2 * Copyright (C) 1995-2011 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
32 #include "irgraph_t.h"
33 #include "iredges_t.h"
39 #include "iroptimize.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
47 #include "opt_polymorphy.h"
52 #include "firm_types.h"
53 #include "bitfiddle.h"
56 /* Make types visible to allow most efficient access */
60 * Returns the tarval of a Const node or tarval_bad for all other nodes.
62 static ir_tarval *default_value_of(const ir_node *n)
65 return get_Const_tarval(n); /* might return tarval_bad */
70 value_of_func value_of_ptr = default_value_of;
72 /* * Set a new value_of function. */
73 void set_value_of_func(value_of_func func)
78 value_of_ptr = default_value_of;
82 * Return the value of a Constant.
84 static ir_tarval *computed_value_Const(const ir_node *n)
86 return get_Const_tarval(n);
87 } /* computed_value_Const */
90 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
92 static ir_tarval *computed_value_SymConst(const ir_node *n)
97 switch (get_SymConst_kind(n)) {
98 case symconst_type_size:
99 type = get_SymConst_type(n);
100 if (get_type_state(type) == layout_fixed)
101 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
103 case symconst_type_align:
104 type = get_SymConst_type(n);
105 if (get_type_state(type) == layout_fixed)
106 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
108 case symconst_ofs_ent:
109 ent = get_SymConst_entity(n);
110 type = get_entity_owner(ent);
111 if (get_type_state(type) == layout_fixed)
112 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
118 } /* computed_value_SymConst */
121 * Return the value of an Add.
123 static ir_tarval *computed_value_Add(const ir_node *n)
125 ir_node *a = get_Add_left(n);
126 ir_node *b = get_Add_right(n);
128 ir_tarval *ta = value_of(a);
129 ir_tarval *tb = value_of(b);
131 if ((ta != tarval_bad) && (tb != tarval_bad))
132 return tarval_add(ta, tb);
135 } /* computed_value_Add */
138 * Return the value of a Sub.
139 * Special case: a - a
141 static ir_tarval *computed_value_Sub(const ir_node *n)
143 ir_mode *mode = get_irn_mode(n);
144 ir_node *a = get_Sub_left(n);
145 ir_node *b = get_Sub_right(n);
150 if (! mode_is_float(mode)) {
153 return get_mode_null(mode);
159 if ((ta != tarval_bad) && (tb != tarval_bad))
160 return tarval_sub(ta, tb, mode);
163 } /* computed_value_Sub */
166 * Return the value of a Carry.
167 * Special : a op 0, 0 op b
169 static ir_tarval *computed_value_Carry(const ir_node *n)
171 ir_node *a = get_binop_left(n);
172 ir_node *b = get_binop_right(n);
173 ir_mode *m = get_irn_mode(n);
174 ir_tarval *ta = value_of(a);
175 ir_tarval *tb = value_of(b);
177 if ((ta != tarval_bad) && (tb != tarval_bad)) {
179 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
181 if (tarval_is_null(ta) || tarval_is_null(tb))
182 return get_mode_null(m);
185 } /* computed_value_Carry */
188 * Return the value of a Borrow.
191 static ir_tarval *computed_value_Borrow(const ir_node *n)
193 ir_node *a = get_binop_left(n);
194 ir_node *b = get_binop_right(n);
195 ir_mode *m = get_irn_mode(n);
196 ir_tarval *ta = value_of(a);
197 ir_tarval *tb = value_of(b);
199 if ((ta != tarval_bad) && (tb != tarval_bad)) {
200 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
201 } else if (tarval_is_null(ta)) {
202 return get_mode_null(m);
205 } /* computed_value_Borrow */
208 * Return the value of an unary Minus.
210 static ir_tarval *computed_value_Minus(const ir_node *n)
212 ir_node *a = get_Minus_op(n);
213 ir_tarval *ta = value_of(a);
215 if (ta != tarval_bad)
216 return tarval_neg(ta);
219 } /* computed_value_Minus */
222 * Return the value of a Mul.
224 static ir_tarval *computed_value_Mul(const ir_node *n)
226 ir_node *a = get_Mul_left(n);
227 ir_node *b = get_Mul_right(n);
228 ir_tarval *ta = value_of(a);
229 ir_tarval *tb = value_of(b);
232 mode = get_irn_mode(n);
233 if (mode != get_irn_mode(a)) {
234 /* n * n = 2n bit multiplication */
235 ta = tarval_convert_to(ta, mode);
236 tb = tarval_convert_to(tb, mode);
239 if (ta != tarval_bad && tb != tarval_bad) {
240 return tarval_mul(ta, tb);
242 /* a * 0 != 0 if a == NaN or a == Inf */
243 if (!mode_is_float(mode)) {
244 /* a*0 = 0 or 0*b = 0 */
245 if (ta == get_mode_null(mode))
247 if (tb == get_mode_null(mode))
252 } /* computed_value_Mul */
255 * Return the value of an And.
256 * Special case: a & 0, 0 & b
258 static ir_tarval *computed_value_And(const ir_node *n)
260 ir_node *a = get_And_left(n);
261 ir_node *b = get_And_right(n);
262 ir_tarval *ta = value_of(a);
263 ir_tarval *tb = value_of(b);
265 if ((ta != tarval_bad) && (tb != tarval_bad)) {
266 return tarval_and (ta, tb);
268 if (tarval_is_null(ta)) return ta;
269 if (tarval_is_null(tb)) return tb;
272 } /* computed_value_And */
275 * Return the value of an Or.
276 * Special case: a | 1...1, 1...1 | b
278 static ir_tarval *computed_value_Or(const ir_node *n)
280 ir_node *a = get_Or_left(n);
281 ir_node *b = get_Or_right(n);
282 ir_tarval *ta = value_of(a);
283 ir_tarval *tb = value_of(b);
285 if ((ta != tarval_bad) && (tb != tarval_bad)) {
286 return tarval_or (ta, tb);
288 if (tarval_is_all_one(ta)) return ta;
289 if (tarval_is_all_one(tb)) return tb;
292 } /* computed_value_Or */
295 * Return the value of an Eor.
297 static ir_tarval *computed_value_Eor(const ir_node *n)
299 ir_node *a = get_Eor_left(n);
300 ir_node *b = get_Eor_right(n);
305 return get_mode_null(get_irn_mode(n));
310 if ((ta != tarval_bad) && (tb != tarval_bad)) {
311 return tarval_eor(ta, tb);
314 } /* computed_value_Eor */
317 * Return the value of a Not.
319 static ir_tarval *computed_value_Not(const ir_node *n)
321 ir_node *a = get_Not_op(n);
322 ir_tarval *ta = value_of(a);
324 if (ta != tarval_bad)
325 return tarval_not(ta);
328 } /* computed_value_Not */
331 * Return the value of a Shl.
333 static ir_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 ir_tarval *ta = value_of(a);
339 ir_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 ir_tarval *computed_value_Shr(const ir_node *n)
352 ir_node *a = get_Shr_left(n);
353 ir_node *b = get_Shr_right(n);
355 ir_tarval *ta = value_of(a);
356 ir_tarval *tb = value_of(b);
358 if ((ta != tarval_bad) && (tb != tarval_bad)) {
359 return tarval_shr(ta, tb);
362 } /* computed_value_Shr */
365 * Return the value of a Shrs.
367 static ir_tarval *computed_value_Shrs(const ir_node *n)
369 ir_node *a = get_Shrs_left(n);
370 ir_node *b = get_Shrs_right(n);
372 ir_tarval *ta = value_of(a);
373 ir_tarval *tb = value_of(b);
375 if ((ta != tarval_bad) && (tb != tarval_bad)) {
376 return tarval_shrs(ta, tb);
379 } /* computed_value_Shrs */
382 * Return the value of a Rotl.
384 static ir_tarval *computed_value_Rotl(const ir_node *n)
386 ir_node *a = get_Rotl_left(n);
387 ir_node *b = get_Rotl_right(n);
389 ir_tarval *ta = value_of(a);
390 ir_tarval *tb = value_of(b);
392 if ((ta != tarval_bad) && (tb != tarval_bad)) {
393 return tarval_rotl(ta, tb);
396 } /* computed_value_Rotl */
399 * Return the value of a Conv.
401 static ir_tarval *computed_value_Conv(const ir_node *n)
403 ir_node *a = get_Conv_op(n);
404 ir_tarval *ta = value_of(a);
406 if (ta != tarval_bad)
407 return tarval_convert_to(ta, get_irn_mode(n));
410 } /* computed_value_Conv */
413 * Calculate the value of a Mux: can be evaluated, if the
414 * sel and the right input are known.
416 static ir_tarval *computed_value_Mux(const ir_node *n)
418 ir_node *sel = get_Mux_sel(n);
419 ir_tarval *ts = value_of(sel);
421 if (ts == get_tarval_b_true()) {
422 ir_node *v = get_Mux_true(n);
425 else if (ts == get_tarval_b_false()) {
426 ir_node *v = get_Mux_false(n);
430 } /* computed_value_Mux */
433 * Calculate the value of a Confirm: can be evaluated,
434 * if it has the form Confirm(x, '=', Const).
436 static ir_tarval *computed_value_Confirm(const ir_node *n)
438 if (get_Confirm_cmp(n) == pn_Cmp_Eq) {
439 ir_tarval *tv = value_of(get_Confirm_bound(n));
440 if (tv != tarval_bad)
443 return value_of(get_Confirm_value(n));
444 } /* computed_value_Confirm */
447 * Return the value of a Proj(Cmp).
449 * This performs a first step of unreachable code elimination.
450 * Proj can not be computed, but folding a Cmp above the Proj here is
451 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
453 * There are several case where we can evaluate a Cmp node, see later.
455 static ir_tarval *computed_value_Proj_Cmp(const ir_node *n)
457 ir_node *cmp = get_Proj_pred(n);
458 ir_node *left = get_Cmp_left(cmp);
459 ir_node *right = get_Cmp_right(cmp);
460 pn_Cmp pn_cmp = get_Proj_pn_cmp(n);
461 ir_mode *mode = get_irn_mode(left);
462 ir_tarval *tv_l, *tv_r;
465 * BEWARE: a == a is NOT always True for floating Point values, as
466 * NaN != NaN is defined, so we must check this here.
468 if (left == right && (!mode_is_float(mode) || pn_cmp == pn_Cmp_Lt || pn_cmp == pn_Cmp_Gt)) {
469 /* This is a trick with the bits used for encoding the Cmp
470 Proj numbers, the following statement is not the same:
471 return new_tarval_from_long(pn_cmp == pn_Cmp_Eq, mode_b) */
472 return new_tarval_from_long(pn_cmp & pn_Cmp_Eq, mode_b);
474 tv_l = value_of(left);
475 tv_r = value_of(right);
477 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
479 * The predecessors of Cmp are target values. We can evaluate
482 pn_Cmp flags = tarval_cmp(tv_l, tv_r);
483 if (flags != pn_Cmp_False) {
484 return new_tarval_from_long (pn_cmp & flags, mode_b);
486 } else if (mode_is_int(mode)) {
487 /* for integer values, we can check against MIN/MAX */
490 if (tv_l == get_mode_min(mode)) {
491 /* MIN <=/> x. This results in true/false. */
492 if (pn_cmp == pn_Cmp_Le)
493 return tarval_b_true;
494 else if (pn_cmp == pn_Cmp_Gt)
495 return tarval_b_false;
496 } else if (tv_r == get_mode_min(mode)) {
497 /* x >=/< MIN. This results in true/false. */
498 if (pn_cmp == pn_Cmp_Ge)
499 return tarval_b_true;
500 else if (pn_cmp == pn_Cmp_Lt)
501 return tarval_b_false;
502 } else if (tv_l == get_mode_max(mode)) {
503 /* MAX >=/< x. This results in true/false. */
504 if (pn_cmp == pn_Cmp_Ge)
505 return tarval_b_true;
506 else if (pn_cmp == pn_Cmp_Lt)
507 return tarval_b_false;
508 } else if (tv_r == get_mode_max(mode)) {
509 /* x <=/> MAX. This results in true/false. */
510 if (pn_cmp == pn_Cmp_Le)
511 return tarval_b_true;
512 else if (pn_cmp == pn_Cmp_Gt)
513 return tarval_b_false;
516 cmp_result = vrp_cmp(left, right);
517 if (cmp_result != pn_Cmp_False) {
518 if (cmp_result == pn_Cmp_Lg) {
519 if (pn_cmp == pn_Cmp_Eq) {
520 return tarval_b_false;
521 } else if (pn_cmp == pn_Cmp_Lg) {
522 return tarval_b_true;
525 return new_tarval_from_long(cmp_result & pn_cmp, mode_b);
528 } else if (mode_is_reference(mode)) {
529 /* pointer compare */
530 ir_node *s_l = skip_Proj(left);
531 ir_node *s_r = skip_Proj(right);
533 if ((is_Alloc(s_l) && tarval_is_null(tv_r)) ||
534 (tarval_is_null(tv_l) && is_Alloc(s_r))) {
536 * The predecessors are Allocs and (void*)(0) constants. In Firm Allocs never
537 * return NULL, they raise an exception. Therefore we can predict
540 return new_tarval_from_long(pn_cmp & pn_Cmp_Lg, mode_b);
543 return computed_value_Cmp_Confirm(cmp, left, right, pn_cmp);
544 } /* computed_value_Proj_Cmp */
547 * Calculate the value of an integer Div.
548 * Special case: 0 / b
550 static ir_tarval *do_computed_value_Div(const ir_node *div)
552 const ir_node *a = get_Div_left(div);
553 const ir_node *b = get_Div_right(div);
554 const ir_mode *mode = get_Div_resmode(div);
555 ir_tarval *ta = value_of(a);
557 const ir_node *dummy;
559 /* cannot optimize 0 / b = 0 because of NaN */
560 if (!mode_is_float(mode)) {
561 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
562 return ta; /* 0 / b == 0 if b != 0 */
565 if (ta != tarval_bad && tb != tarval_bad)
566 return tarval_div(ta, tb);
568 } /* do_computed_value_Div */
571 * Calculate the value of an integer Mod of two nodes.
572 * Special case: a % 1
574 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
576 ir_tarval *ta = value_of(a);
577 ir_tarval *tb = value_of(b);
579 /* Compute a % 1 or c1 % c2 */
580 if (tarval_is_one(tb))
581 return get_mode_null(get_irn_mode(a));
582 if (ta != tarval_bad && tb != tarval_bad)
583 return tarval_mod(ta, tb);
585 } /* do_computed_value_Mod */
588 * Return the value of a Proj(Div).
590 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
592 long proj_nr = get_Proj_proj(n);
593 if (proj_nr != pn_Div_res)
596 return do_computed_value_Div(get_Proj_pred(n));
597 } /* computed_value_Proj_Div */
600 * Return the value of a Proj(Mod).
602 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
604 long proj_nr = get_Proj_proj(n);
606 if (proj_nr == pn_Mod_res) {
607 const ir_node *mod = get_Proj_pred(n);
608 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
611 } /* computed_value_Proj_Mod */
614 * Return the value of a Proj.
616 static ir_tarval *computed_value_Proj(const ir_node *proj)
618 ir_node *n = get_Proj_pred(proj);
620 if (n->op->ops.computed_value_Proj != NULL)
621 return n->op->ops.computed_value_Proj(proj);
623 } /* computed_value_Proj */
626 * If the parameter n can be computed, return its value, else tarval_bad.
627 * Performs constant folding.
629 * @param n The node this should be evaluated
631 ir_tarval *computed_value(const ir_node *n)
633 vrp_attr *vrp = vrp_get_info(n);
634 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == pn_Cmp_Eq) {
635 return vrp->bits_set;
637 if (n->op->ops.computed_value)
638 return n->op->ops.computed_value(n);
640 } /* computed_value */
643 * Set the default computed_value evaluator in an ir_op_ops.
645 * @param code the opcode for the default operation
646 * @param ops the operations initialized
651 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
655 ops->computed_value = computed_value_##a; \
657 #define CASE_PROJ(a) \
659 ops->computed_value_Proj = computed_value_Proj_##a; \
694 } /* firm_set_default_computed_value */
697 * Returns a equivalent block for another block.
698 * If the block has only one predecessor, this is
699 * the equivalent one. If the only predecessor of a block is
700 * the block itself, this is a dead block.
702 * If both predecessors of a block are the branches of a binary
703 * Cond, the equivalent block is Cond's block.
705 * If all predecessors of a block are bad or lies in a dead
706 * block, the current block is dead as well.
708 * Note, that blocks are NEVER turned into Bad's, instead
709 * the dead_block flag is set. So, never test for is_Bad(block),
710 * always use is_dead_Block(block).
712 static ir_node *equivalent_node_Block(ir_node *n)
718 /* don't optimize dead or labeled blocks */
719 if (is_Block_dead(n) || has_Block_entity(n))
722 n_preds = get_Block_n_cfgpreds(n);
724 /* The Block constructor does not call optimize, but mature_immBlock()
725 calls the optimization. */
726 assert(get_Block_matured(n));
728 irg = get_irn_irg(n);
730 /* Straightening: a single entry Block following a single exit Block
731 can be merged, if it is not the Start block. */
732 /* !!! Beware, all Phi-nodes of n must have been optimized away.
733 This should be true, as the block is matured before optimize is called.
734 But what about Phi-cycles with the Phi0/Id that could not be resolved?
735 Remaining Phi nodes are just Ids. */
737 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
740 ir_node *predblock = get_nodes_block(pred);
741 if (predblock == oldn) {
742 /* Jmp jumps into the block it is in -- deal self cycle. */
743 n = set_Block_dead(n);
744 DBG_OPT_DEAD_BLOCK(oldn, n);
747 DBG_OPT_STG(oldn, n);
749 } else if (is_Cond(pred)) {
750 ir_node *predblock = get_nodes_block(pred);
751 if (predblock == oldn) {
752 /* Jmp jumps into the block it is in -- deal self cycle. */
753 n = set_Block_dead(n);
754 DBG_OPT_DEAD_BLOCK(oldn, n);
757 } else if (n_preds == 2) {
758 /* Test whether Cond jumps twice to this block
759 * The more general case which more than 2 predecessors is handles
760 * in optimize_cf(), we handle only this special case for speed here.
762 ir_node *a = get_Block_cfgpred(n, 0);
763 ir_node *b = get_Block_cfgpred(n, 1);
765 if (is_Proj(a) && is_Proj(b)) {
766 ir_node *cond = get_Proj_pred(a);
768 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
769 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
770 /* Also a single entry Block following a single exit Block. Phis have
771 twice the same operand and will be optimized away. */
772 n = get_nodes_block(cond);
773 DBG_OPT_IFSIM1(oldn, a, b, n);
776 } else if (get_opt_unreachable_code() &&
777 (n != get_irg_start_block(irg)) &&
778 (n != get_irg_end_block(irg))) {
781 /* If all inputs are dead, this block is dead too, except if it is
782 the start or end block. This is one step of unreachable code
784 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
785 ir_node *pred = get_Block_cfgpred(n, i);
788 if (is_Bad(pred)) continue;
789 pred_blk = get_nodes_block(skip_Proj(pred));
791 if (is_Block_dead(pred_blk)) continue;
794 /* really found a living input */
799 n = set_Block_dead(n);
800 DBG_OPT_DEAD_BLOCK(oldn, n);
805 } /* equivalent_node_Block */
808 * Returns a equivalent node for a Jmp, a Bad :-)
809 * Of course this only happens if the Block of the Jmp is dead.
