X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=d7ab0b261c3d3564440b63ea989fed2e06970298;hb=f4e5f09589f68012dbd8007389366dd992f7a7c5;hp=ad239d506749e446ff5f9a38bf87cd2d5a932b40;hpb=634ecc6eaeb44e57c6a3cff03a5431a43f14ff69;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index ad239d506..d7ab0b261 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -53,17 +53,37 @@ /* Make types visible to allow most efficient access */ #include "entity_t.h" +/** + * Returns the tarval of a Const node or tarval_bad for all other nodes. + */ +static tarval *default_value_of(const ir_node *n) { + if (is_Const(n)) + return get_Const_tarval(n); /* might return tarval_bad */ + else + return tarval_bad; +} + +value_of_func value_of_ptr = default_value_of; + +/* * Set a new value_of function. */ +void set_value_of_func(value_of_func func) { + if (func != NULL) + value_of_ptr = func; + else + value_of_ptr = default_value_of; +} + /** * Return the value of a Constant. */ -static tarval *computed_value_Const(ir_node *n) { +static tarval *computed_value_Const(const ir_node *n) { return get_Const_tarval(n); } /* computed_value_Const */ /** * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst. */ -static tarval *computed_value_SymConst(ir_node *n) { +static tarval *computed_value_SymConst(const ir_node *n) { ir_type *type; ir_entity *ent; @@ -93,14 +113,14 @@ static tarval *computed_value_SymConst(ir_node *n) { /** * Return the value of an Add. */ -static tarval *computed_value_Add(ir_node *n) { +static tarval *computed_value_Add(const ir_node *n) { ir_node *a = get_Add_left(n); ir_node *b = get_Add_right(n); tarval *ta = value_of(a); tarval *tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) + if ((ta != tarval_bad) && (tb != tarval_bad)) return tarval_add(ta, tb); return tarval_bad; @@ -110,21 +130,22 @@ static tarval *computed_value_Add(ir_node *n) { * Return the value of a Sub. * Special case: a - a */ -static tarval *computed_value_Sub(ir_node *n) { - ir_node *a = get_Sub_left(n); - ir_node *b = get_Sub_right(n); - tarval *ta; - tarval *tb; +static tarval *computed_value_Sub(const ir_node *n) { + ir_mode *mode = get_irn_mode(n); + ir_node *a = get_Sub_left(n); + ir_node *b = get_Sub_right(n); + tarval *ta; + tarval *tb; /* a - a */ if (a == b && !is_Bad(a)) - return get_mode_null(get_irn_mode(n)); + return get_mode_null(mode); ta = value_of(a); tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) - return tarval_sub(ta, tb); + if ((ta != tarval_bad) && (tb != tarval_bad)) + return tarval_sub(ta, tb, mode); return tarval_bad; } /* computed_value_Sub */ @@ -133,7 +154,7 @@ static tarval *computed_value_Sub(ir_node *n) { * Return the value of a Carry. * Special : a op 0, 0 op b */ -static tarval *computed_value_Carry(ir_node *n) { +static tarval *computed_value_Carry(const ir_node *n) { ir_node *a = get_binop_left(n); ir_node *b = get_binop_right(n); ir_mode *m = get_irn_mode(n); @@ -155,7 +176,7 @@ static tarval *computed_value_Carry(ir_node *n) { * Return the value of a Borrow. * Special : a op 0 */ -static tarval *computed_value_Borrow(ir_node *n) { +static tarval *computed_value_Borrow(const ir_node *n) { ir_node *a = get_binop_left(n); ir_node *b = get_binop_right(n); ir_mode *m = get_irn_mode(n); @@ -174,7 +195,7 @@ static tarval *computed_value_Borrow(ir_node *n) { /** * Return the value of an unary Minus. */ -static tarval *computed_value_Minus(ir_node *n) { +static tarval *computed_value_Minus(const ir_node *n) { ir_node *a = get_Minus_op(n); tarval *ta = value_of(a); @@ -187,7 +208,7 @@ static tarval *computed_value_Minus(ir_node *n) { /** * Return the value of a Mul. */ -static tarval *computed_value_Mul(ir_node *n) { +static tarval *computed_value_Mul(const ir_node *n) { ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); ir_mode *mode; @@ -214,78 +235,10 @@ static tarval *computed_value_Mul(ir_node *n) { return tarval_bad; } /* computed_value_Mul */ -/** - * Return the value of a floating point Quot. - */ -static tarval *computed_value_Quot(ir_node *n) { - ir_node *a = get_Quot_left(n); - ir_node *b = get_Quot_right(n); - - tarval *ta = value_of(a); - tarval *tb = value_of(b); - - /* This was missing in original implementation. Why? */ - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_quo(ta, tb); - } - return tarval_bad; -} /* computed_value_Quot */ - -/** - * 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) { - tarval *ta = value_of(a); - tarval *tb = value_of(b); - - /* Compute c1 / c2 or 0 / a, a != 0 */ - if (ta != tarval_bad) { - if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */ - return tarval_div(ta, tb); - else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */ - return ta; - } - return tarval_bad; -} /* do_computed_value_Div */ - -/** - * Return the value of an integer Div. - */ -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. - * Special case: a % 1 - */ -static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) { - tarval *ta = value_of(a); - tarval *tb = value_of(b); - - /* Compute c1 % c2 or a % 1 */ - if (tb != tarval_bad) { - if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */ - return tarval_mod(ta, tb); - 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. - */ -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. */ -static tarval *computed_value_Abs(ir_node *n) { +static tarval *computed_value_Abs(const ir_node *n) { ir_node *a = get_Abs_op(n); tarval *ta = value_of(a); @@ -299,7 +252,7 @@ static tarval *computed_value_Abs(ir_node *n) { * 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(const ir_node *n) { ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); @@ -319,7 +272,7 @@ static tarval *computed_value_And(ir_node *n) { * 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(const ir_node *n) { ir_node *a = get_Or_left(n); ir_node *b = get_Or_right(n); @@ -338,7 +291,7 @@ static tarval *computed_value_Or(ir_node *n) { /** * Return the value of an Eor. */ -static tarval *computed_value_Eor(ir_node *n) { +static tarval *computed_value_Eor(const ir_node *n) { ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); @@ -359,7 +312,7 @@ static tarval *computed_value_Eor(ir_node *n) { /** * Return the value of a Not. */ -static tarval *computed_value_Not(ir_node *n) { +static tarval *computed_value_Not(const ir_node *n) { ir_node *a = get_Not_op(n); tarval *ta = value_of(a); @@ -372,7 +325,7 @@ static tarval *computed_value_Not(ir_node *n) { /** * Return the value of a Shl. */ -static tarval *computed_value_Shl(ir_node *n) { +static tarval *computed_value_Shl(const ir_node *n) { ir_node *a = get_Shl_left(n); ir_node *b = get_Shl_right(n); @@ -388,7 +341,7 @@ static tarval *computed_value_Shl(ir_node *n) { /** * Return the value of a Shr. */ -static tarval *computed_value_Shr(ir_node *n) { +static tarval *computed_value_Shr(const ir_node *n) { ir_node *a = get_Shr_left(n); ir_node *b = get_Shr_right(n); @@ -404,7 +357,7 @@ static tarval *computed_value_Shr(ir_node *n) { /** * Return the value of a Shrs. */ -static tarval *computed_value_Shrs(ir_node *n) { +static tarval *computed_value_Shrs(const ir_node *n) { ir_node *a = get_Shrs_left(n); ir_node *b = get_Shrs_right(n); @@ -418,25 +371,25 @@ static tarval *computed_value_Shrs(ir_node *n) { } /* computed_value_Shrs */ /** - * Return the value of a Rot. + * Return the value of a Rotl. */ -static tarval *computed_value_Rot(ir_node *n) { - ir_node *a = get_Rot_left(n); - ir_node *b = get_Rot_right(n); +static tarval *computed_value_Rotl(const ir_node *n) { + ir_node *a = get_Rotl_left(n); + ir_node *b = get_Rotl_right(n); tarval *ta = value_of(a); tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - return tarval_rot (ta, tb); + return tarval_rotl(ta, tb); } return tarval_bad; -} /* computed_value_Rot */ +} /* computed_value_Rotl */ /** * Return the value of a Conv. */ -static tarval *computed_value_Conv(ir_node *n) { +static tarval *computed_value_Conv(const ir_node *n) { ir_node *a = get_Conv_op(n); tarval *ta = value_of(a); @@ -446,6 +399,45 @@ static tarval *computed_value_Conv(ir_node *n) { return tarval_bad; } /* computed_value_Conv */ +/** + * Calculate the value of a Mux: can be evaluated, if the + * sel and the right input are known. + */ +static tarval *computed_value_Mux(const 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 Confirm: can be evaluated, + * if it has the form Confirm(x, '=', Const). + */ +static tarval *computed_value_Confirm(const ir_node *n) { + /* + * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)). + * Do NOT optimize them away (CondEval wants them), so wait until + * remove_confirm is activated. + */ + if (get_opt_remove_confirm()) { + if (get_Confirm_cmp(n) == pn_Cmp_Eq) { + tarval *tv = value_of(get_Confirm_bound(n)); + if (tv != tarval_bad) + return tv; + } + } + return value_of(get_Confirm_value(n)); +} /* computed_value_Confirm */ + /** * Return the value of a Proj(Cmp). * @@ -455,7 +447,7 @@ static tarval *computed_value_Conv(ir_node *n) { * only 1 is used. * There are several case where we can evaluate a Cmp node, see later. */ -static tarval *computed_value_Proj_Cmp(ir_node *n) { +static tarval *computed_value_Proj_Cmp(const ir_node *n) { ir_node *a = get_Proj_pred(n); ir_node *aa = get_Cmp_left(a); ir_node *ab = get_Cmp_right(a); @@ -530,116 +522,147 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) { * the Cmp result. */ else { - ir_node *aaa = skip_Id(skip_Proj(aa)); - ir_node *aba = skip_Id(skip_Proj(ab)); + ir_node *aaa = skip_Proj(aa); + ir_node *aba = skip_Proj(ab); if ( ( (/* aa is ProjP and aaa is Alloc */ is_Proj(aa) && mode_is_reference(get_irn_mode(aa)) && is_Alloc(aaa)) && ( (/* ab is NULL */ - is_Const(ab) - && mode_is_reference(get_irn_mode(ab)) - && is_Const_null(ab)) + mode_is_reference(get_irn_mode(ab)) + && tarval_is_null(tab)) || (/* ab is other Alloc */ is_Proj(ab) && mode_is_reference(get_irn_mode(ab)) && is_Alloc(aba) && (aaa != aba)))) || (/* aa is NULL and aba is Alloc */ - is_Const(aa) - && mode_is_reference(get_irn_mode(aa)) - && is_Const_null(aa) + mode_is_reference(get_irn_mode(aa)) + && tarval_is_null(taa) && is_Proj(ab) && mode_is_reference(get_irn_mode(ab)) && is_Alloc(aba))) /* 3.: */ - return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); + return new_tarval_from_long(proj_nr & pn_Cmp_Lg, 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). + * Return the value of a floating point Quot. */ -static tarval *computed_value_Proj(ir_node *n) { - ir_node *a = get_Proj_pred(n); - long proj_nr; +static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) { + tarval *ta = value_of(a); + tarval *tb = value_of(b); - switch (get_irn_opcode(a)) { - case iro_Cmp: - return computed_value_Proj_Cmp(n); - - case iro_DivMod: - /* compute either the Div or the Mod part */ - proj_nr = get_Proj_proj(n); - if (proj_nr == pn_DivMod_res_div) - return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); - else if (proj_nr == pn_DivMod_res_mod) - return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); - break; + /* cannot optimize 0 / b = 0 because of NaN */ + if (ta != tarval_bad && tb != tarval_bad) + return tarval_quo(ta, tb); + return tarval_bad; +} /* do_computed_value_Quot */ - case iro_Div: - if (get_Proj_proj(n) == pn_Div_res) - return computed_value(a); - break; +/** + * Calculate the value of an integer Div of two nodes. + * Special case: 0 / b + */ +static tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) { + tarval *ta = value_of(a); + tarval *tb; + const ir_node *dummy; - case iro_Mod: - if (get_Proj_proj(n) == pn_Mod_res) - return computed_value(a); - break; + /* Compute c1 / c2 or 0 / a, a != 0 */ + if (tarval_is_null(ta) && value_not_zero(b, &dummy)) + return ta; /* 0 / b == 0 */ + tb = value_of(b); + if (ta != tarval_bad && tb != tarval_bad) + return tarval_div(ta, tb); + return tarval_bad; +} /* do_computed_value_Div */ - case iro_Quot: - if (get_Proj_proj(n) == pn_Quot_res) - return computed_value(a); - break; +/** + * Calculate the value of an integer Mod of two nodes. + * Special case: a % 1 + */ +static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) { + tarval *ta = value_of(a); + tarval *tb = value_of(b); - default: - return tarval_bad; + /* Compute a % 1 or c1 % c2 */ + if (tarval_is_one(tb)) + return get_mode_null(get_irn_mode(a)); + if (ta != tarval_bad && tb != tarval_bad) + return tarval_mod(ta, tb); + return tarval_bad; +} /* do_computed_value_Mod */ + +/** + * Return the value of a Proj(DivMod). + */ +static tarval *computed_value_Proj_DivMod(const ir_node *n) { + long proj_nr = get_Proj_proj(n); + + /* compute either the Div or the Mod part */ + if (proj_nr == pn_DivMod_res_div) { + const ir_node *a = get_Proj_pred(n); + return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); + } else if (proj_nr == pn_DivMod_res_mod) { + const ir_node *a = get_Proj_pred(n); + return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); } return tarval_bad; -} /* computed_value_Proj */ +} /* computed_value_Proj_DivMod */ /** - * Calculate the value of a Mux: can be evaluated, if the - * sel and the right input are known. + * Return the value of a Proj(Div). */ -static tarval *computed_value_Mux(ir_node *n) { - ir_node *sel = get_Mux_sel(n); - tarval *ts = value_of(sel); +static tarval *computed_value_Proj_Div(const ir_node *n) { + long proj_nr = get_Proj_proj(n); - if (ts == get_tarval_b_true()) { - ir_node *v = get_Mux_true(n); - return value_of(v); + if (proj_nr == pn_Div_res) { + const ir_node *a = get_Proj_pred(n); + return do_computed_value_Div(get_Div_left(a), get_Div_right(a)); } - else if (ts == get_tarval_b_false()) { - ir_node *v = get_Mux_false(n); - return value_of(v); + return tarval_bad; +} /* computed_value_Proj_Div */ + +/** + * Return the value of a Proj(Mod). + */ +static tarval *computed_value_Proj_Mod(const ir_node *n) { + long proj_nr = get_Proj_proj(n); + + if (proj_nr == pn_Mod_res) { + const ir_node *a = get_Proj_pred(n); + return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a)); } return tarval_bad; -} /* computed_value_Mux */ +} /* computed_value_Proj_Mod */ /** - * 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. + * Return the value of a Proj(Quot). */ -static tarval *computed_value_Psi(ir_node *n) { - if (is_Mux(n)) - return computed_value_Mux(n); +static tarval *computed_value_Proj_Quot(const ir_node *n) { + long proj_nr = get_Proj_proj(n); + + if (proj_nr == pn_Quot_res) { + const ir_node *a = get_Proj_pred(n); + return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a)); + } return tarval_bad; -} /* computed_value_Psi */ +} /* computed_value_Proj_Quot */ /** - * Calculate the value of a Confirm: can be evaluated, - * if it has the form Confirm(x, '=', Const). + * Return the value of a Proj. */ -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 */ +static tarval *computed_value_Proj(const ir_node *proj) { + ir_node *n = get_Proj_pred(proj); + + if (n->op->ops.computed_value_Proj != NULL) + return n->op->ops.computed_value_Proj(proj); + return tarval_bad; +} /* computed_value_Proj */ /** * If the parameter n can be computed, return its value, else tarval_bad. @@ -647,7 +670,7 @@ static tarval *computed_value_Confirm(ir_node *n) { * * @param n The node this should be evaluated */ -tarval *computed_value(ir_node *n) { +tarval *computed_value(const ir_node *n) { if (n->op->ops.computed_value) return n->op->ops.computed_value(n); return tarval_bad; @@ -664,9 +687,13 @@ tarval *computed_value(ir_node *n) { */ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops) { -#define CASE(a) \ - case iro_##a: \ - ops->computed_value = computed_value_##a; \ +#define CASE(a) \ + case iro_##a: \ + ops->computed_value = computed_value_##a; \ + break +#define CASE_PROJ(a) \ + case iro_##a: \ + ops->computed_value_Proj = computed_value_Proj_##a; \ break switch (code) { @@ -674,11 +701,10 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops CASE(SymConst); CASE(Add); CASE(Sub); + CASE(Carry); + CASE(Borrow); CASE(Minus); CASE(Mul); - CASE(Quot); - CASE(Div); - CASE(Mod); CASE(Abs); CASE(And); CASE(Or); @@ -687,19 +713,22 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Rot); - CASE(Carry); - CASE(Borrow); + CASE(Rotl); CASE(Conv); - CASE(Proj); CASE(Mux); - CASE(Psi); CASE(Confirm); + CASE_PROJ(Cmp); + CASE_PROJ(DivMod); + CASE_PROJ(Div); + CASE_PROJ(Mod); + CASE_PROJ(Quot); + CASE(Proj); default: /* leave NULL */; } return ops; +#undef CASE_PROJ #undef CASE } /* firm_set_default_computed_value */ @@ -722,10 +751,16 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; - int n_preds = get_Block_n_cfgpreds(n); + int n_preds; + + /* don't optimize dead blocks */ + if (is_Block_dead(n)) + return n; + + n_preds = get_Block_n_cfgpreds(n); - /* The Block constructor does not call optimize, but mature_immBlock - calls the optimization. */ + /* The Block constructor does not call optimize, but mature_immBlock() + calls the optimization. */ assert(get_Block_matured(n)); /* Straightening: a single entry Block following a single exit Block @@ -734,22 +769,26 @@ 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 (n_preds == 1 && is_Jmp(get_Block_cfgpred(n, 0))) { - 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_BLOCK(oldn, n); - } else if (get_opt_control_flow_straightening()) { - n = predblock; - DBG_OPT_STG(oldn, n); - } - } else if (n_preds == 1 && is_Cond(skip_Proj(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_BLOCK(oldn, n); + if (n_preds == 1) { + ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0)); + + if (is_Jmp(pred)) { + ir_node *predblock = get_nodes_block(pred); + if (predblock == oldn) { + /* Jmp jumps into the block it is in -- deal self cycle. */ + n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); + } else if (get_opt_control_flow_straightening()) { + n = predblock; + DBG_OPT_STG(oldn, n); + } + } else if (is_Cond(pred)) { + ir_node *predblock = get_nodes_block(pred); + if (predblock == oldn) { + /* Jmp jumps into the block it is in -- deal self cycle. */ + n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); + } } } else if ((n_preds == 2) && (get_opt_control_flow_weak_simplification())) { @@ -760,15 +799,16 @@ static ir_node *equivalent_node_Block(ir_node *n) ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); - if (is_Proj(a) && - is_Proj(b) && - (get_Proj_pred(a) == get_Proj_pred(b)) && - is_Cond(get_Proj_pred(a)) && - (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(get_Proj_pred(a)); - DBG_OPT_IFSIM1(oldn, a, b, n); + if (is_Proj(a) && is_Proj(b)) { + ir_node *cond = get_Proj_pred(a); + + if (cond == get_Proj_pred(b) && is_Cond(cond) && + get_irn_mode(get_Cond_selector(cond)) == 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(cond); + DBG_OPT_IFSIM1(oldn, a, b, n); + } } } else if (get_opt_unreachable_code() && (n != get_irg_start_block(current_ir_graph)) && @@ -806,10 +846,13 @@ static ir_node *equivalent_node_Block(ir_node *n) * Of course this only happens if the Block of the Jmp is dead. */ static ir_node *equivalent_node_Jmp(ir_node *n) { - /* unreachable code elimination */ - if (is_Block_dead(get_nodes_block(n))) - n = new_Bad(); + ir_node *oldn = n; + /* unreachable code elimination */ + if (is_Block_dead(get_nodes_block(n))) { + n = get_irg_bad(current_ir_graph); + DBG_OPT_DEAD_BLOCK(oldn, n); + } return n; } /* equivalent_node_Jmp */ @@ -824,8 +867,7 @@ static ir_node *equivalent_node_Jmp(ir_node *n) { * 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) -{ +static ir_node *equivalent_node_neutral_zero(ir_node *n) { ir_node *oldn = n; ir_node *a = get_binop_left(n); @@ -862,7 +904,51 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /** * Eor is commutative and has neutral 0. */ -#define equivalent_node_Eor equivalent_node_neutral_zero +static ir_node *equivalent_node_Eor(ir_node *n) { + ir_node *oldn = n; + ir_node *a; + ir_node *b; + + n = equivalent_node_neutral_zero(n); + if (n != oldn) return n; + + a = get_Eor_left(n); + b = get_Eor_right(n); + + if (is_Eor(a)) { + ir_node *aa = get_Eor_left(a); + ir_node *ab = get_Eor_right(a); + + if (aa == b) { + /* (a ^ b) ^ a -> b */ + n = ab; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } else if (ab == b) { + /* (a ^ b) ^ b -> a */ + n = aa; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } + } + if (is_Eor(b)) { + ir_node *ba = get_Eor_left(b); + ir_node *bb = get_Eor_right(b); + + if (ba == a) { + /* a ^ (a ^ b) -> b */ + n = bb; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } else if (bb == a) { + /* a ^ (b ^ a) -> b */ + n = ba; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } + } + return n; +} /* * Optimize a - 0 and (a - x) + x (for modes with wrap-around). @@ -878,14 +964,14 @@ static ir_node *equivalent_node_Add(ir_node *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; + /* 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; + left = get_Add_left(n); right = get_Add_right(n); @@ -921,10 +1007,11 @@ static ir_node *equivalent_node_Add(ir_node *n) { static ir_node *equivalent_node_left_zero(ir_node *n) { ir_node *oldn = n; - ir_node *a = get_binop_left(n); - ir_node *b = get_binop_right(n); + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + tarval *tb = value_of(b); - if (is_Const(b) && is_Const_null(b)) { + if (tarval_is_null(tb)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); @@ -935,7 +1022,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) { #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 +#define equivalent_node_Rotl equivalent_node_left_zero /** * Optimize a - 0 and (a + x) - x (for modes with wrap-around). @@ -950,15 +1037,17 @@ static ir_node *equivalent_node_Sub(ir_node *n) { ir_node *oldn = n; ir_node *b; ir_mode *mode = get_irn_mode(n); + tarval *tb; /* 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; - b = get_Sub_right(n); + b = get_Sub_right(n); + tb = value_of(b); /* Beware: modes might be different */ - if (is_Const(b) && is_Const_null(b)) { + if (tarval_is_null(tb)) { ir_node *a = get_Sub_left(n); if (mode == get_irn_mode(a)) { n = a; @@ -971,7 +1060,7 @@ static ir_node *equivalent_node_Sub(ir_node *n) { /** - * Optimize an "idempotent unary op", ie op(op(n)) = n. + * Optimize an "self-inverse unary op", ie op(op(n)) = n. * * @todo * -(-a) == a, but might overflow two times. @@ -1007,85 +1096,27 @@ static ir_node *equivalent_node_Mul(ir_node *n) { /* we can handle here only the n * n = n bit cases */ if (get_irn_mode(n) == get_irn_mode(a)) { ir_node *b = get_Mul_right(n); + tarval *tv; - /* Mul is commutative and has again an other neutral element. */ - if (is_Const(a) && is_Const_one(a)) { - n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); - } else if (is_Const(b) && is_Const_one(b)) { + /* + * Mul is commutative and has again an other neutral element. + * Constants are place right, so check this case first. + */ + tv = value_of(b); + if (tarval_is_one(tv)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + } else { + tv = value_of(a); + if (tarval_is_one(tv)) { + n = b; + 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) { - ir_node *a = get_Div_left(n); - ir_node *b = get_Div_right(n); - - /* Div is not commutative. */ - if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */ - /* Turn Div into a tuple (mem, bad, a) */ - ir_node *mem = get_Div_mem(n); - ir_node *blk = get_irn_n(n, -1); - turn_into_tuple(n, pn_Div_max); - set_Tuple_pred(n, pn_Div_M, mem); - set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk)); - 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 (is_Const(b) && is_Const_one(b)) { /* Quot(x, 1) == x */ - /* Turn Quot into a tuple (mem, jmp, bad, a) */ - ir_node *mem = get_Quot_mem(n); - ir_node *blk = get_irn_n(n, -1); - turn_into_tuple(n, pn_Quot_max); - set_Tuple_pred(n, pn_Quot_M, mem); - set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk)); - 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) { - ir_node *b = get_DivMod_right(n); - - /* Div is not commutative. */ - if (is_Const(b) && is_Const_one(b)) { /* div(x, 1) == x */ - /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */ - ir_node *a = get_DivMod_left(n); - ir_node *mem = get_Div_mem(n); - ir_node *blk = get_irn_n(n, -1); - ir_mode *mode = get_DivMod_resmode(n); - - turn_into_tuple(n, pn_DivMod_max); - set_Tuple_pred(n, pn_DivMod_M, mem); - set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk)); - 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. */ @@ -1094,16 +1125,25 @@ static ir_node *equivalent_node_Or(ir_node *n) { ir_node *a = get_Or_left(n); ir_node *b = get_Or_right(n); + tarval *tv; if (a == b) { n = a; /* Or has it's own neutral element */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); - } else if (is_Const(a) && is_Const_null(a)) { - n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); - } else if (is_Const(b) && is_Const_null(b)) { + return n; + } + /* constants are cormalized to right, check this site first */ + tv = value_of(b); + if (tarval_is_null(tv)) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); + return n; + } + tv = value_of(a); + if (tarval_is_null(tv)) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); + return n; } return n; @@ -1117,19 +1157,23 @@ static ir_node *equivalent_node_And(ir_node *n) { ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); + tarval *tv; if (a == b) { n = a; /* And has it's own neutral element */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); return n; } - if (is_Const(a) && is_Const_all_one(a)) { - n = b; + /* constants are cormalized to right, check this site first */ + tv = value_of(b); + if (tarval_is_all_one(tv)) { + n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; } - if (is_Const(b) && is_Const_all_one(b)) { - n = a; + tv = value_of(a); + if (tarval_is_all_one(tv)) { + n = b; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; } @@ -1149,7 +1193,6 @@ static ir_node *equivalent_node_And(ir_node *n) { return n; } } - return n; } /* equivalent_node_And */ @@ -1163,27 +1206,31 @@ static ir_node *equivalent_node_Conv(ir_node *n) { ir_mode *n_mode = get_irn_mode(n); ir_mode *a_mode = get_irn_mode(a); +restart: if (n_mode == a_mode) { /* No Conv necessary */ if (get_Conv_strict(n)) { /* special case: the predecessor might be a also a Conv */ if (is_Conv(a)) { if (! get_Conv_strict(a)) { /* first one is not strict, kick it */ - set_Conv_op(n, get_Conv_op(a)); - return n; + a = get_Conv_op(a); + a_mode = get_irn_mode(a); + set_Conv_op(n, a); + goto restart; } - /* else both are strict conv, second is superflous */ - } else if(is_Proj(a)) { - ir_node *pred = get_Proj_pred(a); - if(is_Load(pred)) { - /* loads always return with the exact precision of n_mode */ - assert(get_Load_mode(pred) == n_mode); - return a; + /* else both are strict conv, second is superfluous */ + } else { + if (is_Proj(a)) { + ir_node *pred = get_Proj_pred(a); + if (is_Load(pred)) { + /* loads always return with the exact precision of n_mode */ + assert(get_Load_mode(pred) == n_mode); + return a; + } } + /* leave strict floating point Conv's */ + return n; } - - /* leave strict floating point Conv's */ - return n; } n = a; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); @@ -1191,13 +1238,44 @@ static ir_node *equivalent_node_Conv(ir_node *n) { ir_node *b = get_Conv_op(a); ir_mode *b_mode = get_irn_mode(b); + if (get_Conv_strict(n) && get_Conv_strict(a)) { + /* both are strict conv */ + if (smaller_mode(a_mode, n_mode)) { + /* both are strict, but the first is smaller, so + the second cannot remove more precision, remove the + strict bit */ + set_Conv_strict(n, 0); + } + } if (n_mode == b_mode) { - if (n_mode == mode_b) { - n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - } else if (mode_is_int(n_mode)) { - if (get_mode_size_bits(b_mode) <= get_mode_size_bits(a_mode)) { - n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + if (! get_Conv_strict(n) && ! get_Conv_strict(a)) { + if (n_mode == mode_b) { + n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) { + if (smaller_mode(b_mode, a_mode)) { + n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + } + } + } + if (get_mode_arithmetic(n_mode) == irma_twos_complement && + get_mode_arithmetic(a_mode) == irma_ieee754) { + /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */ + unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0); + unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode); + + if (float_mantissa >= int_mantissa) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + return n; + } + } + if (is_Conv(b)) { + if (smaller_mode(b_mode, a_mode)) { + if (get_Conv_strict(n)) + set_Conv_strict(b, 1); + n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } } @@ -1222,9 +1300,6 @@ static ir_node *equivalent_node_Cast(ir_node *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. */ @@ -1232,7 +1307,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) { int i, n_preds; ir_node *oldn = n; - ir_node *block = NULL; /* to shutup gcc */ + ir_node *block; ir_node *first_val = NULL; /* to shutup gcc */ if (!get_opt_normalize()) return n; @@ -1240,25 +1315,25 @@ static ir_node *equivalent_node_Phi(ir_node *n) { n_preds = get_Phi_n_preds(n); block = get_nodes_block(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 == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */ - return new_Bad(); /* in the Start Block. */ + if (is_Block_dead(block)) /* Control dead */ + return get_irg_bad(current_ir_graph); if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ - /* 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))) - set_Phi_pred(n, i, new_Bad()); - /* Find first non-self-referencing input */ for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ -#if 1 - && (! is_Bad(first_val)) +#if 0 + /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad + * predecessors. Then, Phi nodes in dead code might be removed, causing + * nodes pointing to themself (Add's for instance). + * This is really bad and causes endless recursions in several + * code pathes, so we do NOT optimize such a code. + * This is not that bad as it sounds, optimize_cf() removes bad control flow + * (and bad Phi predecessors), so live code is optimized later. + */ + && (! is_Bad(get_Block_cfgpred(block, i))) #endif ) { /* value not dead */ break; /* then found first value. */ @@ -1267,7 +1342,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) { if (i >= n_preds) { /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - return new_Bad(); + return get_irg_bad(current_ir_graph); } /* search for rest of inputs, determine if any of these @@ -1276,8 +1351,9 @@ static ir_node *equivalent_node_Phi(ir_node *n) { ir_node *scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) && (scnd_val != first_val) -#if 1 - && (! is_Bad(scnd_val)) +#if 0 + /* see above */ + && (! is_Bad(get_Block_cfgpred(block, i))) #endif ) { break; @@ -1294,114 +1370,294 @@ static ir_node *equivalent_node_Phi(ir_node *n) { /** * Several optimizations: - * - no Sync in start block. * - fold Sync-nodes, iff they have only one predecessor except * themselves. */ static ir_node *equivalent_node_Sync(ir_node *n) { - int i, n_preds; + int arity = get_Sync_n_preds(n); + int i; - ir_node *oldn = n; - ir_node *first_val = NULL; /* to shutup gcc */ + for (i = 0; i < arity;) { + ir_node *pred = get_Sync_pred(n, i); + int j; - if (!get_opt_normalize()) return n; + /* Remove Bad predecessors */ + if (is_Bad(pred)) { + del_Sync_n(n, i); + --arity; + continue; + } + + /* Remove duplicate predecessors */ + for (j = 0;; ++j) { + if (j >= i) { + ++i; + break; + } + if (get_Sync_pred(n, j) == pred) { + del_Sync_n(n, i); + --arity; + break; + } + } + } - n_preds = get_Sync_n_preds(n); + if (arity == 0) return get_irg_bad(current_ir_graph); + if (arity == 1) return get_Sync_pred(n, 0); + return n; +} /* equivalent_node_Sync */ - /* 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. */ +/** + * Optimize Proj(Tuple). + */ +static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) { + ir_node *oldn = proj; + ir_node *tuple = get_Proj_pred(proj); + + /* Remove the Tuple/Proj combination. */ + proj = get_Tuple_pred(tuple, get_Proj_proj(proj)); + DBG_OPT_TUPLE(oldn, tuple, proj); + + return proj; +} /* equivalent_node_Proj_Tuple */ + +/** + * Optimize a / 1 = a. + */ +static ir_node *equivalent_node_Proj_Div(ir_node *proj) { + ir_node *oldn = proj; + ir_node *div = get_Proj_pred(proj); + ir_node *b = get_Div_right(div); + tarval *tb = value_of(b); + + /* Div is not commutative. */ + if (tarval_is_one(tb)) { /* div(x, 1) == x */ + switch (get_Proj_proj(proj)) { + case pn_Div_M: + proj = get_Div_mem(div); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + case pn_Div_res: + proj = get_Div_left(div); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + default: + /* we cannot replace the exception Proj's here, this is done in + transform_node_Proj_Div() */ + return proj; } } + return proj; +} /* equivalent_node_Proj_Div */ - if (i >= n_preds) - /* A totally Bad or self-referencing Sync (we didn't break the above loop) */ - return new_Bad(); +/** + * Optimize a / 1.0 = a. + */ +static ir_node *equivalent_node_Proj_Quot(ir_node *proj) { + ir_node *oldn = proj; + ir_node *quot = get_Proj_pred(proj); + ir_node *b = get_Quot_right(quot); + tarval *tb = value_of(b); - /* 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; + /* Div is not commutative. */ + if (tarval_is_one(tb)) { /* Quot(x, 1) == x */ + switch (get_Proj_proj(proj)) { + case pn_Quot_M: + proj = get_Quot_mem(quot); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + case pn_Quot_res: + proj = get_Quot_left(quot); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + default: + /* we cannot replace the exception Proj's here, this is done in + transform_node_Proj_Quot() */ + return proj; + } } + return proj; +} /* equivalent_node_Proj_Quot */ - if (i >= n_preds) { - /* Fold, if no multiple distinct non-self-referencing inputs */ - n = first_val; - DBG_OPT_SYNC(oldn, n); +/** + * Optimize a / 1 = a. + */ +static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) { + ir_node *oldn = proj; + ir_node *divmod = get_Proj_pred(proj); + ir_node *b = get_DivMod_right(divmod); + tarval *tb = value_of(b); + + /* Div is not commutative. */ + if (tarval_is_one(tb)) { /* div(x, 1) == x */ + switch (get_Proj_proj(proj)) { + case pn_DivMod_M: + proj = get_DivMod_mem(divmod); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + case pn_DivMod_res_div: + proj = get_DivMod_left(divmod); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); + return proj; + + default: + /* we cannot replace the exception Proj's here, this is done in + transform_node_Proj_DivMod(). + Note further that the pn_DivMod_res_div case is handled in + computed_value_Proj(). */ + return proj; + } } - return n; -} /* equivalent_node_Sync */ + return proj; +} /* equivalent_node_Proj_DivMod */ /** - * Optimize Proj(Tuple) and gigo() for ProjX in Bad block, - * ProjX(Load) and ProjX(Store). + * Optimize CopyB(mem, x, x) into a Nop. */ -static ir_node *equivalent_node_Proj(ir_node *proj) { - ir_node *oldn = proj; - ir_node *a = get_Proj_pred(proj); +static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) { + ir_node *oldn = proj; + ir_node *copyb = get_Proj_pred(proj); + ir_node *a = get_CopyB_dst(copyb); + ir_node *b = get_CopyB_src(copyb); - if (is_Tuple(a)) { - /* Remove the Tuple/Proj combination. */ - if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) { - proj = get_Tuple_pred(a, get_Proj_proj(proj)); - DBG_OPT_TUPLE(oldn, a, proj); - } else { - /* This should not happen! */ - assert(! "found a Proj with higher number than Tuple predecessors"); - proj = new_Bad(); + if (a == b) { + /* Turn CopyB into a tuple (mem, jmp, bad, bad) */ + switch (get_Proj_proj(proj)) { + case pn_CopyB_M_regular: + proj = get_CopyB_mem(copyb); + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); + break; + + case pn_CopyB_M_except: + case pn_CopyB_X_except: + DBG_OPT_EXC_REM(proj); + proj = get_irg_bad(current_ir_graph); + break; } - } else if (get_irn_mode(proj) == mode_X) { - if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) { - /* Remove dead control flow -- early gigo(). */ - proj = new_Bad(); - } else if (get_opt_ldst_only_null_ptr_exceptions()) { - if (is_Load(a)) { - /* get the Load address */ - ir_node *addr = get_Load_ptr(a); - ir_node *blk = 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); - } - if (get_Proj_proj(proj) == pn_Load_X_except) { - DBG_OPT_EXC_REM(proj); - return new_Bad(); - } else - return new_r_Jmp(current_ir_graph, blk); + } + return proj; +} /* equivalent_node_Proj_CopyB */ + +/** + * Optimize Bounds(idx, idx, upper) into idx. + */ +static ir_node *equivalent_node_Proj_Bound(ir_node *proj) { + ir_node *oldn = proj; + ir_node *bound = get_Proj_pred(proj); + ir_node *idx = get_Bound_index(bound); + ir_node *pred = skip_Proj(idx); + int ret_tuple = 0; + + if (idx == get_Bound_lower(bound)) + ret_tuple = 1; + else if (is_Bound(pred)) { + /* + * idx was Bounds checked in the same MacroBlock previously, + * it is still valid if lower <= pred_lower && pred_upper <= upper. + */ + ir_node *lower = get_Bound_lower(bound); + ir_node *upper = get_Bound_upper(bound); + if (get_Bound_lower(pred) == lower && + get_Bound_upper(pred) == upper && + get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) { + /* + * 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 then. + * So, we must turn in into a tuple. + */ + ret_tuple = 1; + } + } + if (ret_tuple) { + /* Turn Bound into a tuple (mem, jmp, bad, idx) */ + switch (get_Proj_proj(proj)) { + case pn_Bound_M: + DBG_OPT_EXC_REM(proj); + proj = get_Bound_mem(bound); + break; + case pn_Bound_X_except: + DBG_OPT_EXC_REM(proj); + proj = get_irg_bad(current_ir_graph); + break; + case pn_Bound_res: + proj = idx; + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); + break; + default: + /* cannot optimize pn_Bound_X_regular, handled in transform ... */ + ; + } + } + return proj; +} /* equivalent_node_Proj_Bound */ + +/** + * Optimize an Exception Proj(Load) with a non-null address. + */ +static ir_node *equivalent_node_Proj_Load(ir_node *proj) { + if (get_opt_ldst_only_null_ptr_exceptions()) { + if (get_irn_mode(proj) == mode_X) { + ir_node *load = get_Proj_pred(proj); + + /* get the Load address */ + const ir_node *addr = get_Load_ptr(load); + const ir_node *confirm; + + if (value_not_null(addr, &confirm)) { + if (get_Proj_proj(proj) == pn_Load_X_except) { + DBG_OPT_EXC_REM(proj); + return get_irg_bad(current_ir_graph); } - } else if (is_Store(a)) { - /* get the load/store address */ - ir_node *addr = get_Store_ptr(a); - ir_node *blk = 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); - } - if (get_Proj_proj(proj) == pn_Store_X_except) { - DBG_OPT_EXC_REM(proj); - return new_Bad(); - } else - return new_r_Jmp(current_ir_graph, blk); + } + } + } + return proj; +} /* equivalent_node_Proj_Load */ + +/** + * Optimize an Exception Proj(Store) with a non-null address. + */ +static ir_node *equivalent_node_Proj_Store(ir_node *proj) { + if (get_opt_ldst_only_null_ptr_exceptions()) { + if (get_irn_mode(proj) == mode_X) { + ir_node *store = get_Proj_pred(proj); + + /* get the load/store address */ + const ir_node *addr = get_Store_ptr(store); + const ir_node *confirm; + + if (value_not_null(addr, &confirm)) { + if (get_Proj_proj(proj) == pn_Store_X_except) { + DBG_OPT_EXC_REM(proj); + return get_irg_bad(current_ir_graph); } } } } + return proj; +} /* equivalent_node_Proj_Store */ + +/** + * Does all optimizations on nodes that must be done on it's Proj's + * because of creating new nodes. + */ +static ir_node *equivalent_node_Proj(ir_node *proj) { + ir_node *n = get_Proj_pred(proj); + if (get_irn_mode(proj) == mode_X) { + if (is_Block_dead(get_nodes_block(n))) { + /* Remove dead control flow -- early gigo(). */ + return get_irg_bad(current_ir_graph); + } + } + if (n->op->ops.equivalent_node_Proj) + return n->op->ops.equivalent_node_Proj(proj); return proj; } /* equivalent_node_Proj */ @@ -1413,7 +1669,7 @@ static ir_node *equivalent_node_Id(ir_node *n) { do { n = get_Id_pred(n); - } while (get_irn_op(n) == op_Id); + } while (is_Id(n)); DBG_OPT_ID(oldn, n); return n; @@ -1445,41 +1701,54 @@ static ir_node *equivalent_node_Mux(ir_node *n) else if (is_Proj(sel) && !