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_integer_Div(const ir_node *a, const ir_node *b)
552 ir_tarval *ta = value_of(a);
554 const ir_node *dummy;
556 /* Compute c1 / c2 or 0 / a, a != 0 */
557 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
558 return ta; /* 0 / b == 0 */
560 if (ta != tarval_bad && tb != tarval_bad)
561 return tarval_div(ta, tb);
563 } /* do_computed_value_integer_Div */
566 * Return the value of a floating point Div.
568 static ir_tarval *do_computed_value_float_Div(const ir_node *a, const ir_node *b)
570 ir_tarval *ta = value_of(a);
571 ir_tarval *tb = value_of(b);
573 /* cannot optimize 0 / b = 0 because of NaN */
574 if (ta != tarval_bad && tb != tarval_bad)
575 return tarval_quo(ta, tb);
577 } /* do_computed_value_float_Div */
580 * Calculate the value of an integer Mod of two nodes.
581 * Special case: a % 1
583 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
585 ir_tarval *ta = value_of(a);
586 ir_tarval *tb = value_of(b);
588 /* Compute a % 1 or c1 % c2 */
589 if (tarval_is_one(tb))
590 return get_mode_null(get_irn_mode(a));
591 if (ta != tarval_bad && tb != tarval_bad)
592 return tarval_mod(ta, tb);
594 } /* do_computed_value_Mod */
597 * Return the value of a Proj(Div).
599 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
601 long proj_nr = get_Proj_proj(n);
603 if (proj_nr == pn_Div_res) {
604 const ir_node *div = get_Proj_pred(n);
605 const ir_mode *mode = get_irn_mode(get_Div_resmode(div));
606 if (mode_is_int(mode))
607 return do_computed_value_integer_Div(get_Div_left(div), get_Div_right(div));
608 else if (mode_is_float(mode))
609 return do_computed_value_float_Div(get_Div_left(div), get_Div_right(div));
612 } /* computed_value_Proj_Div */
615 * Return the value of a Proj(Mod).
617 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
619 long proj_nr = get_Proj_proj(n);
621 if (proj_nr == pn_Mod_res) {
622 const ir_node *mod = get_Proj_pred(n);
623 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
626 } /* computed_value_Proj_Mod */
629 * Return the value of a Proj.
631 static ir_tarval *computed_value_Proj(const ir_node *proj)
633 ir_node *n = get_Proj_pred(proj);
635 if (n->op->ops.computed_value_Proj != NULL)
636 return n->op->ops.computed_value_Proj(proj);
638 } /* computed_value_Proj */
641 * If the parameter n can be computed, return its value, else tarval_bad.
642 * Performs constant folding.
644 * @param n The node this should be evaluated
646 ir_tarval *computed_value(const ir_node *n)
648 vrp_attr *vrp = vrp_get_info(n);
649 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == pn_Cmp_Eq) {
650 return vrp->bits_set;
652 if (n->op->ops.computed_value)
653 return n->op->ops.computed_value(n);
655 } /* computed_value */
658 * Set the default computed_value evaluator in an ir_op_ops.
660 * @param code the opcode for the default operation
661 * @param ops the operations initialized
666 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
670 ops->computed_value = computed_value_##a; \
672 #define CASE_PROJ(a) \
674 ops->computed_value_Proj = computed_value_Proj_##a; \
709 } /* firm_set_default_computed_value */
712 * Returns a equivalent block for another block.
713 * If the block has only one predecessor, this is
714 * the equivalent one. If the only predecessor of a block is
715 * the block itself, this is a dead block.
717 * If both predecessors of a block are the branches of a binary
718 * Cond, the equivalent block is Cond's block.
720 * If all predecessors of a block are bad or lies in a dead
721 * block, the current block is dead as well.
723 * Note, that blocks are NEVER turned into Bad's, instead
724 * the dead_block flag is set. So, never test for is_Bad(block),
725 * always use is_dead_Block(block).
727 static ir_node *equivalent_node_Block(ir_node *n)
733 /* don't optimize dead or labeled blocks */
734 if (is_Block_dead(n) || has_Block_entity(n))
737 n_preds = get_Block_n_cfgpreds(n);
739 /* The Block constructor does not call optimize, but mature_immBlock()
740 calls the optimization. */
741 assert(get_Block_matured(n));
743 irg = get_irn_irg(n);
745 /* Straightening: a single entry Block following a single exit Block
746 can be merged, if it is not the Start block. */
747 /* !!! Beware, all Phi-nodes of n must have been optimized away.
748 This should be true, as the block is matured before optimize is called.
749 But what about Phi-cycles with the Phi0/Id that could not be resolved?
750 Remaining Phi nodes are just Ids. */
752 ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0));
755 ir_node *predblock = get_nodes_block(pred);
756 if (predblock == oldn) {
757 /* Jmp jumps into the block it is in -- deal self cycle. */
758 n = set_Block_dead(n);
759 DBG_OPT_DEAD_BLOCK(oldn, n);
762 DBG_OPT_STG(oldn, n);
764 } else if (is_Cond(pred)) {
765 ir_node *predblock = get_nodes_block(pred);
766 if (predblock == oldn) {
767 /* Jmp jumps into the block it is in -- deal self cycle. */
768 n = set_Block_dead(n);
769 DBG_OPT_DEAD_BLOCK(oldn, n);
772 } else if (n_preds == 2) {
773 /* Test whether Cond jumps twice to this block
774 * The more general case which more than 2 predecessors is handles
775 * in optimize_cf(), we handle only this special case for speed here.
777 ir_node *a = get_Block_cfgpred(n, 0);
778 ir_node *b = get_Block_cfgpred(n, 1);
780 if (is_Proj(a) && is_Proj(b)) {
781 ir_node *cond = get_Proj_pred(a);
783 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
784 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
785 /* Also a single entry Block following a single exit Block. Phis have
786 twice the same operand and will be optimized away. */
787 n = get_nodes_block(cond);
788 DBG_OPT_IFSIM1(oldn, a, b, n);
791 } else if (get_opt_unreachable_code() &&
792 (n != get_irg_start_block(irg)) &&
793 (n != get_irg_end_block(irg))) {
796 /* If all inputs are dead, this block is dead too, except if it is
797 the start or end block. This is one step of unreachable code
799 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
800 ir_node *pred = get_Block_cfgpred(n, i);
803 if (is_Bad(pred)) continue;
804 pred_blk = get_nodes_block(skip_Proj(pred));
806 if (is_Block_dead(pred_blk)) continue;
809 /* really found a living input */
814 n = set_Block_dead(n);
815 DBG_OPT_DEAD_BLOCK(oldn, n);
820 } /* equivalent_node_Block */
823 * Returns a equivalent node for a Jmp, a Bad :-)
824 * Of course this only happens if the Block of the Jmp is dead.
826 static ir_node *equivalent_node_Jmp(ir_node *n)
830 /* unreachable code elimination */
831 if (is_Block_dead(get_nodes_block(n))) {
832 ir_graph *irg = get_irn_irg(n);
833 n = get_irg_bad(irg);
834 DBG_OPT_DEAD_BLOCK(oldn, n);
837 } /* equivalent_node_Jmp */
839 /** Raise is handled in the same way as Jmp. */
840 #define equivalent_node_Raise equivalent_node_Jmp
843 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
844 See transform_node_Proj_Cond(). */
847 * Optimize operations that are commutative and have neutral 0,
848 * so a op 0 = 0 op a = a.
850 static ir_node *equivalent_node_neutral_zero(ir_node *n)
854 ir_node *a = get_binop_left(n);
855 ir_node *b = get_binop_right(n);
860 /* After running compute_node there is only one constant predecessor.
861 Find this predecessors value and remember the other node: */
862 if ((tv = value_of(a)) != tarval_bad) {
864 } else if ((tv = value_of(b)) != tarval_bad) {
869 /* If this predecessors constant value is zero, the operation is
870 * unnecessary. Remove it.
872 * Beware: If n is a Add, the mode of on and n might be different
873 * which happens in this rare construction: NULL + 3.
874 * Then, a Conv would be needed which we cannot include here.
876 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
879 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
883 } /* equivalent_node_neutral_zero */
886 * Eor is commutative and has neutral 0.
888 static ir_node *equivalent_node_Eor(ir_node *n)
894 n = equivalent_node_neutral_zero(n);
895 if (n != oldn) return n;
898 b = get_Eor_right(n);
901 ir_node *aa = get_Eor_left(a);
902 ir_node *ab = get_Eor_right(a);
905 /* (a ^ b) ^ a -> b */
907 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
909 } else if (ab == b) {
910 /* (a ^ b) ^ b -> a */
912 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
917 ir_node *ba = get_Eor_left(b);
918 ir_node *bb = get_Eor_right(b);
921 /* a ^ (a ^ b) -> b */
923 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
925 } else if (bb == a) {
926 /* a ^ (b ^ a) -> b */
928 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
936 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
938 * The second one looks strange, but this construct
939 * is used heavily in the LCC sources :-).
941 * Beware: The Mode of an Add may be different than the mode of its
942 * predecessors, so we could not return a predecessors in all cases.
944 static ir_node *equivalent_node_Add(ir_node *n)
947 ir_node *left, *right;
948 ir_mode *mode = get_irn_mode(n);
950 n = equivalent_node_neutral_zero(n);
954 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
955 if (mode_is_float(mode)) {
956 ir_graph *irg = get_irn_irg(n);
957 if (get_irg_fp_model(irg) & fp_strict_algebraic)
961 left = get_Add_left(n);
962 right = get_Add_right(n);
965 if (get_Sub_right(left) == right) {
968 n = get_Sub_left(left);
969 if (mode == get_irn_mode(n)) {
970 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
976 if (get_Sub_right(right) == left) {
979 n = get_Sub_left(right);
980 if (mode == get_irn_mode(n)) {
981 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
987 } /* equivalent_node_Add */
990 * optimize operations that are not commutative but have neutral 0 on left,
993 static ir_node *equivalent_node_left_zero(ir_node *n)
997 ir_node *a = get_binop_left(n);
998 ir_node *b = get_binop_right(n);
999 ir_tarval *tb = value_of(b);
1001 if (tarval_is_null(tb)) {
1004 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1007 } /* equivalent_node_left_zero */
1009 #define equivalent_node_Shl equivalent_node_left_zero
1010 #define equivalent_node_Shr equivalent_node_left_zero
1011 #define equivalent_node_Shrs equivalent_node_left_zero
1012 #define equivalent_node_Rotl equivalent_node_left_zero
1015 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
1017 * The second one looks strange, but this construct
1018 * is used heavily in the LCC sources :-).
1020 * Beware: The Mode of a Sub may be different than the mode of its
1021 * predecessors, so we could not return a predecessors in all cases.
1023 static ir_node *equivalent_node_Sub(ir_node *n)
1027 ir_mode *mode = get_irn_mode(n);
1030 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1031 if (mode_is_float(mode)) {
1032 ir_graph *irg = get_irn_irg(n);
1033 if (get_irg_fp_model(irg) & fp_strict_algebraic)
1037 b = get_Sub_right(n);
1040 /* Beware: modes might be different */
1041 if (tarval_is_null(tb)) {
1042 ir_node *a = get_Sub_left(n);
1043 if (mode == get_irn_mode(a)) {
1046 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
1050 } /* equivalent_node_Sub */
1054 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
1057 * -(-a) == a, but might overflow two times.
1058 * We handle it anyway here but the better way would be a
1059 * flag. This would be needed for Pascal for instance.
1061 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
1064 ir_node *pred = get_unop_op(n);
1066 /* optimize symmetric unop */
1067 if (get_irn_op(pred) == get_irn_op(n)) {
1068 n = get_unop_op(pred);
1069 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
1072 } /* equivalent_node_idempotent_unop */
1074 /** Optimize Not(Not(x)) == x. */
1075 #define equivalent_node_Not equivalent_node_idempotent_unop
1077 /** -(-x) == x ??? Is this possible or can --x raise an
1078 out of bounds exception if min =! max? */
1079 #define equivalent_node_Minus equivalent_node_idempotent_unop
1082 * Optimize a * 1 = 1 * a = a.
1084 static ir_node *equivalent_node_Mul(ir_node *n)
1087 ir_node *a = get_Mul_left(n);
1089 /* we can handle here only the n * n = n bit cases */
1090 if (get_irn_mode(n) == get_irn_mode(a)) {
1091 ir_node *b = get_Mul_right(n);
1095 * Mul is commutative and has again an other neutral element.
1096 * Constants are place right, so check this case first.
1099 if (tarval_is_one(tv)) {
1101 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1104 if (tarval_is_one(tv)) {
1106 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1111 } /* equivalent_node_Mul */
1114 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1116 static ir_node *equivalent_node_Or(ir_node *n)
1120 ir_node *a = get_Or_left(n);
1121 ir_node *b = get_Or_right(n);
1125 n = a; /* Or has it's own neutral element */
1126 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1129 /* constants are cormalized to right, check this site first */
1131 if (tarval_is_null(tv)) {
1133 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1137 if (tarval_is_null(tv)) {
1139 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1144 } /* equivalent_node_Or */
1147 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1149 static ir_node *equivalent_node_And(ir_node *n)
1153 ir_node *a = get_And_left(n);
1154 ir_node *b = get_And_right(n);
1158 n = a; /* And has it's own neutral element */
1159 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1162 /* constants are normalized to right, check this site first */
1164 if (tarval_is_all_one(tv)) {
1166 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1169 if (tv != get_tarval_bad()) {
1170 ir_mode *mode = get_irn_mode(n);
1171 if (!mode_is_signed(mode) && is_Conv(a)) {
1172 ir_node *convop = get_Conv_op(a);
1173 ir_mode *convopmode = get_irn_mode(convop);
1174 if (!mode_is_signed(convopmode)) {
1175 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1176 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1178 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1185 if (tarval_is_all_one(tv)) {
1187 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1191 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1194 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1199 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1202 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1207 } /* equivalent_node_And */
1210 * Try to remove useless Conv's:
1212 static ir_node *equivalent_node_Conv(ir_node *n)
1215 ir_node *a = get_Conv_op(n);
1217 ir_mode *n_mode = get_irn_mode(n);
1218 ir_mode *a_mode = get_irn_mode(a);
1221 if (n_mode == a_mode) { /* No Conv necessary */
1222 if (get_Conv_strict(n)) {
1225 /* neither Minus nor Confirm change the precision,
1226 so we can "look-through" */
1229 p = get_Minus_op(p);
1230 } else if (is_Confirm(p)) {
1231 p = get_Confirm_value(p);
1237 if (is_Conv(p) && get_Conv_strict(p)) {
1238 /* we known already, that a_mode == n_mode, and neither
1239 Minus change the mode, so the second Conv
1241 assert(get_irn_mode(p) == n_mode);
1243 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1247 ir_node *pred = get_Proj_pred(p);
1248 if (is_Load(pred)) {
1249 /* Loads always return with the exact precision of n_mode */
1250 assert(get_Load_mode(pred) == n_mode);
1252 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1255 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1256 pred = get_Proj_pred(pred);
1257 if (is_Start(pred)) {
1258 /* Arguments always return with the exact precision,
1259 as strictConv's are place before Call -- if the
1260 caller was compiled with the same setting.
1261 Otherwise, the semantics is probably still right. */
1262 assert(get_irn_mode(p) == n_mode);
1264 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1270 /* special case: the immediate predecessor is also a Conv */
1271 if (! get_Conv_strict(a)) {
1272 /* first one is not strict, kick it */
1274 a_mode = get_irn_mode(a);
1278 /* else both are strict conv, second is superfluous */
1280 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1285 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1288 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1289 ir_node *b = get_Conv_op(a);
1290 ir_mode *b_mode = get_irn_mode(b);
1292 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1293 /* both are strict conv */
1294 if (smaller_mode(a_mode, n_mode)) {
1295 /* both are strict, but the first is smaller, so
1296 the second cannot remove more precision, remove the
1298 set_Conv_strict(n, 0);
1301 if (n_mode == b_mode) {
1302 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1303 if (n_mode == mode_b) {
1304 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1305 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1307 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1308 if (values_in_mode(b_mode, a_mode)) {
1309 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1310 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1315 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1316 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1317 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1318 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1320 if (float_mantissa >= int_mantissa) {
1322 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1327 if (smaller_mode(b_mode, a_mode)) {
1328 if (get_Conv_strict(n))
1329 set_Conv_strict(b, 1);
1330 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1331 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1338 } /* equivalent_node_Conv */
1341 * - fold Phi-nodes, iff they have only one predecessor except
1344 static ir_node *equivalent_node_Phi(ir_node *n)
1350 ir_node *first_val = NULL; /* to shutup gcc */
1352 if (!get_opt_optimize() &&
1353 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1356 n_preds = get_Phi_n_preds(n);
1358 block = get_nodes_block(n);
1360 if (is_Block_dead(block)) {
1361 ir_graph *irg = get_irn_irg(n);
1362 return get_irg_bad(irg);
1365 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1367 /* Find first non-self-referencing input */
1368 for (i = 0; i < n_preds; ++i) {
1369 first_val = get_Phi_pred(n, i);
1370 if ( (first_val != n) /* not self pointer */
1372 /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad
1373 * predecessors. Then, Phi nodes in dead code might be removed, causing
1374 * nodes pointing to themself (Add's for instance).
1375 * This is really bad and causes endless recursions in several
1376 * code pathes, so we do NOT optimize such a code.
1377 * This is not that bad as it sounds, optimize_cf() removes bad control flow
1378 * (and bad Phi predecessors), so live code is optimized later.