811 static ir_node *equivalent_node_Jmp(ir_node *n)
815 /* unreachable code elimination */
816 if (is_Block_dead(get_nodes_block(n))) {
817 ir_graph *irg = get_irn_irg(n);
818 n = get_irg_bad(irg);
819 DBG_OPT_DEAD_BLOCK(oldn, n);
822 } /* equivalent_node_Jmp */
824 /** Raise is handled in the same way as Jmp. */
825 #define equivalent_node_Raise equivalent_node_Jmp
828 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
829 See transform_node_Proj_Cond(). */
832 * Optimize operations that are commutative and have neutral 0,
833 * so a op 0 = 0 op a = a.
835 static ir_node *equivalent_node_neutral_zero(ir_node *n)
839 ir_node *a = get_binop_left(n);
840 ir_node *b = get_binop_right(n);
845 /* After running compute_node there is only one constant predecessor.
846 Find this predecessors value and remember the other node: */
847 if ((tv = value_of(a)) != tarval_bad) {
849 } else if ((tv = value_of(b)) != tarval_bad) {
854 /* If this predecessors constant value is zero, the operation is
855 * unnecessary. Remove it.
857 * Beware: If n is a Add, the mode of on and n might be different
858 * which happens in this rare construction: NULL + 3.
859 * Then, a Conv would be needed which we cannot include here.
861 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
864 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
868 } /* equivalent_node_neutral_zero */
871 * Eor is commutative and has neutral 0.
873 static ir_node *equivalent_node_Eor(ir_node *n)
879 n = equivalent_node_neutral_zero(n);
880 if (n != oldn) return n;
883 b = get_Eor_right(n);
886 ir_node *aa = get_Eor_left(a);
887 ir_node *ab = get_Eor_right(a);
890 /* (a ^ b) ^ a -> b */
892 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
894 } else if (ab == b) {
895 /* (a ^ b) ^ b -> a */
897 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
902 ir_node *ba = get_Eor_left(b);
903 ir_node *bb = get_Eor_right(b);
906 /* a ^ (a ^ b) -> b */
908 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
910 } else if (bb == a) {
911 /* a ^ (b ^ a) -> b */
913 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
921 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
923 * The second one looks strange, but this construct
924 * is used heavily in the LCC sources :-).
926 * Beware: The Mode of an Add may be different than the mode of its
927 * predecessors, so we could not return a predecessors in all cases.
929 static ir_node *equivalent_node_Add(ir_node *n)
932 ir_node *left, *right;
933 ir_mode *mode = get_irn_mode(n);
935 n = equivalent_node_neutral_zero(n);
939 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
940 if (mode_is_float(mode)) {
941 ir_graph *irg = get_irn_irg(n);
942 if (get_irg_fp_model(irg) & fp_strict_algebraic)
946 left = get_Add_left(n);
947 right = get_Add_right(n);
950 if (get_Sub_right(left) == right) {
953 n = get_Sub_left(left);
954 if (mode == get_irn_mode(n)) {
955 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
961 if (get_Sub_right(right) == left) {
964 n = get_Sub_left(right);
965 if (mode == get_irn_mode(n)) {
966 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
972 } /* equivalent_node_Add */
975 * optimize operations that are not commutative but have neutral 0 on left,
978 static ir_node *equivalent_node_left_zero(ir_node *n)
982 ir_node *a = get_binop_left(n);
983 ir_node *b = get_binop_right(n);
984 ir_tarval *tb = value_of(b);
986 if (tarval_is_null(tb)) {
989 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
992 } /* equivalent_node_left_zero */
994 #define equivalent_node_Shl equivalent_node_left_zero
995 #define equivalent_node_Shr equivalent_node_left_zero
996 #define equivalent_node_Shrs equivalent_node_left_zero
997 #define equivalent_node_Rotl equivalent_node_left_zero
1000 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1002 * The second one looks strange, but this construct
1003 * is used heavily in the LCC sources :-).
1005 * Beware: The Mode of a Sub may be different than the mode of its
1006 * predecessors, so we could not return a predecessors in all cases.
1008 static ir_node *equivalent_node_Sub(ir_node *n)
1012 ir_mode *mode = get_irn_mode(n);
1015 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1016 if (mode_is_float(mode)) {
1017 ir_graph *irg = get_irn_irg(n);
1018 if (get_irg_fp_model(irg) & fp_strict_algebraic)
1022 b = get_Sub_right(n);
1025 /* Beware: modes might be different */
1026 if (tarval_is_null(tb)) {
1027 ir_node *a = get_Sub_left(n);
1028 if (mode == get_irn_mode(a)) {
1031 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1035 } /* equivalent_node_Sub */
1039 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1042 * -(-a) == a, but might overflow two times.
1043 * We handle it anyway here but the better way would be a
1044 * flag. This would be needed for Pascal for instance.
1046 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1049 ir_node *pred = get_unop_op(n);
1051 /* optimize symmetric unop */
1052 if (get_irn_op(pred) == get_irn_op(n)) {
1053 n = get_unop_op(pred);
1054 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1057 } /* equivalent_node_idempotent_unop */
1059 /** Optimize Not(Not(x)) == x. */
1060 #define equivalent_node_Not equivalent_node_idempotent_unop
1062 /** -(-x) == x ??? Is this possible or can --x raise an
1063 out of bounds exception if min =! max? */
1064 #define equivalent_node_Minus equivalent_node_idempotent_unop
1067 * Optimize a * 1 = 1 * a = a.
1069 static ir_node *equivalent_node_Mul(ir_node *n)
1072 ir_node *a = get_Mul_left(n);
1074 /* we can handle here only the n * n = n bit cases */
1075 if (get_irn_mode(n) == get_irn_mode(a)) {
1076 ir_node *b = get_Mul_right(n);
1080 * Mul is commutative and has again an other neutral element.
1081 * Constants are place right, so check this case first.
1084 if (tarval_is_one(tv)) {
1086 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1089 if (tarval_is_one(tv)) {
1091 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1096 } /* equivalent_node_Mul */
1099 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1101 static ir_node *equivalent_node_Or(ir_node *n)
1105 ir_node *a = get_Or_left(n);
1106 ir_node *b = get_Or_right(n);
1110 n = a; /* Or has it's own neutral element */
1111 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1114 /* constants are cormalized to right, check this site first */
1116 if (tarval_is_null(tv)) {
1118 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1122 if (tarval_is_null(tv)) {
1124 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1129 } /* equivalent_node_Or */
1132 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1134 static ir_node *equivalent_node_And(ir_node *n)
1138 ir_node *a = get_And_left(n);
1139 ir_node *b = get_And_right(n);
1143 n = a; /* And has it's own neutral element */
1144 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1147 /* constants are normalized to right, check this site first */
1149 if (tarval_is_all_one(tv)) {
1151 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1154 if (tv != get_tarval_bad()) {
1155 ir_mode *mode = get_irn_mode(n);
1156 if (!mode_is_signed(mode) && is_Conv(a)) {
1157 ir_node *convop = get_Conv_op(a);
1158 ir_mode *convopmode = get_irn_mode(convop);
1159 if (!mode_is_signed(convopmode)) {
1160 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1161 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1163 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1170 if (tarval_is_all_one(tv)) {
1172 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1176 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1179 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1184 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1187 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1192 } /* equivalent_node_And */
1195 * Try to remove useless Conv's:
1197 static ir_node *equivalent_node_Conv(ir_node *n)
1200 ir_node *a = get_Conv_op(n);
1202 ir_mode *n_mode = get_irn_mode(n);
1203 ir_mode *a_mode = get_irn_mode(a);
1206 if (n_mode == a_mode) { /* No Conv necessary */
1207 if (get_Conv_strict(n)) {
1210 /* neither Minus nor Confirm change the precision,
1211 so we can "look-through" */
1214 p = get_Minus_op(p);
1215 } else if (is_Confirm(p)) {
1216 p = get_Confirm_value(p);
1222 if (is_Conv(p) && get_Conv_strict(p)) {
1223 /* we known already, that a_mode == n_mode, and neither
1224 Minus change the mode, so the second Conv
1226 assert(get_irn_mode(p) == n_mode);
1228 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1232 ir_node *pred = get_Proj_pred(p);
1233 if (is_Load(pred)) {
1234 /* Loads always return with the exact precision of n_mode */
1235 assert(get_Load_mode(pred) == n_mode);
1237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1240 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1241 pred = get_Proj_pred(pred);
1242 if (is_Start(pred)) {
1243 /* Arguments always return with the exact precision,
1244 as strictConv's are place before Call -- if the
1245 caller was compiled with the same setting.
1246 Otherwise, the semantics is probably still right. */
1247 assert(get_irn_mode(p) == n_mode);
1249 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1255 /* special case: the immediate predecessor is also a Conv */
1256 if (! get_Conv_strict(a)) {
1257 /* first one is not strict, kick it */
1259 a_mode = get_irn_mode(a);
1263 /* else both are strict conv, second is superfluous */
1265 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1270 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1273 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1274 ir_node *b = get_Conv_op(a);
1275 ir_mode *b_mode = get_irn_mode(b);
1277 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1278 /* both are strict conv */
1279 if (smaller_mode(a_mode, n_mode)) {
1280 /* both are strict, but the first is smaller, so
1281 the second cannot remove more precision, remove the
1283 set_Conv_strict(n, 0);
1286 if (n_mode == b_mode) {
1287 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1288 if (n_mode == mode_b) {
1289 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1290 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1292 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1293 if (values_in_mode(b_mode, a_mode)) {
1294 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1295 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1300 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1301 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1302 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1303 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1305 if (float_mantissa >= int_mantissa) {
1307 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1312 if (smaller_mode(b_mode, a_mode)) {
1313 if (get_Conv_strict(n))
1314 set_Conv_strict(b, 1);
1315 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1316 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1323 } /* equivalent_node_Conv */
1326 * - fold Phi-nodes, iff they have only one predecessor except
1329 static ir_node *equivalent_node_Phi(ir_node *n)
1335 ir_node *first_val = NULL; /* to shutup gcc */
1337 if (!get_opt_optimize() &&
1338 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1341 n_preds = get_Phi_n_preds(n);
1343 block = get_nodes_block(n);
1345 if (is_Block_dead(block)) {
1346 ir_graph *irg = get_irn_irg(n);
1347 return get_irg_bad(irg);
1350 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1352 /* Find first non-self-referencing input */
1353 for (i = 0; i < n_preds; ++i) {
1354 first_val = get_Phi_pred(n, i);
1355 if ( (first_val != n) /* not self pointer */
1357 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1358 * predecessors. Then, Phi nodes in dead code might be removed, causing
1359 * nodes pointing to themself (Add's for instance).
1360 * This is really bad and causes endless recursions in several
1361 * code pathes, so we do NOT optimize such a code.
1362 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1363 * (and bad Phi predecessors), so live code is optimized later.
1365 && (! is_Bad(get_Block_cfgpred(block, i)))
1367 ) { /* value not dead */
1368 break; /* then found first value. */
1373 ir_graph *irg = get_irn_irg(n);
1374 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1375 return get_irg_bad(irg);
1378 /* search for rest of inputs, determine if any of these
1379 are non-self-referencing */
1380 while (++i < n_preds) {
1381 ir_node *scnd_val = get_Phi_pred(n, i);
1382 if ( (scnd_val != n)
1383 && (scnd_val != first_val)
1386 && (! is_Bad(get_Block_cfgpred(block, i)))
1393 if (i >= n_preds && !is_Dummy(first_val)) {
1394 /* Fold, if no multiple distinct non-self-referencing inputs */
1396 DBG_OPT_PHI(oldn, n);
1399 } /* equivalent_node_Phi */
1402 * Several optimizations:
1403 * - fold Sync-nodes, iff they have only one predecessor except
1406 static ir_node *equivalent_node_Sync(ir_node *n)
1408 int arity = get_Sync_n_preds(n);
1411 for (i = 0; i < arity;) {
1412 ir_node *pred = get_Sync_pred(n, i);
1415 /* Remove Bad predecessors */
1422 /* Remove duplicate predecessors */
1428 if (get_Sync_pred(n, j) == pred) {
1437 ir_graph *irg = get_irn_irg(n);
1438 return get_irg_bad(irg);
1440 if (arity == 1) return get_Sync_pred(n, 0);
1442 } /* equivalent_node_Sync */
1445 * Optimize Proj(Tuple).
1447 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1449 ir_node *oldn = proj;
1450 ir_node *tuple = get_Proj_pred(proj);
1452 /* Remove the Tuple/Proj combination. */
1453 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1454 DBG_OPT_TUPLE(oldn, tuple, proj);
1457 } /* equivalent_node_Proj_Tuple */
1460 * Optimize a / 1 = a.
1462 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1464 ir_node *oldn = proj;
1465 ir_node *div = get_Proj_pred(proj);
1466 ir_node *b = get_Div_right(div);
1467 ir_tarval *tb = value_of(b);
1469 /* Div is not commutative. */
1470 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1471 switch (get_Proj_proj(proj)) {
1473 proj = get_Div_mem(div);
1474 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1478 proj = get_Div_left(div);
1479 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1483 /* we cannot replace the exception Proj's here, this is done in
1484 transform_node_Proj_Div() */
1489 } /* equivalent_node_Proj_Div */
1492 * Optimize CopyB(mem, x, x) into a Nop.
1494 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1496 ir_node *oldn = proj;
1497 ir_node *copyb = get_Proj_pred(proj);
1498 ir_node *a = get_CopyB_dst(copyb);
1499 ir_node *b = get_CopyB_src(copyb);
1502 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1503 switch (get_Proj_proj(proj)) {
1505 proj = get_CopyB_mem(copyb);
1506 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1509 case pn_CopyB_X_except: {
1510 ir_graph *irg = get_irn_irg(proj);
1511 DBG_OPT_EXC_REM(proj);
1512 proj = get_irg_bad(irg);
1518 } /* equivalent_node_Proj_CopyB */
1521 * Optimize Bounds(idx, idx, upper) into idx.
1523 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1525 ir_node *oldn = proj;
1526 ir_node *bound = get_Proj_pred(proj);
1527 ir_node *idx = get_Bound_index(bound);
1528 ir_node *pred = skip_Proj(idx);
1531 if (idx == get_Bound_lower(bound))
1533 else if (is_Bound(pred)) {
1535 * idx was Bounds checked previously, it is still valid if
1536 * lower <= pred_lower && pred_upper <= upper.
1538 ir_node *lower = get_Bound_lower(bound);
1539 ir_node *upper = get_Bound_upper(bound);
1540 if (get_Bound_lower(pred) == lower &&
1541 get_Bound_upper(pred) == upper) {
1543 * One could expect that we simply return the previous
1544 * Bound here. However, this would be wrong, as we could
1545 * add an exception Proj to a new location then.
1546 * So, we must turn in into a tuple.
1552 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1553 switch (get_Proj_proj(proj)) {
1555 DBG_OPT_EXC_REM(proj);
1556 proj = get_Bound_mem(bound);
1558 case pn_Bound_X_except: {
1559 ir_graph *irg = get_irn_irg(proj);
1560 DBG_OPT_EXC_REM(proj);
1561 proj = get_irg_bad(irg);
1566 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1569 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1574 } /* equivalent_node_Proj_Bound */
1577 * Optimize an Exception Proj(Load) with a non-null address.
1579 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1581 if (get_opt_ldst_only_null_ptr_exceptions()) {
1582 if (get_irn_mode(proj) == mode_X) {
1583 ir_node *load = get_Proj_pred(proj);
1585 /* get the Load address */
1586 const ir_node *addr = get_Load_ptr(load);
1587 const ir_node *confirm;
1589 if (value_not_null(addr, &confirm)) {
1590 if (get_Proj_proj(proj) == pn_Load_X_except) {
1591 ir_graph *irg = get_irn_irg(proj);
1592 DBG_OPT_EXC_REM(proj);
1593 return get_irg_bad(irg);
1599 } /* equivalent_node_Proj_Load */
1602 * Optimize an Exception Proj(Store) with a non-null address.
1604 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1606 if (get_opt_ldst_only_null_ptr_exceptions()) {
1607 if (get_irn_mode(proj) == mode_X) {
1608 ir_node *store = get_Proj_pred(proj);
1610 /* get the load/store address */
1611 const ir_node *addr = get_Store_ptr(store);
1612 const ir_node *confirm;
1614 if (value_not_null(addr, &confirm)) {
1615 if (get_Proj_proj(proj) == pn_Store_X_except) {
1616 ir_graph *irg = get_irn_irg(proj);
1617 DBG_OPT_EXC_REM(proj);
1618 return get_irg_bad(irg);
1624 } /* equivalent_node_Proj_Store */
1627 * Does all optimizations on nodes that must be done on it's Proj's
1628 * because of creating new nodes.
1630 static ir_node *equivalent_node_Proj(ir_node *proj)
1632 ir_node *n = get_Proj_pred(proj);
1634 if (get_irn_mode(proj) == mode_X) {
1635 if (is_Block_dead(get_nodes_block(n))) {
1636 /* Remove dead control flow -- early gigo(). */
1637 ir_graph *irg = get_irn_irg(proj);
1638 return get_irg_bad(irg);
1641 if (n->op->ops.equivalent_node_Proj)
1642 return n->op->ops.equivalent_node_Proj(proj);
1644 } /* equivalent_node_Proj */
1649 static ir_node *equivalent_node_Id(ir_node *n)
1657 DBG_OPT_ID(oldn, n);
1659 } /* equivalent_node_Id */
1664 static ir_node *equivalent_node_Mux(ir_node *n)
1666 ir_node *oldn = n, *sel = get_Mux_sel(n);
1668 ir_tarval *ts = value_of(sel);
1670 /* Mux(true, f, t) == t */
1671 if (ts == tarval_b_true) {
1672 n = get_Mux_true(n);
1673 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1676 /* Mux(false, f, t) == f */
1677 if (ts == tarval_b_false) {
1678 n = get_Mux_false(n);
1679 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1682 n_t = get_Mux_true(n);
1683 n_f = get_Mux_false(n);
1685 /* Mux(v, x, T) == x */
1686 if (is_Unknown(n_f)) {
1688 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1691 /* Mux(v, T, x) == x */
1692 if (is_Unknown(n_t)) {
1694 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1698 /* Mux(v, x, x) == x */
1701 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1704 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1705 ir_node *cmp = get_Proj_pred(sel);
1706 long proj_nr = get_Proj_proj(sel);
1707 ir_node *f = get_Mux_false(n);
1708 ir_node *t = get_Mux_true(n);
1711 * Note further that these optimization work even for floating point
1712 * with NaN's because -NaN == NaN.
1713 * However, if +0 and -0 is handled differently, we cannot use the first one.
1716 ir_node *const cmp_l = get_Cmp_left(cmp);
1717 ir_node *const cmp_r = get_Cmp_right(cmp);
1721 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1722 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1724 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1731 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1732 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1734 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1741 * Note: normalization puts the constant on the right side,
1742 * so we check only one case.
1744 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1745 /* Mux(t CMP 0, X, t) */
1746 if (is_Minus(f) && get_Minus_op(f) == t) {
1747 /* Mux(t CMP 0, -t, t) */
1748 if (proj_nr == pn_Cmp_Eq) {
1749 /* Mux(t == 0, -t, t) ==> -t */
1751 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1752 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1753 /* Mux(t != 0, -t, t) ==> t */
1755 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1762 } /* equivalent_node_Mux */
1765 * Remove Confirm nodes if setting is on.
