */
2896 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2898 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2905 if (tv != tarval_bad) {
2906 value = new_Const(get_tarval_mode(tv), tv);
2908 DBG_OPT_CSTEVAL(n, value);
2911 ir_node *a = get_Mod_left(n);
2912 ir_node *b = get_Mod_right(n);
2913 const ir_node *dummy;
2915 if (a == b && value_not_zero(a, &dummy)) {
2916 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2917 value = new_Const(mode, get_mode_null(mode));
2918 DBG_OPT_CSTEVAL(n, value);
2921 if (mode_is_signed(mode) && is_Const(b)) {
2922 tarval *tv = get_Const_tarval(b);
2924 if (tv == get_mode_minus_one(mode)) {
2926 value = new_Const(mode, get_mode_null(mode));
2927 DBG_OPT_CSTEVAL(n, value);
2931 /* Try architecture dependent optimization */
2932 value = arch_dep_replace_mod_by_const(n);
2940 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2941 mem = get_Mod_mem(n);
2942 blk = get_nodes_block(n);
2944 /* skip a potential Pin */
2945 mem = skip_Pin(mem);
2946 turn_into_tuple(n, pn_Mod_max);
2947 set_Tuple_pred(n, pn_Mod_M, mem);
2948 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2949 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2950 set_Tuple_pred(n, pn_Mod_res, value);
2953 } /* transform_node_Mod */
2956 * Transform a DivMod node.
2958 static ir_node *transform_node_DivMod(ir_node *n) {
2959 const ir_node *dummy;
2960 ir_node *a = get_DivMod_left(n);
2961 ir_node *b = get_DivMod_right(n);
2962 ir_mode *mode = get_DivMod_resmode(n);
2967 if (is_Const(b) && is_const_Phi(a)) {
2968 /* check for Div(Phi, Const) */
2969 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2970 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2972 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2973 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2977 else if (is_Const(a) && is_const_Phi(b)) {
2978 /* check for Div(Const, Phi) */
2979 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2980 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2982 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2983 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2987 else if (is_const_Phi(a) && is_const_Phi(b)) {
2988 /* check for Div(Phi, Phi) */
2989 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2990 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2992 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2993 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3000 if (tb != tarval_bad) {
3001 if (tb == get_mode_one(get_tarval_mode(tb))) {
3003 vb = new_Const(mode, get_mode_null(mode));
3004 DBG_OPT_CSTEVAL(n, vb);
3006 } else if (ta != tarval_bad) {
3007 tarval *resa, *resb;
3008 resa = tarval_div(ta, tb);
3009 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3010 Jmp for X result!? */
3011 resb = tarval_mod(ta, tb);
3012 if (resb == tarval_bad) return n; /* Causes exception! */
3013 va = new_Const(mode, resa);
3014 vb = new_Const(mode, resb);
3015 DBG_OPT_CSTEVAL(n, va);
3016 DBG_OPT_CSTEVAL(n, vb);
3018 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3019 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
3020 vb = new_Const(mode, get_mode_null(mode));
3021 DBG_OPT_CSTEVAL(n, va);
3022 DBG_OPT_CSTEVAL(n, vb);
3024 } else { /* Try architecture dependent optimization */
3027 arch_dep_replace_divmod_by_const(&va, &vb, n);
3028 evaluated = va != NULL;
3030 } else if (a == b) {
3031 if (value_not_zero(a, &dummy)) {
3033 va = new_Const(mode, get_mode_one(mode));
3034 vb = new_Const(mode, get_mode_null(mode));
3035 DBG_OPT_CSTEVAL(n, va);
3036 DBG_OPT_CSTEVAL(n, vb);
3039 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3042 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3043 /* 0 / non-Const = 0 */
3048 if (evaluated) { /* replace by tuple */
3052 mem = get_DivMod_mem(n);
3053 /* skip a potential Pin */
3054 mem = skip_Pin(mem);
3056 blk = get_nodes_block(n);
3057 turn_into_tuple(n, pn_DivMod_max);
3058 set_Tuple_pred(n, pn_DivMod_M, mem);
3059 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3060 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3061 set_Tuple_pred(n, pn_DivMod_res_div, va);
3062 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3066 } /* transform_node_DivMod */
3069 * Optimize x / c to x * (1/c)
3071 static ir_node *transform_node_Quot(ir_node *n) {
3072 ir_mode *mode = get_Quot_resmode(n);
3075 if (get_mode_arithmetic(mode) == irma_ieee754) {
3076 ir_node *b = get_Quot_right(n);
3077 tarval *tv = value_of(b);
3079 if (tv != tarval_bad) {
3083 * Floating point constant folding might be disabled here to
3085 * However, as we check for exact result, doing it is safe.
3088 rem = tarval_enable_fp_ops(1);
3089 tv = tarval_quo(get_mode_one(mode), tv);
3090 (void)tarval_enable_fp_ops(rem);
3092 /* Do the transformation if the result is either exact or we are not
3093 using strict rules. */
3094 if (tv != tarval_bad &&
3095 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3096 ir_node *blk = get_nodes_block(n);
3097 ir_node *c = new_r_Const(current_ir_graph, mode, tv);
3098 ir_node *a = get_Quot_left(n);
3099 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3100 ir_node *mem = get_Quot_mem(n);
3102 /* skip a potential Pin */
3103 mem = skip_Pin(mem);
3104 turn_into_tuple(n, pn_Quot_max);
3105 set_Tuple_pred(n, pn_Quot_M, mem);
3106 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3107 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3108 set_Tuple_pred(n, pn_Quot_res, m);
3109 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3114 } /* transform_node_Quot */
3117 * Optimize Abs(x) into x if x is Confirmed >= 0
3118 * Optimize Abs(x) into -x if x is Confirmed <= 0
3119 * Optimize Abs(-x) int Abs(x)
3121 static ir_node *transform_node_Abs(ir_node *n) {
3122 ir_node *c, *oldn = n;
3123 ir_node *a = get_Abs_op(n);
3126 HANDLE_UNOP_PHI(tarval_abs, a, c);
3128 switch (classify_value_sign(a)) {
3129 case value_classified_negative:
3130 mode = get_irn_mode(n);
3133 * We can replace the Abs by -x here.
3134 * We even could add a new Confirm here
3135 * (if not twos complement)
3137 * Note that -x would create a new node, so we could
3138 * not run it in the equivalent_node() context.
3140 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3141 get_nodes_block(n), a, mode);
3143 DBG_OPT_CONFIRM(oldn, n);
3145 case value_classified_positive:
3146 /* n is positive, Abs is not needed */
3149 DBG_OPT_CONFIRM(oldn, n);
3155 /* Abs(-x) = Abs(x) */
3156 mode = get_irn_mode(n);
3157 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3158 get_nodes_block(n), get_Minus_op(a), mode);
3159 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3163 } /* transform_node_Abs */
3166 * Optimize -a CMP -b into b CMP a.
3167 * This works only for for modes where unary Minus
3169 * Note that two-complement integers can Overflow
3170 * so it will NOT work.
3172 * For == and != can be handled in Proj(Cmp)
3174 static ir_node *transform_node_Cmp(ir_node *n) {
3176 ir_node *left = get_Cmp_left(n);
3177 ir_node *right = get_Cmp_right(n);
3179 if (is_Minus(left) && is_Minus(right) &&
3180 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3181 ir_node *const new_left = get_Minus_op(right);
3182 ir_node *const new_right = get_Minus_op(left);
3183 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3184 get_nodes_block(n), new_left, new_right);
3185 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3188 } /* transform_node_Cmp */
3192 * Transform a Cond node.
3194 * Replace the Cond by a Jmp if it branches on a constant
3197 static ir_node *transform_node_Cond(ir_node *n) {
3200 ir_node *a = get_Cond_selector(n);
3201 tarval *ta = value_of(a);
3203 /* we need block info which is not available in floating irgs */
3204 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3207 if ((ta != tarval_bad) &&
3208 (get_irn_mode(a) == mode_b) &&
3209 (get_opt_unreachable_code())) {
3210 /* It's a boolean Cond, branching on a boolean constant.
3211 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3212 ir_node *blk = get_nodes_block(n);
3213 jmp = new_r_Jmp(current_ir_graph, blk);
3214 turn_into_tuple(n, pn_Cond_max);
3215 if (ta == tarval_b_true) {
3216 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3217 set_Tuple_pred(n, pn_Cond_true, jmp);
3219 set_Tuple_pred(n, pn_Cond_false, jmp);
3220 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3222 /* We might generate an endless loop, so keep it alive. */
3223 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3226 } /* transform_node_Cond */
3229 * Prototype of a recursive transform function
3230 * for bitwise distributive transformations.
