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 #include "firm_types.h"
53 /* Make types visible to allow most efficient access */
57 * Returns the tarval of a Const node or tarval_bad for all other nodes.
59 static tarval *default_value_of(const ir_node *n) {
61 return get_Const_tarval(n); /* might return tarval_bad */
66 value_of_func value_of_ptr = default_value_of;
68 /* * Set a new value_of function. */
69 void set_value_of_func(value_of_func func) {
73 value_of_ptr = default_value_of;
77 * Return the value of a Constant.
79 static tarval *computed_value_Const(const ir_node *n) {
80 return get_Const_tarval(n);
81 } /* computed_value_Const */
84 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
86 static tarval *computed_value_SymConst(const ir_node *n) {
90 switch (get_SymConst_kind(n)) {
91 case symconst_type_size:
92 type = get_SymConst_type(n);
93 if (get_type_state(type) == layout_fixed)
94 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
96 case symconst_type_align:
97 type = get_SymConst_type(n);
98 if (get_type_state(type) == layout_fixed)
99 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
101 case symconst_ofs_ent:
102 ent = get_SymConst_entity(n);
103 type = get_entity_owner(ent);
104 if (get_type_state(type) == layout_fixed)
105 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
111 } /* computed_value_SymConst */
114 * Return the value of an Add.
116 static tarval *computed_value_Add(const ir_node *n) {
117 ir_node *a = get_Add_left(n);
118 ir_node *b = get_Add_right(n);
120 tarval *ta = value_of(a);
121 tarval *tb = value_of(b);
123 if ((ta != tarval_bad) && (tb != tarval_bad))
124 return tarval_add(ta, tb);
127 } /* computed_value_Add */
130 * Return the value of a Sub.
131 * Special case: a - a
133 static tarval *computed_value_Sub(const ir_node *n) {
134 ir_mode *mode = get_irn_mode(n);
135 ir_node *a = get_Sub_left(n);
136 ir_node *b = get_Sub_right(n);
141 if (! mode_is_float(mode)) {
144 return get_mode_null(mode);
150 if ((ta != tarval_bad) && (tb != tarval_bad))
151 return tarval_sub(ta, tb, mode);
154 } /* computed_value_Sub */
157 * Return the value of a Carry.
158 * Special : a op 0, 0 op b
160 static tarval *computed_value_Carry(const ir_node *n) {
161 ir_node *a = get_binop_left(n);
162 ir_node *b = get_binop_right(n);
163 ir_mode *m = get_irn_mode(n);
165 tarval *ta = value_of(a);
166 tarval *tb = value_of(b);
168 if ((ta != tarval_bad) && (tb != tarval_bad)) {
170 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
172 if (tarval_is_null(ta) || tarval_is_null(tb))
173 return get_mode_null(m);
176 } /* computed_value_Carry */
179 * Return the value of a Borrow.
182 static tarval *computed_value_Borrow(const ir_node *n) {
183 ir_node *a = get_binop_left(n);
184 ir_node *b = get_binop_right(n);
185 ir_mode *m = get_irn_mode(n);
187 tarval *ta = value_of(a);
188 tarval *tb = value_of(b);
190 if ((ta != tarval_bad) && (tb != tarval_bad)) {
191 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
192 } else if (tarval_is_null(ta)) {
193 return get_mode_null(m);
196 } /* computed_value_Borrow */
199 * Return the value of an unary Minus.
201 static tarval *computed_value_Minus(const ir_node *n) {
202 ir_node *a = get_Minus_op(n);
203 tarval *ta = value_of(a);
205 if (ta != tarval_bad)
206 return tarval_neg(ta);
209 } /* computed_value_Minus */
212 * Return the value of a Mul.
214 static tarval *computed_value_Mul(const ir_node *n) {
215 ir_node *a = get_Mul_left(n);
216 ir_node *b = get_Mul_right(n);
219 tarval *ta = value_of(a);
220 tarval *tb = value_of(b);
222 mode = get_irn_mode(n);
223 if (mode != get_irn_mode(a)) {
224 /* n * n = 2n bit multiplication */
225 ta = tarval_convert_to(ta, mode);
226 tb = tarval_convert_to(tb, mode);
229 if (ta != tarval_bad && tb != tarval_bad) {
230 return tarval_mul(ta, tb);
232 /* a * 0 != 0 if a == NaN or a == Inf */
233 if (!mode_is_float(mode)) {
234 /* a*0 = 0 or 0*b = 0 */
235 if (ta == get_mode_null(mode))
237 if (tb == get_mode_null(mode))
242 } /* computed_value_Mul */
245 * Return the value of an Abs.
247 static tarval *computed_value_Abs(const ir_node *n) {
248 ir_node *a = get_Abs_op(n);
249 tarval *ta = value_of(a);
251 if (ta != tarval_bad)
252 return tarval_abs(ta);
255 } /* computed_value_Abs */
258 * Return the value of an And.
259 * Special case: a & 0, 0 & b
261 static tarval *computed_value_And(const ir_node *n) {
262 ir_node *a = get_And_left(n);
263 ir_node *b = get_And_right(n);
265 tarval *ta = value_of(a);
266 tarval *tb = value_of(b);
268 if ((ta != tarval_bad) && (tb != tarval_bad)) {
269 return tarval_and (ta, tb);
271 if (tarval_is_null(ta)) return ta;
272 if (tarval_is_null(tb)) return tb;
275 } /* computed_value_And */
278 * Return the value of an Or.
279 * Special case: a | 1...1, 1...1 | b
281 static tarval *computed_value_Or(const ir_node *n) {
282 ir_node *a = get_Or_left(n);
283 ir_node *b = get_Or_right(n);
285 tarval *ta = value_of(a);
286 tarval *tb = value_of(b);
288 if ((ta != tarval_bad) && (tb != tarval_bad)) {
289 return tarval_or (ta, tb);
291 if (tarval_is_all_one(ta)) return ta;
292 if (tarval_is_all_one(tb)) return tb;
295 } /* computed_value_Or */
298 * Return the value of an Eor.
300 static tarval *computed_value_Eor(const ir_node *n) {
301 ir_node *a = get_Eor_left(n);
302 ir_node *b = get_Eor_right(n);
307 return get_mode_null(get_irn_mode(n));
312 if ((ta != tarval_bad) && (tb != tarval_bad)) {
313 return tarval_eor(ta, tb);
316 } /* computed_value_Eor */
319 * Return the value of a Not.
321 static tarval *computed_value_Not(const ir_node *n) {
322 ir_node *a = get_Not_op(n);
323 tarval *ta = value_of(a);
325 if (ta != tarval_bad)
326 return tarval_not(ta);
329 } /* computed_value_Not */
332 * Return the value of a Shl.
334 static tarval *computed_value_Shl(const ir_node *n) {
335 ir_node *a = get_Shl_left(n);
336 ir_node *b = get_Shl_right(n);
338 tarval *ta = value_of(a);
339 tarval *tb = value_of(b);
341 if ((ta != tarval_bad) && (tb != tarval_bad)) {
342 return tarval_shl(ta, tb);
345 } /* computed_value_Shl */
348 * Return the value of a Shr.
350 static tarval *computed_value_Shr(const ir_node *n) {
351 ir_node *a = get_Shr_left(n);
352 ir_node *b = get_Shr_right(n);
354 tarval *ta = value_of(a);
355 tarval *tb = value_of(b);
357 if ((ta != tarval_bad) && (tb != tarval_bad)) {
358 return tarval_shr(ta, tb);
361 } /* computed_value_Shr */
364 * Return the value of a Shrs.
366 static tarval *computed_value_Shrs(const ir_node *n) {
367 ir_node *a = get_Shrs_left(n);
368 ir_node *b = get_Shrs_right(n);
370 tarval *ta = value_of(a);
371 tarval *tb = value_of(b);
373 if ((ta != tarval_bad) && (tb != tarval_bad)) {
374 return tarval_shrs(ta, tb);
377 } /* computed_value_Shrs */
380 * Return the value of a Rotl.
382 static tarval *computed_value_Rotl(const ir_node *n) {
383 ir_node *a = get_Rotl_left(n);
384 ir_node *b = get_Rotl_right(n);
386 tarval *ta = value_of(a);
387 tarval *tb = value_of(b);
389 if ((ta != tarval_bad) && (tb != tarval_bad)) {
390 return tarval_rotl(ta, tb);
393 } /* computed_value_Rotl */
396 * Return the value of a Conv.
398 static tarval *computed_value_Conv(const ir_node *n) {
399 ir_node *a = get_Conv_op(n);
400 tarval *ta = value_of(a);
402 if (ta != tarval_bad)
403 return tarval_convert_to(ta, get_irn_mode(n));
406 } /* computed_value_Conv */
409 * Calculate the value of a Mux: can be evaluated, if the
410 * sel and the right input are known.
412 static tarval *computed_value_Mux(const ir_node *n) {
413 ir_node *sel = get_Mux_sel(n);
414 tarval *ts = value_of(sel);
416 if (ts == get_tarval_b_true()) {
417 ir_node *v = get_Mux_true(n);
420 else if (ts == get_tarval_b_false()) {
421 ir_node *v = get_Mux_false(n);
425 } /* computed_value_Mux */
428 * Calculate the value of a Confirm: can be evaluated,
429 * if it has the form Confirm(x, '=', Const).
431 static tarval *computed_value_Confirm(const ir_node *n) {
433 * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)).
434 * Do NOT optimize them away (jump threading wants them), so wait until
435 * remove_confirm is activated.
437 if (get_opt_remove_confirm()) {
438 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
439 tarval *tv = value_of(get_Confirm_bound(n));
440 if (tv != tarval_bad)
444 return value_of(get_Confirm_value(n));
445 } /* computed_value_Confirm */
448 * Return the value of a Proj(Cmp).
450 * This performs a first step of unreachable code elimination.
451 * Proj can not be computed, but folding a Cmp above the Proj here is
452 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
454 * There are several case where we can evaluate a Cmp node, see later.
456 static tarval *computed_value_Proj_Cmp(const ir_node *n) {
457 ir_node *a = get_Proj_pred(n);
458 ir_node *aa = get_Cmp_left(a);
459 ir_node *ab = get_Cmp_right(a);
460 long proj_nr = get_Proj_proj(n);
463 * BEWARE: a == a is NOT always True for floating Point values, as
464 * NaN != NaN is defined, so we must check this here.
467 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
470 /* This is a trick with the bits used for encoding the Cmp
471 Proj numbers, the following statement is not the same:
472 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
473 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
476 tarval *taa = value_of(aa);
477 tarval *tab = value_of(ab);
478 ir_mode *mode = get_irn_mode(aa);
481 * The predecessors of Cmp are target values. We can evaluate
484 if ((taa != tarval_bad) && (tab != tarval_bad)) {
485 /* strange checks... */
486 pn_Cmp flags = tarval_cmp(taa, tab);
487 if (flags != pn_Cmp_False) {
488 return new_tarval_from_long (proj_nr & flags, mode_b);
491 /* for integer values, we can check against MIN/MAX */
492 else if (mode_is_int(mode)) {
493 /* MIN <=/> x. This results in true/false. */
494 if (taa == get_mode_min(mode)) {
495 /* a compare with the MIN value */
496 if (proj_nr == pn_Cmp_Le)
497 return get_tarval_b_true();
498 else if (proj_nr == pn_Cmp_Gt)
499 return get_tarval_b_false();
501 /* x >=/< MIN. This results in true/false. */
503 if (tab == get_mode_min(mode)) {
504 /* a compare with the MIN value */
505 if (proj_nr == pn_Cmp_Ge)
506 return get_tarval_b_true();
507 else if (proj_nr == pn_Cmp_Lt)
508 return get_tarval_b_false();
510 /* MAX >=/< x. This results in true/false. */
511 else if (taa == get_mode_max(mode)) {
512 if (proj_nr == pn_Cmp_Ge)
513 return get_tarval_b_true();
514 else if (proj_nr == pn_Cmp_Lt)
515 return get_tarval_b_false();
517 /* x <=/> MAX. This results in true/false. */
518 else if (tab == get_mode_max(mode)) {
519 if (proj_nr == pn_Cmp_Le)
520 return get_tarval_b_true();
521 else if (proj_nr == pn_Cmp_Gt)
522 return get_tarval_b_false();
526 * The predecessors are Allocs or (void*)(0) constants. Allocs never
527 * return NULL, they raise an exception. Therefore we can predict
531 ir_node *aaa = skip_Proj(aa);
532 ir_node *aba = skip_Proj(ab);
534 if ( ( (/* aa is ProjP and aaa is Alloc */
536 && mode_is_reference(get_irn_mode(aa))
538 && ( (/* ab is NULL */
539 mode_is_reference(get_irn_mode(ab))
540 && tarval_is_null(tab))
541 || (/* ab is other Alloc */
543 && mode_is_reference(get_irn_mode(ab))
546 || (/* aa is NULL and aba is Alloc */
547 mode_is_reference(get_irn_mode(aa))
548 && tarval_is_null(taa)
550 && mode_is_reference(get_irn_mode(ab))
553 return new_tarval_from_long(proj_nr & pn_Cmp_Lg, mode_b);
556 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
557 } /* computed_value_Proj_Cmp */
560 * Return the value of a floating point Quot.
562 static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) {
563 tarval *ta = value_of(a);
564 tarval *tb = value_of(b);
566 /* cannot optimize 0 / b = 0 because of NaN */
567 if (ta != tarval_bad && tb != tarval_bad)
568 return tarval_quo(ta, tb);
570 } /* do_computed_value_Quot */
573 * Calculate the value of an integer Div of two nodes.
574 * Special case: 0 / b
576 static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) {
577 tarval *ta = value_of(a);
579 const ir_node *dummy;
581 /* Compute c1 / c2 or 0 / a, a != 0 */
582 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
583 return ta; /* 0 / b == 0 */
585 if (ta != tarval_bad && tb != tarval_bad)
586 return tarval_div(ta, tb);
588 } /* do_computed_value_Div */
591 * Calculate the value of an integer Mod of two nodes.
592 * Special case: a % 1
594 static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) {
595 tarval *ta = value_of(a);
596 tarval *tb = value_of(b);
598 /* Compute a % 1 or c1 % c2 */
599 if (tarval_is_one(tb))
600 return get_mode_null(get_irn_mode(a));
601 if (ta != tarval_bad && tb != tarval_bad)
602 return tarval_mod(ta, tb);
604 } /* do_computed_value_Mod */
607 * Return the value of a Proj(DivMod).
609 static tarval *computed_value_Proj_DivMod(const ir_node *n) {
610 long proj_nr = get_Proj_proj(n);
612 /* compute either the Div or the Mod part */
613 if (proj_nr == pn_DivMod_res_div) {
614 const ir_node *a = get_Proj_pred(n);
615 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
616 } else if (proj_nr == pn_DivMod_res_mod) {
617 const ir_node *a = get_Proj_pred(n);
618 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
621 } /* computed_value_Proj_DivMod */
624 * Return the value of a Proj(Div).
626 static tarval *computed_value_Proj_Div(const ir_node *n) {
627 long proj_nr = get_Proj_proj(n);
629 if (proj_nr == pn_Div_res) {
630 const ir_node *a = get_Proj_pred(n);
631 return do_computed_value_Div(get_Div_left(a), get_Div_right(a));
634 } /* computed_value_Proj_Div */
637 * Return the value of a Proj(Mod).
639 static tarval *computed_value_Proj_Mod(const ir_node *n) {
640 long proj_nr = get_Proj_proj(n);
642 if (proj_nr == pn_Mod_res) {
643 const ir_node *a = get_Proj_pred(n);
644 return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a));
647 } /* computed_value_Proj_Mod */
650 * Return the value of a Proj(Quot).
652 static tarval *computed_value_Proj_Quot(const ir_node *n) {
653 long proj_nr = get_Proj_proj(n);
655 if (proj_nr == pn_Quot_res) {
656 const ir_node *a = get_Proj_pred(n);
657 return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a));
660 } /* computed_value_Proj_Quot */
663 * Return the value of a Proj.
665 static tarval *computed_value_Proj(const ir_node *proj) {
666 ir_node *n = get_Proj_pred(proj);
668 if (n->op->ops.computed_value_Proj != NULL)
669 return n->op->ops.computed_value_Proj(proj);
671 } /* computed_value_Proj */
674 * If the parameter n can be computed, return its value, else tarval_bad.
675 * Performs constant folding.
677 * @param n The node this should be evaluated
679 tarval *computed_value(const ir_node *n) {
680 if(mode_is_int(get_irn_mode(n)) && tarval_is_all_one(
681 tarval_or(n->vrp.bits_set, n->vrp.bits_not_set))) {
682 return n->vrp.bits_set;
684 if (n->op->ops.computed_value)
685 return n->op->ops.computed_value(n);
687 } /* computed_value */
690 * Set the default computed_value evaluator in an ir_op_ops.
692 * @param code the opcode for the default operation
693 * @param ops the operations initialized
698 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
702 ops->computed_value = computed_value_##a; \
704 #define CASE_PROJ(a) \
706 ops->computed_value_Proj = computed_value_Proj_##a; \
743 } /* firm_set_default_computed_value */
746 * Returns a equivalent block for another block.
747 * If the block has only one predecessor, this is
748 * the equivalent one. If the only predecessor of a block is
749 * the block itself, this is a dead block.
751 * If both predecessors of a block are the branches of a binary
752 * Cond, the equivalent block is Cond's block.
754 * If all predecessors of a block are bad or lies in a dead
755 * block, the current block is dead as well.
757 * Note, that blocks are NEVER turned into Bad's, instead
758 * the dead_block flag is set. So, never test for is_Bad(block),
759 * always use is_dead_Block(block).
761 static ir_node *equivalent_node_Block(ir_node *n)
766 /* don't optimize dead or labeled blocks */
767 if (is_Block_dead(n) || has_Block_entity(n))
770 n_preds = get_Block_n_cfgpreds(n);
772 /* The Block constructor does not call optimize, but mature_immBlock()
773 calls the optimization. */
774 assert(get_Block_matured(n));
776 /* Straightening: a single entry Block following a single exit Block
777 can be merged, if it is not the Start block. */
778 /* !!! Beware, all Phi-nodes of n must have been optimized away.
779 This should be true, as the block is matured before optimize is called.
780 But what about Phi-cycles with the Phi0/Id that could not be resolved?
781 Remaining Phi nodes are just Ids. */
783 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
786 ir_node *predblock = get_nodes_block(pred);
787 if (predblock == oldn) {
788 /* Jmp jumps into the block it is in -- deal self cycle. */
789 n = set_Block_dead(n);
790 DBG_OPT_DEAD_BLOCK(oldn, n);
791 } else if (get_opt_control_flow_straightening()) {
793 DBG_OPT_STG(oldn, n);
795 } else if (is_Cond(pred)) {
796 ir_node *predblock = get_nodes_block(pred);
797 if (predblock == oldn) {
798 /* Jmp jumps into the block it is in -- deal self cycle. */
799 n = set_Block_dead(n);
800 DBG_OPT_DEAD_BLOCK(oldn, n);
803 } else if ((n_preds == 2) &&
804 (get_opt_control_flow_weak_simplification())) {
805 /* Test whether Cond jumps twice to this block
806 * The more general case which more than 2 predecessors is handles
807 * in optimize_cf(), we handle only this special case for speed here.
809 ir_node *a = get_Block_cfgpred(n, 0);
810 ir_node *b = get_Block_cfgpred(n, 1);
812 if (is_Proj(a) && is_Proj(b)) {
813 ir_node *cond = get_Proj_pred(a);
815 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
816 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
817 /* Also a single entry Block following a single exit Block. Phis have
818 twice the same operand and will be optimized away. */
819 n = get_nodes_block(cond);
820 DBG_OPT_IFSIM1(oldn, a, b, n);
823 } else if (get_opt_unreachable_code() &&
824 (n != get_irg_start_block(current_ir_graph)) &&
825 (n != get_irg_end_block(current_ir_graph)) ) {
828 /* If all inputs are dead, this block is dead too, except if it is
829 the start or end block. This is one step of unreachable code
831 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
832 ir_node *pred = get_Block_cfgpred(n, i);
835 if (is_Bad(pred)) continue;
836 pred_blk = get_nodes_block(skip_Proj(pred));
838 if (is_Block_dead(pred_blk)) continue;
841 /* really found a living input */
846 n = set_Block_dead(n);
847 DBG_OPT_DEAD_BLOCK(oldn, n);
852 } /* equivalent_node_Block */
855 * Returns a equivalent node for a Jmp, a Bad :-)
856 * Of course this only happens if the Block of the Jmp is dead.
