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 (jump threading 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_entity(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_X_except:
1600 DBG_OPT_EXC_REM(proj);
1601 proj = get_irg_bad(current_ir_graph);
1606 } /* equivalent_node_Proj_CopyB */
1609 * Optimize Bounds(idx, idx, upper) into idx.
1611 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1612 ir_node *oldn = proj;
1613 ir_node *bound = get_Proj_pred(proj);
1614 ir_node *idx = get_Bound_index(bound);
1615 ir_node *pred = skip_Proj(idx);
1618 if (idx == get_Bound_lower(bound))
1620 else if (is_Bound(pred)) {
1622 * idx was Bounds checked in the same MacroBlock previously,
1623 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1625 ir_node *lower = get_Bound_lower(bound);
1626 ir_node *upper = get_Bound_upper(bound);
1627 if (get_Bound_lower(pred) == lower &&
1628 get_Bound_upper(pred) == upper &&
1629 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1631 * One could expect that we simply return the previous
1632 * Bound here. However, this would be wrong, as we could
1633 * add an exception Proj to a new location then.
1634 * So, we must turn in into a tuple.
1640 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1641 switch (get_Proj_proj(proj)) {
1643 DBG_OPT_EXC_REM(proj);
1644 proj = get_Bound_mem(bound);
1646 case pn_Bound_X_except:
1647 DBG_OPT_EXC_REM(proj);
1648 proj = get_irg_bad(current_ir_graph);
1652 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1655 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1660 } /* equivalent_node_Proj_Bound */
1663 * Optimize an Exception Proj(Load) with a non-null address.
1665 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1666 if (get_opt_ldst_only_null_ptr_exceptions()) {
1667 if (get_irn_mode(proj) == mode_X) {
1668 ir_node *load = get_Proj_pred(proj);
1670 /* get the Load address */
1671 const ir_node *addr = get_Load_ptr(load);
1672 const ir_node *confirm;
1674 if (value_not_null(addr, &confirm)) {
1675 if (get_Proj_proj(proj) == pn_Load_X_except) {
1676 DBG_OPT_EXC_REM(proj);
1677 return get_irg_bad(current_ir_graph);
1683 } /* equivalent_node_Proj_Load */
1686 * Optimize an Exception Proj(Store) with a non-null address.
1688 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1689 if (get_opt_ldst_only_null_ptr_exceptions()) {
1690 if (get_irn_mode(proj) == mode_X) {
1691 ir_node *store = get_Proj_pred(proj);
1693 /* get the load/store address */
1694 const ir_node *addr = get_Store_ptr(store);
1695 const ir_node *confirm;
1697 if (value_not_null(addr, &confirm)) {
1698 if (get_Proj_proj(proj) == pn_Store_X_except) {
1699 DBG_OPT_EXC_REM(proj);
1700 return get_irg_bad(current_ir_graph);
1706 } /* equivalent_node_Proj_Store */
1709 * Does all optimizations on nodes that must be done on it's Proj's
1710 * because of creating new nodes.
1712 static ir_node *equivalent_node_Proj(ir_node *proj) {
1713 ir_node *n = get_Proj_pred(proj);
1715 if (get_irn_mode(proj) == mode_X) {
1716 if (is_Block_dead(get_nodes_block(n))) {
1717 /* Remove dead control flow -- early gigo(). */
1718 return get_irg_bad(current_ir_graph);
1721 if (n->op->ops.equivalent_node_Proj)
1722 return n->op->ops.equivalent_node_Proj(proj);
1724 } /* equivalent_node_Proj */
1729 static ir_node *equivalent_node_Id(ir_node *n) {
1736 DBG_OPT_ID(oldn, n);
1738 } /* equivalent_node_Id */
1743 static ir_node *equivalent_node_Mux(ir_node *n)
1745 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);
1755 /* Mux(false, f, t) == f */
1756 if (ts == tarval_b_false) {
1757 n = get_Mux_false(n);
1758 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1761 n_t = get_Mux_true(n);
1762 n_f = get_Mux_false(n);
1764 /* Mux(v, x, T) == x */
1765 if (is_Unknown(n_f)) {
1767 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1770 /* Mux(v, T, x) == x */
1771 if (is_Unknown(n_t)) {
1773 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1777 /* Mux(v, x, x) == x */
1780 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1783 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1784 ir_node *cmp = get_Proj_pred(sel);
1785 long proj_nr = get_Proj_proj(sel);
1786 ir_node *f = get_Mux_false(n);
1787 ir_node *t = get_Mux_true(n);
1790 * Note further that these optimization work even for floating point
1791 * with NaN's because -NaN == NaN.
1792 * However, if +0 and -0 is handled differently, we cannot use the first one.
1795 ir_node *const cmp_l = get_Cmp_left(cmp);
1796 ir_node *const cmp_r = get_Cmp_right(cmp);
1800 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1801 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1803 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1810 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1811 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1813 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1820 * Note: normalization puts the constant on the right side,
1821 * so we check only one case.
1823 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1824 /* Mux(t CMP 0, X, t) */
1825 if (is_Minus(f) && get_Minus_op(f) == t) {
1826 /* Mux(t CMP 0, -t, t) */
1827 if (proj_nr == pn_Cmp_Eq) {
1828 /* Mux(t == 0, -t, t) ==> -t */
1830 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1831 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1832 /* Mux(t != 0, -t, t) ==> t */
1834 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1841 } /* equivalent_node_Mux */
1844 * Remove Confirm nodes if setting is on.
1845 * Replace Confirms(x, '=', Constlike) by Constlike.
1847 static ir_node *equivalent_node_Confirm(ir_node *n) {
1848 ir_node *pred = get_Confirm_value(n);
1849 pn_Cmp pnc = get_Confirm_cmp(n);
1851 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1853 * rare case: two identical Confirms one after another,
1854 * replace the second one with the first.
1857 pred = get_Confirm_value(n);
1858 pnc = get_Confirm_cmp(n);
1860 if (get_opt_remove_confirm())
1861 return get_Confirm_value(n);
1866 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1867 * perform no actual computation, as, e.g., the Id nodes. It does not create
1868 * new nodes. It is therefore safe to free n if the node returned is not n.
1869 * If a node returns a Tuple we can not just skip it. If the size of the
1870 * in array fits, we transform n into a tuple (e.g., Div).
1872 ir_node *equivalent_node(ir_node *n) {
1873 if (n->op->ops.equivalent_node)
1874 return n->op->ops.equivalent_node(n);
1876 } /* equivalent_node */
1879 * Sets the default equivalent node operation for an ir_op_ops.
1881 * @param code the opcode for the default operation
1882 * @param ops the operations initialized
1887 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1891 ops->equivalent_node = equivalent_node_##a; \
1893 #define CASE_PROJ(a) \
1895 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1937 } /* firm_set_default_equivalent_node */
1940 * Returns non-zero if a node is a Phi node
1941 * with all predecessors constant.
1943 static int is_const_Phi(ir_node *n) {
1946 if (! is_Phi(n) || get_irn_arity(n) == 0)
1948 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1949 if (! is_Const(get_irn_n(n, i)))
1953 } /* is_const_Phi */
1955 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1956 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1959 * in reality eval_func should be tarval (*eval_func)() but incomplete
1960 * declarations are bad style and generate noisy warnings
1962 typedef void (*eval_func)(void);
1965 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1967 static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode)
1969 if (eval == (eval_func) tarval_sub) {
1970 tarval_sub_type func = (tarval_sub_type)eval;
1972 return func(a, b, mode);
1974 tarval_binop_type func = (tarval_binop_type)eval;
1981 * Apply an evaluator on a binop with a constant operators (and one Phi).
1983 * @param phi the Phi node
1984 * @param other the other operand
1985 * @param eval an evaluator function
1986 * @param mode the mode of the result, may be different from the mode of the Phi!
1987 * @param left if non-zero, other is the left operand, else the right
1989 * @return a new Phi node if the conversion was successful, NULL else
1991 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, ir_mode *mode, int left) {
1996 int i, n = get_irn_arity(phi);
1998 NEW_ARR_A(void *, res, n);
2000 for (i = 0; i < n; ++i) {
2001 pred = get_irn_n(phi, i);
2002 tv = get_Const_tarval(pred);
2003 tv = do_eval(eval, other, tv, mode);
2005 if (tv == tarval_bad) {
2006 /* folding failed, bad */
2012 for (i = 0; i < n; ++i) {
2013 pred = get_irn_n(phi, i);
2014 tv = get_Const_tarval(pred);
2015 tv = do_eval(eval, tv, other, mode);
2017 if (tv == tarval_bad) {
2018 /* folding failed, bad */
2024 irg = current_ir_graph;
2025 for (i = 0; i < n; ++i) {
2026 pred = get_irn_n(phi, i);
2027 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2029 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2030 } /* apply_binop_on_phi */
2033 * Apply an evaluator on a binop with two constant Phi.
2035 * @param a the left Phi node
2036 * @param b the right Phi node
2037 * @param eval an evaluator function
2038 * @param mode the mode of the result, may be different from the mode of the Phi!
2040 * @return a new Phi node if the conversion was successful, NULL else
2042 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode) {
2043 tarval *tv_l, *tv_r, *tv;
2049 if (get_nodes_block(a) != get_nodes_block(b))
2052 n = get_irn_arity(a);
2053 NEW_ARR_A(void *, res, n);
2055 for (i = 0; i < n; ++i) {
2056 pred = get_irn_n(a, i);
2057 tv_l = get_Const_tarval(pred);
2058 pred = get_irn_n(b, i);
2059 tv_r = get_Const_tarval(pred);
2060 tv = do_eval(eval, tv_l, tv_r, mode);
2062 if (tv == tarval_bad) {
2063 /* folding failed, bad */
2068 irg = current_ir_graph;
2069 for (i = 0; i < n; ++i) {
2070 pred = get_irn_n(a, i);
2071 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2073 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
2074 } /* apply_binop_on_2_phis */
2077 * Apply an evaluator on a unop with a constant operator (a Phi).
2079 * @param phi the Phi node
2080 * @param eval an evaluator function
2082 * @return a new Phi node if the conversion was successful, NULL else
2084 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2090 int i, n = get_irn_arity(phi);
2092 NEW_ARR_A(void *, res, n);
2093 for (i = 0; i < n; ++i) {
2094 pred = get_irn_n(phi, i);
2095 tv = get_Const_tarval(pred);
2098 if (tv == tarval_bad) {
2099 /* folding failed, bad */
2104 mode = get_irn_mode(phi);
2105 irg = current_ir_graph;
2106 for (i = 0; i < n; ++i) {
2107 pred = get_irn_n(phi, i);
2108 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2110 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2111 } /* apply_unop_on_phi */
2114 * Apply a conversion on a constant operator (a Phi).
2116 * @param phi the Phi node
2118 * @return a new Phi node if the conversion was successful, NULL else
2120 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2125 int i, n = get_irn_arity(phi);
2127 NEW_ARR_A(void *, res, n);
2128 for (i = 0; i < n; ++i) {
2129 pred = get_irn_n(phi, i);
2130 tv = get_Const_tarval(pred);
2131 tv = tarval_convert_to(tv, mode);
2133 if (tv == tarval_bad) {
2134 /* folding failed, bad */
2139 irg = current_ir_graph;
2140 for (i = 0; i < n; ++i) {
2141 pred = get_irn_n(phi, i);
2142 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2144 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2145 } /* apply_conv_on_phi */
2148 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2149 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2150 * If possible, remove the Conv's.
2152 static ir_node *transform_node_AddSub(ir_node *n) {
2153 ir_mode *mode = get_irn_mode(n);
2155 if (mode_is_reference(mode)) {
2156 ir_node *left = get_binop_left(n);
2157 ir_node *right = get_binop_right(n);
2158 unsigned ref_bits = get_mode_size_bits(mode);
2160 if (is_Conv(left)) {
2161 ir_mode *lmode = get_irn_mode(left);
2162 unsigned bits = get_mode_size_bits(lmode);
2164 if (ref_bits == bits &&
2165 mode_is_int(lmode) &&
2166 get_mode_arithmetic(lmode) == irma_twos_complement) {
2167 ir_node *pre = get_Conv_op(left);
2168 ir_mode *pre_mode = get_irn_mode(pre);
2170 if (mode_is_int(pre_mode) &&
2171 get_mode_size_bits(pre_mode) == bits &&
2172 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2173 /* ok, this conv just changes to sign, moreover the calculation
2174 * is done with same number of bits as our address mode, so
2175 * we can ignore the conv as address calculation can be viewed
2176 * as either signed or unsigned
2178 set_binop_left(n, pre);
2183 if (is_Conv(right)) {
2184 ir_mode *rmode = get_irn_mode(right);
2185 unsigned bits = get_mode_size_bits(rmode);
2187 if (ref_bits == bits &&
2188 mode_is_int(rmode) &&
2189 get_mode_arithmetic(rmode) == irma_twos_complement) {
2190 ir_node *pre = get_Conv_op(right);
2191 ir_mode *pre_mode = get_irn_mode(pre);
2193 if (mode_is_int(pre_mode) &&
2194 get_mode_size_bits(pre_mode) == bits &&
2195 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2196 /* ok, this conv just changes to sign, moreover the calculation
2197 * is done with same number of bits as our address mode, so
2198 * we can ignore the conv as address calculation can be viewed
2199 * as either signed or unsigned
2201 set_binop_right(n, pre);
2206 /* let address arithmetic use unsigned modes */
2207 if (is_Const(right)) {
2208 ir_mode *rmode = get_irn_mode(right);
2210 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2211 /* convert a AddP(P, *s) into AddP(P, *u) */
2212 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2214 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2215 set_binop_right(n, pre);
2221 } /* transform_node_AddSub */
2223 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2225 if (is_Const(b) && is_const_Phi(a)) { \
2226 /* check for Op(Phi, Const) */ \
2227 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2229 else if (is_Const(a) && is_const_Phi(b)) { \
2230 /* check for Op(Const, Phi) */ \
2231 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2233 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2234 /* check for Op(Phi, Phi) */ \
2235 c = apply_binop_on_2_phis(a, b, eval, mode); \
2238 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2242 #define HANDLE_UNOP_PHI(eval, a, c) \
2244 if (is_const_Phi(a)) { \
2245 /* check for Op(Phi) */ \
2246 c = apply_unop_on_phi(a, eval); \
2248 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2254 * Do the AddSub optimization, then Transform
2255 * Constant folding on Phi
2256 * Add(a,a) -> Mul(a, 2)
2257 * Add(Mul(a, x), a) -> Mul(a, x+1)
2258 * if the mode is integer or float.
2259 * Transform Add(a,-b) into Sub(a,b).
2260 * Reassociation might fold this further.