3232 typedef ir_node* (*recursive_transform)(ir_node *n);
3235 * makes use of distributive laws for and, or, eor
3236 * and(a OP c, b OP c) -> and(a, b) OP c
3237 * note, might return a different op than n
3239 static ir_node *transform_bitwise_distributive(ir_node *n,
3240 recursive_transform trans_func)
3243 ir_node *a = get_binop_left(n);
3244 ir_node *b = get_binop_right(n);
3245 ir_op *op = get_irn_op(a);
3246 ir_op *op_root = get_irn_op(n);
3248 if(op != get_irn_op(b))
3251 if (op == op_Conv) {
3252 ir_node *a_op = get_Conv_op(a);
3253 ir_node *b_op = get_Conv_op(b);
3254 ir_mode *a_mode = get_irn_mode(a_op);
3255 ir_mode *b_mode = get_irn_mode(b_op);
3256 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3257 ir_node *blk = get_nodes_block(n);
3260 set_binop_left(n, a_op);
3261 set_binop_right(n, b_op);
3262 set_irn_mode(n, a_mode);
3264 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3266 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3272 /* nothing to gain here */
3276 if (op == op_Shrs || op == op_Shr || op == op_Shl
3277 || op == op_And || op == op_Or || op == op_Eor) {
3278 ir_node *a_left = get_binop_left(a);
3279 ir_node *a_right = get_binop_right(a);
3280 ir_node *b_left = get_binop_left(b);
3281 ir_node *b_right = get_binop_right(b);
3283 ir_node *op1 = NULL;
3284 ir_node *op2 = NULL;
3286 if (is_op_commutative(op)) {
3287 if (a_left == b_left) {
3291 } else if(a_left == b_right) {
3295 } else if(a_right == b_left) {
3301 if(a_right == b_right) {
3308 /* (a sop c) & (b sop c) => (a & b) sop c */
3309 ir_node *blk = get_nodes_block(n);
3311 ir_node *new_n = exact_copy(n);
3312 set_binop_left(new_n, op1);
3313 set_binop_right(new_n, op2);
3314 new_n = trans_func(new_n);
3316 if(op_root == op_Eor && op == op_Or) {
3317 dbg_info *dbgi = get_irn_dbg_info(n);
3318 ir_graph *irg = current_ir_graph;
3319 ir_mode *mode = get_irn_mode(c);
3321 c = new_rd_Not(dbgi, irg, blk, c, mode);
3322 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3325 set_nodes_block(n, blk);
3326 set_binop_left(n, new_n);
3327 set_binop_right(n, c);
3328 add_identities(current_ir_graph->value_table, n);
3331 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3342 static ir_node *transform_node_And(ir_node *n) {
3343 ir_node *c, *oldn = n;
3344 ir_node *a = get_And_left(n);
3345 ir_node *b = get_And_right(n);
3348 mode = get_irn_mode(n);
3349 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3351 /* we can evaluate 2 Projs of the same Cmp */
3352 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3353 ir_node *pred_a = get_Proj_pred(a);
3354 ir_node *pred_b = get_Proj_pred(b);
3355 if (pred_a == pred_b) {
3356 dbg_info *dbgi = get_irn_dbg_info(n);
3357 ir_node *block = get_nodes_block(pred_a);
3358 pn_Cmp pn_a = get_Proj_proj(a);
3359 pn_Cmp pn_b = get_Proj_proj(b);
3360 /* yes, we can simply calculate with pncs */
3361 pn_Cmp new_pnc = pn_a & pn_b;
3363 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3368 ir_node *op = get_Not_op(b);
3370 ir_node *ba = get_And_left(op);
3371 ir_node *bb = get_And_right(op);
3373 /* it's enough to test the following cases due to normalization! */
3374 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3375 /* (a|b) & ~(a&b) = a^b */
3376 ir_node *block = get_nodes_block(n);
3378 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3379 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3387 ir_node *op = get_Not_op(a);
3389 ir_node *aa = get_And_left(op);
3390 ir_node *ab = get_And_right(op);
3392 /* it's enough to test the following cases due to normalization! */
3393 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3394 /* (a|b) & ~(a&b) = a^b */
3395 ir_node *block = get_nodes_block(n);
3397 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3398 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3405 ir_node *al = get_Eor_left(a);
3406 ir_node *ar = get_Eor_right(a);
3409 /* (b ^ a) & b -> ~a & b */
3410 dbg_info *dbg = get_irn_dbg_info(n);
3411 ir_node *block = get_nodes_block(n);
3413 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3414 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3415 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3419 /* (a ^ b) & b -> ~a & b */
3420 dbg_info *dbg = get_irn_dbg_info(n);
3421 ir_node *block = get_nodes_block(n);
3423 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3424 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3425 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3430 ir_node *bl = get_Eor_left(b);
3431 ir_node *br = get_Eor_right(b);
3434 /* a & (a ^ b) -> a & ~b */
3435 dbg_info *dbg = get_irn_dbg_info(n);
3436 ir_node *block = get_nodes_block(n);
3438 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3439 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3440 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3444 /* a & (b ^ a) -> a & ~b */
3445 dbg_info *dbg = get_irn_dbg_info(n);
3446 ir_node *block = get_nodes_block(n);
3448 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3449 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3450 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3454 if (is_Not(a) && is_Not(b)) {
3455 /* ~a & ~b = ~(a|b) */
3456 ir_node *block = get_nodes_block(n);
3457 ir_mode *mode = get_irn_mode(n);
3461 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3462 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3463 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3467 n = transform_bitwise_distributive(n, transform_node_And);
3470 } /* transform_node_And */
3475 static ir_node *transform_node_Eor(ir_node *n) {
3476 ir_node *c, *oldn = n;
3477 ir_node *a = get_Eor_left(n);
3478 ir_node *b = get_Eor_right(n);
3479 ir_mode *mode = get_irn_mode(n);
3481 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3483 /* we can evaluate 2 Projs of the same Cmp */
3484 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3485 ir_node *pred_a = get_Proj_pred(a);
3486 ir_node *pred_b = get_Proj_pred(b);
3487 if(pred_a == pred_b) {
3488 dbg_info *dbgi = get_irn_dbg_info(n);
3489 ir_node *block = get_nodes_block(pred_a);
3490 pn_Cmp pn_a = get_Proj_proj(a);
3491 pn_Cmp pn_b = get_Proj_proj(b);
3492 /* yes, we can simply calculate with pncs */
3493 pn_Cmp new_pnc = pn_a ^ pn_b;
3495 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3502 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3503 mode, get_mode_null(mode));
3504 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3505 } else if (mode == mode_b &&
3507 is_Const(b) && is_Const_one(b) &&
3508 is_Cmp(get_Proj_pred(a))) {
3509 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3510 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3511 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3513 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3514 } else if (is_Const(b)) {
3515 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3516 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b)));
3517 ir_node *not_op = get_Not_op(a);
3518 dbg_info *dbg = get_irn_dbg_info(n);
3519 ir_graph *irg = current_ir_graph;
3520 ir_node *block = get_nodes_block(n);
3521 ir_mode *mode = get_irn_mode(n);
3522 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3524 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3525 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3526 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3529 n = transform_bitwise_distributive(n, transform_node_Eor);
3533 } /* transform_node_Eor */
3538 static ir_node *transform_node_Not(ir_node *n) {
3539 ir_node *c, *oldn = n;
3540 ir_node *a = get_Not_op(n);
3541 ir_mode *mode = get_irn_mode(n);
3543 HANDLE_UNOP_PHI(tarval_not,a,c);
3545 /* check for a boolean Not */
3546 if (mode == mode_b &&
3548 is_Cmp(get_Proj_pred(a))) {
3549 /* We negate a Cmp. The Cmp has the negated result anyways! */
3550 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3551 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3552 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3556 ir_node *eor_b = get_Eor_right(a);
3557 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3558 ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b)));
3559 ir_node *eor_a = get_Eor_left(a);
3560 dbg_info *dbg = get_irn_dbg_info(n);
3561 ir_graph *irg = current_ir_graph;
3562 ir_node *block = get_nodes_block(n);
3563 ir_mode *mode = get_irn_mode(n);
3564 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3568 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3569 if (is_Minus(a)) { /* ~-x -> x + -1 */
3570 dbg_info *dbg = get_irn_dbg_info(n);
3571 ir_graph *irg = current_ir_graph;
3572 ir_node *block = get_nodes_block(n);
3573 ir_node *add_l = get_Minus_op(a);
3574 ir_node *add_r = new_rd_Const(dbg, irg, mode, get_mode_minus_one(mode));
3575 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3576 } else if (is_Add(a)) {
3577 ir_node *add_r = get_Add_right(a);
3578 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3579 /* ~(x + -1) = -x */
3580 ir_node *op = get_Add_left(a);
3581 ir_node *blk = get_nodes_block(n);
3582 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3583 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3588 } /* transform_node_Not */
3591 * Transform a Minus.
3595 * -(a >>u (size-1)) = a >>s (size-1)
3596 * -(a >>s (size-1)) = a >>u (size-1)
3597 * -(a * const) -> a * -const
3599 static ir_node *transform_node_Minus(ir_node *n) {
3600 ir_node *c, *oldn = n;
3601 ir_node *a = get_Minus_op(n);
3604 HANDLE_UNOP_PHI(tarval_neg,a,c);
3606 mode = get_irn_mode(a);
3607 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3608 /* the following rules are only to twos-complement */
3611 ir_node *op = get_Not_op(a);
3612 tarval *tv = get_mode_one(mode);
3613 ir_node *blk = get_nodes_block(n);
3614 ir_node *c = new_r_Const(current_ir_graph, mode, tv);
3615 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3616 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3620 ir_node *c = get_Shr_right(a);
3623 tarval *tv = get_Const_tarval(c);
3625 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3626 /* -(a >>u (size-1)) = a >>s (size-1) */
3627 ir_node *v = get_Shr_left(a);
3629 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3630 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3636 ir_node *c = get_Shrs_right(a);
3639 tarval *tv = get_Const_tarval(c);
3641 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3642 /* -(a >>s (size-1)) = a >>u (size-1) */
3643 ir_node *v = get_Shrs_left(a);
3645 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3646 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3653 /* - (a-b) = b - a */
3654 ir_node *la = get_Sub_left(a);
3655 ir_node *ra = get_Sub_right(a);
3656 ir_node *blk = get_nodes_block(n);
3658 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3659 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3663 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3664 ir_node *mul_l = get_Mul_left(a);
3665 ir_node *mul_r = get_Mul_right(a);
3666 tarval *tv = value_of(mul_r);
3667 if (tv != tarval_bad) {
3668 tv = tarval_neg(tv);
3669 if (tv != tarval_bad) {
3670 ir_node *cnst = new_Const(mode, tv);
3671 dbg_info *dbg = get_irn_dbg_info(a);
3672 ir_graph *irg = current_ir_graph;
3673 ir_node *block = get_nodes_block(a);
3674 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3675 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3682 } /* transform_node_Minus */
3685 * Transform a Cast_type(Const) into a new Const_type
3687 static ir_node *transform_node_Cast(ir_node *n) {
3689 ir_node *pred = get_Cast_op(n);
3690 ir_type *tp = get_irn_type(n);
3692 if (is_Const(pred) && get_Const_type(pred) != tp) {
3693 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_mode(pred),
3694 get_Const_tarval(pred), tp);
3695 DBG_OPT_CSTEVAL(oldn, n);
3696 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3697 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3698 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3699 DBG_OPT_CSTEVAL(oldn, n);
3703 } /* transform_node_Cast */
3706 * Transform a Proj(Load) with a non-null address.
3708 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3709 if (get_opt_ldst_only_null_ptr_exceptions()) {
3710 if (get_irn_mode(proj) == mode_X) {
3711 ir_node *load = get_Proj_pred(proj);
3713 /* get the Load address */
3714 const ir_node *addr = get_Load_ptr(load);
3715 const ir_node *confirm;
3717 if (value_not_null(addr, &confirm)) {
3718 if (confirm == NULL) {
3719 /* this node may float if it did not depend on a Confirm */
3720 set_irn_pinned(load, op_pin_state_floats);
3722 if (get_Proj_proj(proj) == pn_Load_X_except) {
3723 DBG_OPT_EXC_REM(proj);
3724 return get_irg_bad(current_ir_graph);
3726 ir_node *blk = get_nodes_block(load);
3727 return new_r_Jmp(current_ir_graph, blk);
3733 } /* transform_node_Proj_Load */
3736 * Transform a Proj(Store) with a non-null address.
3738 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3739 if (get_opt_ldst_only_null_ptr_exceptions()) {
3740 if (get_irn_mode(proj) == mode_X) {
3741 ir_node *store = get_Proj_pred(proj);
3743 /* get the load/store address */
3744 const ir_node *addr = get_Store_ptr(store);
3745 const ir_node *confirm;
3747 if (value_not_null(addr, &confirm)) {
3748 if (confirm == NULL) {
3749 /* this node may float if it did not depend on a Confirm */
3750 set_irn_pinned(store, op_pin_state_floats);
3752 if (get_Proj_proj(proj) == pn_Store_X_except) {
3753 DBG_OPT_EXC_REM(proj);
3754 return get_irg_bad(current_ir_graph);
3756 ir_node *blk = get_nodes_block(store);
3757 return new_r_Jmp(current_ir_graph, blk);
3763 } /* transform_node_Proj_Store */
3766 * Transform a Proj(Div) with a non-zero value.
3767 * Removes the exceptions and routes the memory to the NoMem node.
