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"
45 #include "opt_confirms.h"
46 #include "opt_polymorphy.h"
51 /* Make types visible to allow most efficient access */
55 * Returns the tarval of a Const node or tarval_bad for all other nodes.
57 static tarval *default_value_of(const ir_node *n) {
59 return get_Const_tarval(n); /* might return tarval_bad */
64 value_of_func value_of_ptr = default_value_of;
66 /* * Set a new value_of function. */
67 void set_value_of_func(value_of_func func) {
71 value_of_ptr = default_value_of;
75 * Return the value of a Constant.
77 static tarval *computed_value_Const(const ir_node *n) {
78 return get_Const_tarval(n);
79 } /* computed_value_Const */
82 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
84 static tarval *computed_value_SymConst(const ir_node *n) {
88 switch (get_SymConst_kind(n)) {
89 case symconst_type_size:
90 type = get_SymConst_type(n);
91 if (get_type_state(type) == layout_fixed)
92 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
94 case symconst_type_align:
95 type = get_SymConst_type(n);
96 if (get_type_state(type) == layout_fixed)
97 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
99 case symconst_ofs_ent:
100 ent = get_SymConst_entity(n);
101 type = get_entity_owner(ent);
102 if (get_type_state(type) == layout_fixed)
103 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
109 } /* computed_value_SymConst */
112 * Return the value of an Add.
114 static tarval *computed_value_Add(const ir_node *n) {
115 ir_node *a = get_Add_left(n);
116 ir_node *b = get_Add_right(n);
118 tarval *ta = value_of(a);
119 tarval *tb = value_of(b);
121 if ((ta != tarval_bad) && (tb != tarval_bad))
122 return tarval_add(ta, tb);
125 } /* computed_value_Add */
128 * Return the value of a Sub.
129 * Special case: a - a
131 static tarval *computed_value_Sub(const ir_node *n) {
132 ir_mode *mode = get_irn_mode(n);
133 ir_node *a = get_Sub_left(n);
134 ir_node *b = get_Sub_right(n);
139 if (! mode_is_float(mode)) {
142 return get_mode_null(mode);
148 if ((ta != tarval_bad) && (tb != tarval_bad))
149 return tarval_sub(ta, tb, mode);
152 } /* computed_value_Sub */
155 * Return the value of a Carry.
156 * Special : a op 0, 0 op b
158 static tarval *computed_value_Carry(const ir_node *n) {
159 ir_node *a = get_binop_left(n);
160 ir_node *b = get_binop_right(n);
161 ir_mode *m = get_irn_mode(n);
163 tarval *ta = value_of(a);
164 tarval *tb = value_of(b);
166 if ((ta != tarval_bad) && (tb != tarval_bad)) {
168 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
170 if (tarval_is_null(ta) || tarval_is_null(tb))
171 return get_mode_null(m);
174 } /* computed_value_Carry */
177 * Return the value of a Borrow.
180 static tarval *computed_value_Borrow(const ir_node *n) {
181 ir_node *a = get_binop_left(n);
182 ir_node *b = get_binop_right(n);
183 ir_mode *m = get_irn_mode(n);
185 tarval *ta = value_of(a);
186 tarval *tb = value_of(b);
188 if ((ta != tarval_bad) && (tb != tarval_bad)) {
189 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
190 } else if (tarval_is_null(ta)) {
191 return get_mode_null(m);
194 } /* computed_value_Borrow */
197 * Return the value of an unary Minus.
199 static tarval *computed_value_Minus(const ir_node *n) {
200 ir_node *a = get_Minus_op(n);
201 tarval *ta = value_of(a);
203 if (ta != tarval_bad)
204 return tarval_neg(ta);
207 } /* computed_value_Minus */
210 * Return the value of a Mul.
212 static tarval *computed_value_Mul(const ir_node *n) {
213 ir_node *a = get_Mul_left(n);
214 ir_node *b = get_Mul_right(n);
217 tarval *ta = value_of(a);
218 tarval *tb = value_of(b);
220 mode = get_irn_mode(n);
221 if (mode != get_irn_mode(a)) {
222 /* n * n = 2n bit multiplication */
223 ta = tarval_convert_to(ta, mode);
224 tb = tarval_convert_to(tb, mode);
227 if (ta != tarval_bad && tb != tarval_bad) {
228 return tarval_mul(ta, tb);
230 /* a * 0 != 0 if a == NaN or a == Inf */
231 if (!mode_is_float(mode)) {
232 /* a*0 = 0 or 0*b = 0 */
233 if (ta == get_mode_null(mode))
235 if (tb == get_mode_null(mode))
240 } /* computed_value_Mul */
243 * Return the value of an Abs.
245 static tarval *computed_value_Abs(const ir_node *n) {
246 ir_node *a = get_Abs_op(n);
247 tarval *ta = value_of(a);
249 if (ta != tarval_bad)
250 return tarval_abs(ta);
253 } /* computed_value_Abs */
256 * Return the value of an And.
257 * Special case: a & 0, 0 & b
259 static tarval *computed_value_And(const ir_node *n) {
260 ir_node *a = get_And_left(n);
261 ir_node *b = get_And_right(n);
263 tarval *ta = value_of(a);
264 tarval *tb = value_of(b);
266 if ((ta != tarval_bad) && (tb != tarval_bad)) {
267 return tarval_and (ta, tb);
269 if (tarval_is_null(ta)) return ta;
270 if (tarval_is_null(tb)) return tb;
273 } /* computed_value_And */
276 * Return the value of an Or.
277 * Special case: a | 1...1, 1...1 | b
279 static tarval *computed_value_Or(const ir_node *n) {
280 ir_node *a = get_Or_left(n);
281 ir_node *b = get_Or_right(n);
283 tarval *ta = value_of(a);
284 tarval *tb = value_of(b);
286 if ((ta != tarval_bad) && (tb != tarval_bad)) {
287 return tarval_or (ta, tb);
289 if (tarval_is_all_one(ta)) return ta;
290 if (tarval_is_all_one(tb)) return tb;
293 } /* computed_value_Or */
296 * Return the value of an Eor.
298 static tarval *computed_value_Eor(const ir_node *n) {
299 ir_node *a = get_Eor_left(n);
300 ir_node *b = get_Eor_right(n);
305 return get_mode_null(get_irn_mode(n));
310 if ((ta != tarval_bad) && (tb != tarval_bad)) {
311 return tarval_eor (ta, tb);
314 } /* computed_value_Eor */
317 * Return the value of a Not.
319 static tarval *computed_value_Not(const ir_node *n) {
320 ir_node *a = get_Not_op(n);
321 tarval *ta = value_of(a);
323 if (ta != tarval_bad)
324 return tarval_not(ta);
327 } /* computed_value_Not */
330 * Return the value of a Shl.
332 static tarval *computed_value_Shl(const ir_node *n) {
333 ir_node *a = get_Shl_left(n);
334 ir_node *b = get_Shl_right(n);
336 tarval *ta = value_of(a);
337 tarval *tb = value_of(b);
339 if ((ta != tarval_bad) && (tb != tarval_bad)) {
340 return tarval_shl (ta, tb);
343 } /* computed_value_Shl */
346 * Return the value of a Shr.
348 static tarval *computed_value_Shr(const ir_node *n) {
349 ir_node *a = get_Shr_left(n);
350 ir_node *b = get_Shr_right(n);
352 tarval *ta = value_of(a);
353 tarval *tb = value_of(b);
355 if ((ta != tarval_bad) && (tb != tarval_bad)) {
356 return tarval_shr (ta, tb);
359 } /* computed_value_Shr */
362 * Return the value of a Shrs.
364 static tarval *computed_value_Shrs(const ir_node *n) {
365 ir_node *a = get_Shrs_left(n);
366 ir_node *b = get_Shrs_right(n);
368 tarval *ta = value_of(a);
369 tarval *tb = value_of(b);
371 if ((ta != tarval_bad) && (tb != tarval_bad)) {
372 return tarval_shrs (ta, tb);
375 } /* computed_value_Shrs */
378 * Return the value of a Rotl.
380 static tarval *computed_value_Rotl(const ir_node *n) {
381 ir_node *a = get_Rotl_left(n);
382 ir_node *b = get_Rotl_right(n);
384 tarval *ta = value_of(a);
385 tarval *tb = value_of(b);
387 if ((ta != tarval_bad) && (tb != tarval_bad)) {
388 return tarval_rotl(ta, tb);
391 } /* computed_value_Rotl */
394 * Return the value of a Conv.
396 static tarval *computed_value_Conv(const ir_node *n) {
397 ir_node *a = get_Conv_op(n);
398 tarval *ta = value_of(a);
400 if (ta != tarval_bad)
401 return tarval_convert_to(ta, get_irn_mode(n));
404 } /* computed_value_Conv */
407 * Calculate the value of a Mux: can be evaluated, if the
408 * sel and the right input are known.
410 static tarval *computed_value_Mux(const ir_node *n) {
411 ir_node *sel = get_Mux_sel(n);
412 tarval *ts = value_of(sel);
414 if (ts == get_tarval_b_true()) {
415 ir_node *v = get_Mux_true(n);
418 else if (ts == get_tarval_b_false()) {
419 ir_node *v = get_Mux_false(n);
423 } /* computed_value_Mux */
426 * Calculate the value of a Confirm: can be evaluated,
427 * if it has the form Confirm(x, '=', Const).
429 static tarval *computed_value_Confirm(const ir_node *n) {
431 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
432 * Do NOT optimize them away (CondEval wants them), so wait until
433 * remove_confirm is activated.
435 if (get_opt_remove_confirm()) {
436 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
437 tarval *tv = value_of(get_Confirm_bound(n));
438 if (tv != tarval_bad)
442 return value_of(get_Confirm_value(n));
443 } /* computed_value_Confirm */
446 * Return the value of a Proj(Cmp).
448 * This performs a first step of unreachable code elimination.
449 * Proj can not be computed, but folding a Cmp above the Proj here is
450 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
452 * There are several case where we can evaluate a Cmp node, see later.
454 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
455 ir_node *a = get_Proj_pred(n);
456 ir_node *aa = get_Cmp_left(a);
457 ir_node *ab = get_Cmp_right(a);
458 long proj_nr = get_Proj_proj(n);
461 * BEWARE: a == a is NOT always True for floating Point values, as
462 * NaN != NaN is defined, so we must check this here.
465 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
468 /* This is a trick with the bits used for encoding the Cmp
469 Proj numbers, the following statement is not the same:
470 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
471 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
474 tarval *taa = value_of(aa);
475 tarval *tab = value_of(ab);
476 ir_mode *mode = get_irn_mode(aa);
479 * The predecessors of Cmp are target values. We can evaluate
482 if ((taa != tarval_bad) && (tab != tarval_bad)) {
483 /* strange checks... */
484 pn_Cmp flags = tarval_cmp(taa, tab);
485 if (flags != pn_Cmp_False) {
486 return new_tarval_from_long (proj_nr & flags, mode_b);
489 /* for integer values, we can check against MIN/MAX */
490 else if (mode_is_int(mode)) {
491 /* MIN <=/> x. This results in true/false. */
492 if (taa == get_mode_min(mode)) {
493 /* a compare with the MIN value */
494 if (proj_nr == pn_Cmp_Le)
495 return get_tarval_b_true();
496 else if (proj_nr == pn_Cmp_Gt)
497 return get_tarval_b_false();
499 /* x >=/< MIN. This results in true/false. */
501 if (tab == get_mode_min(mode)) {
502 /* a compare with the MIN value */
503 if (proj_nr == pn_Cmp_Ge)
504 return get_tarval_b_true();
505 else if (proj_nr == pn_Cmp_Lt)
506 return get_tarval_b_false();
508 /* MAX >=/< x. This results in true/false. */
509 else if (taa == get_mode_max(mode)) {
510 if (proj_nr == pn_Cmp_Ge)
511 return get_tarval_b_true();
512 else if (proj_nr == pn_Cmp_Lt)
513 return get_tarval_b_false();
515 /* x <=/> MAX. This results in true/false. */
516 else if (tab == get_mode_max(mode)) {
517 if (proj_nr == pn_Cmp_Le)
518 return get_tarval_b_true();
519 else if (proj_nr == pn_Cmp_Gt)
520 return get_tarval_b_false();
524 * The predecessors are Allocs or (void*)(0) constants. Allocs never
525 * return NULL, they raise an exception. Therefore we can predict
529 ir_node *aaa = skip_Proj(aa);
530 ir_node *aba = skip_Proj(ab);
532 if ( ( (/* aa is ProjP and aaa is Alloc */
534 && mode_is_reference(get_irn_mode(aa))
536 && ( (/* ab is NULL */
537 mode_is_reference(get_irn_mode(ab))
538 && tarval_is_null(tab))
539 || (/* ab is other Alloc */
541 && mode_is_reference(get_irn_mode(ab))
544 || (/* aa is NULL and aba is Alloc */
545 mode_is_reference(get_irn_mode(aa))
546 && tarval_is_null(taa)
548 && mode_is_reference(get_irn_mode(ab))
551 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
554 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
555 } /* computed_value_Proj_Cmp */
558 * Return the value of a floating point Quot.
560 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
561 tarval *ta = value_of(a);
562 tarval *tb = value_of(b);
564 /* cannot optimize 0 / b = 0 because of NaN */
565 if (ta != tarval_bad && tb != tarval_bad)
566 return tarval_quo(ta, tb);
568 } /* do_computed_value_Quot */
571 * Calculate the value of an integer Div of two nodes.
572 * Special case: 0 / b
574 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
575 tarval *ta = value_of(a);
577 const ir_node *dummy;
579 /* Compute c1 / c2 or 0 / a, a != 0 */
580 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
581 return ta; /* 0 / b == 0 */
583 if (ta != tarval_bad && tb != tarval_bad)
584 return tarval_div(ta, tb);
586 } /* do_computed_value_Div */
589 * Calculate the value of an integer Mod of two nodes.
590 * Special case: a % 1
592 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
593 tarval *ta = value_of(a);
594 tarval *tb = value_of(b);
596 /* Compute a % 1 or c1 % c2 */
597 if (tarval_is_one(tb))
598 return get_mode_null(get_irn_mode(a));
599 if (ta != tarval_bad && tb != tarval_bad)
600 return tarval_mod(ta, tb);
602 } /* do_computed_value_Mod */
605 * Return the value of a Proj(DivMod).
607 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
608 long proj_nr = get_Proj_proj(n);
610 /* compute either the Div or the Mod part */
611 if (proj_nr == pn_DivMod_res_div) {
612 const ir_node *a = get_Proj_pred(n);
613 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
614 } else if (proj_nr == pn_DivMod_res_mod) {
615 const ir_node *a = get_Proj_pred(n);
616 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
619 } /* computed_value_Proj_DivMod */
622 * Return the value of a Proj(Div).
624 static tarval *computed_value_Proj_Div(const ir_node *n) {
625 long proj_nr = get_Proj_proj(n);
627 if (proj_nr == pn_Div_res) {
628 const ir_node *a = get_Proj_pred(n);
629 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
632 } /* computed_value_Proj_Div */
635 * Return the value of a Proj(Mod).
637 static tarval *computed_value_Proj_Mod(const ir_node *n) {
638 long proj_nr = get_Proj_proj(n);
640 if (proj_nr == pn_Mod_res) {
641 const ir_node *a = get_Proj_pred(n);
642 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
645 } /* computed_value_Proj_Mod */
648 * Return the value of a Proj(Quot).
650 static tarval *computed_value_Proj_Quot(const ir_node *n) {
651 long proj_nr = get_Proj_proj(n);
653 if (proj_nr == pn_Quot_res) {
654 const ir_node *a = get_Proj_pred(n);
655 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
658 } /* computed_value_Proj_Quot */
661 * Return the value of a Proj.
663 static tarval *computed_value_Proj(const ir_node *proj) {
664 ir_node *n = get_Proj_pred(proj);
666 if (n->op->ops.computed_value_Proj != NULL)
667 return n->op->ops.computed_value_Proj(proj);
669 } /* computed_value_Proj */
672 * If the parameter n can be computed, return its value, else tarval_bad.
673 * Performs constant folding.
675 * @param n The node this should be evaluated
677 tarval *computed_value(const ir_node *n) {
678 if (n->op->ops.computed_value)
679 return n->op->ops.computed_value(n);
681 } /* computed_value */
684 * Set the default computed_value evaluator in an ir_op_ops.
686 * @param code the opcode for the default operation
687 * @param ops the operations initialized
692 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
696 ops->computed_value = computed_value_##a; \
698 #define CASE_PROJ(a) \
700 ops->computed_value_Proj = computed_value_Proj_##a; \
737 } /* firm_set_default_computed_value */
740 * Returns a equivalent block for another block.
741 * If the block has only one predecessor, this is
742 * the equivalent one. If the only predecessor of a block is
743 * the block itself, this is a dead block.
745 * If both predecessors of a block are the branches of a binary
746 * Cond, the equivalent block is Cond's block.
748 * If all predecessors of a block are bad or lies in a dead
749 * block, the current block is dead as well.
751 * Note, that blocks are NEVER turned into Bad's, instead
752 * the dead_block flag is set. So, never test for is_Bad(block),
753 * always use is_dead_Block(block).
755 static ir_node *equivalent_node_Block(ir_node *n)
760 /* don't optimize dead or labeled blocks */
761 if (is_Block_dead(n) || has_Block_label(n))
764 n_preds = get_Block_n_cfgpreds(n);
766 /* The Block constructor does not call optimize, but mature_immBlock()
767 calls the optimization. */
768 assert(get_Block_matured(n));
770 /* Straightening: a single entry Block following a single exit Block
771 can be merged, if it is not the Start block. */
772 /* !!! Beware, all Phi-nodes of n must have been optimized away.
773 This should be true, as the block is matured before optimize is called.
774 But what about Phi-cycles with the Phi0/Id that could not be resolved?
775 Remaining Phi nodes are just Ids. */
777 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
780 ir_node *predblock = get_nodes_block(pred);
781 if (predblock == oldn) {
782 /* Jmp jumps into the block it is in -- deal self cycle. */
783 n = set_Block_dead(n);
784 DBG_OPT_DEAD_BLOCK(oldn, n);
785 } else if (get_opt_control_flow_straightening()) {
787 DBG_OPT_STG(oldn, n);
789 } else if (is_Cond(pred)) {
790 ir_node *predblock = get_nodes_block(pred);
791 if (predblock == oldn) {
792 /* Jmp jumps into the block it is in -- deal self cycle. */
793 n = set_Block_dead(n);
794 DBG_OPT_DEAD_BLOCK(oldn, n);
797 } else if ((n_preds == 2) &&
798 (get_opt_control_flow_weak_simplification())) {
799 /* Test whether Cond jumps twice to this block
800 * The more general case which more than 2 predecessors is handles
801 * in optimize_cf(), we handle only this special case for speed here.
