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) == ir_relation_less ? 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_relation(n) == ir_relation_equal) {
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 * gives a (conservative) estimation of possible relation when comparing
450 ir_relation ir_get_possible_cmp_relations(const ir_node *left,
451 const ir_node *right)
453 ir_relation possible = ir_relation_true;
454 ir_tarval *tv_l = value_of(left);
455 ir_tarval *tv_r = value_of(right);
456 ir_mode *mode = get_irn_mode(left);
457 ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode);
458 ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode);
460 /* both values known - evaluate them */
461 if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) {
462 possible = tarval_cmp(tv_l, tv_r);
463 /* we can return now, won't get any better */
466 /* a == a is never less or greater (but might be equal or unordered) */
468 possible &= ~ir_relation_less_greater;
469 /* unordered results only happen for float compares */
470 if (!mode_is_float(mode))
471 possible &= ~ir_relation_unordered;
472 /* values can never be less than the least representable number or
473 * greater than the greatest representable number */
475 possible &= ~ir_relation_greater;
477 possible &= ~ir_relation_less;
479 possible &= ~ir_relation_greater;
481 possible &= ~ir_relation_less;
482 /* maybe vrp can tell us more */
483 possible &= vrp_cmp(left, right);
484 /* Alloc nodes never return null (but throw an exception) */
485 if (is_Alloc(left) && tarval_is_null(tv_r))
486 possible &= ~ir_relation_equal;
492 * Return the value of a Cmp.
494 * The basic idea here is to determine which relations are possible and which
495 * one are definitely impossible.
497 static ir_tarval *computed_value_Cmp(const ir_node *cmp)
499 ir_node *left = get_Cmp_left(cmp);
500 ir_node *right = get_Cmp_right(cmp);
501 ir_relation possible = ir_get_possible_cmp_relations(left, right);
502 ir_relation relation = get_Cmp_relation(cmp);
504 /* if none of the requested relations is possible, return false */
505 if ((possible & relation) == ir_relation_false)
506 return tarval_b_false;
507 /* if possible relations are a subset of the requested ones return true */
508 if ((possible & ~relation) == ir_relation_false)
509 return tarval_b_true;
511 return computed_value_Cmp_Confirm(cmp, left, right, relation);
515 * Calculate the value of an integer Div.
516 * Special case: 0 / b
518 static ir_tarval *do_computed_value_Div(const ir_node *div)
520 const ir_node *a = get_Div_left(div);
521 const ir_node *b = get_Div_right(div);
522 const ir_mode *mode = get_Div_resmode(div);
523 ir_tarval *ta = value_of(a);
525 const ir_node *dummy;
527 /* cannot optimize 0 / b = 0 because of NaN */
528 if (!mode_is_float(mode)) {
529 if (tarval_is_null(ta) && value_not_zero(b, &dummy))
530 return ta; /* 0 / b == 0 if b != 0 */
533 if (ta != tarval_bad && tb != tarval_bad)
534 return tarval_div(ta, tb);
536 } /* do_computed_value_Div */
539 * Calculate the value of an integer Mod of two nodes.
540 * Special case: a % 1
542 static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b)
544 ir_tarval *ta = value_of(a);
545 ir_tarval *tb = value_of(b);
547 /* Compute a % 1 or c1 % c2 */
548 if (tarval_is_one(tb))
549 return get_mode_null(get_irn_mode(a));
550 if (ta != tarval_bad && tb != tarval_bad)
551 return tarval_mod(ta, tb);
553 } /* do_computed_value_Mod */
556 * Return the value of a Proj(Div).
558 static ir_tarval *computed_value_Proj_Div(const ir_node *n)
560 long proj_nr = get_Proj_proj(n);
561 if (proj_nr != pn_Div_res)
564 return do_computed_value_Div(get_Proj_pred(n));
565 } /* computed_value_Proj_Div */
568 * Return the value of a Proj(Mod).
570 static ir_tarval *computed_value_Proj_Mod(const ir_node *n)
572 long proj_nr = get_Proj_proj(n);
574 if (proj_nr == pn_Mod_res) {
575 const ir_node *mod = get_Proj_pred(n);
576 return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod));
579 } /* computed_value_Proj_Mod */
582 * Return the value of a Proj.
584 static ir_tarval *computed_value_Proj(const ir_node *proj)
586 ir_node *n = get_Proj_pred(proj);
588 if (n->op->ops.computed_value_Proj != NULL)
589 return n->op->ops.computed_value_Proj(proj);
591 } /* computed_value_Proj */
594 * If the parameter n can be computed, return its value, else tarval_bad.
595 * Performs constant folding.
597 * @param n The node this should be evaluated
599 ir_tarval *computed_value(const ir_node *n)
601 vrp_attr *vrp = vrp_get_info(n);
602 if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == ir_relation_equal) {
603 return vrp->bits_set;
605 if (n->op->ops.computed_value)
606 return n->op->ops.computed_value(n);
608 } /* computed_value */
611 * Set the default computed_value evaluator in an ir_op_ops.
613 * @param code the opcode for the default operation
614 * @param ops the operations initialized
619 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
623 ops->computed_value = computed_value_##a; \
625 #define CASE_PROJ(a) \
627 ops->computed_value_Proj = computed_value_Proj_##a; \
662 } /* firm_set_default_computed_value */
665 * Returns a equivalent block for another block.
666 * If the block has only one predecessor, this is
667 * the equivalent one. If the only predecessor of a block is
668 * the block itself, this is a dead block.
670 * If both predecessors of a block are the branches of a binary
671 * Cond, the equivalent block is Cond's block.
673 * If all predecessors of a block are bad or lies in a dead
674 * block, the current block is dead as well.
676 static ir_node *equivalent_node_Block(ir_node *n)
682 /* don't optimize dead or labeled blocks */
683 if (has_Block_entity(n))
686 n_preds = get_Block_n_cfgpreds(n);
688 /* The Block constructor does not call optimize, but mature_immBlock()
689 calls the optimization. */
690 assert(get_Block_matured(n));
692 irg = get_irn_irg(n);
694 /* if all predecessors of a block are unreachable, then the block is
696 if (is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) {
698 int n_cfgpreds = get_Block_n_cfgpreds(n);
700 for (i = 0; i < n_cfgpreds; ++i) {
701 ir_node *pred = get_Block_cfgpred(n, i);
705 /* only bad inputs? It's unreachable code (unless it is the start or
707 if (i >= n_cfgpreds && n != get_irg_start_block(irg)
708 && n != get_irg_end_block(irg)) {
709 return get_irg_bad(irg);
713 /* Straightening: a single entry Block following a single exit Block
716 ir_node *pred = get_Block_cfgpred(n, 0);
719 ir_node *pred_block = get_nodes_block(pred);
720 DBG_OPT_STG(n, pred_block);
723 } else if (n_preds == 2) {
724 /* Test whether Cond jumps twice to this block
725 * The more general case which more than 2 predecessors is handles
726 * in optimize_cf(), we handle only this special case for speed here.
728 ir_node *a = get_Block_cfgpred(n, 0);
729 ir_node *b = get_Block_cfgpred(n, 1);
731 if (is_Proj(a) && is_Proj(b)) {
732 ir_node *cond = get_Proj_pred(a);
734 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
735 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
736 /* Also a single entry Block following a single exit Block.
737 * Phis have twice the same operand and will be optimized away.
739 n = get_nodes_block(cond);
740 DBG_OPT_IFSIM1(oldn, a, b, n);
746 } /* equivalent_node_Block */
748 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
749 See transform_node_Proj_Cond(). */
752 * Optimize operations that are commutative and have neutral 0,
753 * so a op 0 = 0 op a = a.
755 static ir_node *equivalent_node_neutral_zero(ir_node *n)
759 ir_node *a = get_binop_left(n);
760 ir_node *b = get_binop_right(n);
765 /* After running compute_node there is only one constant predecessor.
766 Find this predecessors value and remember the other node: */
767 if ((tv = value_of(a)) != tarval_bad) {
769 } else if ((tv = value_of(b)) != tarval_bad) {
774 /* If this predecessors constant value is zero, the operation is
775 * unnecessary. Remove it.
777 * Beware: If n is a Add, the mode of on and n might be different
778 * which happens in this rare construction: NULL + 3.
779 * Then, a Conv would be needed which we cannot include here.
781 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
784 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
788 } /* equivalent_node_neutral_zero */
791 * Eor is commutative and has neutral 0.
793 static ir_node *equivalent_node_Eor(ir_node *n)
799 n = equivalent_node_neutral_zero(n);
800 if (n != oldn) return n;
803 b = get_Eor_right(n);
806 ir_node *aa = get_Eor_left(a);
807 ir_node *ab = get_Eor_right(a);
810 /* (a ^ b) ^ a -> b */
812 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
814 } else if (ab == b) {
815 /* (a ^ b) ^ b -> a */
817 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
822 ir_node *ba = get_Eor_left(b);
823 ir_node *bb = get_Eor_right(b);
826 /* a ^ (a ^ b) -> b */
828 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
830 } else if (bb == a) {
831 /* a ^ (b ^ a) -> b */
833 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
841 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
843 * The second one looks strange, but this construct
844 * is used heavily in the LCC sources :-).
846 * Beware: The Mode of an Add may be different than the mode of its
847 * predecessors, so we could not return a predecessors in all cases.
849 static ir_node *equivalent_node_Add(ir_node *n)
852 ir_node *left, *right;
853 ir_mode *mode = get_irn_mode(n);
855 n = equivalent_node_neutral_zero(n);
859 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
860 if (mode_is_float(mode)) {
861 ir_graph *irg = get_irn_irg(n);
862 if (get_irg_fp_model(irg) & fp_strict_algebraic)
866 left = get_Add_left(n);
867 right = get_Add_right(n);
870 if (get_Sub_right(left) == right) {
873 n = get_Sub_left(left);
874 if (mode == get_irn_mode(n)) {
875 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
881 if (get_Sub_right(right) == left) {
884 n = get_Sub_left(right);
885 if (mode == get_irn_mode(n)) {
886 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
892 } /* equivalent_node_Add */
895 * optimize operations that are not commutative but have neutral 0 on left,
898 static ir_node *equivalent_node_left_zero(ir_node *n)
902 ir_node *a = get_binop_left(n);
903 ir_node *b = get_binop_right(n);
904 ir_tarval *tb = value_of(b);
906 if (tarval_is_null(tb)) {
909 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
912 } /* equivalent_node_left_zero */
914 #define equivalent_node_Shl equivalent_node_left_zero
915 #define equivalent_node_Shr equivalent_node_left_zero
916 #define equivalent_node_Shrs equivalent_node_left_zero
917 #define equivalent_node_Rotl equivalent_node_left_zero
920 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
922 * The second one looks strange, but this construct
923 * is used heavily in the LCC sources :-).
925 * Beware: The Mode of a Sub may be different than the mode of its
926 * predecessors, so we could not return a predecessors in all cases.
928 static ir_node *equivalent_node_Sub(ir_node *n)
932 ir_mode *mode = get_irn_mode(n);
935 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
936 if (mode_is_float(mode)) {
937 ir_graph *irg = get_irn_irg(n);
938 if (get_irg_fp_model(irg) & fp_strict_algebraic)
942 b = get_Sub_right(n);
945 /* Beware: modes might be different */
946 if (tarval_is_null(tb)) {
947 ir_node *a = get_Sub_left(n);
948 if (mode == get_irn_mode(a)) {
951 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
955 } /* equivalent_node_Sub */
959 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
962 * -(-a) == a, but might overflow two times.
963 * We handle it anyway here but the better way would be a
964 * flag. This would be needed for Pascal for instance.
966 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
969 ir_node *pred = get_unop_op(n);
971 /* optimize symmetric unop */
972 if (get_irn_op(pred) == get_irn_op(n)) {
973 n = get_unop_op(pred);
974 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
977 } /* equivalent_node_idempotent_unop */
979 /** Optimize Not(Not(x)) == x. */
980 #define equivalent_node_Not equivalent_node_idempotent_unop
982 /** -(-x) == x ??? Is this possible or can --x raise an
983 out of bounds exception if min =! max? */
984 #define equivalent_node_Minus equivalent_node_idempotent_unop
987 * Optimize a * 1 = 1 * a = a.
989 static ir_node *equivalent_node_Mul(ir_node *n)
992 ir_node *a = get_Mul_left(n);
994 /* we can handle here only the n * n = n bit cases */
995 if (get_irn_mode(n) == get_irn_mode(a)) {
996 ir_node *b = get_Mul_right(n);
1000 * Mul is commutative and has again an other neutral element.
1001 * Constants are place right, so check this case first.
1004 if (tarval_is_one(tv)) {
1006 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1009 if (tarval_is_one(tv)) {
1011 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1016 } /* equivalent_node_Mul */
1019 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1021 static ir_node *equivalent_node_Or(ir_node *n)
1025 ir_node *a = get_Or_left(n);
1026 ir_node *b = get_Or_right(n);
1030 n = a; /* idempotence */
1031 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1034 /* constants are normalized to right, check this side first */
1036 if (tarval_is_null(tv)) {
1038 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1042 if (tarval_is_null(tv)) {
1044 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1049 } /* equivalent_node_Or */
1052 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1054 static ir_node *equivalent_node_And(ir_node *n)
1058 ir_node *a = get_And_left(n);
1059 ir_node *b = get_And_right(n);
1063 n = a; /* idempotence */
1064 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1067 /* constants are normalized to right, check this side first */
1069 if (tarval_is_all_one(tv)) {
1071 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1074 if (tv != get_tarval_bad()) {
1075 ir_mode *mode = get_irn_mode(n);
1076 if (!mode_is_signed(mode) && is_Conv(a)) {
1077 ir_node *convop = get_Conv_op(a);
1078 ir_mode *convopmode = get_irn_mode(convop);
1079 if (!mode_is_signed(convopmode)) {
1080 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1081 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1083 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1090 if (tarval_is_all_one(tv)) {
1092 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1096 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1099 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1104 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1107 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1112 } /* equivalent_node_And */
1115 * Try to remove useless Conv's:
1117 static ir_node *equivalent_node_Conv(ir_node *n)
1120 ir_node *a = get_Conv_op(n);
1122 ir_mode *n_mode = get_irn_mode(n);
1123 ir_mode *a_mode = get_irn_mode(a);
1126 if (n_mode == a_mode) { /* No Conv necessary */
1127 if (get_Conv_strict(n)) {
1130 /* neither Minus nor Confirm change the precision,
1131 so we can "look-through" */
1134 p = get_Minus_op(p);
1135 } else if (is_Confirm(p)) {
1136 p = get_Confirm_value(p);
1142 if (is_Conv(p) && get_Conv_strict(p)) {
1143 /* we known already, that a_mode == n_mode, and neither
1144 Minus change the mode, so the second Conv
1146 assert(get_irn_mode(p) == n_mode);
1148 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1152 ir_node *pred = get_Proj_pred(p);
1153 if (is_Load(pred)) {
1154 /* Loads always return with the exact precision of n_mode */
1155 assert(get_Load_mode(pred) == n_mode);
1157 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1160 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1161 pred = get_Proj_pred(pred);
1162 if (is_Start(pred)) {
1163 /* Arguments always return with the exact precision,
1164 as strictConv's are place before Call -- if the
1165 caller was compiled with the same setting.
1166 Otherwise, the semantics is probably still right. */
1167 assert(get_irn_mode(p) == n_mode);
1169 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1175 /* special case: the immediate predecessor is also a Conv */
1176 if (! get_Conv_strict(a)) {
1177 /* first one is not strict, kick it */
1179 a_mode = get_irn_mode(a);
1183 /* else both are strict conv, second is superfluous */
1185 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1190 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1193 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1194 ir_node *b = get_Conv_op(a);
1195 ir_mode *b_mode = get_irn_mode(b);
1197 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1198 /* both are strict conv */
1199 if (smaller_mode(a_mode, n_mode)) {
1200 /* both are strict, but the first is smaller, so
1201 the second cannot remove more precision, remove the
1203 set_Conv_strict(n, 0);
1206 if (n_mode == b_mode) {
1207 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1208 if (n_mode == mode_b) {
1209 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1210 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1212 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1213 if (values_in_mode(b_mode, a_mode)) {
1214 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1215 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1220 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1221 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1222 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1223 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1225 if (float_mantissa >= int_mantissa) {
1227 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1232 if (smaller_mode(b_mode, a_mode)) {
1233 if (get_Conv_strict(n))
1234 set_Conv_strict(b, 1);
1235 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1236 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1243 } /* equivalent_node_Conv */
1246 * - fold Phi-nodes, iff they have only one predecessor except
1249 static ir_node *equivalent_node_Phi(ir_node *n)
1255 ir_node *first_val = NULL; /* to shutup gcc */
1257 if (!get_opt_optimize() &&
1258 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1261 n_preds = get_Phi_n_preds(n);
1263 block = get_nodes_block(n);
1265 /* Phi of dead Region without predecessors. */
1269 /* Find first non-self-referencing input */
1270 for (i = 0; i < n_preds; ++i) {
1271 first_val = get_Phi_pred(n, i);
1272 /* not self pointer */
1273 if (first_val != n) {
1274 /* then found first value. */
1280 ir_graph *irg = get_irn_irg(n);
1281 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1282 return get_irg_bad(irg);
1285 /* search for rest of inputs, determine if any of these
1286 are non-self-referencing */
1287 while (++i < n_preds) {
1288 ir_node *scnd_val = get_Phi_pred(n, i);
1289 if (scnd_val != n && scnd_val != first_val) {
1294 if (i >= n_preds && !is_Dummy(first_val)) {
1295 /* Fold, if no multiple distinct non-self-referencing inputs */
1297 DBG_OPT_PHI(oldn, n);
1300 } /* equivalent_node_Phi */
1303 * Several optimizations:
1304 * - fold Sync-nodes, iff they have only one predecessor except
1307 static ir_node *equivalent_node_Sync(ir_node *n)
1309 int arity = get_Sync_n_preds(n);
1312 for (i = 0; i < arity;) {
1313 ir_node *pred = get_Sync_pred(n, i);
1316 /* Remove Bad predecessors */
1323 /* Remove duplicate predecessors */
1329 if (get_Sync_pred(n, j) == pred) {
1338 ir_graph *irg = get_irn_irg(n);
1339 return get_irg_bad(irg);
1341 if (arity == 1) return get_Sync_pred(n, 0);
1343 } /* equivalent_node_Sync */
1346 * Optimize Proj(Tuple).