3769 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3770 ir_node *div = get_Proj_pred(proj);
3771 ir_node *b = get_Div_right(div);
3772 ir_node *res, *new_mem;
3773 const ir_node *confirm;
3776 if (value_not_zero(b, &confirm)) {
3777 /* div(x, y) && y != 0 */
3778 if (confirm == NULL) {
3779 /* we are sure we have a Const != 0 */
3780 new_mem = get_Div_mem(div);
3781 new_mem = skip_Pin(new_mem);
3782 set_Div_mem(div, new_mem);
3783 set_irn_pinned(div, op_pin_state_floats);
3786 proj_nr = get_Proj_proj(proj);
3788 case pn_Div_X_regular:
3789 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3791 case pn_Div_X_except:
3792 /* we found an exception handler, remove it */
3793 DBG_OPT_EXC_REM(proj);
3797 res = get_Div_mem(div);
3798 new_mem = get_irg_no_mem(current_ir_graph);
3801 /* This node can only float up to the Confirm block */
3802 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3804 set_irn_pinned(div, op_pin_state_floats);
3805 /* this is a Div without exception, we can remove the memory edge */
3806 set_Div_mem(div, new_mem);
3811 } /* transform_node_Proj_Div */
3814 * Transform a Proj(Mod) with a non-zero value.
3815 * Removes the exceptions and routes the memory to the NoMem node.
3817 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3818 ir_node *mod = get_Proj_pred(proj);
3819 ir_node *b = get_Mod_right(mod);
3820 ir_node *res, *new_mem;
3821 const ir_node *confirm;
3824 if (value_not_zero(b, &confirm)) {
3825 /* mod(x, y) && y != 0 */
3826 proj_nr = get_Proj_proj(proj);
3828 if (confirm == NULL) {
3829 /* we are sure we have a Const != 0 */
3830 new_mem = get_Mod_mem(mod);
3831 new_mem = skip_Pin(new_mem);
3832 set_Mod_mem(mod, new_mem);
3833 set_irn_pinned(mod, op_pin_state_floats);
3838 case pn_Mod_X_regular:
3839 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3841 case pn_Mod_X_except:
3842 /* we found an exception handler, remove it */
3843 DBG_OPT_EXC_REM(proj);
3847 res = get_Mod_mem(mod);
3848 new_mem = get_irg_no_mem(current_ir_graph);
3851 /* This node can only float up to the Confirm block */
3852 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3854 /* this is a Mod without exception, we can remove the memory edge */
3855 set_Mod_mem(mod, new_mem);
3858 if (get_Mod_left(mod) == b) {
3859 /* a % a = 0 if a != 0 */
3860 ir_mode *mode = get_irn_mode(proj);
3861 ir_node *res = new_Const(mode, get_mode_null(mode));
3863 DBG_OPT_CSTEVAL(mod, res);
3869 } /* transform_node_Proj_Mod */
3872 * Transform a Proj(DivMod) with a non-zero value.
3873 * Removes the exceptions and routes the memory to the NoMem node.
3875 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3876 ir_node *divmod = get_Proj_pred(proj);
3877 ir_node *b = get_DivMod_right(divmod);
3878 ir_node *res, *new_mem;
3879 const ir_node *confirm;
3882 if (value_not_zero(b, &confirm)) {
3883 /* DivMod(x, y) && y != 0 */
3884 proj_nr = get_Proj_proj(proj);
3886 if (confirm == NULL) {
3887 /* we are sure we have a Const != 0 */
3888 new_mem = get_DivMod_mem(divmod);
3889 new_mem = skip_Pin(new_mem);
3890 set_DivMod_mem(divmod, new_mem);
3891 set_irn_pinned(divmod, op_pin_state_floats);
3896 case pn_DivMod_X_regular:
3897 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3899 case pn_DivMod_X_except:
3900 /* we found an exception handler, remove it */
3901 DBG_OPT_EXC_REM(proj);
3905 res = get_DivMod_mem(divmod);
3906 new_mem = get_irg_no_mem(current_ir_graph);
3909 /* This node can only float up to the Confirm block */
3910 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3912 /* this is a DivMod without exception, we can remove the memory edge */
3913 set_DivMod_mem(divmod, new_mem);
3916 case pn_DivMod_res_mod:
3917 if (get_DivMod_left(divmod) == b) {
3918 /* a % a = 0 if a != 0 */
3919 ir_mode *mode = get_irn_mode(proj);
3920 ir_node *res = new_Const(mode, get_mode_null(mode));
3922 DBG_OPT_CSTEVAL(divmod, res);
3928 } /* transform_node_Proj_DivMod */
3931 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3933 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3934 if (get_opt_unreachable_code()) {
3935 ir_node *n = get_Proj_pred(proj);
3936 ir_node *b = get_Cond_selector(n);
3938 if (mode_is_int(get_irn_mode(b))) {
3939 tarval *tb = value_of(b);
3941 if (tb != tarval_bad) {
3942 /* we have a constant switch */
3943 long num = get_Proj_proj(proj);
3945 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3946 if (get_tarval_long(tb) == num) {
3947 /* Do NOT create a jump here, or we will have 2 control flow ops
3948 * in a block. This case is optimized away in optimize_cf(). */
3951 /* this case will NEVER be taken, kill it */
3952 return get_irg_bad(current_ir_graph);
3959 } /* transform_node_Proj_Cond */
3962 * Create a 0 constant of given mode.
3964 static ir_node *create_zero_const(ir_mode *mode) {
3965 tarval *tv = get_mode_null(mode);
3966 ir_node *cnst = new_Const(mode, tv);
3971 /* the order of the values is important! */
3972 typedef enum const_class {
3978 static const_class classify_const(const ir_node* n)
3980 if (is_Const(n)) return const_const;
3981 if (is_irn_constlike(n)) return const_like;
3986 * Determines whether r is more constlike or has a larger index (in that order)
3989 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3991 const const_class l_order = classify_const(l);
3992 const const_class r_order = classify_const(r);
3994 l_order > r_order ||
3995 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3999 * Normalizes and optimizes Cmp nodes.
4001 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4002 ir_node *n = get_Proj_pred(proj);
4003 ir_node *left = get_Cmp_left(n);
4004 ir_node *right = get_Cmp_right(n);
4007 ir_mode *mode = NULL;
4008 long proj_nr = get_Proj_proj(proj);
4010 /* we can evaluate some cases directly */
4013 return new_Const(mode_b, get_tarval_b_false());
4015 return new_Const(mode_b, get_tarval_b_true());
4017 if (!mode_is_float(get_irn_mode(left)))
4018 return new_Const(mode_b, get_tarval_b_true());
4024 /* remove Casts of both sides */
4025 left = skip_Cast(left);
4026 right = skip_Cast(right);
4028 /* Remove unnecessary conversions */
4029 /* TODO handle constants */
4030 if (is_Conv(left) && is_Conv(right)) {
4031 ir_mode *mode = get_irn_mode(left);
4032 ir_node *op_left = get_Conv_op(left);
4033 ir_node *op_right = get_Conv_op(right);
4034 ir_mode *mode_left = get_irn_mode(op_left);
4035 ir_mode *mode_right = get_irn_mode(op_right);
4037 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4038 && mode_left != mode_b && mode_right != mode_b) {
4039 ir_graph *irg = current_ir_graph;
4040 ir_node *block = get_nodes_block(n);
4042 if (mode_left == mode_right) {
4046 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4047 } else if (smaller_mode(mode_left, mode_right)) {
4048 left = new_r_Conv(irg, block, op_left, mode_right);
4051 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4052 } else if (smaller_mode(mode_right, mode_left)) {
4054 right = new_r_Conv(irg, block, op_right, mode_left);
4056 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4061 /* remove operation on both sides if possible */
4062 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4064 * The following operations are NOT safe for floating point operations, for instance
4065 * 1.0 + inf == 2.0 + inf, =/=> x == y
4067 if (mode_is_int(get_irn_mode(left))) {
4068 unsigned lop = get_irn_opcode(left);
4070 if (lop == get_irn_opcode(right)) {
4071 ir_node *ll, *lr, *rl, *rr;
4073 /* same operation on both sides, try to remove */
4077 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4078 left = get_unop_op(left);
4079 right = get_unop_op(right);
4081 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4084 ll = get_Add_left(left);
4085 lr = get_Add_right(left);
4086 rl = get_Add_left(right);
4087 rr = get_Add_right(right);
4090 /* X + a CMP X + b ==> a CMP b */
4094 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4095 } else if (ll == rr) {
4096 /* X + a CMP b + X ==> a CMP b */
4100 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4101 } else if (lr == rl) {
4102 /* a + X CMP X + b ==> a CMP b */
4106 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4107 } else if (lr == rr) {
4108 /* a + X CMP b + X ==> a CMP b */
4112 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4116 ll = get_Sub_left(left);
4117 lr = get_Sub_right(left);
4118 rl = get_Sub_left(right);
4119 rr = get_Sub_right(right);
4122 /* X - a CMP X - b ==> a CMP b */
4126 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4127 } else if (lr == rr) {
4128 /* a - X CMP b - X ==> a CMP b */
4132 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4136 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4137 /* a ROTL X CMP b ROTL X ==> a CMP b */
4138 left = get_Rotl_left(left);
4139 right = get_Rotl_left(right);
4141 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4149 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4150 if (is_Add(left) || is_Sub(left)) {
4151 ir_node *ll = get_binop_left(left);
4152 ir_node *lr = get_binop_right(left);
4154 if (lr == right && is_Add(left)) {
4161 right = create_zero_const(get_irn_mode(left));
4163 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4166 if (is_Add(right) || is_Sub(right)) {
4167 ir_node *rl = get_binop_left(right);
4168 ir_node *rr = get_binop_right(right);
4170 if (rr == left && is_Add(right)) {
4177 right = create_zero_const(get_irn_mode(left));
4179 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4182 if (is_And(left) && is_Const(right)) {
4183 ir_node *ll = get_binop_left(left);
4184 ir_node *lr = get_binop_right(left);
4185 if (is_Shr(ll) && is_Const(lr)) {
4186 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4187 ir_node *block = get_nodes_block(n);
4188 ir_mode *mode = get_irn_mode(left);
4190 ir_node *llr = get_Shr_right(ll);
4191 if (is_Const(llr)) {
4192 ir_graph *irg = current_ir_graph;
4193 dbg_info *dbg = get_irn_dbg_info(left);
4195 tarval *c1 = get_Const_tarval(llr);
4196 tarval *c2 = get_Const_tarval(lr);
4197 tarval *c3 = get_Const_tarval(right);
4198 tarval *mask = tarval_shl(c2, c1);
4199 tarval *value = tarval_shl(c3, c1);
4201 left = new_rd_And(dbg, irg, block, get_Shr_left(ll), new_Const(mode, mask), mode);
4202 right = new_Const(mode, value);
4207 } /* mode_is_int(...) */
4208 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4210 /* replace mode_b compares with ands/ors */
4211 if (get_irn_mode(left) == mode_b) {
4212 ir_graph *irg = current_ir_graph;
4213 ir_node *block = get_nodes_block(n);
4217 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4218 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4219 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4220 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4221 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4222 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4223 default: bres = NULL;
4226 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4232 * First step: normalize the compare op
4233 * by placing the constant on the right side
4234 * or moving the lower address node to the left.
4236 if (!operands_are_normalized(left, right)) {
4242 proj_nr = get_inversed_pnc(proj_nr);
4247 * Second step: Try to reduce the magnitude
4248 * of a constant. This may help to generate better code
4249 * later and may help to normalize more compares.