803 ir_node *a = get_Block_cfgpred(n, 0);
804 ir_node *b = get_Block_cfgpred(n, 1);
806 if (is_Proj(a) && is_Proj(b)) {
807 ir_node *cond = get_Proj_pred(a);
809 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
810 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
811 /* Also a single entry Block following a single exit Block. Phis have
812 twice the same operand and will be optimized away. */
813 n = get_nodes_block(cond);
814 DBG_OPT_IFSIM1(oldn, a, b, n);
817 } else if (get_opt_unreachable_code() &&
818 (n != get_irg_start_block(current_ir_graph)) &&
819 (n != get_irg_end_block(current_ir_graph)) ) {
822 /* If all inputs are dead, this block is dead too, except if it is
823 the start or end block. This is one step of unreachable code
825 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
826 ir_node *pred = get_Block_cfgpred(n, i);
829 if (is_Bad(pred)) continue;
830 pred_blk = get_nodes_block(skip_Proj(pred));
832 if (is_Block_dead(pred_blk)) continue;
835 /* really found a living input */
840 n = set_Block_dead(n);
841 DBG_OPT_DEAD_BLOCK(oldn, n);
846 } /* equivalent_node_Block */
849 * Returns a equivalent node for a Jmp, a Bad :-)
850 * Of course this only happens if the Block of the Jmp is dead.
852 static ir_node *equivalent_node_Jmp(ir_node *n) {
855 /* unreachable code elimination */
856 if (is_Block_dead(get_nodes_block(n))) {
857 n = get_irg_bad(current_ir_graph);
858 DBG_OPT_DEAD_BLOCK(oldn, n);
861 } /* equivalent_node_Jmp */
863 /** Raise is handled in the same way as Jmp. */
864 #define equivalent_node_Raise equivalent_node_Jmp
867 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
868 See transform_node_Proj_Cond(). */
871 * Optimize operations that are commutative and have neutral 0,
872 * so a op 0 = 0 op a = a.
874 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
877 ir_node *a = get_binop_left(n);
878 ir_node *b = get_binop_right(n);
883 /* After running compute_node there is only one constant predecessor.
884 Find this predecessors value and remember the other node: */
885 if ((tv = value_of(a)) != tarval_bad) {
887 } else if ((tv = value_of(b)) != tarval_bad) {
892 /* If this predecessors constant value is zero, the operation is
893 * unnecessary. Remove it.
895 * Beware: If n is a Add, the mode of on and n might be different
896 * which happens in this rare construction: NULL + 3.
897 * Then, a Conv would be needed which we cannot include here.
899 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
902 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
906 } /* equivalent_node_neutral_zero */
909 * Eor is commutative and has neutral 0.
911 static ir_node *equivalent_node_Eor(ir_node *n) {
916 n = equivalent_node_neutral_zero(n);
917 if (n != oldn) return n;
920 b = get_Eor_right(n);
923 ir_node *aa = get_Eor_left(a);
924 ir_node *ab = get_Eor_right(a);
927 /* (a ^ b) ^ a -> b */
929 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
931 } else if (ab == b) {
932 /* (a ^ b) ^ b -> a */
934 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
939 ir_node *ba = get_Eor_left(b);
940 ir_node *bb = get_Eor_right(b);
943 /* a ^ (a ^ b) -> b */
945 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
947 } else if (bb == a) {
948 /* a ^ (b ^ a) -> b */
950 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
958 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
960 * The second one looks strange, but this construct
961 * is used heavily in the LCC sources :-).
963 * Beware: The Mode of an Add may be different than the mode of its
964 * predecessors, so we could not return a predecessors in all cases.
966 static ir_node *equivalent_node_Add(ir_node *n) {
968 ir_node *left, *right;
969 ir_mode *mode = get_irn_mode(n);
971 n = equivalent_node_neutral_zero(n);
975 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
976 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
979 left = get_Add_left(n);
980 right = get_Add_right(n);
983 if (get_Sub_right(left) == right) {
986 n = get_Sub_left(left);
987 if (mode == get_irn_mode(n)) {
988 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
994 if (get_Sub_right(right) == left) {
997 n = get_Sub_left(right);
998 if (mode == get_irn_mode(n)) {
999 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1005 } /* equivalent_node_Add */
1008 * optimize operations that are not commutative but have neutral 0 on left,
1011 static ir_node *equivalent_node_left_zero(ir_node *n) {
1014 ir_node *a = get_binop_left(n);
1015 ir_node *b = get_binop_right(n);
1016 tarval *tb = value_of(b);
1018 if (tarval_is_null(tb)) {
1021 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1024 } /* equivalent_node_left_zero */
1026 #define equivalent_node_Shl equivalent_node_left_zero
1027 #define equivalent_node_Shr equivalent_node_left_zero
1028 #define equivalent_node_Shrs equivalent_node_left_zero
1029 #define equivalent_node_Rotl equivalent_node_left_zero
1032 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1034 * The second one looks strange, but this construct
1035 * is used heavily in the LCC sources :-).
1037 * Beware: The Mode of a Sub may be different than the mode of its
1038 * predecessors, so we could not return a predecessors in all cases.
1040 static ir_node *equivalent_node_Sub(ir_node *n) {
1043 ir_mode *mode = get_irn_mode(n);
1046 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1047 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1050 b = get_Sub_right(n);
1053 /* Beware: modes might be different */
1054 if (tarval_is_null(tb)) {
1055 ir_node *a = get_Sub_left(n);
1056 if (mode == get_irn_mode(a)) {
1059 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1063 } /* equivalent_node_Sub */
1067 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1070 * -(-a) == a, but might overflow two times.
1071 * We handle it anyway here but the better way would be a
1072 * flag. This would be needed for Pascal for instance.
1074 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1076 ir_node *pred = get_unop_op(n);
1078 /* optimize symmetric unop */
1079 if (get_irn_op(pred) == get_irn_op(n)) {
1080 n = get_unop_op(pred);
1081 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1084 } /* equivalent_node_idempotent_unop */
1086 /** Optimize Not(Not(x)) == x. */
1087 #define equivalent_node_Not equivalent_node_idempotent_unop
1089 /** -(-x) == x ??? Is this possible or can --x raise an
1090 out of bounds exception if min =! max? */
1091 #define equivalent_node_Minus equivalent_node_idempotent_unop
1094 * Optimize a * 1 = 1 * a = a.
1096 static ir_node *equivalent_node_Mul(ir_node *n) {
1098 ir_node *a = get_Mul_left(n);
1100 /* we can handle here only the n * n = n bit cases */
1101 if (get_irn_mode(n) == get_irn_mode(a)) {
1102 ir_node *b = get_Mul_right(n);
1106 * Mul is commutative and has again an other neutral element.
1107 * Constants are place right, so check this case first.
1110 if (tarval_is_one(tv)) {
1112 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1115 if (tarval_is_one(tv)) {
1117 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1122 } /* equivalent_node_Mul */
1125 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1127 static ir_node *equivalent_node_Or(ir_node *n) {
1130 ir_node *a = get_Or_left(n);
1131 ir_node *b = get_Or_right(n);
1135 n = a; /* Or has it's own neutral element */
1136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1139 /* constants are cormalized to right, check this site first */
1141 if (tarval_is_null(tv)) {
1143 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1147 if (tarval_is_null(tv)) {
1149 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1154 } /* equivalent_node_Or */
1157 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1159 static ir_node *equivalent_node_And(ir_node *n) {
1162 ir_node *a = get_And_left(n);
1163 ir_node *b = get_And_right(n);
1167 n = a; /* And has it's own neutral element */
1168 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1171 /* constants are normalized to right, check this site first */
1173 if (tarval_is_all_one(tv)) {
1175 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1178 if (tv != get_tarval_bad()) {
1179 ir_mode *mode = get_irn_mode(n);
1180 if (!mode_is_signed(mode) && is_Conv(a)) {
1181 ir_node *convop = get_Conv_op(a);
1182 ir_mode *convopmode = get_irn_mode(convop);
1183 if (!mode_is_signed(convopmode)) {
1184 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1185 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1187 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1194 if (tarval_is_all_one(tv)) {
1196 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1200 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1203 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1208 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1211 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1216 } /* equivalent_node_And */
1219 * Try to remove useless Conv's:
1221 static ir_node *equivalent_node_Conv(ir_node *n) {
1223 ir_node *a = get_Conv_op(n);
1225 ir_mode *n_mode = get_irn_mode(n);
1226 ir_mode *a_mode = get_irn_mode(a);
1229 if (n_mode == a_mode) { /* No Conv necessary */
1230 if (get_Conv_strict(n)) {
1233 /* neither Minus nor Abs nor Confirm change the precision,
1234 so we can "look-through" */
1237 p = get_Minus_op(p);
1238 } else if (is_Abs(p)) {
1240 } else if (is_Confirm(p)) {
1241 p = get_Confirm_value(p);
1247 if (is_Conv(p) && get_Conv_strict(p)) {
1248 /* we known already, that a_mode == n_mode, and neither
1249 Abs nor Minus change the mode, so the second Conv
1251 assert(get_irn_mode(p) == n_mode);
1253 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1257 ir_node *pred = get_Proj_pred(p);
1258 if (is_Load(pred)) {
1259 /* Loads always return with the exact precision of n_mode */
1260 assert(get_Load_mode(pred) == n_mode);
1262 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1265 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1266 pred = get_Proj_pred(pred);
1267 if (is_Start(pred)) {
1268 /* Arguments always return with the exact precision,
1269 as strictConv's are place before Call -- if the
1270 caller was compiled with the same setting.
1271 Otherwise, the semantics is probably still right. */
1272 assert(get_irn_mode(p) == n_mode);
1274 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1280 /* special case: the immediate predecessor is also a Conv */
1281 if (! get_Conv_strict(a)) {
1282 /* first one is not strict, kick it */
1284 a_mode = get_irn_mode(a);
1288 /* else both are strict conv, second is superfluous */
1290 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1295 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1298 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1299 ir_node *b = get_Conv_op(a);
1300 ir_mode *b_mode = get_irn_mode(b);
1302 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1303 /* both are strict conv */
1304 if (smaller_mode(a_mode, n_mode)) {
1305 /* both are strict, but the first is smaller, so
1306 the second cannot remove more precision, remove the
1308 set_Conv_strict(n, 0);
1311 if (n_mode == b_mode) {
1312 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1313 if (n_mode == mode_b) {
1314 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1315 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1317 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1318 if (values_in_mode(b_mode, a_mode)) {
1319 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1320 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1325 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1326 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1327 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1328 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1330 if (float_mantissa >= int_mantissa) {
1332 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1337 if (smaller_mode(b_mode, a_mode)) {
1338 if (get_Conv_strict(n))
1339 set_Conv_strict(b, 1);
1340 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1341 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1348 } /* equivalent_node_Conv */
1351 * A Cast may be removed if the type of the previous node
1352 * is already the type of the Cast.
1354 static ir_node *equivalent_node_Cast(ir_node *n) {
1356 ir_node *pred = get_Cast_op(n);
1358 if (get_irn_type(pred) == get_Cast_type(n)) {
1360 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1363 } /* equivalent_node_Cast */
1366 * - fold Phi-nodes, iff they have only one predecessor except
1369 static ir_node *equivalent_node_Phi(ir_node *n) {
1374 ir_node *first_val = NULL; /* to shutup gcc */
1376 if (!get_opt_normalize()) return n;
1378 n_preds = get_Phi_n_preds(n);
1380 block = get_nodes_block(n);
1381 if (is_Block_dead(block)) /* Control dead */
1382 return get_irg_bad(current_ir_graph);
1384 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1386 /* Find first non-self-referencing input */
1387 for (i = 0; i < n_preds; ++i) {
1388 first_val = get_Phi_pred(n, i);
1389 if ( (first_val != n) /* not self pointer */
1391 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1392 * predecessors. Then, Phi nodes in dead code might be removed, causing
1393 * nodes pointing to themself (Add's for instance).
1394 * This is really bad and causes endless recursions in several
1395 * code pathes, so we do NOT optimize such a code.
1396 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1397 * (and bad Phi predecessors), so live code is optimized later.
1399 && (! is_Bad(get_Block_cfgpred(block, i)))
1401 ) { /* value not dead */
1402 break; /* then found first value. */
1407 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1408 return get_irg_bad(current_ir_graph);
1411 /* search for rest of inputs, determine if any of these
1412 are non-self-referencing */
1413 while (++i < n_preds) {
1414 ir_node *scnd_val = get_Phi_pred(n, i);
1415 if ( (scnd_val != n)
1416 && (scnd_val != first_val)
1419 && (! is_Bad(get_Block_cfgpred(block, i)))
1427 /* Fold, if no multiple distinct non-self-referencing inputs */
1429 DBG_OPT_PHI(oldn, n);
1432 } /* equivalent_node_Phi */
1435 * Several optimizations:
1436 * - fold Sync-nodes, iff they have only one predecessor except
1439 static ir_node *equivalent_node_Sync(ir_node *n) {
1440 int arity = get_Sync_n_preds(n);
1443 for (i = 0; i < arity;) {
1444 ir_node *pred = get_Sync_pred(n, i);
1447 /* Remove Bad predecessors */
1454 /* Remove duplicate predecessors */
1460 if (get_Sync_pred(n, j) == pred) {
1468 if (arity == 0) return get_irg_bad(current_ir_graph);
1469 if (arity == 1) return get_Sync_pred(n, 0);
1471 } /* equivalent_node_Sync */
1474 * Optimize Proj(Tuple).
1476 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1477 ir_node *oldn = proj;
1478 ir_node *tuple = get_Proj_pred(proj);
1480 /* Remove the Tuple/Proj combination. */
1481 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1482 DBG_OPT_TUPLE(oldn, tuple, proj);
1485 } /* equivalent_node_Proj_Tuple */
1488 * Optimize a / 1 = a.
1490 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1491 ir_node *oldn = proj;
1492 ir_node *div = get_Proj_pred(proj);
1493 ir_node *b = get_Div_right(div);
1494 tarval *tb = value_of(b);
1496 /* Div is not commutative. */
1497 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1498 switch (get_Proj_proj(proj)) {
1500 proj = get_Div_mem(div);
1501 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1505 proj = get_Div_left(div);
1506 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1510 /* we cannot replace the exception Proj's here, this is done in
1511 transform_node_Proj_Div() */
1516 } /* equivalent_node_Proj_Div */
1519 * Optimize a / 1.0 = a.
1521 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1522 ir_node *oldn = proj;
1523 ir_node *quot = get_Proj_pred(proj);
1524 ir_node *b = get_Quot_right(quot);
1525 tarval *tb = value_of(b);
1527 /* Div is not commutative. */
1528 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1529 switch (get_Proj_proj(proj)) {
1531 proj = get_Quot_mem(quot);
1532 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1536 proj = get_Quot_left(quot);
1537 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1541 /* we cannot replace the exception Proj's here, this is done in
1542 transform_node_Proj_Quot() */
1547 } /* equivalent_node_Proj_Quot */
1550 * Optimize a / 1 = a.
1552 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1553 ir_node *oldn = proj;
1554 ir_node *divmod = get_Proj_pred(proj);
1555 ir_node *b = get_DivMod_right(divmod);
1556 tarval *tb = value_of(b);
1558 /* Div is not commutative. */
1559 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1560 switch (get_Proj_proj(proj)) {
1562 proj = get_DivMod_mem(divmod);
1563 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1566 case pn_DivMod_res_div:
1567 proj = get_DivMod_left(divmod);
1568 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1572 /* we cannot replace the exception Proj's here, this is done in
1573 transform_node_Proj_DivMod().
1574 Note further that the pn_DivMod_res_div case is handled in
1575 computed_value_Proj(). */
1580 } /* equivalent_node_Proj_DivMod */
1583 * Optimize CopyB(mem, x, x) into a Nop.
1585 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1586 ir_node *oldn = proj;
1587 ir_node *copyb = get_Proj_pred(proj);
1588 ir_node *a = get_CopyB_dst(copyb);
1589 ir_node *b = get_CopyB_src(copyb);
1592 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1593 switch (get_Proj_proj(proj)) {
1594 case pn_CopyB_M_regular:
1595 proj = get_CopyB_mem(copyb);
1596 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1599 case pn_CopyB_M_except:
1600 case pn_CopyB_X_except:
1601 DBG_OPT_EXC_REM(proj);
1602 proj = get_irg_bad(current_ir_graph);
1607 } /* equivalent_node_Proj_CopyB */
1610 * Optimize Bounds(idx, idx, upper) into idx.
1612 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1613 ir_node *oldn = proj;
1614 ir_node *bound = get_Proj_pred(proj);
1615 ir_node *idx = get_Bound_index(bound);
1616 ir_node *pred = skip_Proj(idx);
1619 if (idx == get_Bound_lower(bound))
1621 else if (is_Bound(pred)) {
1623 * idx was Bounds checked in the same MacroBlock previously,
1624 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1626 ir_node *lower = get_Bound_lower(bound);
1627 ir_node *upper = get_Bound_upper(bound);
1628 if (get_Bound_lower(pred) == lower &&
1629 get_Bound_upper(pred) == upper &&
1630 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1632 * One could expect that we simply return the previous
1633 * Bound here. However, this would be wrong, as we could
1634 * add an exception Proj to a new location then.
1635 * So, we must turn in into a tuple.
1641 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1642 switch (get_Proj_proj(proj)) {
1644 DBG_OPT_EXC_REM(proj);
1645 proj = get_Bound_mem(bound);
1647 case pn_Bound_X_except:
1648 DBG_OPT_EXC_REM(proj);
1649 proj = get_irg_bad(current_ir_graph);
1653 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1656 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1661 } /* equivalent_node_Proj_Bound */
1664 * Optimize an Exception Proj(Load) with a non-null address.
1666 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1667 if (get_opt_ldst_only_null_ptr_exceptions()) {
1668 if (get_irn_mode(proj) == mode_X) {
1669 ir_node *load = get_Proj_pred(proj);
1671 /* get the Load address */
1672 const ir_node *addr = get_Load_ptr(load);
1673 const ir_node *confirm;
1675 if (value_not_null(addr, &confirm)) {
1676 if (get_Proj_proj(proj) == pn_Load_X_except) {
1677 DBG_OPT_EXC_REM(proj);
1678 return get_irg_bad(current_ir_graph);
1684 } /* equivalent_node_Proj_Load */
1687 * Optimize an Exception Proj(Store) with a non-null address.
1689 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1690 if (get_opt_ldst_only_null_ptr_exceptions()) {
1691 if (get_irn_mode(proj) == mode_X) {
1692 ir_node *store = get_Proj_pred(proj);
1694 /* get the load/store address */
1695 const ir_node *addr = get_Store_ptr(store);
1696 const ir_node *confirm;
1698 if (value_not_null(addr, &confirm)) {
1699 if (get_Proj_proj(proj) == pn_Store_X_except) {
1700 DBG_OPT_EXC_REM(proj);
1701 return get_irg_bad(current_ir_graph);
1707 } /* equivalent_node_Proj_Store */
1710 * Does all optimizations on nodes that must be done on it's Proj's
1711 * because of creating new nodes.