1348 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1350 ir_node *oldn = proj;
1351 ir_node *tuple = get_Proj_pred(proj);
1353 /* Remove the Tuple/Proj combination. */
1354 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1355 DBG_OPT_TUPLE(oldn, tuple, proj);
1358 } /* equivalent_node_Proj_Tuple */
1361 * Optimize a / 1 = a.
1363 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1365 ir_node *oldn = proj;
1366 ir_node *div = get_Proj_pred(proj);
1367 ir_node *b = get_Div_right(div);
1368 ir_tarval *tb = value_of(b);
1370 /* Div is not commutative. */
1371 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1372 switch (get_Proj_proj(proj)) {
1374 proj = get_Div_mem(div);
1375 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1379 proj = get_Div_left(div);
1380 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1384 /* we cannot replace the exception Proj's here, this is done in
1385 transform_node_Proj_Div() */
1390 } /* equivalent_node_Proj_Div */
1393 * Optimize CopyB(mem, x, x) into a Nop.
1395 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1397 ir_node *oldn = proj;
1398 ir_node *copyb = get_Proj_pred(proj);
1399 ir_node *a = get_CopyB_dst(copyb);
1400 ir_node *b = get_CopyB_src(copyb);
1403 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1404 switch (get_Proj_proj(proj)) {
1406 proj = get_CopyB_mem(copyb);
1407 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1410 case pn_CopyB_X_except: {
1411 ir_graph *irg = get_irn_irg(proj);
1412 DBG_OPT_EXC_REM(proj);
1413 proj = get_irg_bad(irg);
1419 } /* equivalent_node_Proj_CopyB */
1422 * Optimize Bounds(idx, idx, upper) into idx.
1424 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1426 ir_node *oldn = proj;
1427 ir_node *bound = get_Proj_pred(proj);
1428 ir_node *idx = get_Bound_index(bound);
1429 ir_node *pred = skip_Proj(idx);
1432 if (idx == get_Bound_lower(bound))
1434 else if (is_Bound(pred)) {
1436 * idx was Bounds checked previously, it is still valid if
1437 * lower <= pred_lower && pred_upper <= upper.
1439 ir_node *lower = get_Bound_lower(bound);
1440 ir_node *upper = get_Bound_upper(bound);
1441 if (get_Bound_lower(pred) == lower &&
1442 get_Bound_upper(pred) == upper) {
1444 * One could expect that we simply return the previous
1445 * Bound here. However, this would be wrong, as we could
1446 * add an exception Proj to a new location then.
1447 * So, we must turn in into a tuple.
1453 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1454 switch (get_Proj_proj(proj)) {
1456 DBG_OPT_EXC_REM(proj);
1457 proj = get_Bound_mem(bound);
1459 case pn_Bound_X_except: {
1460 ir_graph *irg = get_irn_irg(proj);
1461 DBG_OPT_EXC_REM(proj);
1462 proj = get_irg_bad(irg);
1467 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1470 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1475 } /* equivalent_node_Proj_Bound */
1478 * Optimize an Exception Proj(Load) with a non-null address.
1480 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1482 if (get_opt_ldst_only_null_ptr_exceptions()) {
1483 if (get_irn_mode(proj) == mode_X) {
1484 ir_node *load = get_Proj_pred(proj);
1486 /* get the Load address */
1487 const ir_node *addr = get_Load_ptr(load);
1488 const ir_node *confirm;
1490 if (value_not_null(addr, &confirm)) {
1491 if (get_Proj_proj(proj) == pn_Load_X_except) {
1492 ir_graph *irg = get_irn_irg(proj);
1493 DBG_OPT_EXC_REM(proj);
1494 return get_irg_bad(irg);
1500 } /* equivalent_node_Proj_Load */
1503 * Optimize an Exception Proj(Store) with a non-null address.
1505 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1507 if (get_opt_ldst_only_null_ptr_exceptions()) {
1508 if (get_irn_mode(proj) == mode_X) {
1509 ir_node *store = get_Proj_pred(proj);
1511 /* get the load/store address */
1512 const ir_node *addr = get_Store_ptr(store);
1513 const ir_node *confirm;
1515 if (value_not_null(addr, &confirm)) {
1516 if (get_Proj_proj(proj) == pn_Store_X_except) {
1517 ir_graph *irg = get_irn_irg(proj);
1518 DBG_OPT_EXC_REM(proj);
1519 return get_irg_bad(irg);
1525 } /* equivalent_node_Proj_Store */
1528 * Does all optimizations on nodes that must be done on its Projs
1529 * because of creating new nodes.
1531 static ir_node *equivalent_node_Proj(ir_node *proj)
1533 ir_node *n = get_Proj_pred(proj);
1534 if (n->op->ops.equivalent_node_Proj)
1535 return n->op->ops.equivalent_node_Proj(proj);
1537 } /* equivalent_node_Proj */
1542 static ir_node *equivalent_node_Id(ir_node *n)
1550 DBG_OPT_ID(oldn, n);
1552 } /* equivalent_node_Id */
1557 static ir_node *equivalent_node_Mux(ir_node *n)
1559 ir_node *oldn = n, *sel = get_Mux_sel(n);
1561 ir_tarval *ts = value_of(sel);
1563 /* Mux(true, f, t) == t */
1564 if (ts == tarval_b_true) {
1565 n = get_Mux_true(n);
1566 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1569 /* Mux(false, f, t) == f */
1570 if (ts == tarval_b_false) {
1571 n = get_Mux_false(n);
1572 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1575 n_t = get_Mux_true(n);
1576 n_f = get_Mux_false(n);
1578 /* Mux(v, x, T) == x */
1579 if (is_Unknown(n_f)) {
1581 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1584 /* Mux(v, T, x) == x */
1585 if (is_Unknown(n_t)) {
1587 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1591 /* Mux(v, x, x) == x */
1594 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1597 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1598 ir_relation relation = get_Cmp_relation(sel);
1599 ir_node *f = get_Mux_false(n);
1600 ir_node *t = get_Mux_true(n);
1603 * Note further that these optimization work even for floating point
1604 * with NaN's because -NaN == NaN.
1605 * However, if +0 and -0 is handled differently, we cannot use the first one.
1607 ir_node *const cmp_l = get_Cmp_left(sel);
1608 ir_node *const cmp_r = get_Cmp_right(sel);
1611 case ir_relation_equal:
1612 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1613 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1615 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1620 case ir_relation_less_greater:
1621 case ir_relation_unordered_less_greater:
1622 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1623 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1625 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1634 * Note: normalization puts the constant on the right side,
1635 * so we check only one case.
1637 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1638 /* Mux(t CMP 0, X, t) */
1639 if (is_Minus(f) && get_Minus_op(f) == t) {
1640 /* Mux(t CMP 0, -t, t) */
1641 if (relation == ir_relation_equal) {
1642 /* Mux(t == 0, -t, t) ==> -t */
1644 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1645 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1646 /* Mux(t != 0, -t, t) ==> t */
1648 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1658 * Remove Confirm nodes if setting is on.
1659 * Replace Confirms(x, '=', Constlike) by Constlike.
1661 static ir_node *equivalent_node_Confirm(ir_node *n)
1663 ir_node *pred = get_Confirm_value(n);
1664 ir_relation relation = get_Confirm_relation(n);
1666 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1668 * rare case: two identical Confirms one after another,
1669 * replace the second one with the first.
1672 pred = get_Confirm_value(n);
1678 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1679 * perform no actual computation, as, e.g., the Id nodes. It does not create
1680 * new nodes. It is therefore safe to free n if the node returned is not n.
1681 * If a node returns a Tuple we can not just skip it. If the size of the
1682 * in array fits, we transform n into a tuple (e.g., Div).
1684 ir_node *equivalent_node(ir_node *n)
1686 if (n->op->ops.equivalent_node)
1687 return n->op->ops.equivalent_node(n);
1689 } /* equivalent_node */
1692 * Sets the default equivalent node operation for an ir_op_ops.
1694 * @param code the opcode for the default operation
1695 * @param ops the operations initialized
1700 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1704 ops->equivalent_node = equivalent_node_##a; \
1706 #define CASE_PROJ(a) \
1708 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1746 } /* firm_set_default_equivalent_node */
1749 * Returns non-zero if a node is a Phi node
1750 * with all predecessors constant.
1752 static int is_const_Phi(ir_node *n)
1756 if (! is_Phi(n) || get_irn_arity(n) == 0)
1758 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1759 if (! is_Const(get_irn_n(n, i)))
1763 } /* is_const_Phi */
1765 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1766 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1769 * in reality eval_func should be tarval (*eval_func)() but incomplete
1770 * declarations are bad style and generate noisy warnings
1772 typedef void (*eval_func)(void);
1775 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1777 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1779 if (eval == (eval_func) tarval_sub) {
1780 tarval_sub_type func = (tarval_sub_type)eval;
1782 return func(a, b, mode);
1784 tarval_binop_type func = (tarval_binop_type)eval;
1791 * Apply an evaluator on a binop with a constant operators (and one Phi).
1793 * @param phi the Phi node
1794 * @param other the other operand
1795 * @param eval an evaluator function
1796 * @param mode the mode of the result, may be different from the mode of the Phi!
1797 * @param left if non-zero, other is the left operand, else the right
1799 * @return a new Phi node if the conversion was successful, NULL else
1801 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1807 int i, n = get_irn_arity(phi);
1809 NEW_ARR_A(void *, res, n);
1811 for (i = 0; i < n; ++i) {
1812 pred = get_irn_n(phi, i);
1813 tv = get_Const_tarval(pred);
1814 tv = do_eval(eval, other, tv, mode);
1816 if (tv == tarval_bad) {
1817 /* folding failed, bad */
1823 for (i = 0; i < n; ++i) {
1824 pred = get_irn_n(phi, i);
1825 tv = get_Const_tarval(pred);
1826 tv = do_eval(eval, tv, other, mode);
1828 if (tv == tarval_bad) {
1829 /* folding failed, bad */
1835 irg = get_irn_irg(phi);
1836 for (i = 0; i < n; ++i) {
1837 pred = get_irn_n(phi, i);
1838 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1840 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1841 } /* apply_binop_on_phi */
1844 * Apply an evaluator on a binop with two constant Phi.
1846 * @param a the left Phi node
1847 * @param b the right Phi node
1848 * @param eval an evaluator function
1849 * @param mode the mode of the result, may be different from the mode of the Phi!
1851 * @return a new Phi node if the conversion was successful, NULL else
1853 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1855 ir_tarval *tv_l, *tv_r, *tv;
1861 if (get_nodes_block(a) != get_nodes_block(b))
1864 n = get_irn_arity(a);
1865 NEW_ARR_A(void *, res, n);
1867 for (i = 0; i < n; ++i) {
1868 pred = get_irn_n(a, i);
1869 tv_l = get_Const_tarval(pred);
1870 pred = get_irn_n(b, i);
1871 tv_r = get_Const_tarval(pred);
1872 tv = do_eval(eval, tv_l, tv_r, mode);
1874 if (tv == tarval_bad) {
1875 /* folding failed, bad */
1880 irg = get_irn_irg(a);
1881 for (i = 0; i < n; ++i) {
1882 pred = get_irn_n(a, i);
1883 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1885 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1886 } /* apply_binop_on_2_phis */
1889 * Apply an evaluator on a unop with a constant operator (a Phi).
1891 * @param phi the Phi node
1892 * @param eval an evaluator function
1894 * @return a new Phi node if the conversion was successful, NULL else
1896 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1903 int i, n = get_irn_arity(phi);
1905 NEW_ARR_A(void *, res, n);
1906 for (i = 0; i < n; ++i) {
1907 pred = get_irn_n(phi, i);
1908 tv = get_Const_tarval(pred);
1911 if (tv == tarval_bad) {
1912 /* folding failed, bad */
1917 mode = get_irn_mode(phi);
1918 irg = get_irn_irg(phi);
1919 for (i = 0; i < n; ++i) {
1920 pred = get_irn_n(phi, i);
1921 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1923 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1924 } /* apply_unop_on_phi */
1927 * Apply a conversion on a constant operator (a Phi).
1929 * @param phi the Phi node
1931 * @return a new Phi node if the conversion was successful, NULL else
1933 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1939 int i, n = get_irn_arity(phi);
1941 NEW_ARR_A(void *, res, n);
1942 for (i = 0; i < n; ++i) {
1943 pred = get_irn_n(phi, i);
1944 tv = get_Const_tarval(pred);
1945 tv = tarval_convert_to(tv, mode);
1947 if (tv == tarval_bad) {
1948 /* folding failed, bad */
1953 irg = get_irn_irg(phi);
1954 for (i = 0; i < n; ++i) {
1955 pred = get_irn_n(phi, i);
1956 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1958 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1959 } /* apply_conv_on_phi */
1962 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1963 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1964 * If possible, remove the Conv's.
1966 static ir_node *transform_node_AddSub(ir_node *n)
1968 ir_mode *mode = get_irn_mode(n);
1970 if (mode_is_reference(mode)) {
1971 ir_node *left = get_binop_left(n);
1972 ir_node *right = get_binop_right(n);
1973 unsigned ref_bits = get_mode_size_bits(mode);
1975 if (is_Conv(left)) {
1976 ir_mode *lmode = get_irn_mode(left);
1977 unsigned bits = get_mode_size_bits(lmode);
1979 if (ref_bits == bits &&
1980 mode_is_int(lmode) &&
1981 get_mode_arithmetic(lmode) == irma_twos_complement) {
1982 ir_node *pre = get_Conv_op(left);
1983 ir_mode *pre_mode = get_irn_mode(pre);
1985 if (mode_is_int(pre_mode) &&
1986 get_mode_size_bits(pre_mode) == bits &&
1987 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1988 /* ok, this conv just changes to sign, moreover the calculation
1989 * is done with same number of bits as our address mode, so
1990 * we can ignore the conv as address calculation can be viewed
1991 * as either signed or unsigned
1993 set_binop_left(n, pre);
1998 if (is_Conv(right)) {
1999 ir_mode *rmode = get_irn_mode(right);
2000 unsigned bits = get_mode_size_bits(rmode);
2002 if (ref_bits == bits &&
2003 mode_is_int(rmode) &&
2004 get_mode_arithmetic(rmode) == irma_twos_complement) {
2005 ir_node *pre = get_Conv_op(right);
2006 ir_mode *pre_mode = get_irn_mode(pre);
2008 if (mode_is_int(pre_mode) &&
2009 get_mode_size_bits(pre_mode) == bits &&
2010 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2011 /* ok, this conv just changes to sign, moreover the calculation
2012 * is done with same number of bits as our address mode, so
2013 * we can ignore the conv as address calculation can be viewed
2014 * as either signed or unsigned
2016 set_binop_right(n, pre);
2021 /* let address arithmetic use unsigned modes */
2022 if (is_Const(right)) {
2023 ir_mode *rmode = get_irn_mode(right);
2025 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2026 /* convert a AddP(P, *s) into AddP(P, *u) */
2027 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2029 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2030 set_binop_right(n, pre);
2036 } /* transform_node_AddSub */
2038 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2041 if (is_Const(b) && is_const_Phi(a)) { \
2042 /* check for Op(Phi, Const) */ \
2043 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2045 else if (is_Const(a) && is_const_Phi(b)) { \
2046 /* check for Op(Const, Phi) */ \
2047 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2049 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2050 /* check for Op(Phi, Phi) */ \
2051 c = apply_binop_on_2_phis(a, b, eval, mode); \
2054 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2059 #define HANDLE_UNOP_PHI(eval, a, c) \
2062 if (is_const_Phi(a)) { \
2063 /* check for Op(Phi) */ \
2064 c = apply_unop_on_phi(a, eval); \
2066 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2073 * Do the AddSub optimization, then Transform
2074 * Constant folding on Phi
2075 * Add(a,a) -> Mul(a, 2)
2076 * Add(Mul(a, x), a) -> Mul(a, x+1)
2077 * if the mode is integer or float.
2078 * Transform Add(a,-b) into Sub(a,b).
2079 * Reassociation might fold this further.