4250 * Of course this is only possible for integer values.
4252 tv = value_of(right);
4253 if (tv != tarval_bad) {
4254 mode = get_irn_mode(right);
4256 /* TODO extend to arbitrary constants */
4257 if (is_Conv(left) && tarval_is_null(tv)) {
4258 ir_node *op = get_Conv_op(left);
4259 ir_mode *op_mode = get_irn_mode(op);
4262 * UpConv(x) REL 0 ==> x REL 0
4264 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4265 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4266 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4267 tv = get_mode_null(op_mode);
4271 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4275 if (tv != tarval_bad) {
4276 /* the following optimization is possible on modes without Overflow
4277 * on Unary Minus or on == and !=:
4278 * -a CMP c ==> a swap(CMP) -c
4280 * Beware: for two-complement Overflow may occur, so only == and != can
4281 * be optimized, see this:
4282 * -MININT < 0 =/=> MININT > 0 !!!
4284 if (is_Minus(left) &&
4285 (!mode_overflow_on_unary_Minus(mode) ||
4286 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4287 tv = tarval_neg(tv);
4289 if (tv != tarval_bad) {
4290 left = get_Minus_op(left);
4291 proj_nr = get_inversed_pnc(proj_nr);
4293 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4295 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4296 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4297 tv = tarval_not(tv);
4299 if (tv != tarval_bad) {
4300 left = get_Not_op(left);
4302 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4306 /* for integer modes, we have more */
4307 if (mode_is_int(mode)) {
4308 /* Ne includes Unordered which is not possible on integers.
4309 * However, frontends often use this wrong, so fix it here */
4310 if (proj_nr & pn_Cmp_Uo) {
4311 proj_nr &= ~pn_Cmp_Uo;
4312 set_Proj_proj(proj, proj_nr);
4315 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4316 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4317 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4318 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4320 if (tv != tarval_bad) {
4321 proj_nr ^= pn_Cmp_Eq;
4323 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4326 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4327 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4328 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4329 tv = tarval_add(tv, get_mode_one(mode));
4331 if (tv != tarval_bad) {
4332 proj_nr ^= pn_Cmp_Eq;
4334 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4338 /* the following reassociations work only for == and != */
4339 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4341 #if 0 /* Might be not that good in general */
4342 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4343 if (tarval_is_null(tv) && is_Sub(left)) {
4344 right = get_Sub_right(left);
4345 left = get_Sub_left(left);
4347 tv = value_of(right);
4349 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4353 if (tv != tarval_bad) {
4354 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4356 ir_node *c1 = get_Sub_right(left);
4357 tarval *tv2 = value_of(c1);
4359 if (tv2 != tarval_bad) {
4360 tv2 = tarval_add(tv, value_of(c1));
4362 if (tv2 != tarval_bad) {
4363 left = get_Sub_left(left);
4366 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4370 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4371 else if (is_Add(left)) {
4372 ir_node *a_l = get_Add_left(left);
4373 ir_node *a_r = get_Add_right(left);
4377 if (is_Const(a_l)) {
4379 tv2 = value_of(a_l);
4382 tv2 = value_of(a_r);
4385 if (tv2 != tarval_bad) {
4386 tv2 = tarval_sub(tv, tv2, NULL);
4388 if (tv2 != tarval_bad) {
4392 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4396 /* -a == c ==> a == -c, -a != c ==> a != -c */
4397 else if (is_Minus(left)) {
4398 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4400 if (tv2 != tarval_bad) {
4401 left = get_Minus_op(left);
4404 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4409 /* the following reassociations work only for <= */
4410 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4411 if (tv != tarval_bad) {
4412 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4413 if (is_Abs(left)) { // TODO something is missing here
4419 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4420 switch (get_irn_opcode(left)) {
4424 c1 = get_And_right(left);
4427 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4428 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4430 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4432 /* TODO: move to constant evaluation */
4433 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4434 c1 = new_Const(mode_b, tv);
4435 DBG_OPT_CSTEVAL(proj, c1);
4439 if (tarval_is_single_bit(tv)) {
4441 * optimization for AND:
4443 * And(x, C) == C ==> And(x, C) != 0
4444 * And(x, C) != C ==> And(X, C) == 0
4446 * if C is a single Bit constant.
4449 /* check for Constant's match. We have check hare the tarvals,
4450 because our const might be changed */
4451 if (get_Const_tarval(c1) == tv) {
4452 /* fine: do the transformation */
4453 tv = get_mode_null(get_tarval_mode(tv));
4454 proj_nr ^= pn_Cmp_Leg;
4456 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4462 c1 = get_Or_right(left);
4463 if (is_Const(c1) && tarval_is_null(tv)) {
4465 * Or(x, C) == 0 && C != 0 ==> FALSE
4466 * Or(x, C) != 0 && C != 0 ==> TRUE
4468 if (! tarval_is_null(get_Const_tarval(c1))) {
4469 /* TODO: move to constant evaluation */
4470 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4471 c1 = new_Const(mode_b, tv);
4472 DBG_OPT_CSTEVAL(proj, c1);
4479 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4481 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4484 c1 = get_Shl_right(left);
4486 tarval *tv1 = get_Const_tarval(c1);
4487 ir_mode *mode = get_irn_mode(left);
4488 tarval *minus1 = get_mode_all_one(mode);
4489 tarval *amask = tarval_shr(minus1, tv1);
4490 tarval *cmask = tarval_shl(minus1, tv1);
4493 if (tarval_and(tv, cmask) != tv) {
4494 /* condition not met */
4495 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4496 c1 = new_Const(mode_b, tv);
4497 DBG_OPT_CSTEVAL(proj, c1);
4500 sl = get_Shl_left(left);
4501 blk = get_nodes_block(n);
4502 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4503 tv = tarval_shr(tv, tv1);
4505 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4510 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4512 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4515 c1 = get_Shr_right(left);
4517 tarval *tv1 = get_Const_tarval(c1);
4518 ir_mode *mode = get_irn_mode(left);
4519 tarval *minus1 = get_mode_all_one(mode);
4520 tarval *amask = tarval_shl(minus1, tv1);
4521 tarval *cmask = tarval_shr(minus1, tv1);
4524 if (tarval_and(tv, cmask) != tv) {
4525 /* condition not met */
4526 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4527 c1 = new_Const(mode_b, tv);
4528 DBG_OPT_CSTEVAL(proj, c1);
4531 sl = get_Shr_left(left);
4532 blk = get_nodes_block(n);
4533 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4534 tv = tarval_shl(tv, tv1);
4536 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4541 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4543 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4546 c1 = get_Shrs_right(left);
4548 tarval *tv1 = get_Const_tarval(c1);
4549 ir_mode *mode = get_irn_mode(left);
4550 tarval *minus1 = get_mode_all_one(mode);
4551 tarval *amask = tarval_shl(minus1, tv1);
4552 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4555 cond = tarval_sub(cond, tv1, NULL);
4556 cond = tarval_shrs(tv, cond);
4558 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4559 /* condition not met */
4560 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4561 c1 = new_Const(mode_b, tv);
4562 DBG_OPT_CSTEVAL(proj, c1);
4565 sl = get_Shrs_left(left);
4566 blk = get_nodes_block(n);
4567 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode);
4568 tv = tarval_shl(tv, tv1);
4570 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4575 } /* tarval != bad */
4578 if (changed & 2) /* need a new Const */
4579 right = new_Const(mode, tv);
4581 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4582 ir_node *op = get_Proj_pred(left);
4584 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4585 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4586 ir_node *c = get_binop_right(op);
4589 tarval *tv = get_Const_tarval(c);
4591 if (tarval_is_single_bit(tv)) {
4592 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4593 ir_node *v = get_binop_left(op);
4594 ir_node *blk = get_irn_n(op, -1);
4595 ir_mode *mode = get_irn_mode(v);
4597 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4598 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode);
4600 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4607 ir_node *block = get_nodes_block(n);
4609 /* create a new compare */
4610 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4611 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4615 } /* transform_node_Proj_Cmp */
4618 * Optimize CopyB(mem, x, x) into a Nop.
4620 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4621 ir_node *copyb = get_Proj_pred(proj);
4622 ir_node *a = get_CopyB_dst(copyb);
4623 ir_node *b = get_CopyB_src(copyb);
4626 switch (get_Proj_proj(proj)) {
4627 case pn_CopyB_X_regular:
4628 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4629 DBG_OPT_EXC_REM(proj);
4630 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4632 case pn_CopyB_M_except:
4633 case pn_CopyB_X_except:
4634 DBG_OPT_EXC_REM(proj);
4635 proj = get_irg_bad(current_ir_graph);
4642 } /* transform_node_Proj_CopyB */
4645 * Optimize Bounds(idx, idx, upper) into idx.
4647 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4648 ir_node *oldn = proj;
4649 ir_node *bound = get_Proj_pred(proj);
4650 ir_node *idx = get_Bound_index(bound);
4651 ir_node *pred = skip_Proj(idx);
4654 if (idx == get_Bound_lower(bound))
4656 else if (is_Bound(pred)) {
4658 * idx was Bounds checked in the same MacroBlock previously,
4659 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4661 ir_node *lower = get_Bound_lower(bound);
4662 ir_node *upper = get_Bound_upper(bound);
4663 if (get_Bound_lower(pred) == lower &&
4664 get_Bound_upper(pred) == upper &&
4665 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4667 * One could expect that we simply return the previous
4668 * Bound here. However, this would be wrong, as we could
4669 * add an exception Proj to a new location then.
4670 * So, we must turn in into a tuple.
4676 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4677 switch (get_Proj_proj(proj)) {
4679 DBG_OPT_EXC_REM(proj);
4680 proj = get_Bound_mem(bound);
4682 case pn_Bound_X_except:
4683 DBG_OPT_EXC_REM(proj);
4684 proj = get_irg_bad(current_ir_graph);
4688 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4690 case pn_Bound_X_regular:
4691 DBG_OPT_EXC_REM(proj);
4692 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4699 } /* transform_node_Proj_Bound */
4702 * Does all optimizations on nodes that must be done on it's Proj's
4703 * because of creating new nodes.
4705 static ir_node *transform_node_Proj(ir_node *proj) {
4706 ir_node *n = get_Proj_pred(proj);
4708 if (n->op->ops.transform_node_Proj)
4709 return n->op->ops.transform_node_Proj(proj);
4711 } /* transform_node_Proj */
4714 * Move Confirms down through Phi nodes.
4716 static ir_node *transform_node_Phi(ir_node *phi) {
4718 ir_mode *mode = get_irn_mode(phi);
4720 if (mode_is_reference(mode)) {
4721 n = get_irn_arity(phi);
4723 /* Beware of Phi0 */
4725 ir_node *pred = get_irn_n(phi, 0);
4726 ir_node *bound, *new_Phi, *block, **in;
4729 if (! is_Confirm(pred))
4732 bound = get_Confirm_bound(pred);
4733 pnc = get_Confirm_cmp(pred);
4735 NEW_ARR_A(ir_node *, in, n);
4736 in[0] = get_Confirm_value(pred);
4738 for (i = 1; i < n; ++i) {
4739 pred = get_irn_n(phi, i);
4741 if (! is_Confirm(pred) ||
4742 get_Confirm_bound(pred) != bound ||
4743 get_Confirm_cmp(pred) != pnc)
4745 in[i] = get_Confirm_value(pred);
4747 /* move the Confirm nodes "behind" the Phi */
4748 block = get_irn_n(phi, -1);
4749 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4750 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4754 } /* transform_node_Phi */
4757 * Returns the operands of a commutative bin-op, if one operand is
4758 * a const, it is returned as the second one.