1713 static ir_node *equivalent_node_Proj(ir_node *proj) {
1714 ir_node *n = get_Proj_pred(proj);
1716 if (get_irn_mode(proj) == mode_X) {
1717 if (is_Block_dead(get_nodes_block(n))) {
1718 /* Remove dead control flow -- early gigo(). */
1719 return get_irg_bad(current_ir_graph);
1722 if (n->op->ops.equivalent_node_Proj)
1723 return n->op->ops.equivalent_node_Proj(proj);
1725 } /* equivalent_node_Proj */
1730 static ir_node *equivalent_node_Id(ir_node *n) {
1737 DBG_OPT_ID(oldn, n);
1739 } /* equivalent_node_Id */
1744 static ir_node *equivalent_node_Mux(ir_node *n)
1746 ir_node *oldn = n, *sel = get_Mux_sel(n);
1747 tarval *ts = value_of(sel);
1749 /* Mux(true, f, t) == t */
1750 if (ts == tarval_b_true) {
1751 n = get_Mux_true(n);
1752 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1754 /* Mux(false, f, t) == f */
1755 else if (ts == tarval_b_false) {
1756 n = get_Mux_false(n);
1757 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1759 /* Mux(v, x, x) == x */
1760 else if (get_Mux_false(n) == get_Mux_true(n)) {
1761 n = get_Mux_true(n);
1762 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1764 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1765 ir_node *cmp = get_Proj_pred(sel);
1766 long proj_nr = get_Proj_proj(sel);
1767 ir_node *f = get_Mux_false(n);
1768 ir_node *t = get_Mux_true(n);
1771 * Note further that these optimization work even for floating point
1772 * with NaN's because -NaN == NaN.
1773 * However, if +0 and -0 is handled differently, we cannot use the first one.
1776 ir_node *const cmp_l = get_Cmp_left(cmp);
1777 ir_node *const cmp_r = get_Cmp_right(cmp);
1781 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1782 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1784 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1791 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1792 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1794 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1801 * Note: normalization puts the constant on the right side,
1802 * so we check only one case.
1804 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1805 /* Mux(t CMP 0, X, t) */
1806 if (is_Minus(f) && get_Minus_op(f) == t) {
1807 /* Mux(t CMP 0, -t, t) */
1808 if (proj_nr == pn_Cmp_Eq) {
1809 /* Mux(t == 0, -t, t) ==> -t */
1811 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1812 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1813 /* Mux(t != 0, -t, t) ==> t */
1815 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1822 } /* equivalent_node_Mux */
1825 * Remove Confirm nodes if setting is on.
1826 * Replace Confirms(x, '=', Constlike) by Constlike.
1828 static ir_node *equivalent_node_Confirm(ir_node *n) {
1829 ir_node *pred = get_Confirm_value(n);
1830 pn_Cmp pnc = get_Confirm_cmp(n);
1832 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1834 * rare case: two identical Confirms one after another,
1835 * replace the second one with the first.
1838 pred = get_Confirm_value(n);
1839 pnc = get_Confirm_cmp(n);
1841 if (get_opt_remove_confirm())
1842 return get_Confirm_value(n);
1847 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1848 * perform no actual computation, as, e.g., the Id nodes. It does not create
1849 * new nodes. It is therefore safe to free n if the node returned is not n.
1850 * If a node returns a Tuple we can not just skip it. If the size of the
1851 * in array fits, we transform n into a tuple (e.g., Div).
1853 ir_node *equivalent_node(ir_node *n) {
1854 if (n->op->ops.equivalent_node)
1855 return n->op->ops.equivalent_node(n);
1857 } /* equivalent_node */
1860 * Sets the default equivalent node operation for an ir_op_ops.
1862 * @param code the opcode for the default operation
1863 * @param ops the operations initialized
1868 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1872 ops->equivalent_node = equivalent_node_##a; \
1874 #define CASE_PROJ(a) \
1876 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1918 } /* firm_set_default_equivalent_node */
1921 * Returns non-zero if a node is a Phi node
1922 * with all predecessors constant.
1924 static int is_const_Phi(ir_node *n) {
1927 if (! is_Phi(n) || get_irn_arity(n) == 0)
1929 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1930 if (! is_Const(get_irn_n(n, i)))
1934 } /* is_const_Phi */
1936 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1937 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1940 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1942 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1943 if (eval == tarval_sub) {
1944 tarval_sub_type func = (tarval_sub_type)eval;
1946 return func(a, b, mode);
1948 tarval_binop_type func = (tarval_binop_type)eval;
1955 * Apply an evaluator on a binop with a constant operators (and one Phi).
1957 * @param phi the Phi node
1958 * @param other the other operand
1959 * @param eval an evaluator function
1960 * @param mode the mode of the result, may be different from the mode of the Phi!
1961 * @param left if non-zero, other is the left operand, else the right
1963 * @return a new Phi node if the conversion was successful, NULL else
1965 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1970 int i, n = get_irn_arity(phi);
1972 NEW_ARR_A(void *, res, n);
1974 for (i = 0; i < n; ++i) {
1975 pred = get_irn_n(phi, i);
1976 tv = get_Const_tarval(pred);
1977 tv = do_eval(eval, other, tv, mode);
1979 if (tv == tarval_bad) {
1980 /* folding failed, bad */
1986 for (i = 0; i < n; ++i) {
1987 pred = get_irn_n(phi, i);
1988 tv = get_Const_tarval(pred);
1989 tv = do_eval(eval, tv, other, mode);
1991 if (tv == tarval_bad) {
1992 /* folding failed, bad */
1998 irg = current_ir_graph;
1999 for (i = 0; i < n; ++i) {
2000 pred = get_irn_n(phi, i);
2001 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2003 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2004 } /* apply_binop_on_phi */
2007 * Apply an evaluator on a binop with two constant Phi.
2009 * @param a the left Phi node
2010 * @param b the right Phi node
2011 * @param eval an evaluator function
2012 * @param mode the mode of the result, may be different from the mode of the Phi!
2014 * @return a new Phi node if the conversion was successful, NULL else
2016 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
2017 tarval *tv_l, *tv_r, *tv;
2023 if (get_nodes_block(a) != get_nodes_block(b))
2026 n = get_irn_arity(a);
2027 NEW_ARR_A(void *, res, n);
2029 for (i = 0; i < n; ++i) {
2030 pred = get_irn_n(a, i);
2031 tv_l = get_Const_tarval(pred);
2032 pred = get_irn_n(b, i);
2033 tv_r = get_Const_tarval(pred);
2034 tv = do_eval(eval, tv_l, tv_r, mode);
2036 if (tv == tarval_bad) {
2037 /* folding failed, bad */
2042 irg = current_ir_graph;
2043 for (i = 0; i < n; ++i) {
2044 pred = get_irn_n(a, i);
2045 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2047 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2048 } /* apply_binop_on_2_phis */
2051 * Apply an evaluator on a unop with a constant operator (a Phi).
2053 * @param phi the Phi node
2054 * @param eval an evaluator function
2056 * @return a new Phi node if the conversion was successful, NULL else
2058 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2064 int i, n = get_irn_arity(phi);
2066 NEW_ARR_A(void *, res, n);
2067 for (i = 0; i < n; ++i) {
2068 pred = get_irn_n(phi, i);
2069 tv = get_Const_tarval(pred);
2072 if (tv == tarval_bad) {
2073 /* folding failed, bad */
2078 mode = get_irn_mode(phi);
2079 irg = current_ir_graph;
2080 for (i = 0; i < n; ++i) {
2081 pred = get_irn_n(phi, i);
2082 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2084 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2085 } /* apply_unop_on_phi */
2088 * Apply a conversion on a constant operator (a Phi).
2090 * @param phi the Phi node
2092 * @return a new Phi node if the conversion was successful, NULL else
2094 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2099 int i, n = get_irn_arity(phi);
2101 NEW_ARR_A(void *, res, n);
2102 for (i = 0; i < n; ++i) {
2103 pred = get_irn_n(phi, i);
2104 tv = get_Const_tarval(pred);
2105 tv = tarval_convert_to(tv, mode);
2107 if (tv == tarval_bad) {
2108 /* folding failed, bad */
2113 irg = current_ir_graph;
2114 for (i = 0; i < n; ++i) {
2115 pred = get_irn_n(phi, i);
2116 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2118 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2119 } /* apply_conv_on_phi */
2122 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2123 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2124 * If possible, remove the Conv's.
2126 static ir_node *transform_node_AddSub(ir_node *n) {
2127 ir_mode *mode = get_irn_mode(n);
2129 if (mode_is_reference(mode)) {
2130 ir_node *left = get_binop_left(n);
2131 ir_node *right = get_binop_right(n);
2132 unsigned ref_bits = get_mode_size_bits(mode);
2134 if (is_Conv(left)) {
2135 ir_mode *lmode = get_irn_mode(left);
2136 unsigned bits = get_mode_size_bits(lmode);
2138 if (ref_bits == bits &&
2139 mode_is_int(lmode) &&
2140 get_mode_arithmetic(lmode) == irma_twos_complement) {
2141 ir_node *pre = get_Conv_op(left);
2142 ir_mode *pre_mode = get_irn_mode(pre);
2144 if (mode_is_int(pre_mode) &&
2145 get_mode_size_bits(pre_mode) == bits &&
2146 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2147 /* ok, this conv just changes to sign, moreover the calculation
2148 * is done with same number of bits as our address mode, so
2149 * we can ignore the conv as address calculation can be viewed
2150 * as either signed or unsigned
2152 set_binop_left(n, pre);
2157 if (is_Conv(right)) {
2158 ir_mode *rmode = get_irn_mode(right);
2159 unsigned bits = get_mode_size_bits(rmode);
2161 if (ref_bits == bits &&
2162 mode_is_int(rmode) &&
2163 get_mode_arithmetic(rmode) == irma_twos_complement) {
2164 ir_node *pre = get_Conv_op(right);
2165 ir_mode *pre_mode = get_irn_mode(pre);
2167 if (mode_is_int(pre_mode) &&
2168 get_mode_size_bits(pre_mode) == bits &&
2169 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2170 /* ok, this conv just changes to sign, moreover the calculation
2171 * is done with same number of bits as our address mode, so
2172 * we can ignore the conv as address calculation can be viewed
2173 * as either signed or unsigned
2175 set_binop_right(n, pre);
2180 /* let address arithmetic use unsigned modes */
2181 if (is_Const(right)) {
2182 ir_mode *rmode = get_irn_mode(right);
2184 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2185 /* convert a AddP(P, *s) into AddP(P, *u) */
2186 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2188 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2189 set_binop_right(n, pre);
2195 } /* transform_node_AddSub */
2197 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2199 if (is_Const(b) && is_const_Phi(a)) { \
2200 /* check for Op(Phi, Const) */ \
2201 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2203 else if (is_Const(a) && is_const_Phi(b)) { \
2204 /* check for Op(Const, Phi) */ \
2205 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2207 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2208 /* check for Op(Phi, Phi) */ \
2209 c = apply_binop_on_2_phis(a, b, eval, mode); \
2212 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2216 #define HANDLE_UNOP_PHI(eval, a, c) \
2218 if (is_const_Phi(a)) { \
2219 /* check for Op(Phi) */ \
2220 c = apply_unop_on_phi(a, eval); \
2222 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2228 * Do the AddSub optimization, then Transform
2229 * Constant folding on Phi
2230 * Add(a,a) -> Mul(a, 2)
2231 * Add(Mul(a, x), a) -> Mul(a, x+1)
2232 * if the mode is integer or float.
2233 * Transform Add(a,-b) into Sub(a,b).
2234 * Reassociation might fold this further.
2236 static ir_node *transform_node_Add(ir_node *n) {
2238 ir_node *a, *b, *c, *oldn = n;
2240 n = transform_node_AddSub(n);
2242 a = get_Add_left(n);
2243 b = get_Add_right(n);
2245 mode = get_irn_mode(n);
2247 if (mode_is_reference(mode)) {
2248 ir_mode *lmode = get_irn_mode(a);
2250 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2251 /* an Add(a, NULL) is a hidden Conv */
2252 dbg_info *dbg = get_irn_dbg_info(n);
2253 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2257 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2259 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2260 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2263 if (mode_is_num(mode)) {
2264 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2265 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2266 ir_node *block = get_nodes_block(n);
2269 get_irn_dbg_info(n),
2273 new_Const_long(mode, 2),
2275 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2280 get_irn_dbg_info(n),
2286 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2291 get_irn_dbg_info(n),
2297 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2300 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2301 /* Here we rely on constants be on the RIGHT side */
2303 ir_node *op = get_Not_op(a);
2305 if (is_Const(b) && is_Const_one(b)) {
2307 ir_node *blk = get_nodes_block(n);
2308 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2309 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2314 n = new_Const(get_mode_minus_one(mode));
2315 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2320 ir_node *op = get_Not_op(b);
2324 n = new_Const(get_mode_minus_one(mode));
2325 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2332 } /* transform_node_Add */
2335 * returns -cnst or NULL if impossible
2337 static ir_node *const_negate(ir_node *cnst) {
2338 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2339 dbg_info *dbgi = get_irn_dbg_info(cnst);
2340 ir_graph *irg = get_irn_irg(cnst);
2341 if (tv == tarval_bad) return NULL;
2342 return new_rd_Const(dbgi, irg, tv);
2346 * Do the AddSub optimization, then Transform
2347 * Constant folding on Phi
2348 * Sub(0,a) -> Minus(a)
2349 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2350 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2351 * Sub(Add(a, x), x) -> a
2352 * Sub(x, Add(x, a)) -> -a
2353 * Sub(x, Const) -> Add(x, -Const)
2355 static ir_node *transform_node_Sub(ir_node *n) {
2360 n = transform_node_AddSub(n);
2362 a = get_Sub_left(n);
2363 b = get_Sub_right(n);
2365 mode = get_irn_mode(n);
2367 if (mode_is_int(mode)) {
2368 ir_mode *lmode = get_irn_mode(a);
2370 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2371 /* a Sub(a, NULL) is a hidden Conv */
2372 dbg_info *dbg = get_irn_dbg_info(n);
2373 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2374 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2378 if (mode == lmode &&
2379 get_mode_arithmetic(mode) == irma_twos_complement &&
2381 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2383 dbg_info *dbg = get_irn_dbg_info(n);
2384 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2385 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2391 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2393 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2394 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2397 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2398 /* a - C -> a + (-C) */
2399 ir_node *cnst = const_negate(b);
2401 ir_node *block = get_nodes_block(n);
2402 dbg_info *dbgi = get_irn_dbg_info(n);
2403 ir_graph *irg = get_irn_irg(n);
2405 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2406 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2411 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2412 ir_graph *irg = current_ir_graph;
2413 dbg_info *dbg = get_irn_dbg_info(n);
2414 ir_node *block = get_nodes_block(n);
2415 ir_node *left = get_Minus_op(a);
2416 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2418 n = new_rd_Minus(dbg, irg, block, add, mode);
2419 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2421 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2422 ir_graph *irg = current_ir_graph;
2423 dbg_info *dbg = get_irn_dbg_info(n);
2424 ir_node *block = get_nodes_block(n);
2425 ir_node *right = get_Minus_op(b);
2427 n = new_rd_Add(dbg, irg, block, a, right, mode);
2428 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2430 } else if (is_Sub(b)) {
2431 /* a - (b - c) -> a + (c - b)
2432 * -> (a - b) + c iff (b - c) is a pointer */
2433 ir_graph *irg = current_ir_graph;
2434 dbg_info *s_dbg = get_irn_dbg_info(b);
2435 ir_node *s_block = get_nodes_block(b);
2436 ir_node *s_left = get_Sub_left(b);
2437 ir_node *s_right = get_Sub_right(b);
2438 ir_mode *s_mode = get_irn_mode(b);
2439 if (mode_is_reference(s_mode)) {
2440 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2441 dbg_info *a_dbg = get_irn_dbg_info(n);
2442 ir_node *a_block = get_nodes_block(n);
2445 s_right = new_r_Conv(irg, a_block, s_right, mode);
2446 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2448 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2449 dbg_info *a_dbg = get_irn_dbg_info(n);
2450 ir_node *a_block = get_nodes_block(n);
2452 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2454 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2456 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2457 ir_node *m_right = get_Mul_right(b);
2458 if (is_Const(m_right)) {
2459 ir_node *cnst2 = const_negate(m_right);
2460 if (cnst2 != NULL) {
2461 ir_graph *irg = current_ir_graph;
2462 dbg_info *m_dbg = get_irn_dbg_info(b);
2463 ir_node *m_block = get_nodes_block(b);
2464 ir_node *m_left = get_Mul_left(b);
2465 ir_mode *m_mode = get_irn_mode(b);
2466 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2467 dbg_info *a_dbg = get_irn_dbg_info(n);
2468 ir_node *a_block = get_nodes_block(n);
2470 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2471 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2477 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2478 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2480 get_irn_dbg_info(n),
2485 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2489 if (mode_wrap_around(mode)) {
2490 ir_node *left = get_Add_left(a);
2491 ir_node *right = get_Add_right(a);
2493 /* FIXME: Does the Conv's work only for two complement or generally? */
2495 if (mode != get_irn_mode(right)) {
2496 /* This Sub is an effective Cast */
2497 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2500 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2502 } else if (right == b) {
2503 if (mode != get_irn_mode(left)) {
2504 /* This Sub is an effective Cast */
2505 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2508 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2514 if (mode_wrap_around(mode)) {
2515 ir_node *left = get_Add_left(b);
2516 ir_node *right = get_Add_right(b);
2518 /* FIXME: Does the Conv's work only for two complement or generally? */
2520 ir_mode *r_mode = get_irn_mode(right);
2522 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2523 if (mode != r_mode) {
2524 /* This Sub is an effective Cast */
2525 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2527 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2529 } else if (right == a) {
2530 ir_mode *l_mode = get_irn_mode(left);
2532 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2533 if (mode != l_mode) {
2534 /* This Sub is an effective Cast */
2535 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2537 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2542 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2543 ir_mode *mode = get_irn_mode(a);
2545 if (mode == get_irn_mode(b)) {
2547 ir_node *op_a = get_Conv_op(a);
2548 ir_node *op_b = get_Conv_op(b);
2550 /* check if it's allowed to skip the conv */
2551 ma = get_irn_mode(op_a);
2552 mb = get_irn_mode(op_b);
2554 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2555 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2558 set_Sub_right(n, b);
2564 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2565 if (!is_reassoc_running() && is_Mul(a)) {
2566 ir_node *ma = get_Mul_left(a);
2567 ir_node *mb = get_Mul_right(a);
2570 ir_node *blk = get_nodes_block(n);
2572 get_irn_dbg_info(n),
2573 current_ir_graph, blk,
2576 get_irn_dbg_info(n),
2577 current_ir_graph, blk,
2579 new_Const_long(mode, 1),
2582 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2584 } else if (mb == b) {
2585 ir_node *blk = get_nodes_block(n);
2587 get_irn_dbg_info(n),
2588 current_ir_graph, blk,
2591 get_irn_dbg_info(n),
2592 current_ir_graph, blk,
2594 new_Const_long(mode, 1),
2597 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2601 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2602 ir_node *x = get_Sub_left(a);
2603 ir_node *y = get_Sub_right(a);
2604 ir_node *blk = get_nodes_block(n);
2605 ir_mode *m_b = get_irn_mode(b);
2606 ir_mode *m_y = get_irn_mode(y);
2610 /* Determine the right mode for the Add. */
2613 else if (mode_is_reference(m_b))
2615 else if (mode_is_reference(m_y))
2619 * Both modes are different but none is reference,
2620 * happens for instance in SubP(SubP(P, Iu), Is).