858 static ir_node *equivalent_node_Jmp(ir_node *n) {
861 /* unreachable code elimination */
862 if (is_Block_dead(get_nodes_block(n))) {
863 n = get_irg_bad(current_ir_graph);
864 DBG_OPT_DEAD_BLOCK(oldn, n);
867 } /* equivalent_node_Jmp */
869 /** Raise is handled in the same way as Jmp. */
870 #define equivalent_node_Raise equivalent_node_Jmp
873 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
874 See transform_node_Proj_Cond(). */
877 * Optimize operations that are commutative and have neutral 0,
878 * so a op 0 = 0 op a = a.
880 static ir_node *equivalent_node_neutral_zero(ir_node *n) {
883 ir_node *a = get_binop_left(n);
884 ir_node *b = get_binop_right(n);
889 /* After running compute_node there is only one constant predecessor.
890 Find this predecessors value and remember the other node: */
891 if ((tv = value_of(a)) != tarval_bad) {
893 } else if ((tv = value_of(b)) != tarval_bad) {
898 /* If this predecessors constant value is zero, the operation is
899 * unnecessary. Remove it.
901 * Beware: If n is a Add, the mode of on and n might be different
902 * which happens in this rare construction: NULL + 3.
903 * Then, a Conv would be needed which we cannot include here.
905 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
908 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
912 } /* equivalent_node_neutral_zero */
915 * Eor is commutative and has neutral 0.
917 static ir_node *equivalent_node_Eor(ir_node *n) {
922 n = equivalent_node_neutral_zero(n);
923 if (n != oldn) return n;
926 b = get_Eor_right(n);
929 ir_node *aa = get_Eor_left(a);
930 ir_node *ab = get_Eor_right(a);
933 /* (a ^ b) ^ a -> b */
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
937 } else if (ab == b) {
938 /* (a ^ b) ^ b -> a */
940 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
945 ir_node *ba = get_Eor_left(b);
946 ir_node *bb = get_Eor_right(b);
949 /* a ^ (a ^ b) -> b */
951 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
953 } else if (bb == a) {
954 /* a ^ (b ^ a) -> b */
956 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
964 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
966 * The second one looks strange, but this construct
967 * is used heavily in the LCC sources :-).
969 * Beware: The Mode of an Add may be different than the mode of its
970 * predecessors, so we could not return a predecessors in all cases.
972 static ir_node *equivalent_node_Add(ir_node *n) {
974 ir_node *left, *right;
975 ir_mode *mode = get_irn_mode(n);
977 n = equivalent_node_neutral_zero(n);
981 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
982 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
985 left = get_Add_left(n);
986 right = get_Add_right(n);
989 if (get_Sub_right(left) == right) {
992 n = get_Sub_left(left);
993 if (mode == get_irn_mode(n)) {
994 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1000 if (get_Sub_right(right) == left) {
1003 n = get_Sub_left(right);
1004 if (mode == get_irn_mode(n)) {
1005 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
1011 } /* equivalent_node_Add */
1014 * optimize operations that are not commutative but have neutral 0 on left,
1017 static ir_node *equivalent_node_left_zero(ir_node *n) {
1020 ir_node *a = get_binop_left(n);
1021 ir_node *b = get_binop_right(n);
1022 tarval *tb = value_of(b);
1024 if (tarval_is_null(tb)) {
1027 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1030 } /* equivalent_node_left_zero */
1032 #define equivalent_node_Shl equivalent_node_left_zero
1033 #define equivalent_node_Shr equivalent_node_left_zero
1034 #define equivalent_node_Shrs equivalent_node_left_zero
1035 #define equivalent_node_Rotl equivalent_node_left_zero
1038 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1040 * The second one looks strange, but this construct
1041 * is used heavily in the LCC sources :-).
1043 * Beware: The Mode of a Sub may be different than the mode of its
1044 * predecessors, so we could not return a predecessors in all cases.
1046 static ir_node *equivalent_node_Sub(ir_node *n) {
1049 ir_mode *mode = get_irn_mode(n);
1052 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1053 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1056 b = get_Sub_right(n);
1059 /* Beware: modes might be different */
1060 if (tarval_is_null(tb)) {
1061 ir_node *a = get_Sub_left(n);
1062 if (mode == get_irn_mode(a)) {
1065 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1069 } /* equivalent_node_Sub */
1073 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1076 * -(-a) == a, but might overflow two times.
1077 * We handle it anyway here but the better way would be a
1078 * flag. This would be needed for Pascal for instance.
1080 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1082 ir_node *pred = get_unop_op(n);
1084 /* optimize symmetric unop */
1085 if (get_irn_op(pred) == get_irn_op(n)) {
1086 n = get_unop_op(pred);
1087 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1090 } /* equivalent_node_idempotent_unop */
1092 /** Optimize Not(Not(x)) == x. */
1093 #define equivalent_node_Not equivalent_node_idempotent_unop
1095 /** -(-x) == x ??? Is this possible or can --x raise an
1096 out of bounds exception if min =! max? */
1097 #define equivalent_node_Minus equivalent_node_idempotent_unop
1100 * Optimize a * 1 = 1 * a = a.
1102 static ir_node *equivalent_node_Mul(ir_node *n) {
1104 ir_node *a = get_Mul_left(n);
1106 /* we can handle here only the n * n = n bit cases */
1107 if (get_irn_mode(n) == get_irn_mode(a)) {
1108 ir_node *b = get_Mul_right(n);
1112 * Mul is commutative and has again an other neutral element.
1113 * Constants are place right, so check this case first.
1116 if (tarval_is_one(tv)) {
1118 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1121 if (tarval_is_one(tv)) {
1123 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1128 } /* equivalent_node_Mul */
1131 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1133 static ir_node *equivalent_node_Or(ir_node *n) {
1136 ir_node *a = get_Or_left(n);
1137 ir_node *b = get_Or_right(n);
1141 n = a; /* Or has it's own neutral element */
1142 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1145 /* constants are cormalized to right, check this site first */
1147 if (tarval_is_null(tv)) {
1149 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1153 if (tarval_is_null(tv)) {
1155 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1160 } /* equivalent_node_Or */
1163 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1165 static ir_node *equivalent_node_And(ir_node *n) {
1168 ir_node *a = get_And_left(n);
1169 ir_node *b = get_And_right(n);
1173 n = a; /* And has it's own neutral element */
1174 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1177 /* constants are normalized to right, check this site first */
1179 if (tarval_is_all_one(tv)) {
1181 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1184 if (tv != get_tarval_bad()) {
1185 ir_mode *mode = get_irn_mode(n);
1186 if (!mode_is_signed(mode) && is_Conv(a)) {
1187 ir_node *convop = get_Conv_op(a);
1188 ir_mode *convopmode = get_irn_mode(convop);
1189 if (!mode_is_signed(convopmode)) {
1190 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1191 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1193 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1200 if (tarval_is_all_one(tv)) {
1202 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1206 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1209 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1214 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1217 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1222 } /* equivalent_node_And */
1225 * Try to remove useless Conv's:
1227 static ir_node *equivalent_node_Conv(ir_node *n) {
1229 ir_node *a = get_Conv_op(n);
1231 ir_mode *n_mode = get_irn_mode(n);
1232 ir_mode *a_mode = get_irn_mode(a);
1235 if (n_mode == a_mode) { /* No Conv necessary */
1236 if (get_Conv_strict(n)) {
1239 /* neither Minus nor Abs nor Confirm change the precision,
1240 so we can "look-through" */
1243 p = get_Minus_op(p);
1244 } else if (is_Abs(p)) {
1246 } else if (is_Confirm(p)) {
1247 p = get_Confirm_value(p);
1253 if (is_Conv(p) && get_Conv_strict(p)) {
1254 /* we known already, that a_mode == n_mode, and neither
1255 Abs nor Minus change the mode, so the second Conv
1257 assert(get_irn_mode(p) == n_mode);
1259 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1263 ir_node *pred = get_Proj_pred(p);
1264 if (is_Load(pred)) {
1265 /* Loads always return with the exact precision of n_mode */
1266 assert(get_Load_mode(pred) == n_mode);
1268 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1271 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1272 pred = get_Proj_pred(pred);
1273 if (is_Start(pred)) {
1274 /* Arguments always return with the exact precision,
1275 as strictConv's are place before Call -- if the
1276 caller was compiled with the same setting.
1277 Otherwise, the semantics is probably still right. */
1278 assert(get_irn_mode(p) == n_mode);
1280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1286 /* special case: the immediate predecessor is also a Conv */
1287 if (! get_Conv_strict(a)) {
1288 /* first one is not strict, kick it */
1290 a_mode = get_irn_mode(a);
1294 /* else both are strict conv, second is superfluous */
1296 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1301 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1304 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1305 ir_node *b = get_Conv_op(a);
1306 ir_mode *b_mode = get_irn_mode(b);
1308 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1309 /* both are strict conv */
1310 if (smaller_mode(a_mode, n_mode)) {
1311 /* both are strict, but the first is smaller, so
1312 the second cannot remove more precision, remove the
1314 set_Conv_strict(n, 0);
1317 if (n_mode == b_mode) {
1318 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1319 if (n_mode == mode_b) {
1320 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1321 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1323 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1324 if (values_in_mode(b_mode, a_mode)) {
1325 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1326 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1331 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1332 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1333 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1334 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1336 if (float_mantissa >= int_mantissa) {
1338 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1343 if (smaller_mode(b_mode, a_mode)) {
1344 if (get_Conv_strict(n))
1345 set_Conv_strict(b, 1);
1346 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1347 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1354 } /* equivalent_node_Conv */
1357 * A Cast may be removed if the type of the previous node
1358 * is already the type of the Cast.
1360 static ir_node *equivalent_node_Cast(ir_node *n) {
1362 ir_node *pred = get_Cast_op(n);
1364 if (get_irn_type(pred) == get_Cast_type(n)) {
1366 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1369 } /* equivalent_node_Cast */
1372 * - fold Phi-nodes, iff they have only one predecessor except
1375 static ir_node *equivalent_node_Phi(ir_node *n) {
1380 ir_node *first_val = NULL; /* to shutup gcc */
1382 if (!get_opt_normalize()) return n;
1384 n_preds = get_Phi_n_preds(n);
1386 block = get_nodes_block(n);
1387 if (is_Block_dead(block)) /* Control dead */
1388 return get_irg_bad(current_ir_graph);
1390 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1392 /* Find first non-self-referencing input */
1393 for (i = 0; i < n_preds; ++i) {
1394 first_val = get_Phi_pred(n, i);
1395 if ( (first_val != n) /* not self pointer */
1397 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1398 * predecessors. Then, Phi nodes in dead code might be removed, causing
1399 * nodes pointing to themself (Add's for instance).
1400 * This is really bad and causes endless recursions in several
1401 * code pathes, so we do NOT optimize such a code.
1402 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1403 * (and bad Phi predecessors), so live code is optimized later.
1405 && (! is_Bad(get_Block_cfgpred(block, i)))
1407 ) { /* value not dead */
1408 break; /* then found first value. */
1413 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1414 return get_irg_bad(current_ir_graph);
1417 /* search for rest of inputs, determine if any of these
1418 are non-self-referencing */
1419 while (++i < n_preds) {
1420 ir_node *scnd_val = get_Phi_pred(n, i);
1421 if ( (scnd_val != n)
1422 && (scnd_val != first_val)
1425 && (! is_Bad(get_Block_cfgpred(block, i)))
1433 /* Fold, if no multiple distinct non-self-referencing inputs */
1435 DBG_OPT_PHI(oldn, n);
1438 } /* equivalent_node_Phi */
1441 * Several optimizations:
1442 * - fold Sync-nodes, iff they have only one predecessor except
1445 static ir_node *equivalent_node_Sync(ir_node *n) {
1446 int arity = get_Sync_n_preds(n);
1449 for (i = 0; i < arity;) {
1450 ir_node *pred = get_Sync_pred(n, i);
1453 /* Remove Bad predecessors */
1460 /* Remove duplicate predecessors */
1466 if (get_Sync_pred(n, j) == pred) {
1474 if (arity == 0) return get_irg_bad(current_ir_graph);
1475 if (arity == 1) return get_Sync_pred(n, 0);
1477 } /* equivalent_node_Sync */
1480 * Optimize Proj(Tuple).
1482 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) {
1483 ir_node *oldn = proj;
1484 ir_node *tuple = get_Proj_pred(proj);
1486 /* Remove the Tuple/Proj combination. */
1487 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1488 DBG_OPT_TUPLE(oldn, tuple, proj);
1491 } /* equivalent_node_Proj_Tuple */
1494 * Optimize a / 1 = a.
1496 static ir_node *equivalent_node_Proj_Div(ir_node *proj) {
1497 ir_node *oldn = proj;
1498 ir_node *div = get_Proj_pred(proj);
1499 ir_node *b = get_Div_right(div);
1500 tarval *tb = value_of(b);
1502 /* Div is not commutative. */
1503 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1504 switch (get_Proj_proj(proj)) {
1506 proj = get_Div_mem(div);
1507 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1511 proj = get_Div_left(div);
1512 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1516 /* we cannot replace the exception Proj's here, this is done in
1517 transform_node_Proj_Div() */
1522 } /* equivalent_node_Proj_Div */
1525 * Optimize a / 1.0 = a.
1527 static ir_node *equivalent_node_Proj_Quot(ir_node *proj) {
1528 ir_node *oldn = proj;
1529 ir_node *quot = get_Proj_pred(proj);
1530 ir_node *b = get_Quot_right(quot);
1531 tarval *tb = value_of(b);
1533 /* Div is not commutative. */
1534 if (tarval_is_one(tb)) { /* Quot(x, 1) == x */
1535 switch (get_Proj_proj(proj)) {
1537 proj = get_Quot_mem(quot);
1538 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1542 proj = get_Quot_left(quot);
1543 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1547 /* we cannot replace the exception Proj's here, this is done in
1548 transform_node_Proj_Quot() */
1553 } /* equivalent_node_Proj_Quot */
1556 * Optimize a / 1 = a.
1558 static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) {
1559 ir_node *oldn = proj;
1560 ir_node *divmod = get_Proj_pred(proj);
1561 ir_node *b = get_DivMod_right(divmod);
1562 tarval *tb = value_of(b);
1564 /* Div is not commutative. */
1565 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1566 switch (get_Proj_proj(proj)) {
1568 proj = get_DivMod_mem(divmod);
1569 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1572 case pn_DivMod_res_div:
1573 proj = get_DivMod_left(divmod);
1574 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1578 /* we cannot replace the exception Proj's here, this is done in
1579 transform_node_Proj_DivMod().
1580 Note further that the pn_DivMod_res_div case is handled in
1581 computed_value_Proj(). */
1586 } /* equivalent_node_Proj_DivMod */
1589 * Optimize CopyB(mem, x, x) into a Nop.
1591 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) {
1592 ir_node *oldn = proj;
1593 ir_node *copyb = get_Proj_pred(proj);
1594 ir_node *a = get_CopyB_dst(copyb);
1595 ir_node *b = get_CopyB_src(copyb);
1598 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1599 switch (get_Proj_proj(proj)) {
1600 case pn_CopyB_M_regular:
1601 proj = get_CopyB_mem(copyb);
1602 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1605 case pn_CopyB_X_except:
1606 DBG_OPT_EXC_REM(proj);
1607 proj = get_irg_bad(current_ir_graph);
1612 } /* equivalent_node_Proj_CopyB */
1615 * Optimize Bounds(idx, idx, upper) into idx.
1617 static ir_node *equivalent_node_Proj_Bound(ir_node *proj) {
1618 ir_node *oldn = proj;
1619 ir_node *bound = get_Proj_pred(proj);
1620 ir_node *idx = get_Bound_index(bound);
1621 ir_node *pred = skip_Proj(idx);
1624 if (idx == get_Bound_lower(bound))
1626 else if (is_Bound(pred)) {
1628 * idx was Bounds checked in the same MacroBlock previously,
1629 * it is still valid if lower <= pred_lower && pred_upper <= upper.
1631 ir_node *lower = get_Bound_lower(bound);
1632 ir_node *upper = get_Bound_upper(bound);
1633 if (get_Bound_lower(pred) == lower &&
1634 get_Bound_upper(pred) == upper &&
1635 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
1637 * One could expect that we simply return the previous
1638 * Bound here. However, this would be wrong, as we could
1639 * add an exception Proj to a new location then.
1640 * So, we must turn in into a tuple.
1646 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1647 switch (get_Proj_proj(proj)) {
1649 DBG_OPT_EXC_REM(proj);
1650 proj = get_Bound_mem(bound);
1652 case pn_Bound_X_except:
1653 DBG_OPT_EXC_REM(proj);
1654 proj = get_irg_bad(current_ir_graph);
1658 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1661 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1666 } /* equivalent_node_Proj_Bound */
1669 * Optimize an Exception Proj(Load) with a non-null address.
1671 static ir_node *equivalent_node_Proj_Load(ir_node *proj) {
1672 if (get_opt_ldst_only_null_ptr_exceptions()) {
1673 if (get_irn_mode(proj) == mode_X) {
1674 ir_node *load = get_Proj_pred(proj);
1676 /* get the Load address */
1677 const ir_node *addr = get_Load_ptr(load);
1678 const ir_node *confirm;
1680 if (value_not_null(addr, &confirm)) {
1681 if (get_Proj_proj(proj) == pn_Load_X_except) {
1682 DBG_OPT_EXC_REM(proj);
1683 return get_irg_bad(current_ir_graph);
1689 } /* equivalent_node_Proj_Load */
1692 * Optimize an Exception Proj(Store) with a non-null address.
1694 static ir_node *equivalent_node_Proj_Store(ir_node *proj) {
1695 if (get_opt_ldst_only_null_ptr_exceptions()) {
1696 if (get_irn_mode(proj) == mode_X) {
1697 ir_node *store = get_Proj_pred(proj);
1699 /* get the load/store address */
1700 const ir_node *addr = get_Store_ptr(store);
1701 const ir_node *confirm;
1703 if (value_not_null(addr, &confirm)) {
1704 if (get_Proj_proj(proj) == pn_Store_X_except) {
1705 DBG_OPT_EXC_REM(proj);
1706 return get_irg_bad(current_ir_graph);
1712 } /* equivalent_node_Proj_Store */
1715 * Does all optimizations on nodes that must be done on it's Proj's
1716 * because of creating new nodes.
1718 static ir_node *equivalent_node_Proj(ir_node *proj) {
1719 ir_node *n = get_Proj_pred(proj);
1721 if (get_irn_mode(proj) == mode_X) {
1722 if (is_Block_dead(get_nodes_block(n))) {
1723 /* Remove dead control flow -- early gigo(). */
1724 return get_irg_bad(current_ir_graph);
1727 if (n->op->ops.equivalent_node_Proj)
1728 return n->op->ops.equivalent_node_Proj(proj);
1730 } /* equivalent_node_Proj */
1735 static ir_node *equivalent_node_Id(ir_node *n) {
1742 DBG_OPT_ID(oldn, n);
1744 } /* equivalent_node_Id */
1749 static ir_node *equivalent_node_Mux(ir_node *n)
1751 ir_node *oldn = n, *sel = get_Mux_sel(n);
1753 tarval *ts = value_of(sel);
1755 /* Mux(true, f, t) == t */
1756 if (ts == tarval_b_true) {
1757 n = get_Mux_true(n);
1758 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1761 /* Mux(false, f, t) == f */
1762 if (ts == tarval_b_false) {
1763 n = get_Mux_false(n);
1764 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1767 n_t = get_Mux_true(n);
1768 n_f = get_Mux_false(n);
1770 /* Mux(v, x, T) == x */
1771 if (is_Unknown(n_f)) {
1773 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1776 /* Mux(v, T, x) == x */
1777 if (is_Unknown(n_t)) {
1779 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1783 /* Mux(v, x, x) == x */
1786 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1789 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1790 ir_node *cmp = get_Proj_pred(sel);
1791 long proj_nr = get_Proj_proj(sel);
1792 ir_node *f = get_Mux_false(n);
1793 ir_node *t = get_Mux_true(n);
1796 * Note further that these optimization work even for floating point
1797 * with NaN's because -NaN == NaN.
1798 * However, if +0 and -0 is handled differently, we cannot use the first one.
1801 ir_node *const cmp_l = get_Cmp_left(cmp);
1802 ir_node *const cmp_r = get_Cmp_right(cmp);
1806 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1807 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1809 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1816 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1817 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1819 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1826 * Note: normalization puts the constant on the right side,
1827 * so we check only one case.
