2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
31 #include "irgraph_t.h"
32 #include "iredges_t.h"
39 #include "dbginfo_t.h"
40 #include "iropt_dbg.h"
45 #include "opt_confirms.h"
46 #include "opt_polymorphy.h"
51 /* Make types visible to allow most efficient access */
55 * Returns the tarval of a Const node or tarval_bad for all other nodes.
57 static tarval *default_value_of(const ir_node *n) {
59 return get_Const_tarval(n); /* might return tarval_bad */
64 value_of_func value_of_ptr = default_value_of;
66 /* * Set a new value_of function. */
67 void set_value_of_func(value_of_func func) {
71 value_of_ptr = default_value_of;
75 * Return the value of a Constant.
77 static tarval *computed_value_Const(const ir_node *n) {
78 return get_Const_tarval(n);
79 } /* computed_value_Const */
82 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
84 static tarval *computed_value_SymConst(const ir_node *n) {
88 switch (get_SymConst_kind(n)) {
89 case symconst_type_size:
90 type = get_SymConst_type(n);
91 if (get_type_state(type) == layout_fixed)
92 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
94 case symconst_type_align:
95 type = get_SymConst_type(n);
96 if (get_type_state(type) == layout_fixed)
97 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
99 case symconst_ofs_ent:
100 ent = get_SymConst_entity(n);
101 type = get_entity_owner(ent);
102 if (get_type_state(type) == layout_fixed)
103 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
109 } /* computed_value_SymConst */
112 * Return the value of an Add.
114 static tarval *computed_value_Add(const ir_node *n) {
115 ir_node *a = get_Add_left(n);
116 ir_node *b = get_Add_right(n);
118 tarval *ta = value_of(a);
119 tarval *tb = value_of(b);
121 if ((ta != tarval_bad) && (tb != tarval_bad))
122 return tarval_add(ta, tb);
125 } /* computed_value_Add */
128 * Return the value of a Sub.
129 * Special case: a - a
131 static tarval *computed_value_Sub(const ir_node *n) {
132 ir_mode *mode = get_irn_mode(n);
133 ir_node *a = get_Sub_left(n);
134 ir_node *b = get_Sub_right(n);
139 if (! mode_is_float(mode)) {
142 return get_mode_null(mode);
148 if ((ta != tarval_bad) && (tb != tarval_bad))
149 return tarval_sub(ta, tb, mode);
152 } /* computed_value_Sub */
155 * Return the value of a Carry.
156 * Special : a op 0, 0 op b
158 static tarval *computed_value_Carry(const ir_node *n) {
159 ir_node *a = get_binop_left(n);
160 ir_node *b = get_binop_right(n);
161 ir_mode *m = get_irn_mode(n);
163 tarval *ta = value_of(a);
164 tarval *tb = value_of(b);
166 if ((ta != tarval_bad) && (tb != tarval_bad)) {
168 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
170 if (tarval_is_null(ta) || tarval_is_null(tb))
171 return get_mode_null(m);
174 } /* computed_value_Carry */
177 * Return the value of a Borrow.
180 static tarval *computed_value_Borrow(const ir_node *n) {
181 ir_node *a = get_binop_left(n);
182 ir_node *b = get_binop_right(n);
183 ir_mode *m = get_irn_mode(n);
185 tarval *ta = value_of(a);
186 tarval *tb = value_of(b);
188 if ((ta != tarval_bad) && (tb != tarval_bad)) {
189 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
190 } else if (tarval_is_null(ta)) {
191 return get_mode_null(m);
194 } /* computed_value_Borrow */
197 * Return the value of an unary Minus.
199 static tarval *computed_value_Minus(const ir_node *n) {
200 ir_node *a = get_Minus_op(n);
201 tarval *ta = value_of(a);
203 if (ta != tarval_bad)
204 return tarval_neg(ta);
207 } /* computed_value_Minus */
210 * Return the value of a Mul.
212 static tarval *computed_value_Mul(const ir_node *n) {
213 ir_node *a = get_Mul_left(n);
214 ir_node *b = get_Mul_right(n);
217 tarval *ta = value_of(a);
218 tarval *tb = value_of(b);
220 mode = get_irn_mode(n);
221 if (mode != get_irn_mode(a)) {
222 /* n * n = 2n bit multiplication */
223 ta = tarval_convert_to(ta, mode);
224 tb = tarval_convert_to(tb, mode);
227 if (ta != tarval_bad && tb != tarval_bad) {
228 return tarval_mul(ta, tb);
230 /* a * 0 != 0 if a == NaN or a == Inf */
231 if (!mode_is_float(mode)) {
232 /* a*0 = 0 or 0*b = 0 */
233 if (ta == get_mode_null(mode))
235 if (tb == get_mode_null(mode))
240 } /* computed_value_Mul */
243 * Return the value of an Abs.
245 static tarval *computed_value_Abs(const ir_node *n) {
246 ir_node *a = get_Abs_op(n);
247 tarval *ta = value_of(a);
249 if (ta != tarval_bad)
250 return tarval_abs(ta);
253 } /* computed_value_Abs */
256 * Return the value of an And.
257 * Special case: a & 0, 0 & b
259 static tarval *computed_value_And(const ir_node *n) {
260 ir_node *a = get_And_left(n);
261 ir_node *b = get_And_right(n);
263 tarval *ta = value_of(a);
264 tarval *tb = value_of(b);
266 if ((ta != tarval_bad) && (tb != tarval_bad)) {
267 return tarval_and (ta, tb);
269 if (tarval_is_null(ta)) return ta;
270 if (tarval_is_null(tb)) return tb;
273 } /* computed_value_And */
276 * Return the value of an Or.
277 * Special case: a | 1...1, 1...1 | b
279 static tarval *computed_value_Or(const ir_node *n) {
280 ir_node *a = get_Or_left(n);
281 ir_node *b = get_Or_right(n);
283 tarval *ta = value_of(a);
284 tarval *tb = value_of(b);
286 if ((ta != tarval_bad) && (tb != tarval_bad)) {
287 return tarval_or (ta, tb);
289 if (tarval_is_all_one(ta)) return ta;
290 if (tarval_is_all_one(tb)) return tb;
293 } /* computed_value_Or */
296 * Return the value of an Eor.
298 static tarval *computed_value_Eor(const ir_node *n) {
299 ir_node *a = get_Eor_left(n);
300 ir_node *b = get_Eor_right(n);
305 return get_mode_null(get_irn_mode(n));
310 if ((ta != tarval_bad) && (tb != tarval_bad)) {
311 return tarval_eor (ta, tb);
314 } /* computed_value_Eor */
317 * Return the value of a Not.
319 static tarval *computed_value_Not(const ir_node *n) {
320 ir_node *a = get_Not_op(n);
321 tarval *ta = value_of(a);
323 if (ta != tarval_bad)
324 return tarval_not(ta);
327 } /* computed_value_Not */
330 * Return the value of a Shl.
332 static tarval *computed_value_Shl(const ir_node *n) {
333 ir_node *a = get_Shl_left(n);
334 ir_node *b = get_Shl_right(n);
336 tarval *ta = value_of(a);
337 tarval *tb = value_of(b);
339 if ((ta != tarval_bad) && (tb != tarval_bad)) {
340 return tarval_shl (ta, tb);
343 } /* computed_value_Shl */
346 * Return the value of a Shr.
348 static tarval *computed_value_Shr(const ir_node *n) {
349 ir_node *a = get_Shr_left(n);
350 ir_node *b = get_Shr_right(n);
352 tarval *ta = value_of(a);
353 tarval *tb = value_of(b);
355 if ((ta != tarval_bad) && (tb != tarval_bad)) {
356 return tarval_shr (ta, tb);
359 } /* computed_value_Shr */
362 * Return the value of a Shrs.
364 static tarval *computed_value_Shrs(const ir_node *n) {
365 ir_node *a = get_Shrs_left(n);
366 ir_node *b = get_Shrs_right(n);
368 tarval *ta = value_of(a);
369 tarval *tb = value_of(b);
371 if ((ta != tarval_bad) && (tb != tarval_bad)) {
372 return tarval_shrs (ta, tb);
375 } /* computed_value_Shrs */
378 * Return the value of a Rotl.
380 static tarval *computed_value_Rotl(const ir_node *n) {
381 ir_node *a = get_Rotl_left(n);
382 ir_node *b = get_Rotl_right(n);
384 tarval *ta = value_of(a);
385 tarval *tb = value_of(b);
387 if ((ta != tarval_bad) && (tb != tarval_bad)) {
388 return tarval_rotl(ta, tb);
391 } /* computed_value_Rotl */
394 * Return the value of a Conv.
396 static tarval *computed_value_Conv(const ir_node *n) {
397 ir_node *a = get_Conv_op(n);
398 tarval *ta = value_of(a);
400 if (ta != tarval_bad)
401 return tarval_convert_to(ta, get_irn_mode(n));
404 } /* computed_value_Conv */
407 * Calculate the value of a Mux: can be evaluated, if the
408 * sel and the right input are known.
410 static tarval *computed_value_Mux(const ir_node *n) {
411 ir_node *sel = get_Mux_sel(n);
412 tarval *ts = value_of(sel);
414 if (ts == get_tarval_b_true()) {
415 ir_node *v = get_Mux_true(n);
418 else if (ts == get_tarval_b_false()) {
419 ir_node *v = get_Mux_false(n);
423 } /* computed_value_Mux */
426 * Calculate the value of a Confirm: can be evaluated,
427 * if it has the form Confirm(x, '=', Const).
429 static tarval *computed_value_Confirm(const ir_node *n) {
431 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
432 * Do NOT optimize them away (CondEval wants them), so wait until
433 * remove_confirm is activated.
435 if (get_opt_remove_confirm()) {
436 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
437 tarval *tv = value_of(get_Confirm_bound(n));
438 if (tv != tarval_bad)
442 return value_of(get_Confirm_value(n));
443 } /* computed_value_Confirm */
446 * Return the value of a Proj(Cmp).
448 * This performs a first step of unreachable code elimination.
449 * Proj can not be computed, but folding a Cmp above the Proj here is
450 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
452 * There are several case where we can evaluate a Cmp node, see later.
454 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
455 ir_node *a = get_Proj_pred(n);
456 ir_node *aa = get_Cmp_left(a);
457 ir_node *ab = get_Cmp_right(a);
458 long proj_nr = get_Proj_proj(n);
461 * BEWARE: a == a is NOT always True for floating Point values, as
462 * NaN != NaN is defined, so we must check this here.
465 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
468 /* This is a trick with the bits used for encoding the Cmp
469 Proj numbers, the following statement is not the same:
470 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
471 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
474 tarval *taa = value_of(aa);
475 tarval *tab = value_of(ab);
476 ir_mode *mode = get_irn_mode(aa);
479 * The predecessors of Cmp are target values. We can evaluate
482 if ((taa != tarval_bad) && (tab != tarval_bad)) {
483 /* strange checks... */
484 pn_Cmp flags = tarval_cmp(taa, tab);
485 if (flags != pn_Cmp_False) {
486 return new_tarval_from_long (proj_nr & flags, mode_b);
489 /* for integer values, we can check against MIN/MAX */
490 else if (mode_is_int(mode)) {
491 /* MIN <=/> x. This results in true/false. */
492 if (taa == get_mode_min(mode)) {
493 /* a compare with the MIN value */
494 if (proj_nr == pn_Cmp_Le)
495 return get_tarval_b_true();
496 else if (proj_nr == pn_Cmp_Gt)
497 return get_tarval_b_false();
499 /* x >=/< MIN. This results in true/false. */
501 if (tab == get_mode_min(mode)) {
502 /* a compare with the MIN value */
503 if (proj_nr == pn_Cmp_Ge)
504 return get_tarval_b_true();
505 else if (proj_nr == pn_Cmp_Lt)
506 return get_tarval_b_false();
508 /* MAX >=/< x. This results in true/false. */
509 else if (taa == get_mode_max(mode)) {
510 if (proj_nr == pn_Cmp_Ge)
511 return get_tarval_b_true();
512 else if (proj_nr == pn_Cmp_Lt)
513 return get_tarval_b_false();
515 /* x <=/> MAX. This results in true/false. */
516 else if (tab == get_mode_max(mode)) {
517 if (proj_nr == pn_Cmp_Le)
518 return get_tarval_b_true();
519 else if (proj_nr == pn_Cmp_Gt)
520 return get_tarval_b_false();
524 * The predecessors are Allocs or (void*)(0) constants. Allocs never
525 * return NULL, they raise an exception. Therefore we can predict
529 ir_node *aaa = skip_Proj(aa);
530 ir_node *aba = skip_Proj(ab);
532 if ( ( (/* aa is ProjP and aaa is Alloc */
534 && mode_is_reference(get_irn_mode(aa))
536 && ( (/* ab is NULL */
537 mode_is_reference(get_irn_mode(ab))
538 && tarval_is_null(tab))
539 || (/* ab is other Alloc */
541 && mode_is_reference(get_irn_mode(ab))
544 || (/* aa is NULL and aba is Alloc */
545 mode_is_reference(get_irn_mode(aa))
546 && tarval_is_null(taa)
548 && mode_is_reference(get_irn_mode(ab))
551 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
554 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
555 } /* computed_value_Proj_Cmp */
558 * Return the value of a floating point Quot.
560 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
561 tarval *ta = value_of(a);
562 tarval *tb = value_of(b);
564 /* cannot optimize 0 / b = 0 because of NaN */
565 if (ta != tarval_bad && tb != tarval_bad)
566 return tarval_quo(ta, tb);
568 } /* do_computed_value_Quot */
571 * Calculate the value of an integer Div of two nodes.
572 * Special case: 0 / b
574 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
575 tarval *ta = value_of(a);
577 const ir_node *dummy;
579 /* Compute c1 / c2 or 0 / a, a != 0 */
580 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
581 return ta; /* 0 / b == 0 */
583 if (ta != tarval_bad && tb != tarval_bad)
584 return tarval_div(ta, tb);
586 } /* do_computed_value_Div */
589 * Calculate the value of an integer Mod of two nodes.
590 * Special case: a % 1
592 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
593 tarval *ta = value_of(a);
594 tarval *tb = value_of(b);
596 /* Compute a % 1 or c1 % c2 */
597 if (tarval_is_one(tb))
598 return get_mode_null(get_irn_mode(a));
599 if (ta != tarval_bad && tb != tarval_bad)
600 return tarval_mod(ta, tb);
602 } /* do_computed_value_Mod */
605 * Return the value of a Proj(DivMod).
607 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
608 long proj_nr = get_Proj_proj(n);
610 /* compute either the Div or the Mod part */
611 if (proj_nr == pn_DivMod_res_div) {
612 const ir_node *a = get_Proj_pred(n);
613 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
614 } else if (proj_nr == pn_DivMod_res_mod) {
615 const ir_node *a = get_Proj_pred(n);
616 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
619 } /* computed_value_Proj_DivMod */
622 * Return the value of a Proj(Div).
624 static tarval *computed_value_Proj_Div(const ir_node *n) {
625 long proj_nr = get_Proj_proj(n);
627 if (proj_nr == pn_Div_res) {
628 const ir_node *a = get_Proj_pred(n);
629 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
632 } /* computed_value_Proj_Div */
635 * Return the value of a Proj(Mod).
637 static tarval *computed_value_Proj_Mod(const ir_node *n) {
638 long proj_nr = get_Proj_proj(n);
640 if (proj_nr == pn_Mod_res) {
641 const ir_node *a = get_Proj_pred(n);
642 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
645 } /* computed_value_Proj_Mod */
648 * Return the value of a Proj(Quot).
650 static tarval *computed_value_Proj_Quot(const ir_node *n) {
651 long proj_nr = get_Proj_proj(n);
653 if (proj_nr == pn_Quot_res) {
654 const ir_node *a = get_Proj_pred(n);
655 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
658 } /* computed_value_Proj_Quot */
661 * Return the value of a Proj.
663 static tarval *computed_value_Proj(const ir_node *proj) {
664 ir_node *n = get_Proj_pred(proj);
666 if (n->op->ops.computed_value_Proj != NULL)
667 return n->op->ops.computed_value_Proj(proj);
669 } /* computed_value_Proj */
672 * If the parameter n can be computed, return its value, else tarval_bad.
673 * Performs constant folding.
675 * @param n The node this should be evaluated
677 tarval *computed_value(const ir_node *n) {
678 if (n->op->ops.computed_value)
679 return n->op->ops.computed_value(n);
681 } /* computed_value */
684 * Set the default computed_value evaluator in an ir_op_ops.
686 * @param code the opcode for the default operation
687 * @param ops the operations initialized
692 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
696 ops->computed_value = computed_value_##a; \
698 #define CASE_PROJ(a) \
700 ops->computed_value_Proj = computed_value_Proj_##a; \
737 } /* firm_set_default_computed_value */
740 * Returns a equivalent block for another block.
741 * If the block has only one predecessor, this is
742 * the equivalent one. If the only predecessor of a block is
743 * the block itself, this is a dead block.
745 * If both predecessors of a block are the branches of a binary
746 * Cond, the equivalent block is Cond's block.
748 * If all predecessors of a block are bad or lies in a dead
749 * block, the current block is dead as well.
751 * Note, that blocks are NEVER turned into Bad's, instead
752 * the dead_block flag is set. So, never test for is_Bad(block),
753 * always use is_dead_Block(block).
755 static ir_node *equivalent_node_Block(ir_node *n)
760 /* don't optimize dead or labeled blocks */
761 if (is_Block_dead(n) || has_Block_label(n))
764 n_preds = get_Block_n_cfgpreds(n);
766 /* The Block constructor does not call optimize, but mature_immBlock()
767 calls the optimization. */
768 assert(get_Block_matured(n));
770 /* Straightening: a single entry Block following a single exit Block
771 can be merged, if it is not the Start block. */
772 /* !!! Beware, all Phi-nodes of n must have been optimized away.
773 This should be true, as the block is matured before optimize is called.
774 But what about Phi-cycles with the Phi0/Id that could not be resolved?
775 Remaining Phi nodes are just Ids. */
777 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
780 ir_node *predblock = get_nodes_block(pred);
781 if (predblock == oldn) {
782 /* Jmp jumps into the block it is in -- deal self cycle. */
783 n = set_Block_dead(n);
784 DBG_OPT_DEAD_BLOCK(oldn, n);
785 } else if (get_opt_control_flow_straightening()) {
787 DBG_OPT_STG(oldn, n);
789 } else if (is_Cond(pred)) {
790 ir_node *predblock = get_nodes_block(pred);
791 if (predblock == oldn) {
792 /* Jmp jumps into the block it is in -- deal self cycle. */
793 n = set_Block_dead(n);
794 DBG_OPT_DEAD_BLOCK(oldn, n);
797 } else if ((n_preds == 2) &&
798 (get_opt_control_flow_weak_simplification())) {
799 /* Test whether Cond jumps twice to this block
800 * The more general case which more than 2 predecessors is handles
801 * in optimize_cf(), we handle only this special case for speed here.