2262 static ir_node *transform_node_Add(ir_node *n) {
2264 ir_node *a, *b, *c, *oldn = n;
2266 n = transform_node_AddSub(n);
2268 a = get_Add_left(n);
2269 b = get_Add_right(n);
2271 mode = get_irn_mode(n);
2273 if (mode_is_reference(mode)) {
2274 ir_mode *lmode = get_irn_mode(a);
2276 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2277 /* an Add(a, NULL) is a hidden Conv */
2278 dbg_info *dbg = get_irn_dbg_info(n);
2279 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2283 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2285 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2286 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2289 if (mode_is_num(mode)) {
2290 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2291 if (!is_irg_state(current_ir_graph, IR_GRAPH_STATE_ARCH_DEP)
2292 && a == b && mode_is_int(mode)) {
2293 ir_node *block = get_nodes_block(n);
2296 get_irn_dbg_info(n),
2299 new_Const_long(mode, 2),
2301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2306 get_irn_dbg_info(n),
2311 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2316 get_irn_dbg_info(n),
2321 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2324 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2325 /* Here we rely on constants be on the RIGHT side */
2327 ir_node *op = get_Not_op(a);
2329 if (is_Const(b) && is_Const_one(b)) {
2331 ir_node *blk = get_nodes_block(n);
2332 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2333 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2338 n = new_Const(get_mode_minus_one(mode));
2339 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2344 ir_node *op = get_Not_op(b);
2348 n = new_Const(get_mode_minus_one(mode));
2349 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2356 } /* transform_node_Add */
2359 * returns -cnst or NULL if impossible
2361 static ir_node *const_negate(ir_node *cnst) {
2362 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2363 dbg_info *dbgi = get_irn_dbg_info(cnst);
2364 ir_graph *irg = get_irn_irg(cnst);
2365 if (tv == tarval_bad) return NULL;
2366 return new_rd_Const(dbgi, irg, tv);
2370 * Do the AddSub optimization, then Transform
2371 * Constant folding on Phi
2372 * Sub(0,a) -> Minus(a)
2373 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2374 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2375 * Sub(Add(a, x), x) -> a
2376 * Sub(x, Add(x, a)) -> -a
2377 * Sub(x, Const) -> Add(x, -Const)
2379 static ir_node *transform_node_Sub(ir_node *n) {
2384 n = transform_node_AddSub(n);
2386 a = get_Sub_left(n);
2387 b = get_Sub_right(n);
2389 mode = get_irn_mode(n);
2391 if (mode_is_int(mode)) {
2392 ir_mode *lmode = get_irn_mode(a);
2394 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2395 /* a Sub(a, NULL) is a hidden Conv */
2396 dbg_info *dbg = get_irn_dbg_info(n);
2397 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2398 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2402 if (mode == lmode &&
2403 get_mode_arithmetic(mode) == irma_twos_complement &&
2405 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2407 dbg_info *dbg = get_irn_dbg_info(n);
2408 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2409 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2415 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2417 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2418 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2421 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2422 /* a - C -> a + (-C) */
2423 ir_node *cnst = const_negate(b);
2425 ir_node *block = get_nodes_block(n);
2426 dbg_info *dbgi = get_irn_dbg_info(n);
2428 n = new_rd_Add(dbgi, block, a, cnst, mode);
2429 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2434 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2435 dbg_info *dbg = get_irn_dbg_info(n);
2436 ir_node *block = get_nodes_block(n);
2437 ir_node *left = get_Minus_op(a);
2438 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2440 n = new_rd_Minus(dbg, block, add, mode);
2441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2443 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2444 dbg_info *dbg = get_irn_dbg_info(n);
2445 ir_node *block = get_nodes_block(n);
2446 ir_node *right = get_Minus_op(b);
2448 n = new_rd_Add(dbg, block, a, right, mode);
2449 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2451 } else if (is_Sub(b)) {
2452 /* a - (b - c) -> a + (c - b)
2453 * -> (a - b) + c iff (b - c) is a pointer */
2454 dbg_info *s_dbg = get_irn_dbg_info(b);
2455 ir_node *s_block = get_nodes_block(b);
2456 ir_node *s_left = get_Sub_left(b);
2457 ir_node *s_right = get_Sub_right(b);
2458 ir_mode *s_mode = get_irn_mode(b);
2459 if (mode_is_reference(s_mode)) {
2460 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2461 dbg_info *a_dbg = get_irn_dbg_info(n);
2462 ir_node *a_block = get_nodes_block(n);
2465 s_right = new_r_Conv(a_block, s_right, mode);
2466 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2468 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2469 dbg_info *a_dbg = get_irn_dbg_info(n);
2470 ir_node *a_block = get_nodes_block(n);
2472 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2474 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2476 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2477 ir_node *m_right = get_Mul_right(b);
2478 if (is_Const(m_right)) {
2479 ir_node *cnst2 = const_negate(m_right);
2480 if (cnst2 != NULL) {
2481 dbg_info *m_dbg = get_irn_dbg_info(b);
2482 ir_node *m_block = get_nodes_block(b);
2483 ir_node *m_left = get_Mul_left(b);
2484 ir_mode *m_mode = get_irn_mode(b);
2485 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2486 dbg_info *a_dbg = get_irn_dbg_info(n);
2487 ir_node *a_block = get_nodes_block(n);
2489 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2490 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2496 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2497 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2499 get_irn_dbg_info(n),
2503 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2507 if (mode_wrap_around(mode)) {
2508 ir_node *left = get_Add_left(a);
2509 ir_node *right = get_Add_right(a);
2511 /* FIXME: Does the Conv's work only for two complement or generally? */
2513 if (mode != get_irn_mode(right)) {
2514 /* This Sub is an effective Cast */
2515 right = new_r_Conv(get_nodes_block(n), right, mode);
2518 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2520 } else if (right == b) {
2521 if (mode != get_irn_mode(left)) {
2522 /* This Sub is an effective Cast */
2523 left = new_r_Conv(get_nodes_block(n), left, mode);
2526 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2532 if (mode_wrap_around(mode)) {
2533 ir_node *left = get_Add_left(b);
2534 ir_node *right = get_Add_right(b);
2536 /* FIXME: Does the Conv's work only for two complement or generally? */
2538 ir_mode *r_mode = get_irn_mode(right);
2540 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2541 if (mode != r_mode) {
2542 /* This Sub is an effective Cast */
2543 n = new_r_Conv(get_nodes_block(n), n, mode);
2545 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2547 } else if (right == a) {
2548 ir_mode *l_mode = get_irn_mode(left);
2550 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2551 if (mode != l_mode) {
2552 /* This Sub is an effective Cast */
2553 n = new_r_Conv(get_nodes_block(n), n, mode);
2555 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2560 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2561 ir_mode *mode = get_irn_mode(a);
2563 if (mode == get_irn_mode(b)) {
2565 ir_node *op_a = get_Conv_op(a);
2566 ir_node *op_b = get_Conv_op(b);
2568 /* check if it's allowed to skip the conv */
2569 ma = get_irn_mode(op_a);
2570 mb = get_irn_mode(op_b);
2572 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2573 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2576 set_Sub_right(n, b);
2582 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2583 if (!is_reassoc_running() && is_Mul(a)) {
2584 ir_node *ma = get_Mul_left(a);
2585 ir_node *mb = get_Mul_right(a);
2588 ir_node *blk = get_nodes_block(n);
2590 get_irn_dbg_info(n),
2594 get_irn_dbg_info(n),
2597 new_Const_long(mode, 1),
2600 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2602 } else if (mb == b) {
2603 ir_node *blk = get_nodes_block(n);
2605 get_irn_dbg_info(n),
2609 get_irn_dbg_info(n),
2612 new_Const_long(mode, 1),
2615 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2619 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2620 ir_node *x = get_Sub_left(a);
2621 ir_node *y = get_Sub_right(a);
2622 ir_node *blk = get_nodes_block(n);
2623 ir_mode *m_b = get_irn_mode(b);
2624 ir_mode *m_y = get_irn_mode(y);
2628 /* Determine the right mode for the Add. */
2631 else if (mode_is_reference(m_b))
2633 else if (mode_is_reference(m_y))
2637 * Both modes are different but none is reference,
2638 * happens for instance in SubP(SubP(P, Iu), Is).
2639 * We have two possibilities here: Cast or ignore.
2640 * Currently we ignore this case.
2645 add = new_r_Add(blk, y, b, add_mode);
2647 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2648 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2652 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2653 if (is_Const(a) && is_Not(b)) {
2654 /* c - ~X = X + (c+1) */
2655 tarval *tv = get_Const_tarval(a);
2657 tv = tarval_add(tv, get_mode_one(mode));
2658 if (tv != tarval_bad) {
2659 ir_node *blk = get_nodes_block(n);
2660 ir_node *c = new_Const(tv);
2661 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2662 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2668 } /* transform_node_Sub */
2671 * Several transformation done on n*n=2n bits mul.
2672 * These transformations must be done here because new nodes may be produced.
2674 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2676 ir_node *a = get_Mul_left(n);
2677 ir_node *b = get_Mul_right(n);
2678 tarval *ta = value_of(a);
2679 tarval *tb = value_of(b);
2680 ir_mode *smode = get_irn_mode(a);
2682 if (ta == get_mode_one(smode)) {
2683 /* (L)1 * (L)b = (L)b */
2684 ir_node *blk = get_nodes_block(n);
2685 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2686 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2689 else if (ta == get_mode_minus_one(smode)) {
2690 /* (L)-1 * (L)b = (L)b */
2691 ir_node *blk = get_nodes_block(n);
2692 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2693 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2694 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2697 if (tb == get_mode_one(smode)) {
2698 /* (L)a * (L)1 = (L)a */
2699 ir_node *blk = get_irn_n(a, -1);
2700 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2701 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2704 else if (tb == get_mode_minus_one(smode)) {
2705 /* (L)a * (L)-1 = (L)-a */
2706 ir_node *blk = get_nodes_block(n);
2707 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2708 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2709 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2716 * Transform Mul(a,-1) into -a.
2717 * Do constant evaluation of Phi nodes.
2718 * Do architecture dependent optimizations on Mul nodes
2720 static ir_node *transform_node_Mul(ir_node *n) {
2721 ir_node *c, *oldn = n;
2722 ir_mode *mode = get_irn_mode(n);
2723 ir_node *a = get_Mul_left(n);
2724 ir_node *b = get_Mul_right(n);
2726 if (is_Bad(a) || is_Bad(b))
2729 if (mode != get_irn_mode(a))
2730 return transform_node_Mul2n(n, mode);
2732 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2734 if (mode_is_signed(mode)) {
2737 if (value_of(a) == get_mode_minus_one(mode))
2739 else if (value_of(b) == get_mode_minus_one(mode))
2742 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2743 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2748 if (is_Const(b)) { /* (-a) * const -> a * -const */
2749 ir_node *cnst = const_negate(b);
2751 dbg_info *dbgi = get_irn_dbg_info(n);
2752 ir_node *block = get_nodes_block(n);
2753 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2754 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2757 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2758 dbg_info *dbgi = get_irn_dbg_info(n);
2759 ir_node *block = get_nodes_block(n);
2760 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2761 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2763 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2764 ir_node *sub_l = get_Sub_left(b);
2765 ir_node *sub_r = get_Sub_right(b);
2766 dbg_info *dbgi = get_irn_dbg_info(n);
2767 ir_node *block = get_nodes_block(n);
2768 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2769 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2770 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2773 } else if (is_Minus(b)) {
2774 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2775 ir_node *sub_l = get_Sub_left(a);
2776 ir_node *sub_r = get_Sub_right(a);
2777 dbg_info *dbgi = get_irn_dbg_info(n);
2778 ir_node *block = get_nodes_block(n);
2779 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2780 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2781 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2784 } else if (is_Shl(a)) {
2785 ir_node *const shl_l = get_Shl_left(a);
2786 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2787 /* (1 << x) * b -> b << x */
2788 dbg_info *const dbgi = get_irn_dbg_info(n);
2789 ir_node *const block = get_nodes_block(n);
2790 ir_node *const shl_r = get_Shl_right(a);
2791 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2792 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2795 } else if (is_Shl(b)) {
2796 ir_node *const shl_l = get_Shl_left(b);
2797 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2798 /* a * (1 << x) -> a << x */
2799 dbg_info *const dbgi = get_irn_dbg_info(n);
2800 ir_node *const block = get_nodes_block(n);
2801 ir_node *const shl_r = get_Shl_right(b);
2802 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2803 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2807 if (get_mode_arithmetic(mode) == irma_ieee754) {
2809 tarval *tv = get_Const_tarval(a);
2810 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2811 && !tarval_is_negative(tv)) {
2812 /* 2.0 * b = b + b */
2813 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2814 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2818 else if (is_Const(b)) {
2819 tarval *tv = get_Const_tarval(b);
2820 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2821 && !tarval_is_negative(tv)) {
2822 /* a * 2.0 = a + a */
2823 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2824 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2829 return arch_dep_replace_mul_with_shifts(n);
2830 } /* transform_node_Mul */
2833 * Transform a Div Node.
2835 static ir_node *transform_node_Div(ir_node *n) {
2836 ir_mode *mode = get_Div_resmode(n);
2837 ir_node *a = get_Div_left(n);
2838 ir_node *b = get_Div_right(n);
2840 const ir_node *dummy;
2842 if (is_Const(b) && is_const_Phi(a)) {
2843 /* check for Div(Phi, Const) */
2844 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2846 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2850 else if (is_Const(a) && is_const_Phi(b)) {
2851 /* check for Div(Const, Phi) */
2852 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2854 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2858 else if (is_const_Phi(a) && is_const_Phi(b)) {
2859 /* check for Div(Phi, Phi) */
2860 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2862 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2869 if (a == b && value_not_zero(a, &dummy)) {
2870 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2871 value = new_Const(get_mode_one(mode));
2872 DBG_OPT_CSTEVAL(n, value);
2875 if (mode_is_signed(mode) && is_Const(b)) {
2876 tarval *tv = get_Const_tarval(b);
2878 if (tv == get_mode_minus_one(mode)) {
2880 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2881 DBG_OPT_CSTEVAL(n, value);
2885 /* Try architecture dependent optimization */
2886 value = arch_dep_replace_div_by_const(n);
2893 /* Turn Div into a tuple (mem, jmp, bad, value) */
2894 mem = get_Div_mem(n);
2895 blk = get_nodes_block(n);
2897 /* skip a potential Pin */
2898 mem = skip_Pin(mem);
2899 turn_into_tuple(n, pn_Div_max);
2900 set_Tuple_pred(n, pn_Div_M, mem);
2901 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2902 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2903 set_Tuple_pred(n, pn_Div_res, value);
2906 } /* transform_node_Div */
2909 * Transform a Mod node.
2911 static ir_node *transform_node_Mod(ir_node *n) {
2912 ir_mode *mode = get_Mod_resmode(n);
2913 ir_node *a = get_Mod_left(n);
2914 ir_node *b = get_Mod_right(n);
2918 if (is_Const(b) && is_const_Phi(a)) {
2919 /* check for Div(Phi, Const) */
2920 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2922 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2926 else if (is_Const(a) && is_const_Phi(b)) {
2927 /* check for Div(Const, Phi) */
2928 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2930 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2934 else if (is_const_Phi(a) && is_const_Phi(b)) {
2935 /* check for Div(Phi, Phi) */
2936 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2938 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2945 if (tv != tarval_bad) {
2946 value = new_Const(tv);
2948 DBG_OPT_CSTEVAL(n, value);
2951 ir_node *a = get_Mod_left(n);
2952 ir_node *b = get_Mod_right(n);
2953 const ir_node *dummy;
2955 if (a == b && value_not_zero(a, &dummy)) {
2956 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2957 value = new_Const(get_mode_null(mode));
2958 DBG_OPT_CSTEVAL(n, value);
2961 if (mode_is_signed(mode) && is_Const(b)) {
2962 tarval *tv = get_Const_tarval(b);
2964 if (tv == get_mode_minus_one(mode)) {
2966 value = new_Const(get_mode_null(mode));
2967 DBG_OPT_CSTEVAL(n, value);
2971 /* Try architecture dependent optimization */
2972 value = arch_dep_replace_mod_by_const(n);
2980 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2981 mem = get_Mod_mem(n);
2982 blk = get_nodes_block(n);
2984 /* skip a potential Pin */
2985 mem = skip_Pin(mem);
2986 turn_into_tuple(n, pn_Mod_max);
2987 set_Tuple_pred(n, pn_Mod_M, mem);
2988 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2989 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2990 set_Tuple_pred(n, pn_Mod_res, value);
2993 } /* transform_node_Mod */
2996 * Transform a DivMod node.
