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 /* Straightening: a single entry Block following a single exit Block
697 ir_node *pred = get_Block_cfgpred(n, 0);
700 ir_node *pred_block = get_nodes_block(pred);
701 DBG_OPT_STG(n, pred_block);
704 } else if (n_preds == 2) {
705 /* Test whether Cond jumps twice to this block
706 * The more general case which more than 2 predecessors is handles
707 * in optimize_cf(), we handle only this special case for speed here.
709 ir_node *a = get_Block_cfgpred(n, 0);
710 ir_node *b = get_Block_cfgpred(n, 1);
712 if (is_Proj(a) && is_Proj(b)) {
713 ir_node *cond = get_Proj_pred(a);
715 if (cond == get_Proj_pred(b) && is_Cond(cond) &&
716 get_irn_mode(get_Cond_selector(cond)) == mode_b) {
717 /* Also a single entry Block following a single exit Block.
718 * Phis have twice the same operand and will be optimized away.
720 n = get_nodes_block(cond);
721 DBG_OPT_IFSIM1(oldn, a, b, n);
724 } else if (is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) {
726 int n_cfgpreds = get_Block_n_cfgpreds(n);
728 for (i = 0; i < n_cfgpreds; ++i) {
729 ir_node *pred = get_Block_cfgpred(n, i);
733 /* only bad unreachable inputs? It's unreachable code (unless it is the
734 * start or end block) */
735 if (i >= n_cfgpreds && n != get_irg_start_block(irg)
736 && n != get_irg_end_block(irg)) {
737 return get_irg_bad(irg);
742 } /* equivalent_node_Block */
744 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
745 See transform_node_Proj_Cond(). */
748 * Optimize operations that are commutative and have neutral 0,
749 * so a op 0 = 0 op a = a.
751 static ir_node *equivalent_node_neutral_zero(ir_node *n)
755 ir_node *a = get_binop_left(n);
756 ir_node *b = get_binop_right(n);
761 /* After running compute_node there is only one constant predecessor.
762 Find this predecessors value and remember the other node: */
763 if ((tv = value_of(a)) != tarval_bad) {
765 } else if ((tv = value_of(b)) != tarval_bad) {
770 /* If this predecessors constant value is zero, the operation is
771 * unnecessary. Remove it.
773 * Beware: If n is a Add, the mode of on and n might be different
774 * which happens in this rare construction: NULL + 3.
775 * Then, a Conv would be needed which we cannot include here.
777 if (tarval_is_null(tv) && get_irn_mode(on) == get_irn_mode(n)) {
780 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
784 } /* equivalent_node_neutral_zero */
787 * Eor is commutative and has neutral 0.
789 static ir_node *equivalent_node_Eor(ir_node *n)
795 n = equivalent_node_neutral_zero(n);
796 if (n != oldn) return n;
799 b = get_Eor_right(n);
802 ir_node *aa = get_Eor_left(a);
803 ir_node *ab = get_Eor_right(a);
806 /* (a ^ b) ^ a -> b */
808 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
810 } else if (ab == b) {
811 /* (a ^ b) ^ b -> a */
813 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
818 ir_node *ba = get_Eor_left(b);
819 ir_node *bb = get_Eor_right(b);
822 /* a ^ (a ^ b) -> b */
824 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
826 } else if (bb == a) {
827 /* a ^ (b ^ a) -> b */
829 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A);
837 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
839 * The second one looks strange, but this construct
840 * is used heavily in the LCC sources :-).
842 * Beware: The Mode of an Add may be different than the mode of its
843 * predecessors, so we could not return a predecessors in all cases.
845 static ir_node *equivalent_node_Add(ir_node *n)
848 ir_node *left, *right;
849 ir_mode *mode = get_irn_mode(n);
851 n = equivalent_node_neutral_zero(n);
855 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
856 if (mode_is_float(mode)) {
857 ir_graph *irg = get_irn_irg(n);
858 if (get_irg_fp_model(irg) & fp_strict_algebraic)
862 left = get_Add_left(n);
863 right = get_Add_right(n);
866 if (get_Sub_right(left) == right) {
869 n = get_Sub_left(left);
870 if (mode == get_irn_mode(n)) {
871 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
877 if (get_Sub_right(right) == left) {
880 n = get_Sub_left(right);
881 if (mode == get_irn_mode(n)) {
882 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
888 } /* equivalent_node_Add */
891 * optimize operations that are not commutative but have neutral 0 on left,
894 static ir_node *equivalent_node_left_zero(ir_node *n)
898 ir_node *a = get_binop_left(n);
899 ir_node *b = get_binop_right(n);
900 ir_tarval *tb = value_of(b);
902 if (tarval_is_null(tb)) {
905 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
908 } /* equivalent_node_left_zero */
910 #define equivalent_node_Shl equivalent_node_left_zero
911 #define equivalent_node_Shr equivalent_node_left_zero
912 #define equivalent_node_Shrs equivalent_node_left_zero
913 #define equivalent_node_Rotl equivalent_node_left_zero
916 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
918 * The second one looks strange, but this construct
919 * is used heavily in the LCC sources :-).
921 * Beware: The Mode of a Sub may be different than the mode of its
922 * predecessors, so we could not return a predecessors in all cases.
924 static ir_node *equivalent_node_Sub(ir_node *n)
928 ir_mode *mode = get_irn_mode(n);
931 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
932 if (mode_is_float(mode)) {
933 ir_graph *irg = get_irn_irg(n);
934 if (get_irg_fp_model(irg) & fp_strict_algebraic)
938 b = get_Sub_right(n);
941 /* Beware: modes might be different */
942 if (tarval_is_null(tb)) {
943 ir_node *a = get_Sub_left(n);
944 if (mode == get_irn_mode(a)) {
947 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
951 } /* equivalent_node_Sub */
955 * Optimize an "self-inverse unary op", ie op(op(n)) = n.
958 * -(-a) == a, but might overflow two times.
959 * We handle it anyway here but the better way would be a
960 * flag. This would be needed for Pascal for instance.
962 static ir_node *equivalent_node_idempotent_unop(ir_node *n)
965 ir_node *pred = get_unop_op(n);
967 /* optimize symmetric unop */
968 if (get_irn_op(pred) == get_irn_op(n)) {
969 n = get_unop_op(pred);
970 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
973 } /* equivalent_node_idempotent_unop */
975 /** Optimize Not(Not(x)) == x. */
976 #define equivalent_node_Not equivalent_node_idempotent_unop
978 /** -(-x) == x ??? Is this possible or can --x raise an
979 out of bounds exception if min =! max? */
980 #define equivalent_node_Minus equivalent_node_idempotent_unop
983 * Optimize a * 1 = 1 * a = a.
985 static ir_node *equivalent_node_Mul(ir_node *n)
988 ir_node *a = get_Mul_left(n);
990 /* we can handle here only the n * n = n bit cases */
991 if (get_irn_mode(n) == get_irn_mode(a)) {
992 ir_node *b = get_Mul_right(n);
996 * Mul is commutative and has again an other neutral element.
997 * Constants are place right, so check this case first.
1000 if (tarval_is_one(tv)) {
1002 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1005 if (tarval_is_one(tv)) {
1007 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1012 } /* equivalent_node_Mul */
1015 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1017 static ir_node *equivalent_node_Or(ir_node *n)
1021 ir_node *a = get_Or_left(n);
1022 ir_node *b = get_Or_right(n);
1026 n = a; /* idempotence */
1027 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1030 /* constants are normalized to right, check this side first */
1032 if (tarval_is_null(tv)) {
1034 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1038 if (tarval_is_null(tv)) {
1040 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1045 } /* equivalent_node_Or */
1048 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a.
1050 static ir_node *equivalent_node_And(ir_node *n)
1054 ir_node *a = get_And_left(n);
1055 ir_node *b = get_And_right(n);
1059 n = a; /* idempotence */
1060 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1063 /* constants are normalized to right, check this side first */
1065 if (tarval_is_all_one(tv)) {
1067 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1070 if (tv != get_tarval_bad()) {
1071 ir_mode *mode = get_irn_mode(n);
1072 if (!mode_is_signed(mode) && is_Conv(a)) {
1073 ir_node *convop = get_Conv_op(a);
1074 ir_mode *convopmode = get_irn_mode(convop);
1075 if (!mode_is_signed(convopmode)) {
1076 if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) {
1077 /* Conv(X) & all_one(mode(X)) = Conv(X) */
1079 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1086 if (tarval_is_all_one(tv)) {
1088 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1092 if (b == get_Or_left(a) || b == get_Or_right(a)) {
1095 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1100 if (a == get_Or_left(b) || a == get_Or_right(b)) {
1103 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1108 } /* equivalent_node_And */
1111 * Try to remove useless Conv's:
1113 static ir_node *equivalent_node_Conv(ir_node *n)
1116 ir_node *a = get_Conv_op(n);
1118 ir_mode *n_mode = get_irn_mode(n);
1119 ir_mode *a_mode = get_irn_mode(a);
1122 if (n_mode == a_mode) { /* No Conv necessary */
1123 if (get_Conv_strict(n)) {
1126 /* neither Minus nor Confirm change the precision,
1127 so we can "look-through" */
1130 p = get_Minus_op(p);
1131 } else if (is_Confirm(p)) {
1132 p = get_Confirm_value(p);
1138 if (is_Conv(p) && get_Conv_strict(p)) {
1139 /* we known already, that a_mode == n_mode, and neither
1140 Minus change the mode, so the second Conv
1142 assert(get_irn_mode(p) == n_mode);
1144 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1148 ir_node *pred = get_Proj_pred(p);
1149 if (is_Load(pred)) {
1150 /* Loads always return with the exact precision of n_mode */
1151 assert(get_Load_mode(pred) == n_mode);
1153 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1156 if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) {
1157 pred = get_Proj_pred(pred);
1158 if (is_Start(pred)) {
1159 /* Arguments always return with the exact precision,
1160 as strictConv's are place before Call -- if the
1161 caller was compiled with the same setting.
1162 Otherwise, the semantics is probably still right. */
1163 assert(get_irn_mode(p) == n_mode);
1165 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1171 /* special case: the immediate predecessor is also a Conv */
1172 if (! get_Conv_strict(a)) {
1173 /* first one is not strict, kick it */
1175 a_mode = get_irn_mode(a);
1179 /* else both are strict conv, second is superfluous */
1181 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1186 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1189 } else if (is_Conv(a)) { /* Conv(Conv(b)) */
1190 ir_node *b = get_Conv_op(a);
1191 ir_mode *b_mode = get_irn_mode(b);
1193 if (get_Conv_strict(n) && get_Conv_strict(a)) {
1194 /* both are strict conv */
1195 if (smaller_mode(a_mode, n_mode)) {
1196 /* both are strict, but the first is smaller, so
1197 the second cannot remove more precision, remove the
1199 set_Conv_strict(n, 0);
1202 if (n_mode == b_mode) {
1203 if (! get_Conv_strict(n) && ! get_Conv_strict(a)) {
1204 if (n_mode == mode_b) {
1205 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1206 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1208 } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) {
1209 if (values_in_mode(b_mode, a_mode)) {
1210 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1211 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1216 if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) {
1217 /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */
1218 unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0);
1219 unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode);
1221 if (float_mantissa >= int_mantissa) {
1223 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1228 if (smaller_mode(b_mode, a_mode)) {
1229 if (get_Conv_strict(n))
1230 set_Conv_strict(b, 1);
1231 n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */
1232 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1239 } /* equivalent_node_Conv */
1242 * - fold Phi-nodes, iff they have only one predecessor except
1245 static ir_node *equivalent_node_Phi(ir_node *n)
1251 ir_node *first_val = NULL; /* to shutup gcc */
1253 if (!get_opt_optimize() &&
1254 get_irg_phase_state(get_irn_irg(n)) != phase_building)
1257 n_preds = get_Phi_n_preds(n);
1259 block = get_nodes_block(n);
1261 /* Phi of dead Region without predecessors. */
1265 /* Find first non-self-referencing input */
1266 for (i = 0; i < n_preds; ++i) {
1267 first_val = get_Phi_pred(n, i);
1268 /* not self pointer */
1269 if (first_val != n) {
1270 /* then found first value. */
1276 ir_graph *irg = get_irn_irg(n);
1277 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1278 return get_irg_bad(irg);
1281 /* search for rest of inputs, determine if any of these
1282 are non-self-referencing */
1283 while (++i < n_preds) {
1284 ir_node *scnd_val = get_Phi_pred(n, i);
1285 if (scnd_val != n && scnd_val != first_val) {
1290 if (i >= n_preds && !is_Dummy(first_val)) {
1291 /* Fold, if no multiple distinct non-self-referencing inputs */
1293 DBG_OPT_PHI(oldn, n);
1296 } /* equivalent_node_Phi */
1299 * Several optimizations:
1300 * - fold Sync-nodes, iff they have only one predecessor except
1303 static ir_node *equivalent_node_Sync(ir_node *n)
1305 int arity = get_Sync_n_preds(n);
1308 for (i = 0; i < arity;) {
1309 ir_node *pred = get_Sync_pred(n, i);
1312 /* Remove Bad predecessors */
1319 /* Remove duplicate predecessors */
1325 if (get_Sync_pred(n, j) == pred) {
1334 ir_graph *irg = get_irn_irg(n);
1335 return get_irg_bad(irg);
1337 if (arity == 1) return get_Sync_pred(n, 0);
1339 } /* equivalent_node_Sync */
1342 * Optimize Proj(Tuple).
1344 static ir_node *equivalent_node_Proj_Tuple(ir_node *proj)
1346 ir_node *oldn = proj;
1347 ir_node *tuple = get_Proj_pred(proj);
1349 /* Remove the Tuple/Proj combination. */
1350 proj = get_Tuple_pred(tuple, get_Proj_proj(proj));
1351 DBG_OPT_TUPLE(oldn, tuple, proj);
1354 } /* equivalent_node_Proj_Tuple */
1357 * Optimize a / 1 = a.
1359 static ir_node *equivalent_node_Proj_Div(ir_node *proj)
1361 ir_node *oldn = proj;
1362 ir_node *div = get_Proj_pred(proj);
1363 ir_node *b = get_Div_right(div);
1364 ir_tarval *tb = value_of(b);
1366 /* Div is not commutative. */
1367 if (tarval_is_one(tb)) { /* div(x, 1) == x */
1368 switch (get_Proj_proj(proj)) {
1370 proj = get_Div_mem(div);
1371 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1375 proj = get_Div_left(div);
1376 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1);
1380 /* we cannot replace the exception Proj's here, this is done in
1381 transform_node_Proj_Div() */
1386 } /* equivalent_node_Proj_Div */
1389 * Optimize CopyB(mem, x, x) into a Nop.
1391 static ir_node *equivalent_node_Proj_CopyB(ir_node *proj)
1393 ir_node *oldn = proj;
1394 ir_node *copyb = get_Proj_pred(proj);
1395 ir_node *a = get_CopyB_dst(copyb);
1396 ir_node *b = get_CopyB_src(copyb);
1399 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1400 switch (get_Proj_proj(proj)) {
1402 proj = get_CopyB_mem(copyb);
1403 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1406 case pn_CopyB_X_except: {
1407 ir_graph *irg = get_irn_irg(proj);
1408 DBG_OPT_EXC_REM(proj);
1409 proj = get_irg_bad(irg);
1415 } /* equivalent_node_Proj_CopyB */
1418 * Optimize Bounds(idx, idx, upper) into idx.
1420 static ir_node *equivalent_node_Proj_Bound(ir_node *proj)
1422 ir_node *oldn = proj;
1423 ir_node *bound = get_Proj_pred(proj);
1424 ir_node *idx = get_Bound_index(bound);
1425 ir_node *pred = skip_Proj(idx);
1428 if (idx == get_Bound_lower(bound))
1430 else if (is_Bound(pred)) {
1432 * idx was Bounds checked previously, it is still valid if
1433 * lower <= pred_lower && pred_upper <= upper.
1435 ir_node *lower = get_Bound_lower(bound);
1436 ir_node *upper = get_Bound_upper(bound);
1437 if (get_Bound_lower(pred) == lower &&
1438 get_Bound_upper(pred) == upper) {
1440 * One could expect that we simply return the previous
1441 * Bound here. However, this would be wrong, as we could
1442 * add an exception Proj to a new location then.
1443 * So, we must turn in into a tuple.
1449 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1450 switch (get_Proj_proj(proj)) {
1452 DBG_OPT_EXC_REM(proj);
1453 proj = get_Bound_mem(bound);
1455 case pn_Bound_X_except: {
1456 ir_graph *irg = get_irn_irg(proj);
1457 DBG_OPT_EXC_REM(proj);
1458 proj = get_irg_bad(irg);
1463 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
1466 /* cannot optimize pn_Bound_X_regular, handled in transform ... */
1471 } /* equivalent_node_Proj_Bound */
1474 * Optimize an Exception Proj(Load) with a non-null address.
1476 static ir_node *equivalent_node_Proj_Load(ir_node *proj)
1478 if (get_opt_ldst_only_null_ptr_exceptions()) {
1479 if (get_irn_mode(proj) == mode_X) {
1480 ir_node *load = get_Proj_pred(proj);
1482 /* get the Load address */
1483 const ir_node *addr = get_Load_ptr(load);
1484 const ir_node *confirm;
1486 if (value_not_null(addr, &confirm)) {
1487 if (get_Proj_proj(proj) == pn_Load_X_except) {
1488 ir_graph *irg = get_irn_irg(proj);
1489 DBG_OPT_EXC_REM(proj);
1490 return get_irg_bad(irg);
1496 } /* equivalent_node_Proj_Load */
1499 * Optimize an Exception Proj(Store) with a non-null address.
1501 static ir_node *equivalent_node_Proj_Store(ir_node *proj)
1503 if (get_opt_ldst_only_null_ptr_exceptions()) {
1504 if (get_irn_mode(proj) == mode_X) {
1505 ir_node *store = get_Proj_pred(proj);
1507 /* get the load/store address */
1508 const ir_node *addr = get_Store_ptr(store);
1509 const ir_node *confirm;
1511 if (value_not_null(addr, &confirm)) {
1512 if (get_Proj_proj(proj) == pn_Store_X_except) {
1513 ir_graph *irg = get_irn_irg(proj);
1514 DBG_OPT_EXC_REM(proj);
1515 return get_irg_bad(irg);
1521 } /* equivalent_node_Proj_Store */
1524 * Does all optimizations on nodes that must be done on its Projs
1525 * because of creating new nodes.