803 ir_node *a = get_Block_cfgpred(n, 0);
804 ir_node *b = get_Block_cfgpred(n, 1);
806 if (is_Proj(a) && is_Proj(b)) {
807 ir_node *cond = get_Proj_pred(a);
809 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
810 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
811 /* Also a single entry Block following a single exit Block. Phis have
812 twice the same operand and will be optimized away. */
813 n = get_nodes_block(cond);
814 DBG_OPT_IFSIM1(oldn, a, b, n);
817 } else if (get_opt_unreachable_code() &&
818 (n != get_irg_start_block(current_ir_graph)) &&
819 (n != get_irg_end_block(current_ir_graph)) ) {
822 /* If all inputs are dead, this block is dead too, except if it is
823 the start or end block. This is one step of unreachable code
825 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
826 ir_node *pred = get_Block_cfgpred(n, i);
829 if (is_Bad(pred)) continue;
830 pred_blk = get_nodes_block(skip_Proj(pred));
832 if (is_Block_dead(pred_blk)) continue;
835 /* really found a living input */
840 n = set_Block_dead(n);
841 DBG_OPT_DEAD_BLOCK(oldn, n);
846 } /* equivalent_node_Block */
849 * Returns a equivalent node for a Jmp, a Bad :-)
850 * Of course this only happens if the Block of the Jmp is dead.
852 static ir_node *equivalent_node_Jmp(ir_node *n) {
855 /* unreachable code elimination */
856 if (is_Block_dead(get_nodes_block(n))) {
857 n = get_irg_bad(current_ir_graph);
858 DBG_OPT_DEAD_BLOCK(oldn, n);
861 } /* equivalent_node_Jmp */
863 /** Raise is handled in the same way as Jmp. */
864 #define equivalent_node_Raise equivalent_node_Jmp
867 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
868 See transform_node_Proj_Cond(). */
871 * Optimize operations that are commutative and have neutral 0,
872 * so a op 0 = 0 op a = a.
874 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
877 ir_node *a = get_binop_left(n);
878 ir_node *b = get_binop_right(n);
883 /* After running compute_node there is only one constant predecessor.
884 Find this predecessors value and remember the other node: */
885 if ((tv = value_of(a)) != tarval_bad) {
887 } else if ((tv = value_of(b)) != tarval_bad) {
892 /* If this predecessors constant value is zero, the operation is
893 * unnecessary. Remove it.
895 * Beware: If n is a Add, the mode of on and n might be different
896 * which happens in this rare construction: NULL + 3.
897 * Then, a Conv would be needed which we cannot include here.
899 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
902 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
906 } /* equivalent_node_neutral_zero */
909 * Eor is commutative and has neutral 0.
911 static ir_node *equivalent_node_Eor(ir_node *n) {
916 n = equivalent_node_neutral_zero(n);
917 if (n != oldn) return n;
920 b = get_Eor_right(n);
923 ir_node *aa = get_Eor_left(a);
924 ir_node *ab = get_Eor_right(a);
927 /* (a ^ b) ^ a -> b */
929 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
931 } else if (ab == b) {
932 /* (a ^ b) ^ b -> a */
934 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
939 ir_node *ba = get_Eor_left(b);
940 ir_node *bb = get_Eor_right(b);
943 /* a ^ (a ^ b) -> b */
945 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
947 } else if (bb == a) {
948 /* a ^ (b ^ a) -> b */
950 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
958 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
960 * The second one looks strange, but this construct
961 * is used heavily in the LCC sources :-).
963 * Beware: The Mode of an Add may be different than the mode of its
964 * predecessors, so we could not return a predecessors in all cases.
966 static ir_node *equivalent_node_Add(ir_node *n) {
968 ir_node *left, *right;
969 ir_mode *mode = get_irn_mode(n);
971 n = equivalent_node_neutral_zero(n);
975 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
976 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
979 left = get_Add_left(n);
980 right = get_Add_right(n);
983 if (get_Sub_right(left) == right) {
986 n = get_Sub_left(left);
987 if (mode == get_irn_mode(n)) {
988 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
994 if (get_Sub_right(right) == left) {
997 n = get_Sub_left(right);
998 if (mode == get_irn_mode(n)) {
999 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1005 } /* equivalent_node_Add */
1008 * optimize operations that are not commutative but have neutral 0 on left,
1011 static ir_node *equivalent_node_left_zero(ir_node *n) {
1014 ir_node *a = get_binop_left(n);
1015 ir_node *b = get_binop_right(n);
1016 tarval *tb = value_of(b);
1018 if (tarval_is_null(tb)) {
1021 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1024 } /* equivalent_node_left_zero */
1026 #define equivalent_node_Shl equivalent_node_left_zero
1027 #define equivalent_node_Shr equivalent_node_left_zero
1028 #define equivalent_node_Shrs equivalent_node_left_zero
1029 #define equivalent_node_Rotl equivalent_node_left_zero
1032 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1034 * The second one looks strange, but this construct
1035 * is used heavily in the LCC sources :-).
1037 * Beware: The Mode of a Sub may be different than the mode of its
1038 * predecessors, so we could not return a predecessors in all cases.
1040 static ir_node *equivalent_node_Sub(ir_node *n) {
1043 ir_mode *mode = get_irn_mode(n);
1046 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1047 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1050 b = get_Sub_right(n);
1053 /* Beware: modes might be different */
1054 if (tarval_is_null(tb)) {
1055 ir_node *a = get_Sub_left(n);
1056 if (mode == get_irn_mode(a)) {
1059 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1063 } /* equivalent_node_Sub */
1067 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1070 * -(-a) == a, but might overflow two times.
1071 * We handle it anyway here but the better way would be a
1072 * flag. This would be needed for Pascal for instance.
1074 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1076 ir_node *pred = get_unop_op(n);
1078 /* optimize symmetric unop */
1079 if (get_irn_op(pred) == get_irn_op(n)) {
1080 n = get_unop_op(pred);
1081 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1084 } /* equivalent_node_idempotent_unop */
1086 /** Optimize Not(Not(x)) == x. */
1087 #define equivalent_node_Not equivalent_node_idempotent_unop
1089 /** -(-x) == x ??? Is this possible or can --x raise an
1090 out of bounds exception if min =! max? */
1091 #define equivalent_node_Minus equivalent_node_idempotent_unop
1094 * Optimize a * 1 = 1 * a = a.
1096 static ir_node *equivalent_node_Mul(ir_node *n) {
1098 ir_node *a = get_Mul_left(n);
1100 /* we can handle here only the n * n = n bit cases */
1101 if (get_irn_mode(n) == get_irn_mode(a)) {
1102 ir_node *b = get_Mul_right(n);
1106 * Mul is commutative and has again an other neutral element.
1107 * Constants are place right, so check this case first.
1110 if (tarval_is_one(tv)) {
1112 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1115 if (tarval_is_one(tv)) {
1117 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1122 } /* equivalent_node_Mul */
1125 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1127 static ir_node *equivalent_node_Or(ir_node *n) {
1130 ir_node *a = get_Or_left(n);
1131 ir_node *b = get_Or_right(n);
1135 n = a; /* Or has it's own neutral element */
1136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1139 /* constants are cormalized to right, check this site first */
1141 if (tarval_is_null(tv)) {
1143 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1147 if (tarval_is_null(tv)) {
1149 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1154 } /* equivalent_node_Or */
1157 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1159 static ir_node *equivalent_node_And(ir_node *n) {
1162 ir_node *a = get_And_left(n);
1163 ir_node *b = get_And_right(n);
1167 n = a; /* And has it's own neutral element */
1168 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1171 /* constants are cormalized to right, check this site first */
1173 if (tarval_is_all_one(tv)) {
1175 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1179 if (tarval_is_all_one(tv)) {
1181 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1185 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1188 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1193 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1196 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1201 } /* equivalent_node_And */
1204 * Try to remove useless Conv's:
1206 static ir_node *equivalent_node_Conv(ir_node *n) {
1208 ir_node *a = get_Conv_op(n);
1210 ir_mode *n_mode = get_irn_mode(n);
1211 ir_mode *a_mode = get_irn_mode(a);
1214 if (n_mode == a_mode) { /* No Conv necessary */
1215 if (get_Conv_strict(n)) {
1218 /* neither Minus nor Abs nor Confirm change the precision,
1219 so we can "look-through" */
1222 p = get_Minus_op(p);
1223 } else if (is_Abs(p)) {
1225 } else if (is_Confirm(p)) {
1226 p = get_Confirm_value(p);
1232 if (is_Conv(p) && get_Conv_strict(p)) {
1233 /* we known already, that a_mode == n_mode, and neither
1234 Abs nor Minus change the mode, so the second Conv
1236 assert(get_irn_mode(p) == n_mode);
1238 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1242 ir_node *pred = get_Proj_pred(p);
1243 if (is_Load(pred)) {
1244 /* Loads always return with the exact precision of n_mode */
1245 assert(get_Load_mode(pred) == n_mode);
1247 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1250 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1251 pred = get_Proj_pred(pred);
1252 if (is_Start(pred)) {
1253 /* Arguments always return with the exact precision,
1254 as strictConv's are place before Call -- if the
1255 caller was compiled with the same setting.
1256 Otherwise, the semantics is probably still right. */
1257 assert(get_irn_mode(p) == n_mode);
1259 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1265 /* special case: the immediate predecessor is also a Conv */
1266 if (! get_Conv_strict(a)) {
1267 /* first one is not strict, kick it */
1269 a_mode = get_irn_mode(a);
1273 /* else both are strict conv, second is superfluous */
1275 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1283 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1284 ir_node *b = get_Conv_op(a);
1285 ir_mode *b_mode = get_irn_mode(b);
1287 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1288 /* both are strict conv */
1289 if (smaller_mode(a_mode, n_mode)) {
1290 /* both are strict, but the first is smaller, so
1291 the second cannot remove more precision, remove the
1293 set_Conv_strict(n, 0);
1296 if (n_mode == b_mode) {
1297 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1298 if (n_mode == mode_b) {
1299 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1300 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1302 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1303 if (values_in_mode(b_mode, a_mode)) {
1304 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1305 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1310 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1311 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1312 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1313 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1315 if (float_mantissa >= int_mantissa) {
1317 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1322 if (smaller_mode(b_mode, a_mode)) {
1323 if (get_Conv_strict(n))
1324 set_Conv_strict(b, 1);
1325 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1326 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1333 } /* equivalent_node_Conv */
1336 * A Cast may be removed if the type of the previous node
1337 * is already the type of the Cast.
1339 static ir_node *equivalent_node_Cast(ir_node *n) {
1341 ir_node *pred = get_Cast_op(n);
1343 if (get_irn_type(pred) == get_Cast_type(n)) {
1345 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1348 } /* equivalent_node_Cast */
1351 * - fold Phi-nodes, iff they have only one predecessor except
1354 static ir_node *equivalent_node_Phi(ir_node *n) {
1359 ir_node *first_val = NULL; /* to shutup gcc */
1361 if (!get_opt_normalize()) return n;
1363 n_preds = get_Phi_n_preds(n);
1365 block = get_nodes_block(n);
1366 if (is_Block_dead(block)) /* Control dead */
1367 return get_irg_bad(current_ir_graph);
1369 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1371 /* Find first non-self-referencing input */
1372 for (i = 0; i < n_preds; ++i) {
1373 first_val = get_Phi_pred(n, i);
1374 if ( (first_val != n) /* not self pointer */
1376 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1377 * predecessors. Then, Phi nodes in dead code might be removed, causing
1378 * nodes pointing to themself (Add's for instance).
1379 * This is really bad and causes endless recursions in several
1380 * code pathes, so we do NOT optimize such a code.
1381 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1382 * (and bad Phi predecessors), so live code is optimized later.
1384 && (! is_Bad(get_Block_cfgpred(block, i)))
1386 ) { /* value not dead */
1387 break; /* then found first value. */
1392 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1393 return get_irg_bad(current_ir_graph);
1396 /* search for rest of inputs, determine if any of these
1397 are non-self-referencing */
1398 while (++i < n_preds) {
1399 ir_node *scnd_val = get_Phi_pred(n, i);
1400 if ( (scnd_val != n)
1401 && (scnd_val != first_val)
1404 && (! is_Bad(get_Block_cfgpred(block, i)))
1412 /* Fold, if no multiple distinct non-self-referencing inputs */
1414 DBG_OPT_PHI(oldn, n);
1417 } /* equivalent_node_Phi */
1420 * Several optimizations:
1421 * - fold Sync-nodes, iff they have only one predecessor except
1424 static ir_node *equivalent_node_Sync(ir_node *n) {
1425 int arity = get_Sync_n_preds(n);
1428 for (i = 0; i < arity;) {
1429 ir_node *pred = get_Sync_pred(n, i);
1432 /* Remove Bad predecessors */
1439 /* Remove duplicate predecessors */
1445 if (get_Sync_pred(n, j) == pred) {
1453 if (arity == 0) return get_irg_bad(current_ir_graph);
1454 if (arity == 1) return get_Sync_pred(n, 0);
1456 } /* equivalent_node_Sync */
1459 * Optimize Proj(Tuple).
1461 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1462 ir_node *oldn = proj;
1463 ir_node *tuple = get_Proj_pred(proj);
1465 /* Remove the Tuple/Proj combination. */
1466 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1467 DBG_OPT_TUPLE(oldn, tuple, proj);
1470 } /* equivalent_node_Proj_Tuple */
1473 * Optimize a / 1 = a.
1475 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1476 ir_node *oldn = proj;
1477 ir_node *div = get_Proj_pred(proj);
1478 ir_node *b = get_Div_right(div);
1479 tarval *tb = value_of(b);
1481 /* Div is not commutative. */
1482 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1483 switch (get_Proj_proj(proj)) {
1485 proj = get_Div_mem(div);
1486 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1490 proj = get_Div_left(div);
1491 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1495 /* we cannot replace the exception Proj's here, this is done in
1496 transform_node_Proj_Div() */
1501 } /* equivalent_node_Proj_Div */
1504 * Optimize a / 1.0 = a.
1506 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1507 ir_node *oldn = proj;
1508 ir_node *quot = get_Proj_pred(proj);
1509 ir_node *b = get_Quot_right(quot);
1510 tarval *tb = value_of(b);
1512 /* Div is not commutative. */
1513 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1514 switch (get_Proj_proj(proj)) {
1516 proj = get_Quot_mem(quot);
1517 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1521 proj = get_Quot_left(quot);
1522 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1526 /* we cannot replace the exception Proj's here, this is done in
1527 transform_node_Proj_Quot() */
1532 } /* equivalent_node_Proj_Quot */
1535 * Optimize a / 1 = a.
1537 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1538 ir_node *oldn = proj;
1539 ir_node *divmod = get_Proj_pred(proj);
1540 ir_node *b = get_DivMod_right(divmod);
1541 tarval *tb = value_of(b);
1543 /* Div is not commutative. */
1544 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1545 switch (get_Proj_proj(proj)) {
1547 proj = get_DivMod_mem(divmod);
1548 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1551 case pn_DivMod_res_div:
1552 proj = get_DivMod_left(divmod);
1553 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1557 /* we cannot replace the exception Proj's here, this is done in
1558 transform_node_Proj_DivMod().
1559 Note further that the pn_DivMod_res_div case is handled in
1560 computed_value_Proj(). */
1565 } /* equivalent_node_Proj_DivMod */
1568 * Optimize CopyB(mem, x, x) into a Nop.
1570 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1571 ir_node *oldn = proj;
1572 ir_node *copyb = get_Proj_pred(proj);
1573 ir_node *a = get_CopyB_dst(copyb);
1574 ir_node *b = get_CopyB_src(copyb);
1577 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1578 switch (get_Proj_proj(proj)) {
1579 case pn_CopyB_M_regular:
1580 proj = get_CopyB_mem(copyb);
1581 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1584 case pn_CopyB_M_except:
1585 case pn_CopyB_X_except:
1586 DBG_OPT_EXC_REM(proj);
1587 proj = get_irg_bad(current_ir_graph);
1592 } /* equivalent_node_Proj_CopyB */
1595 * Optimize Bounds(idx, idx, upper) into idx.
1597 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1598 ir_node *oldn = proj;
1599 ir_node *bound = get_Proj_pred(proj);
1600 ir_node *idx = get_Bound_index(bound);
1601 ir_node *pred = skip_Proj(idx);
1604 if (idx == get_Bound_lower(bound))
1606 else if (is_Bound(pred)) {
1608 * idx was Bounds checked in the same MacroBlock previously,
1609 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1611 ir_node *lower = get_Bound_lower(bound);
1612 ir_node *upper = get_Bound_upper(bound);
1613 if (get_Bound_lower(pred) == lower &&
1614 get_Bound_upper(pred) == upper &&
1615 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1617 * One could expect that we simply return the previous
1618 * Bound here. However, this would be wrong, as we could
1619 * add an exception Proj to a new location then.
1620 * So, we must turn in into a tuple.
1626 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1627 switch (get_Proj_proj(proj)) {
1629 DBG_OPT_EXC_REM(proj);
1630 proj = get_Bound_mem(bound);
1632 case pn_Bound_X_except:
1633 DBG_OPT_EXC_REM(proj);
1634 proj = get_irg_bad(current_ir_graph);
1638 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1641 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1646 } /* equivalent_node_Proj_Bound */
1649 * Optimize an Exception Proj(Load) with a non-null address.
1651 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1652 if (get_opt_ldst_only_null_ptr_exceptions()) {
1653 if (get_irn_mode(proj) == mode_X) {
1654 ir_node *load = get_Proj_pred(proj);
1656 /* get the Load address */
1657 const ir_node *addr = get_Load_ptr(load);
1658 const ir_node *confirm;
1660 if (value_not_null(addr, &confirm)) {
1661 if (get_Proj_proj(proj) == pn_Load_X_except) {
1662 DBG_OPT_EXC_REM(proj);
1663 return get_irg_bad(current_ir_graph);
1669 } /* equivalent_node_Proj_Load */
1672 * Optimize an Exception Proj(Store) with a non-null address.
1674 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1675 if (get_opt_ldst_only_null_ptr_exceptions()) {
1676 if (get_irn_mode(proj) == mode_X) {
1677 ir_node *store = get_Proj_pred(proj);
1679 /* get the load/store address */
1680 const ir_node *addr = get_Store_ptr(store);
1681 const ir_node *confirm;
1683 if (value_not_null(addr, &confirm)) {
1684 if (get_Proj_proj(proj) == pn_Store_X_except) {
1685 DBG_OPT_EXC_REM(proj);
1686 return get_irg_bad(current_ir_graph);
1692 } /* equivalent_node_Proj_Store */
1695 * Does all optimizations on nodes that must be done on it's Proj's
1696 * because of creating new nodes.
1698 static ir_node *equivalent_node_Proj(ir_node *proj) {
1699 ir_node *n = get_Proj_pred(proj);
1701 if (get_irn_mode(proj) == mode_X) {
1702 if (is_Block_dead(get_nodes_block(n))) {
1703 /* Remove dead control flow -- early gigo(). */
1704 return get_irg_bad(current_ir_graph);
1707 if (n->op->ops.equivalent_node_Proj)
1708 return n->op->ops.equivalent_node_Proj(proj);
1710 } /* equivalent_node_Proj */
1715 static ir_node *equivalent_node_Id(ir_node *n) {
1722 DBG_OPT_ID(oldn, n);
1724 } /* equivalent_node_Id */
1729 static ir_node *equivalent_node_Mux(ir_node *n)
1731 ir_node *oldn = n, *sel = get_Mux_sel(n);
1732 tarval *ts = value_of(sel);
1734 /* Mux(true, f, t) == t */
1735 if (ts == tarval_b_true) {
1736 n = get_Mux_true(n);
1737 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1739 /* Mux(false, f, t) == f */
1740 else if (ts == tarval_b_false) {
1741 n = get_Mux_false(n);
1742 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1744 /* Mux(v, x, x) == x */
1745 else if (get_Mux_false(n) == get_Mux_true(n)) {
1746 n = get_Mux_true(n);
1747 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1749 else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1750 ir_node *cmp = get_Proj_pred(sel);
1751 long proj_nr = get_Proj_proj(sel);
1752 ir_node *f = get_Mux_false(n);
1753 ir_node *t = get_Mux_true(n);
1756 * Note further that these optimization work even for floating point
1757 * with NaN's because -NaN == NaN.