1766 * Replace Confirms(x, '=', Constlike) by Constlike.
1768 static ir_node *equivalent_node_Confirm(ir_node *n)
1770 ir_node *pred = get_Confirm_value(n);
1771 pn_Cmp pnc = get_Confirm_cmp(n);
1773 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1775 * rare case: two identical Confirms one after another,
1776 * replace the second one with the first.
1779 pred = get_Confirm_value(n);
1780 pnc = get_Confirm_cmp(n);
1786 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1787 * perform no actual computation, as, e.g., the Id nodes. It does not create
1788 * new nodes. It is therefore safe to free n if the node returned is not n.
1789 * If a node returns a Tuple we can not just skip it. If the size of the
1790 * in array fits, we transform n into a tuple (e.g., Div).
1792 ir_node *equivalent_node(ir_node *n)
1794 if (n->op->ops.equivalent_node)
1795 return n->op->ops.equivalent_node(n);
1797 } /* equivalent_node */
1800 * Sets the default equivalent node operation for an ir_op_ops.
1802 * @param code the opcode for the default operation
1803 * @param ops the operations initialized
1808 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1812 ops->equivalent_node = equivalent_node_##a; \
1814 #define CASE_PROJ(a) \
1816 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1856 } /* firm_set_default_equivalent_node */
1859 * Returns non-zero if a node is a Phi node
1860 * with all predecessors constant.
1862 static int is_const_Phi(ir_node *n)
1866 if (! is_Phi(n) || get_irn_arity(n) == 0)
1868 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1869 if (! is_Const(get_irn_n(n, i)))
1873 } /* is_const_Phi */
1875 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1876 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1879 * in reality eval_func should be tarval (*eval_func)() but incomplete
1880 * declarations are bad style and generate noisy warnings
1882 typedef void (*eval_func)(void);
1885 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1887 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1889 if (eval == (eval_func) tarval_sub) {
1890 tarval_sub_type func = (tarval_sub_type)eval;
1892 return func(a, b, mode);
1894 tarval_binop_type func = (tarval_binop_type)eval;
1901 * Apply an evaluator on a binop with a constant operators (and one Phi).
1903 * @param phi the Phi node
1904 * @param other the other operand
1905 * @param eval an evaluator function
1906 * @param mode the mode of the result, may be different from the mode of the Phi!
1907 * @param left if non-zero, other is the left operand, else the right
1909 * @return a new Phi node if the conversion was successful, NULL else
1911 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1917 int i, n = get_irn_arity(phi);
1919 NEW_ARR_A(void *, res, n);
1921 for (i = 0; i < n; ++i) {
1922 pred = get_irn_n(phi, i);
1923 tv = get_Const_tarval(pred);
1924 tv = do_eval(eval, other, tv, mode);
1926 if (tv == tarval_bad) {
1927 /* folding failed, bad */
1933 for (i = 0; i < n; ++i) {
1934 pred = get_irn_n(phi, i);
1935 tv = get_Const_tarval(pred);
1936 tv = do_eval(eval, tv, other, mode);
1938 if (tv == tarval_bad) {
1939 /* folding failed, bad */
1945 irg = get_irn_irg(phi);
1946 for (i = 0; i < n; ++i) {
1947 pred = get_irn_n(phi, i);
1948 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1950 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1951 } /* apply_binop_on_phi */
1954 * Apply an evaluator on a binop with two constant Phi.
1956 * @param a the left Phi node
1957 * @param b the right Phi node
1958 * @param eval an evaluator function
1959 * @param mode the mode of the result, may be different from the mode of the Phi!
1961 * @return a new Phi node if the conversion was successful, NULL else
1963 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1965 ir_tarval *tv_l, *tv_r, *tv;
1971 if (get_nodes_block(a) != get_nodes_block(b))
1974 n = get_irn_arity(a);
1975 NEW_ARR_A(void *, res, n);
1977 for (i = 0; i < n; ++i) {
1978 pred = get_irn_n(a, i);
1979 tv_l = get_Const_tarval(pred);
1980 pred = get_irn_n(b, i);
1981 tv_r = get_Const_tarval(pred);
1982 tv = do_eval(eval, tv_l, tv_r, mode);
1984 if (tv == tarval_bad) {
1985 /* folding failed, bad */
1990 irg = get_irn_irg(a);
1991 for (i = 0; i < n; ++i) {
1992 pred = get_irn_n(a, i);
1993 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1995 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1996 } /* apply_binop_on_2_phis */
1999 * Apply an evaluator on a unop with a constant operator (a Phi).
2001 * @param phi the Phi node
2002 * @param eval an evaluator function
2004 * @return a new Phi node if the conversion was successful, NULL else
2006 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
2013 int i, n = get_irn_arity(phi);
2015 NEW_ARR_A(void *, res, n);
2016 for (i = 0; i < n; ++i) {
2017 pred = get_irn_n(phi, i);
2018 tv = get_Const_tarval(pred);
2021 if (tv == tarval_bad) {
2022 /* folding failed, bad */
2027 mode = get_irn_mode(phi);
2028 irg = get_irn_irg(phi);
2029 for (i = 0; i < n; ++i) {
2030 pred = get_irn_n(phi, i);
2031 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2033 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2034 } /* apply_unop_on_phi */
2037 * Apply a conversion on a constant operator (a Phi).
2039 * @param phi the Phi node
2041 * @return a new Phi node if the conversion was successful, NULL else
2043 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
2049 int i, n = get_irn_arity(phi);
2051 NEW_ARR_A(void *, res, n);
2052 for (i = 0; i < n; ++i) {
2053 pred = get_irn_n(phi, i);
2054 tv = get_Const_tarval(pred);
2055 tv = tarval_convert_to(tv, mode);
2057 if (tv == tarval_bad) {
2058 /* folding failed, bad */
2063 irg = get_irn_irg(phi);
2064 for (i = 0; i < n; ++i) {
2065 pred = get_irn_n(phi, i);
2066 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2068 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2069 } /* apply_conv_on_phi */
2072 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2073 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2074 * If possible, remove the Conv's.
2076 static ir_node *transform_node_AddSub(ir_node *n)
2078 ir_mode *mode = get_irn_mode(n);
2080 if (mode_is_reference(mode)) {
2081 ir_node *left = get_binop_left(n);
2082 ir_node *right = get_binop_right(n);
2083 unsigned ref_bits = get_mode_size_bits(mode);
2085 if (is_Conv(left)) {
2086 ir_mode *lmode = get_irn_mode(left);
2087 unsigned bits = get_mode_size_bits(lmode);
2089 if (ref_bits == bits &&
2090 mode_is_int(lmode) &&
2091 get_mode_arithmetic(lmode) == irma_twos_complement) {
2092 ir_node *pre = get_Conv_op(left);
2093 ir_mode *pre_mode = get_irn_mode(pre);
2095 if (mode_is_int(pre_mode) &&
2096 get_mode_size_bits(pre_mode) == bits &&
2097 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2098 /* ok, this conv just changes to sign, moreover the calculation
2099 * is done with same number of bits as our address mode, so
2100 * we can ignore the conv as address calculation can be viewed
2101 * as either signed or unsigned
2103 set_binop_left(n, pre);
2108 if (is_Conv(right)) {
2109 ir_mode *rmode = get_irn_mode(right);
2110 unsigned bits = get_mode_size_bits(rmode);
2112 if (ref_bits == bits &&
2113 mode_is_int(rmode) &&
2114 get_mode_arithmetic(rmode) == irma_twos_complement) {
2115 ir_node *pre = get_Conv_op(right);
2116 ir_mode *pre_mode = get_irn_mode(pre);
2118 if (mode_is_int(pre_mode) &&
2119 get_mode_size_bits(pre_mode) == bits &&
2120 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2121 /* ok, this conv just changes to sign, moreover the calculation
2122 * is done with same number of bits as our address mode, so
2123 * we can ignore the conv as address calculation can be viewed
2124 * as either signed or unsigned
2126 set_binop_right(n, pre);
2131 /* let address arithmetic use unsigned modes */
2132 if (is_Const(right)) {
2133 ir_mode *rmode = get_irn_mode(right);
2135 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2136 /* convert a AddP(P, *s) into AddP(P, *u) */
2137 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2139 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2140 set_binop_right(n, pre);
2146 } /* transform_node_AddSub */
2148 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2151 if (is_Const(b) && is_const_Phi(a)) { \
2152 /* check for Op(Phi, Const) */ \
2153 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2155 else if (is_Const(a) && is_const_Phi(b)) { \
2156 /* check for Op(Const, Phi) */ \
2157 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2159 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2160 /* check for Op(Phi, Phi) */ \
2161 c = apply_binop_on_2_phis(a, b, eval, mode); \
2164 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2169 #define HANDLE_UNOP_PHI(eval, a, c) \
2172 if (is_const_Phi(a)) { \
2173 /* check for Op(Phi) */ \
2174 c = apply_unop_on_phi(a, eval); \
2176 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2183 * Do the AddSub optimization, then Transform
2184 * Constant folding on Phi
2185 * Add(a,a) -> Mul(a, 2)
2186 * Add(Mul(a, x), a) -> Mul(a, x+1)
2187 * if the mode is integer or float.
2188 * Transform Add(a,-b) into Sub(a,b).
2189 * Reassociation might fold this further.
2191 static ir_node *transform_node_Add(ir_node *n)
2194 ir_node *a, *b, *c, *oldn = n;
2195 vrp_attr *a_vrp, *b_vrp;
2197 n = transform_node_AddSub(n);
2199 a = get_Add_left(n);
2200 b = get_Add_right(n);
2202 mode = get_irn_mode(n);
2204 if (mode_is_reference(mode)) {
2205 ir_mode *lmode = get_irn_mode(a);
2207 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2208 /* an Add(a, NULL) is a hidden Conv */
2209 dbg_info *dbg = get_irn_dbg_info(n);
2210 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2214 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2216 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2217 if (mode_is_float(mode)) {
2218 ir_graph *irg = get_irn_irg(n);
2219 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2223 if (mode_is_num(mode)) {
2224 ir_graph *irg = get_irn_irg(n);
2225 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2226 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2227 && a == b && mode_is_int(mode)) {
2228 ir_node *block = get_nodes_block(n);
2231 get_irn_dbg_info(n),
2234 new_r_Const_long(irg, mode, 2),
2236 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2241 get_irn_dbg_info(n),
2246 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2251 get_irn_dbg_info(n),
2256 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2259 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2260 /* Here we rely on constants be on the RIGHT side */
2262 ir_node *op = get_Not_op(a);
2264 if (is_Const(b) && is_Const_one(b)) {
2266 ir_node *blk = get_nodes_block(n);
2267 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2268 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2273 n = new_r_Const(irg, get_mode_minus_one(mode));
2274 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2279 ir_node *op = get_Not_op(b);
2283 n = new_r_Const(irg, get_mode_minus_one(mode));
2284 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2291 a_vrp = vrp_get_info(a);
2292 b_vrp = vrp_get_info(b);
2294 if (a_vrp && b_vrp) {
2295 ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2297 if (tarval_is_null(c)) {
2298 dbg_info *dbgi = get_irn_dbg_info(n);
2299 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2303 } /* transform_node_Add */
2306 * returns -cnst or NULL if impossible
2308 static ir_node *const_negate(ir_node *cnst)
2310 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2311 dbg_info *dbgi = get_irn_dbg_info(cnst);
2312 ir_graph *irg = get_irn_irg(cnst);
2313 if (tv == tarval_bad) return NULL;
2314 return new_rd_Const(dbgi, irg, tv);
2318 * Do the AddSub optimization, then Transform
2319 * Constant folding on Phi
2320 * Sub(0,a) -> Minus(a)
2321 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2322 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2323 * Sub(Add(a, x), x) -> a
2324 * Sub(x, Add(x, a)) -> -a
2325 * Sub(x, Const) -> Add(x, -Const)
2327 static ir_node *transform_node_Sub(ir_node *n)
2333 n = transform_node_AddSub(n);
2335 a = get_Sub_left(n);
2336 b = get_Sub_right(n);
2338 mode = get_irn_mode(n);
2340 if (mode_is_int(mode)) {
2341 ir_mode *lmode = get_irn_mode(a);
2343 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2344 /* a Sub(a, NULL) is a hidden Conv */
2345 dbg_info *dbg = get_irn_dbg_info(n);
2346 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2347 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2351 if (mode == lmode &&
2352 get_mode_arithmetic(mode) == irma_twos_complement &&
2354 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2356 dbg_info *dbg = get_irn_dbg_info(n);
2357 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2358 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2364 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2366 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2367 if (mode_is_float(mode)) {
2368 ir_graph *irg = get_irn_irg(n);
2369 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2373 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2374 /* a - C -> a + (-C) */
2375 ir_node *cnst = const_negate(b);
2377 ir_node *block = get_nodes_block(n);
2378 dbg_info *dbgi = get_irn_dbg_info(n);
2380 n = new_rd_Add(dbgi, block, a, cnst, mode);
2381 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2386 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2387 dbg_info *dbg = get_irn_dbg_info(n);
2388 ir_node *block = get_nodes_block(n);
2389 ir_node *left = get_Minus_op(a);
2390 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2392 n = new_rd_Minus(dbg, block, add, mode);
2393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2395 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2396 dbg_info *dbg = get_irn_dbg_info(n);
2397 ir_node *block = get_nodes_block(n);
2398 ir_node *right = get_Minus_op(b);
2400 n = new_rd_Add(dbg, block, a, right, mode);
2401 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2403 } else if (is_Sub(b)) {
2404 /* a - (b - c) -> a + (c - b)
2405 * -> (a - b) + c iff (b - c) is a pointer */
2406 dbg_info *s_dbg = get_irn_dbg_info(b);
2407 ir_node *s_block = get_nodes_block(b);
2408 ir_node *s_left = get_Sub_left(b);
2409 ir_node *s_right = get_Sub_right(b);
2410 ir_mode *s_mode = get_irn_mode(b);
2411 if (mode_is_reference(s_mode)) {
2412 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2413 dbg_info *a_dbg = get_irn_dbg_info(n);
2414 ir_node *a_block = get_nodes_block(n);
2417 s_right = new_r_Conv(a_block, s_right, mode);
2418 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2420 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2421 dbg_info *a_dbg = get_irn_dbg_info(n);
2422 ir_node *a_block = get_nodes_block(n);
2424 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2426 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2428 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2429 ir_node *m_right = get_Mul_right(b);
2430 if (is_Const(m_right)) {
2431 ir_node *cnst2 = const_negate(m_right);
2432 if (cnst2 != NULL) {
2433 dbg_info *m_dbg = get_irn_dbg_info(b);
2434 ir_node *m_block = get_nodes_block(b);
2435 ir_node *m_left = get_Mul_left(b);
2436 ir_mode *m_mode = get_irn_mode(b);
2437 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2438 dbg_info *a_dbg = get_irn_dbg_info(n);
2439 ir_node *a_block = get_nodes_block(n);
2441 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2442 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2448 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2449 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2451 get_irn_dbg_info(n),
2455 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2459 if (mode_wrap_around(mode)) {
2460 ir_node *left = get_Add_left(a);
2461 ir_node *right = get_Add_right(a);
2463 /* FIXME: Does the Conv's work only for two complement or generally? */
2465 if (mode != get_irn_mode(right)) {
2466 /* This Sub is an effective Cast */
2467 right = new_r_Conv(get_nodes_block(n), right, mode);
2470 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2472 } else if (right == b) {
2473 if (mode != get_irn_mode(left)) {
2474 /* This Sub is an effective Cast */
2475 left = new_r_Conv(get_nodes_block(n), left, mode);
2478 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2484 if (mode_wrap_around(mode)) {
2485 ir_node *left = get_Add_left(b);
2486 ir_node *right = get_Add_right(b);
2488 /* FIXME: Does the Conv's work only for two complement or generally? */
2490 ir_mode *r_mode = get_irn_mode(right);
2492 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2493 if (mode != r_mode) {
2494 /* This Sub is an effective Cast */
2495 n = new_r_Conv(get_nodes_block(n), n, mode);
2497 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2499 } else if (right == a) {
2500 ir_mode *l_mode = get_irn_mode(left);
2502 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2503 if (mode != l_mode) {
2504 /* This Sub is an effective Cast */
2505 n = new_r_Conv(get_nodes_block(n), n, mode);
2507 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2512 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2513 ir_mode *mode = get_irn_mode(a);
2515 if (mode == get_irn_mode(b)) {
2517 ir_node *op_a = get_Conv_op(a);
2518 ir_node *op_b = get_Conv_op(b);
2520 /* check if it's allowed to skip the conv */
2521 ma = get_irn_mode(op_a);
2522 mb = get_irn_mode(op_b);
2524 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2525 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2528 set_Sub_right(n, b);
2534 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2535 if (!is_reassoc_running() && is_Mul(a)) {
2536 ir_node *ma = get_Mul_left(a);
2537 ir_node *mb = get_Mul_right(a);
2540 ir_node *blk = get_nodes_block(n);
2541 ir_graph *irg = get_irn_irg(n);
2543 get_irn_dbg_info(n),
2547 get_irn_dbg_info(n),
2550 new_r_Const(irg, get_mode_one(mode)),
2553 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2555 } else if (mb == b) {
2556 ir_node *blk = get_nodes_block(n);
2557 ir_graph *irg = get_irn_irg(n);
2559 get_irn_dbg_info(n),
2563 get_irn_dbg_info(n),
2566 new_r_Const(irg, get_mode_one(mode)),
2569 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2573 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2574 ir_node *x = get_Sub_left(a);
2575 ir_node *y = get_Sub_right(a);
2576 ir_node *blk = get_nodes_block(n);
2577 ir_mode *m_b = get_irn_mode(b);
2578 ir_mode *m_y = get_irn_mode(y);
2582 /* Determine the right mode for the Add. */
2585 else if (mode_is_reference(m_b))
2587 else if (mode_is_reference(m_y))
2591 * Both modes are different but none is reference,
2592 * happens for instance in SubP(SubP(P, Iu), Is).
2593 * We have two possibilities here: Cast or ignore.
2594 * Currently we ignore this case.
2599 add = new_r_Add(blk, y, b, add_mode);
2601 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2602 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2606 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2607 if (is_Const(a) && is_Not(b)) {
2608 /* c - ~X = X + (c+1) */
2609 ir_tarval *tv = get_Const_tarval(a);
2611 tv = tarval_add(tv, get_mode_one(mode));
2612 if (tv != tarval_bad) {
2613 ir_node *blk = get_nodes_block(n);
2614 ir_graph *irg = get_irn_irg(n);
2615 ir_node *c = new_r_Const(irg, tv);
2616 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2617 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2623 } /* transform_node_Sub */
2626 * Several transformation done on n*n=2n bits mul.
2627 * These transformations must be done here because new nodes may be produced.