1527 static ir_node *equivalent_node_Proj(ir_node *proj)
1529 ir_node *n = get_Proj_pred(proj);
1530 if (n->op->ops.equivalent_node_Proj)
1531 return n->op->ops.equivalent_node_Proj(proj);
1533 } /* equivalent_node_Proj */
1538 static ir_node *equivalent_node_Id(ir_node *n)
1546 DBG_OPT_ID(oldn, n);
1548 } /* equivalent_node_Id */
1553 static ir_node *equivalent_node_Mux(ir_node *n)
1555 ir_node *oldn = n, *sel = get_Mux_sel(n);
1557 ir_tarval *ts = value_of(sel);
1559 /* Mux(true, f, t) == t */
1560 if (ts == tarval_b_true) {
1561 n = get_Mux_true(n);
1562 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1565 /* Mux(false, f, t) == f */
1566 if (ts == tarval_b_false) {
1567 n = get_Mux_false(n);
1568 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1571 n_t = get_Mux_true(n);
1572 n_f = get_Mux_false(n);
1574 /* Mux(v, x, T) == x */
1575 if (is_Unknown(n_f)) {
1577 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1580 /* Mux(v, T, x) == x */
1581 if (is_Unknown(n_t)) {
1583 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1587 /* Mux(v, x, x) == x */
1590 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1593 if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
1594 ir_relation relation = get_Cmp_relation(sel);
1595 ir_node *f = get_Mux_false(n);
1596 ir_node *t = get_Mux_true(n);
1599 * Note further that these optimization work even for floating point
1600 * with NaN's because -NaN == NaN.
1601 * However, if +0 and -0 is handled differently, we cannot use the first one.
1603 ir_node *const cmp_l = get_Cmp_left(sel);
1604 ir_node *const cmp_r = get_Cmp_right(sel);
1607 case ir_relation_equal:
1608 if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */
1609 (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */
1611 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1616 case ir_relation_less_greater:
1617 case ir_relation_unordered_less_greater:
1618 if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */
1619 (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */
1621 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1630 * Note: normalization puts the constant on the right side,
1631 * so we check only one case.
1633 if (cmp_l == t && tarval_is_null(value_of(cmp_r))) {
1634 /* Mux(t CMP 0, X, t) */
1635 if (is_Minus(f) && get_Minus_op(f) == t) {
1636 /* Mux(t CMP 0, -t, t) */
1637 if (relation == ir_relation_equal) {
1638 /* Mux(t == 0, -t, t) ==> -t */
1640 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1641 } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) {
1642 /* Mux(t != 0, -t, t) ==> t */
1644 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1654 * Remove Confirm nodes if setting is on.
1655 * Replace Confirms(x, '=', Constlike) by Constlike.
1657 static ir_node *equivalent_node_Confirm(ir_node *n)
1659 ir_node *pred = get_Confirm_value(n);
1660 ir_relation relation = get_Confirm_relation(n);
1662 while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) {
1664 * rare case: two identical Confirms one after another,
1665 * replace the second one with the first.
1668 pred = get_Confirm_value(n);
1674 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1675 * perform no actual computation, as, e.g., the Id nodes. It does not create
1676 * new nodes. It is therefore safe to free n if the node returned is not n.
1677 * If a node returns a Tuple we can not just skip it. If the size of the
1678 * in array fits, we transform n into a tuple (e.g., Div).
1680 ir_node *equivalent_node(ir_node *n)
1682 if (n->op->ops.equivalent_node)
1683 return n->op->ops.equivalent_node(n);
1685 } /* equivalent_node */
1688 * Sets the default equivalent node operation for an ir_op_ops.
1690 * @param code the opcode for the default operation
1691 * @param ops the operations initialized
1696 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1700 ops->equivalent_node = equivalent_node_##a; \
1702 #define CASE_PROJ(a) \
1704 ops->equivalent_node_Proj = equivalent_node_Proj_##a; \
1742 } /* firm_set_default_equivalent_node */
1745 * Returns non-zero if a node is a Phi node
1746 * with all predecessors constant.
1748 static int is_const_Phi(ir_node *n)
1752 if (! is_Phi(n) || get_irn_arity(n) == 0)
1754 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
1755 if (! is_Const(get_irn_n(n, i)))
1759 } /* is_const_Phi */
1761 typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode);
1762 typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b);
1765 * in reality eval_func should be tarval (*eval_func)() but incomplete
1766 * declarations are bad style and generate noisy warnings
1768 typedef void (*eval_func)(void);
1771 * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter.
1773 static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode)
1775 if (eval == (eval_func) tarval_sub) {
1776 tarval_sub_type func = (tarval_sub_type)eval;
1778 return func(a, b, mode);
1780 tarval_binop_type func = (tarval_binop_type)eval;
1787 * Apply an evaluator on a binop with a constant operators (and one Phi).
1789 * @param phi the Phi node
1790 * @param other the other operand
1791 * @param eval an evaluator function
1792 * @param mode the mode of the result, may be different from the mode of the Phi!
1793 * @param left if non-zero, other is the left operand, else the right
1795 * @return a new Phi node if the conversion was successful, NULL else
1797 static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left)
1803 int i, n = get_irn_arity(phi);
1805 NEW_ARR_A(void *, res, n);
1807 for (i = 0; i < n; ++i) {
1808 pred = get_irn_n(phi, i);
1809 tv = get_Const_tarval(pred);
1810 tv = do_eval(eval, other, tv, mode);
1812 if (tv == tarval_bad) {
1813 /* folding failed, bad */
1819 for (i = 0; i < n; ++i) {
1820 pred = get_irn_n(phi, i);
1821 tv = get_Const_tarval(pred);
1822 tv = do_eval(eval, tv, other, mode);
1824 if (tv == tarval_bad) {
1825 /* folding failed, bad */
1831 irg = get_irn_irg(phi);
1832 for (i = 0; i < n; ++i) {
1833 pred = get_irn_n(phi, i);
1834 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1836 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1837 } /* apply_binop_on_phi */
1840 * Apply an evaluator on a binop with two constant Phi.
1842 * @param a the left Phi node
1843 * @param b the right Phi node
1844 * @param eval an evaluator function
1845 * @param mode the mode of the result, may be different from the mode of the Phi!
1847 * @return a new Phi node if the conversion was successful, NULL else
1849 static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode)
1851 ir_tarval *tv_l, *tv_r, *tv;
1857 if (get_nodes_block(a) != get_nodes_block(b))
1860 n = get_irn_arity(a);
1861 NEW_ARR_A(void *, res, n);
1863 for (i = 0; i < n; ++i) {
1864 pred = get_irn_n(a, i);
1865 tv_l = get_Const_tarval(pred);
1866 pred = get_irn_n(b, i);
1867 tv_r = get_Const_tarval(pred);
1868 tv = do_eval(eval, tv_l, tv_r, mode);
1870 if (tv == tarval_bad) {
1871 /* folding failed, bad */
1876 irg = get_irn_irg(a);
1877 for (i = 0; i < n; ++i) {
1878 pred = get_irn_n(a, i);
1879 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1881 return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode);
1882 } /* apply_binop_on_2_phis */
1885 * Apply an evaluator on a unop with a constant operator (a Phi).
1887 * @param phi the Phi node
1888 * @param eval an evaluator function
1890 * @return a new Phi node if the conversion was successful, NULL else
1892 static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *))
1899 int i, n = get_irn_arity(phi);
1901 NEW_ARR_A(void *, res, n);
1902 for (i = 0; i < n; ++i) {
1903 pred = get_irn_n(phi, i);
1904 tv = get_Const_tarval(pred);
1907 if (tv == tarval_bad) {
1908 /* folding failed, bad */
1913 mode = get_irn_mode(phi);
1914 irg = get_irn_irg(phi);
1915 for (i = 0; i < n; ++i) {
1916 pred = get_irn_n(phi, i);
1917 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1919 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1920 } /* apply_unop_on_phi */
1923 * Apply a conversion on a constant operator (a Phi).
1925 * @param phi the Phi node
1927 * @return a new Phi node if the conversion was successful, NULL else
1929 static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode)
1935 int i, n = get_irn_arity(phi);
1937 NEW_ARR_A(void *, res, n);
1938 for (i = 0; i < n; ++i) {
1939 pred = get_irn_n(phi, i);
1940 tv = get_Const_tarval(pred);
1941 tv = tarval_convert_to(tv, mode);
1943 if (tv == tarval_bad) {
1944 /* folding failed, bad */
1949 irg = get_irn_irg(phi);
1950 for (i = 0; i < n; ++i) {
1951 pred = get_irn_n(phi, i);
1952 res[i] = new_r_Const(irg, (ir_tarval*)res[i]);
1954 return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode);
1955 } /* apply_conv_on_phi */
1958 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1959 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1960 * If possible, remove the Conv's.
1962 static ir_node *transform_node_AddSub(ir_node *n)
1964 ir_mode *mode = get_irn_mode(n);
1966 if (mode_is_reference(mode)) {
1967 ir_node *left = get_binop_left(n);
1968 ir_node *right = get_binop_right(n);
1969 unsigned ref_bits = get_mode_size_bits(mode);
1971 if (is_Conv(left)) {
1972 ir_mode *lmode = get_irn_mode(left);
1973 unsigned bits = get_mode_size_bits(lmode);
1975 if (ref_bits == bits &&
1976 mode_is_int(lmode) &&
1977 get_mode_arithmetic(lmode) == irma_twos_complement) {
1978 ir_node *pre = get_Conv_op(left);
1979 ir_mode *pre_mode = get_irn_mode(pre);
1981 if (mode_is_int(pre_mode) &&
1982 get_mode_size_bits(pre_mode) == bits &&
1983 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1984 /* ok, this conv just changes to sign, moreover the calculation
1985 * is done with same number of bits as our address mode, so
1986 * we can ignore the conv as address calculation can be viewed
1987 * as either signed or unsigned
1989 set_binop_left(n, pre);
1994 if (is_Conv(right)) {
1995 ir_mode *rmode = get_irn_mode(right);
1996 unsigned bits = get_mode_size_bits(rmode);
1998 if (ref_bits == bits &&
1999 mode_is_int(rmode) &&
2000 get_mode_arithmetic(rmode) == irma_twos_complement) {
2001 ir_node *pre = get_Conv_op(right);
2002 ir_mode *pre_mode = get_irn_mode(pre);
2004 if (mode_is_int(pre_mode) &&
2005 get_mode_size_bits(pre_mode) == bits &&
2006 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
2007 /* ok, this conv just changes to sign, moreover the calculation
2008 * is done with same number of bits as our address mode, so
2009 * we can ignore the conv as address calculation can be viewed
2010 * as either signed or unsigned
2012 set_binop_right(n, pre);
2017 /* let address arithmetic use unsigned modes */
2018 if (is_Const(right)) {
2019 ir_mode *rmode = get_irn_mode(right);
2021 if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
2022 /* convert a AddP(P, *s) into AddP(P, *u) */
2023 ir_mode *nm = get_reference_mode_unsigned_eq(mode);
2025 ir_node *pre = new_r_Conv(get_nodes_block(n), right, nm);
2026 set_binop_right(n, pre);
2032 } /* transform_node_AddSub */
2034 #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \
2037 if (is_Const(b) && is_const_Phi(a)) { \
2038 /* check for Op(Phi, Const) */ \
2039 c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\
2041 else if (is_Const(a) && is_const_Phi(b)) { \
2042 /* check for Op(Const, Phi) */ \
2043 c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\
2045 else if (is_const_Phi(a) && is_const_Phi(b)) { \
2046 /* check for Op(Phi, Phi) */ \
2047 c = apply_binop_on_2_phis(a, b, eval, mode); \
2050 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2055 #define HANDLE_UNOP_PHI(eval, a, c) \
2058 if (is_const_Phi(a)) { \
2059 /* check for Op(Phi) */ \
2060 c = apply_unop_on_phi(a, eval); \
2062 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
2069 * Do the AddSub optimization, then Transform
2070 * Constant folding on Phi
2071 * Add(a,a) -> Mul(a, 2)
2072 * Add(Mul(a, x), a) -> Mul(a, x+1)
2073 * if the mode is integer or float.
2074 * Transform Add(a,-b) into Sub(a,b).
2075 * Reassociation might fold this further.
2077 static ir_node *transform_node_Add(ir_node *n)
2080 ir_node *a, *b, *c, *oldn = n;
2081 vrp_attr *a_vrp, *b_vrp;
2083 n = transform_node_AddSub(n);
2085 a = get_Add_left(n);
2086 b = get_Add_right(n);
2088 mode = get_irn_mode(n);
2090 if (mode_is_reference(mode)) {
2091 ir_mode *lmode = get_irn_mode(a);
2093 if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) {
2094 /* an Add(a, NULL) is a hidden Conv */
2095 dbg_info *dbg = get_irn_dbg_info(n);
2096 return new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2100 HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode);
2102 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2103 if (mode_is_float(mode)) {
2104 ir_graph *irg = get_irn_irg(n);
2105 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2109 if (mode_is_num(mode)) {
2110 ir_graph *irg = get_irn_irg(n);
2111 /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */
2112 if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP)
2113 && a == b && mode_is_int(mode)) {
2114 ir_node *block = get_nodes_block(n);
2117 get_irn_dbg_info(n),
2120 new_r_Const_long(irg, mode, 2),
2122 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
2127 get_irn_dbg_info(n),
2132 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2137 get_irn_dbg_info(n),
2142 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
2145 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2146 /* Here we rely on constants be on the RIGHT side */
2148 ir_node *op = get_Not_op(a);
2150 if (is_Const(b) && is_Const_one(b)) {
2152 ir_node *blk = get_nodes_block(n);
2153 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode);
2154 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2159 n = new_r_Const(irg, get_mode_minus_one(mode));
2160 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2165 ir_node *op = get_Not_op(b);
2169 n = new_r_Const(irg, get_mode_minus_one(mode));
2170 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X);
2177 a_vrp = vrp_get_info(a);
2178 b_vrp = vrp_get_info(b);
2180 if (a_vrp && b_vrp) {
2181 ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set);
2183 if (tarval_is_null(c)) {
2184 dbg_info *dbgi = get_irn_dbg_info(n);
2185 return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode);
2189 } /* transform_node_Add */
2192 * returns -cnst or NULL if impossible
2194 static ir_node *const_negate(ir_node *cnst)
2196 ir_tarval *tv = tarval_neg(get_Const_tarval(cnst));
2197 dbg_info *dbgi = get_irn_dbg_info(cnst);
2198 ir_graph *irg = get_irn_irg(cnst);
2199 if (tv == tarval_bad) return NULL;
2200 return new_rd_Const(dbgi, irg, tv);
2204 * Do the AddSub optimization, then Transform
2205 * Constant folding on Phi
2206 * Sub(0,a) -> Minus(a)
2207 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2208 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2209 * Sub(Add(a, x), x) -> a
2210 * Sub(x, Add(x, a)) -> -a
2211 * Sub(x, Const) -> Add(x, -Const)
2213 static ir_node *transform_node_Sub(ir_node *n)
2219 n = transform_node_AddSub(n);
2221 a = get_Sub_left(n);
2222 b = get_Sub_right(n);
2224 mode = get_irn_mode(n);
2226 if (mode_is_int(mode)) {
2227 ir_mode *lmode = get_irn_mode(a);
2229 if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) {
2230 /* a Sub(a, NULL) is a hidden Conv */
2231 dbg_info *dbg = get_irn_dbg_info(n);
2232 n = new_rd_Conv(dbg, get_nodes_block(n), a, mode);
2233 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV);
2237 if (mode == lmode &&
2238 get_mode_arithmetic(mode) == irma_twos_complement &&
2240 get_Const_tarval(a) == get_mode_minus_one(mode)) {
2242 dbg_info *dbg = get_irn_dbg_info(n);
2243 n = new_rd_Not(dbg, get_nodes_block(n), b, mode);
2244 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT);
2250 HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode);
2252 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2253 if (mode_is_float(mode)) {
2254 ir_graph *irg = get_irn_irg(n);
2255 if (get_irg_fp_model(irg) & fp_strict_algebraic)
2259 if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) {
2260 /* a - C -> a + (-C) */
2261 ir_node *cnst = const_negate(b);
2263 ir_node *block = get_nodes_block(n);
2264 dbg_info *dbgi = get_irn_dbg_info(n);
2266 n = new_rd_Add(dbgi, block, a, cnst, mode);
2267 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2272 if (is_Minus(a)) { /* (-a) - b -> -(a + b) */
2273 dbg_info *dbg = get_irn_dbg_info(n);
2274 ir_node *block = get_nodes_block(n);
2275 ir_node *left = get_Minus_op(a);
2276 ir_node *add = new_rd_Add(dbg, block, left, b, mode);
2278 n = new_rd_Minus(dbg, block, add, mode);
2279 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2281 } else if (is_Minus(b)) { /* a - (-b) -> a + b */
2282 dbg_info *dbg = get_irn_dbg_info(n);
2283 ir_node *block = get_nodes_block(n);
2284 ir_node *right = get_Minus_op(b);
2286 n = new_rd_Add(dbg, block, a, right, mode);
2287 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS);
2289 } else if (is_Sub(b)) {
2290 /* a - (b - c) -> a + (c - b)
2291 * -> (a - b) + c iff (b - c) is a pointer */
2292 dbg_info *s_dbg = get_irn_dbg_info(b);
2293 ir_node *s_left = get_Sub_left(b);
2294 ir_node *s_right = get_Sub_right(b);
2295 ir_mode *s_mode = get_irn_mode(b);
2296 if (mode_is_reference(s_mode)) {
2297 ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */
2298 ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode);
2299 dbg_info *a_dbg = get_irn_dbg_info(n);
2302 s_right = new_r_Conv(lowest_block, s_right, mode);
2303 n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode);
2305 ir_node *s_block = get_nodes_block(b);
2306 ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode);
2307 dbg_info *a_dbg = get_irn_dbg_info(n);
2308 ir_node *a_block = get_nodes_block(n);
2310 n = new_rd_Add(a_dbg, a_block, a, sub, mode);
2312 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2314 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
2315 ir_node *m_right = get_Mul_right(b);
2316 if (is_Const(m_right)) {
2317 ir_node *cnst2 = const_negate(m_right);
2318 if (cnst2 != NULL) {
2319 dbg_info *m_dbg = get_irn_dbg_info(b);
2320 ir_node *m_block = get_nodes_block(b);
2321 ir_node *m_left = get_Mul_left(b);
2322 ir_mode *m_mode = get_irn_mode(b);
2323 ir_node *mul = new_rd_Mul(m_dbg, m_block, m_left, cnst2, m_mode);
2324 dbg_info *a_dbg = get_irn_dbg_info(n);
2325 ir_node *a_block = get_nodes_block(n);
2327 n = new_rd_Add(a_dbg, a_block, a, mul, mode);
2328 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
2334 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2335 if (mode_is_num(mode) && mode == get_irn_mode(a) && is_Const(a) && is_Const_null(a)) {
2337 get_irn_dbg_info(n),
2341 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2345 if (mode_wrap_around(mode)) {
2346 ir_node *left = get_Add_left(a);
2347 ir_node *right = get_Add_right(a);
2349 /* FIXME: Does the Conv's work only for two complement or generally? */
2351 if (mode != get_irn_mode(right)) {
2352 /* This Sub is an effective Cast */
2353 right = new_r_Conv(get_nodes_block(n), right, mode);
2356 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2358 } else if (right == b) {
2359 if (mode != get_irn_mode(left)) {
2360 /* This Sub is an effective Cast */
2361 left = new_r_Conv(get_nodes_block(n), left, mode);
2364 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2370 if (mode_wrap_around(mode)) {
2371 ir_node *left = get_Add_left(b);
2372 ir_node *right = get_Add_right(b);
2374 /* FIXME: Does the Conv's work only for two complement or generally? */
2376 ir_mode *r_mode = get_irn_mode(right);
2378 n = new_r_Minus(get_nodes_block(n), right, r_mode);
2379 if (mode != r_mode) {
2380 /* This Sub is an effective Cast */
2381 n = new_r_Conv(get_nodes_block(n), n, mode);
2383 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2385 } else if (right == a) {
2386 ir_mode *l_mode = get_irn_mode(left);
2388 n = new_r_Minus(get_nodes_block(n), left, l_mode);
2389 if (mode != l_mode) {
2390 /* This Sub is an effective Cast */
2391 n = new_r_Conv(get_nodes_block(n), n, mode);
2393 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2398 if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2399 ir_mode *mode = get_irn_mode(a);
2401 if (mode == get_irn_mode(b)) {
2403 ir_node *op_a = get_Conv_op(a);
2404 ir_node *op_b = get_Conv_op(b);
2406 /* check if it's allowed to skip the conv */
2407 ma = get_irn_mode(op_a);
2408 mb = get_irn_mode(op_b);
2410 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2411 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2414 set_Sub_right(n, b);
2420 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2421 if (!is_reassoc_running() && is_Mul(a)) {
2422 ir_node *ma = get_Mul_left(a);
2423 ir_node *mb = get_Mul_right(a);
2426 ir_node *blk = get_nodes_block(n);
2427 ir_graph *irg = get_irn_irg(n);
2429 get_irn_dbg_info(n),
2433 get_irn_dbg_info(n),
2436 new_r_Const(irg, get_mode_one(mode)),
2439 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2441 } else if (mb == b) {
2442 ir_node *blk = get_nodes_block(n);
2443 ir_graph *irg = get_irn_irg(n);
2445 get_irn_dbg_info(n),
2449 get_irn_dbg_info(n),
2452 new_r_Const(irg, get_mode_one(mode)),
2455 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2459 if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */
2460 ir_node *x = get_Sub_left(a);
2461 ir_node *y = get_Sub_right(a);
2462 ir_node *blk = get_nodes_block(n);
2463 ir_mode *m_b = get_irn_mode(b);
2464 ir_mode *m_y = get_irn_mode(y);
2468 /* Determine the right mode for the Add. */
2471 else if (mode_is_reference(m_b))
2473 else if (mode_is_reference(m_y))
2477 * Both modes are different but none is reference,
2478 * happens for instance in SubP(SubP(P, Iu), Is).