2081 static ir_node *transform_node_Add(ir_node *n)
2084 ir_node *a, *b, *c, *oldn = n;
2085 vrp_attr *a_vrp, *b_vrp;
2087 n = transform_node_AddSub(n);
2089 a = get_Add_left(n);
2090 b = get_Add_right(n);
2092 mode = get_irn_mode(n);
2094 if (mode_is_reference(mode)) {
2095 ir_mode *lmode = get_irn_mode(a);
2097 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2098 /* an Add(a, NULL) is a hidden Conv */
2099 dbg_info *dbg = get_irn_dbg_info(n);
2100 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2104 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2106 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2107 if (mode_is_float(mode)) {
2108 ir_graph *irg = get_irn_irg(n);
2109 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2113 if (mode_is_num(mode)) {
2114 ir_graph *irg = get_irn_irg(n);
2115 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2116 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2117 && a == b && mode_is_int(mode)) {
2118 ir_node *block = get_nodes_block(n);
2121 get_irn_dbg_info(n),
2124 new_r_Const_long(irg, mode, 2),
2126 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2131 get_irn_dbg_info(n),
2136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2141 get_irn_dbg_info(n),
2146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2149 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2150 /* Here we rely on constants be on the RIGHT side */
2152 ir_node *op = get_Not_op(a);
2154 if (is_Const(b) && is_Const_one(b)) {
2156 ir_node *blk = get_nodes_block(n);
2157 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2158 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2163 n = new_r_Const(irg, get_mode_minus_one(mode));
2164 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2169 ir_node *op = get_Not_op(b);
2173 n = new_r_Const(irg, get_mode_minus_one(mode));
2174 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2181 a_vrp = vrp_get_info(a);
2182 b_vrp = vrp_get_info(b);
2184 if (a_vrp && b_vrp) {
2185 ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2187 if (tarval_is_null(c)) {
2188 dbg_info *dbgi = get_irn_dbg_info(n);
2189 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2193 } /* transform_node_Add */
2196 * returns -cnst or NULL if impossible
2198 static ir_node *const_negate(ir_node *cnst)
2200 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2201 dbg_info *dbgi = get_irn_dbg_info(cnst);
2202 ir_graph *irg = get_irn_irg(cnst);
2203 if (tv == tarval_bad) return NULL;
2204 return new_rd_Const(dbgi, irg, tv);
2208 * Do the AddSub optimization, then Transform
2209 * Constant folding on Phi
2210 * Sub(0,a) -> Minus(a)
2211 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2212 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2213 * Sub(Add(a, x), x) -> a
2214 * Sub(x, Add(x, a)) -> -a
2215 * Sub(x, Const) -> Add(x, -Const)
2217 static ir_node *transform_node_Sub(ir_node *n)
2223 n = transform_node_AddSub(n);
2225 a = get_Sub_left(n);
2226 b = get_Sub_right(n);
2228 mode = get_irn_mode(n);
2230 if (mode_is_int(mode)) {
2231 ir_mode *lmode = get_irn_mode(a);
2233 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2234 /* a Sub(a, NULL) is a hidden Conv */
2235 dbg_info *dbg = get_irn_dbg_info(n);
2236 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2237 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2241 if (mode == lmode &&
2242 get_mode_arithmetic(mode) == irma_twos_complement &&
2244 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2246 dbg_info *dbg = get_irn_dbg_info(n);
2247 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2248 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2254 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2256 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2257 if (mode_is_float(mode)) {
2258 ir_graph *irg = get_irn_irg(n);
2259 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2263 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2264 /* a - C -> a + (-C) */
2265 ir_node *cnst = const_negate(b);
2267 ir_node *block = get_nodes_block(n);
2268 dbg_info *dbgi = get_irn_dbg_info(n);
2270 n = new_rd_Add(dbgi, block, a, cnst, mode);
2271 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2276 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2277 dbg_info *dbg = get_irn_dbg_info(n);
2278 ir_node *block = get_nodes_block(n);
2279 ir_node *left = get_Minus_op(a);
2280 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2282 n = new_rd_Minus(dbg, block, add, mode);
2283 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2285 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2286 dbg_info *dbg = get_irn_dbg_info(n);
2287 ir_node *block = get_nodes_block(n);
2288 ir_node *right = get_Minus_op(b);
2290 n = new_rd_Add(dbg, block, a, right, mode);
2291 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2293 } else if (is_Sub(b)) {
2294 /* a - (b - c) -> a + (c - b)
2295 * -> (a - b) + c iff (b - c) is a pointer */
2296 dbg_info *s_dbg = get_irn_dbg_info(b);
2297 ir_node *s_left = get_Sub_left(b);
2298 ir_node *s_right = get_Sub_right(b);
2299 ir_mode *s_mode = get_irn_mode(b);
2300 if (mode_is_reference(s_mode)) {
2301 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2302 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2303 dbg_info *a_dbg = get_irn_dbg_info(n);
2306 s_right = new_r_Conv(lowest_block, s_right, mode);
2307 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2309 ir_node *s_block = get_nodes_block(b);
2310 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2311 dbg_info *a_dbg = get_irn_dbg_info(n);
2312 ir_node *a_block = get_nodes_block(n);
2314 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2316 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2318 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2319 ir_node *m_right = get_Mul_right(b);
2320 if (is_Const(m_right)) {
2321 ir_node *cnst2 = const_negate(m_right);
2322 if (cnst2 != NULL) {
2323 dbg_info *m_dbg = get_irn_dbg_info(b);
2324 ir_node *m_block = get_nodes_block(b);
2325 ir_node *m_left = get_Mul_left(b);
2326 ir_mode *m_mode = get_irn_mode(b);
2327 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2328 dbg_info *a_dbg = get_irn_dbg_info(n);
2329 ir_node *a_block = get_nodes_block(n);
2331 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2332 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2338 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2339 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2341 get_irn_dbg_info(n),
2345 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2349 if (mode_wrap_around(mode)) {
2350 ir_node *left = get_Add_left(a);
2351 ir_node *right = get_Add_right(a);
2353 /* FIXME: Does the Conv's work only for two complement or generally? */
2355 if (mode != get_irn_mode(right)) {
2356 /* This Sub is an effective Cast */
2357 right = new_r_Conv(get_nodes_block(n), right, mode);
2360 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2362 } else if (right == b) {
2363 if (mode != get_irn_mode(left)) {
2364 /* This Sub is an effective Cast */
2365 left = new_r_Conv(get_nodes_block(n), left, mode);
2368 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2374 if (mode_wrap_around(mode)) {
2375 ir_node *left = get_Add_left(b);
2376 ir_node *right = get_Add_right(b);
2378 /* FIXME: Does the Conv's work only for two complement or generally? */
2380 ir_mode *r_mode = get_irn_mode(right);
2382 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2383 if (mode != r_mode) {
2384 /* This Sub is an effective Cast */
2385 n = new_r_Conv(get_nodes_block(n), n, mode);
2387 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2389 } else if (right == a) {
2390 ir_mode *l_mode = get_irn_mode(left);
2392 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2393 if (mode != l_mode) {
2394 /* This Sub is an effective Cast */
2395 n = new_r_Conv(get_nodes_block(n), n, mode);
2397 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2402 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2403 ir_mode *mode = get_irn_mode(a);
2405 if (mode == get_irn_mode(b)) {
2407 ir_node *op_a = get_Conv_op(a);
2408 ir_node *op_b = get_Conv_op(b);
2410 /* check if it's allowed to skip the conv */
2411 ma = get_irn_mode(op_a);
2412 mb = get_irn_mode(op_b);
2414 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2415 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2418 set_Sub_right(n, b);
2424 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2425 if (!is_reassoc_running() && is_Mul(a)) {
2426 ir_node *ma = get_Mul_left(a);
2427 ir_node *mb = get_Mul_right(a);
2430 ir_node *blk = get_nodes_block(n);
2431 ir_graph *irg = get_irn_irg(n);
2433 get_irn_dbg_info(n),
2437 get_irn_dbg_info(n),
2440 new_r_Const(irg, get_mode_one(mode)),
2443 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2445 } else if (mb == b) {
2446 ir_node *blk = get_nodes_block(n);
2447 ir_graph *irg = get_irn_irg(n);
2449 get_irn_dbg_info(n),
2453 get_irn_dbg_info(n),
2456 new_r_Const(irg, get_mode_one(mode)),
2459 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2463 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2464 ir_node *x = get_Sub_left(a);
2465 ir_node *y = get_Sub_right(a);
2466 ir_node *blk = get_nodes_block(n);
2467 ir_mode *m_b = get_irn_mode(b);
2468 ir_mode *m_y = get_irn_mode(y);
2472 /* Determine the right mode for the Add. */
2475 else if (mode_is_reference(m_b))
2477 else if (mode_is_reference(m_y))
2481 * Both modes are different but none is reference,
2482 * happens for instance in SubP(SubP(P, Iu), Is).
2483 * We have two possibilities here: Cast or ignore.
2484 * Currently we ignore this case.
2489 add = new_r_Add(blk, y, b, add_mode);
2491 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2492 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2496 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2497 if (is_Const(a) && is_Not(b)) {
2498 /* c - ~X = X + (c+1) */
2499 ir_tarval *tv = get_Const_tarval(a);
2501 tv = tarval_add(tv, get_mode_one(mode));
2502 if (tv != tarval_bad) {
2503 ir_node *blk = get_nodes_block(n);
2504 ir_graph *irg = get_irn_irg(n);
2505 ir_node *c = new_r_Const(irg, tv);
2506 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2507 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2513 } /* transform_node_Sub */
2516 * Several transformation done on n*n=2n bits mul.
2517 * These transformations must be done here because new nodes may be produced.
2519 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2522 ir_node *a = get_Mul_left(n);
2523 ir_node *b = get_Mul_right(n);
2524 ir_tarval *ta = value_of(a);
2525 ir_tarval *tb = value_of(b);
2526 ir_mode *smode = get_irn_mode(a);
2528 if (ta == get_mode_one(smode)) {
2529 /* (L)1 * (L)b = (L)b */
2530 ir_node *blk = get_nodes_block(n);
2531 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2532 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2535 else if (ta == get_mode_minus_one(smode)) {
2536 /* (L)-1 * (L)b = (L)b */
2537 ir_node *blk = get_nodes_block(n);
2538 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2539 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2540 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2543 if (tb == get_mode_one(smode)) {
2544 /* (L)a * (L)1 = (L)a */
2545 ir_node *blk = get_irn_n(a, -1);
2546 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2547 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2550 else if (tb == get_mode_minus_one(smode)) {
2551 /* (L)a * (L)-1 = (L)-a */
2552 ir_node *blk = get_nodes_block(n);
2553 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2554 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2555 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2562 * Transform Mul(a,-1) into -a.
2563 * Do constant evaluation of Phi nodes.
2564 * Do architecture dependent optimizations on Mul nodes
2566 static ir_node *transform_node_Mul(ir_node *n)
2568 ir_node *c, *oldn = n;
2569 ir_mode *mode = get_irn_mode(n);
2570 ir_node *a = get_Mul_left(n);
2571 ir_node *b = get_Mul_right(n);
2573 if (is_Bad(a) || is_Bad(b))
2576 if (mode != get_irn_mode(a))
2577 return transform_node_Mul2n(n, mode);
2579 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2581 if (mode_is_signed(mode)) {
2584 if (value_of(a) == get_mode_minus_one(mode))
2586 else if (value_of(b) == get_mode_minus_one(mode))
2589 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2590 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2595 if (is_Const(b)) { /* (-a) * const -> a * -const */
2596 ir_node *cnst = const_negate(b);
2598 dbg_info *dbgi = get_irn_dbg_info(n);
2599 ir_node *block = get_nodes_block(n);
2600 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2601 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2604 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2605 dbg_info *dbgi = get_irn_dbg_info(n);
2606 ir_node *block = get_nodes_block(n);
2607 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2608 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2610 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2611 ir_node *sub_l = get_Sub_left(b);
2612 ir_node *sub_r = get_Sub_right(b);
2613 dbg_info *dbgi = get_irn_dbg_info(n);
2614 ir_node *block = get_nodes_block(n);
2615 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2616 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2617 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2620 } else if (is_Minus(b)) {
2621 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2622 ir_node *sub_l = get_Sub_left(a);
2623 ir_node *sub_r = get_Sub_right(a);
2624 dbg_info *dbgi = get_irn_dbg_info(n);
2625 ir_node *block = get_nodes_block(n);
2626 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2627 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2628 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2631 } else if (is_Shl(a)) {
2632 ir_node *const shl_l = get_Shl_left(a);
2633 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2634 /* (1 << x) * b -> b << x */
2635 dbg_info *const dbgi = get_irn_dbg_info(n);
2636 ir_node *const block = get_nodes_block(n);
2637 ir_node *const shl_r = get_Shl_right(a);
2638 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2639 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2642 } else if (is_Shl(b)) {
2643 ir_node *const shl_l = get_Shl_left(b);
2644 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2645 /* a * (1 << x) -> a << x */
2646 dbg_info *const dbgi = get_irn_dbg_info(n);
2647 ir_node *const block = get_nodes_block(n);
2648 ir_node *const shl_r = get_Shl_right(b);
2649 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2650 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2654 if (get_mode_arithmetic(mode) == irma_ieee754) {
2656 ir_tarval *tv = get_Const_tarval(a);
2657 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2658 && !tarval_is_negative(tv)) {
2659 /* 2.0 * b = b + b */
2660 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2661 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2665 else if (is_Const(b)) {
2666 ir_tarval *tv = get_Const_tarval(b);
2667 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2668 && !tarval_is_negative(tv)) {
2669 /* a * 2.0 = a + a */
2670 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2671 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2676 return arch_dep_replace_mul_with_shifts(n);
2677 } /* transform_node_Mul */
2680 * Transform a Div Node.
2682 static ir_node *transform_node_Div(ir_node *n)
2684 ir_mode *mode = get_Div_resmode(n);
2685 ir_node *a = get_Div_left(n);
2686 ir_node *b = get_Div_right(n);
2688 const ir_node *dummy;
2690 if (mode_is_int(mode)) {
2691 if (is_Const(b) && is_const_Phi(a)) {
2692 /* check for Div(Phi, Const) */
2693 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2695 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2698 } else if (is_Const(a) && is_const_Phi(b)) {
2699 /* check for Div(Const, Phi) */
2700 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2702 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2705 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2706 /* check for Div(Phi, Phi) */
2707 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2709 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2714 if (a == b && value_not_zero(a, &dummy)) {
2715 ir_graph *irg = get_irn_irg(n);
2716 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2717 value = new_r_Const(irg, get_mode_one(mode));
2718 DBG_OPT_CSTEVAL(n, value);
2721 if (mode_is_signed(mode) && is_Const(b)) {
2722 ir_tarval *tv = get_Const_tarval(b);
2724 if (tv == get_mode_minus_one(mode)) {
2726 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2727 DBG_OPT_CSTEVAL(n, value);
2731 /* Try architecture dependent optimization */
2732 value = arch_dep_replace_div_by_const(n);
2735 assert(mode_is_float(mode));
2737 /* Optimize x/c to x*(1/c) */
2738 if (get_mode_arithmetic(mode) == irma_ieee754) {
2739 ir_tarval *tv = value_of(b);
2741 if (tv != tarval_bad) {
2742 int rem = tarval_fp_ops_enabled();
2745 * Floating point constant folding might be disabled here to
2747 * However, as we check for exact result, doing it is safe.
2750 tarval_enable_fp_ops(1);
2751 tv = tarval_div(get_mode_one(mode), tv);
2752 tarval_enable_fp_ops(rem);
2754 /* Do the transformation if the result is either exact or we are
2755 not using strict rules. */
2756 if (tv != tarval_bad &&
2757 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2758 ir_node *block = get_nodes_block(n);
2759 ir_graph *irg = get_irn_irg(block);
2760 ir_node *c = new_r_Const(irg, tv);
2761 dbg_info *dbgi = get_irn_dbg_info(n);
2762 value = new_rd_Mul(dbgi, block, a, c, mode);
2775 /* Turn Div into a tuple (mem, jmp, bad, value) */
2776 mem = get_Div_mem(n);
2777 blk = get_nodes_block(n);
2778 irg = get_irn_irg(blk);
2780 /* skip a potential Pin */
2781 mem = skip_Pin(mem);
2782 turn_into_tuple(n, pn_Div_max);
2783 set_Tuple_pred(n, pn_Div_M, mem);
2784 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2785 set_Tuple_pred(n, pn_Div_X_except, get_irg_bad(irg));
2786 set_Tuple_pred(n, pn_Div_res, value);
2789 } /* transform_node_Div */
2792 * Transform a Mod node.
2794 static ir_node *transform_node_Mod(ir_node *n)
2796 ir_mode *mode = get_Mod_resmode(n);
2797 ir_node *a = get_Mod_left(n);
2798 ir_node *b = get_Mod_right(n);
2803 if (is_Const(b) && is_const_Phi(a)) {
2804 /* check for Div(Phi, Const) */
2805 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2807 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2811 else if (is_Const(a) && is_const_Phi(b)) {
2812 /* check for Div(Const, Phi) */
2813 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2815 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2819 else if (is_const_Phi(a) && is_const_Phi(b)) {
2820 /* check for Div(Phi, Phi) */
2821 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2823 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2830 irg = get_irn_irg(n);
2831 if (tv != tarval_bad) {
2832 value = new_r_Const(irg, tv);
2834 DBG_OPT_CSTEVAL(n, value);
2837 ir_node *a = get_Mod_left(n);
2838 ir_node *b = get_Mod_right(n);
2839 const ir_node *dummy;
2841 if (a == b && value_not_zero(a, &dummy)) {
2842 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2843 value = new_r_Const(irg, get_mode_null(mode));
2844 DBG_OPT_CSTEVAL(n, value);
2847 if (mode_is_signed(mode) && is_Const(b)) {
2848 ir_tarval *tv = get_Const_tarval(b);
2850 if (tv == get_mode_minus_one(mode)) {
2852 value = new_r_Const(irg, get_mode_null(mode));
2853 DBG_OPT_CSTEVAL(n, value);
2857 /* Try architecture dependent optimization */
2858 value = arch_dep_replace_mod_by_const(n);
2867 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2868 mem = get_Mod_mem(n);
2869 blk = get_nodes_block(n);
2870 irg = get_irn_irg(blk);
2872 /* skip a potential Pin */
2873 mem = skip_Pin(mem);
2874 turn_into_tuple(n, pn_Mod_max);
2875 set_Tuple_pred(n, pn_Mod_M, mem);
2876 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2877 set_Tuple_pred(n, pn_Mod_X_except, get_irg_bad(irg));
2878 set_Tuple_pred(n, pn_Mod_res, value);
2881 } /* transform_node_Mod */
2884 * Transform a Cond node.
