X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=1711aa07dbcf580419fa7648fbd31ea1e7cf1ca7;hb=87f7c13645e762f21991ad79985a7f3e6ef612cf;hp=6b3e6b4187810cd51166aff5f6eadbea1e9e20f3;hpb=9fe9c20a1570729f1bdd21a6f174ec045d2ff986;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 6b3e6b418..1711aa07d 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -3,10 +3,10 @@ * File name: ir/ir/iropt.c * Purpose: iropt --- optimizations intertwined with IR construction. * Author: Christian Schaefer - * Modified by: Goetz Lindenmaier + * Modified by: Goetz Lindenmaier, Michael Beck * Created: * CVS-ID: $Id$ - * Copyright: (c) 1998-2003 Universität Karlsruhe + * Copyright: (c) 1998-2006 Universität Karlsruhe * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. */ @@ -24,65 +24,70 @@ #include #endif -# include "irnode_t.h" -# include "irgraph_t.h" -# include "irmode_t.h" -# include "iropt_t.h" -# include "ircons_t.h" -# include "irgmod.h" -# include "irvrfy.h" -# include "tv_t.h" -# include "dbginfo_t.h" -# include "iropt_dbg.h" -# include "irflag_t.h" -# include "firmstat.h" -# include "irarch.h" -# include "hashptr.h" +#include "irnode_t.h" +#include "irgraph_t.h" +#include "iredges_t.h" +#include "irmode_t.h" +#include "iropt_t.h" +#include "ircons_t.h" +#include "irgmod.h" +#include "irvrfy.h" +#include "tv_t.h" +#include "dbginfo_t.h" +#include "iropt_dbg.h" +#include "irflag_t.h" +#include "irhooks.h" +#include "irarch.h" +#include "hashptr.h" +#include "archop.h" +#include "opt_polymorphy.h" +#include "opt_confirms.h" +#include "irtools.h" /* Make types visible to allow most efficient access */ -# include "entity_t.h" - -# ifdef DO_HEAPANALYSIS -/* heapanal can't cope with NoMems */ -# else /* if defined DO_HEAPANALYSIS */ -# define USE_NOMEM -# endif /* defined DO_HEAPANALYSIS */ - -/** - * Trivial INLINEable routine for copy propagation. - * Does follow Ids, needed to optimize INLINEd code. - */ -static INLINE ir_node * -follow_Id (ir_node *n) -{ - while (get_irn_op (n) == op_Id) n = get_Id_pred (n); - return n; -} +#include "entity_t.h" /** - * return the value of a Constant + * Return the value of a Constant. */ -static tarval *computed_value_Const(ir_node *n) -{ - return get_Const_tarval(n); -} +static tarval *computed_value_Const(ir_node *n) { + return get_Const_tarval(n); +} /* computed_value_Const */ /** - * return the value of a 'sizeof' SymConst + * Return the value of a 'sizeof' or 'alignof' SymConst. */ -static tarval *computed_value_SymConst(ir_node *n) -{ - if ((get_SymConst_kind(n) == symconst_size) && - (get_type_state(get_SymConst_type(n))) == layout_fixed) - return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(n)); +static tarval *computed_value_SymConst(ir_node *n) { + ir_type *type; + entity *ent; + + switch (get_SymConst_kind(n)) { + case symconst_type_size: + type = get_SymConst_type(n); + if (get_type_state(type) == layout_fixed) + return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n)); + break; + case symconst_type_align: + type = get_SymConst_type(n); + if (get_type_state(type) == layout_fixed) + return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n)); + break; + case symconst_ofs_ent: + ent = get_SymConst_entity(n); + type = get_entity_owner(ent); + if (get_type_state(type) == layout_fixed) + return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n)); + break; + default: + break; + } return tarval_bad; -} +} /* computed_value_SymConst */ /** - * return the value of an Add + * Return the value of an Add. */ -static tarval *computed_value_Add(ir_node *n) -{ +static tarval *computed_value_Add(ir_node *n) { ir_node *a = get_Add_left(n); ir_node *b = get_Add_right(n); @@ -93,14 +98,13 @@ static tarval *computed_value_Add(ir_node *n) return tarval_add(ta, tb); return tarval_bad; -} +} /* computed_value_Add */ /** - * return the value of a Sub + * Return the value of a Sub. * Special case: a - a */ -static tarval *computed_value_Sub(ir_node *n) -{ +static tarval *computed_value_Sub(ir_node *n) { ir_node *a = get_Sub_left(n); ir_node *b = get_Sub_right(n); tarval *ta; @@ -108,7 +112,7 @@ static tarval *computed_value_Sub(ir_node *n) /* a - a */ if (a == b && !is_Bad(a)) - return get_tarval_null(get_irn_mode(n)); + return get_mode_null(get_irn_mode(n)); ta = value_of(a); tb = value_of(b); @@ -117,13 +121,55 @@ static tarval *computed_value_Sub(ir_node *n) return tarval_sub(ta, tb); return tarval_bad; -} +} /* computed_value_Sub */ /** - * return the value of an unary Minus + * Return the value of a Carry. + * Special : a op 0, 0 op b */ -static tarval *computed_value_Minus(ir_node *n) -{ +static tarval *computed_value_Carry(ir_node *n) { + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *m = get_irn_mode(n); + + tarval *ta = value_of(a); + tarval *tb = value_of(b); + + if ((ta != tarval_bad) && (tb != tarval_bad)) { + tarval_add(ta, tb); + return tarval_carry() ? get_mode_one(m) : get_mode_null(m); + } else { + if ( (classify_tarval(ta) == TV_CLASSIFY_NULL) + || (classify_tarval(tb) == TV_CLASSIFY_NULL)) + return get_mode_null(m); + } + return tarval_bad; +} /* computed_value_Carry */ + +/** + * Return the value of a Borrow. + * Special : a op 0 + */ +static tarval *computed_value_Borrow(ir_node *n) { + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *m = get_irn_mode(n); + + tarval *ta = value_of(a); + tarval *tb = value_of(b); + + if ((ta != tarval_bad) && (tb != tarval_bad)) { + return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m); + } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) { + return get_mode_null(m); + } + return tarval_bad; +} /* computed_value_Borrow */ + +/** + * Return the value of an unary Minus. + */ +static tarval *computed_value_Minus(ir_node *n) { ir_node *a = get_Minus_op(n); tarval *ta = value_of(a); @@ -131,13 +177,12 @@ static tarval *computed_value_Minus(ir_node *n) return tarval_neg(ta); return tarval_bad; -} +} /* computed_value_Minus */ /** - * return the value of a Mul + * Return the value of a Mul. */ -static tarval *computed_value_Mul(ir_node *n) -{ +static tarval *computed_value_Mul(ir_node *n) { ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); @@ -156,13 +201,12 @@ static tarval *computed_value_Mul(ir_node *n) return tb; } return tarval_bad; -} +} /* computed_value_Mul */ /** - * return the value of a floating point Quot + * Return the value of a floating point Quot. */ -static tarval *computed_value_Quot(ir_node *n) -{ +static tarval *computed_value_Quot(ir_node *n) { ir_node *a = get_Quot_left(n); ir_node *b = get_Quot_right(n); @@ -175,14 +219,13 @@ static tarval *computed_value_Quot(ir_node *n) return tarval_quo(ta, tb); } return tarval_bad; -} +} /* computed_value_Quot */ /** - * calculate the value of an integer Div of two nodes + * Calculate the value of an integer Div of two nodes. * Special case: 0 / b */ -static tarval *do_computed_value_Div(ir_node *a, ir_node *b) -{ +static tarval *do_computed_value_Div(ir_node *a, ir_node *b) { tarval *ta = value_of(a); tarval *tb = value_of(b); @@ -194,22 +237,20 @@ static tarval *do_computed_value_Div(ir_node *a, ir_node *b) return ta; } return tarval_bad; -} +} /* do_computed_value_Div */ /** - * return the value of an integer Div + * Return the value of an integer Div. */ -static tarval *computed_value_Div(ir_node *n) -{ +static tarval *computed_value_Div(ir_node *n) { return do_computed_value_Div(get_Div_left(n), get_Div_right(n)); -} +} /* computed_value_Div */ /** - * calculate the value of an integer Mod of two nodes + * Calculate the value of an integer Mod of two nodes. * Special case: a % 1 */ -static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) -{ +static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) { tarval *ta = value_of(a); tarval *tb = value_of(b); @@ -220,23 +261,20 @@ static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */ return get_mode_null(get_irn_mode(a)); } - return tarval_bad; -} +} /* do_computed_value_Mod */ /** - * return the value of an integer Mod + * Return the value of an integer Mod. */ -static tarval *computed_value_Mod(ir_node *n) -{ +static tarval *computed_value_Mod(ir_node *n) { return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n)); -} +} /* computed_value_Mod */ /** - * return the value of an Abs + * Return the value of an Abs. */ -static tarval *computed_value_Abs(ir_node *n) -{ +static tarval *computed_value_Abs(ir_node *n) { ir_node *a = get_Abs_op(n); tarval *ta = value_of(a); @@ -244,14 +282,13 @@ static tarval *computed_value_Abs(ir_node *n) return tarval_abs(ta); return tarval_bad; -} +} /* computed_value_Abs */ /** - * return the value of an And + * Return the value of an And. * Special case: a & 0, 0 & b */ -static tarval *computed_value_And(ir_node *n) -{ +static tarval *computed_value_And(ir_node *n) { ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); @@ -269,14 +306,13 @@ static tarval *computed_value_And(ir_node *n) } } return tarval_bad; -} +} /* computed_value_And */ /** - * return the value of an Or + * Return the value of an Or. * Special case: a | 1...1, 1...1 | b */ -static tarval *computed_value_Or(ir_node *n) -{ +static tarval *computed_value_Or(ir_node *n) { ir_node *a = get_Or_left(n); ir_node *b = get_Or_right(n); @@ -293,20 +329,19 @@ static tarval *computed_value_Or(ir_node *n) } } return tarval_bad; -} +} /* computed_value_Or */ /** - * return the value of an Eor + * Return the value of an Eor. */ -static tarval *computed_value_Eor(ir_node *n) -{ +static tarval *computed_value_Eor(ir_node *n) { ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); tarval *ta, *tb; if (a == b) - return get_tarval_null(get_irn_mode(n)); + return get_mode_null(get_irn_mode(n)); ta = value_of(a); tb = value_of(b); @@ -315,13 +350,12 @@ static tarval *computed_value_Eor(ir_node *n) return tarval_eor (ta, tb); } return tarval_bad; -} +} /* computed_value_Eor */ /** - * return the value of a Not + * Return the value of a Not. */ -static tarval *computed_value_Not(ir_node *n) -{ +static tarval *computed_value_Not(ir_node *n) { ir_node *a = get_Not_op(n); tarval *ta = value_of(a); @@ -329,13 +363,12 @@ static tarval *computed_value_Not(ir_node *n) return tarval_not(ta); return tarval_bad; -} +} /* computed_value_Not */ /** - * return the value of a Shl + * Return the value of a Shl. */ -static tarval *computed_value_Shl(ir_node *n) -{ +static tarval *computed_value_Shl(ir_node *n) { ir_node *a = get_Shl_left(n); ir_node *b = get_Shl_right(n); @@ -346,13 +379,12 @@ static tarval *computed_value_Shl(ir_node *n) return tarval_shl (ta, tb); } return tarval_bad; -} +} /* computed_value_Shl */ /** - * return the value of a Shr + * Return the value of a Shr. */ -static tarval *computed_value_Shr(ir_node *n) -{ +static tarval *computed_value_Shr(ir_node *n) { ir_node *a = get_Shr_left(n); ir_node *b = get_Shr_right(n); @@ -363,13 +395,12 @@ static tarval *computed_value_Shr(ir_node *n) return tarval_shr (ta, tb); } return tarval_bad; -} +} /* computed_value_Shr */ /** - * return the value of a Shrs + * Return the value of a Shrs. */ -static tarval *computed_value_Shrs(ir_node *n) -{ +static tarval *computed_value_Shrs(ir_node *n) { ir_node *a = get_Shrs_left(n); ir_node *b = get_Shrs_right(n); @@ -380,10 +411,10 @@ static tarval *computed_value_Shrs(ir_node *n) return tarval_shrs (ta, tb); } return tarval_bad; -} +} /* computed_value_Shrs */ /** - * return the value of a Rot + * Return the value of a Rot. */ static tarval *computed_value_Rot(ir_node *n) { @@ -397,10 +428,10 @@ static tarval *computed_value_Rot(ir_node *n) return tarval_rot (ta, tb); } return tarval_bad; -} +} /* computed_value_Rot */ /** - * return the value of a Conv + * Return the value of a Conv. */ static tarval *computed_value_Conv(ir_node *n) { @@ -411,82 +442,134 @@ static tarval *computed_value_Conv(ir_node *n) return tarval_convert_to(ta, get_irn_mode(n)); return tarval_bad; -} +} /* computed_value_Conv */ /** - * return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), Proj(DivMod) + * Return the value of a Proj(Cmp). + * + * This performs a first step of unreachable code elimination. + * Proj can not be computed, but folding a Cmp above the Proj here is + * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which + * only 1 is used. + * There are several case where we can evaluate a Cmp node, see later. */ -static tarval *computed_value_Proj(ir_node *n) +static tarval *computed_value_Proj_Cmp(ir_node *n) { + ir_node *a = get_Proj_pred(n); + ir_node *aa = get_Cmp_left(a); + ir_node *ab = get_Cmp_right(a); + long proj_nr = get_Proj_proj(n); + + /* + * BEWARE: a == a is NOT always True for floating Point values, as + * NaN != NaN is defined, so we must check this here. + */ + if (aa == ab && ( + !