2479 * We have two possibilities here: Cast or ignore.
2480 * Currently we ignore this case.
2485 add = new_r_Add(blk, y, b, add_mode);
2487 n = new_rd_Sub(get_irn_dbg_info(n), blk, x, add, mode);
2488 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z);
2492 if (get_mode_arithmetic(mode) == irma_twos_complement) {
2493 if (is_Const(a) && is_Not(b)) {
2494 /* c - ~X = X + (c+1) */
2495 ir_tarval *tv = get_Const_tarval(a);
2497 tv = tarval_add(tv, get_mode_one(mode));
2498 if (tv != tarval_bad) {
2499 ir_node *blk = get_nodes_block(n);
2500 ir_graph *irg = get_irn_irg(n);
2501 ir_node *c = new_r_Const(irg, tv);
2502 n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode);
2503 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X);
2509 } /* transform_node_Sub */
2512 * Several transformation done on n*n=2n bits mul.
2513 * These transformations must be done here because new nodes may be produced.
2515 static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode)
2518 ir_node *a = get_Mul_left(n);
2519 ir_node *b = get_Mul_right(n);
2520 ir_tarval *ta = value_of(a);
2521 ir_tarval *tb = value_of(b);
2522 ir_mode *smode = get_irn_mode(a);
2524 if (ta == get_mode_one(smode)) {
2525 /* (L)1 * (L)b = (L)b */
2526 ir_node *blk = get_nodes_block(n);
2527 n = new_rd_Conv(get_irn_dbg_info(n), blk, b, mode);
2528 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2531 else if (ta == get_mode_minus_one(smode)) {
2532 /* (L)-1 * (L)b = (L)b */
2533 ir_node *blk = get_nodes_block(n);
2534 n = new_rd_Minus(get_irn_dbg_info(n), blk, b, smode);
2535 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2536 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2539 if (tb == get_mode_one(smode)) {
2540 /* (L)a * (L)1 = (L)a */
2541 ir_node *blk = get_irn_n(a, -1);
2542 n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode);
2543 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
2546 else if (tb == get_mode_minus_one(smode)) {
2547 /* (L)a * (L)-1 = (L)-a */
2548 ir_node *blk = get_nodes_block(n);
2549 n = new_rd_Minus(get_irn_dbg_info(n), blk, a, smode);
2550 n = new_rd_Conv(get_irn_dbg_info(n), blk, n, mode);
2551 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2558 * Transform Mul(a,-1) into -a.
2559 * Do constant evaluation of Phi nodes.
2560 * Do architecture dependent optimizations on Mul nodes
2562 static ir_node *transform_node_Mul(ir_node *n)
2564 ir_node *c, *oldn = n;
2565 ir_mode *mode = get_irn_mode(n);
2566 ir_node *a = get_Mul_left(n);
2567 ir_node *b = get_Mul_right(n);
2569 if (is_Bad(a) || is_Bad(b))
2572 if (mode != get_irn_mode(a))
2573 return transform_node_Mul2n(n, mode);
2575 HANDLE_BINOP_PHI((eval_func) tarval_mul, a, b, c, mode);
2577 if (mode_is_signed(mode)) {
2580 if (value_of(a) == get_mode_minus_one(mode))
2582 else if (value_of(b) == get_mode_minus_one(mode))
2585 n = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), r, mode);
2586 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2591 if (is_Const(b)) { /* (-a) * const -> a * -const */
2592 ir_node *cnst = const_negate(b);
2594 dbg_info *dbgi = get_irn_dbg_info(n);
2595 ir_node *block = get_nodes_block(n);
2596 n = new_rd_Mul(dbgi, block, get_Minus_op(a), cnst, mode);
2597 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2600 } else if (is_Minus(b)) { /* (-a) * (-b) -> a * b */
2601 dbg_info *dbgi = get_irn_dbg_info(n);
2602 ir_node *block = get_nodes_block(n);
2603 n = new_rd_Mul(dbgi, block, get_Minus_op(a), get_Minus_op(b), mode);
2604 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_MINUS);
2606 } else if (is_Sub(b)) { /* (-a) * (b - c) -> a * (c - b) */
2607 ir_node *sub_l = get_Sub_left(b);
2608 ir_node *sub_r = get_Sub_right(b);
2609 dbg_info *dbgi = get_irn_dbg_info(n);
2610 ir_node *block = get_nodes_block(n);
2611 ir_node *new_b = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2612 n = new_rd_Mul(dbgi, block, get_Minus_op(a), new_b, mode);
2613 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2616 } else if (is_Minus(b)) {
2617 if (is_Sub(a)) { /* (a - b) * (-c) -> (b - a) * c */
2618 ir_node *sub_l = get_Sub_left(a);
2619 ir_node *sub_r = get_Sub_right(a);
2620 dbg_info *dbgi = get_irn_dbg_info(n);
2621 ir_node *block = get_nodes_block(n);
2622 ir_node *new_a = new_rd_Sub(dbgi, block, sub_r, sub_l, mode);
2623 n = new_rd_Mul(dbgi, block, new_a, get_Minus_op(b), mode);
2624 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS);
2627 } else if (is_Shl(a)) {
2628 ir_node *const shl_l = get_Shl_left(a);
2629 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2630 /* (1 << x) * b -> b << x */
2631 dbg_info *const dbgi = get_irn_dbg_info(n);
2632 ir_node *const block = get_nodes_block(n);
2633 ir_node *const shl_r = get_Shl_right(a);
2634 n = new_rd_Shl(dbgi, block, b, shl_r, mode);
2635 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2638 } else if (is_Shl(b)) {
2639 ir_node *const shl_l = get_Shl_left(b);
2640 if (is_Const(shl_l) && is_Const_one(shl_l)) {
2641 /* a * (1 << x) -> a << x */
2642 dbg_info *const dbgi = get_irn_dbg_info(n);
2643 ir_node *const block = get_nodes_block(n);
2644 ir_node *const shl_r = get_Shl_right(b);
2645 n = new_rd_Shl(dbgi, block, a, shl_r, mode);
2646 // TODO add me DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_SHIFT);
2650 if (get_mode_arithmetic(mode) == irma_ieee754) {
2652 ir_tarval *tv = get_Const_tarval(a);
2653 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2654 && !tarval_is_negative(tv)) {
2655 /* 2.0 * b = b + b */
2656 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode);
2657 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2661 else if (is_Const(b)) {
2662 ir_tarval *tv = get_Const_tarval(b);
2663 if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)
2664 && !tarval_is_negative(tv)) {
2665 /* a * 2.0 = a + a */
2666 n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode);
2667 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A);
2672 return arch_dep_replace_mul_with_shifts(n);
2673 } /* transform_node_Mul */
2676 * Transform a Div Node.
2678 static ir_node *transform_node_Div(ir_node *n)
2680 ir_mode *mode = get_Div_resmode(n);
2681 ir_node *a = get_Div_left(n);
2682 ir_node *b = get_Div_right(n);
2684 const ir_node *dummy;
2686 if (mode_is_int(mode)) {
2687 if (is_Const(b) && is_const_Phi(a)) {
2688 /* check for Div(Phi, Const) */
2689 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0);
2691 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2694 } else if (is_Const(a) && is_const_Phi(b)) {
2695 /* check for Div(Const, Phi) */
2696 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1);
2698 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2701 } else if (is_const_Phi(a) && is_const_Phi(b)) {
2702 /* check for Div(Phi, Phi) */
2703 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode);
2705 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2710 if (a == b && value_not_zero(a, &dummy)) {
2711 ir_graph *irg = get_irn_irg(n);
2712 /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */
2713 value = new_r_Const(irg, get_mode_one(mode));
2714 DBG_OPT_CSTEVAL(n, value);
2717 if (mode_is_signed(mode) && is_Const(b)) {
2718 ir_tarval *tv = get_Const_tarval(b);
2720 if (tv == get_mode_minus_one(mode)) {
2722 value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode);
2723 DBG_OPT_CSTEVAL(n, value);
2727 /* Try architecture dependent optimization */
2728 value = arch_dep_replace_div_by_const(n);
2731 assert(mode_is_float(mode));
2733 /* Optimize x/c to x*(1/c) */
2734 if (get_mode_arithmetic(mode) == irma_ieee754) {
2735 ir_tarval *tv = value_of(b);
2737 if (tv != tarval_bad) {
2738 int rem = tarval_fp_ops_enabled();
2741 * Floating point constant folding might be disabled here to
2743 * However, as we check for exact result, doing it is safe.
2746 tarval_enable_fp_ops(1);
2747 tv = tarval_div(get_mode_one(mode), tv);
2748 tarval_enable_fp_ops(rem);
2750 /* Do the transformation if the result is either exact or we are
2751 not using strict rules. */
2752 if (tv != tarval_bad &&
2753 (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) {
2754 ir_node *block = get_nodes_block(n);
2755 ir_graph *irg = get_irn_irg(block);
2756 ir_node *c = new_r_Const(irg, tv);
2757 dbg_info *dbgi = get_irn_dbg_info(n);
2758 value = new_rd_Mul(dbgi, block, a, c, mode);
2771 /* Turn Div into a tuple (mem, jmp, bad, value) */
2772 mem = get_Div_mem(n);
2773 blk = get_nodes_block(n);
2774 irg = get_irn_irg(blk);
2776 /* skip a potential Pin */
2777 mem = skip_Pin(mem);
2778 turn_into_tuple(n, pn_Div_max);
2779 set_Tuple_pred(n, pn_Div_M, mem);
2780 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
2781 set_Tuple_pred(n, pn_Div_X_except, get_irg_bad(irg));
2782 set_Tuple_pred(n, pn_Div_res, value);
2785 } /* transform_node_Div */
2788 * Transform a Mod node.
2790 static ir_node *transform_node_Mod(ir_node *n)
2792 ir_mode *mode = get_Mod_resmode(n);
2793 ir_node *a = get_Mod_left(n);
2794 ir_node *b = get_Mod_right(n);
2799 if (is_Const(b) && is_const_Phi(a)) {
2800 /* check for Div(Phi, Const) */
2801 value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0);
2803 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2807 else if (is_Const(a) && is_const_Phi(b)) {
2808 /* check for Div(Const, Phi) */
2809 value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1);
2811 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2815 else if (is_const_Phi(a) && is_const_Phi(b)) {
2816 /* check for Div(Phi, Phi) */
2817 value = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode);
2819 DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI);
2826 irg = get_irn_irg(n);
2827 if (tv != tarval_bad) {
2828 value = new_r_Const(irg, tv);
2830 DBG_OPT_CSTEVAL(n, value);
2833 ir_node *a = get_Mod_left(n);
2834 ir_node *b = get_Mod_right(n);
2835 const ir_node *dummy;
2837 if (a == b && value_not_zero(a, &dummy)) {
2838 /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */
2839 value = new_r_Const(irg, get_mode_null(mode));
2840 DBG_OPT_CSTEVAL(n, value);
2843 if (mode_is_signed(mode) && is_Const(b)) {
2844 ir_tarval *tv = get_Const_tarval(b);
2846 if (tv == get_mode_minus_one(mode)) {
2848 value = new_r_Const(irg, get_mode_null(mode));
2849 DBG_OPT_CSTEVAL(n, value);
2853 /* Try architecture dependent optimization */
2854 value = arch_dep_replace_mod_by_const(n);
2863 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2864 mem = get_Mod_mem(n);
2865 blk = get_nodes_block(n);
2866 irg = get_irn_irg(blk);
2868 /* skip a potential Pin */
2869 mem = skip_Pin(mem);
2870 turn_into_tuple(n, pn_Mod_max);
2871 set_Tuple_pred(n, pn_Mod_M, mem);
2872 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
2873 set_Tuple_pred(n, pn_Mod_X_except, get_irg_bad(irg));
2874 set_Tuple_pred(n, pn_Mod_res, value);
2877 } /* transform_node_Mod */
2880 * Transform a Cond node.
2882 * Replace the Cond by a Jmp if it branches on a constant
2885 static ir_node *transform_node_Cond(ir_node *n)
2888 ir_node *a = get_Cond_selector(n);
2889 ir_tarval *ta = value_of(a);
2890 ir_graph *irg = get_irn_irg(n);
2893 /* we need block info which is not available in floating irgs */
2894 if (get_irg_pinned(irg) == op_pin_state_floats)
2897 if ((ta != tarval_bad) &&
2898 (get_irn_mode(a) == mode_b) &&
2899 (get_opt_unreachable_code())) {
2900 /* It's a boolean Cond, branching on a boolean constant.
2901 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2902 ir_node *blk = get_nodes_block(n);
2903 jmp = new_r_Jmp(blk);
2904 turn_into_tuple(n, pn_Cond_max);
2905 if (ta == tarval_b_true) {
2906 set_Tuple_pred(n, pn_Cond_false, get_irg_bad(irg));
2907 set_Tuple_pred(n, pn_Cond_true, jmp);
2909 set_Tuple_pred(n, pn_Cond_false, jmp);
2910 set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg));
2912 /* We might generate an endless loop, so keep it alive. */
2913 add_End_keepalive(get_irg_end(irg), blk);
2916 } /* transform_node_Cond */
2919 * Prototype of a recursive transform function
2920 * for bitwise distributive transformations.
2922 typedef ir_node* (*recursive_transform)(ir_node *n);
2925 * makes use of distributive laws for and, or, eor
2926 * and(a OP c, b OP c) -> and(a, b) OP c
2927 * note, might return a different op than n
2929 static ir_node *transform_bitwise_distributive(ir_node *n,
2930 recursive_transform trans_func)
2933 ir_node *a = get_binop_left(n);
2934 ir_node *b = get_binop_right(n);
2935 ir_op *op = get_irn_op(a);
2936 ir_op *op_root = get_irn_op(n);
2938 if (op != get_irn_op(b))
2941 /* and(conv(a), conv(b)) -> conv(and(a,b)) */
2942 if (op == op_Conv) {
2943 ir_node *a_op = get_Conv_op(a);
2944 ir_node *b_op = get_Conv_op(b);
2945 ir_mode *a_mode = get_irn_mode(a_op);
2946 ir_mode *b_mode = get_irn_mode(b_op);
2947 if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2948 ir_node *blk = get_nodes_block(n);
2951 set_binop_left(n, a_op);
2952 set_binop_right(n, b_op);
2953 set_irn_mode(n, a_mode);
2955 n = new_r_Conv(blk, n, get_irn_mode(oldn));
2957 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
2963 /* nothing to gain here */
2967 if (op == op_Shrs || op == op_Shr || op == op_Shl
2968 || op == op_And || op == op_Or || op == op_Eor) {
2969 ir_node *a_left = get_binop_left(a);
2970 ir_node *a_right = get_binop_right(a);
2971 ir_node *b_left = get_binop_left(b);
2972 ir_node *b_right = get_binop_right(b);
2974 ir_node *op1 = NULL;
2975 ir_node *op2 = NULL;
2977 if (is_op_commutative(op)) {
2978 if (a_left == b_left) {
2982 } else if (a_left == b_right) {
2986 } else if (a_right == b_left) {
2992 if (a_right == b_right) {
2999 /* (a sop c) & (b sop c) => (a & b) sop c */
3000 ir_node *blk = get_nodes_block(n);
3002 ir_node *new_n = exact_copy(n);
3003 set_binop_left(new_n, op1);
3004 set_binop_right(new_n, op2);
3005 new_n = trans_func(new_n);
3007 if (op_root == op_Eor && op == op_Or) {
3008 dbg_info *dbgi = get_irn_dbg_info(n);
3009 ir_mode *mode = get_irn_mode(c);
3011 c = new_rd_Not(dbgi, blk, c, mode);
3012 n = new_rd_And(dbgi, blk, new_n, c, mode);
3015 set_nodes_block(n, blk);
3016 set_binop_left(n, new_n);
3017 set_binop_right(n, c);
3021 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
3030 * Create a 0 constant of given mode.