1829 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1830 /* Mux(t CMP 0, X, t) */
1831 if (is_Minus(f) && get_Minus_op(f) == t) {
1832 /* Mux(t CMP 0, -t, t) */
1833 if (proj_nr == pn_Cmp_Eq) {
1834 /* Mux(t == 0, -t, t) ==> -t */
1836 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1837 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1838 /* Mux(t != 0, -t, t) ==> t */
1840 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1847 } /* equivalent_node_Mux */
1850 * Remove Confirm nodes if setting is on.
1851 * Replace Confirms(x, '=', Constlike) by Constlike.
1853 static ir_node *equivalent_node_Confirm(ir_node *n) {
1854 ir_node *pred = get_Confirm_value(n);
1855 pn_Cmp pnc = get_Confirm_cmp(n);
1857 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1859 * rare case: two identical Confirms one after another,
1860 * replace the second one with the first.
1863 pred = get_Confirm_value(n);
1864 pnc = get_Confirm_cmp(n);
1866 if (get_opt_remove_confirm())
1867 return get_Confirm_value(n);
1872 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1873 * perform no actual computation, as, e.g., the Id nodes. It does not create
1874 * new nodes. It is therefore safe to free n if the node returned is not n.
1875 * If a node returns a Tuple we can not just skip it. If the size of the
1876 * in array fits, we transform n into a tuple (e.g., Div).
1878 ir_node *equivalent_node(ir_node *n) {
1879 if (n->op->ops.equivalent_node)
1880 return n->op->ops.equivalent_node(n);
1882 } /* equivalent_node */
1885 * Sets the default equivalent node operation for an ir_op_ops.
1887 * @param code the opcode for the default operation
1888 * @param ops the operations initialized
1893 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1897 ops->equivalent_node = equivalent_node_##a; \
1899 #define CASE_PROJ(a) \
1901 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1943 } /* firm_set_default_equivalent_node */
1946 * Returns non-zero if a node is a Phi node
1947 * with all predecessors constant.
1949 static int is_const_Phi(ir_node *n) {
1952 if (! is_Phi(n) || get_irn_arity(n) == 0)
1954 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1955 if (! is_Const(get_irn_n(n, i)))
1959 } /* is_const_Phi */
1961 typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode);
1962 typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b);
1965 * in reality eval_func should be tarval (*eval_func)() but incomplete
1966 * declarations are bad style and generate noisy warnings
1968 typedef void (*eval_func)(void);
1971 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1973 static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode)
1975 if (eval == (eval_func) tarval_sub) {
1976 tarval_sub_type func = (tarval_sub_type)eval;
1978 return func(a, b, mode);
1980 tarval_binop_type func = (tarval_binop_type)eval;
1987 * Apply an evaluator on a binop with a constant operators (and one Phi).
1989 * @param phi the Phi node
1990 * @param other the other operand
1991 * @param eval an evaluator function
1992 * @param mode the mode of the result, may be different from the mode of the Phi!
1993 * @param left if non-zero, other is the left operand, else the right
1995 * @return a new Phi node if the conversion was successful, NULL else
1997 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, ir_mode *mode, int left) {
2002 int i, n = get_irn_arity(phi);
2004 NEW_ARR_A(void *, res, n);
2006 for (i = 0; i < n; ++i) {
2007 pred = get_irn_n(phi, i);
2008 tv = get_Const_tarval(pred);
2009 tv = do_eval(eval, other, tv, mode);
2011 if (tv == tarval_bad) {
2012 /* folding failed, bad */
2018 for (i = 0; i < n; ++i) {
2019 pred = get_irn_n(phi, i);
2020 tv = get_Const_tarval(pred);
2021 tv = do_eval(eval, tv, other, mode);
2023 if (tv == tarval_bad) {
2024 /* folding failed, bad */
2030 irg = current_ir_graph;
2031 for (i = 0; i < n; ++i) {
2032 pred = get_irn_n(phi, i);
2033 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2035 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2036 } /* apply_binop_on_phi */
2039 * Apply an evaluator on a binop with two constant Phi.
2041 * @param a the left Phi node
2042 * @param b the right Phi node
2043 * @param eval an evaluator function
2044 * @param mode the mode of the result, may be different from the mode of the Phi!
2046 * @return a new Phi node if the conversion was successful, NULL else
2048 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode) {
2049 tarval *tv_l, *tv_r, *tv;
2055 if (get_nodes_block(a) != get_nodes_block(b))
2058 n = get_irn_arity(a);
2059 NEW_ARR_A(void *, res, n);
2061 for (i = 0; i < n; ++i) {
2062 pred = get_irn_n(a, i);
2063 tv_l = get_Const_tarval(pred);
2064 pred = get_irn_n(b, i);
2065 tv_r = get_Const_tarval(pred);
2066 tv = do_eval(eval, tv_l, tv_r, mode);
2068 if (tv == tarval_bad) {
2069 /* folding failed, bad */
2074 irg = current_ir_graph;
2075 for (i = 0; i < n; ++i) {
2076 pred = get_irn_n(a, i);
2077 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2079 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
2080 } /* apply_binop_on_2_phis */
2083 * Apply an evaluator on a unop with a constant operator (a Phi).
2085 * @param phi the Phi node
2086 * @param eval an evaluator function
2088 * @return a new Phi node if the conversion was successful, NULL else
2090 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
2096 int i, n = get_irn_arity(phi);
2098 NEW_ARR_A(void *, res, n);
2099 for (i = 0; i < n; ++i) {
2100 pred = get_irn_n(phi, i);
2101 tv = get_Const_tarval(pred);
2104 if (tv == tarval_bad) {
2105 /* folding failed, bad */
2110 mode = get_irn_mode(phi);
2111 irg = current_ir_graph;
2112 for (i = 0; i < n; ++i) {
2113 pred = get_irn_n(phi, i);
2114 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2116 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2117 } /* apply_unop_on_phi */
2120 * Apply a conversion on a constant operator (a Phi).
2122 * @param phi the Phi node
2124 * @return a new Phi node if the conversion was successful, NULL else
2126 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) {
2131 int i, n = get_irn_arity(phi);
2133 NEW_ARR_A(void *, res, n);
2134 for (i = 0; i < n; ++i) {
2135 pred = get_irn_n(phi, i);
2136 tv = get_Const_tarval(pred);
2137 tv = tarval_convert_to(tv, mode);
2139 if (tv == tarval_bad) {
2140 /* folding failed, bad */
2145 irg = current_ir_graph;
2146 for (i = 0; i < n; ++i) {
2147 pred = get_irn_n(phi, i);
2148 res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred));
2150 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2151 } /* apply_conv_on_phi */
2154 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2155 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2156 * If possible, remove the Conv's.
2158 static ir_node *transform_node_AddSub(ir_node *n) {
2159 ir_mode *mode = get_irn_mode(n);
2161 if (mode_is_reference(mode)) {
2162 ir_node *left = get_binop_left(n);
2163 ir_node *right = get_binop_right(n);
2164 unsigned ref_bits = get_mode_size_bits(mode);
2166 if (is_Conv(left)) {
2167 ir_mode *lmode = get_irn_mode(left);
2168 unsigned bits = get_mode_size_bits(lmode);
2170 if (ref_bits == bits &&
2171 mode_is_int(lmode) &&
2172 get_mode_arithmetic(lmode) == irma_twos_complement) {
2173 ir_node *pre = get_Conv_op(left);
2174 ir_mode *pre_mode = get_irn_mode(pre);
2176 if (mode_is_int(pre_mode) &&
2177 get_mode_size_bits(pre_mode) == bits &&
2178 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2179 /* ok, this conv just changes to sign, moreover the calculation
2180 * is done with same number of bits as our address mode, so
2181 * we can ignore the conv as address calculation can be viewed
2182 * as either signed or unsigned
2184 set_binop_left(n, pre);
2189 if (is_Conv(right)) {
2190 ir_mode *rmode = get_irn_mode(right);
2191 unsigned bits = get_mode_size_bits(rmode);
2193 if (ref_bits == bits &&
2194 mode_is_int(rmode) &&
2195 get_mode_arithmetic(rmode) == irma_twos_complement) {
2196 ir_node *pre = get_Conv_op(right);
2197 ir_mode *pre_mode = get_irn_mode(pre);
2199 if (mode_is_int(pre_mode) &&
2200 get_mode_size_bits(pre_mode) == bits &&
2201 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2202 /* ok, this conv just changes to sign, moreover the calculation
2203 * is done with same number of bits as our address mode, so
2204 * we can ignore the conv as address calculation can be viewed
2205 * as either signed or unsigned
2207 set_binop_right(n, pre);
2212 /* let address arithmetic use unsigned modes */
2213 if (is_Const(right)) {
2214 ir_mode *rmode = get_irn_mode(right);
2216 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2217 /* convert a AddP(P, *s) into AddP(P, *u) */
2218 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2220 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2221 set_binop_right(n, pre);
2227 } /* transform_node_AddSub */
2229 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2231 if (is_Const(b) && is_const_Phi(a)) { \
2232 /* check for Op(Phi, Const) */ \
2233 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2235 else if (is_Const(a) && is_const_Phi(b)) { \
2236 /* check for Op(Const, Phi) */ \
2237 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2239 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2240 /* check for Op(Phi, Phi) */ \
2241 c = apply_binop_on_2_phis(a, b, eval, mode); \
2244 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2248 #define HANDLE_UNOP_PHI(eval, a, c) \
2250 if (is_const_Phi(a)) { \
2251 /* check for Op(Phi) */ \
2252 c = apply_unop_on_phi(a, eval); \
2254 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2260 * Do the AddSub optimization, then Transform
2261 * Constant folding on Phi
2262 * Add(a,a) -> Mul(a, 2)
2263 * Add(Mul(a, x), a) -> Mul(a, x+1)
2264 * if the mode is integer or float.
2265 * Transform Add(a,-b) into Sub(a,b).
2266 * Reassociation might fold this further.
2268 static ir_node *transform_node_Add(ir_node *n) {
2270 ir_node *a, *b, *c, *oldn = n;
2272 n = transform_node_AddSub(n);
2274 a = get_Add_left(n);
2275 b = get_Add_right(n);
2277 mode = get_irn_mode(n);
2279 if (mode_is_reference(mode)) {
2280 ir_mode *lmode = get_irn_mode(a);
2282 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2283 /* an Add(a, NULL) is a hidden Conv */
2284 dbg_info *dbg = get_irn_dbg_info(n);
2285 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2289 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2291 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2292 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2295 if (mode_is_num(mode)) {
2296 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2297 if (!is_irg_state(current_ir_graph, IR_GRAPH_STATE_ARCH_DEP)
2298 && a == b && mode_is_int(mode)) {
2299 ir_node *block = get_nodes_block(n);
2302 get_irn_dbg_info(n),
2305 new_Const_long(mode, 2),
2307 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2312 get_irn_dbg_info(n),
2317 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2322 get_irn_dbg_info(n),
2327 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2330 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2331 /* Here we rely on constants be on the RIGHT side */
2333 ir_node *op = get_Not_op(a);
2335 if (is_Const(b) && is_Const_one(b)) {
2337 ir_node *blk = get_nodes_block(n);
2338 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2339 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2344 n = new_Const(get_mode_minus_one(mode));
2345 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2350 ir_node *op = get_Not_op(b);
2354 n = new_Const(get_mode_minus_one(mode));
2355 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2361 if (mode_is_int(mode)) {
2362 tarval *c = tarval_and(
2363 tarval_not(a->vrp.bits_not_set),
2364 tarval_not(b->vrp.bits_not_set)
2367 if(tarval_is_null(c)) {
2368 dbg_info *dbgi = get_irn_dbg_info(n);
2369 return new_rd_Or(dbgi, get_nodes_block(n),
2374 } /* transform_node_Add */
2377 * returns -cnst or NULL if impossible
2379 static ir_node *const_negate(ir_node *cnst) {
2380 tarval *tv = tarval_neg(get_Const_tarval(cnst));
2381 dbg_info *dbgi = get_irn_dbg_info(cnst);
2382 ir_graph *irg = get_irn_irg(cnst);
2383 if (tv == tarval_bad) return NULL;
2384 return new_rd_Const(dbgi, irg, tv);
2388 * Do the AddSub optimization, then Transform
2389 * Constant folding on Phi
2390 * Sub(0,a) -> Minus(a)
2391 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2392 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2393 * Sub(Add(a, x), x) -> a
2394 * Sub(x, Add(x, a)) -> -a
2395 * Sub(x, Const) -> Add(x, -Const)
2397 static ir_node *transform_node_Sub(ir_node *n) {
2402 n = transform_node_AddSub(n);
2404 a = get_Sub_left(n);
2405 b = get_Sub_right(n);
2407 mode = get_irn_mode(n);
2409 if (mode_is_int(mode)) {
2410 ir_mode *lmode = get_irn_mode(a);
2412 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2413 /* a Sub(a, NULL) is a hidden Conv */
2414 dbg_info *dbg = get_irn_dbg_info(n);
2415 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2420 if (mode == lmode &&
2421 get_mode_arithmetic(mode) == irma_twos_complement &&
2423 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2425 dbg_info *dbg = get_irn_dbg_info(n);
2426 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2427 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2433 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2435 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2436 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2439 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2440 /* a - C -> a + (-C) */
2441 ir_node *cnst = const_negate(b);
2443 ir_node *block = get_nodes_block(n);
2444 dbg_info *dbgi = get_irn_dbg_info(n);
2446 n = new_rd_Add(dbgi, block, a, cnst, mode);
2447 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2452 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2453 dbg_info *dbg = get_irn_dbg_info(n);
2454 ir_node *block = get_nodes_block(n);
2455 ir_node *left = get_Minus_op(a);
2456 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2458 n = new_rd_Minus(dbg, block, add, mode);
2459 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2461 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2462 dbg_info *dbg = get_irn_dbg_info(n);
2463 ir_node *block = get_nodes_block(n);
2464 ir_node *right = get_Minus_op(b);
2466 n = new_rd_Add(dbg, block, a, right, mode);
2467 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2469 } else if (is_Sub(b)) {
2470 /* a - (b - c) -> a + (c - b)
2471 * -> (a - b) + c iff (b - c) is a pointer */
2472 dbg_info *s_dbg = get_irn_dbg_info(b);
2473 ir_node *s_block = get_nodes_block(b);
2474 ir_node *s_left = get_Sub_left(b);
2475 ir_node *s_right = get_Sub_right(b);
2476 ir_mode *s_mode = get_irn_mode(b);
2477 if (mode_is_reference(s_mode)) {
2478 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2479 dbg_info *a_dbg = get_irn_dbg_info(n);
2480 ir_node *a_block = get_nodes_block(n);
2483 s_right = new_r_Conv(a_block, s_right, mode);
2484 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2486 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2487 dbg_info *a_dbg = get_irn_dbg_info(n);
2488 ir_node *a_block = get_nodes_block(n);
2490 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2492 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2494 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2495 ir_node *m_right = get_Mul_right(b);
2496 if (is_Const(m_right)) {
2497 ir_node *cnst2 = const_negate(m_right);
2498 if (cnst2 != NULL) {
2499 dbg_info *m_dbg = get_irn_dbg_info(b);
2500 ir_node *m_block = get_nodes_block(b);
2501 ir_node *m_left = get_Mul_left(b);
2502 ir_mode *m_mode = get_irn_mode(b);
2503 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2504 dbg_info *a_dbg = get_irn_dbg_info(n);
2505 ir_node *a_block = get_nodes_block(n);
2507 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2508 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2514 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2515 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2517 get_irn_dbg_info(n),
2521 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2525 if (mode_wrap_around(mode)) {
2526 ir_node *left = get_Add_left(a);
2527 ir_node *right = get_Add_right(a);
2529 /* FIXME: Does the Conv's work only for two complement or generally? */
2531 if (mode != get_irn_mode(right)) {
2532 /* This Sub is an effective Cast */
2533 right = new_r_Conv(get_nodes_block(n), right, mode);
2536 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2538 } else if (right == b) {
2539 if (mode != get_irn_mode(left)) {
2540 /* This Sub is an effective Cast */
2541 left = new_r_Conv(get_nodes_block(n), left, mode);
2544 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2550 if (mode_wrap_around(mode)) {
2551 ir_node *left = get_Add_left(b);
2552 ir_node *right = get_Add_right(b);
2554 /* FIXME: Does the Conv's work only for two complement or generally? */
2556 ir_mode *r_mode = get_irn_mode(right);
2558 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2559 if (mode != r_mode) {
2560 /* This Sub is an effective Cast */
2561 n = new_r_Conv(get_nodes_block(n), n, mode);
2563 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2565 } else if (right == a) {
2566 ir_mode *l_mode = get_irn_mode(left);
2568 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2569 if (mode != l_mode) {
2570 /* This Sub is an effective Cast */
2571 n = new_r_Conv(get_nodes_block(n), n, mode);
2573 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2578 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2579 ir_mode *mode = get_irn_mode(a);
2581 if (mode == get_irn_mode(b)) {
2583 ir_node *op_a = get_Conv_op(a);
2584 ir_node *op_b = get_Conv_op(b);
2586 /* check if it's allowed to skip the conv */
2587 ma = get_irn_mode(op_a);
2588 mb = get_irn_mode(op_b);
2590 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2591 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2594 set_Sub_right(n, b);
2600 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2601 if (!is_reassoc_running() && is_Mul(a)) {
2602 ir_node *ma = get_Mul_left(a);
2603 ir_node *mb = get_Mul_right(a);
2606 ir_node *blk = get_nodes_block(n);
2608 get_irn_dbg_info(n),
2612 get_irn_dbg_info(n),
2615 new_Const_long(mode, 1),
2618 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2620 } else if (mb == b) {
2621 ir_node *blk = get_nodes_block(n);
2623 get_irn_dbg_info(n),
2627 get_irn_dbg_info(n),
2630 new_Const_long(mode, 1),
2633 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2637 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2638 ir_node *x = get_Sub_left(a);
2639 ir_node *y = get_Sub_right(a);
2640 ir_node *blk = get_nodes_block(n);
2641 ir_mode *m_b = get_irn_mode(b);
2642 ir_mode *m_y = get_irn_mode(y);
2646 /* Determine the right mode for the Add. */
2649 else if (mode_is_reference(m_b))
2651 else if (mode_is_reference(m_y))
2655 * Both modes are different but none is reference,
2656 * happens for instance in SubP(SubP(P, Iu), Is).
2657 * We have two possibilities here: Cast or ignore.
2658 * Currently we ignore this case.
2663 add = new_r_Add(blk, y, b, add_mode);
2665 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2666 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2670 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2671 if (is_Const(a) && is_Not(b)) {
2672 /* c - ~X = X + (c+1) */
2673 tarval *tv = get_Const_tarval(a);
2675 tv = tarval_add(tv, get_mode_one(mode));
2676 if (tv != tarval_bad) {
2677 ir_node *blk = get_nodes_block(n);
2678 ir_node *c = new_Const(tv);
2679 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2680 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2686 } /* transform_node_Sub */
2689 * Several transformation done on n*n=2n bits mul.
2690 * These transformations must be done here because new nodes may be produced.
2692 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) {
2694 ir_node *a = get_Mul_left(n);
2695 ir_node *b = get_Mul_right(n);
2696 tarval *ta = value_of(a);
2697 tarval *tb = value_of(b);
2698 ir_mode *smode = get_irn_mode(a);
2700 if (ta == get_mode_one(smode)) {
2701 /* (L)1 * (L)b = (L)b */
2702 ir_node *blk = get_nodes_block(n);
2703 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2704 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2707 else if (ta == get_mode_minus_one(smode)) {
2708 /* (L)-1 * (L)b = (L)b */
2709 ir_node *blk = get_nodes_block(n);
2710 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2711 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2712 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2715 if (tb == get_mode_one(smode)) {
2716 /* (L)a * (L)1 = (L)a */
2717 ir_node *blk = get_irn_n(a, -1);
2718 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2719 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2722 else if (tb == get_mode_minus_one(smode)) {
2723 /* (L)a * (L)-1 = (L)-a */
2724 ir_node *blk = get_nodes_block(n);
2725 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2726 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2727 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2734 * Transform Mul(a,-1) into -a.
2735 * Do constant evaluation of Phi nodes.