1380 && (! is_Bad(get_Block_cfgpred(block, i)))
1382 ) { /* value not dead */
1383 break; /* then found first value. */
1388 ir_graph *irg = get_irn_irg(n);
1389 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1390 return get_irg_bad(irg);
1393 /* search for rest of inputs, determine if any of these
1394 are non-self-referencing */
1395 while (++i < n_preds) {
1396 ir_node *scnd_val = get_Phi_pred(n, i);
1397 if ( (scnd_val != n)
1398 && (scnd_val != first_val)
1401 && (! is_Bad(get_Block_cfgpred(block, i)))
1408 if (i >= n_preds && !is_Dummy(first_val)) {
1409 /* Fold, if no multiple distinct non-self-referencing inputs */
1411 DBG_OPT_PHI(oldn, n);
1414 } /* equivalent_node_Phi */
1417 * Several optimizations:
1418 * - fold Sync-nodes, iff they have only one predecessor except
1421 static ir_node *equivalent_node_Sync(ir_node *n)
1423 int arity = get_Sync_n_preds(n);
1426 for (i = 0; i < arity;) {
1427 ir_node *pred = get_Sync_pred(n, i);
1430 /* Remove Bad predecessors */
1437 /* Remove duplicate predecessors */
1443 if (get_Sync_pred(n, j) == pred) {
1452 ir_graph *irg = get_irn_irg(n);
1453 return get_irg_bad(irg);
1455 if (arity == 1) return get_Sync_pred(n, 0);
1457 } /* equivalent_node_Sync */
1460 * Optimize Proj(Tuple).
1462 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1464 ir_node *oldn = proj;
1465 ir_node *tuple = get_Proj_pred(proj);
1467 /* Remove the Tuple/Proj combination. */
1468 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1469 DBG_OPT_TUPLE(oldn, tuple, proj);
1472 } /* equivalent_node_Proj_Tuple */
1475 * Optimize a / 1 = a.
1477 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1479 ir_node *oldn = proj;
1480 ir_node *div = get_Proj_pred(proj);
1481 ir_node *b = get_Div_right(div);
1482 ir_tarval *tb = value_of(b);
1484 /* Div is not commutative. */
1485 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1486 switch (get_Proj_proj(proj)) {
1488 proj = get_Div_mem(div);
1489 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1493 proj = get_Div_left(div);
1494 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1498 /* we cannot replace the exception Proj's here, this is done in
1499 transform_node_Proj_Div() */
1504 } /* equivalent_node_Proj_Div */
1507 * Optimize CopyB(mem, x, x) into a Nop.
1509 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1511 ir_node *oldn = proj;
1512 ir_node *copyb = get_Proj_pred(proj);
1513 ir_node *a = get_CopyB_dst(copyb);
1514 ir_node *b = get_CopyB_src(copyb);
1517 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1518 switch (get_Proj_proj(proj)) {
1520 proj = get_CopyB_mem(copyb);
1521 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1524 case pn_CopyB_X_except: {
1525 ir_graph *irg = get_irn_irg(proj);
1526 DBG_OPT_EXC_REM(proj);
1527 proj = get_irg_bad(irg);
1533 } /* equivalent_node_Proj_CopyB */
1536 * Optimize Bounds(idx, idx, upper) into idx.
1538 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1540 ir_node *oldn = proj;
1541 ir_node *bound = get_Proj_pred(proj);
1542 ir_node *idx = get_Bound_index(bound);
1543 ir_node *pred = skip_Proj(idx);
1546 if (idx == get_Bound_lower(bound))
1548 else if (is_Bound(pred)) {
1550 * idx was Bounds checked previously, it is still valid if
1551 * lower <= pred_lower && pred_upper <= upper.
1553 ir_node *lower = get_Bound_lower(bound);
1554 ir_node *upper = get_Bound_upper(bound);
1555 if (get_Bound_lower(pred) == lower &&
1556 get_Bound_upper(pred) == upper) {
1558 * One could expect that we simply return the previous
1559 * Bound here. However, this would be wrong, as we could
1560 * add an exception Proj to a new location then.
1561 * So, we must turn in into a tuple.
1567 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1568 switch (get_Proj_proj(proj)) {
1570 DBG_OPT_EXC_REM(proj);
1571 proj = get_Bound_mem(bound);
1573 case pn_Bound_X_except: {
1574 ir_graph *irg = get_irn_irg(proj);
1575 DBG_OPT_EXC_REM(proj);
1576 proj = get_irg_bad(irg);
1581 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1584 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1589 } /* equivalent_node_Proj_Bound */
1592 * Optimize an Exception Proj(Load) with a non-null address.
1594 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1596 if (get_opt_ldst_only_null_ptr_exceptions()) {
1597 if (get_irn_mode(proj) == mode_X) {
1598 ir_node *load = get_Proj_pred(proj);
1600 /* get the Load address */
1601 const ir_node *addr = get_Load_ptr(load);
1602 const ir_node *confirm;
1604 if (value_not_null(addr, &confirm)) {
1605 if (get_Proj_proj(proj) == pn_Load_X_except) {
1606 ir_graph *irg = get_irn_irg(proj);
1607 DBG_OPT_EXC_REM(proj);
1608 return get_irg_bad(irg);
1614 } /* equivalent_node_Proj_Load */
1617 * Optimize an Exception Proj(Store) with a non-null address.
1619 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1621 if (get_opt_ldst_only_null_ptr_exceptions()) {
1622 if (get_irn_mode(proj) == mode_X) {
1623 ir_node *store = get_Proj_pred(proj);
1625 /* get the load/store address */
1626 const ir_node *addr = get_Store_ptr(store);
1627 const ir_node *confirm;
1629 if (value_not_null(addr, &confirm)) {
1630 if (get_Proj_proj(proj) == pn_Store_X_except) {
1631 ir_graph *irg = get_irn_irg(proj);
1632 DBG_OPT_EXC_REM(proj);
1633 return get_irg_bad(irg);
1639 } /* equivalent_node_Proj_Store */
1642 * Does all optimizations on nodes that must be done on it's Proj's
1643 * because of creating new nodes.
1645 static ir_node *equivalent_node_Proj(ir_node *proj)
1647 ir_node *n = get_Proj_pred(proj);
1649 if (get_irn_mode(proj) == mode_X) {
1650 if (is_Block_dead(get_nodes_block(n))) {
1651 /* Remove dead control flow -- early gigo(). */
1652 ir_graph *irg = get_irn_irg(proj);
1653 return get_irg_bad(irg);
1656 if (n->op->ops.equivalent_node_Proj)
1657 return n->op->ops.equivalent_node_Proj(proj);
1659 } /* equivalent_node_Proj */
1664 static ir_node *equivalent_node_Id(ir_node *n)
1672 DBG_OPT_ID(oldn, n);
1674 } /* equivalent_node_Id */
1679 static ir_node *equivalent_node_Mux(ir_node *n)
1681 ir_node *oldn = n, *sel = get_Mux_sel(n);
1683 ir_tarval *ts = value_of(sel);
1685 /* Mux(true, f, t) == t */
1686 if (ts == tarval_b_true) {
1687 n = get_Mux_true(n);
1688 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1691 /* Mux(false, f, t) == f */
1692 if (ts == tarval_b_false) {
1693 n = get_Mux_false(n);
1694 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1697 n_t = get_Mux_true(n);
1698 n_f = get_Mux_false(n);
1700 /* Mux(v, x, T) == x */
1701 if (is_Unknown(n_f)) {
1703 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1706 /* Mux(v, T, x) == x */
1707 if (is_Unknown(n_t)) {
1709 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1713 /* Mux(v, x, x) == x */
1716 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1719 if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1720 ir_node *cmp = get_Proj_pred(sel);
1721 long proj_nr = get_Proj_proj(sel);
1722 ir_node *f = get_Mux_false(n);
1723 ir_node *t = get_Mux_true(n);
1726 * Note further that these optimization work even for floating point
1727 * with NaN's because -NaN == NaN.
1728 * However, if +0 and -0 is handled differently, we cannot use the first one.
1731 ir_node *const cmp_l = get_Cmp_left(cmp);
1732 ir_node *const cmp_r = get_Cmp_right(cmp);
1736 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1737 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1739 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1746 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1747 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1749 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1756 * Note: normalization puts the constant on the right side,
1757 * so we check only one case.
1759 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1760 /* Mux(t CMP 0, X, t) */
1761 if (is_Minus(f) && get_Minus_op(f) == t) {
1762 /* Mux(t CMP 0, -t, t) */
1763 if (proj_nr == pn_Cmp_Eq) {
1764 /* Mux(t == 0, -t, t) ==> -t */
1766 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1767 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1768 /* Mux(t != 0, -t, t) ==> t */
1770 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1777 } /* equivalent_node_Mux */
1780 * Remove Confirm nodes if setting is on.
1781 * Replace Confirms(x, '=', Constlike) by Constlike.
1783 static ir_node *equivalent_node_Confirm(ir_node *n)
1785 ir_node *pred = get_Confirm_value(n);
1786 pn_Cmp pnc = get_Confirm_cmp(n);
1788 while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) {
1790 * rare case: two identical Confirms one after another,
1791 * replace the second one with the first.
1794 pred = get_Confirm_value(n);
1795 pnc = get_Confirm_cmp(n);
1801 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1802 * perform no actual computation, as, e.g., the Id nodes. It does not create
1803 * new nodes. It is therefore safe to free n if the node returned is not n.
1804 * If a node returns a Tuple we can not just skip it. If the size of the
1805 * in array fits, we transform n into a tuple (e.g., Div).
1807 ir_node *equivalent_node(ir_node *n)
1809 if (n->op->ops.equivalent_node)
1810 return n->op->ops.equivalent_node(n);
1812 } /* equivalent_node */
1815 * Sets the default equivalent node operation for an ir_op_ops.
1817 * @param code the opcode for the default operation
1818 * @param ops the operations initialized
1823 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1827 ops->equivalent_node = equivalent_node_##a; \
1829 #define CASE_PROJ(a) \
1831 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1871 } /* firm_set_default_equivalent_node */
1874 * Returns non-zero if a node is a Phi node
1875 * with all predecessors constant.
1877 static int is_const_Phi(ir_node *n)
1881 if (! is_Phi(n) || get_irn_arity(n) == 0)
1883 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1884 if (! is_Const(get_irn_n(n, i)))
1888 } /* is_const_Phi */
1890 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1891 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1894 * in reality eval_func should be tarval (*eval_func)() but incomplete
1895 * declarations are bad style and generate noisy warnings
1897 typedef void (*eval_func)(void);
1900 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1902 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1904 if (eval == (eval_func) tarval_sub) {
1905 tarval_sub_type func = (tarval_sub_type)eval;
1907 return func(a, b, mode);
1909 tarval_binop_type func = (tarval_binop_type)eval;
1916 * Apply an evaluator on a binop with a constant operators (and one Phi).
1918 * @param phi the Phi node
1919 * @param other the other operand
1920 * @param eval an evaluator function
1921 * @param mode the mode of the result, may be different from the mode of the Phi!
1922 * @param left if non-zero, other is the left operand, else the right
1924 * @return a new Phi node if the conversion was successful, NULL else
1926 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1932 int i, n = get_irn_arity(phi);
1934 NEW_ARR_A(void *, res, n);
1936 for (i = 0; i < n; ++i) {
1937 pred = get_irn_n(phi, i);
1938 tv = get_Const_tarval(pred);
1939 tv = do_eval(eval, other, tv, mode);
1941 if (tv == tarval_bad) {
1942 /* folding failed, bad */
1948 for (i = 0; i < n; ++i) {
1949 pred = get_irn_n(phi, i);
1950 tv = get_Const_tarval(pred);
1951 tv = do_eval(eval, tv, other, mode);
1953 if (tv == tarval_bad) {
1954 /* folding failed, bad */
1960 irg = get_irn_irg(phi);
1961 for (i = 0; i < n; ++i) {
1962 pred = get_irn_n(phi, i);
1963 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1965 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1966 } /* apply_binop_on_phi */
1969 * Apply an evaluator on a binop with two constant Phi.
1971 * @param a the left Phi node
1972 * @param b the right Phi node
1973 * @param eval an evaluator function
1974 * @param mode the mode of the result, may be different from the mode of the Phi!
1976 * @return a new Phi node if the conversion was successful, NULL else
1978 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1980 ir_tarval *tv_l, *tv_r, *tv;
1986 if (get_nodes_block(a) != get_nodes_block(b))
1989 n = get_irn_arity(a);
1990 NEW_ARR_A(void *, res, n);
1992 for (i = 0; i < n; ++i) {
1993 pred = get_irn_n(a, i);
1994 tv_l = get_Const_tarval(pred);
1995 pred = get_irn_n(b, i);
1996 tv_r = get_Const_tarval(pred);
1997 tv = do_eval(eval, tv_l, tv_r, mode);
1999 if (tv == tarval_bad) {
2000 /* folding failed, bad */
2005 irg = get_irn_irg(a);
2006 for (i = 0; i < n; ++i) {
2007 pred = get_irn_n(a, i);
2008 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2010 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
2011 } /* apply_binop_on_2_phis */
2014 * Apply an evaluator on a unop with a constant operator (a Phi).
2016 * @param phi the Phi node
2017 * @param eval an evaluator function
2019 * @return a new Phi node if the conversion was successful, NULL else
2021 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
2028 int i, n = get_irn_arity(phi);
2030 NEW_ARR_A(void *, res, n);
2031 for (i = 0; i < n; ++i) {
2032 pred = get_irn_n(phi, i);
2033 tv = get_Const_tarval(pred);
2036 if (tv == tarval_bad) {
2037 /* folding failed, bad */
2042 mode = get_irn_mode(phi);
2043 irg = get_irn_irg(phi);
2044 for (i = 0; i < n; ++i) {
2045 pred = get_irn_n(phi, i);
2046 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2048 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2049 } /* apply_unop_on_phi */
2052 * Apply a conversion on a constant operator (a Phi).
2054 * @param phi the Phi node
2056 * @return a new Phi node if the conversion was successful, NULL else
2058 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
2064 int i, n = get_irn_arity(phi);
2066 NEW_ARR_A(void *, res, n);
2067 for (i = 0; i < n; ++i) {
2068 pred = get_irn_n(phi, i);
2069 tv = get_Const_tarval(pred);
2070 tv = tarval_convert_to(tv, mode);
2072 if (tv == tarval_bad) {
2073 /* folding failed, bad */
2078 irg = get_irn_irg(phi);
2079 for (i = 0; i < n; ++i) {
2080 pred = get_irn_n(phi, i);
2081 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
2083 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
2084 } /* apply_conv_on_phi */
2087 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
2088 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
2089 * If possible, remove the Conv's.
2091 static ir_node *transform_node_AddSub(ir_node *n)
2093 ir_mode *mode = get_irn_mode(n);
2095 if (mode_is_reference(mode)) {
2096 ir_node *left = get_binop_left(n);
2097 ir_node *right = get_binop_right(n);
2098 unsigned ref_bits = get_mode_size_bits(mode);
2100 if (is_Conv(left)) {
2101 ir_mode *lmode = get_irn_mode(left);
2102 unsigned bits = get_mode_size_bits(lmode);
2104 if (ref_bits == bits &&
2105 mode_is_int(lmode) &&
2106 get_mode_arithmetic(lmode) == irma_twos_complement) {
2107 ir_node *pre = get_Conv_op(left);
2108 ir_mode *pre_mode = get_irn_mode(pre);
2110 if (mode_is_int(pre_mode) &&
2111 get_mode_size_bits(pre_mode) == bits &&
2112 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2113 /* ok, this conv just changes to sign, moreover the calculation
2114 * is done with same number of bits as our address mode, so
2115 * we can ignore the conv as address calculation can be viewed
2116 * as either signed or unsigned
2118 set_binop_left(n, pre);
2123 if (is_Conv(right)) {
2124 ir_mode *rmode = get_irn_mode(right);
2125 unsigned bits = get_mode_size_bits(rmode);
2127 if (ref_bits == bits &&
2128 mode_is_int(rmode) &&
2129 get_mode_arithmetic(rmode) == irma_twos_complement) {
2130 ir_node *pre = get_Conv_op(right);
2131 ir_mode *pre_mode = get_irn_mode(pre);
2133 if (mode_is_int(pre_mode) &&
2134 get_mode_size_bits(pre_mode) == bits &&
2135 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2136 /* ok, this conv just changes to sign, moreover the calculation
2137 * is done with same number of bits as our address mode, so
2138 * we can ignore the conv as address calculation can be viewed
2139 * as either signed or unsigned
2141 set_binop_right(n, pre);
2146 /* let address arithmetic use unsigned modes */
2147 if (is_Const(right)) {
2148 ir_mode *rmode = get_irn_mode(right);
2150 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2151 /* convert a AddP(P, *s) into AddP(P, *u) */
2152 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2154 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2155 set_binop_right(n, pre);
2161 } /* transform_node_AddSub */
2163 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2166 if (is_Const(b) && is_const_Phi(a)) { \
2167 /* check for Op(Phi, Const) */ \
2168 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2170 else if (is_Const(a) && is_const_Phi(b)) { \
2171 /* check for Op(Const, Phi) */ \
2172 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2174 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2175 /* check for Op(Phi, Phi) */ \
2176 c = apply_binop_on_2_phis(a, b, eval, mode); \
2179 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2184 #define HANDLE_UNOP_PHI(eval, a, c) \
2187 if (is_const_Phi(a)) { \
2188 /* check for Op(Phi) */ \
2189 c = apply_unop_on_phi(a, eval); \
2191 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2198 * Do the AddSub optimization, then Transform
2199 * Constant folding on Phi
2200 * Add(a,a) -> Mul(a, 2)
2201 * Add(Mul(a, x), a) -> Mul(a, x+1)
2202 * if the mode is integer or float.
2203 * Transform Add(a,-b) into Sub(a,b).
2204 * Reassociation might fold this further.