3032 static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode)
3034 ir_tarval *tv = get_mode_null(mode);
3035 ir_node *cnst = new_r_Const(irg, tv);
3043 static ir_node *transform_node_And(ir_node *n)
3045 ir_node *c, *oldn = n;
3046 ir_node *a = get_And_left(n);
3047 ir_node *b = get_And_right(n);
3049 vrp_attr *a_vrp, *b_vrp;
3051 if (is_Cmp(a) && is_Cmp(b)) {
3052 ir_node *a_left = get_Cmp_left(a);
3053 ir_node *a_right = get_Cmp_right(a);
3054 ir_node *b_left = get_Cmp_left(b);
3055 ir_node *b_right = get_Cmp_right(b);
3056 ir_relation a_relation = get_Cmp_relation(a);
3057 ir_relation b_relation = get_Cmp_relation(b);
3058 /* we can combine the relations of two compares with the same
3060 if (a_left == b_left && b_left == b_right) {
3061 dbg_info *dbgi = get_irn_dbg_info(n);
3062 ir_node *block = get_nodes_block(n);
3063 ir_relation new_relation = a_relation & b_relation;
3064 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3066 /* Cmp(a==0) and Cmp(b==0) can be optimized to Cmp(a|b==0) */
3067 if (is_Const(a_right) && is_Const_null(a_right)
3068 && is_Const(b_right) && is_Const_null(b_right)
3069 && a_relation == b_relation && a_relation == ir_relation_equal
3070 && !mode_is_float(get_irn_mode(a_left))
3071 && !mode_is_float(get_irn_mode(b_left))) {
3072 dbg_info *dbgi = get_irn_dbg_info(n);
3073 ir_node *block = get_nodes_block(n);
3074 ir_mode *mode = get_irn_mode(a_left);
3075 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
3076 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
3077 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
3078 ir_graph *irg = get_irn_irg(n);
3079 ir_node *zero = create_zero_const(irg, mode);
3080 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal);
3084 mode = get_irn_mode(n);
3085 HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode);
3089 ir_node *op = get_Not_op(b);
3091 ir_node *ba = get_And_left(op);
3092 ir_node *bb = get_And_right(op);
3094 /* it's enough to test the following cases due to normalization! */
3095 if (get_Or_left(a) == ba && get_Or_right(a) == bb) {
3096 /* (a|b) & ~(a&b) = a^b */
3097 ir_node *block = get_nodes_block(n);
3099 n = new_rd_Eor(get_irn_dbg_info(n), block, ba, bb, mode);
3100 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3108 ir_node *op = get_Not_op(a);
3110 ir_node *aa = get_And_left(op);
3111 ir_node *ab = get_And_right(op);
3113 /* it's enough to test the following cases due to normalization! */
3114 if (get_Or_left(b) == aa && get_Or_right(b) == ab) {
3115 /* (a|b) & ~(a&b) = a^b */
3116 ir_node *block = get_nodes_block(n);
3118 n = new_rd_Eor(get_irn_dbg_info(n), block, aa, ab, mode);
3119 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_TO_EOR);
3126 ir_node *al = get_Eor_left(a);
3127 ir_node *ar = get_Eor_right(a);
3130 /* (b ^ a) & b -> ~a & b */
3131 dbg_info *dbg = get_irn_dbg_info(n);
3132 ir_node *block = get_nodes_block(n);
3134 ar = new_rd_Not(dbg, block, ar, mode);
3135 n = new_rd_And(dbg, block, ar, b, mode);
3136 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3140 /* (a ^ b) & b -> ~a & b */
3141 dbg_info *dbg = get_irn_dbg_info(n);
3142 ir_node *block = get_nodes_block(n);
3144 al = new_rd_Not(dbg, block, al, mode);
3145 n = new_rd_And(dbg, block, al, b, mode);
3146 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3151 ir_node *bl = get_Eor_left(b);
3152 ir_node *br = get_Eor_right(b);
3155 /* a & (a ^ b) -> a & ~b */
3156 dbg_info *dbg = get_irn_dbg_info(n);
3157 ir_node *block = get_nodes_block(n);
3159 br = new_rd_Not(dbg, block, br, mode);
3160 n = new_rd_And(dbg, block, br, a, mode);
3161 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3165 /* a & (b ^ a) -> a & ~b */
3166 dbg_info *dbg = get_irn_dbg_info(n);
3167 ir_node *block = get_nodes_block(n);
3169 bl = new_rd_Not(dbg, block, bl, mode);
3170 n = new_rd_And(dbg, block, bl, a, mode);
3171 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3175 if (is_Not(a) && is_Not(b)) {
3176 /* ~a & ~b = ~(a|b) */
3177 ir_node *block = get_nodes_block(n);
3178 ir_mode *mode = get_irn_mode(n);
3182 n = new_rd_Or(get_irn_dbg_info(n), block, a, b, mode);
3183 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
3184 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
3188 b_vrp = vrp_get_info(b);
3189 if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a),
3190 b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) {
3196 a_vrp = vrp_get_info(a);
3197 if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b),
3198 a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) {
3202 n = transform_bitwise_distributive(n, transform_node_And);
3205 } /* transform_node_And */
3207 /* the order of the values is important! */
3208 typedef enum const_class {
3214 static const_class classify_const(const ir_node* n)
3216 if (is_Const(n)) return const_const;
3217 if (is_irn_constlike(n)) return const_like;
3222 * Determines whether r is more constlike or has a larger index (in that order)
3225 static bool operands_are_normalized(const ir_node *l, const ir_node *r)
3227 const const_class l_order = classify_const(l);
3228 const const_class r_order = classify_const(r);
3230 l_order > r_order ||
3231 (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r));
3237 static ir_node *transform_node_Eor(ir_node *n)
3239 ir_node *c, *oldn = n;
3240 ir_node *a = get_Eor_left(n);
3241 ir_node *b = get_Eor_right(n);
3242 ir_mode *mode = get_irn_mode(n);
3244 /* we can combine the relations of two compares with the same operands */
3245 if (is_Cmp(a) && is_Cmp(b)) {
3246 ir_node *a_left = get_Cmp_left(a);
3247 ir_node *a_right = get_Cmp_left(a);
3248 ir_node *b_left = get_Cmp_left(b);
3249 ir_node *b_right = get_Cmp_right(b);
3250 if (a_left == b_left && b_left == b_right) {
3251 dbg_info *dbgi = get_irn_dbg_info(n);
3252 ir_node *block = get_nodes_block(n);
3253 ir_relation a_relation = get_Cmp_relation(a);
3254 ir_relation b_relation = get_Cmp_relation(b);
3255 ir_relation new_relation = a_relation ^ b_relation;
3256 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
3260 HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode);
3262 /* normalize not nodes... ~a ^ b <=> a ^ ~b */
3263 if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) {
3264 dbg_info *dbg = get_irn_dbg_info(n);
3265 ir_node *block = get_nodes_block(n);
3266 ir_node *new_not = new_rd_Not(dbg, block, b, mode);
3267 ir_node *new_left = get_Not_op(a);
3268 n = new_rd_Eor(dbg, block, new_left, new_not, mode);
3269 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3271 } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) {
3272 dbg_info *dbg = get_irn_dbg_info(n);
3273 ir_node *block = get_nodes_block(n);
3274 ir_node *new_not = new_rd_Not(dbg, block, a, mode);
3275 ir_node *new_right = get_Not_op(b);
3276 n = new_rd_Eor(dbg, block, new_not, new_right, mode);
3277 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3281 /* x ^ 1...1 -> ~1 */
3282 if (is_Const(b) && is_Const_all_one(b)) {
3283 n = new_r_Not(get_nodes_block(n), a, mode);
3284 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
3288 n = transform_bitwise_distributive(n, transform_node_Eor);
3290 } /* transform_node_Eor */
3295 static ir_node *transform_node_Not(ir_node *n)
3297 ir_node *c, *oldn = n;
3298 ir_node *a = get_Not_op(n);
3299 ir_mode *mode = get_irn_mode(n);
3301 HANDLE_UNOP_PHI(tarval_not,a,c);
3303 /* check for a boolean Not */
3305 dbg_info *dbgi = get_irn_dbg_info(a);
3306 ir_node *block = get_nodes_block(a);
3307 ir_relation relation = get_Cmp_relation(a);
3308 relation = get_negated_relation(relation);
3309 n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation);
3310 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
3314 /* normalize ~(a ^ b) => a ^ ~b */
3316 dbg_info *dbg = get_irn_dbg_info(n);
3317 ir_node *block = get_nodes_block(n);
3318 ir_node *eor_right = get_Eor_right(a);
3319 ir_node *eor_left = get_Eor_left(a);
3320 eor_right = new_rd_Not(dbg, block, eor_right, mode);
3321 n = new_rd_Eor(dbg, block, eor_left, eor_right, mode);
3325 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3326 if (is_Minus(a)) { /* ~-x -> x + -1 */
3327 dbg_info *dbg = get_irn_dbg_info(n);
3328 ir_graph *irg = get_irn_irg(n);
3329 ir_node *block = get_nodes_block(n);
3330 ir_node *add_l = get_Minus_op(a);
3331 ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode));
3332 n = new_rd_Add(dbg, block, add_l, add_r, mode);
3333 } else if (is_Add(a)) {
3334 ir_node *add_r = get_Add_right(a);
3335 if (is_Const(add_r) && is_Const_all_one(add_r)) {
3336 /* ~(x + -1) = -x */
3337 ir_node *op = get_Add_left(a);
3338 ir_node *blk = get_nodes_block(n);
3339 n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n));
3340 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
3348 * Transform a Minus.
3352 * -(a >>u (size-1)) = a >>s (size-1)
3353 * -(a >>s (size-1)) = a >>u (size-1)
3354 * -(a * const) -> a * -const
3356 static ir_node *transform_node_Minus(ir_node *n)
3358 ir_node *c, *oldn = n;
3359 ir_node *a = get_Minus_op(n);
3362 HANDLE_UNOP_PHI(tarval_neg,a,c);
3364 mode = get_irn_mode(a);
3365 if (get_mode_arithmetic(mode) == irma_twos_complement) {
3366 /* the following rules are only to twos-complement */
3369 ir_node *op = get_Not_op(a);
3370 ir_tarval *tv = get_mode_one(mode);
3371 ir_node *blk = get_nodes_block(n);
3372 ir_graph *irg = get_irn_irg(blk);
3373 ir_node *c = new_r_Const(irg, tv);
3374 n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode);
3375 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
3379 ir_node *c = get_Shr_right(a);
3382 ir_tarval *tv = get_Const_tarval(c);
3384 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3385 /* -(a >>u (size-1)) = a >>s (size-1) */
3386 ir_node *v = get_Shr_left(a);
3388 n = new_rd_Shrs(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3389 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3395 ir_node *c = get_Shrs_right(a);
3398 ir_tarval *tv = get_Const_tarval(c);
3400 if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) {
3401 /* -(a >>s (size-1)) = a >>u (size-1) */
3402 ir_node *v = get_Shrs_left(a);
3404 n = new_rd_Shr(get_irn_dbg_info(n), get_nodes_block(n), v, c, mode);
3405 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE);
3412 /* - (a-b) = b - a */
3413 ir_node *la = get_Sub_left(a);
3414 ir_node *ra = get_Sub_right(a);
3415 ir_node *blk = get_nodes_block(n);
3417 n = new_rd_Sub(get_irn_dbg_info(n), blk, ra, la, mode);
3418 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB);
3422 if (is_Mul(a)) { /* -(a * const) -> a * -const */
3423 ir_node *mul_l = get_Mul_left(a);
3424 ir_node *mul_r = get_Mul_right(a);
3425 ir_tarval *tv = value_of(mul_r);
3426 if (tv != tarval_bad) {
3427 tv = tarval_neg(tv);
3428 if (tv != tarval_bad) {
3429 ir_graph *irg = get_irn_irg(n);
3430 ir_node *cnst = new_r_Const(irg, tv);
3431 dbg_info *dbg = get_irn_dbg_info(a);
3432 ir_node *block = get_nodes_block(a);
3433 n = new_rd_Mul(dbg, block, mul_l, cnst, mode);
3434 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
3441 } /* transform_node_Minus */
3444 * Transform a Proj(Load) with a non-null address.
3446 static ir_node *transform_node_Proj_Load(ir_node *proj)
3448 if (get_opt_ldst_only_null_ptr_exceptions()) {
3449 if (get_irn_mode(proj) == mode_X) {
3450 ir_node *load = get_Proj_pred(proj);
3452 /* get the Load address */
3453 const ir_node *addr = get_Load_ptr(load);
3454 const ir_node *confirm;
3456 if (value_not_null(addr, &confirm)) {
3457 if (confirm == NULL) {
3458 /* this node may float if it did not depend on a Confirm */
3459 set_irn_pinned(load, op_pin_state_floats);
3461 if (get_Proj_proj(proj) == pn_Load_X_except) {
3462 ir_graph *irg = get_irn_irg(proj);
3463 DBG_OPT_EXC_REM(proj);
3464 return get_irg_bad(irg);
3466 ir_node *blk = get_nodes_block(load);
3467 return new_r_Jmp(blk);
3473 } /* transform_node_Proj_Load */
3476 * Transform a Proj(Store) with a non-null address.
3478 static ir_node *transform_node_Proj_Store(ir_node *proj)
3480 if (get_opt_ldst_only_null_ptr_exceptions()) {
3481 if (get_irn_mode(proj) == mode_X) {
3482 ir_node *store = get_Proj_pred(proj);
3484 /* get the load/store address */
3485 const ir_node *addr = get_Store_ptr(store);
3486 const ir_node *confirm;
3488 if (value_not_null(addr, &confirm)) {
3489 if (confirm == NULL) {
3490 /* this node may float if it did not depend on a Confirm */
3491 set_irn_pinned(store, op_pin_state_floats);
3493 if (get_Proj_proj(proj) == pn_Store_X_except) {
3494 ir_graph *irg = get_irn_irg(proj);
3495 DBG_OPT_EXC_REM(proj);
3496 return get_irg_bad(irg);
3498 ir_node *blk = get_nodes_block(store);
3499 return new_r_Jmp(blk);
3505 } /* transform_node_Proj_Store */
3508 * Transform a Proj(Div) with a non-zero value.
3509 * Removes the exceptions and routes the memory to the NoMem node.
3511 static ir_node *transform_node_Proj_Div(ir_node *proj)
3513 ir_node *div = get_Proj_pred(proj);
3514 ir_node *b = get_Div_right(div);
3515 ir_node *res, *new_mem;
3516 const ir_node *confirm;
3519 if (value_not_zero(b, &confirm)) {
3520 /* div(x, y) && y != 0 */
3521 if (confirm == NULL) {
3522 /* we are sure we have a Const != 0 */
3523 new_mem = get_Div_mem(div);
3524 new_mem = skip_Pin(new_mem);
3525 set_Div_mem(div, new_mem);
3526 set_irn_pinned(div, op_pin_state_floats);
3529 proj_nr = get_Proj_proj(proj);
3531 case pn_Div_X_regular:
3532 return new_r_Jmp(get_nodes_block(div));
3534 case pn_Div_X_except: {
3535 ir_graph *irg = get_irn_irg(proj);
3536 /* we found an exception handler, remove it */
3537 DBG_OPT_EXC_REM(proj);
3538 return get_irg_bad(irg);
3542 ir_graph *irg = get_irn_irg(proj);
3543 res = get_Div_mem(div);
3544 new_mem = get_irg_no_mem(irg);
3547 /* This node can only float up to the Confirm block */
3548 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3550 set_irn_pinned(div, op_pin_state_floats);
3551 /* this is a Div without exception, we can remove the memory edge */
3552 set_Div_mem(div, new_mem);
3558 } /* transform_node_Proj_Div */
3561 * Transform a Proj(Mod) with a non-zero value.
3562 * Removes the exceptions and routes the memory to the NoMem node.
3564 static ir_node *transform_node_Proj_Mod(ir_node *proj)
3566 ir_node *mod = get_Proj_pred(proj);
3567 ir_node *b = get_Mod_right(mod);
3568 ir_node *res, *new_mem;
3569 const ir_node *confirm;
3572 if (value_not_zero(b, &confirm)) {
3573 /* mod(x, y) && y != 0 */
3574 proj_nr = get_Proj_proj(proj);
3576 if (confirm == NULL) {
3577 /* we are sure we have a Const != 0 */
3578 new_mem = get_Mod_mem(mod);
3579 new_mem = skip_Pin(new_mem);
3580 set_Mod_mem(mod, new_mem);
3581 set_irn_pinned(mod, op_pin_state_floats);
3586 case pn_Mod_X_regular:
3587 return new_r_Jmp(get_irn_n(mod, -1));
3589 case pn_Mod_X_except: {
3590 ir_graph *irg = get_irn_irg(proj);
3591 /* we found an exception handler, remove it */
3592 DBG_OPT_EXC_REM(proj);
3593 return get_irg_bad(irg);
3597 ir_graph *irg = get_irn_irg(proj);
3598 res = get_Mod_mem(mod);
3599 new_mem = get_irg_no_mem(irg);
3602 /* This node can only float up to the Confirm block */
3603 new_mem = new_r_Pin(get_nodes_block(confirm), new_mem);
3605 /* this is a Mod without exception, we can remove the memory edge */
3606 set_Mod_mem(mod, new_mem);
3610 if (get_Mod_left(mod) == b) {
3611 /* a % a = 0 if a != 0 */
3612 ir_graph *irg = get_irn_irg(proj);
3613 ir_mode *mode = get_irn_mode(proj);
3614 ir_node *res = new_r_Const(irg, get_mode_null(mode));
3616 DBG_OPT_CSTEVAL(mod, res);
3622 } /* transform_node_Proj_Mod */
3625 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
3627 static ir_node *transform_node_Proj_Cond(ir_node *proj)
3629 ir_node *n = get_Proj_pred(proj);
3630 ir_node *b = get_Cond_selector(n);
3632 if (!get_opt_unreachable_code())
3635 if (mode_is_int(get_irn_mode(b))) {
3636 ir_tarval *tb = value_of(b);
3638 if (tb != tarval_bad) {
3639 /* we have a constant switch */
3640 long num = get_Proj_proj(proj);
3642 if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */
3643 if (get_tarval_long(tb) == num) {
3644 /* Do NOT create a jump here, or we will have 2 control flow ops
3645 * in a block. This case is optimized away in optimize_cf(). */
3648 ir_graph *irg = get_irn_irg(proj);
3649 /* this case will NEVER be taken, kill it */
3650 return get_irg_bad(irg);
3654 long num = get_Proj_proj(proj);
3655 vrp_attr *b_vrp = vrp_get_info(b);
3656 if (num != get_Cond_default_proj(n) && b_vrp) {
3657 /* Try handling with vrp data. We only remove dead parts. */
3658 ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b));
3660 if (b_vrp->range_type == VRP_RANGE) {
3661 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3662 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3664 if ((cmp_result & ir_relation_greater) == cmp_result && (cmp_result2
3665 & ir_relation_less) == cmp_result2) {
3666 ir_graph *irg = get_irn_irg(proj);
3667 return get_irg_bad(irg);
3669 } else if (b_vrp->range_type == VRP_ANTIRANGE) {
3670 ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp);
3671 ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp);
3673 if ((cmp_result & ir_relation_less_equal) == cmp_result && (cmp_result2
3674 & ir_relation_greater_equal) == cmp_result2) {
3675 ir_graph *irg = get_irn_irg(proj);
3676 return get_irg_bad(irg);
3681 tarval_and( b_vrp->bits_set, tp),
3683 ) == ir_relation_equal)) {
3684 ir_graph *irg = get_irn_irg(proj);
3685 return get_irg_bad(irg);
3691 tarval_not(b_vrp->bits_not_set)),
3692 tarval_not(b_vrp->bits_not_set))
3693 == ir_relation_equal)) {
3694 ir_graph *irg = get_irn_irg(proj);
3695 return get_irg_bad(irg);
3706 * return true if the operation returns a value with exactly 1 bit set
3708 static bool is_single_bit(const ir_node *node)
3710 /* a first implementation, could be extended with vrp and others... */
3712 ir_node *shl_l = get_Shl_left(node);
3713 ir_mode *mode = get_irn_mode(node);
3714 int modulo = get_mode_modulo_shift(mode);
3715 /* this works if we shift a 1 and we have modulo shift */
3716 if (is_Const(shl_l) && is_Const_one(shl_l)
3717 && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) {
3720 } else if (is_Const(node)) {
3721 ir_tarval *tv = get_Const_tarval(node);
3722 return tarval_is_single_bit(tv);
3728 * Normalizes and optimizes Cmp nodes.