2621 * We have two possibilities here: Cast or ignore.
2622 * Currently we ignore this case.
2627 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2629 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2630 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2634 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2635 if (is_Const(a) && is_Not(b)) {
2636 /* c - ~X = X + (c+1) */
2637 tarval *tv = get_Const_tarval(a);
2639 tv = tarval_add(tv, get_mode_one(mode));
2640 if (tv != tarval_bad) {
2641 ir_node *blk = get_nodes_block(n);
2642 ir_node *c = new_Const(tv);
2643 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2644 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2650 } /* transform_node_Sub */
2653 * Several transformation done on n*n=2n bits mul.
2654 * These transformations must be done here because new nodes may be produced.
2656 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2658 ir_node *a = get_Mul_left(n);
2659 ir_node *b = get_Mul_right(n);
2660 tarval *ta = value_of(a);
2661 tarval *tb = value_of(b);
2662 ir_mode *smode = get_irn_mode(a);
2664 if (ta == get_mode_one(smode)) {
2665 /* (L)1 * (L)b = (L)b */
2666 ir_node *blk = get_nodes_block(n);
2667 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2668 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2671 else if (ta == get_mode_minus_one(smode)) {
2672 /* (L)-1 * (L)b = (L)b */
2673 ir_node *blk = get_nodes_block(n);
2674 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2675 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2676 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2679 if (tb == get_mode_one(smode)) {
2680 /* (L)a * (L)1 = (L)a */
2681 ir_node *blk = get_irn_n(a, -1);
2682 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2683 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2686 else if (tb == get_mode_minus_one(smode)) {
2687 /* (L)a * (L)-1 = (L)-a */
2688 ir_node *blk = get_nodes_block(n);
2689 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2690 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2691 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2698 * Transform Mul(a,-1) into -a.
2699 * Do constant evaluation of Phi nodes.
2700 * Do architecture dependent optimizations on Mul nodes
2702 static ir_node *transform_node_Mul(ir_node *n) {
2703 ir_node *c, *oldn = n;
2704 ir_mode *mode = get_irn_mode(n);
2705 ir_node *a = get_Mul_left(n);
2706 ir_node *b = get_Mul_right(n);
2708 if (is_Bad(a) || is_Bad(b))
2711 if (mode != get_irn_mode(a))
2712 return transform_node_Mul2n(n, mode);
2714 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2716 if (mode_is_signed(mode)) {
2719 if (value_of(a) == get_mode_minus_one(mode))
2721 else if (value_of(b) == get_mode_minus_one(mode))
2724 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2725 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2730 if (is_Const(b)) { /* (-a) * const -> a * -const */
2731 ir_node *cnst = const_negate(b);
2733 dbg_info *dbgi = get_irn_dbg_info(n);
2734 ir_node *block = get_nodes_block(n);
2735 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2736 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2739 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2740 dbg_info *dbgi = get_irn_dbg_info(n);
2741 ir_node *block = get_nodes_block(n);
2742 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2743 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2745 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2746 ir_node *sub_l = get_Sub_left(b);
2747 ir_node *sub_r = get_Sub_right(b);
2748 dbg_info *dbgi = get_irn_dbg_info(n);
2749 ir_graph *irg = current_ir_graph;
2750 ir_node *block = get_nodes_block(n);
2751 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2752 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2753 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2756 } else if (is_Minus(b)) {
2757 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2758 ir_node *sub_l = get_Sub_left(a);
2759 ir_node *sub_r = get_Sub_right(a);
2760 dbg_info *dbgi = get_irn_dbg_info(n);
2761 ir_graph *irg = current_ir_graph;
2762 ir_node *block = get_nodes_block(n);
2763 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2764 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2765 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2769 if (get_mode_arithmetic(mode) == irma_ieee754) {
2771 tarval *tv = get_Const_tarval(a);
2772 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2773 && !tarval_is_negative(tv)) {
2774 /* 2.0 * b = b + b */
2775 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2776 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2780 else if (is_Const(b)) {
2781 tarval *tv = get_Const_tarval(b);
2782 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2783 && !tarval_is_negative(tv)) {
2784 /* a * 2.0 = a + a */
2785 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2786 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2791 return arch_dep_replace_mul_with_shifts(n);
2792 } /* transform_node_Mul */
2795 * Transform a Div Node.
2797 static ir_node *transform_node_Div(ir_node *n) {
2798 ir_mode *mode = get_Div_resmode(n);
2799 ir_node *a = get_Div_left(n);
2800 ir_node *b = get_Div_right(n);
2804 if (is_Const(b) && is_const_Phi(a)) {
2805 /* check for Div(Phi, Const) */
2806 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2808 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2812 else if (is_Const(a) && is_const_Phi(b)) {
2813 /* check for Div(Const, Phi) */
2814 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2816 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2820 else if (is_const_Phi(a) && is_const_Phi(b)) {
2821 /* check for Div(Phi, Phi) */
2822 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2824 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2831 if (tv != tarval_bad) {
2832 value = new_Const(tv);
2834 DBG_OPT_CSTEVAL(n, value);
2837 ir_node *a = get_Div_left(n);
2838 ir_node *b = get_Div_right(n);
2839 const ir_node *dummy;
2841 if (a == b && value_not_zero(a, &dummy)) {
2842 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2843 value = new_Const(get_mode_one(mode));
2844 DBG_OPT_CSTEVAL(n, value);
2847 if (mode_is_signed(mode) && is_Const(b)) {
2848 tarval *tv = get_Const_tarval(b);
2850 if (tv == get_mode_minus_one(mode)) {
2852 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2853 DBG_OPT_CSTEVAL(n, value);
2857 /* Try architecture dependent optimization */
2858 value = arch_dep_replace_div_by_const(n);
2866 /* Turn Div into a tuple (mem, jmp, bad, value) */
2867 mem = get_Div_mem(n);
2868 blk = get_nodes_block(n);
2870 /* skip a potential Pin */
2871 mem = skip_Pin(mem);
2872 turn_into_tuple(n, pn_Div_max);
2873 set_Tuple_pred(n, pn_Div_M, mem);
2874 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2875 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2876 set_Tuple_pred(n, pn_Div_res, value);
2879 } /* transform_node_Div */
2882 * Transform a Mod node.
2884 static ir_node *transform_node_Mod(ir_node *n) {
2885 ir_mode *mode = get_Mod_resmode(n);
2886 ir_node *a = get_Mod_left(n);
2887 ir_node *b = get_Mod_right(n);
2891 if (is_Const(b) && is_const_Phi(a)) {
2892 /* check for Div(Phi, Const) */
2893 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2895 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2899 else if (is_Const(a) && is_const_Phi(b)) {
2900 /* check for Div(Const, Phi) */
2901 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2903 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2907 else if (is_const_Phi(a) && is_const_Phi(b)) {
2908 /* check for Div(Phi, Phi) */
2909 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2911 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2918 if (tv != tarval_bad) {
2919 value = new_Const(tv);
2921 DBG_OPT_CSTEVAL(n, value);
2924 ir_node *a = get_Mod_left(n);
2925 ir_node *b = get_Mod_right(n);
2926 const ir_node *dummy;
2928 if (a == b && value_not_zero(a, &dummy)) {
2929 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2930 value = new_Const(get_mode_null(mode));
2931 DBG_OPT_CSTEVAL(n, value);
2934 if (mode_is_signed(mode) && is_Const(b)) {
2935 tarval *tv = get_Const_tarval(b);
2937 if (tv == get_mode_minus_one(mode)) {
2939 value = new_Const(get_mode_null(mode));
2940 DBG_OPT_CSTEVAL(n, value);
2944 /* Try architecture dependent optimization */
2945 value = arch_dep_replace_mod_by_const(n);
2953 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2954 mem = get_Mod_mem(n);
2955 blk = get_nodes_block(n);
2957 /* skip a potential Pin */
2958 mem = skip_Pin(mem);
2959 turn_into_tuple(n, pn_Mod_max);
2960 set_Tuple_pred(n, pn_Mod_M, mem);
2961 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2962 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2963 set_Tuple_pred(n, pn_Mod_res, value);
2966 } /* transform_node_Mod */
2969 * Transform a DivMod node.
2971 static ir_node *transform_node_DivMod(ir_node *n) {
2972 const ir_node *dummy;
2973 ir_node *a = get_DivMod_left(n);
2974 ir_node *b = get_DivMod_right(n);
2975 ir_mode *mode = get_DivMod_resmode(n);
2980 if (is_Const(b) && is_const_Phi(a)) {
2981 /* check for Div(Phi, Const) */
2982 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2983 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2985 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2986 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2990 else if (is_Const(a) && is_const_Phi(b)) {
2991 /* check for Div(Const, Phi) */
2992 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2993 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2995 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2996 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3000 else if (is_const_Phi(a) && is_const_Phi(b)) {
3001 /* check for Div(Phi, Phi) */
3002 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
3003 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
3005 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3006 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3013 if (tb != tarval_bad) {
3014 if (tb == get_mode_one(get_tarval_mode(tb))) {
3016 vb = new_Const(get_mode_null(mode));
3017 DBG_OPT_CSTEVAL(n, vb);
3019 } else if (ta != tarval_bad) {
3020 tarval *resa, *resb;
3021 resa = tarval_div(ta, tb);
3022 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3023 Jmp for X result!? */
3024 resb = tarval_mod(ta, tb);
3025 if (resb == tarval_bad) return n; /* Causes exception! */
3026 va = new_Const(resa);
3027 vb = new_Const(resb);
3028 DBG_OPT_CSTEVAL(n, va);
3029 DBG_OPT_CSTEVAL(n, vb);
3031 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3032 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
3033 vb = new_Const(get_mode_null(mode));
3034 DBG_OPT_CSTEVAL(n, va);
3035 DBG_OPT_CSTEVAL(n, vb);
3037 } else { /* Try architecture dependent optimization */
3040 arch_dep_replace_divmod_by_const(&va, &vb, n);
3041 evaluated = va != NULL;
3043 } else if (a == b) {
3044 if (value_not_zero(a, &dummy)) {
3046 va = new_Const(get_mode_one(mode));
3047 vb = new_Const(get_mode_null(mode));
3048 DBG_OPT_CSTEVAL(n, va);
3049 DBG_OPT_CSTEVAL(n, vb);
3052 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3055 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3056 /* 0 / non-Const = 0 */
3061 if (evaluated) { /* replace by tuple */
3065 mem = get_DivMod_mem(n);
3066 /* skip a potential Pin */
3067 mem = skip_Pin(mem);
3069 blk = get_nodes_block(n);
3070 turn_into_tuple(n, pn_DivMod_max);
3071 set_Tuple_pred(n, pn_DivMod_M, mem);
3072 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3073 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3074 set_Tuple_pred(n, pn_DivMod_res_div, va);
3075 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3079 } /* transform_node_DivMod */
3082 * Optimize x / c to x * (1/c)
3084 static ir_node *transform_node_Quot(ir_node *n) {
3085 ir_mode *mode = get_Quot_resmode(n);
3088 if (get_mode_arithmetic(mode) == irma_ieee754) {
3089 ir_node *b = get_Quot_right(n);
3090 tarval *tv = value_of(b);
3092 if (tv != tarval_bad) {
3093 int rem = tarval_fp_ops_enabled();
3096 * Floating point constant folding might be disabled here to
3098 * However, as we check for exact result, doing it is safe.
3101 tarval_enable_fp_ops(1);
3102 tv = tarval_quo(get_mode_one(mode), tv);
3103 tarval_enable_fp_ops(rem);
3105 /* Do the transformation if the result is either exact or we are not
3106 using strict rules. */
3107 if (tv != tarval_bad &&
3108 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3109 ir_node *blk = get_nodes_block(n);
3110 ir_node *c = new_Const(tv);
3111 ir_node *a = get_Quot_left(n);
3112 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3113 ir_node *mem = get_Quot_mem(n);
3115 /* skip a potential Pin */
3116 mem = skip_Pin(mem);
3117 turn_into_tuple(n, pn_Quot_max);
3118 set_Tuple_pred(n, pn_Quot_M, mem);
3119 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3120 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3121 set_Tuple_pred(n, pn_Quot_res, m);
3122 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3127 } /* transform_node_Quot */
3130 * Optimize Abs(x) into x if x is Confirmed >= 0
3131 * Optimize Abs(x) into -x if x is Confirmed <= 0
3132 * Optimize Abs(-x) int Abs(x)
3134 static ir_node *transform_node_Abs(ir_node *n) {
3135 ir_node *c, *oldn = n;
3136 ir_node *a = get_Abs_op(n);
3139 HANDLE_UNOP_PHI(tarval_abs, a, c);
3141 switch (classify_value_sign(a)) {
3142 case value_classified_negative:
3143 mode = get_irn_mode(n);
3146 * We can replace the Abs by -x here.
3147 * We even could add a new Confirm here
3148 * (if not twos complement)
3150 * Note that -x would create a new node, so we could
3151 * not run it in the equivalent_node() context.
3153 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3154 get_nodes_block(n), a, mode);
3156 DBG_OPT_CONFIRM(oldn, n);
3158 case value_classified_positive:
3159 /* n is positive, Abs is not needed */
3162 DBG_OPT_CONFIRM(oldn, n);
3168 /* Abs(-x) = Abs(x) */
3169 mode = get_irn_mode(n);
3170 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3171 get_nodes_block(n), get_Minus_op(a), mode);
3172 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3176 } /* transform_node_Abs */
3179 * Optimize -a CMP -b into b CMP a.
3180 * This works only for for modes where unary Minus
3182 * Note that two-complement integers can Overflow
3183 * so it will NOT work.
3185 * For == and != can be handled in Proj(Cmp)
3187 static ir_node *transform_node_Cmp(ir_node *n) {
3189 ir_node *left = get_Cmp_left(n);
3190 ir_node *right = get_Cmp_right(n);
3192 if (is_Minus(left) && is_Minus(right) &&
3193 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3194 ir_node *const new_left = get_Minus_op(right);
3195 ir_node *const new_right = get_Minus_op(left);
3196 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3197 get_nodes_block(n), new_left, new_right);
3198 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3201 } /* transform_node_Cmp */
3205 * Transform a Cond node.
3207 * Replace the Cond by a Jmp if it branches on a constant
3210 static ir_node *transform_node_Cond(ir_node *n) {
3213 ir_node *a = get_Cond_selector(n);
3214 tarval *ta = value_of(a);
3216 /* we need block info which is not available in floating irgs */
3217 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3220 if ((ta != tarval_bad) &&
3221 (get_irn_mode(a) == mode_b) &&
3222 (get_opt_unreachable_code())) {
3223 /* It's a boolean Cond, branching on a boolean constant.
3224 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3225 ir_node *blk = get_nodes_block(n);
3226 jmp = new_r_Jmp(current_ir_graph, blk);
3227 turn_into_tuple(n, pn_Cond_max);
3228 if (ta == tarval_b_true) {
3229 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3230 set_Tuple_pred(n, pn_Cond_true, jmp);
3232 set_Tuple_pred(n, pn_Cond_false, jmp);
3233 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3235 /* We might generate an endless loop, so keep it alive. */
3236 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3239 } /* transform_node_Cond */
3242 * Prototype of a recursive transform function
3243 * for bitwise distributive transformations.