1758 * However, if +0 and -0 is handled differently, we cannot use the first one.
1761 ir_node *const cmp_l = get_Cmp_left(cmp);
1762 ir_node *const cmp_r = get_Cmp_right(cmp);
1766 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1767 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1769 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1776 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1777 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1779 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1786 * Note: normalization puts the constant on the right side,
1787 * so we check only one case.
1789 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1790 /* Mux(t CMP 0, X, t) */
1791 if (is_Minus(f) && get_Minus_op(f) == t) {
1792 /* Mux(t CMP 0, -t, t) */
1793 if (proj_nr == pn_Cmp_Eq) {
1794 /* Mux(t == 0, -t, t) ==> -t */
1796 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1797 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1798 /* Mux(t != 0, -t, t) ==> t */
1800 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1807 } /* equivalent_node_Mux */
1810 * Remove Confirm nodes if setting is on.
1811 * Replace Confirms(x, '=', Constlike) by Constlike.
1813 static ir_node *equivalent_node_Confirm(ir_node *n) {
1814 ir_node *pred = get_Confirm_value(n);
1815 pn_Cmp pnc = get_Confirm_cmp(n);
1817 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1819 * rare case: two identical Confirms one after another,
1820 * replace the second one with the first.
1823 pred = get_Confirm_value(n);
1824 pnc = get_Confirm_cmp(n);
1826 if (get_opt_remove_confirm())
1827 return get_Confirm_value(n);
1832 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1833 * perform no actual computation, as, e.g., the Id nodes. It does not create
1834 * new nodes. It is therefore safe to free n if the node returned is not n.
1835 * If a node returns a Tuple we can not just skip it. If the size of the
1836 * in array fits, we transform n into a tuple (e.g., Div).
1838 ir_node *equivalent_node(ir_node *n) {
1839 if (n->op->ops.equivalent_node)
1840 return n->op->ops.equivalent_node(n);
1842 } /* equivalent_node */
1845 * Sets the default equivalent node operation for an ir_op_ops.
1847 * @param code the opcode for the default operation
1848 * @param ops the operations initialized
1853 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1857 ops->equivalent_node = equivalent_node_##a; \
1859 #define CASE_PROJ(a) \
1861 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1903 } /* firm_set_default_equivalent_node */
1906 * Returns non-zero if a node is a Phi node
1907 * with all predecessors constant.
1909 static int is_const_Phi(ir_node *n) {
1912 if (! is_Phi(n) || get_irn_arity(n) == 0)
1914 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1915 if (! is_Const(get_irn_n(n, i)))
1919 } /* is_const_Phi */
1921 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1922 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1925 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1927 static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) {
1928 if (eval == tarval_sub) {
1929 tarval_sub_type func = (tarval_sub_type)eval;
1931 return func(a, b, mode);
1933 tarval_binop_type func = (tarval_binop_type)eval;
1940 * Apply an evaluator on a binop with a constant operators (and one Phi).
1942 * @param phi the Phi node
1943 * @param other the other operand
1944 * @param eval an evaluator function
1945 * @param mode the mode of the result, may be different from the mode of the Phi!
1946 * @param left if non-zero, other is the left operand, else the right
1948 * @return a new Phi node if the conversion was successful, NULL else
1950 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) {
1955 int i, n = get_irn_arity(phi);
1957 NEW_ARR_A(void *, res, n);
1959 for (i = 0; i < n; ++i) {
1960 pred = get_irn_n(phi, i);
1961 tv = get_Const_tarval(pred);
1962 tv = do_eval(eval, other, tv, mode);
1964 if (tv == tarval_bad) {
1965 /* folding failed, bad */
1971 for (i = 0; i < n; ++i) {
1972 pred = get_irn_n(phi, i);
1973 tv = get_Const_tarval(pred);
1974 tv = do_eval(eval, tv, other, mode);
1976 if (tv == tarval_bad) {
1977 /* folding failed, bad */
1983 irg = current_ir_graph;
1984 for (i = 0; i < n; ++i) {
1985 pred = get_irn_n(phi, i);
1986 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
1988 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1989 } /* apply_binop_on_phi */
1992 * Apply an evaluator on a binop with two constant Phi.
1994 * @param a the left Phi node
1995 * @param b the right Phi node
1996 * @param eval an evaluator function
1997 * @param mode the mode of the result, may be different from the mode of the Phi!
1999 * @return a new Phi node if the conversion was successful, NULL else
2001 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) {
2002 tarval *tv_l, *tv_r, *tv;
2008 if (get_nodes_block(a) != get_nodes_block(b))
2011 n = get_irn_arity(a);
2012 NEW_ARR_A(void *, res, n);
2014 for (i = 0; i < n; ++i) {
2015 pred = get_irn_n(a, i);
2016 tv_l = get_Const_tarval(pred);
2017 pred = get_irn_n(b, i);
2018 tv_r = get_Const_tarval(pred);
2019 tv = do_eval(eval, tv_l, tv_r, mode);
2021 if (tv == tarval_bad) {
2022 /* folding failed, bad */
2027 irg = current_ir_graph;
2028 for (i = 0; i < n; ++i) {
2029 pred = get_irn_n(a, i);
2030 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2032 return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode);
2033 } /* apply_binop_on_2_phis */
2036 * Apply an evaluator on a unop with a constant operator (a Phi).
2038 * @param phi the Phi node
2039 * @param eval an evaluator function
2041 * @return a new Phi node if the conversion was successful, NULL else
2043 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2049 int i, n = get_irn_arity(phi);
2051 NEW_ARR_A(void *, res, n);
2052 for (i = 0; i < n; ++i) {
2053 pred = get_irn_n(phi, i);
2054 tv = get_Const_tarval(pred);
2057 if (tv == tarval_bad) {
2058 /* folding failed, bad */
2063 mode = get_irn_mode(phi);
2064 irg = current_ir_graph;
2065 for (i = 0; i < n; ++i) {
2066 pred = get_irn_n(phi, i);
2067 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2069 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2070 } /* apply_unop_on_phi */
2073 * Apply a conversion on a constant operator (a Phi).
2075 * @param phi the Phi node
2077 * @return a new Phi node if the conversion was successful, NULL else
2079 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2084 int i, n = get_irn_arity(phi);
2086 NEW_ARR_A(void *, res, n);
2087 for (i = 0; i < n; ++i) {
2088 pred = get_irn_n(phi, i);
2089 tv = get_Const_tarval(pred);
2090 tv = tarval_convert_to(tv, mode);
2092 if (tv == tarval_bad) {
2093 /* folding failed, bad */
2098 irg = current_ir_graph;
2099 for (i = 0; i < n; ++i) {
2100 pred = get_irn_n(phi, i);
2101 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2103 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
2104 } /* apply_conv_on_phi */
2107 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2108 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2109 * If possible, remove the Conv's.
2111 static ir_node *transform_node_AddSub(ir_node *n) {
2112 ir_mode *mode = get_irn_mode(n);
2114 if (mode_is_reference(mode)) {
2115 ir_node *left = get_binop_left(n);
2116 ir_node *right = get_binop_right(n);
2117 unsigned ref_bits = get_mode_size_bits(mode);
2119 if (is_Conv(left)) {
2120 ir_mode *lmode = get_irn_mode(left);
2121 unsigned bits = get_mode_size_bits(lmode);
2123 if (ref_bits == bits &&
2124 mode_is_int(lmode) &&
2125 get_mode_arithmetic(lmode) == irma_twos_complement) {
2126 ir_node *pre = get_Conv_op(left);
2127 ir_mode *pre_mode = get_irn_mode(pre);
2129 if (mode_is_int(pre_mode) &&
2130 get_mode_size_bits(pre_mode) == bits &&
2131 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2132 /* ok, this conv just changes to sign, moreover the calculation
2133 * is done with same number of bits as our address mode, so
2134 * we can ignore the conv as address calculation can be viewed
2135 * as either signed or unsigned
2137 set_binop_left(n, pre);
2142 if (is_Conv(right)) {
2143 ir_mode *rmode = get_irn_mode(right);
2144 unsigned bits = get_mode_size_bits(rmode);
2146 if (ref_bits == bits &&
2147 mode_is_int(rmode) &&
2148 get_mode_arithmetic(rmode) == irma_twos_complement) {
2149 ir_node *pre = get_Conv_op(right);
2150 ir_mode *pre_mode = get_irn_mode(pre);
2152 if (mode_is_int(pre_mode) &&
2153 get_mode_size_bits(pre_mode) == bits &&
2154 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2155 /* ok, this conv just changes to sign, moreover the calculation
2156 * is done with same number of bits as our address mode, so
2157 * we can ignore the conv as address calculation can be viewed
2158 * as either signed or unsigned
2160 set_binop_right(n, pre);
2165 /* let address arithmetic use unsigned modes */
2166 if (is_Const(right)) {
2167 ir_mode *rmode = get_irn_mode(right);
2169 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2170 /* convert a AddP(P, *s) into AddP(P, *u) */
2171 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2173 ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
2174 set_binop_right(n, pre);
2180 } /* transform_node_AddSub */
2182 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2184 if (is_Const(b) && is_const_Phi(a)) { \
2185 /* check for Op(Phi, Const) */ \
2186 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2188 else if (is_Const(a) && is_const_Phi(b)) { \
2189 /* check for Op(Const, Phi) */ \
2190 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2192 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2193 /* check for Op(Phi, Phi) */ \
2194 c = apply_binop_on_2_phis(a, b, eval, mode); \
2197 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2201 #define HANDLE_UNOP_PHI(eval, a, c) \
2203 if (is_const_Phi(a)) { \
2204 /* check for Op(Phi) */ \
2205 c = apply_unop_on_phi(a, eval); \
2207 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2213 * Do the AddSub optimization, then Transform
2214 * Constant folding on Phi
2215 * Add(a,a) -> Mul(a, 2)
2216 * Add(Mul(a, x), a) -> Mul(a, x+1)
2217 * if the mode is integer or float.
2218 * Transform Add(a,-b) into Sub(a,b).
2219 * Reassociation might fold this further.
2221 static ir_node *transform_node_Add(ir_node *n) {
2223 ir_node *a, *b, *c, *oldn = n;
2225 n = transform_node_AddSub(n);
2227 a = get_Add_left(n);
2228 b = get_Add_right(n);
2230 mode = get_irn_mode(n);
2232 if (mode_is_reference(mode)) {
2233 ir_mode *lmode = get_irn_mode(a);
2235 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2236 /* an Add(a, NULL) is a hidden Conv */
2237 dbg_info *dbg = get_irn_dbg_info(n);
2238 return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2242 HANDLE_BINOP_PHI(tarval_add, a, b, c, mode);
2244 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2245 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2248 if (mode_is_num(mode)) {
2249 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2250 if (!is_arch_dep_running() && a == b && mode_is_int(mode)) {
2251 ir_node *block = get_nodes_block(n);
2254 get_irn_dbg_info(n),
2258 new_Const_long(mode, 2),
2260 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2265 get_irn_dbg_info(n),
2271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2276 get_irn_dbg_info(n),
2282 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2285 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2286 /* Here we rely on constants be on the RIGHT side */
2288 ir_node *op = get_Not_op(a);
2290 if (is_Const(b) && is_Const_one(b)) {
2292 ir_node *blk = get_nodes_block(n);
2293 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2294 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2299 n = new_Const(get_mode_minus_one(mode));
2300 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2305 ir_node *op = get_Not_op(b);
2309 n = new_Const(get_mode_minus_one(mode));
2310 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2317 } /* transform_node_Add */
2320 * returns -cnst or NULL if impossible
2322 static ir_node *const_negate(ir_node *cnst) {
2323 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2324 dbg_info *dbgi = get_irn_dbg_info(cnst);
2325 ir_graph *irg = get_irn_irg(cnst);
2326 if (tv == tarval_bad) return NULL;
2327 return new_rd_Const(dbgi, irg, tv);
2331 * Do the AddSub optimization, then Transform
2332 * Constant folding on Phi
2333 * Sub(0,a) -> Minus(a)
2334 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2335 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2336 * Sub(Add(a, x), x) -> a
2337 * Sub(x, Add(x, a)) -> -a
2338 * Sub(x, Const) -> Add(x, -Const)
2340 static ir_node *transform_node_Sub(ir_node *n) {
2345 n = transform_node_AddSub(n);
2347 a = get_Sub_left(n);
2348 b = get_Sub_right(n);
2350 mode = get_irn_mode(n);
2352 if (mode_is_int(mode)) {
2353 ir_mode *lmode = get_irn_mode(a);
2355 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2356 /* a Sub(a, NULL) is a hidden Conv */
2357 dbg_info *dbg = get_irn_dbg_info(n);
2358 n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode);
2359 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2363 if (mode == lmode &&
2364 get_mode_arithmetic(mode) == irma_twos_complement &&
2366 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2368 dbg_info *dbg = get_irn_dbg_info(n);
2369 n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode);
2370 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2376 HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode);
2378 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2379 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2382 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2383 /* a - C -> a + (-C) */
2384 ir_node *cnst = const_negate(b);
2386 ir_node *block = get_nodes_block(n);
2387 dbg_info *dbgi = get_irn_dbg_info(n);
2388 ir_graph *irg = get_irn_irg(n);
2390 n = new_rd_Add(dbgi, irg, block, a, cnst, mode);
2391 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2396 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2397 ir_graph *irg = current_ir_graph;
2398 dbg_info *dbg = get_irn_dbg_info(n);
2399 ir_node *block = get_nodes_block(n);
2400 ir_node *left = get_Minus_op(a);
2401 ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode);
2403 n = new_rd_Minus(dbg, irg, block, add, mode);
2404 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2406 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2407 ir_graph *irg = current_ir_graph;
2408 dbg_info *dbg = get_irn_dbg_info(n);
2409 ir_node *block = get_nodes_block(n);
2410 ir_node *right = get_Minus_op(b);
2412 n = new_rd_Add(dbg, irg, block, a, right, mode);
2413 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2415 } else if (is_Sub(b)) {
2416 /* a - (b - c) -> a + (c - b)
2417 * -> (a - b) + c iff (b - c) is a pointer */
2418 ir_graph *irg = current_ir_graph;
2419 dbg_info *s_dbg = get_irn_dbg_info(b);
2420 ir_node *s_block = get_nodes_block(b);
2421 ir_node *s_left = get_Sub_left(b);
2422 ir_node *s_right = get_Sub_right(b);
2423 ir_mode *s_mode = get_irn_mode(b);
2424 if (mode_is_reference(s_mode)) {
2425 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode);
2426 dbg_info *a_dbg = get_irn_dbg_info(n);
2427 ir_node *a_block = get_nodes_block(n);
2430 s_right = new_r_Conv(irg, a_block, s_right, mode);
2431 n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode);
2433 ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode);
2434 dbg_info *a_dbg = get_irn_dbg_info(n);
2435 ir_node *a_block = get_nodes_block(n);
2437 n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode);
2439 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2441 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2442 ir_node *m_right = get_Mul_right(b);
2443 if (is_Const(m_right)) {
2444 ir_node *cnst2 = const_negate(m_right);
2445 if (cnst2 != NULL) {
2446 ir_graph *irg = current_ir_graph;
2447 dbg_info *m_dbg = get_irn_dbg_info(b);
2448 ir_node *m_block = get_nodes_block(b);
2449 ir_node *m_left = get_Mul_left(b);
2450 ir_mode *m_mode = get_irn_mode(b);
2451 ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode);
2452 dbg_info *a_dbg = get_irn_dbg_info(n);
2453 ir_node *a_block = get_nodes_block(n);
2455 n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode);
2456 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2462 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2463 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2465 get_irn_dbg_info(n),
2470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2474 if (mode_wrap_around(mode)) {
2475 ir_node *left = get_Add_left(a);
2476 ir_node *right = get_Add_right(a);
2478 /* FIXME: Does the Conv's work only for two complement or generally? */
2480 if (mode != get_irn_mode(right)) {
2481 /* This Sub is an effective Cast */
2482 right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode);
2485 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2487 } else if (right == b) {
2488 if (mode != get_irn_mode(left)) {
2489 /* This Sub is an effective Cast */
2490 left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode);
2493 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2499 if (mode_wrap_around(mode)) {
2500 ir_node *left = get_Add_left(b);
2501 ir_node *right = get_Add_right(b);
2503 /* FIXME: Does the Conv's work only for two complement or generally? */
2505 ir_mode *r_mode = get_irn_mode(right);
2507 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode);
2508 if (mode != r_mode) {
2509 /* This Sub is an effective Cast */
2510 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2512 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2514 } else if (right == a) {
2515 ir_mode *l_mode = get_irn_mode(left);
2517 n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode);
2518 if (mode != l_mode) {
2519 /* This Sub is an effective Cast */
2520 n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode);
2522 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2527 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2528 ir_mode *mode = get_irn_mode(a);
2530 if (mode == get_irn_mode(b)) {
2532 ir_node *op_a = get_Conv_op(a);
2533 ir_node *op_b = get_Conv_op(b);
2535 /* check if it's allowed to skip the conv */
2536 ma = get_irn_mode(op_a);
2537 mb = get_irn_mode(op_b);
2539 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2540 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2543 set_Sub_right(n, b);
2549 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2550 if (!is_reassoc_running() && is_Mul(a)) {
2551 ir_node *ma = get_Mul_left(a);
2552 ir_node *mb = get_Mul_right(a);
2555 ir_node *blk = get_nodes_block(n);
2557 get_irn_dbg_info(n),
2558 current_ir_graph, blk,
2561 get_irn_dbg_info(n),
2562 current_ir_graph, blk,
2564 new_Const_long(mode, 1),
2567 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2569 } else if (mb == b) {
2570 ir_node *blk = get_nodes_block(n);
2572 get_irn_dbg_info(n),
2573 current_ir_graph, blk,
2576 get_irn_dbg_info(n),
2577 current_ir_graph, blk,
2579 new_Const_long(mode, 1),
2582 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2586 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2587 ir_node *x = get_Sub_left(a);
2588 ir_node *y = get_Sub_right(a);
2589 ir_node *blk = get_nodes_block(n);
2590 ir_mode *m_b = get_irn_mode(b);
2591 ir_mode *m_y = get_irn_mode(y);
2595 /* Determine the right mode for the Add. */
2598 else if (mode_is_reference(m_b))
2600 else if (mode_is_reference(m_y))
2604 * Both modes are different but none is reference,
2605 * happens for instance in SubP(SubP(P, Iu), Is).
2606 * We have two possibilities here: Cast or ignore.
2607 * Currently we ignore this case.
2612 add = new_r_Add(current_ir_graph, blk, y, b, add_mode);
2614 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode);
2615 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2619 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2620 if (is_Const(a) && is_Not(b)) {
2621 /* c - ~X = X + (c+1) */
2622 tarval *tv = get_Const_tarval(a);
2624 tv = tarval_add(tv, get_mode_one(mode));
2625 if (tv != tarval_bad) {
2626 ir_node *blk = get_nodes_block(n);
2627 ir_node *c = new_Const(tv);
2628 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode);
2629 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2635 } /* transform_node_Sub */
2638 * Several transformation done on n*n=2n bits mul.
2639 * These transformations must be done here because new nodes may be produced.