3730 static ir_node *transform_node_Cmp(ir_node *n)
3732 ir_node *left = get_Cmp_left(n);
3733 ir_node *right = get_Cmp_right(n);
3734 ir_mode *mode = get_irn_mode(left);
3735 ir_tarval *tv = NULL;
3736 bool changed = false;
3737 bool changedc = false;
3738 ir_relation relation = get_Cmp_relation(n);
3739 ir_relation possible = ir_get_possible_cmp_relations(left, right);
3741 /* mask out impossible relations */
3742 ir_relation new_relation = relation & possible;
3743 if (new_relation != relation) {
3744 relation = new_relation;
3748 /* Remove unnecessary conversions */
3749 /* TODO handle conv+constant */
3750 if (is_Conv(left) && is_Conv(right)) {
3751 ir_node *op_left = get_Conv_op(left);
3752 ir_node *op_right = get_Conv_op(right);
3753 ir_mode *mode_left = get_irn_mode(op_left);
3754 ir_mode *mode_right = get_irn_mode(op_right);
3756 if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)
3757 && mode_left != mode_b && mode_right != mode_b) {
3758 ir_node *block = get_nodes_block(n);
3760 if (mode_left == mode_right) {
3764 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV);
3765 } else if (smaller_mode(mode_left, mode_right)) {
3766 left = new_r_Conv(block, op_left, mode_right);
3769 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3770 } else if (smaller_mode(mode_right, mode_left)) {
3772 right = new_r_Conv(block, op_right, mode_left);
3774 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
3780 * Optimize -a CMP -b into b CMP a.
3781 * This works only for modes where unary Minus cannot Overflow.
3782 * Note that two-complement integers can Overflow so it will NOT work.
3784 if (!mode_overflow_on_unary_Minus(mode) &&
3785 is_Minus(left) && is_Minus(right)) {
3786 left = get_Minus_op(left);
3787 right = get_Minus_op(right);
3788 relation = get_inversed_relation(relation);
3790 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3793 /* remove operation on both sides if possible */
3794 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
3796 * The following operations are NOT safe for floating point operations, for instance
3797 * 1.0 + inf == 2.0 + inf, =/=> x == y
3799 if (mode_is_int(mode)) {
3800 unsigned lop = get_irn_opcode(left);
3802 if (lop == get_irn_opcode(right)) {
3803 ir_node *ll, *lr, *rl, *rr;
3805 /* same operation on both sides, try to remove */
3809 /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */
3810 left = get_unop_op(left);
3811 right = get_unop_op(right);
3813 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3816 ll = get_Add_left(left);
3817 lr = get_Add_right(left);
3818 rl = get_Add_left(right);
3819 rr = get_Add_right(right);
3822 /* X + a CMP X + b ==> a CMP b */
3826 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3827 } else if (ll == rr) {
3828 /* X + a CMP b + X ==> a CMP b */
3832 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3833 } else if (lr == rl) {
3834 /* a + X CMP X + b ==> a CMP b */
3838 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3839 } else if (lr == rr) {
3840 /* a + X CMP b + X ==> a CMP b */
3844 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3848 ll = get_Sub_left(left);
3849 lr = get_Sub_right(left);
3850 rl = get_Sub_left(right);
3851 rr = get_Sub_right(right);
3854 /* X - a CMP X - b ==> a CMP b */
3858 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3859 } else if (lr == rr) {
3860 /* a - X CMP b - X ==> a CMP b */
3864 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3868 if (get_Rotl_right(left) == get_Rotl_right(right)) {
3869 /* a ROTL X CMP b ROTL X ==> a CMP b */
3870 left = get_Rotl_left(left);
3871 right = get_Rotl_left(right);
3873 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3881 /* X+A == A, A+X == A, A-X == A -> X == 0 */
3882 if (is_Add(left) || is_Sub(left)) {
3883 ir_node *ll = get_binop_left(left);
3884 ir_node *lr = get_binop_right(left);
3886 if (lr == right && is_Add(left)) {
3892 ir_graph *irg = get_irn_irg(n);
3894 right = create_zero_const(irg, mode);
3896 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3899 if (is_Add(right) || is_Sub(right)) {
3900 ir_node *rl = get_binop_left(right);
3901 ir_node *rr = get_binop_right(right);
3903 if (rr == left && is_Add(right)) {
3909 ir_graph *irg = get_irn_irg(n);
3911 right = create_zero_const(irg, mode);
3913 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP);
3917 if (is_And(left) && is_Const(right)) {
3918 ir_node *ll = get_binop_left(left);
3919 ir_node *lr = get_binop_right(left);
3920 if (is_Shr(ll) && is_Const(lr)) {
3921 /* Cmp((x >>u c1) & c2, c3) = Cmp(x & (c2 << c1), c3 << c1) */
3922 ir_node *block = get_nodes_block(n);
3923 ir_mode *mode = get_irn_mode(left);
3925 ir_node *llr = get_Shr_right(ll);
3926 if (is_Const(llr)) {
3927 dbg_info *dbg = get_irn_dbg_info(left);
3928 ir_graph *irg = get_irn_irg(left);
3930 ir_tarval *c1 = get_Const_tarval(llr);
3931 ir_tarval *c2 = get_Const_tarval(lr);
3932 ir_tarval *c3 = get_Const_tarval(right);
3933 ir_tarval *mask = tarval_shl(c2, c1);
3934 ir_tarval *value = tarval_shl(c3, c1);
3936 left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode);
3937 right = new_r_Const(irg, value);
3942 /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0
3944 if (is_Const(right) && is_Const_null(right) && is_Eor(left)) {
3945 right = get_Eor_right(left);
3946 left = get_Eor_left(left);
3949 } /* mode_is_int(...) */
3952 /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced
3953 * by the simpler Cmp(And(1bit), val), 0) negated pnc */
3954 if (mode_is_int(mode) && is_And(left)
3955 && (relation == ir_relation_equal
3956 || (mode_is_signed(mode) && relation == ir_relation_less_greater)
3957 || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) {
3958 ir_node *and0 = get_And_left(left);
3959 ir_node *and1 = get_And_right(left);
3960 if (and1 == right) {
3961 ir_node *tmp = and0;
3965 if (and0 == right && is_single_bit(and0)) {
3966 ir_graph *irg = get_irn_irg(n);
3968 relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal;
3969 right = create_zero_const(irg, mode);
3974 /* replace mode_b compares with ands/ors */
3975 if (mode == mode_b) {
3976 ir_node *block = get_nodes_block(n);
3980 case ir_relation_less_equal:
3981 bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b);
3983 case ir_relation_less:
3984 bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b);
3986 case ir_relation_greater_equal:
3987 bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b);
3989 case ir_relation_greater:
3990 bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b);
3992 case ir_relation_less_greater:
3993 bres = new_r_Eor(block, left, right, mode_b);
3995 case ir_relation_equal:
3996 bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b);
3999 #ifdef DEBUG_libfirm
4000 ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n);
4005 DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL);
4011 * First step: normalize the compare op
4012 * by placing the constant on the right side
4013 * or moving the lower address node to the left.
4015 if (!operands_are_normalized(left, right)) {
4020 relation = get_inversed_relation(relation);
4025 * Second step: Try to reduce the magnitude
4026 * of a constant. This may help to generate better code
4027 * later and may help to normalize more compares.
4028 * Of course this is only possible for integer values.
4030 tv = value_of(right);
4031 if (tv != tarval_bad) {
4032 ir_mode *mode = get_irn_mode(right);
4034 /* TODO extend to arbitrary constants */
4035 if (is_Conv(left) && tarval_is_null(tv)) {
4036 ir_node *op = get_Conv_op(left);
4037 ir_mode *op_mode = get_irn_mode(op);
4040 * UpConv(x) REL 0 ==> x REL 0
4041 * Don't do this for float values as it's unclear whether it is a
4042 * win. (on the other side it makes detection/creation of fabs hard)
4044 if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) &&
4045 ((relation == ir_relation_equal || relation == ir_relation_less_greater) ||
4046 mode_is_signed(mode) || !mode_is_signed(op_mode)) &&
4047 !mode_is_float(mode)) {
4048 tv = get_mode_null(op_mode);
4052 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
4056 if (tv != tarval_bad) {
4057 /* the following optimization is possible on modes without Overflow
4058 * on Unary Minus or on == and !=:
4059 * -a CMP c ==> a swap(CMP) -c
4061 * Beware: for two-complement Overflow may occur, so only == and != can
4062 * be optimized, see this:
4063 * -MININT < 0 =/=> MININT > 0 !!!
4065 if (is_Minus(left) &&
4066 (!mode_overflow_on_unary_Minus(mode) ||
4067 (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) {
4068 tv = tarval_neg(tv);
4070 if (tv != tarval_bad) {
4071 left = get_Minus_op(left);
4072 relation = get_inversed_relation(relation);
4074 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4076 } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) {
4077 /* Not(a) ==/!= c ==> a ==/!= Not(c) */
4078 tv = tarval_not(tv);
4080 if (tv != tarval_bad) {
4081 left = get_Not_op(left);
4083 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4087 /* for integer modes, we have more */
4088 if (mode_is_int(mode)) {
4089 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
4090 if ((relation == ir_relation_less || relation == ir_relation_greater_equal) &&
4091 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) {
4092 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4094 if (tv != tarval_bad) {
4095 relation ^= ir_relation_equal;
4097 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4100 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
4101 else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) &&
4102 tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) {
4103 tv = tarval_add(tv, get_mode_one(mode));
4105 if (tv != tarval_bad) {
4106 relation ^= ir_relation_equal;
4108 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4112 /* the following reassociations work only for == and != */
4113 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4115 #if 0 /* Might be not that good in general */
4116 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
4117 if (tarval_is_null(tv) && is_Sub(left)) {
4118 right = get_Sub_right(left);
4119 left = get_Sub_left(left);
4121 tv = value_of(right);
4123 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4127 if (tv != tarval_bad) {
4128 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
4130 ir_node *c1 = get_Sub_right(left);
4131 ir_tarval *tv2 = value_of(c1);
4133 if (tv2 != tarval_bad) {
4134 tv2 = tarval_add(tv, value_of(c1));
4136 if (tv2 != tarval_bad) {
4137 left = get_Sub_left(left);
4140 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4144 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
4145 else if (is_Add(left)) {
4146 ir_node *a_l = get_Add_left(left);
4147 ir_node *a_r = get_Add_right(left);
4151 if (is_Const(a_l)) {
4153 tv2 = value_of(a_l);
4156 tv2 = value_of(a_r);
4159 if (tv2 != tarval_bad) {
4160 tv2 = tarval_sub(tv, tv2, NULL);
4162 if (tv2 != tarval_bad) {
4166 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4170 /* -a == c ==> a == -c, -a != c ==> a != -c */
4171 else if (is_Minus(left)) {
4172 ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL);
4174 if (tv2 != tarval_bad) {
4175 left = get_Minus_op(left);
4178 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C);
4185 if (relation == ir_relation_equal || relation == ir_relation_less_greater) {
4186 switch (get_irn_opcode(left)) {
4190 c1 = get_And_right(left);
4193 * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2
4194 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2
4196 ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv);
4198 /* TODO: move to constant evaluation */
4199 ir_graph *irg = get_irn_irg(n);
4200 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4201 c1 = new_r_Const(irg, tv);
4202 DBG_OPT_CSTEVAL(n, c1);
4206 if (tarval_is_single_bit(tv)) {
4208 * optimization for AND:
4210 * And(x, C) == C ==> And(x, C) != 0
4211 * And(x, C) != C ==> And(X, C) == 0
4213 * if C is a single Bit constant.
4216 /* check for Constant's match. We have check hare the tarvals,
4217 because our const might be changed */
4218 if (get_Const_tarval(c1) == tv) {
4219 /* fine: do the transformation */
4220 tv = get_mode_null(get_tarval_mode(tv));
4221 relation ^= ir_relation_less_equal_greater;
4223 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN);
4229 c1 = get_Or_right(left);
4230 if (is_Const(c1) && tarval_is_null(tv)) {
4232 * Or(x, C) == 0 && C != 0 ==> FALSE
4233 * Or(x, C) != 0 && C != 0 ==> TRUE
4235 if (! tarval_is_null(get_Const_tarval(c1))) {
4236 /* TODO: move to constant evaluation */
4237 ir_graph *irg = get_irn_irg(n);
4238 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4239 c1 = new_r_Const(irg, tv);
4240 DBG_OPT_CSTEVAL(n, c1);
4247 * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c
4249 * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c
4252 c1 = get_Shl_right(left);
4254 ir_graph *irg = get_irn_irg(c1);
4255 ir_tarval *tv1 = get_Const_tarval(c1);
4256 ir_mode *mode = get_irn_mode(left);
4257 ir_tarval *minus1 = get_mode_all_one(mode);
4258 ir_tarval *amask = tarval_shr(minus1, tv1);
4259 ir_tarval *cmask = tarval_shl(minus1, tv1);
4262 if (tarval_and(tv, cmask) != tv) {
4263 /* condition not met */
4264 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4265 c1 = new_r_Const(irg, tv);
4266 DBG_OPT_CSTEVAL(n, c1);
4269 sl = get_Shl_left(left);
4270 blk = get_nodes_block(n);
4271 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4272 tv = tarval_shr(tv, tv1);
4274 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4279 * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c
4281 * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c
4284 c1 = get_Shr_right(left);
4286 ir_graph *irg = get_irn_irg(c1);
4287 ir_tarval *tv1 = get_Const_tarval(c1);
4288 ir_mode *mode = get_irn_mode(left);
4289 ir_tarval *minus1 = get_mode_all_one(mode);
4290 ir_tarval *amask = tarval_shl(minus1, tv1);
4291 ir_tarval *cmask = tarval_shr(minus1, tv1);
4294 if (tarval_and(tv, cmask) != tv) {
4295 /* condition not met */
4296 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4297 c1 = new_r_Const(irg, tv);
4298 DBG_OPT_CSTEVAL(n, c1);
4301 sl = get_Shr_left(left);
4302 blk = get_nodes_block(n);
4303 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4304 tv = tarval_shl(tv, tv1);
4306 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4311 * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4313 * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1}
4316 c1 = get_Shrs_right(left);
4318 ir_graph *irg = get_irn_irg(c1);
4319 ir_tarval *tv1 = get_Const_tarval(c1);
4320 ir_mode *mode = get_irn_mode(left);
4321 ir_tarval *minus1 = get_mode_all_one(mode);
4322 ir_tarval *amask = tarval_shl(minus1, tv1);
4323 ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1));
4326 cond = tarval_sub(cond, tv1, NULL);
4327 cond = tarval_shrs(tv, cond);
4329 if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) {
4330 /* condition not met */
4331 tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true();
4332 c1 = new_r_Const(irg, tv);
4333 DBG_OPT_CSTEVAL(n, c1);
4336 sl = get_Shrs_left(left);
4337 blk = get_nodes_block(n);
4338 left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode);
4339 tv = tarval_shl(tv, tv1);
4341 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND);
4346 } /* tarval != bad */
4349 if (changedc) { /* need a new Const */
4350 ir_graph *irg = get_irn_irg(n);
4351 right = new_r_Const(irg, tv);
4355 if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) {
4356 ir_node *op = get_Proj_pred(left);
4358 if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) {
4359 ir_node *c = get_binop_right(op);
4362 ir_tarval *tv = get_Const_tarval(c);
4364 if (tarval_is_single_bit(tv)) {
4365 /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */
4366 ir_node *v = get_binop_left(op);
4367 ir_node *blk = get_irn_n(op, -1);
4368 ir_graph *irg = get_irn_irg(op);
4369 ir_mode *mode = get_irn_mode(v);
4371 tv = tarval_sub(tv, get_mode_one(mode), NULL);
4372 left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode);
4374 DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND);
4381 dbg_info *dbgi = get_irn_dbg_info(n);
4382 ir_node *block = get_nodes_block(n);
4384 /* create a new compare */
4385 n = new_rd_Cmp(dbgi, block, left, right, relation);
4392 * Optimize CopyB(mem, x, x) into a Nop.
4394 static ir_node *transform_node_Proj_CopyB(ir_node *proj)
4396 ir_node *copyb = get_Proj_pred(proj);
4397 ir_node *a = get_CopyB_dst(copyb);
4398 ir_node *b = get_CopyB_src(copyb);
4401 switch (get_Proj_proj(proj)) {
4402 case pn_CopyB_X_regular:
4403 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
4404 DBG_OPT_EXC_REM(proj);
4405 proj = new_r_Jmp(get_nodes_block(copyb));
4407 case pn_CopyB_X_except:
4408 DBG_OPT_EXC_REM(proj);
4409 proj = get_irg_bad(get_irn_irg(proj));
4416 } /* transform_node_Proj_CopyB */
4419 * Optimize Bounds(idx, idx, upper) into idx.