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt) + ) { /* 1.: */ + + /* This is a trick with the bits used for encoding the Cmp + Proj numbers, the following statement is not the same: + return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */ + return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b); + } + else { + tarval *taa = value_of(aa); + tarval *tab = value_of(ab); + ir_mode *mode = get_irn_mode(aa); + + /* + * The predecessors of Cmp are target values. We can evaluate + * the Cmp. + */ + if ((taa != tarval_bad) && (tab != tarval_bad)) { + /* strange checks... */ + pn_Cmp flags = tarval_cmp(taa, tab); + if (flags != pn_Cmp_False) { + return new_tarval_from_long (proj_nr & flags, mode_b); + } + } + /* for integer values, we can check against MIN/MAX */ + else if (mode_is_int(mode)) { + /* MIN <=/> x. This results in true/false. */ + if (taa == get_mode_min(mode)) { + /* a compare with the MIN value */ + if (proj_nr == pn_Cmp_Le) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Gt) + return get_tarval_b_false(); + } + /* x >=/< MIN. This results in true/false. */ + else + if (tab == get_mode_min(mode)) { + /* a compare with the MIN value */ + if (proj_nr == pn_Cmp_Ge) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Lt) + return get_tarval_b_false(); + } + /* MAX >=/< x. This results in true/false. */ + else if (taa == get_mode_max(mode)) { + if (proj_nr == pn_Cmp_Ge) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Lt) + return get_tarval_b_false(); + } + /* x <=/> MAX. This results in true/false. */ + else if (tab == get_mode_max(mode)) { + if (proj_nr == pn_Cmp_Le) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Gt) + return get_tarval_b_false(); + } + } + /* + * The predecessors are Allocs or (void*)(0) constants. Allocs never + * return NULL, they raise an exception. Therefore we can predict + * the Cmp result. + */ + else { + ir_node *aaa = skip_Id(skip_Proj(aa)); + ir_node *aba = skip_Id(skip_Proj(ab)); + + if ( ( (/* aa is ProjP and aaa is Alloc */ + (get_irn_op(aa) == op_Proj) + && (mode_is_reference(get_irn_mode(aa))) + && (get_irn_op(aaa) == op_Alloc)) + && ( (/* ab is NULL */ + (get_irn_op(ab) == op_Const) + && (mode_is_reference(get_irn_mode(ab))) + && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) + || (/* ab is other Alloc */ + (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc) + && (aaa != aba)))) + || (/* aa is NULL and aba is Alloc */ + (get_irn_op(aa) == op_Const) + && (mode_is_reference(get_irn_mode(aa))) + && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) + && (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc))) + /* 3.: */ + return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); + } + } + return computed_value_Cmp_Confirm(a, aa, ab, proj_nr); +} /* computed_value_Proj_Cmp */ + +/** + * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), + * Proj(DivMod) and Proj(Quot). + */ +static tarval *computed_value_Proj(ir_node *n) { ir_node *a = get_Proj_pred(n); - ir_node *aa, *ab; long proj_nr; - /* Optimize Cmp nodes. - This performs a first step of unreachable code elimination. - Proj can not be computed, but folding a Cmp above the Proj here is - not as wasteful as folding a Cmp into a Tuple of 16 Consts of which - only 1 is used. - There are several case where we can evaluate a Cmp node: - 1. The nodes compared are both the same. If we compare for - equal, greater equal, ... this will return true, else it - will return false. This step relies on cse. - 2. The predecessors of Cmp are target values. We can evaluate - the Cmp. - 3. The predecessors are Allocs or void* constants. Allocs never - return NULL, they raise an exception. Therefore we can predict - the Cmp result. */ switch (get_irn_opcode(a)) { case iro_Cmp: - aa = get_Cmp_left(a); - ab = get_Cmp_right(a); - proj_nr = get_Proj_proj(n); - - if (aa == ab && !mode_is_float(get_irn_mode(aa))) { /* 1.: */ - /* BEWARE: a == a is NOT always True for floating Point!!! */ - /* This is a trick with the bits used for encoding the Cmp - Proj numbers, the following statement is not the same: - return new_tarval_from_long (proj_nr == Eq, mode_b) */ - return new_tarval_from_long (proj_nr & Eq, mode_b); - } else { - tarval *taa = value_of(aa); - tarval *tab = value_of(ab); - - if ((taa != tarval_bad) && (tab != tarval_bad)) { /* 2.: */ - /* strange checks... */ - pnc_number flags = tarval_cmp (taa, tab); - if (flags != False) { - return new_tarval_from_long (proj_nr & flags, mode_b); - } - } else { /* check for 3.: */ - ir_node *aaa = skip_Id(skip_Proj(aa)); - ir_node *aba = skip_Id(skip_Proj(ab)); - - if ( ( (/* aa is ProjP and aaa is Alloc */ - (get_irn_op(aa) == op_Proj) - && (mode_is_reference(get_irn_mode(aa))) - && (get_irn_op(aaa) == op_Alloc)) - && ( (/* ab is constant void */ - (get_irn_op(ab) == op_Const) - && (mode_is_reference(get_irn_mode(ab))) - && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) - || (/* ab is other Alloc */ - (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc) - && (aaa != aba)))) - || (/* aa is void and aba is Alloc */ - (get_irn_op(aa) == op_Const) - && (mode_is_reference(get_irn_mode(aa))) - && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) - && (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc))) - /* 3.: */ - return new_tarval_from_long (proj_nr & Ne, mode_b); - } - } - break; + return computed_value_Proj_Cmp(n); case iro_DivMod: /* compute either the Div or the Mod part */ @@ -507,11 +590,55 @@ static tarval *computed_value_Proj(ir_node *n) return computed_value(a); break; + case iro_Quot: + if (get_Proj_proj(n) == pn_Quot_res) + return computed_value(a); + break; + default: return tarval_bad; } return tarval_bad; -} +} /* computed_value_Proj */ + +/** + * Calculate the value of a Mux: can be evaluated, if the + * sel and the right input are known. + */ +static tarval *computed_value_Mux(ir_node *n) { + ir_node *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); + + if (ts == get_tarval_b_true()) { + ir_node *v = get_Mux_true(n); + return value_of(v); + } + else if (ts == get_tarval_b_false()) { + ir_node *v = get_Mux_false(n); + return value_of(v); + } + return tarval_bad; +} /* computed_value_Mux */ + +/** + * Calculate the value of a Psi: can be evaluated, if a condition is true + * and all previous conditions are false. If all conditions are false + * we evaluate to the default one. + */ +static tarval *computed_value_Psi(ir_node *n) { + if (is_Mux(n)) + return computed_value_Mux(n); + return tarval_bad; +} /* computed_value_Psi */ + +/** + * Calculate the value of a Confirm: can be evaluated, + * if it has the form Confirm(x, '=', Const). + */ +static tarval *computed_value_Confirm(ir_node *n) { + return get_Confirm_cmp(n) == pn_Cmp_Eq ? + value_of(get_Confirm_bound(n)) : tarval_bad; +} /* computed_value_Confirm */ /** * If the parameter n can be computed, return its value, else tarval_bad. @@ -519,24 +646,29 @@ static tarval *computed_value_Proj(ir_node *n) * * @param n The node this should be evaluated */ -tarval *computed_value(ir_node *n) -{ - if (n->op->computed_value) - return n->op->computed_value(n); +tarval *computed_value(ir_node *n) { + if (n->op->ops.computed_value) + return n->op->ops.computed_value(n); return tarval_bad; -} +} /* computed_value */ /** - * set the default computed_value evaluator + * Set the default computed_value evaluator in an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_computed_value(ir_op *op) +static ir_op_ops *firm_set_default_computed_value(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->computed_value = computed_value_##a; \ + ops->computed_value = computed_value_##a; \ break - switch (op->code) { + switch (code) { CASE(Const); CASE(SymConst); CASE(Add); @@ -555,38 +687,41 @@ static ir_op *firm_set_default_computed_value(ir_op *op) CASE(Shr); CASE(Shrs); CASE(Rot); + CASE(Carry); + CASE(Borrow); CASE(Conv); CASE(Proj); + CASE(Mux); + CASE(Psi); + CASE(Confirm); default: - op->computed_value = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE -} - -#if 0 -/* returns 1 if the a and b are pointers to different locations. */ -static bool -different_identity (ir_node *a, ir_node *b) -{ - assert (mode_is_reference(get_irn_mode (a)) - && mode_is_reference(get_irn_mode (b))); - - if (get_irn_op (a) == op_Proj && get_irn_op(b) == op_Proj) { - ir_node *a1 = get_Proj_pred (a); - ir_node *b1 = get_Proj_pred (b); - if (a1 != b1 && get_irn_op (a1) == op_Alloc - && get_irn_op (b1) == op_Alloc) - return 1; - } - return 0; -} -#endif +} /* firm_set_default_computed_value */ +/** + * Returns a equivalent block for another block. + * If the block has only one predecessor, this is + * the equivalent one. If the only predecessor of a block is + * the block itself, this is a dead block. + * + * If both predecessors of a block are the branches of a binary + * Cond, the equivalent block is Cond's block. + * + * If all predecessors of a block are bad or lies in a dead + * block, the current block is dead as well. + * + * Note, that blocks are NEVER turned into Bad's, instead + * the dead_block flag is set. So, never test for is_Bad(block), + * always use is_dead_Block(block). + */ static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; + int n_preds = get_Block_n_cfgpreds(n); /* The Block constructor does not call optimize, but mature_immBlock calls the optimization. */ @@ -598,31 +733,32 @@ static ir_node *equivalent_node_Block(ir_node *n) This should be true, as the block is matured before optimize is called. But what about Phi-cycles with the Phi0/Id that could not be resolved? Remaining Phi nodes are just Ids. */ - if ((get_Block_n_cfgpreds(n) == 1) && - (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { + if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = set_Block_dead(n); - DBG_OPT_DEAD(oldn, n); + DBG_OPT_DEAD_BLOCK(oldn, n); } else if (get_opt_control_flow_straightening()) { n = predblock; DBG_OPT_STG(oldn, n); } } - else if ((get_Block_n_cfgpreds(n) == 1) && + else if ((n_preds == 1) && (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { - ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); + ir_node *predblock = get_Block_cfgpred_block(n, 0); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ n = set_Block_dead(n); - DBG_OPT_DEAD(oldn, n); + DBG_OPT_DEAD_BLOCK(oldn, n); } } - else if ((get_Block_n_cfgpreds(n) == 2) && + else if ((n_preds == 2) && (get_opt_control_flow_weak_simplification())) { /* Test whether Cond jumps twice to this block - @@@ we could do this also with two loops finding two preds from several ones. */ + * The more general case which more than 2 predecessors is handles + * in optimize_cf(), we handle only this special case for speed here. + */ ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); @@ -633,23 +769,24 @@ static ir_node *equivalent_node_Block(ir_node *n) (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { /* Also a single entry Block following a single exit Block. Phis have twice the same operand and will be optimized away. */ - n = get_nodes_block(a); - DBG_OPT_IFSIM(oldn, a, b, n); + n = get_nodes_block(get_Proj_pred(a)); + DBG_OPT_IFSIM1(oldn, a, b, n); } - } else if (get_opt_unreachable_code() && - (n != current_ir_graph->start_block) && - (n != current_ir_graph->end_block) ) { - int i, n_cfg = get_Block_n_cfgpreds(n); + } + else if (get_opt_unreachable_code() && + (n != get_irg_start_block(current_ir_graph)) && + (n != get_irg_end_block(current_ir_graph)) ) { + int i; /* If all inputs are dead, this block is dead too, except if it is - the start or end block. This is a step of unreachable code + the start or end block. This is one step of unreachable code elimination */ - for (i = 0; i < n_cfg; i++) { + for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) { ir_node *pred = get_Block_cfgpred(n, i); ir_node *pred_blk; if (is_Bad(pred)) continue; - pred_blk = get_nodes_block(pred); + pred_blk = get_nodes_block(skip_Proj(pred)); if (is_Block_dead(pred_blk)) continue; @@ -658,36 +795,36 @@ static ir_node *equivalent_node_Block(ir_node *n) break; } } - if (i == n_cfg) + if (i < 0) { n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); + } } return n; -} +} /* equivalent_node_Block */ /** * Returns a equivalent node for a Jmp, a Bad :-) - * Of course this only happens if the Block of the Jmp is Bad. + * Of course this only happens if the Block of the Jmp is dead. */ -static ir_node *equivalent_node_Jmp(ir_node *n) -{ - /* GL: Why not same for op_Raise?? */ +static ir_node *equivalent_node_Jmp(ir_node *n) { /* unreachable code elimination */ if (is_Block_dead(get_nodes_block(n))) n = new_Bad(); return n; -} +} /* equivalent_node_Jmp */ -static ir_node *equivalent_node_Cond(ir_node *n) -{ - /* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See cases for iro_Cond and iro_Proj in transform_node. */ - return n; -} +/** Raise is handled in the same way as Jmp. */ +#define equivalent_node_Raise equivalent_node_Jmp + + +/* We do not evaluate Cond here as we replace it by a new node, a Jmp. + See transform_node_Proj_Cond(). */ /** - * optimize operations that are commutative and have neutral 0, + * Optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. */ static ir_node *equivalent_node_neutral_zero(ir_node *n) @@ -710,25 +847,84 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) return n; /* If this predecessors constant value is zero, the operation is - unnecessary. Remove it: */ + * unnecessary. Remove it. + * + * Beware: If n is a Add, the mode of on and n might be different + * which happens in this rare construction: NULL + 3. + * Then, a Conv would be needed which we cannot include here. + */ if (classify_tarval (tv) == TV_CLASSIFY_NULL) { - n = on; + if (get_irn_mode(on) == get_irn_mode(n)) { + n = on; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); + } } return n; -} +} /* equivalent_node_neutral_zero */ -#define equivalent_node_Add equivalent_node_neutral_zero +/** + * Eor is commutative and has neutral 0. + */ #define equivalent_node_Eor equivalent_node_neutral_zero +/* + * Optimize a - 0 and (a - x) + x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of an Add may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. + */ +static ir_node *equivalent_node_Add(ir_node *n) +{ + ir_node *oldn = n; + ir_node *left, *right; + ir_mode *mode = get_irn_mode(n); + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + n = equivalent_node_neutral_zero(n); + if (n != oldn) + return n; + + left = get_Add_left(n); + right = get_Add_right(n); + + if (get_irn_op(left) == op_Sub) { + if (get_Sub_right(left) == right) { + /* (a - x) + x */ + + n = get_Sub_left(left); + if (mode == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + if (get_irn_op(right) == op_Sub) { + if (get_Sub_right(right) == left) { + /* x + (a - x) */ + + n = get_Sub_left(right); + if (mode == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + return n; +} /* equivalent_node_Add */ + /** * optimize operations that are not commutative but have neutral 0 on left, * so a op 0 = a. */ -static ir_node *equivalent_node_left_zero(ir_node *n) -{ +static ir_node *equivalent_node_left_zero(ir_node *n) { ir_node *oldn = n; ir_node *a = get_binop_left(n); @@ -737,22 +933,78 @@ static ir_node *equivalent_node_left_zero(ir_node *n) if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } - return n; -} +} /* equivalent_node_left_zero */ -#define equivalent_node_Sub equivalent_node_left_zero #define equivalent_node_Shl equivalent_node_left_zero #define equivalent_node_Shr equivalent_node_left_zero #define equivalent_node_Shrs equivalent_node_left_zero #define equivalent_node_Rot equivalent_node_left_zero /** - * Er, a "symmetic unop", ie op(op(n)) = n. + * Optimize a - 0 and (a + x) - x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of a Sub may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. + */ +static ir_node *equivalent_node_Sub(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a, *b; + ir_mode *mode = get_irn_mode(n); + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + a = get_Sub_left(n); + b = get_Sub_right(n); + + /* Beware: modes might be different */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + if (mode == get_irn_mode(a)) { + n = a; + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); + } + } + else if (get_irn_op(a) == op_Add) { + if (mode_wrap_around(mode)) { + ir_node *left = get_Add_left(a); + ir_node *right = get_Add_right(a); + + if (left == b) { + if (mode == get_irn_mode(right)) { + n = right; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + } + } + else if (right == b) { + if (mode == get_irn_mode(left)) { + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + } + } + } + } + return n; +} /* equivalent_node_Sub */ + + +/** + * Optimize an "idempotent unary op", ie op(op(n)) = n. + * + * @todo + * -(-a) == a, but might overflow two times. + * We handle it anyway here but the better way would be a + * flag. This would be needed for Pascal for instance. */ -static ir_node *equivalent_node_symmetric_unop(ir_node *n) +static ir_node *equivalent_node_idempotent_unop(ir_node *n) { ir_node *oldn = n; ir_node *pred = get_unop_op(n); @@ -763,14 +1015,14 @@ static ir_node *equivalent_node_symmetric_unop(ir_node *n) DBG_OPT_ALGSIM2(oldn, pred, n); } return n; -} +} /* equivalent_node_idempotent_unop */ -/* NotNot x == x */ -#define equivalent_node_Not equivalent_node_symmetric_unop +/** Optimize Not(Not(x)) == x. */ +#define equivalent_node_Not equivalent_node_idempotent_unop -/* --x == x */ /* ??? Is this possible or can --x raise an +/** --x == x ??? Is this possible or can --x raise an out of bounds exception if min =! max? */ -#define equivalent_node_Minus equivalent_node_symmetric_unop +#define equivalent_node_Minus equivalent_node_idempotent_unop /** * Optimize a * 1 = 1 * a = a. @@ -778,26 +1030,24 @@ static ir_node *equivalent_node_symmetric_unop(ir_node *n) static ir_node *equivalent_node_Mul(ir_node *n) { ir_node *oldn = n; - ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); /* Mul is commutative and has again an other neutral element. */ if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); } return n; -} +} /* equivalent_node_Mul */ /** * Optimize a / 1 = a. */ -static ir_node *equivalent_node_Div(ir_node *n) -{ +static ir_node *equivalent_node_Div(ir_node *n) { ir_node *a = get_Div_left(n); ir_node *b = get_Div_right(n); @@ -805,19 +1055,37 @@ static ir_node *equivalent_node_Div(ir_node *n) if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); - turn_into_tuple(n, 3); + turn_into_tuple(n, pn_Div_max); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_Div_res, a); } return n; -} +} /* equivalent_node_Div */ + +/** + * Optimize a / 1.0 = a. + */ +static ir_node *equivalent_node_Quot(ir_node *n) { + ir_node *a = get_Quot_left(n); + ir_node *b = get_Quot_right(n); + + /* Div is not commutative. */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */ + /* Turn Quot into a tuple (mem, bad, a) */ + ir_node *mem = get_Quot_mem(n); + turn_into_tuple(n, pn_Quot_max); + set_Tuple_pred(n, pn_Quot_M, mem); + set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Quot_res, a); + } + return n; +} /* equivalent_node_Quot */ /** * Optimize a / 1 = a. */ -static ir_node *equivalent_node_DivMod(ir_node *n) -{ +static ir_node *equivalent_node_DivMod(ir_node *n) { ir_node *a = get_DivMod_left(n); ir_node *b = get_DivMod_right(n); @@ -827,20 +1095,19 @@ static ir_node *equivalent_node_DivMod(ir_node *n) ir_node *mem = get_Div_mem(n); ir_mode *mode = get_irn_mode(b); - turn_into_tuple(n, 4); + turn_into_tuple(n, pn_DivMod_max); set_Tuple_pred(n, pn_DivMod_M, mem); set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_DivMod_res_div, a); set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode))); } return n; -} +} /* equivalent_node_DivMod */ /** * Use algebraic simplification a | a = a | 0 = 0 | a = a. */ -static ir_node *equivalent_node_Or(ir_node *n) -{ +static ir_node *equivalent_node_Or(ir_node *n) { ir_node *oldn = n; ir_node *a = get_Or_left(n); @@ -848,22 +1115,22 @@ static ir_node *equivalent_node_Or(ir_node *n) if (a == b) { n = a; /* Or has it's own neutral element */ + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); } return n; -} +} /* equivalent_node_Or */ /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. */ -static ir_node *equivalent_node_And(ir_node *n) -{ +static ir_node *equivalent_node_And(ir_node *n) { ir_node *oldn = n; ir_node *a = get_And_left(n); @@ -871,21 +1138,21 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) { n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); } return n; -} +} /* equivalent_node_And */ /** - * Try to remove useless conv's: + * Try to remove useless Conv's: */ -static ir_node *equivalent_node_Conv(ir_node *n) -{ +static ir_node *equivalent_node_Conv(ir_node *n) { ir_node *oldn = n; ir_node *a = get_Conv_op(n); ir_node *b; @@ -894,8 +1161,11 @@ static ir_node *equivalent_node_Conv(ir_node *n) ir_mode *a_mode = get_irn_mode(a); if (n_mode == a_mode) { /* No Conv necessary */ + /* leave strict floating point Conv's */ + if (get_Conv_strict(n)) + return n; n = a; - DBG_OPT_ALGSIM3(oldn, a, n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ ir_mode *b_mode; @@ -906,36 +1176,41 @@ static ir_node *equivalent_node_Conv(ir_node *n) if (n_mode == b_mode) { if (n_mode == mode_b) { n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { if (smaller_mode(b_mode, a_mode)){ n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } } } } return n; -} +} /* equivalent_node_Conv */ /** * A Cast may be removed if the type of the previous node - * is already to type of the Cast. + * is already the type of the Cast. */ static ir_node *equivalent_node_Cast(ir_node *n) { + ir_node *oldn = n; ir_node *pred = get_Cast_op(n); - if (get_irn_type(pred) == get_Cast_type(n)) + + if (get_irn_type(pred) == get_Cast_type(n)) { n = pred; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST); + } return n; -} +} /* equivalent_node_Cast */ -/* Several optimizations: - - no Phi in start block. - - remove Id operators that are inputs to Phi - - fold Phi-nodes, iff they have only one predecessor except - themselves. -*/ +/** + * Several optimizations: + * - no Phi in start block. + * - remove Id operators that are inputs to Phi + * - fold Phi-nodes, iff they have only one predecessor except + * themselves. + */ static ir_node *equivalent_node_Phi(ir_node *n) { int i, n_preds; @@ -943,7 +1218,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) ir_node *oldn = n; ir_node *block = NULL; /* to shutup gcc */ ir_node *first_val = NULL; /* to shutup gcc */ - ir_node *scnd_val = NULL; /* to shutup gcc */ if (!get_opt_normalize()) return n; @@ -953,85 +1227,114 @@ static ir_node *equivalent_node_Phi(ir_node *n) /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!! assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */ if ((is_Block_dead(block)) || /* Control dead */ - (block == current_ir_graph->start_block)) /* There should be no Phi nodes */ - return new_Bad(); /* in the Start Block. */ + (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */ + return new_Bad(); /* in the Start Block. */ if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ -#if 0 - /* first we test for a special case: */ - /* Confirm is a special node fixing additional information for a - value that is known at a certain point. This is useful for - dataflow analysis. */ - if (n_preds == 2) { - ir_node *a = get_Phi_pred(n, 0); - ir_node *b = get_Phi_pred(n, 1); - if ( (get_irn_op(a) == op_Confirm) - && (get_irn_op(b) == op_Confirm) - && follow_Id (get_irn_n(a, 0) == get_irn_n(b, 0)) - && (get_irn_n(a, 1) == get_irn_n (b, 1)) - && (a->data.num == (~b->data.num & irpn_True) )) { - return get_irn_n(a, 0); - } - } -#endif - /* If the Block has a Bad pred, we also have one. */ for (i = 0; i < n_preds; ++i) - if (is_Bad (get_Block_cfgpred(block, i))) + if (is_Bad(get_Block_cfgpred(block, i))) set_Phi_pred(n, i, new_Bad()); /* Find first non-self-referencing input */ - for (i = 0; i < n_preds; ++i) { + for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ #if 1 - && (get_irn_op(first_val) != op_Bad) + && (! is_Bad(first_val)) #endif ) { /* value not dead */ break; /* then found first value. */ } } - /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - if (i >= n_preds) { return new_Bad(); } - - scnd_val = NULL; + if (i >= n_preds) { + /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ + return new_Bad(); + } - /* follow_Id () for rest of inputs, determine if any of these + /* search for rest of inputs, determine if any of these are non-self-referencing */ while (++i < n_preds) { - scnd_val = get_Phi_pred(n, i); + ir_node *scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) && (scnd_val != first_val) #if 1 - && (get_irn_op(scnd_val) != op_Bad) + && (! is_Bad(scnd_val)) #endif ) { break; } } - /* Fold, if no multiple distinct non-self-referencing inputs */ if (i >= n_preds) { + /* Fold, if no multiple distinct non-self-referencing inputs */ n = first_val; - DBG_OPT_PHI(oldn, first_val, n); - } else { - /* skip the remaining Ids (done in get_Phi_pred). */ - /* superfluous, since we walk all to propagate Block's Bads. - while (++i < n_preds) get_Phi_pred(n, i); */ + DBG_OPT_PHI(oldn, n); } return n; -} +} /* equivalent_node_Phi */ /** - * optimize Proj(Tuple) and gigo for ProjX in Bad block + * Several optimizations: + * - no Sync in start block. + * - fold Sync-nodes, iff they have only one predecessor except + * themselves. */ -static ir_node *equivalent_node_Proj(ir_node *n) +static ir_node *equivalent_node_Sync(ir_node *n) { + int i, n_preds; + ir_node *oldn = n; + ir_node *first_val = NULL; /* to shutup gcc */ - ir_node *a = get_Proj_pred(n); + if (!get_opt_normalize()) return n; + + n_preds = get_Sync_n_preds(n); + + /* Find first non-self-referencing input */ + for (i = 0; i < n_preds; ++i) { + first_val = get_Sync_pred(n, i); + if ((first_val != n) /* not self pointer */ && + (! is_Bad(first_val)) + ) { /* value not dead */ + break; /* then found first value. */ + } + } + + if (i >= n_preds) + /* A totally Bad or self-referencing Sync (we didn't break the above loop) */ + return new_Bad(); + + /* search the rest of inputs, determine if any of these + are non-self-referencing */ + while (++i < n_preds) { + ir_node *scnd_val = get_Sync_pred(n, i); + if ((scnd_val != n) && + (scnd_val != first_val) && + (! is_Bad(scnd_val)) + ) + break; + } + + if (i >= n_preds) { + /* Fold, if no multiple distinct non-self-referencing inputs */ + n = first_val; + DBG_OPT_SYNC(oldn, n); + } + return n; +} /* equivalent_node_Sync */ + +/** + * Optimize Proj(Tuple) and gigo() for ProjX in Bad block, + * ProjX(Load) and ProjX(Store). + */ +static ir_node *equivalent_node_Proj(ir_node *n) +{ + ir_node *oldn = n; + + ir_node *a = get_Proj_pred(n); if ( get_irn_op(a) == op_Tuple) { /* Remove the Tuple/Proj combination. */ @@ -1042,26 +1345,248 @@ static ir_node *equivalent_node_Proj(ir_node *n) assert(0); /* This should not happen! */ n = new_Bad(); } - } else if (get_irn_mode(n) == mode_X && - is_Block_dead(get_nodes_block(n))) { - /* Remove dead control flow -- early gigo. */ - n = new_Bad(); } + else if (get_irn_mode(n) == mode_X) { + if (is_Block_dead(get_nodes_block(skip_Proj(n)))) { + /* Remove dead control flow -- early gigo(). */ + n = new_Bad(); + } + else if (get_opt_ldst_only_null_ptr_exceptions()) { + ir_op *op = get_irn_op(a); + + if (op == op_Load || op == op_Store) { + /* get the load/store address */ + ir_node *addr = get_irn_n(a, 1); + ir_node *confirm; + + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(a, op_pin_state_floats); + } + DBG_OPT_EXC_REM(n); + return new_Bad(); + } + } + } + } + return n; -} +} /* equivalent_node_Proj */ /** * Remove Id's. */ -static ir_node *equivalent_node_Id(ir_node *n) -{ +static ir_node *equivalent_node_Id(ir_node *n) { ir_node *oldn = n; - n = follow_Id(n); + do { + n = get_Id_pred(n); + } while (get_irn_op(n) == op_Id); + DBG_OPT_ID(oldn, n); return n; +} /* equivalent_node_Id */ + +/** + * Optimize a Mux. + */ +static ir_node *equivalent_node_Mux(ir_node *n) +{ + ir_node *oldn = n, *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); + + /* Mux(true, f, t) == t */ + if (ts == tarval_b_true) { + n = get_Mux_true(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); + } + /* Mux(false, f, t) == f */ + else if (ts == tarval_b_false) { + n = get_Mux_false(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); + } + /* Mux(v, x, x) == x */ + else if (get_Mux_false(n) == get_Mux_true(n)) { + n = get_Mux_true(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ); + } + else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) { + ir_node *cmp = get_Proj_pred(sel); + long proj_nr = get_Proj_proj(sel); + ir_node *b = get_Mux_false(n); + ir_node *a = get_Mux_true(n); + + /* + * Note: normalization puts the constant on the right site, + * so we check only one case. + * + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the first one. + */ + if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) { + if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { + /* Mux(a CMP 0, X, a) */ + if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) { + /* Mux(a CMP 0, -a, a) */ + if (proj_nr == pn_Cmp_Eq) { + /* Mux(a == 0, -a, a) ==> -a */ + n = b; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(a != 0, -a, a) ==> a */ + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + } + else if (classify_Const(b) == CNST_NULL) { + /* Mux(a CMP 0, 0, a) */ + if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(a != 0, 0, a) ==> a */ + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + else if (proj_nr == pn_Cmp_Eq) { + /* Mux(a == 0, 0, a) ==> 0 */ + n = b; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + } + } + } + } + return n; +} /* equivalent_node_Mux */ + +/** + * Returns a equivalent node of a Psi: if a condition is true + * and all previous conditions are false we know its value. + * If all conditions are false its value is the default one. + */ +static ir_node *equivalent_node_Psi(ir_node *n) { + if (is_Mux(n)) + return equivalent_node_Mux(n); + return n; +} /* equivalent_node_Psi */ + +/** + * Optimize -a CMP -b into b CMP a. + * This works only for for modes where unary Minus + * cannot Overflow. + * Note that two-complement integers can Overflow + * so it will NOT work. + */ +static ir_node *equivalent_node_Cmp(ir_node *n) +{ + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + + if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus && + !mode_overflow_on_unary_Minus(get_irn_mode(left))) { + left = get_Minus_op(left); + right = get_Minus_op(right); + set_Cmp_left(n, right); + set_Cmp_right(n, left); + } + return n; +} /* equivalent_node_Cmp */ + +/** + * Remove Confirm nodes if setting is on. + * Replace Confirms(x, '=', Constlike) by Constlike. + */ +static ir_node *equivalent_node_Confirm(ir_node *n) { + ir_node *pred = get_Confirm_value(n); + pn_Cmp pnc = get_Confirm_cmp(n); + + if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) { + /* + * rare case: two identical Confirms one after another, + * replace the second one with the first. + */ + n = pred; + } + if (pnc == pn_Cmp_Eq) { + ir_node *bound = get_Confirm_bound(n); + + /* + * Optimize a rare case: + * Confirm(x, '=', Constlike) ==> Constlike + */ + if (is_irn_constlike(bound)) { + DBG_OPT_CONFIRM(n, bound); + return bound; + } + } + return get_opt_remove_confirm() ? get_Confirm_value(n) : n; } +/** + * Optimize CopyB(mem, x, x) into a Nop. + */ +static ir_node *equivalent_node_CopyB(ir_node *n) { + ir_node *a = get_CopyB_dst(n); + ir_node *b = get_CopyB_src(n); + + if (a == b) { + /* Turn CopyB into a tuple (mem, bad, bad) */ + ir_node *mem = get_CopyB_mem(n); + turn_into_tuple(n, pn_CopyB_max); + set_Tuple_pred(n, pn_CopyB_M, mem); + set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_CopyB_M_except, new_Bad()); + } + return n; +} /* equivalent_node_CopyB */ + +/** + * Optimize Bounds(idx, idx, upper) into idx. + */ +static ir_node *equivalent_node_Bound(ir_node *n) { + ir_node *idx = get_Bound_index(n); + ir_node *lower = get_Bound_lower(n); + int ret_tuple = 0; + + /* By definition lower < upper, so if idx == lower --> + lower <= idx && idx < upper */ + if (idx == lower) { + /* Turn Bound into a tuple (mem, bad, idx) */ + ret_tuple = 1; + } + else { + ir_node *pred = skip_Proj(idx); + + if (get_irn_op(pred) == op_Bound) { + /* + * idx was Bounds_check previously, it is still valid if + * lower <= pred_lower && pred_upper <= upper. + */ + ir_node *upper = get_Bound_upper(n); + if (get_Bound_lower(pred) == lower && + get_Bound_upper(pred) == upper) { + /* + * One could expect that we simply return the previous + * Bound here. However, this would be wrong, as we could + * add an exception Proj to a new location than. + * So, we must turn in into a tuple + */ + ret_tuple = 1; + } + } + } + if (ret_tuple) { + /* Turn Bound into a tuple (mem, bad, idx) */ + ir_node *mem = get_Bound_mem(n); + turn_into_tuple(n, pn_Bound_max); + set_Tuple_pred(n, pn_Bound_M, mem); + set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Bound_res, idx); + } + return n; +} /* equivalent_node_Bound */ + /** * equivalent_node() returns a node equivalent to input n. It skips all nodes that * perform no actual computation, as, e.g., the Id nodes. It does not create @@ -1069,28 +1594,32 @@ static ir_node *equivalent_node_Id(ir_node *n) * If a node returns a Tuple we can not just skip it. If the size of the * in array fits, we transform n into a tuple (e.g., Div). */ -ir_node * -equivalent_node(ir_node *n) -{ - if (n->op->equivalent_node) - return n->op->equivalent_node(n); +ir_node *equivalent_node(ir_node *n) { + if (n->op->ops.equivalent_node) + return n->op->ops.equivalent_node(n); return n; -} +} /* equivalent_node */ /** - * set the default equivalent node operation + * Sets the default equivalent node operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_equivalent_node(ir_op *op) +static ir_op_ops *firm_set_default_equivalent_node(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->equivalent_node = equivalent_node_##a; \ + ops->equivalent_node = equivalent_node_##a; \ break - switch (op->code) { + switch (code) { CASE(Block); CASE(Jmp); - CASE(Cond); + CASE(Raise); CASE(Or); CASE(Add); CASE(Eor); @@ -1103,26 +1632,33 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(Minus); CASE(Mul); CASE(Div); + CASE(Quot); CASE(DivMod); CASE(And); CASE(Conv); CASE(Cast); CASE(Phi); + CASE(Sync); CASE(Proj); CASE(Id); + CASE(Mux); + CASE(Psi); + CASE(Cmp); + CASE(Confirm); + CASE(CopyB); + CASE(Bound); default: - op->equivalent_node = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE -} +} /* firm_set_default_equivalent_node */ /** * Do node specific optimizations of nodes predecessors. */ -static void -optimize_preds(ir_node *n) { +static void optimize_preds(ir_node *n) { ir_node *a = NULL, *b = NULL; /* get the operands we will work on for simple cases. */ @@ -1149,11 +1685,120 @@ optimize_preds(ir_node *n) { default: break; } /* end switch */ -} +} /* optimize_preds */ + +/** + * Returns non-zero if a node is a Phi node + * with all predecessors constant. + */ +static int is_const_Phi(ir_node *n) { + int i; + + if (! is_Phi(n)) + return 0; + for (i = get_irn_arity(n) - 1; i >= 0; --i) + if (! is_Const(get_irn_n(n, i))) + return 0; + return 1; +} /* is_const_Phi */ + +/** + * Apply an evaluator on a binop with a constant operators (and one Phi). + * + * @param phi the Phi node + * @param other the other operand + * @param eval an evaluator function + * @param left if non-zero, other is the left operand, else the right + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) { + tarval *tv; + void **res; + ir_node *pred; + ir_mode *mode; + ir_graph *irg; + int i, n = get_irn_arity(phi); + + NEW_ARR_A(void *, res, n); + if (left) { + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(other, tv); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return NULL; + } + res[i] = tv; + } + } + else { + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(tv, other); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return 0; + } + res[i] = tv; + } + } + mode = get_irn_mode(phi); + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + res[i] = new_r_Const_type(irg, get_irg_start_block(irg), + mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); +} /* apply_binop_on_phi */ + +/** + * Apply an evaluator on a unop with a constant operator (a Phi). + * + * @param phi the Phi node + * @param eval an evaluator function + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) { + tarval *tv; + void **res; + ir_node *pred; + ir_mode *mode; + ir_graph *irg; + int i, n = get_irn_arity(phi); + + NEW_ARR_A(void *, res, n); + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = eval(tv); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return 0; + } + res[i] = tv; + } + mode = get_irn_mode(phi); + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + res[i] = new_r_Const_type(irg, get_irg_start_block(irg), + mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); +} /* apply_unop_on_phi */ /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and - * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)) if possible. + * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)). + * If possible, remove the Conv's. */ static ir_node *transform_node_AddSub(ir_node *n) { @@ -1211,16 +1856,303 @@ static ir_node *transform_node_AddSub(ir_node *n) } } return n; -} +} /* transform_node_AddSub */ + +#define HANDLE_BINOP_PHI(op,a,b,c) \ + c = NULL; \ + if (is_Const(b) && is_const_Phi(a)) { \ + /* check for Op(Phi, Const) */ \ + c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \ + } \ + else if (is_Const(a) && is_const_Phi(b)) { \ + /* check for Op(Const, Phi) */ \ + c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \ + } \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ + } + +#define HANDLE_UNOP_PHI(op,a,c) \ + c = NULL; \ + if (is_const_Phi(a)) { \ + /* check for Op(Phi) */ \ + c = apply_unop_on_phi(a, op); \ + } \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ + } + -#define transform_node_Add transform_node_AddSub -#define transform_node_Sub transform_node_AddSub +/** + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Add(a,a) -> Mul(a, 2) + * Add(Mul(a, x), a) -> Mul(a, x+1) + * if the mode is integer or float. + * Transform Add(a,-b) into Sub(a,b). + * Reassociation might fold this further. + */ +static ir_node *transform_node_Add(ir_node *n) +{ + ir_mode *mode; + ir_node *a, *b, *c, *oldn = n; + + n = transform_node_AddSub(n); + + a = get_Add_left(n); + b = get_Add_right(n); + + HANDLE_BINOP_PHI(tarval_add, a,b,c); + + mode = get_irn_mode(n); -/** Do architecture dependend optimizations on Mul nodes */ + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + if (mode_is_num(mode)) { + if (a == b) { + ir_node *block = get_irn_n(n, -1); + + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, + block, + a, + new_r_Const_long(current_ir_graph, block, mode, 2), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); + } + else if (get_irn_op(a) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + b, + get_Minus_op(a), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + else if (get_irn_op(b) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + a, + get_Minus_op(b), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) { + ir_node *ma = get_Mul_left(a); + ir_node *mb = get_Mul_right(a); + + if (b == ma) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + else if (b == mb) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) { + ir_node *ma = get_Mul_left(b); + ir_node *mb = get_Mul_right(b); + + if (a == ma) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + else if (a == mb) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), current_ir_graph, blk, + mb, + new_rd_Add( + get_irn_dbg_info(n), current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); + } + } + } + return n; +} /* transform_node_Add */ + +/** + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Sub(0,a) -> Minus(a) + * Sub(Mul(a, x), a) -> Mul(a, x-1) + * Sub(Sub(x, y), b) -> Sub(x, Add(y,b)) + */ +static ir_node *transform_node_Sub(ir_node *n) +{ + ir_mode *mode; + ir_node *oldn = n; + ir_node *a, *b, *c; + + n = transform_node_AddSub(n); + + a = get_Sub_left(n); + b = get_Sub_right(n); + + HANDLE_BINOP_PHI(tarval_sub, a,b,c); + + mode = get_irn_mode(n); + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + + /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ + if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) { + n = new_rd_Minus( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + b, + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); + } + /* do NOT execute this code if reassociation is enabled, it does the inverse! */ + else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) { + ir_node *ma = get_Mul_left(a); + ir_node *mb = get_Mul_right(a); + + if (ma == b) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, blk, + ma, + new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, blk, + mb, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + } + else if (mb == b) { + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, blk, + mb, + new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, blk, + ma, + new_r_Const_long(current_ir_graph, blk, mode, 1), + mode), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); + } + } + else if (get_irn_op(a) == op_Sub) { + ir_node *x = get_Sub_left(a); + ir_node *y = get_Sub_right(a); + ir_node *blk = get_irn_n(n, -1); + ir_mode *m_b = get_irn_mode(b); + ir_mode *m_y = get_irn_mode(y); + ir_node *add; + + /* Determine the right mode for the Add. */ + if (m_b == m_y) + mode = m_b; + else if (mode_is_reference(m_b)) + mode = m_b; + else if (mode_is_reference(m_y)) + mode = m_y; + else { + /* + * Both modes are different but none is reference, + * happens for instance in SubP(SubP(P, Iu), Is). + * We have two