2886 * Replace the Cond by a Jmp if it branches on a constant
2889 static ir_node *transform_node_Cond(ir_node *n)
2892 ir_node *a = get_Cond_selector(n);
2893 ir_tarval *ta = value_of(a);
2894 ir_graph *irg = get_irn_irg(n);
2897 /* we need block info which is not available in floating irgs */
2898 if (get_irg_pinned(irg) == op_pin_state_floats)
2901 if ((ta != tarval_bad) &&
2902 (get_irn_mode(a) == mode_b) &&
2903 (get_opt_unreachable_code())) {
2904 /* It's a boolean Cond, branching on a boolean constant.
2905 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2906 ir_node *blk = get_nodes_block(n);
2907 jmp = new_r_Jmp(blk);
2908 turn_into_tuple(n, pn_Cond_max);
2909 if (ta == tarval_b_true) {
2910 set_Tuple_pred(n, pn_Cond_false, get_irg_bad(irg));
2911 set_Tuple_pred(n, pn_Cond_true, jmp);
2913 set_Tuple_pred(n, pn_Cond_false, jmp);
2914 set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg));
2916 /* We might generate an endless loop, so keep it alive. */
2917 add_End_keepalive(get_irg_end(irg), blk);
2920 } /* transform_node_Cond */
2923 * Prototype of a recursive transform function
2924 * for bitwise distributive transformations.
2926 typedef ir_node* (*recursive_transform)(ir_node *n);
2929 * makes use of distributive laws for and, or, eor
2930 * and(a OP c, b OP c) -> and(a, b) OP c
2931 * note, might return a different op than n
2933 static ir_node *transform_bitwise_distributive(ir_node *n,
2934 recursive_transform trans_func)
2937 ir_node *a = get_binop_left(n);
2938 ir_node *b = get_binop_right(n);
2939 ir_op *op = get_irn_op(a);
2940 ir_op *op_root = get_irn_op(n);
2942 if (op != get_irn_op(b))
2945 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2946 if (op == op_Conv) {
2947 ir_node *a_op = get_Conv_op(a);
2948 ir_node *b_op = get_Conv_op(b);
2949 ir_mode *a_mode = get_irn_mode(a_op);
2950 ir_mode *b_mode = get_irn_mode(b_op);
2951 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2952 ir_node *blk = get_nodes_block(n);
2955 set_binop_left(n, a_op);
2956 set_binop_right(n, b_op);
2957 set_irn_mode(n, a_mode);
2959 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2961 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2967 /* nothing to gain here */
2971 if (op == op_Shrs || op == op_Shr || op == op_Shl
2972 || op == op_And || op == op_Or || op == op_Eor) {
2973 ir_node *a_left = get_binop_left(a);
2974 ir_node *a_right = get_binop_right(a);
2975 ir_node *b_left = get_binop_left(b);
2976 ir_node *b_right = get_binop_right(b);
2978 ir_node *op1 = NULL;
2979 ir_node *op2 = NULL;
2981 if (is_op_commutative(op)) {
2982 if (a_left == b_left) {
2986 } else if (a_left == b_right) {
2990 } else if (a_right == b_left) {
2996 if (a_right == b_right) {
3003 /* (a sop c) & (b sop c) => (a & b) sop c */
3004 ir_node *blk = get_nodes_block(n);
3006 ir_node *new_n = exact_copy(n);
3007 set_binop_left(new_n, op1);
3008 set_binop_right(new_n, op2);
3009 new_n = trans_func(new_n);
3011 if (op_root == op_Eor && op == op_Or) {
3012 dbg_info *dbgi = get_irn_dbg_info(n);
3013 ir_mode *mode = get_irn_mode(c);
3015 c = new_rd_Not(dbgi, blk, c, mode);
3016 n = new_rd_And(dbgi, blk, new_n, c, mode);
3019 set_nodes_block(n, blk);
3020 set_binop_left(n, new_n);
3021 set_binop_right(n, c);
3025 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3034 * Create a 0 constant of given mode.
3036 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
3038 ir_tarval *tv = get_mode_null(mode);
3039 ir_node *cnst = new_r_Const(irg, tv);
3047 static ir_node *transform_node_And(ir_node *n)
3049 ir_node *c, *oldn = n;
3050 ir_node *a = get_And_left(n);
3051 ir_node *b = get_And_right(n);
3053 vrp_attr *a_vrp, *b_vrp;
3055 if (is_Cmp(a) && is_Cmp(b)) {
3056 ir_node *a_left = get_Cmp_left(a);
3057 ir_node *a_right = get_Cmp_right(a);
3058 ir_node *b_left = get_Cmp_left(b);
3059 ir_node *b_right = get_Cmp_right(b);
3060 ir_relation a_relation = get_Cmp_relation(a);
3061 ir_relation b_relation = get_Cmp_relation(b);
3062 /* we can combine the relations of two compares with the same
3064 if (a_left == b_left && b_left == b_right) {
3065 dbg_info *dbgi = get_irn_dbg_info(n);
3066 ir_node *block = get_nodes_block(n);
3067 ir_relation new_relation = a_relation & b_relation;
3068 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3070 /* Cmp(a==0) and Cmp(b==0) can be optimized to Cmp(a|b==0) */
3071 if (is_Const(a_right) && is_Const_null(a_right)
3072 && is_Const(b_right) && is_Const_null(b_right)
3073 && a_relation == b_relation && a_relation == ir_relation_equal
3074 && !mode_is_float(get_irn_mode(a_left))
3075 && !mode_is_float(get_irn_mode(b_left))) {
3076 dbg_info *dbgi = get_irn_dbg_info(n);
3077 ir_node *block = get_nodes_block(n);
3078 ir_mode *mode = get_irn_mode(a_left);
3079 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
3080 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
3081 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
3082 ir_graph *irg = get_irn_irg(n);
3083 ir_node *zero = create_zero_const(irg, mode);
3084 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3088 mode = get_irn_mode(n);
3089 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3093 ir_node *op = get_Not_op(b);
3095 ir_node *ba = get_And_left(op);
3096 ir_node *bb = get_And_right(op);
3098 /* it's enough to test the following cases due to normalization! */
3099 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3100 /* (a|b) & ~(a&b) = a^b */
3101 ir_node *block = get_nodes_block(n);
3103 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3104 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3112 ir_node *op = get_Not_op(a);
3114 ir_node *aa = get_And_left(op);
3115 ir_node *ab = get_And_right(op);
3117 /* it's enough to test the following cases due to normalization! */
3118 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3119 /* (a|b) & ~(a&b) = a^b */
3120 ir_node *block = get_nodes_block(n);
3122 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3123 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3130 ir_node *al = get_Eor_left(a);
3131 ir_node *ar = get_Eor_right(a);
3134 /* (b ^ a) & b -> ~a & b */
3135 dbg_info *dbg = get_irn_dbg_info(n);
3136 ir_node *block = get_nodes_block(n);
3138 ar = new_rd_Not(dbg, block, ar, mode);
3139 n = new_rd_And(dbg, block, ar, b, mode);
3140 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3144 /* (a ^ b) & b -> ~a & b */
3145 dbg_info *dbg = get_irn_dbg_info(n);
3146 ir_node *block = get_nodes_block(n);
3148 al = new_rd_Not(dbg, block, al, mode);
3149 n = new_rd_And(dbg, block, al, b, mode);
3150 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3155 ir_node *bl = get_Eor_left(b);
3156 ir_node *br = get_Eor_right(b);
3159 /* a & (a ^ b) -> a & ~b */
3160 dbg_info *dbg = get_irn_dbg_info(n);
3161 ir_node *block = get_nodes_block(n);
3163 br = new_rd_Not(dbg, block, br, mode);
3164 n = new_rd_And(dbg, block, br, a, mode);
3165 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3169 /* a & (b ^ a) -> a & ~b */
3170 dbg_info *dbg = get_irn_dbg_info(n);
3171 ir_node *block = get_nodes_block(n);
3173 bl = new_rd_Not(dbg, block, bl, mode);
3174 n = new_rd_And(dbg, block, bl, a, mode);
3175 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3179 if (is_Not(a) && is_Not(b)) {
3180 /* ~a & ~b = ~(a|b) */
3181 ir_node *block = get_nodes_block(n);
3182 ir_mode *mode = get_irn_mode(n);
3186 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3187 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3188 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3192 b_vrp = vrp_get_info(b);
3193 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3194 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3200 a_vrp = vrp_get_info(a);
3201 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3202 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3206 n = transform_bitwise_distributive(n, transform_node_And);
3209 } /* transform_node_And */
3211 /* the order of the values is important! */
3212 typedef enum const_class {
3218 static const_class classify_const(const ir_node* n)
3220 if (is_Const(n)) return const_const;
3221 if (is_irn_constlike(n)) return const_like;
3226 * Determines whether r is more constlike or has a larger index (in that order)
3229 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3231 const const_class l_order = classify_const(l);
3232 const const_class r_order = classify_const(r);
3234 l_order > r_order ||
3235 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3241 static ir_node *transform_node_Eor(ir_node *n)
3243 ir_node *c, *oldn = n;
3244 ir_node *a = get_Eor_left(n);
3245 ir_node *b = get_Eor_right(n);
3246 ir_mode *mode = get_irn_mode(n);
3248 /* we can combine the relations of two compares with the same operands */
3249 if (is_Cmp(a) && is_Cmp(b)) {
3250 ir_node *a_left = get_Cmp_left(a);
3251 ir_node *a_right = get_Cmp_left(a);
3252 ir_node *b_left = get_Cmp_left(b);
3253 ir_node *b_right = get_Cmp_right(b);
3254 if (a_left == b_left && b_left == b_right) {
3255 dbg_info *dbgi = get_irn_dbg_info(n);
3256 ir_node *block = get_nodes_block(n);
3257 ir_relation a_relation = get_Cmp_relation(a);
3258 ir_relation b_relation = get_Cmp_relation(b);
3259 ir_relation new_relation = a_relation ^ b_relation;
3260 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3264 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3266 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3267 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3268 dbg_info *dbg = get_irn_dbg_info(n);
3269 ir_node *block = get_nodes_block(n);
3270 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3271 ir_node *new_left = get_Not_op(a);
3272 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3273 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3275 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3276 dbg_info *dbg = get_irn_dbg_info(n);
3277 ir_node *block = get_nodes_block(n);
3278 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3279 ir_node *new_right = get_Not_op(b);
3280 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3281 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3285 /* x ^ 1...1 -> ~1 */
3286 if (is_Const(b) && is_Const_all_one(b)) {
3287 n = new_r_Not(get_nodes_block(n), a, mode);
3288 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3292 n = transform_bitwise_distributive(n, transform_node_Eor);
3294 } /* transform_node_Eor */
3299 static ir_node *transform_node_Not(ir_node *n)
3301 ir_node *c, *oldn = n;
3302 ir_node *a = get_Not_op(n);
3303 ir_mode *mode = get_irn_mode(n);
3305 HANDLE_UNOP_PHI(tarval_not,a,c);
3307 /* check for a boolean Not */
3309 dbg_info *dbgi = get_irn_dbg_info(a);
3310 ir_node *block = get_nodes_block(a);
3311 ir_relation relation = get_Cmp_relation(a);
3312 relation = get_negated_relation(relation);
3313 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3314 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3318 /* normalize ~(a ^ b) => a ^ ~b */
3320 dbg_info *dbg = get_irn_dbg_info(n);
3321 ir_node *block = get_nodes_block(n);
3322 ir_node *eor_right = get_Eor_right(a);
3323 ir_node *eor_left = get_Eor_left(a);
3324 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3325 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3329 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3330 if (is_Minus(a)) { /* ~-x -> x + -1 */
3331 dbg_info *dbg = get_irn_dbg_info(n);
3332 ir_graph *irg = get_irn_irg(n);
3333 ir_node *block = get_nodes_block(n);
3334 ir_node *add_l = get_Minus_op(a);
3335 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3336 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3337 } else if (is_Add(a)) {
3338 ir_node *add_r = get_Add_right(a);
3339 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3340 /* ~(x + -1) = -x */
3341 ir_node *op = get_Add_left(a);
3342 ir_node *blk = get_nodes_block(n);
3343 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3344 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3352 * Transform a Minus.
3356 * -(a >>u (size-1)) = a >>s (size-1)
3357 * -(a >>s (size-1)) = a >>u (size-1)
3358 * -(a * const) -> a * -const
3360 static ir_node *transform_node_Minus(ir_node *n)
3362 ir_node *c, *oldn = n;
3363 ir_node *a = get_Minus_op(n);
3366 HANDLE_UNOP_PHI(tarval_neg,a,c);
3368 mode = get_irn_mode(a);
3369 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3370 /* the following rules are only to twos-complement */
3373 ir_node *op = get_Not_op(a);
3374 ir_tarval *tv = get_mode_one(mode);
3375 ir_node *blk = get_nodes_block(n);
3376 ir_graph *irg = get_irn_irg(blk);
3377 ir_node *c = new_r_Const(irg, tv);
3378 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3379 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3383 ir_node *c = get_Shr_right(a);
3386 ir_tarval *tv = get_Const_tarval(c);
3388 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3389 /* -(a >>u (size-1)) = a >>s (size-1) */
3390 ir_node *v = get_Shr_left(a);
3392 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3393 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3399 ir_node *c = get_Shrs_right(a);
3402 ir_tarval *tv = get_Const_tarval(c);
3404 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3405 /* -(a >>s (size-1)) = a >>u (size-1) */
3406 ir_node *v = get_Shrs_left(a);
3408 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3409 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3416 /* - (a-b) = b - a */
3417 ir_node *la = get_Sub_left(a);
3418 ir_node *ra = get_Sub_right(a);
3419 ir_node *blk = get_nodes_block(n);
3421 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3422 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3426 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3427 ir_node *mul_l = get_Mul_left(a);
3428 ir_node *mul_r = get_Mul_right(a);
3429 ir_tarval *tv = value_of(mul_r);
3430 if (tv != tarval_bad) {
3431 tv = tarval_neg(tv);
3432 if (tv != tarval_bad) {
3433 ir_graph *irg = get_irn_irg(n);
3434 ir_node *cnst = new_r_Const(irg, tv);
3435 dbg_info *dbg = get_irn_dbg_info(a);
3436 ir_node *block = get_nodes_block(a);
3437 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3438 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3445 } /* transform_node_Minus */
3448 * Transform a Proj(Load) with a non-null address.
3450 static ir_node *transform_node_Proj_Load(ir_node *proj)
3452 if (get_opt_ldst_only_null_ptr_exceptions()) {
3453 if (get_irn_mode(proj) == mode_X) {
3454 ir_node *load = get_Proj_pred(proj);
3456 /* get the Load address */
3457 const ir_node *addr = get_Load_ptr(load);
3458 const ir_node *confirm;
3460 if (value_not_null(addr, &confirm)) {
3461 if (confirm == NULL) {
3462 /* this node may float if it did not depend on a Confirm */
3463 set_irn_pinned(load, op_pin_state_floats);
3465 if (get_Proj_proj(proj) == pn_Load_X_except) {
3466 ir_graph *irg = get_irn_irg(proj);
3467 DBG_OPT_EXC_REM(proj);
3468 return get_irg_bad(irg);
3470 ir_node *blk = get_nodes_block(load);
3471 return new_r_Jmp(blk);
3477 } /* transform_node_Proj_Load */
3480 * Transform a Proj(Store) with a non-null address.
3482 static ir_node *transform_node_Proj_Store(ir_node *proj)
3484 if (get_opt_ldst_only_null_ptr_exceptions()) {
3485 if (get_irn_mode(proj) == mode_X) {
3486 ir_node *store = get_Proj_pred(proj);
3488 /* get the load/store address */
3489 const ir_node *addr = get_Store_ptr(store);
3490 const ir_node *confirm;
3492 if (value_not_null(addr, &confirm)) {
3493 if (confirm == NULL) {
3494 /* this node may float if it did not depend on a Confirm */
3495 set_irn_pinned(store, op_pin_state_floats);
3497 if (get_Proj_proj(proj) == pn_Store_X_except) {
3498 ir_graph *irg = get_irn_irg(proj);
3499 DBG_OPT_EXC_REM(proj);
3500 return get_irg_bad(irg);
3502 ir_node *blk = get_nodes_block(store);
3503 return new_r_Jmp(blk);
3509 } /* transform_node_Proj_Store */
3512 * Transform a Proj(Div) with a non-zero value.
3513 * Removes the exceptions and routes the memory to the NoMem node.
3515 static ir_node *transform_node_Proj_Div(ir_node *proj)
3517 ir_node *div = get_Proj_pred(proj);
3518 ir_node *b = get_Div_right(div);
3519 ir_node *res, *new_mem;
3520 const ir_node *confirm;
3523 if (value_not_zero(b, &confirm)) {
3524 /* div(x, y) && y != 0 */
3525 if (confirm == NULL) {
3526 /* we are sure we have a Const != 0 */
3527 new_mem = get_Div_mem(div);
3528 new_mem = skip_Pin(new_mem);
3529 set_Div_mem(div, new_mem);
3530 set_irn_pinned(div, op_pin_state_floats);
3533 proj_nr = get_Proj_proj(proj);
3535 case pn_Div_X_regular:
3536 return new_r_Jmp(get_nodes_block(div));
3538 case pn_Div_X_except: {
3539 ir_graph *irg = get_irn_irg(proj);
3540 /* we found an exception handler, remove it */
3541 DBG_OPT_EXC_REM(proj);
3542 return get_irg_bad(irg);
3546 ir_graph *irg = get_irn_irg(proj);
3547 res = get_Div_mem(div);
3548 new_mem = get_irg_no_mem(irg);
3551 /* This node can only float up to the Confirm block */
3552 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3554 set_irn_pinned(div, op_pin_state_floats);
3555 /* this is a Div without exception, we can remove the memory edge */
3556 set_Div_mem(div, new_mem);
3562 } /* transform_node_Proj_Div */
3565 * Transform a Proj(Mod) with a non-zero value.