2629 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2632 ir_node *a = get_Mul_left(n);
2633 ir_node *b = get_Mul_right(n);
2634 ir_tarval *ta = value_of(a);
2635 ir_tarval *tb = value_of(b);
2636 ir_mode *smode = get_irn_mode(a);
2638 if (ta == get_mode_one(smode)) {
2639 /* (L)1 * (L)b = (L)b */
2640 ir_node *blk = get_nodes_block(n);
2641 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2642 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2645 else if (ta == get_mode_minus_one(smode)) {
2646 /* (L)-1 * (L)b = (L)b */
2647 ir_node *blk = get_nodes_block(n);
2648 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2649 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2650 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2653 if (tb == get_mode_one(smode)) {
2654 /* (L)a * (L)1 = (L)a */
2655 ir_node *blk = get_irn_n(a, -1);
2656 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2657 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2660 else if (tb == get_mode_minus_one(smode)) {
2661 /* (L)a * (L)-1 = (L)-a */
2662 ir_node *blk = get_nodes_block(n);
2663 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2664 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2665 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2672 * Transform Mul(a,-1) into -a.
2673 * Do constant evaluation of Phi nodes.
2674 * Do architecture dependent optimizations on Mul nodes
2676 static ir_node *transform_node_Mul(ir_node *n)
2678 ir_node *c, *oldn = n;
2679 ir_mode *mode = get_irn_mode(n);
2680 ir_node *a = get_Mul_left(n);
2681 ir_node *b = get_Mul_right(n);
2683 if (is_Bad(a) || is_Bad(b))
2686 if (mode != get_irn_mode(a))
2687 return transform_node_Mul2n(n, mode);
2689 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2691 if (mode_is_signed(mode)) {
2694 if (value_of(a) == get_mode_minus_one(mode))
2696 else if (value_of(b) == get_mode_minus_one(mode))
2699 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2700 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2705 if (is_Const(b)) { /* (-a) * const -> a * -const */
2706 ir_node *cnst = const_negate(b);
2708 dbg_info *dbgi = get_irn_dbg_info(n);
2709 ir_node *block = get_nodes_block(n);
2710 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2711 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2714 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2715 dbg_info *dbgi = get_irn_dbg_info(n);
2716 ir_node *block = get_nodes_block(n);
2717 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2718 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2720 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2721 ir_node *sub_l = get_Sub_left(b);
2722 ir_node *sub_r = get_Sub_right(b);
2723 dbg_info *dbgi = get_irn_dbg_info(n);
2724 ir_node *block = get_nodes_block(n);
2725 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2726 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2727 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2730 } else if (is_Minus(b)) {
2731 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2732 ir_node *sub_l = get_Sub_left(a);
2733 ir_node *sub_r = get_Sub_right(a);
2734 dbg_info *dbgi = get_irn_dbg_info(n);
2735 ir_node *block = get_nodes_block(n);
2736 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2737 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2738 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2741 } else if (is_Shl(a)) {
2742 ir_node *const shl_l = get_Shl_left(a);
2743 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2744 /* (1 << x) * b -> b << x */
2745 dbg_info *const dbgi = get_irn_dbg_info(n);
2746 ir_node *const block = get_nodes_block(n);
2747 ir_node *const shl_r = get_Shl_right(a);
2748 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2749 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2752 } else if (is_Shl(b)) {
2753 ir_node *const shl_l = get_Shl_left(b);
2754 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2755 /* a * (1 << x) -> a << x */
2756 dbg_info *const dbgi = get_irn_dbg_info(n);
2757 ir_node *const block = get_nodes_block(n);
2758 ir_node *const shl_r = get_Shl_right(b);
2759 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2760 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2764 if (get_mode_arithmetic(mode) == irma_ieee754) {
2766 ir_tarval *tv = get_Const_tarval(a);
2767 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2768 && !tarval_is_negative(tv)) {
2769 /* 2.0 * b = b + b */
2770 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2771 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2775 else if (is_Const(b)) {
2776 ir_tarval *tv = get_Const_tarval(b);
2777 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2778 && !tarval_is_negative(tv)) {
2779 /* a * 2.0 = a + a */
2780 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2781 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2786 return arch_dep_replace_mul_with_shifts(n);
2787 } /* transform_node_Mul */
2790 * Transform a Div Node.
2792 static ir_node *transform_node_Div(ir_node *n)
2794 ir_mode *mode = get_Div_resmode(n);
2795 ir_node *a = get_Div_left(n);
2796 ir_node *b = get_Div_right(n);
2798 const ir_node *dummy;
2800 if (mode_is_int(mode)) {
2801 if (is_Const(b) && is_const_Phi(a)) {
2802 /* check for Div(Phi, Const) */
2803 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2805 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2808 } else if (is_Const(a) && is_const_Phi(b)) {
2809 /* check for Div(Const, Phi) */
2810 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2812 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2815 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2816 /* check for Div(Phi, Phi) */
2817 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2819 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2824 if (a == b && value_not_zero(a, &dummy)) {
2825 ir_graph *irg = get_irn_irg(n);
2826 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2827 value = new_r_Const(irg, get_mode_one(mode));
2828 DBG_OPT_CSTEVAL(n, value);
2831 if (mode_is_signed(mode) && is_Const(b)) {
2832 ir_tarval *tv = get_Const_tarval(b);
2834 if (tv == get_mode_minus_one(mode)) {
2836 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2837 DBG_OPT_CSTEVAL(n, value);
2841 /* Try architecture dependent optimization */
2842 value = arch_dep_replace_div_by_const(n);
2845 assert(mode_is_float(mode));
2847 /* Optimize x/c to x*(1/c) */
2848 if (get_mode_arithmetic(mode) == irma_ieee754) {
2849 ir_tarval *tv = value_of(b);
2851 if (tv != tarval_bad) {
2852 int rem = tarval_fp_ops_enabled();
2855 * Floating point constant folding might be disabled here to
2857 * However, as we check for exact result, doing it is safe.
2860 tarval_enable_fp_ops(1);
2861 tv = tarval_div(get_mode_one(mode), tv);
2862 tarval_enable_fp_ops(rem);
2864 /* Do the transformation if the result is either exact or we are
2865 not using strict rules. */
2866 if (tv != tarval_bad &&
2867 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2868 ir_node *block = get_nodes_block(n);
2869 ir_graph *irg = get_irn_irg(block);
2870 ir_node *c = new_r_Const(irg, tv);
2871 dbg_info *dbgi = get_irn_dbg_info(n);
2872 value = new_rd_Mul(dbgi, block, a, c, mode);
2885 /* Turn Div into a tuple (mem, jmp, bad, value) */
2886 mem = get_Div_mem(n);
2887 blk = get_nodes_block(n);
2888 irg = get_irn_irg(blk);
2890 /* skip a potential Pin */
2891 mem = skip_Pin(mem);
2892 turn_into_tuple(n, pn_Div_max);
2893 set_Tuple_pred(n, pn_Div_M, mem);
2894 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2895 set_Tuple_pred(n, pn_Div_X_except, get_irg_bad(irg));
2896 set_Tuple_pred(n, pn_Div_res, value);
2899 } /* transform_node_Div */
2902 * Transform a Mod node.
2904 static ir_node *transform_node_Mod(ir_node *n)
2906 ir_mode *mode = get_Mod_resmode(n);
2907 ir_node *a = get_Mod_left(n);
2908 ir_node *b = get_Mod_right(n);
2913 if (is_Const(b) && is_const_Phi(a)) {
2914 /* check for Div(Phi, Const) */
2915 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2917 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2921 else if (is_Const(a) && is_const_Phi(b)) {
2922 /* check for Div(Const, Phi) */
2923 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2925 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2929 else if (is_const_Phi(a) && is_const_Phi(b)) {
2930 /* check for Div(Phi, Phi) */
2931 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2933 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2940 irg = get_irn_irg(n);
2941 if (tv != tarval_bad) {
2942 value = new_r_Const(irg, tv);
2944 DBG_OPT_CSTEVAL(n, value);
2947 ir_node *a = get_Mod_left(n);
2948 ir_node *b = get_Mod_right(n);
2949 const ir_node *dummy;
2951 if (a == b && value_not_zero(a, &dummy)) {
2952 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2953 value = new_r_Const(irg, get_mode_null(mode));
2954 DBG_OPT_CSTEVAL(n, value);
2957 if (mode_is_signed(mode) && is_Const(b)) {
2958 ir_tarval *tv = get_Const_tarval(b);
2960 if (tv == get_mode_minus_one(mode)) {
2962 value = new_r_Const(irg, get_mode_null(mode));
2963 DBG_OPT_CSTEVAL(n, value);
2967 /* Try architecture dependent optimization */
2968 value = arch_dep_replace_mod_by_const(n);
2977 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2978 mem = get_Mod_mem(n);
2979 blk = get_nodes_block(n);
2980 irg = get_irn_irg(blk);
2982 /* skip a potential Pin */
2983 mem = skip_Pin(mem);
2984 turn_into_tuple(n, pn_Mod_max);
2985 set_Tuple_pred(n, pn_Mod_M, mem);
2986 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2987 set_Tuple_pred(n, pn_Mod_X_except, get_irg_bad(irg));
2988 set_Tuple_pred(n, pn_Mod_res, value);
2991 } /* transform_node_Mod */
2994 * Optimize -a CMP -b into b CMP a.
2995 * This works only for for modes where unary Minus
2997 * Note that two-complement integers can Overflow
2998 * so it will NOT work.
3000 * For == and != can be handled in Proj(Cmp)
3002 static ir_node *transform_node_Cmp(ir_node *n)
3005 ir_node *left = get_Cmp_left(n);
3006 ir_node *right = get_Cmp_right(n);
3008 if (is_Minus(left) && is_Minus(right) &&
3009 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3010 ir_node *const new_left = get_Minus_op(right);
3011 ir_node *const new_right = get_Minus_op(left);
3012 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3013 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3016 } /* transform_node_Cmp */
3020 * Transform a Cond node.
3022 * Replace the Cond by a Jmp if it branches on a constant
3025 static ir_node *transform_node_Cond(ir_node *n)
3028 ir_node *a = get_Cond_selector(n);
3029 ir_tarval *ta = value_of(a);
3030 ir_graph *irg = get_irn_irg(n);
3033 /* we need block info which is not available in floating irgs */
3034 if (get_irg_pinned(irg) == op_pin_state_floats)
3037 if ((ta != tarval_bad) &&
3038 (get_irn_mode(a) == mode_b) &&
3039 (get_opt_unreachable_code())) {
3040 /* It's a boolean Cond, branching on a boolean constant.
3041 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3042 ir_node *blk = get_nodes_block(n);
3043 jmp = new_r_Jmp(blk);
3044 turn_into_tuple(n, pn_Cond_max);
3045 if (ta == tarval_b_true) {
3046 set_Tuple_pred(n, pn_Cond_false, get_irg_bad(irg));
3047 set_Tuple_pred(n, pn_Cond_true, jmp);
3049 set_Tuple_pred(n, pn_Cond_false, jmp);
3050 set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg));
3052 /* We might generate an endless loop, so keep it alive. */
3053 add_End_keepalive(get_irg_end(irg), blk);
3056 } /* transform_node_Cond */
3059 * Prototype of a recursive transform function
3060 * for bitwise distributive transformations.
3062 typedef ir_node* (*recursive_transform)(ir_node *n);
3065 * makes use of distributive laws for and, or, eor
3066 * and(a OP c, b OP c) -> and(a, b) OP c
3067 * note, might return a different op than n
3069 static ir_node *transform_bitwise_distributive(ir_node *n,
3070 recursive_transform trans_func)
3073 ir_node *a = get_binop_left(n);
3074 ir_node *b = get_binop_right(n);
3075 ir_op *op = get_irn_op(a);
3076 ir_op *op_root = get_irn_op(n);
3078 if (op != get_irn_op(b))
3081 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3082 if (op == op_Conv) {
3083 ir_node *a_op = get_Conv_op(a);
3084 ir_node *b_op = get_Conv_op(b);
3085 ir_mode *a_mode = get_irn_mode(a_op);
3086 ir_mode *b_mode = get_irn_mode(b_op);
3087 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3088 ir_node *blk = get_nodes_block(n);
3091 set_binop_left(n, a_op);
3092 set_binop_right(n, b_op);
3093 set_irn_mode(n, a_mode);
3095 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3097 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3103 /* nothing to gain here */
3107 if (op == op_Shrs || op == op_Shr || op == op_Shl
3108 || op == op_And || op == op_Or || op == op_Eor) {
3109 ir_node *a_left = get_binop_left(a);
3110 ir_node *a_right = get_binop_right(a);
3111 ir_node *b_left = get_binop_left(b);
3112 ir_node *b_right = get_binop_right(b);
3114 ir_node *op1 = NULL;
3115 ir_node *op2 = NULL;
3117 if (is_op_commutative(op)) {
3118 if (a_left == b_left) {
3122 } else if (a_left == b_right) {
3126 } else if (a_right == b_left) {
3132 if (a_right == b_right) {
3139 /* (a sop c) & (b sop c) => (a & b) sop c */
3140 ir_node *blk = get_nodes_block(n);
3142 ir_node *new_n = exact_copy(n);
3143 set_binop_left(new_n, op1);
3144 set_binop_right(new_n, op2);
3145 new_n = trans_func(new_n);
3147 if (op_root == op_Eor && op == op_Or) {
3148 dbg_info *dbgi = get_irn_dbg_info(n);
3149 ir_mode *mode = get_irn_mode(c);
3151 c = new_rd_Not(dbgi, blk, c, mode);
3152 n = new_rd_And(dbgi, blk, new_n, c, mode);
3155 set_nodes_block(n, blk);
3156 set_binop_left(n, new_n);
3157 set_binop_right(n, c);
3161 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3172 static ir_node *transform_node_And(ir_node *n)
3174 ir_node *c, *oldn = n;
3175 ir_node *a = get_And_left(n);
3176 ir_node *b = get_And_right(n);
3178 vrp_attr *a_vrp, *b_vrp;
3180 mode = get_irn_mode(n);
3181 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3183 /* we can evaluate 2 Projs of the same Cmp */
3184 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3185 ir_node *pred_a = get_Proj_pred(a);
3186 ir_node *pred_b = get_Proj_pred(b);
3187 if (pred_a == pred_b) {
3188 dbg_info *dbgi = get_irn_dbg_info(n);
3189 pn_Cmp pn_a = get_Proj_pn_cmp(a);
3190 pn_Cmp pn_b = get_Proj_pn_cmp(b);
3191 /* yes, we can simply calculate with pncs */
3192 pn_Cmp new_pnc = pn_a & pn_b;
3194 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3199 ir_node *op = get_Not_op(b);
3201 ir_node *ba = get_And_left(op);
3202 ir_node *bb = get_And_right(op);
3204 /* it's enough to test the following cases due to normalization! */
3205 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3206 /* (a|b) & ~(a&b) = a^b */
3207 ir_node *block = get_nodes_block(n);
3209 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3210 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3218 ir_node *op = get_Not_op(a);
3220 ir_node *aa = get_And_left(op);
3221 ir_node *ab = get_And_right(op);
3223 /* it's enough to test the following cases due to normalization! */
3224 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3225 /* (a|b) & ~(a&b) = a^b */
3226 ir_node *block = get_nodes_block(n);
3228 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3229 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3236 ir_node *al = get_Eor_left(a);
3237 ir_node *ar = get_Eor_right(a);
3240 /* (b ^ a) & b -> ~a & b */
3241 dbg_info *dbg = get_irn_dbg_info(n);
3242 ir_node *block = get_nodes_block(n);
3244 ar = new_rd_Not(dbg, block, ar, mode);
3245 n = new_rd_And(dbg, block, ar, b, mode);
3246 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3250 /* (a ^ b) & b -> ~a & b */
3251 dbg_info *dbg = get_irn_dbg_info(n);
3252 ir_node *block = get_nodes_block(n);
3254 al = new_rd_Not(dbg, block, al, mode);
3255 n = new_rd_And(dbg, block, al, b, mode);
3256 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3261 ir_node *bl = get_Eor_left(b);
3262 ir_node *br = get_Eor_right(b);
3265 /* a & (a ^ b) -> a & ~b */
3266 dbg_info *dbg = get_irn_dbg_info(n);
3267 ir_node *block = get_nodes_block(n);
3269 br = new_rd_Not(dbg, block, br, mode);
3270 n = new_rd_And(dbg, block, br, a, mode);
3271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3275 /* a & (b ^ a) -> a & ~b */
3276 dbg_info *dbg = get_irn_dbg_info(n);
3277 ir_node *block = get_nodes_block(n);
3279 bl = new_rd_Not(dbg, block, bl, mode);
3280 n = new_rd_And(dbg, block, bl, a, mode);
3281 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3285 if (is_Not(a) && is_Not(b)) {
3286 /* ~a & ~b = ~(a|b) */
3287 ir_node *block = get_nodes_block(n);
3288 ir_mode *mode = get_irn_mode(n);
3292 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3293 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3294 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3298 b_vrp = vrp_get_info(b);
3299 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3300 b_vrp->bits_not_set), get_Const_tarval(a)) == pn_Cmp_Eq)) {
3306 a_vrp = vrp_get_info(a);
3307 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3308 a_vrp->bits_not_set), get_Const_tarval(b)) == pn_Cmp_Eq)) {
3312 n = transform_bitwise_distributive(n, transform_node_And);
3315 } /* transform_node_And */
3317 /* the order of the values is important! */
3318 typedef enum const_class {
3324 static const_class classify_const(const ir_node* n)
3326 if (is_Const(n)) return const_const;
3327 if (is_irn_constlike(n)) return const_like;
3332 * Determines whether r is more constlike or has a larger index (in that order)
3335 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3337 const const_class l_order = classify_const(l);
3338 const const_class r_order = classify_const(r);
3340 l_order > r_order ||
3341 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3347 static ir_node *transform_node_Eor(ir_node *n)
3349 ir_node *c, *oldn = n;
3350 ir_node *a = get_Eor_left(n);
3351 ir_node *b = get_Eor_right(n);
3352 ir_mode *mode = get_irn_mode(n);
3354 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3356 /* we can evaluate 2 Projs of the same Cmp */
3357 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3358 ir_node *pred_a = get_Proj_pred(a);
3359 ir_node *pred_b = get_Proj_pred(b);
3360 if (pred_a == pred_b) {
3361 dbg_info *dbgi = get_irn_dbg_info(n);
3362 pn_Cmp pn_a = get_Proj_pn_cmp(a);
3363 pn_Cmp pn_b = get_Proj_pn_cmp(b);
3364 /* yes, we can simply calculate with pncs */
3365 pn_Cmp new_pnc = pn_a ^ pn_b;
3367 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3371 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3372 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3373 dbg_info *dbg = get_irn_dbg_info(n);
3374 ir_node *block = get_nodes_block(n);
3375 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3376 ir_node *new_left = get_Not_op(a);
3377 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3378 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3380 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3381 dbg_info *dbg = get_irn_dbg_info(n);
3382 ir_node *block = get_nodes_block(n);
3383 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3384 ir_node *new_right = get_Not_op(b);
3385 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3386 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3390 /* x ^ 1...1 -> ~1 */
3391 if (is_Const(b) && is_Const_all_one(b)) {
3392 n = new_r_Not(get_nodes_block(n), a, mode);
3393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3397 n = transform_bitwise_distributive(n, transform_node_Eor);
3399 } /* transform_node_Eor */
3404 static ir_node *transform_node_Not(ir_node *n)
3406 ir_node *c, *oldn = n;
3407 ir_node *a = get_Not_op(n);
3408 ir_mode *mode = get_irn_mode(n);
3410 HANDLE_UNOP_PHI(tarval_not,a,c);
3412 /* check for a boolean Not */
3413 if (mode == mode_b && is_Proj(a)) {
3414 ir_node *a_pred = get_Proj_pred(a);
3415 if (is_Cmp(a_pred)) {
3416 /* We negate a Cmp. The Cmp has the negated result anyways! */
3417 n = new_r_Proj(get_Proj_pred(a),
3418 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3419 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3424 /* normalize ~(a ^ b) => a ^ ~b */
3426 dbg_info *dbg = get_irn_dbg_info(n);
3427 ir_node *block = get_nodes_block(n);
3428 ir_node *eor_right = get_Eor_right(a);
3429 ir_node *eor_left = get_Eor_left(a);
3430 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3431 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3435 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3436 if (is_Minus(a)) { /* ~-x -> x + -1 */
3437 dbg_info *dbg = get_irn_dbg_info(n);
3438 ir_graph *irg = get_irn_irg(n);
3439 ir_node *block = get_nodes_block(n);
3440 ir_node *add_l = get_Minus_op(a);
3441 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3442 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3443 } else if (is_Add(a)) {
3444 ir_node *add_r = get_Add_right(a);
3445 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3446 /* ~(x + -1) = -x */
3447 ir_node *op = get_Add_left(a);
3448 ir_node *blk = get_nodes_block(n);
3449 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3450 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3455 } /* transform_node_Not */
3458 * Transform a Minus.