4760 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4761 ir_node *op_a = get_binop_left(binop);
4762 ir_node *op_b = get_binop_right(binop);
4764 assert(is_op_commutative(get_irn_op(binop)));
4766 if (is_Const(op_a)) {
4773 } /* get_comm_Binop_Ops */
4776 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4777 * Such pattern may arise in bitfield stores.
4779 * value c4 value c4 & c2
4780 * AND c3 AND c1 | c3
4787 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4790 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4793 ir_node *and_l, *c3;
4794 ir_node *value, *c4;
4795 ir_node *new_and, *new_const, *block;
4796 ir_mode *mode = get_irn_mode(or);
4798 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4801 get_comm_Binop_Ops(or, &and, &c1);
4802 if (!is_Const(c1) || !is_And(and))
4805 get_comm_Binop_Ops(and, &or_l, &c2);
4809 tv1 = get_Const_tarval(c1);
4810 tv2 = get_Const_tarval(c2);
4812 tv = tarval_or(tv1, tv2);
4813 if (tarval_is_all_one(tv)) {
4814 /* the AND does NOT clear a bit with isn't set by the OR */
4815 set_Or_left(or, or_l);
4816 set_Or_right(or, c1);
4818 /* check for more */
4825 get_comm_Binop_Ops(or_l, &and_l, &c3);
4826 if (!is_Const(c3) || !is_And(and_l))
4829 get_comm_Binop_Ops(and_l, &value, &c4);
4833 /* ok, found the pattern, check for conditions */
4834 assert(mode == get_irn_mode(and));
4835 assert(mode == get_irn_mode(or_l));
4836 assert(mode == get_irn_mode(and_l));
4838 tv3 = get_Const_tarval(c3);
4839 tv4 = get_Const_tarval(c4);
4841 tv = tarval_or(tv4, tv2);
4842 if (!tarval_is_all_one(tv)) {
4843 /* have at least one 0 at the same bit position */
4847 n_tv4 = tarval_not(tv4);
4848 if (tv3 != tarval_and(tv3, n_tv4)) {
4849 /* bit in the or_mask is outside the and_mask */
4853 n_tv2 = tarval_not(tv2);
4854 if (tv1 != tarval_and(tv1, n_tv2)) {
4855 /* bit in the or_mask is outside the and_mask */
4859 /* ok, all conditions met */
4860 block = get_irn_n(or, -1);
4862 new_and = new_r_And(current_ir_graph, block,
4863 value, new_r_Const(current_ir_graph, mode, tarval_and(tv4, tv2)), mode);
4865 new_const = new_r_Const(current_ir_graph, mode, tarval_or(tv3, tv1));
4867 set_Or_left(or, new_and);
4868 set_Or_right(or, new_const);
4870 /* check for more */
4872 } /* transform_node_Or_bf_store */
4875 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4877 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4878 ir_mode *mode = get_irn_mode(or);
4879 ir_node *shl, *shr, *block;
4880 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4883 if (! mode_is_int(mode))
4886 shl = get_binop_left(or);
4887 shr = get_binop_right(or);
4896 } else if (!is_Shl(shl)) {
4898 } else if (!is_Shr(shr)) {
4901 x = get_Shl_left(shl);
4902 if (x != get_Shr_left(shr))
4905 c1 = get_Shl_right(shl);
4906 c2 = get_Shr_right(shr);
4907 if (is_Const(c1) && is_Const(c2)) {
4908 tv1 = get_Const_tarval(c1);
4909 if (! tarval_is_long(tv1))
4912 tv2 = get_Const_tarval(c2);
4913 if (! tarval_is_long(tv2))
4916 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4917 != (int) get_mode_size_bits(mode))
4920 /* yet, condition met */
4921 block = get_nodes_block(or);
4923 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4925 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4932 rotval = sub; /* a Rot right is not supported, so use a rot left */
4933 } else if (is_Sub(c2)) {
4939 if (get_Sub_right(sub) != v)
4942 c1 = get_Sub_left(sub);
4946 tv1 = get_Const_tarval(c1);
4947 if (! tarval_is_long(tv1))
4950 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4953 /* yet, condition met */
4954 block = get_nodes_block(or);
4956 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4958 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4960 } /* transform_node_Or_Rotl */
4965 static ir_node *transform_node_Or(ir_node *n) {
4966 ir_node *c, *oldn = n;
4967 ir_node *a = get_Or_left(n);
4968 ir_node *b = get_Or_right(n);
4971 if (is_Not(a) && is_Not(b)) {
4972 /* ~a | ~b = ~(a&b) */
4973 ir_node *block = get_nodes_block(n);
4975 mode = get_irn_mode(n);
4978 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4979 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4980 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4984 /* we can evaluate 2 Projs of the same Cmp */
4985 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4986 ir_node *pred_a = get_Proj_pred(a);
4987 ir_node *pred_b = get_Proj_pred(b);
4988 if (pred_a == pred_b) {
4989 dbg_info *dbgi = get_irn_dbg_info(n);
4990 ir_node *block = get_nodes_block(pred_a);
4991 pn_Cmp pn_a = get_Proj_proj(a);
4992 pn_Cmp pn_b = get_Proj_proj(b);
4993 /* yes, we can simply calculate with pncs */
4994 pn_Cmp new_pnc = pn_a | pn_b;
4996 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
5001 mode = get_irn_mode(n);
5002 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
5004 n = transform_node_Or_bf_store(n);
5005 n = transform_node_Or_Rotl(n);
5009 n = transform_bitwise_distributive(n, transform_node_Or);
5012 } /* transform_node_Or */
5016 static ir_node *transform_node(ir_node *n);
5019 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5021 * Should be moved to reassociation?
5023 static ir_node *transform_node_shift(ir_node *n) {
5024 ir_node *left, *right;
5026 tarval *tv1, *tv2, *res;
5027 ir_node *in[2], *irn, *block;
5029 left = get_binop_left(n);
5031 /* different operations */
5032 if (get_irn_op(left) != get_irn_op(n))
5035 right = get_binop_right(n);
5036 tv1 = value_of(right);
5037 if (tv1 == tarval_bad)
5040 tv2 = value_of(get_binop_right(left));
5041 if (tv2 == tarval_bad)
5044 res = tarval_add(tv1, tv2);
5045 mode = get_irn_mode(n);
5047 /* beware: a simple replacement works only, if res < modulo shift */
5049 int modulo_shf = get_mode_modulo_shift(mode);
5050 if (modulo_shf > 0) {
5051 tarval *modulo = new_tarval_from_long(modulo_shf,
5052 get_tarval_mode(res));
5054 assert(modulo_shf >= (int) get_mode_size_bits(mode));
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_mode *smode = get_irn_mode(right);
5063 ir_node *cnst = new_Const(smode, new_tarval_from_long(get_mode_size_bits(mode) - 1, smode));
5064 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5068 return new_Const(mode, get_mode_null(mode));
5072 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5075 /* ok, we can replace it */
5076 block = get_nodes_block(n);
5078 in[0] = get_binop_left(left);
5079 in[1] = new_r_Const(current_ir_graph, get_tarval_mode(res), res);
5081 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5083 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5085 return transform_node(irn);
5086 } /* transform_node_shift */
5089 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5091 * - and, or, xor instead of &
5092 * - Shl, Shr, Shrs, rotl instead of >>
5093 * (with a special case for Or/Xor + Shrs)
5095 static ir_node *transform_node_bitop_shift(ir_node *n) {
5097 ir_node *right = get_binop_right(n);
5098 ir_mode *mode = get_irn_mode(n);
5099 ir_node *bitop_left;
5100 ir_node *bitop_right;
5112 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5114 if (!is_Const(right))
5117 left = get_binop_left(n);
5118 op_left = get_irn_op(left);
5119 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5122 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5123 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5124 /* TODO: test if sign bit is affectes */
5128 bitop_right = get_binop_right(left);
5129 if (!is_Const(bitop_right))
5132 bitop_left = get_binop_left(left);
5134 irg = get_irn_irg(n);
5135 block = get_nodes_block(n);
5136 dbgi = get_irn_dbg_info(n);
5137 tv1 = get_Const_tarval(bitop_right);
5138 tv2 = get_Const_tarval(right);
5140 assert(get_tarval_mode(tv1) == mode);
5143 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5144 tv_shift = tarval_shl(tv1, tv2);
5145 } else if(is_Shr(n)) {
5146 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5147 tv_shift = tarval_shr(tv1, tv2);
5148 } else if(is_Shrs(n)) {
5149 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5150 tv_shift = tarval_shrs(tv1, tv2);
5153 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5154 tv_shift = tarval_rotl(tv1, tv2);
5157 assert(get_tarval_mode(tv_shift) == mode);
5158 new_const = new_Const(mode, tv_shift);
5160 if (op_left == op_And) {
5161 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5162 } else if(op_left == op_Or) {
5163 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5165 assert(op_left == op_Eor);
5166 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5174 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5176 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5177 * (also with x >>s c1 when c1>=c2)
5179 static ir_node *transform_node_shl_shr(ir_node *n) {
5181 ir_node *right = get_binop_right(n);
5197 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5199 if (!is_Const(right))
5202 left = get_binop_left(n);
5203 mode = get_irn_mode(n);
5204 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5205 ir_node *shr_right = get_binop_right(left);
5207 if (!is_Const(shr_right))
5210 x = get_binop_left(left);
5211 tv_shr = get_Const_tarval(shr_right);
5212 tv_shl = get_Const_tarval(right);
5214 if (is_Shrs(left)) {
5215 /* shrs variant only allowed if c1 >= c2 */
5216 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5219 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5222 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5224 tv_mask = tarval_shl(tv_mask, tv_shl);
5225 } else if(is_Shr(n) && is_Shl(left)) {
5226 ir_node *shl_right = get_Shl_right(left);
5228 if (!is_Const(shl_right))
5231 x = get_Shl_left(left);
5232 tv_shr = get_Const_tarval(right);
5233 tv_shl = get_Const_tarval(shl_right);
5235 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5236 tv_mask = tarval_shr(tv_mask, tv_shr);
5241 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5242 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5245 assert(tv_mask != tarval_bad);
5246 assert(get_tarval_mode(tv_mask) == mode);
5248 irg = get_irn_irg(n);
5249 block = get_nodes_block(n);
5250 dbgi = get_irn_dbg_info(n);
5252 pnc = tarval_cmp(tv_shl, tv_shr);
5253 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5254 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5255 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5257 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5259 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5262 assert(pnc == pn_Cmp_Gt);
5263 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5264 new_const = new_Const(get_tarval_mode(tv_shift), tv_shift);
5265 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5268 new_const = new_Const(mode, tv_mask);
5269 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5277 static ir_node *transform_node_Shr(ir_node *n) {
5278 ir_node *c, *oldn = n;
5279 ir_node *left = get_Shr_left(n);
5280 ir_node *right = get_Shr_right(n);
5281 ir_mode *mode = get_irn_mode(n);
5283 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5284 n = transform_node_shift(n);
5287 n = transform_node_shl_shr(n);
5289 n = transform_node_bitop_shift(n);
5292 } /* transform_node_Shr */
5297 static ir_node *transform_node_Shrs(ir_node *n) {
5298 ir_node *c, *oldn = n;
5299 ir_node *a = get_Shrs_left(n);
5300 ir_node *b = get_Shrs_right(n);
5301 ir_mode *mode = get_irn_mode(n);
5303 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5304 n = transform_node_shift(n);
5307 n = transform_node_bitop_shift(n);
5310 } /* transform_node_Shrs */
5315 static ir_node *transform_node_Shl(ir_node *n) {
5316 ir_node *c, *oldn = n;
5317 ir_node *a = get_Shl_left(n);
5318 ir_node *b = get_Shl_right(n);
5319 ir_mode *mode = get_irn_mode(n);
5321 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5322 n = transform_node_shift(n);
5325 n = transform_node_shl_shr(n);
5327 n = transform_node_bitop_shift(n);
5330 } /* transform_node_Shl */
5335 static ir_node *transform_node_Rotl(ir_node *n) {
5336 ir_node *c, *oldn = n;
5337 ir_node *a = get_Rotl_left(n);
5338 ir_node *b = get_Rotl_right(n);
5339 ir_mode *mode = get_irn_mode(n);
5341 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5342 n = transform_node_shift(n);
5345 n = transform_node_bitop_shift(n);
5348 } /* transform_node_Rotl */
5353 static ir_node *transform_node_Conv(ir_node *n) {
5354 ir_node *c, *oldn = n;
5355 ir_mode *mode = get_irn_mode(n);
5356 ir_node *a = get_Conv_op(n);
5358 if (mode != mode_b && is_const_Phi(a)) {
5359 /* Do NOT optimize mode_b Conv's, this leads to remaining
5360 * Phib nodes later, because the conv_b_lower operation
5361 * is instantly reverted, when it tries to insert a Convb.