2206 static ir_node *transform_node_Add(ir_node *n)
2209 ir_node *a, *b, *c, *oldn = n;
2210 vrp_attr *a_vrp, *b_vrp;
2212 n = transform_node_AddSub(n);
2214 a = get_Add_left(n);
2215 b = get_Add_right(n);
2217 mode = get_irn_mode(n);
2219 if (mode_is_reference(mode)) {
2220 ir_mode *lmode = get_irn_mode(a);
2222 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2223 /* an Add(a, NULL) is a hidden Conv */
2224 dbg_info *dbg = get_irn_dbg_info(n);
2225 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2229 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2231 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2232 if (mode_is_float(mode)) {
2233 ir_graph *irg = get_irn_irg(n);
2234 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2238 if (mode_is_num(mode)) {
2239 ir_graph *irg = get_irn_irg(n);
2240 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2241 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2242 && a == b && mode_is_int(mode)) {
2243 ir_node *block = get_nodes_block(n);
2246 get_irn_dbg_info(n),
2249 new_r_Const_long(irg, mode, 2),
2251 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2256 get_irn_dbg_info(n),
2261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2266 get_irn_dbg_info(n),
2271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2274 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2275 /* Here we rely on constants be on the RIGHT side */
2277 ir_node *op = get_Not_op(a);
2279 if (is_Const(b) && is_Const_one(b)) {
2281 ir_node *blk = get_nodes_block(n);
2282 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2283 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2288 n = new_r_Const(irg, get_mode_minus_one(mode));
2289 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2294 ir_node *op = get_Not_op(b);
2298 n = new_r_Const(irg, get_mode_minus_one(mode));
2299 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2306 a_vrp = vrp_get_info(a);
2307 b_vrp = vrp_get_info(b);
2309 if (a_vrp && b_vrp) {
2310 ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2312 if (tarval_is_null(c)) {
2313 dbg_info *dbgi = get_irn_dbg_info(n);
2314 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2318 } /* transform_node_Add */
2321 * returns -cnst or NULL if impossible
2323 static ir_node *const_negate(ir_node *cnst)
2325 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2326 dbg_info *dbgi = get_irn_dbg_info(cnst);
2327 ir_graph *irg = get_irn_irg(cnst);
2328 if (tv == tarval_bad) return NULL;
2329 return new_rd_Const(dbgi, irg, tv);
2333 * Do the AddSub optimization, then Transform
2334 * Constant folding on Phi
2335 * Sub(0,a) -> Minus(a)
2336 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2337 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2338 * Sub(Add(a, x), x) -> a
2339 * Sub(x, Add(x, a)) -> -a
2340 * Sub(x, Const) -> Add(x, -Const)
2342 static ir_node *transform_node_Sub(ir_node *n)
2348 n = transform_node_AddSub(n);
2350 a = get_Sub_left(n);
2351 b = get_Sub_right(n);
2353 mode = get_irn_mode(n);
2355 if (mode_is_int(mode)) {
2356 ir_mode *lmode = get_irn_mode(a);
2358 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2359 /* a Sub(a, NULL) is a hidden Conv */
2360 dbg_info *dbg = get_irn_dbg_info(n);
2361 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2362 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2366 if (mode == lmode &&
2367 get_mode_arithmetic(mode) == irma_twos_complement &&
2369 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2371 dbg_info *dbg = get_irn_dbg_info(n);
2372 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2373 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2379 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2381 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2382 if (mode_is_float(mode)) {
2383 ir_graph *irg = get_irn_irg(n);
2384 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2388 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2389 /* a - C -> a + (-C) */
2390 ir_node *cnst = const_negate(b);
2392 ir_node *block = get_nodes_block(n);
2393 dbg_info *dbgi = get_irn_dbg_info(n);
2395 n = new_rd_Add(dbgi, block, a, cnst, mode);
2396 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2401 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2402 dbg_info *dbg = get_irn_dbg_info(n);
2403 ir_node *block = get_nodes_block(n);
2404 ir_node *left = get_Minus_op(a);
2405 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2407 n = new_rd_Minus(dbg, block, add, mode);
2408 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2410 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2411 dbg_info *dbg = get_irn_dbg_info(n);
2412 ir_node *block = get_nodes_block(n);
2413 ir_node *right = get_Minus_op(b);
2415 n = new_rd_Add(dbg, block, a, right, mode);
2416 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2418 } else if (is_Sub(b)) {
2419 /* a - (b - c) -> a + (c - b)
2420 * -> (a - b) + c iff (b - c) is a pointer */
2421 dbg_info *s_dbg = get_irn_dbg_info(b);
2422 ir_node *s_block = get_nodes_block(b);
2423 ir_node *s_left = get_Sub_left(b);
2424 ir_node *s_right = get_Sub_right(b);
2425 ir_mode *s_mode = get_irn_mode(b);
2426 if (mode_is_reference(s_mode)) {
2427 ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode);
2428 dbg_info *a_dbg = get_irn_dbg_info(n);
2429 ir_node *a_block = get_nodes_block(n);
2432 s_right = new_r_Conv(a_block, s_right, mode);
2433 n = new_rd_Add(a_dbg, a_block, sub, s_right, mode);
2435 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2436 dbg_info *a_dbg = get_irn_dbg_info(n);
2437 ir_node *a_block = get_nodes_block(n);
2439 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2441 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2443 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2444 ir_node *m_right = get_Mul_right(b);
2445 if (is_Const(m_right)) {
2446 ir_node *cnst2 = const_negate(m_right);
2447 if (cnst2 != NULL) {
2448 dbg_info *m_dbg = get_irn_dbg_info(b);
2449 ir_node *m_block = get_nodes_block(b);
2450 ir_node *m_left = get_Mul_left(b);
2451 ir_mode *m_mode = get_irn_mode(b);
2452 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2453 dbg_info *a_dbg = get_irn_dbg_info(n);
2454 ir_node *a_block = get_nodes_block(n);
2456 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2457 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2463 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2464 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2466 get_irn_dbg_info(n),
2470 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2474 if (mode_wrap_around(mode)) {
2475 ir_node *left = get_Add_left(a);
2476 ir_node *right = get_Add_right(a);
2478 /* FIXME: Does the Conv's work only for two complement or generally? */
2480 if (mode != get_irn_mode(right)) {
2481 /* This Sub is an effective Cast */
2482 right = new_r_Conv(get_nodes_block(n), right, mode);
2485 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2487 } else if (right == b) {
2488 if (mode != get_irn_mode(left)) {
2489 /* This Sub is an effective Cast */
2490 left = new_r_Conv(get_nodes_block(n), left, mode);
2493 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2499 if (mode_wrap_around(mode)) {
2500 ir_node *left = get_Add_left(b);
2501 ir_node *right = get_Add_right(b);
2503 /* FIXME: Does the Conv's work only for two complement or generally? */
2505 ir_mode *r_mode = get_irn_mode(right);
2507 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2508 if (mode != r_mode) {
2509 /* This Sub is an effective Cast */
2510 n = new_r_Conv(get_nodes_block(n), n, mode);
2512 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2514 } else if (right == a) {
2515 ir_mode *l_mode = get_irn_mode(left);
2517 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2518 if (mode != l_mode) {
2519 /* This Sub is an effective Cast */
2520 n = new_r_Conv(get_nodes_block(n), n, mode);
2522 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2527 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2528 ir_mode *mode = get_irn_mode(a);
2530 if (mode == get_irn_mode(b)) {
2532 ir_node *op_a = get_Conv_op(a);
2533 ir_node *op_b = get_Conv_op(b);
2535 /* check if it's allowed to skip the conv */
2536 ma = get_irn_mode(op_a);
2537 mb = get_irn_mode(op_b);
2539 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2540 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2543 set_Sub_right(n, b);
2549 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2550 if (!is_reassoc_running() && is_Mul(a)) {
2551 ir_node *ma = get_Mul_left(a);
2552 ir_node *mb = get_Mul_right(a);
2555 ir_node *blk = get_nodes_block(n);
2556 ir_graph *irg = get_irn_irg(n);
2558 get_irn_dbg_info(n),
2562 get_irn_dbg_info(n),
2565 new_r_Const(irg, get_mode_one(mode)),
2568 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2570 } else if (mb == b) {
2571 ir_node *blk = get_nodes_block(n);
2572 ir_graph *irg = get_irn_irg(n);
2574 get_irn_dbg_info(n),
2578 get_irn_dbg_info(n),
2581 new_r_Const(irg, get_mode_one(mode)),
2584 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2588 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2589 ir_node *x = get_Sub_left(a);
2590 ir_node *y = get_Sub_right(a);
2591 ir_node *blk = get_nodes_block(n);
2592 ir_mode *m_b = get_irn_mode(b);
2593 ir_mode *m_y = get_irn_mode(y);
2597 /* Determine the right mode for the Add. */
2600 else if (mode_is_reference(m_b))
2602 else if (mode_is_reference(m_y))
2606 * Both modes are different but none is reference,
2607 * happens for instance in SubP(SubP(P, Iu), Is).
2608 * We have two possibilities here: Cast or ignore.
2609 * Currently we ignore this case.
2614 add = new_r_Add(blk, y, b, add_mode);
2616 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2617 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2621 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2622 if (is_Const(a) && is_Not(b)) {
2623 /* c - ~X = X + (c+1) */
2624 ir_tarval *tv = get_Const_tarval(a);
2626 tv = tarval_add(tv, get_mode_one(mode));
2627 if (tv != tarval_bad) {
2628 ir_node *blk = get_nodes_block(n);
2629 ir_graph *irg = get_irn_irg(n);
2630 ir_node *c = new_r_Const(irg, tv);
2631 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2632 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2638 } /* transform_node_Sub */
2641 * Several transformation done on n*n=2n bits mul.
2642 * These transformations must be done here because new nodes may be produced.
2644 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2647 ir_node *a = get_Mul_left(n);
2648 ir_node *b = get_Mul_right(n);
2649 ir_tarval *ta = value_of(a);
2650 ir_tarval *tb = value_of(b);
2651 ir_mode *smode = get_irn_mode(a);
2653 if (ta == get_mode_one(smode)) {
2654 /* (L)1 * (L)b = (L)b */
2655 ir_node *blk = get_nodes_block(n);
2656 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2657 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2660 else if (ta == get_mode_minus_one(smode)) {
2661 /* (L)-1 * (L)b = (L)b */
2662 ir_node *blk = get_nodes_block(n);
2663 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, 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);
2668 if (tb == get_mode_one(smode)) {
2669 /* (L)a * (L)1 = (L)a */
2670 ir_node *blk = get_irn_n(a, -1);
2671 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2672 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2675 else if (tb == get_mode_minus_one(smode)) {
2676 /* (L)a * (L)-1 = (L)-a */
2677 ir_node *blk = get_nodes_block(n);
2678 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2679 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2680 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2687 * Transform Mul(a,-1) into -a.
2688 * Do constant evaluation of Phi nodes.
2689 * Do architecture dependent optimizations on Mul nodes
2691 static ir_node *transform_node_Mul(ir_node *n)
2693 ir_node *c, *oldn = n;
2694 ir_mode *mode = get_irn_mode(n);
2695 ir_node *a = get_Mul_left(n);
2696 ir_node *b = get_Mul_right(n);
2698 if (is_Bad(a) || is_Bad(b))
2701 if (mode != get_irn_mode(a))
2702 return transform_node_Mul2n(n, mode);
2704 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2706 if (mode_is_signed(mode)) {
2709 if (value_of(a) == get_mode_minus_one(mode))
2711 else if (value_of(b) == get_mode_minus_one(mode))
2714 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2715 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2720 if (is_Const(b)) { /* (-a) * const -> a * -const */
2721 ir_node *cnst = const_negate(b);
2723 dbg_info *dbgi = get_irn_dbg_info(n);
2724 ir_node *block = get_nodes_block(n);
2725 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2726 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2729 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2730 dbg_info *dbgi = get_irn_dbg_info(n);
2731 ir_node *block = get_nodes_block(n);
2732 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2733 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2735 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2736 ir_node *sub_l = get_Sub_left(b);
2737 ir_node *sub_r = get_Sub_right(b);
2738 dbg_info *dbgi = get_irn_dbg_info(n);
2739 ir_node *block = get_nodes_block(n);
2740 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2741 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2742 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2745 } else if (is_Minus(b)) {
2746 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2747 ir_node *sub_l = get_Sub_left(a);
2748 ir_node *sub_r = get_Sub_right(a);
2749 dbg_info *dbgi = get_irn_dbg_info(n);
2750 ir_node *block = get_nodes_block(n);
2751 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2752 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2753 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2756 } else if (is_Shl(a)) {
2757 ir_node *const shl_l = get_Shl_left(a);
2758 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2759 /* (1 << x) * b -> b << x */
2760 dbg_info *const dbgi = get_irn_dbg_info(n);
2761 ir_node *const block = get_nodes_block(n);
2762 ir_node *const shl_r = get_Shl_right(a);
2763 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2764 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2767 } else if (is_Shl(b)) {
2768 ir_node *const shl_l = get_Shl_left(b);
2769 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2770 /* a * (1 << x) -> a << x */
2771 dbg_info *const dbgi = get_irn_dbg_info(n);
2772 ir_node *const block = get_nodes_block(n);
2773 ir_node *const shl_r = get_Shl_right(b);
2774 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2775 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2779 if (get_mode_arithmetic(mode) == irma_ieee754) {
2781 ir_tarval *tv = get_Const_tarval(a);
2782 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2783 && !tarval_is_negative(tv)) {
2784 /* 2.0 * b = b + b */
2785 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2786 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2790 else if (is_Const(b)) {
2791 ir_tarval *tv = get_Const_tarval(b);
2792 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2793 && !tarval_is_negative(tv)) {
2794 /* a * 2.0 = a + a */
2795 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2796 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2801 return arch_dep_replace_mul_with_shifts(n);
2802 } /* transform_node_Mul */
2805 * Transform a Div Node.
2807 static ir_node *transform_node_Div(ir_node *n)
2809 ir_mode *mode = get_Div_resmode(n);
2810 ir_node *a = get_Div_left(n);
2811 ir_node *b = get_Div_right(n);
2813 const ir_node *dummy;
2815 if (mode_is_int(mode)) {
2816 if (is_Const(b) && is_const_Phi(a)) {
2817 /* check for Div(Phi, Const) */
2818 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2820 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2823 } else if (is_Const(a) && is_const_Phi(b)) {
2824 /* check for Div(Const, Phi) */
2825 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2827 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2830 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2831 /* check for Div(Phi, Phi) */
2832 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2834 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2839 if (a == b && value_not_zero(a, &dummy)) {
2840 ir_graph *irg = get_irn_irg(n);
2841 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2842 value = new_r_Const(irg, get_mode_one(mode));
2843 DBG_OPT_CSTEVAL(n, value);
2846 if (mode_is_signed(mode) && is_Const(b)) {
2847 ir_tarval *tv = get_Const_tarval(b);
2849 if (tv == get_mode_minus_one(mode)) {
2851 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2852 DBG_OPT_CSTEVAL(n, value);
2856 /* Try architecture dependent optimization */
2857 value = arch_dep_replace_div_by_const(n);
2860 assert(mode_is_float(mode));
2862 /* Optimize x/c to x*(1/c) */
2863 if (get_mode_arithmetic(mode) == irma_ieee754) {
2864 ir_tarval *tv = value_of(b);
2866 if (tv != tarval_bad) {
2867 int rem = tarval_fp_ops_enabled();
2870 * Floating point constant folding might be disabled here to
2872 * However, as we check for exact result, doing it is safe.
2875 tarval_enable_fp_ops(1);
2876 tv = tarval_div(get_mode_one(mode), tv);
2877 tarval_enable_fp_ops(rem);
2879 /* Do the transformation if the result is either exact or we are
2880 not using strict rules. */
2881 if (tv != tarval_bad &&
2882 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2883 ir_node *block = get_nodes_block(n);
2884 ir_graph *irg = get_irn_irg(block);
2885 ir_node *c = new_r_Const(irg, tv);
2886 dbg_info *dbgi = get_irn_dbg_info(n);
2887 value = new_rd_Mul(dbgi, block, a, c, mode);
2900 /* Turn Div into a tuple (mem, jmp, bad, value) */
2901 mem = get_Div_mem(n);
2902 blk = get_nodes_block(n);
2903 irg = get_irn_irg(blk);
2905 /* skip a potential Pin */
2906 mem = skip_Pin(mem);
2907 turn_into_tuple(n, pn_Div_max);
2908 set_Tuple_pred(n, pn_Div_M, mem);
2909 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2910 set_Tuple_pred(n, pn_Div_X_except, get_irg_bad(irg));
2911 set_Tuple_pred(n, pn_Div_res, value);
2914 } /* transform_node_Div */
2917 * Transform a Mod node.
2919 static ir_node *transform_node_Mod(ir_node *n)
2921 ir_mode *mode = get_Mod_resmode(n);
2922 ir_node *a = get_Mod_left(n);
2923 ir_node *b = get_Mod_right(n);
2928 if (is_Const(b) && is_const_Phi(a)) {
2929 /* check for Div(Phi, Const) */
2930 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2932 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2936 else if (is_Const(a) && is_const_Phi(b)) {
2937 /* check for Div(Const, Phi) */
2938 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2940 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2944 else if (is_const_Phi(a) && is_const_Phi(b)) {
2945 /* check for Div(Phi, Phi) */
2946 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2948 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2955 irg = get_irn_irg(n);
2956 if (tv != tarval_bad) {
2957 value = new_r_Const(irg, tv);
2959 DBG_OPT_CSTEVAL(n, value);
2962 ir_node *a = get_Mod_left(n);
2963 ir_node *b = get_Mod_right(n);
2964 const ir_node *dummy;
2966 if (a == b && value_not_zero(a, &dummy)) {
2967 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2968 value = new_r_Const(irg, get_mode_null(mode));
2969 DBG_OPT_CSTEVAL(n, value);
2972 if (mode_is_signed(mode) && is_Const(b)) {
2973 ir_tarval *tv = get_Const_tarval(b);
2975 if (tv == get_mode_minus_one(mode)) {
2977 value = new_r_Const(irg, get_mode_null(mode));
2978 DBG_OPT_CSTEVAL(n, value);
2982 /* Try architecture dependent optimization */
2983 value = arch_dep_replace_mod_by_const(n);
2992 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2993 mem = get_Mod_mem(n);
2994 blk = get_nodes_block(n);
2995 irg = get_irn_irg(blk);
2997 /* skip a potential Pin */
2998 mem = skip_Pin(mem);
2999 turn_into_tuple(n, pn_Mod_max);
3000 set_Tuple_pred(n, pn_Mod_M, mem);
3001 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
3002 set_Tuple_pred(n, pn_Mod_X_except, get_irg_bad(irg));
3003 set_Tuple_pred(n, pn_Mod_res, value);
3006 } /* transform_node_Mod */
3009 * Optimize -a CMP -b into b CMP a.