3566 * Removes the exceptions and routes the memory to the NoMem node.
3568 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3570 ir_node *mod = get_Proj_pred(proj);
3571 ir_node *b = get_Mod_right(mod);
3572 ir_node *res, *new_mem;
3573 const ir_node *confirm;
3576 if (value_not_zero(b, &confirm)) {
3577 /* mod(x, y) && y != 0 */
3578 proj_nr = get_Proj_proj(proj);
3580 if (confirm == NULL) {
3581 /* we are sure we have a Const != 0 */
3582 new_mem = get_Mod_mem(mod);
3583 new_mem = skip_Pin(new_mem);
3584 set_Mod_mem(mod, new_mem);
3585 set_irn_pinned(mod, op_pin_state_floats);
3590 case pn_Mod_X_regular:
3591 return new_r_Jmp(get_irn_n(mod, -1));
3593 case pn_Mod_X_except: {
3594 ir_graph *irg = get_irn_irg(proj);
3595 /* we found an exception handler, remove it */
3596 DBG_OPT_EXC_REM(proj);
3597 return get_irg_bad(irg);
3601 ir_graph *irg = get_irn_irg(proj);
3602 res = get_Mod_mem(mod);
3603 new_mem = get_irg_no_mem(irg);
3606 /* This node can only float up to the Confirm block */
3607 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3609 /* this is a Mod without exception, we can remove the memory edge */
3610 set_Mod_mem(mod, new_mem);
3614 if (get_Mod_left(mod) == b) {
3615 /* a % a = 0 if a != 0 */
3616 ir_graph *irg = get_irn_irg(proj);
3617 ir_mode *mode = get_irn_mode(proj);
3618 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3620 DBG_OPT_CSTEVAL(mod, res);
3626 } /* transform_node_Proj_Mod */
3629 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3631 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3633 ir_node *n = get_Proj_pred(proj);
3634 ir_node *b = get_Cond_selector(n);
3636 if (!get_opt_unreachable_code())
3639 if (mode_is_int(get_irn_mode(b))) {
3640 ir_tarval *tb = value_of(b);
3642 if (tb != tarval_bad) {
3643 /* we have a constant switch */
3644 long num = get_Proj_proj(proj);
3646 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3647 if (get_tarval_long(tb) == num) {
3648 /* Do NOT create a jump here, or we will have 2 control flow ops
3649 * in a block. This case is optimized away in optimize_cf(). */
3652 ir_graph *irg = get_irn_irg(proj);
3653 /* this case will NEVER be taken, kill it */
3654 return get_irg_bad(irg);
3658 long num = get_Proj_proj(proj);
3659 vrp_attr *b_vrp = vrp_get_info(b);
3660 if (num != get_Cond_default_proj(n) && b_vrp) {
3661 /* Try handling with vrp data. We only remove dead parts. */
3662 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3664 if (b_vrp->range_type == VRP_RANGE) {
3665 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3666 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3668 if ((cmp_result & ir_relation_greater) == cmp_result && (cmp_result2
3669 & ir_relation_less) == cmp_result2) {
3670 ir_graph *irg = get_irn_irg(proj);
3671 return get_irg_bad(irg);
3673 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3674 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3675 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3677 if ((cmp_result & ir_relation_less_equal) == cmp_result && (cmp_result2
3678 & ir_relation_greater_equal) == cmp_result2) {
3679 ir_graph *irg = get_irn_irg(proj);
3680 return get_irg_bad(irg);
3685 tarval_and( b_vrp->bits_set, tp),
3687 ) == ir_relation_equal)) {
3688 ir_graph *irg = get_irn_irg(proj);
3689 return get_irg_bad(irg);
3695 tarval_not(b_vrp->bits_not_set)),
3696 tarval_not(b_vrp->bits_not_set))
3697 == ir_relation_equal)) {
3698 ir_graph *irg = get_irn_irg(proj);
3699 return get_irg_bad(irg);
3710 * return true if the operation returns a value with exactly 1 bit set
3712 static bool is_single_bit(const ir_node *node)
3714 /* a first implementation, could be extended with vrp and others... */
3716 ir_node *shl_l = get_Shl_left(node);
3717 ir_mode *mode = get_irn_mode(node);
3718 int modulo = get_mode_modulo_shift(mode);
3719 /* this works if we shift a 1 and we have modulo shift */
3720 if (is_Const(shl_l) && is_Const_one(shl_l)
3721 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3724 } else if (is_Const(node)) {
3725 ir_tarval *tv = get_Const_tarval(node);
3726 return tarval_is_single_bit(tv);
3732 * Normalizes and optimizes Cmp nodes.
3734 static ir_node *transform_node_Cmp(ir_node *n)
3736 ir_node *left = get_Cmp_left(n);
3737 ir_node *right = get_Cmp_right(n);
3738 ir_mode *mode = get_irn_mode(left);
3739 ir_tarval *tv = NULL;
3740 bool changed = false;
3741 bool changedc = false;
3742 ir_relation relation = get_Cmp_relation(n);
3743 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3745 /* mask out impossible relations */
3746 ir_relation new_relation = relation & possible;
3747 if (new_relation != relation) {
3748 relation = new_relation;
3752 /* Remove unnecessary conversions */
3753 /* TODO handle conv+constant */
3754 if (is_Conv(left) && is_Conv(right)) {
3755 ir_node *op_left = get_Conv_op(left);
3756 ir_node *op_right = get_Conv_op(right);
3757 ir_mode *mode_left = get_irn_mode(op_left);
3758 ir_mode *mode_right = get_irn_mode(op_right);
3760 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3761 && mode_left != mode_b && mode_right != mode_b) {
3762 ir_node *block = get_nodes_block(n);
3764 if (mode_left == mode_right) {
3768 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3769 } else if (smaller_mode(mode_left, mode_right)) {
3770 left = new_r_Conv(block, op_left, mode_right);
3773 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3774 } else if (smaller_mode(mode_right, mode_left)) {
3776 right = new_r_Conv(block, op_right, mode_left);
3778 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3784 * Optimize -a CMP -b into b CMP a.
3785 * This works only for modes where unary Minus cannot Overflow.
3786 * Note that two-complement integers can Overflow so it will NOT work.
3788 if (!mode_overflow_on_unary_Minus(mode) &&
3789 is_Minus(left) && is_Minus(right)) {
3790 left = get_Minus_op(left);
3791 right = get_Minus_op(right);
3792 relation = get_inversed_relation(relation);
3794 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3797 /* remove operation on both sides if possible */
3798 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3800 * The following operations are NOT safe for floating point operations, for instance
3801 * 1.0 + inf == 2.0 + inf, =/=> x == y
3803 if (mode_is_int(mode)) {
3804 unsigned lop = get_irn_opcode(left);
3806 if (lop == get_irn_opcode(right)) {
3807 ir_node *ll, *lr, *rl, *rr;
3809 /* same operation on both sides, try to remove */
3813 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3814 left = get_unop_op(left);
3815 right = get_unop_op(right);
3817 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3820 ll = get_Add_left(left);
3821 lr = get_Add_right(left);
3822 rl = get_Add_left(right);
3823 rr = get_Add_right(right);
3826 /* X + a CMP X + b ==> a CMP b */
3830 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3831 } else if (ll == rr) {
3832 /* X + a CMP b + X ==> a CMP b */
3836 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3837 } else if (lr == rl) {
3838 /* a + X CMP X + b ==> a CMP b */
3842 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3843 } else if (lr == rr) {
3844 /* a + X CMP b + X ==> a CMP b */
3848 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3852 ll = get_Sub_left(left);
3853 lr = get_Sub_right(left);
3854 rl = get_Sub_left(right);
3855 rr = get_Sub_right(right);
3858 /* X - a CMP X - b ==> a CMP b */
3862 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3863 } else if (lr == rr) {
3864 /* a - X CMP b - X ==> a CMP b */
3868 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3872 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3873 /* a ROTL X CMP b ROTL X ==> a CMP b */
3874 left = get_Rotl_left(left);
3875 right = get_Rotl_left(right);
3877 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3885 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3886 if (is_Add(left) || is_Sub(left)) {
3887 ir_node *ll = get_binop_left(left);
3888 ir_node *lr = get_binop_right(left);
3890 if (lr == right && is_Add(left)) {
3896 ir_graph *irg = get_irn_irg(n);
3898 right = create_zero_const(irg, mode);
3900 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3903 if (is_Add(right) || is_Sub(right)) {
3904 ir_node *rl = get_binop_left(right);
3905 ir_node *rr = get_binop_right(right);
3907 if (rr == left && is_Add(right)) {
3913 ir_graph *irg = get_irn_irg(n);
3915 right = create_zero_const(irg, mode);
3917 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3921 if (is_And(left) && is_Const(right)) {
3922 ir_node *ll = get_binop_left(left);
3923 ir_node *lr = get_binop_right(left);
3924 if (is_Shr(ll) && is_Const(lr)) {
3925 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
3926 ir_node *block = get_nodes_block(n);
3927 ir_mode *mode = get_irn_mode(left);
3929 ir_node *llr = get_Shr_right(ll);
3930 if (is_Const(llr)) {
3931 dbg_info *dbg = get_irn_dbg_info(left);
3932 ir_graph *irg = get_irn_irg(left);
3934 ir_tarval *c1 = get_Const_tarval(llr);
3935 ir_tarval *c2 = get_Const_tarval(lr);
3936 ir_tarval *c3 = get_Const_tarval(right);
3937 ir_tarval *mask = tarval_shl(c2, c1);
3938 ir_tarval *value = tarval_shl(c3, c1);
3940 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
3941 right = new_r_Const(irg, value);
3946 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
3948 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
3949 right = get_Eor_right(left);
3950 left = get_Eor_left(left);
3953 } /* mode_is_int(...) */
3956 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
3957 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
3958 if (mode_is_int(mode) && is_And(left)
3959 && (relation == ir_relation_equal
3960 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
3961 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
3962 ir_node *and0 = get_And_left(left);
3963 ir_node *and1 = get_And_right(left);
3964 if (and1 == right) {
3965 ir_node *tmp = and0;
3969 if (and0 == right && is_single_bit(and0)) {
3970 ir_graph *irg = get_irn_irg(n);
3972 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
3973 right = create_zero_const(irg, mode);
3978 /* replace mode_b compares with ands/ors */
3979 if (mode == mode_b) {
3980 ir_node *block = get_nodes_block(n);
3984 case ir_relation_less_equal:
3985 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
3987 case ir_relation_less:
3988 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
3990 case ir_relation_greater_equal:
3991 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
3993 case ir_relation_greater:
3994 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
3996 case ir_relation_less_greater:
3997 bres = new_r_Eor(block, left, right, mode_b);
3999 case ir_relation_equal:
4000 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
4003 #ifdef DEBUG_libfirm
4004 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4009 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4015 * First step: normalize the compare op
4016 * by placing the constant on the right side
4017 * or moving the lower address node to the left.
4019 if (!operands_are_normalized(left, right)) {
4024 relation = get_inversed_relation(relation);
4029 * Second step: Try to reduce the magnitude
4030 * of a constant. This may help to generate better code
4031 * later and may help to normalize more compares.
4032 * Of course this is only possible for integer values.
4034 tv = value_of(right);
4035 if (tv != tarval_bad) {
4036 ir_mode *mode = get_irn_mode(right);
4038 /* TODO extend to arbitrary constants */
4039 if (is_Conv(left) && tarval_is_null(tv)) {
4040 ir_node *op = get_Conv_op(left);
4041 ir_mode *op_mode = get_irn_mode(op);
4044 * UpConv(x) REL 0 ==> x REL 0
4045 * Don't do this for float values as it's unclear whether it is a
4046 * win. (on the other side it makes detection/creation of fabs hard)
4048 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4049 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4050 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4051 !mode_is_float(mode)) {
4052 tv = get_mode_null(op_mode);
4056 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4060 if (tv != tarval_bad) {
4061 /* the following optimization is possible on modes without Overflow
4062 * on Unary Minus or on == and !=:
4063 * -a CMP c ==> a swap(CMP) -c
4065 * Beware: for two-complement Overflow may occur, so only == and != can
4066 * be optimized, see this:
4067 * -MININT < 0 =/=> MININT > 0 !!!
4069 if (is_Minus(left) &&
4070 (!mode_overflow_on_unary_Minus(mode) ||
4071 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4072 tv = tarval_neg(tv);
4074 if (tv != tarval_bad) {
4075 left = get_Minus_op(left);
4076 relation = get_inversed_relation(relation);
4078 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4080 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4081 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4082 tv = tarval_not(tv);
4084 if (tv != tarval_bad) {
4085 left = get_Not_op(left);
4087 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4091 /* for integer modes, we have more */
4092 if (mode_is_int(mode)) {
4093 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4094 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4095 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4096 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4098 if (tv != tarval_bad) {
4099 relation ^= ir_relation_equal;
4101 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4104 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4105 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4106 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4107 tv = tarval_add(tv, get_mode_one(mode));
4109 if (tv != tarval_bad) {
4110 relation ^= ir_relation_equal;
4112 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4116 /* the following reassociations work only for == and != */
4117 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4119 #if 0 /* Might be not that good in general */
4120 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4121 if (tarval_is_null(tv) && is_Sub(left)) {
4122 right = get_Sub_right(left);
4123 left = get_Sub_left(left);
4125 tv = value_of(right);
4127 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4131 if (tv != tarval_bad) {
4132 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4134 ir_node *c1 = get_Sub_right(left);
4135 ir_tarval *tv2 = value_of(c1);
4137 if (tv2 != tarval_bad) {
4138 tv2 = tarval_add(tv, value_of(c1));
4140 if (tv2 != tarval_bad) {
4141 left = get_Sub_left(left);
4144 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4148 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4149 else if (is_Add(left)) {
4150 ir_node *a_l = get_Add_left(left);
4151 ir_node *a_r = get_Add_right(left);
4155 if (is_Const(a_l)) {
4157 tv2 = value_of(a_l);
4160 tv2 = value_of(a_r);
4163 if (tv2 != tarval_bad) {
4164 tv2 = tarval_sub(tv, tv2, NULL);
4166 if (tv2 != tarval_bad) {
4170 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4174 /* -a == c ==> a == -c, -a != c ==> a != -c */
4175 else if (is_Minus(left)) {
4176 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4178 if (tv2 != tarval_bad) {
4179 left = get_Minus_op(left);
4182 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4189 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4190 switch (get_irn_opcode(left)) {
4194 c1 = get_And_right(left);
4197 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4198 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4200 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4202 /* TODO: move to constant evaluation */
4203 ir_graph *irg = get_irn_irg(n);
4204 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4205 c1 = new_r_Const(irg, tv);
4206 DBG_OPT_CSTEVAL(n, c1);
4210 if (tarval_is_single_bit(tv)) {
4212 * optimization for AND:
4214 * And(x, C) == C ==> And(x, C) != 0
4215 * And(x, C) != C ==> And(X, C) == 0
4217 * if C is a single Bit constant.
4220 /* check for Constant's match. We have check hare the tarvals,
4221 because our const might be changed */
4222 if (get_Const_tarval(c1) == tv) {
4223 /* fine: do the transformation */
4224 tv = get_mode_null(get_tarval_mode(tv));
4225 relation ^= ir_relation_less_equal_greater;
4227 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4233 c1 = get_Or_right(left);
4234 if (is_Const(c1) && tarval_is_null(tv)) {
4236 * Or(x, C) == 0 && C != 0 ==> FALSE
4237 * Or(x, C) != 0 && C != 0 ==> TRUE
4239 if (! tarval_is_null(get_Const_tarval(c1))) {
4240 /* TODO: move to constant evaluation */
4241 ir_graph *irg = get_irn_irg(n);
4242 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4243 c1 = new_r_Const(irg, tv);
4244 DBG_OPT_CSTEVAL(n, c1);
4251 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4253 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4256 c1 = get_Shl_right(left);
4258 ir_graph *irg = get_irn_irg(c1);
4259 ir_tarval *tv1 = get_Const_tarval(c1);
4260 ir_mode *mode = get_irn_mode(left);
4261 ir_tarval *minus1 = get_mode_all_one(mode);
4262 ir_tarval *amask = tarval_shr(minus1, tv1);
4263 ir_tarval *cmask = tarval_shl(minus1, tv1);
4266 if (tarval_and(tv, cmask) != tv) {
4267 /* condition not met */
4268 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4269 c1 = new_r_Const(irg, tv);
4270 DBG_OPT_CSTEVAL(n, c1);
4273 sl = get_Shl_left(left);
4274 blk = get_nodes_block(n);
4275 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4276 tv = tarval_shr(tv, tv1);
4278 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4283 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4285 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4288 c1 = get_Shr_right(left);
4290 ir_graph *irg = get_irn_irg(c1);
4291 ir_tarval *tv1 = get_Const_tarval(c1);
4292 ir_mode *mode = get_irn_mode(left);
4293 ir_tarval *minus1 = get_mode_all_one(mode);
4294 ir_tarval *amask = tarval_shl(minus1, tv1);
4295 ir_tarval *cmask = tarval_shr(minus1, tv1);
4298 if (tarval_and(tv, cmask) != tv) {
4299 /* condition not met */
4300 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4301 c1 = new_r_Const(irg, tv);
4302 DBG_OPT_CSTEVAL(n, c1);
4305 sl = get_Shr_left(left);
4306 blk = get_nodes_block(n);
4307 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4308 tv = tarval_shl(tv, tv1);
4310 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4315 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4317 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4320 c1 = get_Shrs_right(left);
4322 ir_graph *irg = get_irn_irg(c1);
4323 ir_tarval *tv1 = get_Const_tarval(c1);
4324 ir_mode *mode = get_irn_mode(left);
4325 ir_tarval *minus1 = get_mode_all_one(mode);
4326 ir_tarval *amask = tarval_shl(minus1, tv1);
4327 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4330 cond = tarval_sub(cond, tv1, NULL);
4331 cond = tarval_shrs(tv, cond);
4333 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4334 /* condition not met */
4335 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4336 c1 = new_r_Const(irg, tv);
4337 DBG_OPT_CSTEVAL(n, c1);
4340 sl = get_Shrs_left(left);
4341 blk = get_nodes_block(n);
4342 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4343 tv = tarval_shl(tv, tv1);
4345 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4350 } /* tarval != bad */
4353 if (changedc) { /* need a new Const */
4354 ir_graph *irg = get_irn_irg(n);
4355 right = new_r_Const(irg, tv);
4359 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4360 ir_node *op = get_Proj_pred(left);
4362 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4363 ir_node *c = get_binop_right(op);
4366 ir_tarval *tv = get_Const_tarval(c);
4368 if (tarval_is_single_bit(tv)) {
4369 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4370 ir_node *v = get_binop_left(op);
4371 ir_node *blk = get_irn_n(op, -1);
4372 ir_graph *irg = get_irn_irg(op);
4373 ir_mode *mode = get_irn_mode(v);
4375 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4376 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4378 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4385 dbg_info *dbgi = get_irn_dbg_info(n);
4386 ir_node *block = get_nodes_block(n);
4388 /* create a new compare */
4389 n = new_rd_Cmp(dbgi, block, left, right, relation);
4396 * Optimize CopyB(mem, x, x) into a Nop.