4421 static ir_node *transform_node_Proj_Bound(ir_node *proj)
4423 ir_node *oldn = proj;
4424 ir_node *bound = get_Proj_pred(proj);
4425 ir_node *idx = get_Bound_index(bound);
4426 ir_node *pred = skip_Proj(idx);
4429 if (idx == get_Bound_lower(bound))
4431 else if (is_Bound(pred)) {
4433 * idx was Bounds checked previously, it is still valid if
4434 * lower <= pred_lower && pred_upper <= upper.
4436 ir_node *lower = get_Bound_lower(bound);
4437 ir_node *upper = get_Bound_upper(bound);
4438 if (get_Bound_lower(pred) == lower &&
4439 get_Bound_upper(pred) == upper) {
4441 * One could expect that we simply return the previous
4442 * Bound here. However, this would be wrong, as we could
4443 * add an exception Proj to a new location then.
4444 * So, we must turn in into a tuple.
4450 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
4451 switch (get_Proj_proj(proj)) {
4453 DBG_OPT_EXC_REM(proj);
4454 proj = get_Bound_mem(bound);
4456 case pn_Bound_X_except:
4457 DBG_OPT_EXC_REM(proj);
4458 proj = get_irg_bad(get_irn_irg(proj));
4462 DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP);
4464 case pn_Bound_X_regular:
4465 DBG_OPT_EXC_REM(proj);
4466 proj = new_r_Jmp(get_nodes_block(bound));
4473 } /* transform_node_Proj_Bound */
4476 * Does all optimizations on nodes that must be done on its Projs
4477 * because of creating new nodes.
4479 static ir_node *transform_node_Proj(ir_node *proj)
4481 ir_node *n = get_Proj_pred(proj);
4483 if (n->op->ops.transform_node_Proj)
4484 return n->op->ops.transform_node_Proj(proj);
4486 } /* transform_node_Proj */
4488 static ir_node *transform_node_Phi(ir_node *phi)
4490 int n = get_irn_arity(phi);
4491 ir_mode *mode = get_irn_mode(phi);
4492 ir_node *block = get_nodes_block(phi);
4493 ir_graph *irg = get_irn_irg(phi);
4494 ir_node *bad = get_irg_bad(irg);
4497 /* Set phi-operands for bad-block inputs to bad */
4498 for (i = 0; i < n; ++i) {
4499 ir_node *pred = get_Block_cfgpred(block, i);
4502 set_irn_n(phi, i, bad);
4505 /* Move Confirms down through Phi nodes. */
4506 if (mode_is_reference(mode)) {
4507 n = get_irn_arity(phi);
4509 /* Beware of Phi0 */
4511 ir_node *pred = get_irn_n(phi, 0);
4512 ir_node *bound, *new_phi, *block, **in;
4513 ir_relation relation;
4515 if (! is_Confirm(pred))
4518 bound = get_Confirm_bound(pred);
4519 relation = get_Confirm_relation(pred);
4521 NEW_ARR_A(ir_node *, in, n);
4522 in[0] = get_Confirm_value(pred);
4524 for (i = 1; i < n; ++i) {
4525 pred = get_irn_n(phi, i);
4527 if (! is_Confirm(pred) ||
4528 get_Confirm_bound(pred) != bound ||
4529 get_Confirm_relation(pred) != relation)
4531 in[i] = get_Confirm_value(pred);
4533 /* move the Confirm nodes "behind" the Phi */
4534 block = get_irn_n(phi, -1);
4535 new_phi = new_r_Phi(block, n, in, get_irn_mode(phi));
4536 return new_r_Confirm(block, new_phi, bound, relation);
4543 * Returns the operands of a commutative bin-op, if one operand is
4544 * a const, it is returned as the second one.
4546 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c)
4548 ir_node *op_a = get_binop_left(binop);
4549 ir_node *op_b = get_binop_right(binop);
4551 assert(is_op_commutative(get_irn_op(binop)));
4553 if (is_Const(op_a)) {
4560 } /* get_comm_Binop_Ops */
4563 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
4564 * Such pattern may arise in bitfield stores.
4566 * value c4 value c4 & c2
4567 * AND c3 AND c1 | c3
4574 * AND c1 ===> OR if (c1 | c2) == 0x111..11
4577 static ir_node *transform_node_Or_bf_store(ir_node *irn_or)
4579 ir_node *irn_and, *c1;
4581 ir_node *and_l, *c3;
4582 ir_node *value, *c4;
4583 ir_node *new_and, *new_const, *block;
4584 ir_mode *mode = get_irn_mode(irn_or);
4586 ir_tarval *tv1, *tv2, *tv3, *tv4, *tv;
4590 get_comm_Binop_Ops(irn_or, &irn_and, &c1);
4591 if (!is_Const(c1) || !is_And(irn_and))
4594 get_comm_Binop_Ops(irn_and, &or_l, &c2);
4598 tv1 = get_Const_tarval(c1);
4599 tv2 = get_Const_tarval(c2);
4601 tv = tarval_or(tv1, tv2);
4602 if (tarval_is_all_one(tv)) {
4603 /* the AND does NOT clear a bit with isn't set by the OR */
4604 set_Or_left(irn_or, or_l);
4605 set_Or_right(irn_or, c1);
4607 /* check for more */
4614 get_comm_Binop_Ops(or_l, &and_l, &c3);
4615 if (!is_Const(c3) || !is_And(and_l))
4618 get_comm_Binop_Ops(and_l, &value, &c4);
4622 /* ok, found the pattern, check for conditions */
4623 assert(mode == get_irn_mode(irn_and));
4624 assert(mode == get_irn_mode(or_l));
4625 assert(mode == get_irn_mode(and_l));
4627 tv3 = get_Const_tarval(c3);
4628 tv4 = get_Const_tarval(c4);
4630 tv = tarval_or(tv4, tv2);
4631 if (!tarval_is_all_one(tv)) {
4632 /* have at least one 0 at the same bit position */
4636 if (tv3 != tarval_andnot(tv3, tv4)) {
4637 /* bit in the or_mask is outside the and_mask */
4641 if (tv1 != tarval_andnot(tv1, tv2)) {
4642 /* bit in the or_mask is outside the and_mask */
4646 /* ok, all conditions met */
4647 block = get_irn_n(irn_or, -1);
4648 irg = get_irn_irg(block);
4650 new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode);
4652 new_const = new_r_Const(irg, tarval_or(tv3, tv1));
4654 set_Or_left(irn_or, new_and);
4655 set_Or_right(irn_or, new_const);
4657 /* check for more */
4659 } /* transform_node_Or_bf_store */
4662 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl
4664 static ir_node *transform_node_Or_Rotl(ir_node *irn_or)
4666 ir_mode *mode = get_irn_mode(irn_or);
4667 ir_node *shl, *shr, *block;
4668 ir_node *irn, *x, *c1, *c2, *n;
4669 ir_tarval *tv1, *tv2;
4671 /* some backends can't handle rotl */
4672 if (!be_get_backend_param()->support_rotl)
4675 if (! mode_is_int(mode))
4678 shl = get_binop_left(irn_or);
4679 shr = get_binop_right(irn_or);
4688 } else if (!is_Shl(shl)) {
4690 } else if (!is_Shr(shr)) {
4693 x = get_Shl_left(shl);
4694 if (x != get_Shr_left(shr))
4697 c1 = get_Shl_right(shl);
4698 c2 = get_Shr_right(shr);
4699 if (is_Const(c1) && is_Const(c2)) {
4700 tv1 = get_Const_tarval(c1);
4701 if (! tarval_is_long(tv1))
4704 tv2 = get_Const_tarval(c2);
4705 if (! tarval_is_long(tv2))
4708 if (get_tarval_long(tv1) + get_tarval_long(tv2)
4709 != (int) get_mode_size_bits(mode))
4712 /* yet, condition met */
4713 block = get_nodes_block(irn_or);
4715 n = new_r_Rotl(block, x, c1, mode);
4717 DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL);
4721 /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets
4722 * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */
4723 if (!ir_is_negated_value(c1, c2)) {
4727 /* yet, condition met */
4728 block = get_nodes_block(irn_or);
4729 n = new_r_Rotl(block, x, c1, mode);
4730 DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL);
4732 } /* transform_node_Or_Rotl */
4734 static bool is_cmp_unequal_zero(const ir_node *node)
4736 ir_relation relation = get_Cmp_relation(node);
4737 ir_node *left = get_Cmp_left(node);
4738 ir_node *right = get_Cmp_right(node);
4739 ir_mode *mode = get_irn_mode(left);
4741 if (!is_Const(right) || !is_Const_null(right))
4743 if (mode_is_signed(mode)) {
4744 return relation == ir_relation_less_greater;
4746 return relation == ir_relation_greater;
4753 static ir_node *transform_node_Or(ir_node *n)
4755 ir_node *c, *oldn = n;
4756 ir_node *a = get_Or_left(n);
4757 ir_node *b = get_Or_right(n);
4760 if (is_Not(a) && is_Not(b)) {
4761 /* ~a | ~b = ~(a&b) */
4762 ir_node *block = get_nodes_block(n);
4764 mode = get_irn_mode(n);
4767 n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode);
4768 n = new_rd_Not(get_irn_dbg_info(n), block, n, mode);
4769 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN);
4773 /* we can combine the relations of two compares with the same operands */
4774 if (is_Cmp(a) && is_Cmp(b)) {
4775 ir_node *a_left = get_Cmp_left(a);
4776 ir_node *a_right = get_Cmp_left(a);
4777 ir_node *b_left = get_Cmp_left(b);
4778 ir_node *b_right = get_Cmp_right(b);
4779 if (a_left == b_left && b_left == b_right) {
4780 dbg_info *dbgi = get_irn_dbg_info(n);
4781 ir_node *block = get_nodes_block(n);
4782 ir_relation a_relation = get_Cmp_relation(a);
4783 ir_relation b_relation = get_Cmp_relation(b);
4784 ir_relation new_relation = a_relation | b_relation;
4785 return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation);
4787 /* Cmp(a!=0) or Cmp(b!=0) => Cmp(a|b != 0) */
4788 if (is_cmp_unequal_zero(a) && is_cmp_unequal_zero(b)
4789 && !mode_is_float(get_irn_mode(a_left))
4790 && !mode_is_float(get_irn_mode(b_left))) {
4791 ir_graph *irg = get_irn_irg(n);
4792 dbg_info *dbgi = get_irn_dbg_info(n);
4793 ir_node *block = get_nodes_block(n);
4794 ir_mode *mode = get_irn_mode(a_left);
4795 ir_node *n_b_left = get_irn_mode(b_left) != mode ?
4796 new_rd_Conv(dbgi, block, b_left, mode) : b_left;
4797 ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode);
4798 ir_node *zero = create_zero_const(irg, mode);
4799 return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater);
4803 mode = get_irn_mode(n);
4804 HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode);
4806 n = transform_node_Or_bf_store(n);
4807 n = transform_node_Or_Rotl(n);
4811 n = transform_bitwise_distributive(n, transform_node_Or);
4814 } /* transform_node_Or */
4818 static ir_node *transform_node(ir_node *n);
4821 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl.
4823 * Should be moved to reassociation?
4825 static ir_node *transform_node_shift(ir_node *n)
4827 ir_node *left, *right;
4829 ir_tarval *tv1, *tv2, *res;
4830 ir_node *in[2], *irn, *block;
4833 left = get_binop_left(n);
4835 /* different operations */
4836 if (get_irn_op(left) != get_irn_op(n))
4839 right = get_binop_right(n);
4840 tv1 = value_of(right);
4841 if (tv1 == tarval_bad)
4844 tv2 = value_of(get_binop_right(left));
4845 if (tv2 == tarval_bad)
4848 res = tarval_add(tv1, tv2);
4849 mode = get_irn_mode(n);
4850 irg = get_irn_irg(n);
4852 /* beware: a simple replacement works only, if res < modulo shift */
4854 int modulo_shf = get_mode_modulo_shift(mode);
4855 if (modulo_shf > 0) {
4856 ir_tarval *modulo = new_tarval_from_long(modulo_shf,
4857 get_tarval_mode(res));
4859 assert(modulo_shf >= (int) get_mode_size_bits(mode));
4861 /* shifting too much */
4862 if (!(tarval_cmp(res, modulo) & ir_relation_less)) {
4864 ir_node *block = get_nodes_block(n);
4865 dbg_info *dbgi = get_irn_dbg_info(n);
4866 ir_mode *smode = get_irn_mode(right);
4867 ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1);
4868 return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode);
4871 return new_r_Const(irg, get_mode_null(mode));
4875 res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res)));
4878 /* ok, we can replace it */
4879 block = get_nodes_block(n);
4881 in[0] = get_binop_left(left);
4882 in[1] = new_r_Const(irg, res);
4884 irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in);
4886 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
4888 return transform_node(irn);
4889 } /* transform_node_shift */
4892 * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2)
4894 * - and, or, xor instead of &
4895 * - Shl, Shr, Shrs, rotl instead of >>
4896 * (with a special case for Or/Xor + Shrs)
4898 static ir_node *transform_node_bitop_shift(ir_node *n)
4901 ir_node *right = get_binop_right(n);
4902 ir_mode *mode = get_irn_mode(n);
4903 ir_node *bitop_left;
4904 ir_node *bitop_right;
4914 ir_tarval *tv_shift;
4916 assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
4918 if (!is_Const(right))
4921 left = get_binop_left(n);
4922 op_left = get_irn_op(left);
4923 if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
4926 /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
4927 if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
4928 /* TODO: test if sign bit is affectes */
4932 bitop_right = get_binop_right(left);
4933 if (!is_Const(bitop_right))
4936 bitop_left = get_binop_left(left);
4938 block = get_nodes_block(n);
4939 dbgi = get_irn_dbg_info(n);
4940 tv1 = get_Const_tarval(bitop_right);
4941 tv2 = get_Const_tarval(right);
4943 assert(get_tarval_mode(tv1) == mode);
4946 new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
4947 tv_shift = tarval_shl(tv1, tv2);
4948 } else if (is_Shr(n)) {
4949 new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
4950 tv_shift = tarval_shr(tv1, tv2);
4951 } else if (is_Shrs(n)) {
4952 new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
4953 tv_shift = tarval_shrs(tv1, tv2);
4956 new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
4957 tv_shift = tarval_rotl(tv1, tv2);
4960 assert(get_tarval_mode(tv_shift) == mode);
4961 irg = get_irn_irg(n);
4962 new_const = new_r_Const(irg, tv_shift);
4964 if (op_left == op_And) {
4965 new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
4966 } else if (op_left == op_Or) {
4967 new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
4969 assert(op_left == op_Eor);
4970 new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
4978 * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2)
4980 * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2)
4981 * (also with x >>s c1 when c1>=c2)
4983 static ir_node *transform_node_shl_shr(ir_node *n)
4986 ir_node *right = get_binop_right(n);
4996 ir_tarval *tv_shift;
4999 ir_relation relation;
5002 assert(is_Shl(n) || is_Shr(n) || is_Shrs(n));
5004 if (!is_Const(right))
5007 left = get_binop_left(n);
5008 mode = get_irn_mode(n);
5009 if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) {
5010 ir_node *shr_right = get_binop_right(left);
5012 if (!is_Const(shr_right))
5015 x = get_binop_left(left);
5016 tv_shr = get_Const_tarval(shr_right);
5017 tv_shl = get_Const_tarval(right);
5019 if (is_Shrs(left)) {
5020 /* shrs variant only allowed if c1 >= c2 */
5021 if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal))
5024 tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr);
5027 tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr);
5029 tv_mask = tarval_shl(tv_mask, tv_shl);
5030 } else if (is_Shr(n) && is_Shl(left)) {
5031 ir_node *shl_right = get_Shl_right(left);
5033 if (!is_Const(shl_right))
5036 x = get_Shl_left(left);
5037 tv_shr = get_Const_tarval(right);
5038 tv_shl = get_Const_tarval(shl_right);
5040 tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl);
5041 tv_mask = tarval_shr(tv_mask, tv_shr);
5046 if (get_tarval_mode(tv_shl) != get_tarval_mode(tv_shr)) {
5047 tv_shl = tarval_convert_to(tv_shl, get_tarval_mode(tv_shr));
5050 assert(tv_mask != tarval_bad);
5051 assert(get_tarval_mode(tv_mask) == mode);
5053 block = get_nodes_block(n);
5054 irg = get_irn_irg(block);
5055 dbgi = get_irn_dbg_info(n);
5057 relation = tarval_cmp(tv_shl, tv_shr);
5058 if (relation == ir_relation_less || relation == ir_relation_equal) {
5059 tv_shift = tarval_sub(tv_shr, tv_shl, NULL);
5060 new_const = new_r_Const(irg, tv_shift);
5062 new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode);
5064 new_shift = new_rd_Shr(dbgi, block, x, new_const, mode);
5067 assert(relation == ir_relation_greater);
5068 tv_shift = tarval_sub(tv_shl, tv_shr, NULL);
5069 new_const = new_r_Const(irg, tv_shift);
5070 new_shift = new_rd_Shl(dbgi, block, x, new_const, mode);
5073 new_const = new_r_Const(irg, tv_mask);
5074 new_and = new_rd_And(dbgi, block, new_shift, new_const, mode);
5079 static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val)
5081 ir_mode *mode = get_tarval_mode(tv);
5082 ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode);
5083 return tarval_mod(tv, modulo_tv);
5086 typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
5087 ir_node *left, ir_node *right, ir_mode *mode);
5090 * Normalisation: if we have a shl/shr with modulo_shift behaviour
5091 * then we can use that to minimize the value of Add(x, const) or
5092 * Sub(Const, x). In particular this often avoids 1 instruction in some
5093 * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
5094 * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
5096 static ir_node *transform_node_shift_modulo(ir_node *n,
5097 new_shift_func new_shift)
5099 ir_mode *mode = get_irn_mode(n);
5100 int modulo = get_mode_modulo_shift(mode);
5101 ir_node *newop = NULL;
5102 ir_mode *mode_right;
5109 if (get_mode_arithmetic(mode) != irma_twos_complement)
5111 if (!is_po2(modulo))
5114 irg = get_irn_irg(n);
5115 block = get_nodes_block(n);
5116 right = get_binop_right(n);
5117 mode_right = get_irn_mode(right);
5118 if (is_Const(right)) {
5119 ir_tarval *tv = get_Const_tarval(right);
5120 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5125 newop = new_r_Const(irg, tv_mod);
5126 } else if (is_Add(right)) {
5127 ir_node *add_right = get_Add_right(right);
5128 if (is_Const(add_right)) {
5129 ir_tarval *tv = get_Const_tarval(add_right);
5130 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5135 newconst = new_r_Const(irg, tv_mod);
5136 newop = new_r_Add(block, get_Add_left(right), newconst,
5139 } else if (is_Sub(right)) {
5140 ir_node *sub_left = get_Sub_left(right);
5141 if (is_Const(sub_left)) {
5142 ir_tarval *tv = get_Const_tarval(sub_left);
5143 ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo);
5148 newconst = new_r_Const(irg, tv_mod);
5149 newop = new_r_Sub(block, newconst, get_Sub_right(right),
5156 if (newop != NULL) {
5157 dbg_info *dbgi = get_irn_dbg_info(n);
5158 ir_node *left = get_binop_left(n);
5159 return new_shift(dbgi, block, left, newop, mode);
5167 static ir_node *transform_node_Shr(ir_node *n)
5169 ir_node *c, *oldn = n;
5170 ir_node *left = get_Shr_left(n);
5171 ir_node *right = get_Shr_right(n);
5172 ir_mode *mode = get_irn_mode(n);
5174 HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode);
5175 n = transform_node_shift(n);
5178 n = transform_node_shift_modulo(n, new_rd_Shr);
5180 n = transform_node_shl_shr(n);
5182 n = transform_node_bitop_shift(n);
5185 } /* transform_node_Shr */
5190 static ir_node *transform_node_Shrs(ir_node *n)
5192 ir_node *c, *oldn = n;
5193 ir_node *a = get_Shrs_left(n);
5194 ir_node *b = get_Shrs_right(n);
5195 ir_mode *mode = get_irn_mode(n);
5197 HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode);
5198 n = transform_node_shift(n);
5201 n = transform_node_shift_modulo(n, new_rd_Shrs);
5203 n = transform_node_bitop_shift(n);
5206 } /* transform_node_Shrs */
5211 static ir_node *transform_node_Shl(ir_node *n)
5213 ir_node *c, *oldn = n;
5214 ir_node *a = get_Shl_left(n);
5215 ir_node *b = get_Shl_right(n);
5216 ir_mode *mode = get_irn_mode(n);
5218 HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode);
5219 n = transform_node_shift(n);
5222 n = transform_node_shift_modulo(n, new_rd_Shl);
5224 n = transform_node_shl_shr(n);
5226 n = transform_node_bitop_shift(n);
5229 } /* transform_node_Shl */
5234 static ir_node *transform_node_Rotl(ir_node *n)
5236 ir_node *c, *oldn = n;
5237 ir_node *a = get_Rotl_left(n);
5238 ir_node *b = get_Rotl_right(n);
5239 ir_mode *mode = get_irn_mode(n);
5241 HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
5242 n = transform_node_shift(n);
5245 n = transform_node_bitop_shift(n);
5248 } /* transform_node_Rotl */
5253 static ir_node *transform_node_Conv(ir_node *n)
5255 ir_node *c, *oldn = n;
5256 ir_mode *mode = get_irn_mode(n);
5257 ir_node *a = get_Conv_op(n);
5259 if (mode != mode_b && is_const_Phi(a)) {
5260 /* Do NOT optimize mode_b Conv's, this leads to remaining
5261 * Phib nodes later, because the conv_b_lower operation
5262 * is instantly reverted, when it tries to insert a Convb.