2641 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2643 ir_node *a = get_Mul_left(n);
2644 ir_node *b = get_Mul_right(n);
2645 tarval *ta = value_of(a);
2646 tarval *tb = value_of(b);
2647 ir_mode *smode = get_irn_mode(a);
2649 if (ta == get_mode_one(smode)) {
2650 /* (L)1 * (L)b = (L)b */
2651 ir_node *blk = get_nodes_block(n);
2652 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode);
2653 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2656 else if (ta == get_mode_minus_one(smode)) {
2657 /* (L)-1 * (L)b = (L)b */
2658 ir_node *blk = get_nodes_block(n);
2659 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode);
2660 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2661 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2664 if (tb == get_mode_one(smode)) {
2665 /* (L)a * (L)1 = (L)a */
2666 ir_node *blk = get_irn_n(a, -1);
2667 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode);
2668 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2671 else if (tb == get_mode_minus_one(smode)) {
2672 /* (L)a * (L)-1 = (L)-a */
2673 ir_node *blk = get_nodes_block(n);
2674 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode);
2675 n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode);
2676 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2683 * Transform Mul(a,-1) into -a.
2684 * Do constant evaluation of Phi nodes.
2685 * Do architecture dependent optimizations on Mul nodes
2687 static ir_node *transform_node_Mul(ir_node *n) {
2688 ir_node *c, *oldn = n;
2689 ir_mode *mode = get_irn_mode(n);
2690 ir_node *a = get_Mul_left(n);
2691 ir_node *b = get_Mul_right(n);
2693 if (is_Bad(a) || is_Bad(b))
2696 if (mode != get_irn_mode(a))
2697 return transform_node_Mul2n(n, mode);
2699 HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode);
2701 if (mode_is_signed(mode)) {
2704 if (value_of(a) == get_mode_minus_one(mode))
2706 else if (value_of(b) == get_mode_minus_one(mode))
2709 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode);
2710 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2715 if (is_Const(b)) { /* (-a) * const -> a * -const */
2716 ir_node *cnst = const_negate(b);
2718 dbg_info *dbgi = get_irn_dbg_info(n);
2719 ir_node *block = get_nodes_block(n);
2720 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), cnst, mode);
2721 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2724 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2725 dbg_info *dbgi = get_irn_dbg_info(n);
2726 ir_node *block = get_nodes_block(n);
2727 n = new_rd_Mul(dbgi, current_ir_graph, block, get_Minus_op(a), get_Minus_op(b), mode);
2728 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2730 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2731 ir_node *sub_l = get_Sub_left(b);
2732 ir_node *sub_r = get_Sub_right(b);
2733 dbg_info *dbgi = get_irn_dbg_info(n);
2734 ir_graph *irg = current_ir_graph;
2735 ir_node *block = get_nodes_block(n);
2736 ir_node *new_b = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2737 n = new_rd_Mul(dbgi, irg, block, get_Minus_op(a), new_b, mode);
2738 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2741 } else if (is_Minus(b)) {
2742 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2743 ir_node *sub_l = get_Sub_left(a);
2744 ir_node *sub_r = get_Sub_right(a);
2745 dbg_info *dbgi = get_irn_dbg_info(n);
2746 ir_graph *irg = current_ir_graph;
2747 ir_node *block = get_nodes_block(n);
2748 ir_node *new_a = new_rd_Sub(dbgi, irg, block, sub_r, sub_l, mode);
2749 n = new_rd_Mul(dbgi, irg, block, new_a, get_Minus_op(b), mode);
2750 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2754 if (get_mode_arithmetic(mode) == irma_ieee754) {
2756 tarval *tv = get_Const_tarval(a);
2757 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2758 && !tarval_is_negative(tv)) {
2759 /* 2.0 * b = b + b */
2760 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode);
2761 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2765 else if (is_Const(b)) {
2766 tarval *tv = get_Const_tarval(b);
2767 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2768 && !tarval_is_negative(tv)) {
2769 /* a * 2.0 = a + a */
2770 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode);
2771 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2776 return arch_dep_replace_mul_with_shifts(n);
2777 } /* transform_node_Mul */
2780 * Transform a Div Node.
2782 static ir_node *transform_node_Div(ir_node *n) {
2783 ir_mode *mode = get_Div_resmode(n);
2784 ir_node *a = get_Div_left(n);
2785 ir_node *b = get_Div_right(n);
2789 if (is_Const(b) && is_const_Phi(a)) {
2790 /* check for Div(Phi, Const) */
2791 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2793 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2797 else if (is_Const(a) && is_const_Phi(b)) {
2798 /* check for Div(Const, Phi) */
2799 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2801 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2805 else if (is_const_Phi(a) && is_const_Phi(b)) {
2806 /* check for Div(Phi, Phi) */
2807 value = apply_binop_on_2_phis(a, b, tarval_div, mode);
2809 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2816 if (tv != tarval_bad) {
2817 value = new_Const(tv);
2819 DBG_OPT_CSTEVAL(n, value);
2822 ir_node *a = get_Div_left(n);
2823 ir_node *b = get_Div_right(n);
2824 const ir_node *dummy;
2826 if (a == b && value_not_zero(a, &dummy)) {
2827 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2828 value = new_Const(get_mode_one(mode));
2829 DBG_OPT_CSTEVAL(n, value);
2832 if (mode_is_signed(mode) && is_Const(b)) {
2833 tarval *tv = get_Const_tarval(b);
2835 if (tv == get_mode_minus_one(mode)) {
2837 value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
2838 DBG_OPT_CSTEVAL(n, value);
2842 /* Try architecture dependent optimization */
2843 value = arch_dep_replace_div_by_const(n);
2851 /* Turn Div into a tuple (mem, jmp, bad, value) */
2852 mem = get_Div_mem(n);
2853 blk = get_nodes_block(n);
2855 /* skip a potential Pin */
2856 mem = skip_Pin(mem);
2857 turn_into_tuple(n, pn_Div_max);
2858 set_Tuple_pred(n, pn_Div_M, mem);
2859 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2860 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2861 set_Tuple_pred(n, pn_Div_res, value);
2864 } /* transform_node_Div */
2867 * Transform a Mod node.
2869 static ir_node *transform_node_Mod(ir_node *n) {
2870 ir_mode *mode = get_Mod_resmode(n);
2871 ir_node *a = get_Mod_left(n);
2872 ir_node *b = get_Mod_right(n);
2876 if (is_Const(b) && is_const_Phi(a)) {
2877 /* check for Div(Phi, Const) */
2878 value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2880 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2884 else if (is_Const(a) && is_const_Phi(b)) {
2885 /* check for Div(Const, Phi) */
2886 value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2888 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2892 else if (is_const_Phi(a) && is_const_Phi(b)) {
2893 /* check for Div(Phi, Phi) */
2894 value = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2896 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2903 if (tv != tarval_bad) {
2904 value = new_Const(tv);
2906 DBG_OPT_CSTEVAL(n, value);
2909 ir_node *a = get_Mod_left(n);
2910 ir_node *b = get_Mod_right(n);
2911 const ir_node *dummy;
2913 if (a == b && value_not_zero(a, &dummy)) {
2914 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2915 value = new_Const(get_mode_null(mode));
2916 DBG_OPT_CSTEVAL(n, value);
2919 if (mode_is_signed(mode) && is_Const(b)) {
2920 tarval *tv = get_Const_tarval(b);
2922 if (tv == get_mode_minus_one(mode)) {
2924 value = new_Const(get_mode_null(mode));
2925 DBG_OPT_CSTEVAL(n, value);
2929 /* Try architecture dependent optimization */
2930 value = arch_dep_replace_mod_by_const(n);
2938 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2939 mem = get_Mod_mem(n);
2940 blk = get_nodes_block(n);
2942 /* skip a potential Pin */
2943 mem = skip_Pin(mem);
2944 turn_into_tuple(n, pn_Mod_max);
2945 set_Tuple_pred(n, pn_Mod_M, mem);
2946 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2947 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2948 set_Tuple_pred(n, pn_Mod_res, value);
2951 } /* transform_node_Mod */
2954 * Transform a DivMod node.
2956 static ir_node *transform_node_DivMod(ir_node *n) {
2957 const ir_node *dummy;
2958 ir_node *a = get_DivMod_left(n);
2959 ir_node *b = get_DivMod_right(n);
2960 ir_mode *mode = get_DivMod_resmode(n);
2965 if (is_Const(b) && is_const_Phi(a)) {
2966 /* check for Div(Phi, Const) */
2967 va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0);
2968 vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0);
2970 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2971 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2975 else if (is_Const(a) && is_const_Phi(b)) {
2976 /* check for Div(Const, Phi) */
2977 va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1);
2978 vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1);
2980 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2981 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2985 else if (is_const_Phi(a) && is_const_Phi(b)) {
2986 /* check for Div(Phi, Phi) */
2987 va = apply_binop_on_2_phis(a, b, tarval_div, mode);
2988 vb = apply_binop_on_2_phis(a, b, tarval_mod, mode);
2990 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
2991 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
2998 if (tb != tarval_bad) {
2999 if (tb == get_mode_one(get_tarval_mode(tb))) {
3001 vb = new_Const(get_mode_null(mode));
3002 DBG_OPT_CSTEVAL(n, vb);
3004 } else if (ta != tarval_bad) {
3005 tarval *resa, *resb;
3006 resa = tarval_div(ta, tb);
3007 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3008 Jmp for X result!? */
3009 resb = tarval_mod(ta, tb);
3010 if (resb == tarval_bad) return n; /* Causes exception! */
3011 va = new_Const(resa);
3012 vb = new_Const(resb);
3013 DBG_OPT_CSTEVAL(n, va);
3014 DBG_OPT_CSTEVAL(n, vb);
3016 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3017 va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode);
3018 vb = new_Const(get_mode_null(mode));
3019 DBG_OPT_CSTEVAL(n, va);
3020 DBG_OPT_CSTEVAL(n, vb);
3022 } else { /* Try architecture dependent optimization */
3025 arch_dep_replace_divmod_by_const(&va, &vb, n);
3026 evaluated = va != NULL;
3028 } else if (a == b) {
3029 if (value_not_zero(a, &dummy)) {
3031 va = new_Const(get_mode_one(mode));
3032 vb = new_Const(get_mode_null(mode));
3033 DBG_OPT_CSTEVAL(n, va);
3034 DBG_OPT_CSTEVAL(n, vb);
3037 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3040 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3041 /* 0 / non-Const = 0 */
3046 if (evaluated) { /* replace by tuple */
3050 mem = get_DivMod_mem(n);
3051 /* skip a potential Pin */
3052 mem = skip_Pin(mem);
3054 blk = get_nodes_block(n);
3055 turn_into_tuple(n, pn_DivMod_max);
3056 set_Tuple_pred(n, pn_DivMod_M, mem);
3057 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
3058 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3059 set_Tuple_pred(n, pn_DivMod_res_div, va);
3060 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3064 } /* transform_node_DivMod */
3067 * Optimize x / c to x * (1/c)
3069 static ir_node *transform_node_Quot(ir_node *n) {
3070 ir_mode *mode = get_Quot_resmode(n);
3073 if (get_mode_arithmetic(mode) == irma_ieee754) {
3074 ir_node *b = get_Quot_right(n);
3075 tarval *tv = value_of(b);
3077 if (tv != tarval_bad) {
3078 int rem = tarval_fp_ops_enabled();
3081 * Floating point constant folding might be disabled here to
3083 * However, as we check for exact result, doing it is safe.
3086 tarval_enable_fp_ops(1);
3087 tv = tarval_quo(get_mode_one(mode), tv);
3088 tarval_enable_fp_ops(rem);
3090 /* Do the transformation if the result is either exact or we are not
3091 using strict rules. */
3092 if (tv != tarval_bad &&
3093 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3094 ir_node *blk = get_nodes_block(n);
3095 ir_node *c = new_Const(tv);
3096 ir_node *a = get_Quot_left(n);
3097 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode);
3098 ir_node *mem = get_Quot_mem(n);
3100 /* skip a potential Pin */
3101 mem = skip_Pin(mem);
3102 turn_into_tuple(n, pn_Quot_max);
3103 set_Tuple_pred(n, pn_Quot_M, mem);
3104 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
3105 set_Tuple_pred(n, pn_Quot_X_except, new_r_Bad(current_ir_graph));
3106 set_Tuple_pred(n, pn_Quot_res, m);
3107 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3112 } /* transform_node_Quot */
3115 * Optimize Abs(x) into x if x is Confirmed >= 0
3116 * Optimize Abs(x) into -x if x is Confirmed <= 0
3117 * Optimize Abs(-x) int Abs(x)
3119 static ir_node *transform_node_Abs(ir_node *n) {
3120 ir_node *c, *oldn = n;
3121 ir_node *a = get_Abs_op(n);
3124 HANDLE_UNOP_PHI(tarval_abs, a, c);
3126 switch (classify_value_sign(a)) {
3127 case value_classified_negative:
3128 mode = get_irn_mode(n);
3131 * We can replace the Abs by -x here.
3132 * We even could add a new Confirm here
3133 * (if not twos complement)
3135 * Note that -x would create a new node, so we could
3136 * not run it in the equivalent_node() context.
3138 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
3139 get_nodes_block(n), a, mode);
3141 DBG_OPT_CONFIRM(oldn, n);
3143 case value_classified_positive:
3144 /* n is positive, Abs is not needed */
3147 DBG_OPT_CONFIRM(oldn, n);
3153 /* Abs(-x) = Abs(x) */
3154 mode = get_irn_mode(n);
3155 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph,
3156 get_nodes_block(n), get_Minus_op(a), mode);
3157 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3161 } /* transform_node_Abs */
3164 * Optimize -a CMP -b into b CMP a.
3165 * This works only for for modes where unary Minus
3167 * Note that two-complement integers can Overflow
3168 * so it will NOT work.
3170 * For == and != can be handled in Proj(Cmp)
3172 static ir_node *transform_node_Cmp(ir_node *n) {
3174 ir_node *left = get_Cmp_left(n);
3175 ir_node *right = get_Cmp_right(n);
3177 if (is_Minus(left) && is_Minus(right) &&
3178 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3179 ir_node *const new_left = get_Minus_op(right);
3180 ir_node *const new_right = get_Minus_op(left);
3181 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph,
3182 get_nodes_block(n), new_left, new_right);
3183 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3186 } /* transform_node_Cmp */
3190 * Transform a Cond node.
3192 * Replace the Cond by a Jmp if it branches on a constant
3195 static ir_node *transform_node_Cond(ir_node *n) {
3198 ir_node *a = get_Cond_selector(n);
3199 tarval *ta = value_of(a);
3201 /* we need block info which is not available in floating irgs */
3202 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3205 if ((ta != tarval_bad) &&
3206 (get_irn_mode(a) == mode_b) &&
3207 (get_opt_unreachable_code())) {
3208 /* It's a boolean Cond, branching on a boolean constant.
3209 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3210 ir_node *blk = get_nodes_block(n);
3211 jmp = new_r_Jmp(current_ir_graph, blk);
3212 turn_into_tuple(n, pn_Cond_max);
3213 if (ta == tarval_b_true) {
3214 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3215 set_Tuple_pred(n, pn_Cond_true, jmp);
3217 set_Tuple_pred(n, pn_Cond_false, jmp);
3218 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3220 /* We might generate an endless loop, so keep it alive. */
3221 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3224 } /* transform_node_Cond */
3227 * Prototype of a recursive transform function
3228 * for bitwise distributive transformations.