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); + ir_node *f = get_Mux_false(n); + ir_node *t = 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 (is_Cmp(cmp) && get_Cmp_left(cmp) == a) { - ir_node *cmp_r = get_Cmp_right(cmp); - if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - /* Mux(a CMP 0, X, a) */ - if (is_Minus(b) && 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; + if (is_Cmp(cmp)) { + ir_node *const cmp_l = get_Cmp_left(cmp); + ir_node *const cmp_r = get_Cmp_right(cmp); + + switch (proj_nr) { + case pn_Cmp_Eq: + if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */ + (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */ + n = f; 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; + return n; + } + break; + + case pn_Cmp_Lg: + case pn_Cmp_Ne: + if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */ + (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */ + n = t; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + return n; } - } else if (is_Const(b) && is_Const_null(b)) { - /* 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; + break; + } + + /* + * Note: normalization puts the constant on the right side, + * so we check only one case. + */ + if (cmp_l == t && tarval_is_null(value_of(cmp_r))) { + /* Mux(t CMP 0, X, t) */ + if (is_Minus(f) && get_Minus_op(f) == t) { + /* Mux(t CMP 0, -t, t) */ + if (proj_nr == pn_Cmp_Eq) { + /* Mux(t == 0, -t, t) ==> -t */ + n = f; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } else if (proj_nr == pn_Cmp_Eq) { - /* Mux(a == 0, 0, a) ==> 0 */ - n = b; + } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(t != 0, -t, t) ==> t */ + n = t; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } @@ -1489,40 +1758,6 @@ static ir_node *equivalent_node_Mux(ir_node *n) 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. - * - * For == and != can be handled in Proj(Cmp) - */ -static ir_node *equivalent_node_Cmp(ir_node *n) { - ir_node *left = get_Cmp_left(n); - ir_node *right = get_Cmp_right(n); - - if (is_Minus(left) && is_Minus(right) && - !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. @@ -1531,94 +1766,19 @@ static ir_node *equivalent_node_Confirm(ir_node *n) { ir_node *pred = get_Confirm_value(n); pn_Cmp pnc = get_Confirm_cmp(n); - if (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) { + while (is_Confirm(pred) && 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, jmp, bad, bad) */ - ir_node *mem = get_CopyB_mem(n); - ir_node *blk = get_nodes_block(n); - turn_into_tuple(n, pn_CopyB_max); - set_Tuple_pred(n, pn_CopyB_M, mem); - set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk)); - 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, jmp, 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 then. - * So, we must turn in into a tuple. - */ - ret_tuple = 1; - } - } - } - if (ret_tuple) { - /* Turn Bound into a tuple (mem, jmp, bad, idx) */ - ir_node *mem = get_Bound_mem(n); - ir_node *blk = get_nodes_block(n); - turn_into_tuple(n, pn_Bound_max); - set_Tuple_pred(n, pn_Bound_M, mem); - set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */ - set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */ - set_Tuple_pred(n, pn_Bound_res, idx); + n = pred; + pred = get_Confirm_value(n); + pnc = get_Confirm_cmp(n); } + if (get_opt_remove_confirm()) + return get_Confirm_value(n); return n; -} /* equivalent_node_Bound */ +} /** * equivalent_node() returns a node equivalent to input n. It skips all nodes that @@ -1648,44 +1808,50 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op case iro_##a: \ ops->equivalent_node = equivalent_node_##a; \ break +#define CASE_PROJ(a) \ + case iro_##a: \ + ops->equivalent_node_Proj = equivalent_node_Proj_##a; \ + break switch (code) { CASE(Block); CASE(Jmp); CASE(Raise); - CASE(Or); - CASE(Add); CASE(Eor); - CASE(Sub); + CASE(Add); CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Rot); + CASE(Rotl); + CASE(Sub); CASE(Not); CASE(Minus); CASE(Mul); - CASE(Div); - CASE(Quot); - CASE(DivMod); + CASE(Or); CASE(And); CASE(Conv); CASE(Cast); CASE(Phi); CASE(Sync); + CASE_PROJ(Tuple); + CASE_PROJ(Div); + CASE_PROJ(Quot); + CASE_PROJ(DivMod); + CASE_PROJ(CopyB); + CASE_PROJ(Bound); + CASE_PROJ(Load); + CASE_PROJ(Store); CASE(Proj); CASE(Id); CASE(Mux); - CASE(Psi); - CASE(Cmp); CASE(Confirm); - CASE(CopyB); - CASE(Bound); default: /* leave NULL */; } return ops; #undef CASE +#undef CASE_PROJ } /* firm_set_default_equivalent_node */ /** @@ -1697,12 +1863,31 @@ static int is_const_Phi(ir_node *n) { if (! is_Phi(n) || get_irn_arity(n) == 0) return 0; - for (i = get_irn_arity(n) - 1; i >= 0; --i) + for (i = get_irn_arity(n) - 1; i >= 0; --i) { if (! is_Const(get_irn_n(n, i))) return 0; - return 1; + } + return 1; } /* is_const_Phi */ +typedef tarval *(*tarval_sub_type)(tarval *a, tarval *b, ir_mode *mode); +typedef tarval *(*tarval_binop_type)(tarval *a, tarval *b); + +/** + * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter. + */ +static tarval *do_eval(tarval *(*eval)(), tarval *a, tarval *b, ir_mode *mode) { + if (eval == tarval_sub) { + tarval_sub_type func = (tarval_sub_type)eval; + + return func(a, b, mode); + } else { + tarval_binop_type func = (tarval_binop_type)eval; + + return func(a, b); + } +} + /** * Apply an evaluator on a binop with a constant operators (and one Phi). * @@ -1714,7 +1899,7 @@ static int is_const_Phi(ir_node *n) { * * @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 *), ir_mode *mode, int left) { +static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(), ir_mode *mode, int left) { tarval *tv; void **res; ir_node *pred; @@ -1726,7 +1911,7 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)( for (i = 0; i < n; ++i) { pred = get_irn_n(phi, i); tv = get_Const_tarval(pred); - tv = eval(other, tv); + tv = do_eval(eval, other, tv, mode); if (tv == tarval_bad) { /* folding failed, bad */ @@ -1738,7 +1923,7 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)( for (i = 0; i < n; ++i) { pred = get_irn_n(phi, i); tv = get_Const_tarval(pred); - tv = eval(tv, other); + tv = do_eval(eval, tv, other, mode); if (tv == tarval_bad) { /* folding failed, bad */ @@ -1756,6 +1941,50 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)( return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); } /* apply_binop_on_phi */ +/** + * Apply an evaluator on a binop with two constant Phi. + * + * @param a the left Phi node + * @param b the right Phi node + * @param eval an evaluator function + * @param mode the mode of the result, may be different from the mode of the Phi! + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(), ir_mode *mode) { + tarval *tv_l, *tv_r, *tv; + void **res; + ir_node *pred; + ir_graph *irg; + int i, n; + + if (get_nodes_block(a) != get_nodes_block(b)) + return NULL; + + n = get_irn_arity(a); + NEW_ARR_A(void *, res, n); + + for (i = 0; i < n; ++i) { + pred = get_irn_n(a, i); + tv_l = get_Const_tarval(pred); + pred = get_irn_n(b, i); + tv_r = get_Const_tarval(pred); + tv = do_eval(eval, tv_l, tv_r, mode); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return NULL; + } + res[i] = tv; + } + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(a, 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(a), n, (ir_node **)res, mode); +} /* apply_binop_on_2_phis */ + /** * Apply an evaluator on a unop with a constant operator (a Phi). * @@ -1838,17 +2067,17 @@ static ir_node *transform_node_AddSub(ir_node *n) { ir_mode *mode = get_irn_mode(n); if (mode_is_reference(mode)) { - ir_node *left = get_binop_left(n); - ir_node *right = get_binop_right(n); - int ref_bits = get_mode_size_bits(mode); + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + unsigned ref_bits = get_mode_size_bits(mode); if (is_Conv(left)) { - ir_mode *mode = get_irn_mode(left); - int bits = get_mode_size_bits(mode); + ir_mode *lmode = get_irn_mode(left); + unsigned bits = get_mode_size_bits(lmode); if (ref_bits == bits && - mode_is_int(mode) && - get_mode_arithmetic(mode) == irma_twos_complement) { + mode_is_int(lmode) && + get_mode_arithmetic(lmode) == irma_twos_complement) { ir_node *pre = get_Conv_op(left); ir_mode *pre_mode = get_irn_mode(pre); @@ -1866,12 +2095,12 @@ static ir_node *transform_node_AddSub(ir_node *n) { } if (is_Conv(right)) { - ir_mode *mode = get_irn_mode(right); - int bits = get_mode_size_bits(mode); + ir_mode *rmode = get_irn_mode(right); + unsigned bits = get_mode_size_bits(rmode); if (ref_bits == bits && - mode_is_int(mode) && - get_mode_arithmetic(mode) == irma_twos_complement) { + mode_is_int(rmode) && + get_mode_arithmetic(rmode) == irma_twos_complement) { ir_node *pre = get_Conv_op(right); ir_mode *pre_mode = get_irn_mode(pre); @@ -1887,7 +2116,21 @@ static ir_node *transform_node_AddSub(ir_node *n) { } } } + + /* let address arithmetic use unsigned modes */ + if (is_Const(right)) { + ir_mode *rmode = get_irn_mode(right); + + if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) { + /* convert a AddP(P, *s) into AddP(P, *u) */ + ir_mode *nm = get_reference_mode_unsigned_eq(mode); + + ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm); + set_binop_right(n, pre); + } + } } + return n; } /* transform_node_AddSub */ @@ -1901,6 +2144,10 @@ static ir_node *transform_node_AddSub(ir_node *n) { /* check for Op(Const, Phi) */ \ c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\ } \ + else if (is_const_Phi(a) && is_const_Phi(b)) { \ + /* check for Op(Phi, Phi) */ \ + c = apply_binop_on_2_phis(a, b, eval, mode); \ + } \ if (c) { \ DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ return c; \ @@ -1936,6 +2183,17 @@ static ir_node *transform_node_Add(ir_node *n) { b = get_Add_right(n); mode = get_irn_mode(n); + + if (mode_is_reference(mode)) { + ir_mode *lmode = get_irn_mode(a); + + if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) { + /* an Add(a, NULL) is a hidden Conv */ + dbg_info *dbg = get_irn_dbg_info(n); + return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode); + } + } + HANDLE_BINOP_PHI(tarval_add, a, b, c, mode); /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ @@ -1945,7 +2203,7 @@ static ir_node *transform_node_Add(ir_node *n) { if (mode_is_num(mode)) { /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */ if (!is_arch_dep_running() && a == b && mode_is_int(mode)) { - ir_node *block = get_irn_n(n, -1); + ir_node *block = get_nodes_block(n); n = new_rd_Mul( get_irn_dbg_info(n), @@ -1961,7 +2219,7 @@ static ir_node *transform_node_Add(ir_node *n) { n = new_rd_Sub( get_irn_dbg_info(n), current_ir_graph, - get_irn_n(n, -1), + get_nodes_block(n), b, get_Minus_op(a), mode); @@ -1972,81 +2230,13 @@ static ir_node *transform_node_Add(ir_node *n) { n = new_rd_Sub( get_irn_dbg_info(n), current_ir_graph, - get_irn_n(n, -1), + get_nodes_block(n), a, get_Minus_op(b), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); return n; } - if (! is_reassoc_running()) { - /* do NOT execute this code if reassociation is enabled, it