5363 c = apply_conv_on_phi(a, mode);
5365 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5370 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5371 return new_r_Unknown(current_ir_graph, mode);
5374 if (mode_is_reference(mode) &&
5375 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5377 ir_node *l = get_Add_left(a);
5378 ir_node *r = get_Add_right(a);
5379 dbg_info *dbgi = get_irn_dbg_info(a);
5380 ir_node *block = get_nodes_block(n);
5382 ir_node *lop = get_Conv_op(l);
5383 if(get_irn_mode(lop) == mode) {
5384 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5385 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5390 ir_node *rop = get_Conv_op(r);
5391 if(get_irn_mode(rop) == mode) {
5392 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5393 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5400 } /* transform_node_Conv */
5403 * Remove dead blocks and nodes in dead blocks
5404 * in keep alive list. We do not generate a new End node.
5406 static ir_node *transform_node_End(ir_node *n) {
5407 int i, j, n_keepalives = get_End_n_keepalives(n);
5410 NEW_ARR_A(ir_node *, in, n_keepalives);
5412 for (i = j = 0; i < n_keepalives; ++i) {
5413 ir_node *ka = get_End_keepalive(n, i);
5415 if (! is_Block_dead(ka)) {
5419 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5421 } else if (is_Bad(ka)) {
5422 /* no need to keep Bad */
5427 if (j != n_keepalives)
5428 set_End_keepalives(n, j, in);
5430 } /* transform_node_End */
5432 /** returns 1 if a == -b */
5433 static int is_negated_value(ir_node *a, ir_node *b) {
5434 if (is_Minus(a) && get_Minus_op(a) == b)
5436 if (is_Minus(b) && get_Minus_op(b) == a)
5438 if (is_Sub(a) && is_Sub(b)) {
5439 ir_node *a_left = get_Sub_left(a);
5440 ir_node *a_right = get_Sub_right(a);
5441 ir_node *b_left = get_Sub_left(b);
5442 ir_node *b_right = get_Sub_right(b);
5444 if (a_left == b_right && a_right == b_left)
5452 * Optimize a Mux into some simpler cases.
5454 static ir_node *transform_node_Mux(ir_node *n) {
5455 ir_node *oldn = n, *sel = get_Mux_sel(n);
5456 ir_mode *mode = get_irn_mode(n);
5457 ir_node *t = get_Mux_true(n);
5458 ir_node *f = get_Mux_false(n);
5459 ir_graph *irg = current_ir_graph;
5461 /* first normalization step: move a possible zero to the false case */
5463 ir_node *cmp = get_Proj_pred(sel);
5466 if (is_Const(t) && is_Const_null(t)) {
5469 /* Mux(x, 0, y) => Mux(x, y, 0) */
5470 pn_Cmp pnc = get_Proj_proj(sel);
5471 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5472 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5473 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5481 /* note: after normalization, false can only happen on default */
5482 if (mode == mode_b) {
5483 dbg_info *dbg = get_irn_dbg_info(n);
5484 ir_node *block = get_nodes_block(n);
5485 ir_graph *irg = current_ir_graph;
5488 tarval *tv_t = get_Const_tarval(t);
5489 if (tv_t == tarval_b_true) {
5491 /* Muxb(sel, true, false) = sel */
5492 assert(get_Const_tarval(f) == tarval_b_false);
5493 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5496 /* Muxb(sel, true, x) = Or(sel, x) */
5497 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5498 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5502 } else if (is_Const(f)) {
5503 tarval *tv_f = get_Const_tarval(f);
5504 if (tv_f == tarval_b_true) {
5505 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5506 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5507 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5508 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5511 /* Muxb(sel, x, false) = And(sel, x) */
5512 assert(tv_f == tarval_b_false);
5513 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5514 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5520 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5521 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5522 tarval *a = get_Const_tarval(t);
5523 tarval *b = get_Const_tarval(f);
5524 tarval *null = get_tarval_null(mode);
5527 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5528 diff = tarval_sub(a, b, NULL);
5531 diff = tarval_sub(b, a, NULL);
5535 if (diff == get_tarval_one(mode) && min != null) {
5536 dbg_info *dbg = get_irn_dbg_info(n);
5537 ir_node *block = get_nodes_block(n);
5538 ir_graph *irg = current_ir_graph;
5539 ir_node *t = new_Const(mode, tarval_sub(a, min, NULL));
5540 ir_node *f = new_Const(mode, tarval_sub(b, min, NULL));
5541 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5542 n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode);
5548 ir_node *cmp = get_Proj_pred(sel);
5549 long pn = get_Proj_proj(sel);
5552 * Note: normalization puts the constant on the right side,
5553 * so we check only one case.
5555 * Note further that these optimization work even for floating point
5556 * with NaN's because -NaN == NaN.
5557 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5561 ir_node *cmp_r = get_Cmp_right(cmp);
5562 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5563 ir_node *block = get_nodes_block(n);
5564 ir_node *cmp_l = get_Cmp_left(cmp);
5566 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5569 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5570 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5572 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5573 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5575 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5577 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5578 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5580 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5581 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5583 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5585 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5590 if (mode_is_int(mode)) {
5592 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5593 /* Mux((a & b) != 0, c, 0) */
5594 ir_node *and_r = get_And_right(cmp_l);
5597 if (and_r == t && f == cmp_r) {
5598 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5599 if (pn == pn_Cmp_Lg) {
5600 /* Mux((a & 2^C) != 0, 2^C, 0) */
5602 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5604 /* Mux((a & 2^C) == 0, 2^C, 0) */
5605 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5606 block, cmp_l, t, mode);
5607 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5612 if (is_Shl(and_r)) {
5613 ir_node *shl_l = get_Shl_left(and_r);
5614 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5615 if (and_r == t && f == cmp_r) {
5616 if (pn == pn_Cmp_Lg) {
5617 /* (a & (1 << n)) != 0, (1 << n), 0) */
5619 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5621 /* (a & (1 << n)) == 0, (1 << n), 0) */
5622 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5623 block, cmp_l, t, mode);
5624 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5630 and_l = get_And_left(cmp_l);
5631 if (is_Shl(and_l)) {
5632 ir_node *shl_l = get_Shl_left(and_l);
5633 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5634 if (and_l == t && f == cmp_r) {
5635 if (pn == pn_Cmp_Lg) {
5636 /* ((1 << n) & a) != 0, (1 << n), 0) */
5638 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5640 /* ((1 << n) & a) == 0, (1 << n), 0) */
5641 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5642 block, cmp_l, t, mode);
5643 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5654 return arch_transform_node_Mux(n);
5655 } /* transform_node_Mux */
5658 * optimize Sync nodes that have other syncs as input we simply add the inputs
5659 * of the other sync to our own inputs
5661 static ir_node *transform_node_Sync(ir_node *n) {
5662 int arity = get_Sync_n_preds(n);
5665 for (i = 0; i < arity;) {
5666 ir_node *pred = get_Sync_pred(n, i);
5670 if (!is_Sync(pred)) {
5678 pred_arity = get_Sync_n_preds(pred);
5679 for (j = 0; j < pred_arity; ++j) {
5680 ir_node *pred_pred = get_Sync_pred(pred, j);
5685 add_irn_n(n, pred_pred);
5689 if (get_Sync_pred(n, k) == pred_pred) break;
5694 /* rehash the sync node */
5695 add_identities(current_ir_graph->value_table, n);
5701 * Tries several [inplace] [optimizing] transformations and returns an
5702 * equivalent node. The difference to equivalent_node() is that these
5703 * transformations _do_ generate new nodes, and thus the old node must
5704 * not be freed even if the equivalent node isn't the old one.
5706 static ir_node *transform_node(ir_node *n) {
5710 * Transform_node is the only "optimizing transformation" that might
5711 * return a node with a different opcode. We iterate HERE until fixpoint
5712 * to get the final result.
5716 if (n->op->ops.transform_node)
5717 n = n->op->ops.transform_node(n);
5718 } while (oldn != n);
5721 } /* transform_node */
5724 * Sets the default transform node operation for an ir_op_ops.