2998 static ir_node *transform_node_DivMod(ir_node *n) {
2999 const ir_node *dummy;
3000 ir_node *a = get_DivMod_left(n);
3001 ir_node *b = get_DivMod_right(n);
3002 ir_mode *mode = get_DivMod_resmode(n);
3007 if (is_Const(b) && is_const_Phi(a)) {
3008 /* check for Div(Phi, Const) */
3009 va = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3010 vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3012 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3013 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3017 else if (is_Const(a) && is_const_Phi(b)) {
3018 /* check for Div(Const, Phi) */
3019 va = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3020 vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3022 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3023 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3027 else if (is_const_Phi(a) && is_const_Phi(b)) {
3028 /* check for Div(Phi, Phi) */
3029 va = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3030 vb = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3032 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3033 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3040 if (tb != tarval_bad) {
3041 if (tb == get_mode_one(get_tarval_mode(tb))) {
3043 vb = new_Const(get_mode_null(mode));
3044 DBG_OPT_CSTEVAL(n, vb);
3046 } else if (ta != tarval_bad) {
3047 tarval *resa, *resb;
3048 resa = tarval_div(ta, tb);
3049 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3050 Jmp for X result!? */
3051 resb = tarval_mod(ta, tb);
3052 if (resb == tarval_bad) return n; /* Causes exception! */
3053 va = new_Const(resa);
3054 vb = new_Const(resb);
3055 DBG_OPT_CSTEVAL(n, va);
3056 DBG_OPT_CSTEVAL(n, vb);
3058 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3059 va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3060 vb = new_Const(get_mode_null(mode));
3061 DBG_OPT_CSTEVAL(n, va);
3062 DBG_OPT_CSTEVAL(n, vb);
3064 } else { /* Try architecture dependent optimization */
3067 arch_dep_replace_divmod_by_const(&va, &vb, n);
3068 evaluated = va != NULL;
3070 } else if (a == b) {
3071 if (value_not_zero(a, &dummy)) {
3073 va = new_Const(get_mode_one(mode));
3074 vb = new_Const(get_mode_null(mode));
3075 DBG_OPT_CSTEVAL(n, va);
3076 DBG_OPT_CSTEVAL(n, vb);
3079 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3082 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3083 /* 0 / non-Const = 0 */
3088 if (evaluated) { /* replace by tuple */
3092 mem = get_DivMod_mem(n);
3093 /* skip a potential Pin */
3094 mem = skip_Pin(mem);
3096 blk = get_nodes_block(n);
3097 turn_into_tuple(n, pn_DivMod_max);
3098 set_Tuple_pred(n, pn_DivMod_M, mem);
3099 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(blk));
3100 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3101 set_Tuple_pred(n, pn_DivMod_res_div, va);
3102 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3106 } /* transform_node_DivMod */
3109 * Optimize x / c to x * (1/c)
3111 static ir_node *transform_node_Quot(ir_node *n) {
3112 ir_mode *mode = get_Quot_resmode(n);
3115 if (get_mode_arithmetic(mode) == irma_ieee754) {
3116 ir_node *b = get_Quot_right(n);
3117 tarval *tv = value_of(b);
3119 if (tv != tarval_bad) {
3120 int rem = tarval_fp_ops_enabled();
3123 * Floating point constant folding might be disabled here to
3125 * However, as we check for exact result, doing it is safe.
3128 tarval_enable_fp_ops(1);
3129 tv = tarval_quo(get_mode_one(mode), tv);
3130 tarval_enable_fp_ops(rem);
3132 /* Do the transformation if the result is either exact or we are not
3133 using strict rules. */
3134 if (tv != tarval_bad &&
3135 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3136 ir_node *blk = get_nodes_block(n);
3137 ir_node *c = new_Const(tv);
3138 ir_node *a = get_Quot_left(n);
3139 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode);
3140 ir_node *mem = get_Quot_mem(n);
3142 /* skip a potential Pin */
3143 mem = skip_Pin(mem);
3144 turn_into_tuple(n, pn_Quot_max);
3145 set_Tuple_pred(n, pn_Quot_M, mem);
3146 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(blk));
3147 set_Tuple_pred(n, pn_Quot_X_except, new_Bad());
3148 set_Tuple_pred(n, pn_Quot_res, m);
3149 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3154 } /* transform_node_Quot */
3157 * Optimize Abs(x) into x if x is Confirmed >= 0
3158 * Optimize Abs(x) into -x if x is Confirmed <= 0
3159 * Optimize Abs(-x) int Abs(x)
3161 static ir_node *transform_node_Abs(ir_node *n) {
3162 ir_node *c, *oldn = n;
3163 ir_node *a = get_Abs_op(n);
3166 HANDLE_UNOP_PHI(tarval_abs, a, c);
3168 switch (classify_value_sign(a)) {
3169 case value_classified_negative:
3170 mode = get_irn_mode(n);
3173 * We can replace the Abs by -x here.
3174 * We even could add a new Confirm here
3175 * (if not twos complement)
3177 * Note that -x would create a new node, so we could
3178 * not run it in the equivalent_node() context.
3180 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3182 DBG_OPT_CONFIRM(oldn, n);
3184 case value_classified_positive:
3185 /* n is positive, Abs is not needed */
3188 DBG_OPT_CONFIRM(oldn, n);
3194 /* Abs(-x) = Abs(x) */
3195 mode = get_irn_mode(n);
3196 n = new_rd_Abs(get_irn_dbg_info(n), get_nodes_block(n), get_Minus_op(a), mode);
3197 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3201 } /* transform_node_Abs */
3204 * Optimize -a CMP -b into b CMP a.
3205 * This works only for for modes where unary Minus
3207 * Note that two-complement integers can Overflow
3208 * so it will NOT work.
3210 * For == and != can be handled in Proj(Cmp)
3212 static ir_node *transform_node_Cmp(ir_node *n) {
3214 ir_node *left = get_Cmp_left(n);
3215 ir_node *right = get_Cmp_right(n);
3217 if (is_Minus(left) && is_Minus(right) &&
3218 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3219 ir_node *const new_left = get_Minus_op(right);
3220 ir_node *const new_right = get_Minus_op(left);
3221 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3222 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3225 } /* transform_node_Cmp */
3229 * Transform a Cond node.
3231 * Replace the Cond by a Jmp if it branches on a constant
3234 static ir_node *transform_node_Cond(ir_node *n) {
3237 ir_node *a = get_Cond_selector(n);
3238 tarval *ta = value_of(a);
3240 /* we need block info which is not available in floating irgs */
3241 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3244 if ((ta != tarval_bad) &&
3245 (get_irn_mode(a) == mode_b) &&
3246 (get_opt_unreachable_code())) {
3247 /* It's a boolean Cond, branching on a boolean constant.
3248 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3249 ir_node *blk = get_nodes_block(n);
3250 jmp = new_r_Jmp(blk);
3251 turn_into_tuple(n, pn_Cond_max);
3252 if (ta == tarval_b_true) {
3253 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3254 set_Tuple_pred(n, pn_Cond_true, jmp);
3256 set_Tuple_pred(n, pn_Cond_false, jmp);
3257 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3259 /* We might generate an endless loop, so keep it alive. */
3260 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3263 } /* transform_node_Cond */
3266 * Prototype of a recursive transform function
3267 * for bitwise distributive transformations.
3269 typedef ir_node* (*recursive_transform)(ir_node *n);
3272 * makes use of distributive laws for and, or, eor
3273 * and(a OP c, b OP c) -> and(a, b) OP c
3274 * note, might return a different op than n
3276 static ir_node *transform_bitwise_distributive(ir_node *n,
3277 recursive_transform trans_func)
3280 ir_node *a = get_binop_left(n);
3281 ir_node *b = get_binop_right(n);
3282 ir_op *op = get_irn_op(a);
3283 ir_op *op_root = get_irn_op(n);
3285 if(op != get_irn_op(b))
3288 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3289 if (op == op_Conv) {
3290 ir_node *a_op = get_Conv_op(a);
3291 ir_node *b_op = get_Conv_op(b);
3292 ir_mode *a_mode = get_irn_mode(a_op);
3293 ir_mode *b_mode = get_irn_mode(b_op);
3294 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3295 ir_node *blk = get_nodes_block(n);
3298 set_binop_left(n, a_op);
3299 set_binop_right(n, b_op);
3300 set_irn_mode(n, a_mode);
3302 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3304 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3310 /* nothing to gain here */
3314 if (op == op_Shrs || op == op_Shr || op == op_Shl
3315 || op == op_And || op == op_Or || op == op_Eor) {
3316 ir_node *a_left = get_binop_left(a);
3317 ir_node *a_right = get_binop_right(a);
3318 ir_node *b_left = get_binop_left(b);
3319 ir_node *b_right = get_binop_right(b);
3321 ir_node *op1 = NULL;
3322 ir_node *op2 = NULL;
3324 if (is_op_commutative(op)) {
3325 if (a_left == b_left) {
3329 } else if(a_left == b_right) {
3333 } else if(a_right == b_left) {
3339 if(a_right == b_right) {
3346 /* (a sop c) & (b sop c) => (a & b) sop c */
3347 ir_node *blk = get_nodes_block(n);
3349 ir_node *new_n = exact_copy(n);
3350 set_binop_left(new_n, op1);
3351 set_binop_right(new_n, op2);
3352 new_n = trans_func(new_n);
3354 if(op_root == op_Eor && op == op_Or) {
3355 dbg_info *dbgi = get_irn_dbg_info(n);
3356 ir_mode *mode = get_irn_mode(c);
3358 c = new_rd_Not(dbgi, blk, c, mode);
3359 n = new_rd_And(dbgi, blk, new_n, c, mode);
3362 set_nodes_block(n, blk);
3363 set_binop_left(n, new_n);
3364 set_binop_right(n, c);
3365 add_identities(current_ir_graph->value_table, n);
3368 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3379 static ir_node *transform_node_And(ir_node *n) {
3380 ir_node *c, *oldn = n;
3381 ir_node *a = get_And_left(n);
3382 ir_node *b = get_And_right(n);
3385 mode = get_irn_mode(n);
3386 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3388 /* we can evaluate 2 Projs of the same Cmp */
3389 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3390 ir_node *pred_a = get_Proj_pred(a);
3391 ir_node *pred_b = get_Proj_pred(b);
3392 if (pred_a == pred_b) {
3393 dbg_info *dbgi = get_irn_dbg_info(n);
3394 ir_node *block = get_nodes_block(pred_a);
3395 pn_Cmp pn_a = get_Proj_proj(a);
3396 pn_Cmp pn_b = get_Proj_proj(b);
3397 /* yes, we can simply calculate with pncs */
3398 pn_Cmp new_pnc = pn_a & pn_b;
3400 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3405 ir_node *op = get_Not_op(b);
3407 ir_node *ba = get_And_left(op);
3408 ir_node *bb = get_And_right(op);
3410 /* it's enough to test the following cases due to normalization! */
3411 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3412 /* (a|b) & ~(a&b) = a^b */
3413 ir_node *block = get_nodes_block(n);
3415 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3416 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3424 ir_node *op = get_Not_op(a);
3426 ir_node *aa = get_And_left(op);
3427 ir_node *ab = get_And_right(op);
3429 /* it's enough to test the following cases due to normalization! */
3430 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3431 /* (a|b) & ~(a&b) = a^b */
3432 ir_node *block = get_nodes_block(n);
3434 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3435 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3442 ir_node *al = get_Eor_left(a);
3443 ir_node *ar = get_Eor_right(a);
3446 /* (b ^ a) & b -> ~a & b */
3447 dbg_info *dbg = get_irn_dbg_info(n);
3448 ir_node *block = get_nodes_block(n);
3450 ar = new_rd_Not(dbg, block, ar, mode);
3451 n = new_rd_And(dbg, block, ar, b, mode);
3452 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3456 /* (a ^ b) & b -> ~a & b */
3457 dbg_info *dbg = get_irn_dbg_info(n);
3458 ir_node *block = get_nodes_block(n);
3460 al = new_rd_Not(dbg, block, al, mode);
3461 n = new_rd_And(dbg, block, al, b, mode);
3462 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3467 ir_node *bl = get_Eor_left(b);
3468 ir_node *br = get_Eor_right(b);
3471 /* a & (a ^ b) -> a & ~b */
3472 dbg_info *dbg = get_irn_dbg_info(n);
3473 ir_node *block = get_nodes_block(n);
3475 br = new_rd_Not(dbg, block, br, mode);
3476 n = new_rd_And(dbg, block, br, a, mode);
3477 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3481 /* a & (b ^ a) -> a & ~b */
3482 dbg_info *dbg = get_irn_dbg_info(n);
3483 ir_node *block = get_nodes_block(n);
3485 bl = new_rd_Not(dbg, block, bl, mode);
3486 n = new_rd_And(dbg, block, bl, a, mode);
3487 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3491 if (is_Not(a) && is_Not(b)) {
3492 /* ~a & ~b = ~(a|b) */
3493 ir_node *block = get_nodes_block(n);
3494 ir_mode *mode = get_irn_mode(n);
3498 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3499 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3500 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3504 n = transform_bitwise_distributive(n, transform_node_And);
3507 } /* transform_node_And */
3512 static ir_node *transform_node_Eor(ir_node *n) {
3513 ir_node *c, *oldn = n;
3514 ir_node *a = get_Eor_left(n);
3515 ir_node *b = get_Eor_right(n);
3516 ir_mode *mode = get_irn_mode(n);
3518 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3520 /* we can evaluate 2 Projs of the same Cmp */
3521 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3522 ir_node *pred_a = get_Proj_pred(a);
3523 ir_node *pred_b = get_Proj_pred(b);
3524 if(pred_a == pred_b) {
3525 dbg_info *dbgi = get_irn_dbg_info(n);
3526 ir_node *block = get_nodes_block(pred_a);
3527 pn_Cmp pn_a = get_Proj_proj(a);
3528 pn_Cmp pn_b = get_Proj_proj(b);
3529 /* yes, we can simply calculate with pncs */
3530 pn_Cmp new_pnc = pn_a ^ pn_b;
3532 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3538 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3539 get_mode_null(mode));
3540 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3541 } else if (is_Const(b)) {
3542 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3543 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b)));
3544 ir_node *not_op = get_Not_op(a);
3545 dbg_info *dbg = get_irn_dbg_info(n);
3546 ir_node *block = get_nodes_block(n);
3547 ir_mode *mode = get_irn_mode(n);
3548 n = new_rd_Eor(dbg, block, not_op, cnst, mode);
3550 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3551 n = new_r_Not(get_nodes_block(n), a, mode);
3552 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3555 n = transform_bitwise_distributive(n, transform_node_Eor);
3559 } /* transform_node_Eor */
3564 static ir_node *transform_node_Not(ir_node *n) {
3565 ir_node *c, *oldn = n;
3566 ir_node *a = get_Not_op(n);
3567 ir_mode *mode = get_irn_mode(n);
3569 HANDLE_UNOP_PHI(tarval_not,a,c);
3571 /* check for a boolean Not */
3572 if (mode == mode_b && is_Proj(a)) {
3573 ir_node *a_pred = get_Proj_pred(a);
3574 if (is_Cmp(a_pred)) {
3575 ir_node *cmp_block = get_nodes_block(a_pred);
3576 /* We negate a Cmp. The Cmp has the negated result anyways! */
3577 n = new_r_Proj(cmp_block, get_Proj_pred(a),
3578 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3579 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3584 ir_node *eor_b = get_Eor_right(a);
3585 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3586 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b)));
3587 ir_node *eor_a = get_Eor_left(a);
3588 dbg_info *dbg = get_irn_dbg_info(n);
3589 ir_node *block = get_nodes_block(n);
3590 ir_mode *mode = get_irn_mode(n);
3591 n = new_rd_Eor(dbg, block, eor_a, cnst, mode);
3595 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3596 if (is_Minus(a)) { /* ~-x -> x + -1 */
3597 dbg_info *dbg = get_irn_dbg_info(n);
3598 ir_graph *irg = current_ir_graph;
3599 ir_node *block = get_nodes_block(n);
3600 ir_node *add_l = get_Minus_op(a);
3601 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3602 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3603 } else if (is_Add(a)) {
3604 ir_node *add_r = get_Add_right(a);
3605 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3606 /* ~(x + -1) = -x */
3607 ir_node *op = get_Add_left(a);
3608 ir_node *blk = get_nodes_block(n);
3609 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3610 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3615 } /* transform_node_Not */
3618 * Transform a Minus.