does the inverse! */ - if (is_Mul(a)) { - 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); - return n; - } 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); - return n; - } - } - if (is_Mul(b)) { - 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); - return n; - } - 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; - } - } - } if (get_mode_arithmetic(mode) == irma_twos_complement) { /* Here we rely on constants be on the RIGHT side */ if (is_Not(a)) { @@ -2054,14 +2244,14 @@ static ir_node *transform_node_Add(ir_node *n) { if (is_Const(b) && is_Const_one(b)) { /* ~x + 1 = -x */ - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); return n; } if (op == b) { /* ~x + x = -1 */ - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); return n; @@ -2072,7 +2262,7 @@ static ir_node *transform_node_Add(ir_node *n) { if (op == a) { /* x + ~x = -1 */ - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); return n; @@ -2118,6 +2308,29 @@ static ir_node *transform_node_Sub(ir_node *n) { mode = get_irn_mode(n); + if (mode_is_int(mode)) { + ir_mode *lmode = get_irn_mode(a); + + if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) { + /* a Sub(a, NULL) is a hidden Conv */ + dbg_info *dbg = get_irn_dbg_info(n); + n = new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_CONV); + return n; + } + + if (mode == lmode && + get_mode_arithmetic(mode) == irma_twos_complement && + is_Const(a) && + get_Const_tarval(a) == get_mode_minus_one(mode)) { + /* -1 - x -> ~x */ + dbg_info *dbg = get_irn_dbg_info(n); + n = new_rd_Not(dbg, current_ir_graph, get_nodes_block(n), b, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_NOT); + return n; + } + } + restart: HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode); @@ -2158,18 +2371,30 @@ restart: n = new_rd_Add(dbg, irg, block, a, right, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS); return n; - } else if (is_Sub(b)) { /* a - (b - c) -> a + (c - b) */ + } else if (is_Sub(b)) { + /* a - (b - c) -> a + (c - b) + * -> (a - b) + c iff (b - c) is a pointer */ ir_graph *irg = current_ir_graph; dbg_info *s_dbg = get_irn_dbg_info(b); ir_node *s_block = get_nodes_block(b); - ir_node *s_left = get_Sub_right(b); - ir_node *s_right = get_Sub_left(b); + ir_node *s_left = get_Sub_left(b); + ir_node *s_right = get_Sub_right(b); ir_mode *s_mode = get_irn_mode(b); - ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode); - dbg_info *a_dbg = get_irn_dbg_info(n); - ir_node *a_block = get_nodes_block(n); + if (s_mode == mode_P) { + ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, a, s_left, mode); + dbg_info *a_dbg = get_irn_dbg_info(n); + ir_node *a_block = get_nodes_block(n); + + if (s_mode != mode) + s_right = new_r_Conv(irg, a_block, s_right, mode); + n = new_rd_Add(a_dbg, irg, a_block, sub, s_right, mode); + } else { + ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_right, s_left, s_mode); + dbg_info *a_dbg = get_irn_dbg_info(n); + ir_node *a_block = get_nodes_block(n); - n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode); + n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode); + } DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ @@ -2198,7 +2423,7 @@ restart: n = new_rd_Minus( get_irn_dbg_info(n), current_ir_graph, - get_irn_n(n, -1), + get_nodes_block(n), b, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); @@ -2213,7 +2438,7 @@ restart: if (left == b) { if (mode != get_irn_mode(right)) { /* This Sub is an effective Cast */ - right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode); + right = new_r_Conv(get_irn_irg(n), get_nodes_block(n), right, mode); } n = right; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); @@ -2221,7 +2446,7 @@ restart: } else if (right == b) { if (mode != get_irn_mode(left)) { /* This Sub is an effective Cast */ - left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode); + left = new_r_Conv(get_irn_irg(n), get_nodes_block(n), left, mode); } n = left; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); @@ -2238,20 +2463,20 @@ restart: if (left == a) { ir_mode *r_mode = get_irn_mode(right); - n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), right, r_mode); + n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), right, r_mode); if (mode != r_mode) { /* This Sub is an effective Cast */ - n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode); + n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode); } DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); return n; } else if (right == a) { ir_mode *l_mode = get_irn_mode(left); - n = new_r_Minus(get_irn_irg(n), get_irn_n(n, -1), left, l_mode); + n = new_r_Minus(get_irn_irg(n), get_nodes_block(n), left, l_mode); if (mode != l_mode) { /* This Sub is an effective Cast */ - n = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), n, mode); + n = new_r_Conv(get_irn_irg(n), get_nodes_block(n), n, mode); } DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); return n; @@ -2263,16 +2488,16 @@ restart: if (mode == get_irn_mode(b)) { ir_mode *ma, *mb; - - a = get_Conv_op(a); - b = get_Conv_op(b); + ir_node *op_a = get_Conv_op(a); + ir_node *op_b = get_Conv_op(b); /* check if it's allowed to skip the conv */ - ma = get_irn_mode(a); - mb = get_irn_mode(b); + ma = get_irn_mode(op_a); + mb = get_irn_mode(op_b); if (mode_is_reference(ma) && mode_is_reference(mb)) { /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */ + a = op_a; b = op_b; set_Sub_left(n, a); set_Sub_right(n, b); @@ -2286,7 +2511,7 @@ restart: ir_node *mb = get_Mul_right(a); if (ma == b) { - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_rd_Mul( get_irn_dbg_info(n), current_ir_graph, blk, @@ -2301,7 +2526,7 @@ restart: DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); return n; } else if (mb == b) { - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_rd_Mul( get_irn_dbg_info(n), current_ir_graph, blk, @@ -2320,7 +2545,7 @@ restart: if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */ ir_node *x = get_Sub_left(a); ir_node *y = get_Sub_right(a); - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); ir_mode *m_b = get_irn_mode(b); ir_mode *m_y = get_irn_mode(y); ir_mode *add_mode; @@ -2357,7 +2582,7 @@ restart: tv = tarval_add(tv, get_mode_one(mode)); if (tv != tarval_bad) { - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, get_Not_op(b), c, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X); @@ -2381,26 +2606,30 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { ir_mode *smode = get_irn_mode(a); if (ta == get_mode_one(smode)) { - ir_node *blk = get_irn_n(n, -1); + /* (L)1 * (L)b = (L)b */ + ir_node *blk = get_nodes_block(n); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; } else if (ta == get_mode_minus_one(smode)) { - ir_node *blk = get_irn_n(n, -1); + /* (L)-1 * (L)b = (L)b */ + ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); return n; } if (tb == get_mode_one(smode)) { + /* (L)a * (L)1 = (L)a */ ir_node *blk = get_irn_n(a, -1); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; } else if (tb == get_mode_minus_one(smode)) { - ir_node *blk = get_irn_n(n, -1); + /* (L)a * (L)-1 = (L)-a */ + ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); @@ -2436,7 +2665,7 @@ static ir_node *transform_node_Mul(ir_node *n) { 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); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), r, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); return n; } @@ -2484,16 +2713,20 @@ static ir_node *transform_node_Mul(ir_node *n) { if (get_mode_arithmetic(mode) == irma_ieee754) { if (is_Const(a)) { tarval *tv = get_Const_tarval(a); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { - n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), b, b, mode); + if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + && !tarval_is_negative(tv)) { + /* 2.0 * b = b + b */ + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); return n; } } else if (is_Const(b)) { tarval *tv = get_Const_tarval(b); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { - n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, a, mode); + if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + && !tarval_is_negative(tv)) { + /* a * 2.0 = a + a */ + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); return n; } @@ -2506,19 +2739,48 @@ static ir_node *transform_node_Mul(ir_node *n) { * Transform a Div Node. */ static ir_node *transform_node_Div(ir_node *n) { - tarval *tv = value_of(n); ir_mode *mode = get_Div_resmode(n); - ir_node *value = n; + ir_node *a = get_Div_left(n); + ir_node *b = get_Div_right(n); + ir_node *value; + tarval *tv; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + value = apply_binop_on_2_phis(a, b, tarval_div, mode); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + value = n; + tv = value_of(n); if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); DBG_OPT_CSTEVAL(n, value); goto make_tuple; } else { - ir_node *a = get_Div_left(n); - ir_node *b = get_Div_right(n); - ir_node *dummy; + ir_node *a = get_Div_left(n); + ir_node *b = get_Div_right(n); + const ir_node *dummy; if (a == b && value_not_zero(a, &dummy)) { /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ @@ -2531,7 +2793,7 @@ static ir_node *transform_node_Div(ir_node *n) { if (tv == get_mode_minus_one(mode)) { /* a / -1 */ - value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); + value = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode); DBG_OPT_CSTEVAL(n, value); goto make_tuple; } @@ -2547,11 +2809,10 @@ static ir_node *transform_node_Div(ir_node *n) { make_tuple: /* Turn Div into a tuple (mem, jmp, bad, value) */ mem = get_Div_mem(n); - blk = get_irn_n(n, -1); + blk = get_nodes_block(n); /* skip a potential Pin */ - if (is_Pin(mem)) - mem = get_Pin_op(mem); + mem = skip_Pin(mem); turn_into_tuple(n, pn_Div_max); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk)); @@ -2565,19 +2826,48 @@ make_tuple: * Transform a Mod node. */ static ir_node *transform_node_Mod(ir_node *n) { - tarval *tv = value_of(n); ir_mode *mode = get_Mod_resmode(n); - ir_node *value = n; + ir_node *a = get_Mod_left(n); + ir_node *b = get_Mod_right(n); + ir_node *value; + tarval *tv; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + value = apply_binop_on_2_phis(a, b, tarval_mod, mode); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + value = n; + tv = value_of(n); if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); DBG_OPT_CSTEVAL(n, value); goto make_tuple; } else { - ir_node *a = get_Mod_left(n); - ir_node *b = get_Mod_right(n); - ir_node *dummy; + ir_node *a = get_Mod_left(n); + ir_node *b = get_Mod_right(n); + const ir_node *dummy; if (a == b && value_not_zero(a, &dummy)) { /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */ @@ -2606,11 +2896,10 @@ static ir_node *transform_node_Mod(ir_node *n) { make_tuple: /* Turn Mod into a tuple (mem, jmp, bad, value) */ mem = get_Mod_mem(n); - blk = get_irn_n(n, -1); + blk = get_nodes_block(n); /* skip a potential Pin */ - if (is_Pin(mem)) - mem = get_Pin_op(mem); + mem = skip_Pin(mem); turn_into_tuple(n, pn_Mod_max); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk)); @@ -2624,18 +2913,52 @@ make_tuple: * Transform a DivMod node. */ static ir_node *transform_node_DivMod(ir_node *n) { - ir_node *dummy; - ir_node *a = get_DivMod_left(n); - ir_node *b = get_DivMod_right(n); - ir_mode *mode = get_DivMod_resmode(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); - int evaluated = 0; + const ir_node *dummy; + ir_node *a = get_DivMod_left(n); + ir_node *b = get_DivMod_right(n); + ir_mode *mode = get_DivMod_resmode(n); + ir_node *va, *vb; + tarval *ta, *tb; + int evaluated = 0; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0); + vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1); + vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + va = apply_binop_on_2_phis(a, b, tarval_div, mode); + vb = apply_binop_on_2_phis(a, b, tarval_mod, mode); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + ta = value_of(a); + tb = value_of(b); if (tb != tarval_bad) { if (tb == get_mode_one(get_tarval_mode(tb))) { - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, b); + va = a; + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else if (ta != tarval_bad) { tarval *resa, *resb; @@ -2644,28 +2967,30 @@ static ir_node *transform_node_DivMod(ir_node *n) { Jmp for X result!? */ resb = tarval_mod(ta, tb); if (resb == tarval_bad) return n; /* Causes exception! */ - a = new_Const(mode, resa); - b = new_Const(mode, resb); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_Const(mode, resa); + vb = new_Const(mode, resb); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) { - a = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, mode); + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else { /* Try architecture dependent optimization */ - arch_dep_replace_divmod_by_const(&a, &b, n); - evaluated = a != NULL; + va = a; + vb = b; + arch_dep_replace_divmod_by_const(&va, &vb, n); + evaluated = va != NULL; } } else if (a == b) { if (value_not_zero(a, &dummy)) { /* a/a && a != 0 */ - a = new_Const(mode, get_mode_one(mode)); - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_Const(mode, get_mode_one(mode)); + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else { /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ @@ -2673,7 +2998,7 @@ static ir_node *transform_node_DivMod(ir_node *n) { } } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) { /* 0 / non-Const = 0 */ - b = a; + vb = va = a; goto make_tuple; } @@ -2683,16 +3008,15 @@ static ir_node *transform_node_DivMod(ir_node *n) { make_tuple: mem = get_DivMod_mem(n); /* skip a potential Pin */ - if (is_Pin(mem)) - mem = get_Pin_op(mem); + mem = skip_Pin(mem); - blk = get_irn_n(n, -1); + blk = get_nodes_block(n); turn_into_tuple(n, pn_DivMod_max); set_Tuple_pred(n, pn_DivMod_M, mem); set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk)); 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); + set_Tuple_pred(n, pn_DivMod_res_div, va); + set_Tuple_pred(n, pn_DivMod_res_mod, vb); } return n; @@ -2707,25 +3031,33 @@ static ir_node *transform_node_Quot(ir_node *n) { if (get_mode_arithmetic(mode) == irma_ieee754) { ir_node *b = get_Quot_right(n); + tarval *tv = value_of(b); - if (is_Const(b)) { - tarval *tv = get_Const_tarval(b); + if (tv != tarval_bad) { + int rem; + /* + * Floating point constant folding might be disabled here to + * prevent rounding. + * However, as we check for exact result, doing it is safe. + * Switch it on. + */ + rem = tarval_enable_fp_ops(1); tv = tarval_quo(get_mode_one(mode), tv); + (void)tarval_enable_fp_ops(rem); /* Do the transformation if the result is either exact or we are not using strict rules. */ if (tv != tarval_bad && (tarval_ieee754_get_exact() || (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic) == 0)) { - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); ir_node *a = get_Quot_left(n); ir_node *m = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, a, c, mode); ir_node *mem = get_Quot_mem(n); /* skip a potential Pin */ - if (is_Pin(mem)) - mem = get_Pin_op(mem); + mem = skip_Pin(mem); turn_into_tuple(n, pn_Quot_max); set_Tuple_pred(n, pn_Quot_M, mem); set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk)); @@ -2741,6 +3073,7 @@ static ir_node *transform_node_Quot(ir_node *n) { /** * Optimize Abs(x) into x if x is Confirmed >= 0 * Optimize Abs(x) into -x if x is Confirmed <= 0 + * Optimize Abs(-x) int Abs(x) */ static ir_node *transform_node_Abs(ir_node *n) { ir_node *c, *oldn = n; @@ -2755,13 +3088,14 @@ static ir_node *transform_node_Abs(ir_node *n) { /* * We can replace the Abs by -x here. - * We even could add a new Confirm here. + * We even could add a new Confirm here + * (if not twos complement) * * 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); + get_nodes_block(n), a, mode); DBG_OPT_CONFIRM(oldn, n); return n; @@ -2772,10 +3106,45 @@ static ir_node *transform_node_Abs(ir_node *n) { DBG_OPT_CONFIRM(oldn, n); return n; default: + break; + } + if (is_Minus(a)) { + /* Abs(-x) = Abs(x) */ + mode = get_irn_mode(n); + n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, + get_nodes_block(n), get_Minus_op(a), mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X); return n; } + return n; } /* transform_node_Abs */ +/** + * 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. + * + * For == and != can be handled in Proj(Cmp) + */ +static ir_node *transform_node_Cmp(ir_node *n) { + ir_node *oldn = n; + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + + if (is_Minus(left) && is_Minus(right) && + !mode_overflow_on_unary_Minus(get_irn_mode(left))) { + ir_node *const new_left = get_Minus_op(right); + ir_node *const new_right = get_Minus_op(left); + n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, + get_nodes_block(n), new_left, new_right); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP); + } + return n; +} /* transform_node_Cmp */ + + /** * Transform a Cond node. * @@ -2797,7 +3166,8 @@ static ir_node *transform_node_Cond(ir_node *n) { (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)); + ir_node *blk = get_nodes_block(n); + jmp = new_r_Jmp(current_ir_graph, blk); turn_into_tuple(n, pn_Cond_max); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); @@ -2807,12 +3177,16 @@ static ir_node *transform_node_Cond(ir_node *n) { set_Tuple_pred(n, pn_Cond_true, new_Bad()); } /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); + add_End_keepalive(get_irg_end(current_ir_graph), blk); } return n; } /* transform_node_Cond */ -typedef ir_node* (*recursive_transform) (ir_node *n); +/** + * Prototype of a recursive transform function + * for bitwise distributive transformations. + */ +typedef ir_node* (*recursive_transform)(ir_node *n); /** * makes use of distributive laws for and, or, eor @@ -2837,7 +3211,7 @@ static ir_node *transform_bitwise_distributive(ir_node *n, ir_mode *a_mode = get_irn_mode(a_op); ir_mode *b_mode = get_irn_mode(b_op); if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) { - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = exact_copy(n); set_binop_left(n, a_op); @@ -2863,7 +3237,8 @@ static ir_node *transform_bitwise_distributive(ir_node *n, ir_node *b_left = get_binop_left(b); ir_node *b_right = get_binop_right(b); ir_node *c = NULL; - ir_node *op1, *op2; + ir_node *op1 = NULL; + ir_node *op2 = NULL; if (is_op_commutative(op)) { if (a_left == b_left) { @@ -2888,7 +3263,7 @@ static ir_node *transform_bitwise_distributive(ir_node *n, if (c != NULL) { /* (a sop c) & (b sop c) => (a & b) sop c */ - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); ir_node *new_n = exact_copy(n); set_binop_left(new_n, op1); @@ -2904,7 +3279,7 @@ static ir_node *transform_bitwise_distributive(ir_node *n, n = new_rd_And(dbgi, irg, blk, new_n, c, mode); } else { n = exact_copy(a); - set_irn_n(n, -1, blk); + set_nodes_block(n, blk); set_binop_left(n, new_n); set_binop_right(n, c); add_identities(current_ir_graph->value_table, n); @@ -3081,7 +3456,7 @@ static ir_node *transform_node_Eor(ir_node *n) { if (a == b) { /* a ^ a = 0 */ - n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), + n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), mode, get_mode_null(mode)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A); } else if (mode == mode_b && @@ -3089,7 +3464,7 @@ static ir_node *transform_node_Eor(ir_node *n) { is_Const(b) && is_Const_one(b) && is_Cmp(get_Proj_pred(a))) { /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + n = new_r_Proj(current_ir_graph, get_nodes_block(n), 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); @@ -3129,7 +3504,7 @@ static ir_node *transform_node_Not(ir_node *n) { is_Proj(a) && is_Cmp(get_Proj_pred(a))) { /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + n = new_r_Proj(current_ir_graph, get_nodes_block(n), 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; @@ -3160,7 +3535,7 @@ static ir_node *transform_node_Not(ir_node *n) { if (is_Const(add_r) && is_Const_all_one(add_r)) { /* ~(x + -1) = -x */ ir_node *op = get_Add_left(a); - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); } @@ -3192,7 +3567,7 @@ static ir_node *transform_node_Minus(ir_node *n) { /* -(~x) = x + 1 */ ir_node *op = get_Not_op(a); tarval *tv = get_mode_one(mode); - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv); n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode); DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT); @@ -3204,11 +3579,11 @@ static ir_node *transform_node_Minus(ir_node *n) { if (is_Const(c)) { tarval *tv = get_Const_tarval(c); - if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>u (size-1)) = a >>s (size-1) */ ir_node *v = get_Shr_left(a); - n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode); + n = new_rd_Shrs(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode); DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE); return n; } @@ -3220,11 +3595,11 @@ static ir_node *transform_node_Minus(ir_node *n) { if (is_Const(c)) { tarval *tv = get_Const_tarval(c); - if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>s (size-1)) = a >>u (size-1) */ ir_node *v = get_Shrs_left(a); - n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), v, c, mode); + n = new_rd_Shr(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), v, c, mode); DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_PREDICATE); return n; } @@ -3235,7 +3610,7 @@ static ir_node *transform_node_Minus(ir_node *n) { /* - (a-b) = b - a */ ir_node *la = get_Sub_left(a); ir_node *ra = get_Sub_right(a); - ir_node *blk = get_irn_n(n, -1); + ir_node *blk = get_nodes_block(n); n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, ra, la, mode); DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_SUB); @@ -3245,15 +3620,18 @@ static ir_node *transform_node_Minus(ir_node *n) { if (is_Mul(a)) { /* -(a * const) -> a * -const */ ir_node *mul_l = get_Mul_left(a); ir_node *mul_r = get_Mul_right(a); - if (is_Const(mul_r)) { - tarval *tv = tarval_neg(get_Const_tarval(mul_r)); - ir_node *cnst = new_Const(mode, tv); - dbg_info *dbg = get_irn_dbg_info(a); - ir_graph *irg = current_ir_graph; - ir_node *block = get_nodes_block(a); - n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode); - DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C); - return n; + tarval *tv = value_of(mul_r); + if (tv != tarval_bad) { + tv = tarval_neg(tv); + if (tv != tarval_bad) { + ir_node *cnst = new_Const(mode, tv); + dbg_info *dbg = get_irn_dbg_info(a); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(a); + n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C); + return n; + } } } @@ -3273,14 +3651,74 @@ static ir_node *transform_node_Cast(ir_node *n) { get_Const_tarval(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) { - n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred), - get_SymConst_kind(pred), tp); + n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), + get_SymConst_symbol(pred), get_SymConst_kind(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } return n; } /* transform_node_Cast */ +/** + * Transform a Proj(Load) with a non-null address. + */ +static ir_node *transform_node_Proj_Load(ir_node *proj) { + if (get_opt_ldst_only_null_ptr_exceptions()) { + if (get_irn_mode(proj) == mode_X) { + ir_node *load = get_Proj_pred(proj); + + /* get the Load address */ + const ir_node *addr = get_Load_ptr(load); + const 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(load, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Load_X_except) { + DBG_OPT_EXC_REM(proj); + return get_irg_bad(current_ir_graph); + } else { + ir_node *blk = get_nodes_block(load); + return new_r_Jmp(current_ir_graph, blk); + } + } + } + } + return proj; +} /* transform_node_Proj_Load */ + +/** + * Transform a Proj(Store) with a non-null address. + */ +static ir_node *transform_node_Proj_Store(ir_node *proj) { + if (get_opt_ldst_only_null_ptr_exceptions()) { + if (get_irn_mode(proj) == mode_X) { + ir_node *store = get_Proj_pred(proj); + + /* get the load/store address */ + const ir_node *addr = get_Store_ptr(store); + const 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(store, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Store_X_except) { + DBG_OPT_EXC_REM(proj); + return get_irg_bad(current_ir_graph); + } else { + ir_node *blk = get_nodes_block(store); + return new_r_Jmp(current_ir_graph, blk); + } + } + } + } + return proj; +} /* transform_node_Proj_Store */ + /** * Transform a Proj(Div) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. @@ -3288,7 +3726,8 @@ static ir_node *transform_node_Cast(ir_node *n) { static ir_node *transform_node_Proj_Div(ir_node *proj) { ir_node *div = get_Proj_pred(proj); ir_node *b = get_Div_right(div); - ir_node *confirm, *res, *new_mem; + ir_node *res, *new_mem; + const ir_node *confirm; long proj_nr; if (value_not_zero(b, &confirm)) { @@ -3296,8 +3735,7 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { if (confirm == NULL) { /* we are sure we have a Const != 0 */ new_mem = get_Div_mem(div); - if (is_Pin(new_mem)) - new_mem = get_Pin_op(new_mem); + new_mem = skip_Pin(new_mem); set_Div_mem(div, new_mem); set_irn_pinned(div, op_pin_state_floats); } @@ -3336,7 +3774,8 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { static ir_node *transform_node_Proj_Mod(ir_node *proj) { ir_node *mod = get_Proj_pred(proj); ir_node *b = get_Mod_right(mod); - ir_node *confirm, *res, *new_mem; + ir_node *res, *new_mem; + const ir_node *confirm; long proj_nr; if (value_not_zero(b, &confirm)) { @@ -3346,8 +3785,7 