5264 c = apply_conv_on_phi(a, mode);
5266 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI);
5271 if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */
5272 ir_graph *irg = get_irn_irg(n);
5273 return new_r_Unknown(irg, mode);
5276 if (mode_is_reference(mode) &&
5277 get_mode_size_bits(mode) == get_mode_size_bits(get_irn_mode(a)) &&
5279 ir_node *l = get_Add_left(a);
5280 ir_node *r = get_Add_right(a);
5281 dbg_info *dbgi = get_irn_dbg_info(a);
5282 ir_node *block = get_nodes_block(n);
5284 ir_node *lop = get_Conv_op(l);
5285 if (get_irn_mode(lop) == mode) {
5286 /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */
5287 n = new_rd_Add(dbgi, block, lop, r, mode);
5292 ir_node *rop = get_Conv_op(r);
5293 if (get_irn_mode(rop) == mode) {
5294 /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */
5295 n = new_rd_Add(dbgi, block, l, rop, mode);
5302 } /* transform_node_Conv */
5305 * Remove dead blocks and nodes in dead blocks
5306 * in keep alive list. We do not generate a new End node.
5308 static ir_node *transform_node_End(ir_node *n)
5310 int i, j, n_keepalives = get_End_n_keepalives(n);
5313 NEW_ARR_A(ir_node *, in, n_keepalives);
5315 for (i = j = 0; i < n_keepalives; ++i) {
5316 ir_node *ka = get_End_keepalive(n, i);
5318 /* no need to keep Bad */
5323 if (j != n_keepalives)
5324 set_End_keepalives(n, j, in);
5326 } /* transform_node_End */
5328 int ir_is_negated_value(const ir_node *a, const ir_node *b)
5330 if (is_Minus(a) && get_Minus_op(a) == b)
5332 if (is_Minus(b) && get_Minus_op(b) == a)
5334 if (is_Sub(a) && is_Sub(b)) {
5335 ir_node *a_left = get_Sub_left(a);
5336 ir_node *a_right = get_Sub_right(a);
5337 ir_node *b_left = get_Sub_left(b);
5338 ir_node *b_right = get_Sub_right(b);
5340 if (a_left == b_right && a_right == b_left)
5348 * Optimize a Mux into some simpler cases.
5350 static ir_node *transform_node_Mux(ir_node *n)
5352 ir_node *oldn = n, *sel = get_Mux_sel(n);
5353 ir_mode *mode = get_irn_mode(n);
5354 ir_node *t = get_Mux_true(n);
5355 ir_node *f = get_Mux_false(n);
5356 ir_graph *irg = get_irn_irg(n);
5358 if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
5362 ir_node* block = get_nodes_block(n);
5364 ir_node* c1 = get_Mux_sel(t);
5365 ir_node* t1 = get_Mux_true(t);
5366 ir_node* f1 = get_Mux_false(t);
5368 /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */
5369 ir_node* and_ = new_r_And(block, c0, c1, mode_b);
5370 ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode);
5375 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5376 } else if (f == t1) {
5377 /* Mux(cond0, Mux(cond1, x, y), x) */
5378 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5379 ir_node* and_ = new_r_And(block, c0, not_c1, mode_b);
5380 ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode);
5385 DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE);
5387 } else if (is_Mux(f)) {
5388 ir_node* block = get_nodes_block(n);
5390 ir_node* c1 = get_Mux_sel(f);
5391 ir_node* t1 = get_Mux_true(f);
5392 ir_node* f1 = get_Mux_false(f);
5394 /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */
5395 ir_node* or_ = new_r_Or(block, c0, c1, mode_b);
5396 ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode);
5401 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5402 } else if (t == f1) {
5403 /* Mux(cond0, x, Mux(cond1, y, x)) */
5404 ir_node* not_c1 = new_r_Not(block, c1, mode_b);
5405 ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b);
5406 ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode);
5411 DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE);
5415 /* first normalization step: try to move a constant to the false side,
5416 * 0 preferred on false side too */
5417 if (is_Cmp(sel) && is_Const(t) &&
5418 (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) {
5419 dbg_info *seldbgi = get_irn_dbg_info(sel);
5420 ir_node *block = get_nodes_block(sel);
5421 ir_relation relation = get_Cmp_relation(sel);
5426 /* Mux(x, a, b) => Mux(not(x), b, a) */
5427 relation = get_negated_relation(relation);
5428 sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel),
5429 get_Cmp_right(sel), relation);
5430 n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode);
5433 /* note: after normalization, false can only happen on default */
5434 if (mode == mode_b) {
5435 dbg_info *dbg = get_irn_dbg_info(n);
5436 ir_node *block = get_nodes_block(n);
5439 ir_tarval *tv_t = get_Const_tarval(t);
5440 if (tv_t == tarval_b_true) {
5442 /* Muxb(sel, true, false) = sel */
5443 assert(get_Const_tarval(f) == tarval_b_false);
5444 DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL);
5447 /* Muxb(sel, true, x) = Or(sel, x) */
5448 n = new_rd_Or(dbg, block, sel, f, mode_b);
5449 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL);
5453 } else if (is_Const(f)) {
5454 ir_tarval *tv_f = get_Const_tarval(f);
5455 if (tv_f == tarval_b_true) {
5456 /* Muxb(sel, x, true) = Or(Not(sel), x) */
5457 ir_node* not_sel = new_rd_Not(dbg, block, sel, mode_b);
5458 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL);
5459 n = new_rd_Or(dbg, block, not_sel, t, mode_b);
5462 /* Muxb(sel, x, false) = And(sel, x) */
5463 assert(tv_f == tarval_b_false);
5464 n = new_rd_And(dbg, block, sel, t, mode_b);
5465 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL);
5471 /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b
5472 * value to integer. */
5473 if (is_Const(t) && is_Const(f) && mode_is_int(mode)) {
5474 ir_tarval *a = get_Const_tarval(t);
5475 ir_tarval *b = get_Const_tarval(f);
5477 if (tarval_is_one(a) && tarval_is_null(b)) {
5478 ir_node *block = get_nodes_block(n);
5479 ir_node *conv = new_r_Conv(block, sel, mode);
5481 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5483 } else if (tarval_is_null(a) && tarval_is_one(b)) {
5484 ir_node *block = get_nodes_block(n);
5485 ir_node *not_ = new_r_Not(block, sel, mode_b);
5486 ir_node *conv = new_r_Conv(block, not_, mode);
5488 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV);
5494 ir_node *cmp_r = get_Cmp_right(sel);
5495 if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
5496 ir_node *block = get_nodes_block(n);
5497 ir_node *cmp_l = get_Cmp_left(sel);
5499 if (mode_is_int(mode)) {
5500 ir_relation relation = get_Cmp_relation(sel);
5502 if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) {
5503 /* Mux((a & b) != 0, c, 0) */
5504 ir_node *and_r = get_And_right(cmp_l);
5507 if (and_r == t && f == cmp_r) {
5508 if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) {
5509 if (relation == ir_relation_less_greater) {
5510 /* Mux((a & 2^C) != 0, 2^C, 0) */
5512 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5514 /* Mux((a & 2^C) == 0, 2^C, 0) */
5515 n = new_rd_Eor(get_irn_dbg_info(n),
5516 block, cmp_l, t, mode);
5517 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5522 if (is_Shl(and_r)) {
5523 ir_node *shl_l = get_Shl_left(and_r);
5524 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5525 if (and_r == t && f == cmp_r) {
5526 if (relation == ir_relation_less_greater) {
5527 /* (a & (1 << n)) != 0, (1 << n), 0) */
5529 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5531 /* (a & (1 << n)) == 0, (1 << n), 0) */
5532 n = new_rd_Eor(get_irn_dbg_info(n),
5533 block, cmp_l, t, mode);
5534 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5540 and_l = get_And_left(cmp_l);
5541 if (is_Shl(and_l)) {
5542 ir_node *shl_l = get_Shl_left(and_l);
5543 if (is_Const(shl_l) && is_Const_one(shl_l)) {
5544 if (and_l == t && f == cmp_r) {
5545 if (relation == ir_relation_less_greater) {
5546 /* ((1 << n) & a) != 0, (1 << n), 0) */
5548 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5550 /* ((1 << n) & a) == 0, (1 << n), 0) */
5551 n = new_rd_Eor(get_irn_dbg_info(n),
5552 block, cmp_l, t, mode);
5553 DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP);
5568 * optimize Sync nodes that have other syncs as input we simply add the inputs
5569 * of the other sync to our own inputs
5571 static ir_node *transform_node_Sync(ir_node *n)
5573 int arity = get_Sync_n_preds(n);
5576 for (i = 0; i < arity;) {
5577 ir_node *pred = get_Sync_pred(n, i);
5581 if (!is_Sync(pred)) {
5589 pred_arity = get_Sync_n_preds(pred);
5590 for (j = 0; j < pred_arity; ++j) {
5591 ir_node *pred_pred = get_Sync_pred(pred, j);
5596 add_irn_n(n, pred_pred);
5600 if (get_Sync_pred(n, k) == pred_pred) break;
5605 /* rehash the sync node */
5609 } /* transform_node_Sync */
5611 static ir_node *transform_node_Load(ir_node *n)
5613 /* if our memory predecessor is a load from the same address, then reuse the
5614 * previous result */
5615 ir_node *mem = get_Load_mem(n);
5620 /* don't touch volatile loads */
5621 if (get_Load_volatility(n) == volatility_is_volatile)
5623 mem_pred = get_Proj_pred(mem);
5624 if (is_Load(mem_pred)) {
5625 ir_node *pred_load = mem_pred;
5627 /* conservatively compare the 2 loads. TODO: This could be less strict
5628 * with fixup code in some situations (like smaller/bigger modes) */
5629 if (get_Load_ptr(pred_load) != get_Load_ptr(n))
5631 if (get_Load_mode(pred_load) != get_Load_mode(n))
5633 /* all combinations of aligned/unaligned pred/n should be fine so we do
5634 * not compare the unaligned attribute */
5636 ir_node *block = get_nodes_block(n);
5637 ir_node *jmp = new_r_Jmp(block);
5638 ir_graph *irg = get_irn_irg(n);
5639 ir_node *bad = get_irg_bad(irg);
5640 ir_mode *mode = get_Load_mode(n);
5641 ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res);
5642 ir_node *in[pn_Load_max] = { mem, jmp, bad, res };
5643 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5646 } else if (is_Store(mem_pred)) {
5647 ir_node *pred_store = mem_pred;
5648 ir_node *value = get_Store_value(pred_store);
5650 if (get_Store_ptr(pred_store) != get_Load_ptr(n))
5652 if (get_irn_mode(value) != get_Load_mode(n))
5654 /* all combinations of aligned/unaligned pred/n should be fine so we do
5655 * not compare the unaligned attribute */
5657 ir_node *block = get_nodes_block(n);
5658 ir_node *jmp = new_r_Jmp(block);
5659 ir_graph *irg = get_irn_irg(n);
5660 ir_node *bad = get_irg_bad(irg);
5661 ir_node *res = value;
5662 ir_node *in[pn_Load_max] = { mem, jmp, bad, res };
5663 ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
5672 * optimize a trampoline Call into a direct Call
5674 static ir_node *transform_node_Call(ir_node *call)
5676 ir_node *callee = get_Call_ptr(call);
5677 ir_node *adr, *mem, *res, *bl, **in;
5678 ir_type *ctp, *mtp, *tp;
5682 size_t i, n_res, n_param;
5685 if (! is_Proj(callee))
5687 callee = get_Proj_pred(callee);
5688 if (! is_Builtin(callee))
5690 if (get_Builtin_kind(callee) != ir_bk_inner_trampoline)
5693 mem = get_Call_mem(call);
5695 if (skip_Proj(mem) == callee) {
5696 /* memory is routed to the trampoline, skip */
5697 mem = get_Builtin_mem(callee);
5700 /* build a new call type */
5701 mtp = get_Call_type(call);
5702 tdb = get_type_dbg_info(mtp);
5704 n_res = get_method_n_ress(mtp);
5705 n_param = get_method_n_params(mtp);
5706 ctp = new_d_type_method(n_param + 1, n_res, tdb);
5708 for (i = 0; i < n_res; ++i)
5709 set_method_res_type(ctp, i, get_method_res_type(mtp, i));
5711 NEW_ARR_A(ir_node *, in, n_param + 1);
5713 /* FIXME: we don't need a new pointer type in every step */
5714 irg = get_irn_irg(call);
5715 tp = get_irg_frame_type(irg);
5716 tp = new_type_pointer(tp);
5717 set_method_param_type(ctp, 0, tp);
5719 in[0] = get_Builtin_param(callee, 2);
5720 for (i = 0; i < n_param; ++i) {
5721 set_method_param_type(ctp, i + 1, get_method_param_type(mtp, i));
5722 in[i + 1] = get_Call_param(call, i);
5724 var = get_method_variadicity(mtp);
5725 set_method_variadicity(ctp, var);
5726 /* When we resolve a trampoline, the function must be called by a this-call */
5727 set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call);
5728 set_method_additional_properties(ctp, get_method_additional_properties(mtp));
5730 adr = get_Builtin_param(callee, 1);
5732 db = get_irn_dbg_info(call);
5733 bl = get_nodes_block(call);
5735 res = new_rd_Call(db, bl, mem, adr, n_param + 1, in, ctp);
5736 if (get_irn_pinned(call) == op_pin_state_floats)
5737 set_irn_pinned(res, op_pin_state_floats);
5739 } /* transform_node_Call */
5742 * Tries several [inplace] [optimizing] transformations and returns an
5743 * equivalent node. The difference to equivalent_node() is that these
5744 * transformations _do_ generate new nodes, and thus the old node must
5745 * not be freed even if the equivalent node isn't the old one.
5747 static ir_node *transform_node(ir_node *n)
5752 * Transform_node is the only "optimizing transformation" that might
5753 * return a node with a different opcode. We iterate HERE until fixpoint
5754 * to get the final result.
5758 if (n->op->ops.transform_node != NULL)
5759 n = n->op->ops.transform_node(n);
5760 } while (oldn != n);
5763 } /* transform_node */
5766 * Sets the default transform node operation for an ir_op_ops.