5726 * @param code the opcode for the default operation
5727 * @param ops the operations initialized
5732 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5736 ops->transform_node = transform_node_##a; \
5738 #define CASE_PROJ(a) \
5740 ops->transform_node_Proj = transform_node_Proj_##a; \
5742 #define CASE_PROJ_EX(a) \
5744 ops->transform_node = transform_node_##a; \
5745 ops->transform_node_Proj = transform_node_Proj_##a; \
5754 CASE_PROJ_EX(DivMod);
5788 } /* firm_set_default_transform_node */
5791 /* **************** Common Subexpression Elimination **************** */
5793 /** The size of the hash table used, should estimate the number of nodes
5795 #define N_IR_NODES 512
5797 /** Compares the attributes of two Const nodes. */
5798 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5799 return (get_Const_tarval(a) != get_Const_tarval(b))
5800 || (get_Const_type(a) != get_Const_type(b));
5801 } /* node_cmp_attr_Const */
5803 /** Compares the attributes of two Proj nodes. */
5804 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5805 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5806 } /* node_cmp_attr_Proj */
5808 /** Compares the attributes of two Filter nodes. */
5809 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5810 return get_Filter_proj(a) != get_Filter_proj(b);
5811 } /* node_cmp_attr_Filter */
5813 /** Compares the attributes of two Alloc nodes. */
5814 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5815 const alloc_attr *pa = get_irn_alloc_attr(a);
5816 const alloc_attr *pb = get_irn_alloc_attr(b);
5817 return (pa->where != pb->where) || (pa->type != pb->type);
5818 } /* node_cmp_attr_Alloc */
5820 /** Compares the attributes of two Free nodes. */
5821 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5822 const free_attr *pa = get_irn_free_attr(a);
5823 const free_attr *pb = get_irn_free_attr(b);
5824 return (pa->where != pb->where) || (pa->type != pb->type);
5825 } /* node_cmp_attr_Free */
5827 /** Compares the attributes of two SymConst nodes. */
5828 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5829 const symconst_attr *pa = get_irn_symconst_attr(a);
5830 const symconst_attr *pb = get_irn_symconst_attr(b);
5831 return (pa->kind != pb->kind)
5832 || (pa->sym.type_p != pb->sym.type_p)
5833 || (pa->tp != pb->tp);
5834 } /* node_cmp_attr_SymConst */
5836 /** Compares the attributes of two Call nodes. */
5837 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5838 return get_irn_call_attr(a) != get_irn_call_attr(b);
5839 } /* node_cmp_attr_Call */
5841 /** Compares the attributes of two Sel nodes. */
5842 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5843 const ir_entity *a_ent = get_Sel_entity(a);
5844 const ir_entity *b_ent = get_Sel_entity(b);
5847 (a_ent->kind != b_ent->kind) ||
5848 (a_ent->name != b_ent->name) ||
5849 (a_ent->owner != b_ent->owner) ||
5850 (a_ent->ld_name != b_ent->ld_name) ||
5851 (a_ent->type != b_ent->type);
5853 /* Matze: inlining of functions can produce 2 entities with same type,
5855 return a_ent != b_ent;
5856 } /* node_cmp_attr_Sel */
5858 /** Compares the attributes of two Phi nodes. */
5859 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5860 /* we can only enter this function if both nodes have the same number of inputs,
5861 hence it is enough to check if one of them is a Phi0 */
5863 /* check the Phi0 pos attribute */
5864 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5867 } /* node_cmp_attr_Phi */
5869 /** Compares the attributes of two Conv nodes. */
5870 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5871 return get_Conv_strict(a) != get_Conv_strict(b);
5872 } /* node_cmp_attr_Conv */
5874 /** Compares the attributes of two Cast nodes. */
5875 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5876 return get_Cast_type(a) != get_Cast_type(b);
5877 } /* node_cmp_attr_Cast */
5879 /** Compares the attributes of two Load nodes. */
5880 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5881 if (get_Load_volatility(a) == volatility_is_volatile ||
5882 get_Load_volatility(b) == volatility_is_volatile)
5883 /* NEVER do CSE on volatile Loads */
5885 /* do not CSE Loads with different alignment. Be conservative. */
5886 if (get_Load_align(a) != get_Load_align(b))
5889 return get_Load_mode(a) != get_Load_mode(b);
5890 } /* node_cmp_attr_Load */
5892 /** Compares the attributes of two Store nodes. */
5893 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5894 /* do not CSE Stores with different alignment. Be conservative. */
5895 if (get_Store_align(a) != get_Store_align(b))
5898 /* NEVER do CSE on volatile Stores */
5899 return (get_Store_volatility(a) == volatility_is_volatile ||
5900 get_Store_volatility(b) == volatility_is_volatile);
5901 } /* node_cmp_attr_Store */
5903 /** Compares two exception attributes */
5904 static int node_cmp_exception(ir_node *a, ir_node *b) {
5905 const except_attr *ea = get_irn_except_attr(a);
5906 const except_attr *eb = get_irn_except_attr(b);
5908 return ea->pin_state != eb->pin_state;
5911 #define node_cmp_attr_Bound node_cmp_exception
5913 /** Compares the attributes of two Div nodes. */
5914 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5915 const divmod_attr *ma = get_irn_divmod_attr(a);
5916 const divmod_attr *mb = get_irn_divmod_attr(b);
5917 return ma->exc.pin_state != mb->exc.pin_state ||
5918 ma->res_mode != mb->res_mode ||
5919 ma->no_remainder != mb->no_remainder;
5920 } /* node_cmp_attr_Div */
5922 /** Compares the attributes of two DivMod nodes. */
5923 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5924 const divmod_attr *ma = get_irn_divmod_attr(a);
5925 const divmod_attr *mb = get_irn_divmod_attr(b);
5926 return ma->exc.pin_state != mb->exc.pin_state ||
5927 ma->res_mode != mb->res_mode;
5928 } /* node_cmp_attr_DivMod */
5930 /** Compares the attributes of two Mod nodes. */
5931 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5932 const divmod_attr *ma = get_irn_divmod_attr(a);
5933 const divmod_attr *mb = get_irn_divmod_attr(b);
5934 return ma->exc.pin_state != mb->exc.pin_state ||
5935 ma->res_mode != mb->res_mode;
5936 } /* node_cmp_attr_Mod */
5938 /** Compares the attributes of two Quot nodes. */
5939 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5940 const divmod_attr *ma = get_irn_divmod_attr(a);
5941 const divmod_attr *mb = get_irn_divmod_attr(b);
5942 return ma->exc.pin_state != mb->exc.pin_state ||
5943 ma->res_mode != mb->res_mode;
5944 } /* node_cmp_attr_Quot */
5946 /** Compares the attributes of two Confirm nodes. */
5947 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5948 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5949 } /* node_cmp_attr_Confirm */
5951 /** Compares the attributes of two ASM nodes. */
5952 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5954 const ir_asm_constraint *ca;
5955 const ir_asm_constraint *cb;
5958 if (get_ASM_text(a) != get_ASM_text(b))
5961 /* Should we really check the constraints here? Should be better, but is strange. */
5962 n = get_ASM_n_input_constraints(a);
5963 if (n != get_ASM_n_input_constraints(b))
5966 ca = get_ASM_input_constraints(a);
5967 cb = get_ASM_input_constraints(b);
5968 for (i = 0; i < n; ++i) {
5969 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5973 n = get_ASM_n_output_constraints(a);
5974 if (n != get_ASM_n_output_constraints(b))
5977 ca = get_ASM_output_constraints(a);
5978 cb = get_ASM_output_constraints(b);
5979 for (i = 0; i < n; ++i) {
5980 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5984 n = get_ASM_n_clobbers(a);
5985 if (n != get_ASM_n_clobbers(b))
5988 cla = get_ASM_clobbers(a);
5989 clb = get_ASM_clobbers(b);
5990 for (i = 0; i < n; ++i) {
5991 if (cla[i] != clb[i])
5995 } /* node_cmp_attr_ASM */
5998 * Set the default node attribute compare operation for an ir_op_ops.
6000 * @param code the opcode for the default operation
6001 * @param ops the operations initialized
6006 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6010 ops->node_cmp_attr = node_cmp_attr_##a; \
6041 } /* firm_set_default_node_cmp_attr */
6044 * Compare function for two nodes in the value table. Gets two
6045 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6047 int identities_cmp(const void *elt, const void *key) {
6048 ir_node *a = (ir_node *)elt;
6049 ir_node *b = (ir_node *)key;
6052 if (a == b) return 0;
6054 if ((get_irn_op(a) != get_irn_op(b)) ||
6055 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6057 /* compare if a's in and b's in are of equal length */
6058 irn_arity_a = get_irn_intra_arity(a);
6059 if (irn_arity_a != get_irn_intra_arity(b))
6062 if (get_irn_pinned(a) == op_pin_state_pinned) {
6063 /* for pinned nodes, the block inputs must be equal */
6064 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6066 } else if (! get_opt_global_cse()) {
6067 /* for block-local CSE both nodes must be in the same MacroBlock */
6068 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6072 /* compare a->in[0..ins] with b->in[0..ins] */
6073 for (i = 0; i < irn_arity_a; i++)
6074 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6078 * here, we already now that the nodes are identical except their
6081 if (a->op->ops.node_cmp_attr)
6082 return a->op->ops.node_cmp_attr(a, b);
6085 } /* identities_cmp */
6088 * Calculate a hash value of a node.
6090 * @param node The IR-node
6092 unsigned ir_node_hash(const ir_node *node) {
6093 return node->op->ops.hash(node);
6094 } /* ir_node_hash */
6097 pset *new_identities(void) {
6098 return new_pset(identities_cmp, N_IR_NODES);
6099 } /* new_identities */
6101 void del_identities(pset *value_table) {
6102 del_pset(value_table);
6103 } /* del_identities */
6105 /* Normalize a node by putting constants (and operands with larger
6106 * node index) on the right (operator side). */
6107 void ir_normalize_node(ir_node *n) {
6108 if (is_op_commutative(get_irn_op(n))) {
6109 ir_node *l = get_binop_left(n);
6110 ir_node *r = get_binop_right(n);
6112 /* For commutative operators perform a OP b == b OP a but keep
6113 * constants on the RIGHT side. This helps greatly in some
6114 * optimizations. Moreover we use the idx number to make the form
6116 if (!operands_are_normalized(l, r)) {
6117 set_binop_left(n, r);
6118 set_binop_right(n, l);
6122 } /* ir_normalize_node */
6125 * Update the nodes after a match in the value table. If both nodes have
6126 * the same MacroBlock but different Blocks, we must ensure that the node
6127 * with the dominating Block (the node that is near to the MacroBlock header
6128 * is stored in the table.
6129 * Because a MacroBlock has only one "non-exception" flow, we don't need
6130 * dominance info here: We known, that one block must dominate the other and
6131 * following the only block input will allow to find it.
6133 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6134 ir_node *known_blk, *new_block, *block, *mbh;
6136 if (get_opt_global_cse()) {
6137 /* Block inputs are meaning less */
6140 known_blk = get_irn_n(known_irn, -1);
6141 new_block = get_irn_n(new_ir_node, -1);
6142 if (known_blk == new_block) {
6143 /* already in the same block */
6147 * We expect the typical case when we built the graph. In that case, the
6148 * known_irn is already the upper one, so checking this should be faster.
6151 mbh = get_Block_MacroBlock(new_block);
6153 if (block == known_blk) {
6154 /* ok, we have found it: known_block dominates new_block as expected */
6159 * We have reached the MacroBlock header NOT founding
6160 * the known_block. new_block must dominate known_block.
6163 set_irn_n(known_irn, -1, new_block);
6166 assert(get_Block_n_cfgpreds(block) == 1);
6167 block = get_Block_cfgpred_block(block, 0);
6169 } /* update_value_table */
6172 * Return the canonical node computing the same value as n.
6173 * Looks up the node in a hash table, enters it in the table
6174 * if it isn't there yet.
6176 * @param value_table the HashSet containing all nodes in the
6178 * @param n the node to look up
6180 * @return a node that computes the same value as n or n if no such
6181 * node could be found
6183 ir_node *identify_remember(pset *value_table, ir_node *n) {
6186 if (!value_table) return n;
6188 ir_normalize_node(n);
6189 /* lookup or insert in hash table with given hash key. */
6190 o = pset_insert(value_table, n, ir_node_hash(n));
6193 update_known_irn(o, n);
6197 } /* identify_remember */
6200 * During construction we set the op_pin_state_pinned flag in the graph right when the
6201 * optimization is performed. The flag turning on procedure global cse could
6202 * be changed between two allocations. This way we are safe.