3245 typedef ir_node* (*recursive_transform)(ir_node *n);
3248 * makes use of distributive laws for and, or, eor
3249 * and(a OP c, b OP c) -> and(a, b) OP c
3250 * note, might return a different op than n
3252 static ir_node *transform_bitwise_distributive(ir_node *n,
3253 recursive_transform trans_func)
3256 ir_node *a = get_binop_left(n);
3257 ir_node *b = get_binop_right(n);
3258 ir_op *op = get_irn_op(a);
3259 ir_op *op_root = get_irn_op(n);
3261 if(op != get_irn_op(b))
3264 if (op == op_Conv) {
3265 ir_node *a_op = get_Conv_op(a);
3266 ir_node *b_op = get_Conv_op(b);
3267 ir_mode *a_mode = get_irn_mode(a_op);
3268 ir_mode *b_mode = get_irn_mode(b_op);
3269 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3270 ir_node *blk = get_nodes_block(n);
3273 set_binop_left(n, a_op);
3274 set_binop_right(n, b_op);
3275 set_irn_mode(n, a_mode);
3277 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3279 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3285 /* nothing to gain here */
3289 if (op == op_Shrs || op == op_Shr || op == op_Shl
3290 || op == op_And || op == op_Or || op == op_Eor) {
3291 ir_node *a_left = get_binop_left(a);
3292 ir_node *a_right = get_binop_right(a);
3293 ir_node *b_left = get_binop_left(b);
3294 ir_node *b_right = get_binop_right(b);
3296 ir_node *op1 = NULL;
3297 ir_node *op2 = NULL;
3299 if (is_op_commutative(op)) {
3300 if (a_left == b_left) {
3304 } else if(a_left == b_right) {
3308 } else if(a_right == b_left) {
3314 if(a_right == b_right) {
3321 /* (a sop c) & (b sop c) => (a & b) sop c */
3322 ir_node *blk = get_nodes_block(n);
3324 ir_node *new_n = exact_copy(n);
3325 set_binop_left(new_n, op1);
3326 set_binop_right(new_n, op2);
3327 new_n = trans_func(new_n);
3329 if(op_root == op_Eor && op == op_Or) {
3330 dbg_info *dbgi = get_irn_dbg_info(n);
3331 ir_graph *irg = current_ir_graph;
3332 ir_mode *mode = get_irn_mode(c);
3334 c = new_rd_Not(dbgi, irg, blk, c, mode);
3335 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3338 set_nodes_block(n, blk);
3339 set_binop_left(n, new_n);
3340 set_binop_right(n, c);
3341 add_identities(current_ir_graph->value_table, n);
3344 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3355 static ir_node *transform_node_And(ir_node *n) {
3356 ir_node *c, *oldn = n;
3357 ir_node *a = get_And_left(n);
3358 ir_node *b = get_And_right(n);
3361 mode = get_irn_mode(n);
3362 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3364 /* we can evaluate 2 Projs of the same Cmp */
3365 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3366 ir_node *pred_a = get_Proj_pred(a);
3367 ir_node *pred_b = get_Proj_pred(b);
3368 if (pred_a == pred_b) {
3369 dbg_info *dbgi = get_irn_dbg_info(n);
3370 ir_node *block = get_nodes_block(pred_a);
3371 pn_Cmp pn_a = get_Proj_proj(a);
3372 pn_Cmp pn_b = get_Proj_proj(b);
3373 /* yes, we can simply calculate with pncs */
3374 pn_Cmp new_pnc = pn_a & pn_b;
3376 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3381 ir_node *op = get_Not_op(b);
3383 ir_node *ba = get_And_left(op);
3384 ir_node *bb = get_And_right(op);
3386 /* it's enough to test the following cases due to normalization! */
3387 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3388 /* (a|b) & ~(a&b) = a^b */
3389 ir_node *block = get_nodes_block(n);
3391 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3392 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3400 ir_node *op = get_Not_op(a);
3402 ir_node *aa = get_And_left(op);
3403 ir_node *ab = get_And_right(op);
3405 /* it's enough to test the following cases due to normalization! */
3406 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3407 /* (a|b) & ~(a&b) = a^b */
3408 ir_node *block = get_nodes_block(n);
3410 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3411 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3418 ir_node *al = get_Eor_left(a);
3419 ir_node *ar = get_Eor_right(a);
3422 /* (b ^ a) & b -> ~a & b */
3423 dbg_info *dbg = get_irn_dbg_info(n);
3424 ir_node *block = get_nodes_block(n);
3426 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3427 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3428 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3432 /* (a ^ b) & b -> ~a & b */
3433 dbg_info *dbg = get_irn_dbg_info(n);
3434 ir_node *block = get_nodes_block(n);
3436 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3437 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3438 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3443 ir_node *bl = get_Eor_left(b);
3444 ir_node *br = get_Eor_right(b);
3447 /* a & (a ^ b) -> a & ~b */
3448 dbg_info *dbg = get_irn_dbg_info(n);
3449 ir_node *block = get_nodes_block(n);
3451 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3452 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3453 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3457 /* a & (b ^ a) -> a & ~b */
3458 dbg_info *dbg = get_irn_dbg_info(n);
3459 ir_node *block = get_nodes_block(n);
3461 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3462 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3463 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3467 if (is_Not(a) && is_Not(b)) {
3468 /* ~a & ~b = ~(a|b) */
3469 ir_node *block = get_nodes_block(n);
3470 ir_mode *mode = get_irn_mode(n);
3474 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3475 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3476 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3480 n = transform_bitwise_distributive(n, transform_node_And);
3483 } /* transform_node_And */
3488 static ir_node *transform_node_Eor(ir_node *n) {
3489 ir_node *c, *oldn = n;
3490 ir_node *a = get_Eor_left(n);
3491 ir_node *b = get_Eor_right(n);
3492 ir_mode *mode = get_irn_mode(n);
3494 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3496 /* we can evaluate 2 Projs of the same Cmp */
3497 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3498 ir_node *pred_a = get_Proj_pred(a);
3499 ir_node *pred_b = get_Proj_pred(b);
3500 if(pred_a == pred_b) {
3501 dbg_info *dbgi = get_irn_dbg_info(n);
3502 ir_node *block = get_nodes_block(pred_a);
3503 pn_Cmp pn_a = get_Proj_proj(a);
3504 pn_Cmp pn_b = get_Proj_proj(b);
3505 /* yes, we can simply calculate with pncs */
3506 pn_Cmp new_pnc = pn_a ^ pn_b;
3508 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3515 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3516 get_mode_null(mode));
3517 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3518 } else if (mode == mode_b &&
3520 is_Const(b) && is_Const_one(b) &&
3521 is_Cmp(get_Proj_pred(a))) {
3522 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3523 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3524 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3526 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3527 } else if (is_Const(b)) {
3528 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3529 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b)));
3530 ir_node *not_op = get_Not_op(a);
3531 dbg_info *dbg = get_irn_dbg_info(n);
3532 ir_graph *irg = current_ir_graph;
3533 ir_node *block = get_nodes_block(n);
3534 ir_mode *mode = get_irn_mode(n);
3535 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3537 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3538 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3539 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3542 n = transform_bitwise_distributive(n, transform_node_Eor);
3546 } /* transform_node_Eor */
3551 static ir_node *transform_node_Not(ir_node *n) {
3552 ir_node *c, *oldn = n;
3553 ir_node *a = get_Not_op(n);
3554 ir_mode *mode = get_irn_mode(n);
3556 HANDLE_UNOP_PHI(tarval_not,a,c);
3558 /* check for a boolean Not */
3559 if (mode == mode_b &&
3561 is_Cmp(get_Proj_pred(a))) {
3562 /* We negate a Cmp. The Cmp has the negated result anyways! */
3563 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3564 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3565 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3569 ir_node *eor_b = get_Eor_right(a);
3570 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3571 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b)));
3572 ir_node *eor_a = get_Eor_left(a);
3573 dbg_info *dbg = get_irn_dbg_info(n);
3574 ir_graph *irg = current_ir_graph;
3575 ir_node *block = get_nodes_block(n);
3576 ir_mode *mode = get_irn_mode(n);
3577 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3581 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3582 if (is_Minus(a)) { /* ~-x -> x + -1 */
3583 dbg_info *dbg = get_irn_dbg_info(n);
3584 ir_graph *irg = current_ir_graph;
3585 ir_node *block = get_nodes_block(n);
3586 ir_node *add_l = get_Minus_op(a);
3587 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3588 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3589 } else if (is_Add(a)) {
3590 ir_node *add_r = get_Add_right(a);
3591 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3592 /* ~(x + -1) = -x */
3593 ir_node *op = get_Add_left(a);
3594 ir_node *blk = get_nodes_block(n);
3595 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3596 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3601 } /* transform_node_Not */
3604 * Transform a Minus.
3608 * -(a >>u (size-1)) = a >>s (size-1)
3609 * -(a >>s (size-1)) = a >>u (size-1)
3610 * -(a * const) -> a * -const
3612 static ir_node *transform_node_Minus(ir_node *n) {
3613 ir_node *c, *oldn = n;
3614 ir_node *a = get_Minus_op(n);
3617 HANDLE_UNOP_PHI(tarval_neg,a,c);
3619 mode = get_irn_mode(a);
3620 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3621 /* the following rules are only to twos-complement */
3624 ir_node *op = get_Not_op(a);
3625 tarval *tv = get_mode_one(mode);
3626 ir_node *blk = get_nodes_block(n);
3627 ir_node *c = new_Const(tv);
3628 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3629 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3633 ir_node *c = get_Shr_right(a);
3636 tarval *tv = get_Const_tarval(c);
3638 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3639 /* -(a >>u (size-1)) = a >>s (size-1) */
3640 ir_node *v = get_Shr_left(a);
3642 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3643 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3649 ir_node *c = get_Shrs_right(a);
3652 tarval *tv = get_Const_tarval(c);
3654 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3655 /* -(a >>s (size-1)) = a >>u (size-1) */
3656 ir_node *v = get_Shrs_left(a);
3658 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3659 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3666 /* - (a-b) = b - a */
3667 ir_node *la = get_Sub_left(a);
3668 ir_node *ra = get_Sub_right(a);
3669 ir_node *blk = get_nodes_block(n);
3671 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3672 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3676 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3677 ir_node *mul_l = get_Mul_left(a);
3678 ir_node *mul_r = get_Mul_right(a);
3679 tarval *tv = value_of(mul_r);
3680 if (tv != tarval_bad) {
3681 tv = tarval_neg(tv);
3682 if (tv != tarval_bad) {
3683 ir_node *cnst = new_Const(tv);
3684 dbg_info *dbg = get_irn_dbg_info(a);
3685 ir_graph *irg = current_ir_graph;
3686 ir_node *block = get_nodes_block(a);
3687 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3688 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3695 } /* transform_node_Minus */
3698 * Transform a Cast_type(Const) into a new Const_type
3700 static ir_node *transform_node_Cast(ir_node *n) {
3702 ir_node *pred = get_Cast_op(n);
3703 ir_type *tp = get_irn_type(n);
3705 if (is_Const(pred) && get_Const_type(pred) != tp) {
3706 n = new_rd_Const_type(NULL, current_ir_graph, get_Const_tarval(pred), tp);
3707 DBG_OPT_CSTEVAL(oldn, n);
3708 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3709 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3710 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3711 DBG_OPT_CSTEVAL(oldn, n);
3715 } /* transform_node_Cast */
3718 * Transform a Proj(Load) with a non-null address.
3720 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3721 if (get_opt_ldst_only_null_ptr_exceptions()) {
3722 if (get_irn_mode(proj) == mode_X) {
3723 ir_node *load = get_Proj_pred(proj);
3725 /* get the Load address */
3726 const ir_node *addr = get_Load_ptr(load);
3727 const ir_node *confirm;
3729 if (value_not_null(addr, &confirm)) {
3730 if (confirm == NULL) {
3731 /* this node may float if it did not depend on a Confirm */
3732 set_irn_pinned(load, op_pin_state_floats);
3734 if (get_Proj_proj(proj) == pn_Load_X_except) {
3735 DBG_OPT_EXC_REM(proj);
3736 return get_irg_bad(current_ir_graph);
3738 ir_node *blk = get_nodes_block(load);
3739 return new_r_Jmp(current_ir_graph, blk);
3745 } /* transform_node_Proj_Load */
3748 * Transform a Proj(Store) with a non-null address.
3750 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3751 if (get_opt_ldst_only_null_ptr_exceptions()) {
3752 if (get_irn_mode(proj) == mode_X) {
3753 ir_node *store = get_Proj_pred(proj);
3755 /* get the load/store address */
3756 const ir_node *addr = get_Store_ptr(store);
3757 const ir_node *confirm;
3759 if (value_not_null(addr, &confirm)) {
3760 if (confirm == NULL) {
3761 /* this node may float if it did not depend on a Confirm */
3762 set_irn_pinned(store, op_pin_state_floats);
3764 if (get_Proj_proj(proj) == pn_Store_X_except) {
3765 DBG_OPT_EXC_REM(proj);
3766 return get_irg_bad(current_ir_graph);
3768 ir_node *blk = get_nodes_block(store);
3769 return new_r_Jmp(current_ir_graph, blk);
3775 } /* transform_node_Proj_Store */
3778 * Transform a Proj(Div) with a non-zero value.
3779 * Removes the exceptions and routes the memory to the NoMem node.
3781 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3782 ir_node *div = get_Proj_pred(proj);
3783 ir_node *b = get_Div_right(div);
3784 ir_node *res, *new_mem;
3785 const ir_node *confirm;
3788 if (value_not_zero(b, &confirm)) {
3789 /* div(x, y) && y != 0 */
3790 if (confirm == NULL) {
3791 /* we are sure we have a Const != 0 */
3792 new_mem = get_Div_mem(div);
3793 new_mem = skip_Pin(new_mem);
3794 set_Div_mem(div, new_mem);
3795 set_irn_pinned(div, op_pin_state_floats);
3798 proj_nr = get_Proj_proj(proj);
3800 case pn_Div_X_regular:
3801 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3803 case pn_Div_X_except:
3804 /* we found an exception handler, remove it */
3805 DBG_OPT_EXC_REM(proj);
3809 res = get_Div_mem(div);
3810 new_mem = get_irg_no_mem(current_ir_graph);
3813 /* This node can only float up to the Confirm block */
3814 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3816 set_irn_pinned(div, op_pin_state_floats);
3817 /* this is a Div without exception, we can remove the memory edge */
3818 set_Div_mem(div, new_mem);
3823 } /* transform_node_Proj_Div */
3826 * Transform a Proj(Mod) with a non-zero value.
3827 * Removes the exceptions and routes the memory to the NoMem node.
3829 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3830 ir_node *mod = get_Proj_pred(proj);
3831 ir_node *b = get_Mod_right(mod);
3832 ir_node *res, *new_mem;
3833 const ir_node *confirm;
3836 if (value_not_zero(b, &confirm)) {
3837 /* mod(x, y) && y != 0 */
3838 proj_nr = get_Proj_proj(proj);
3840 if (confirm == NULL) {
3841 /* we are sure we have a Const != 0 */
3842 new_mem = get_Mod_mem(mod);
3843 new_mem = skip_Pin(new_mem);
3844 set_Mod_mem(mod, new_mem);
3845 set_irn_pinned(mod, op_pin_state_floats);
3850 case pn_Mod_X_regular:
3851 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3853 case pn_Mod_X_except:
3854 /* we found an exception handler, remove it */
3855 DBG_OPT_EXC_REM(proj);
3859 res = get_Mod_mem(mod);
3860 new_mem = get_irg_no_mem(current_ir_graph);
3863 /* This node can only float up to the Confirm block */
3864 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3866 /* this is a Mod without exception, we can remove the memory edge */
3867 set_Mod_mem(mod, new_mem);
3870 if (get_Mod_left(mod) == b) {
3871 /* a % a = 0 if a != 0 */
3872 ir_mode *mode = get_irn_mode(proj);
3873 ir_node *res = new_Const(get_mode_null(mode));
3875 DBG_OPT_CSTEVAL(mod, res);
3881 } /* transform_node_Proj_Mod */
3884 * Transform a Proj(DivMod) with a non-zero value.
3885 * Removes the exceptions and routes the memory to the NoMem node.
3887 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3888 ir_node *divmod = get_Proj_pred(proj);
3889 ir_node *b = get_DivMod_right(divmod);
3890 ir_node *res, *new_mem;
3891 const ir_node *confirm;
3894 if (value_not_zero(b, &confirm)) {
3895 /* DivMod(x, y) && y != 0 */
3896 proj_nr = get_Proj_proj(proj);
3898 if (confirm == NULL) {
3899 /* we are sure we have a Const != 0 */
3900 new_mem = get_DivMod_mem(divmod);
3901 new_mem = skip_Pin(new_mem);
3902 set_DivMod_mem(divmod, new_mem);
3903 set_irn_pinned(divmod, op_pin_state_floats);
3908 case pn_DivMod_X_regular:
3909 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3911 case pn_DivMod_X_except:
3912 /* we found an exception handler, remove it */
3913 DBG_OPT_EXC_REM(proj);
3917 res = get_DivMod_mem(divmod);
3918 new_mem = get_irg_no_mem(current_ir_graph);
3921 /* This node can only float up to the Confirm block */
3922 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3924 /* this is a DivMod without exception, we can remove the memory edge */
3925 set_DivMod_mem(divmod, new_mem);
3928 case pn_DivMod_res_mod:
3929 if (get_DivMod_left(divmod) == b) {
3930 /* a % a = 0 if a != 0 */
3931 ir_mode *mode = get_irn_mode(proj);
3932 ir_node *res = new_Const(get_mode_null(mode));
3934 DBG_OPT_CSTEVAL(divmod, res);
3940 } /* transform_node_Proj_DivMod */
3943 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3945 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3946 if (get_opt_unreachable_code()) {
3947 ir_node *n = get_Proj_pred(proj);
3948 ir_node *b = get_Cond_selector(n);
3950 if (mode_is_int(get_irn_mode(b))) {
3951 tarval *tb = value_of(b);
3953 if (tb != tarval_bad) {
3954 /* we have a constant switch */
3955 long num = get_Proj_proj(proj);
3957 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3958 if (get_tarval_long(tb) == num) {
3959 /* Do NOT create a jump here, or we will have 2 control flow ops
3960 * in a block. This case is optimized away in optimize_cf(). */
3963 /* this case will NEVER be taken, kill it */
3964 return get_irg_bad(current_ir_graph);
3971 } /* transform_node_Proj_Cond */
3974 * Create a 0 constant of given mode.
3976 static ir_node *create_zero_const(ir_mode *mode) {
3977 tarval *tv = get_mode_null(mode);
3978 ir_node *cnst = new_Const(tv);
3983 /* the order of the values is important! */
3984 typedef enum const_class {
3990 static const_class classify_const(const ir_node* n)
3992 if (is_Const(n)) return const_const;
3993 if (is_irn_constlike(n)) return const_like;
3998 * Determines whether r is more constlike or has a larger index (in that order)
4001 static int operands_are_normalized(const ir_node *l, const ir_node *r)
4003 const const_class l_order = classify_const(l);
4004 const const_class r_order = classify_const(r);
4006 l_order > r_order ||
4007 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4011 * Normalizes and optimizes Cmp nodes.
4013 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4014 ir_node *n = get_Proj_pred(proj);
4015 ir_node *left = get_Cmp_left(n);
4016 ir_node *right = get_Cmp_right(n);
4019 ir_mode *mode = NULL;
4020 long proj_nr = get_Proj_proj(proj);
4022 /* we can evaluate some cases directly */
4025 return new_Const(get_tarval_b_false());
4027 return new_Const(get_tarval_b_true());
4029 if (!mode_is_float(get_irn_mode(left)))
4030 return new_Const(get_tarval_b_true());
4036 /* remove Casts of both sides */
4037 left = skip_Cast(left);
4038 right = skip_Cast(right);
4040 /* Remove unnecessary conversions */
4041 /* TODO handle constants */
4042 if (is_Conv(left) && is_Conv(right)) {
4043 ir_mode *mode = get_irn_mode(left);
4044 ir_node *op_left = get_Conv_op(left);
4045 ir_node *op_right = get_Conv_op(right);
4046 ir_mode *mode_left = get_irn_mode(op_left);
4047 ir_mode *mode_right = get_irn_mode(op_right);
4049 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4050 && mode_left != mode_b && mode_right != mode_b) {
4051 ir_graph *irg = current_ir_graph;
4052 ir_node *block = get_nodes_block(n);
4054 if (mode_left == mode_right) {
4058 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4059 } else if (smaller_mode(mode_left, mode_right)) {
4060 left = new_r_Conv(irg, block, op_left, mode_right);
4063 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4064 } else if (smaller_mode(mode_right, mode_left)) {
4066 right = new_r_Conv(irg, block, op_right, mode_left);
4068 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4073 /* remove operation on both sides if possible */
4074 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4076 * The following operations are NOT safe for floating point operations, for instance
4077 * 1.0 + inf == 2.0 + inf, =/=> x == y
4079 if (mode_is_int(get_irn_mode(left))) {
4080 unsigned lop = get_irn_opcode(left);
4082 if (lop == get_irn_opcode(right)) {
4083 ir_node *ll, *lr, *rl, *rr;
4085 /* same operation on both sides, try to remove */
4089 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4090 left = get_unop_op(left);
4091 right = get_unop_op(right);
4093 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4096 ll = get_Add_left(left);
4097 lr = get_Add_right(left);
4098 rl = get_Add_left(right);
4099 rr = get_Add_right(right);
4102 /* X + a CMP X + b ==> a CMP b */
4106 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4107 } else if (ll == rr) {
4108 /* X + a CMP b + X ==> a CMP b */
4112 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4113 } else if (lr == rl) {
4114 /* a + X CMP X + b ==> a CMP b */
4118 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4119 } else if (lr == rr) {
4120 /* a + X CMP b + X ==> a CMP b */
4124 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4128 ll = get_Sub_left(left);
4129 lr = get_Sub_right(left);
4130 rl = get_Sub_left(right);
4131 rr = get_Sub_right(right);
4134 /* X - a CMP X - b ==> a CMP b */
4138 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4139 } else if (lr == rr) {
4140 /* a - X CMP b - X ==> a CMP b */
4144 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4148 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4149 /* a ROTL X CMP b ROTL X ==> a CMP b */
4150 left = get_Rotl_left(left);
4151 right = get_Rotl_left(right);
4153 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4161 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4162 if (is_Add(left) || is_Sub(left)) {
4163 ir_node *ll = get_binop_left(left);
4164 ir_node *lr = get_binop_right(left);
4166 if (lr == right && is_Add(left)) {
4173 right = create_zero_const(get_irn_mode(left));
4175 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4178 if (is_Add(right) || is_Sub(right)) {
4179 ir_node *rl = get_binop_left(right);
4180 ir_node *rr = get_binop_right(right);
4182 if (rr == left && is_Add(right)) {
4189 right = create_zero_const(get_irn_mode(left));
4191 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4194 if (is_And(left) && is_Const(right)) {
4195 ir_node *ll = get_binop_left(left);
4196 ir_node *lr = get_binop_right(left);
4197 if (is_Shr(ll) && is_Const(lr)) {
4198 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4199 ir_node *block = get_nodes_block(n);
4200 ir_mode *mode = get_irn_mode(left);
4202 ir_node *llr = get_Shr_right(ll);
4203 if (is_Const(llr)) {
4204 ir_graph *irg = current_ir_graph;
4205 dbg_info *dbg = get_irn_dbg_info(left);
4207 tarval *c1 = get_Const_tarval(llr);
4208 tarval *c2 = get_Const_tarval(lr);
4209 tarval *c3 = get_Const_tarval(right);
4210 tarval *mask = tarval_shl(c2, c1);
4211 tarval *value = tarval_shl(c3, c1);
4213 left = new_rd_And(dbg, irg, block, get_Shr_left(ll), new_Const(mask), mode);
4214 right = new_Const(value);
4219 } /* mode_is_int(...) */
4220 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4222 /* replace mode_b compares with ands/ors */
4223 if (get_irn_mode(left) == mode_b) {
4224 ir_graph *irg = current_ir_graph;
4225 ir_node *block = get_nodes_block(n);
4229 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4230 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4231 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4232 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4233 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4234 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4235 default: bres = NULL;
4238 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4244 * First step: normalize the compare op
4245 * by placing the constant on the right side
4246 * or moving the lower address node to the left.