4398 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4400 ir_node *copyb = get_Proj_pred(proj);
4401 ir_node *a = get_CopyB_dst(copyb);
4402 ir_node *b = get_CopyB_src(copyb);
4405 switch (get_Proj_proj(proj)) {
4406 case pn_CopyB_X_regular:
4407 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4408 DBG_OPT_EXC_REM(proj);
4409 proj = new_r_Jmp(get_nodes_block(copyb));
4411 case pn_CopyB_X_except:
4412 DBG_OPT_EXC_REM(proj);
4413 proj = get_irg_bad(get_irn_irg(proj));
4420 } /* transform_node_Proj_CopyB */
4423 * Optimize Bounds(idx, idx, upper) into idx.
4425 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4427 ir_node *oldn = proj;
4428 ir_node *bound = get_Proj_pred(proj);
4429 ir_node *idx = get_Bound_index(bound);
4430 ir_node *pred = skip_Proj(idx);
4433 if (idx == get_Bound_lower(bound))
4435 else if (is_Bound(pred)) {
4437 * idx was Bounds checked previously, it is still valid if
4438 * lower <= pred_lower && pred_upper <= upper.
4440 ir_node *lower = get_Bound_lower(bound);
4441 ir_node *upper = get_Bound_upper(bound);
4442 if (get_Bound_lower(pred) == lower &&
4443 get_Bound_upper(pred) == upper) {
4445 * One could expect that we simply return the previous
4446 * Bound here. However, this would be wrong, as we could
4447 * add an exception Proj to a new location then.
4448 * So, we must turn in into a tuple.
4454 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4455 switch (get_Proj_proj(proj)) {
4457 DBG_OPT_EXC_REM(proj);
4458 proj = get_Bound_mem(bound);
4460 case pn_Bound_X_except:
4461 DBG_OPT_EXC_REM(proj);
4462 proj = get_irg_bad(get_irn_irg(proj));
4466 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4468 case pn_Bound_X_regular:
4469 DBG_OPT_EXC_REM(proj);
4470 proj = new_r_Jmp(get_nodes_block(bound));
4477 } /* transform_node_Proj_Bound */
4480 * Does all optimizations on nodes that must be done on its Projs
4481 * because of creating new nodes.
4483 static ir_node *transform_node_Proj(ir_node *proj)
4485 ir_node *n = get_Proj_pred(proj);
4487 if (n->op->ops.transform_node_Proj)
4488 return n->op->ops.transform_node_Proj(proj);
4490 } /* transform_node_Proj */
4492 static ir_node *transform_node_Phi(ir_node *phi)
4494 int n = get_irn_arity(phi);
4495 ir_mode *mode = get_irn_mode(phi);
4496 ir_node *block = get_nodes_block(phi);
4497 ir_graph *irg = get_irn_irg(phi);
4498 ir_node *bad = get_irg_bad(irg);
4501 /* Set phi-operands for bad-block inputs to bad */
4502 for (i = 0; i < n; ++i) {
4503 ir_node *pred = get_Block_cfgpred(block, i);
4506 set_irn_n(phi, i, bad);
4509 /* Move Confirms down through Phi nodes. */
4510 if (mode_is_reference(mode)) {
4511 n = get_irn_arity(phi);
4513 /* Beware of Phi0 */
4515 ir_node *pred = get_irn_n(phi, 0);
4516 ir_node *bound, *new_phi, *block, **in;
4517 ir_relation relation;
4519 if (! is_Confirm(pred))
4522 bound = get_Confirm_bound(pred);
4523 relation = get_Confirm_relation(pred);
4525 NEW_ARR_A(ir_node *, in, n);
4526 in[0] = get_Confirm_value(pred);
4528 for (i = 1; i < n; ++i) {
4529 pred = get_irn_n(phi, i);
4531 if (! is_Confirm(pred) ||
4532 get_Confirm_bound(pred) != bound ||
4533 get_Confirm_relation(pred) != relation)
4535 in[i] = get_Confirm_value(pred);
4537 /* move the Confirm nodes "behind" the Phi */
4538 block = get_irn_n(phi, -1);
4539 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4540 return new_r_Confirm(block, new_phi, bound, relation);
4547 * Returns the operands of a commutative bin-op, if one operand is
4548 * a const, it is returned as the second one.
4550 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4552 ir_node *op_a = get_binop_left(binop);
4553 ir_node *op_b = get_binop_right(binop);
4555 assert(is_op_commutative(get_irn_op(binop)));
4557 if (is_Const(op_a)) {
4564 } /* get_comm_Binop_Ops */
4567 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4568 * Such pattern may arise in bitfield stores.
4570 * value c4 value c4 & c2
4571 * AND c3 AND c1 | c3
4578 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4581 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4583 ir_node *irn_and, *c1;
4585 ir_node *and_l, *c3;
4586 ir_node *value, *c4;
4587 ir_node *new_and, *new_const, *block;
4588 ir_mode *mode = get_irn_mode(irn_or);
4590 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4594 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4595 if (!is_Const(c1) || !is_And(irn_and))
4598 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4602 tv1 = get_Const_tarval(c1);
4603 tv2 = get_Const_tarval(c2);
4605 tv = tarval_or(tv1, tv2);
4606 if (tarval_is_all_one(tv)) {
4607 /* the AND does NOT clear a bit with isn't set by the OR */
4608 set_Or_left(irn_or, or_l);
4609 set_Or_right(irn_or, c1);
4611 /* check for more */
4618 get_comm_Binop_Ops(or_l, &and_l, &c3);
4619 if (!is_Const(c3) || !is_And(and_l))
4622 get_comm_Binop_Ops(and_l, &value, &c4);
4626 /* ok, found the pattern, check for conditions */
4627 assert(mode == get_irn_mode(irn_and));
4628 assert(mode == get_irn_mode(or_l));
4629 assert(mode == get_irn_mode(and_l));
4631 tv3 = get_Const_tarval(c3);
4632 tv4 = get_Const_tarval(c4);
4634 tv = tarval_or(tv4, tv2);
4635 if (!tarval_is_all_one(tv)) {
4636 /* have at least one 0 at the same bit position */
4640 if (tv3 != tarval_andnot(tv3, tv4)) {
4641 /* bit in the or_mask is outside the and_mask */
4645 if (tv1 != tarval_andnot(tv1, tv2)) {
4646 /* bit in the or_mask is outside the and_mask */
4650 /* ok, all conditions met */
4651 block = get_irn_n(irn_or, -1);
4652 irg = get_irn_irg(block);
4654 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4656 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4658 set_Or_left(irn_or, new_and);
4659 set_Or_right(irn_or, new_const);
4661 /* check for more */
4663 } /* transform_node_Or_bf_store */
4666 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4668 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4670 ir_mode *mode = get_irn_mode(irn_or);
4671 ir_node *shl, *shr, *block;
4672 ir_node *irn, *x, *c1, *c2, *n;
4673 ir_tarval *tv1, *tv2;
4675 /* some backends can't handle rotl */
4676 if (!be_get_backend_param()->support_rotl)
4679 if (! mode_is_int(mode))
4682 shl = get_binop_left(irn_or);
4683 shr = get_binop_right(irn_or);
4692 } else if (!is_Shl(shl)) {
4694 } else if (!is_Shr(shr)) {
4697 x = get_Shl_left(shl);
4698 if (x != get_Shr_left(shr))
4701 c1 = get_Shl_right(shl);
4702 c2 = get_Shr_right(shr);
4703 if (is_Const(c1) && is_Const(c2)) {
4704 tv1 = get_Const_tarval(c1);
4705 if (! tarval_is_long(tv1))
4708 tv2 = get_Const_tarval(c2);
4709 if (! tarval_is_long(tv2))
4712 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4713 != (int) get_mode_size_bits(mode))
4716 /* yet, condition met */
4717 block = get_nodes_block(irn_or);
4719 n = new_r_Rotl(block, x, c1, mode);
4721 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4725 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4726 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4727 if (!ir_is_negated_value(c1, c2)) {
4731 /* yet, condition met */
4732 block = get_nodes_block(irn_or);
4733 n = new_r_Rotl(block, x, c1, mode);
4734 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4736 } /* transform_node_Or_Rotl */
4738 static bool is_cmp_unequal_zero(const ir_node *node)
4740 ir_relation relation = get_Cmp_relation(node);
4741 ir_node *left = get_Cmp_left(node);
4742 ir_node *right = get_Cmp_right(node);
4743 ir_mode *mode = get_irn_mode(left);
4745 if (!is_Const(right) || !is_Const_null(right))
4747 if (mode_is_signed(mode)) {
4748 return relation == ir_relation_less_greater;
4750 return relation == ir_relation_greater;
4757 static ir_node *transform_node_Or(ir_node *n)
4759 ir_node *c, *oldn = n;
4760 ir_node *a = get_Or_left(n);
4761 ir_node *b = get_Or_right(n);
4764 if (is_Not(a) && is_Not(b)) {
4765 /* ~a | ~b = ~(a&b) */
4766 ir_node *block = get_nodes_block(n);
4768 mode = get_irn_mode(n);
4771 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4772 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4773 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4777 /* we can combine the relations of two compares with the same operands */
4778 if (is_Cmp(a) && is_Cmp(b)) {
4779 ir_node *a_left = get_Cmp_left(a);
4780 ir_node *a_right = get_Cmp_left(a);
4781 ir_node *b_left = get_Cmp_left(b);
4782 ir_node *b_right = get_Cmp_right(b);
4783 if (a_left == b_left && b_left == b_right) {
4784 dbg_info *dbgi = get_irn_dbg_info(n);
4785 ir_node *block = get_nodes_block(n);
4786 ir_relation a_relation = get_Cmp_relation(a);
4787 ir_relation b_relation = get_Cmp_relation(b);
4788 ir_relation new_relation = a_relation | b_relation;
4789 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4791 /* Cmp(a!=0) or Cmp(b!=0) => Cmp(a|b != 0) */
4792 if (is_cmp_unequal_zero(a) && is_cmp_unequal_zero(b)
4793 && !mode_is_float(get_irn_mode(a_left))
4794 && !mode_is_float(get_irn_mode(b_left))) {
4795 ir_graph *irg = get_irn_irg(n);
4796 dbg_info *dbgi = get_irn_dbg_info(n);
4797 ir_node *block = get_nodes_block(n);
4798 ir_mode *mode = get_irn_mode(a_left);
4799 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
4800 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
4801 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
4802 ir_node *zero = create_zero_const(irg, mode);
4803 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4807 mode = get_irn_mode(n);
4808 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4810 n = transform_node_Or_bf_store(n);
4811 n = transform_node_Or_Rotl(n);
4815 n = transform_bitwise_distributive(n, transform_node_Or);
4818 } /* transform_node_Or */
4822 static ir_node *transform_node(ir_node *n);
4825 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4827 * Should be moved to reassociation?
4829 static ir_node *transform_node_shift(ir_node *n)
4831 ir_node *left, *right;
4833 ir_tarval *tv1, *tv2, *res;
4834 ir_node *in[2], *irn, *block;
4837 left = get_binop_left(n);
4839 /* different operations */
4840 if (get_irn_op(left) != get_irn_op(n))
4843 right = get_binop_right(n);
4844 tv1 = value_of(right);
4845 if (tv1 == tarval_bad)
4848 tv2 = value_of(get_binop_right(left));
4849 if (tv2 == tarval_bad)
4852 res = tarval_add(tv1, tv2);
4853 mode = get_irn_mode(n);
4854 irg = get_irn_irg(n);
4856 /* beware: a simple replacement works only, if res < modulo shift */
4858 int modulo_shf = get_mode_modulo_shift(mode);
4859 if (modulo_shf > 0) {
4860 ir_tarval *modulo = new_tarval_from_long(modulo_shf,
4861 get_tarval_mode(res));
4863 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4865 /* shifting too much */
4866 if (!(tarval_cmp(res, modulo) & ir_relation_less)) {
4868 ir_node *block = get_nodes_block(n);
4869 dbg_info *dbgi = get_irn_dbg_info(n);
4870 ir_mode *smode = get_irn_mode(right);
4871 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
4872 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
4875 return new_r_Const(irg, get_mode_null(mode));
4879 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
4882 /* ok, we can replace it */
4883 block = get_nodes_block(n);
4885 in[0] = get_binop_left(left);
4886 in[1] = new_r_Const(irg, res);
4888 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
4890 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4892 return transform_node(irn);
4893 } /* transform_node_shift */
4896 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
4898 * - and, or, xor instead of &
4899 * - Shl, Shr, Shrs, rotl instead of >>
4900 * (with a special case for Or/Xor + Shrs)
4902 static ir_node *transform_node_bitop_shift(ir_node *n)
4905 ir_node *right = get_binop_right(n);
4906 ir_mode *mode = get_irn_mode(n);
4907 ir_node *bitop_left;
4908 ir_node *bitop_right;
4918 ir_tarval *tv_shift;
4920 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
4922 if (!is_Const(right))
4925 left = get_binop_left(n);
4926 op_left = get_irn_op(left);
4927 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
4930 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
4931 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
4932 /* TODO: test if sign bit is affectes */
4936 bitop_right = get_binop_right(left);
4937 if (!is_Const(bitop_right))
4940 bitop_left = get_binop_left(left);
4942 block = get_nodes_block(n);
4943 dbgi = get_irn_dbg_info(n);
4944 tv1 = get_Const_tarval(bitop_right);
4945 tv2 = get_Const_tarval(right);
4947 assert(get_tarval_mode(tv1) == mode);
4950 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
4951 tv_shift = tarval_shl(tv1, tv2);
4952 } else if (is_Shr(n)) {
4953 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
4954 tv_shift = tarval_shr(tv1, tv2);
4955 } else if (is_Shrs(n)) {
4956 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
4957 tv_shift = tarval_shrs(tv1, tv2);
4960 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
4961 tv_shift = tarval_rotl(tv1, tv2);
4964 assert(get_tarval_mode(tv_shift) == mode);
4965 irg = get_irn_irg(n);
4966 new_const = new_r_Const(irg, tv_shift);
4968 if (op_left == op_And) {
4969 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
4970 } else if (op_left == op_Or) {
4971 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
4973 assert(op_left == op_Eor);
4974 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
4982 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
4984 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
4985 * (also with x >>s c1 when c1>=c2)
4987 static ir_node *transform_node_shl_shr(ir_node *n)
4990 ir_node *right = get_binop_right(n);
5000 ir_tarval *tv_shift;
5003 ir_relation relation;
5006 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5008 if (!is_Const(right))
5011 left = get_binop_left(n);
5012 mode = get_irn_mode(n);
5013 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5014 ir_node *shr_right = get_binop_right(left);
5016 if (!is_Const(shr_right))
5019 x = get_binop_left(left);
5020 tv_shr = get_Const_tarval(shr_right);
5021 tv_shl = get_Const_tarval(right);
5023 if (is_Shrs(left)) {
5024 /* shrs variant only allowed if c1 >= c2 */
5025 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5028 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5031 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5033 tv_mask = tarval_shl(tv_mask, tv_shl);
5034 } else if (is_Shr(n) && is_Shl(left)) {
5035 ir_node *shl_right = get_Shl_right(left);
5037 if (!is_Const(shl_right))
5040 x = get_Shl_left(left);
5041 tv_shr = get_Const_tarval(right);
5042 tv_shl = get_Const_tarval(shl_right);
5044 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5045 tv_mask = tarval_shr(tv_mask, tv_shr);
5050 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5051 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5054 assert(tv_mask != tarval_bad);
5055 assert(get_tarval_mode(tv_mask) == mode);
5057 block = get_nodes_block(n);
5058 irg = get_irn_irg(block);
5059 dbgi = get_irn_dbg_info(n);
5061 relation = tarval_cmp(tv_shl, tv_shr);
5062 if (relation == ir_relation_less || relation == ir_relation_equal) {
5063 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5064 new_const = new_r_Const(irg, tv_shift);
5066 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5068 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5071 assert(relation == ir_relation_greater);
5072 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5073 new_const = new_r_Const(irg, tv_shift);
5074 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5077 new_const = new_r_Const(irg, tv_mask);
5078 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5083 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5085 ir_mode *mode = get_tarval_mode(tv);
5086 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5087 return tarval_mod(tv, modulo_tv);
5090 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5091 ir_node *left, ir_node *right, ir_mode *mode);
5094 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5095 * then we can use that to minimize the value of Add(x, const) or
5096 * Sub(Const, x). In particular this often avoids 1 instruction in some
5097 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5098 * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
5100 static ir_node *transform_node_shift_modulo(ir_node *n,
5101 new_shift_func new_shift)
5103 ir_mode *mode = get_irn_mode(n);
5104 int modulo = get_mode_modulo_shift(mode);
5105 ir_node *newop = NULL;
5106 ir_mode *mode_right;
5113 if (get_mode_arithmetic(mode) != irma_twos_complement)
5115 if (!is_po2(modulo))
5118 irg = get_irn_irg(n);
5119 block = get_nodes_block(n);
5120 right = get_binop_right(n);
5121 mode_right = get_irn_mode(right);
5122 if (is_Const(right)) {
5123 ir_tarval *tv = get_Const_tarval(right);
5124 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5129 newop = new_r_Const(irg, tv_mod);
5130 } else if (is_Add(right)) {
5131 ir_node *add_right = get_Add_right(right);
5132 if (is_Const(add_right)) {
5133 ir_tarval *tv = get_Const_tarval(add_right);
5134 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5139 newconst = new_r_Const(irg, tv_mod);
5140 newop = new_r_Add(block, get_Add_left(right), newconst,
5143 } else if (is_Sub(right)) {
5144 ir_node *sub_left = get_Sub_left(right);
5145 if (is_Const(sub_left)) {
5146 ir_tarval *tv = get_Const_tarval(sub_left);
5147 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5152 newconst = new_r_Const(irg, tv_mod);
5153 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5160 if (newop != NULL) {
5161 dbg_info *dbgi = get_irn_dbg_info(n);
5162 ir_node *left = get_binop_left(n);
5163 return new_shift(dbgi, block, left, newop, mode);
5171 static ir_node *transform_node_Shr(ir_node *n)
5173 ir_node *c, *oldn = n;
5174 ir_node *left = get_Shr_left(n);
5175 ir_node *right = get_Shr_right(n);
5176 ir_mode *mode = get_irn_mode(n);
5178 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5179 n = transform_node_shift(n);
5182 n = transform_node_shift_modulo(n, new_rd_Shr);
5184 n = transform_node_shl_shr(n);
5186 n = transform_node_bitop_shift(n);
5189 } /* transform_node_Shr */
5194 static ir_node *transform_node_Shrs(ir_node *n)
5196 ir_node *c, *oldn = n;
5197 ir_node *a = get_Shrs_left(n);
5198 ir_node *b = get_Shrs_right(n);
5199 ir_mode *mode = get_irn_mode(n);
5201 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5202 n = transform_node_shift(n);
5205 n = transform_node_shift_modulo(n, new_rd_Shrs);
5207 n = transform_node_bitop_shift(n);
5210 } /* transform_node_Shrs */
5215 static ir_node *transform_node_Shl(ir_node *n)
5217 ir_node *c, *oldn = n;
5218 ir_node *a = get_Shl_left(n);
5219 ir_node *b = get_Shl_right(n);
5220 ir_mode *mode = get_irn_mode(n);
5222 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5223 n = transform_node_shift(n);
5226 n = transform_node_shift_modulo(n, new_rd_Shl);
5228 n = transform_node_shl_shr(n);
5230 n = transform_node_bitop_shift(n);
5233 } /* transform_node_Shl */
5238 static ir_node *transform_node_Rotl(ir_node *n)
5240 ir_node *c, *oldn = n;
5241 ir_node *a = get_Rotl_left(n);
5242 ir_node *b = get_Rotl_right(n);
5243 ir_mode *mode = get_irn_mode(n);
5245 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5246 n = transform_node_shift(n);
5249 n = transform_node_bitop_shift(n);
5252 } /* transform_node_Rotl */
5257 static ir_node *transform_node_Conv(ir_node *n)
5259 ir_node *c, *oldn = n;
5260 ir_mode *mode = get_irn_mode(n);
5261 ir_node *a = get_Conv_op(n);
5263 if (mode != mode_b && is_const_Phi(a)) {
5264 /* Do NOT optimize mode_b Conv's, this leads to remaining
5265 * Phib nodes later, because the conv_b_lower operation
5266 * is instantly reverted, when it tries to insert a Convb.