3230 typedef ir_node* (*recursive_transform)(ir_node *n);
3233 * makes use of distributive laws for and, or, eor
3234 * and(a OP c, b OP c) -> and(a, b) OP c
3235 * note, might return a different op than n
3237 static ir_node *transform_bitwise_distributive(ir_node *n,
3238 recursive_transform trans_func)
3241 ir_node *a = get_binop_left(n);
3242 ir_node *b = get_binop_right(n);
3243 ir_op *op = get_irn_op(a);
3244 ir_op *op_root = get_irn_op(n);
3246 if(op != get_irn_op(b))
3249 if (op == op_Conv) {
3250 ir_node *a_op = get_Conv_op(a);
3251 ir_node *b_op = get_Conv_op(b);
3252 ir_mode *a_mode = get_irn_mode(a_op);
3253 ir_mode *b_mode = get_irn_mode(b_op);
3254 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3255 ir_node *blk = get_nodes_block(n);
3258 set_binop_left(n, a_op);
3259 set_binop_right(n, b_op);
3260 set_irn_mode(n, a_mode);
3262 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
3264 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3270 /* nothing to gain here */
3274 if (op == op_Shrs || op == op_Shr || op == op_Shl
3275 || op == op_And || op == op_Or || op == op_Eor) {
3276 ir_node *a_left = get_binop_left(a);
3277 ir_node *a_right = get_binop_right(a);
3278 ir_node *b_left = get_binop_left(b);
3279 ir_node *b_right = get_binop_right(b);
3281 ir_node *op1 = NULL;
3282 ir_node *op2 = NULL;
3284 if (is_op_commutative(op)) {
3285 if (a_left == b_left) {
3289 } else if(a_left == b_right) {
3293 } else if(a_right == b_left) {
3299 if(a_right == b_right) {
3306 /* (a sop c) & (b sop c) => (a & b) sop c */
3307 ir_node *blk = get_nodes_block(n);
3309 ir_node *new_n = exact_copy(n);
3310 set_binop_left(new_n, op1);
3311 set_binop_right(new_n, op2);
3312 new_n = trans_func(new_n);
3314 if(op_root == op_Eor && op == op_Or) {
3315 dbg_info *dbgi = get_irn_dbg_info(n);
3316 ir_graph *irg = current_ir_graph;
3317 ir_mode *mode = get_irn_mode(c);
3319 c = new_rd_Not(dbgi, irg, blk, c, mode);
3320 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
3323 set_nodes_block(n, blk);
3324 set_binop_left(n, new_n);
3325 set_binop_right(n, c);
3326 add_identities(current_ir_graph->value_table, n);
3329 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3340 static ir_node *transform_node_And(ir_node *n) {
3341 ir_node *c, *oldn = n;
3342 ir_node *a = get_And_left(n);
3343 ir_node *b = get_And_right(n);
3346 mode = get_irn_mode(n);
3347 HANDLE_BINOP_PHI(tarval_and, a, b, c, mode);
3349 /* we can evaluate 2 Projs of the same Cmp */
3350 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3351 ir_node *pred_a = get_Proj_pred(a);
3352 ir_node *pred_b = get_Proj_pred(b);
3353 if (pred_a == pred_b) {
3354 dbg_info *dbgi = get_irn_dbg_info(n);
3355 ir_node *block = get_nodes_block(pred_a);
3356 pn_Cmp pn_a = get_Proj_proj(a);
3357 pn_Cmp pn_b = get_Proj_proj(b);
3358 /* yes, we can simply calculate with pncs */
3359 pn_Cmp new_pnc = pn_a & pn_b;
3361 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b, new_pnc);
3366 ir_node *op = get_Not_op(b);
3368 ir_node *ba = get_And_left(op);
3369 ir_node *bb = get_And_right(op);
3371 /* it's enough to test the following cases due to normalization! */
3372 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3373 /* (a|b) & ~(a&b) = a^b */
3374 ir_node *block = get_nodes_block(n);
3376 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, ba, bb, mode);
3377 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3385 ir_node *op = get_Not_op(a);
3387 ir_node *aa = get_And_left(op);
3388 ir_node *ab = get_And_right(op);
3390 /* it's enough to test the following cases due to normalization! */
3391 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3392 /* (a|b) & ~(a&b) = a^b */
3393 ir_node *block = get_nodes_block(n);
3395 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, block, aa, ab, mode);
3396 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3403 ir_node *al = get_Eor_left(a);
3404 ir_node *ar = get_Eor_right(a);
3407 /* (b ^ a) & b -> ~a & b */
3408 dbg_info *dbg = get_irn_dbg_info(n);
3409 ir_node *block = get_nodes_block(n);
3411 ar = new_rd_Not(dbg, current_ir_graph, block, ar, mode);
3412 n = new_rd_And(dbg, current_ir_graph, block, ar, b, mode);
3413 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3417 /* (a ^ b) & b -> ~a & b */
3418 dbg_info *dbg = get_irn_dbg_info(n);
3419 ir_node *block = get_nodes_block(n);
3421 al = new_rd_Not(dbg, current_ir_graph, block, al, mode);
3422 n = new_rd_And(dbg, current_ir_graph, block, al, b, mode);
3423 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3428 ir_node *bl = get_Eor_left(b);
3429 ir_node *br = get_Eor_right(b);
3432 /* a & (a ^ b) -> a & ~b */
3433 dbg_info *dbg = get_irn_dbg_info(n);
3434 ir_node *block = get_nodes_block(n);
3436 br = new_rd_Not(dbg, current_ir_graph, block, br, mode);
3437 n = new_rd_And(dbg, current_ir_graph, block, br, a, mode);
3438 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3442 /* a & (b ^ a) -> a & ~b */
3443 dbg_info *dbg = get_irn_dbg_info(n);
3444 ir_node *block = get_nodes_block(n);
3446 bl = new_rd_Not(dbg, current_ir_graph, block, bl, mode);
3447 n = new_rd_And(dbg, current_ir_graph, block, bl, a, mode);
3448 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3452 if (is_Not(a) && is_Not(b)) {
3453 /* ~a & ~b = ~(a|b) */
3454 ir_node *block = get_nodes_block(n);
3455 ir_mode *mode = get_irn_mode(n);
3459 n = new_rd_Or(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
3460 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
3461 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3465 n = transform_bitwise_distributive(n, transform_node_And);
3468 } /* transform_node_And */
3473 static ir_node *transform_node_Eor(ir_node *n) {
3474 ir_node *c, *oldn = n;
3475 ir_node *a = get_Eor_left(n);
3476 ir_node *b = get_Eor_right(n);
3477 ir_mode *mode = get_irn_mode(n);
3479 HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode);
3481 /* we can evaluate 2 Projs of the same Cmp */
3482 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3483 ir_node *pred_a = get_Proj_pred(a);
3484 ir_node *pred_b = get_Proj_pred(b);
3485 if(pred_a == pred_b) {
3486 dbg_info *dbgi = get_irn_dbg_info(n);
3487 ir_node *block = get_nodes_block(pred_a);
3488 pn_Cmp pn_a = get_Proj_proj(a);
3489 pn_Cmp pn_b = get_Proj_proj(b);
3490 /* yes, we can simply calculate with pncs */
3491 pn_Cmp new_pnc = pn_a ^ pn_b;
3493 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
3500 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3501 get_mode_null(mode));
3502 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3503 } else if (mode == mode_b &&
3505 is_Const(b) && is_Const_one(b) &&
3506 is_Cmp(get_Proj_pred(a))) {
3507 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
3508 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3509 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
3511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
3512 } else if (is_Const(b)) {
3513 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3514 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b)));
3515 ir_node *not_op = get_Not_op(a);
3516 dbg_info *dbg = get_irn_dbg_info(n);
3517 ir_graph *irg = current_ir_graph;
3518 ir_node *block = get_nodes_block(n);
3519 ir_mode *mode = get_irn_mode(n);
3520 n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode);
3522 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3523 n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode);
3524 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3527 n = transform_bitwise_distributive(n, transform_node_Eor);
3531 } /* transform_node_Eor */
3536 static ir_node *transform_node_Not(ir_node *n) {
3537 ir_node *c, *oldn = n;
3538 ir_node *a = get_Not_op(n);
3539 ir_mode *mode = get_irn_mode(n);
3541 HANDLE_UNOP_PHI(tarval_not,a,c);
3543 /* check for a boolean Not */
3544 if (mode == mode_b &&
3546 is_Cmp(get_Proj_pred(a))) {
3547 /* We negate a Cmp. The Cmp has the negated result anyways! */
3548 n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a),
3549 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3550 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3554 ir_node *eor_b = get_Eor_right(a);
3555 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3556 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b)));
3557 ir_node *eor_a = get_Eor_left(a);
3558 dbg_info *dbg = get_irn_dbg_info(n);
3559 ir_graph *irg = current_ir_graph;
3560 ir_node *block = get_nodes_block(n);
3561 ir_mode *mode = get_irn_mode(n);
3562 n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode);
3566 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3567 if (is_Minus(a)) { /* ~-x -> x + -1 */
3568 dbg_info *dbg = get_irn_dbg_info(n);
3569 ir_graph *irg = current_ir_graph;
3570 ir_node *block = get_nodes_block(n);
3571 ir_node *add_l = get_Minus_op(a);
3572 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3573 n = new_rd_Add(dbg, irg, block, add_l, add_r, mode);
3574 } else if (is_Add(a)) {
3575 ir_node *add_r = get_Add_right(a);
3576 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3577 /* ~(x + -1) = -x */
3578 ir_node *op = get_Add_left(a);
3579 ir_node *blk = get_nodes_block(n);
3580 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
3581 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3586 } /* transform_node_Not */
3589 * Transform a Minus.
3593 * -(a >>u (size-1)) = a >>s (size-1)
3594 * -(a >>s (size-1)) = a >>u (size-1)
3595 * -(a * const) -> a * -const
3597 static ir_node *transform_node_Minus(ir_node *n) {
3598 ir_node *c, *oldn = n;
3599 ir_node *a = get_Minus_op(n);
3602 HANDLE_UNOP_PHI(tarval_neg,a,c);
3604 mode = get_irn_mode(a);
3605 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3606 /* the following rules are only to twos-complement */
3609 ir_node *op = get_Not_op(a);
3610 tarval *tv = get_mode_one(mode);
3611 ir_node *blk = get_nodes_block(n);
3612 ir_node *c = new_Const(tv);
3613 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
3614 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3618 ir_node *c = get_Shr_right(a);
3621 tarval *tv = get_Const_tarval(c);
3623 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3624 /* -(a >>u (size-1)) = a >>s (size-1) */
3625 ir_node *v = get_Shr_left(a);
3627 n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3628 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3634 ir_node *c = get_Shrs_right(a);
3637 tarval *tv = get_Const_tarval(c);
3639 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3640 /* -(a >>s (size-1)) = a >>u (size-1) */
3641 ir_node *v = get_Shrs_left(a);
3643 n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode);
3644 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3651 /* - (a-b) = b - a */
3652 ir_node *la = get_Sub_left(a);
3653 ir_node *ra = get_Sub_right(a);
3654 ir_node *blk = get_nodes_block(n);
3656 n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode);
3657 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3661 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3662 ir_node *mul_l = get_Mul_left(a);
3663 ir_node *mul_r = get_Mul_right(a);
3664 tarval *tv = value_of(mul_r);
3665 if (tv != tarval_bad) {
3666 tv = tarval_neg(tv);
3667 if (tv != tarval_bad) {
3668 ir_node *cnst = new_Const(tv);
3669 dbg_info *dbg = get_irn_dbg_info(a);
3670 ir_graph *irg = current_ir_graph;
3671 ir_node *block = get_nodes_block(a);
3672 n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
3673 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3680 } /* transform_node_Minus */
3683 * Transform a Cast_type(Const) into a new Const_type
3685 static ir_node *transform_node_Cast(ir_node *n) {
3687 ir_node *pred = get_Cast_op(n);
3688 ir_type *tp = get_irn_type(n);
3690 if (is_Const(pred) && get_Const_type(pred) != tp) {
3691 n = new_rd_Const_type(NULL, current_ir_graph, get_Const_tarval(pred), tp);
3692 DBG_OPT_CSTEVAL(oldn, n);
3693 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3694 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
3695 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3696 DBG_OPT_CSTEVAL(oldn, n);
3700 } /* transform_node_Cast */
3703 * Transform a Proj(Load) with a non-null address.
3705 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3706 if (get_opt_ldst_only_null_ptr_exceptions()) {
3707 if (get_irn_mode(proj) == mode_X) {
3708 ir_node *load = get_Proj_pred(proj);
3710 /* get the Load address */
3711 const ir_node *addr = get_Load_ptr(load);
3712 const ir_node *confirm;
3714 if (value_not_null(addr, &confirm)) {
3715 if (confirm == NULL) {
3716 /* this node may float if it did not depend on a Confirm */
3717 set_irn_pinned(load, op_pin_state_floats);
3719 if (get_Proj_proj(proj) == pn_Load_X_except) {
3720 DBG_OPT_EXC_REM(proj);
3721 return get_irg_bad(current_ir_graph);
3723 ir_node *blk = get_nodes_block(load);
3724 return new_r_Jmp(current_ir_graph, blk);
3730 } /* transform_node_Proj_Load */
3733 * Transform a Proj(Store) with a non-null address.
3735 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3736 if (get_opt_ldst_only_null_ptr_exceptions()) {
3737 if (get_irn_mode(proj) == mode_X) {
3738 ir_node *store = get_Proj_pred(proj);
3740 /* get the load/store address */
3741 const ir_node *addr = get_Store_ptr(store);
3742 const ir_node *confirm;
3744 if (value_not_null(addr, &confirm)) {
3745 if (confirm == NULL) {
3746 /* this node may float if it did not depend on a Confirm */
3747 set_irn_pinned(store, op_pin_state_floats);
3749 if (get_Proj_proj(proj) == pn_Store_X_except) {
3750 DBG_OPT_EXC_REM(proj);
3751 return get_irg_bad(current_ir_graph);
3753 ir_node *blk = get_nodes_block(store);
3754 return new_r_Jmp(current_ir_graph, blk);
3760 } /* transform_node_Proj_Store */
3763 * Transform a Proj(Div) with a non-zero value.
3764 * Removes the exceptions and routes the memory to the NoMem node.
3766 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3767 ir_node *div = get_Proj_pred(proj);
3768 ir_node *b = get_Div_right(div);
3769 ir_node *res, *new_mem;
3770 const ir_node *confirm;
3773 if (value_not_zero(b, &confirm)) {
3774 /* div(x, y) && y != 0 */
3775 if (confirm == NULL) {
3776 /* we are sure we have a Const != 0 */
3777 new_mem = get_Div_mem(div);
3778 new_mem = skip_Pin(new_mem);
3779 set_Div_mem(div, new_mem);
3780 set_irn_pinned(div, op_pin_state_floats);
3783 proj_nr = get_Proj_proj(proj);
3785 case pn_Div_X_regular:
3786 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
3788 case pn_Div_X_except:
3789 /* we found an exception handler, remove it */
3790 DBG_OPT_EXC_REM(proj);
3794 res = get_Div_mem(div);
3795 new_mem = get_irg_no_mem(current_ir_graph);
3798 /* This node can only float up to the Confirm block */
3799 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3801 set_irn_pinned(div, op_pin_state_floats);
3802 /* this is a Div without exception, we can remove the memory edge */
3803 set_Div_mem(div, new_mem);
3808 } /* transform_node_Proj_Div */
3811 * Transform a Proj(Mod) with a non-zero value.
3812 * Removes the exceptions and routes the memory to the NoMem node.
3814 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3815 ir_node *mod = get_Proj_pred(proj);
3816 ir_node *b = get_Mod_right(mod);
3817 ir_node *res, *new_mem;
3818 const ir_node *confirm;
3821 if (value_not_zero(b, &confirm)) {
3822 /* mod(x, y) && y != 0 */
3823 proj_nr = get_Proj_proj(proj);
3825 if (confirm == NULL) {
3826 /* we are sure we have a Const != 0 */
3827 new_mem = get_Mod_mem(mod);
3828 new_mem = skip_Pin(new_mem);
3829 set_Mod_mem(mod, new_mem);
3830 set_irn_pinned(mod, op_pin_state_floats);
3835 case pn_Mod_X_regular:
3836 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
3838 case pn_Mod_X_except:
3839 /* we found an exception handler, remove it */
3840 DBG_OPT_EXC_REM(proj);
3844 res = get_Mod_mem(mod);
3845 new_mem = get_irg_no_mem(current_ir_graph);
3848 /* This node can only float up to the Confirm block */
3849 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3851 /* this is a Mod without exception, we can remove the memory edge */
3852 set_Mod_mem(mod, new_mem);
3855 if (get_Mod_left(mod) == b) {
3856 /* a % a = 0 if a != 0 */
3857 ir_mode *mode = get_irn_mode(proj);
3858 ir_node *res = new_Const(get_mode_null(mode));
3860 DBG_OPT_CSTEVAL(mod, res);
3866 } /* transform_node_Proj_Mod */
3869 * Transform a Proj(DivMod) with a non-zero value.
3870 * Removes the exceptions and routes the memory to the NoMem node.
3872 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3873 ir_node *divmod = get_Proj_pred(proj);
3874 ir_node *b = get_DivMod_right(divmod);
3875 ir_node *res, *new_mem;
3876 const ir_node *confirm;
3879 if (value_not_zero(b, &confirm)) {
3880 /* DivMod(x, y) && y != 0 */
3881 proj_nr = get_Proj_proj(proj);
3883 if (confirm == NULL) {
3884 /* we are sure we have a Const != 0 */
3885 new_mem = get_DivMod_mem(divmod);
3886 new_mem = skip_Pin(new_mem);
3887 set_DivMod_mem(divmod, new_mem);
3888 set_irn_pinned(divmod, op_pin_state_floats);
3893 case pn_DivMod_X_regular:
3894 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
3896 case pn_DivMod_X_except:
3897 /* we found an exception handler, remove it */
3898 DBG_OPT_EXC_REM(proj);
3902 res = get_DivMod_mem(divmod);
3903 new_mem = get_irg_no_mem(current_ir_graph);
3906 /* This node can only float up to the Confirm block */
3907 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
3909 /* this is a DivMod without exception, we can remove the memory edge */
3910 set_DivMod_mem(divmod, new_mem);
3913 case pn_DivMod_res_mod:
3914 if (get_DivMod_left(divmod) == b) {
3915 /* a % a = 0 if a != 0 */
3916 ir_mode *mode = get_irn_mode(proj);
3917 ir_node *res = new_Const(get_mode_null(mode));
3919 DBG_OPT_CSTEVAL(divmod, res);
3925 } /* transform_node_Proj_DivMod */
3928 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3930 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3931 if (get_opt_unreachable_code()) {
3932 ir_node *n = get_Proj_pred(proj);
3933 ir_node *b = get_Cond_selector(n);
3935 if (mode_is_int(get_irn_mode(b))) {
3936 tarval *tb = value_of(b);
3938 if (tb != tarval_bad) {
3939 /* we have a constant switch */
3940 long num = get_Proj_proj(proj);
3942 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
3943 if (get_tarval_long(tb) == num) {
3944 /* Do NOT create a jump here, or we will have 2 control flow ops
3945 * in a block. This case is optimized away in optimize_cf(). */
3948 /* this case will NEVER be taken, kill it */
3949 return get_irg_bad(current_ir_graph);
3956 } /* transform_node_Proj_Cond */
3959 * Create a 0 constant of given mode.
3961 static ir_node *create_zero_const(ir_mode *mode) {
3962 tarval *tv = get_mode_null(mode);
3963 ir_node *cnst = new_Const(tv);
3968 /* the order of the values is important! */
3969 typedef enum const_class {
3975 static const_class classify_const(const ir_node* n)
3977 if (is_Const(n)) return const_const;
3978 if (is_irn_constlike(n)) return const_like;
3983 * Determines whether r is more constlike or has a larger index (in that order)
3986 static int operands_are_normalized(const ir_node *l, const ir_node *r)
3988 const const_class l_order = classify_const(l);
3989 const const_class r_order = classify_const(r);
3991 l_order > r_order ||
3992 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3996 * Normalizes and optimizes Cmp nodes.
3998 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
3999 ir_node *n = get_Proj_pred(proj);
4000 ir_node *left = get_Cmp_left(n);
4001 ir_node *right = get_Cmp_right(n);
4004 ir_mode *mode = NULL;
4005 long proj_nr = get_Proj_proj(proj);
4007 /* we can evaluate some cases directly */
4010 return new_Const(get_tarval_b_false());
4012 return new_Const(get_tarval_b_true());
4014 if (!mode_is_float(get_irn_mode(left)))
4015 return new_Const(get_tarval_b_true());
4021 /* remove Casts of both sides */
4022 left = skip_Cast(left);
4023 right = skip_Cast(right);
4025 /* Remove unnecessary conversions */
4026 /* TODO handle constants */
4027 if (is_Conv(left) && is_Conv(right)) {
4028 ir_mode *mode = get_irn_mode(left);
4029 ir_node *op_left = get_Conv_op(left);
4030 ir_node *op_right = get_Conv_op(right);
4031 ir_mode *mode_left = get_irn_mode(op_left);
4032 ir_mode *mode_right = get_irn_mode(op_right);
4034 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4035 && mode_left != mode_b && mode_right != mode_b) {
4036 ir_graph *irg = current_ir_graph;
4037 ir_node *block = get_nodes_block(n);
4039 if (mode_left == mode_right) {
4043 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4044 } else if (smaller_mode(mode_left, mode_right)) {
4045 left = new_r_Conv(irg, block, op_left, mode_right);
4048 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4049 } else if (smaller_mode(mode_right, mode_left)) {
4051 right = new_r_Conv(irg, block, op_right, mode_left);
4053 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4058 /* remove operation on both sides if possible */
4059 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4061 * The following operations are NOT safe for floating point operations, for instance
4062 * 1.0 + inf == 2.0 + inf, =/=> x == y
4064 if (mode_is_int(get_irn_mode(left))) {
4065 unsigned lop = get_irn_opcode(left);
4067 if (lop == get_irn_opcode(right)) {
4068 ir_node *ll, *lr, *rl, *rr;
4070 /* same operation on both sides, try to remove */
4074 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4075 left = get_unop_op(left);
4076 right = get_unop_op(right);
4078 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4081 ll = get_Add_left(left);
4082 lr = get_Add_right(left);
4083 rl = get_Add_left(right);
4084 rr = get_Add_right(right);
4087 /* X + a CMP X + b ==> a CMP b */
4091 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4092 } else if (ll == rr) {
4093 /* X + a CMP b + X ==> a CMP b */
4097 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4098 } else if (lr == rl) {
4099 /* a + X CMP X + b ==> a CMP b */
4103 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4104 } else if (lr == rr) {
4105 /* a + X CMP b + X ==> a CMP b */
4109 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4113 ll = get_Sub_left(left);
4114 lr = get_Sub_right(left);
4115 rl = get_Sub_left(right);
4116 rr = get_Sub_right(right);
4119 /* X - a CMP X - b ==> a CMP b */
4123 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4124 } else if (lr == rr) {
4125 /* a - X CMP b - X ==> a CMP b */
4129 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4133 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4134 /* a ROTL X CMP b ROTL X ==> a CMP b */
4135 left = get_Rotl_left(left);
4136 right = get_Rotl_left(right);
4138 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4146 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4147 if (is_Add(left) || is_Sub(left)) {
4148 ir_node *ll = get_binop_left(left);
4149 ir_node *lr = get_binop_right(left);
4151 if (lr == right && is_Add(left)) {
4158 right = create_zero_const(get_irn_mode(left));
4160 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4163 if (is_Add(right) || is_Sub(right)) {
4164 ir_node *rl = get_binop_left(right);
4165 ir_node *rr = get_binop_right(right);
4167 if (rr == left && is_Add(right)) {
4174 right = create_zero_const(get_irn_mode(left));
4176 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4179 if (is_And(left) && is_Const(right)) {
4180 ir_node *ll = get_binop_left(left);
4181 ir_node *lr = get_binop_right(left);
4182 if (is_Shr(ll) && is_Const(lr)) {
4183 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4184 ir_node *block = get_nodes_block(n);
4185 ir_mode *mode = get_irn_mode(left);
4187 ir_node *llr = get_Shr_right(ll);
4188 if (is_Const(llr)) {
4189 ir_graph *irg = current_ir_graph;
4190 dbg_info *dbg = get_irn_dbg_info(left);
4192 tarval *c1 = get_Const_tarval(llr);
4193 tarval *c2 = get_Const_tarval(lr);
4194 tarval *c3 = get_Const_tarval(right);
4195 tarval *mask = tarval_shl(c2, c1);
4196 tarval *value = tarval_shl(c3, c1);
4198 left = new_rd_And(dbg, irg, block, get_Shr_left(ll), new_Const(mask), mode);
4199 right = new_Const(value);
4204 } /* mode_is_int(...) */
4205 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4207 /* replace mode_b compares with ands/ors */
4208 if (get_irn_mode(left) == mode_b) {
4209 ir_graph *irg = current_ir_graph;
4210 ir_node *block = get_nodes_block(n);
4214 case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4215 case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break;
4216 case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4217 case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break;
4218 case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break;
4219 case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break;
4220 default: bres = NULL;
4223 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4229 * First step: normalize the compare op
4230 * by placing the constant on the right side
4231 * or moving the lower address node to the left.