possibilities here: Cast or ignore. + * Currently we ignore this case. + */ + return n; + } + + add = new_r_Add(current_ir_graph, blk, y, b, mode); + + set_Sub_left(n, x); + set_Sub_right(n, add); + DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z); + } + + return n; +} /* transform_node_Sub */ + +/** + * Transform Mul(a,-1) into -a. + * Do constant evaluation of Phi nodes. + * Do architecture dependent optimizations on Mul nodes + */ static ir_node *transform_node_Mul(ir_node *n) { + ir_node *c, *oldn = n; + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + ir_mode *mode; + + HANDLE_BINOP_PHI(tarval_mul, a,b,c); + + mode = get_irn_mode(n); + if (mode_is_signed(mode)) { + ir_node *r = NULL; + + if (value_of(a) == get_mode_minus_one(mode)) + r = b; + else if (value_of(b) == get_mode_minus_one(mode)) + r = a; + if (r) { + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); + return n; + } + } return arch_dep_replace_mul_with_shifts(n); -} +} /* transform_node_Mul */ +/** + * Transform a Div Node. + */ static ir_node *transform_node_Div(ir_node *n) { tarval *tv = value_of(n); @@ -1228,23 +2160,29 @@ static ir_node *transform_node_Div(ir_node *n) /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ - if (tv != tarval_bad) + if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); - else /* Try architecture dependand optimization */ + + DBG_OPT_CSTEVAL(n, value); + } + else /* Try architecture dependent optimization */ value = arch_dep_replace_div_by_const(n); if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ ir_node *mem = get_Div_mem(n); - turn_into_tuple(n, 3); + turn_into_tuple(n, pn_Div_max); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_except, new_Bad()); set_Tuple_pred(n, pn_Div_res, value); } return n; -} +} /* transform_node_Div */ +/** + * Transform a Mod node. + */ static ir_node *transform_node_Mod(ir_node *n) { tarval *tv = value_of(n); @@ -1252,23 +2190,29 @@ static ir_node *transform_node_Mod(ir_node *n) /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */ - if (tv != tarval_bad) + if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); - else /* Try architecture dependand optimization */ + + DBG_OPT_CSTEVAL(n, value); + } + else /* Try architecture dependent optimization */ value = arch_dep_replace_mod_by_const(n); if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ ir_node *mem = get_Mod_mem(n); - turn_into_tuple(n, 3); + turn_into_tuple(n, pn_Mod_max); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); set_Tuple_pred(n, pn_Mod_res, value); } return n; -} +} /* transform_node_Mod */ +/** + * Transform a DivMod node. + */ static ir_node *transform_node_DivMod(ir_node *n) { int evaluated = 0; @@ -1288,7 +2232,10 @@ static ir_node *transform_node_DivMod(ir_node *n) if (tb == get_mode_one(get_tarval_mode(tb))) { b = new_Const (mode, get_mode_null(mode)); evaluated = 1; - } else if (ta != tarval_bad) { + + DBG_OPT_CSTEVAL(n, b); + } + else if (ta != tarval_bad) { tarval *resa, *resb; resa = tarval_div (ta, tb); if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through @@ -1298,8 +2245,11 @@ static ir_node *transform_node_DivMod(ir_node *n) a = new_Const (mode, resa); b = new_Const (mode, resb); evaluated = 1; + + DBG_OPT_CSTEVAL(n, a); + DBG_OPT_CSTEVAL(n, b); } - else { /* Try architecture dependand optimization */ + else { /* Try architecture dependent optimization */ arch_dep_replace_divmod_by_const(&a, &b, n); evaluated = a != NULL; } @@ -1311,17 +2261,54 @@ static ir_node *transform_node_DivMod(ir_node *n) if (evaluated) { /* replace by tuple */ ir_node *mem = get_DivMod_mem(n); - turn_into_tuple(n, 4); + turn_into_tuple(n, pn_DivMod_max); set_Tuple_pred(n, pn_DivMod_M, mem); set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_DivMod_res_div, a); set_Tuple_pred(n, pn_DivMod_res_mod, b); - assert(get_nodes_block(n)); } return n; -} +} /* transform_node_DivMod */ +/** + * Optimize Abs(x) into x if x is Confirmed >= 0 + * Optimize Abs(x) into -x if x is Confirmed <= 0 + */ +static ir_node *transform_node_Abs(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_Abs_op(n); + value_classify_sign sign = classify_value_sign(a); + + if (sign == value_classified_negative) { + ir_mode *mode = get_irn_mode(n); + + /* + * We can replace the Abs by -x here. + * We even could add a new Confirm here. + * + * Note that -x would create a new node, so we could + * not run it in the equivalent_node() context. + */ + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, + get_irn_n(n, -1), a, mode); + + DBG_OPT_CONFIRM(oldn, n); + } + else if (sign == value_classified_positive) { + /* n is positive, Abs is not needed */ + n = a; + + DBG_OPT_CONFIRM(oldn, n); + } + + return n; +} /* transform_node_Abs */ + +/** + * Transform a Cond node. + */ static ir_node *transform_node_Cond(ir_node *n) { /* Replace the Cond by a Jmp if it branches on a constant @@ -1330,13 +2317,17 @@ static ir_node *transform_node_Cond(ir_node *n) ir_node *a = get_Cond_selector(n); tarval *ta = value_of(a); + /* we need block info which is not available in floating irgs */ + if (get_irg_pinned(current_ir_graph) == op_pin_state_floats) + return n; + if ((ta != tarval_bad) && (get_irn_mode(a) == mode_b) && (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); - turn_into_tuple(n, 2); + turn_into_tuple(n, pn_Cond_max); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); set_Tuple_pred(n, pn_Cond_true, jmp); @@ -1346,221 +2337,578 @@ static ir_node *transform_node_Cond(ir_node *n) } /* We might generate an endless loop, so keep it alive. */ add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); - } else if ((ta != tarval_bad) && - (get_irn_mode(a) == mode_Iu) && - (get_Cond_kind(n) == dense) && - (get_opt_unreachable_code())) { - /* I don't want to allow Tuples smaller than the biggest Proj. - Also this tuple might get really big... - I generate the Jmp here, and remember it in link. Link is used - when optimizing Proj. */ - set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n))); - /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); - } else if ((get_irn_op(a) == op_Eor) - && (get_irn_mode(a) == mode_b) - && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { - /* The Eor is a negate. Generate a new Cond without the negate, - simulate the negate by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), - get_Eor_left(a))); - } else if ((get_irn_op(a) == op_Not) - && (get_irn_mode(a) == mode_b)) { - /* A Not before the Cond. Generate a new Cond without the Not, - simulate the Not by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), - get_Not_op(a))); } return n; -} +} /* transform_node_Cond */ + +/** + * Transform an And. + */ +static ir_node *transform_node_And(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_And_left(n); + ir_node *b = get_And_right(n); + + HANDLE_BINOP_PHI(tarval_and, a,b,c); + return n; +} /* transform_node_And */ /** * Transform an Eor. */ static ir_node *transform_node_Eor(ir_node *n) { + ir_node *c, *oldn = n; ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); + ir_mode *mode = get_irn_mode(n); - if ((get_irn_mode(n) == mode_b) + HANDLE_BINOP_PHI(tarval_eor, a,b,c); + + if (a == b) { + /* a ^ a = 0 */ + n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), + mode, get_mode_null(mode)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A); + } + else if ((mode == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) + && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), - mode_b, get_negated_pnc(get_Proj_proj(a))); - else if ((get_irn_mode(n) == mode_b) - && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode)); + + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL); + } + else if ((mode == mode_b) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) { /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ - n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); + n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b); + + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + } return n; -} +} /* transform_node_Eor */ /** - * Transform a boolean Not. + * Transform a Not. */ static ir_node *transform_node_Not(ir_node *n) { + ir_node *c, *oldn = n; ir_node *a = get_Not_op(n); + HANDLE_UNOP_PHI(tarval_not,a,c); + + /* check for a boolean Not */ if ( (get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) + && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), - mode_b, get_negated_pnc(get_Proj_proj(a))); + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); + } + return n; +} /* transform_node_Not */ +/** + * Transform a Minus. + */ +static ir_node *transform_node_Minus(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Minus_op(n); + + HANDLE_UNOP_PHI(tarval_neg,a,c); return n; -} +} /* transform_node_Minus */ /** - * Transform a Cast of a Const into a new Const + * Transform a Cast_type(Const) into a new Const_type */ static ir_node *transform_node_Cast(ir_node *n) { + ir_node *oldn = n; ir_node *pred = get_Cast_op(n); - type *tp = get_irn_type(pred); + ir_type *tp = get_irn_type(n); if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { - n = new_rd_Const_type(NULL, current_ir_graph, get_nodes_block(pred), get_irn_mode(pred), + n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), get_Const_tarval(pred), tp); + DBG_OPT_CSTEVAL(oldn, n); } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) { - n = new_rd_SymConst_type(NULL, current_ir_graph, get_nodes_block(pred), get_SymConst_symbol(pred), + n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred), get_SymConst_kind(pred), tp); + DBG_OPT_CSTEVAL(oldn, n); } + return n; -} +} /* transform_node_Cast */ /** - * Transform a Div/Mod/DivMod with a non-zero constant. Must be - * done here instead of equivalent node because it creates new - * nodes. + * Transform a Proj(Div) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. - * - * Further, it optimizes jump tables by removing all impossible cases. */ -static ir_node *transform_node_Proj(ir_node *proj) +static ir_node *transform_node_Proj_Div(ir_node *proj) { ir_node *n = get_Proj_pred(proj); - ir_node *b; - tarval *tb; + ir_node *b = get_Div_right(n); + ir_node *confirm; long proj_nr; - switch (get_irn_opcode(n)) { - case iro_Div: - b = get_Div_right(n); - tb = value_of(b); - - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ - proj_nr = get_Proj_proj(proj); + if (value_not_zero(b, &confirm)) { + /* div(x, y) && y != 0 */ + proj_nr = get_Proj_proj(proj); + if (proj_nr == pn_Div_X_except) { + /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); + return new_Bad(); + } + else if (proj_nr == pn_Div_M) { + ir_node *res = get_Div_mem(n); + ir_node *new_mem = get_irg_no_mem(current_ir_graph); - /* this node may float */ + if (confirm) { + /* This node can only float up to the Confirm block */ + new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); + } set_irn_pinned(n, op_pin_state_floats); + /* this is a Div without exception, we can remove the memory edge */ + set_Div_mem(n, new_mem); + return res; + } + } + return proj; +} /* transform_node_Proj_Div */ - if (proj_nr == pn_Div_X_except) { - /* we found an exception handler, remove it */ - return new_Bad(); - } else { - /* the memory Proj can be removed */ - ir_node *res = get_Div_mem(n); -# ifdef USE_NOMEM - set_Div_mem(n, get_irg_no_mem(current_ir_graph)); -# endif /* defined USE_NOMEM */ - if (proj_nr == pn_Div_M) - return res; +/** + * Transform a Proj(Mod) with a non-zero value. + * Removes the exceptions and routes the memory to the NoMem node. + */ +static ir_node *transform_node_Proj_Mod(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + ir_node *b = get_Mod_right(n); + ir_node *confirm; + long proj_nr; + + if (value_not_zero(b, &confirm)) { + /* mod(x, y) && y != 0 */ + proj_nr = get_Proj_proj(proj); + + if (proj_nr == pn_Mod_X_except) { + /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); + return new_Bad(); + } else if (proj_nr == pn_Mod_M) { + ir_node *res = get_Mod_mem(n); + ir_node *new_mem = get_irg_no_mem(current_ir_graph); + + if (confirm) { + /* This node can only float up to the Confirm block */ + new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } + set_irn_pinned(n, op_pin_state_floats); + /* this is a Mod without exception, we can remove the memory edge */ + set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); + return res; } - break; - case iro_Mod: - b = get_Mod_right(n); - tb = value_of(b); + else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ - proj_nr = get_Proj_proj(proj); + DBG_OPT_CSTEVAL(n, res); + return res; + } + } + return proj; +} /* transform_node_Proj_Mod */ - /* this node may float */ - set_irn_pinned(n, op_pin_state_floats); +/** + * Transform a Proj(DivMod) with a non-zero value. + * Removes the exceptions and routes the memory to the NoMem node. + */ +static ir_node *transform_node_Proj_DivMod(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + ir_node *b = get_DivMod_right(n); + ir_node *confirm; + long proj_nr; - if (proj_nr == pn_Mod_X_except) { - /* we found an exception handler, remove it */ - return new_Bad(); - } else { - /* the memory Proj can be removed */ - ir_node *res = get_Mod_mem(n); -# ifdef USE_NOMEM - set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); -# endif /* defined USE_NOMEM */ - if (proj_nr == pn_Mod_M) - return res; + if (value_not_zero(b, &confirm)) { + /* DivMod(x, y) && y != 0 */ + proj_nr = get_Proj_proj(proj); + + if (proj_nr == pn_DivMod_X_except) { + /* we found an exception handler, remove it */ + DBG_OPT_EXC_REM(proj); + return new_Bad(); + } + else if (proj_nr == pn_DivMod_M) { + ir_node *res = get_DivMod_mem(n); + ir_node *new_mem = get_irg_no_mem(current_ir_graph); + + if (confirm) { + /* This node can only float up to the Confirm block */ + new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } + set_irn_pinned(n, op_pin_state_floats); + /* this is a DivMod without exception, we can remove the memory edge */ + set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); + return res; } - break; - case iro_DivMod: - b = get_DivMod_right(n); - tb = value_of(b); + else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); - if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ - proj_nr = get_Proj_proj(proj); + DBG_OPT_CSTEVAL(n, res); + return res; + } + } + return proj; +} /* transform_node_Proj_DivMod */ - /* this node may float */ - set_irn_pinned(n, op_pin_state_floats); +/** + * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases. + */ +static ir_node *transform_node_Proj_Cond(ir_node *proj) +{ + if (get_opt_unreachable_code()) { + ir_node *n = get_Proj_pred(proj); + ir_node *b = get_Cond_selector(n); - if (proj_nr == pn_DivMod_X_except) { - /* we found an exception handler, remove it */ - return new_Bad(); - } - else { - /* the memory Proj can be removed */ - ir_node *res = get_DivMod_mem(n); -# ifdef USE_NOMEM - set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); -# endif /* defined USE_NOMEM */ - if (proj_nr == pn_DivMod_M) - return res; + if (mode_is_int(get_irn_mode(b))) { + tarval *tb = value_of(b); + + if (tb != tarval_bad) { + /* we have a constant switch */ + long num = get_Proj_proj(proj); + + if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */ + if (get_tarval_long(tb) == num) { + /* Do NOT create a jump here, or we will have 2 control flow ops + * in a block. This case is optimized away in optimize_cf(). */ + return proj; + } + else { + /* this case will NEVER be taken, kill it */ + return new_Bad(); + } + } } } - break; + } + return proj; +} /* transform_node_Proj_Cond */ + +/** + * Normalizes and optimizes Cmp nodes. + */ +static ir_node *transform_node_Proj_Cmp(ir_node *proj) +{ + if (get_opt_reassociation()) { + ir_node *n = get_Proj_pred(proj); + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + ir_node *c = NULL; + tarval *tv = NULL; + int changed = 0; + ir_mode *mode = NULL; + long proj_nr = get_Proj_proj(proj); + + /* + * First step: normalize the compare op + * by placing the constant on the right site + * or moving the lower address node to the left. + * We ignore the case that both are constants + * this case should be optimized away. + */ + if (get_irn_op(right) == op_Const) + c = right; + else if (get_irn_op(left) == op_Const) { + c = left; + left = right; + right = c; + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; + } + else if (get_irn_idx(left) > get_irn_idx(right)) { + ir_node *t = left; + + left = right; + right = t; + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; + } + + /* + * Second step: Try to reduce the magnitude + * of a constant. This may help to generate better code + * later and may help to normalize more compares. + * Of course this is only possible for integer values. + */ + if (c) { + mode = get_irn_mode(c); + tv = get_Const_tarval(c); + + if (tv != tarval_bad) { + /* the following optimization is possible on modes without Overflow + * on Unary Minus or on == and !=: + * -a CMP c ==> a swap(CMP) -c + * + * Beware: for two-complement Overflow may occur, so only == and != can + * be optimized, see this: + * -MININT < 0 =/=> MININT > 0 !!! + */ + if (get_opt_constant_folding() && get_irn_op(left) == op_Minus && + (!mode_overflow_on_unary_Minus(mode) || + (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { + left = get_Minus_op(left); + tv = tarval_sub(get_mode_null(mode), tv); + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 2; + } + + /* for integer modes, we have more */ + if (mode_is_int(mode)) { + /* Ne includes Unordered which is not possible on integers. + * However, frontends often use this wrong, so fix it here */ + if (proj_nr & pn_Cmp_Uo) { + proj_nr &= ~pn_Cmp_Uo; + set_Proj_proj(proj, proj_nr); + } + + /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ + if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { + tv = tarval_sub(tv, get_mode_one(mode)); + + proj_nr ^= pn_Cmp_Eq; + changed |= 2; + } + /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ + else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { + tv = tarval_add(tv, get_mode_one(mode)); + + proj_nr ^= pn_Cmp_Eq; + changed |= 2; + } + + /* the following reassociations work only for == and != */ + if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + + /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ + if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) { + right = get_Sub_right(left); + left = get_Sub_left(left); + + tv = value_of(right); + changed = 1; + } + + if (tv != tarval_bad) { + ir_op *op = get_irn_op(left); + + /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ + if (op == op_Sub) { + ir_node *c1 = get_Sub_right(left); + tarval *tv2 = value_of(c1); + + if (tv2 != tarval_bad) { + tv2 = tarval_add(tv, value_of(c1)); + + if (tv2 != tarval_bad) { + left = get_Sub_left(left); + tv = tv2; + changed |= 2; + } + } + } + /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ + else if (op == op_Add) { + ir_node *a_l = get_Add_left(left); + ir_node *a_r = get_Add_right(left); + ir_node *a; + tarval *tv2; + + if (get_irn_op(a_l) == op_Const) { + a = a_r; + tv2 = value_of(a_l); + } + else { + a = a_l; + tv2 = value_of(a_r); + } + + if (tv2 != tarval_bad) { + tv2 = tarval_sub(tv, tv2); + + if (tv2 != tarval_bad) { + left = a; + tv = tv2; + changed |= 2; + } + } + } + /* -a == c ==> a == -c, -a != c ==> a != -c */ + else if (op == op_Minus) { + tarval *tv2 = tarval_sub(get_mode_null(mode), tv); + + if (tv2 != tarval_bad) { + left = get_Minus_op(left); + tv = tv2; + changed |= 2; + } + } + } + } /* == or != */ + /* the following reassociations work only for <= */ + else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + if (tv != tarval_bad) { + ir_op *op = get_irn_op(left); + + /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */ + if (op == op_Abs) { + } + } + } + } /* mode_is_int */ + + /* + * optimization for AND: + * Optimize: + * And(x, C) == C ==> And(x, C) != 0 + * And(x, C) != C ==> And(X, C) == 0 + * + * if C is a single Bit constant. + */ + if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && + (get_irn_op(left) == op_And)) { + if (is_single_bit_tarval(tv)) { + /* check for Constant's match. We have check hare the tarvals, + because our const might be changed */ + ir_node *la = get_And_left(left); + ir_node *ra = get_And_right(left); + if ((is_Const(la) && get_Const_tarval(la) == tv) || + (is_Const(ra) && get_Const_tarval(ra) == tv)) { + /* fine: do the transformation */ + tv = get_mode_null(get_tarval_mode(tv)); + proj_nr ^= pn_Cmp_Leg; + changed |= 2; + } + } + } + } /* tarval != bad */ + } + + if (changed) { + ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ + + if (changed & 2) /* need a new Const */ + right = new_Const(mode, tv); + + /* create a new compare */ + n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, + left, right); + + set_Proj_pred(proj, n); + set_Proj_proj(proj, proj_nr); + } + } + return proj; +} /* transform_node_Proj_Cmp */ + +/** + * Does all optimizations on nodes that must be done on it's Proj's + * because of creating new nodes. + */ +static ir_node *transform_node_Proj(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + + switch (get_irn_opcode(n)) { + case iro_Div: + return transform_node_Proj_Div(proj); - case iro_Cond: - if (get_opt_unreachable_code()) { - b = get_Cond_selector(n); - tb = value_of(b); + case iro_Mod: + return transform_node_Proj_Mod(proj); - if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) { - /* we have a constant switch */ - long num = get_Proj_proj(proj); + case iro_DivMod: + return transform_node_Proj_DivMod(proj); - if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */ - if (get_tarval_long(tb) == num) { - /* Do NOT create a jump here, or we will have 2 control flow ops - * in a block. This case is optimized away in optimize_cf(). */ - return proj; - } - else - return new_Bad(); - } - } - } - return proj; + case iro_Cond: + return transform_node_Proj_Cond(proj); + + case iro_Cmp: + return transform_node_Proj_Cmp(proj); case iro_Tuple: /* should not happen, but if it does will be optimized away */ - break; + return equivalent_node_Proj(proj); default: /* do nothing */ return proj; } +} /* transform_node_Proj */ - /* we have added a Tuple, optimize it for the current Proj away */ - return equivalent_node_Proj(proj); -} +/** + * Move Confirms down through Phi nodes. + */ +static ir_node *transform_node_Phi(ir_node *phi) { + int i, n; + ir_mode *mode = get_irn_mode(phi); + + if (mode_is_reference(mode)) { + n = get_irn_arity(phi); + + /* Beware of Phi0 */ + if (n > 0) { + ir_node *pred = get_irn_n(phi, 0); + ir_node *bound, *new_Phi, *block, **in; + pn_Cmp pnc; + + if (! is_Confirm(pred)) + return phi; + + bound = get_Confirm_bound(pred); + pnc = get_Confirm_cmp(pred); + + NEW_ARR_A(ir_node *, in, n); + in[0] = get_Confirm_value(pred); + + for (i = 1; i < n; ++i) { + pred = get_irn_n(phi, i); + + if (! is_Confirm(pred) || + get_Confirm_bound(pred) != bound || + get_Confirm_cmp(pred) != pnc) + return phi; + in[i] = get_Confirm_value(pred); + } + /* move the Confirm nodes "behind" the Phi */ + block = get_irn_n(phi, -1); + new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi)); + return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc); + } + } + return phi; +} /* transform_node_Phi */ /** - * returns the operands of a commutative bin-op, if one operand is + * Returns the operands of a commutative bin-op, if one operand is * a const, it is returned as the second one. */ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) @@ -1578,7 +2926,7 @@ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) *a = op_a; *c = op_b; } -} +} /* get_comm_Binop_Ops */ /** * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. @@ -1590,7 +2938,7 @@ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) * AND c1 * OR */ -static ir_node *transform_node_Or(ir_node *or) +static ir_node *transform_node_Or_bf_store(ir_node *or) { ir_node *and, *c1; ir_node *or_l, *c2; @@ -1646,7 +2994,7 @@ static ir_node *transform_node_Or(ir_node *or) } /* ok, all conditions met */ - block = get_nodes_block(or); + block = get_irn_n(or, -1); new_and = new_r_And(current_ir_graph, block, value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); @@ -1657,18 +3005,149 @@ static ir_node *transform_node_Or(ir_node *or) set_Or_right(or, new_const); /* check for more */ - return transform_node_Or(or); -} + return transform_node_Or_bf_store(or); +} /* transform_node_Or_bf_store */ + +/** + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot + */ +static ir_node *transform_node_Or_Rot(ir_node *or) +{ + ir_mode *mode = get_irn_mode(or); + ir_node *shl, *shr, *block; + ir_node *irn, *x, *c1, *c2, *v, *sub, *n; + tarval *tv1, *tv2; + + if (! mode_is_int(mode)) + return or; + + shl = get_binop_left(or); + shr = get_binop_right(or); + + if (get_irn_op(shl) == op_Shr) { + if (get_irn_op(shr) != op_Shl) + return or; + + irn = shl; + shl = shr; + shr = irn; + } + else if (get_irn_op(shl) != op_Shl) + return or; + else if (get_irn_op(shr) != op_Shr) + return or; + + x = get_Shl_left(shl); + if (x != get_Shr_left(shr)) + return or; + + c1 = get_Shl_right(shl); + c2 = get_Shr_right(shr); + if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) { + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + tv2 = get_Const_tarval(c2); + if (! tarval_is_long(tv2)) + return or; + + if (get_tarval_long(tv1) + get_tarval_long(tv2) + != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_irn_n(or, -1); + + n = new_r_Rot(current_ir_graph, block, x, c1, mode); + + DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + else if (get_irn_op(c1) == op_Sub) { + v = c2; + sub = c1; + + if (get_Sub_right(sub) != v) + return or; + + c1 = get_Sub_left(sub); + if (get_irn_op(c1) != op_Const) + return or; + + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_nodes_block(or); + + /* a Rot right is not supported, so use a rot left */ + n = new_r_Rot(current_ir_graph, block, x, sub, mode); + + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + else if (get_irn_op(c2) == op_Sub) { + v = c1; + sub = c2; + + c1 = get_Sub_left(sub); + if (get_irn_op(c1) != op_Const) + return or; + + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_irn_n(or, -1); + + /* a Rot Left */ + n = new_r_Rot(current_ir_graph, block, x, v, mode); + + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + + return or; +} /* transform_node_Or_Rot */ + +/** + * Transform an Or. + */ +static ir_node *transform_node_Or(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Or_left(n); + ir_node *b = get_Or_right(n); + + HANDLE_BINOP_PHI(tarval_or, a,b,c); + + n = transform_node_Or_bf_store(n); + n = transform_node_Or_Rot(n); + + return n; +} /* transform_node_Or */ + /* forward */ static ir_node *transform_node(ir_node *n); /** - * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl. + * + * Should be moved to reassociation? */ -static ir_node * transform_node_shift(ir_node *n) +static ir_node *transform_node_shift(ir_node *n) { - ir_node *left; + ir_node *left, *right; tarval *tv1, *tv2, *res; ir_mode *mode; int modulo_shf, flag; @@ -1679,7 +3158,8 @@ static ir_node * transform_node_shift(ir_node *n) if (get_irn_op(left) != get_irn_op(n)) return n; - tv1 = value_of(get_binop_right(n)); + right = get_binop_right(n); + tv1 = value_of(right); if (tv1 == tarval_bad) return n; @@ -1698,7 +3178,7 @@ static ir_node * transform_node_shift(ir_node *n) if (modulo_shf > 0) { tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res)); - if (tarval_cmp(res, modulo) & Lt) + if (tarval_cmp(res, modulo) & pn_Cmp_Lt) flag = 1; } else @@ -1706,18 +3186,222 @@ static ir_node * transform_node_shift(ir_node *n) if (flag) { /* ok, we can replace it */ - ir_node *in[2], *irn, *block = get_nodes_block(n); + ir_node *in[2], *irn, *block = get_irn_n(n, -1); in[0] = get_binop_left(left); in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + return transform_node(irn); } return n; -} +} /* transform_node_shift */ + +/** + * Transform a Shr. + */ +static ir_node *transform_node_Shr(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shr_left(n); + ir_node *b = get_Shr_right(n); + + HANDLE_BINOP_PHI(tarval_shr, a, b, c); + return transform_node_shift(n); +} /* transform_node_Shr */ + +/** + * Transform a Shrs. + */ +static ir_node *transform_node_Shrs(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shrs_left(n); + ir_node *b = get_Shrs_right(n); + + HANDLE_BINOP_PHI(tarval_shrs, a, b, c); + return transform_node_shift(n); +} /* transform_node_Shrs */ + +/** + * Transform a Shl. + */ +static ir_node *transform_node_Shl(ir_node *n) +{ + ir_node *c, *oldn = n; + ir_node *a = get_Shl_left(n); + ir_node *b = get_Shl_right(n); + + HANDLE_BINOP_PHI(tarval_shl, a, b, c); + return transform_node_shift(n); +} /* transform_node_Shl */ + +/** + * Remove dead blocks and nodes in dead blocks + * in keep alive list. We do not generate a new End node. + */ +static ir_node *transform_node_End(ir_node *n) { + int i, n_keepalives = get_End_n_keepalives(n); + + for (i = 0; i < n_keepalives; ++i) { + ir_node *ka = get_End_keepalive(n, i); + if (is_Block(ka)) { + if (is_Block_dead(ka)) { + set_End_keepalive(n, i, new_Bad()); + } + } + else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) + set_End_keepalive(n, i, new_Bad()); + } + return n; +} /* transform_node_End */ + +/** + * Optimize a Mux into some simpler cases. + */ +static ir_node *transform_node_Mux(ir_node *n) +{ + ir_node *oldn = n, *sel = get_Mux_sel(n); + ir_mode *mode = get_irn_mode(n); + + if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) { + ir_node *cmp = get_Proj_pred(sel); + long proj_nr = get_Proj_proj(sel); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(n); + + if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { + ir_node *block = get_irn_n(n, -1); + + /* + * Note: normalization puts the constant on the right site, + * so we check only one case. + * + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the first one. + */ + if (get_irn_op(f) == op_Minus && + get_Minus_op(f) == t && + get_Cmp_left(cmp) == t) { + + if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { + /* Mux(a >=/> 0, -a, a) ==> Abs(a) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + t, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + t, mode); + n = new_rd_Minus(get_irn_dbg_info(n), + current_ir_graph, + block, + n, mode); + + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + } + else if (get_irn_op(t) == op_Minus && + get_Minus_op(t) == f && + get_Cmp_left(cmp) == f) { + + if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + /* Mux(a <=/< 0, a, -a) ==> Abs(a) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + f, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { + /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + f, mode); + n = new_rd_Minus(get_irn_dbg_info(n), + current_ir_graph, + block, + n, mode); + + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + } + + if (mode_is_int(mode) && mode_is_signed(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *x = get_Cmp_left(cmp); + + /* the following optimization works only with signed integer two-complement mode */ + + if (mode == get_irn_mode(x)) { + /* + * FIXME: this restriction is two rigid, as it would still + * work if mode(x) = Hs and mode == Is, but at least it removes + * all wrong cases. + */ + if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) && + classify_Const(t) == CNST_ALL_ONE && + classify_Const(f) == CNST_NULL) { + /* + * Mux(x:T Shrs(x, sizeof_bits(T) - 1) + * Conditions: + * T must be signed. + */ + n = new_rd_Shrs(get_irn_dbg_info(n), + current_ir_graph, block, x, + new_r_Const_long(current_ir_graph, block, mode_Iu, + get_mode_size_bits(mode) - 1), + mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + return n; + } + else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) && + classify_Const(t) == CNST_ONE && + classify_Const(f) == CNST_NULL) { + /* + * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1) + * Conditions: + * T must be signed. + */ + n = new_rd_Shr(get_irn_dbg_info(n), + current_ir_graph, block, + new_r_Minus(current_ir_graph, block, x, mode), + new_r_Const_long(current_ir_graph, block, mode_Iu, + get_mode_size_bits(mode) - 1), + mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + return n; + } + } + } + } + } + return arch_transform_node_Mux(n); +} /* transform_node_Mux */ + +/** + * Optimize a Psi into some simpler cases. + */ +static ir_node *transform_node_Psi(ir_node *n) { + if (is_Mux(n)) + return transform_node_Mux(n); + return n; +} /* transform_node_Psi */ /** * Tries several [inplace] [optimizing] transformations and returns an @@ -1727,46 +3411,58 @@ static ir_node * transform_node_shift(ir_node *n) */ static ir_node *transform_node(ir_node *n) { - if (n->op->transform_node) - n = n->op->transform_node(n); + if (n->op->ops.transform_node) + n = n->op->ops.transform_node(n); return n; -} +} /* transform_node */ /** - * set the default transform node operation + * sSets the default transform node operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_transform_node(ir_op *op) +static ir_op_ops *firm_set_default_transform_node(opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ - op->transform_node = transform_node_##a; \ + ops->transform_node = transform_node_##a; \ break - switch (op->code) { + switch (code) { CASE(Add); CASE(Sub); CASE(Mul); CASE(Div); CASE(Mod); CASE(DivMod); + CASE(Abs); CASE(Cond); + CASE(And); + CASE(Or); CASE(Eor); + CASE(Minus); CASE(Not); CASE(Cast); CASE(Proj); - CASE(Or); - case iro_Shr: - case iro_Shrs: - case iro_Shl: - op->transform_node = transform_node_shift; - break; + CASE(Phi); + CASE(Sel); + CASE(Shr); + CASE(Shrs); + CASE(Shl); + CASE(End); + CASE(Mux); + CASE(Psi); default: - op->transform_node = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE -} +} /* firm_set_default_transform_node */ /* **************** Common Subexpression Elimination **************** */ @@ -1776,103 +3472,108 @@ static ir_op *firm_set_default_transform_node(ir_op *op) #define N_IR_NODES 512 /** Compares the attributes of two Const nodes. */ -static int node_cmp_attr_Const(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Const(ir_node *a, ir_node *b) { return (get_Const_tarval(a) != get_Const_tarval(b)) || (get_Const_type(a) != get_Const_type(b)); -} +} /* node_cmp_attr_Const */ /** Compares the attributes of two Proj nodes. */ -static int node_cmp_attr_Proj(ir_node *a, ir_node *b) -{ - return get_irn_proj_attr (a) != get_irn_proj_attr (b); -} +static int node_cmp_attr_Proj(ir_node *a, ir_node *b) { + return get_irn_proj_attr (a) != get_irn_proj_attr (b); +} /* node_cmp_attr_Proj */ /** Compares the attributes of two Filter nodes. */ -static int node_cmp_attr_Filter(ir_node *a, ir_node *b) -{ - return get_Filter_proj(a) != get_Filter_proj(b); -} +static int node_cmp_attr_Filter(ir_node *a, ir_node *b) { + return get_Filter_proj(a) != get_Filter_proj(b); +} /* node_cmp_attr_Filter */ /** Compares the attributes of two Alloc nodes. */ -static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) -{ - return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) - || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); -} +static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) { + return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) + || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); +} /* node_cmp_attr_Alloc */ /** Compares the attributes of two Free nodes. */ -static int node_cmp_attr_Free(ir_node *a, ir_node *b) -{ - return (get_irn_free_attr(a) != get_irn_free_attr(b)); -} +static int node_cmp_attr_Free(ir_node *a, ir_node *b) { + return (get_irn_free_attr(a).where != get_irn_free_attr(b).where) + || (get_irn_free_attr(a).type != get_irn_free_attr(b).type); +} /* node_cmp_attr_Free */ /** Compares the attributes of two SymConst nodes. */ -static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) -{ - return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) +static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { + return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p) || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp); -} +} /* node_cmp_attr_SymConst */ /** Compares the attributes of two Call nodes. */ -static int node_cmp_attr_Call(ir_node *a, ir_node *b) -{ - return (get_irn_call_attr(a) != get_irn_call_attr(b)); -} +static int node_cmp_attr_Call(ir_node *a, ir_node *b) { + return (get_irn_call_attr(a) != get_irn_call_attr(b)); +} /* node_cmp_attr_Call */ /** Compares the attributes of two Sel nodes. */ -static int node_cmp_attr_Sel(ir_node *a, ir_node *b) -{ - return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) +static int node_cmp_attr_Sel(ir_node *a, ir_node *b) { + return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name) || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); -} +} /* node_cmp_attr_Sel */ /** Compares the attributes of two Phi nodes. */ -static int node_cmp_attr_Phi(ir_node *a, ir_node *b) -{ - return get_irn_phi_attr (a) != get_irn_phi_attr (b); -} +static int node_cmp_attr_Phi(ir_node *a, ir_node *b) { + return get_irn_phi_attr (a) != get_irn_phi_attr (b); +} /* node_cmp_attr_Phi */ + +/** Compares the attributes of two Conv nodes. */ +static int node_cmp_attr_Conv(ir_node *a, ir_node *b) { + return get_Conv_strict(a) != get_Conv_strict(b); +} /* node_cmp_attr_Conv */ /** Compares the attributes of two Cast nodes. */ -static int node_cmp_attr_Cast(ir_node *a, ir_node *b) -{ - return get_Cast_type(a) != get_Cast_type(b); -} +static int node_cmp_attr_Cast(ir_node *a, ir_node *b) { + return get_Cast_type(a) != get_Cast_type(b); +} /* node_cmp_attr_Cast */ /** Compares the attributes of two Load nodes. */ -static int node_cmp_attr_Load(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Load(ir_node *a, ir_node *b) { if (get_Load_volatility(a) == volatility_is_volatile || get_Load_volatility(b) == volatility_is_volatile) /* NEVER do CSE on volatile Loads */ return 1; return get_Load_mode(a) != get_Load_mode(b); -} +} /* node_cmp_attr_Load */ /** Compares the attributes of two Store nodes. */ -static int node_cmp_attr_Store(ir_node *a, ir_node *b) -{ +static int node_cmp_attr_Store(ir_node *a, ir_node *b) { /* NEVER do CSE on volatile Stores */ return (get_Store_volatility(a) == volatility_is_volatile || - get_Store_volatility(b) == volatility_is_volatile); -} + get_Store_volatility(b) == volatility_is_volatile); +} /* node_cmp_attr_Store */ + +/** Compares the attributes of two Confirm nodes. */ +static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) { + return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); +} /* node_cmp_attr_Confirm */ /** - * set the default node attribute compare operation + * Set the default node attribute compare operation for an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. */ -static ir_op *firm_set_default_node_cmp_attr(ir_op *op) +static ir_op_ops *firm_set_default_node_cmp_attr(opcode code, ir_op_ops *ops) { -#define CASE(a) \ - case iro_##a: \ - op->node_cmp_attr = node_cmp_attr_##a; \ +#define CASE(a) \ + case iro_##a: \ + ops->node_cmp_attr = node_cmp_attr_##a; \ break - switch (op->code) { + switch (code) { CASE(Const); CASE(Proj); CASE(Filter); @@ -1882,23 +3583,24 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) CASE(Call); CASE(Sel); CASE(Phi); + CASE(Conv); CASE(Cast); CASE(Load); CASE(Store); + CASE(Confirm); default: - op->node_cmp_attr = NULL; + /* leave NULL */; } - return op; + return ops; #undef CASE -} +} /* firm_set_default_node_cmp_attr */ -/** +/* * Compare function for two nodes in the hash table. Gets two * nodes as parameters. Returns 0 if the nodes are a cse. */ -static int -vt_cmp (const void *elt, const void *key) +int identities_cmp(const void *elt, const void *key) { ir_node *a, *b; int i, irn_arity_a; @@ -1931,17 +3633,16 @@ vt_cmp (const void *elt, const void *key) * here, we already now that the nodes are identical except their * attributes */ - if (a->op->node_cmp_attr) - return a->op->node_cmp_attr(a, b); + if (a->op->ops.node_cmp_attr) + return a->op->ops.node_cmp_attr(a, b); return 0; -} +} /* identities_cmp */ /* * Calculate a hash value of a node. */ -unsigned -ir_node_hash (ir_node *node) +unsigned ir_node_hash(ir_node *node) { unsigned h; int i, irn_arity; @@ -1952,7 +3653,7 @@ ir_node_hash (ir_node *node) h = 9*h + HASH_PTR(get_irn_mode(node)); } else if (node->op == op_SymConst) { /* special value for const, as they only differ in their symbol. */ - h = HASH_PTR(node->attr.i.sym.type_p); + h = HASH_PTR(node->attr.symc.sym.type_p); h = 9*h + HASH_PTR(get_irn_mode(node)); } else { @@ -1960,7 +3661,7 @@ ir_node_hash (ir_node *node) h = irn_arity = get_irn_intra_arity(node); /* consider all in nodes... except the block if not a control flow. */ - for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { + for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { h = 9*h + HASH_PTR(get_irn_intra_n(node, i)); } @@ -1971,28 +3672,29 @@ ir_node_hash (ir_node *node) } return h; -} +} /* ir_node_hash */ -pset * -new_identities(void) { - return