4248 if (!operands_are_normalized(left, right)) {
4254 proj_nr = get_inversed_pnc(proj_nr);
4259 * Second step: Try to reduce the magnitude
4260 * of a constant. This may help to generate better code
4261 * later and may help to normalize more compares.
4262 * Of course this is only possible for integer values.
4264 tv = value_of(right);
4265 if (tv != tarval_bad) {
4266 mode = get_irn_mode(right);
4268 /* TODO extend to arbitrary constants */
4269 if (is_Conv(left) && tarval_is_null(tv)) {
4270 ir_node *op = get_Conv_op(left);
4271 ir_mode *op_mode = get_irn_mode(op);
4274 * UpConv(x) REL 0 ==> x REL 0
4276 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4277 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4278 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4279 tv = get_mode_null(op_mode);
4283 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4287 if (tv != tarval_bad) {
4288 /* the following optimization is possible on modes without Overflow
4289 * on Unary Minus or on == and !=:
4290 * -a CMP c ==> a swap(CMP) -c
4292 * Beware: for two-complement Overflow may occur, so only == and != can
4293 * be optimized, see this:
4294 * -MININT < 0 =/=> MININT > 0 !!!
4296 if (is_Minus(left) &&
4297 (!mode_overflow_on_unary_Minus(mode) ||
4298 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4299 tv = tarval_neg(tv);
4301 if (tv != tarval_bad) {
4302 left = get_Minus_op(left);
4303 proj_nr = get_inversed_pnc(proj_nr);
4305 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4307 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4308 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4309 tv = tarval_not(tv);
4311 if (tv != tarval_bad) {
4312 left = get_Not_op(left);
4314 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4318 /* for integer modes, we have more */
4319 if (mode_is_int(mode)) {
4320 /* Ne includes Unordered which is not possible on integers.
4321 * However, frontends often use this wrong, so fix it here */
4322 if (proj_nr & pn_Cmp_Uo) {
4323 proj_nr &= ~pn_Cmp_Uo;
4324 set_Proj_proj(proj, proj_nr);
4327 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4328 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4329 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4330 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4332 if (tv != tarval_bad) {
4333 proj_nr ^= pn_Cmp_Eq;
4335 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4338 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4339 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4340 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4341 tv = tarval_add(tv, get_mode_one(mode));
4343 if (tv != tarval_bad) {
4344 proj_nr ^= pn_Cmp_Eq;
4346 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4350 /* the following reassociations work only for == and != */
4351 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4353 #if 0 /* Might be not that good in general */
4354 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4355 if (tarval_is_null(tv) && is_Sub(left)) {
4356 right = get_Sub_right(left);
4357 left = get_Sub_left(left);
4359 tv = value_of(right);
4361 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4365 if (tv != tarval_bad) {
4366 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4368 ir_node *c1 = get_Sub_right(left);
4369 tarval *tv2 = value_of(c1);
4371 if (tv2 != tarval_bad) {
4372 tv2 = tarval_add(tv, value_of(c1));
4374 if (tv2 != tarval_bad) {
4375 left = get_Sub_left(left);
4378 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4382 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4383 else if (is_Add(left)) {
4384 ir_node *a_l = get_Add_left(left);
4385 ir_node *a_r = get_Add_right(left);
4389 if (is_Const(a_l)) {
4391 tv2 = value_of(a_l);
4394 tv2 = value_of(a_r);
4397 if (tv2 != tarval_bad) {
4398 tv2 = tarval_sub(tv, tv2, NULL);
4400 if (tv2 != tarval_bad) {
4404 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4408 /* -a == c ==> a == -c, -a != c ==> a != -c */
4409 else if (is_Minus(left)) {
4410 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4412 if (tv2 != tarval_bad) {
4413 left = get_Minus_op(left);
4416 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4421 /* the following reassociations work only for <= */
4422 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4423 if (tv != tarval_bad) {
4424 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4425 if (is_Abs(left)) { // TODO something is missing here
4431 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4432 switch (get_irn_opcode(left)) {
4436 c1 = get_And_right(left);
4439 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4440 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4442 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4444 /* TODO: move to constant evaluation */
4445 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4447 DBG_OPT_CSTEVAL(proj, c1);
4451 if (tarval_is_single_bit(tv)) {
4453 * optimization for AND:
4455 * And(x, C) == C ==> And(x, C) != 0
4456 * And(x, C) != C ==> And(X, C) == 0
4458 * if C is a single Bit constant.
4461 /* check for Constant's match. We have check hare the tarvals,
4462 because our const might be changed */
4463 if (get_Const_tarval(c1) == tv) {
4464 /* fine: do the transformation */
4465 tv = get_mode_null(get_tarval_mode(tv));
4466 proj_nr ^= pn_Cmp_Leg;
4468 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4474 c1 = get_Or_right(left);
4475 if (is_Const(c1) && tarval_is_null(tv)) {
4477 * Or(x, C) == 0 && C != 0 ==> FALSE
4478 * Or(x, C) != 0 && C != 0 ==> TRUE
4480 if (! tarval_is_null(get_Const_tarval(c1))) {
4481 /* TODO: move to constant evaluation */
4482 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4484 DBG_OPT_CSTEVAL(proj, c1);
4491 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4493 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4496 c1 = get_Shl_right(left);
4498 tarval *tv1 = get_Const_tarval(c1);
4499 ir_mode *mode = get_irn_mode(left);
4500 tarval *minus1 = get_mode_all_one(mode);
4501 tarval *amask = tarval_shr(minus1, tv1);
4502 tarval *cmask = tarval_shl(minus1, tv1);
4505 if (tarval_and(tv, cmask) != tv) {
4506 /* condition not met */
4507 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4509 DBG_OPT_CSTEVAL(proj, c1);
4512 sl = get_Shl_left(left);
4513 blk = get_nodes_block(n);
4514 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4515 tv = tarval_shr(tv, tv1);
4517 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4522 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4524 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4527 c1 = get_Shr_right(left);
4529 tarval *tv1 = get_Const_tarval(c1);
4530 ir_mode *mode = get_irn_mode(left);
4531 tarval *minus1 = get_mode_all_one(mode);
4532 tarval *amask = tarval_shl(minus1, tv1);
4533 tarval *cmask = tarval_shr(minus1, tv1);
4536 if (tarval_and(tv, cmask) != tv) {
4537 /* condition not met */
4538 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4540 DBG_OPT_CSTEVAL(proj, c1);
4543 sl = get_Shr_left(left);
4544 blk = get_nodes_block(n);
4545 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4546 tv = tarval_shl(tv, tv1);
4548 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4553 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4555 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4558 c1 = get_Shrs_right(left);
4560 tarval *tv1 = get_Const_tarval(c1);
4561 ir_mode *mode = get_irn_mode(left);
4562 tarval *minus1 = get_mode_all_one(mode);
4563 tarval *amask = tarval_shl(minus1, tv1);
4564 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4567 cond = tarval_sub(cond, tv1, NULL);
4568 cond = tarval_shrs(tv, cond);
4570 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4571 /* condition not met */
4572 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4574 DBG_OPT_CSTEVAL(proj, c1);
4577 sl = get_Shrs_left(left);
4578 blk = get_nodes_block(n);
4579 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4580 tv = tarval_shl(tv, tv1);
4582 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4587 } /* tarval != bad */
4590 if (changed & 2) /* need a new Const */
4591 right = new_Const(tv);
4593 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4594 ir_node *op = get_Proj_pred(left);
4596 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4597 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4598 ir_node *c = get_binop_right(op);
4601 tarval *tv = get_Const_tarval(c);
4603 if (tarval_is_single_bit(tv)) {
4604 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4605 ir_node *v = get_binop_left(op);
4606 ir_node *blk = get_irn_n(op, -1);
4607 ir_mode *mode = get_irn_mode(v);
4609 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4610 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(tv), mode);
4612 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4619 ir_node *block = get_nodes_block(n);
4621 /* create a new compare */
4622 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4623 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4627 } /* transform_node_Proj_Cmp */
4630 * Optimize CopyB(mem, x, x) into a Nop.
4632 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4633 ir_node *copyb = get_Proj_pred(proj);
4634 ir_node *a = get_CopyB_dst(copyb);
4635 ir_node *b = get_CopyB_src(copyb);
4638 switch (get_Proj_proj(proj)) {
4639 case pn_CopyB_X_regular:
4640 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4641 DBG_OPT_EXC_REM(proj);
4642 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4644 case pn_CopyB_M_except:
4645 case pn_CopyB_X_except:
4646 DBG_OPT_EXC_REM(proj);
4647 proj = get_irg_bad(current_ir_graph);
4654 } /* transform_node_Proj_CopyB */
4657 * Optimize Bounds(idx, idx, upper) into idx.
4659 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4660 ir_node *oldn = proj;
4661 ir_node *bound = get_Proj_pred(proj);
4662 ir_node *idx = get_Bound_index(bound);
4663 ir_node *pred = skip_Proj(idx);
4666 if (idx == get_Bound_lower(bound))
4668 else if (is_Bound(pred)) {
4670 * idx was Bounds checked in the same MacroBlock previously,
4671 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4673 ir_node *lower = get_Bound_lower(bound);
4674 ir_node *upper = get_Bound_upper(bound);
4675 if (get_Bound_lower(pred) == lower &&
4676 get_Bound_upper(pred) == upper &&
4677 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4679 * One could expect that we simply return the previous
4680 * Bound here. However, this would be wrong, as we could
4681 * add an exception Proj to a new location then.
4682 * So, we must turn in into a tuple.
4688 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4689 switch (get_Proj_proj(proj)) {
4691 DBG_OPT_EXC_REM(proj);
4692 proj = get_Bound_mem(bound);
4694 case pn_Bound_X_except:
4695 DBG_OPT_EXC_REM(proj);
4696 proj = get_irg_bad(current_ir_graph);
4700 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4702 case pn_Bound_X_regular:
4703 DBG_OPT_EXC_REM(proj);
4704 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4711 } /* transform_node_Proj_Bound */
4714 * Does all optimizations on nodes that must be done on it's Proj's
4715 * because of creating new nodes.
4717 static ir_node *transform_node_Proj(ir_node *proj) {
4718 ir_node *n = get_Proj_pred(proj);
4720 if (n->op->ops.transform_node_Proj)
4721 return n->op->ops.transform_node_Proj(proj);
4723 } /* transform_node_Proj */
4726 * Move Confirms down through Phi nodes.
4728 static ir_node *transform_node_Phi(ir_node *phi) {
4730 ir_mode *mode = get_irn_mode(phi);
4732 if (mode_is_reference(mode)) {
4733 n = get_irn_arity(phi);
4735 /* Beware of Phi0 */
4737 ir_node *pred = get_irn_n(phi, 0);
4738 ir_node *bound, *new_Phi, *block, **in;
4741 if (! is_Confirm(pred))
4744 bound = get_Confirm_bound(pred);
4745 pnc = get_Confirm_cmp(pred);
4747 NEW_ARR_A(ir_node *, in, n);
4748 in[0] = get_Confirm_value(pred);
4750 for (i = 1; i < n; ++i) {
4751 pred = get_irn_n(phi, i);
4753 if (! is_Confirm(pred) ||
4754 get_Confirm_bound(pred) != bound ||
4755 get_Confirm_cmp(pred) != pnc)
4757 in[i] = get_Confirm_value(pred);
4759 /* move the Confirm nodes "behind" the Phi */
4760 block = get_irn_n(phi, -1);
4761 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4762 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4766 } /* transform_node_Phi */
4769 * Returns the operands of a commutative bin-op, if one operand is
4770 * a const, it is returned as the second one.
4772 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4773 ir_node *op_a = get_binop_left(binop);
4774 ir_node *op_b = get_binop_right(binop);
4776 assert(is_op_commutative(get_irn_op(binop)));
4778 if (is_Const(op_a)) {
4785 } /* get_comm_Binop_Ops */
4788 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4789 * Such pattern may arise in bitfield stores.
4791 * value c4 value c4 & c2
4792 * AND c3 AND c1 | c3
4799 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4802 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4805 ir_node *and_l, *c3;
4806 ir_node *value, *c4;
4807 ir_node *new_and, *new_const, *block;
4808 ir_mode *mode = get_irn_mode(or);
4810 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4813 get_comm_Binop_Ops(or, &and, &c1);
4814 if (!is_Const(c1) || !is_And(and))
4817 get_comm_Binop_Ops(and, &or_l, &c2);
4821 tv1 = get_Const_tarval(c1);
4822 tv2 = get_Const_tarval(c2);
4824 tv = tarval_or(tv1, tv2);
4825 if (tarval_is_all_one(tv)) {
4826 /* the AND does NOT clear a bit with isn't set by the OR */
4827 set_Or_left(or, or_l);
4828 set_Or_right(or, c1);
4830 /* check for more */
4837 get_comm_Binop_Ops(or_l, &and_l, &c3);
4838 if (!is_Const(c3) || !is_And(and_l))
4841 get_comm_Binop_Ops(and_l, &value, &c4);
4845 /* ok, found the pattern, check for conditions */
4846 assert(mode == get_irn_mode(and));
4847 assert(mode == get_irn_mode(or_l));
4848 assert(mode == get_irn_mode(and_l));
4850 tv3 = get_Const_tarval(c3);
4851 tv4 = get_Const_tarval(c4);
4853 tv = tarval_or(tv4, tv2);
4854 if (!tarval_is_all_one(tv)) {
4855 /* have at least one 0 at the same bit position */
4859 n_tv4 = tarval_not(tv4);
4860 if (tv3 != tarval_and(tv3, n_tv4)) {
4861 /* bit in the or_mask is outside the and_mask */
4865 n_tv2 = tarval_not(tv2);
4866 if (tv1 != tarval_and(tv1, n_tv2)) {
4867 /* bit in the or_mask is outside the and_mask */
4871 /* ok, all conditions met */
4872 block = get_irn_n(or, -1);
4874 new_and = new_r_And(current_ir_graph, block,
4875 value, new_Const(tarval_and(tv4, tv2)), mode);
4877 new_const = new_Const(tarval_or(tv3, tv1));
4879 set_Or_left(or, new_and);
4880 set_Or_right(or, new_const);
4882 /* check for more */
4884 } /* transform_node_Or_bf_store */
4887 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4889 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4890 ir_mode *mode = get_irn_mode(or);
4891 ir_node *shl, *shr, *block;
4892 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4895 if (! mode_is_int(mode))
4898 shl = get_binop_left(or);
4899 shr = get_binop_right(or);
4908 } else if (!is_Shl(shl)) {
4910 } else if (!is_Shr(shr)) {
4913 x = get_Shl_left(shl);
4914 if (x != get_Shr_left(shr))
4917 c1 = get_Shl_right(shl);
4918 c2 = get_Shr_right(shr);
4919 if (is_Const(c1) && is_Const(c2)) {
4920 tv1 = get_Const_tarval(c1);
4921 if (! tarval_is_long(tv1))
4924 tv2 = get_Const_tarval(c2);
4925 if (! tarval_is_long(tv2))
4928 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4929 != (int) get_mode_size_bits(mode))
4932 /* yet, condition met */
4933 block = get_nodes_block(or);
4935 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4937 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4944 rotval = sub; /* a Rot right is not supported, so use a rot left */
4945 } else if (is_Sub(c2)) {
4951 if (get_Sub_right(sub) != v)
4954 c1 = get_Sub_left(sub);
4958 tv1 = get_Const_tarval(c1);
4959 if (! tarval_is_long(tv1))
4962 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4965 /* yet, condition met */
4966 block = get_nodes_block(or);
4968 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4970 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4972 } /* transform_node_Or_Rotl */
4977 static ir_node *transform_node_Or(ir_node *n) {
4978 ir_node *c, *oldn = n;
4979 ir_node *a = get_Or_left(n);
4980 ir_node *b = get_Or_right(n);
4983 if (is_Not(a) && is_Not(b)) {
4984 /* ~a | ~b = ~(a&b) */
4985 ir_node *block = get_nodes_block(n);
4987 mode = get_irn_mode(n);
4990 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4991 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4992 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4996 /* we can evaluate 2 Projs of the same Cmp */
4997 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4998 ir_node *pred_a = get_Proj_pred(a);
4999 ir_node *pred_b = get_Proj_pred(b);
5000 if (pred_a == pred_b) {
5001 dbg_info *dbgi = get_irn_dbg_info(n);
5002 ir_node *block = get_nodes_block(pred_a);
5003 pn_Cmp pn_a = get_Proj_proj(a);
5004 pn_Cmp pn_b = get_Proj_proj(b);
5005 /* yes, we can simply calculate with pncs */
5006 pn_Cmp new_pnc = pn_a | pn_b;
5008 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
5013 mode = get_irn_mode(n);
5014 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
5016 n = transform_node_Or_bf_store(n);
5017 n = transform_node_Or_Rotl(n);
5021 n = transform_bitwise_distributive(n, transform_node_Or);
5024 } /* transform_node_Or */
5028 static ir_node *transform_node(ir_node *n);
5031 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5033 * Should be moved to reassociation?