3622 * -(a >>u (size-1)) = a >>s (size-1)
3623 * -(a >>s (size-1)) = a >>u (size-1)
3624 * -(a * const) -> a * -const
3626 static ir_node *transform_node_Minus(ir_node *n) {
3627 ir_node *c, *oldn = n;
3628 ir_node *a = get_Minus_op(n);
3631 HANDLE_UNOP_PHI(tarval_neg,a,c);
3633 mode = get_irn_mode(a);
3634 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3635 /* the following rules are only to twos-complement */
3638 ir_node *op = get_Not_op(a);
3639 tarval *tv = get_mode_one(mode);
3640 ir_node *blk = get_nodes_block(n);
3641 ir_node *c = new_Const(tv);
3642 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3643 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3647 ir_node *c = get_Shr_right(a);
3650 tarval *tv = get_Const_tarval(c);
3652 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3653 /* -(a >>u (size-1)) = a >>s (size-1) */
3654 ir_node *v = get_Shr_left(a);
3656 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3657 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3663 ir_node *c = get_Shrs_right(a);
3666 tarval *tv = get_Const_tarval(c);
3668 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3669 /* -(a >>s (size-1)) = a >>u (size-1) */
3670 ir_node *v = get_Shrs_left(a);
3672 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3673 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3680 /* - (a-b) = b - a */
3681 ir_node *la = get_Sub_left(a);
3682 ir_node *ra = get_Sub_right(a);
3683 ir_node *blk = get_nodes_block(n);
3685 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3686 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3690 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3691 ir_node *mul_l = get_Mul_left(a);
3692 ir_node *mul_r = get_Mul_right(a);
3693 tarval *tv = value_of(mul_r);
3694 if (tv != tarval_bad) {
3695 tv = tarval_neg(tv);
3696 if (tv != tarval_bad) {
3697 ir_node *cnst = new_Const(tv);
3698 dbg_info *dbg = get_irn_dbg_info(a);
3699 ir_node *block = get_nodes_block(a);
3700 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3701 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3708 } /* transform_node_Minus */
3711 * Transform a Cast_type(Const) into a new Const_type
3713 static ir_node *transform_node_Cast(ir_node *n) {
3715 ir_node *pred = get_Cast_op(n);
3716 ir_type *tp = get_irn_type(n);
3718 if (is_Const(pred) && get_Const_type(pred) != tp) {
3719 n = new_rd_Const_type(NULL, current_ir_graph, get_Const_tarval(pred), tp);
3720 DBG_OPT_CSTEVAL(oldn, n);
3721 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3722 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_mode(pred),
3723 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3724 DBG_OPT_CSTEVAL(oldn, n);
3728 } /* transform_node_Cast */
3731 * Transform a Proj(Load) with a non-null address.
3733 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3734 if (get_opt_ldst_only_null_ptr_exceptions()) {
3735 if (get_irn_mode(proj) == mode_X) {
3736 ir_node *load = get_Proj_pred(proj);
3738 /* get the Load address */
3739 const ir_node *addr = get_Load_ptr(load);
3740 const ir_node *confirm;
3742 if (value_not_null(addr, &confirm)) {
3743 if (confirm == NULL) {
3744 /* this node may float if it did not depend on a Confirm */
3745 set_irn_pinned(load, op_pin_state_floats);
3747 if (get_Proj_proj(proj) == pn_Load_X_except) {
3748 DBG_OPT_EXC_REM(proj);
3749 return get_irg_bad(current_ir_graph);
3751 ir_node *blk = get_nodes_block(load);
3752 return new_r_Jmp(blk);
3758 } /* transform_node_Proj_Load */
3761 * Transform a Proj(Store) with a non-null address.
3763 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3764 if (get_opt_ldst_only_null_ptr_exceptions()) {
3765 if (get_irn_mode(proj) == mode_X) {
3766 ir_node *store = get_Proj_pred(proj);
3768 /* get the load/store address */
3769 const ir_node *addr = get_Store_ptr(store);
3770 const ir_node *confirm;
3772 if (value_not_null(addr, &confirm)) {
3773 if (confirm == NULL) {
3774 /* this node may float if it did not depend on a Confirm */
3775 set_irn_pinned(store, op_pin_state_floats);
3777 if (get_Proj_proj(proj) == pn_Store_X_except) {
3778 DBG_OPT_EXC_REM(proj);
3779 return get_irg_bad(current_ir_graph);
3781 ir_node *blk = get_nodes_block(store);
3782 return new_r_Jmp(blk);
3788 } /* transform_node_Proj_Store */
3791 * Transform a Proj(Div) with a non-zero value.
3792 * Removes the exceptions and routes the memory to the NoMem node.
3794 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3795 ir_node *div = get_Proj_pred(proj);
3796 ir_node *b = get_Div_right(div);
3797 ir_node *res, *new_mem;
3798 const ir_node *confirm;
3801 if (value_not_zero(b, &confirm)) {
3802 /* div(x, y) && y != 0 */
3803 if (confirm == NULL) {
3804 /* we are sure we have a Const != 0 */
3805 new_mem = get_Div_mem(div);
3806 new_mem = skip_Pin(new_mem);
3807 set_Div_mem(div, new_mem);
3808 set_irn_pinned(div, op_pin_state_floats);
3811 proj_nr = get_Proj_proj(proj);
3813 case pn_Div_X_regular:
3814 return new_r_Jmp(get_nodes_block(div));
3816 case pn_Div_X_except:
3817 /* we found an exception handler, remove it */
3818 DBG_OPT_EXC_REM(proj);
3822 res = get_Div_mem(div);
3823 new_mem = get_irg_no_mem(current_ir_graph);
3826 /* This node can only float up to the Confirm block */
3827 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3829 set_irn_pinned(div, op_pin_state_floats);
3830 /* this is a Div without exception, we can remove the memory edge */
3831 set_Div_mem(div, new_mem);
3836 } /* transform_node_Proj_Div */
3839 * Transform a Proj(Mod) with a non-zero value.
3840 * Removes the exceptions and routes the memory to the NoMem node.
3842 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3843 ir_node *mod = get_Proj_pred(proj);
3844 ir_node *b = get_Mod_right(mod);
3845 ir_node *res, *new_mem;
3846 const ir_node *confirm;
3849 if (value_not_zero(b, &confirm)) {
3850 /* mod(x, y) && y != 0 */
3851 proj_nr = get_Proj_proj(proj);
3853 if (confirm == NULL) {
3854 /* we are sure we have a Const != 0 */
3855 new_mem = get_Mod_mem(mod);
3856 new_mem = skip_Pin(new_mem);
3857 set_Mod_mem(mod, new_mem);
3858 set_irn_pinned(mod, op_pin_state_floats);
3863 case pn_Mod_X_regular:
3864 return new_r_Jmp(get_irn_n(mod, -1));
3866 case pn_Mod_X_except:
3867 /* we found an exception handler, remove it */
3868 DBG_OPT_EXC_REM(proj);
3872 res = get_Mod_mem(mod);
3873 new_mem = get_irg_no_mem(current_ir_graph);
3876 /* This node can only float up to the Confirm block */
3877 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3879 /* this is a Mod without exception, we can remove the memory edge */
3880 set_Mod_mem(mod, new_mem);
3883 if (get_Mod_left(mod) == b) {
3884 /* a % a = 0 if a != 0 */
3885 ir_mode *mode = get_irn_mode(proj);
3886 ir_node *res = new_Const(get_mode_null(mode));
3888 DBG_OPT_CSTEVAL(mod, res);
3894 } /* transform_node_Proj_Mod */
3897 * Transform a Proj(DivMod) with a non-zero value.
3898 * Removes the exceptions and routes the memory to the NoMem node.
3900 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3901 ir_node *divmod = get_Proj_pred(proj);
3902 ir_node *b = get_DivMod_right(divmod);
3903 ir_node *res, *new_mem;
3904 const ir_node *confirm;
3907 if (value_not_zero(b, &confirm)) {
3908 /* DivMod(x, y) && y != 0 */
3909 proj_nr = get_Proj_proj(proj);
3911 if (confirm == NULL) {
3912 /* we are sure we have a Const != 0 */
3913 new_mem = get_DivMod_mem(divmod);
3914 new_mem = skip_Pin(new_mem);
3915 set_DivMod_mem(divmod, new_mem);
3916 set_irn_pinned(divmod, op_pin_state_floats);
3921 case pn_DivMod_X_regular:
3922 return new_r_Jmp(get_nodes_block(divmod));
3924 case pn_DivMod_X_except:
3925 /* we found an exception handler, remove it */
3926 DBG_OPT_EXC_REM(proj);
3930 res = get_DivMod_mem(divmod);
3931 new_mem = get_irg_no_mem(current_ir_graph);
3934 /* This node can only float up to the Confirm block */
3935 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3937 /* this is a DivMod without exception, we can remove the memory edge */
3938 set_DivMod_mem(divmod, new_mem);
3941 case pn_DivMod_res_mod:
3942 if (get_DivMod_left(divmod) == b) {
3943 /* a % a = 0 if a != 0 */
3944 ir_mode *mode = get_irn_mode(proj);
3945 ir_node *res = new_Const(get_mode_null(mode));
3947 DBG_OPT_CSTEVAL(divmod, res);
3953 } /* transform_node_Proj_DivMod */
3956 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3958 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3959 if (get_opt_unreachable_code()) {
3960 ir_node *n = get_Proj_pred(proj);
3961 ir_node *b = get_Cond_selector(n);
3963 if (mode_is_int(get_irn_mode(b))) {
3964 tarval *tb = value_of(b);
3966 if (tb != tarval_bad) {
3967 /* we have a constant switch */
3968 long num = get_Proj_proj(proj);
3970 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3971 if (get_tarval_long(tb) == num) {
3972 /* Do NOT create a jump here, or we will have 2 control flow ops
3973 * in a block. This case is optimized away in optimize_cf(). */
3976 /* this case will NEVER be taken, kill it */
3977 return get_irg_bad(current_ir_graph);
3984 } /* transform_node_Proj_Cond */
3987 * Create a 0 constant of given mode.
3989 static ir_node *create_zero_const(ir_mode *mode) {
3990 tarval *tv = get_mode_null(mode);
3991 ir_node *cnst = new_Const(tv);
3996 /* the order of the values is important! */
3997 typedef enum const_class {
4003 static const_class classify_const(const ir_node* n)
4005 if (is_Const(n)) return const_const;
4006 if (is_irn_constlike(n)) return const_like;
4011 * Determines whether r is more constlike or has a larger index (in that order)
4014 static int operands_are_normalized(const ir_node *l, const ir_node *r)
4016 const const_class l_order = classify_const(l);
4017 const const_class r_order = classify_const(r);
4019 l_order > r_order ||
4020 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4024 * Normalizes and optimizes Cmp nodes.
4026 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4027 ir_node *n = get_Proj_pred(proj);
4028 ir_node *left = get_Cmp_left(n);
4029 ir_node *right = get_Cmp_right(n);
4032 ir_mode *mode = NULL;
4033 long proj_nr = get_Proj_proj(proj);
4035 /* we can evaluate some cases directly */
4038 return new_Const(get_tarval_b_false());
4040 return new_Const(get_tarval_b_true());
4042 if (!mode_is_float(get_irn_mode(left)))
4043 return new_Const(get_tarval_b_true());
4049 /* remove Casts of both sides */
4050 left = skip_Cast(left);
4051 right = skip_Cast(right);
4053 /* Remove unnecessary conversions */
4054 /* TODO handle constants */
4055 if (is_Conv(left) && is_Conv(right)) {
4056 ir_mode *mode = get_irn_mode(left);
4057 ir_node *op_left = get_Conv_op(left);
4058 ir_node *op_right = get_Conv_op(right);
4059 ir_mode *mode_left = get_irn_mode(op_left);
4060 ir_mode *mode_right = get_irn_mode(op_right);
4062 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4063 && mode_left != mode_b && mode_right != mode_b) {
4064 ir_node *block = get_nodes_block(n);
4066 if (mode_left == mode_right) {
4070 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4071 } else if (smaller_mode(mode_left, mode_right)) {
4072 left = new_r_Conv(block, op_left, mode_right);
4075 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4076 } else if (smaller_mode(mode_right, mode_left)) {
4078 right = new_r_Conv(block, op_right, mode_left);
4080 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4085 /* remove operation on both sides if possible */
4086 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4088 * The following operations are NOT safe for floating point operations, for instance
4089 * 1.0 + inf == 2.0 + inf, =/=> x == y
4091 if (mode_is_int(get_irn_mode(left))) {
4092 unsigned lop = get_irn_opcode(left);
4094 if (lop == get_irn_opcode(right)) {
4095 ir_node *ll, *lr, *rl, *rr;
4097 /* same operation on both sides, try to remove */
4101 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4102 left = get_unop_op(left);
4103 right = get_unop_op(right);
4105 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4108 ll = get_Add_left(left);
4109 lr = get_Add_right(left);
4110 rl = get_Add_left(right);
4111 rr = get_Add_right(right);
4114 /* X + a CMP X + b ==> a CMP b */
4118 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4119 } else if (ll == rr) {
4120 /* X + a CMP b + X ==> a CMP b */
4124 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4125 } else if (lr == rl) {
4126 /* a + X CMP X + b ==> a CMP b */
4130 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4131 } else if (lr == rr) {
4132 /* a + X CMP b + X ==> a CMP b */
4136 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4140 ll = get_Sub_left(left);
4141 lr = get_Sub_right(left);
4142 rl = get_Sub_left(right);
4143 rr = get_Sub_right(right);
4146 /* X - a CMP X - b ==> a CMP b */
4150 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4151 } else if (lr == rr) {
4152 /* a - X CMP b - X ==> a CMP b */
4156 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4160 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4161 /* a ROTL X CMP b ROTL X ==> a CMP b */
4162 left = get_Rotl_left(left);
4163 right = get_Rotl_left(right);
4165 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4173 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4174 if (is_Add(left) || is_Sub(left)) {
4175 ir_node *ll = get_binop_left(left);
4176 ir_node *lr = get_binop_right(left);
4178 if (lr == right && is_Add(left)) {
4185 right = create_zero_const(get_irn_mode(left));
4187 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4190 if (is_Add(right) || is_Sub(right)) {
4191 ir_node *rl = get_binop_left(right);
4192 ir_node *rr = get_binop_right(right);
4194 if (rr == left && is_Add(right)) {
4201 right = create_zero_const(get_irn_mode(left));
4203 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4206 if (is_And(left) && is_Const(right)) {
4207 ir_node *ll = get_binop_left(left);
4208 ir_node *lr = get_binop_right(left);
4209 if (is_Shr(ll) && is_Const(lr)) {
4210 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4211 ir_node *block = get_nodes_block(n);
4212 ir_mode *mode = get_irn_mode(left);
4214 ir_node *llr = get_Shr_right(ll);
4215 if (is_Const(llr)) {
4216 dbg_info *dbg = get_irn_dbg_info(left);
4218 tarval *c1 = get_Const_tarval(llr);
4219 tarval *c2 = get_Const_tarval(lr);
4220 tarval *c3 = get_Const_tarval(right);
4221 tarval *mask = tarval_shl(c2, c1);
4222 tarval *value = tarval_shl(c3, c1);
4224 left = new_rd_And(dbg, block, get_Shr_left(ll), new_Const(mask), mode);
4225 right = new_Const(value);
4230 } /* mode_is_int(...) */
4231 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4233 /* replace mode_b compares with ands/ors */
4234 if (get_irn_mode(left) == mode_b) {
4235 ir_node *block = get_nodes_block(n);
4239 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4240 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4241 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4242 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4243 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4244 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4245 default: bres = NULL;
4248 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4254 * First step: normalize the compare op
4255 * by placing the constant on the right side
4256 * or moving the lower address node to the left.