2736 * Do architecture dependent optimizations on Mul nodes
2738 static ir_node *transform_node_Mul(ir_node *n) {
2739 ir_node *c, *oldn = n;
2740 ir_mode *mode = get_irn_mode(n);
2741 ir_node *a = get_Mul_left(n);
2742 ir_node *b = get_Mul_right(n);
2744 if (is_Bad(a) || is_Bad(b))
2747 if (mode != get_irn_mode(a))
2748 return transform_node_Mul2n(n, mode);
2750 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2752 if (mode_is_signed(mode)) {
2755 if (value_of(a) == get_mode_minus_one(mode))
2757 else if (value_of(b) == get_mode_minus_one(mode))
2760 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2761 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2766 if (is_Const(b)) { /* (-a) * const -> a * -const */
2767 ir_node *cnst = const_negate(b);
2769 dbg_info *dbgi = get_irn_dbg_info(n);
2770 ir_node *block = get_nodes_block(n);
2771 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2772 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2775 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2776 dbg_info *dbgi = get_irn_dbg_info(n);
2777 ir_node *block = get_nodes_block(n);
2778 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2779 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2781 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2782 ir_node *sub_l = get_Sub_left(b);
2783 ir_node *sub_r = get_Sub_right(b);
2784 dbg_info *dbgi = get_irn_dbg_info(n);
2785 ir_node *block = get_nodes_block(n);
2786 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2787 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2788 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2791 } else if (is_Minus(b)) {
2792 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2793 ir_node *sub_l = get_Sub_left(a);
2794 ir_node *sub_r = get_Sub_right(a);
2795 dbg_info *dbgi = get_irn_dbg_info(n);
2796 ir_node *block = get_nodes_block(n);
2797 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2798 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2799 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2802 } else if (is_Shl(a)) {
2803 ir_node *const shl_l = get_Shl_left(a);
2804 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2805 /* (1 << x) * b -> b << x */
2806 dbg_info *const dbgi = get_irn_dbg_info(n);
2807 ir_node *const block = get_nodes_block(n);
2808 ir_node *const shl_r = get_Shl_right(a);
2809 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2810 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2813 } else if (is_Shl(b)) {
2814 ir_node *const shl_l = get_Shl_left(b);
2815 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2816 /* a * (1 << x) -> a << x */
2817 dbg_info *const dbgi = get_irn_dbg_info(n);
2818 ir_node *const block = get_nodes_block(n);
2819 ir_node *const shl_r = get_Shl_right(b);
2820 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2821 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2825 if (get_mode_arithmetic(mode) == irma_ieee754) {
2827 tarval *tv = get_Const_tarval(a);
2828 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2829 && !tarval_is_negative(tv)) {
2830 /* 2.0 * b = b + b */
2831 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2832 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2836 else if (is_Const(b)) {
2837 tarval *tv = get_Const_tarval(b);
2838 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2839 && !tarval_is_negative(tv)) {
2840 /* a * 2.0 = a + a */
2841 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2842 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2847 return arch_dep_replace_mul_with_shifts(n);
2848 } /* transform_node_Mul */
2851 * Transform a Div Node.
2853 static ir_node *transform_node_Div(ir_node *n) {
2854 ir_mode *mode = get_Div_resmode(n);
2855 ir_node *a = get_Div_left(n);
2856 ir_node *b = get_Div_right(n);
2858 const ir_node *dummy;
2860 if (is_Const(b) && is_const_Phi(a)) {
2861 /* check for Div(Phi, Const) */
2862 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2864 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2868 else if (is_Const(a) && is_const_Phi(b)) {
2869 /* check for Div(Const, Phi) */
2870 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2872 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2876 else if (is_const_Phi(a) && is_const_Phi(b)) {
2877 /* check for Div(Phi, Phi) */
2878 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2880 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2887 if (a == b && value_not_zero(a, &dummy)) {
2888 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2889 value = new_Const(get_mode_one(mode));
2890 DBG_OPT_CSTEVAL(n, value);
2893 if (mode_is_signed(mode) && is_Const(b)) {
2894 tarval *tv = get_Const_tarval(b);
2896 if (tv == get_mode_minus_one(mode)) {
2898 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2899 DBG_OPT_CSTEVAL(n, value);
2903 /* Try architecture dependent optimization */
2904 value = arch_dep_replace_div_by_const(n);
2911 /* Turn Div into a tuple (mem, jmp, bad, value) */
2912 mem = get_Div_mem(n);
2913 blk = get_nodes_block(n);
2915 /* skip a potential Pin */
2916 mem = skip_Pin(mem);
2917 turn_into_tuple(n, pn_Div_max);
2918 set_Tuple_pred(n, pn_Div_M, mem);
2919 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2920 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2921 set_Tuple_pred(n, pn_Div_res, value);
2924 } /* transform_node_Div */
2927 * Transform a Mod node.
2929 static ir_node *transform_node_Mod(ir_node *n) {
2930 ir_mode *mode = get_Mod_resmode(n);
2931 ir_node *a = get_Mod_left(n);
2932 ir_node *b = get_Mod_right(n);
2936 if (is_Const(b) && is_const_Phi(a)) {
2937 /* check for Div(Phi, Const) */
2938 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2940 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2944 else if (is_Const(a) && is_const_Phi(b)) {
2945 /* check for Div(Const, Phi) */
2946 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2948 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2952 else if (is_const_Phi(a) && is_const_Phi(b)) {
2953 /* check for Div(Phi, Phi) */
2954 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2956 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2963 if (tv != tarval_bad) {
2964 value = new_Const(tv);
2966 DBG_OPT_CSTEVAL(n, value);
2969 ir_node *a = get_Mod_left(n);
2970 ir_node *b = get_Mod_right(n);
2971 const ir_node *dummy;
2973 if (a == b && value_not_zero(a, &dummy)) {
2974 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2975 value = new_Const(get_mode_null(mode));
2976 DBG_OPT_CSTEVAL(n, value);
2979 if (mode_is_signed(mode) && is_Const(b)) {
2980 tarval *tv = get_Const_tarval(b);
2982 if (tv == get_mode_minus_one(mode)) {
2984 value = new_Const(get_mode_null(mode));
2985 DBG_OPT_CSTEVAL(n, value);
2989 /* Try architecture dependent optimization */
2990 value = arch_dep_replace_mod_by_const(n);
2998 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2999 mem = get_Mod_mem(n);
3000 blk = get_nodes_block(n);
3002 /* skip a potential Pin */
3003 mem = skip_Pin(mem);
3004 turn_into_tuple(n, pn_Mod_max);
3005 set_Tuple_pred(n, pn_Mod_M, mem);
3006 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3007 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
3008 set_Tuple_pred(n, pn_Mod_res, value);
3011 } /* transform_node_Mod */
3014 * Transform a DivMod node.
3016 static ir_node *transform_node_DivMod(ir_node *n) {
3017 const ir_node *dummy;
3018 ir_node *a = get_DivMod_left(n);
3019 ir_node *b = get_DivMod_right(n);
3020 ir_mode *mode = get_DivMod_resmode(n);
3025 if (is_Const(b) && is_const_Phi(a)) {
3026 /* check for Div(Phi, Const) */
3027 va = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
3028 vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
3030 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3031 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3035 else if (is_Const(a) && is_const_Phi(b)) {
3036 /* check for Div(Const, Phi) */
3037 va = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
3038 vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
3040 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3041 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3045 else if (is_const_Phi(a) && is_const_Phi(b)) {
3046 /* check for Div(Phi, Phi) */
3047 va = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
3048 vb = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
3050 DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI);
3051 DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI);
3058 if (tb != tarval_bad) {
3059 if (tb == get_mode_one(get_tarval_mode(tb))) {
3061 vb = new_Const(get_mode_null(mode));
3062 DBG_OPT_CSTEVAL(n, vb);
3064 } else if (ta != tarval_bad) {
3065 tarval *resa, *resb;
3066 resa = tarval_div(ta, tb);
3067 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
3068 Jmp for X result!? */
3069 resb = tarval_mod(ta, tb);
3070 if (resb == tarval_bad) return n; /* Causes exception! */
3071 va = new_Const(resa);
3072 vb = new_Const(resb);
3073 DBG_OPT_CSTEVAL(n, va);
3074 DBG_OPT_CSTEVAL(n, vb);
3076 } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) {
3077 va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3078 vb = new_Const(get_mode_null(mode));
3079 DBG_OPT_CSTEVAL(n, va);
3080 DBG_OPT_CSTEVAL(n, vb);
3082 } else { /* Try architecture dependent optimization */
3085 arch_dep_replace_divmod_by_const(&va, &vb, n);
3086 evaluated = va != NULL;
3088 } else if (a == b) {
3089 if (value_not_zero(a, &dummy)) {
3091 va = new_Const(get_mode_one(mode));
3092 vb = new_Const(get_mode_null(mode));
3093 DBG_OPT_CSTEVAL(n, va);
3094 DBG_OPT_CSTEVAL(n, vb);
3097 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
3100 } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) {
3101 /* 0 / non-Const = 0 */
3106 if (evaluated) { /* replace by tuple */
3110 mem = get_DivMod_mem(n);
3111 /* skip a potential Pin */
3112 mem = skip_Pin(mem);
3114 blk = get_nodes_block(n);
3115 turn_into_tuple(n, pn_DivMod_max);
3116 set_Tuple_pred(n, pn_DivMod_M, mem);
3117 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(blk));
3118 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
3119 set_Tuple_pred(n, pn_DivMod_res_div, va);
3120 set_Tuple_pred(n, pn_DivMod_res_mod, vb);
3124 } /* transform_node_DivMod */
3127 * Optimize x / c to x * (1/c)
3129 static ir_node *transform_node_Quot(ir_node *n) {
3130 ir_mode *mode = get_Quot_resmode(n);
3133 if (get_mode_arithmetic(mode) == irma_ieee754) {
3134 ir_node *b = get_Quot_right(n);
3135 tarval *tv = value_of(b);
3137 if (tv != tarval_bad) {
3138 int rem = tarval_fp_ops_enabled();
3141 * Floating point constant folding might be disabled here to
3143 * However, as we check for exact result, doing it is safe.
3146 tarval_enable_fp_ops(1);
3147 tv = tarval_quo(get_mode_one(mode), tv);
3148 tarval_enable_fp_ops(rem);
3150 /* Do the transformation if the result is either exact or we are not
3151 using strict rules. */
3152 if (tv != tarval_bad &&
3153 (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) {
3154 ir_node *blk = get_nodes_block(n);
3155 ir_node *c = new_Const(tv);
3156 ir_node *a = get_Quot_left(n);
3157 ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode);
3158 ir_node *mem = get_Quot_mem(n);
3160 /* skip a potential Pin */
3161 mem = skip_Pin(mem);
3162 turn_into_tuple(n, pn_Quot_max);
3163 set_Tuple_pred(n, pn_Quot_M, mem);
3164 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(blk));
3165 set_Tuple_pred(n, pn_Quot_X_except, new_Bad());
3166 set_Tuple_pred(n, pn_Quot_res, m);
3167 DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL);
3172 } /* transform_node_Quot */
3175 * Optimize Abs(x) into x if x is Confirmed >= 0
3176 * Optimize Abs(x) into -x if x is Confirmed <= 0
3177 * Optimize Abs(-x) int Abs(x)
3179 static ir_node *transform_node_Abs(ir_node *n) {
3180 ir_node *c, *oldn = n;
3181 ir_node *a = get_Abs_op(n);
3184 HANDLE_UNOP_PHI(tarval_abs, a, c);
3186 switch (classify_value_sign(a)) {
3187 case value_classified_negative:
3188 mode = get_irn_mode(n);
3191 * We can replace the Abs by -x here.
3192 * We even could add a new Confirm here
3193 * (if not twos complement)
3195 * Note that -x would create a new node, so we could
3196 * not run it in the equivalent_node() context.
3198 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
3200 DBG_OPT_CONFIRM(oldn, n);
3202 case value_classified_positive:
3203 /* n is positive, Abs is not needed */
3206 DBG_OPT_CONFIRM(oldn, n);
3212 /* Abs(-x) = Abs(x) */
3213 mode = get_irn_mode(n);
3214 n = new_rd_Abs(get_irn_dbg_info(n), get_nodes_block(n), get_Minus_op(a), mode);
3215 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X);
3219 } /* transform_node_Abs */
3222 * Optimize -a CMP -b into b CMP a.
3223 * This works only for for modes where unary Minus
3225 * Note that two-complement integers can Overflow
3226 * so it will NOT work.
3228 * For == and != can be handled in Proj(Cmp)
3230 static ir_node *transform_node_Cmp(ir_node *n) {
3232 ir_node *left = get_Cmp_left(n);
3233 ir_node *right = get_Cmp_right(n);
3235 if (is_Minus(left) && is_Minus(right) &&
3236 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3237 ir_node *const new_left = get_Minus_op(right);
3238 ir_node *const new_right = get_Minus_op(left);
3239 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3240 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3243 } /* transform_node_Cmp */
3247 * Transform a Cond node.
3249 * Replace the Cond by a Jmp if it branches on a constant
3252 static ir_node *transform_node_Cond(ir_node *n) {
3255 ir_node *a = get_Cond_selector(n);
3256 tarval *ta = value_of(a);
3258 /* we need block info which is not available in floating irgs */
3259 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
3262 if ((ta != tarval_bad) &&
3263 (get_irn_mode(a) == mode_b) &&
3264 (get_opt_unreachable_code())) {
3265 /* It's a boolean Cond, branching on a boolean constant.
3266 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3267 ir_node *blk = get_nodes_block(n);
3268 jmp = new_r_Jmp(blk);
3269 turn_into_tuple(n, pn_Cond_max);
3270 if (ta == tarval_b_true) {
3271 set_Tuple_pred(n, pn_Cond_false, new_Bad());
3272 set_Tuple_pred(n, pn_Cond_true, jmp);
3274 set_Tuple_pred(n, pn_Cond_false, jmp);
3275 set_Tuple_pred(n, pn_Cond_true, new_Bad());
3277 /* We might generate an endless loop, so keep it alive. */
3278 add_End_keepalive(get_irg_end(current_ir_graph), blk);
3281 } /* transform_node_Cond */
3284 * Prototype of a recursive transform function
3285 * for bitwise distributive transformations.
3287 typedef ir_node* (*recursive_transform)(ir_node *n);
3290 * makes use of distributive laws for and, or, eor
3291 * and(a OP c, b OP c) -> and(a, b) OP c
3292 * note, might return a different op than n
3294 static ir_node *transform_bitwise_distributive(ir_node *n,
3295 recursive_transform trans_func)
3298 ir_node *a = get_binop_left(n);
3299 ir_node *b = get_binop_right(n);
3300 ir_op *op = get_irn_op(a);
3301 ir_op *op_root = get_irn_op(n);
3303 if(op != get_irn_op(b))
3306 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3307 if (op == op_Conv) {
3308 ir_node *a_op = get_Conv_op(a);
3309 ir_node *b_op = get_Conv_op(b);
3310 ir_mode *a_mode = get_irn_mode(a_op);
3311 ir_mode *b_mode = get_irn_mode(b_op);
3312 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3313 ir_node *blk = get_nodes_block(n);
3316 set_binop_left(n, a_op);
3317 set_binop_right(n, b_op);
3318 set_irn_mode(n, a_mode);
3320 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3322 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3328 /* nothing to gain here */
3332 if (op == op_Shrs || op == op_Shr || op == op_Shl
3333 || op == op_And || op == op_Or || op == op_Eor) {
3334 ir_node *a_left = get_binop_left(a);
3335 ir_node *a_right = get_binop_right(a);
3336 ir_node *b_left = get_binop_left(b);
3337 ir_node *b_right = get_binop_right(b);
3339 ir_node *op1 = NULL;
3340 ir_node *op2 = NULL;
3342 if (is_op_commutative(op)) {
3343 if (a_left == b_left) {
3347 } else if(a_left == b_right) {
3351 } else if(a_right == b_left) {
3357 if(a_right == b_right) {
3364 /* (a sop c) & (b sop c) => (a & b) sop c */
3365 ir_node *blk = get_nodes_block(n);
3367 ir_node *new_n = exact_copy(n);
3368 set_binop_left(new_n, op1);
3369 set_binop_right(new_n, op2);
3370 new_n = trans_func(new_n);
3372 if(op_root == op_Eor && op == op_Or) {
3373 dbg_info *dbgi = get_irn_dbg_info(n);
3374 ir_mode *mode = get_irn_mode(c);
3376 c = new_rd_Not(dbgi, blk, c, mode);
3377 n = new_rd_And(dbgi, blk, new_n, c, mode);
3380 set_nodes_block(n, blk);
3381 set_binop_left(n, new_n);
3382 set_binop_right(n, c);
3383 add_identities(current_ir_graph->value_table, n);
3386 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3397 static ir_node *transform_node_And(ir_node *n) {
3398 ir_node *c, *oldn = n;
3399 ir_node *a = get_And_left(n);
3400 ir_node *b = get_And_right(n);
3403 mode = get_irn_mode(n);
3404 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3406 /* we can evaluate 2 Projs of the same Cmp */
3407 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3408 ir_node *pred_a = get_Proj_pred(a);
3409 ir_node *pred_b = get_Proj_pred(b);
3410 if (pred_a == pred_b) {
3411 dbg_info *dbgi = get_irn_dbg_info(n);
3412 ir_node *block = get_nodes_block(pred_a);
3413 pn_Cmp pn_a = get_Proj_proj(a);
3414 pn_Cmp pn_b = get_Proj_proj(b);
3415 /* yes, we can simply calculate with pncs */
3416 pn_Cmp new_pnc = pn_a & pn_b;
3418 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3423 ir_node *op = get_Not_op(b);
3425 ir_node *ba = get_And_left(op);
3426 ir_node *bb = get_And_right(op);
3428 /* it's enough to test the following cases due to normalization! */
3429 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3430 /* (a|b) & ~(a&b) = a^b */
3431 ir_node *block = get_nodes_block(n);
3433 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3434 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3442 ir_node *op = get_Not_op(a);
3444 ir_node *aa = get_And_left(op);
3445 ir_node *ab = get_And_right(op);
3447 /* it's enough to test the following cases due to normalization! */
3448 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3449 /* (a|b) & ~(a&b) = a^b */
3450 ir_node *block = get_nodes_block(n);
3452 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3453 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3460 ir_node *al = get_Eor_left(a);
3461 ir_node *ar = get_Eor_right(a);
3464 /* (b ^ a) & b -> ~a & b */
3465 dbg_info *dbg = get_irn_dbg_info(n);
3466 ir_node *block = get_nodes_block(n);
3468 ar = new_rd_Not(dbg, block, ar, mode);
3469 n = new_rd_And(dbg, block, ar, b, mode);
3470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3474 /* (a ^ b) & b -> ~a & b */
3475 dbg_info *dbg = get_irn_dbg_info(n);
3476 ir_node *block = get_nodes_block(n);
3478 al = new_rd_Not(dbg, block, al, mode);
3479 n = new_rd_And(dbg, block, al, b, mode);
3480 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3485 ir_node *bl = get_Eor_left(b);
3486 ir_node *br = get_Eor_right(b);
3489 /* a & (a ^ b) -> a & ~b */
3490 dbg_info *dbg = get_irn_dbg_info(n);
3491 ir_node *block = get_nodes_block(n);
3493 br = new_rd_Not(dbg, block, br, mode);
3494 n = new_rd_And(dbg, block, br, a, mode);
3495 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3499 /* a & (b ^ a) -> a & ~b */
3500 dbg_info *dbg = get_irn_dbg_info(n);
3501 ir_node *block = get_nodes_block(n);
3503 bl = new_rd_Not(dbg, block, bl, mode);
3504 n = new_rd_And(dbg, block, bl, a, mode);
3505 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3509 if (is_Not(a) && is_Not(b)) {
3510 /* ~a & ~b = ~(a|b) */
3511 ir_node *block = get_nodes_block(n);
3512 ir_mode *mode = get_irn_mode(n);
3516 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3517 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3518 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3522 if (is_Const(a) && (tarval_is_all_one(tarval_or(get_Const_tarval(a),
3523 b->vrp.bits_not_set)))) {
3524 return new_rd_Id(get_irn_dbg_info(n), get_nodes_block(n),
3525 b, get_irn_mode(n));
3529 if (is_Const(b) && (tarval_is_all_one(tarval_or(get_Const_tarval(b),
3530 a->vrp.bits_not_set)))) {
3531 return new_rd_Id(get_irn_dbg_info(n), get_nodes_block(n),
3532 a, get_irn_mode(n));
3536 n = transform_bitwise_distributive(n, transform_node_And);
3539 } /* transform_node_And */
3544 static ir_node *transform_node_Eor(ir_node *n) {
3545 ir_node *c, *oldn = n;
3546 ir_node *a = get_Eor_left(n);
3547 ir_node *b = get_Eor_right(n);
3548 ir_mode *mode = get_irn_mode(n);
3550 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3552 /* we can evaluate 2 Projs of the same Cmp */
3553 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3554 ir_node *pred_a = get_Proj_pred(a);
3555 ir_node *pred_b = get_Proj_pred(b);
3556 if(pred_a == pred_b) {
3557 dbg_info *dbgi = get_irn_dbg_info(n);
3558 ir_node *block = get_nodes_block(pred_a);
3559 pn_Cmp pn_a = get_Proj_proj(a);
3560 pn_Cmp pn_b = get_Proj_proj(b);
3561 /* yes, we can simply calculate with pncs */
3562 pn_Cmp new_pnc = pn_a ^ pn_b;
3564 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
3570 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph,
3571 get_mode_null(mode));
3572 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
3573 } else if (is_Const(b)) {
3574 if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */
3575 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b)));
3576 ir_node *not_op = get_Not_op(a);
3577 dbg_info *dbg = get_irn_dbg_info(n);
3578 ir_node *block = get_nodes_block(n);
3579 ir_mode *mode = get_irn_mode(n);
3580 n = new_rd_Eor(dbg, block, not_op, cnst, mode);
3582 } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */
3583 n = new_r_Not(get_nodes_block(n), a, mode);
3584 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3587 n = transform_bitwise_distributive(n, transform_node_Eor);
3591 } /* transform_node_Eor */
3596 static ir_node *transform_node_Not(ir_node *n) {
3597 ir_node *c, *oldn = n;
3598 ir_node *a = get_Not_op(n);
3599 ir_mode *mode = get_irn_mode(n);
3601 HANDLE_UNOP_PHI(tarval_not,a,c);
3603 /* check for a boolean Not */
3604 if (mode == mode_b && is_Proj(a)) {
3605 ir_node *a_pred = get_Proj_pred(a);
3606 if (is_Cmp(a_pred)) {
3607 ir_node *cmp_block = get_nodes_block(a_pred);
3608 /* We negate a Cmp. The Cmp has the negated result anyways! */
3609 n = new_r_Proj(cmp_block, get_Proj_pred(a),
3610 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3611 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3616 ir_node *eor_b = get_Eor_right(a);
3617 if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */
3618 ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b)));
3619 ir_node *eor_a = get_Eor_left(a);
3620 dbg_info *dbg = get_irn_dbg_info(n);
3621 ir_node *block = get_nodes_block(n);
3622 ir_mode *mode = get_irn_mode(n);
3623 n = new_rd_Eor(dbg, block, eor_a, cnst, mode);
3627 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3628 if (is_Minus(a)) { /* ~-x -> x + -1 */
3629 dbg_info *dbg = get_irn_dbg_info(n);
3630 ir_graph *irg = current_ir_graph;
3631 ir_node *block = get_nodes_block(n);
3632 ir_node *add_l = get_Minus_op(a);
3633 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3634 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3635 } else if (is_Add(a)) {
3636 ir_node *add_r = get_Add_right(a);
3637 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3638 /* ~(x + -1) = -x */
3639 ir_node *op = get_Add_left(a);
3640 ir_node *blk = get_nodes_block(n);
3641 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3642 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3647 } /* transform_node_Not */
3650 * Transform a Minus.