3010 * This works only for for modes where unary Minus
3012 * Note that two-complement integers can Overflow
3013 * so it will NOT work.
3015 * For == and != can be handled in Proj(Cmp)
3017 static ir_node *transform_node_Cmp(ir_node *n)
3020 ir_node *left = get_Cmp_left(n);
3021 ir_node *right = get_Cmp_right(n);
3023 if (is_Minus(left) && is_Minus(right) &&
3024 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
3025 ir_node *const new_left = get_Minus_op(right);
3026 ir_node *const new_right = get_Minus_op(left);
3027 n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right);
3028 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP);
3031 } /* transform_node_Cmp */
3035 * Transform a Cond node.
3037 * Replace the Cond by a Jmp if it branches on a constant
3040 static ir_node *transform_node_Cond(ir_node *n)
3043 ir_node *a = get_Cond_selector(n);
3044 ir_tarval *ta = value_of(a);
3045 ir_graph *irg = get_irn_irg(n);
3048 /* we need block info which is not available in floating irgs */
3049 if (get_irg_pinned(irg) == op_pin_state_floats)
3052 if ((ta != tarval_bad) &&
3053 (get_irn_mode(a) == mode_b) &&
3054 (get_opt_unreachable_code())) {
3055 /* It's a boolean Cond, branching on a boolean constant.
3056 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
3057 ir_node *blk = get_nodes_block(n);
3058 jmp = new_r_Jmp(blk);
3059 turn_into_tuple(n, pn_Cond_max);
3060 if (ta == tarval_b_true) {
3061 set_Tuple_pred(n, pn_Cond_false, get_irg_bad(irg));
3062 set_Tuple_pred(n, pn_Cond_true, jmp);
3064 set_Tuple_pred(n, pn_Cond_false, jmp);
3065 set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg));
3067 /* We might generate an endless loop, so keep it alive. */
3068 add_End_keepalive(get_irg_end(irg), blk);
3071 } /* transform_node_Cond */
3074 * Prototype of a recursive transform function
3075 * for bitwise distributive transformations.
3077 typedef ir_node* (*recursive_transform)(ir_node *n);
3080 * makes use of distributive laws for and, or, eor
3081 * and(a OP c, b OP c) -> and(a, b) OP c
3082 * note, might return a different op than n
3084 static ir_node *transform_bitwise_distributive(ir_node *n,
3085 recursive_transform trans_func)
3088 ir_node *a = get_binop_left(n);
3089 ir_node *b = get_binop_right(n);
3090 ir_op *op = get_irn_op(a);
3091 ir_op *op_root = get_irn_op(n);
3093 if (op != get_irn_op(b))
3096 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
3097 if (op == op_Conv) {
3098 ir_node *a_op = get_Conv_op(a);
3099 ir_node *b_op = get_Conv_op(b);
3100 ir_mode *a_mode = get_irn_mode(a_op);
3101 ir_mode *b_mode = get_irn_mode(b_op);
3102 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
3103 ir_node *blk = get_nodes_block(n);
3106 set_binop_left(n, a_op);
3107 set_binop_right(n, b_op);
3108 set_irn_mode(n, a_mode);
3110 n = new_r_Conv(blk, n, get_irn_mode(oldn));
3112 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
3118 /* nothing to gain here */
3122 if (op == op_Shrs || op == op_Shr || op == op_Shl
3123 || op == op_And || op == op_Or || op == op_Eor) {
3124 ir_node *a_left = get_binop_left(a);
3125 ir_node *a_right = get_binop_right(a);
3126 ir_node *b_left = get_binop_left(b);
3127 ir_node *b_right = get_binop_right(b);
3129 ir_node *op1 = NULL;
3130 ir_node *op2 = NULL;
3132 if (is_op_commutative(op)) {
3133 if (a_left == b_left) {
3137 } else if (a_left == b_right) {
3141 } else if (a_right == b_left) {
3147 if (a_right == b_right) {
3154 /* (a sop c) & (b sop c) => (a & b) sop c */
3155 ir_node *blk = get_nodes_block(n);
3157 ir_node *new_n = exact_copy(n);
3158 set_binop_left(new_n, op1);
3159 set_binop_right(new_n, op2);
3160 new_n = trans_func(new_n);
3162 if (op_root == op_Eor && op == op_Or) {
3163 dbg_info *dbgi = get_irn_dbg_info(n);
3164 ir_mode *mode = get_irn_mode(c);
3166 c = new_rd_Not(dbgi, blk, c, mode);
3167 n = new_rd_And(dbgi, blk, new_n, c, mode);
3170 set_nodes_block(n, blk);
3171 set_binop_left(n, new_n);
3172 set_binop_right(n, c);
3176 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3187 static ir_node *transform_node_And(ir_node *n)
3189 ir_node *c, *oldn = n;
3190 ir_node *a = get_And_left(n);
3191 ir_node *b = get_And_right(n);
3193 vrp_attr *a_vrp, *b_vrp;
3195 mode = get_irn_mode(n);
3196 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3198 /* we can evaluate 2 Projs of the same Cmp */
3199 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3200 ir_node *pred_a = get_Proj_pred(a);
3201 ir_node *pred_b = get_Proj_pred(b);
3202 if (pred_a == pred_b) {
3203 dbg_info *dbgi = get_irn_dbg_info(n);
3204 pn_Cmp pn_a = get_Proj_pn_cmp(a);
3205 pn_Cmp pn_b = get_Proj_pn_cmp(b);
3206 /* yes, we can simply calculate with pncs */
3207 pn_Cmp new_pnc = pn_a & pn_b;
3209 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3214 ir_node *op = get_Not_op(b);
3216 ir_node *ba = get_And_left(op);
3217 ir_node *bb = get_And_right(op);
3219 /* it's enough to test the following cases due to normalization! */
3220 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3221 /* (a|b) & ~(a&b) = a^b */
3222 ir_node *block = get_nodes_block(n);
3224 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3225 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3233 ir_node *op = get_Not_op(a);
3235 ir_node *aa = get_And_left(op);
3236 ir_node *ab = get_And_right(op);
3238 /* it's enough to test the following cases due to normalization! */
3239 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3240 /* (a|b) & ~(a&b) = a^b */
3241 ir_node *block = get_nodes_block(n);
3243 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3244 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3251 ir_node *al = get_Eor_left(a);
3252 ir_node *ar = get_Eor_right(a);
3255 /* (b ^ a) & b -> ~a & b */
3256 dbg_info *dbg = get_irn_dbg_info(n);
3257 ir_node *block = get_nodes_block(n);
3259 ar = new_rd_Not(dbg, block, ar, mode);
3260 n = new_rd_And(dbg, block, ar, b, mode);
3261 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3265 /* (a ^ b) & b -> ~a & b */
3266 dbg_info *dbg = get_irn_dbg_info(n);
3267 ir_node *block = get_nodes_block(n);
3269 al = new_rd_Not(dbg, block, al, mode);
3270 n = new_rd_And(dbg, block, al, b, mode);
3271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3276 ir_node *bl = get_Eor_left(b);
3277 ir_node *br = get_Eor_right(b);
3280 /* a & (a ^ b) -> a & ~b */
3281 dbg_info *dbg = get_irn_dbg_info(n);
3282 ir_node *block = get_nodes_block(n);
3284 br = new_rd_Not(dbg, block, br, mode);
3285 n = new_rd_And(dbg, block, br, a, mode);
3286 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3290 /* a & (b ^ a) -> a & ~b */
3291 dbg_info *dbg = get_irn_dbg_info(n);
3292 ir_node *block = get_nodes_block(n);
3294 bl = new_rd_Not(dbg, block, bl, mode);
3295 n = new_rd_And(dbg, block, bl, a, mode);
3296 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3300 if (is_Not(a) && is_Not(b)) {
3301 /* ~a & ~b = ~(a|b) */
3302 ir_node *block = get_nodes_block(n);
3303 ir_mode *mode = get_irn_mode(n);
3307 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3308 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3309 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3313 b_vrp = vrp_get_info(b);
3314 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3315 b_vrp->bits_not_set), get_Const_tarval(a)) == pn_Cmp_Eq)) {
3321 a_vrp = vrp_get_info(a);
3322 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3323 a_vrp->bits_not_set), get_Const_tarval(b)) == pn_Cmp_Eq)) {
3327 n = transform_bitwise_distributive(n, transform_node_And);
3330 } /* transform_node_And */
3332 /* the order of the values is important! */
3333 typedef enum const_class {
3339 static const_class classify_const(const ir_node* n)
3341 if (is_Const(n)) return const_const;
3342 if (is_irn_constlike(n)) return const_like;
3347 * Determines whether r is more constlike or has a larger index (in that order)
3350 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3352 const const_class l_order = classify_const(l);
3353 const const_class r_order = classify_const(r);
3355 l_order > r_order ||
3356 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3362 static ir_node *transform_node_Eor(ir_node *n)
3364 ir_node *c, *oldn = n;
3365 ir_node *a = get_Eor_left(n);
3366 ir_node *b = get_Eor_right(n);
3367 ir_mode *mode = get_irn_mode(n);
3369 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3371 /* we can evaluate 2 Projs of the same Cmp */
3372 if (mode == mode_b && is_Proj(a) && is_Proj(b)) {
3373 ir_node *pred_a = get_Proj_pred(a);
3374 ir_node *pred_b = get_Proj_pred(b);
3375 if (pred_a == pred_b) {
3376 dbg_info *dbgi = get_irn_dbg_info(n);
3377 pn_Cmp pn_a = get_Proj_pn_cmp(a);
3378 pn_Cmp pn_b = get_Proj_pn_cmp(b);
3379 /* yes, we can simply calculate with pncs */
3380 pn_Cmp new_pnc = pn_a ^ pn_b;
3382 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
3386 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3387 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3388 dbg_info *dbg = get_irn_dbg_info(n);
3389 ir_node *block = get_nodes_block(n);
3390 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3391 ir_node *new_left = get_Not_op(a);
3392 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3393 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3395 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3396 dbg_info *dbg = get_irn_dbg_info(n);
3397 ir_node *block = get_nodes_block(n);
3398 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3399 ir_node *new_right = get_Not_op(b);
3400 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3401 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3405 /* x ^ 1...1 -> ~1 */
3406 if (is_Const(b) && is_Const_all_one(b)) {
3407 n = new_r_Not(get_nodes_block(n), a, mode);
3408 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3412 n = transform_bitwise_distributive(n, transform_node_Eor);
3414 } /* transform_node_Eor */
3419 static ir_node *transform_node_Not(ir_node *n)
3421 ir_node *c, *oldn = n;
3422 ir_node *a = get_Not_op(n);
3423 ir_mode *mode = get_irn_mode(n);
3425 HANDLE_UNOP_PHI(tarval_not,a,c);
3427 /* check for a boolean Not */
3428 if (mode == mode_b && is_Proj(a)) {
3429 ir_node *a_pred = get_Proj_pred(a);
3430 if (is_Cmp(a_pred)) {
3431 /* We negate a Cmp. The Cmp has the negated result anyways! */
3432 n = new_r_Proj(get_Proj_pred(a),
3433 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
3434 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3439 /* normalize ~(a ^ b) => a ^ ~b */
3441 dbg_info *dbg = get_irn_dbg_info(n);
3442 ir_node *block = get_nodes_block(n);
3443 ir_node *eor_right = get_Eor_right(a);
3444 ir_node *eor_left = get_Eor_left(a);
3445 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3446 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3450 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3451 if (is_Minus(a)) { /* ~-x -> x + -1 */
3452 dbg_info *dbg = get_irn_dbg_info(n);
3453 ir_graph *irg = get_irn_irg(n);
3454 ir_node *block = get_nodes_block(n);
3455 ir_node *add_l = get_Minus_op(a);
3456 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3457 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3458 } else if (is_Add(a)) {
3459 ir_node *add_r = get_Add_right(a);
3460 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3461 /* ~(x + -1) = -x */
3462 ir_node *op = get_Add_left(a);
3463 ir_node *blk = get_nodes_block(n);
3464 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3465 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3470 } /* transform_node_Not */
3473 * Transform a Minus.
3477 * -(a >>u (size-1)) = a >>s (size-1)
3478 * -(a >>s (size-1)) = a >>u (size-1)
3479 * -(a * const) -> a * -const
3481 static ir_node *transform_node_Minus(ir_node *n)
3483 ir_node *c, *oldn = n;
3484 ir_node *a = get_Minus_op(n);
3487 HANDLE_UNOP_PHI(tarval_neg,a,c);
3489 mode = get_irn_mode(a);
3490 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3491 /* the following rules are only to twos-complement */
3494 ir_node *op = get_Not_op(a);
3495 ir_tarval *tv = get_mode_one(mode);
3496 ir_node *blk = get_nodes_block(n);
3497 ir_graph *irg = get_irn_irg(blk);
3498 ir_node *c = new_r_Const(irg, tv);
3499 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3500 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3504 ir_node *c = get_Shr_right(a);
3507 ir_tarval *tv = get_Const_tarval(c);
3509 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3510 /* -(a >>u (size-1)) = a >>s (size-1) */
3511 ir_node *v = get_Shr_left(a);
3513 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3514 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3520 ir_node *c = get_Shrs_right(a);
3523 ir_tarval *tv = get_Const_tarval(c);
3525 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3526 /* -(a >>s (size-1)) = a >>u (size-1) */
3527 ir_node *v = get_Shrs_left(a);
3529 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3530 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3537 /* - (a-b) = b - a */
3538 ir_node *la = get_Sub_left(a);
3539 ir_node *ra = get_Sub_right(a);
3540 ir_node *blk = get_nodes_block(n);
3542 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3543 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3547 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3548 ir_node *mul_l = get_Mul_left(a);
3549 ir_node *mul_r = get_Mul_right(a);
3550 ir_tarval *tv = value_of(mul_r);
3551 if (tv != tarval_bad) {
3552 tv = tarval_neg(tv);
3553 if (tv != tarval_bad) {
3554 ir_graph *irg = get_irn_irg(n);
3555 ir_node *cnst = new_r_Const(irg, tv);
3556 dbg_info *dbg = get_irn_dbg_info(a);
3557 ir_node *block = get_nodes_block(a);
3558 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3559 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3566 } /* transform_node_Minus */
3569 * Transform a Proj(Load) with a non-null address.
3571 static ir_node *transform_node_Proj_Load(ir_node *proj)
3573 if (get_opt_ldst_only_null_ptr_exceptions()) {
3574 if (get_irn_mode(proj) == mode_X) {
3575 ir_node *load = get_Proj_pred(proj);
3577 /* get the Load address */
3578 const ir_node *addr = get_Load_ptr(load);
3579 const ir_node *confirm;
3581 if (value_not_null(addr, &confirm)) {
3582 if (confirm == NULL) {
3583 /* this node may float if it did not depend on a Confirm */
3584 set_irn_pinned(load, op_pin_state_floats);
3586 if (get_Proj_proj(proj) == pn_Load_X_except) {
3587 ir_graph *irg = get_irn_irg(proj);
3588 DBG_OPT_EXC_REM(proj);
3589 return get_irg_bad(irg);
3591 ir_node *blk = get_nodes_block(load);
3592 return new_r_Jmp(blk);
3598 } /* transform_node_Proj_Load */
3601 * Transform a Proj(Store) with a non-null address.
3603 static ir_node *transform_node_Proj_Store(ir_node *proj)
3605 if (get_opt_ldst_only_null_ptr_exceptions()) {
3606 if (get_irn_mode(proj) == mode_X) {
3607 ir_node *store = get_Proj_pred(proj);
3609 /* get the load/store address */
3610 const ir_node *addr = get_Store_ptr(store);
3611 const ir_node *confirm;
3613 if (value_not_null(addr, &confirm)) {
3614 if (confirm == NULL) {
3615 /* this node may float if it did not depend on a Confirm */
3616 set_irn_pinned(store, op_pin_state_floats);
3618 if (get_Proj_proj(proj) == pn_Store_X_except) {
3619 ir_graph *irg = get_irn_irg(proj);
3620 DBG_OPT_EXC_REM(proj);
3621 return get_irg_bad(irg);
3623 ir_node *blk = get_nodes_block(store);
3624 return new_r_Jmp(blk);
3630 } /* transform_node_Proj_Store */
3633 * Transform a Proj(Div) with a non-zero value.
3634 * Removes the exceptions and routes the memory to the NoMem node.