3462 * -(a >>u (size-1)) = a >>s (size-1)
3463 * -(a >>s (size-1)) = a >>u (size-1)
3464 * -(a * const) -> a * -const
3466 static ir_node *transform_node_Minus(ir_node *n)
3468 ir_node *c, *oldn = n;
3469 ir_node *a = get_Minus_op(n);
3472 HANDLE_UNOP_PHI(tarval_neg,a,c);
3474 mode = get_irn_mode(a);
3475 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3476 /* the following rules are only to twos-complement */
3479 ir_node *op = get_Not_op(a);
3480 ir_tarval *tv = get_mode_one(mode);
3481 ir_node *blk = get_nodes_block(n);
3482 ir_graph *irg = get_irn_irg(blk);
3483 ir_node *c = new_r_Const(irg, tv);
3484 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3485 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3489 ir_node *c = get_Shr_right(a);
3492 ir_tarval *tv = get_Const_tarval(c);
3494 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3495 /* -(a >>u (size-1)) = a >>s (size-1) */
3496 ir_node *v = get_Shr_left(a);
3498 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3499 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3505 ir_node *c = get_Shrs_right(a);
3508 ir_tarval *tv = get_Const_tarval(c);
3510 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3511 /* -(a >>s (size-1)) = a >>u (size-1) */
3512 ir_node *v = get_Shrs_left(a);
3514 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3515 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3522 /* - (a-b) = b - a */
3523 ir_node *la = get_Sub_left(a);
3524 ir_node *ra = get_Sub_right(a);
3525 ir_node *blk = get_nodes_block(n);
3527 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3528 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3532 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3533 ir_node *mul_l = get_Mul_left(a);
3534 ir_node *mul_r = get_Mul_right(a);
3535 ir_tarval *tv = value_of(mul_r);
3536 if (tv != tarval_bad) {
3537 tv = tarval_neg(tv);
3538 if (tv != tarval_bad) {
3539 ir_graph *irg = get_irn_irg(n);
3540 ir_node *cnst = new_r_Const(irg, tv);
3541 dbg_info *dbg = get_irn_dbg_info(a);
3542 ir_node *block = get_nodes_block(a);
3543 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3544 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3551 } /* transform_node_Minus */
3554 * Transform a Proj(Load) with a non-null address.
3556 static ir_node *transform_node_Proj_Load(ir_node *proj)
3558 if (get_opt_ldst_only_null_ptr_exceptions()) {
3559 if (get_irn_mode(proj) == mode_X) {
3560 ir_node *load = get_Proj_pred(proj);
3562 /* get the Load address */
3563 const ir_node *addr = get_Load_ptr(load);
3564 const ir_node *confirm;
3566 if (value_not_null(addr, &confirm)) {
3567 if (confirm == NULL) {
3568 /* this node may float if it did not depend on a Confirm */
3569 set_irn_pinned(load, op_pin_state_floats);
3571 if (get_Proj_proj(proj) == pn_Load_X_except) {
3572 ir_graph *irg = get_irn_irg(proj);
3573 DBG_OPT_EXC_REM(proj);
3574 return get_irg_bad(irg);
3576 ir_node *blk = get_nodes_block(load);
3577 return new_r_Jmp(blk);
3583 } /* transform_node_Proj_Load */
3586 * Transform a Proj(Store) with a non-null address.
3588 static ir_node *transform_node_Proj_Store(ir_node *proj)
3590 if (get_opt_ldst_only_null_ptr_exceptions()) {
3591 if (get_irn_mode(proj) == mode_X) {
3592 ir_node *store = get_Proj_pred(proj);
3594 /* get the load/store address */
3595 const ir_node *addr = get_Store_ptr(store);
3596 const ir_node *confirm;
3598 if (value_not_null(addr, &confirm)) {
3599 if (confirm == NULL) {
3600 /* this node may float if it did not depend on a Confirm */
3601 set_irn_pinned(store, op_pin_state_floats);
3603 if (get_Proj_proj(proj) == pn_Store_X_except) {
3604 ir_graph *irg = get_irn_irg(proj);
3605 DBG_OPT_EXC_REM(proj);
3606 return get_irg_bad(irg);
3608 ir_node *blk = get_nodes_block(store);
3609 return new_r_Jmp(blk);
3615 } /* transform_node_Proj_Store */
3618 * Transform a Proj(Div) with a non-zero value.
3619 * Removes the exceptions and routes the memory to the NoMem node.
3621 static ir_node *transform_node_Proj_Div(ir_node *proj)
3623 ir_node *div = get_Proj_pred(proj);
3624 ir_node *b = get_Div_right(div);
3625 ir_node *res, *new_mem;
3626 const ir_node *confirm;
3629 if (value_not_zero(b, &confirm)) {
3630 /* div(x, y) && y != 0 */
3631 if (confirm == NULL) {
3632 /* we are sure we have a Const != 0 */
3633 new_mem = get_Div_mem(div);
3634 new_mem = skip_Pin(new_mem);
3635 set_Div_mem(div, new_mem);
3636 set_irn_pinned(div, op_pin_state_floats);
3639 proj_nr = get_Proj_proj(proj);
3641 case pn_Div_X_regular:
3642 return new_r_Jmp(get_nodes_block(div));
3644 case pn_Div_X_except: {
3645 ir_graph *irg = get_irn_irg(proj);
3646 /* we found an exception handler, remove it */
3647 DBG_OPT_EXC_REM(proj);
3648 return get_irg_bad(irg);
3652 ir_graph *irg = get_irn_irg(proj);
3653 res = get_Div_mem(div);
3654 new_mem = get_irg_no_mem(irg);
3657 /* This node can only float up to the Confirm block */
3658 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3660 set_irn_pinned(div, op_pin_state_floats);
3661 /* this is a Div without exception, we can remove the memory edge */
3662 set_Div_mem(div, new_mem);
3668 } /* transform_node_Proj_Div */
3671 * Transform a Proj(Mod) with a non-zero value.
3672 * Removes the exceptions and routes the memory to the NoMem node.
3674 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3676 ir_node *mod = get_Proj_pred(proj);
3677 ir_node *b = get_Mod_right(mod);
3678 ir_node *res, *new_mem;
3679 const ir_node *confirm;
3682 if (value_not_zero(b, &confirm)) {
3683 /* mod(x, y) && y != 0 */
3684 proj_nr = get_Proj_proj(proj);
3686 if (confirm == NULL) {
3687 /* we are sure we have a Const != 0 */
3688 new_mem = get_Mod_mem(mod);
3689 new_mem = skip_Pin(new_mem);
3690 set_Mod_mem(mod, new_mem);
3691 set_irn_pinned(mod, op_pin_state_floats);
3696 case pn_Mod_X_regular:
3697 return new_r_Jmp(get_irn_n(mod, -1));
3699 case pn_Mod_X_except: {
3700 ir_graph *irg = get_irn_irg(proj);
3701 /* we found an exception handler, remove it */
3702 DBG_OPT_EXC_REM(proj);
3703 return get_irg_bad(irg);
3707 ir_graph *irg = get_irn_irg(proj);
3708 res = get_Mod_mem(mod);
3709 new_mem = get_irg_no_mem(irg);
3712 /* This node can only float up to the Confirm block */
3713 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3715 /* this is a Mod without exception, we can remove the memory edge */
3716 set_Mod_mem(mod, new_mem);
3720 if (get_Mod_left(mod) == b) {
3721 /* a % a = 0 if a != 0 */
3722 ir_graph *irg = get_irn_irg(proj);
3723 ir_mode *mode = get_irn_mode(proj);
3724 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3726 DBG_OPT_CSTEVAL(mod, res);
3732 } /* transform_node_Proj_Mod */
3735 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3737 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3739 if (get_opt_unreachable_code()) {
3740 ir_node *n = get_Proj_pred(proj);
3741 ir_node *b = get_Cond_selector(n);
3743 if (mode_is_int(get_irn_mode(b))) {
3744 ir_tarval *tb = value_of(b);
3746 if (tb != tarval_bad) {
3747 /* we have a constant switch */
3748 long num = get_Proj_proj(proj);
3750 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3751 if (get_tarval_long(tb) == num) {
3752 /* Do NOT create a jump here, or we will have 2 control flow ops
3753 * in a block. This case is optimized away in optimize_cf(). */
3756 ir_graph *irg = get_irn_irg(proj);
3757 /* this case will NEVER be taken, kill it */
3758 return get_irg_bad(irg);
3762 long num = get_Proj_proj(proj);
3763 vrp_attr *b_vrp = vrp_get_info(b);
3764 if (num != get_Cond_default_proj(n) && b_vrp) {
3765 /* Try handling with vrp data. We only remove dead parts. */
3766 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3768 if (b_vrp->range_type == VRP_RANGE) {
3769 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3770 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3772 if ((cmp_result & pn_Cmp_Gt) == cmp_result && (cmp_result2
3773 & pn_Cmp_Lt) == cmp_result2) {
3774 ir_graph *irg = get_irn_irg(proj);
3775 return get_irg_bad(irg);
3777 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3778 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3779 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3781 if ((cmp_result & pn_Cmp_Le) == cmp_result && (cmp_result2
3782 & pn_Cmp_Ge) == cmp_result2) {
3783 ir_graph *irg = get_irn_irg(proj);
3784 return get_irg_bad(irg);
3789 tarval_and( b_vrp->bits_set, tp),
3792 ir_graph *irg = get_irn_irg(proj);
3793 return get_irg_bad(irg);
3799 tarval_not(b_vrp->bits_not_set)),
3800 tarval_not(b_vrp->bits_not_set))
3802 ir_graph *irg = get_irn_irg(proj);
3803 return get_irg_bad(irg);
3812 } /* transform_node_Proj_Cond */
3815 * return true if the operation returns a value with exactly 1 bit set
3817 static bool is_single_bit(const ir_node *node)
3819 /* a first implementation, could be extended with vrp and others... */
3821 ir_node *shl_l = get_Shl_left(node);
3822 ir_mode *mode = get_irn_mode(node);
3823 int modulo = get_mode_modulo_shift(mode);
3824 /* this works if we shift a 1 and we have modulo shift */
3825 if (is_Const(shl_l) && is_Const_one(shl_l)
3826 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3829 } else if (is_Const(node)) {
3830 ir_tarval *tv = get_Const_tarval(node);
3831 return tarval_is_single_bit(tv);
3837 * Create a 0 constant of given mode.
3839 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
3841 ir_tarval *tv = get_mode_null(mode);
3842 ir_node *cnst = new_r_Const(irg, tv);
3848 * Normalizes and optimizes Cmp nodes.
3850 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
3852 ir_node *n = get_Proj_pred(proj);
3853 ir_node *left = get_Cmp_left(n);
3854 ir_node *right = get_Cmp_right(n);
3855 ir_tarval *tv = NULL;
3857 ir_mode *mode = get_irn_mode(left);
3858 long proj_nr = get_Proj_proj(proj);
3860 /* we can evaluate some cases directly */
3862 case pn_Cmp_False: {
3863 ir_graph *irg = get_irn_irg(proj);
3864 return new_r_Const(irg, get_tarval_b_false());
3867 ir_graph *irg = get_irn_irg(proj);
3868 return new_r_Const(irg, get_tarval_b_true());
3871 if (!mode_is_float(mode)) {
3872 ir_graph *irg = get_irn_irg(proj);
3873 return new_r_Const(irg, get_tarval_b_true());
3880 /* remove Casts of both sides */
3881 left = skip_Cast(left);
3882 right = skip_Cast(right);
3884 /* Remove unnecessary conversions */
3885 /* TODO handle constants */
3886 if (is_Conv(left) && is_Conv(right)) {
3887 ir_node *op_left = get_Conv_op(left);
3888 ir_node *op_right = get_Conv_op(right);
3889 ir_mode *mode_left = get_irn_mode(op_left);
3890 ir_mode *mode_right = get_irn_mode(op_right);
3892 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3893 && mode_left != mode_b && mode_right != mode_b) {
3894 ir_node *block = get_nodes_block(n);
3896 if (mode_left == mode_right) {
3900 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3901 } else if (smaller_mode(mode_left, mode_right)) {
3902 left = new_r_Conv(block, op_left, mode_right);
3905 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3906 } else if (smaller_mode(mode_right, mode_left)) {
3908 right = new_r_Conv(block, op_right, mode_left);
3910 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3915 /* remove operation on both sides if possible */
3916 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3918 * The following operations are NOT safe for floating point operations, for instance
3919 * 1.0 + inf == 2.0 + inf, =/=> x == y
3921 if (mode_is_int(mode)) {
3922 unsigned lop = get_irn_opcode(left);
3924 if (lop == get_irn_opcode(right)) {
3925 ir_node *ll, *lr, *rl, *rr;
3927 /* same operation on both sides, try to remove */
3931 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3932 left = get_unop_op(left);
3933 right = get_unop_op(right);
3935 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3938 ll = get_Add_left(left);
3939 lr = get_Add_right(left);
3940 rl = get_Add_left(right);
3941 rr = get_Add_right(right);
3944 /* X + a CMP X + b ==> a CMP b */
3948 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3949 } else if (ll == rr) {
3950 /* X + a CMP b + X ==> a CMP b */
3954 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3955 } else if (lr == rl) {
3956 /* a + X CMP X + b ==> a CMP b */
3960 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3961 } else if (lr == rr) {
3962 /* a + X CMP b + X ==> a CMP b */
3966 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3970 ll = get_Sub_left(left);
3971 lr = get_Sub_right(left);
3972 rl = get_Sub_left(right);
3973 rr = get_Sub_right(right);
3976 /* X - a CMP X - b ==> a CMP b */
3980 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3981 } else if (lr == rr) {
3982 /* a - X CMP b - X ==> a CMP b */
3986 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3990 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3991 /* a ROTL X CMP b ROTL X ==> a CMP b */
3992 left = get_Rotl_left(left);
3993 right = get_Rotl_left(right);
3995 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4003 /* X+A == A, A+X == A, A-X == A -> X == 0 */
4004 if (is_Add(left) || is_Sub(left)) {
4005 ir_node *ll = get_binop_left(left);
4006 ir_node *lr = get_binop_right(left);
4008 if (lr == right && is_Add(left)) {
4014 ir_graph *irg = get_irn_irg(n);
4016 right = create_zero_const(irg, mode);
4018 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4021 if (is_Add(right) || is_Sub(right)) {
4022 ir_node *rl = get_binop_left(right);
4023 ir_node *rr = get_binop_right(right);
4025 if (rr == left && is_Add(right)) {
4031 ir_graph *irg = get_irn_irg(n);
4033 right = create_zero_const(irg, mode);
4035 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4038 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
4039 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
4041 ir_node *and0 = get_And_left(left);
4042 ir_node *and1 = get_And_right(left);
4043 if (and1 == right) {
4044 ir_node *tmp = and0;
4048 if (and0 == right && is_single_bit(and0)) {
4049 ir_graph *irg = get_irn_irg(n);
4050 proj_nr = get_negated_pnc(proj_nr, mode);
4051 right = create_zero_const(irg, mode);
4056 if (is_And(left) && is_Const(right)) {
4057 ir_node *ll = get_binop_left(left);
4058 ir_node *lr = get_binop_right(left);
4059 if (is_Shr(ll) && is_Const(lr)) {
4060 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4061 ir_node *block = get_nodes_block(n);
4062 ir_mode *mode = get_irn_mode(left);
4064 ir_node *llr = get_Shr_right(ll);
4065 if (is_Const(llr)) {
4066 dbg_info *dbg = get_irn_dbg_info(left);
4067 ir_graph *irg = get_irn_irg(left);
4069 ir_tarval *c1 = get_Const_tarval(llr);
4070 ir_tarval *c2 = get_Const_tarval(lr);
4071 ir_tarval *c3 = get_Const_tarval(right);
4072 ir_tarval *mask = tarval_shl(c2, c1);
4073 ir_tarval *value = tarval_shl(c3, c1);
4075 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4076 right = new_r_Const(irg, value);
4081 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4083 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4084 right = get_Eor_right(left);
4085 left = get_Eor_left(left);
4088 } /* mode_is_int(...) */
4089 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4091 /* replace mode_b compares with ands/ors */
4092 if (get_irn_mode(left) == mode_b) {
4093 ir_node *block = get_nodes_block(n);
4097 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4098 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4099 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4100 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4101 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4102 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4103 default: bres = NULL;
4106 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4112 * First step: normalize the compare op
4113 * by placing the constant on the right side
4114 * or moving the lower address node to the left.
4116 if (!operands_are_normalized(left, right)) {
4122 proj_nr = get_inversed_pnc(proj_nr);
4127 * Second step: Try to reduce the magnitude
4128 * of a constant. This may help to generate better code
4129 * later and may help to normalize more compares.
4130 * Of course this is only possible for integer values.
4132 tv = value_of(right);
4133 if (tv != tarval_bad) {
4134 mode = get_irn_mode(right);
4136 /* TODO extend to arbitrary constants */
4137 if (is_Conv(left) && tarval_is_null(tv)) {
4138 ir_node *op = get_Conv_op(left);
4139 ir_mode *op_mode = get_irn_mode(op);
4142 * UpConv(x) REL 0 ==> x REL 0
4143 * Don't do this for float values as it's unclear whether it is a
4144 * win. (on the other side it makes detection/creation of fabs hard)
4146 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4147 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4148 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4149 !mode_is_float(mode)) {
4150 tv = get_mode_null(op_mode);
4154 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4158 if (tv != tarval_bad) {
4159 /* the following optimization is possible on modes without Overflow
4160 * on Unary Minus or on == and !=:
4161 * -a CMP c ==> a swap(CMP) -c
4163 * Beware: for two-complement Overflow may occur, so only == and != can
4164 * be optimized, see this:
4165 * -MININT < 0 =/=> MININT > 0 !!!