5268 c = apply_conv_on_phi(a, mode);
5270 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5275 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5276 ir_graph *irg = get_irn_irg(n);
5277 return new_r_Unknown(irg, mode);
5280 if (mode_is_reference(mode) &&
5281 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5283 ir_node *l = get_Add_left(a);
5284 ir_node *r = get_Add_right(a);
5285 dbg_info *dbgi = get_irn_dbg_info(a);
5286 ir_node *block = get_nodes_block(n);
5288 ir_node *lop = get_Conv_op(l);
5289 if (get_irn_mode(lop) == mode) {
5290 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5291 n = new_rd_Add(dbgi, block, lop, r, mode);
5296 ir_node *rop = get_Conv_op(r);
5297 if (get_irn_mode(rop) == mode) {
5298 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5299 n = new_rd_Add(dbgi, block, l, rop, mode);
5306 } /* transform_node_Conv */
5309 * Remove dead blocks and nodes in dead blocks
5310 * in keep alive list. We do not generate a new End node.
5312 static ir_node *transform_node_End(ir_node *n)
5314 int i, j, n_keepalives = get_End_n_keepalives(n);
5317 NEW_ARR_A(ir_node *, in, n_keepalives);
5319 for (i = j = 0; i < n_keepalives; ++i) {
5320 ir_node *ka = get_End_keepalive(n, i);
5322 /* no need to keep Bad */
5327 if (j != n_keepalives)
5328 set_End_keepalives(n, j, in);
5330 } /* transform_node_End */
5332 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5334 if (is_Minus(a) && get_Minus_op(a) == b)
5336 if (is_Minus(b) && get_Minus_op(b) == a)
5338 if (is_Sub(a) && is_Sub(b)) {
5339 ir_node *a_left = get_Sub_left(a);
5340 ir_node *a_right = get_Sub_right(a);
5341 ir_node *b_left = get_Sub_left(b);
5342 ir_node *b_right = get_Sub_right(b);
5344 if (a_left == b_right && a_right == b_left)
5352 * Optimize a Mux into some simpler cases.
5354 static ir_node *transform_node_Mux(ir_node *n)
5356 ir_node *oldn = n, *sel = get_Mux_sel(n);
5357 ir_mode *mode = get_irn_mode(n);
5358 ir_node *t = get_Mux_true(n);
5359 ir_node *f = get_Mux_false(n);
5360 ir_graph *irg = get_irn_irg(n);
5362 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5366 ir_node* block = get_nodes_block(n);
5368 ir_node* c1 = get_Mux_sel(t);
5369 ir_node* t1 = get_Mux_true(t);
5370 ir_node* f1 = get_Mux_false(t);
5372 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5373 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5374 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5379 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5380 } else if (f == t1) {
5381 /* Mux(cond0, Mux(cond1, x, y), x) */
5382 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5383 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5384 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5389 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5391 } else if (is_Mux(f)) {
5392 ir_node* block = get_nodes_block(n);
5394 ir_node* c1 = get_Mux_sel(f);
5395 ir_node* t1 = get_Mux_true(f);
5396 ir_node* f1 = get_Mux_false(f);
5398 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5399 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5400 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5405 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5406 } else if (t == f1) {
5407 /* Mux(cond0, x, Mux(cond1, y, x)) */
5408 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5409 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5410 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5415 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5419 /* first normalization step: try to move a constant to the false side,
5420 * 0 preferred on false side too */
5421 if (is_Cmp(sel) && is_Const(t) &&
5422 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5423 dbg_info *seldbgi = get_irn_dbg_info(sel);
5424 ir_node *block = get_nodes_block(sel);
5425 ir_relation relation = get_Cmp_relation(sel);
5430 /* Mux(x, a, b) => Mux(not(x), b, a) */
5431 relation = get_negated_relation(relation);
5432 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5433 get_Cmp_right(sel), relation);
5434 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5437 /* note: after normalization, false can only happen on default */
5438 if (mode == mode_b) {
5439 dbg_info *dbg = get_irn_dbg_info(n);
5440 ir_node *block = get_nodes_block(n);
5443 ir_tarval *tv_t = get_Const_tarval(t);
5444 if (tv_t == tarval_b_true) {
5446 /* Muxb(sel, true, false) = sel */
5447 assert(get_Const_tarval(f) == tarval_b_false);
5448 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5451 /* Muxb(sel, true, x) = Or(sel, x) */
5452 n = new_rd_Or(dbg, block, sel, f, mode_b);
5453 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5457 } else if (is_Const(f)) {
5458 ir_tarval *tv_f = get_Const_tarval(f);
5459 if (tv_f == tarval_b_true) {
5460 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5461 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5462 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5463 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5466 /* Muxb(sel, x, false) = And(sel, x) */
5467 assert(tv_f == tarval_b_false);
5468 n = new_rd_And(dbg, block, sel, t, mode_b);
5469 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5475 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5476 * value to integer. */
5477 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5478 ir_tarval *a = get_Const_tarval(t);
5479 ir_tarval *b = get_Const_tarval(f);
5481 if (tarval_is_one(a) && tarval_is_null(b)) {
5482 ir_node *block = get_nodes_block(n);
5483 ir_node *conv = new_r_Conv(block, sel, mode);
5485 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5487 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5488 ir_node *block = get_nodes_block(n);
5489 ir_node *not_ = new_r_Not(block, sel, mode_b);
5490 ir_node *conv = new_r_Conv(block, not_, mode);
5492 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5498 ir_node *cmp_r = get_Cmp_right(sel);
5499 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5500 ir_node *block = get_nodes_block(n);
5501 ir_node *cmp_l = get_Cmp_left(sel);
5503 if (mode_is_int(mode)) {
5504 ir_relation relation = get_Cmp_relation(sel);
5506 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5507 /* Mux((a & b) != 0, c, 0) */
5508 ir_node *and_r = get_And_right(cmp_l);
5511 if (and_r == t && f == cmp_r) {
5512 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5513 if (relation == ir_relation_less_greater) {
5514 /* Mux((a & 2^C) != 0, 2^C, 0) */
5516 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5518 /* Mux((a & 2^C) == 0, 2^C, 0) */
5519 n = new_rd_Eor(get_irn_dbg_info(n),
5520 block, cmp_l, t, mode);
5521 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5526 if (is_Shl(and_r)) {
5527 ir_node *shl_l = get_Shl_left(and_r);
5528 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5529 if (and_r == t && f == cmp_r) {
5530 if (relation == ir_relation_less_greater) {
5531 /* (a & (1 << n)) != 0, (1 << n), 0) */
5533 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5535 /* (a & (1 << n)) == 0, (1 << n), 0) */
5536 n = new_rd_Eor(get_irn_dbg_info(n),
5537 block, cmp_l, t, mode);
5538 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5544 and_l = get_And_left(cmp_l);
5545 if (is_Shl(and_l)) {
5546 ir_node *shl_l = get_Shl_left(and_l);
5547 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5548 if (and_l == t && f == cmp_r) {
5549 if (relation == ir_relation_less_greater) {
5550 /* ((1 << n) & a) != 0, (1 << n), 0) */
5552 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5554 /* ((1 << n) & a) == 0, (1 << n), 0) */
5555 n = new_rd_Eor(get_irn_dbg_info(n),
5556 block, cmp_l, t, mode);
5557 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5572 * optimize Sync nodes that have other syncs as input we simply add the inputs
5573 * of the other sync to our own inputs
5575 static ir_node *transform_node_Sync(ir_node *n)
5577 int arity = get_Sync_n_preds(n);
5580 for (i = 0; i < arity;) {
5581 ir_node *pred = get_Sync_pred(n, i);
5585 if (!is_Sync(pred)) {
5593 pred_arity = get_Sync_n_preds(pred);
5594 for (j = 0; j < pred_arity; ++j) {
5595 ir_node *pred_pred = get_Sync_pred(pred, j);
5600 add_irn_n(n, pred_pred);
5604 if (get_Sync_pred(n, k) == pred_pred) break;
5609 /* rehash the sync node */
5613 } /* transform_node_Sync */
5615 static ir_node *transform_node_Load(ir_node *n)
5617 /* if our memory predecessor is a load from the same address, then reuse the
5618 * previous result */
5619 ir_node *mem = get_Load_mem(n);
5624 /* don't touch volatile loads */
5625 if (get_Load_volatility(n) == volatility_is_volatile)
5627 mem_pred = get_Proj_pred(mem);
5628 if (is_Load(mem_pred)) {
5629 ir_node *pred_load = mem_pred;
5631 /* conservatively compare the 2 loads. TODO: This could be less strict
5632 * with fixup code in some situations (like smaller/bigger modes) */
5633 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
5635 if (get_Load_mode(pred_load) != get_Load_mode(n))
5637 /* all combinations of aligned/unaligned pred/n should be fine so we do
5638 * not compare the unaligned attribute */
5640 ir_node *block = get_nodes_block(n);
5641 ir_node *jmp = new_r_Jmp(block);
5642 ir_graph *irg = get_irn_irg(n);
5643 ir_node *bad = get_irg_bad(irg);
5644 ir_mode *mode = get_Load_mode(n);
5645 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
5646 ir_node *in[pn_Load_max] = { mem, jmp, bad, res };
5647 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5650 } else if (is_Store(mem_pred)) {
5651 ir_node *pred_store = mem_pred;
5652 ir_node *value = get_Store_value(pred_store);
5654 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
5656 if (get_irn_mode(value) != get_Load_mode(n))
5658 /* all combinations of aligned/unaligned pred/n should be fine so we do
5659 * not compare the unaligned attribute */
5661 ir_node *block = get_nodes_block(n);
5662 ir_node *jmp = new_r_Jmp(block);
5663 ir_graph *irg = get_irn_irg(n);
5664 ir_node *bad = get_irg_bad(irg);
5665 ir_node *res = value;
5666 ir_node *in[pn_Load_max] = { mem, jmp, bad, res };
5667 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5676 * optimize a trampoline Call into a direct Call
5678 static ir_node *transform_node_Call(ir_node *call)
5680 ir_node *callee = get_Call_ptr(call);
5681 ir_node *adr, *mem, *res, *bl, **in;
5682 ir_type *ctp, *mtp, *tp;
5686 size_t i, n_res, n_param;
5689 if (! is_Proj(callee))
5691 callee = get_Proj_pred(callee);
5692 if (! is_Builtin(callee))
5694 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5697 mem = get_Call_mem(call);
5699 if (skip_Proj(mem) == callee) {
5700 /* memory is routed to the trampoline, skip */
5701 mem = get_Builtin_mem(callee);
5704 /* build a new call type */
5705 mtp = get_Call_type(call);
5706 tdb = get_type_dbg_info(mtp);
5708 n_res = get_method_n_ress(mtp);
5709 n_param = get_method_n_params(mtp);
5710 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5712 for (i = 0; i < n_res; ++i)
5713 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5715 NEW_ARR_A(ir_node *, in, n_param + 1);
5717 /* FIXME: we don't need a new pointer type in every step */
5718 irg = get_irn_irg(call);
5719 tp = get_irg_frame_type(irg);
5720 tp = new_type_pointer(tp);
5721 set_method_param_type(ctp, 0, tp);
5723 in[0] = get_Builtin_param(callee, 2);
5724 for (i = 0; i < n_param; ++i) {
5725 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5726 in[i + 1] = get_Call_param(call, i);
5728 var = get_method_variadicity(mtp);
5729 set_method_variadicity(ctp, var);
5730 /* When we resolve a trampoline, the function must be called by a this-call */
5731 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5732 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5734 adr = get_Builtin_param(callee, 1);
5736 db = get_irn_dbg_info(call);
5737 bl = get_nodes_block(call);
5739 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5740 if (get_irn_pinned(call) == op_pin_state_floats)
5741 set_irn_pinned(res, op_pin_state_floats);
5743 } /* transform_node_Call */
5746 * Tries several [inplace] [optimizing] transformations and returns an
5747 * equivalent node. The difference to equivalent_node() is that these
5748 * transformations _do_ generate new nodes, and thus the old node must
5749 * not be freed even if the equivalent node isn't the old one.
5751 static ir_node *transform_node(ir_node *n)
5756 * Transform_node is the only "optimizing transformation" that might
5757 * return a node with a different opcode. We iterate HERE until fixpoint
5758 * to get the final result.
5762 if (n->op->ops.transform_node != NULL)
5763 n = n->op->ops.transform_node(n);
5764 } while (oldn != n);
5767 } /* transform_node */
5770 * Sets the default transform node operation for an ir_op_ops.