5035 static ir_node *transform_node_shift(ir_node *n) {
5036 ir_node *left, *right;
5038 tarval *tv1, *tv2, *res;
5039 ir_node *in[2], *irn, *block;
5041 left = get_binop_left(n);
5043 /* different operations */
5044 if (get_irn_op(left) != get_irn_op(n))
5047 right = get_binop_right(n);
5048 tv1 = value_of(right);
5049 if (tv1 == tarval_bad)
5052 tv2 = value_of(get_binop_right(left));
5053 if (tv2 == tarval_bad)
5056 res = tarval_add(tv1, tv2);
5057 mode = get_irn_mode(n);
5059 /* beware: a simple replacement works only, if res < modulo shift */
5061 int modulo_shf = get_mode_modulo_shift(mode);
5062 if (modulo_shf > 0) {
5063 tarval *modulo = new_tarval_from_long(modulo_shf,
5064 get_tarval_mode(res));
5066 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5068 /* shifting too much */
5069 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5071 ir_graph *irg = get_irn_irg(n);
5072 ir_node *block = get_nodes_block(n);
5073 dbg_info *dbgi = get_irn_dbg_info(n);
5074 ir_mode *smode = get_irn_mode(right);
5075 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5076 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5080 return new_Const(get_mode_null(mode));
5084 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5087 /* ok, we can replace it */
5088 block = get_nodes_block(n);
5090 in[0] = get_binop_left(left);
5091 in[1] = new_Const(res);
5093 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5095 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5097 return transform_node(irn);
5098 } /* transform_node_shift */
5101 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5103 * - and, or, xor instead of &
5104 * - Shl, Shr, Shrs, rotl instead of >>
5105 * (with a special case for Or/Xor + Shrs)
5107 static ir_node *transform_node_bitop_shift(ir_node *n) {
5109 ir_node *right = get_binop_right(n);
5110 ir_mode *mode = get_irn_mode(n);
5111 ir_node *bitop_left;
5112 ir_node *bitop_right;
5124 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5126 if (!is_Const(right))
5129 left = get_binop_left(n);
5130 op_left = get_irn_op(left);
5131 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5134 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5135 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5136 /* TODO: test if sign bit is affectes */
5140 bitop_right = get_binop_right(left);
5141 if (!is_Const(bitop_right))
5144 bitop_left = get_binop_left(left);
5146 irg = get_irn_irg(n);
5147 block = get_nodes_block(n);
5148 dbgi = get_irn_dbg_info(n);
5149 tv1 = get_Const_tarval(bitop_right);
5150 tv2 = get_Const_tarval(right);
5152 assert(get_tarval_mode(tv1) == mode);
5155 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5156 tv_shift = tarval_shl(tv1, tv2);
5157 } else if(is_Shr(n)) {
5158 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5159 tv_shift = tarval_shr(tv1, tv2);
5160 } else if(is_Shrs(n)) {
5161 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5162 tv_shift = tarval_shrs(tv1, tv2);
5165 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5166 tv_shift = tarval_rotl(tv1, tv2);
5169 assert(get_tarval_mode(tv_shift) == mode);
5170 new_const = new_Const(tv_shift);
5172 if (op_left == op_And) {
5173 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5174 } else if(op_left == op_Or) {
5175 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5177 assert(op_left == op_Eor);
5178 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5186 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5188 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5189 * (also with x >>s c1 when c1>=c2)
5191 static ir_node *transform_node_shl_shr(ir_node *n) {
5193 ir_node *right = get_binop_right(n);
5209 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5211 if (!is_Const(right))
5214 left = get_binop_left(n);
5215 mode = get_irn_mode(n);
5216 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5217 ir_node *shr_right = get_binop_right(left);
5219 if (!is_Const(shr_right))
5222 x = get_binop_left(left);
5223 tv_shr = get_Const_tarval(shr_right);
5224 tv_shl = get_Const_tarval(right);
5226 if (is_Shrs(left)) {
5227 /* shrs variant only allowed if c1 >= c2 */
5228 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5231 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5234 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5236 tv_mask = tarval_shl(tv_mask, tv_shl);
5237 } else if(is_Shr(n) && is_Shl(left)) {
5238 ir_node *shl_right = get_Shl_right(left);
5240 if (!is_Const(shl_right))
5243 x = get_Shl_left(left);
5244 tv_shr = get_Const_tarval(right);
5245 tv_shl = get_Const_tarval(shl_right);
5247 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5248 tv_mask = tarval_shr(tv_mask, tv_shr);
5253 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5254 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5257 assert(tv_mask != tarval_bad);
5258 assert(get_tarval_mode(tv_mask) == mode);
5260 irg = get_irn_irg(n);
5261 block = get_nodes_block(n);
5262 dbgi = get_irn_dbg_info(n);
5264 pnc = tarval_cmp(tv_shl, tv_shr);
5265 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5266 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5267 new_const = new_Const(tv_shift);
5269 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5271 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5274 assert(pnc == pn_Cmp_Gt);
5275 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5276 new_const = new_Const(tv_shift);
5277 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5280 new_const = new_Const(tv_mask);
5281 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5289 static ir_node *transform_node_Shr(ir_node *n) {
5290 ir_node *c, *oldn = n;
5291 ir_node *left = get_Shr_left(n);
5292 ir_node *right = get_Shr_right(n);
5293 ir_mode *mode = get_irn_mode(n);
5295 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5296 n = transform_node_shift(n);
5299 n = transform_node_shl_shr(n);
5301 n = transform_node_bitop_shift(n);
5304 } /* transform_node_Shr */
5309 static ir_node *transform_node_Shrs(ir_node *n) {
5310 ir_node *c, *oldn = n;
5311 ir_node *a = get_Shrs_left(n);
5312 ir_node *b = get_Shrs_right(n);
5313 ir_mode *mode = get_irn_mode(n);
5315 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5316 n = transform_node_shift(n);
5319 n = transform_node_bitop_shift(n);
5322 } /* transform_node_Shrs */
5327 static ir_node *transform_node_Shl(ir_node *n) {
5328 ir_node *c, *oldn = n;
5329 ir_node *a = get_Shl_left(n);
5330 ir_node *b = get_Shl_right(n);
5331 ir_mode *mode = get_irn_mode(n);
5333 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5334 n = transform_node_shift(n);
5337 n = transform_node_shl_shr(n);
5339 n = transform_node_bitop_shift(n);
5342 } /* transform_node_Shl */
5347 static ir_node *transform_node_Rotl(ir_node *n) {
5348 ir_node *c, *oldn = n;
5349 ir_node *a = get_Rotl_left(n);
5350 ir_node *b = get_Rotl_right(n);
5351 ir_mode *mode = get_irn_mode(n);
5353 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5354 n = transform_node_shift(n);
5357 n = transform_node_bitop_shift(n);
5360 } /* transform_node_Rotl */
5365 static ir_node *transform_node_Conv(ir_node *n) {
5366 ir_node *c, *oldn = n;
5367 ir_mode *mode = get_irn_mode(n);
5368 ir_node *a = get_Conv_op(n);
5370 if (mode != mode_b && is_const_Phi(a)) {
5371 /* Do NOT optimize mode_b Conv's, this leads to remaining
5372 * Phib nodes later, because the conv_b_lower operation
5373 * is instantly reverted, when it tries to insert a Convb.
5375 c = apply_conv_on_phi(a, mode);
5377 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5382 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5383 return new_r_Unknown(current_ir_graph, mode);
5386 if (mode_is_reference(mode) &&
5387 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5389 ir_node *l = get_Add_left(a);
5390 ir_node *r = get_Add_right(a);
5391 dbg_info *dbgi = get_irn_dbg_info(a);
5392 ir_node *block = get_nodes_block(n);
5394 ir_node *lop = get_Conv_op(l);
5395 if(get_irn_mode(lop) == mode) {
5396 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5397 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5402 ir_node *rop = get_Conv_op(r);
5403 if(get_irn_mode(rop) == mode) {
5404 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5405 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5412 } /* transform_node_Conv */
5415 * Remove dead blocks and nodes in dead blocks
5416 * in keep alive list. We do not generate a new End node.
5418 static ir_node *transform_node_End(ir_node *n) {
5419 int i, j, n_keepalives = get_End_n_keepalives(n);
5422 NEW_ARR_A(ir_node *, in, n_keepalives);
5424 for (i = j = 0; i < n_keepalives; ++i) {
5425 ir_node *ka = get_End_keepalive(n, i);
5427 if (! is_Block_dead(ka)) {
5431 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5433 } else if (is_Bad(ka)) {
5434 /* no need to keep Bad */
5439 if (j != n_keepalives)
5440 set_End_keepalives(n, j, in);
5442 } /* transform_node_End */
5444 /** returns 1 if a == -b */
5445 static int is_negated_value(ir_node *a, ir_node *b) {
5446 if (is_Minus(a) && get_Minus_op(a) == b)
5448 if (is_Minus(b) && get_Minus_op(b) == a)
5450 if (is_Sub(a) && is_Sub(b)) {
5451 ir_node *a_left = get_Sub_left(a);
5452 ir_node *a_right = get_Sub_right(a);
5453 ir_node *b_left = get_Sub_left(b);
5454 ir_node *b_right = get_Sub_right(b);
5456 if (a_left == b_right && a_right == b_left)
5464 * Optimize a Mux into some simpler cases.
5466 static ir_node *transform_node_Mux(ir_node *n) {
5467 ir_node *oldn = n, *sel = get_Mux_sel(n);
5468 ir_mode *mode = get_irn_mode(n);
5469 ir_node *t = get_Mux_true(n);
5470 ir_node *f = get_Mux_false(n);
5471 ir_graph *irg = current_ir_graph;
5473 /* first normalization step: move a possible zero to the false case */
5475 ir_node *cmp = get_Proj_pred(sel);
5478 if (is_Const(t) && is_Const_null(t)) {
5481 /* Mux(x, 0, y) => Mux(x, y, 0) */
5482 pn_Cmp pnc = get_Proj_proj(sel);
5483 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5484 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5485 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5493 /* note: after normalization, false can only happen on default */
5494 if (mode == mode_b) {
5495 dbg_info *dbg = get_irn_dbg_info(n);
5496 ir_node *block = get_nodes_block(n);
5497 ir_graph *irg = current_ir_graph;
5500 tarval *tv_t = get_Const_tarval(t);
5501 if (tv_t == tarval_b_true) {
5503 /* Muxb(sel, true, false) = sel */
5504 assert(get_Const_tarval(f) == tarval_b_false);
5505 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5508 /* Muxb(sel, true, x) = Or(sel, x) */
5509 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5510 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5514 } else if (is_Const(f)) {
5515 tarval *tv_f = get_Const_tarval(f);
5516 if (tv_f == tarval_b_true) {
5517 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5518 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5519 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5520 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5523 /* Muxb(sel, x, false) = And(sel, x) */
5524 assert(tv_f == tarval_b_false);
5525 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5526 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5532 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5533 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5534 tarval *a = get_Const_tarval(t);
5535 tarval *b = get_Const_tarval(f);
5536 tarval *null = get_tarval_null(mode);
5539 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5540 diff = tarval_sub(a, b, NULL);
5543 diff = tarval_sub(b, a, NULL);
5547 if (diff == get_tarval_one(mode) && min != null) {
5548 dbg_info *dbg = get_irn_dbg_info(n);
5549 ir_node *block = get_nodes_block(n);
5550 ir_graph *irg = current_ir_graph;
5551 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5552 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5553 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5554 n = new_rd_Add(dbg, irg, block, n, new_Const(min), mode);
5560 ir_node *cmp = get_Proj_pred(sel);
5561 long pn = get_Proj_proj(sel);
5564 * Note: normalization puts the constant on the right side,
5565 * so we check only one case.
5567 * Note further that these optimization work even for floating point
5568 * with NaN's because -NaN == NaN.
5569 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5573 ir_node *cmp_r = get_Cmp_right(cmp);
5574 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5575 ir_node *block = get_nodes_block(n);
5576 ir_node *cmp_l = get_Cmp_left(cmp);
5578 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5581 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5582 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5584 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5585 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5587 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5589 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5590 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5592 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5593 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5595 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5597 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5602 if (mode_is_int(mode)) {
5604 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5605 /* Mux((a & b) != 0, c, 0) */
5606 ir_node *and_r = get_And_right(cmp_l);
5609 if (and_r == t && f == cmp_r) {
5610 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5611 if (pn == pn_Cmp_Lg) {
5612 /* Mux((a & 2^C) != 0, 2^C, 0) */
5614 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5616 /* Mux((a & 2^C) == 0, 2^C, 0) */
5617 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5618 block, cmp_l, t, mode);
5619 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5624 if (is_Shl(and_r)) {
5625 ir_node *shl_l = get_Shl_left(and_r);
5626 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5627 if (and_r == t && f == cmp_r) {
5628 if (pn == pn_Cmp_Lg) {
5629 /* (a & (1 << n)) != 0, (1 << n), 0) */
5631 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5633 /* (a & (1 << n)) == 0, (1 << n), 0) */
5634 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5635 block, cmp_l, t, mode);
5636 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5642 and_l = get_And_left(cmp_l);
5643 if (is_Shl(and_l)) {
5644 ir_node *shl_l = get_Shl_left(and_l);
5645 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5646 if (and_l == t && f == cmp_r) {
5647 if (pn == pn_Cmp_Lg) {
5648 /* ((1 << n) & a) != 0, (1 << n), 0) */
5650 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5652 /* ((1 << n) & a) == 0, (1 << n), 0) */
5653 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5654 block, cmp_l, t, mode);
5655 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5668 } /* transform_node_Mux */
5671 * optimize Sync nodes that have other syncs as input we simply add the inputs
5672 * of the other sync to our own inputs
5674 static ir_node *transform_node_Sync(ir_node *n) {
5675 int arity = get_Sync_n_preds(n);
5678 for (i = 0; i < arity;) {
5679 ir_node *pred = get_Sync_pred(n, i);
5683 if (!is_Sync(pred)) {
5691 pred_arity = get_Sync_n_preds(pred);
5692 for (j = 0; j < pred_arity; ++j) {
5693 ir_node *pred_pred = get_Sync_pred(pred, j);
5698 add_irn_n(n, pred_pred);
5702 if (get_Sync_pred(n, k) == pred_pred) break;
5707 /* rehash the sync node */
5708 add_identities(current_ir_graph->value_table, n);
5714 * Tries several [inplace] [optimizing] transformations and returns an
5715 * equivalent node. The difference to equivalent_node() is that these
5716 * transformations _do_ generate new nodes, and thus the old node must
5717 * not be freed even if the equivalent node isn't the old one.
5719 static ir_node *transform_node(ir_node *n) {
5723 * Transform_node is the only "optimizing transformation" that might
5724 * return a node with a different opcode. We iterate HERE until fixpoint
5725 * to get the final result.
5729 if (n->op->ops.transform_node)
5730 n = n->op->ops.transform_node(n);
5731 } while (oldn != n);
5734 } /* transform_node */
5737 * Sets the default transform node operation for an ir_op_ops.
5739 * @param code the opcode for the default operation
5740 * @param ops the operations initialized
5745 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5749 ops->transform_node = transform_node_##a; \
5751 #define CASE_PROJ(a) \
5753 ops->transform_node_Proj = transform_node_Proj_##a; \
5755 #define CASE_PROJ_EX(a) \
5757 ops->transform_node = transform_node_##a; \
5758 ops->transform_node_Proj = transform_node_Proj_##a; \
5767 CASE_PROJ_EX(DivMod);
5801 } /* firm_set_default_transform_node */
5804 /* **************** Common Subexpression Elimination **************** */
5806 /** The size of the hash table used, should estimate the number of nodes
5808 #define N_IR_NODES 512
5810 /** Compares the attributes of two Const nodes. */
5811 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5812 return (get_Const_tarval(a) != get_Const_tarval(b))
5813 || (get_Const_type(a) != get_Const_type(b));
5814 } /* node_cmp_attr_Const */
5816 /** Compares the attributes of two Proj nodes. */
5817 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5818 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5819 } /* node_cmp_attr_Proj */
5821 /** Compares the attributes of two Filter nodes. */
5822 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5823 return get_Filter_proj(a) != get_Filter_proj(b);
5824 } /* node_cmp_attr_Filter */
5826 /** Compares the attributes of two Alloc nodes. */
5827 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5828 const alloc_attr *pa = get_irn_alloc_attr(a);
5829 const alloc_attr *pb = get_irn_alloc_attr(b);
5830 return (pa->where != pb->where) || (pa->type != pb->type);
5831 } /* node_cmp_attr_Alloc */
5833 /** Compares the attributes of two Free nodes. */
5834 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5835 const free_attr *pa = get_irn_free_attr(a);
5836 const free_attr *pb = get_irn_free_attr(b);
5837 return (pa->where != pb->where) || (pa->type != pb->type);
5838 } /* node_cmp_attr_Free */
5840 /** Compares the attributes of two SymConst nodes. */
5841 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5842 const symconst_attr *pa = get_irn_symconst_attr(a);
5843 const symconst_attr *pb = get_irn_symconst_attr(b);
5844 return (pa->kind != pb->kind)
5845 || (pa->sym.type_p != pb->sym.type_p)
5846 || (pa->tp != pb->tp);
5847 } /* node_cmp_attr_SymConst */
5849 /** Compares the attributes of two Call nodes. */
5850 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5851 return get_irn_call_attr(a) != get_irn_call_attr(b);
5852 } /* node_cmp_attr_Call */
5854 /** Compares the attributes of two Sel nodes. */
5855 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5856 const ir_entity *a_ent = get_Sel_entity(a);
5857 const ir_entity *b_ent = get_Sel_entity(b);
5860 (a_ent->kind != b_ent->kind) ||
5861 (a_ent->name != b_ent->name) ||
5862 (a_ent->owner != b_ent->owner) ||
5863 (a_ent->ld_name != b_ent->ld_name) ||
5864 (a_ent->type != b_ent->type);
5866 /* Matze: inlining of functions can produce 2 entities with same type,
5868 return a_ent != b_ent;
5869 } /* node_cmp_attr_Sel */
5871 /** Compares the attributes of two Phi nodes. */
5872 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5873 /* we can only enter this function if both nodes have the same number of inputs,
5874 hence it is enough to check if one of them is a Phi0 */
5876 /* check the Phi0 pos attribute */
5877 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5880 } /* node_cmp_attr_Phi */
5882 /** Compares the attributes of two Conv nodes. */
5883 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5884 return get_Conv_strict(a) != get_Conv_strict(b);
5885 } /* node_cmp_attr_Conv */
5887 /** Compares the attributes of two Cast nodes. */
5888 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5889 return get_Cast_type(a) != get_Cast_type(b);
5890 } /* node_cmp_attr_Cast */
5892 /** Compares the attributes of two Load nodes. */
5893 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5894 if (get_Load_volatility(a) == volatility_is_volatile ||
5895 get_Load_volatility(b) == volatility_is_volatile)
5896 /* NEVER do CSE on volatile Loads */
5898 /* do not CSE Loads with different alignment. Be conservative. */
5899 if (get_Load_align(a) != get_Load_align(b))
5902 return get_Load_mode(a) != get_Load_mode(b);
5903 } /* node_cmp_attr_Load */
5905 /** Compares the attributes of two Store nodes. */
5906 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5907 /* do not CSE Stores with different alignment. Be conservative. */
5908 if (get_Store_align(a) != get_Store_align(b))
5911 /* NEVER do CSE on volatile Stores */
5912 return (get_Store_volatility(a) == volatility_is_volatile ||
5913 get_Store_volatility(b) == volatility_is_volatile);
5914 } /* node_cmp_attr_Store */
5916 /** Compares two exception attributes */
5917 static int node_cmp_exception(ir_node *a, ir_node *b) {
5918 const except_attr *ea = get_irn_except_attr(a);
5919 const except_attr *eb = get_irn_except_attr(b);
5921 return ea->pin_state != eb->pin_state;
5924 #define node_cmp_attr_Bound node_cmp_exception
5926 /** Compares the attributes of two Div nodes. */
5927 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5928 const divmod_attr *ma = get_irn_divmod_attr(a);
5929 const divmod_attr *mb = get_irn_divmod_attr(b);
5930 return ma->exc.pin_state != mb->exc.pin_state ||
5931 ma->res_mode != mb->res_mode ||
5932 ma->no_remainder != mb->no_remainder;
5933 } /* node_cmp_attr_Div */
5935 /** Compares the attributes of two DivMod nodes. */
5936 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5937 const divmod_attr *ma = get_irn_divmod_attr(a);
5938 const divmod_attr *mb = get_irn_divmod_attr(b);
5939 return ma->exc.pin_state != mb->exc.pin_state ||
5940 ma->res_mode != mb->res_mode;
5941 } /* node_cmp_attr_DivMod */
5943 /** Compares the attributes of two Mod nodes. */
5944 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5945 const divmod_attr *ma = get_irn_divmod_attr(a);
5946 const divmod_attr *mb = get_irn_divmod_attr(b);
5947 return ma->exc.pin_state != mb->exc.pin_state ||
5948 ma->res_mode != mb->res_mode;
5949 } /* node_cmp_attr_Mod */
5951 /** Compares the attributes of two Quot nodes. */
5952 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5953 const divmod_attr *ma = get_irn_divmod_attr(a);
5954 const divmod_attr *mb = get_irn_divmod_attr(b);
5955 return ma->exc.pin_state != mb->exc.pin_state ||
5956 ma->res_mode != mb->res_mode;
5957 } /* node_cmp_attr_Quot */
5959 /** Compares the attributes of two Confirm nodes. */
5960 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5961 /* no need to compare the bound, as this is a input */
5962 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5963 } /* node_cmp_attr_Confirm */
5965 /** Compares the attributes of two Builtin nodes. */
5966 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b) {
5967 const builtin_attr *ma = get_irn_builtin_attr(a);
5968 const builtin_attr *mb = get_irn_builtin_attr(b);
5970 /* no need to compare the type, equal kind means equal type */
5971 return ma->kind != mb->kind;
5972 } /* node_cmp_attr_Builtin */
5974 /** Compares the attributes of two ASM nodes. */
5975 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5977 const ir_asm_constraint *ca;
5978 const ir_asm_constraint *cb;
5981 if (get_ASM_text(a) != get_ASM_text(b))
5984 /* Should we really check the constraints here? Should be better, but is strange. */
5985 n = get_ASM_n_input_constraints(a);
5986 if (n != get_ASM_n_input_constraints(b))
5989 ca = get_ASM_input_constraints(a);
5990 cb = get_ASM_input_constraints(b);
5991 for (i = 0; i < n; ++i) {
5992 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5996 n = get_ASM_n_output_constraints(a);
5997 if (n != get_ASM_n_output_constraints(b))
6000 ca = get_ASM_output_constraints(a);
6001 cb = get_ASM_output_constraints(b);
6002 for (i = 0; i < n; ++i) {
6003 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6007 n = get_ASM_n_clobbers(a);
6008 if (n != get_ASM_n_clobbers(b))
6011 cla = get_ASM_clobbers(a);
6012 clb = get_ASM_clobbers(b);
6013 for (i = 0; i < n; ++i) {
6014 if (cla[i] != clb[i])
6018 } /* node_cmp_attr_ASM */
6021 * Set the default node attribute compare operation for an ir_op_ops.