4167 if (is_Minus(left) &&
4168 (!mode_overflow_on_unary_Minus(mode) ||
4169 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4170 tv = tarval_neg(tv);
4172 if (tv != tarval_bad) {
4173 left = get_Minus_op(left);
4174 proj_nr = get_inversed_pnc(proj_nr);
4176 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4178 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4179 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4180 tv = tarval_not(tv);
4182 if (tv != tarval_bad) {
4183 left = get_Not_op(left);
4185 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4189 /* for integer modes, we have more */
4190 if (mode_is_int(mode)) {
4191 /* Ne includes Unordered which is not possible on integers.
4192 * However, frontends often use this wrong, so fix it here */
4193 if (proj_nr & pn_Cmp_Uo) {
4194 proj_nr &= ~pn_Cmp_Uo;
4195 set_Proj_proj(proj, proj_nr);
4198 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4199 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4200 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4201 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4203 if (tv != tarval_bad) {
4204 proj_nr ^= pn_Cmp_Eq;
4206 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4209 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4210 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4211 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4212 tv = tarval_add(tv, get_mode_one(mode));
4214 if (tv != tarval_bad) {
4215 proj_nr ^= pn_Cmp_Eq;
4217 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4221 /* the following reassociations work only for == and != */
4222 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4224 #if 0 /* Might be not that good in general */
4225 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4226 if (tarval_is_null(tv) && is_Sub(left)) {
4227 right = get_Sub_right(left);
4228 left = get_Sub_left(left);
4230 tv = value_of(right);
4232 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4236 if (tv != tarval_bad) {
4237 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4239 ir_node *c1 = get_Sub_right(left);
4240 ir_tarval *tv2 = value_of(c1);
4242 if (tv2 != tarval_bad) {
4243 tv2 = tarval_add(tv, value_of(c1));
4245 if (tv2 != tarval_bad) {
4246 left = get_Sub_left(left);
4249 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4253 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4254 else if (is_Add(left)) {
4255 ir_node *a_l = get_Add_left(left);
4256 ir_node *a_r = get_Add_right(left);
4260 if (is_Const(a_l)) {
4262 tv2 = value_of(a_l);
4265 tv2 = value_of(a_r);
4268 if (tv2 != tarval_bad) {
4269 tv2 = tarval_sub(tv, tv2, NULL);
4271 if (tv2 != tarval_bad) {
4275 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4279 /* -a == c ==> a == -c, -a != c ==> a != -c */
4280 else if (is_Minus(left)) {
4281 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4283 if (tv2 != tarval_bad) {
4284 left = get_Minus_op(left);
4287 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4294 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4295 switch (get_irn_opcode(left)) {
4299 c1 = get_And_right(left);
4302 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4303 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4305 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4307 /* TODO: move to constant evaluation */
4308 ir_graph *irg = get_irn_irg(n);
4309 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4310 c1 = new_r_Const(irg, tv);
4311 DBG_OPT_CSTEVAL(proj, c1);
4315 if (tarval_is_single_bit(tv)) {
4317 * optimization for AND:
4319 * And(x, C) == C ==> And(x, C) != 0
4320 * And(x, C) != C ==> And(X, C) == 0
4322 * if C is a single Bit constant.
4325 /* check for Constant's match. We have check hare the tarvals,
4326 because our const might be changed */
4327 if (get_Const_tarval(c1) == tv) {
4328 /* fine: do the transformation */
4329 tv = get_mode_null(get_tarval_mode(tv));
4330 proj_nr ^= pn_Cmp_Leg;
4332 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4338 c1 = get_Or_right(left);
4339 if (is_Const(c1) && tarval_is_null(tv)) {
4341 * Or(x, C) == 0 && C != 0 ==> FALSE
4342 * Or(x, C) != 0 && C != 0 ==> TRUE
4344 if (! tarval_is_null(get_Const_tarval(c1))) {
4345 /* TODO: move to constant evaluation */
4346 ir_graph *irg = get_irn_irg(n);
4347 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4348 c1 = new_r_Const(irg, tv);
4349 DBG_OPT_CSTEVAL(proj, c1);
4356 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4358 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4361 c1 = get_Shl_right(left);
4363 ir_graph *irg = get_irn_irg(c1);
4364 ir_tarval *tv1 = get_Const_tarval(c1);
4365 ir_mode *mode = get_irn_mode(left);
4366 ir_tarval *minus1 = get_mode_all_one(mode);
4367 ir_tarval *amask = tarval_shr(minus1, tv1);
4368 ir_tarval *cmask = tarval_shl(minus1, tv1);
4371 if (tarval_and(tv, cmask) != tv) {
4372 /* condition not met */
4373 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4374 c1 = new_r_Const(irg, tv);
4375 DBG_OPT_CSTEVAL(proj, c1);
4378 sl = get_Shl_left(left);
4379 blk = get_nodes_block(n);
4380 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4381 tv = tarval_shr(tv, tv1);
4383 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4388 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4390 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4393 c1 = get_Shr_right(left);
4395 ir_graph *irg = get_irn_irg(c1);
4396 ir_tarval *tv1 = get_Const_tarval(c1);
4397 ir_mode *mode = get_irn_mode(left);
4398 ir_tarval *minus1 = get_mode_all_one(mode);
4399 ir_tarval *amask = tarval_shl(minus1, tv1);
4400 ir_tarval *cmask = tarval_shr(minus1, tv1);
4403 if (tarval_and(tv, cmask) != tv) {
4404 /* condition not met */
4405 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4406 c1 = new_r_Const(irg, tv);
4407 DBG_OPT_CSTEVAL(proj, c1);
4410 sl = get_Shr_left(left);
4411 blk = get_nodes_block(n);
4412 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4413 tv = tarval_shl(tv, tv1);
4415 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4420 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4422 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4425 c1 = get_Shrs_right(left);
4427 ir_graph *irg = get_irn_irg(c1);
4428 ir_tarval *tv1 = get_Const_tarval(c1);
4429 ir_mode *mode = get_irn_mode(left);
4430 ir_tarval *minus1 = get_mode_all_one(mode);
4431 ir_tarval *amask = tarval_shl(minus1, tv1);
4432 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4435 cond = tarval_sub(cond, tv1, NULL);
4436 cond = tarval_shrs(tv, cond);
4438 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4439 /* condition not met */
4440 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4441 c1 = new_r_Const(irg, tv);
4442 DBG_OPT_CSTEVAL(proj, c1);
4445 sl = get_Shrs_left(left);
4446 blk = get_nodes_block(n);
4447 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4448 tv = tarval_shl(tv, tv1);
4450 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4455 } /* tarval != bad */
4458 if (changed & 2) { /* need a new Const */
4459 ir_graph *irg = get_irn_irg(n);
4460 right = new_r_Const(irg, tv);
4463 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4464 ir_node *op = get_Proj_pred(left);
4466 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4467 ir_node *c = get_binop_right(op);
4470 ir_tarval *tv = get_Const_tarval(c);
4472 if (tarval_is_single_bit(tv)) {
4473 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4474 ir_node *v = get_binop_left(op);
4475 ir_node *blk = get_irn_n(op, -1);
4476 ir_graph *irg = get_irn_irg(op);
4477 ir_mode *mode = get_irn_mode(v);
4479 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4480 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4482 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4489 ir_node *block = get_nodes_block(n);
4491 /* create a new compare */
4492 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4493 proj = new_rd_Proj(get_irn_dbg_info(proj), n, get_irn_mode(proj), proj_nr);
4497 } /* transform_node_Proj_Cmp */
4500 * Optimize CopyB(mem, x, x) into a Nop.
4502 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4504 ir_node *copyb = get_Proj_pred(proj);
4505 ir_node *a = get_CopyB_dst(copyb);
4506 ir_node *b = get_CopyB_src(copyb);
4509 switch (get_Proj_proj(proj)) {
4510 case pn_CopyB_X_regular:
4511 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4512 DBG_OPT_EXC_REM(proj);
4513 proj = new_r_Jmp(get_nodes_block(copyb));
4515 case pn_CopyB_X_except:
4516 DBG_OPT_EXC_REM(proj);
4517 proj = get_irg_bad(get_irn_irg(proj));
4524 } /* transform_node_Proj_CopyB */
4527 * Optimize Bounds(idx, idx, upper) into idx.
4529 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4531 ir_node *oldn = proj;
4532 ir_node *bound = get_Proj_pred(proj);
4533 ir_node *idx = get_Bound_index(bound);
4534 ir_node *pred = skip_Proj(idx);
4537 if (idx == get_Bound_lower(bound))
4539 else if (is_Bound(pred)) {
4541 * idx was Bounds checked previously, it is still valid if
4542 * lower <= pred_lower && pred_upper <= upper.
4544 ir_node *lower = get_Bound_lower(bound);
4545 ir_node *upper = get_Bound_upper(bound);
4546 if (get_Bound_lower(pred) == lower &&
4547 get_Bound_upper(pred) == upper) {
4549 * One could expect that we simply return the previous
4550 * Bound here. However, this would be wrong, as we could
4551 * add an exception Proj to a new location then.
4552 * So, we must turn in into a tuple.
4558 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4559 switch (get_Proj_proj(proj)) {
4561 DBG_OPT_EXC_REM(proj);
4562 proj = get_Bound_mem(bound);
4564 case pn_Bound_X_except:
4565 DBG_OPT_EXC_REM(proj);
4566 proj = get_irg_bad(get_irn_irg(proj));
4570 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4572 case pn_Bound_X_regular:
4573 DBG_OPT_EXC_REM(proj);
4574 proj = new_r_Jmp(get_nodes_block(bound));
4581 } /* transform_node_Proj_Bound */
4584 * Does all optimizations on nodes that must be done on it's Proj's
4585 * because of creating new nodes.
4587 static ir_node *transform_node_Proj(ir_node *proj)
4589 ir_node *n = get_Proj_pred(proj);
4591 if (n->op->ops.transform_node_Proj)
4592 return n->op->ops.transform_node_Proj(proj);
4594 } /* transform_node_Proj */
4597 * Move Confirms down through Phi nodes.
4599 static ir_node *transform_node_Phi(ir_node *phi)
4602 ir_mode *mode = get_irn_mode(phi);
4604 if (mode_is_reference(mode)) {
4605 n = get_irn_arity(phi);
4607 /* Beware of Phi0 */
4609 ir_node *pred = get_irn_n(phi, 0);
4610 ir_node *bound, *new_phi, *block, **in;
4613 if (! is_Confirm(pred))
4616 bound = get_Confirm_bound(pred);
4617 pnc = get_Confirm_cmp(pred);
4619 NEW_ARR_A(ir_node *, in, n);
4620 in[0] = get_Confirm_value(pred);
4622 for (i = 1; i < n; ++i) {
4623 pred = get_irn_n(phi, i);
4625 if (! is_Confirm(pred) ||
4626 get_Confirm_bound(pred) != bound ||
4627 get_Confirm_cmp(pred) != pnc)
4629 in[i] = get_Confirm_value(pred);
4631 /* move the Confirm nodes "behind" the Phi */
4632 block = get_irn_n(phi, -1);
4633 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4634 return new_r_Confirm(block, new_phi, bound, pnc);
4638 } /* transform_node_Phi */
4641 * Returns the operands of a commutative bin-op, if one operand is
4642 * a const, it is returned as the second one.
4644 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4646 ir_node *op_a = get_binop_left(binop);
4647 ir_node *op_b = get_binop_right(binop);
4649 assert(is_op_commutative(get_irn_op(binop)));
4651 if (is_Const(op_a)) {
4658 } /* get_comm_Binop_Ops */
4661 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4662 * Such pattern may arise in bitfield stores.
4664 * value c4 value c4 & c2
4665 * AND c3 AND c1 | c3
4672 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4675 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4677 ir_node *irn_and, *c1;
4679 ir_node *and_l, *c3;
4680 ir_node *value, *c4;
4681 ir_node *new_and, *new_const, *block;
4682 ir_mode *mode = get_irn_mode(irn_or);
4684 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4688 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4689 if (!is_Const(c1) || !is_And(irn_and))
4692 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4696 tv1 = get_Const_tarval(c1);
4697 tv2 = get_Const_tarval(c2);
4699 tv = tarval_or(tv1, tv2);
4700 if (tarval_is_all_one(tv)) {
4701 /* the AND does NOT clear a bit with isn't set by the OR */
4702 set_Or_left(irn_or, or_l);
4703 set_Or_right(irn_or, c1);
4705 /* check for more */
4712 get_comm_Binop_Ops(or_l, &and_l, &c3);
4713 if (!is_Const(c3) || !is_And(and_l))
4716 get_comm_Binop_Ops(and_l, &value, &c4);
4720 /* ok, found the pattern, check for conditions */
4721 assert(mode == get_irn_mode(irn_and));
4722 assert(mode == get_irn_mode(or_l));
4723 assert(mode == get_irn_mode(and_l));
4725 tv3 = get_Const_tarval(c3);
4726 tv4 = get_Const_tarval(c4);
4728 tv = tarval_or(tv4, tv2);
4729 if (!tarval_is_all_one(tv)) {
4730 /* have at least one 0 at the same bit position */
4734 if (tv3 != tarval_andnot(tv3, tv4)) {
4735 /* bit in the or_mask is outside the and_mask */
4739 if (tv1 != tarval_andnot(tv1, tv2)) {
4740 /* bit in the or_mask is outside the and_mask */
4744 /* ok, all conditions met */
4745 block = get_irn_n(irn_or, -1);
4746 irg = get_irn_irg(block);
4748 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4750 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4752 set_Or_left(irn_or, new_and);
4753 set_Or_right(irn_or, new_const);
4755 /* check for more */
4757 } /* transform_node_Or_bf_store */
4760 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4762 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4764 ir_mode *mode = get_irn_mode(irn_or);
4765 ir_node *shl, *shr, *block;
4766 ir_node *irn, *x, *c1, *c2, *n;
4767 ir_tarval *tv1, *tv2;
4769 /* some backends can't handle rotl */
4770 if (!be_get_backend_param()->support_rotl)
4773 if (! mode_is_int(mode))
4776 shl = get_binop_left(irn_or);
4777 shr = get_binop_right(irn_or);
4786 } else if (!is_Shl(shl)) {
4788 } else if (!is_Shr(shr)) {
4791 x = get_Shl_left(shl);
4792 if (x != get_Shr_left(shr))
4795 c1 = get_Shl_right(shl);
4796 c2 = get_Shr_right(shr);
4797 if (is_Const(c1) && is_Const(c2)) {
4798 tv1 = get_Const_tarval(c1);
4799 if (! tarval_is_long(tv1))
4802 tv2 = get_Const_tarval(c2);
4803 if (! tarval_is_long(tv2))
4806 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4807 != (int) get_mode_size_bits(mode))
4810 /* yet, condition met */
4811 block = get_nodes_block(irn_or);
4813 n = new_r_Rotl(block, x, c1, mode);
4815 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4819 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4820 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4821 if (!is_negated_value(c1, c2)) {
4825 /* yet, condition met */
4826 block = get_nodes_block(irn_or);
4827 n = new_r_Rotl(block, x, c1, mode);
4828 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4830 } /* transform_node_Or_Rotl */
4835 static ir_node *transform_node_Or(ir_node *n)
4837 ir_node *c, *oldn = n;
4838 ir_node *a = get_Or_left(n);
4839 ir_node *b = get_Or_right(n);
4842 if (is_Not(a) && is_Not(b)) {
4843 /* ~a | ~b = ~(a&b) */
4844 ir_node *block = get_nodes_block(n);
4846 mode = get_irn_mode(n);
4849 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4850 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4851 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4855 /* we can evaluate 2 Projs of the same Cmp */
4856 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4857 ir_node *pred_a = get_Proj_pred(a);
4858 ir_node *pred_b = get_Proj_pred(b);
4859 if (pred_a == pred_b) {
4860 dbg_info *dbgi = get_irn_dbg_info(n);
4861 pn_Cmp pn_a = get_Proj_pn_cmp(a);
4862 pn_Cmp pn_b = get_Proj_pn_cmp(b);
4863 /* yes, we can simply calculate with pncs */
4864 pn_Cmp new_pnc = pn_a | pn_b;
4866 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
4870 mode = get_irn_mode(n);
4871 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4873 n = transform_node_Or_bf_store(n);
4874 n = transform_node_Or_Rotl(n);
4878 n = transform_bitwise_distributive(n, transform_node_Or);
4881 } /* transform_node_Or */
4885 static ir_node *transform_node(ir_node *n);
4888 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4890 * Should be moved to reassociation?