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { if (confirm == NULL) { /* we are sure we have a Const != 0 */ new_mem = get_Mod_mem(mod); - if (is_Pin(new_mem)) - new_mem = get_Pin_op(new_mem); + new_mem = skip_Pin(new_mem); set_Mod_mem(mod, new_mem); set_irn_pinned(mod, op_pin_state_floats); } @@ -3394,7 +3832,8 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { static ir_node *transform_node_Proj_DivMod(ir_node *proj) { ir_node *divmod = get_Proj_pred(proj); ir_node *b = get_DivMod_right(divmod); - ir_node *confirm, *res, *new_mem; + ir_node *res, *new_mem; + const ir_node *confirm; long proj_nr; if (value_not_zero(b, &confirm)) { @@ -3404,8 +3843,7 @@ static ir_node *transform_node_Proj_DivMod(ir_node *proj) { if (confirm == NULL) { /* we are sure we have a Const != 0 */ new_mem = get_DivMod_mem(divmod); - if (is_Pin(new_mem)) - new_mem = get_Pin_op(new_mem); + new_mem = skip_Pin(new_mem); set_DivMod_mem(divmod, new_mem); set_irn_pinned(divmod, op_pin_state_floats); } @@ -3468,7 +3906,7 @@ static ir_node *transform_node_Proj_Cond(ir_node *proj) { return proj; } else { /* this case will NEVER be taken, kill it */ - return new_Bad(); + return get_irg_bad(current_ir_graph); } } } @@ -3487,18 +3925,44 @@ static ir_node *create_zero_const(ir_mode *mode) { return cnst; } +/* the order of the values is important! */ +typedef enum const_class { + const_const = 0, + const_like = 1, + const_other = 2 +} const_class; + +static const_class classify_const(const ir_node* n) +{ + if (is_Const(n)) return const_const; + if (is_irn_constlike(n)) return const_like; + return const_other; +} + +/** + * Determines whether r is more constlike or has a larger index (in that order) + * than l. + */ +static int operands_are_normalized(const ir_node *l, const ir_node *r) +{ + const const_class l_order = classify_const(l); + const const_class r_order = classify_const(r); + return + l_order > r_order || + (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r)); +} + /** * Normalizes and optimizes Cmp nodes. */ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { - 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); + ir_node *n = get_Proj_pred(proj); + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + tarval *tv = NULL; + int changed = 0; + ir_mode *mode = NULL; + long proj_nr = get_Proj_proj(proj); /* we can evaluate some cases directly */ switch (proj_nr) { @@ -3514,11 +3978,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { break; } - /* remove Casts */ - if (is_Cast(left)) - left = get_Cast_op(left); - if (is_Cast(right)) - right = get_Cast_op(right); + /* remove Casts of both sides */ + left = skip_Cast(left); + right = skip_Cast(right); /* Remove unnecessary conversions */ /* TODO handle constants */ @@ -3529,7 +3991,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { ir_mode *mode_left = get_irn_mode(op_left); ir_mode *mode_right = get_irn_mode(op_right); - if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)) { + if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode) + && mode_left != mode_b && mode_right != mode_b) { ir_graph *irg = current_ir_graph; ir_node *block = get_nodes_block(n); @@ -3552,7 +4015,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } } - /* remove operation of both sides if possible */ + /* remove operation on both sides if possible */ if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { /* * The following operations are NOT safe for floating point operations, for instance @@ -3626,11 +4089,11 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; - case iro_Rot: - if (get_Rot_right(left) == get_Rot_right(right)) { - /* a ROT X CMP b ROT X ==> a CMP b */ - left = get_Rot_left(left); - right = get_Rot_left(right); + case iro_Rotl: + if (get_Rotl_right(left) == get_Rotl_right(right)) { + /* a ROTL X CMP b ROTL X ==> a CMP b */ + left = get_Rotl_left(left); + right = get_Rotl_left(right); changed |= 1; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } @@ -3697,28 +4160,12 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } } - if (!get_opt_reassociation()) - return proj; - /* * First step: normalize the compare op * by placing the constant on the right side * or moving the lower address node to the left. - * We ignore the case that both are constants - * this case should be optimized away. */ - if (is_Const(right)) { - c = right; - } else if (is_irn_constlike(left)) { - c = left; - left = right; - right = c; - if(!is_Const(c)) - c = NULL; - - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; - } else if (get_irn_idx(left) > get_irn_idx(right)) { + if (!operands_are_normalized(left, right)) { ir_node *t = left; left = right; @@ -3734,9 +4181,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { * 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); + tv = value_of(right); + if (tv != tarval_bad) { + mode = get_irn_mode(right); /* TODO extend to arbitrary constants */ if (is_Conv(left) && tarval_is_null(tv)) { @@ -3800,7 +4247,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* 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)); + tv = tarval_sub(tv, get_mode_one(mode), NULL); if (tv != tarval_bad) { proj_nr ^= pn_Cmp_Eq; @@ -3868,7 +4315,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } if (tv2 != tarval_bad) { - tv2 = tarval_sub(tv, tv2); + tv2 = tarval_sub(tv, tv2, NULL); if (tv2 != tarval_bad) { left = a; @@ -3880,7 +4327,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } /* -a == c ==> a == -c, -a != c ==> a != -c */ else if (is_Minus(left)) { - tarval *tv2 = tarval_sub(get_mode_null(mode), tv); + tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL); if (tv2 != tarval_bad) { left = get_Minus_op(left); @@ -3895,7 +4342,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { if (tv != tarval_bad) { /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */ - if (get_irn_op(left) == op_Abs) { // TODO something is missing here + if (is_Abs(left)) { // TODO something is missing here } } } @@ -4037,7 +4484,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1)); ir_node *sl, *blk; - cond = tarval_sub(cond, tv1); + cond = tarval_sub(cond, tv1, NULL); cond = tarval_shrs(tv, cond); if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) { @@ -4079,7 +4526,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { ir_node *blk = get_irn_n(op, -1); ir_mode *mode = get_irn_mode(v); - tv = tarval_sub(tv, get_mode_one(mode)); + tv = tarval_sub(tv, get_mode_one(mode), NULL); left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode); changed |= 1; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND); @@ -4089,7 +4536,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } if (changed) { - ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ + ir_node *block = get_nodes_block(n); /* create a new compare */ n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right); @@ -4099,6 +4546,90 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { return proj; } /* transform_node_Proj_Cmp */ +/** + * Optimize CopyB(mem, x, x) into a Nop. + */ +static ir_node *transform_node_Proj_CopyB(ir_node *proj) { + ir_node *copyb = get_Proj_pred(proj); + ir_node *a = get_CopyB_dst(copyb); + ir_node *b = get_CopyB_src(copyb); + + if (a == b) { + switch (get_Proj_proj(proj)) { + case pn_CopyB_X_regular: + /* Turn CopyB into a tuple (mem, jmp, bad, bad) */ + DBG_OPT_EXC_REM(proj); + proj = new_r_Jmp(current_ir_graph, get_nodes_block(copyb)); + break; + case pn_CopyB_M_except: + case pn_CopyB_X_except: + DBG_OPT_EXC_REM(proj); + proj = get_irg_bad(current_ir_graph); + break; + default: + break; + } + } + return proj; +} /* transform_node_Proj_CopyB */ + +/** + * Optimize Bounds(idx, idx, upper) into idx. + */ +static ir_node *transform_node_Proj_Bound(ir_node *proj) { + ir_node *oldn = proj; + ir_node *bound = get_Proj_pred(proj); + ir_node *idx = get_Bound_index(bound); + ir_node *pred = skip_Proj(idx); + int ret_tuple = 0; + + if (idx == get_Bound_lower(bound)) + ret_tuple = 1; + else if (is_Bound(pred)) { + /* + * idx was Bounds checked in the same MacroBlock previously, + * it is still valid if lower <= pred_lower && pred_upper <= upper. + */ + ir_node *lower = get_Bound_lower(bound); + ir_node *upper = get_Bound_upper(bound); + if (get_Bound_lower(pred) == lower && + get_Bound_upper(pred) == upper && + get_irn_MacroBlock(bound) == get_irn_MacroBlock(pred)) { + /* + * 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 then. + * So, we must turn in into a tuple. + */ + ret_tuple = 1; + } + } + if (ret_tuple) { + /* Turn Bound into a tuple (mem, jmp, bad, idx) */ + switch (get_Proj_proj(proj)) { + case pn_Bound_M: + DBG_OPT_EXC_REM(proj); + proj = get_Bound_mem(bound); + break; + case pn_Bound_X_except: + DBG_OPT_EXC_REM(proj); + proj = get_irg_bad(current_ir_graph); + break; + case pn_Bound_res: + proj = idx; + DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); + break; + case pn_Bound_X_regular: + DBG_OPT_EXC_REM(proj); + proj = new_r_Jmp(current_ir_graph, get_nodes_block(bound)); + break; + default: + break; + } + } + return proj; +} /* transform_node_Proj_Bound */ + /** * Does all optimizations on nodes that must be done on it's Proj's * because of creating new nodes. @@ -4106,30 +4637,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { 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_Mod: - return transform_node_Proj_Mod(proj); - - case iro_DivMod: - return transform_node_Proj_DivMod(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 */ - return equivalent_node_Proj(proj); - - default: - /* do nothing */ - return proj; - } + if (n->op->ops.transform_node_Proj) + return n->op->ops.transform_node_Proj(proj); + return proj; } /* transform_node_Proj */ /** @@ -4294,12 +4804,12 @@ static ir_node *transform_node_Or_bf_store(ir_node *or) { } /* transform_node_Or_bf_store */ /** - * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl */ -static ir_node *transform_node_Or_Rot(ir_node *or) { +static ir_node *transform_node_Or_Rotl(ir_node *or) { ir_mode *mode = get_irn_mode(or); ir_node *shl, *shr, *block; - ir_node *irn, *x, *c1, *c2, *v, *sub, *n; + ir_node *irn, *x, *c1, *c2, *v, *sub, *n, *rotval; tarval *tv1, *tv2; if (! mode_is_int(mode)) @@ -4336,69 +4846,50 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { 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 (is_Sub(c1)) { - v = c2; - sub = c1; - - if (get_Sub_right(sub) != v) - return or; - - c1 = get_Sub_left(sub); - if (!is_Const(c1)) - return or; - - tv1 = get_Const_tarval(c1); - if (! tarval_is_long(tv1)) - return or; - - if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + != (int) 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); + n = new_r_Rotl(current_ir_graph, block, x, c1, mode); - DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); return n; - } else if (is_Sub(c2)) { - v = c1; - sub = c2; + } - c1 = get_Sub_left(sub); - if (!is_Const(c1)) - return or; + if (is_Sub(c1)) { + v = c2; + sub = c1; + rotval = sub; /* a Rot right is not supported, so use a rot left */ + } else if (is_Sub(c2)) { + v = c1; + sub = c2; + rotval = v; + } else return or; - tv1 = get_Const_tarval(c1); - if (! tarval_is_long(tv1)) - return or; + if (get_Sub_right(sub) != v) + return or; - if (get_tarval_long(tv1) != get_mode_size_bits(mode)) - return or; + c1 = get_Sub_left(sub); + if (!is_Const(c1)) + return or; - /* yet, condition met */ - block = get_irn_n(or, -1); + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; - /* a Rot Left */ - n = new_r_Rot(current_ir_graph, block, x, v, mode); + if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode)) + return or; - DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); - return n; - } + /* yet, condition met */ + block = get_nodes_block(or); - return or; -} /* transform_node_Or_Rot */ + n = new_r_Rotl(current_ir_graph, block, x, rotval, mode); + + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL); + return n; +} /* transform_node_Or_Rotl */ /** * Transform an Or. @@ -4443,88 +4934,288 @@ static ir_node *transform_node_Or(ir_node *n) { HANDLE_BINOP_PHI(tarval_or, a, b, c, mode); n = transform_node_Or_bf_store(n); - n = transform_node_Or_Rot(n); + n = transform_node_Or_Rotl(n); if (n != oldn) return n; - n = transform_bitwise_distributive(n, transform_node_Or); + n = transform_bitwise_distributive(n, transform_node_Or); + + return n; +} /* transform_node_Or */ + + +/* forward */ +static ir_node *transform_node(ir_node *n); + +/** + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl. + * + * Should be moved to reassociation? + */ +static ir_node *transform_node_shift(ir_node *n) { + ir_node *left, *right; + ir_mode *mode; + tarval *tv1, *tv2, *res; + ir_node *in[2], *irn, *block; + + left = get_binop_left(n); + + /* different operations */ + if (get_irn_op(left) != get_irn_op(n)) + return