6023 * @param code the opcode for the default operation
6024 * @param ops the operations initialized
6029 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6033 ops->node_cmp_attr = node_cmp_attr_##a; \
6065 } /* firm_set_default_node_cmp_attr */
6068 * Compare function for two nodes in the value table. Gets two
6069 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6071 int identities_cmp(const void *elt, const void *key) {
6072 ir_node *a = (ir_node *)elt;
6073 ir_node *b = (ir_node *)key;
6076 if (a == b) return 0;
6078 if ((get_irn_op(a) != get_irn_op(b)) ||
6079 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6081 /* compare if a's in and b's in are of equal length */
6082 irn_arity_a = get_irn_intra_arity(a);
6083 if (irn_arity_a != get_irn_intra_arity(b))
6086 if (get_irn_pinned(a) == op_pin_state_pinned) {
6087 /* for pinned nodes, the block inputs must be equal */
6088 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6090 } else if (! get_opt_global_cse()) {
6091 /* for block-local CSE both nodes must be in the same MacroBlock */
6092 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6096 /* compare a->in[0..ins] with b->in[0..ins] */
6097 for (i = 0; i < irn_arity_a; i++)
6098 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6102 * here, we already now that the nodes are identical except their
6105 if (a->op->ops.node_cmp_attr)
6106 return a->op->ops.node_cmp_attr(a, b);
6109 } /* identities_cmp */
6112 * Calculate a hash value of a node.
6114 * @param node The IR-node
6116 unsigned ir_node_hash(const ir_node *node) {
6117 return node->op->ops.hash(node);
6118 } /* ir_node_hash */
6121 pset *new_identities(void) {
6122 return new_pset(identities_cmp, N_IR_NODES);
6123 } /* new_identities */
6125 void del_identities(pset *value_table) {
6126 del_pset(value_table);
6127 } /* del_identities */
6129 /* Normalize a node by putting constants (and operands with larger
6130 * node index) on the right (operator side). */
6131 void ir_normalize_node(ir_node *n) {
6132 if (is_op_commutative(get_irn_op(n))) {
6133 ir_node *l = get_binop_left(n);
6134 ir_node *r = get_binop_right(n);
6136 /* For commutative operators perform a OP b == b OP a but keep
6137 * constants on the RIGHT side. This helps greatly in some
6138 * optimizations. Moreover we use the idx number to make the form
6140 if (!operands_are_normalized(l, r)) {
6141 set_binop_left(n, r);
6142 set_binop_right(n, l);
6146 } /* ir_normalize_node */
6149 * Update the nodes after a match in the value table. If both nodes have
6150 * the same MacroBlock but different Blocks, we must ensure that the node
6151 * with the dominating Block (the node that is near to the MacroBlock header
6152 * is stored in the table.
6153 * Because a MacroBlock has only one "non-exception" flow, we don't need
6154 * dominance info here: We known, that one block must dominate the other and
6155 * following the only block input will allow to find it.
6157 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6158 ir_node *known_blk, *new_block, *block, *mbh;
6160 if (get_opt_global_cse()) {
6161 /* Block inputs are meaning less */
6164 known_blk = get_irn_n(known_irn, -1);
6165 new_block = get_irn_n(new_ir_node, -1);
6166 if (known_blk == new_block) {
6167 /* already in the same block */
6171 * We expect the typical case when we built the graph. In that case, the
6172 * known_irn is already the upper one, so checking this should be faster.
6175 mbh = get_Block_MacroBlock(new_block);
6177 if (block == known_blk) {
6178 /* ok, we have found it: known_block dominates new_block as expected */
6183 * We have reached the MacroBlock header NOT founding
6184 * the known_block. new_block must dominate known_block.
6187 set_irn_n(known_irn, -1, new_block);
6190 assert(get_Block_n_cfgpreds(block) == 1);
6191 block = get_Block_cfgpred_block(block, 0);
6193 } /* update_value_table */
6196 * Return the canonical node computing the same value as n.
6197 * Looks up the node in a hash table, enters it in the table
6198 * if it isn't there yet.
6200 * @param value_table the HashSet containing all nodes in the
6202 * @param n the node to look up
6204 * @return a node that computes the same value as n or n if no such
6205 * node could be found
6207 ir_node *identify_remember(pset *value_table, ir_node *n) {
6210 if (!value_table) return n;
6212 ir_normalize_node(n);
6213 /* lookup or insert in hash table with given hash key. */
6214 o = pset_insert(value_table, n, ir_node_hash(n));
6217 update_known_irn(o, n);
6221 } /* identify_remember */
6224 * During construction we set the op_pin_state_pinned flag in the graph right when the
6225 * optimization is performed. The flag turning on procedure global cse could
6226 * be changed between two allocations. This way we are safe.
6228 * @param value_table The value table
6229 * @param n The node to lookup
6231 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6234 n = identify_remember(value_table, n);
6235 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6236 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6238 } /* identify_cons */
6240 /* Add a node to the identities value table. */
6241 void add_identities(pset *value_table, ir_node *node) {
6242 if (get_opt_cse() && is_no_Block(node))
6243 identify_remember(value_table, node);
6244 } /* add_identities */
6246 /* Visit each node in the value table of a graph. */
6247 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6249 ir_graph *rem = current_ir_graph;
6251 current_ir_graph = irg;
6252 foreach_pset(irg->value_table, node)
6254 current_ir_graph = rem;
6255 } /* visit_all_identities */
6258 * Garbage in, garbage out. If a node has a dead input, i.e., the
6259 * Bad node is input to the node, return the Bad node.
6261 static ir_node *gigo(ir_node *node) {
6263 ir_op *op = get_irn_op(node);
6265 /* remove garbage blocks by looking at control flow that leaves the block
6266 and replacing the control flow by Bad. */
6267 if (get_irn_mode(node) == mode_X) {
6268 ir_node *block = get_nodes_block(skip_Proj(node));
6270 /* Don't optimize nodes in immature blocks. */
6271 if (!get_Block_matured(block))
6273 /* Don't optimize End, may have Bads. */
6274 if (op == op_End) return node;
6276 if (is_Block(block)) {
6277 if (is_Block_dead(block)) {
6278 /* control flow from dead block is dead */
6282 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6283 if (!is_Bad(get_irn_n(block, i)))
6287 ir_graph *irg = get_irn_irg(block);
6288 /* the start block is never dead */
6289 if (block != get_irg_start_block(irg)
6290 && block != get_irg_end_block(irg)) {
6292 * Do NOT kill control flow without setting
6293 * the block to dead of bad things can happen:
6294 * We get a Block that is not reachable be irg_block_walk()
6295 * but can be found by irg_walk()!
6297 set_Block_dead(block);
6304 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6305 blocks predecessors is dead. */
6306 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6307 irn_arity = get_irn_arity(node);
6310 * Beware: we can only read the block of a non-floating node.
6312 if (is_irn_pinned_in_irg(node) &&
6313 is_Block_dead(get_nodes_block(skip_Proj(node))))
6316 for (i = 0; i < irn_arity; i++) {
6317 ir_node *pred = get_irn_n(node, i);
6322 /* Propagating Unknowns here seems to be a bad idea, because
6323 sometimes we need a node as a input and did not want that
6325 However, it might be useful to move this into a later phase
6326 (if you think that optimizing such code is useful). */
6327 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6328 return new_Unknown(get_irn_mode(node));
6333 /* With this code we violate the agreement that local_optimize
6334 only leaves Bads in Block, Phi and Tuple nodes. */
6335 /* If Block has only Bads as predecessors it's garbage. */
6336 /* If Phi has only Bads as predecessors it's garbage. */
6337 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6338 irn_arity = get_irn_arity(node);
6339 for (i = 0; i < irn_arity; i++) {
6340 if (!is_Bad(get_irn_n(node, i))) break;
6342 if (i == irn_arity) node = new_Bad();
6349 * These optimizations deallocate nodes from the obstack.
6350 * It can only be called if it is guaranteed that no other nodes
6351 * reference this one, i.e., right after construction of a node.
6353 * @param n The node to optimize
6355 * current_ir_graph must be set to the graph of the node!
6357 ir_node *optimize_node(ir_node *n) {
6360 ir_opcode iro = get_irn_opcode(n);
6362 /* Always optimize Phi nodes: part of the construction. */
6363 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6365 /* constant expression evaluation / constant folding */
6366 if (get_opt_constant_folding()) {
6367 /* neither constants nor Tuple values can be evaluated */
6368 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6369 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6370 int old_fp_mode = tarval_fp_ops_enabled();
6372 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6374 /* try to evaluate */
6375 tv = computed_value(n);
6376 if (tv != tarval_bad) {
6378 ir_type *old_tp = get_irn_type(n);
6379 int i, arity = get_irn_arity(n);
6383 * Try to recover the type of the new expression.
6385 for (i = 0; i < arity && !old_tp; ++i)
6386 old_tp = get_irn_type(get_irn_n(n, i));
6389 * we MUST copy the node here temporary, because it's still needed
6390 * for DBG_OPT_CSTEVAL
6392 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6393 oldn = alloca(node_size);
6395 memcpy(oldn, n, node_size);
6396 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6398 /* ARG, copy the in array, we need it for statistics */
6399 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6401 /* note the inplace edges module */
6402 edges_node_deleted(n, current_ir_graph);
6404 /* evaluation was successful -- replace the node. */
6405 irg_kill_node(current_ir_graph, n);
6408 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6409 set_Const_type(nw, old_tp);
6410 DBG_OPT_CSTEVAL(oldn, nw);
6411 tarval_enable_fp_ops(old_fp_mode);
6414 tarval_enable_fp_ops(old_fp_mode);
6418 /* remove unnecessary nodes */
6419 if (get_opt_algebraic_simplification() ||
6420 (iro == iro_Phi) || /* always optimize these nodes. */
6422 (iro == iro_Proj) ||
6423 (iro == iro_Block) ) /* Flags tested local. */
6424 n = equivalent_node(n);
6426 /* Common Subexpression Elimination.
6428 * Checks whether n is already available.
6429 * The block input is used to distinguish different subexpressions. Right
6430 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6431 * subexpressions within a block.
6434 n = identify_cons(current_ir_graph->value_table, n);
6437 edges_node_deleted(oldn, current_ir_graph);
6439 /* We found an existing, better node, so we can deallocate the old node. */
6440 irg_kill_node(current_ir_graph, oldn);
6444 /* Some more constant expression evaluation that does not allow to
6446 iro = get_irn_opcode(n);
6447 if (get_opt_algebraic_simplification() ||
6448 (iro == iro_Cond) ||
6449 (iro == iro_Proj)) /* Flags tested local. */
6450 n = transform_node(n);
6452 /* Remove nodes with dead (Bad) input.
6453 Run always for transformation induced Bads. */
6456 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6457 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6459 n = identify_remember(current_ir_graph->value_table, o);
6465 } /* optimize_node */
6469 * These optimizations never deallocate nodes (in place). This can cause dead
6470 * nodes lying on the obstack. Remove these by a dead node elimination,
6471 * i.e., a copying garbage collection.
6473 ir_node *optimize_in_place_2(ir_node *n) {
6476 ir_opcode iro = get_irn_opcode(n);
6478 if (!get_opt_optimize() && !is_Phi(n)) return n;
6480 /* constant expression evaluation / constant folding */
6481 if (get_opt_constant_folding()) {
6482 /* neither constants nor Tuple values can be evaluated */
6483 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6484 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6485 int old_fp_mode = tarval_fp_ops_enabled();
6487 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6488 /* try to evaluate */
6489 tv = computed_value(n);
6490 if (tv != tarval_bad) {
6491 /* evaluation was successful -- replace the node. */
6492 ir_type *old_tp = get_irn_type(n);
6493 int i, arity = get_irn_arity(n);
6496 * Try to recover the type of the new expression.
6498 for (i = 0; i < arity && !old_tp; ++i)
6499 old_tp = get_irn_type(get_irn_n(n, i));
6503 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6504 set_Const_type(n, old_tp);
6506 DBG_OPT_CSTEVAL(oldn, n);
6507 tarval_enable_fp_ops(old_fp_mode);
6510 tarval_enable_fp_ops(old_fp_mode);
6514 /* remove unnecessary nodes */
6515 if (get_opt_constant_folding() ||
6516 (iro == iro_Phi) || /* always optimize these nodes. */
6517 (iro == iro_Id) || /* ... */
6518 (iro == iro_Proj) || /* ... */
6519 (iro == iro_Block) ) /* Flags tested local. */
6520 n = equivalent_node(n);
6522 /** common subexpression elimination **/
6523 /* Checks whether n is already available. */
6524 /* The block input is used to distinguish different subexpressions. Right
6525 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6526 subexpressions within a block. */
6527 if (get_opt_cse()) {
6529 n = identify_remember(current_ir_graph->value_table, o);
6534 /* Some more constant expression evaluation. */
6535 iro = get_irn_opcode(n);
6536 if (get_opt_constant_folding() ||
6537 (iro == iro_Cond) ||
6538 (iro == iro_Proj)) /* Flags tested local. */
6539 n = transform_node(n);
6541 /* Remove nodes with dead (Bad) input.
6542 Run always for transformation induced Bads. */
6545 /* Now we can verify the node, as it has no dead inputs any more. */
6548 /* Now we have a legal, useful node. Enter it in hash table for cse.
6549 Blocks should be unique anyways. (Except the successor of start:
6550 is cse with the start block!) */
6551 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6553 n = identify_remember(current_ir_graph->value_table, o);
6559 } /* optimize_in_place_2 */
6562 * Wrapper for external use, set proper status bits after optimization.
6564 ir_node *optimize_in_place(ir_node *n) {
6565 /* Handle graph state */
6566 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6568 if (get_opt_global_cse())
6569 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6570 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6571 set_irg_outs_inconsistent(current_ir_graph);
6573 /* FIXME: Maybe we could also test whether optimizing the node can
6574 change the control graph. */
6575 set_irg_doms_inconsistent(current_ir_graph);
6576 return optimize_in_place_2(n);
6577 } /* optimize_in_place */
6580 * Calculate a hash value of a Const node.
6582 static unsigned hash_Const(const ir_node *node) {
6585 /* special value for const, as they only differ in their tarval. */
6586 h = HASH_PTR(node->attr.con.tv);
6587 h = 9*h + HASH_PTR(get_irn_mode(node));
6593 * Calculate a hash value of a SymConst node.
6595 static unsigned hash_SymConst(const ir_node *node) {
6598 /* special value for const, as they only differ in their symbol. */
6599 h = HASH_PTR(node->attr.symc.sym.type_p);
6600 h = 9*h + HASH_PTR(get_irn_mode(node));
6603 } /* hash_SymConst */
6606 * Set the default hash operation in an ir_op_ops.
6608 * @param code the opcode for the default operation
6609 * @param ops the operations initialized
6614 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6618 ops->hash = hash_##a; \
6621 /* hash function already set */
6622 if (ops->hash != NULL)
6629 /* use input/mode default hash if no function was given */
6630 ops->hash = firm_default_hash;
6638 * Sets the default operation for an ir_ops.
6640 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6641 ops = firm_set_default_hash(code, ops);
6642 ops = firm_set_default_computed_value(code, ops);
6643 ops = firm_set_default_equivalent_node(code, ops);
6644 ops = firm_set_default_transform_node(code, ops);
6645 ops = firm_set_default_node_cmp_attr(code, ops);
6646 ops = firm_set_default_get_type(code, ops);
6647 ops = firm_set_default_get_type_attr(code, ops);
6648 ops = firm_set_default_get_entity_attr(code, ops);
6651 } /* firm_set_default_operations */