4258 if (!operands_are_normalized(left, right)) {
4264 proj_nr = get_inversed_pnc(proj_nr);
4269 * Second step: Try to reduce the magnitude
4270 * of a constant. This may help to generate better code
4271 * later and may help to normalize more compares.
4272 * Of course this is only possible for integer values.
4274 tv = value_of(right);
4275 if (tv != tarval_bad) {
4276 mode = get_irn_mode(right);
4278 /* TODO extend to arbitrary constants */
4279 if (is_Conv(left) && tarval_is_null(tv)) {
4280 ir_node *op = get_Conv_op(left);
4281 ir_mode *op_mode = get_irn_mode(op);
4284 * UpConv(x) REL 0 ==> x REL 0
4285 * Don't do this for float values as it's unclear whether it is a
4286 * win. (on the other side it makes detection/creation of fabs hard)
4288 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4289 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4290 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4291 !mode_is_float(mode)) {
4292 tv = get_mode_null(op_mode);
4296 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4300 if (tv != tarval_bad) {
4301 /* the following optimization is possible on modes without Overflow
4302 * on Unary Minus or on == and !=:
4303 * -a CMP c ==> a swap(CMP) -c
4305 * Beware: for two-complement Overflow may occur, so only == and != can
4306 * be optimized, see this:
4307 * -MININT < 0 =/=> MININT > 0 !!!
4309 if (is_Minus(left) &&
4310 (!mode_overflow_on_unary_Minus(mode) ||
4311 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4312 tv = tarval_neg(tv);
4314 if (tv != tarval_bad) {
4315 left = get_Minus_op(left);
4316 proj_nr = get_inversed_pnc(proj_nr);
4318 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4320 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4321 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4322 tv = tarval_not(tv);
4324 if (tv != tarval_bad) {
4325 left = get_Not_op(left);
4327 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4331 /* for integer modes, we have more */
4332 if (mode_is_int(mode)) {
4333 /* Ne includes Unordered which is not possible on integers.
4334 * However, frontends often use this wrong, so fix it here */
4335 if (proj_nr & pn_Cmp_Uo) {
4336 proj_nr &= ~pn_Cmp_Uo;
4337 set_Proj_proj(proj, proj_nr);
4340 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4341 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4342 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4343 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4345 if (tv != tarval_bad) {
4346 proj_nr ^= pn_Cmp_Eq;
4348 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4351 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4352 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4353 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4354 tv = tarval_add(tv, get_mode_one(mode));
4356 if (tv != tarval_bad) {
4357 proj_nr ^= pn_Cmp_Eq;
4359 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4363 /* the following reassociations work only for == and != */
4364 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4366 #if 0 /* Might be not that good in general */
4367 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4368 if (tarval_is_null(tv) && is_Sub(left)) {
4369 right = get_Sub_right(left);
4370 left = get_Sub_left(left);
4372 tv = value_of(right);
4374 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4378 if (tv != tarval_bad) {
4379 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4381 ir_node *c1 = get_Sub_right(left);
4382 tarval *tv2 = value_of(c1);
4384 if (tv2 != tarval_bad) {
4385 tv2 = tarval_add(tv, value_of(c1));
4387 if (tv2 != tarval_bad) {
4388 left = get_Sub_left(left);
4391 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4395 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4396 else if (is_Add(left)) {
4397 ir_node *a_l = get_Add_left(left);
4398 ir_node *a_r = get_Add_right(left);
4402 if (is_Const(a_l)) {
4404 tv2 = value_of(a_l);
4407 tv2 = value_of(a_r);
4410 if (tv2 != tarval_bad) {
4411 tv2 = tarval_sub(tv, tv2, NULL);
4413 if (tv2 != tarval_bad) {
4417 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4421 /* -a == c ==> a == -c, -a != c ==> a != -c */
4422 else if (is_Minus(left)) {
4423 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4425 if (tv2 != tarval_bad) {
4426 left = get_Minus_op(left);
4429 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4434 /* the following reassociations work only for <= */
4435 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4436 if (tv != tarval_bad) {
4437 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4438 if (is_Abs(left)) { // TODO something is missing here
4444 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4445 switch (get_irn_opcode(left)) {
4449 c1 = get_And_right(left);
4452 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4453 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4455 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4457 /* TODO: move to constant evaluation */
4458 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4460 DBG_OPT_CSTEVAL(proj, c1);
4464 if (tarval_is_single_bit(tv)) {
4466 * optimization for AND:
4468 * And(x, C) == C ==> And(x, C) != 0
4469 * And(x, C) != C ==> And(X, C) == 0
4471 * if C is a single Bit constant.
4474 /* check for Constant's match. We have check hare the tarvals,
4475 because our const might be changed */
4476 if (get_Const_tarval(c1) == tv) {
4477 /* fine: do the transformation */
4478 tv = get_mode_null(get_tarval_mode(tv));
4479 proj_nr ^= pn_Cmp_Leg;
4481 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4487 c1 = get_Or_right(left);
4488 if (is_Const(c1) && tarval_is_null(tv)) {
4490 * Or(x, C) == 0 && C != 0 ==> FALSE
4491 * Or(x, C) != 0 && C != 0 ==> TRUE
4493 if (! tarval_is_null(get_Const_tarval(c1))) {
4494 /* TODO: move to constant evaluation */
4495 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4497 DBG_OPT_CSTEVAL(proj, c1);
4504 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4506 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4509 c1 = get_Shl_right(left);
4511 tarval *tv1 = get_Const_tarval(c1);
4512 ir_mode *mode = get_irn_mode(left);
4513 tarval *minus1 = get_mode_all_one(mode);
4514 tarval *amask = tarval_shr(minus1, tv1);
4515 tarval *cmask = tarval_shl(minus1, tv1);
4518 if (tarval_and(tv, cmask) != tv) {
4519 /* condition not met */
4520 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4522 DBG_OPT_CSTEVAL(proj, c1);
4525 sl = get_Shl_left(left);
4526 blk = get_nodes_block(n);
4527 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4528 tv = tarval_shr(tv, tv1);
4530 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4535 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4537 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4540 c1 = get_Shr_right(left);
4542 tarval *tv1 = get_Const_tarval(c1);
4543 ir_mode *mode = get_irn_mode(left);
4544 tarval *minus1 = get_mode_all_one(mode);
4545 tarval *amask = tarval_shl(minus1, tv1);
4546 tarval *cmask = tarval_shr(minus1, tv1);
4549 if (tarval_and(tv, cmask) != tv) {
4550 /* condition not met */
4551 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4553 DBG_OPT_CSTEVAL(proj, c1);
4556 sl = get_Shr_left(left);
4557 blk = get_nodes_block(n);
4558 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4559 tv = tarval_shl(tv, tv1);
4561 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4566 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4568 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4571 c1 = get_Shrs_right(left);
4573 tarval *tv1 = get_Const_tarval(c1);
4574 ir_mode *mode = get_irn_mode(left);
4575 tarval *minus1 = get_mode_all_one(mode);
4576 tarval *amask = tarval_shl(minus1, tv1);
4577 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4580 cond = tarval_sub(cond, tv1, NULL);
4581 cond = tarval_shrs(tv, cond);
4583 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4584 /* condition not met */
4585 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4587 DBG_OPT_CSTEVAL(proj, c1);
4590 sl = get_Shrs_left(left);
4591 blk = get_nodes_block(n);
4592 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4593 tv = tarval_shl(tv, tv1);
4595 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4600 } /* tarval != bad */
4603 if (changed & 2) /* need a new Const */
4604 right = new_Const(tv);
4606 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4607 ir_node *op = get_Proj_pred(left);
4609 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4610 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4611 ir_node *c = get_binop_right(op);
4614 tarval *tv = get_Const_tarval(c);
4616 if (tarval_is_single_bit(tv)) {
4617 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4618 ir_node *v = get_binop_left(op);
4619 ir_node *blk = get_irn_n(op, -1);
4620 ir_mode *mode = get_irn_mode(v);
4622 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4623 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_Const(tv), mode);
4625 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4632 ir_node *block = get_nodes_block(n);
4634 /* create a new compare */
4635 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4636 proj = new_rd_Proj(get_irn_dbg_info(proj), block, n, get_irn_mode(proj), proj_nr);
4640 } /* transform_node_Proj_Cmp */
4643 * Optimize CopyB(mem, x, x) into a Nop.
4645 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4646 ir_node *copyb = get_Proj_pred(proj);
4647 ir_node *a = get_CopyB_dst(copyb);
4648 ir_node *b = get_CopyB_src(copyb);
4651 switch (get_Proj_proj(proj)) {
4652 case pn_CopyB_X_regular:
4653 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4654 DBG_OPT_EXC_REM(proj);
4655 proj = new_r_Jmp(get_nodes_block(copyb));
4657 case pn_CopyB_X_except:
4658 DBG_OPT_EXC_REM(proj);
4659 proj = get_irg_bad(get_irn_irg(proj));
4666 } /* transform_node_Proj_CopyB */
4669 * Optimize Bounds(idx, idx, upper) into idx.
4671 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4672 ir_node *oldn = proj;
4673 ir_node *bound = get_Proj_pred(proj);
4674 ir_node *idx = get_Bound_index(bound);
4675 ir_node *pred = skip_Proj(idx);
4678 if (idx == get_Bound_lower(bound))
4680 else if (is_Bound(pred)) {
4682 * idx was Bounds checked in the same MacroBlock previously,
4683 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4685 ir_node *lower = get_Bound_lower(bound);
4686 ir_node *upper = get_Bound_upper(bound);
4687 if (get_Bound_lower(pred) == lower &&
4688 get_Bound_upper(pred) == upper &&
4689 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4691 * One could expect that we simply return the previous
4692 * Bound here. However, this would be wrong, as we could
4693 * add an exception Proj to a new location then.
4694 * So, we must turn in into a tuple.
4700 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4701 switch (get_Proj_proj(proj)) {
4703 DBG_OPT_EXC_REM(proj);
4704 proj = get_Bound_mem(bound);
4706 case pn_Bound_X_except:
4707 DBG_OPT_EXC_REM(proj);
4708 proj = get_irg_bad(get_irn_irg(proj));
4712 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4714 case pn_Bound_X_regular:
4715 DBG_OPT_EXC_REM(proj);
4716 proj = new_r_Jmp(get_nodes_block(bound));
4723 } /* transform_node_Proj_Bound */
4726 * Does all optimizations on nodes that must be done on it's Proj's
4727 * because of creating new nodes.
4729 static ir_node *transform_node_Proj(ir_node *proj) {
4730 ir_node *n = get_Proj_pred(proj);
4732 if (n->op->ops.transform_node_Proj)
4733 return n->op->ops.transform_node_Proj(proj);
4735 } /* transform_node_Proj */
4738 * Move Confirms down through Phi nodes.
4740 static ir_node *transform_node_Phi(ir_node *phi) {
4742 ir_mode *mode = get_irn_mode(phi);
4744 if (mode_is_reference(mode)) {
4745 n = get_irn_arity(phi);
4747 /* Beware of Phi0 */
4749 ir_node *pred = get_irn_n(phi, 0);
4750 ir_node *bound, *new_Phi, *block, **in;
4753 if (! is_Confirm(pred))
4756 bound = get_Confirm_bound(pred);
4757 pnc = get_Confirm_cmp(pred);
4759 NEW_ARR_A(ir_node *, in, n);
4760 in[0] = get_Confirm_value(pred);
4762 for (i = 1; i < n; ++i) {
4763 pred = get_irn_n(phi, i);
4765 if (! is_Confirm(pred) ||
4766 get_Confirm_bound(pred) != bound ||
4767 get_Confirm_cmp(pred) != pnc)
4769 in[i] = get_Confirm_value(pred);
4771 /* move the Confirm nodes "behind" the Phi */
4772 block = get_irn_n(phi, -1);
4773 new_Phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4774 return new_r_Confirm(block, new_Phi, bound, pnc);
4778 } /* transform_node_Phi */
4781 * Returns the operands of a commutative bin-op, if one operand is
4782 * a const, it is returned as the second one.
4784 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4785 ir_node *op_a = get_binop_left(binop);
4786 ir_node *op_b = get_binop_right(binop);
4788 assert(is_op_commutative(get_irn_op(binop)));
4790 if (is_Const(op_a)) {
4797 } /* get_comm_Binop_Ops */
4800 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4801 * Such pattern may arise in bitfield stores.