6204 * @param value_table The value table
6205 * @param n The node to lookup
6207 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6210 n = identify_remember(value_table, n);
6211 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6212 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6214 } /* identify_cons */
6216 /* Add a node to the identities value table. */
6217 void add_identities(pset *value_table, ir_node *node) {
6218 if (get_opt_cse() && is_no_Block(node))
6219 identify_remember(value_table, node);
6220 } /* add_identities */
6222 /* Visit each node in the value table of a graph. */
6223 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6225 ir_graph *rem = current_ir_graph;
6227 current_ir_graph = irg;
6228 foreach_pset(irg->value_table, node)
6230 current_ir_graph = rem;
6231 } /* visit_all_identities */
6234 * Garbage in, garbage out. If a node has a dead input, i.e., the
6235 * Bad node is input to the node, return the Bad node.
6237 static ir_node *gigo(ir_node *node) {
6239 ir_op *op = get_irn_op(node);
6241 /* remove garbage blocks by looking at control flow that leaves the block
6242 and replacing the control flow by Bad. */
6243 if (get_irn_mode(node) == mode_X) {
6244 ir_node *block = get_nodes_block(skip_Proj(node));
6246 /* Don't optimize nodes in immature blocks. */
6247 if (!get_Block_matured(block))
6249 /* Don't optimize End, may have Bads. */
6250 if (op == op_End) return node;
6252 if (is_Block(block)) {
6253 if (is_Block_dead(block)) {
6254 /* control flow from dead block is dead */
6258 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6259 if (!is_Bad(get_irn_n(block, i)))
6263 ir_graph *irg = get_irn_irg(block);
6264 /* the start block is never dead */
6265 if (block != get_irg_start_block(irg)
6266 && block != get_irg_end_block(irg)) {
6268 * Do NOT kill control flow without setting
6269 * the block to dead of bad things can happen:
6270 * We get a Block that is not reachable be irg_block_walk()
6271 * but can be found by irg_walk()!
6273 set_Block_dead(block);
6280 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6281 blocks predecessors is dead. */
6282 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6283 irn_arity = get_irn_arity(node);
6286 * Beware: we can only read the block of a non-floating node.
6288 if (is_irn_pinned_in_irg(node) &&
6289 is_Block_dead(get_nodes_block(skip_Proj(node))))
6292 for (i = 0; i < irn_arity; i++) {
6293 ir_node *pred = get_irn_n(node, i);
6298 /* Propagating Unknowns here seems to be a bad idea, because
6299 sometimes we need a node as a input and did not want that
6301 However, it might be useful to move this into a later phase
6302 (if you think that optimizing such code is useful). */
6303 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6304 return new_Unknown(get_irn_mode(node));
6309 /* With this code we violate the agreement that local_optimize
6310 only leaves Bads in Block, Phi and Tuple nodes. */
6311 /* If Block has only Bads as predecessors it's garbage. */
6312 /* If Phi has only Bads as predecessors it's garbage. */
6313 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6314 irn_arity = get_irn_arity(node);
6315 for (i = 0; i < irn_arity; i++) {
6316 if (!is_Bad(get_irn_n(node, i))) break;
6318 if (i == irn_arity) node = new_Bad();
6325 * These optimizations deallocate nodes from the obstack.
6326 * It can only be called if it is guaranteed that no other nodes
6327 * reference this one, i.e., right after construction of a node.
6329 * @param n The node to optimize
6331 * current_ir_graph must be set to the graph of the node!
6333 ir_node *optimize_node(ir_node *n) {
6336 ir_opcode iro = get_irn_opcode(n);
6338 /* Always optimize Phi nodes: part of the construction. */
6339 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6341 /* constant expression evaluation / constant folding */
6342 if (get_opt_constant_folding()) {
6343 /* neither constants nor Tuple values can be evaluated */
6344 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6345 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6346 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6347 /* try to evaluate */
6348 tv = computed_value(n);
6349 if (tv != tarval_bad) {
6351 ir_type *old_tp = get_irn_type(n);
6352 int i, arity = get_irn_arity(n);
6356 * Try to recover the type of the new expression.
6358 for (i = 0; i < arity && !old_tp; ++i)
6359 old_tp = get_irn_type(get_irn_n(n, i));
6362 * we MUST copy the node here temporary, because it's still needed
6363 * for DBG_OPT_CSTEVAL
6365 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6366 oldn = alloca(node_size);
6368 memcpy(oldn, n, node_size);
6369 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6371 /* ARG, copy the in array, we need it for statistics */
6372 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6374 /* note the inplace edges module */
6375 edges_node_deleted(n, current_ir_graph);
6377 /* evaluation was successful -- replace the node. */
6378 irg_kill_node(current_ir_graph, n);
6379 nw = new_Const(get_tarval_mode(tv), tv);
6381 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6382 set_Const_type(nw, old_tp);
6383 DBG_OPT_CSTEVAL(oldn, nw);
6384 tarval_enable_fp_ops(old_fp_mode);
6387 tarval_enable_fp_ops(old_fp_mode);
6391 /* remove unnecessary nodes */
6392 if (get_opt_algebraic_simplification() ||
6393 (iro == iro_Phi) || /* always optimize these nodes. */
6395 (iro == iro_Proj) ||
6396 (iro == iro_Block) ) /* Flags tested local. */
6397 n = equivalent_node(n);
6399 /* Common Subexpression Elimination.
6401 * Checks whether n is already available.
6402 * The block input is used to distinguish different subexpressions. Right
6403 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6404 * subexpressions within a block.
6407 n = identify_cons(current_ir_graph->value_table, n);
6410 edges_node_deleted(oldn, current_ir_graph);
6412 /* We found an existing, better node, so we can deallocate the old node. */
6413 irg_kill_node(current_ir_graph, oldn);
6417 /* Some more constant expression evaluation that does not allow to
6419 iro = get_irn_opcode(n);
6420 if (get_opt_algebraic_simplification() ||
6421 (iro == iro_Cond) ||
6422 (iro == iro_Proj)) /* Flags tested local. */
6423 n = transform_node(n);
6425 /* Remove nodes with dead (Bad) input.
6426 Run always for transformation induced Bads. */
6429 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6430 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6432 n = identify_remember(current_ir_graph->value_table, o);
6438 } /* optimize_node */
6442 * These optimizations never deallocate nodes (in place). This can cause dead
6443 * nodes lying on the obstack. Remove these by a dead node elimination,
6444 * i.e., a copying garbage collection.
6446 ir_node *optimize_in_place_2(ir_node *n) {
6449 ir_opcode iro = get_irn_opcode(n);
6451 if (!get_opt_optimize() && !is_Phi(n)) return n;
6453 /* constant expression evaluation / constant folding */
6454 if (get_opt_constant_folding()) {
6455 /* neither constants nor Tuple values can be evaluated */
6456 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6457 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6458 int old_fp_mode = tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6459 /* try to evaluate */
6460 tv = computed_value(n);
6461 if (tv != tarval_bad) {
6462 /* evaluation was successful -- replace the node. */
6463 ir_type *old_tp = get_irn_type(n);
6464 int i, arity = get_irn_arity(n);
6467 * Try to recover the type of the new expression.
6469 for (i = 0; i < arity && !old_tp; ++i)
6470 old_tp = get_irn_type(get_irn_n(n, i));
6472 n = new_Const(get_tarval_mode(tv), tv);
6474 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6475 set_Const_type(n, old_tp);
6477 DBG_OPT_CSTEVAL(oldn, n);
6478 tarval_enable_fp_ops(old_fp_mode);
6481 tarval_enable_fp_ops(old_fp_mode);
6485 /* remove unnecessary nodes */
6486 if (get_opt_constant_folding() ||
6487 (iro == iro_Phi) || /* always optimize these nodes. */
6488 (iro == iro_Id) || /* ... */
6489 (iro == iro_Proj) || /* ... */
6490 (iro == iro_Block) ) /* Flags tested local. */
6491 n = equivalent_node(n);
6493 /** common subexpression elimination **/
6494 /* Checks whether n is already available. */
6495 /* The block input is used to distinguish different subexpressions. Right
6496 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6497 subexpressions within a block. */
6498 if (get_opt_cse()) {
6500 n = identify_remember(current_ir_graph->value_table, o);
6505 /* Some more constant expression evaluation. */
6506 iro = get_irn_opcode(n);
6507 if (get_opt_constant_folding() ||
6508 (iro == iro_Cond) ||
6509 (iro == iro_Proj)) /* Flags tested local. */
6510 n = transform_node(n);
6512 /* Remove nodes with dead (Bad) input.
6513 Run always for transformation induced Bads. */
6516 /* Now we can verify the node, as it has no dead inputs any more. */
6519 /* Now we have a legal, useful node. Enter it in hash table for cse.
6520 Blocks should be unique anyways. (Except the successor of start:
6521 is cse with the start block!) */
6522 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6524 n = identify_remember(current_ir_graph->value_table, o);
6530 } /* optimize_in_place_2 */
6533 * Wrapper for external use, set proper status bits after optimization.
6535 ir_node *optimize_in_place(ir_node *n) {
6536 /* Handle graph state */
6537 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6539 if (get_opt_global_cse())
6540 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6541 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6542 set_irg_outs_inconsistent(current_ir_graph);
6544 /* FIXME: Maybe we could also test whether optimizing the node can
6545 change the control graph. */
6546 set_irg_doms_inconsistent(current_ir_graph);
6547 return optimize_in_place_2(n);
6548 } /* optimize_in_place */
6551 * Calculate a hash value of a Const node.
6553 static unsigned hash_Const(const ir_node *node) {
6556 /* special value for const, as they only differ in their tarval. */
6557 h = HASH_PTR(node->attr.con.tv);
6558 h = 9*h + HASH_PTR(get_irn_mode(node));
6564 * Calculate a hash value of a SymConst node.
6566 static unsigned hash_SymConst(const ir_node *node) {
6569 /* special value for const, as they only differ in their symbol. */
6570 h = HASH_PTR(node->attr.symc.sym.type_p);
6571 h = 9*h + HASH_PTR(get_irn_mode(node));
6574 } /* hash_SymConst */
6577 * Set the default hash operation in an ir_op_ops.
6579 * @param code the opcode for the default operation
6580 * @param ops the operations initialized
6585 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6589 ops->hash = hash_##a; \
6592 /* hash function already set */
6593 if (ops->hash != NULL)
6600 /* use input/mode default hash if no function was given */
6601 ops->hash = firm_default_hash;
6609 * Sets the default operation for an ir_ops.
6611 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6612 ops = firm_set_default_hash(code, ops);
6613 ops = firm_set_default_computed_value(code, ops);
6614 ops = firm_set_default_equivalent_node(code, ops);
6615 ops = firm_set_default_transform_node(code, ops);
6616 ops = firm_set_default_node_cmp_attr(code, ops);
6617 ops = firm_set_default_get_type(code, ops);
6618 ops = firm_set_default_get_type_attr(code, ops);
6619 ops = firm_set_default_get_entity_attr(code, ops);
6622 } /* firm_set_default_operations */