5772 * @param code the opcode for the default operation
5773 * @param ops the operations initialized
5778 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5782 ops->transform_node = transform_node_##a; \
5784 #define CASE_PROJ(a) \
5786 ops->transform_node_Proj = transform_node_Proj_##a; \
5788 #define CASE_PROJ_EX(a) \
5790 ops->transform_node = transform_node_##a; \
5791 ops->transform_node_Proj = transform_node_Proj_##a; \
5831 } /* firm_set_default_transform_node */
5834 /* **************** Common Subexpression Elimination **************** */
5836 /** The size of the hash table used, should estimate the number of nodes
5838 #define N_IR_NODES 512
5840 /** Compares the attributes of two Const nodes. */
5841 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
5843 return get_Const_tarval(a) != get_Const_tarval(b);
5846 /** Compares the attributes of two Proj nodes. */
5847 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
5849 return a->attr.proj.proj != b->attr.proj.proj;
5852 /** Compares the attributes of two Alloc nodes. */
5853 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
5855 const alloc_attr *pa = &a->attr.alloc;
5856 const alloc_attr *pb = &b->attr.alloc;
5857 return (pa->where != pb->where) || (pa->type != pb->type);
5860 /** Compares the attributes of two Free nodes. */
5861 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
5863 const free_attr *pa = &a->attr.free;
5864 const free_attr *pb = &b->attr.free;
5865 return (pa->where != pb->where) || (pa->type != pb->type);
5868 /** Compares the attributes of two SymConst nodes. */
5869 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
5871 const symconst_attr *pa = &a->attr.symc;
5872 const symconst_attr *pb = &b->attr.symc;
5873 return (pa->kind != pb->kind)
5874 || (pa->sym.type_p != pb->sym.type_p);
5877 /** Compares the attributes of two Call nodes. */
5878 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
5880 const call_attr *pa = &a->attr.call;
5881 const call_attr *pb = &b->attr.call;
5882 return (pa->type != pb->type)
5883 || (pa->tail_call != pb->tail_call);
5886 /** Compares the attributes of two Sel nodes. */
5887 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
5889 const ir_entity *a_ent = get_Sel_entity(a);
5890 const ir_entity *b_ent = get_Sel_entity(b);
5891 return a_ent != b_ent;
5894 /** Compares the attributes of two Phi nodes. */
5895 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
5897 /* we can only enter this function if both nodes have the same number of inputs,
5898 hence it is enough to check if one of them is a Phi0 */
5900 /* check the Phi0 pos attribute */
5901 return a->attr.phi.u.pos != b->attr.phi.u.pos;
5906 /** Compares the attributes of two Conv nodes. */
5907 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
5909 return get_Conv_strict(a) != get_Conv_strict(b);
5912 /** Compares the attributes of two Cast nodes. */
5913 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
5915 return get_Cast_type(a) != get_Cast_type(b);
5918 /** Compares the attributes of two Load nodes. */
5919 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
5921 if (get_Load_volatility(a) == volatility_is_volatile ||
5922 get_Load_volatility(b) == volatility_is_volatile)
5923 /* NEVER do CSE on volatile Loads */
5925 /* do not CSE Loads with different alignment. Be conservative. */
5926 if (get_Load_unaligned(a) != get_Load_unaligned(b))
5929 return get_Load_mode(a) != get_Load_mode(b);
5932 /** Compares the attributes of two Store nodes. */
5933 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
5935 /* do not CSE Stores with different alignment. Be conservative. */
5936 if (get_Store_unaligned(a) != get_Store_unaligned(b))
5939 /* NEVER do CSE on volatile Stores */
5940 return (get_Store_volatility(a) == volatility_is_volatile ||
5941 get_Store_volatility(b) == volatility_is_volatile);
5944 /** Compares two exception attributes */
5945 static int node_cmp_exception(const ir_node *a, const ir_node *b)
5947 const except_attr *ea = &a->attr.except;
5948 const except_attr *eb = &b->attr.except;
5950 return ea->pin_state != eb->pin_state;
5953 #define node_cmp_attr_Bound node_cmp_exception
5955 /** Compares the attributes of two Div nodes. */
5956 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
5958 const div_attr *ma = &a->attr.div;
5959 const div_attr *mb = &b->attr.div;
5960 return ma->exc.pin_state != mb->exc.pin_state ||
5961 ma->resmode != mb->resmode ||
5962 ma->no_remainder != mb->no_remainder;
5965 /** Compares the attributes of two Mod nodes. */
5966 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
5968 const mod_attr *ma = &a->attr.mod;
5969 const mod_attr *mb = &b->attr.mod;
5970 return ma->exc.pin_state != mb->exc.pin_state ||
5971 ma->resmode != mb->resmode;
5974 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
5976 const cmp_attr *ma = &a->attr.cmp;
5977 const cmp_attr *mb = &b->attr.cmp;
5978 return ma->relation != mb->relation;
5981 /** Compares the attributes of two Confirm nodes. */
5982 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
5984 const confirm_attr *ma = &a->attr.confirm;
5985 const confirm_attr *mb = &b->attr.confirm;
5986 return ma->relation != mb->relation;
5989 /** Compares the attributes of two Builtin nodes. */
5990 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
5992 /* no need to compare the type, equal kind means equal type */
5993 return get_Builtin_kind(a) != get_Builtin_kind(b);
5996 /** Compares the attributes of two ASM nodes. */
5997 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
6000 const ir_asm_constraint *ca;
6001 const ir_asm_constraint *cb;
6004 if (get_ASM_text(a) != get_ASM_text(b))
6007 /* Should we really check the constraints here? Should be better, but is strange. */
6008 n = get_ASM_n_input_constraints(a);
6009 if (n != get_ASM_n_input_constraints(b))
6012 ca = get_ASM_input_constraints(a);
6013 cb = get_ASM_input_constraints(b);
6014 for (i = 0; i < n; ++i) {
6015 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6016 || ca[i].mode != cb[i].mode)
6020 n = get_ASM_n_output_constraints(a);
6021 if (n != get_ASM_n_output_constraints(b))
6024 ca = get_ASM_output_constraints(a);
6025 cb = get_ASM_output_constraints(b);
6026 for (i = 0; i < n; ++i) {
6027 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6028 || ca[i].mode != cb[i].mode)
6032 n = get_ASM_n_clobbers(a);
6033 if (n != get_ASM_n_clobbers(b))
6036 cla = get_ASM_clobbers(a);
6037 clb = get_ASM_clobbers(b);
6038 for (i = 0; i < n; ++i) {
6039 if (cla[i] != clb[i])
6045 /** Compares the inexistent attributes of two Dummy nodes. */
6046 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6054 * Set the default node attribute compare operation for an ir_op_ops.
6056 * @param code the opcode for the default operation
6057 * @param ops the operations initialized
6062 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6066 ops->node_cmp_attr = node_cmp_attr_##a; \
6098 } /* firm_set_default_node_cmp_attr */
6101 * Compare function for two nodes in the value table. Gets two
6102 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6104 int identities_cmp(const void *elt, const void *key)
6106 ir_node *a = (ir_node *)elt;
6107 ir_node *b = (ir_node *)key;
6110 if (a == b) return 0;
6112 if ((get_irn_op(a) != get_irn_op(b)) ||
6113 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6115 /* compare if a's in and b's in are of equal length */
6116 irn_arity_a = get_irn_arity(a);
6117 if (irn_arity_a != get_irn_arity(b))
6120 /* blocks are never the same */
6124 if (get_irn_pinned(a) == op_pin_state_pinned) {
6125 /* for pinned nodes, the block inputs must be equal */
6126 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6128 } else if (! get_opt_global_cse()) {
6129 /* for block-local CSE both nodes must be in the same Block */
6130 if (get_nodes_block(a) != get_nodes_block(b))
6134 /* compare a->in[0..ins] with b->in[0..ins] */
6135 for (i = 0; i < irn_arity_a; ++i) {
6136 ir_node *pred_a = get_irn_n(a, i);
6137 ir_node *pred_b = get_irn_n(b, i);
6138 if (pred_a != pred_b) {
6139 /* if both predecessors are CSE neutral they might be different */
6140 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6146 * here, we already now that the nodes are identical except their
6149 if (a->op->ops.node_cmp_attr)
6150 return a->op->ops.node_cmp_attr(a, b);
6153 } /* identities_cmp */
6156 * Calculate a hash value of a node.
6158 * @param node The IR-node
6160 unsigned ir_node_hash(const ir_node *node)
6162 return node->op->ops.hash(node);
6163 } /* ir_node_hash */
6166 void new_identities(ir_graph *irg)
6168 if (irg->value_table != NULL)
6169 del_pset(irg->value_table);
6170 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6171 } /* new_identities */
6173 void del_identities(ir_graph *irg)
6175 if (irg->value_table != NULL)
6176 del_pset(irg->value_table);
6177 } /* del_identities */
6179 /* Normalize a node by putting constants (and operands with larger
6180 * node index) on the right (operator side). */
6181 void ir_normalize_node(ir_node *n)
6183 if (is_op_commutative(get_irn_op(n))) {
6184 ir_node *l = get_binop_left(n);
6185 ir_node *r = get_binop_right(n);
6187 /* For commutative operators perform a OP b == b OP a but keep
6188 * constants on the RIGHT side. This helps greatly in some
6189 * optimizations. Moreover we use the idx number to make the form
6191 if (!operands_are_normalized(l, r)) {
6192 set_binop_left(n, r);
6193 set_binop_right(n, l);
6197 } /* ir_normalize_node */
6200 * Return the canonical node computing the same value as n.
6201 * Looks up the node in a hash table, enters it in the table
6202 * if it isn't there yet.
6204 * @param n the node to look up
6206 * @return a node that computes the same value as n or n if no such
6207 * node could be found
6209 ir_node *identify_remember(ir_node *n)
6211 ir_graph *irg = get_irn_irg(n);
6212 pset *value_table = irg->value_table;
6215 if (value_table == NULL)
6218 ir_normalize_node(n);
6219 /* lookup or insert in hash table with given hash key. */
6220 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6223 /* n is reachable again */
6224 edges_node_revival(nn);
6228 } /* identify_remember */
6231 * During construction we set the op_pin_state_pinned flag in the graph right
6232 * when the optimization is performed. The flag turning on procedure global
6233 * cse could be changed between two allocations. This way we are safe.
6235 * @param n The node to lookup
6237 static inline ir_node *identify_cons(ir_node *n)
6241 n = identify_remember(n);
6242 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6243 ir_graph *irg = get_irn_irg(n);
6244 set_irg_pinned(irg, op_pin_state_floats);
6247 } /* identify_cons */
6249 /* Add a node to the identities value table. */
6250 void add_identities(ir_node *node)
6257 identify_remember(node);
6260 /* Visit each node in the value table of a graph. */
6261 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6264 ir_graph *rem = current_ir_graph;
6266 current_ir_graph = irg;
6267 foreach_pset(irg->value_table, ir_node*, node) {
6270 current_ir_graph = rem;
6271 } /* visit_all_identities */
6274 * Garbage in, garbage out. If a node has a dead input, i.e., the
6275 * Bad node is input to the node, return the Bad node.
6277 static ir_node *gigo(ir_node *node)
6279 ir_op *op = get_irn_op(node);
6281 /* Code in "Bad" blocks is unreachable and can be replaced by Bad */
6282 if (op != op_Block && is_Bad(get_nodes_block(node))) {
6283 ir_graph *irg = get_irn_irg(node);
6284 return get_irg_bad(irg);
6291 * These optimizations deallocate nodes from the obstack.
6292 * It can only be called if it is guaranteed that no other nodes
6293 * reference this one, i.e., right after construction of a node.
6295 * @param n The node to optimize
6297 ir_node *optimize_node(ir_node *n)
6300 ir_graph *irg = get_irn_irg(n);
6301 unsigned iro = get_irn_opcode(n);
6304 /* Always optimize Phi nodes: part of the construction. */
6305 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6307 /* Remove nodes with dead (Bad) input.
6308 Run always for transformation induced Bads. */
6311 edges_node_deleted(oldn);
6313 /* We found an existing, better node, so we can deallocate the old node. */
6314 irg_kill_node(irg, oldn);
6318 /* constant expression evaluation / constant folding */
6319 if (get_opt_constant_folding()) {
6320 /* neither constants nor Tuple values can be evaluated */
6321 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6322 /* try to evaluate */
6323 tv = computed_value(n);
6324 if (tv != tarval_bad) {
6329 * we MUST copy the node here temporarily, because it's still
6330 * needed for DBG_OPT_CSTEVAL
6332 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6333 oldn = (ir_node*)alloca(node_size);
6335 memcpy(oldn, n, node_size);
6336 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6338 /* ARG, copy the in array, we need it for statistics */
6339 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6341 /* note the inplace edges module */
6342 edges_node_deleted(n);
6344 /* evaluation was successful -- replace the node. */
6345 irg_kill_node(irg, n);
6346 nw = new_r_Const(irg, tv);
6348 DBG_OPT_CSTEVAL(oldn, nw);
6354 /* remove unnecessary nodes */
6355 if (get_opt_algebraic_simplification() ||
6356 (iro == iro_Phi) || /* always optimize these nodes. */
6358 (iro == iro_Proj) ||
6359 (iro == iro_Block) ) /* Flags tested local. */
6360 n = equivalent_node(n);
6362 /* Common Subexpression Elimination.
6364 * Checks whether n is already available.
6365 * The block input is used to distinguish different subexpressions. Right
6366 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6367 * subexpressions within a block.
6370 n = identify_cons(n);
6373 edges_node_deleted(oldn);
6375 /* We found an existing, better node, so we can deallocate the old node. */
6376 irg_kill_node(irg, oldn);
6380 /* Some more constant expression evaluation that does not allow to
6382 iro = get_irn_opcode(n);
6383 if (get_opt_algebraic_simplification() ||
6384 (iro == iro_Cond) ||
6385 (iro == iro_Proj)) /* Flags tested local. */
6386 n = transform_node(n);
6388 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6389 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6391 n = identify_remember(o);
6397 } /* optimize_node */
6401 * These optimizations never deallocate nodes (in place). This can cause dead
6402 * nodes lying on the obstack. Remove these by a dead node elimination,
6403 * i.e., a copying garbage collection.
6405 ir_node *optimize_in_place_2(ir_node *n)
6409 unsigned iro = get_irn_opcode(n);
6411 if (!get_opt_optimize() && !is_Phi(n)) return n;
6413 if (iro == iro_Deleted)
6416 /* Remove nodes with dead (Bad) input.
6417 Run always for transformation induced Bads. */
6422 /* constant expression evaluation / constant folding */
6423 if (get_opt_constant_folding()) {
6424 /* neither constants nor Tuple values can be evaluated */
6425 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6426 /* try to evaluate */
6427 tv = computed_value(n);
6428 if (tv != tarval_bad) {
6429 /* evaluation was successful -- replace the node. */
6430 ir_graph *irg = get_irn_irg(n);
6432 n = new_r_Const(irg, tv);
6434 DBG_OPT_CSTEVAL(oldn, n);
6440 /* remove unnecessary nodes */
6441 if (get_opt_constant_folding() ||
6442 (iro == iro_Phi) || /* always optimize these nodes. */
6443 (iro == iro_Id) || /* ... */
6444 (iro == iro_Proj) || /* ... */
6445 (iro == iro_Block) ) /* Flags tested local. */
6446 n = equivalent_node(n);
6448 /** common subexpression elimination **/
6449 /* Checks whether n is already available. */
6450 /* The block input is used to distinguish different subexpressions. Right
6451 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6452 subexpressions within a block. */
6453 if (get_opt_cse()) {
6455 n = identify_remember(o);
6460 /* Some more constant expression evaluation. */
6461 iro = get_irn_opcode(n);
6462 if (get_opt_constant_folding() ||
6463 (iro == iro_Cond) ||
6464 (iro == iro_Proj)) /* Flags tested local. */
6465 n = transform_node(n);
6467 /* Now we can verify the node, as it has no dead inputs any more. */
6470 /* Now we have a legal, useful node. Enter it in hash table for cse.
6471 Blocks should be unique anyways. (Except the successor of start:
6472 is cse with the start block!) */
6473 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6475 n = identify_remember(o);
6481 } /* optimize_in_place_2 */
6484 * Wrapper for external use, set proper status bits after optimization.
6486 ir_node *optimize_in_place(ir_node *n)
6488 ir_graph *irg = get_irn_irg(n);
6489 /* Handle graph state */
6490 assert(get_irg_phase_state(irg) != phase_building);
6492 if (get_opt_global_cse())
6493 set_irg_pinned(irg, op_pin_state_floats);
6494 if (get_irg_outs_state(irg) == outs_consistent)
6495 set_irg_outs_inconsistent(irg);
6497 /* FIXME: Maybe we could also test whether optimizing the node can
6498 change the control graph. */
6499 set_irg_doms_inconsistent(irg);
6500 return optimize_in_place_2(n);
6501 } /* optimize_in_place */
6504 * Calculate a hash value of a Const node.
6506 static unsigned hash_Const(const ir_node *node)
6510 /* special value for const, as they only differ in their tarval. */
6511 h = HASH_PTR(node->attr.con.tarval);
6517 * Calculate a hash value of a SymConst node.
6519 static unsigned hash_SymConst(const ir_node *node)
6523 /* all others are pointers */
6524 h = HASH_PTR(node->attr.symc.sym.type_p);
6527 } /* hash_SymConst */
6530 * Set the default hash operation in an ir_op_ops.
6532 * @param code the opcode for the default operation
6533 * @param ops the operations initialized
6538 static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops)
6542 ops->hash = hash_##a; \
6545 /* hash function already set */
6546 if (ops->hash != NULL)
6553 /* use input/mode default hash if no function was given */
6554 ops->hash = firm_default_hash;
6562 * Sets the default operation for an ir_ops.
6564 ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops)
6566 ops = firm_set_default_hash(code, ops);
6567 ops = firm_set_default_computed_value(code, ops);
6568 ops = firm_set_default_equivalent_node(code, ops);
6569 ops = firm_set_default_transform_node(code, ops);
6570 ops = firm_set_default_node_cmp_attr(code, ops);
6571 ops = firm_set_default_get_type_attr(code, ops);
6572 ops = firm_set_default_get_entity_attr(code, ops);
6575 } /* firm_set_default_operations */