4803 * value c4 value c4 & c2
4804 * AND c3 AND c1 | c3
4811 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4814 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4817 ir_node *and_l, *c3;
4818 ir_node *value, *c4;
4819 ir_node *new_and, *new_const, *block;
4820 ir_mode *mode = get_irn_mode(or);
4822 tarval *tv1, *tv2, *tv3, *tv4, *tv;
4825 get_comm_Binop_Ops(or, &and, &c1);
4826 if (!is_Const(c1) || !is_And(and))
4829 get_comm_Binop_Ops(and, &or_l, &c2);
4833 tv1 = get_Const_tarval(c1);
4834 tv2 = get_Const_tarval(c2);
4836 tv = tarval_or(tv1, tv2);
4837 if (tarval_is_all_one(tv)) {
4838 /* the AND does NOT clear a bit with isn't set by the OR */
4839 set_Or_left(or, or_l);
4840 set_Or_right(or, c1);
4842 /* check for more */
4849 get_comm_Binop_Ops(or_l, &and_l, &c3);
4850 if (!is_Const(c3) || !is_And(and_l))
4853 get_comm_Binop_Ops(and_l, &value, &c4);
4857 /* ok, found the pattern, check for conditions */
4858 assert(mode == get_irn_mode(and));
4859 assert(mode == get_irn_mode(or_l));
4860 assert(mode == get_irn_mode(and_l));
4862 tv3 = get_Const_tarval(c3);
4863 tv4 = get_Const_tarval(c4);
4865 tv = tarval_or(tv4, tv2);
4866 if (!tarval_is_all_one(tv)) {
4867 /* have at least one 0 at the same bit position */
4871 if (tv3 != tarval_andnot(tv3, tv4)) {
4872 /* bit in the or_mask is outside the and_mask */
4876 if (tv1 != tarval_andnot(tv1, tv2)) {
4877 /* bit in the or_mask is outside the and_mask */
4881 /* ok, all conditions met */
4882 block = get_irn_n(or, -1);
4884 new_and = new_r_And(block, value, new_Const(tarval_and(tv4, tv2)), mode);
4886 new_const = new_Const(tarval_or(tv3, tv1));
4888 set_Or_left(or, new_and);
4889 set_Or_right(or, new_const);
4891 /* check for more */
4893 } /* transform_node_Or_bf_store */
4896 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4898 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4899 ir_mode *mode = get_irn_mode(or);
4900 ir_node *shl, *shr, *block;
4901 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4904 if (! mode_is_int(mode))
4907 shl = get_binop_left(or);
4908 shr = get_binop_right(or);
4917 } else if (!is_Shl(shl)) {
4919 } else if (!is_Shr(shr)) {
4922 x = get_Shl_left(shl);
4923 if (x != get_Shr_left(shr))
4926 c1 = get_Shl_right(shl);
4927 c2 = get_Shr_right(shr);
4928 if (is_Const(c1) && is_Const(c2)) {
4929 tv1 = get_Const_tarval(c1);
4930 if (! tarval_is_long(tv1))
4933 tv2 = get_Const_tarval(c2);
4934 if (! tarval_is_long(tv2))
4937 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4938 != (int) get_mode_size_bits(mode))
4941 /* yet, condition met */
4942 block = get_nodes_block(or);
4944 n = new_r_Rotl(block, x, c1, mode);
4946 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4953 rotval = sub; /* a Rot right is not supported, so use a rot left */
4954 } else if (is_Sub(c2)) {
4960 if (get_Sub_right(sub) != v)
4963 c1 = get_Sub_left(sub);
4967 tv1 = get_Const_tarval(c1);
4968 if (! tarval_is_long(tv1))
4971 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4974 /* yet, condition met */
4975 block = get_nodes_block(or);
4977 n = new_r_Rotl(block, x, rotval, mode);
4979 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4981 } /* transform_node_Or_Rotl */
4986 static ir_node *transform_node_Or(ir_node *n) {
4987 ir_node *c, *oldn = n;
4988 ir_node *a = get_Or_left(n);
4989 ir_node *b = get_Or_right(n);
4992 if (is_Not(a) && is_Not(b)) {
4993 /* ~a | ~b = ~(a&b) */
4994 ir_node *block = get_nodes_block(n);
4996 mode = get_irn_mode(n);
4999 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
5000 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
5001 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
5005 /* we can evaluate 2 Projs of the same Cmp */
5006 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
5007 ir_node *pred_a = get_Proj_pred(a);
5008 ir_node *pred_b = get_Proj_pred(b);
5009 if (pred_a == pred_b) {
5010 dbg_info *dbgi = get_irn_dbg_info(n);
5011 ir_node *block = get_nodes_block(pred_a);
5012 pn_Cmp pn_a = get_Proj_proj(a);
5013 pn_Cmp pn_b = get_Proj_proj(b);
5014 /* yes, we can simply calculate with pncs */
5015 pn_Cmp new_pnc = pn_a | pn_b;
5017 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
5021 mode = get_irn_mode(n);
5022 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
5024 n = transform_node_Or_bf_store(n);
5025 n = transform_node_Or_Rotl(n);
5029 n = transform_bitwise_distributive(n, transform_node_Or);
5032 } /* transform_node_Or */
5036 static ir_node *transform_node(ir_node *n);
5039 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5041 * Should be moved to reassociation?
5043 static ir_node *transform_node_shift(ir_node *n) {
5044 ir_node *left, *right;
5046 tarval *tv1, *tv2, *res;
5047 ir_node *in[2], *irn, *block;
5049 left = get_binop_left(n);
5051 /* different operations */
5052 if (get_irn_op(left) != get_irn_op(n))
5055 right = get_binop_right(n);
5056 tv1 = value_of(right);
5057 if (tv1 == tarval_bad)
5060 tv2 = value_of(get_binop_right(left));
5061 if (tv2 == tarval_bad)
5064 res = tarval_add(tv1, tv2);
5065 mode = get_irn_mode(n);
5067 /* beware: a simple replacement works only, if res < modulo shift */
5069 int modulo_shf = get_mode_modulo_shift(mode);
5070 if (modulo_shf > 0) {
5071 tarval *modulo = new_tarval_from_long(modulo_shf,
5072 get_tarval_mode(res));
5074 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5076 /* shifting too much */
5077 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5079 ir_node *block = get_nodes_block(n);
5080 dbg_info *dbgi = get_irn_dbg_info(n);
5081 ir_mode *smode = get_irn_mode(right);
5082 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5083 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5086 return new_Const(get_mode_null(mode));
5090 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5093 /* ok, we can replace it */
5094 block = get_nodes_block(n);
5096 in[0] = get_binop_left(left);
5097 in[1] = new_Const(res);
5099 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5101 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5103 return transform_node(irn);
5104 } /* transform_node_shift */
5107 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5109 * - and, or, xor instead of &
5110 * - Shl, Shr, Shrs, rotl instead of >>
5111 * (with a special case for Or/Xor + Shrs)
5113 static ir_node *transform_node_bitop_shift(ir_node *n) {
5115 ir_node *right = get_binop_right(n);
5116 ir_mode *mode = get_irn_mode(n);
5117 ir_node *bitop_left;
5118 ir_node *bitop_right;
5129 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5131 if (!is_Const(right))
5134 left = get_binop_left(n);
5135 op_left = get_irn_op(left);
5136 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5139 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5140 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5141 /* TODO: test if sign bit is affectes */
5145 bitop_right = get_binop_right(left);
5146 if (!is_Const(bitop_right))
5149 bitop_left = get_binop_left(left);
5151 block = get_nodes_block(n);
5152 dbgi = get_irn_dbg_info(n);
5153 tv1 = get_Const_tarval(bitop_right);
5154 tv2 = get_Const_tarval(right);
5156 assert(get_tarval_mode(tv1) == mode);
5159 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5160 tv_shift = tarval_shl(tv1, tv2);
5161 } else if(is_Shr(n)) {
5162 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5163 tv_shift = tarval_shr(tv1, tv2);
5164 } else if(is_Shrs(n)) {
5165 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5166 tv_shift = tarval_shrs(tv1, tv2);
5169 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5170 tv_shift = tarval_rotl(tv1, tv2);
5173 assert(get_tarval_mode(tv_shift) == mode);
5174 new_const = new_Const(tv_shift);
5176 if (op_left == op_And) {
5177 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5178 } else if(op_left == op_Or) {
5179 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5181 assert(op_left == op_Eor);
5182 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5190 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5192 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5193 * (also with x >>s c1 when c1>=c2)
5195 static ir_node *transform_node_shl_shr(ir_node *n) {
5197 ir_node *right = get_binop_right(n);
5213 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5215 if (!is_Const(right))
5218 left = get_binop_left(n);
5219 mode = get_irn_mode(n);
5220 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5221 ir_node *shr_right = get_binop_right(left);
5223 if (!is_Const(shr_right))
5226 x = get_binop_left(left);
5227 tv_shr = get_Const_tarval(shr_right);
5228 tv_shl = get_Const_tarval(right);
5230 if (is_Shrs(left)) {
5231 /* shrs variant only allowed if c1 >= c2 */
5232 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5235 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5238 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5240 tv_mask = tarval_shl(tv_mask, tv_shl);
5241 } else if(is_Shr(n) && is_Shl(left)) {
5242 ir_node *shl_right = get_Shl_right(left);
5244 if (!is_Const(shl_right))
5247 x = get_Shl_left(left);
5248 tv_shr = get_Const_tarval(right);
5249 tv_shl = get_Const_tarval(shl_right);
5251 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5252 tv_mask = tarval_shr(tv_mask, tv_shr);
5257 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5258 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5261 assert(tv_mask != tarval_bad);
5262 assert(get_tarval_mode(tv_mask) == mode);
5264 irg = get_irn_irg(n);
5265 block = get_nodes_block(n);
5266 dbgi = get_irn_dbg_info(n);
5268 pnc = tarval_cmp(tv_shl, tv_shr);
5269 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5270 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5271 new_const = new_Const(tv_shift);
5273 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5275 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5278 assert(pnc == pn_Cmp_Gt);
5279 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5280 new_const = new_Const(tv_shift);
5281 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5284 new_const = new_Const(tv_mask);
5285 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5293 static ir_node *transform_node_Shr(ir_node *n) {
5294 ir_node *c, *oldn = n;
5295 ir_node *left = get_Shr_left(n);
5296 ir_node *right = get_Shr_right(n);
5297 ir_mode *mode = get_irn_mode(n);
5299 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5300 n = transform_node_shift(n);
5303 n = transform_node_shl_shr(n);
5305 n = transform_node_bitop_shift(n);
5308 } /* transform_node_Shr */
5313 static ir_node *transform_node_Shrs(ir_node *n) {
5314 ir_node *c, *oldn = n;
5315 ir_node *a = get_Shrs_left(n);
5316 ir_node *b = get_Shrs_right(n);
5317 ir_mode *mode = get_irn_mode(n);
5319 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5320 n = transform_node_shift(n);
5323 n = transform_node_bitop_shift(n);
5326 } /* transform_node_Shrs */
5331 static ir_node *transform_node_Shl(ir_node *n) {
5332 ir_node *c, *oldn = n;
5333 ir_node *a = get_Shl_left(n);
5334 ir_node *b = get_Shl_right(n);
5335 ir_mode *mode = get_irn_mode(n);
5337 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5338 n = transform_node_shift(n);
5341 n = transform_node_shl_shr(n);
5343 n = transform_node_bitop_shift(n);
5346 } /* transform_node_Shl */
5351 static ir_node *transform_node_Rotl(ir_node *n) {
5352 ir_node *c, *oldn = n;
5353 ir_node *a = get_Rotl_left(n);
5354 ir_node *b = get_Rotl_right(n);
5355 ir_mode *mode = get_irn_mode(n);
5357 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5358 n = transform_node_shift(n);
5361 n = transform_node_bitop_shift(n);
5364 } /* transform_node_Rotl */
5369 static ir_node *transform_node_Conv(ir_node *n) {
5370 ir_node *c, *oldn = n;
5371 ir_mode *mode = get_irn_mode(n);
5372 ir_node *a = get_Conv_op(n);
5374 if (mode != mode_b && is_const_Phi(a)) {
5375 /* Do NOT optimize mode_b Conv's, this leads to remaining
5376 * Phib nodes later, because the conv_b_lower operation
5377 * is instantly reverted, when it tries to insert a Convb.
5379 c = apply_conv_on_phi(a, mode);
5381 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5386 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5387 return new_r_Unknown(current_ir_graph, mode);
5390 if (mode_is_reference(mode) &&
5391 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5393 ir_node *l = get_Add_left(a);
5394 ir_node *r = get_Add_right(a);
5395 dbg_info *dbgi = get_irn_dbg_info(a);
5396 ir_node *block = get_nodes_block(n);
5398 ir_node *lop = get_Conv_op(l);
5399 if(get_irn_mode(lop) == mode) {
5400 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5401 n = new_rd_Add(dbgi, block, lop, r, mode);
5406 ir_node *rop = get_Conv_op(r);
5407 if(get_irn_mode(rop) == mode) {
5408 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5409 n = new_rd_Add(dbgi, block, l, rop, mode);
5416 } /* transform_node_Conv */
5419 * Remove dead blocks and nodes in dead blocks
5420 * in keep alive list. We do not generate a new End node.
5422 static ir_node *transform_node_End(ir_node *n) {
5423 int i, j, n_keepalives = get_End_n_keepalives(n);
5426 NEW_ARR_A(ir_node *, in, n_keepalives);
5428 for (i = j = 0; i < n_keepalives; ++i) {
5429 ir_node *ka = get_End_keepalive(n, i);
5431 if (! is_Block_dead(ka)) {
5435 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5437 } else if (is_Bad(ka)) {
5438 /* no need to keep Bad */
5443 if (j != n_keepalives)
5444 set_End_keepalives(n, j, in);
5446 } /* transform_node_End */
5448 bool is_negated_value(ir_node *a, ir_node *b)
5450 if (is_Minus(a) && get_Minus_op(a) == b)
5452 if (is_Minus(b) && get_Minus_op(b) == a)
5454 if (is_Sub(a) && is_Sub(b)) {
5455 ir_node *a_left = get_Sub_left(a);
5456 ir_node *a_right = get_Sub_right(a);
5457 ir_node *b_left = get_Sub_left(b);
5458 ir_node *b_right = get_Sub_right(b);
5460 if (a_left == b_right && a_right == b_left)
5468 * Optimize a Mux into some simpler cases.
5470 static ir_node *transform_node_Mux(ir_node *n) {
5471 ir_node *oldn = n, *sel = get_Mux_sel(n);
5472 ir_mode *mode = get_irn_mode(n);
5473 ir_node *t = get_Mux_true(n);
5474 ir_node *f = get_Mux_false(n);
5475 ir_graph *irg = current_ir_graph;
5477 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5481 ir_node* block = get_nodes_block(n);
5483 ir_node* c1 = get_Mux_sel(t);
5484 ir_node* t1 = get_Mux_true(t);
5485 ir_node* f1 = get_Mux_false(t);
5487 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5488 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5489 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5494 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5495 } else if (f == t1) {
5496 /* Mux(cond0, Mux(cond1, x, y), x) */
5497 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5498 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5499 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5504 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5506 } else if (is_Mux(f)) {
5507 ir_node* block = get_nodes_block(n);
5509 ir_node* c1 = get_Mux_sel(f);
5510 ir_node* t1 = get_Mux_true(f);
5511 ir_node* f1 = get_Mux_false(f);
5513 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5514 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5515 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5520 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5521 } else if (t == f1) {
5522 /* Mux(cond0, x, Mux(cond1, y, x)) */
5523 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5524 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5525 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5530 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5534 /* first normalization step: move a possible zero to the false case */
5536 ir_node *cmp = get_Proj_pred(sel);
5539 if (is_Const(t) && is_Const_null(t)) {
5542 /* Mux(x, 0, y) => Mux(x, y, 0) */
5543 pn_Cmp pnc = get_Proj_proj(sel);
5544 sel = new_r_Proj(get_nodes_block(cmp), cmp, mode_b,
5545 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5546 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, t, f, mode);
5554 /* note: after normalization, false can only happen on default */
5555 if (mode == mode_b) {
5556 dbg_info *dbg = get_irn_dbg_info(n);
5557 ir_node *block = get_nodes_block(n);
5560 tarval *tv_t = get_Const_tarval(t);
5561 if (tv_t == tarval_b_true) {
5563 /* Muxb(sel, true, false) = sel */
5564 assert(get_Const_tarval(f) == tarval_b_false);
5565 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5568 /* Muxb(sel, true, x) = Or(sel, x) */
5569 n = new_rd_Or(dbg, block, sel, f, mode_b);
5570 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5574 } else if (is_Const(f)) {
5575 tarval *tv_f = get_Const_tarval(f);
5576 if (tv_f == tarval_b_true) {
5577 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5578 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5579 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5580 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5583 /* Muxb(sel, x, false) = And(sel, x) */
5584 assert(tv_f == tarval_b_false);
5585 n = new_rd_And(dbg, block, sel, t, mode_b);
5586 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5592 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5593 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5594 tarval *a = get_Const_tarval(t);
5595 tarval *b = get_Const_tarval(f);
5598 if (tarval_is_one(a) && tarval_is_null(b)) {
5599 ir_node *block = get_nodes_block(n);
5600 ir_node *conv = new_r_Conv(block, sel, mode);
5602 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5604 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5605 ir_node *block = get_nodes_block(n);
5606 ir_node *not_ = new_r_Not(block, sel, mode_b);
5607 ir_node *conv = new_r_Conv(block, not_, mode);
5609 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5612 /* TODO: it's not really clear if that helps in general or should be moved
5613 * to backend, especially with the MUX->Conv transformation above */
5614 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5615 diff = tarval_sub(a, b, NULL);
5618 diff = tarval_sub(b, a, NULL);
5622 if (diff == get_tarval_one(mode)) {
5623 dbg_info *dbg = get_irn_dbg_info(n);
5624 ir_node *block = get_nodes_block(n);
5625 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5626 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5627 n = new_rd_Mux(dbg, block, sel, f, t, mode);
5628 n = new_rd_Add(dbg, block, n, new_Const(min), mode);
5634 ir_node *cmp = get_Proj_pred(sel);
5635 long pn = get_Proj_proj(sel);
5638 * Note: normalization puts the constant on the right side,
5639 * so we check only one case.