5768 * @param code the opcode for the default operation
5769 * @param ops the operations initialized
5774 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
5778 ops->transform_node = transform_node_##a; \
5780 #define CASE_PROJ(a) \
5782 ops->transform_node_Proj = transform_node_Proj_##a; \
5784 #define CASE_PROJ_EX(a) \
5786 ops->transform_node = transform_node_##a; \
5787 ops->transform_node_Proj = transform_node_Proj_##a; \
5827 } /* firm_set_default_transform_node */
5830 /* **************** Common Subexpression Elimination **************** */
5832 /** The size of the hash table used, should estimate the number of nodes
5834 #define N_IR_NODES 512
5836 /** Compares the attributes of two Const nodes. */
5837 static int node_cmp_attr_Const(const ir_node *a, const ir_node *b)
5839 return get_Const_tarval(a) != get_Const_tarval(b);
5842 /** Compares the attributes of two Proj nodes. */
5843 static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b)
5845 return a->attr.proj.proj != b->attr.proj.proj;
5848 /** Compares the attributes of two Alloc nodes. */
5849 static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b)
5851 const alloc_attr *pa = &a->attr.alloc;
5852 const alloc_attr *pb = &b->attr.alloc;
5853 return (pa->where != pb->where) || (pa->type != pb->type);
5856 /** Compares the attributes of two Free nodes. */
5857 static int node_cmp_attr_Free(const ir_node *a, const ir_node *b)
5859 const free_attr *pa = &a->attr.free;
5860 const free_attr *pb = &b->attr.free;
5861 return (pa->where != pb->where) || (pa->type != pb->type);
5864 /** Compares the attributes of two SymConst nodes. */
5865 static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b)
5867 const symconst_attr *pa = &a->attr.symc;
5868 const symconst_attr *pb = &b->attr.symc;
5869 return (pa->kind != pb->kind)
5870 || (pa->sym.type_p != pb->sym.type_p);
5873 /** Compares the attributes of two Call nodes. */
5874 static int node_cmp_attr_Call(const ir_node *a, const ir_node *b)
5876 const call_attr *pa = &a->attr.call;
5877 const call_attr *pb = &b->attr.call;
5878 return (pa->type != pb->type)
5879 || (pa->tail_call != pb->tail_call);
5882 /** Compares the attributes of two Sel nodes. */
5883 static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b)
5885 const ir_entity *a_ent = get_Sel_entity(a);
5886 const ir_entity *b_ent = get_Sel_entity(b);
5887 return a_ent != b_ent;
5890 /** Compares the attributes of two Phi nodes. */
5891 static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b)
5893 /* we can only enter this function if both nodes have the same number of inputs,
5894 hence it is enough to check if one of them is a Phi0 */
5896 /* check the Phi0 pos attribute */
5897 return a->attr.phi.u.pos != b->attr.phi.u.pos;
5902 /** Compares the attributes of two Conv nodes. */
5903 static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b)
5905 return get_Conv_strict(a) != get_Conv_strict(b);
5908 /** Compares the attributes of two Cast nodes. */
5909 static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b)
5911 return get_Cast_type(a) != get_Cast_type(b);
5914 /** Compares the attributes of two Load nodes. */
5915 static int node_cmp_attr_Load(const ir_node *a, const ir_node *b)
5917 if (get_Load_volatility(a) == volatility_is_volatile ||
5918 get_Load_volatility(b) == volatility_is_volatile)
5919 /* NEVER do CSE on volatile Loads */
5921 /* do not CSE Loads with different alignment. Be conservative. */
5922 if (get_Load_unaligned(a) != get_Load_unaligned(b))
5925 return get_Load_mode(a) != get_Load_mode(b);
5928 /** Compares the attributes of two Store nodes. */
5929 static int node_cmp_attr_Store(const ir_node *a, const ir_node *b)
5931 /* do not CSE Stores with different alignment. Be conservative. */
5932 if (get_Store_unaligned(a) != get_Store_unaligned(b))
5935 /* NEVER do CSE on volatile Stores */
5936 return (get_Store_volatility(a) == volatility_is_volatile ||
5937 get_Store_volatility(b) == volatility_is_volatile);
5940 /** Compares two exception attributes */
5941 static int node_cmp_exception(const ir_node *a, const ir_node *b)
5943 const except_attr *ea = &a->attr.except;
5944 const except_attr *eb = &b->attr.except;
5946 return ea->pin_state != eb->pin_state;
5949 #define node_cmp_attr_Bound node_cmp_exception
5951 /** Compares the attributes of two Div nodes. */
5952 static int node_cmp_attr_Div(const ir_node *a, const ir_node *b)
5954 const div_attr *ma = &a->attr.div;
5955 const div_attr *mb = &b->attr.div;
5956 return ma->exc.pin_state != mb->exc.pin_state ||
5957 ma->resmode != mb->resmode ||
5958 ma->no_remainder != mb->no_remainder;
5961 /** Compares the attributes of two Mod nodes. */
5962 static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b)
5964 const mod_attr *ma = &a->attr.mod;
5965 const mod_attr *mb = &b->attr.mod;
5966 return ma->exc.pin_state != mb->exc.pin_state ||
5967 ma->resmode != mb->resmode;
5970 static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b)
5972 const cmp_attr *ma = &a->attr.cmp;
5973 const cmp_attr *mb = &b->attr.cmp;
5974 return ma->relation != mb->relation;
5977 /** Compares the attributes of two Confirm nodes. */
5978 static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b)
5980 const confirm_attr *ma = &a->attr.confirm;
5981 const confirm_attr *mb = &b->attr.confirm;
5982 return ma->relation != mb->relation;
5985 /** Compares the attributes of two Builtin nodes. */
5986 static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b)
5988 /* no need to compare the type, equal kind means equal type */
5989 return get_Builtin_kind(a) != get_Builtin_kind(b);
5992 /** Compares the attributes of two ASM nodes. */
5993 static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b)
5996 const ir_asm_constraint *ca;
5997 const ir_asm_constraint *cb;
6000 if (get_ASM_text(a) != get_ASM_text(b))
6003 /* Should we really check the constraints here? Should be better, but is strange. */
6004 n = get_ASM_n_input_constraints(a);
6005 if (n != get_ASM_n_input_constraints(b))
6008 ca = get_ASM_input_constraints(a);
6009 cb = get_ASM_input_constraints(b);
6010 for (i = 0; i < n; ++i) {
6011 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6012 || ca[i].mode != cb[i].mode)
6016 n = get_ASM_n_output_constraints(a);
6017 if (n != get_ASM_n_output_constraints(b))
6020 ca = get_ASM_output_constraints(a);
6021 cb = get_ASM_output_constraints(b);
6022 for (i = 0; i < n; ++i) {
6023 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint
6024 || ca[i].mode != cb[i].mode)
6028 n = get_ASM_n_clobbers(a);
6029 if (n != get_ASM_n_clobbers(b))
6032 cla = get_ASM_clobbers(a);
6033 clb = get_ASM_clobbers(b);
6034 for (i = 0; i < n; ++i) {
6035 if (cla[i] != clb[i])
6041 /** Compares the inexistent attributes of two Dummy nodes. */
6042 static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b)
6050 * Set the default node attribute compare operation for an ir_op_ops.
6052 * @param code the opcode for the default operation
6053 * @param ops the operations initialized
6058 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
6062 ops->node_cmp_attr = node_cmp_attr_##a; \
6094 } /* firm_set_default_node_cmp_attr */
6097 * Compare function for two nodes in the value table. Gets two
6098 * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression.
6100 int identities_cmp(const void *elt, const void *key)
6102 ir_node *a = (ir_node *)elt;
6103 ir_node *b = (ir_node *)key;
6106 if (a == b) return 0;
6108 if ((get_irn_op(a) != get_irn_op(b)) ||
6109 (get_irn_mode(a) != get_irn_mode(b))) return 1;
6111 /* compare if a's in and b's in are of equal length */
6112 irn_arity_a = get_irn_arity(a);
6113 if (irn_arity_a != get_irn_arity(b))
6116 /* blocks are never the same */
6120 if (get_irn_pinned(a) == op_pin_state_pinned) {
6121 /* for pinned nodes, the block inputs must be equal */
6122 if (get_irn_n(a, -1) != get_irn_n(b, -1))
6124 } else if (! get_opt_global_cse()) {
6125 /* for block-local CSE both nodes must be in the same Block */
6126 if (get_nodes_block(a) != get_nodes_block(b))
6130 /* compare a->in[0..ins] with b->in[0..ins] */
6131 for (i = 0; i < irn_arity_a; ++i) {
6132 ir_node *pred_a = get_irn_n(a, i);
6133 ir_node *pred_b = get_irn_n(b, i);
6134 if (pred_a != pred_b) {
6135 /* if both predecessors are CSE neutral they might be different */
6136 if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b))
6142 * here, we already now that the nodes are identical except their
6145 if (a->op->ops.node_cmp_attr)
6146 return a->op->ops.node_cmp_attr(a, b);
6149 } /* identities_cmp */
6152 * Calculate a hash value of a node.
6154 * @param node The IR-node
6156 unsigned ir_node_hash(const ir_node *node)
6158 return node->op->ops.hash(node);
6159 } /* ir_node_hash */
6162 void new_identities(ir_graph *irg)
6164 if (irg->value_table != NULL)
6165 del_pset(irg->value_table);
6166 irg->value_table = new_pset(identities_cmp, N_IR_NODES);
6167 } /* new_identities */
6169 void del_identities(ir_graph *irg)
6171 if (irg->value_table != NULL)
6172 del_pset(irg->value_table);
6173 } /* del_identities */
6175 /* Normalize a node by putting constants (and operands with larger
6176 * node index) on the right (operator side). */
6177 void ir_normalize_node(ir_node *n)
6179 if (is_op_commutative(get_irn_op(n))) {
6180 ir_node *l = get_binop_left(n);
6181 ir_node *r = get_binop_right(n);
6183 /* For commutative operators perform a OP b == b OP a but keep
6184 * constants on the RIGHT side. This helps greatly in some
6185 * optimizations. Moreover we use the idx number to make the form
6187 if (!operands_are_normalized(l, r)) {
6188 set_binop_left(n, r);
6189 set_binop_right(n, l);
6193 } /* ir_normalize_node */
6196 * Return the canonical node computing the same value as n.
6197 * Looks up the node in a hash table, enters it in the table
6198 * if it isn't there yet.
6200 * @param n the node to look up
6202 * @return a node that computes the same value as n or n if no such
6203 * node could be found
6205 ir_node *identify_remember(ir_node *n)
6207 ir_graph *irg = get_irn_irg(n);
6208 pset *value_table = irg->value_table;
6211 if (value_table == NULL)
6214 ir_normalize_node(n);
6215 /* lookup or insert in hash table with given hash key. */
6216 nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n));
6219 /* n is reachable again */
6220 edges_node_revival(nn);
6224 } /* identify_remember */
6227 * During construction we set the op_pin_state_pinned flag in the graph right
6228 * when the optimization is performed. The flag turning on procedure global
6229 * cse could be changed between two allocations. This way we are safe.
6231 * @param n The node to lookup
6233 static inline ir_node *identify_cons(ir_node *n)
6237 n = identify_remember(n);
6238 if (n != old && get_nodes_block(old) != get_nodes_block(n)) {
6239 ir_graph *irg = get_irn_irg(n);
6240 set_irg_pinned(irg, op_pin_state_floats);
6243 } /* identify_cons */
6245 /* Add a node to the identities value table. */
6246 void add_identities(ir_node *node)
6253 identify_remember(node);
6256 /* Visit each node in the value table of a graph. */
6257 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env)
6260 ir_graph *rem = current_ir_graph;
6262 current_ir_graph = irg;
6263 foreach_pset(irg->value_table, ir_node*, node) {
6266 current_ir_graph = rem;
6267 } /* visit_all_identities */
6270 * Garbage in, garbage out. If a node has a dead input, i.e., the
6271 * Bad node is input to the node, return the Bad node.
6273 static ir_node *gigo(ir_node *node)
6275 ir_op *op = get_irn_op(node);
6277 /* Code in "Bad" blocks is unreachable and can be replaced by Bad */
6278 if (op != op_Block && is_Bad(get_nodes_block(node))) {
6279 ir_graph *irg = get_irn_irg(node);
6280 return get_irg_bad(irg);
6283 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
6284 blocks predecessors is dead. */
6285 if (op != op_Block && op != op_Phi && op != op_Tuple && op != op_Anchor
6286 && op != op_Sync && op != op_End) {
6287 ir_graph *irg = get_irn_irg(node);
6288 int irn_arity = get_irn_arity(node);
6291 for (i = 0; i < irn_arity; i++) {
6292 ir_node *pred = get_irn_n(node, i);
6295 /* be careful not to kill cfopts too early or we might violate
6296 * the 1 cfop per block property */
6298 || is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK))
6299 return get_irg_bad(irg);
6308 * These optimizations deallocate nodes from the obstack.
6309 * It can only be called if it is guaranteed that no other nodes
6310 * reference this one, i.e., right after construction of a node.
6312 * @param n The node to optimize
6314 ir_node *optimize_node(ir_node *n)
6317 ir_graph *irg = get_irn_irg(n);
6318 unsigned iro = get_irn_opcode(n);
6321 /* Always optimize Phi nodes: part of the construction. */
6322 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
6324 /* Remove nodes with dead (Bad) input.
6325 Run always for transformation induced Bads. */
6328 edges_node_deleted(oldn);
6330 /* We found an existing, better node, so we can deallocate the old node. */
6331 irg_kill_node(irg, oldn);
6335 /* constant expression evaluation / constant folding */
6336 if (get_opt_constant_folding()) {
6337 /* neither constants nor Tuple values can be evaluated */
6338 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
6339 /* try to evaluate */
6340 tv = computed_value(n);
6341 if (tv != tarval_bad) {
6346 * we MUST copy the node here temporarily, because it's still
6347 * needed for DBG_OPT_CSTEVAL
6349 node_size = offsetof(ir_node, attr) + n->op->attr_size;
6350 oldn = (ir_node*)alloca(node_size);
6352 memcpy(oldn, n, node_size);
6353 CLONE_ARR_A(ir_node *, oldn->in, n->in);
6355 /* ARG, copy the in array, we need it for statistics */
6356 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
6358 /* note the inplace edges module */
6359 edges_node_deleted(n);
6361 /* evaluation was successful -- replace the node. */
6362 irg_kill_node(irg, n);
6363 nw = new_r_Const(irg, tv);
6365 DBG_OPT_CSTEVAL(oldn, nw);
6371 /* remove unnecessary nodes */
6372 if (get_opt_algebraic_simplification() ||
6373 (iro == iro_Phi) || /* always optimize these nodes. */
6375 (iro == iro_Proj) ||
6376 (iro == iro_Block) ) /* Flags tested local. */
6377 n = equivalent_node(n);
6379 /* Common Subexpression Elimination.
6381 * Checks whether n is already available.
6382 * The block input is used to distinguish different subexpressions. Right
6383 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
6384 * subexpressions within a block.
6387 n = identify_cons(n);
6390 edges_node_deleted(oldn);
6392 /* We found an existing, better node, so we can deallocate the old node. */
6393 irg_kill_node(irg, oldn);
6397 /* Some more constant expression evaluation that does not allow to
6399 iro = get_irn_opcode(n);
6400 if (get_opt_algebraic_simplification() ||
6401 (iro == iro_Cond) ||
6402 (iro == iro_Proj)) /* Flags tested local. */
6403 n = transform_node(n);
6405 /* Now we have a legal, useful node. Enter it in hash table for CSE */
6406 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6408 n = identify_remember(o);
6414 } /* optimize_node */
6418 * These optimizations never deallocate nodes (in place). This can cause dead
6419 * nodes lying on the obstack. Remove these by a dead node elimination,
6420 * i.e., a copying garbage collection.
6422 ir_node *optimize_in_place_2(ir_node *n)
6426 unsigned iro = get_irn_opcode(n);
6428 if (!get_opt_optimize() && !is_Phi(n)) return n;
6430 if (iro == iro_Deleted)
6433 /* Remove nodes with dead (Bad) input.
6434 Run always for transformation induced Bads. */
6439 /* constant expression evaluation / constant folding */
6440 if (get_opt_constant_folding()) {
6441 /* neither constants nor Tuple values can be evaluated */
6442 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
6443 /* try to evaluate */
6444 tv = computed_value(n);
6445 if (tv != tarval_bad) {
6446 /* evaluation was successful -- replace the node. */
6447 ir_graph *irg = get_irn_irg(n);
6449 n = new_r_Const(irg, tv);
6451 DBG_OPT_CSTEVAL(oldn, n);
6457 /* remove unnecessary nodes */
6458 if (get_opt_constant_folding() ||
6459 (iro == iro_Phi) || /* always optimize these nodes. */
6460 (iro == iro_Id) || /* ... */
6461 (iro == iro_Proj) || /* ... */
6462 (iro == iro_Block) ) /* Flags tested local. */
6463 n = equivalent_node(n);
6465 /** common subexpression elimination **/
6466 /* Checks whether n is already available. */
6467 /* The block input is used to distinguish different subexpressions. Right
6468 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
6469 subexpressions within a block. */
6470 if (get_opt_cse()) {
6472 n = identify_remember(o);
6477 /* Some more constant expression evaluation. */
6478 iro = get_irn_opcode(n);
6479 if (get_opt_constant_folding() ||
6480 (iro == iro_Cond) ||
6481 (iro == iro_Proj)) /* Flags tested local. */
6482 n = transform_node(n);
6484 /* Now we can verify the node, as it has no dead inputs any more. */
6487 /* Now we have a legal, useful node. Enter it in hash table for cse.
6488 Blocks should be unique anyways. (Except the successor of start:
6489 is cse with the start block!) */
6490 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
6492 n = identify_remember(o);
6498 } /* optimize_in_place_2 */
6501 * Wrapper for external use, set proper status bits after optimization.
6503 ir_node *optimize_in_place(ir_node *n)
6505 ir_graph *irg = get_irn_irg(n);
6506 /* Handle graph state */
6507 assert(get_irg_phase_state(irg) != phase_building);
6509 if (get_opt_global_cse())
6510 set_irg_pinned(irg, op_pin_state_floats);
6511 if (get_irg_outs_state(irg) == outs_consistent)
6512 set_irg_outs_inconsistent(irg);
6514 /* FIXME: Maybe we could also test whether optimizing the node can
6515 change the control graph. */
6516 set_irg_doms_inconsistent(irg);
6517 return optimize_in_place_2(n);
6518 } /* optimize_in_place */
6521 * Calculate a hash value of a Const node.
6523 static unsigned hash_Const(const ir_node *node)
6527 /* special value for const, as they only differ in their tarval. */
6528 h = HASH_PTR(node->attr.con.tarval);
6534 * Calculate a hash value of a SymConst node.
6536 static unsigned hash_SymConst(const ir_node *node)
6540 /* all others are pointers */
6541 h = HASH_PTR(node->attr.symc.sym.type_p);
6544 } /* hash_SymConst */
6547 * Set the default hash operation in an ir_op_ops.
6549 * @param code the opcode for the default operation
6550 * @param ops the operations initialized
6555 static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops)
6559 ops->hash = hash_##a; \
6562 /* hash function already set */
6563 if (ops->hash != NULL)
6570 /* use input/mode default hash if no function was given */
6571 ops->hash = firm_default_hash;
6579 * Sets the default operation for an ir_ops.
6581 ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops)
6583 ops = firm_set_default_hash(code, ops);
6584 ops = firm_set_default_computed_value(code, ops);
6585 ops = firm_set_default_equivalent_node(code, ops);
6586 ops = firm_set_default_transform_node(code, ops);
6587 ops = firm_set_default_node_cmp_attr(code, ops);
6588 ops = firm_set_default_get_type_attr(code, ops);
6589 ops = firm_set_default_get_entity_attr(code, ops);
6592 } /* firm_set_default_operations */