3654 * -(a >>u (size-1)) = a >>s (size-1)
3655 * -(a >>s (size-1)) = a >>u (size-1)
3656 * -(a * const) -> a * -const
3658 static ir_node *transform_node_Minus(ir_node *n) {
3659 ir_node *c, *oldn = n;
3660 ir_node *a = get_Minus_op(n);
3663 HANDLE_UNOP_PHI(tarval_neg,a,c);
3665 mode = get_irn_mode(a);
3666 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3667 /* the following rules are only to twos-complement */
3670 ir_node *op = get_Not_op(a);
3671 tarval *tv = get_mode_one(mode);
3672 ir_node *blk = get_nodes_block(n);
3673 ir_node *c = new_Const(tv);
3674 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3675 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3679 ir_node *c = get_Shr_right(a);
3682 tarval *tv = get_Const_tarval(c);
3684 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3685 /* -(a >>u (size-1)) = a >>s (size-1) */
3686 ir_node *v = get_Shr_left(a);
3688 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3689 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3695 ir_node *c = get_Shrs_right(a);
3698 tarval *tv = get_Const_tarval(c);
3700 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3701 /* -(a >>s (size-1)) = a >>u (size-1) */
3702 ir_node *v = get_Shrs_left(a);
3704 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3705 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3712 /* - (a-b) = b - a */
3713 ir_node *la = get_Sub_left(a);
3714 ir_node *ra = get_Sub_right(a);
3715 ir_node *blk = get_nodes_block(n);
3717 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3718 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3722 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3723 ir_node *mul_l = get_Mul_left(a);
3724 ir_node *mul_r = get_Mul_right(a);
3725 tarval *tv = value_of(mul_r);
3726 if (tv != tarval_bad) {
3727 tv = tarval_neg(tv);
3728 if (tv != tarval_bad) {
3729 ir_node *cnst = new_Const(tv);
3730 dbg_info *dbg = get_irn_dbg_info(a);
3731 ir_node *block = get_nodes_block(a);
3732 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3733 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3740 } /* transform_node_Minus */
3743 * Transform a Cast_type(Const) into a new Const_type
3745 static ir_node *transform_node_Cast(ir_node *n) {
3747 ir_node *pred = get_Cast_op(n);
3748 ir_type *tp = get_irn_type(n);
3750 if (is_Const(pred) && get_Const_type(pred) != tp) {
3751 n = new_rd_Const_type(NULL, current_ir_graph, get_Const_tarval(pred), tp);
3752 DBG_OPT_CSTEVAL(oldn, n);
3753 } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) {
3754 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_mode(pred),
3755 get_SymConst_symbol(pred), get_SymConst_kind(pred), tp);
3756 DBG_OPT_CSTEVAL(oldn, n);
3760 } /* transform_node_Cast */
3763 * Transform a Proj(Load) with a non-null address.
3765 static ir_node *transform_node_Proj_Load(ir_node *proj) {
3766 if (get_opt_ldst_only_null_ptr_exceptions()) {
3767 if (get_irn_mode(proj) == mode_X) {
3768 ir_node *load = get_Proj_pred(proj);
3770 /* get the Load address */
3771 const ir_node *addr = get_Load_ptr(load);
3772 const ir_node *confirm;
3774 if (value_not_null(addr, &confirm)) {
3775 if (confirm == NULL) {
3776 /* this node may float if it did not depend on a Confirm */
3777 set_irn_pinned(load, op_pin_state_floats);
3779 if (get_Proj_proj(proj) == pn_Load_X_except) {
3780 DBG_OPT_EXC_REM(proj);
3781 return get_irg_bad(current_ir_graph);
3783 ir_node *blk = get_nodes_block(load);
3784 return new_r_Jmp(blk);
3790 } /* transform_node_Proj_Load */
3793 * Transform a Proj(Store) with a non-null address.
3795 static ir_node *transform_node_Proj_Store(ir_node *proj) {
3796 if (get_opt_ldst_only_null_ptr_exceptions()) {
3797 if (get_irn_mode(proj) == mode_X) {
3798 ir_node *store = get_Proj_pred(proj);
3800 /* get the load/store address */
3801 const ir_node *addr = get_Store_ptr(store);
3802 const ir_node *confirm;
3804 if (value_not_null(addr, &confirm)) {
3805 if (confirm == NULL) {
3806 /* this node may float if it did not depend on a Confirm */
3807 set_irn_pinned(store, op_pin_state_floats);
3809 if (get_Proj_proj(proj) == pn_Store_X_except) {
3810 DBG_OPT_EXC_REM(proj);
3811 return get_irg_bad(current_ir_graph);
3813 ir_node *blk = get_nodes_block(store);
3814 return new_r_Jmp(blk);
3820 } /* transform_node_Proj_Store */
3823 * Transform a Proj(Div) with a non-zero value.
3824 * Removes the exceptions and routes the memory to the NoMem node.
3826 static ir_node *transform_node_Proj_Div(ir_node *proj) {
3827 ir_node *div = get_Proj_pred(proj);
3828 ir_node *b = get_Div_right(div);
3829 ir_node *res, *new_mem;
3830 const ir_node *confirm;
3833 if (value_not_zero(b, &confirm)) {
3834 /* div(x, y) && y != 0 */
3835 if (confirm == NULL) {
3836 /* we are sure we have a Const != 0 */
3837 new_mem = get_Div_mem(div);
3838 new_mem = skip_Pin(new_mem);
3839 set_Div_mem(div, new_mem);
3840 set_irn_pinned(div, op_pin_state_floats);
3843 proj_nr = get_Proj_proj(proj);
3845 case pn_Div_X_regular:
3846 return new_r_Jmp(get_nodes_block(div));
3848 case pn_Div_X_except:
3849 /* we found an exception handler, remove it */
3850 DBG_OPT_EXC_REM(proj);
3854 res = get_Div_mem(div);
3855 new_mem = get_irg_no_mem(current_ir_graph);
3858 /* This node can only float up to the Confirm block */
3859 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3861 set_irn_pinned(div, op_pin_state_floats);
3862 /* this is a Div without exception, we can remove the memory edge */
3863 set_Div_mem(div, new_mem);
3868 } /* transform_node_Proj_Div */
3871 * Transform a Proj(Mod) with a non-zero value.
3872 * Removes the exceptions and routes the memory to the NoMem node.
3874 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
3875 ir_node *mod = get_Proj_pred(proj);
3876 ir_node *b = get_Mod_right(mod);
3877 ir_node *res, *new_mem;
3878 const ir_node *confirm;
3881 if (value_not_zero(b, &confirm)) {
3882 /* mod(x, y) && y != 0 */
3883 proj_nr = get_Proj_proj(proj);
3885 if (confirm == NULL) {
3886 /* we are sure we have a Const != 0 */
3887 new_mem = get_Mod_mem(mod);
3888 new_mem = skip_Pin(new_mem);
3889 set_Mod_mem(mod, new_mem);
3890 set_irn_pinned(mod, op_pin_state_floats);
3895 case pn_Mod_X_regular:
3896 return new_r_Jmp(get_irn_n(mod, -1));
3898 case pn_Mod_X_except:
3899 /* we found an exception handler, remove it */
3900 DBG_OPT_EXC_REM(proj);
3904 res = get_Mod_mem(mod);
3905 new_mem = get_irg_no_mem(current_ir_graph);
3908 /* This node can only float up to the Confirm block */
3909 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3911 /* this is a Mod without exception, we can remove the memory edge */
3912 set_Mod_mem(mod, new_mem);
3915 if (get_Mod_left(mod) == b) {
3916 /* a % a = 0 if a != 0 */
3917 ir_mode *mode = get_irn_mode(proj);
3918 ir_node *res = new_Const(get_mode_null(mode));
3920 DBG_OPT_CSTEVAL(mod, res);
3926 } /* transform_node_Proj_Mod */
3929 * Transform a Proj(DivMod) with a non-zero value.
3930 * Removes the exceptions and routes the memory to the NoMem node.
3932 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
3933 ir_node *divmod = get_Proj_pred(proj);
3934 ir_node *b = get_DivMod_right(divmod);
3935 ir_node *res, *new_mem;
3936 const ir_node *confirm;
3939 if (value_not_zero(b, &confirm)) {
3940 /* DivMod(x, y) && y != 0 */
3941 proj_nr = get_Proj_proj(proj);
3943 if (confirm == NULL) {
3944 /* we are sure we have a Const != 0 */
3945 new_mem = get_DivMod_mem(divmod);
3946 new_mem = skip_Pin(new_mem);
3947 set_DivMod_mem(divmod, new_mem);
3948 set_irn_pinned(divmod, op_pin_state_floats);
3953 case pn_DivMod_X_regular:
3954 return new_r_Jmp(get_nodes_block(divmod));
3956 case pn_DivMod_X_except:
3957 /* we found an exception handler, remove it */
3958 DBG_OPT_EXC_REM(proj);
3962 res = get_DivMod_mem(divmod);
3963 new_mem = get_irg_no_mem(current_ir_graph);
3966 /* This node can only float up to the Confirm block */
3967 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3969 /* this is a DivMod without exception, we can remove the memory edge */
3970 set_DivMod_mem(divmod, new_mem);
3973 case pn_DivMod_res_mod:
3974 if (get_DivMod_left(divmod) == b) {
3975 /* a % a = 0 if a != 0 */
3976 ir_mode *mode = get_irn_mode(proj);
3977 ir_node *res = new_Const(get_mode_null(mode));
3979 DBG_OPT_CSTEVAL(divmod, res);
3985 } /* transform_node_Proj_DivMod */
3988 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3990 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
3991 if (get_opt_unreachable_code()) {
3992 ir_node *n = get_Proj_pred(proj);
3993 ir_node *b = get_Cond_selector(n);
3995 if (mode_is_int(get_irn_mode(b))) {
3996 tarval *tb = value_of(b);
3998 if (tb != tarval_bad) {
3999 /* we have a constant switch */
4000 long num = get_Proj_proj(proj);
4002 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
4003 if (get_tarval_long(tb) == num) {
4004 /* Do NOT create a jump here, or we will have 2 control flow ops
4005 * in a block. This case is optimized away in optimize_cf(). */
4008 /* this case will NEVER be taken, kill it */
4009 return get_irg_bad(current_ir_graph);
4013 long num = get_Proj_proj(proj);
4014 if (num != get_Cond_default_proj(n)) {
4015 /* Try handling with vrp data. We only remove dead parts. */
4016 tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
4018 if (b->vrp.range_type == VRP_RANGE) {
4019 pn_Cmp cmp_result = tarval_cmp(b->vrp.range_bottom, tp);
4020 pn_Cmp cmp_result2 = tarval_cmp(b->vrp.range_top, tp);
4022 if ((cmp_result & pn_Cmp_Lt) == cmp_result && (cmp_result2
4023 & pn_Cmp_Gt) == cmp_result2) {
4024 return get_irg_bad(current_ir_graph);
4026 } else if (b->vrp.range_type == VRP_ANTIRANGE) {
4027 pn_Cmp cmp_result = tarval_cmp(b->vrp.range_bottom, tp);
4028 pn_Cmp cmp_result2 = tarval_cmp(b->vrp.range_top, tp);
4030 if ((cmp_result & pn_Cmp_Ge) == cmp_result && (cmp_result2
4031 & pn_Cmp_Le) == cmp_result2) {
4032 return get_irg_bad(current_ir_graph);
4037 tarval_and( b->vrp.bits_set, tp),
4041 return get_irg_bad(current_ir_graph);
4047 b->vrp.bits_not_set),
4048 b->vrp.bits_not_set)
4051 return get_irg_bad(current_ir_graph);
4060 } /* transform_node_Proj_Cond */
4063 * Create a 0 constant of given mode.
4065 static ir_node *create_zero_const(ir_mode *mode) {
4066 tarval *tv = get_mode_null(mode);
4067 ir_node *cnst = new_Const(tv);
4072 /* the order of the values is important! */
4073 typedef enum const_class {
4079 static const_class classify_const(const ir_node* n)
4081 if (is_Const(n)) return const_const;
4082 if (is_irn_constlike(n)) return const_like;
4087 * Determines whether r is more constlike or has a larger index (in that order)
4090 static int operands_are_normalized(const ir_node *l, const ir_node *r)
4092 const const_class l_order = classify_const(l);
4093 const const_class r_order = classify_const(r);
4095 l_order > r_order ||
4096 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
4100 * Normalizes and optimizes Cmp nodes.
4102 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
4103 ir_node *n = get_Proj_pred(proj);
4104 ir_node *left = get_Cmp_left(n);
4105 ir_node *right = get_Cmp_right(n);
4108 ir_mode *mode = NULL;
4109 long proj_nr = get_Proj_proj(proj);
4111 /* we can evaluate some cases directly */
4114 return new_Const(get_tarval_b_false());
4116 return new_Const(get_tarval_b_true());
4118 if (!mode_is_float(get_irn_mode(left)))
4119 return new_Const(get_tarval_b_true());
4125 /* remove Casts of both sides */
4126 left = skip_Cast(left);
4127 right = skip_Cast(right);
4129 /* Remove unnecessary conversions */
4130 /* TODO handle constants */
4131 if (is_Conv(left) && is_Conv(right)) {
4132 ir_mode *mode = get_irn_mode(left);
4133 ir_node *op_left = get_Conv_op(left);
4134 ir_node *op_right = get_Conv_op(right);
4135 ir_mode *mode_left = get_irn_mode(op_left);
4136 ir_mode *mode_right = get_irn_mode(op_right);
4138 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
4139 && mode_left != mode_b && mode_right != mode_b) {
4140 ir_node *block = get_nodes_block(n);
4142 if (mode_left == mode_right) {
4146 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
4147 } else if (smaller_mode(mode_left, mode_right)) {
4148 left = new_r_Conv(block, op_left, mode_right);
4151 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4152 } else if (smaller_mode(mode_right, mode_left)) {
4154 right = new_r_Conv(block, op_right, mode_left);
4156 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4161 /* remove operation on both sides if possible */
4162 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4164 * The following operations are NOT safe for floating point operations, for instance
4165 * 1.0 + inf == 2.0 + inf, =/=> x == y
4167 if (mode_is_int(get_irn_mode(left))) {
4168 unsigned lop = get_irn_opcode(left);
4170 if (lop == get_irn_opcode(right)) {
4171 ir_node *ll, *lr, *rl, *rr;
4173 /* same operation on both sides, try to remove */
4177 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
4178 left = get_unop_op(left);
4179 right = get_unop_op(right);
4181 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4184 ll = get_Add_left(left);
4185 lr = get_Add_right(left);
4186 rl = get_Add_left(right);
4187 rr = get_Add_right(right);
4190 /* X + a CMP X + b ==> a CMP b */
4194 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4195 } else if (ll == rr) {
4196 /* X + a CMP b + X ==> a CMP b */
4200 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4201 } else if (lr == rl) {
4202 /* a + X CMP X + b ==> a CMP b */
4206 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4207 } else if (lr == rr) {
4208 /* a + X CMP b + X ==> a CMP b */
4212 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4216 ll = get_Sub_left(left);
4217 lr = get_Sub_right(left);
4218 rl = get_Sub_left(right);
4219 rr = get_Sub_right(right);
4222 /* X - a CMP X - b ==> a CMP b */
4226 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4227 } else if (lr == rr) {
4228 /* a - X CMP b - X ==> a CMP b */
4232 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4236 if (get_Rotl_right(left) == get_Rotl_right(right)) {
4237 /* a ROTL X CMP b ROTL X ==> a CMP b */
4238 left = get_Rotl_left(left);
4239 right = get_Rotl_left(right);
4241 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4249 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4250 if (is_Add(left) || is_Sub(left)) {
4251 ir_node *ll = get_binop_left(left);
4252 ir_node *lr = get_binop_right(left);
4254 if (lr == right && is_Add(left)) {
4261 right = create_zero_const(get_irn_mode(left));
4263 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4266 if (is_Add(right) || is_Sub(right)) {
4267 ir_node *rl = get_binop_left(right);
4268 ir_node *rr = get_binop_right(right);
4270 if (rr == left && is_Add(right)) {
4277 right = create_zero_const(get_irn_mode(left));
4279 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4282 if (is_And(left) && is_Const(right)) {
4283 ir_node *ll = get_binop_left(left);
4284 ir_node *lr = get_binop_right(left);
4285 if (is_Shr(ll) && is_Const(lr)) {
4286 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4287 ir_node *block = get_nodes_block(n);
4288 ir_mode *mode = get_irn_mode(left);
4290 ir_node *llr = get_Shr_right(ll);
4291 if (is_Const(llr)) {
4292 dbg_info *dbg = get_irn_dbg_info(left);
4294 tarval *c1 = get_Const_tarval(llr);
4295 tarval *c2 = get_Const_tarval(lr);
4296 tarval *c3 = get_Const_tarval(right);
4297 tarval *mask = tarval_shl(c2, c1);
4298 tarval *value = tarval_shl(c3, c1);
4300 left = new_rd_And(dbg, block, get_Shr_left(ll), new_Const(mask), mode);
4301 right = new_Const(value);
4306 } /* mode_is_int(...) */
4307 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4309 /* replace mode_b compares with ands/ors */
4310 if (get_irn_mode(left) == mode_b) {
4311 ir_node *block = get_nodes_block(n);
4315 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4316 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4317 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4318 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4319 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4320 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4321 default: bres = NULL;
4324 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4330 * First step: normalize the compare op
4331 * by placing the constant on the right side
4332 * or moving the lower address node to the left.
4334 if (!operands_are_normalized(left, right)) {
4340 proj_nr = get_inversed_pnc(proj_nr);
4345 * Second step: Try to reduce the magnitude
4346 * of a constant. This may help to generate better code
4347 * later and may help to normalize more compares.
4348 * Of course this is only possible for integer values.