4233 if (!operands_are_normalized(left, right)) {
4239 proj_nr = get_inversed_pnc(proj_nr);
4244 * Second step: Try to reduce the magnitude
4245 * of a constant. This may help to generate better code
4246 * later and may help to normalize more compares.
4247 * Of course this is only possible for integer values.
4249 tv = value_of(right);
4250 if (tv != tarval_bad) {
4251 mode = get_irn_mode(right);
4253 /* TODO extend to arbitrary constants */
4254 if (is_Conv(left) && tarval_is_null(tv)) {
4255 ir_node *op = get_Conv_op(left);
4256 ir_mode *op_mode = get_irn_mode(op);
4259 * UpConv(x) REL 0 ==> x REL 0
4261 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4262 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4263 mode_is_signed(mode) || !mode_is_signed(op_mode))) {
4264 tv = get_mode_null(op_mode);
4268 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4272 if (tv != tarval_bad) {
4273 /* the following optimization is possible on modes without Overflow
4274 * on Unary Minus or on == and !=:
4275 * -a CMP c ==> a swap(CMP) -c
4277 * Beware: for two-complement Overflow may occur, so only == and != can
4278 * be optimized, see this:
4279 * -MININT < 0 =/=> MININT > 0 !!!
4281 if (is_Minus(left) &&
4282 (!mode_overflow_on_unary_Minus(mode) ||
4283 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4284 tv = tarval_neg(tv);
4286 if (tv != tarval_bad) {
4287 left = get_Minus_op(left);
4288 proj_nr = get_inversed_pnc(proj_nr);
4290 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4292 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4293 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4294 tv = tarval_not(tv);
4296 if (tv != tarval_bad) {
4297 left = get_Not_op(left);
4299 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4303 /* for integer modes, we have more */
4304 if (mode_is_int(mode)) {
4305 /* Ne includes Unordered which is not possible on integers.
4306 * However, frontends often use this wrong, so fix it here */
4307 if (proj_nr & pn_Cmp_Uo) {
4308 proj_nr &= ~pn_Cmp_Uo;
4309 set_Proj_proj(proj, proj_nr);
4312 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4313 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4314 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4315 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4317 if (tv != tarval_bad) {
4318 proj_nr ^= pn_Cmp_Eq;
4320 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4323 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4324 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4325 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4326 tv = tarval_add(tv, get_mode_one(mode));
4328 if (tv != tarval_bad) {
4329 proj_nr ^= pn_Cmp_Eq;
4331 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4335 /* the following reassociations work only for == and != */
4336 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4338 #if 0 /* Might be not that good in general */
4339 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4340 if (tarval_is_null(tv) && is_Sub(left)) {
4341 right = get_Sub_right(left);
4342 left = get_Sub_left(left);
4344 tv = value_of(right);
4346 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4350 if (tv != tarval_bad) {
4351 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4353 ir_node *c1 = get_Sub_right(left);
4354 tarval *tv2 = value_of(c1);
4356 if (tv2 != tarval_bad) {
4357 tv2 = tarval_add(tv, value_of(c1));
4359 if (tv2 != tarval_bad) {
4360 left = get_Sub_left(left);
4363 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4367 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4368 else if (is_Add(left)) {
4369 ir_node *a_l = get_Add_left(left);
4370 ir_node *a_r = get_Add_right(left);
4374 if (is_Const(a_l)) {
4376 tv2 = value_of(a_l);
4379 tv2 = value_of(a_r);
4382 if (tv2 != tarval_bad) {
4383 tv2 = tarval_sub(tv, tv2, NULL);
4385 if (tv2 != tarval_bad) {
4389 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4393 /* -a == c ==> a == -c, -a != c ==> a != -c */
4394 else if (is_Minus(left)) {
4395 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4397 if (tv2 != tarval_bad) {
4398 left = get_Minus_op(left);
4401 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4406 /* the following reassociations work only for <= */
4407 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4408 if (tv != tarval_bad) {
4409 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4410 if (is_Abs(left)) { // TODO something is missing here
4416 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4417 switch (get_irn_opcode(left)) {
4421 c1 = get_And_right(left);
4424 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4425 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4427 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4429 /* TODO: move to constant evaluation */
4430 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4432 DBG_OPT_CSTEVAL(proj, c1);
4436 if (tarval_is_single_bit(tv)) {
4438 * optimization for AND:
4440 * And(x, C) == C ==> And(x, C) != 0
4441 * And(x, C) != C ==> And(X, C) == 0
4443 * if C is a single Bit constant.
4446 /* check for Constant's match. We have check hare the tarvals,
4447 because our const might be changed */
4448 if (get_Const_tarval(c1) == tv) {
4449 /* fine: do the transformation */
4450 tv = get_mode_null(get_tarval_mode(tv));
4451 proj_nr ^= pn_Cmp_Leg;
4453 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4459 c1 = get_Or_right(left);
4460 if (is_Const(c1) && tarval_is_null(tv)) {
4462 * Or(x, C) == 0 && C != 0 ==> FALSE
4463 * Or(x, C) != 0 && C != 0 ==> TRUE
4465 if (! tarval_is_null(get_Const_tarval(c1))) {
4466 /* TODO: move to constant evaluation */
4467 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4469 DBG_OPT_CSTEVAL(proj, c1);
4476 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4478 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4481 c1 = get_Shl_right(left);
4483 tarval *tv1 = get_Const_tarval(c1);
4484 ir_mode *mode = get_irn_mode(left);
4485 tarval *minus1 = get_mode_all_one(mode);
4486 tarval *amask = tarval_shr(minus1, tv1);
4487 tarval *cmask = tarval_shl(minus1, tv1);
4490 if (tarval_and(tv, cmask) != tv) {
4491 /* condition not met */
4492 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4494 DBG_OPT_CSTEVAL(proj, c1);
4497 sl = get_Shl_left(left);
4498 blk = get_nodes_block(n);
4499 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4500 tv = tarval_shr(tv, tv1);
4502 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4507 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4509 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4512 c1 = get_Shr_right(left);
4514 tarval *tv1 = get_Const_tarval(c1);
4515 ir_mode *mode = get_irn_mode(left);
4516 tarval *minus1 = get_mode_all_one(mode);
4517 tarval *amask = tarval_shl(minus1, tv1);
4518 tarval *cmask = tarval_shr(minus1, tv1);
4521 if (tarval_and(tv, cmask) != tv) {
4522 /* condition not met */
4523 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4525 DBG_OPT_CSTEVAL(proj, c1);
4528 sl = get_Shr_left(left);
4529 blk = get_nodes_block(n);
4530 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4531 tv = tarval_shl(tv, tv1);
4533 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4538 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4540 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4543 c1 = get_Shrs_right(left);
4545 tarval *tv1 = get_Const_tarval(c1);
4546 ir_mode *mode = get_irn_mode(left);
4547 tarval *minus1 = get_mode_all_one(mode);
4548 tarval *amask = tarval_shl(minus1, tv1);
4549 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4552 cond = tarval_sub(cond, tv1, NULL);
4553 cond = tarval_shrs(tv, cond);
4555 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4556 /* condition not met */
4557 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4559 DBG_OPT_CSTEVAL(proj, c1);
4562 sl = get_Shrs_left(left);
4563 blk = get_nodes_block(n);
4564 left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(amask), mode);
4565 tv = tarval_shl(tv, tv1);
4567 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4572 } /* tarval != bad */
4575 if (changed & 2) /* need a new Const */
4576 right = new_Const(tv);
4578 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4579 ir_node *op = get_Proj_pred(left);
4581 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4582 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4583 ir_node *c = get_binop_right(op);
4586 tarval *tv = get_Const_tarval(c);
4588 if (tarval_is_single_bit(tv)) {
4589 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4590 ir_node *v = get_binop_left(op);
4591 ir_node *blk = get_irn_n(op, -1);
4592 ir_mode *mode = get_irn_mode(v);
4594 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4595 left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(tv), mode);
4597 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4604 ir_node *block = get_nodes_block(n);
4606 /* create a new compare */
4607 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right);
4608 proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr);
4612 } /* transform_node_Proj_Cmp */
4615 * Optimize CopyB(mem, x, x) into a Nop.
4617 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4618 ir_node *copyb = get_Proj_pred(proj);
4619 ir_node *a = get_CopyB_dst(copyb);
4620 ir_node *b = get_CopyB_src(copyb);
4623 switch (get_Proj_proj(proj)) {
4624 case pn_CopyB_X_regular:
4625 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4626 DBG_OPT_EXC_REM(proj);
4627 proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb));
4629 case pn_CopyB_M_except:
4630 case pn_CopyB_X_except:
4631 DBG_OPT_EXC_REM(proj);
4632 proj = get_irg_bad(current_ir_graph);
4639 } /* transform_node_Proj_CopyB */
4642 * Optimize Bounds(idx, idx, upper) into idx.
4644 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4645 ir_node *oldn = proj;
4646 ir_node *bound = get_Proj_pred(proj);
4647 ir_node *idx = get_Bound_index(bound);
4648 ir_node *pred = skip_Proj(idx);
4651 if (idx == get_Bound_lower(bound))
4653 else if (is_Bound(pred)) {
4655 * idx was Bounds checked in the same MacroBlock previously,
4656 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4658 ir_node *lower = get_Bound_lower(bound);
4659 ir_node *upper = get_Bound_upper(bound);
4660 if (get_Bound_lower(pred) == lower &&
4661 get_Bound_upper(pred) == upper &&
4662 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4664 * One could expect that we simply return the previous
4665 * Bound here. However, this would be wrong, as we could
4666 * add an exception Proj to a new location then.
4667 * So, we must turn in into a tuple.
4673 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4674 switch (get_Proj_proj(proj)) {
4676 DBG_OPT_EXC_REM(proj);
4677 proj = get_Bound_mem(bound);
4679 case pn_Bound_X_except:
4680 DBG_OPT_EXC_REM(proj);
4681 proj = get_irg_bad(current_ir_graph);
4685 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4687 case pn_Bound_X_regular:
4688 DBG_OPT_EXC_REM(proj);
4689 proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound));
4696 } /* transform_node_Proj_Bound */
4699 * Does all optimizations on nodes that must be done on it's Proj's
4700 * because of creating new nodes.
4702 static ir_node *transform_node_Proj(ir_node *proj) {
4703 ir_node *n = get_Proj_pred(proj);
4705 if (n->op->ops.transform_node_Proj)
4706 return n->op->ops.transform_node_Proj(proj);
4708 } /* transform_node_Proj */
4711 * Move Confirms down through Phi nodes.
4713 static ir_node *transform_node_Phi(ir_node *phi) {
4715 ir_mode *mode = get_irn_mode(phi);
4717 if (mode_is_reference(mode)) {
4718 n = get_irn_arity(phi);
4720 /* Beware of Phi0 */
4722 ir_node *pred = get_irn_n(phi, 0);
4723 ir_node *bound, *new_Phi, *block, **in;
4726 if (! is_Confirm(pred))
4729 bound = get_Confirm_bound(pred);
4730 pnc = get_Confirm_cmp(pred);
4732 NEW_ARR_A(ir_node *, in, n);
4733 in[0] = get_Confirm_value(pred);
4735 for (i = 1; i < n; ++i) {
4736 pred = get_irn_n(phi, i);
4738 if (! is_Confirm(pred) ||
4739 get_Confirm_bound(pred) != bound ||
4740 get_Confirm_cmp(pred) != pnc)
4742 in[i] = get_Confirm_value(pred);
4744 /* move the Confirm nodes "behind" the Phi */
4745 block = get_irn_n(phi, -1);
4746 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
4747 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
4751 } /* transform_node_Phi */
4754 * Returns the operands of a commutative bin-op, if one operand is
4755 * a const, it is returned as the second one.
4757 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4758 ir_node *op_a = get_binop_left(binop);
4759 ir_node *op_b = get_binop_right(binop);
4761 assert(is_op_commutative(get_irn_op(binop)));
4763 if (is_Const(op_a)) {
4770 } /* get_comm_Binop_Ops */
4773 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4774 * Such pattern may arise in bitfield stores.
4776 * value c4 value c4 & c2
4777 * AND c3 AND c1 | c3
4784 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4787 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4790 ir_node *and_l, *c3;
4791 ir_node *value, *c4;
4792 ir_node *new_and, *new_const, *block;
4793 ir_mode *mode = get_irn_mode(or);
4795 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
4798 get_comm_Binop_Ops(or, &and, &c1);
4799 if (!is_Const(c1) || !is_And(and))
4802 get_comm_Binop_Ops(and, &or_l, &c2);
4806 tv1 = get_Const_tarval(c1);
4807 tv2 = get_Const_tarval(c2);
4809 tv = tarval_or(tv1, tv2);
4810 if (tarval_is_all_one(tv)) {
4811 /* the AND does NOT clear a bit with isn't set by the OR */
4812 set_Or_left(or, or_l);
4813 set_Or_right(or, c1);
4815 /* check for more */
4822 get_comm_Binop_Ops(or_l, &and_l, &c3);
4823 if (!is_Const(c3) || !is_And(and_l))
4826 get_comm_Binop_Ops(and_l, &value, &c4);
4830 /* ok, found the pattern, check for conditions */
4831 assert(mode == get_irn_mode(and));
4832 assert(mode == get_irn_mode(or_l));
4833 assert(mode == get_irn_mode(and_l));
4835 tv3 = get_Const_tarval(c3);
4836 tv4 = get_Const_tarval(c4);
4838 tv = tarval_or(tv4, tv2);
4839 if (!tarval_is_all_one(tv)) {
4840 /* have at least one 0 at the same bit position */
4844 n_tv4 = tarval_not(tv4);
4845 if (tv3 != tarval_and(tv3, n_tv4)) {
4846 /* bit in the or_mask is outside the and_mask */
4850 n_tv2 = tarval_not(tv2);
4851 if (tv1 != tarval_and(tv1, n_tv2)) {
4852 /* bit in the or_mask is outside the and_mask */
4856 /* ok, all conditions met */
4857 block = get_irn_n(or, -1);
4859 new_and = new_r_And(current_ir_graph, block,
4860 value, new_Const(tarval_and(tv4, tv2)), mode);
4862 new_const = new_Const(tarval_or(tv3, tv1));
4864 set_Or_left(or, new_and);
4865 set_Or_right(or, new_const);
4867 /* check for more */
4869 } /* transform_node_Or_bf_store */
4872 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4874 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4875 ir_mode *mode = get_irn_mode(or);
4876 ir_node *shl, *shr, *block;
4877 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4880 if (! mode_is_int(mode))
4883 shl = get_binop_left(or);
4884 shr = get_binop_right(or);
4893 } else if (!is_Shl(shl)) {
4895 } else if (!is_Shr(shr)) {
4898 x = get_Shl_left(shl);
4899 if (x != get_Shr_left(shr))
4902 c1 = get_Shl_right(shl);
4903 c2 = get_Shr_right(shr);
4904 if (is_Const(c1) && is_Const(c2)) {
4905 tv1 = get_Const_tarval(c1);
4906 if (! tarval_is_long(tv1))
4909 tv2 = get_Const_tarval(c2);
4910 if (! tarval_is_long(tv2))
4913 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4914 != (int) get_mode_size_bits(mode))
4917 /* yet, condition met */
4918 block = get_nodes_block(or);
4920 n = new_r_Rotl(current_ir_graph, block, x, c1, mode);
4922 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4929 rotval = sub; /* a Rot right is not supported, so use a rot left */
4930 } else if (is_Sub(c2)) {
4936 if (get_Sub_right(sub) != v)
4939 c1 = get_Sub_left(sub);
4943 tv1 = get_Const_tarval(c1);
4944 if (! tarval_is_long(tv1))
4947 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
4950 /* yet, condition met */
4951 block = get_nodes_block(or);
4953 n = new_r_Rotl(current_ir_graph, block, x, rotval, mode);
4955 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
4957 } /* transform_node_Or_Rotl */
4962 static ir_node *transform_node_Or(ir_node *n) {
4963 ir_node *c, *oldn = n;
4964 ir_node *a = get_Or_left(n);
4965 ir_node *b = get_Or_right(n);
4968 if (is_Not(a) && is_Not(b)) {
4969 /* ~a | ~b = ~(a&b) */
4970 ir_node *block = get_nodes_block(n);
4972 mode = get_irn_mode(n);
4975 n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode);
4976 n = new_rd_Not(get_irn_dbg_info(n), current_ir_graph, block, n, mode);
4977 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4981 /* we can evaluate 2 Projs of the same Cmp */
4982 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4983 ir_node *pred_a = get_Proj_pred(a);
4984 ir_node *pred_b = get_Proj_pred(b);
4985 if (pred_a == pred_b) {
4986 dbg_info *dbgi = get_irn_dbg_info(n);
4987 ir_node *block = get_nodes_block(pred_a);
4988 pn_Cmp pn_a = get_Proj_proj(a);
4989 pn_Cmp pn_b = get_Proj_proj(b);
4990 /* yes, we can simply calculate with pncs */
4991 pn_Cmp new_pnc = pn_a | pn_b;
4993 return new_rd_Proj(dbgi, current_ir_graph, block, pred_a, mode_b,
4998 mode = get_irn_mode(n);
4999 HANDLE_BINOP_PHI(tarval_or, a, b, c, mode);
5001 n = transform_node_Or_bf_store(n);
5002 n = transform_node_Or_Rotl(n);
5006 n = transform_bitwise_distributive(n, transform_node_Or);
5009 } /* transform_node_Or */
5013 static ir_node *transform_node(ir_node *n);
5016 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5018 * Should be moved to reassociation?