3636 static ir_node *transform_node_Proj_Div(ir_node *proj)
3638 ir_node *div = get_Proj_pred(proj);
3639 ir_node *b = get_Div_right(div);
3640 ir_node *res, *new_mem;
3641 const ir_node *confirm;
3644 if (value_not_zero(b, &confirm)) {
3645 /* div(x, y) && y != 0 */
3646 if (confirm == NULL) {
3647 /* we are sure we have a Const != 0 */
3648 new_mem = get_Div_mem(div);
3649 new_mem = skip_Pin(new_mem);
3650 set_Div_mem(div, new_mem);
3651 set_irn_pinned(div, op_pin_state_floats);
3654 proj_nr = get_Proj_proj(proj);
3656 case pn_Div_X_regular:
3657 return new_r_Jmp(get_nodes_block(div));
3659 case pn_Div_X_except: {
3660 ir_graph *irg = get_irn_irg(proj);
3661 /* we found an exception handler, remove it */
3662 DBG_OPT_EXC_REM(proj);
3663 return get_irg_bad(irg);
3667 ir_graph *irg = get_irn_irg(proj);
3668 res = get_Div_mem(div);
3669 new_mem = get_irg_no_mem(irg);
3672 /* This node can only float up to the Confirm block */
3673 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3675 set_irn_pinned(div, op_pin_state_floats);
3676 /* this is a Div without exception, we can remove the memory edge */
3677 set_Div_mem(div, new_mem);
3683 } /* transform_node_Proj_Div */
3686 * Transform a Proj(Mod) with a non-zero value.
3687 * Removes the exceptions and routes the memory to the NoMem node.
3689 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3691 ir_node *mod = get_Proj_pred(proj);
3692 ir_node *b = get_Mod_right(mod);
3693 ir_node *res, *new_mem;
3694 const ir_node *confirm;
3697 if (value_not_zero(b, &confirm)) {
3698 /* mod(x, y) && y != 0 */
3699 proj_nr = get_Proj_proj(proj);
3701 if (confirm == NULL) {
3702 /* we are sure we have a Const != 0 */
3703 new_mem = get_Mod_mem(mod);
3704 new_mem = skip_Pin(new_mem);
3705 set_Mod_mem(mod, new_mem);
3706 set_irn_pinned(mod, op_pin_state_floats);
3711 case pn_Mod_X_regular:
3712 return new_r_Jmp(get_irn_n(mod, -1));
3714 case pn_Mod_X_except: {
3715 ir_graph *irg = get_irn_irg(proj);
3716 /* we found an exception handler, remove it */
3717 DBG_OPT_EXC_REM(proj);
3718 return get_irg_bad(irg);
3722 ir_graph *irg = get_irn_irg(proj);
3723 res = get_Mod_mem(mod);
3724 new_mem = get_irg_no_mem(irg);
3727 /* This node can only float up to the Confirm block */
3728 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3730 /* this is a Mod without exception, we can remove the memory edge */
3731 set_Mod_mem(mod, new_mem);
3735 if (get_Mod_left(mod) == b) {
3736 /* a % a = 0 if a != 0 */
3737 ir_graph *irg = get_irn_irg(proj);
3738 ir_mode *mode = get_irn_mode(proj);
3739 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3741 DBG_OPT_CSTEVAL(mod, res);
3747 } /* transform_node_Proj_Mod */
3750 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3752 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3754 if (get_opt_unreachable_code()) {
3755 ir_node *n = get_Proj_pred(proj);
3756 ir_node *b = get_Cond_selector(n);
3758 if (mode_is_int(get_irn_mode(b))) {
3759 ir_tarval *tb = value_of(b);
3761 if (tb != tarval_bad) {
3762 /* we have a constant switch */
3763 long num = get_Proj_proj(proj);
3765 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3766 if (get_tarval_long(tb) == num) {
3767 /* Do NOT create a jump here, or we will have 2 control flow ops
3768 * in a block. This case is optimized away in optimize_cf(). */
3771 ir_graph *irg = get_irn_irg(proj);
3772 /* this case will NEVER be taken, kill it */
3773 return get_irg_bad(irg);
3777 long num = get_Proj_proj(proj);
3778 vrp_attr *b_vrp = vrp_get_info(b);
3779 if (num != get_Cond_default_proj(n) && b_vrp) {
3780 /* Try handling with vrp data. We only remove dead parts. */
3781 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3783 if (b_vrp->range_type == VRP_RANGE) {
3784 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3785 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3787 if ((cmp_result & pn_Cmp_Gt) == cmp_result && (cmp_result2
3788 & pn_Cmp_Lt) == cmp_result2) {
3789 ir_graph *irg = get_irn_irg(proj);
3790 return get_irg_bad(irg);
3792 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3793 pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3794 pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3796 if ((cmp_result & pn_Cmp_Le) == cmp_result && (cmp_result2
3797 & pn_Cmp_Ge) == cmp_result2) {
3798 ir_graph *irg = get_irn_irg(proj);
3799 return get_irg_bad(irg);
3804 tarval_and( b_vrp->bits_set, tp),
3807 ir_graph *irg = get_irn_irg(proj);
3808 return get_irg_bad(irg);
3814 tarval_not(b_vrp->bits_not_set)),
3815 tarval_not(b_vrp->bits_not_set))
3817 ir_graph *irg = get_irn_irg(proj);
3818 return get_irg_bad(irg);
3827 } /* transform_node_Proj_Cond */
3830 * Create a 0 constant of given mode.
3832 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
3834 ir_tarval *tv = get_mode_null(mode);
3835 ir_node *cnst = new_r_Const(irg, tv);
3841 * Normalizes and optimizes Cmp nodes.
3843 static ir_node *transform_node_Proj_Cmp(ir_node *proj)
3845 ir_node *n = get_Proj_pred(proj);
3846 ir_node *left = get_Cmp_left(n);
3847 ir_node *right = get_Cmp_right(n);
3848 ir_tarval *tv = NULL;
3850 ir_mode *mode = NULL;
3851 long proj_nr = get_Proj_proj(proj);
3853 /* we can evaluate some cases directly */
3855 case pn_Cmp_False: {
3856 ir_graph *irg = get_irn_irg(proj);
3857 return new_r_Const(irg, get_tarval_b_false());
3860 ir_graph *irg = get_irn_irg(proj);
3861 return new_r_Const(irg, get_tarval_b_true());
3864 if (!mode_is_float(get_irn_mode(left))) {
3865 ir_graph *irg = get_irn_irg(proj);
3866 return new_r_Const(irg, get_tarval_b_true());
3873 /* remove Casts of both sides */
3874 left = skip_Cast(left);
3875 right = skip_Cast(right);
3877 /* Remove unnecessary conversions */
3878 /* TODO handle constants */
3879 if (is_Conv(left) && is_Conv(right)) {
3880 ir_mode *mode = get_irn_mode(left);
3881 ir_node *op_left = get_Conv_op(left);
3882 ir_node *op_right = get_Conv_op(right);
3883 ir_mode *mode_left = get_irn_mode(op_left);
3884 ir_mode *mode_right = get_irn_mode(op_right);
3886 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3887 && mode_left != mode_b && mode_right != mode_b) {
3888 ir_node *block = get_nodes_block(n);
3890 if (mode_left == mode_right) {
3894 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3895 } else if (smaller_mode(mode_left, mode_right)) {
3896 left = new_r_Conv(block, op_left, mode_right);
3899 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3900 } else if (smaller_mode(mode_right, mode_left)) {
3902 right = new_r_Conv(block, op_right, mode_left);
3904 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3909 /* remove operation on both sides if possible */
3910 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
3912 * The following operations are NOT safe for floating point operations, for instance
3913 * 1.0 + inf == 2.0 + inf, =/=> x == y
3915 if (mode_is_int(get_irn_mode(left))) {
3916 unsigned lop = get_irn_opcode(left);
3918 if (lop == get_irn_opcode(right)) {
3919 ir_node *ll, *lr, *rl, *rr;
3921 /* same operation on both sides, try to remove */
3925 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3926 left = get_unop_op(left);
3927 right = get_unop_op(right);
3929 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3932 ll = get_Add_left(left);
3933 lr = get_Add_right(left);
3934 rl = get_Add_left(right);
3935 rr = get_Add_right(right);
3938 /* X + a CMP X + b ==> a CMP b */
3942 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3943 } else if (ll == rr) {
3944 /* X + a CMP b + X ==> a CMP b */
3948 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3949 } else if (lr == rl) {
3950 /* a + X CMP X + b ==> a CMP b */
3954 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3955 } else if (lr == rr) {
3956 /* a + X CMP b + X ==> a CMP b */
3960 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3964 ll = get_Sub_left(left);
3965 lr = get_Sub_right(left);
3966 rl = get_Sub_left(right);
3967 rr = get_Sub_right(right);
3970 /* X - a CMP X - b ==> a CMP b */
3974 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3975 } else if (lr == rr) {
3976 /* a - X CMP b - X ==> a CMP b */
3980 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3984 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3985 /* a ROTL X CMP b ROTL X ==> a CMP b */
3986 left = get_Rotl_left(left);
3987 right = get_Rotl_left(right);
3989 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3997 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3998 if (is_Add(left) || is_Sub(left)) {
3999 ir_node *ll = get_binop_left(left);
4000 ir_node *lr = get_binop_right(left);
4002 if (lr == right && is_Add(left)) {
4008 ir_graph *irg = get_irn_irg(n);
4010 right = create_zero_const(irg, get_irn_mode(left));
4012 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4015 if (is_Add(right) || is_Sub(right)) {
4016 ir_node *rl = get_binop_left(right);
4017 ir_node *rr = get_binop_right(right);
4019 if (rr == left && is_Add(right)) {
4025 ir_graph *irg = get_irn_irg(n);
4027 right = create_zero_const(irg, get_irn_mode(left));
4029 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
4032 if (is_And(left) && is_Const(right)) {
4033 ir_node *ll = get_binop_left(left);
4034 ir_node *lr = get_binop_right(left);
4035 if (is_Shr(ll) && is_Const(lr)) {
4036 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
4037 ir_node *block = get_nodes_block(n);
4038 ir_mode *mode = get_irn_mode(left);
4040 ir_node *llr = get_Shr_right(ll);
4041 if (is_Const(llr)) {
4042 dbg_info *dbg = get_irn_dbg_info(left);
4043 ir_graph *irg = get_irn_irg(left);
4045 ir_tarval *c1 = get_Const_tarval(llr);
4046 ir_tarval *c2 = get_Const_tarval(lr);
4047 ir_tarval *c3 = get_Const_tarval(right);
4048 ir_tarval *mask = tarval_shl(c2, c1);
4049 ir_tarval *value = tarval_shl(c3, c1);
4051 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
4052 right = new_r_Const(irg, value);
4057 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
4059 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
4060 right = get_Eor_right(left);
4061 left = get_Eor_left(left);
4064 } /* mode_is_int(...) */
4065 } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */
4067 /* replace mode_b compares with ands/ors */
4068 if (get_irn_mode(left) == mode_b) {
4069 ir_node *block = get_nodes_block(n);
4073 case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break;
4074 case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break;
4075 case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break;
4076 case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break;
4077 case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break;
4078 case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break;
4079 default: bres = NULL;
4082 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4088 * First step: normalize the compare op
4089 * by placing the constant on the right side
4090 * or moving the lower address node to the left.
4092 if (!operands_are_normalized(left, right)) {
4098 proj_nr = get_inversed_pnc(proj_nr);
4103 * Second step: Try to reduce the magnitude
4104 * of a constant. This may help to generate better code
4105 * later and may help to normalize more compares.
4106 * Of course this is only possible for integer values.
4108 tv = value_of(right);
4109 if (tv != tarval_bad) {
4110 mode = get_irn_mode(right);
4112 /* TODO extend to arbitrary constants */
4113 if (is_Conv(left) && tarval_is_null(tv)) {
4114 ir_node *op = get_Conv_op(left);
4115 ir_mode *op_mode = get_irn_mode(op);
4118 * UpConv(x) REL 0 ==> x REL 0
4119 * Don't do this for float values as it's unclear whether it is a
4120 * win. (on the other side it makes detection/creation of fabs hard)
4122 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4123 ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) ||
4124 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4125 !mode_is_float(mode)) {
4126 tv = get_mode_null(op_mode);
4130 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4134 if (tv != tarval_bad) {
4135 /* the following optimization is possible on modes without Overflow
4136 * on Unary Minus or on == and !=:
4137 * -a CMP c ==> a swap(CMP) -c
4139 * Beware: for two-complement Overflow may occur, so only == and != can
4140 * be optimized, see this:
4141 * -MININT < 0 =/=> MININT > 0 !!!
4143 if (is_Minus(left) &&
4144 (!mode_overflow_on_unary_Minus(mode) ||
4145 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
4146 tv = tarval_neg(tv);
4148 if (tv != tarval_bad) {
4149 left = get_Minus_op(left);
4150 proj_nr = get_inversed_pnc(proj_nr);
4152 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4154 } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) {
4155 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4156 tv = tarval_not(tv);
4158 if (tv != tarval_bad) {
4159 left = get_Not_op(left);
4161 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4165 /* for integer modes, we have more */
4166 if (mode_is_int(mode)) {
4167 /* Ne includes Unordered which is not possible on integers.
4168 * However, frontends often use this wrong, so fix it here */
4169 if (proj_nr & pn_Cmp_Uo) {
4170 proj_nr &= ~pn_Cmp_Uo;
4171 set_Proj_proj(proj, proj_nr);
4174 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4175 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
4176 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
4177 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4179 if (tv != tarval_bad) {
4180 proj_nr ^= pn_Cmp_Eq;
4182 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4185 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4186 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
4187 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
4188 tv = tarval_add(tv, get_mode_one(mode));
4190 if (tv != tarval_bad) {
4191 proj_nr ^= pn_Cmp_Eq;
4193 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4197 /* the following reassociations work only for == and != */
4198 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4200 #if 0 /* Might be not that good in general */
4201 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4202 if (tarval_is_null(tv) && is_Sub(left)) {
4203 right = get_Sub_right(left);
4204 left = get_Sub_left(left);
4206 tv = value_of(right);
4208 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4212 if (tv != tarval_bad) {
4213 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4215 ir_node *c1 = get_Sub_right(left);
4216 ir_tarval *tv2 = value_of(c1);
4218 if (tv2 != tarval_bad) {
4219 tv2 = tarval_add(tv, value_of(c1));
4221 if (tv2 != tarval_bad) {
4222 left = get_Sub_left(left);
4225 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4229 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4230 else if (is_Add(left)) {
4231 ir_node *a_l = get_Add_left(left);
4232 ir_node *a_r = get_Add_right(left);
4236 if (is_Const(a_l)) {
4238 tv2 = value_of(a_l);
4241 tv2 = value_of(a_r);
4244 if (tv2 != tarval_bad) {
4245 tv2 = tarval_sub(tv, tv2, NULL);
4247 if (tv2 != tarval_bad) {
4251 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4255 /* -a == c ==> a == -c, -a != c ==> a != -c */
4256 else if (is_Minus(left)) {
4257 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4259 if (tv2 != tarval_bad) {
4260 left = get_Minus_op(left);
4263 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4270 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
4271 switch (get_irn_opcode(left)) {
4275 c1 = get_And_right(left);
4278 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4279 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4281 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4283 /* TODO: move to constant evaluation */
4284 ir_graph *irg = get_irn_irg(n);
4285 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4286 c1 = new_r_Const(irg, tv);
4287 DBG_OPT_CSTEVAL(proj, c1);
4291 if (tarval_is_single_bit(tv)) {
4293 * optimization for AND:
4295 * And(x, C) == C ==> And(x, C) != 0
4296 * And(x, C) != C ==> And(X, C) == 0
4298 * if C is a single Bit constant.
4301 /* check for Constant's match. We have check hare the tarvals,
4302 because our const might be changed */
4303 if (get_Const_tarval(c1) == tv) {
4304 /* fine: do the transformation */
4305 tv = get_mode_null(get_tarval_mode(tv));
4306 proj_nr ^= pn_Cmp_Leg;
4308 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4314 c1 = get_Or_right(left);
4315 if (is_Const(c1) && tarval_is_null(tv)) {
4317 * Or(x, C) == 0 && C != 0 ==> FALSE
4318 * Or(x, C) != 0 && C != 0 ==> TRUE
4320 if (! tarval_is_null(get_Const_tarval(c1))) {
4321 /* TODO: move to constant evaluation */
4322 ir_graph *irg = get_irn_irg(n);
4323 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4324 c1 = new_r_Const(irg, tv);
4325 DBG_OPT_CSTEVAL(proj, c1);
4332 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4334 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4337 c1 = get_Shl_right(left);
4339 ir_graph *irg = get_irn_irg(c1);
4340 ir_tarval *tv1 = get_Const_tarval(c1);
4341 ir_mode *mode = get_irn_mode(left);
4342 ir_tarval *minus1 = get_mode_all_one(mode);
4343 ir_tarval *amask = tarval_shr(minus1, tv1);
4344 ir_tarval *cmask = tarval_shl(minus1, tv1);
4347 if (tarval_and(tv, cmask) != tv) {
4348 /* condition not met */
4349 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4350 c1 = new_r_Const(irg, tv);
4351 DBG_OPT_CSTEVAL(proj, c1);
4354 sl = get_Shl_left(left);
4355 blk = get_nodes_block(n);
4356 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4357 tv = tarval_shr(tv, tv1);
4359 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4364 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4366 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4369 c1 = get_Shr_right(left);
4371 ir_graph *irg = get_irn_irg(c1);
4372 ir_tarval *tv1 = get_Const_tarval(c1);
4373 ir_mode *mode = get_irn_mode(left);
4374 ir_tarval *minus1 = get_mode_all_one(mode);
4375 ir_tarval *amask = tarval_shl(minus1, tv1);
4376 ir_tarval *cmask = tarval_shr(minus1, tv1);
4379 if (tarval_and(tv, cmask) != tv) {
4380 /* condition not met */
4381 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4382 c1 = new_r_Const(irg, tv);
4383 DBG_OPT_CSTEVAL(proj, c1);
4386 sl = get_Shr_left(left);
4387 blk = get_nodes_block(n);
4388 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4389 tv = tarval_shl(tv, tv1);
4391 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4396 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4398 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4401 c1 = get_Shrs_right(left);
4403 ir_graph *irg = get_irn_irg(c1);
4404 ir_tarval *tv1 = get_Const_tarval(c1);
4405 ir_mode *mode = get_irn_mode(left);
4406 ir_tarval *minus1 = get_mode_all_one(mode);
4407 ir_tarval *amask = tarval_shl(minus1, tv1);
4408 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4411 cond = tarval_sub(cond, tv1, NULL);
4412 cond = tarval_shrs(tv, cond);
4414 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4415 /* condition not met */
4416 tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true();
4417 c1 = new_r_Const(irg, tv);
4418 DBG_OPT_CSTEVAL(proj, c1);
4421 sl = get_Shrs_left(left);
4422 blk = get_nodes_block(n);
4423 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4424 tv = tarval_shl(tv, tv1);
4426 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4431 } /* tarval != bad */
4434 if (changed & 2) { /* need a new Const */
4435 ir_graph *irg = get_irn_irg(n);
4436 right = new_r_Const(irg, tv);
4439 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4440 ir_node *op = get_Proj_pred(left);
4442 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4443 ir_node *c = get_binop_right(op);
4446 ir_tarval *tv = get_Const_tarval(c);
4448 if (tarval_is_single_bit(tv)) {
4449 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4450 ir_node *v = get_binop_left(op);
4451 ir_node *blk = get_irn_n(op, -1);
4452 ir_graph *irg = get_irn_irg(op);
4453 ir_mode *mode = get_irn_mode(v);
4455 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4456 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4458 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4465 ir_node *block = get_nodes_block(n);
4467 /* create a new compare */
4468 n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right);
4469 proj = new_rd_Proj(get_irn_dbg_info(proj), n, get_irn_mode(proj), proj_nr);
4473 } /* transform_node_Proj_Cmp */
4476 * Optimize CopyB(mem, x, x) into a Nop.