4350 tv = value_of(right);
4351 if (tv != tarval_bad) {
4352 mode = get_irn_mode(right);
4354 /* TODO extend to arbitrary constants */
4355 if (is_Conv(left) && tarval_is_null(tv)) {
4356 ir_node *op = get_Conv_op(left);
4357 ir_mode *op_mode = get_irn_mode(op);
4360 * UpConv(x) REL 0 ==> x REL 0
4361 * Don't do this for float values as it's unclear whether it is a
4362 * win. (on the other side it makes detection/creation of fabs hard)
4364 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4365 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4366 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4367 !mode_is_float(mode)) {
4368 tv = get_mode_null(op_mode);
4372 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4376 if (tv != tarval_bad) {
4377 /* the following optimization is possible on modes without Overflow
4378 * on Unary Minus or on == and !=:
4379 * -a CMP c ==> a swap(CMP) -c
4381 * Beware: for two-complement Overflow may occur, so only == and != can
4382 * be optimized, see this:
4383 * -MININT < 0 =/=> MININT > 0 !!!
4385 if (is_Minus(left) &&
4386 (!mode_overflow_on_unary_Minus(mode) ||
4387 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4388 tv = tarval_neg(tv);
4390 if (tv != tarval_bad) {
4391 left = get_Minus_op(left);
4392 proj_nr = get_inversed_pnc(proj_nr);
4394 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4396 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4397 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4398 tv = tarval_not(tv);
4400 if (tv != tarval_bad) {
4401 left = get_Not_op(left);
4403 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4407 /* for integer modes, we have more */
4408 if (mode_is_int(mode)) {
4409 /* Ne includes Unordered which is not possible on integers.
4410 * However, frontends often use this wrong, so fix it here */
4411 if (proj_nr & pn_Cmp_Uo) {
4412 proj_nr &= ~pn_Cmp_Uo;
4413 set_Proj_proj(proj, proj_nr);
4416 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4417 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4418 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4419 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4421 if (tv != tarval_bad) {
4422 proj_nr ^= pn_Cmp_Eq;
4424 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4427 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4428 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4429 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4430 tv = tarval_add(tv, get_mode_one(mode));
4432 if (tv != tarval_bad) {
4433 proj_nr ^= pn_Cmp_Eq;
4435 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4439 /* the following reassociations work only for == and != */
4440 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4442 #if 0 /* Might be not that good in general */
4443 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4444 if (tarval_is_null(tv) && is_Sub(left)) {
4445 right = get_Sub_right(left);
4446 left = get_Sub_left(left);
4448 tv = value_of(right);
4450 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4454 if (tv != tarval_bad) {
4455 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4457 ir_node *c1 = get_Sub_right(left);
4458 tarval *tv2 = value_of(c1);
4460 if (tv2 != tarval_bad) {
4461 tv2 = tarval_add(tv, value_of(c1));
4463 if (tv2 != tarval_bad) {
4464 left = get_Sub_left(left);
4467 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4471 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4472 else if (is_Add(left)) {
4473 ir_node *a_l = get_Add_left(left);
4474 ir_node *a_r = get_Add_right(left);
4478 if (is_Const(a_l)) {
4480 tv2 = value_of(a_l);
4483 tv2 = value_of(a_r);
4486 if (tv2 != tarval_bad) {
4487 tv2 = tarval_sub(tv, tv2, NULL);
4489 if (tv2 != tarval_bad) {
4493 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4497 /* -a == c ==> a == -c, -a != c ==> a != -c */
4498 else if (is_Minus(left)) {
4499 tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4501 if (tv2 != tarval_bad) {
4502 left = get_Minus_op(left);
4505 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4510 /* the following reassociations work only for <= */
4511 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
4512 if (tv != tarval_bad) {
4513 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
4514 if (is_Abs(left)) { // TODO something is missing here
4520 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4521 switch (get_irn_opcode(left)) {
4525 c1 = get_And_right(left);
4528 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4529 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4531 tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4533 /* TODO: move to constant evaluation */
4534 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4536 DBG_OPT_CSTEVAL(proj, c1);
4540 if (tarval_is_single_bit(tv)) {
4542 * optimization for AND:
4544 * And(x, C) == C ==> And(x, C) != 0
4545 * And(x, C) != C ==> And(X, C) == 0
4547 * if C is a single Bit constant.
4550 /* check for Constant's match. We have check hare the tarvals,
4551 because our const might be changed */
4552 if (get_Const_tarval(c1) == tv) {
4553 /* fine: do the transformation */
4554 tv = get_mode_null(get_tarval_mode(tv));
4555 proj_nr ^= pn_Cmp_Leg;
4557 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4563 c1 = get_Or_right(left);
4564 if (is_Const(c1) && tarval_is_null(tv)) {
4566 * Or(x, C) == 0 && C != 0 ==> FALSE
4567 * Or(x, C) != 0 && C != 0 ==> TRUE
4569 if (! tarval_is_null(get_Const_tarval(c1))) {
4570 /* TODO: move to constant evaluation */
4571 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4573 DBG_OPT_CSTEVAL(proj, c1);
4580 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4582 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4585 c1 = get_Shl_right(left);
4587 tarval *tv1 = get_Const_tarval(c1);
4588 ir_mode *mode = get_irn_mode(left);
4589 tarval *minus1 = get_mode_all_one(mode);
4590 tarval *amask = tarval_shr(minus1, tv1);
4591 tarval *cmask = tarval_shl(minus1, tv1);
4594 if (tarval_and(tv, cmask) != tv) {
4595 /* condition not met */
4596 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4598 DBG_OPT_CSTEVAL(proj, c1);
4601 sl = get_Shl_left(left);
4602 blk = get_nodes_block(n);
4603 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4604 tv = tarval_shr(tv, tv1);
4606 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4611 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4613 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4616 c1 = get_Shr_right(left);
4618 tarval *tv1 = get_Const_tarval(c1);
4619 ir_mode *mode = get_irn_mode(left);
4620 tarval *minus1 = get_mode_all_one(mode);
4621 tarval *amask = tarval_shl(minus1, tv1);
4622 tarval *cmask = tarval_shr(minus1, tv1);
4625 if (tarval_and(tv, cmask) != tv) {
4626 /* condition not met */
4627 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4629 DBG_OPT_CSTEVAL(proj, c1);
4632 sl = get_Shr_left(left);
4633 blk = get_nodes_block(n);
4634 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4635 tv = tarval_shl(tv, tv1);
4637 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4642 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4644 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4647 c1 = get_Shrs_right(left);
4649 tarval *tv1 = get_Const_tarval(c1);
4650 ir_mode *mode = get_irn_mode(left);
4651 tarval *minus1 = get_mode_all_one(mode);
4652 tarval *amask = tarval_shl(minus1, tv1);
4653 tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4656 cond = tarval_sub(cond, tv1, NULL);
4657 cond = tarval_shrs(tv, cond);
4659 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4660 /* condition not met */
4661 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4663 DBG_OPT_CSTEVAL(proj, c1);
4666 sl = get_Shrs_left(left);
4667 blk = get_nodes_block(n);
4668 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode);
4669 tv = tarval_shl(tv, tv1);
4671 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4676 } /* tarval != bad */
4679 if (changed & 2) /* need a new Const */
4680 right = new_Const(tv);
4682 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4683 ir_node *op = get_Proj_pred(left);
4685 if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) ||
4686 (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) {
4687 ir_node *c = get_binop_right(op);
4690 tarval *tv = get_Const_tarval(c);
4692 if (tarval_is_single_bit(tv)) {
4693 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4694 ir_node *v = get_binop_left(op);
4695 ir_node *blk = get_irn_n(op, -1);
4696 ir_mode *mode = get_irn_mode(v);
4698 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4699 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_Const(tv), mode);
4701 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4708 ir_node *block = get_nodes_block(n);
4710 /* create a new compare */
4711 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4712 proj = new_rd_Proj(get_irn_dbg_info(proj), block, n, get_irn_mode(proj), proj_nr);
4716 } /* transform_node_Proj_Cmp */
4719 * Optimize CopyB(mem, x, x) into a Nop.
4721 static ir_node *transform_node_Proj_CopyB(ir_node *proj) {
4722 ir_node *copyb = get_Proj_pred(proj);
4723 ir_node *a = get_CopyB_dst(copyb);
4724 ir_node *b = get_CopyB_src(copyb);
4727 switch (get_Proj_proj(proj)) {
4728 case pn_CopyB_X_regular:
4729 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4730 DBG_OPT_EXC_REM(proj);
4731 proj = new_r_Jmp(get_nodes_block(copyb));
4733 case pn_CopyB_X_except:
4734 DBG_OPT_EXC_REM(proj);
4735 proj = get_irg_bad(get_irn_irg(proj));
4742 } /* transform_node_Proj_CopyB */
4745 * Optimize Bounds(idx, idx, upper) into idx.
4747 static ir_node *transform_node_Proj_Bound(ir_node *proj) {
4748 ir_node *oldn = proj;
4749 ir_node *bound = get_Proj_pred(proj);
4750 ir_node *idx = get_Bound_index(bound);
4751 ir_node *pred = skip_Proj(idx);
4754 if (idx == get_Bound_lower(bound))
4756 else if (is_Bound(pred)) {
4758 * idx was Bounds checked in the same MacroBlock previously,
4759 * it is still valid if lower <= pred_lower && pred_upper <= upper.
4761 ir_node *lower = get_Bound_lower(bound);
4762 ir_node *upper = get_Bound_upper(bound);
4763 if (get_Bound_lower(pred) == lower &&
4764 get_Bound_upper(pred) == upper &&
4765 get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) {
4767 * One could expect that we simply return the previous
4768 * Bound here. However, this would be wrong, as we could
4769 * add an exception Proj to a new location then.
4770 * So, we must turn in into a tuple.
4776 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4777 switch (get_Proj_proj(proj)) {
4779 DBG_OPT_EXC_REM(proj);
4780 proj = get_Bound_mem(bound);
4782 case pn_Bound_X_except:
4783 DBG_OPT_EXC_REM(proj);
4784 proj = get_irg_bad(get_irn_irg(proj));
4788 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4790 case pn_Bound_X_regular:
4791 DBG_OPT_EXC_REM(proj);
4792 proj = new_r_Jmp(get_nodes_block(bound));
4799 } /* transform_node_Proj_Bound */
4802 * Does all optimizations on nodes that must be done on it's Proj's
4803 * because of creating new nodes.
4805 static ir_node *transform_node_Proj(ir_node *proj) {
4806 ir_node *n = get_Proj_pred(proj);
4808 if (n->op->ops.transform_node_Proj)
4809 return n->op->ops.transform_node_Proj(proj);
4811 } /* transform_node_Proj */
4814 * Move Confirms down through Phi nodes.
4816 static ir_node *transform_node_Phi(ir_node *phi) {
4818 ir_mode *mode = get_irn_mode(phi);
4820 if (mode_is_reference(mode)) {
4821 n = get_irn_arity(phi);
4823 /* Beware of Phi0 */
4825 ir_node *pred = get_irn_n(phi, 0);
4826 ir_node *bound, *new_Phi, *block, **in;
4829 if (! is_Confirm(pred))
4832 bound = get_Confirm_bound(pred);
4833 pnc = get_Confirm_cmp(pred);
4835 NEW_ARR_A(ir_node *, in, n);
4836 in[0] = get_Confirm_value(pred);
4838 for (i = 1; i < n; ++i) {
4839 pred = get_irn_n(phi, i);
4841 if (! is_Confirm(pred) ||
4842 get_Confirm_bound(pred) != bound ||
4843 get_Confirm_cmp(pred) != pnc)
4845 in[i] = get_Confirm_value(pred);
4847 /* move the Confirm nodes "behind" the Phi */
4848 block = get_irn_n(phi, -1);
4849 new_Phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4850 return new_r_Confirm(block, new_Phi, bound, pnc);
4854 } /* transform_node_Phi */
4857 * Returns the operands of a commutative bin-op, if one operand is
4858 * a const, it is returned as the second one.
4860 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
4861 ir_node *op_a = get_binop_left(binop);
4862 ir_node *op_b = get_binop_right(binop);
4864 assert(is_op_commutative(get_irn_op(binop)));
4866 if (is_Const(op_a)) {
4873 } /* get_comm_Binop_Ops */
4876 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4877 * Such pattern may arise in bitfield stores.
4879 * value c4 value c4 & c2
4880 * AND c3 AND c1 | c3
4887 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4890 static ir_node *transform_node_Or_bf_store(ir_node *or) {
4893 ir_node *and_l, *c3;
4894 ir_node *value, *c4;
4895 ir_node *new_and, *new_const, *block;
4896 ir_mode *mode = get_irn_mode(or);
4898 tarval *tv1, *tv2, *tv3, *tv4, *tv;
4901 get_comm_Binop_Ops(or, &and, &c1);
4902 if (!is_Const(c1) || !is_And(and))
4905 get_comm_Binop_Ops(and, &or_l, &c2);
4909 tv1 = get_Const_tarval(c1);
4910 tv2 = get_Const_tarval(c2);
4912 tv = tarval_or(tv1, tv2);
4913 if (tarval_is_all_one(tv)) {
4914 /* the AND does NOT clear a bit with isn't set by the OR */
4915 set_Or_left(or, or_l);
4916 set_Or_right(or, c1);
4918 /* check for more */
4925 get_comm_Binop_Ops(or_l, &and_l, &c3);
4926 if (!is_Const(c3) || !is_And(and_l))
4929 get_comm_Binop_Ops(and_l, &value, &c4);
4933 /* ok, found the pattern, check for conditions */
4934 assert(mode == get_irn_mode(and));
4935 assert(mode == get_irn_mode(or_l));
4936 assert(mode == get_irn_mode(and_l));
4938 tv3 = get_Const_tarval(c3);
4939 tv4 = get_Const_tarval(c4);
4941 tv = tarval_or(tv4, tv2);
4942 if (!tarval_is_all_one(tv)) {
4943 /* have at least one 0 at the same bit position */
4947 if (tv3 != tarval_andnot(tv3, tv4)) {
4948 /* bit in the or_mask is outside the and_mask */
4952 if (tv1 != tarval_andnot(tv1, tv2)) {
4953 /* bit in the or_mask is outside the and_mask */
4957 /* ok, all conditions met */
4958 block = get_irn_n(or, -1);
4960 new_and = new_r_And(block, value, new_Const(tarval_and(tv4, tv2)), mode);
4962 new_const = new_Const(tarval_or(tv3, tv1));
4964 set_Or_left(or, new_and);
4965 set_Or_right(or, new_const);
4967 /* check for more */
4969 } /* transform_node_Or_bf_store */
4972 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4974 static ir_node *transform_node_Or_Rotl(ir_node *or) {
4975 ir_mode *mode = get_irn_mode(or);
4976 ir_node *shl, *shr, *block;
4977 ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval;
4980 if (! mode_is_int(mode))
4983 shl = get_binop_left(or);
4984 shr = get_binop_right(or);
4993 } else if (!is_Shl(shl)) {
4995 } else if (!is_Shr(shr)) {
4998 x = get_Shl_left(shl);
4999 if (x != get_Shr_left(shr))
5002 c1 = get_Shl_right(shl);
5003 c2 = get_Shr_right(shr);
5004 if (is_Const(c1) && is_Const(c2)) {
5005 tv1 = get_Const_tarval(c1);
5006 if (! tarval_is_long(tv1))
5009 tv2 = get_Const_tarval(c2);
5010 if (! tarval_is_long(tv2))
5013 if (get_tarval_long(tv1) + get_tarval_long(tv2)
5014 != (int) get_mode_size_bits(mode))
5017 /* yet, condition met */
5018 block = get_nodes_block(or);
5020 n = new_r_Rotl(block, x, c1, mode);
5022 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
5029 rotval = sub; /* a Rot right is not supported, so use a rot left */
5030 } else if (is_Sub(c2)) {
5036 if (get_Sub_right(sub) != v)
5039 c1 = get_Sub_left(sub);
5043 tv1 = get_Const_tarval(c1);
5044 if (! tarval_is_long(tv1))
5047 if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode))
5050 /* yet, condition met */
5051 block = get_nodes_block(or);
5053 n = new_r_Rotl(block, x, rotval, mode);
5055 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL);
5057 } /* transform_node_Or_Rotl */
5062 static ir_node *transform_node_Or(ir_node *n) {
5063 ir_node *c, *oldn = n;
5064 ir_node *a = get_Or_left(n);
5065 ir_node *b = get_Or_right(n);
5068 if (is_Not(a) && is_Not(b)) {
5069 /* ~a | ~b = ~(a&b) */
5070 ir_node *block = get_nodes_block(n);
5072 mode = get_irn_mode(n);
5075 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
5076 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
5077 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
5081 /* we can evaluate 2 Projs of the same Cmp */
5082 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
5083 ir_node *pred_a = get_Proj_pred(a);
5084 ir_node *pred_b = get_Proj_pred(b);
5085 if (pred_a == pred_b) {
5086 dbg_info *dbgi = get_irn_dbg_info(n);
5087 ir_node *block = get_nodes_block(pred_a);
5088 pn_Cmp pn_a = get_Proj_proj(a);
5089 pn_Cmp pn_b = get_Proj_proj(b);
5090 /* yes, we can simply calculate with pncs */
5091 pn_Cmp new_pnc = pn_a | pn_b;
5093 return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc);
5097 mode = get_irn_mode(n);
5098 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
5100 n = transform_node_Or_bf_store(n);
5101 n = transform_node_Or_Rotl(n);
5105 n = transform_bitwise_distributive(n, transform_node_Or);
5108 } /* transform_node_Or */
5112 static ir_node *transform_node(ir_node *n);
5115 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
5117 * Should be moved to reassociation?