new_pset(vt_cmp, N_IR_NODES); -} +pset *new_identities(void) { + return new_pset(identities_cmp, N_IR_NODES); +} /* new_identities */ -void -del_identities(pset *value_table) { +void del_identities(pset *value_table) { del_pset(value_table); -} +} /* del_identities */ /** * Return the canonical node computing the same value as n. + * + * @param value_table The value table + * @param n The node to lookup + * * Looks up the node in a hash table. * * For Const nodes this is performed in the constructor, too. Const * nodes are extremely time critical because of their frequent use in * constant string arrays. */ -static INLINE ir_node * -identify (pset *value_table, ir_node *n) +static INLINE ir_node *identify(pset *value_table, ir_node *n) { ir_node *o = NULL; @@ -2004,43 +3706,41 @@ identify (pset *value_table, ir_node *n) ir_node *r = get_binop_right(n); /* for commutative operators perform a OP b == b OP a */ - if (l > r) { + if (get_irn_idx(l) > get_irn_idx(r)) { set_binop_left(n, r); set_binop_right(n, l); } } } - o = pset_find (value_table, n, ir_node_hash (n)); + o = pset_find(value_table, n, ir_node_hash(n)); if (!o) return n; DBG_OPT_CSE(n, o); return o; -} +} /* identify */ /** * During construction we set the op_pin_state_pinned flag in the graph right when the * optimization is performed. The flag turning on procedure global cse could * be changed between two allocations. This way we are safe. */ -static INLINE ir_node * -identify_cons (pset *value_table, ir_node *n) { +static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) { ir_node *old = n; n = identify(value_table, n); if (get_irn_n(old, -1) != get_irn_n(n, -1)) set_irg_pinned(current_ir_graph, op_pin_state_floats); return n; -} +} /* identify_cons */ -/** +/* * Return the canonical node computing the same value as n. * Looks up the node in a hash table, enters it in the table * if it isn't there yet. */ -static ir_node * -identify_remember (pset *value_table, ir_node *n) +ir_node *identify_remember(pset *value_table, ir_node *n) { ir_node *o = NULL; @@ -2067,34 +3767,49 @@ identify_remember (pset *value_table, ir_node *n) } return o; -} +} /* identify_remember */ -void -add_identities (pset *value_table, ir_node *node) { - if (get_opt_cse() && (get_irn_opcode(node) != iro_Block)) - identify_remember (value_table, node); -} +/* Add a node to the identities value table. */ +void add_identities(pset *value_table, ir_node *node) { + if (get_opt_cse() && is_no_Block(node)) + identify_remember(value_table, node); +} /* add_identities */ + +/* Visit each node in the value table of a graph. */ +void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) { + ir_node *node; + ir_graph *rem = current_ir_graph; + + current_ir_graph = irg; + foreach_pset(irg->value_table, node) + visit(node, env); + current_ir_graph = rem; +} /* visit_all_identities */ /** - * garbage in, garbage out. If a node has a dead input, i.e., the + * Garbage in, garbage out. If a node has a dead input, i.e., the * Bad node is input to the node, return the Bad node. */ -static INLINE ir_node * -gigo (ir_node *node) +static INLINE ir_node *gigo(ir_node *node) { int i, irn_arity; - ir_op* op = get_irn_op(node); + ir_op *op = get_irn_op(node); /* remove garbage blocks by looking at control flow that leaves the block and replacing the control flow by Bad. */ if (get_irn_mode(node) == mode_X) { - ir_node *block = get_nodes_block(node); - if (op == op_End) return node; /* Don't optimize End, may have Bads. */ + ir_node *block = get_nodes_block(skip_Proj(node)); - if (get_irn_op(block) == op_Block && get_Block_matured(block)) { + /* Don't optimize nodes in immature blocks. */ + if (!get_Block_matured(block)) return node; + /* Don't optimize End, may have Bads. */ + if (op == op_End) return node; + + if (is_Block(block)) { irn_arity = get_irn_arity(block); for (i = 0; i < irn_arity; i++) { - if (!is_Bad(get_irn_n(block, i))) break; + if (!is_Bad(get_irn_n(block, i))) + break; } if (i == irn_arity) return new_Bad(); } @@ -2102,16 +3817,30 @@ gigo (ir_node *node) /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the blocks predecessors is dead. */ - if ( op != op_Block && op != op_Phi && op != op_Tuple) { + if (op != op_Block && op != op_Phi && op != op_Tuple) { irn_arity = get_irn_arity(node); - if (is_Block_dead(get_nodes_block(node))) + /* + * Beware: we can only read the block of a non-floating node. + */ + if (is_irn_pinned_in_irg(node) && + is_Block_dead(get_nodes_block(node))) return new_Bad(); for (i = 0; i < irn_arity; i++) { - if (is_Bad(get_irn_n(node, i))) { + ir_node *pred = get_irn_n(node, i); + + if (is_Bad(pred)) return new_Bad(); - } +#if 0 + /* Propagating Unknowns here seems to be a bad idea, because + sometimes we need a node as a input and did not want that + it kills it's user. + However, it might be useful to move this into a later phase + (if you think that optimizing such code is useful). */ + if (is_Unknown(pred) && mode_is_data(get_irn_mode(node))) + return new_Unknown(get_irn_mode(node)); +#endif } } #if 0 @@ -2128,84 +3857,97 @@ gigo (ir_node *node) } #endif return node; -} - +} /* gigo */ /** * These optimizations deallocate nodes from the obstack. * It can only be called if it is guaranteed that no other nodes * reference this one, i.e., right after construction of a node. + * + * @param n The node to optimize + * + * current_ir_graph must be set to the graph of the node! */ -ir_node * -optimize_node (ir_node *n) +ir_node *optimize_node(ir_node *n) { tarval *tv; ir_node *oldn = n; opcode iro = get_irn_opcode(n); - type *old_tp = get_irn_type(n); - { - int i, arity = get_irn_arity(n); - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - } - - /* Allways optimize Phi nodes: part of the construction. */ + /* Always optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && (get_irn_mode(n) != mode_T)) { /* try to evaluate */ tv = computed_value(n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { + if (tv != tarval_bad) { + ir_node *nw; + ir_type *old_tp = get_irn_type(n); + int i, arity = get_irn_arity(n); + int node_size; + + /* + * Try to recover the type of the new expression. + */ + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + /* * we MUST copy the node here temporary, because it's still needed * for DBG_OPT_CSTEVAL */ - int node_size = offsetof(ir_node, attr) + n->op->attr_size; + node_size = offsetof(ir_node, attr) + n->op->attr_size; oldn = alloca(node_size); memcpy(oldn, n, node_size); - CLONE_ARR_A(ir_node *, oldn->in, n->in); + CLONE_ARR_A(ir_node *, oldn->in, n->in); + + /* ARG, copy the in array, we need it for statistics */ + memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); - /* ARG, copy the in array, we need it for statistics */ - memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); + /* note the inplace edges module */ + edges_node_deleted(n, current_ir_graph); /* evaluation was successful -- replace the node. */ - obstack_free (current_ir_graph->obst, n); - n = new_Const (get_tarval_mode (tv), tv); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) - set_Const_type(n, old_tp); - DBG_OPT_CSTEVAL(oldn, n); - return n; + irg_kill_node(current_ir_graph, n); + nw = new_Const(get_tarval_mode (tv), tv); + + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(nw, old_tp); + DBG_OPT_CSTEVAL(oldn, nw); + return nw; } } } /* remove unnecessary nodes */ if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || - (iro == iro_Proj) || - (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node (n); + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || + (iro == iro_Proj) || + (iro == iro_Block) ) /* Flags tested local. */ + n = equivalent_node(n); optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ - /** common subexpression elimination **/ - /* Checks whether n is already available. */ - /* The block input is used to distinguish different subexpressions. Right - now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common - subexpressions within a block. */ + /* Common Subexpression Elimination. + * + * Checks whether n is already available. + * The block input is used to distinguish different subexpressions. Right + * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common + * subexpressions within a block. + */ if (get_opt_cse()) - n = identify_cons (current_ir_graph->value_table, n); + n = identify_cons(current_ir_graph->value_table, n); if (n != oldn) { - /* We found an existing, better node, so we can deallocate the old node. */ - obstack_free (current_ir_graph->obst, oldn); + edges_node_deleted(oldn, current_ir_graph); + /* We found an existing, better node, so we can deallocate the old node. */ + irg_kill_node(current_ir_graph, oldn); return n; } @@ -2213,63 +3955,57 @@ optimize_node (ir_node *n) free the node. */ iro = get_irn_opcode(n); if (get_opt_constant_folding() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ - n = transform_node (n); + (iro == iro_Cond) || + (iro == iro_Proj)) /* Flags tested local. */ + n = transform_node(n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ n = gigo (n); - /* Now we have a legal, useful node. Enter it in hash table for cse */ + /* Now we have a legal, useful node. Enter it in hash table for CSE */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { - n = identify_remember (current_ir_graph->value_table, n); + n = identify_remember(current_ir_graph->value_table, n); } return n; -} +} /* optimize_node */ /** - * These optimizations never deallocate nodes. This can cause dead + * These optimizations never deallocate nodes (in place). This can cause dead * nodes lying on the obstack. Remove these by a dead node elimination, * i.e., a copying garbage collection. */ -ir_node * -optimize_in_place_2 (ir_node *n) +ir_node *optimize_in_place_2(ir_node *n) { tarval *tv; ir_node *oldn = n; opcode iro = get_irn_opcode(n); - type *old_tp = get_irn_type(n); - { - int i, arity = get_irn_arity(n); - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - } - if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; - /* if not optimize return n */ - if (n == NULL) { - assert(0); - /* Here this is possible. Why? */ - return n; - } - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { /* try to evaluate */ tv = computed_value(n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { + if (tv != tarval_bad) { /* evaluation was successful -- replace the node. */ - n = new_Const (get_tarval_mode (tv), tv); + ir_type *old_tp = get_irn_type(n); + int i, arity = get_irn_arity(n); + + /* + * Try to recover the type of the new expression. + */ + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) - set_Const_type(n, old_tp); + n = new_Const(get_tarval_mode(tv), tv); + + if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) + set_Const_type(n, old_tp); DBG_OPT_CSTEVAL(oldn, n); return n; @@ -2283,7 +4019,7 @@ optimize_in_place_2 (ir_node *n) (iro == iro_Id) || /* ... */ (iro == iro_Proj) || /* ... */ (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node (n); + n = equivalent_node(n); optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ @@ -2293,7 +4029,7 @@ optimize_in_place_2 (ir_node *n) now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) { - n = identify (current_ir_graph->value_table, n); + n = identify(current_ir_graph->value_table, n); } /* Some more constant expression evaluation. */ @@ -2301,11 +4037,11 @@ optimize_in_place_2 (ir_node *n) if (get_opt_constant_folding() || (iro == iro_Cond) || (iro == iro_Proj)) /* Flags tested local. */ - n = transform_node (n); + n = transform_node(n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ - n = gigo (n); + n = gigo(n); /* Now we can verify the node, as it has no dead inputs any more. */ irn_vrfy(n); @@ -2314,17 +4050,15 @@ optimize_in_place_2 (ir_node *n) Blocks should be unique anyways. (Except the successor of start: is cse with the start block!) */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) - n = identify_remember (current_ir_graph->value_table, n); + n = identify_remember(current_ir_graph->value_table, n); return n; -} +} /* optimize_in_place_2 */ /** * Wrapper for external use, set proper status bits after optimization. */ -ir_node * -optimize_in_place (ir_node *n) -{ +ir_node *optimize_in_place(ir_node *n) { /* Handle graph state */ assert(get_irg_phase_state(current_ir_graph) != phase_building); @@ -2333,22 +4067,23 @@ optimize_in_place (ir_node *n) if (get_irg_outs_state(current_ir_graph) == outs_consistent) set_irg_outs_inconsistent(current_ir_graph); - /* Maybe we could also test whether optimizing the node can + /* FIXME: Maybe we could also test whether optimizing the node can change the control graph. */ - if (get_irg_dom_state(current_ir_graph) == dom_consistent) - set_irg_dom_inconsistent(current_ir_graph); + set_irg_doms_inconsistent(current_ir_graph); return optimize_in_place_2 (n); -} +} /* optimize_in_place */ -/** - * set the default ir op operations +/* + * Sets the default operation for an ir_ops. */ -ir_op *firm_set_default_operations(ir_op *op) -{ - op = firm_set_default_computed_value(op); - op = firm_set_default_equivalent_node(op); - op = firm_set_default_transform_node(op); - op = firm_set_default_node_cmp_attr(op); - - return op; -} +ir_op_ops *firm_set_default_operations(opcode code, ir_op_ops *ops) { + ops = firm_set_default_computed_value(code, ops); + ops = firm_set_default_equivalent_node(code, ops); + ops = firm_set_default_transform_node(code, ops); + ops = firm_set_default_node_cmp_attr(code, ops); + ops = firm_set_default_get_type(code, ops); + ops = firm_set_default_get_type_attr(code, ops); + ops = firm_set_default_get_entity_attr(code, ops); + + return ops; +} /* firm_set_default_operations */