4478 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4480 ir_node *copyb = get_Proj_pred(proj);
4481 ir_node *a = get_CopyB_dst(copyb);
4482 ir_node *b = get_CopyB_src(copyb);
4485 switch (get_Proj_proj(proj)) {
4486 case pn_CopyB_X_regular:
4487 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4488 DBG_OPT_EXC_REM(proj);
4489 proj = new_r_Jmp(get_nodes_block(copyb));
4491 case pn_CopyB_X_except:
4492 DBG_OPT_EXC_REM(proj);
4493 proj = get_irg_bad(get_irn_irg(proj));
4500 } /* transform_node_Proj_CopyB */
4503 * Optimize Bounds(idx, idx, upper) into idx.
4505 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4507 ir_node *oldn = proj;
4508 ir_node *bound = get_Proj_pred(proj);
4509 ir_node *idx = get_Bound_index(bound);
4510 ir_node *pred = skip_Proj(idx);
4513 if (idx == get_Bound_lower(bound))
4515 else if (is_Bound(pred)) {
4517 * idx was Bounds checked previously, it is still valid if
4518 * lower <= pred_lower && pred_upper <= upper.
4520 ir_node *lower = get_Bound_lower(bound);
4521 ir_node *upper = get_Bound_upper(bound);
4522 if (get_Bound_lower(pred) == lower &&
4523 get_Bound_upper(pred) == upper) {
4525 * One could expect that we simply return the previous
4526 * Bound here. However, this would be wrong, as we could
4527 * add an exception Proj to a new location then.
4528 * So, we must turn in into a tuple.
4534 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4535 switch (get_Proj_proj(proj)) {
4537 DBG_OPT_EXC_REM(proj);
4538 proj = get_Bound_mem(bound);
4540 case pn_Bound_X_except:
4541 DBG_OPT_EXC_REM(proj);
4542 proj = get_irg_bad(get_irn_irg(proj));
4546 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4548 case pn_Bound_X_regular:
4549 DBG_OPT_EXC_REM(proj);
4550 proj = new_r_Jmp(get_nodes_block(bound));
4557 } /* transform_node_Proj_Bound */
4560 * Does all optimizations on nodes that must be done on it's Proj's
4561 * because of creating new nodes.
4563 static ir_node *transform_node_Proj(ir_node *proj)
4565 ir_node *n = get_Proj_pred(proj);
4567 if (n->op->ops.transform_node_Proj)
4568 return n->op->ops.transform_node_Proj(proj);
4570 } /* transform_node_Proj */
4573 * Move Confirms down through Phi nodes.
4575 static ir_node *transform_node_Phi(ir_node *phi)
4578 ir_mode *mode = get_irn_mode(phi);
4580 if (mode_is_reference(mode)) {
4581 n = get_irn_arity(phi);
4583 /* Beware of Phi0 */
4585 ir_node *pred = get_irn_n(phi, 0);
4586 ir_node *bound, *new_phi, *block, **in;
4589 if (! is_Confirm(pred))
4592 bound = get_Confirm_bound(pred);
4593 pnc = get_Confirm_cmp(pred);
4595 NEW_ARR_A(ir_node *, in, n);
4596 in[0] = get_Confirm_value(pred);
4598 for (i = 1; i < n; ++i) {
4599 pred = get_irn_n(phi, i);
4601 if (! is_Confirm(pred) ||
4602 get_Confirm_bound(pred) != bound ||
4603 get_Confirm_cmp(pred) != pnc)
4605 in[i] = get_Confirm_value(pred);
4607 /* move the Confirm nodes "behind" the Phi */
4608 block = get_irn_n(phi, -1);
4609 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4610 return new_r_Confirm(block, new_phi, bound, pnc);
4614 } /* transform_node_Phi */
4617 * Returns the operands of a commutative bin-op, if one operand is
4618 * a const, it is returned as the second one.
4620 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4622 ir_node *op_a = get_binop_left(binop);
4623 ir_node *op_b = get_binop_right(binop);
4625 assert(is_op_commutative(get_irn_op(binop)));
4627 if (is_Const(op_a)) {
4634 } /* get_comm_Binop_Ops */
4637 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4638 * Such pattern may arise in bitfield stores.
4640 * value c4 value c4 & c2
4641 * AND c3 AND c1 | c3
4648 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4651 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4653 ir_node *irn_and, *c1;
4655 ir_node *and_l, *c3;
4656 ir_node *value, *c4;
4657 ir_node *new_and, *new_const, *block;
4658 ir_mode *mode = get_irn_mode(irn_or);
4660 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4664 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4665 if (!is_Const(c1) || !is_And(irn_and))
4668 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4672 tv1 = get_Const_tarval(c1);
4673 tv2 = get_Const_tarval(c2);
4675 tv = tarval_or(tv1, tv2);
4676 if (tarval_is_all_one(tv)) {
4677 /* the AND does NOT clear a bit with isn't set by the OR */
4678 set_Or_left(irn_or, or_l);
4679 set_Or_right(irn_or, c1);
4681 /* check for more */
4688 get_comm_Binop_Ops(or_l, &and_l, &c3);
4689 if (!is_Const(c3) || !is_And(and_l))
4692 get_comm_Binop_Ops(and_l, &value, &c4);
4696 /* ok, found the pattern, check for conditions */
4697 assert(mode == get_irn_mode(irn_and));
4698 assert(mode == get_irn_mode(or_l));
4699 assert(mode == get_irn_mode(and_l));
4701 tv3 = get_Const_tarval(c3);
4702 tv4 = get_Const_tarval(c4);
4704 tv = tarval_or(tv4, tv2);
4705 if (!tarval_is_all_one(tv)) {
4706 /* have at least one 0 at the same bit position */
4710 if (tv3 != tarval_andnot(tv3, tv4)) {
4711 /* bit in the or_mask is outside the and_mask */
4715 if (tv1 != tarval_andnot(tv1, tv2)) {
4716 /* bit in the or_mask is outside the and_mask */
4720 /* ok, all conditions met */
4721 block = get_irn_n(irn_or, -1);
4722 irg = get_irn_irg(block);
4724 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4726 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4728 set_Or_left(irn_or, new_and);
4729 set_Or_right(irn_or, new_const);
4731 /* check for more */
4733 } /* transform_node_Or_bf_store */
4736 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4738 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4740 ir_mode *mode = get_irn_mode(irn_or);
4741 ir_node *shl, *shr, *block;
4742 ir_node *irn, *x, *c1, *c2, *n;
4743 ir_tarval *tv1, *tv2;
4745 /* some backends can't handle rotl */
4746 if (!be_get_backend_param()->support_rotl)
4749 if (! mode_is_int(mode))
4752 shl = get_binop_left(irn_or);
4753 shr = get_binop_right(irn_or);
4762 } else if (!is_Shl(shl)) {
4764 } else if (!is_Shr(shr)) {
4767 x = get_Shl_left(shl);
4768 if (x != get_Shr_left(shr))
4771 c1 = get_Shl_right(shl);
4772 c2 = get_Shr_right(shr);
4773 if (is_Const(c1) && is_Const(c2)) {
4774 tv1 = get_Const_tarval(c1);
4775 if (! tarval_is_long(tv1))
4778 tv2 = get_Const_tarval(c2);
4779 if (! tarval_is_long(tv2))
4782 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4783 != (int) get_mode_size_bits(mode))
4786 /* yet, condition met */
4787 block = get_nodes_block(irn_or);
4789 n = new_r_Rotl(block, x, c1, mode);
4791 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4795 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4796 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4797 if (!is_negated_value(c1, c2)) {
4801 /* yet, condition met */
4802 block = get_nodes_block(irn_or);
4803 n = new_r_Rotl(block, x, c1, mode);
4804 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4806 } /* transform_node_Or_Rotl */
4811 static ir_node *transform_node_Or(ir_node *n)
4813 ir_node *c, *oldn = n;
4814 ir_node *a = get_Or_left(n);
4815 ir_node *b = get_Or_right(n);
4818 if (is_Not(a) && is_Not(b)) {
4819 /* ~a | ~b = ~(a&b) */
4820 ir_node *block = get_nodes_block(n);
4822 mode = get_irn_mode(n);
4825 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4826 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4827 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4831 /* we can evaluate 2 Projs of the same Cmp */
4832 if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) {
4833 ir_node *pred_a = get_Proj_pred(a);
4834 ir_node *pred_b = get_Proj_pred(b);
4835 if (pred_a == pred_b) {
4836 dbg_info *dbgi = get_irn_dbg_info(n);
4837 pn_Cmp pn_a = get_Proj_pn_cmp(a);
4838 pn_Cmp pn_b = get_Proj_pn_cmp(b);
4839 /* yes, we can simply calculate with pncs */
4840 pn_Cmp new_pnc = pn_a | pn_b;
4842 return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc);
4846 mode = get_irn_mode(n);
4847 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4849 n = transform_node_Or_bf_store(n);
4850 n = transform_node_Or_Rotl(n);
4854 n = transform_bitwise_distributive(n, transform_node_Or);
4857 } /* transform_node_Or */
4861 static ir_node *transform_node(ir_node *n);
4864 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4866 * Should be moved to reassociation?
4868 static ir_node *transform_node_shift(ir_node *n)
4870 ir_node *left, *right;
4872 ir_tarval *tv1, *tv2, *res;
4873 ir_node *in[2], *irn, *block;
4876 left = get_binop_left(n);
4878 /* different operations */
4879 if (get_irn_op(left) != get_irn_op(n))
4882 right = get_binop_right(n);
4883 tv1 = value_of(right);
4884 if (tv1 == tarval_bad)
4887 tv2 = value_of(get_binop_right(left));
4888 if (tv2 == tarval_bad)
4891 res = tarval_add(tv1, tv2);
4892 mode = get_irn_mode(n);
4893 irg = get_irn_irg(n);
4895 /* beware: a simple replacement works only, if res < modulo shift */
4897 int modulo_shf = get_mode_modulo_shift(mode);
4898 if (modulo_shf > 0) {
4899 ir_tarval *modulo = new_tarval_from_long(modulo_shf,
4900 get_tarval_mode(res));
4902 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4904 /* shifting too much */
4905 if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) {
4907 ir_node *block = get_nodes_block(n);
4908 dbg_info *dbgi = get_irn_dbg_info(n);
4909 ir_mode *smode = get_irn_mode(right);
4910 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
4911 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
4914 return new_r_Const(irg, get_mode_null(mode));
4918 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
4921 /* ok, we can replace it */
4922 block = get_nodes_block(n);
4924 in[0] = get_binop_left(left);
4925 in[1] = new_r_Const(irg, res);
4927 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
4929 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4931 return transform_node(irn);
4932 } /* transform_node_shift */
4935 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
4937 * - and, or, xor instead of &
4938 * - Shl, Shr, Shrs, rotl instead of >>
4939 * (with a special case for Or/Xor + Shrs)
4941 static ir_node *transform_node_bitop_shift(ir_node *n)
4944 ir_node *right = get_binop_right(n);
4945 ir_mode *mode = get_irn_mode(n);
4946 ir_node *bitop_left;
4947 ir_node *bitop_right;
4957 ir_tarval *tv_shift;
4959 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
4961 if (!is_Const(right))
4964 left = get_binop_left(n);
4965 op_left = get_irn_op(left);
4966 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
4969 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
4970 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
4971 /* TODO: test if sign bit is affectes */
4975 bitop_right = get_binop_right(left);
4976 if (!is_Const(bitop_right))
4979 bitop_left = get_binop_left(left);
4981 block = get_nodes_block(n);
4982 dbgi = get_irn_dbg_info(n);
4983 tv1 = get_Const_tarval(bitop_right);
4984 tv2 = get_Const_tarval(right);
4986 assert(get_tarval_mode(tv1) == mode);
4989 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
4990 tv_shift = tarval_shl(tv1, tv2);
4991 } else if (is_Shr(n)) {
4992 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
4993 tv_shift = tarval_shr(tv1, tv2);
4994 } else if (is_Shrs(n)) {
4995 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
4996 tv_shift = tarval_shrs(tv1, tv2);
4999 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
5000 tv_shift = tarval_rotl(tv1, tv2);
5003 assert(get_tarval_mode(tv_shift) == mode);
5004 irg = get_irn_irg(n);
5005 new_const = new_r_Const(irg, tv_shift);
5007 if (op_left == op_And) {
5008 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
5009 } else if (op_left == op_Or) {
5010 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
5012 assert(op_left == op_Eor);
5013 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
5021 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
5023 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
5024 * (also with x >>s c1 when c1>=c2)
5026 static ir_node *transform_node_shl_shr(ir_node *n)
5029 ir_node *right = get_binop_right(n);
5039 ir_tarval *tv_shift;
5045 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5047 if (!is_Const(right))
5050 left = get_binop_left(n);
5051 mode = get_irn_mode(n);
5052 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5053 ir_node *shr_right = get_binop_right(left);
5055 if (!is_Const(shr_right))
5058 x = get_binop_left(left);
5059 tv_shr = get_Const_tarval(shr_right);
5060 tv_shl = get_Const_tarval(right);
5062 if (is_Shrs(left)) {
5063 /* shrs variant only allowed if c1 >= c2 */
5064 if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge))
5067 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5070 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5072 tv_mask = tarval_shl(tv_mask, tv_shl);
5073 } else if (is_Shr(n) && is_Shl(left)) {
5074 ir_node *shl_right = get_Shl_right(left);
5076 if (!is_Const(shl_right))
5079 x = get_Shl_left(left);
5080 tv_shr = get_Const_tarval(right);
5081 tv_shl = get_Const_tarval(shl_right);
5083 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5084 tv_mask = tarval_shr(tv_mask, tv_shr);
5089 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5090 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5093 assert(tv_mask != tarval_bad);
5094 assert(get_tarval_mode(tv_mask) == mode);
5096 block = get_nodes_block(n);
5097 irg = get_irn_irg(block);
5098 dbgi = get_irn_dbg_info(n);
5100 pnc = tarval_cmp(tv_shl, tv_shr);
5101 if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) {
5102 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5103 new_const = new_r_Const(irg, tv_shift);
5105 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5107 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5110 assert(pnc == pn_Cmp_Gt);
5111 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5112 new_const = new_r_Const(irg, tv_shift);
5113 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5116 new_const = new_r_Const(irg, tv_mask);
5117 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5122 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5124 ir_mode *mode = get_tarval_mode(tv);
5125 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5126 return tarval_mod(tv, modulo_tv);
5129 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5130 ir_node *left, ir_node *right, ir_mode *mode);
5133 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5134 * then we can use that to minimize the value of Add(x, const) or
5135 * Sub(Const, x). In particular this often avoids 1 instruction in some
5136 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5137 * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
5139 static ir_node *transform_node_shift_modulo(ir_node *n,
5140 new_shift_func new_shift)
5142 ir_mode *mode = get_irn_mode(n);
5143 int modulo = get_mode_modulo_shift(mode);
5144 ir_node *newop = NULL;
5145 ir_mode *mode_right;
5152 if (get_mode_arithmetic(mode) != irma_twos_complement)
5154 if (!is_po2(modulo))
5157 irg = get_irn_irg(n);
5158 block = get_nodes_block(n);
5159 right = get_binop_right(n);
5160 mode_right = get_irn_mode(right);
5161 if (is_Const(right)) {
5162 ir_tarval *tv = get_Const_tarval(right);
5163 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5168 newop = new_r_Const(irg, tv_mod);
5169 } else if (is_Add(right)) {
5170 ir_node *add_right = get_Add_right(right);
5171 if (is_Const(add_right)) {
5172 ir_tarval *tv = get_Const_tarval(add_right);
5173 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5178 newconst = new_r_Const(irg, tv_mod);
5179 newop = new_r_Add(block, get_Add_left(right), newconst,
5182 } else if (is_Sub(right)) {
5183 ir_node *sub_left = get_Sub_left(right);
5184 if (is_Const(sub_left)) {
5185 ir_tarval *tv = get_Const_tarval(sub_left);
5186 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5191 newconst = new_r_Const(irg, tv_mod);
5192 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5199 if (newop != NULL) {
5200 dbg_info *dbgi = get_irn_dbg_info(n);
5201 ir_node *left = get_binop_left(n);
5202 return new_shift(dbgi, block, left, newop, mode);
5210 static ir_node *transform_node_Shr(ir_node *n)
5212 ir_node *c, *oldn = n;
5213 ir_node *left = get_Shr_left(n);
5214 ir_node *right = get_Shr_right(n);
5215 ir_mode *mode = get_irn_mode(n);
5217 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5218 n = transform_node_shift(n);
5221 n = transform_node_shift_modulo(n, new_rd_Shr);
5223 n = transform_node_shl_shr(n);
5225 n = transform_node_bitop_shift(n);
5228 } /* transform_node_Shr */
5233 static ir_node *transform_node_Shrs(ir_node *n)
5235 ir_node *c, *oldn = n;
5236 ir_node *a = get_Shrs_left(n);
5237 ir_node *b = get_Shrs_right(n);
5238 ir_mode *mode = get_irn_mode(n);
5240 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5241 n = transform_node_shift(n);
5244 n = transform_node_shift_modulo(n, new_rd_Shrs);
5246 n = transform_node_bitop_shift(n);
5249 } /* transform_node_Shrs */
5254 static ir_node *transform_node_Shl(ir_node *n)
5256 ir_node *c, *oldn = n;
5257 ir_node *a = get_Shl_left(n);
5258 ir_node *b = get_Shl_right(n);
5259 ir_mode *mode = get_irn_mode(n);
5261 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5262 n = transform_node_shift(n);
5265 n = transform_node_shift_modulo(n, new_rd_Shl);
5267 n = transform_node_shl_shr(n);
5269 n = transform_node_bitop_shift(n);
5272 } /* transform_node_Shl */
5277 static ir_node *transform_node_Rotl(ir_node *n)
5279 ir_node *c, *oldn = n;
5280 ir_node *a = get_Rotl_left(n);
5281 ir_node *b = get_Rotl_right(n);
5282 ir_mode *mode = get_irn_mode(n);
5284 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5285 n = transform_node_shift(n);
5288 n = transform_node_bitop_shift(n);
5291 } /* transform_node_Rotl */
5296 static ir_node *transform_node_Conv(ir_node *n)
5298 ir_node *c, *oldn = n;
5299 ir_mode *mode = get_irn_mode(n);
5300 ir_node *a = get_Conv_op(n);
5302 if (mode != mode_b && is_const_Phi(a)) {
5303 /* Do NOT optimize mode_b Conv's, this leads to remaining
5304 * Phib nodes later, because the conv_b_lower operation
5305 * is instantly reverted, when it tries to insert a Convb.