5119 static ir_node *transform_node_shift(ir_node *n) {
5120 ir_node *left, *right;
5122 tarval *tv1, *tv2, *res;
5123 ir_node *in[2], *irn, *block;
5125 left = get_binop_left(n);
5127 /* different operations */
5128 if (get_irn_op(left) != get_irn_op(n))
5131 right = get_binop_right(n);
5132 tv1 = value_of(right);
5133 if (tv1 == tarval_bad)
5136 tv2 = value_of(get_binop_right(left));
5137 if (tv2 == tarval_bad)
5140 res = tarval_add(tv1, tv2);
5141 mode = get_irn_mode(n);
5143 /* beware: a simple replacement works only, if res < modulo shift */
5145 int modulo_shf = get_mode_modulo_shift(mode);
5146 if (modulo_shf > 0) {
5147 tarval *modulo = new_tarval_from_long(modulo_shf,
5148 get_tarval_mode(res));
5150 assert(modulo_shf >= (int) get_mode_size_bits(mode));
5152 /* shifting too much */
5153 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
5155 ir_node *block = get_nodes_block(n);
5156 dbg_info *dbgi = get_irn_dbg_info(n);
5157 ir_mode *smode = get_irn_mode(right);
5158 ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1);
5159 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
5162 return new_Const(get_mode_null(mode));
5166 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
5169 /* ok, we can replace it */
5170 block = get_nodes_block(n);
5172 in[0] = get_binop_left(left);
5173 in[1] = new_Const(res);
5175 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
5177 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
5179 return transform_node(irn);
5180 } /* transform_node_shift */
5183 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
5185 * - and, or, xor instead of &
5186 * - Shl, Shr, Shrs, rotl instead of >>
5187 * (with a special case for Or/Xor + Shrs)
5189 static ir_node *transform_node_bitop_shift(ir_node *n) {
5191 ir_node *right = get_binop_right(n);
5192 ir_mode *mode = get_irn_mode(n);
5193 ir_node *bitop_left;
5194 ir_node *bitop_right;
5205 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
5207 if (!is_Const(right))
5210 left = get_binop_left(n);
5211 op_left = get_irn_op(left);
5212 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
5215 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
5216 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
5217 /* TODO: test if sign bit is affectes */
5221 bitop_right = get_binop_right(left);
5222 if (!is_Const(bitop_right))
5225 bitop_left = get_binop_left(left);
5227 block = get_nodes_block(n);
5228 dbgi = get_irn_dbg_info(n);
5229 tv1 = get_Const_tarval(bitop_right);
5230 tv2 = get_Const_tarval(right);
5232 assert(get_tarval_mode(tv1) == mode);
5235 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5236 tv_shift = tarval_shl(tv1, tv2);
5237 } else if(is_Shr(n)) {
5238 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5239 tv_shift = tarval_shr(tv1, tv2);
5240 } else if(is_Shrs(n)) {
5241 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5242 tv_shift = tarval_shrs(tv1, tv2);
5245 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5246 tv_shift = tarval_rotl(tv1, tv2);
5249 assert(get_tarval_mode(tv_shift) == mode);
5250 new_const = new_Const(tv_shift);
5252 if (op_left == op_And) {
5253 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5254 } else if(op_left == op_Or) {
5255 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5257 assert(op_left == op_Eor);
5258 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5266 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5268 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5269 * (also with x >>s c1 when c1>=c2)
5271 static ir_node *transform_node_shl_shr(ir_node *n) {
5273 ir_node *right = get_binop_right(n);
5289 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5291 if (!is_Const(right))
5294 left = get_binop_left(n);
5295 mode = get_irn_mode(n);
5296 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5297 ir_node *shr_right = get_binop_right(left);
5299 if (!is_Const(shr_right))
5302 x = get_binop_left(left);
5303 tv_shr = get_Const_tarval(shr_right);
5304 tv_shl = get_Const_tarval(right);
5306 if (is_Shrs(left)) {
5307 /* shrs variant only allowed if c1 >= c2 */
5308 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5311 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5314 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5316 tv_mask = tarval_shl(tv_mask, tv_shl);
5317 } else if(is_Shr(n) && is_Shl(left)) {
5318 ir_node *shl_right = get_Shl_right(left);
5320 if (!is_Const(shl_right))
5323 x = get_Shl_left(left);
5324 tv_shr = get_Const_tarval(right);
5325 tv_shl = get_Const_tarval(shl_right);
5327 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5328 tv_mask = tarval_shr(tv_mask, tv_shr);
5333 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5334 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5337 assert(tv_mask != tarval_bad);
5338 assert(get_tarval_mode(tv_mask) == mode);
5340 irg = get_irn_irg(n);
5341 block = get_nodes_block(n);
5342 dbgi = get_irn_dbg_info(n);
5344 pnc = tarval_cmp(tv_shl, tv_shr);
5345 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5346 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5347 new_const = new_Const(tv_shift);
5349 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5351 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5354 assert(pnc == pn_Cmp_Gt);
5355 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5356 new_const = new_Const(tv_shift);
5357 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5360 new_const = new_Const(tv_mask);
5361 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5369 static ir_node *transform_node_Shr(ir_node *n) {
5370 ir_node *c, *oldn = n;
5371 ir_node *left = get_Shr_left(n);
5372 ir_node *right = get_Shr_right(n);
5373 ir_mode *mode = get_irn_mode(n);
5375 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5376 n = transform_node_shift(n);
5379 n = transform_node_shl_shr(n);
5381 n = transform_node_bitop_shift(n);
5384 } /* transform_node_Shr */
5389 static ir_node *transform_node_Shrs(ir_node *n) {
5390 ir_node *c, *oldn = n;
5391 ir_node *a = get_Shrs_left(n);
5392 ir_node *b = get_Shrs_right(n);
5393 ir_mode *mode = get_irn_mode(n);
5395 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5396 n = transform_node_shift(n);
5399 n = transform_node_bitop_shift(n);
5402 } /* transform_node_Shrs */
5407 static ir_node *transform_node_Shl(ir_node *n) {
5408 ir_node *c, *oldn = n;
5409 ir_node *a = get_Shl_left(n);
5410 ir_node *b = get_Shl_right(n);
5411 ir_mode *mode = get_irn_mode(n);
5413 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5414 n = transform_node_shift(n);
5417 n = transform_node_shl_shr(n);
5419 n = transform_node_bitop_shift(n);
5422 } /* transform_node_Shl */
5427 static ir_node *transform_node_Rotl(ir_node *n) {
5428 ir_node *c, *oldn = n;
5429 ir_node *a = get_Rotl_left(n);
5430 ir_node *b = get_Rotl_right(n);
5431 ir_mode *mode = get_irn_mode(n);
5433 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5434 n = transform_node_shift(n);
5437 n = transform_node_bitop_shift(n);
5440 } /* transform_node_Rotl */
5445 static ir_node *transform_node_Conv(ir_node *n) {
5446 ir_node *c, *oldn = n;
5447 ir_mode *mode = get_irn_mode(n);
5448 ir_node *a = get_Conv_op(n);
5450 if (mode != mode_b && is_const_Phi(a)) {
5451 /* Do NOT optimize mode_b Conv's, this leads to remaining
5452 * Phib nodes later, because the conv_b_lower operation
5453 * is instantly reverted, when it tries to insert a Convb.
5455 c = apply_conv_on_phi(a, mode);
5457 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5462 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5463 return new_r_Unknown(current_ir_graph, mode);
5466 if (mode_is_reference(mode) &&
5467 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5469 ir_node *l = get_Add_left(a);
5470 ir_node *r = get_Add_right(a);
5471 dbg_info *dbgi = get_irn_dbg_info(a);
5472 ir_node *block = get_nodes_block(n);
5474 ir_node *lop = get_Conv_op(l);
5475 if(get_irn_mode(lop) == mode) {
5476 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5477 n = new_rd_Add(dbgi, block, lop, r, mode);
5482 ir_node *rop = get_Conv_op(r);
5483 if(get_irn_mode(rop) == mode) {
5484 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5485 n = new_rd_Add(dbgi, block, l, rop, mode);
5492 } /* transform_node_Conv */
5495 * Remove dead blocks and nodes in dead blocks
5496 * in keep alive list. We do not generate a new End node.
5498 static ir_node *transform_node_End(ir_node *n) {
5499 int i, j, n_keepalives = get_End_n_keepalives(n);
5502 NEW_ARR_A(ir_node *, in, n_keepalives);
5504 for (i = j = 0; i < n_keepalives; ++i) {
5505 ir_node *ka = get_End_keepalive(n, i);
5507 if (! is_Block_dead(ka)) {
5511 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5513 } else if (is_Bad(ka)) {
5514 /* no need to keep Bad */
5519 if (j != n_keepalives)
5520 set_End_keepalives(n, j, in);
5522 } /* transform_node_End */
5524 bool is_negated_value(ir_node *a, ir_node *b)
5526 if (is_Minus(a) && get_Minus_op(a) == b)
5528 if (is_Minus(b) && get_Minus_op(b) == a)
5530 if (is_Sub(a) && is_Sub(b)) {
5531 ir_node *a_left = get_Sub_left(a);
5532 ir_node *a_right = get_Sub_right(a);
5533 ir_node *b_left = get_Sub_left(b);
5534 ir_node *b_right = get_Sub_right(b);
5536 if (a_left == b_right && a_right == b_left)
5544 * Optimize a Mux into some simpler cases.
5546 static ir_node *transform_node_Mux(ir_node *n) {
5547 ir_node *oldn = n, *sel = get_Mux_sel(n);
5548 ir_mode *mode = get_irn_mode(n);
5549 ir_node *t = get_Mux_true(n);
5550 ir_node *f = get_Mux_false(n);
5551 ir_graph *irg = current_ir_graph;
5553 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5557 ir_node* block = get_nodes_block(n);
5559 ir_node* c1 = get_Mux_sel(t);
5560 ir_node* t1 = get_Mux_true(t);
5561 ir_node* f1 = get_Mux_false(t);
5563 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5564 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5565 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5570 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5571 } else if (f == t1) {
5572 /* Mux(cond0, Mux(cond1, x, y), x) */
5573 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5574 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5575 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5580 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5582 } else if (is_Mux(f)) {
5583 ir_node* block = get_nodes_block(n);
5585 ir_node* c1 = get_Mux_sel(f);
5586 ir_node* t1 = get_Mux_true(f);
5587 ir_node* f1 = get_Mux_false(f);
5589 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5590 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5591 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5596 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5597 } else if (t == f1) {
5598 /* Mux(cond0, x, Mux(cond1, y, x)) */
5599 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5600 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5601 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5606 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5610 /* first normalization step: move a possible zero to the false case */
5612 ir_node *cmp = get_Proj_pred(sel);
5615 if (is_Const(t) && is_Const_null(t)) {
5618 /* Mux(x, 0, y) => Mux(x, y, 0) */
5619 pn_Cmp pnc = get_Proj_proj(sel);
5620 sel = new_r_Proj(get_nodes_block(cmp), cmp, mode_b,
5621 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5622 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, t, f, mode);
5630 /* note: after normalization, false can only happen on default */
5631 if (mode == mode_b) {
5632 dbg_info *dbg = get_irn_dbg_info(n);
5633 ir_node *block = get_nodes_block(n);
5636 tarval *tv_t = get_Const_tarval(t);
5637 if (tv_t == tarval_b_true) {
5639 /* Muxb(sel, true, false) = sel */
5640 assert(get_Const_tarval(f) == tarval_b_false);
5641 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5644 /* Muxb(sel, true, x) = Or(sel, x) */
5645 n = new_rd_Or(dbg, block, sel, f, mode_b);
5646 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5650 } else if (is_Const(f)) {
5651 tarval *tv_f = get_Const_tarval(f);
5652 if (tv_f == tarval_b_true) {
5653 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5654 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5655 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5656 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5659 /* Muxb(sel, x, false) = And(sel, x) */
5660 assert(tv_f == tarval_b_false);
5661 n = new_rd_And(dbg, block, sel, t, mode_b);
5662 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5668 /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */
5669 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5670 tarval *a = get_Const_tarval(t);
5671 tarval *b = get_Const_tarval(f);
5674 if (tarval_is_one(a) && tarval_is_null(b)) {
5675 ir_node *block = get_nodes_block(n);
5676 ir_node *conv = new_r_Conv(block, sel, mode);
5678 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5680 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5681 ir_node *block = get_nodes_block(n);
5682 ir_node *not_ = new_r_Not(block, sel, mode_b);
5683 ir_node *conv = new_r_Conv(block, not_, mode);
5685 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5688 /* TODO: it's not really clear if that helps in general or should be moved
5689 * to backend, especially with the MUX->Conv transformation above */
5690 if (tarval_cmp(a, b) & pn_Cmp_Gt) {
5691 diff = tarval_sub(a, b, NULL);
5694 diff = tarval_sub(b, a, NULL);
5698 if (diff == get_tarval_one(mode)) {
5699 dbg_info *dbg = get_irn_dbg_info(n);
5700 ir_node *block = get_nodes_block(n);
5701 ir_node *t = new_Const(tarval_sub(a, min, NULL));
5702 ir_node *f = new_Const(tarval_sub(b, min, NULL));
5703 n = new_rd_Mux(dbg, block, sel, f, t, mode);
5704 n = new_rd_Add(dbg, block, n, new_Const(min), mode);
5710 ir_node *cmp = get_Proj_pred(sel);
5711 long pn = get_Proj_proj(sel);
5714 * Note: normalization puts the constant on the right side,
5715 * so we check only one case.
5717 * Note further that these optimization work even for floating point
5718 * with NaN's because -NaN == NaN.
5719 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
5723 ir_node *cmp_r = get_Cmp_right(cmp);
5724 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5725 ir_node *block = get_nodes_block(n);
5726 ir_node *cmp_l = get_Cmp_left(cmp);
5728 if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) {
5731 /* NaN's work fine with abs, so it is ok to remove Uo */
5732 long pnc = pn & ~pn_Cmp_Uo;
5734 if ( (cmp_l == t && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt))
5735 || (cmp_l == f && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt)))
5737 /* Mux(a >/>= 0, a, -a) = Mux(a </<= 0, -a, a) ==> Abs(a) */
5738 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5739 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5741 } else if ((cmp_l == t && (pnc == pn_Cmp_Le || pnc == pn_Cmp_Lt))
5742 || (cmp_l == f && (pnc == pn_Cmp_Ge || pnc == pn_Cmp_Gt)))
5744 /* Mux(a </<= 0, a, -a) = Mux(a >/>= 0, -a, a) ==> -Abs(a) */
5745 n = new_rd_Abs(get_irn_dbg_info(n), block, cmp_l, mode);
5746 n = new_rd_Minus(get_irn_dbg_info(n), block, n, mode);
5747 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
5752 if (mode_is_int(mode)) {
5754 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5755 /* Mux((a & b) != 0, c, 0) */
5756 ir_node *and_r = get_And_right(cmp_l);
5759 if (and_r == t && f == cmp_r) {
5760 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5761 if (pn == pn_Cmp_Lg) {
5762 /* Mux((a & 2^C) != 0, 2^C, 0) */
5764 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5766 /* Mux((a & 2^C) == 0, 2^C, 0) */
5767 n = new_rd_Eor(get_irn_dbg_info(n),
5768 block, cmp_l, t, mode);
5769 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5774 if (is_Shl(and_r)) {
5775 ir_node *shl_l = get_Shl_left(and_r);
5776 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5777 if (and_r == t && f == cmp_r) {
5778 if (pn == pn_Cmp_Lg) {
5779 /* (a & (1 << n)) != 0, (1 << n), 0) */
5781 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5783 /* (a & (1 << n)) == 0, (1 << n), 0) */
5784 n = new_rd_Eor(get_irn_dbg_info(n),
5785 block, cmp_l, t, mode);
5786 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5792 and_l = get_And_left(cmp_l);
5793 if (is_Shl(and_l)) {
5794 ir_node *shl_l = get_Shl_left(and_l);
5795 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5796 if (and_l == t && f == cmp_r) {
5797 if (pn == pn_Cmp_Lg) {
5798 /* ((1 << n) & a) != 0, (1 << n), 0) */
5800 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5802 /* ((1 << n) & a) == 0, (1 << n), 0) */
5803 n = new_rd_Eor(get_irn_dbg_info(n),
5804 block, cmp_l, t, mode);
5805 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5818 } /* transform_node_Mux */
5821 * optimize Sync nodes that have other syncs as input we simply add the inputs
5822 * of the other sync to our own inputs
5824 static ir_node *transform_node_Sync(ir_node *n) {
5825 int arity = get_Sync_n_preds(n);
5828 for (i = 0; i < arity;) {
5829 ir_node *pred = get_Sync_pred(n, i);
5833 if (!is_Sync(pred)) {
5841 pred_arity = get_Sync_n_preds(pred);
5842 for (j = 0; j < pred_arity; ++j) {
5843 ir_node *pred_pred = get_Sync_pred(pred, j);
5848 add_irn_n(n, pred_pred);
5852 if (get_Sync_pred(n, k) == pred_pred) break;
5857 /* rehash the sync node */
5858 add_identities(current_ir_graph->value_table, n);
5861 } /* transform_node_Sync */
5864 * optimize a trampoline Call into a direct Call
5866 static ir_node *transform_node_Call(ir_node *call) {
5867 ir_node *callee = get_Call_ptr(call);
5868 ir_node *adr, *mem, *res, *bl, **in;
5869 ir_type *ctp, *mtp, *tp;
5872 int i, n_res, n_param;
5875 if (! is_Proj(callee))
5877 callee = get_Proj_pred(callee);
5878 if (! is_Builtin(callee))
5880 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5883 mem = get_Call_mem(call);
5885 if (skip_Proj(mem) == callee) {
5886 /* memory is routed to the trampoline, skip */
5887 mem = get_Builtin_mem(callee);
5890 /* build a new call type */
5891 mtp = get_Call_type(call);
5892 tdb = get_type_dbg_info(mtp);
5894 n_res = get_method_n_ress(mtp);
5895 n_param = get_method_n_params(mtp);
5896 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5898 for (i = 0; i < n_res; ++i)
5899 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5901 NEW_ARR_A(ir_node *, in, n_param + 1);
5903 /* FIXME: we don't need a new pointer type in every step */
5904 tp = get_irg_frame_type(current_ir_graph);
5905 tp = new_type_pointer(tp);
5906 set_method_param_type(ctp, 0, tp);
5908 in[0] = get_Builtin_param(callee, 2);
5909 for (i = 0; i < n_param; ++i) {
5910 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5911 in[i + 1] = get_Call_param(call, i);
5913 var = get_method_variadicity(mtp);
5914 set_method_variadicity(ctp, var);
5915 if (var == variadicity_variadic) {
5916 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
5918 /* When we resolve a trampoline, the function must be called by a this-call */
5919 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5920 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5922 adr = get_Builtin_param(callee, 1);
5924 db = get_irn_dbg_info(call);
5925 bl = get_nodes_block(call);
5927 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5928 if (get_irn_pinned(call) == op_pin_state_floats)
5929 set_irn_pinned(res, op_pin_state_floats);
5931 } /* transform_node_Call */
5934 * Tries several [inplace] [optimizing] transformations and returns an
5935 * equivalent node. The difference to equivalent_node() is that these
5936 * transformations _do_ generate new nodes, and thus the old node must
5937 * not be freed even if the equivalent node isn't the old one.
5939 static ir_node *transform_node(ir_node *n) {
5943 * Transform_node is the only "optimizing transformation" that might
5944 * return a node with a different opcode. We iterate HERE until fixpoint
5945 * to get the final result.
5949 if (n->op->ops.transform_node != NULL)
5950 n = n->op->ops.transform_node(n);
5951 } while (oldn != n);
5954 } /* transform_node */
5957 * Sets the default transform node operation for an ir_op_ops.