5020 static ir_node *transform_node_shift(ir_node *n) {
5021 ir_node *left, *right;
5023 tarval *tv1, *tv2, *res;
5024 ir_node *in[2], *irn, *block;
5026 left = get_binop_left(n);
5028 /* different operations */
5029 if (get_irn_op(left) != get_irn_op(n))
5032 right = get_binop_right(n);
5033 tv1 = value_of(right);
5034 if (tv1 == tarval_bad)
5037 tv2 = value_of(get_binop_right(left));
5038 if (tv2 == tarval_bad)
5041 res = tarval_add(tv1, tv2);
5042 mode = get_irn_mode(n);
5044 /* beware: a simple replacement works only, if res < modulo shift */
5046 int modulo_shf = get_mode_modulo_shift(mode);
5047 if (modulo_shf > 0) {
5048 tarval *modulo = new_tarval_from_long(modulo_shf,
5049 get_tarval_mode(res));
5051 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5053 /* shifting too much */
5054 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5056 ir_graph *irg = get_irn_irg(n);
5057 ir_node *block = get_nodes_block(n);
5058 dbg_info *dbgi = get_irn_dbg_info(n);
5059 ir_mode *smode = get_irn_mode(right);
5060 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5061 return new_rd_Shrs(dbgi, irg, block, get_binop_left(left),
5065 return new_Const(get_mode_null(mode));
5069 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5072 /* ok, we can replace it */
5073 block = get_nodes_block(n);
5075 in[0] = get_binop_left(left);
5076 in[1] = new_Const(res);
5078 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
5080 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5082 return transform_node(irn);
5083 } /* transform_node_shift */
5086 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5088 * - and, or, xor instead of &
5089 * - Shl, Shr, Shrs, rotl instead of >>
5090 * (with a special case for Or/Xor + Shrs)
5092 static ir_node *transform_node_bitop_shift(ir_node *n) {
5094 ir_node *right = get_binop_right(n);
5095 ir_mode *mode = get_irn_mode(n);
5096 ir_node *bitop_left;
5097 ir_node *bitop_right;
5109 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5111 if (!is_Const(right))
5114 left = get_binop_left(n);
5115 op_left = get_irn_op(left);
5116 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5119 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5120 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5121 /* TODO: test if sign bit is affectes */
5125 bitop_right = get_binop_right(left);
5126 if (!is_Const(bitop_right))
5129 bitop_left = get_binop_left(left);
5131 irg = get_irn_irg(n);
5132 block = get_nodes_block(n);
5133 dbgi = get_irn_dbg_info(n);
5134 tv1 = get_Const_tarval(bitop_right);
5135 tv2 = get_Const_tarval(right);
5137 assert(get_tarval_mode(tv1) == mode);
5140 new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode);
5141 tv_shift = tarval_shl(tv1, tv2);
5142 } else if(is_Shr(n)) {
5143 new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode);
5144 tv_shift = tarval_shr(tv1, tv2);
5145 } else if(is_Shrs(n)) {
5146 new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode);
5147 tv_shift = tarval_shrs(tv1, tv2);
5150 new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode);
5151 tv_shift = tarval_rotl(tv1, tv2);
5154 assert(get_tarval_mode(tv_shift) == mode);
5155 new_const = new_Const(tv_shift);
5157 if (op_left == op_And) {
5158 new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5159 } else if(op_left == op_Or) {
5160 new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode);
5162 assert(op_left == op_Eor);
5163 new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode);
5171 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5173 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5174 * (also with x >>s c1 when c1>=c2)
5176 static ir_node *transform_node_shl_shr(ir_node *n) {
5178 ir_node *right = get_binop_right(n);
5194 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5196 if (!is_Const(right))
5199 left = get_binop_left(n);
5200 mode = get_irn_mode(n);
5201 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5202 ir_node *shr_right = get_binop_right(left);
5204 if (!is_Const(shr_right))
5207 x = get_binop_left(left);
5208 tv_shr = get_Const_tarval(shr_right);
5209 tv_shl = get_Const_tarval(right);
5211 if (is_Shrs(left)) {
5212 /* shrs variant only allowed if c1 >= c2 */
5213 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5216 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5219 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5221 tv_mask = tarval_shl(tv_mask, tv_shl);
5222 } else if(is_Shr(n) && is_Shl(left)) {
5223 ir_node *shl_right = get_Shl_right(left);
5225 if (!is_Const(shl_right))
5228 x = get_Shl_left(left);
5229 tv_shr = get_Const_tarval(right);
5230 tv_shl = get_Const_tarval(shl_right);
5232 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5233 tv_mask = tarval_shr(tv_mask, tv_shr);
5238 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5239 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5242 assert(tv_mask != tarval_bad);
5243 assert(get_tarval_mode(tv_mask) == mode);
5245 irg = get_irn_irg(n);
5246 block = get_nodes_block(n);
5247 dbgi = get_irn_dbg_info(n);
5249 pnc = tarval_cmp(tv_shl, tv_shr);
5250 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5251 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5252 new_const = new_Const(tv_shift);
5254 new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode);
5256 new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode);
5259 assert(pnc == pn_Cmp_Gt);
5260 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5261 new_const = new_Const(tv_shift);
5262 new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode);
5265 new_const = new_Const(tv_mask);
5266 new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode);
5274 static ir_node *transform_node_Shr(ir_node *n) {
5275 ir_node *c, *oldn = n;
5276 ir_node *left = get_Shr_left(n);
5277 ir_node *right = get_Shr_right(n);
5278 ir_mode *mode = get_irn_mode(n);
5280 HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode);
5281 n = transform_node_shift(n);
5284 n = transform_node_shl_shr(n);
5286 n = transform_node_bitop_shift(n);
5289 } /* transform_node_Shr */
5294 static ir_node *transform_node_Shrs(ir_node *n) {
5295 ir_node *c, *oldn = n;
5296 ir_node *a = get_Shrs_left(n);
5297 ir_node *b = get_Shrs_right(n);
5298 ir_mode *mode = get_irn_mode(n);
5300 HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode);
5301 n = transform_node_shift(n);
5304 n = transform_node_bitop_shift(n);
5307 } /* transform_node_Shrs */
5312 static ir_node *transform_node_Shl(ir_node *n) {
5313 ir_node *c, *oldn = n;
5314 ir_node *a = get_Shl_left(n);
5315 ir_node *b = get_Shl_right(n);
5316 ir_mode *mode = get_irn_mode(n);
5318 HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode);
5319 n = transform_node_shift(n);
5322 n = transform_node_shl_shr(n);
5324 n = transform_node_bitop_shift(n);
5327 } /* transform_node_Shl */
5332 static ir_node *transform_node_Rotl(ir_node *n) {
5333 ir_node *c, *oldn = n;
5334 ir_node *a = get_Rotl_left(n);
5335 ir_node *b = get_Rotl_right(n);
5336 ir_mode *mode = get_irn_mode(n);
5338 HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode);
5339 n = transform_node_shift(n);
5342 n = transform_node_bitop_shift(n);
5345 } /* transform_node_Rotl */
5350 static ir_node *transform_node_Conv(ir_node *n) {
5351 ir_node *c, *oldn = n;
5352 ir_mode *mode = get_irn_mode(n);
5353 ir_node *a = get_Conv_op(n);
5355 if (mode != mode_b && is_const_Phi(a)) {
5356 /* Do NOT optimize mode_b Conv's, this leads to remaining
5357 * Phib nodes later, because the conv_b_lower operation
5358 * is instantly reverted, when it tries to insert a Convb.
5360 c = apply_conv_on_phi(a, mode);
5362 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5367 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5368 return new_r_Unknown(current_ir_graph, mode);
5371 if (mode_is_reference(mode) &&
5372 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5374 ir_node *l = get_Add_left(a);
5375 ir_node *r = get_Add_right(a);
5376 dbg_info *dbgi = get_irn_dbg_info(a);
5377 ir_node *block = get_nodes_block(n);
5379 ir_node *lop = get_Conv_op(l);
5380 if(get_irn_mode(lop) == mode) {
5381 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5382 n = new_rd_Add(dbgi, current_ir_graph, block, lop, r, mode);
5387 ir_node *rop = get_Conv_op(r);
5388 if(get_irn_mode(rop) == mode) {
5389 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5390 n = new_rd_Add(dbgi, current_ir_graph, block, l, rop, mode);
5397 } /* transform_node_Conv */
5400 * Remove dead blocks and nodes in dead blocks
5401 * in keep alive list. We do not generate a new End node.
5403 static ir_node *transform_node_End(ir_node *n) {
5404 int i, j, n_keepalives = get_End_n_keepalives(n);
5407 NEW_ARR_A(ir_node *, in, n_keepalives);
5409 for (i = j = 0; i < n_keepalives; ++i) {
5410 ir_node *ka = get_End_keepalive(n, i);
5412 if (! is_Block_dead(ka)) {
5416 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5418 } else if (is_Bad(ka)) {
5419 /* no need to keep Bad */
5424 if (j != n_keepalives)
5425 set_End_keepalives(n, j, in);
5427 } /* transform_node_End */
5429 /** returns 1 if a == -b */
5430 static int is_negated_value(ir_node *a, ir_node *b) {
5431 if (is_Minus(a) && get_Minus_op(a) == b)
5433 if (is_Minus(b) && get_Minus_op(b) == a)
5435 if (is_Sub(a) && is_Sub(b)) {
5436 ir_node *a_left = get_Sub_left(a);
5437 ir_node *a_right = get_Sub_right(a);
5438 ir_node *b_left = get_Sub_left(b);
5439 ir_node *b_right = get_Sub_right(b);
5441 if (a_left == b_right && a_right == b_left)
5449 * Optimize a Mux into some simpler cases.
5451 static ir_node *transform_node_Mux(ir_node *n) {
5452 ir_node *oldn = n, *sel = get_Mux_sel(n);
5453 ir_mode *mode = get_irn_mode(n);
5454 ir_node *t = get_Mux_true(n);
5455 ir_node *f = get_Mux_false(n);
5456 ir_graph *irg = current_ir_graph;
5458 /* first normalization step: move a possible zero to the false case */
5460 ir_node *cmp = get_Proj_pred(sel);
5463 if (is_Const(t) && is_Const_null(t)) {
5466 /* Mux(x, 0, y) => Mux(x, y, 0) */
5467 pn_Cmp pnc = get_Proj_proj(sel);
5468 sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b,
5469 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5470 n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode);
5478 /* note: after normalization, false can only happen on default */
5479 if (mode == mode_b) {
5480 dbg_info *dbg = get_irn_dbg_info(n);
5481 ir_node *block = get_nodes_block(n);
5482 ir_graph *irg = current_ir_graph;
5485 tarval *tv_t = get_Const_tarval(t);
5486 if (tv_t == tarval_b_true) {
5488 /* Muxb(sel, true, false) = sel */
5489 assert(get_Const_tarval(f) == tarval_b_false);
5490 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5493 /* Muxb(sel, true, x) = Or(sel, x) */
5494 n = new_rd_Or(dbg, irg, block, sel, f, mode_b);
5495 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5499 } else if (is_Const(f)) {
5500 tarval *tv_f = get_Const_tarval(f);
5501 if (tv_f == tarval_b_true) {
5502 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5503 ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b);
5504 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5505 n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b);
5508 /* Muxb(sel, x, false) = And(sel, x) */
5509 assert(tv_f == tarval_b_false);
5510 n = new_rd_And(dbg, irg, block, sel, t, mode_b);
5511 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5517 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5518 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5519 tarval *a = get_Const_tarval(t);
5520 tarval *b = get_Const_tarval(f);
5521 tarval *null = get_tarval_null(mode);
5524 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5525 diff = tarval_sub(a, b, NULL);
5528 diff = tarval_sub(b, a, NULL);
5532 if (diff == get_tarval_one(mode) && min != null) {
5533 dbg_info *dbg = get_irn_dbg_info(n);
5534 ir_node *block = get_nodes_block(n);
5535 ir_graph *irg = current_ir_graph;
5536 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5537 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5538 n = new_rd_Mux(dbg, irg, block, sel, f, t, mode);
5539 n = new_rd_Add(dbg, irg, block, n, new_Const(min), mode);
5545 ir_node *cmp = get_Proj_pred(sel);
5546 long pn = get_Proj_proj(sel);
5549 * Note: normalization puts the constant on the right side,
5550 * so we check only one case.
5552 * Note further that these optimization work even for floating point
5553 * with NaN's because -NaN == NaN.
5554 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5558 ir_node *cmp_r = get_Cmp_right(cmp);
5559 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5560 ir_node *block = get_nodes_block(n);
5561 ir_node *cmp_l = get_Cmp_left(cmp);
5563 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5566 if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))
5567 || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)))
5569 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5570 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5572 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5574 } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))
5575 || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)))
5577 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5578 n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block,
5580 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
5582 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5587 if (mode_is_int(mode)) {
5589 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5590 /* Mux((a & b) != 0, c, 0) */
5591 ir_node *and_r = get_And_right(cmp_l);
5594 if (and_r == t && f == cmp_r) {
5595 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5596 if (pn == pn_Cmp_Lg) {
5597 /* Mux((a & 2^C) != 0, 2^C, 0) */
5599 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5601 /* Mux((a & 2^C) == 0, 2^C, 0) */
5602 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5603 block, cmp_l, t, mode);
5604 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5609 if (is_Shl(and_r)) {
5610 ir_node *shl_l = get_Shl_left(and_r);
5611 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5612 if (and_r == t && f == cmp_r) {
5613 if (pn == pn_Cmp_Lg) {
5614 /* (a & (1 << n)) != 0, (1 << n), 0) */
5616 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5618 /* (a & (1 << n)) == 0, (1 << n), 0) */
5619 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5620 block, cmp_l, t, mode);
5621 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5627 and_l = get_And_left(cmp_l);
5628 if (is_Shl(and_l)) {
5629 ir_node *shl_l = get_Shl_left(and_l);
5630 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5631 if (and_l == t && f == cmp_r) {
5632 if (pn == pn_Cmp_Lg) {
5633 /* ((1 << n) & a) != 0, (1 << n), 0) */
5635 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5637 /* ((1 << n) & a) == 0, (1 << n), 0) */
5638 n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph,
5639 block, cmp_l, t, mode);
5640 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5653 } /* transform_node_Mux */
5656 * optimize Sync nodes that have other syncs as input we simply add the inputs
5657 * of the other sync to our own inputs
5659 static ir_node *transform_node_Sync(ir_node *n) {
5660 int arity = get_Sync_n_preds(n);
5663 for (i = 0; i < arity;) {
5664 ir_node *pred = get_Sync_pred(n, i);
5668 if (!is_Sync(pred)) {
5676 pred_arity = get_Sync_n_preds(pred);
5677 for (j = 0; j < pred_arity; ++j) {
5678 ir_node *pred_pred = get_Sync_pred(pred, j);
5683 add_irn_n(n, pred_pred);
5687 if (get_Sync_pred(n, k) == pred_pred) break;
5692 /* rehash the sync node */
5693 add_identities(current_ir_graph->value_table, n);
5699 * Tries several [inplace] [optimizing] transformations and returns an
5700 * equivalent node. The difference to equivalent_node() is that these
5701 * transformations _do_ generate new nodes, and thus the old node must
5702 * not be freed even if the equivalent node isn't the old one.
5704 static ir_node *transform_node(ir_node *n) {
5708 * Transform_node is the only "optimizing transformation" that might
5709 * return a node with a different opcode. We iterate HERE until fixpoint
5710 * to get the final result.
5714 if (n->op->ops.transform_node)
5715 n = n->op->ops.transform_node(n);
5716 } while (oldn != n);
5719 } /* transform_node */
5722 * Sets the default transform node operation for an ir_op_ops.