5307 c = apply_conv_on_phi(a, mode);
5309 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5314 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5315 ir_graph *irg = get_irn_irg(n);
5316 return new_r_Unknown(irg, mode);
5319 if (mode_is_reference(mode) &&
5320 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5322 ir_node *l = get_Add_left(a);
5323 ir_node *r = get_Add_right(a);
5324 dbg_info *dbgi = get_irn_dbg_info(a);
5325 ir_node *block = get_nodes_block(n);
5327 ir_node *lop = get_Conv_op(l);
5328 if (get_irn_mode(lop) == mode) {
5329 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5330 n = new_rd_Add(dbgi, block, lop, r, mode);
5335 ir_node *rop = get_Conv_op(r);
5336 if (get_irn_mode(rop) == mode) {
5337 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5338 n = new_rd_Add(dbgi, block, l, rop, mode);
5345 } /* transform_node_Conv */
5348 * Remove dead blocks and nodes in dead blocks
5349 * in keep alive list. We do not generate a new End node.
5351 static ir_node *transform_node_End(ir_node *n)
5353 int i, j, n_keepalives = get_End_n_keepalives(n);
5356 NEW_ARR_A(ir_node *, in, n_keepalives);
5358 for (i = j = 0; i < n_keepalives; ++i) {
5359 ir_node *ka = get_End_keepalive(n, i);
5361 if (! is_Block_dead(ka)) {
5365 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
5367 } else if (is_Bad(ka)) {
5368 /* no need to keep Bad */
5373 if (j != n_keepalives)
5374 set_End_keepalives(n, j, in);
5376 } /* transform_node_End */
5378 bool is_negated_value(ir_node *a, ir_node *b)
5380 if (is_Minus(a) && get_Minus_op(a) == b)
5382 if (is_Minus(b) && get_Minus_op(b) == a)
5384 if (is_Sub(a) && is_Sub(b)) {
5385 ir_node *a_left = get_Sub_left(a);
5386 ir_node *a_right = get_Sub_right(a);
5387 ir_node *b_left = get_Sub_left(b);
5388 ir_node *b_right = get_Sub_right(b);
5390 if (a_left == b_right && a_right == b_left)
5398 * Optimize a Mux into some simpler cases.
5400 static ir_node *transform_node_Mux(ir_node *n)
5402 ir_node *oldn = n, *sel = get_Mux_sel(n);
5403 ir_mode *mode = get_irn_mode(n);
5404 ir_node *t = get_Mux_true(n);
5405 ir_node *f = get_Mux_false(n);
5406 ir_graph *irg = get_irn_irg(n);
5408 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5412 ir_node* block = get_nodes_block(n);
5414 ir_node* c1 = get_Mux_sel(t);
5415 ir_node* t1 = get_Mux_true(t);
5416 ir_node* f1 = get_Mux_false(t);
5418 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5419 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5420 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5425 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5426 } else if (f == t1) {
5427 /* Mux(cond0, Mux(cond1, x, y), x) */
5428 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5429 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5430 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5435 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5437 } else if (is_Mux(f)) {
5438 ir_node* block = get_nodes_block(n);
5440 ir_node* c1 = get_Mux_sel(f);
5441 ir_node* t1 = get_Mux_true(f);
5442 ir_node* f1 = get_Mux_false(f);
5444 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5445 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5446 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5451 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5452 } else if (t == f1) {
5453 /* Mux(cond0, x, Mux(cond1, y, x)) */
5454 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5455 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5456 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5461 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5465 /* first normalization step: try to move a constant to the false side,
5466 * 0 preferred on false side too */
5468 ir_node *cmp = get_Proj_pred(sel);
5470 if (is_Cmp(cmp) && is_Const(t) &&
5471 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5472 pn_Cmp pnc = get_Proj_pn_cmp(sel);
5477 /* Mux(x, a, b) => Mux(not(x), b, a) */
5478 sel = new_r_Proj(cmp, mode_b,
5479 get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp))));
5480 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5484 /* note: after normalization, false can only happen on default */
5485 if (mode == mode_b) {
5486 dbg_info *dbg = get_irn_dbg_info(n);
5487 ir_node *block = get_nodes_block(n);
5490 ir_tarval *tv_t = get_Const_tarval(t);
5491 if (tv_t == tarval_b_true) {
5493 /* Muxb(sel, true, false) = sel */
5494 assert(get_Const_tarval(f) == tarval_b_false);
5495 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5498 /* Muxb(sel, true, x) = Or(sel, x) */
5499 n = new_rd_Or(dbg, block, sel, f, mode_b);
5500 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5504 } else if (is_Const(f)) {
5505 ir_tarval *tv_f = get_Const_tarval(f);
5506 if (tv_f == tarval_b_true) {
5507 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5508 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5509 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5510 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5513 /* Muxb(sel, x, false) = And(sel, x) */
5514 assert(tv_f == tarval_b_false);
5515 n = new_rd_And(dbg, block, sel, t, mode_b);
5516 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5522 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5523 * value to integer. */
5524 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5525 ir_tarval *a = get_Const_tarval(t);
5526 ir_tarval *b = get_Const_tarval(f);
5528 if (tarval_is_one(a) && tarval_is_null(b)) {
5529 ir_node *block = get_nodes_block(n);
5530 ir_node *conv = new_r_Conv(block, sel, mode);
5532 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5534 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5535 ir_node *block = get_nodes_block(n);
5536 ir_node *not_ = new_r_Not(block, sel, mode_b);
5537 ir_node *conv = new_r_Conv(block, not_, mode);
5539 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5545 ir_node *cmp = get_Proj_pred(sel);
5546 long pn = get_Proj_proj(sel);
5549 * Note: normalization puts the constant on the right side,
5550 * so we check only one case.
5553 ir_node *cmp_r = get_Cmp_right(cmp);
5554 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5555 ir_node *block = get_nodes_block(n);
5556 ir_node *cmp_l = get_Cmp_left(cmp);
5558 if (mode_is_int(mode)) {
5560 if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) {
5561 /* Mux((a & b) != 0, c, 0) */
5562 ir_node *and_r = get_And_right(cmp_l);
5565 if (and_r == t && f == cmp_r) {
5566 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5567 if (pn == pn_Cmp_Lg) {
5568 /* Mux((a & 2^C) != 0, 2^C, 0) */
5570 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5572 /* Mux((a & 2^C) == 0, 2^C, 0) */
5573 n = new_rd_Eor(get_irn_dbg_info(n),
5574 block, cmp_l, t, mode);
5575 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5580 if (is_Shl(and_r)) {
5581 ir_node *shl_l = get_Shl_left(and_r);
5582 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5583 if (and_r == t && f == cmp_r) {
5584 if (pn == pn_Cmp_Lg) {
5585 /* (a & (1 << n)) != 0, (1 << n), 0) */
5587 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5589 /* (a & (1 << n)) == 0, (1 << n), 0) */
5590 n = new_rd_Eor(get_irn_dbg_info(n),
5591 block, cmp_l, t, mode);
5592 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5598 and_l = get_And_left(cmp_l);
5599 if (is_Shl(and_l)) {
5600 ir_node *shl_l = get_Shl_left(and_l);
5601 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5602 if (and_l == t && f == cmp_r) {
5603 if (pn == pn_Cmp_Lg) {
5604 /* ((1 << n) & a) != 0, (1 << n), 0) */
5606 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5608 /* ((1 << n) & a) == 0, (1 << n), 0) */
5609 n = new_rd_Eor(get_irn_dbg_info(n),
5610 block, cmp_l, t, mode);
5611 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP);
5624 } /* transform_node_Mux */
5627 * optimize Sync nodes that have other syncs as input we simply add the inputs
5628 * of the other sync to our own inputs
5630 static ir_node *transform_node_Sync(ir_node *n)
5632 int arity = get_Sync_n_preds(n);
5635 for (i = 0; i < arity;) {
5636 ir_node *pred = get_Sync_pred(n, i);
5640 if (!is_Sync(pred)) {
5648 pred_arity = get_Sync_n_preds(pred);
5649 for (j = 0; j < pred_arity; ++j) {
5650 ir_node *pred_pred = get_Sync_pred(pred, j);
5655 add_irn_n(n, pred_pred);
5659 if (get_Sync_pred(n, k) == pred_pred) break;
5664 /* rehash the sync node */
5668 } /* transform_node_Sync */
5671 * optimize a trampoline Call into a direct Call
5673 static ir_node *transform_node_Call(ir_node *call)
5675 ir_node *callee = get_Call_ptr(call);
5676 ir_node *adr, *mem, *res, *bl, **in;
5677 ir_type *ctp, *mtp, *tp;
5681 int i, n_res, n_param;
5684 if (! is_Proj(callee))
5686 callee = get_Proj_pred(callee);
5687 if (! is_Builtin(callee))
5689 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5692 mem = get_Call_mem(call);
5694 if (skip_Proj(mem) == callee) {
5695 /* memory is routed to the trampoline, skip */
5696 mem = get_Builtin_mem(callee);
5699 /* build a new call type */
5700 mtp = get_Call_type(call);
5701 tdb = get_type_dbg_info(mtp);
5703 n_res = get_method_n_ress(mtp);
5704 n_param = get_method_n_params(mtp);
5705 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5707 for (i = 0; i < n_res; ++i)
5708 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5710 NEW_ARR_A(ir_node *, in, n_param + 1);
5712 /* FIXME: we don't need a new pointer type in every step */
5713 irg = get_irn_irg(call);
5714 tp = get_irg_frame_type(irg);
5715 tp = new_type_pointer(tp);
5716 set_method_param_type(ctp, 0, tp);
5718 in[0] = get_Builtin_param(callee, 2);
5719 for (i = 0; i < n_param; ++i) {
5720 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5721 in[i + 1] = get_Call_param(call, i);
5723 var = get_method_variadicity(mtp);
5724 set_method_variadicity(ctp, var);
5725 if (var == variadicity_variadic) {
5726 set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1);
5728 /* When we resolve a trampoline, the function must be called by a this-call */
5729 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5730 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5732 adr = get_Builtin_param(callee, 1);
5734 db = get_irn_dbg_info(call);
5735 bl = get_nodes_block(call);
5737 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5738 if (get_irn_pinned(call) == op_pin_state_floats)
5739 set_irn_pinned(res, op_pin_state_floats);
5741 } /* transform_node_Call */
5744 * Tries several [inplace] [optimizing] transformations and returns an
5745 * equivalent node. The difference to equivalent_node() is that these
5746 * transformations _do_ generate new nodes, and thus the old node must
5747 * not be freed even if the equivalent node isn't the old one.
5749 static ir_node *transform_node(ir_node *n)
5754 * Transform_node is the only "optimizing transformation" that might
5755 * return a node with a different opcode. We iterate HERE until fixpoint
5756 * to get the final result.
5760 if (n->op->ops.transform_node != NULL)
5761 n = n->op->ops.transform_node(n);
5762 } while (oldn != n);
5765 } /* transform_node */
5768 * Sets the default transform node operation for an ir_op_ops.
5770 * @param code the opcode for the default operation
5771 * @param ops the operations initialized
5776 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5780 ops->transform_node = transform_node_##a; \
5782 #define CASE_PROJ(a) \
5784 ops->transform_node_Proj = transform_node_Proj_##a; \
5786 #define CASE_PROJ_EX(a) \
5788 ops->transform_node = transform_node_##a; \
5789 ops->transform_node_Proj = transform_node_Proj_##a; \
5829 } /* firm_set_default_transform_node */
5832 /* **************** Common Subexpression Elimination **************** */
5834 /** The size of the hash table used, should estimate the number of nodes
5836 #define N_IR_NODES 512
5838 /** Compares the attributes of two Const nodes. */
5839 static int node_cmp_attr_Const(ir_node *a, ir_node *b)
5841 return get_Const_tarval(a) != get_Const_tarval(b);
5844 /** Compares the attributes of two Proj nodes. */
5845 static int node_cmp_attr_Proj(ir_node *a, ir_node *b)
5847 return a->attr.proj.proj != b->attr.proj.proj;
5848 } /* node_cmp_attr_Proj */
5850 /** Compares the attributes of two Alloc nodes. */
5851 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b)
5853 const alloc_attr *pa = &a->attr.alloc;
5854 const alloc_attr *pb = &b->attr.alloc;
5855 return (pa->where != pb->where) || (pa->type != pb->type);
5856 } /* node_cmp_attr_Alloc */
5858 /** Compares the attributes of two Free nodes. */
5859 static int node_cmp_attr_Free(ir_node *a, ir_node *b)
5861 const free_attr *pa = &a->attr.free;
5862 const free_attr *pb = &b->attr.free;
5863 return (pa->where != pb->where) || (pa->type != pb->type);
5864 } /* node_cmp_attr_Free */
5866 /** Compares the attributes of two SymConst nodes. */
5867 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b)
5869 const symconst_attr *pa = &a->attr.symc;
5870 const symconst_attr *pb = &b->attr.symc;
5871 return (pa->kind != pb->kind)
5872 || (pa->sym.type_p != pb->sym.type_p);
5875 /** Compares the attributes of two Call nodes. */
5876 static int node_cmp_attr_Call(ir_node *a, ir_node *b)
5878 const call_attr *pa = &a->attr.call;
5879 const call_attr *pb = &b->attr.call;
5880 return (pa->type != pb->type)
5881 || (pa->tail_call != pb->tail_call);
5882 } /* node_cmp_attr_Call */
5884 /** Compares the attributes of two Sel nodes. */
5885 static int node_cmp_attr_Sel(ir_node *a, ir_node *b)
5887 const ir_entity *a_ent = get_Sel_entity(a);
5888 const ir_entity *b_ent = get_Sel_entity(b);
5889 return a_ent != b_ent;
5890 } /* node_cmp_attr_Sel */
5892 /** Compares the attributes of two Phi nodes. */
5893 static int node_cmp_attr_Phi(ir_node *a, ir_node *b)
5895 /* we can only enter this function if both nodes have the same number of inputs,
5896 hence it is enough to check if one of them is a Phi0 */
5898 /* check the Phi0 pos attribute */
5899 return a->attr.phi.u.pos != b->attr.phi.u.pos;
5902 } /* node_cmp_attr_Phi */
5904 /** Compares the attributes of two Conv nodes. */
5905 static int node_cmp_attr_Conv(ir_node *a, ir_node *b)
5907 return get_Conv_strict(a) != get_Conv_strict(b);
5908 } /* node_cmp_attr_Conv */
5910 /** Compares the attributes of two Cast nodes. */
5911 static int node_cmp_attr_Cast(ir_node *a, ir_node *b)
5913 return get_Cast_type(a) != get_Cast_type(b);
5914 } /* node_cmp_attr_Cast */
5916 /** Compares the attributes of two Load nodes. */
5917 static int node_cmp_attr_Load(ir_node *a, ir_node *b)
5919 if (get_Load_volatility(a) == volatility_is_volatile ||
5920 get_Load_volatility(b) == volatility_is_volatile)
5921 /* NEVER do CSE on volatile Loads */
5923 /* do not CSE Loads with different alignment. Be conservative. */
5924 if (get_Load_align(a) != get_Load_align(b))
5927 return get_Load_mode(a) != get_Load_mode(b);
5928 } /* node_cmp_attr_Load */
5930 /** Compares the attributes of two Store nodes. */
5931 static int node_cmp_attr_Store(ir_node *a, ir_node *b)
5933 /* do not CSE Stores with different alignment. Be conservative. */
5934 if (get_Store_align(a) != get_Store_align(b))
5937 /* NEVER do CSE on volatile Stores */
5938 return (get_Store_volatility(a) == volatility_is_volatile ||
5939 get_Store_volatility(b) == volatility_is_volatile);
5940 } /* node_cmp_attr_Store */
5942 /** Compares two exception attributes */
5943 static int node_cmp_exception(ir_node *a, ir_node *b)
5945 const except_attr *ea = &a->attr.except;
5946 const except_attr *eb = &b->attr.except;
5948 return ea->pin_state != eb->pin_state;
5951 #define node_cmp_attr_Bound node_cmp_exception
5953 /** Compares the attributes of two Div nodes. */
5954 static int node_cmp_attr_Div(ir_node *a, ir_node *b)
5956 const divmod_attr *ma = &a->attr.divmod;
5957 const divmod_attr *mb = &b->attr.divmod;
5958 return ma->exc.pin_state != mb->exc.pin_state ||
5959 ma->resmode != mb->resmode ||
5960 ma->no_remainder != mb->no_remainder;
5961 } /* node_cmp_attr_Div */
5963 /** Compares the attributes of two Div or Mod nodes. */
5964 static int node_cmp_attr_Div_Mod(ir_node *a, ir_node *b)
5966 const divmod_attr *ma = &a->attr.divmod;
5967 const divmod_attr *mb = &b->attr.divmod;
5968 return ma->exc.pin_state != mb->exc.pin_state ||
5969 ma->resmode != mb->resmode;
5970 } /* node_cmp_attr_Div_Mod */
5972 /** Compares the attributes of two Mod nodes. */
5973 static int node_cmp_attr_Mod(ir_node *a, ir_node *b)
5975 return node_cmp_attr_Div_Mod(a, b);
5976 } /* node_cmp_attr_Mod */
5978 /** Compares the attributes of two Confirm nodes. */
5979 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b)
5981 /* no need to compare the bound, as this is a input */
5982 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
5983 } /* node_cmp_attr_Confirm */
5985 /** Compares the attributes of two Builtin nodes. */
5986 static int node_cmp_attr_Builtin(ir_node *a, ir_node *b)
5988 /* no need to compare the type, equal kind means equal type */
5989 return get_Builtin_kind(a) != get_Builtin_kind(b);
5990 } /* node_cmp_attr_Builtin */
5992 /** Compares the attributes of two ASM nodes. */
5993 static int node_cmp_attr_ASM(ir_node *a, ir_node *b)
5996 const ir_asm_constraint *ca;
5997 const ir_asm_constraint *cb;
6000 if (get_ASM_text(a) != get_ASM_text(b))
6003 /* Should we really check the constraints here? Should be better, but is strange. */
6004 n = get_ASM_n_input_constraints(a);
6005 if (n != get_ASM_n_input_constraints(b))
6008 ca = get_ASM_input_constraints(a);
6009 cb = get_ASM_input_constraints(b);
6010 for (i = 0; i < n; ++i) {
6011 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6015 n = get_ASM_n_output_constraints(a);
6016 if (n != get_ASM_n_output_constraints(b))
6019 ca = get_ASM_output_constraints(a);
6020 cb = get_ASM_output_constraints(b);
6021 for (i = 0; i < n; ++i) {
6022 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
6026 n = get_ASM_n_clobbers(a);
6027 if (n != get_ASM_n_clobbers(b))
6030 cla = get_ASM_clobbers(a);
6031 clb = get_ASM_clobbers(b);
6032 for (i = 0; i < n; ++i) {
6033 if (cla[i] != clb[i])
6037 } /* node_cmp_attr_ASM */
6039 /** Compares the inexistent attributes of two Dummy nodes. */
6040 static int node_cmp_attr_Dummy(ir_node *a, ir_node *b)
6048 * Set the default node attribute compare operation for an ir_op_ops.