5959 * @param code the opcode for the default operation
5960 * @param ops the operations initialized
5965 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5969 ops->transform_node = transform_node_##a; \
5971 #define CASE_PROJ(a) \
5973 ops->transform_node_Proj = transform_node_Proj_##a; \
5975 #define CASE_PROJ_EX(a) \
5977 ops->transform_node = transform_node_##a; \
5978 ops->transform_node_Proj = transform_node_Proj_##a; \
5987 CASE_PROJ_EX(DivMod);
6022 } /* firm_set_default_transform_node */
6025 /* **************** Common Subexpression Elimination **************** */
6027 /** The size of the hash table used, should estimate the number of nodes
6029 #define N_IR_NODES 512
6031 /** Compares the attributes of two Const nodes. */
6032 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
6033 return (get_Const_tarval(a) != get_Const_tarval(b))
6034 || (get_Const_type(a) != get_Const_type(b));
6035 } /* node_cmp_attr_Const */
6037 /** Compares the attributes of two Proj nodes. */
6038 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
6039 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
6040 } /* node_cmp_attr_Proj */
6042 /** Compares the attributes of two Filter nodes. */
6043 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
6044 return get_Filter_proj(a) != get_Filter_proj(b);
6045 } /* node_cmp_attr_Filter */
6047 /** Compares the attributes of two Alloc nodes. */
6048 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
6049 const alloc_attr *pa = get_irn_alloc_attr(a);
6050 const alloc_attr *pb = get_irn_alloc_attr(b);
6051 return (pa->where != pb->where) || (pa->type != pb->type);
6052 } /* node_cmp_attr_Alloc */
6054 /** Compares the attributes of two Free nodes. */
6055 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
6056 const free_attr *pa = get_irn_free_attr(a);
6057 const free_attr *pb = get_irn_free_attr(b);
6058 return (pa->where != pb->where) || (pa->type != pb->type);
6059 } /* node_cmp_attr_Free */
6061 /** Compares the attributes of two SymConst nodes. */
6062 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
6063 const symconst_attr *pa = get_irn_symconst_attr(a);
6064 const symconst_attr *pb = get_irn_symconst_attr(b);
6065 return (pa->kind != pb->kind)
6066 || (pa->sym.type_p != pb->sym.type_p)
6067 || (pa->tp != pb->tp);
6068 } /* node_cmp_attr_SymConst */
6070 /** Compares the attributes of two Call nodes. */
6071 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
6072 const call_attr *pa = get_irn_call_attr(a);
6073 const call_attr *pb = get_irn_call_attr(b);
6074 return (pa->type != pb->type)
6075 || (pa->tail_call != pb->tail_call);
6076 } /* node_cmp_attr_Call */
6078 /** Compares the attributes of two Sel nodes. */
6079 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
6080 const ir_entity *a_ent = get_Sel_entity(a);
6081 const ir_entity *b_ent = get_Sel_entity(b);
6084 (a_ent->kind != b_ent->kind) ||
6085 (a_ent->name != b_ent->name) ||
6086 (a_ent->owner != b_ent->owner) ||
6087 (a_ent->ld_name != b_ent->ld_name) ||
6088 (a_ent->type != b_ent->type);
6090 /* Matze: inlining of functions can produce 2 entities with same type,
6092 return a_ent != b_ent;
6093 } /* node_cmp_attr_Sel */
6095 /** Compares the attributes of two Phi nodes. */
6096 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
6097 /* we can only enter this function if both nodes have the same number of inputs,
6098 hence it is enough to check if one of them is a Phi0 */
6100 /* check the Phi0 pos attribute */
6101 return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
6104 } /* node_cmp_attr_Phi */
6106 /** Compares the attributes of two Conv nodes. */
6107 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
6108 return get_Conv_strict(a) != get_Conv_strict(b);
6109 } /* node_cmp_attr_Conv */
6111 /** Compares the attributes of two Cast nodes. */
6112 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
6113 return get_Cast_type(a) != get_Cast_type(b);
6114 } /* node_cmp_attr_Cast */
6116 /** Compares the attributes of two Load nodes. */
6117 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
6118 if (get_Load_volatility(a) == volatility_is_volatile ||
6119 get_Load_volatility(b) == volatility_is_volatile)
6120 /* NEVER do CSE on volatile Loads */
6122 /* do not CSE Loads with different alignment. Be conservative. */
6123 if (get_Load_align(a) != get_Load_align(b))
6126 return get_Load_mode(a) != get_Load_mode(b);
6127 } /* node_cmp_attr_Load */
6129 /** Compares the attributes of two Store nodes. */
6130 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
6131 /* do not CSE Stores with different alignment. Be conservative. */
6132 if (get_Store_align(a) != get_Store_align(b))
6135 /* NEVER do CSE on volatile Stores */
6136 return (get_Store_volatility(a) == volatility_is_volatile ||
6137 get_Store_volatility(b) == volatility_is_volatile);
6138 } /* node_cmp_attr_Store */
6140 /** Compares two exception attributes */
6141 static int node_cmp_exception(ir_node *a, ir_node *b) {
6142 const except_attr *ea = get_irn_except_attr(a);
6143 const except_attr *eb = get_irn_except_attr(b);
6145 return ea->pin_state != eb->pin_state;
6148 #define node_cmp_attr_Bound node_cmp_exception
6150 /** Compares the attributes of two Div nodes. */
6151 static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
6152 const divmod_attr *ma = get_irn_divmod_attr(a);
6153 const divmod_attr *mb = get_irn_divmod_attr(b);
6154 return ma->exc.pin_state != mb->exc.pin_state ||
6155 ma->resmode != mb->resmode ||
6156 ma->no_remainder != mb->no_remainder;
6157 } /* node_cmp_attr_Div */
6159 /** Compares the attributes of two DivMod nodes. */
6160 static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
6161 const divmod_attr *ma = get_irn_divmod_attr(a);
6162 const divmod_attr *mb = get_irn_divmod_attr(b);
6163 return ma->exc.pin_state != mb->exc.pin_state ||
6164 ma->resmode != mb->resmode;
6165 } /* node_cmp_attr_DivMod */
6167 /** Compares the attributes of two Mod nodes. */
6168 static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
6169 const divmod_attr *ma = get_irn_divmod_attr(a);
6170 const divmod_attr *mb = get_irn_divmod_attr(b);
6171 return ma->exc.pin_state != mb->exc.pin_state ||
6172 ma->resmode != mb->resmode;
6173 } /* node_cmp_attr_Mod */
6175 /** Compares the attributes of two Quot nodes. */
6176 static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
6177 const divmod_attr *ma = get_irn_divmod_attr(a);
6178 const divmod_attr *mb = get_irn_divmod_attr(b);
6179 return ma->exc.pin_state != mb->exc.pin_state ||
6180 ma->resmode != mb->resmode;
6181 } /* node_cmp_attr_Quot */
6183 /** Compares the attributes of two Confirm nodes. */
6184 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
6185 /* no need to compare the bound, as this is a input */
6186 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
6187 } /* node_cmp_attr_Confirm */
6189 /** Compares the attributes of two Builtin nodes. */
6190 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b) {
6191 const builtin_attr *ma = get_irn_builtin_attr(a);
6192 const builtin_attr *mb = get_irn_builtin_attr(b);
6194 /* no need to compare the type, equal kind means equal type */
6195 return ma->kind != mb->kind;
6196 } /* node_cmp_attr_Builtin */
6198 /** Compares the attributes of two ASM nodes. */
6199 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
6201 const ir_asm_constraint *ca;
6202 const ir_asm_constraint *cb;
6205 if (get_ASM_text(a) != get_ASM_text(b))
6208 /* Should we really check the constraints here? Should be better, but is strange. */
6209 n = get_ASM_n_input_constraints(a);
6210 if (n != get_ASM_n_input_constraints(b))
6213 ca = get_ASM_input_constraints(a);
6214 cb = get_ASM_input_constraints(b);
6215 for (i = 0; i < n; ++i) {
6216 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6220 n = get_ASM_n_output_constraints(a);
6221 if (n != get_ASM_n_output_constraints(b))
6224 ca = get_ASM_output_constraints(a);
6225 cb = get_ASM_output_constraints(b);
6226 for (i = 0; i < n; ++i) {
6227 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6231 n = get_ASM_n_clobbers(a);
6232 if (n != get_ASM_n_clobbers(b))
6235 cla = get_ASM_clobbers(a);
6236 clb = get_ASM_clobbers(b);
6237 for (i = 0; i < n; ++i) {
6238 if (cla[i] != clb[i])
6242 } /* node_cmp_attr_ASM */
6244 /** Compares the inexistent attributes of two Dummy nodes. */
6245 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6253 * Set the default node attribute compare operation for an ir_op_ops.
6255 * @param code the opcode for the default operation
6256 * @param ops the operations initialized
6261 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6265 ops->node_cmp_attr = node_cmp_attr_##a; \
6298 } /* firm_set_default_node_cmp_attr */
6301 * Compare function for two nodes in the value table. Gets two
6302 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6304 int identities_cmp(const void *elt, const void *key) {
6305 ir_node *a = (ir_node *)elt;
6306 ir_node *b = (ir_node *)key;
6309 if (a == b) return 0;
6311 if ((get_irn_op(a) != get_irn_op(b)) ||
6312 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6314 /* compare if a's in and b's in are of equal length */
6315 irn_arity_a = get_irn_intra_arity(a);
6316 if (irn_arity_a != get_irn_intra_arity(b))
6319 if (get_irn_pinned(a) == op_pin_state_pinned) {
6320 /* for pinned nodes, the block inputs must be equal */
6321 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
6323 } else if (! get_opt_global_cse()) {
6324 /* for block-local CSE both nodes must be in the same MacroBlock */
6325 if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b))
6329 /* compare a->in[0..ins] with b->in[0..ins] */
6330 for (i = 0; i < irn_arity_a; ++i) {
6331 ir_node *pred_a = get_irn_intra_n(a, i);
6332 ir_node *pred_b = get_irn_intra_n(b, i);
6333 if (pred_a != pred_b) {
6334 /* if both predecessors are CSE neutral they might be different */
6335 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6341 * here, we already now that the nodes are identical except their
6344 if (a->op->ops.node_cmp_attr)
6345 return a->op->ops.node_cmp_attr(a, b);
6348 } /* identities_cmp */
6351 * Calculate a hash value of a node.
6353 * @param node The IR-node
6355 unsigned ir_node_hash(const ir_node *node) {
6356 return node->op->ops.hash(node);
6357 } /* ir_node_hash */
6360 pset *new_identities(void) {
6361 return new_pset(identities_cmp, N_IR_NODES);
6362 } /* new_identities */
6364 void del_identities(pset *value_table) {
6365 del_pset(value_table);
6366 } /* del_identities */
6368 /* Normalize a node by putting constants (and operands with larger
6369 * node index) on the right (operator side). */
6370 void ir_normalize_node(ir_node *n) {
6371 if (is_op_commutative(get_irn_op(n))) {
6372 ir_node *l = get_binop_left(n);
6373 ir_node *r = get_binop_right(n);
6375 /* For commutative operators perform a OP b == b OP a but keep
6376 * constants on the RIGHT side. This helps greatly in some
6377 * optimizations. Moreover we use the idx number to make the form
6379 if (!operands_are_normalized(l, r)) {
6380 set_binop_left(n, r);
6381 set_binop_right(n, l);
6385 } /* ir_normalize_node */
6388 * Update the nodes after a match in the value table. If both nodes have
6389 * the same MacroBlock but different Blocks, we must ensure that the node
6390 * with the dominating Block (the node that is near to the MacroBlock header
6391 * is stored in the table.
6392 * Because a MacroBlock has only one "non-exception" flow, we don't need
6393 * dominance info here: We known, that one block must dominate the other and
6394 * following the only block input will allow to find it.
6396 static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) {
6397 ir_node *known_blk, *new_block, *block, *mbh;
6399 if (get_opt_global_cse()) {
6400 /* Block inputs are meaning less */
6403 known_blk = get_irn_n(known_irn, -1);
6404 new_block = get_irn_n(new_ir_node, -1);
6405 if (known_blk == new_block) {
6406 /* already in the same block */
6410 * We expect the typical case when we built the graph. In that case, the
6411 * known_irn is already the upper one, so checking this should be faster.
6414 mbh = get_Block_MacroBlock(new_block);
6416 if (block == known_blk) {
6417 /* ok, we have found it: known_block dominates new_block as expected */
6422 * We have reached the MacroBlock header NOT founding
6423 * the known_block. new_block must dominate known_block.
6426 set_irn_n(known_irn, -1, new_block);
6429 assert(get_Block_n_cfgpreds(block) == 1);
6430 block = get_Block_cfgpred_block(block, 0);
6432 } /* update_value_table */
6435 * Return the canonical node computing the same value as n.
6436 * Looks up the node in a hash table, enters it in the table
6437 * if it isn't there yet.
6439 * @param value_table the HashSet containing all nodes in the
6441 * @param n the node to look up
6443 * @return a node that computes the same value as n or n if no such
6444 * node could be found
6446 ir_node *identify_remember(pset *value_table, ir_node *n) {
6449 if (!value_table) return n;
6451 ir_normalize_node(n);
6452 /* lookup or insert in hash table with given hash key. */
6453 nn = pset_insert(value_table, n, ir_node_hash(n));
6456 update_known_irn(nn, n);
6458 /* n is reachable again */
6459 edges_node_revival(nn, get_irn_irg(nn));
6463 } /* identify_remember */
6466 * During construction we set the op_pin_state_pinned flag in the graph right when the
6467 * optimization is performed. The flag turning on procedure global cse could
6468 * be changed between two allocations. This way we are safe.
6470 * @param value_table The value table
6471 * @param n The node to lookup
6473 static inline ir_node *identify_cons(pset *value_table, ir_node *n) {
6476 n = identify_remember(value_table, n);
6477 if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n))
6478 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6480 } /* identify_cons */
6482 /* Add a node to the identities value table. */
6483 void add_identities(pset *value_table, ir_node *node) {
6484 if (get_opt_cse() && is_no_Block(node))
6485 identify_remember(value_table, node);
6486 } /* add_identities */
6488 /* Visit each node in the value table of a graph. */
6489 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
6491 ir_graph *rem = current_ir_graph;
6493 current_ir_graph = irg;
6494 foreach_pset(irg->value_table, node)
6496 current_ir_graph = rem;
6497 } /* visit_all_identities */
6500 * Garbage in, garbage out. If a node has a dead input, i.e., the
6501 * Bad node is input to the node, return the Bad node.
6503 static ir_node *gigo(ir_node *node) {
6505 ir_op *op = get_irn_op(node);
6507 /* remove garbage blocks by looking at control flow that leaves the block
6508 and replacing the control flow by Bad. */
6509 if (get_irn_mode(node) == mode_X) {
6510 ir_node *block = get_nodes_block(skip_Proj(node));
6512 /* Don't optimize nodes in immature blocks. */
6513 if (!get_Block_matured(block))
6515 /* Don't optimize End, may have Bads. */
6516 if (op == op_End) return node;
6518 if (is_Block(block)) {
6519 if (is_Block_dead(block)) {
6520 /* control flow from dead block is dead */
6524 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6525 if (!is_Bad(get_irn_n(block, i)))
6529 ir_graph *irg = get_irn_irg(block);
6530 /* the start block is never dead */
6531 if (block != get_irg_start_block(irg)
6532 && block != get_irg_end_block(irg)) {
6534 * Do NOT kill control flow without setting
6535 * the block to dead of bad things can happen:
6536 * We get a Block that is not reachable be irg_block_walk()
6537 * but can be found by irg_walk()!
6539 set_Block_dead(block);
6546 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6547 blocks predecessors is dead. */
6548 if (op != op_Block && op != op_Phi && op != op_Tuple) {
6549 irn_arity = get_irn_arity(node);
6552 * Beware: we can only read the block of a non-floating node.
6554 if (is_irn_pinned_in_irg(node) &&
6555 is_Block_dead(get_nodes_block(skip_Proj(node))))
6558 for (i = 0; i < irn_arity; i++) {
6559 ir_node *pred = get_irn_n(node, i);
6564 /* Propagating Unknowns here seems to be a bad idea, because
6565 sometimes we need a node as a input and did not want that
6567 However, it might be useful to move this into a later phase
6568 (if you think that optimizing such code is useful). */
6569 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6570 return new_Unknown(get_irn_mode(node));
6575 /* With this code we violate the agreement that local_optimize
6576 only leaves Bads in Block, Phi and Tuple nodes. */
6577 /* If Block has only Bads as predecessors it's garbage. */
6578 /* If Phi has only Bads as predecessors it's garbage. */
6579 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6580 irn_arity = get_irn_arity(node);
6581 for (i = 0; i < irn_arity; i++) {
6582 if (!is_Bad(get_irn_n(node, i))) break;
6584 if (i == irn_arity) node = new_Bad();
6591 * These optimizations deallocate nodes from the obstack.
6592 * It can only be called if it is guaranteed that no other nodes
6593 * reference this one, i.e., right after construction of a node.
6595 * @param n The node to optimize
6597 * current_ir_graph must be set to the graph of the node!
6599 ir_node *optimize_node(ir_node *n) {
6602 ir_opcode iro = get_irn_opcode(n);
6604 /* Always optimize Phi nodes: part of the construction. */
6605 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6607 /* constant expression evaluation / constant folding */
6608 if (get_opt_constant_folding()) {
6609 /* neither constants nor Tuple values can be evaluated */
6610 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6611 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6612 int old_fp_mode = tarval_fp_ops_enabled();
6614 tarval_enable_fp_ops(! (fp_model & fp_no_float_fold));
6616 /* try to evaluate */
6617 tv = computed_value(n);
6618 if (tv != tarval_bad) {
6620 ir_type *old_tp = get_irn_type(n);
6621 int i, arity = get_irn_arity(n);
6625 * Try to recover the type of the new expression.
6627 for (i = 0; i < arity && !old_tp; ++i)
6628 old_tp = get_irn_type(get_irn_n(n, i));
6631 * we MUST copy the node here temporary, because it's still needed
6632 * for DBG_OPT_CSTEVAL
6634 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6635 oldn = alloca(node_size);
6637 memcpy(oldn, n, node_size);
6638 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6640 /* ARG, copy the in array, we need it for statistics */
6641 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6643 /* note the inplace edges module */
6644 edges_node_deleted(n, current_ir_graph);
6646 /* evaluation was successful -- replace the node. */
6647 irg_kill_node(current_ir_graph, n);
6650 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6651 set_Const_type(nw, old_tp);
6652 DBG_OPT_CSTEVAL(oldn, nw);
6653 tarval_enable_fp_ops(old_fp_mode);
6656 tarval_enable_fp_ops(old_fp_mode);
6660 /* remove unnecessary nodes */
6661 if (get_opt_algebraic_simplification() ||
6662 (iro == iro_Phi) || /* always optimize these nodes. */
6664 (iro == iro_Proj) ||
6665 (iro == iro_Block) ) /* Flags tested local. */
6666 n = equivalent_node(n);
6668 /* Common Subexpression Elimination.
6670 * Checks whether n is already available.
6671 * The block input is used to distinguish different subexpressions. Right
6672 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6673 * subexpressions within a block.
6676 n = identify_cons(current_ir_graph->value_table, n);
6679 edges_node_deleted(oldn, current_ir_graph);
6681 /* We found an existing, better node, so we can deallocate the old node. */
6682 irg_kill_node(current_ir_graph, oldn);
6686 /* Some more constant expression evaluation that does not allow to
6688 iro = get_irn_opcode(n);
6689 if (get_opt_algebraic_simplification() ||
6690 (iro == iro_Cond) ||
6691 (iro == iro_Proj)) /* Flags tested local. */
6692 n = transform_node(n);
6694 /* Remove nodes with dead (Bad) input.
6695 Run always for transformation induced Bads. */
6698 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6699 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6701 n = identify_remember(current_ir_graph->value_table, o);
6707 } /* optimize_node */
6711 * These optimizations never deallocate nodes (in place). This can cause dead
6712 * nodes lying on the obstack. Remove these by a dead node elimination,
6713 * i.e., a copying garbage collection.
6715 ir_node *optimize_in_place_2(ir_node *n) {
6718 ir_opcode iro = get_irn_opcode(n);
6720 if (!get_opt_optimize() && !is_Phi(n)) return n;
6722 /* constant expression evaluation / constant folding */
6723 if (get_opt_constant_folding()) {
6724 /* neither constants nor Tuple values can be evaluated */
6725 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6726 unsigned fp_model = get_irg_fp_model(current_ir_graph);
6727 int old_fp_mode = tarval_fp_ops_enabled();
6729 tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0);
6730 /* try to evaluate */
6731 tv = computed_value(n);
6732 if (tv != tarval_bad) {
6733 /* evaluation was successful -- replace the node. */
6734 ir_type *old_tp = get_irn_type(n);
6735 int i, arity = get_irn_arity(n);
6738 * Try to recover the type of the new expression.
6740 for (i = 0; i < arity && !old_tp; ++i)
6741 old_tp = get_irn_type(get_irn_n(n, i));
6745 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
6746 set_Const_type(n, old_tp);
6748 DBG_OPT_CSTEVAL(oldn, n);
6749 tarval_enable_fp_ops(old_fp_mode);
6752 tarval_enable_fp_ops(old_fp_mode);
6756 /* remove unnecessary nodes */
6757 if (get_opt_constant_folding() ||
6758 (iro == iro_Phi) || /* always optimize these nodes. */
6759 (iro == iro_Id) || /* ... */
6760 (iro == iro_Proj) || /* ... */
6761 (iro == iro_Block) ) /* Flags tested local. */
6762 n = equivalent_node(n);
6764 /** common subexpression elimination **/
6765 /* Checks whether n is already available. */
6766 /* The block input is used to distinguish different subexpressions. Right
6767 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6768 subexpressions within a block. */
6769 if (get_opt_cse()) {
6771 n = identify_remember(current_ir_graph->value_table, o);
6776 /* Some more constant expression evaluation. */
6777 iro = get_irn_opcode(n);
6778 if (get_opt_constant_folding() ||
6779 (iro == iro_Cond) ||
6780 (iro == iro_Proj)) /* Flags tested local. */
6781 n = transform_node(n);
6783 /* Remove nodes with dead (Bad) input.
6784 Run always for transformation induced Bads. */
6787 /* Now we can verify the node, as it has no dead inputs any more. */
6790 /* Now we have a legal, useful node. Enter it in hash table for cse.
6791 Blocks should be unique anyways. (Except the successor of start:
6792 is cse with the start block!) */
6793 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6795 n = identify_remember(current_ir_graph->value_table, o);
6801 } /* optimize_in_place_2 */
6804 * Wrapper for external use, set proper status bits after optimization.
6806 ir_node *optimize_in_place(ir_node *n) {
6807 /* Handle graph state */
6808 assert(get_irg_phase_state(current_ir_graph) != phase_building);
6810 if (get_opt_global_cse())
6811 set_irg_pinned(current_ir_graph, op_pin_state_floats);
6812 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
6813 set_irg_outs_inconsistent(current_ir_graph);
6815 /* FIXME: Maybe we could also test whether optimizing the node can
6816 change the control graph. */
6817 set_irg_doms_inconsistent(current_ir_graph);
6818 return optimize_in_place_2(n);
6819 } /* optimize_in_place */
6822 * Calculate a hash value of a Const node.
6824 static unsigned hash_Const(const ir_node *node) {
6827 /* special value for const, as they only differ in their tarval. */
6828 h = HASH_PTR(node->attr.con.tv);
6834 * Calculate a hash value of a SymConst node.
6836 static unsigned hash_SymConst(const ir_node *node) {
6839 /* all others are pointers */
6840 h = HASH_PTR(node->attr.symc.sym.type_p);
6843 } /* hash_SymConst */
6846 * Set the default hash operation in an ir_op_ops.
6848 * @param code the opcode for the default operation
6849 * @param ops the operations initialized
6854 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6858 ops->hash = hash_##a; \
6861 /* hash function already set */
6862 if (ops->hash != NULL)
6869 /* use input/mode default hash if no function was given */
6870 ops->hash = firm_default_hash;
6878 * Sets the default operation for an ir_ops.
6880 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
6881 ops = firm_set_default_hash(code, ops);
6882 ops = firm_set_default_computed_value(code, ops);
6883 ops = firm_set_default_equivalent_node(code, ops);
6884 ops = firm_set_default_transform_node(code, ops);
6885 ops = firm_set_default_node_cmp_attr(code, ops);
6886 ops = firm_set_default_get_type(code, ops);
6887 ops = firm_set_default_get_type_attr(code, ops);
6888 ops = firm_set_default_get_entity_attr(code, ops);
6891 } /* firm_set_default_operations */