5724 * @param code the opcode for the default operation
5725 * @param ops the operations initialized
5730 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5734 ops->transform_node = transform_node_##a; \
5736 #define CASE_PROJ(a) \
5738 ops->transform_node_Proj = transform_node_Proj_##a; \
5740 #define CASE_PROJ_EX(a) \
5742 ops->transform_node = transform_node_##a; \
5743 ops->transform_node_Proj = transform_node_Proj_##a; \
5752 CASE_PROJ_EX(DivMod);
5786 } /* firm_set_default_transform_node */
5789 /* **************** Common Subexpression Elimination **************** */
5791 /** The size of the hash table used, should estimate the number of nodes
5793 #define N_IR_NODES 512
5795 /** Compares the attributes of two Const nodes. */
5796 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
5797 return (get_Const_tarval(a) != get_Const_tarval(b))
5798 || (get_Const_type(a) != get_Const_type(b));
5799 } /* node_cmp_attr_Const */
5801 /** Compares the attributes of two Proj nodes. */
5802 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
5803 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
5804 } /* node_cmp_attr_Proj */
5806 /** Compares the attributes of two Filter nodes. */
5807 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
5808 return get_Filter_proj(a) != get_Filter_proj(b);
5809 } /* node_cmp_attr_Filter */
5811 /** Compares the attributes of two Alloc nodes. */
5812 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
5813 const alloc_attr *pa = get_irn_alloc_attr(a);
5814 const alloc_attr *pb = get_irn_alloc_attr(b);
5815 return (pa->where != pb->where) || (pa->type != pb->type);
5816 } /* node_cmp_attr_Alloc */
5818 /** Compares the attributes of two Free nodes. */
5819 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
5820 const free_attr *pa = get_irn_free_attr(a);
5821 const free_attr *pb = get_irn_free_attr(b);
5822 return (pa->where != pb->where) || (pa->type != pb->type);
5823 } /* node_cmp_attr_Free */
5825 /** Compares the attributes of two SymConst nodes. */
5826 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
5827 const symconst_attr *pa = get_irn_symconst_attr(a);
5828 const symconst_attr *pb = get_irn_symconst_attr(b);
5829 return (pa->kind != pb->kind)
5830 || (pa->sym.type_p != pb->sym.type_p)
5831 || (pa->tp != pb->tp);
5832 } /* node_cmp_attr_SymConst */
5834 /** Compares the attributes of two Call nodes. */
5835 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
5836 return get_irn_call_attr(a) != get_irn_call_attr(b);
5837 } /* node_cmp_attr_Call */
5839 /** Compares the attributes of two Sel nodes. */
5840 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
5841 const ir_entity *a_ent = get_Sel_entity(a);
5842 const ir_entity *b_ent = get_Sel_entity(b);
5845 (a_ent->kind != b_ent->kind) ||
5846 (a_ent->name != b_ent->name) ||
5847 (a_ent->owner != b_ent->owner) ||
5848 (a_ent->ld_name != b_ent->ld_name) ||
5849 (a_ent->type != b_ent->type);
5851 /* Matze: inlining of functions can produce 2 entities with same type,
5853 return a_ent != b_ent;
5854 } /* node_cmp_attr_Sel */
5856 /** Compares the attributes of two Phi nodes. */
5857 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
5858 /* we can only enter this function if both nodes have the same number of inputs,
5859 hence it is enough to check if one of them is a Phi0 */
5861 /* check the Phi0 pos attribute */
5862 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
5865 } /* node_cmp_attr_Phi */
5867 /** Compares the attributes of two Conv nodes. */
5868 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
5869 return get_Conv_strict(a) != get_Conv_strict(b);
5870 } /* node_cmp_attr_Conv */
5872 /** Compares the attributes of two Cast nodes. */
5873 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
5874 return get_Cast_type(a) != get_Cast_type(b);
5875 } /* node_cmp_attr_Cast */
5877 /** Compares the attributes of two Load nodes. */
5878 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
5879 if (get_Load_volatility(a) == volatility_is_volatile ||
5880 get_Load_volatility(b) == volatility_is_volatile)
5881 /* NEVER do CSE on volatile Loads */
5883 /* do not CSE Loads with different alignment. Be conservative. */
5884 if (get_Load_align(a) != get_Load_align(b))
5887 return get_Load_mode(a) != get_Load_mode(b);
5888 } /* node_cmp_attr_Load */
5890 /** Compares the attributes of two Store nodes. */
5891 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
5892 /* do not CSE Stores with different alignment. Be conservative. */
5893 if (get_Store_align(a) != get_Store_align(b))
5896 /* NEVER do CSE on volatile Stores */
5897 return (get_Store_volatility(a) == volatility_is_volatile ||
5898 get_Store_volatility(b) == volatility_is_volatile);
5899 } /* node_cmp_attr_Store */
5901 /** Compares two exception attributes */
5902 static int node_cmp_exception(ir_node *a, ir_node *b) {
5903 const except_attr *ea = get_irn_except_attr(a);
5904 const except_attr *eb = get_irn_except_attr(b);
5906 return ea->pin_state != eb->pin_state;
5909 #define node_cmp_attr_Bound node_cmp_exception
5911 /** Compares the attributes of two Div nodes. */
5912 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
5913 const divmod_attr *ma = get_irn_divmod_attr(a);
5914 const divmod_attr *mb = get_irn_divmod_attr(b);
5915 return ma->exc.pin_state != mb->exc.pin_state ||
5916 ma->res_mode != mb->res_mode ||
5917 ma->no_remainder != mb->no_remainder;
5918 } /* node_cmp_attr_Div */
5920 /** Compares the attributes of two DivMod nodes. */
5921 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
5922 const divmod_attr *ma = get_irn_divmod_attr(a);
5923 const divmod_attr *mb = get_irn_divmod_attr(b);
5924 return ma->exc.pin_state != mb->exc.pin_state ||
5925 ma->res_mode != mb->res_mode;
5926 } /* node_cmp_attr_DivMod */
5928 /** Compares the attributes of two Mod nodes. */
5929 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
5930 const divmod_attr *ma = get_irn_divmod_attr(a);
5931 const divmod_attr *mb = get_irn_divmod_attr(b);
5932 return ma->exc.pin_state != mb->exc.pin_state ||
5933 ma->res_mode != mb->res_mode;
5934 } /* node_cmp_attr_Mod */
5936 /** Compares the attributes of two Quot nodes. */
5937 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
5938 const divmod_attr *ma = get_irn_divmod_attr(a);
5939 const divmod_attr *mb = get_irn_divmod_attr(b);
5940 return ma->exc.pin_state != mb->exc.pin_state ||
5941 ma->res_mode != mb->res_mode;
5942 } /* node_cmp_attr_Quot */
5944 /** Compares the attributes of two Confirm nodes. */
5945 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
5946 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5947 } /* node_cmp_attr_Confirm */
5949 /** Compares the attributes of two ASM nodes. */
5950 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
5952 const ir_asm_constraint *ca;
5953 const ir_asm_constraint *cb;
5956 if (get_ASM_text(a) != get_ASM_text(b))
5959 /* Should we really check the constraints here? Should be better, but is strange. */
5960 n = get_ASM_n_input_constraints(a);
5961 if (n != get_ASM_n_input_constraints(b))
5964 ca = get_ASM_input_constraints(a);
5965 cb = get_ASM_input_constraints(b);
5966 for (i = 0; i < n; ++i) {
5967 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5971 n = get_ASM_n_output_constraints(a);
5972 if (n != get_ASM_n_output_constraints(b))
5975 ca = get_ASM_output_constraints(a);
5976 cb = get_ASM_output_constraints(b);
5977 for (i = 0; i < n; ++i) {
5978 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
5982 n = get_ASM_n_clobbers(a);
5983 if (n != get_ASM_n_clobbers(b))
5986 cla = get_ASM_clobbers(a);
5987 clb = get_ASM_clobbers(b);
5988 for (i = 0; i < n; ++i) {
5989 if (cla[i] != clb[i])
5993 } /* node_cmp_attr_ASM */
5996 * Set the default node attribute compare operation for an ir_op_ops.
5998 * @param code the opcode for the default operation
5999 * @param ops the operations initialized
6004 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6008 ops->node_cmp_attr = node_cmp_attr_##a; \
6039 } /* firm_set_default_node_cmp_attr */
6042 * Compare function for two nodes in the value table. Gets two
6043 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6045 int identities_cmp(const void *elt, const void *key) {
6046 ir_node *a = (ir_node *)elt;
6047 ir_node *b = (ir_node *)key;
6050 if (a == b) return 0;
6052 if ((get_irn_op(a) != get_irn_op(b)) ||
6053 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6055 /* compare if a's in and b's in are of equal length */
6056 irn_arity_a = get_irn_intra_arity(a);
6057 if (irn_arity_a != get_irn_intra_arity(b))
6060 if (get_irn_pinned(a) == op_pin_state_pinned) {
6061 /* for pinned nodes, the block inputs must be equal */
6062 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6064 } else if (! get_opt_global_cse()) {
6065 /* for block-local CSE both nodes must be in the same MacroBlock */
6066 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6070 /* compare a->in[0..ins] with b->in[0..ins] */
6071 for (i = 0; i < irn_arity_a; i++)
6072 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
6076 * here, we already now that the nodes are identical except their
6079 if (a->op->ops.node_cmp_attr)
6080 return a->op->ops.node_cmp_attr(a, b);
6083 } /* identities_cmp */
6086 * Calculate a hash value of a node.
6088 * @param node The IR-node
6090 unsigned ir_node_hash(const ir_node *node) {
6091 return node->op->ops.hash(node);
6092 } /* ir_node_hash */
6095 pset *new_identities(void) {
6096 return new_pset(identities_cmp, N_IR_NODES);
6097 } /* new_identities */
6099 void del_identities(pset *value_table) {
6100 del_pset(value_table);
6101 } /* del_identities */
6103 /* Normalize a node by putting constants (and operands with larger
6104 * node index) on the right (operator side). */
6105 void ir_normalize_node(ir_node *n) {
6106 if (is_op_commutative(get_irn_op(n))) {
6107 ir_node *l = get_binop_left(n);
6108 ir_node *r = get_binop_right(n);
6110 /* For commutative operators perform a OP b == b OP a but keep
6111 * constants on the RIGHT side. This helps greatly in some
6112 * optimizations. Moreover we use the idx number to make the form
6114 if (!operands_are_normalized(l, r)) {
6115 set_binop_left(n, r);
6116 set_binop_right(n, l);
6120 } /* ir_normalize_node */
6123 * Update the nodes after a match in the value table. If both nodes have
6124 * the same MacroBlock but different Blocks, we must ensure that the node
6125 * with the dominating Block (the node that is near to the MacroBlock header
6126 * is stored in the table.
6127 * Because a MacroBlock has only one "non-exception" flow, we don't need
6128 * dominance info here: We known, that one block must dominate the other and
6129 * following the only block input will allow to find it.
6131 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6132 ir_node *known_blk, *new_block, *block, *mbh;
6134 if (get_opt_global_cse()) {
6135 /* Block inputs are meaning less */
6138 known_blk = get_irn_n(known_irn, -1);
6139 new_block = get_irn_n(new_ir_node, -1);
6140 if (known_blk == new_block) {
6141 /* already in the same block */
6145 * We expect the typical case when we built the graph. In that case, the
6146 * known_irn is already the upper one, so checking this should be faster.
6149 mbh = get_Block_MacroBlock(new_block);
6151 if (block == known_blk) {
6152 /* ok, we have found it: known_block dominates new_block as expected */
6157 * We have reached the MacroBlock header NOT founding
6158 * the known_block. new_block must dominate known_block.
6161 set_irn_n(known_irn, -1, new_block);
6164 assert(get_Block_n_cfgpreds(block) == 1);
6165 block = get_Block_cfgpred_block(block, 0);
6167 } /* update_value_table */
6170 * Return the canonical node computing the same value as n.
6171 * Looks up the node in a hash table, enters it in the table
6172 * if it isn't there yet.
6174 * @param value_table the HashSet containing all nodes in the
6176 * @param n the node to look up
6178 * @return a node that computes the same value as n or n if no such
6179 * node could be found
6181 ir_node *identify_remember(pset *value_table, ir_node *n) {
6184 if (!value_table) return n;
6186 ir_normalize_node(n);
6187 /* lookup or insert in hash table with given hash key. */
6188 o = pset_insert(value_table, n, ir_node_hash(n));
6191 update_known_irn(o, n);
6195 } /* identify_remember */
6198 * During construction we set the op_pin_state_pinned flag in the graph right when the
6199 * optimization is performed. The flag turning on procedure global cse could
6200 * be changed between two allocations. This way we are safe.
6202 * @param value_table The value table
6203 * @param n The node to lookup
6205 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6208 n = identify_remember(value_table, n);
6209 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6210 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6212 } /* identify_cons */
6214 /* Add a node to the identities value table. */
6215 void add_identities(pset *value_table, ir_node *node) {
6216 if (get_opt_cse() && is_no_Block(node))
6217 identify_remember(value_table, node);
6218 } /* add_identities */
6220 /* Visit each node in the value table of a graph. */
6221 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6223 ir_graph *rem = current_ir_graph;
6225 current_ir_graph = irg;
6226 foreach_pset(irg->value_table, node)
6228 current_ir_graph = rem;
6229 } /* visit_all_identities */
6232 * Garbage in, garbage out. If a node has a dead input, i.e., the
6233 * Bad node is input to the node, return the Bad node.
6235 static ir_node *gigo(ir_node *node) {
6237 ir_op *op = get_irn_op(node);
6239 /* remove garbage blocks by looking at control flow that leaves the block
6240 and replacing the control flow by Bad. */
6241 if (get_irn_mode(node) == mode_X) {
6242 ir_node *block = get_nodes_block(skip_Proj(node));
6244 /* Don't optimize nodes in immature blocks. */
6245 if (!get_Block_matured(block))
6247 /* Don't optimize End, may have Bads. */
6248 if (op == op_End) return node;
6250 if (is_Block(block)) {
6251 if (is_Block_dead(block)) {
6252 /* control flow from dead block is dead */
6256 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6257 if (!is_Bad(get_irn_n(block, i)))
6261 ir_graph *irg = get_irn_irg(block);
6262 /* the start block is never dead */
6263 if (block != get_irg_start_block(irg)
6264 && block != get_irg_end_block(irg)) {
6266 * Do NOT kill control flow without setting
6267 * the block to dead of bad things can happen:
6268 * We get a Block that is not reachable be irg_block_walk()
6269 * but can be found by irg_walk()!
6271 set_Block_dead(block);
6278 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6279 blocks predecessors is dead. */
6280 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6281 irn_arity = get_irn_arity(node);
6284 * Beware: we can only read the block of a non-floating node.
6286 if (is_irn_pinned_in_irg(node) &&
6287 is_Block_dead(get_nodes_block(skip_Proj(node))))
6290 for (i = 0; i < irn_arity; i++) {
6291 ir_node *pred = get_irn_n(node, i);
6296 /* Propagating Unknowns here seems to be a bad idea, because
6297 sometimes we need a node as a input and did not want that
6299 However, it might be useful to move this into a later phase
6300 (if you think that optimizing such code is useful). */
6301 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6302 return new_Unknown(get_irn_mode(node));
6307 /* With this code we violate the agreement that local_optimize
6308 only leaves Bads in Block, Phi and Tuple nodes. */
6309 /* If Block has only Bads as predecessors it's garbage. */
6310 /* If Phi has only Bads as predecessors it's garbage. */
6311 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6312 irn_arity = get_irn_arity(node);
6313 for (i = 0; i < irn_arity; i++) {
6314 if (!is_Bad(get_irn_n(node, i))) break;
6316 if (i == irn_arity) node = new_Bad();
6323 * These optimizations deallocate nodes from the obstack.
6324 * It can only be called if it is guaranteed that no other nodes
6325 * reference this one, i.e., right after construction of a node.
6327 * @param n The node to optimize
6329 * current_ir_graph must be set to the graph of the node!
6331 ir_node *optimize_node(ir_node *n) {
6334 ir_opcode iro = get_irn_opcode(n);
6336 /* Always optimize Phi nodes: part of the construction. */
6337 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6339 /* constant expression evaluation / constant folding */
6340 if (get_opt_constant_folding()) {
6341 /* neither constants nor Tuple values can be evaluated */
6342 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6343 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6344 int old_fp_mode = tarval_fp_ops_enabled();
6346 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6348 /* try to evaluate */
6349 tv = computed_value(n);
6350 if (tv != tarval_bad) {
6352 ir_type *old_tp = get_irn_type(n);
6353 int i, arity = get_irn_arity(n);
6357 * Try to recover the type of the new expression.
6359 for (i = 0; i < arity && !old_tp; ++i)
6360 old_tp = get_irn_type(get_irn_n(n, i));
6363 * we MUST copy the node here temporary, because it's still needed
6364 * for DBG_OPT_CSTEVAL
6366 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6367 oldn = alloca(node_size);
6369 memcpy(oldn, n, node_size);
6370 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6372 /* ARG, copy the in array, we need it for statistics */
6373 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6375 /* note the inplace edges module */
6376 edges_node_deleted(n, current_ir_graph);
6378 /* evaluation was successful -- replace the node. */
6379 irg_kill_node(current_ir_graph, n);
6382 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6383 set_Const_type(nw, old_tp);
6384 DBG_OPT_CSTEVAL(oldn, nw);
6385 tarval_enable_fp_ops(old_fp_mode);
6388 tarval_enable_fp_ops(old_fp_mode);
6392 /* remove unnecessary nodes */
6393 if (get_opt_algebraic_simplification() ||
6394 (iro == iro_Phi) || /* always optimize these nodes. */
6396 (iro == iro_Proj) ||
6397 (iro == iro_Block) ) /* Flags tested local. */
6398 n = equivalent_node(n);
6400 /* Common Subexpression Elimination.
6402 * Checks whether n is already available.
6403 * The block input is used to distinguish different subexpressions. Right
6404 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6405 * subexpressions within a block.
6408 n = identify_cons(current_ir_graph->value_table, n);
6411 edges_node_deleted(oldn, current_ir_graph);
6413 /* We found an existing, better node, so we can deallocate the old node. */
6414 irg_kill_node(current_ir_graph, oldn);
6418 /* Some more constant expression evaluation that does not allow to
6420 iro = get_irn_opcode(n);
6421 if (get_opt_algebraic_simplification() ||
6422 (iro == iro_Cond) ||
6423 (iro == iro_Proj)) /* Flags tested local. */
6424 n = transform_node(n);
6426 /* Remove nodes with dead (Bad) input.
6427 Run always for transformation induced Bads. */
6430 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6431 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6433 n = identify_remember(current_ir_graph->value_table, o);
6439 } /* optimize_node */
6443 * These optimizations never deallocate nodes (in place). This can cause dead
6444 * nodes lying on the obstack. Remove these by a dead node elimination,
6445 * i.e., a copying garbage collection.
6447 ir_node *optimize_in_place_2(ir_node *n) {
6450 ir_opcode iro = get_irn_opcode(n);
6452 if (!get_opt_optimize() && !is_Phi(n)) return n;
6454 /* constant expression evaluation / constant folding */
6455 if (get_opt_constant_folding()) {
6456 /* neither constants nor Tuple values can be evaluated */
6457 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6458 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6459 int old_fp_mode = tarval_fp_ops_enabled();
6461 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6462 /* try to evaluate */
6463 tv = computed_value(n);
6464 if (tv != tarval_bad) {
6465 /* evaluation was successful -- replace the node. */
6466 ir_type *old_tp = get_irn_type(n);
6467 int i, arity = get_irn_arity(n);
6470 * Try to recover the type of the new expression.
6472 for (i = 0; i < arity && !old_tp; ++i)
6473 old_tp = get_irn_type(get_irn_n(n, i));
6477 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6478 set_Const_type(n, old_tp);
6480 DBG_OPT_CSTEVAL(oldn, n);
6481 tarval_enable_fp_ops(old_fp_mode);
6484 tarval_enable_fp_ops(old_fp_mode);
6488 /* remove unnecessary nodes */
6489 if (get_opt_constant_folding() ||
6490 (iro == iro_Phi) || /* always optimize these nodes. */
6491 (iro == iro_Id) || /* ... */
6492 (iro == iro_Proj) || /* ... */
6493 (iro == iro_Block) ) /* Flags tested local. */
6494 n = equivalent_node(n);
6496 /** common subexpression elimination **/
6497 /* Checks whether n is already available. */
6498 /* The block input is used to distinguish different subexpressions. Right
6499 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6500 subexpressions within a block. */
6501 if (get_opt_cse()) {
6503 n = identify_remember(current_ir_graph->value_table, o);
6508 /* Some more constant expression evaluation. */
6509 iro = get_irn_opcode(n);
6510 if (get_opt_constant_folding() ||
6511 (iro == iro_Cond) ||
6512 (iro == iro_Proj)) /* Flags tested local. */
6513 n = transform_node(n);
6515 /* Remove nodes with dead (Bad) input.
6516 Run always for transformation induced Bads. */
6519 /* Now we can verify the node, as it has no dead inputs any more. */
6522 /* Now we have a legal, useful node. Enter it in hash table for cse.
6523 Blocks should be unique anyways. (Except the successor of start:
6524 is cse with the start block!) */
6525 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6527 n = identify_remember(current_ir_graph->value_table, o);
6533 } /* optimize_in_place_2 */
6536 * Wrapper for external use, set proper status bits after optimization.
6538 ir_node *optimize_in_place(ir_node *n) {
6539 /* Handle graph state */
6540 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6542 if (get_opt_global_cse())
6543 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6544 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6545 set_irg_outs_inconsistent(current_ir_graph);
6547 /* FIXME: Maybe we could also test whether optimizing the node can
6548 change the control graph. */
6549 set_irg_doms_inconsistent(current_ir_graph);
6550 return optimize_in_place_2(n);
6551 } /* optimize_in_place */
6554 * Calculate a hash value of a Const node.
6556 static unsigned hash_Const(const ir_node *node) {
6559 /* special value for const, as they only differ in their tarval. */
6560 h = HASH_PTR(node->attr.con.tv);
6561 h = 9*h + HASH_PTR(get_irn_mode(node));
6567 * Calculate a hash value of a SymConst node.
6569 static unsigned hash_SymConst(const ir_node *node) {
6572 /* special value for const, as they only differ in their symbol. */
6573 h = HASH_PTR(node->attr.symc.sym.type_p);
6574 h = 9*h + HASH_PTR(get_irn_mode(node));
6577 } /* hash_SymConst */
6580 * Set the default hash operation in an ir_op_ops.
6582 * @param code the opcode for the default operation
6583 * @param ops the operations initialized
6588 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6592 ops->hash = hash_##a; \
6595 /* hash function already set */
6596 if (ops->hash != NULL)
6603 /* use input/mode default hash if no function was given */
6604 ops->hash = firm_default_hash;
6612 * Sets the default operation for an ir_ops.
6614 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6615 ops = firm_set_default_hash(code, ops);
6616 ops = firm_set_default_computed_value(code, ops);
6617 ops = firm_set_default_equivalent_node(code, ops);
6618 ops = firm_set_default_transform_node(code, ops);
6619 ops = firm_set_default_node_cmp_attr(code, ops);
6620 ops = firm_set_default_get_type(code, ops);
6621 ops = firm_set_default_get_type_attr(code, ops);
6622 ops = firm_set_default_get_entity_attr(code, ops);
6625 } /* firm_set_default_operations */