6050 * @param code the opcode for the default operation
6051 * @param ops the operations initialized
6056 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6060 ops->node_cmp_attr = node_cmp_attr_##a; \
6091 } /* firm_set_default_node_cmp_attr */
6094 * Compare function for two nodes in the value table. Gets two
6095 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6097 int identities_cmp(const void *elt, const void *key)
6099 ir_node *a = (ir_node *)elt;
6100 ir_node *b = (ir_node *)key;
6103 if (a == b) return 0;
6105 if ((get_irn_op(a) != get_irn_op(b)) ||
6106 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6108 /* compare if a's in and b's in are of equal length */
6109 irn_arity_a = get_irn_arity(a);
6110 if (irn_arity_a != get_irn_arity(b))
6113 /* blocks are never the same */
6117 if (get_irn_pinned(a) == op_pin_state_pinned) {
6118 /* for pinned nodes, the block inputs must be equal */
6119 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6121 } else if (! get_opt_global_cse()) {
6122 /* for block-local CSE both nodes must be in the same Block */
6123 if (get_nodes_block(a) != get_nodes_block(b))
6127 /* compare a->in[0..ins] with b->in[0..ins] */
6128 for (i = 0; i < irn_arity_a; ++i) {
6129 ir_node *pred_a = get_irn_n(a, i);
6130 ir_node *pred_b = get_irn_n(b, i);
6131 if (pred_a != pred_b) {
6132 /* if both predecessors are CSE neutral they might be different */
6133 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6139 * here, we already now that the nodes are identical except their
6142 if (a->op->ops.node_cmp_attr)
6143 return a->op->ops.node_cmp_attr(a, b);
6146 } /* identities_cmp */
6149 * Calculate a hash value of a node.
6151 * @param node The IR-node
6153 unsigned ir_node_hash(const ir_node *node)
6155 return node->op->ops.hash(node);
6156 } /* ir_node_hash */
6159 void new_identities(ir_graph *irg)
6161 if (irg->value_table != NULL)
6162 del_pset(irg->value_table);
6163 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6164 } /* new_identities */
6166 void del_identities(ir_graph *irg)
6168 if (irg->value_table != NULL)
6169 del_pset(irg->value_table);
6170 } /* del_identities */
6172 /* Normalize a node by putting constants (and operands with larger
6173 * node index) on the right (operator side). */
6174 void ir_normalize_node(ir_node *n)
6176 if (is_op_commutative(get_irn_op(n))) {
6177 ir_node *l = get_binop_left(n);
6178 ir_node *r = get_binop_right(n);
6180 /* For commutative operators perform a OP b == b OP a but keep
6181 * constants on the RIGHT side. This helps greatly in some
6182 * optimizations. Moreover we use the idx number to make the form
6184 if (!operands_are_normalized(l, r)) {
6185 set_binop_left(n, r);
6186 set_binop_right(n, l);
6190 } /* ir_normalize_node */
6193 * Return the canonical node computing the same value as n.
6194 * Looks up the node in a hash table, enters it in the table
6195 * if it isn't there yet.
6197 * @param n the node to look up
6199 * @return a node that computes the same value as n or n if no such
6200 * node could be found
6202 ir_node *identify_remember(ir_node *n)
6204 ir_graph *irg = get_irn_irg(n);
6205 pset *value_table = irg->value_table;
6208 if (value_table == NULL)
6211 ir_normalize_node(n);
6212 /* lookup or insert in hash table with given hash key. */
6213 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6216 /* n is reachable again */
6217 edges_node_revival(nn, get_irn_irg(nn));
6221 } /* identify_remember */
6224 * During construction we set the op_pin_state_pinned flag in the graph right
6225 * when the optimization is performed. The flag turning on procedure global
6226 * cse could be changed between two allocations. This way we are safe.
6228 * @param n The node to lookup
6230 static inline ir_node *identify_cons(ir_node *n)
6234 n = identify_remember(n);
6235 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6236 ir_graph *irg = get_irn_irg(n);
6237 set_irg_pinned(irg, op_pin_state_floats);
6240 } /* identify_cons */
6242 /* Add a node to the identities value table. */
6243 void add_identities(ir_node *node)
6250 identify_remember(node);
6253 /* Visit each node in the value table of a graph. */
6254 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6257 ir_graph *rem = current_ir_graph;
6259 current_ir_graph = irg;
6260 foreach_pset(irg->value_table, ir_node*, node) {
6263 current_ir_graph = rem;
6264 } /* visit_all_identities */
6267 * Garbage in, garbage out. If a node has a dead input, i.e., the
6268 * Bad node is input to the node, return the Bad node.
6270 static ir_node *gigo(ir_node *node)
6273 ir_op *op = get_irn_op(node);
6275 /* remove garbage blocks by looking at control flow that leaves the block
6276 and replacing the control flow by Bad. */
6277 if (get_irn_mode(node) == mode_X) {
6278 ir_node *block = get_nodes_block(skip_Proj(node));
6279 ir_graph *irg = get_irn_irg(block);
6281 /* Don't optimize nodes in immature blocks. */
6282 if (!get_Block_matured(block))
6284 /* Don't optimize End, may have Bads. */
6285 if (op == op_End) return node;
6287 if (is_Block(block)) {
6288 if (is_Block_dead(block)) {
6289 /* control flow from dead block is dead */
6290 return get_irg_bad(irg);
6293 for (i = get_irn_arity(block) - 1; i >= 0; --i) {
6294 if (!is_Bad(get_irn_n(block, i)))
6298 ir_graph *irg = get_irn_irg(block);
6299 /* the start block is never dead */
6300 if (block != get_irg_start_block(irg)
6301 && block != get_irg_end_block(irg)) {
6303 * Do NOT kill control flow without setting
6304 * the block to dead of bad things can happen:
6305 * We get a Block that is not reachable be irg_block_walk()
6306 * but can be found by irg_walk()!
6308 set_Block_dead(block);
6309 return get_irg_bad(irg);
6315 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6316 blocks predecessors is dead. */
6317 if (op != op_Block && op != op_Phi && op != op_Tuple && op != op_Anchor) {
6318 ir_graph *irg = get_irn_irg(node);
6319 irn_arity = get_irn_arity(node);
6322 * Beware: we can only read the block of a non-floating node.
6324 if (is_irn_pinned_in_irg(node) &&
6325 is_Block_dead(get_nodes_block(skip_Proj(node))))
6326 return get_irg_bad(irg);
6328 for (i = 0; i < irn_arity; i++) {
6329 ir_node *pred = get_irn_n(node, i);
6332 return get_irg_bad(irg);
6334 /* Propagating Unknowns here seems to be a bad idea, because
6335 sometimes we need a node as a input and did not want that
6337 However, it might be useful to move this into a later phase
6338 (if you think that optimizing such code is useful). */
6339 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
6340 return new_r_Unknown(irg, get_irn_mode(node));
6345 /* With this code we violate the agreement that local_optimize
6346 only leaves Bads in Block, Phi and Tuple nodes. */
6347 /* If Block has only Bads as predecessors it's garbage. */
6348 /* If Phi has only Bads as predecessors it's garbage. */
6349 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
6350 irn_arity = get_irn_arity(node);
6351 for (i = 0; i < irn_arity; i++) {
6352 if (!is_Bad(get_irn_n(node, i))) break;
6354 if (i == irn_arity) node = get_irg_bad(irg);
6361 * These optimizations deallocate nodes from the obstack.
6362 * It can only be called if it is guaranteed that no other nodes
6363 * reference this one, i.e., right after construction of a node.
6365 * @param n The node to optimize
6367 ir_node *optimize_node(ir_node *n)
6370 ir_graph *irg = get_irn_irg(n);
6371 unsigned iro = get_irn_opcode(n);
6374 /* Always optimize Phi nodes: part of the construction. */
6375 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6377 /* constant expression evaluation / constant folding */
6378 if (get_opt_constant_folding()) {
6379 /* neither constants nor Tuple values can be evaluated */
6380 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6381 /* try to evaluate */
6382 tv = computed_value(n);
6383 if (tv != tarval_bad) {
6388 * we MUST copy the node here temporary, because it's still
6389 * needed for DBG_OPT_CSTEVAL
6391 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6392 oldn = (ir_node*)alloca(node_size);
6394 memcpy(oldn, n, node_size);
6395 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6397 /* ARG, copy the in array, we need it for statistics */
6398 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6400 /* note the inplace edges module */
6401 edges_node_deleted(n, irg);
6403 /* evaluation was successful -- replace the node. */
6404 irg_kill_node(irg, n);
6405 nw = new_r_Const(irg, tv);
6407 DBG_OPT_CSTEVAL(oldn, nw);
6413 /* remove unnecessary nodes */
6414 if (get_opt_algebraic_simplification() ||
6415 (iro == iro_Phi) || /* always optimize these nodes. */
6417 (iro == iro_Proj) ||
6418 (iro == iro_Block) ) /* Flags tested local. */
6419 n = equivalent_node(n);
6421 /* Common Subexpression Elimination.
6423 * Checks whether n is already available.
6424 * The block input is used to distinguish different subexpressions. Right
6425 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6426 * subexpressions within a block.
6429 n = identify_cons(n);
6432 edges_node_deleted(oldn, irg);
6434 /* We found an existing, better node, so we can deallocate the old node. */
6435 irg_kill_node(irg, oldn);
6439 /* Some more constant expression evaluation that does not allow to
6441 iro = get_irn_opcode(n);
6442 if (get_opt_algebraic_simplification() ||
6443 (iro == iro_Cond) ||
6444 (iro == iro_Proj)) /* Flags tested local. */
6445 n = transform_node(n);
6447 /* Remove nodes with dead (Bad) input.
6448 Run always for transformation induced Bads. */
6451 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6452 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6454 n = identify_remember(o);
6460 } /* optimize_node */
6464 * These optimizations never deallocate nodes (in place). This can cause dead
6465 * nodes lying on the obstack. Remove these by a dead node elimination,
6466 * i.e., a copying garbage collection.
6468 ir_node *optimize_in_place_2(ir_node *n)
6472 unsigned iro = get_irn_opcode(n);
6474 if (!get_opt_optimize() && !is_Phi(n)) return n;
6476 if (iro == iro_Deleted)
6479 /* constant expression evaluation / constant folding */
6480 if (get_opt_constant_folding()) {
6481 /* neither constants nor Tuple values can be evaluated */
6482 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6483 /* try to evaluate */
6484 tv = computed_value(n);
6485 if (tv != tarval_bad) {
6486 /* evaluation was successful -- replace the node. */
6487 ir_graph *irg = get_irn_irg(n);
6489 n = new_r_Const(irg, tv);
6491 DBG_OPT_CSTEVAL(oldn, n);
6497 /* remove unnecessary nodes */
6498 if (get_opt_constant_folding() ||
6499 (iro == iro_Phi) || /* always optimize these nodes. */
6500 (iro == iro_Id) || /* ... */
6501 (iro == iro_Proj) || /* ... */
6502 (iro == iro_Block) ) /* Flags tested local. */
6503 n = equivalent_node(n);
6505 /** common subexpression elimination **/
6506 /* Checks whether n is already available. */
6507 /* The block input is used to distinguish different subexpressions. Right
6508 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6509 subexpressions within a block. */
6510 if (get_opt_cse()) {
6512 n = identify_remember(o);
6517 /* Some more constant expression evaluation. */
6518 iro = get_irn_opcode(n);
6519 if (get_opt_constant_folding() ||
6520 (iro == iro_Cond) ||
6521 (iro == iro_Proj)) /* Flags tested local. */
6522 n = transform_node(n);
6524 /* Remove nodes with dead (Bad) input.
6525 Run always for transformation induced Bads. */
6528 /* Now we can verify the node, as it has no dead inputs any more. */
6531 /* Now we have a legal, useful node. Enter it in hash table for cse.
6532 Blocks should be unique anyways. (Except the successor of start:
6533 is cse with the start block!) */
6534 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6536 n = identify_remember(o);
6542 } /* optimize_in_place_2 */
6545 * Wrapper for external use, set proper status bits after optimization.
6547 ir_node *optimize_in_place(ir_node *n)
6549 ir_graph *irg = get_irn_irg(n);
6550 /* Handle graph state */
6551 assert(get_irg_phase_state(irg) != phase_building);
6553 if (get_opt_global_cse())
6554 set_irg_pinned(irg, op_pin_state_floats);
6555 if (get_irg_outs_state(irg) == outs_consistent)
6556 set_irg_outs_inconsistent(irg);
6558 /* FIXME: Maybe we could also test whether optimizing the node can
6559 change the control graph. */
6560 set_irg_doms_inconsistent(irg);
6561 return optimize_in_place_2(n);
6562 } /* optimize_in_place */
6565 * Calculate a hash value of a Const node.
6567 static unsigned hash_Const(const ir_node *node)
6571 /* special value for const, as they only differ in their tarval. */
6572 h = HASH_PTR(node->attr.con.tarval);
6578 * Calculate a hash value of a SymConst node.
6580 static unsigned hash_SymConst(const ir_node *node)
6584 /* all others are pointers */
6585 h = HASH_PTR(node->attr.symc.sym.type_p);
6588 } /* hash_SymConst */
6591 * Set the default hash operation in an ir_op_ops.
6593 * @param code the opcode for the default operation
6594 * @param ops the operations initialized
6599 static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops)
6603 ops->hash = hash_##a; \
6606 /* hash function already set */
6607 if (ops->hash != NULL)
6614 /* use input/mode default hash if no function was given */
6615 ops->hash = firm_default_hash;
6623 * Sets the default operation for an ir_ops.
6625 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops)
6627 ops = firm_set_default_hash(code, ops);
6628 ops = firm_set_default_computed_value(code, ops);
6629 ops = firm_set_default_equivalent_node(code, ops);
6630 ops = firm_set_default_transform_node(code, ops);
6631 ops = firm_set_default_node_cmp_attr(code, ops);
6632 ops = firm_set_default_get_type_attr(code, ops);
6633 ops = firm_set_default_get_entity_attr(code, ops);
6636 } /* firm_set_default_operations */