5641 * Note further that these optimization work even for floating point
5642 * with NaN's because -NaN == NaN.
5643 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5647 ir_node *cmp_r = get_Cmp_right(cmp);
5648 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5649 ir_node *block = get_nodes_block(n);
5650 ir_node *cmp_l = get_Cmp_left(cmp);
5652 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5655 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5656 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5658 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5659 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5660 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5662 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5663 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5665 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5666 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5667 n = new_rd_Minus(get_irn_dbg_info(n), block, n, mode);
5668 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5673 if (mode_is_int(mode)) {
5675 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5676 /* Mux((a & b) != 0, c, 0) */
5677 ir_node *and_r = get_And_right(cmp_l);
5680 if (and_r == t && f == cmp_r) {
5681 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5682 if (pn == pn_Cmp_Lg) {
5683 /* Mux((a & 2^C) != 0, 2^C, 0) */
5685 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5687 /* Mux((a & 2^C) == 0, 2^C, 0) */
5688 n = new_rd_Eor(get_irn_dbg_info(n),
5689 block, cmp_l, t, mode);
5690 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5695 if (is_Shl(and_r)) {
5696 ir_node *shl_l = get_Shl_left(and_r);
5697 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5698 if (and_r == t && f == cmp_r) {
5699 if (pn == pn_Cmp_Lg) {
5700 /* (a & (1 << n)) != 0, (1 << n), 0) */
5702 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5704 /* (a & (1 << n)) == 0, (1 << n), 0) */
5705 n = new_rd_Eor(get_irn_dbg_info(n),
5706 block, cmp_l, t, mode);
5707 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5713 and_l = get_And_left(cmp_l);
5714 if (is_Shl(and_l)) {
5715 ir_node *shl_l = get_Shl_left(and_l);
5716 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5717 if (and_l == t && f == cmp_r) {
5718 if (pn == pn_Cmp_Lg) {
5719 /* ((1 << n) & a) != 0, (1 << n), 0) */
5721 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5723 /* ((1 << n) & a) == 0, (1 << n), 0) */
5724 n = new_rd_Eor(get_irn_dbg_info(n),
5725 block, cmp_l, t, mode);
5726 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5739 } /* transform_node_Mux */
5742 * optimize Sync nodes that have other syncs as input we simply add the inputs
5743 * of the other sync to our own inputs
5745 static ir_node *transform_node_Sync(ir_node *n) {
5746 int arity = get_Sync_n_preds(n);
5749 for (i = 0; i < arity;) {
5750 ir_node *pred = get_Sync_pred(n, i);
5754 if (!is_Sync(pred)) {
5762 pred_arity = get_Sync_n_preds(pred);
5763 for (j = 0; j < pred_arity; ++j) {
5764 ir_node *pred_pred = get_Sync_pred(pred, j);
5769 add_irn_n(n, pred_pred);
5773 if (get_Sync_pred(n, k) == pred_pred) break;
5778 /* rehash the sync node */
5779 add_identities(current_ir_graph->value_table, n);
5782 } /* transform_node_Sync */
5785 * optimize a trampoline Call into a direct Call
5787 static ir_node *transform_node_Call(ir_node *call) {
5788 ir_node *callee = get_Call_ptr(call);
5789 ir_node *adr, *mem, *res, *bl, **in;
5790 ir_type *ctp, *mtp, *tp;
5793 int i, n_res, n_param;
5796 if (! is_Proj(callee))
5798 callee = get_Proj_pred(callee);
5799 if (! is_Builtin(callee))
5801 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5804 mem = get_Call_mem(call);
5806 if (skip_Proj(mem) == callee) {
5807 /* memory is routed to the trampoline, skip */
5808 mem = get_Builtin_mem(callee);
5811 /* build a new call type */
5812 mtp = get_Call_type(call);
5813 tdb = get_type_dbg_info(mtp);
5815 n_res = get_method_n_ress(mtp);
5816 n_param = get_method_n_params(mtp);
5817 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5819 for (i = 0; i < n_res; ++i)
5820 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5822 NEW_ARR_A(ir_node *, in, n_param + 1);
5824 /* FIXME: we don't need a new pointer type in every step */
5825 tp = get_irg_frame_type(current_ir_graph);
5826 tp = new_type_pointer(tp);
5827 set_method_param_type(ctp, 0, tp);
5829 in[0] = get_Builtin_param(callee, 2);
5830 for (i = 0; i < n_param; ++i) {
5831 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5832 in[i + 1] = get_Call_param(call, i);
5834 var = get_method_variadicity(mtp);
5835 set_method_variadicity(ctp, var);
5836 if (var == variadicity_variadic) {
5837 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
5839 /* When we resolve a trampoline, the function must be called by a this-call */
5840 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5841 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5843 adr = get_Builtin_param(callee, 1);
5845 db = get_irn_dbg_info(call);
5846 bl = get_nodes_block(call);
5848 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5849 if (get_irn_pinned(call) == op_pin_state_floats)
5850 set_irn_pinned(res, op_pin_state_floats);
5852 } /* transform_node_Call */
5855 * Tries several [inplace] [optimizing] transformations and returns an
5856 * equivalent node. The difference to equivalent_node() is that these
5857 * transformations _do_ generate new nodes, and thus the old node must
5858 * not be freed even if the equivalent node isn't the old one.
5860 static ir_node *transform_node(ir_node *n) {
5864 * Transform_node is the only "optimizing transformation" that might
5865 * return a node with a different opcode. We iterate HERE until fixpoint
5866 * to get the final result.
5870 if (n->op->ops.transform_node != NULL)
5871 n = n->op->ops.transform_node(n);
5872 } while (oldn != n);
5875 } /* transform_node */
5878 * Sets the default transform node operation for an ir_op_ops.
5880 * @param code the opcode for the default operation
5881 * @param ops the operations initialized
5886 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5890 ops->transform_node = transform_node_##a; \
5892 #define CASE_PROJ(a) \
5894 ops->transform_node_Proj = transform_node_Proj_##a; \
5896 #define CASE_PROJ_EX(a) \
5898 ops->transform_node = transform_node_##a; \
5899 ops->transform_node_Proj = transform_node_Proj_##a; \
5908 CASE_PROJ_EX(DivMod);
5943 } /* firm_set_default_transform_node */
5946 /* **************** Common Subexpression Elimination **************** */
5948 /** The size of the hash table used, should estimate the number of nodes
5950 #define N_IR_NODES 512
5952 /** Compares the attributes of two Const nodes. */
5953 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5954 return (get_Const_tarval(a) != get_Const_tarval(b))
5955 || (get_Const_type(a) != get_Const_type(b));
5956 } /* node_cmp_attr_Const */
5958 /** Compares the attributes of two Proj nodes. */
5959 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5960 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5961 } /* node_cmp_attr_Proj */
5963 /** Compares the attributes of two Filter nodes. */
5964 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5965 return get_Filter_proj(a) != get_Filter_proj(b);
5966 } /* node_cmp_attr_Filter */
5968 /** Compares the attributes of two Alloc nodes. */
5969 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5970 const alloc_attr *pa = get_irn_alloc_attr(a);
5971 const alloc_attr *pb = get_irn_alloc_attr(b);
5972 return (pa->where != pb->where) || (pa->type != pb->type);
5973 } /* node_cmp_attr_Alloc */
5975 /** Compares the attributes of two Free nodes. */
5976 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5977 const free_attr *pa = get_irn_free_attr(a);
5978 const free_attr *pb = get_irn_free_attr(b);
5979 return (pa->where != pb->where) || (pa->type != pb->type);
5980 } /* node_cmp_attr_Free */
5982 /** Compares the attributes of two SymConst nodes. */
5983 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5984 const symconst_attr *pa = get_irn_symconst_attr(a);
5985 const symconst_attr *pb = get_irn_symconst_attr(b);
5986 return (pa->kind != pb->kind)
5987 || (pa->sym.type_p != pb->sym.type_p)
5988 || (pa->tp != pb->tp);
5989 } /* node_cmp_attr_SymConst */
5991 /** Compares the attributes of two Call nodes. */
5992 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5993 const call_attr *pa = get_irn_call_attr(a);
5994 const call_attr *pb = get_irn_call_attr(b);
5995 return (pa->type != pb->type)
5996 || (pa->tail_call != pb->tail_call);
5997 } /* node_cmp_attr_Call */
5999 /** Compares the attributes of two Sel nodes. */
6000 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
6001 const ir_entity *a_ent = get_Sel_entity(a);
6002 const ir_entity *b_ent = get_Sel_entity(b);
6005 (a_ent->kind != b_ent->kind) ||
6006 (a_ent->name != b_ent->name) ||
6007 (a_ent->owner != b_ent->owner) ||
6008 (a_ent->ld_name != b_ent->ld_name) ||
6009 (a_ent->type != b_ent->type);
6011 /* Matze: inlining of functions can produce 2 entities with same type,
6013 return a_ent != b_ent;
6014 } /* node_cmp_attr_Sel */
6016 /** Compares the attributes of two Phi nodes. */
6017 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
6018 /* we can only enter this function if both nodes have the same number of inputs,
6019 hence it is enough to check if one of them is a Phi0 */
6021 /* check the Phi0 pos attribute */
6022 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
6025 } /* node_cmp_attr_Phi */
6027 /** Compares the attributes of two Conv nodes. */
6028 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
6029 return get_Conv_strict(a) != get_Conv_strict(b);
6030 } /* node_cmp_attr_Conv */
6032 /** Compares the attributes of two Cast nodes. */
6033 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
6034 return get_Cast_type(a) != get_Cast_type(b);
6035 } /* node_cmp_attr_Cast */
6037 /** Compares the attributes of two Load nodes. */
6038 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
6039 if (get_Load_volatility(a) == volatility_is_volatile ||
6040 get_Load_volatility(b) == volatility_is_volatile)
6041 /* NEVER do CSE on volatile Loads */
6043 /* do not CSE Loads with different alignment. Be conservative. */
6044 if (get_Load_align(a) != get_Load_align(b))
6047 return get_Load_mode(a) != get_Load_mode(b);
6048 } /* node_cmp_attr_Load */
6050 /** Compares the attributes of two Store nodes. */
6051 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
6052 /* do not CSE Stores with different alignment. Be conservative. */
6053 if (get_Store_align(a) != get_Store_align(b))
6056 /* NEVER do CSE on volatile Stores */
6057 return (get_Store_volatility(a) == volatility_is_volatile ||
6058 get_Store_volatility(b) == volatility_is_volatile);
6059 } /* node_cmp_attr_Store */
6061 /** Compares two exception attributes */
6062 static int node_cmp_exception(ir_node *a, ir_node *b) {
6063 const except_attr *ea = get_irn_except_attr(a);
6064 const except_attr *eb = get_irn_except_attr(b);
6066 return ea->pin_state != eb->pin_state;
6069 #define node_cmp_attr_Bound node_cmp_exception
6071 /** Compares the attributes of two Div nodes. */
6072 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
6073 const divmod_attr *ma = get_irn_divmod_attr(a);
6074 const divmod_attr *mb = get_irn_divmod_attr(b);
6075 return ma->exc.pin_state != mb->exc.pin_state ||
6076 ma->resmode != mb->resmode ||
6077 ma->no_remainder != mb->no_remainder;
6078 } /* node_cmp_attr_Div */
6080 /** Compares the attributes of two DivMod nodes. */
6081 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
6082 const divmod_attr *ma = get_irn_divmod_attr(a);
6083 const divmod_attr *mb = get_irn_divmod_attr(b);
6084 return ma->exc.pin_state != mb->exc.pin_state ||
6085 ma->resmode != mb->resmode;
6086 } /* node_cmp_attr_DivMod */
6088 /** Compares the attributes of two Mod nodes. */
6089 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
6090 const divmod_attr *ma = get_irn_divmod_attr(a);
6091 const divmod_attr *mb = get_irn_divmod_attr(b);
6092 return ma->exc.pin_state != mb->exc.pin_state ||
6093 ma->resmode != mb->resmode;
6094 } /* node_cmp_attr_Mod */
6096 /** Compares the attributes of two Quot nodes. */
6097 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
6098 const divmod_attr *ma = get_irn_divmod_attr(a);
6099 const divmod_attr *mb = get_irn_divmod_attr(b);
6100 return ma->exc.pin_state != mb->exc.pin_state ||
6101 ma->resmode != mb->resmode;
6102 } /* node_cmp_attr_Quot */
6104 /** Compares the attributes of two Confirm nodes. */
6105 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
6106 /* no need to compare the bound, as this is a input */
6107 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
6108 } /* node_cmp_attr_Confirm */
6110 /** Compares the attributes of two Builtin nodes. */
6111 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b) {
6112 const builtin_attr *ma = get_irn_builtin_attr(a);
6113 const builtin_attr *mb = get_irn_builtin_attr(b);
6115 /* no need to compare the type, equal kind means equal type */
6116 return ma->kind != mb->kind;
6117 } /* node_cmp_attr_Builtin */
6119 /** Compares the attributes of two ASM nodes. */
6120 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
6122 const ir_asm_constraint *ca;
6123 const ir_asm_constraint *cb;
6126 if (get_ASM_text(a) != get_ASM_text(b))
6129 /* Should we really check the constraints here? Should be better, but is strange. */
6130 n = get_ASM_n_input_constraints(a);
6131 if (n != get_ASM_n_input_constraints(b))
6134 ca = get_ASM_input_constraints(a);
6135 cb = get_ASM_input_constraints(b);
6136 for (i = 0; i < n; ++i) {
6137 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6141 n = get_ASM_n_output_constraints(a);
6142 if (n != get_ASM_n_output_constraints(b))
6145 ca = get_ASM_output_constraints(a);
6146 cb = get_ASM_output_constraints(b);
6147 for (i = 0; i < n; ++i) {
6148 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6152 n = get_ASM_n_clobbers(a);
6153 if (n != get_ASM_n_clobbers(b))
6156 cla = get_ASM_clobbers(a);
6157 clb = get_ASM_clobbers(b);
6158 for (i = 0; i < n; ++i) {
6159 if (cla[i] != clb[i])
6163 } /* node_cmp_attr_ASM */
6165 /** Compares the inexistent attributes of two Dummy nodes. */
6166 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6174 * Set the default node attribute compare operation for an ir_op_ops.