4892 static ir_node *transform_node_shift(ir_node *n)
4894 ir_node *left, *right;
4896 ir_tarval *tv1, *tv2, *res;
4897 ir_node *in[2], *irn, *block;
4900 left = get_binop_left(n);
4902 /* different operations */
4903 if (get_irn_op(left) != get_irn_op(n))
4906 right = get_binop_right(n);
4907 tv1 = value_of(right);
4908 if (tv1 == tarval_bad)
4911 tv2 = value_of(get_binop_right(left));
4912 if (tv2 == tarval_bad)
4915 res = tarval_add(tv1, tv2);
4916 mode = get_irn_mode(n);
4917 irg = get_irn_irg(n);
4919 /* beware: a simple replacement works only, if res < modulo shift */
4921 int modulo_shf = get_mode_modulo_shift(mode);
4922 if (modulo_shf > 0) {
4923 ir_tarval *modulo = new_tarval_from_long(modulo_shf,
4924 get_tarval_mode(res));
4926 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4928 /* shifting too much */
4929 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4931 ir_node *block = get_nodes_block(n);
4932 dbg_info *dbgi = get_irn_dbg_info(n);
4933 ir_mode *smode = get_irn_mode(right);
4934 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
4935 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
4938 return new_r_Const(irg, get_mode_null(mode));
4942 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
4945 /* ok, we can replace it */
4946 block = get_nodes_block(n);
4948 in[0] = get_binop_left(left);
4949 in[1] = new_r_Const(irg, res);
4951 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
4953 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4955 return transform_node(irn);
4956 } /* transform_node_shift */
4959 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
4961 * - and, or, xor instead of &
4962 * - Shl, Shr, Shrs, rotl instead of >>
4963 * (with a special case for Or/Xor + Shrs)
4965 static ir_node *transform_node_bitop_shift(ir_node *n)
4968 ir_node *right = get_binop_right(n);
4969 ir_mode *mode = get_irn_mode(n);
4970 ir_node *bitop_left;
4971 ir_node *bitop_right;
4981 ir_tarval *tv_shift;
4983 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
4985 if (!is_Const(right))
4988 left = get_binop_left(n);
4989 op_left = get_irn_op(left);
4990 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
4993 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
4994 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
4995 /* TODO: test if sign bit is affectes */
4999 bitop_right = get_binop_right(left);
5000 if (!is_Const(bitop_right))
5003 bitop_left = get_binop_left(left);
5005 block = get_nodes_block(n);
5006 dbgi = get_irn_dbg_info(n);
5007 tv1 = get_Const_tarval(bitop_right);
5008 tv2 = get_Const_tarval(right);
5010 assert(get_tarval_mode(tv1) == mode);
5013 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
5014 tv_shift = tarval_shl(tv1, tv2);
5015 } else if (is_Shr(n)) {
5016 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
5017 tv_shift = tarval_shr(tv1, tv2);
5018 } else if (is_Shrs(n)) {
5019 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
5020 tv_shift = tarval_shrs(tv1, tv2);
5023 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5024 tv_shift = tarval_rotl(tv1, tv2);
5027 assert(get_tarval_mode(tv_shift) == mode);
5028 irg = get_irn_irg(n);
5029 new_const = new_r_Const(irg, tv_shift);
5031 if (op_left == op_And) {
5032 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5033 } else if (op_left == op_Or) {
5034 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5036 assert(op_left == op_Eor);
5037 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5045 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5047 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5048 * (also with x >>s c1 when c1>=c2)
5050 static ir_node *transform_node_shl_shr(ir_node *n)
5053 ir_node *right = get_binop_right(n);
5063 ir_tarval *tv_shift;
5069 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5071 if (!is_Const(right))
5074 left = get_binop_left(n);
5075 mode = get_irn_mode(n);
5076 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5077 ir_node *shr_right = get_binop_right(left);
5079 if (!is_Const(shr_right))
5082 x = get_binop_left(left);
5083 tv_shr = get_Const_tarval(shr_right);
5084 tv_shl = get_Const_tarval(right);
5086 if (is_Shrs(left)) {
5087 /* shrs variant only allowed if c1 >= c2 */
5088 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5091 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5094 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5096 tv_mask = tarval_shl(tv_mask, tv_shl);
5097 } else if (is_Shr(n) && is_Shl(left)) {
5098 ir_node *shl_right = get_Shl_right(left);
5100 if (!is_Const(shl_right))
5103 x = get_Shl_left(left);
5104 tv_shr = get_Const_tarval(right);
5105 tv_shl = get_Const_tarval(shl_right);
5107 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5108 tv_mask = tarval_shr(tv_mask, tv_shr);
5113 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5114 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5117 assert(tv_mask != tarval_bad);
5118 assert(get_tarval_mode(tv_mask) == mode);
5120 block = get_nodes_block(n);
5121 irg = get_irn_irg(block);
5122 dbgi = get_irn_dbg_info(n);
5124 pnc = tarval_cmp(tv_shl, tv_shr);
5125 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5126 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5127 new_const = new_r_Const(irg, tv_shift);
5129 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5131 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5134 assert(pnc == pn_Cmp_Gt);
5135 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5136 new_const = new_r_Const(irg, tv_shift);
5137 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5140 new_const = new_r_Const(irg, tv_mask);
5141 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5146 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5148 ir_mode *mode = get_tarval_mode(tv);
5149 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5150 return tarval_mod(tv, modulo_tv);
5153 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5154 ir_node *left, ir_node *right, ir_mode *mode);
5157 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5158 * then we can use that to minimize the value of Add(x, const) or
5159 * Sub(Const, x). In particular this often avoids 1 instruction in some
5160 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5161 * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
5163 static ir_node *transform_node_shift_modulo(ir_node *n,
5164 new_shift_func new_shift)
5166 ir_mode *mode = get_irn_mode(n);
5167 int modulo = get_mode_modulo_shift(mode);
5168 ir_node *newop = NULL;
5169 ir_mode *mode_right;
5176 if (get_mode_arithmetic(mode) != irma_twos_complement)
5178 if (!is_po2(modulo))
5181 irg = get_irn_irg(n);
5182 block = get_nodes_block(n);
5183 right = get_binop_right(n);
5184 mode_right = get_irn_mode(right);
5185 if (is_Const(right)) {
5186 ir_tarval *tv = get_Const_tarval(right);
5187 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5192 newop = new_r_Const(irg, tv_mod);
5193 } else if (is_Add(right)) {
5194 ir_node *add_right = get_Add_right(right);
5195 if (is_Const(add_right)) {
5196 ir_tarval *tv = get_Const_tarval(add_right);
5197 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5202 newconst = new_r_Const(irg, tv_mod);
5203 newop = new_r_Add(block, get_Add_left(right), newconst,
5206 } else if (is_Sub(right)) {
5207 ir_node *sub_left = get_Sub_left(right);
5208 if (is_Const(sub_left)) {
5209 ir_tarval *tv = get_Const_tarval(sub_left);
5210 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5215 newconst = new_r_Const(irg, tv_mod);
5216 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5223 if (newop != NULL) {
5224 dbg_info *dbgi = get_irn_dbg_info(n);
5225 ir_node *left = get_binop_left(n);
5226 return new_shift(dbgi, block, left, newop, mode);
5234 static ir_node *transform_node_Shr(ir_node *n)
5236 ir_node *c, *oldn = n;
5237 ir_node *left = get_Shr_left(n);
5238 ir_node *right = get_Shr_right(n);
5239 ir_mode *mode = get_irn_mode(n);
5241 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5242 n = transform_node_shift(n);
5245 n = transform_node_shift_modulo(n, new_rd_Shr);
5247 n = transform_node_shl_shr(n);
5249 n = transform_node_bitop_shift(n);
5252 } /* transform_node_Shr */
5257 static ir_node *transform_node_Shrs(ir_node *n)
5259 ir_node *c, *oldn = n;
5260 ir_node *a = get_Shrs_left(n);
5261 ir_node *b = get_Shrs_right(n);
5262 ir_mode *mode = get_irn_mode(n);
5264 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5265 n = transform_node_shift(n);
5268 n = transform_node_shift_modulo(n, new_rd_Shrs);
5270 n = transform_node_bitop_shift(n);
5273 } /* transform_node_Shrs */
5278 static ir_node *transform_node_Shl(ir_node *n)
5280 ir_node *c, *oldn = n;
5281 ir_node *a = get_Shl_left(n);
5282 ir_node *b = get_Shl_right(n);
5283 ir_mode *mode = get_irn_mode(n);
5285 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5286 n = transform_node_shift(n);
5289 n = transform_node_shift_modulo(n, new_rd_Shl);
5291 n = transform_node_shl_shr(n);
5293 n = transform_node_bitop_shift(n);
5296 } /* transform_node_Shl */
5301 static ir_node *transform_node_Rotl(ir_node *n)
5303 ir_node *c, *oldn = n;
5304 ir_node *a = get_Rotl_left(n);
5305 ir_node *b = get_Rotl_right(n);
5306 ir_mode *mode = get_irn_mode(n);
5308 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5309 n = transform_node_shift(n);
5312 n = transform_node_bitop_shift(n);
5315 } /* transform_node_Rotl */
5320 static ir_node *transform_node_Conv(ir_node *n)
5322 ir_node *c, *oldn = n;
5323 ir_mode *mode = get_irn_mode(n);
5324 ir_node *a = get_Conv_op(n);
5326 if (mode != mode_b && is_const_Phi(a)) {
5327 /* Do NOT optimize mode_b Conv's, this leads to remaining
5328 * Phib nodes later, because the conv_b_lower operation
5329 * is instantly reverted, when it tries to insert a Convb.
5331 c = apply_conv_on_phi(a, mode);
5333 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5338 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5339 ir_graph *irg = get_irn_irg(n);
5340 return new_r_Unknown(irg, mode);
5343 if (mode_is_reference(mode) &&
5344 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5346 ir_node *l = get_Add_left(a);
5347 ir_node *r = get_Add_right(a);
5348 dbg_info *dbgi = get_irn_dbg_info(a);
5349 ir_node *block = get_nodes_block(n);
5351 ir_node *lop = get_Conv_op(l);
5352 if (get_irn_mode(lop) == mode) {
5353 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5354 n = new_rd_Add(dbgi, block, lop, r, mode);
5359 ir_node *rop = get_Conv_op(r);
5360 if (get_irn_mode(rop) == mode) {
5361 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5362 n = new_rd_Add(dbgi, block, l, rop, mode);
5369 } /* transform_node_Conv */
5372 * Remove dead blocks and nodes in dead blocks
5373 * in keep alive list. We do not generate a new End node.
5375 static ir_node *transform_node_End(ir_node *n)
5377 int i, j, n_keepalives = get_End_n_keepalives(n);
5380 NEW_ARR_A(ir_node *, in, n_keepalives);
5382 for (i = j = 0; i < n_keepalives; ++i) {
5383 ir_node *ka = get_End_keepalive(n, i);
5385 if (! is_Block_dead(ka)) {
5389 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5391 } else if (is_Bad(ka)) {
5392 /* no need to keep Bad */
5397 if (j != n_keepalives)
5398 set_End_keepalives(n, j, in);
5400 } /* transform_node_End */
5402 bool is_negated_value(ir_node *a, ir_node *b)
5404 if (is_Minus(a) && get_Minus_op(a) == b)
5406 if (is_Minus(b) && get_Minus_op(b) == a)
5408 if (is_Sub(a) && is_Sub(b)) {
5409 ir_node *a_left = get_Sub_left(a);
5410 ir_node *a_right = get_Sub_right(a);
5411 ir_node *b_left = get_Sub_left(b);
5412 ir_node *b_right = get_Sub_right(b);
5414 if (a_left == b_right && a_right == b_left)
5422 * Optimize a Mux into some simpler cases.
5424 static ir_node *transform_node_Mux(ir_node *n)
5426 ir_node *oldn = n, *sel = get_Mux_sel(n);
5427 ir_mode *mode = get_irn_mode(n);
5428 ir_node *t = get_Mux_true(n);
5429 ir_node *f = get_Mux_false(n);
5430 ir_graph *irg = get_irn_irg(n);
5432 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5436 ir_node* block = get_nodes_block(n);
5438 ir_node* c1 = get_Mux_sel(t);
5439 ir_node* t1 = get_Mux_true(t);
5440 ir_node* f1 = get_Mux_false(t);
5442 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5443 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5444 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5449 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5450 } else if (f == t1) {
5451 /* Mux(cond0, Mux(cond1, x, y), x) */
5452 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5453 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5454 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5459 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5461 } else if (is_Mux(f)) {
5462 ir_node* block = get_nodes_block(n);
5464 ir_node* c1 = get_Mux_sel(f);
5465 ir_node* t1 = get_Mux_true(f);
5466 ir_node* f1 = get_Mux_false(f);
5468 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5469 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5470 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5475 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5476 } else if (t == f1) {
5477 /* Mux(cond0, x, Mux(cond1, y, x)) */
5478 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5479 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5480 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5485 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5489 /* first normalization step: try to move a constant to the false side,
5490 * 0 preferred on false side too */
5492 ir_node *cmp = get_Proj_pred(sel);
5494 if (is_Cmp(cmp) && is_Const(t) &&
5495 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5496 pn_Cmp pnc = get_Proj_pn_cmp(sel);
5501 /* Mux(x, a, b) => Mux(not(x), b, a) */
5502 sel = new_r_Proj(cmp, mode_b,
5503 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5504 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5508 /* note: after normalization, false can only happen on default */
5509 if (mode == mode_b) {
5510 dbg_info *dbg = get_irn_dbg_info(n);
5511 ir_node *block = get_nodes_block(n);
5514 ir_tarval *tv_t = get_Const_tarval(t);
5515 if (tv_t == tarval_b_true) {
5517 /* Muxb(sel, true, false) = sel */
5518 assert(get_Const_tarval(f) == tarval_b_false);
5519 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5522 /* Muxb(sel, true, x) = Or(sel, x) */
5523 n = new_rd_Or(dbg, block, sel, f, mode_b);
5524 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5528 } else if (is_Const(f)) {
5529 ir_tarval *tv_f = get_Const_tarval(f);
5530 if (tv_f == tarval_b_true) {
5531 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5532 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5533 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5534 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5537 /* Muxb(sel, x, false) = And(sel, x) */
5538 assert(tv_f == tarval_b_false);
5539 n = new_rd_And(dbg, block, sel, t, mode_b);
5540 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5546 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5547 * value to integer. */
5548 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5549 ir_tarval *a = get_Const_tarval(t);
5550 ir_tarval *b = get_Const_tarval(f);
5552 if (tarval_is_one(a) && tarval_is_null(b)) {
5553 ir_node *block = get_nodes_block(n);
5554 ir_node *conv = new_r_Conv(block, sel, mode);
5556 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5558 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5559 ir_node *block = get_nodes_block(n);
5560 ir_node *not_ = new_r_Not(block, sel, mode_b);
5561 ir_node *conv = new_r_Conv(block, not_, mode);
5563 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5569 ir_node *cmp = get_Proj_pred(sel);
5570 long pn = get_Proj_proj(sel);
5573 * Note: normalization puts the constant on the right side,
5574 * so we check only one case.
5577 ir_node *cmp_r = get_Cmp_right(cmp);
5578 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5579 ir_node *block = get_nodes_block(n);
5580 ir_node *cmp_l = get_Cmp_left(cmp);
5582 if (mode_is_int(mode)) {
5584 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5585 /* Mux((a & b) != 0, c, 0) */
5586 ir_node *and_r = get_And_right(cmp_l);
5589 if (and_r == t && f == cmp_r) {
5590 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5591 if (pn == pn_Cmp_Lg) {
5592 /* Mux((a & 2^C) != 0, 2^C, 0) */
5594 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5596 /* Mux((a & 2^C) == 0, 2^C, 0) */
5597 n = new_rd_Eor(get_irn_dbg_info(n),
5598 block, cmp_l, t, mode);
5599 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5604 if (is_Shl(and_r)) {
5605 ir_node *shl_l = get_Shl_left(and_r);
5606 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5607 if (and_r == t && f == cmp_r) {
5608 if (pn == pn_Cmp_Lg) {
5609 /* (a & (1 << n)) != 0, (1 << n), 0) */
5611 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5613 /* (a & (1 << n)) == 0, (1 << n), 0) */
5614 n = new_rd_Eor(get_irn_dbg_info(n),
5615 block, cmp_l, t, mode);
5616 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5622 and_l = get_And_left(cmp_l);
5623 if (is_Shl(and_l)) {
5624 ir_node *shl_l = get_Shl_left(and_l);
5625 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5626 if (and_l == t && f == cmp_r) {
5627 if (pn == pn_Cmp_Lg) {
5628 /* ((1 << n) & a) != 0, (1 << n), 0) */
5630 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5632 /* ((1 << n) & a) == 0, (1 << n), 0) */
5633 n = new_rd_Eor(get_irn_dbg_info(n),
5634 block, cmp_l, t, mode);
5635 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5648 } /* transform_node_Mux */
5651 * optimize Sync nodes that have other syncs as input we simply add the inputs
5652 * of the other sync to our own inputs
5654 static ir_node *transform_node_Sync(ir_node *n)
5656 int arity = get_Sync_n_preds(n);
5659 for (i = 0; i < arity;) {
5660 ir_node *pred = get_Sync_pred(n, i);
5664 if (!is_Sync(pred)) {
5672 pred_arity = get_Sync_n_preds(pred);
5673 for (j = 0; j < pred_arity; ++j) {
5674 ir_node *pred_pred = get_Sync_pred(pred, j);
5679 add_irn_n(n, pred_pred);
5683 if (get_Sync_pred(n, k) == pred_pred) break;
5688 /* rehash the sync node */
5692 } /* transform_node_Sync */
5695 * optimize a trampoline Call into a direct Call
5697 static ir_node *transform_node_Call(ir_node *call)
5699 ir_node *callee = get_Call_ptr(call);
5700 ir_node *adr, *mem, *res, *bl, **in;
5701 ir_type *ctp, *mtp, *tp;
5705 int i, n_res, n_param;
5708 if (! is_Proj(callee))
5710 callee = get_Proj_pred(callee);
5711 if (! is_Builtin(callee))
5713 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5716 mem = get_Call_mem(call);
5718 if (skip_Proj(mem) == callee) {
5719 /* memory is routed to the trampoline, skip */
5720 mem = get_Builtin_mem(callee);
5723 /* build a new call type */
5724 mtp = get_Call_type(call);
5725 tdb = get_type_dbg_info(mtp);
5727 n_res = get_method_n_ress(mtp);
5728 n_param = get_method_n_params(mtp);
5729 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5731 for (i = 0; i < n_res; ++i)
5732 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5734 NEW_ARR_A(ir_node *, in, n_param + 1);
5736 /* FIXME: we don't need a new pointer type in every step */
5737 irg = get_irn_irg(call);
5738 tp = get_irg_frame_type(irg);
5739 tp = new_type_pointer(tp);
5740 set_method_param_type(ctp, 0, tp);
5742 in[0] = get_Builtin_param(callee, 2);
5743 for (i = 0; i < n_param; ++i) {
5744 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5745 in[i + 1] = get_Call_param(call, i);
5747 var = get_method_variadicity(mtp);
5748 set_method_variadicity(ctp, var);
5749 if (var == variadicity_variadic) {
5750 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
5752 /* When we resolve a trampoline, the function must be called by a this-call */
5753 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5754 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5756 adr = get_Builtin_param(callee, 1);
5758 db = get_irn_dbg_info(call);
5759 bl = get_nodes_block(call);
5761 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5762 if (get_irn_pinned(call) == op_pin_state_floats)
5763 set_irn_pinned(res, op_pin_state_floats);
5765 } /* transform_node_Call */
5768 * Tries several [inplace] [optimizing] transformations and returns an
5769 * equivalent node. The difference to equivalent_node() is that these
5770 * transformations _do_ generate new nodes, and thus the old node must
5771 * not be freed even if the equivalent node isn't the old one.
5773 static ir_node *transform_node(ir_node *n)
5778 * Transform_node is the only "optimizing transformation" that might
5779 * return a node with a different opcode. We iterate HERE until fixpoint
5780 * to get the final result.
5784 if (n->op->ops.transform_node != NULL)
5785 n = n->op->ops.transform_node(n);
5786 } while (oldn != n);
5789 } /* transform_node */
5792 * Sets the default transform node operation for an ir_op_ops.