6176 * @param code the opcode for the default operation
6177 * @param ops the operations initialized
6182 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6186 ops->node_cmp_attr = node_cmp_attr_##a; \
6219 } /* firm_set_default_node_cmp_attr */
6222 * Compare function for two nodes in the value table. Gets two
6223 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6225 int identities_cmp(const void *elt, const void *key) {
6226 ir_node *a = (ir_node *)elt;
6227 ir_node *b = (ir_node *)key;
6230 if (a == b) return 0;
6232 if ((get_irn_op(a) != get_irn_op(b)) ||
6233 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6235 /* compare if a's in and b's in are of equal length */
6236 irn_arity_a = get_irn_intra_arity(a);
6237 if (irn_arity_a != get_irn_intra_arity(b))
6240 if (get_irn_pinned(a) == op_pin_state_pinned) {
6241 /* for pinned nodes, the block inputs must be equal */
6242 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6244 } else if (! get_opt_global_cse()) {
6245 /* for block-local CSE both nodes must be in the same MacroBlock */
6246 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6250 /* compare a->in[0..ins] with b->in[0..ins] */
6251 for (i = 0; i < irn_arity_a; ++i) {
6252 ir_node *pred_a = get_irn_intra_n(a, i);
6253 ir_node *pred_b = get_irn_intra_n(b, i);
6254 if (pred_a != pred_b) {
6255 /* if both predecessors are CSE neutral they might be different */
6256 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6262 * here, we already now that the nodes are identical except their
6265 if (a->op->ops.node_cmp_attr)
6266 return a->op->ops.node_cmp_attr(a, b);
6269 } /* identities_cmp */
6272 * Calculate a hash value of a node.
6274 * @param node The IR-node
6276 unsigned ir_node_hash(const ir_node *node) {
6277 return node->op->ops.hash(node);
6278 } /* ir_node_hash */
6281 pset *new_identities(void) {
6282 return new_pset(identities_cmp, N_IR_NODES);
6283 } /* new_identities */
6285 void del_identities(pset *value_table) {
6286 del_pset(value_table);
6287 } /* del_identities */
6289 /* Normalize a node by putting constants (and operands with larger
6290 * node index) on the right (operator side). */
6291 void ir_normalize_node(ir_node *n) {
6292 if (is_op_commutative(get_irn_op(n))) {
6293 ir_node *l = get_binop_left(n);
6294 ir_node *r = get_binop_right(n);
6296 /* For commutative operators perform a OP b == b OP a but keep
6297 * constants on the RIGHT side. This helps greatly in some
6298 * optimizations. Moreover we use the idx number to make the form
6300 if (!operands_are_normalized(l, r)) {
6301 set_binop_left(n, r);
6302 set_binop_right(n, l);
6306 } /* ir_normalize_node */
6309 * Update the nodes after a match in the value table. If both nodes have
6310 * the same MacroBlock but different Blocks, we must ensure that the node
6311 * with the dominating Block (the node that is near to the MacroBlock header
6312 * is stored in the table.
6313 * Because a MacroBlock has only one "non-exception" flow, we don't need
6314 * dominance info here: We known, that one block must dominate the other and
6315 * following the only block input will allow to find it.
6317 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6318 ir_node *known_blk, *new_block, *block, *mbh;
6320 if (get_opt_global_cse()) {
6321 /* Block inputs are meaning less */
6324 known_blk = get_irn_n(known_irn, -1);
6325 new_block = get_irn_n(new_ir_node, -1);
6326 if (known_blk == new_block) {
6327 /* already in the same block */
6331 * We expect the typical case when we built the graph. In that case, the
6332 * known_irn is already the upper one, so checking this should be faster.
6335 mbh = get_Block_MacroBlock(new_block);
6337 if (block == known_blk) {
6338 /* ok, we have found it: known_block dominates new_block as expected */
6343 * We have reached the MacroBlock header NOT founding
6344 * the known_block. new_block must dominate known_block.
6347 set_irn_n(known_irn, -1, new_block);
6350 assert(get_Block_n_cfgpreds(block) == 1);
6351 block = get_Block_cfgpred_block(block, 0);
6353 } /* update_value_table */
6356 * Return the canonical node computing the same value as n.
6357 * Looks up the node in a hash table, enters it in the table
6358 * if it isn't there yet.
6360 * @param value_table the HashSet containing all nodes in the
6362 * @param n the node to look up
6364 * @return a node that computes the same value as n or n if no such
6365 * node could be found
6367 ir_node *identify_remember(pset *value_table, ir_node *n) {
6370 if (!value_table) return n;
6372 ir_normalize_node(n);
6373 /* lookup or insert in hash table with given hash key. */
6374 nn = pset_insert(value_table, n, ir_node_hash(n));
6377 update_known_irn(nn, n);
6379 /* n is reachable again */
6380 edges_node_revival(nn, get_irn_irg(nn));
6384 } /* identify_remember */
6387 * During construction we set the op_pin_state_pinned flag in the graph right when the
6388 * optimization is performed. The flag turning on procedure global cse could
6389 * be changed between two allocations. This way we are safe.
6391 * @param value_table The value table
6392 * @param n The node to lookup
6394 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6397 n = identify_remember(value_table, n);
6398 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6399 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6401 } /* identify_cons */
6403 /* Add a node to the identities value table. */
6404 void add_identities(pset *value_table, ir_node *node) {
6405 if (get_opt_cse() && is_no_Block(node))
6406 identify_remember(value_table, node);
6407 } /* add_identities */
6409 /* Visit each node in the value table of a graph. */
6410 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6412 ir_graph *rem = current_ir_graph;
6414 current_ir_graph = irg;
6415 foreach_pset(irg->value_table, node)
6417 current_ir_graph = rem;
6418 } /* visit_all_identities */
6421 * Garbage in, garbage out. If a node has a dead input, i.e., the
6422 * Bad node is input to the node, return the Bad node.
6424 static ir_node *gigo(ir_node *node) {
6426 ir_op *op = get_irn_op(node);
6428 /* remove garbage blocks by looking at control flow that leaves the block
6429 and replacing the control flow by Bad. */
6430 if (get_irn_mode(node) == mode_X) {
6431 ir_node *block = get_nodes_block(skip_Proj(node));
6433 /* Don't optimize nodes in immature blocks. */
6434 if (!get_Block_matured(block))
6436 /* Don't optimize End, may have Bads. */
6437 if (op == op_End) return node;
6439 if (is_Block(block)) {
6440 if (is_Block_dead(block)) {
6441 /* control flow from dead block is dead */
6445 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6446 if (!is_Bad(get_irn_n(block, i)))
6450 ir_graph *irg = get_irn_irg(block);
6451 /* the start block is never dead */
6452 if (block != get_irg_start_block(irg)
6453 && block != get_irg_end_block(irg)) {
6455 * Do NOT kill control flow without setting
6456 * the block to dead of bad things can happen:
6457 * We get a Block that is not reachable be irg_block_walk()
6458 * but can be found by irg_walk()!
6460 set_Block_dead(block);
6467 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6468 blocks predecessors is dead. */
6469 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6470 irn_arity = get_irn_arity(node);
6473 * Beware: we can only read the block of a non-floating node.
6475 if (is_irn_pinned_in_irg(node) &&
6476 is_Block_dead(get_nodes_block(skip_Proj(node))))
6479 for (i = 0; i < irn_arity; i++) {
6480 ir_node *pred = get_irn_n(node, i);
6485 /* Propagating Unknowns here seems to be a bad idea, because
6486 sometimes we need a node as a input and did not want that
6488 However, it might be useful to move this into a later phase
6489 (if you think that optimizing such code is useful). */
6490 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6491 return new_Unknown(get_irn_mode(node));
6496 /* With this code we violate the agreement that local_optimize
6497 only leaves Bads in Block, Phi and Tuple nodes. */
6498 /* If Block has only Bads as predecessors it's garbage. */
6499 /* If Phi has only Bads as predecessors it's garbage. */
6500 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6501 irn_arity = get_irn_arity(node);
6502 for (i = 0; i < irn_arity; i++) {
6503 if (!is_Bad(get_irn_n(node, i))) break;
6505 if (i == irn_arity) node = new_Bad();
6512 * These optimizations deallocate nodes from the obstack.
6513 * It can only be called if it is guaranteed that no other nodes
6514 * reference this one, i.e., right after construction of a node.
6516 * @param n The node to optimize
6518 * current_ir_graph must be set to the graph of the node!
6520 ir_node *optimize_node(ir_node *n) {
6523 ir_opcode iro = get_irn_opcode(n);
6525 /* Always optimize Phi nodes: part of the construction. */
6526 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6528 /* constant expression evaluation / constant folding */
6529 if (get_opt_constant_folding()) {
6530 /* neither constants nor Tuple values can be evaluated */
6531 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6532 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6533 int old_fp_mode = tarval_fp_ops_enabled();
6535 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6537 /* try to evaluate */
6538 tv = computed_value(n);
6539 if (tv != tarval_bad) {
6541 ir_type *old_tp = get_irn_type(n);
6542 int i, arity = get_irn_arity(n);
6546 * Try to recover the type of the new expression.
6548 for (i = 0; i < arity && !old_tp; ++i)
6549 old_tp = get_irn_type(get_irn_n(n, i));
6552 * we MUST copy the node here temporary, because it's still needed
6553 * for DBG_OPT_CSTEVAL
6555 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6556 oldn = alloca(node_size);
6558 memcpy(oldn, n, node_size);
6559 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6561 /* ARG, copy the in array, we need it for statistics */
6562 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6564 /* note the inplace edges module */
6565 edges_node_deleted(n, current_ir_graph);
6567 /* evaluation was successful -- replace the node. */
6568 irg_kill_node(current_ir_graph, n);
6571 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6572 set_Const_type(nw, old_tp);
6573 DBG_OPT_CSTEVAL(oldn, nw);
6574 tarval_enable_fp_ops(old_fp_mode);
6577 tarval_enable_fp_ops(old_fp_mode);
6581 /* remove unnecessary nodes */
6582 if (get_opt_algebraic_simplification() ||
6583 (iro == iro_Phi) || /* always optimize these nodes. */
6585 (iro == iro_Proj) ||
6586 (iro == iro_Block) ) /* Flags tested local. */
6587 n = equivalent_node(n);
6589 /* Common Subexpression Elimination.
6591 * Checks whether n is already available.
6592 * The block input is used to distinguish different subexpressions. Right
6593 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6594 * subexpressions within a block.
6597 n = identify_cons(current_ir_graph->value_table, n);
6600 edges_node_deleted(oldn, current_ir_graph);
6602 /* We found an existing, better node, so we can deallocate the old node. */
6603 irg_kill_node(current_ir_graph, oldn);
6607 /* Some more constant expression evaluation that does not allow to
6609 iro = get_irn_opcode(n);
6610 if (get_opt_algebraic_simplification() ||
6611 (iro == iro_Cond) ||
6612 (iro == iro_Proj)) /* Flags tested local. */
6613 n = transform_node(n);
6615 /* Remove nodes with dead (Bad) input.
6616 Run always for transformation induced Bads. */
6619 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6620 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6622 n = identify_remember(current_ir_graph->value_table, o);
6628 } /* optimize_node */
6632 * These optimizations never deallocate nodes (in place). This can cause dead
6633 * nodes lying on the obstack. Remove these by a dead node elimination,
6634 * i.e., a copying garbage collection.
6636 ir_node *optimize_in_place_2(ir_node *n) {
6639 ir_opcode iro = get_irn_opcode(n);
6641 if (!get_opt_optimize() && !is_Phi(n)) return n;
6643 /* constant expression evaluation / constant folding */
6644 if (get_opt_constant_folding()) {
6645 /* neither constants nor Tuple values can be evaluated */
6646 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6647 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6648 int old_fp_mode = tarval_fp_ops_enabled();
6650 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6651 /* try to evaluate */
6652 tv = computed_value(n);
6653 if (tv != tarval_bad) {
6654 /* evaluation was successful -- replace the node. */
6655 ir_type *old_tp = get_irn_type(n);
6656 int i, arity = get_irn_arity(n);
6659 * Try to recover the type of the new expression.
6661 for (i = 0; i < arity && !old_tp; ++i)
6662 old_tp = get_irn_type(get_irn_n(n, i));
6666 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6667 set_Const_type(n, old_tp);
6669 DBG_OPT_CSTEVAL(oldn, n);
6670 tarval_enable_fp_ops(old_fp_mode);
6673 tarval_enable_fp_ops(old_fp_mode);
6677 /* remove unnecessary nodes */
6678 if (get_opt_constant_folding() ||
6679 (iro == iro_Phi) || /* always optimize these nodes. */
6680 (iro == iro_Id) || /* ... */
6681 (iro == iro_Proj) || /* ... */
6682 (iro == iro_Block) ) /* Flags tested local. */
6683 n = equivalent_node(n);
6685 /** common subexpression elimination **/
6686 /* Checks whether n is already available. */
6687 /* The block input is used to distinguish different subexpressions. Right
6688 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6689 subexpressions within a block. */
6690 if (get_opt_cse()) {
6692 n = identify_remember(current_ir_graph->value_table, o);
6697 /* Some more constant expression evaluation. */
6698 iro = get_irn_opcode(n);
6699 if (get_opt_constant_folding() ||
6700 (iro == iro_Cond) ||
6701 (iro == iro_Proj)) /* Flags tested local. */
6702 n = transform_node(n);
6704 /* Remove nodes with dead (Bad) input.
6705 Run always for transformation induced Bads. */
6708 /* Now we can verify the node, as it has no dead inputs any more. */
6711 /* Now we have a legal, useful node. Enter it in hash table for cse.
6712 Blocks should be unique anyways. (Except the successor of start:
6713 is cse with the start block!) */
6714 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6716 n = identify_remember(current_ir_graph->value_table, o);
6722 } /* optimize_in_place_2 */
6725 * Wrapper for external use, set proper status bits after optimization.
6727 ir_node *optimize_in_place(ir_node *n) {
6728 /* Handle graph state */
6729 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6731 if (get_opt_global_cse())
6732 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6733 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6734 set_irg_outs_inconsistent(current_ir_graph);
6736 /* FIXME: Maybe we could also test whether optimizing the node can
6737 change the control graph. */
6738 set_irg_doms_inconsistent(current_ir_graph);
6739 return optimize_in_place_2(n);
6740 } /* optimize_in_place */
6743 * Calculate a hash value of a Const node.
6745 static unsigned hash_Const(const ir_node *node) {
6748 /* special value for const, as they only differ in their tarval. */
6749 h = HASH_PTR(node->attr.con.tv);
6755 * Calculate a hash value of a SymConst node.
6757 static unsigned hash_SymConst(const ir_node *node) {
6760 /* all others are pointers */
6761 h = HASH_PTR(node->attr.symc.sym.type_p);
6764 } /* hash_SymConst */
6767 * Set the default hash operation in an ir_op_ops.
6769 * @param code the opcode for the default operation
6770 * @param ops the operations initialized
6775 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6779 ops->hash = hash_##a; \
6782 /* hash function already set */
6783 if (ops->hash != NULL)
6790 /* use input/mode default hash if no function was given */
6791 ops->hash = firm_default_hash;
6799 * Sets the default operation for an ir_ops.
6801 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6802 ops = firm_set_default_hash(code, ops);
6803 ops = firm_set_default_computed_value(code, ops);
6804 ops = firm_set_default_equivalent_node(code, ops);
6805 ops = firm_set_default_transform_node(code, ops);
6806 ops = firm_set_default_node_cmp_attr(code, ops);
6807 ops = firm_set_default_get_type(code, ops);
6808 ops = firm_set_default_get_type_attr(code, ops);
6809 ops = firm_set_default_get_entity_attr(code, ops);
6812 } /* firm_set_default_operations */