5794 * @param code the opcode for the default operation
5795 * @param ops the operations initialized
5800 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5804 ops->transform_node = transform_node_##a; \
5806 #define CASE_PROJ(a) \
5808 ops->transform_node_Proj = transform_node_Proj_##a; \
5810 #define CASE_PROJ_EX(a) \
5812 ops->transform_node = transform_node_##a; \
5813 ops->transform_node_Proj = transform_node_Proj_##a; \
5853 } /* firm_set_default_transform_node */
5856 /* **************** Common Subexpression Elimination **************** */
5858 /** The size of the hash table used, should estimate the number of nodes
5860 #define N_IR_NODES 512
5862 /** Compares the attributes of two Const nodes. */
5863 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
5865 return get_Const_tarval(a) != get_Const_tarval(b);
5868 /** Compares the attributes of two Proj nodes. */
5869 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
5871 return a->attr.proj.proj != b->attr.proj.proj;
5872 } /* node_cmp_attr_Proj */
5874 /** Compares the attributes of two Alloc nodes. */
5875 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
5877 const alloc_attr *pa = &a->attr.alloc;
5878 const alloc_attr *pb = &b->attr.alloc;
5879 return (pa->where != pb->where) || (pa->type != pb->type);
5880 } /* node_cmp_attr_Alloc */
5882 /** Compares the attributes of two Free nodes. */
5883 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
5885 const free_attr *pa = &a->attr.free;
5886 const free_attr *pb = &b->attr.free;
5887 return (pa->where != pb->where) || (pa->type != pb->type);
5888 } /* node_cmp_attr_Free */
5890 /** Compares the attributes of two SymConst nodes. */
5891 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
5893 const symconst_attr *pa = &a->attr.symc;
5894 const symconst_attr *pb = &b->attr.symc;
5895 return (pa->kind != pb->kind)
5896 || (pa->sym.type_p != pb->sym.type_p);
5899 /** Compares the attributes of two Call nodes. */
5900 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
5902 const call_attr *pa = &a->attr.call;
5903 const call_attr *pb = &b->attr.call;
5904 return (pa->type != pb->type)
5905 || (pa->tail_call != pb->tail_call);
5906 } /* node_cmp_attr_Call */
5908 /** Compares the attributes of two Sel nodes. */
5909 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
5911 const ir_entity *a_ent = get_Sel_entity(a);
5912 const ir_entity *b_ent = get_Sel_entity(b);
5913 return a_ent != b_ent;
5914 } /* node_cmp_attr_Sel */
5916 /** Compares the attributes of two Phi nodes. */
5917 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
5919 /* we can only enter this function if both nodes have the same number of inputs,
5920 hence it is enough to check if one of them is a Phi0 */
5922 /* check the Phi0 pos attribute */
5923 return a->attr.phi.u.pos != b->attr.phi.u.pos;
5926 } /* node_cmp_attr_Phi */
5928 /** Compares the attributes of two Conv nodes. */
5929 static int node_cmp_attr_Conv(ir_node *a, ir_node *b)
5931 return get_Conv_strict(a) != get_Conv_strict(b);
5932 } /* node_cmp_attr_Conv */
5934 /** Compares the attributes of two Cast nodes. */
5935 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
5937 return get_Cast_type(a) != get_Cast_type(b);
5938 } /* node_cmp_attr_Cast */
5940 /** Compares the attributes of two Load nodes. */
5941 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
5943 if (get_Load_volatility(a) == volatility_is_volatile ||
5944 get_Load_volatility(b) == volatility_is_volatile)
5945 /* NEVER do CSE on volatile Loads */
5947 /* do not CSE Loads with different alignment. Be conservative. */
5948 if (get_Load_align(a) != get_Load_align(b))
5951 return get_Load_mode(a) != get_Load_mode(b);
5952 } /* node_cmp_attr_Load */
5954 /** Compares the attributes of two Store nodes. */
5955 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
5957 /* do not CSE Stores with different alignment. Be conservative. */
5958 if (get_Store_align(a) != get_Store_align(b))
5961 /* NEVER do CSE on volatile Stores */
5962 return (get_Store_volatility(a) == volatility_is_volatile ||
5963 get_Store_volatility(b) == volatility_is_volatile);
5964 } /* node_cmp_attr_Store */
5966 /** Compares two exception attributes */
5967 static int node_cmp_exception(ir_node *a, ir_node *b)
5969 const except_attr *ea = &a->attr.except;
5970 const except_attr *eb = &b->attr.except;
5972 return ea->pin_state != eb->pin_state;
5975 #define node_cmp_attr_Bound node_cmp_exception
5977 /** Compares the attributes of two Div nodes. */
5978 static int node_cmp_attr_Div(ir_node *a, ir_node *b)
5980 const divmod_attr *ma = &a->attr.divmod;
5981 const divmod_attr *mb = &b->attr.divmod;
5982 return ma->exc.pin_state != mb->exc.pin_state ||
5983 ma->resmode != mb->resmode ||
5984 ma->no_remainder != mb->no_remainder;
5985 } /* node_cmp_attr_Div */
5987 /** Compares the attributes of two Div or Mod nodes. */
5988 static int node_cmp_attr_Div_Mod(ir_node *a, ir_node *b)
5990 const divmod_attr *ma = &a->attr.divmod;
5991 const divmod_attr *mb = &b->attr.divmod;
5992 return ma->exc.pin_state != mb->exc.pin_state ||
5993 ma->resmode != mb->resmode;
5994 } /* node_cmp_attr_Div_Mod */
5996 /** Compares the attributes of two Mod nodes. */
5997 static int node_cmp_attr_Mod(ir_node *a, ir_node *b)
5999 return node_cmp_attr_Div_Mod(a, b);
6000 } /* node_cmp_attr_Mod */
6002 /** Compares the attributes of two Confirm nodes. */
6003 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
6005 /* no need to compare the bound, as this is a input */
6006 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
6007 } /* node_cmp_attr_Confirm */
6009 /** Compares the attributes of two Builtin nodes. */
6010 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b)
6012 /* no need to compare the type, equal kind means equal type */
6013 return get_Builtin_kind(a) != get_Builtin_kind(b);
6014 } /* node_cmp_attr_Builtin */
6016 /** Compares the attributes of two ASM nodes. */
6017 static int node_cmp_attr_ASM(ir_node *a, ir_node *b)
6020 const ir_asm_constraint *ca;
6021 const ir_asm_constraint *cb;
6024 if (get_ASM_text(a) != get_ASM_text(b))
6027 /* Should we really check the constraints here? Should be better, but is strange. */
6028 n = get_ASM_n_input_constraints(a);
6029 if (n != get_ASM_n_input_constraints(b))
6032 ca = get_ASM_input_constraints(a);
6033 cb = get_ASM_input_constraints(b);
6034 for (i = 0; i < n; ++i) {
6035 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6039 n = get_ASM_n_output_constraints(a);
6040 if (n != get_ASM_n_output_constraints(b))
6043 ca = get_ASM_output_constraints(a);
6044 cb = get_ASM_output_constraints(b);
6045 for (i = 0; i < n; ++i) {
6046 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6050 n = get_ASM_n_clobbers(a);
6051 if (n != get_ASM_n_clobbers(b))
6054 cla = get_ASM_clobbers(a);
6055 clb = get_ASM_clobbers(b);
6056 for (i = 0; i < n; ++i) {
6057 if (cla[i] != clb[i])
6061 } /* node_cmp_attr_ASM */
6063 /** Compares the inexistent attributes of two Dummy nodes. */
6064 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6072 * Set the default node attribute compare operation for an ir_op_ops.
6074 * @param code the opcode for the default operation
6075 * @param ops the operations initialized
6080 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6084 ops->node_cmp_attr = node_cmp_attr_##a; \
6115 } /* firm_set_default_node_cmp_attr */
6118 * Compare function for two nodes in the value table. Gets two
6119 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6121 int identities_cmp(const void *elt, const void *key)
6123 ir_node *a = (ir_node *)elt;
6124 ir_node *b = (ir_node *)key;
6127 if (a == b) return 0;
6129 if ((get_irn_op(a) != get_irn_op(b)) ||
6130 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6132 /* compare if a's in and b's in are of equal length */
6133 irn_arity_a = get_irn_arity(a);
6134 if (irn_arity_a != get_irn_arity(b))
6137 /* blocks are never the same */
6141 if (get_irn_pinned(a) == op_pin_state_pinned) {
6142 /* for pinned nodes, the block inputs must be equal */
6143 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6145 } else if (! get_opt_global_cse()) {
6146 /* for block-local CSE both nodes must be in the same Block */
6147 if (get_nodes_block(a) != get_nodes_block(b))
6151 /* compare a->in[0..ins] with b->in[0..ins] */
6152 for (i = 0; i < irn_arity_a; ++i) {
6153 ir_node *pred_a = get_irn_n(a, i);
6154 ir_node *pred_b = get_irn_n(b, i);
6155 if (pred_a != pred_b) {
6156 /* if both predecessors are CSE neutral they might be different */
6157 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6163 * here, we already now that the nodes are identical except their
6166 if (a->op->ops.node_cmp_attr)
6167 return a->op->ops.node_cmp_attr(a, b);
6170 } /* identities_cmp */
6173 * Calculate a hash value of a node.
6175 * @param node The IR-node
6177 unsigned ir_node_hash(const ir_node *node)
6179 return node->op->ops.hash(node);
6180 } /* ir_node_hash */
6183 void new_identities(ir_graph *irg)
6185 if (irg->value_table != NULL)
6186 del_pset(irg->value_table);
6187 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6188 } /* new_identities */
6190 void del_identities(ir_graph *irg)
6192 if (irg->value_table != NULL)
6193 del_pset(irg->value_table);
6194 } /* del_identities */
6196 /* Normalize a node by putting constants (and operands with larger
6197 * node index) on the right (operator side). */
6198 void ir_normalize_node(ir_node *n)
6200 if (is_op_commutative(get_irn_op(n))) {
6201 ir_node *l = get_binop_left(n);
6202 ir_node *r = get_binop_right(n);
6204 /* For commutative operators perform a OP b == b OP a but keep
6205 * constants on the RIGHT side. This helps greatly in some
6206 * optimizations. Moreover we use the idx number to make the form
6208 if (!operands_are_normalized(l, r)) {
6209 set_binop_left(n, r);
6210 set_binop_right(n, l);
6214 } /* ir_normalize_node */
6217 * Return the canonical node computing the same value as n.
6218 * Looks up the node in a hash table, enters it in the table
6219 * if it isn't there yet.
6221 * @param n the node to look up
6223 * @return a node that computes the same value as n or n if no such
6224 * node could be found
6226 ir_node *identify_remember(ir_node *n)
6228 ir_graph *irg = get_irn_irg(n);
6229 pset *value_table = irg->value_table;
6232 if (value_table == NULL)
6235 ir_normalize_node(n);
6236 /* lookup or insert in hash table with given hash key. */
6237 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6240 /* n is reachable again */
6241 edges_node_revival(nn, get_irn_irg(nn));
6245 } /* identify_remember */
6248 * During construction we set the op_pin_state_pinned flag in the graph right
6249 * when the optimization is performed. The flag turning on procedure global
6250 * cse could be changed between two allocations. This way we are safe.
6252 * @param n The node to lookup
6254 static inline ir_node *identify_cons(ir_node *n)
6258 n = identify_remember(n);
6259 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6260 ir_graph *irg = get_irn_irg(n);
6261 set_irg_pinned(irg, op_pin_state_floats);
6264 } /* identify_cons */
6266 /* Add a node to the identities value table. */
6267 void add_identities(ir_node *node)
6274 identify_remember(node);
6277 /* Visit each node in the value table of a graph. */
6278 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6281 ir_graph *rem = current_ir_graph;
6283 current_ir_graph = irg;
6284 foreach_pset(irg->value_table, ir_node*, node) {
6287 current_ir_graph = rem;
6288 } /* visit_all_identities */
6291 * Garbage in, garbage out. If a node has a dead input, i.e., the
6292 * Bad node is input to the node, return the Bad node.
6294 static ir_node *gigo(ir_node *node)
6297 ir_op *op = get_irn_op(node);
6299 /* remove garbage blocks by looking at control flow that leaves the block
6300 and replacing the control flow by Bad. */
6301 if (get_irn_mode(node) == mode_X) {
6302 ir_node *block = get_nodes_block(skip_Proj(node));
6303 ir_graph *irg = get_irn_irg(block);
6305 /* Don't optimize nodes in immature blocks. */
6306 if (!get_Block_matured(block))
6308 /* Don't optimize End, may have Bads. */
6309 if (op == op_End) return node;
6311 if (is_Block(block)) {
6312 if (is_Block_dead(block)) {
6313 /* control flow from dead block is dead */
6314 return get_irg_bad(irg);
6317 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6318 if (!is_Bad(get_irn_n(block, i)))
6322 ir_graph *irg = get_irn_irg(block);
6323 /* the start block is never dead */
6324 if (block != get_irg_start_block(irg)
6325 && block != get_irg_end_block(irg)) {
6327 * Do NOT kill control flow without setting
6328 * the block to dead of bad things can happen:
6329 * We get a Block that is not reachable be irg_block_walk()
6330 * but can be found by irg_walk()!
6332 set_Block_dead(block);
6333 return get_irg_bad(irg);
6339 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6340 blocks predecessors is dead. */
6341 if (op != op_Block && op != op_Phi && op != op_Tuple && op != op_Anchor) {
6342 ir_graph *irg = get_irn_irg(node);
6343 irn_arity = get_irn_arity(node);
6346 * Beware: we can only read the block of a non-floating node.
6348 if (is_irn_pinned_in_irg(node) &&
6349 is_Block_dead(get_nodes_block(skip_Proj(node))))
6350 return get_irg_bad(irg);
6352 for (i = 0; i < irn_arity; i++) {
6353 ir_node *pred = get_irn_n(node, i);
6356 return get_irg_bad(irg);
6358 /* Propagating Unknowns here seems to be a bad idea, because
6359 sometimes we need a node as a input and did not want that
6361 However, it might be useful to move this into a later phase
6362 (if you think that optimizing such code is useful). */
6363 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6364 return new_r_Unknown(irg, get_irn_mode(node));
6369 /* With this code we violate the agreement that local_optimize
6370 only leaves Bads in Block, Phi and Tuple nodes. */
6371 /* If Block has only Bads as predecessors it's garbage. */
6372 /* If Phi has only Bads as predecessors it's garbage. */
6373 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6374 irn_arity = get_irn_arity(node);
6375 for (i = 0; i < irn_arity; i++) {
6376 if (!is_Bad(get_irn_n(node, i))) break;
6378 if (i == irn_arity) node = get_irg_bad(irg);
6385 * These optimizations deallocate nodes from the obstack.
6386 * It can only be called if it is guaranteed that no other nodes
6387 * reference this one, i.e., right after construction of a node.
6389 * @param n The node to optimize
6391 ir_node *optimize_node(ir_node *n)
6394 ir_graph *irg = get_irn_irg(n);
6395 unsigned iro = get_irn_opcode(n);
6398 /* Always optimize Phi nodes: part of the construction. */
6399 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6401 /* constant expression evaluation / constant folding */
6402 if (get_opt_constant_folding()) {
6403 /* neither constants nor Tuple values can be evaluated */
6404 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6405 /* try to evaluate */
6406 tv = computed_value(n);
6407 if (tv != tarval_bad) {
6412 * we MUST copy the node here temporary, because it's still
6413 * needed for DBG_OPT_CSTEVAL
6415 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6416 oldn = (ir_node*)alloca(node_size);
6418 memcpy(oldn, n, node_size);
6419 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6421 /* ARG, copy the in array, we need it for statistics */
6422 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6424 /* note the inplace edges module */
6425 edges_node_deleted(n, irg);
6427 /* evaluation was successful -- replace the node. */
6428 irg_kill_node(irg, n);
6429 nw = new_r_Const(irg, tv);
6431 DBG_OPT_CSTEVAL(oldn, nw);
6437 /* remove unnecessary nodes */
6438 if (get_opt_algebraic_simplification() ||
6439 (iro == iro_Phi) || /* always optimize these nodes. */
6441 (iro == iro_Proj) ||
6442 (iro == iro_Block) ) /* Flags tested local. */
6443 n = equivalent_node(n);
6445 /* Common Subexpression Elimination.
6447 * Checks whether n is already available.
6448 * The block input is used to distinguish different subexpressions. Right
6449 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6450 * subexpressions within a block.
6453 n = identify_cons(n);
6456 edges_node_deleted(oldn, irg);
6458 /* We found an existing, better node, so we can deallocate the old node. */
6459 irg_kill_node(irg, oldn);
6463 /* Some more constant expression evaluation that does not allow to
6465 iro = get_irn_opcode(n);
6466 if (get_opt_algebraic_simplification() ||
6467 (iro == iro_Cond) ||
6468 (iro == iro_Proj)) /* Flags tested local. */
6469 n = transform_node(n);
6471 /* Remove nodes with dead (Bad) input.
6472 Run always for transformation induced Bads. */
6475 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6476 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6478 n = identify_remember(o);
6484 } /* optimize_node */
6488 * These optimizations never deallocate nodes (in place). This can cause dead
6489 * nodes lying on the obstack. Remove these by a dead node elimination,
6490 * i.e., a copying garbage collection.
6492 ir_node *optimize_in_place_2(ir_node *n)
6496 unsigned iro = get_irn_opcode(n);
6498 if (!get_opt_optimize() && !is_Phi(n)) return n;
6500 if (iro == iro_Deleted)
6503 /* constant expression evaluation / constant folding */
6504 if (get_opt_constant_folding()) {
6505 /* neither constants nor Tuple values can be evaluated */
6506 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6507 /* try to evaluate */
6508 tv = computed_value(n);
6509 if (tv != tarval_bad) {
6510 /* evaluation was successful -- replace the node. */
6511 ir_graph *irg = get_irn_irg(n);
6513 n = new_r_Const(irg, tv);
6515 DBG_OPT_CSTEVAL(oldn, n);
6521 /* remove unnecessary nodes */
6522 if (get_opt_constant_folding() ||
6523 (iro == iro_Phi) || /* always optimize these nodes. */
6524 (iro == iro_Id) || /* ... */
6525 (iro == iro_Proj) || /* ... */
6526 (iro == iro_Block) ) /* Flags tested local. */
6527 n = equivalent_node(n);
6529 /** common subexpression elimination **/
6530 /* Checks whether n is already available. */
6531 /* The block input is used to distinguish different subexpressions. Right
6532 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6533 subexpressions within a block. */
6534 if (get_opt_cse()) {
6536 n = identify_remember(o);
6541 /* Some more constant expression evaluation. */
6542 iro = get_irn_opcode(n);
6543 if (get_opt_constant_folding() ||
6544 (iro == iro_Cond) ||
6545 (iro == iro_Proj)) /* Flags tested local. */
6546 n = transform_node(n);
6548 /* Remove nodes with dead (Bad) input.
6549 Run always for transformation induced Bads. */
6552 /* Now we can verify the node, as it has no dead inputs any more. */
6555 /* Now we have a legal, useful node. Enter it in hash table for cse.
6556 Blocks should be unique anyways. (Except the successor of start:
6557 is cse with the start block!) */
6558 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6560 n = identify_remember(o);
6566 } /* optimize_in_place_2 */
6569 * Wrapper for external use, set proper status bits after optimization.
6571 ir_node *optimize_in_place(ir_node *n)
6573 ir_graph *irg = get_irn_irg(n);
6574 /* Handle graph state */
6575 assert(get_irg_phase_state(irg) != phase_building);
6577 if (get_opt_global_cse())
6578 set_irg_pinned(irg, op_pin_state_floats);
6579 if (get_irg_outs_state(irg) == outs_consistent)
6580 set_irg_outs_inconsistent(irg);
6582 /* FIXME: Maybe we could also test whether optimizing the node can
6583 change the control graph. */
6584 set_irg_doms_inconsistent(irg);
6585 return optimize_in_place_2(n);
6586 } /* optimize_in_place */
6589 * Calculate a hash value of a Const node.
6591 static unsigned hash_Const(const ir_node *node)
6595 /* special value for const, as they only differ in their tarval. */
6596 h = HASH_PTR(node->attr.con.tarval);
6602 * Calculate a hash value of a SymConst node.
6604 static unsigned hash_SymConst(const ir_node *node)
6608 /* all others are pointers */
6609 h = HASH_PTR(node->attr.symc.sym.type_p);
6612 } /* hash_SymConst */
6615 * Set the default hash operation in an ir_op_ops.
6617 * @param code the opcode for the default operation
6618 * @param ops the operations initialized
6623 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6627 ops->hash = hash_##a; \
6630 /* hash function already set */
6631 if (ops->hash != NULL)
6638 /* use input/mode default hash if no function was given */
6639 ops->hash = firm_default_hash;
6647 * Sets the default operation for an ir_ops.
6649 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops)
6651 ops = firm_set_default_hash(code, ops);
6652 ops = firm_set_default_computed_value(code, ops);
6653 ops = firm_set_default_equivalent_node(code, ops);
6654 ops = firm_set_default_transform_node(code, ops);
6655 ops = firm_set_default_node_cmp_attr(code, ops);
6656 ops = firm_set_default_get_type_attr(code, ops);
6657 ops = firm_set_default_get_entity_attr(code, ops);
6660 } /* firm_set_default_operations */