n; + + right = get_binop_right(n); + tv1 = value_of(right); + if (tv1 == tarval_bad) + return n; + + tv2 = value_of(get_binop_right(left)); + if (tv2 == tarval_bad) + return n; + + res = tarval_add(tv1, tv2); + mode = get_irn_mode(n); + + /* beware: a simple replacement works only, if res < modulo shift */ + if (!is_Rotl(n)) { + int modulo_shf = get_mode_modulo_shift(mode); + assert(modulo_shf >= (int) get_mode_size_bits(mode)); + if (modulo_shf > 0) { + tarval *modulo = new_tarval_from_long(modulo_shf, + get_tarval_mode(res)); + + /* shifting too much */ + if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) { + if (is_Shrs(n)) { + ir_graph *irg = get_irn_irg(n); + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu)); + return new_rd_Shrs(dbgi, irg, block, get_binop_left(left), + cnst, mode); + } + + return new_Const(mode, get_mode_null(mode)); + } + } + } else { + res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res))); + } + + /* ok, we can replace it */ + block = get_nodes_block(n); + + 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); +} /* transform_node_shift */ + +/** + * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + */ +static ir_node *transform_node_bitop_shift(ir_node *n) { + ir_node *left; + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_graph *irg; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + tarval *tv1; + tarval *tv2; + tarval *tv_shift; + + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); + + if (!is_Const(right)) + return n; + + left = get_binop_left(n); + op_left = get_irn_op(left); + if (op_left != op_And && op_left != op_Or && op_left != op_Eor) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { + /* TODO: test if sign bit is affectes */ + return n; + } + + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; + + bitop_left = get_binop_left(left); - return n; -} /* transform_node_Or */ + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); + + assert(get_tarval_mode(tv1) == mode); + + if (is_Shl(n)) { + new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shl(tv1, tv2); + } else if(is_Shr(n)) { + new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shr(tv1, tv2); + } else if(is_Shrs(n)) { + new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shrs(tv1, tv2); + } else { + assert(is_Rotl(n)); + new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_rotl(tv1, tv2); + } + assert(get_tarval_mode(tv_shift) == mode); + new_const = new_Const(mode, tv_shift); -/* forward */ -static ir_node *transform_node(ir_node *n); + if (op_left == op_And) { + new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode); + } else if(op_left == op_Or) { + new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode); + } else { + assert(op_left == op_Eor); + new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode); + } + + return new_bitop; +} /** - * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rot. - * - * Should be moved to reassociation? + * normalisation: + * (x << c1) >> c2 <=> x OP (c2-c1) & ((-1 << c1) >> c2) + * also: + * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2) + * (also with x >>s c1 when c1>=c2) */ -static ir_node *transform_node_shift(ir_node *n) { - ir_node *left, *right; - tarval *tv1, *tv2, *res; - ir_mode *mode; - int modulo_shf, flag; - - left = get_binop_left(n); - - /* different operations */ - if (get_irn_op(left) != get_irn_op(n)) +static ir_node *transform_node_shl_shr(ir_node *n) { + ir_node *left; + ir_node *right = get_binop_right(n); + ir_node *x; + ir_graph *irg; + ir_node *block; + ir_mode *mode; + dbg_info *dbgi; + ir_node *new_const; + ir_node *new_shift; + ir_node *new_and; + tarval *tv_shl; + tarval *tv_shr; + tarval *tv_shift; + tarval *tv_mask; + pn_Cmp pnc; + int need_shrs = 0; + + assert(is_Shl(n) || is_Shr(n) || is_Shrs(n)); + + if (!is_Const(right)) return n; - right = get_binop_right(n); - tv1 = value_of(right); - if (tv1 == tarval_bad) - return n; + left = get_binop_left(n); + mode = get_irn_mode(n); + if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) { + ir_node *shr_right = get_binop_right(left); - tv2 = value_of(get_binop_right(left)); - if (tv2 == tarval_bad) - return n; + if (!is_Const(shr_right)) + return n; - res = tarval_add(tv1, tv2); + x = get_binop_left(left); + tv_shr = get_Const_tarval(shr_right); + tv_shl = get_Const_tarval(right); - /* beware: a simple replacement works only, if res < modulo shift */ - mode = get_irn_mode(n); + if (is_Shrs(left)) { + /* shrs variant only allowed if c1 >= c2 */ + if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge)) + return n; - flag = 0; + tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr); + need_shrs = 1; + } else { + tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr); + } + tv_mask = tarval_shl(tv_mask, tv_shl); + } else if(is_Shr(n) && is_Shl(left)) { + ir_node *shl_right = get_Shl_right(left); - modulo_shf = get_mode_modulo_shift(mode); - if (modulo_shf > 0) { - tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res)); + if (!is_Const(shl_right)) + return n; - if (tarval_cmp(res, modulo) & pn_Cmp_Lt) - flag = 1; - } else - flag = 1; + x = get_Shl_left(left); + tv_shr = get_Const_tarval(right); + tv_shl = get_Const_tarval(shl_right); - if (flag) { - /* ok, we can replace it */ - ir_node *in[2], *irn, *block = get_irn_n(n, -1); + tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl); + tv_mask = tarval_shr(tv_mask, tv_shr); + } else { + return n; + } - in[0] = get_binop_left(left); - in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr)); + assert(tv_mask != tarval_bad); + assert(get_tarval_mode(tv_mask) == mode); - irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + + pnc = tarval_cmp(tv_shl, tv_shr); + if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) { + tv_shift = tarval_sub(tv_shr, tv_shl, NULL); + new_const = new_Const(get_tarval_mode(tv_shift), tv_shift); + if (need_shrs) { + new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode); + } else { + new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode); + } + } else { + assert(pnc == pn_Cmp_Gt); + tv_shift = tarval_sub(tv_shl, tv_shr, NULL); + new_const = new_Const(get_tarval_mode(tv_shift), tv_shift); + new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode); + } - DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + new_const = new_Const(mode, tv_mask); + new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode); - return transform_node(irn); - } - return n; -} /* transform_node_shift */ + return new_and; +} /** * 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); - ir_mode *mode = get_irn_mode(n); + ir_node *left = get_Shr_left(n); + ir_node *right = get_Shr_right(n); + ir_mode *mode = get_irn_mode(n); + + HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode); + n = transform_node_shift(n); - HANDLE_BINOP_PHI(tarval_shr, a, b, c, mode); - return transform_node_shift(n); + if (is_Shr(n)) + n = transform_node_shl_shr(n); + if (is_Shr(n)) + n = transform_node_bitop_shift(n); + + return n; } /* transform_node_Shr */ /** @@ -4537,7 +5228,12 @@ static ir_node *transform_node_Shrs(ir_node *n) { ir_mode *mode = get_irn_mode(n); HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode); - return transform_node_shift(n); + n = transform_node_shift(n); + + if (is_Shrs(n)) + n = transform_node_bitop_shift(n); + + return n; } /* transform_node_Shrs */ /** @@ -4550,21 +5246,33 @@ static ir_node *transform_node_Shl(ir_node *n) { ir_mode *mode = get_irn_mode(n); HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode); - return transform_node_shift(n); + n = transform_node_shift(n); + + if (is_Shl(n)) + n = transform_node_shl_shr(n); + if (is_Shl(n)) + n = transform_node_bitop_shift(n); + + return n; } /* transform_node_Shl */ /** - * Transform a Rot. + * Transform a Rotl. */ -static ir_node *transform_node_Rot(ir_node *n) { +static ir_node *transform_node_Rotl(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Rot_left(n); - ir_node *b = get_Rot_right(n); + ir_node *a = get_Rotl_left(n); + ir_node *b = get_Rotl_right(n); ir_mode *mode = get_irn_mode(n); - HANDLE_BINOP_PHI(tarval_rot, a, b, c, mode); - return transform_node_shift(n); -} /* transform_node_Rot */ + HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode); + n = transform_node_shift(n); + + if (is_Rotl(n)) + n = transform_node_bitop_shift(n); + + return n; +} /* transform_node_Rotl */ /** * Transform a Conv. @@ -4580,6 +5288,12 @@ static ir_node *transform_node_Conv(ir_node *n) { return c; } } + + if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */ + ir_mode *mode = get_irn_mode(n); + return new_r_Unknown(current_ir_graph, mode); + } + return n; } /* transform_node_Conv */ @@ -4603,9 +5317,7 @@ static ir_node *transform_node_End(ir_node *n) { } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) { continue; } - /* FIXME: beabi need to keep a Proj(M) */ - if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka)) - in[j++] = ka; + in[j++] = ka; } if (j != n_keepalives) set_End_keepalives(n, j, in); @@ -4614,17 +5326,17 @@ static ir_node *transform_node_End(ir_node *n) { /** returns 1 if a == -b */ static int is_negated_value(ir_node *a, ir_node *b) { - if(is_Minus(a) && get_Minus_op(a) == b) + if (is_Minus(a) && get_Minus_op(a) == b) return 1; - if(is_Minus(b) && get_Minus_op(b) == a) + if (is_Minus(b) && get_Minus_op(b) == a) return 1; - if(is_Sub(a) && is_Sub(b)) { + if (is_Sub(a) && is_Sub(b)) { ir_node *a_left = get_Sub_left(a); ir_node *a_right = get_Sub_right(a); ir_node *b_left = get_Sub_left(b); ir_node *b_right = get_Sub_right(b); - if(a_left == b_right && a_right == b_left) + if (a_left == b_right && a_right == b_left) return 1; } @@ -4637,50 +5349,99 @@ static int is_negated_value(ir_node *a, ir_node *b) { 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); + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_graph *irg = current_ir_graph; + + /* first normalization step: move a possible zero to the false case */ + if (is_Proj(sel)) { + ir_node *cmp = get_Proj_pred(sel); + + if (is_Cmp(cmp)) { + if (is_Const(t) && is_Const_null(t)) { + ir_node *tmp; + + /* Mux(x, 0, y) => Mux(x, y, 0) */ + pn_Cmp pnc = get_Proj_proj(sel); + sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b, + get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp)))); + n = new_rd_Mux(get_irn_dbg_info(n), irg, get_nodes_block(n), sel, t, f, mode); + tmp = t; + t = f; + f = tmp; + } + } + } + /* note: after normalization, false can only happen on default */ if (mode == mode_b) { - ir_node *t = get_Mux_true(n); - ir_node *f = get_Mux_false(n); dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_irn_n(n, -1); + ir_node *block = get_nodes_block(n); ir_graph *irg = current_ir_graph; if (is_Const(t)) { tarval *tv_t = get_Const_tarval(t); if (tv_t == tarval_b_true) { if (is_Const(f)) { + /* Muxb(sel, true, false) = sel */ assert(get_Const_tarval(f) == tarval_b_false); + DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL); return sel; } else { - return new_rd_Or(dbg, irg, block, sel, f, mode_b); - } - } else { - ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); - assert(tv_t == tarval_b_false); - if (is_Const(f)) { - assert(get_Const_tarval(f) == tarval_b_true); - return not_sel; - } else { - return new_rd_And(dbg, irg, block, not_sel, f, mode_b); + /* Muxb(sel, true, x) = Or(sel, x) */ + n = new_rd_Or(dbg, irg, block, sel, f, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL); + return n; } } } else if (is_Const(f)) { tarval *tv_f = get_Const_tarval(f); if (tv_f == tarval_b_true) { + /* Muxb(sel, x, true) = Or(Not(sel), x) */ ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); - return new_rd_Or(dbg, irg, block, not_sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); + n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b); + return n; } else { + /* Muxb(sel, x, false) = And(sel, x) */ assert(tv_f == tarval_b_false); - return new_rd_And(dbg, irg, block, sel, t, mode_b); + n = new_rd_And(dbg, irg, block, sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); + return n; } } } - if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) { + /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */ + if (is_Const(t) && is_Const(f) && mode_is_int(mode)) { + tarval *a = get_Const_tarval(t); + tarval *b = get_Const_tarval(f); + tarval *null = get_tarval_null(mode); + tarval *diff, *min; + + if (tarval_cmp(a, b) & pn_Cmp_Gt) { + diff = tarval_sub(a, b, NULL); + min = b; + } else { + diff = tarval_sub(b, a, NULL); + min = a; + } + + if (diff == get_tarval_one(mode) && min != null) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_graph *irg = current_ir_graph; + ir_node *t = new_Const(mode, tarval_sub(a, min, NULL)); + ir_node *f = new_Const(mode, tarval_sub(b, min, NULL)); + n = new_rd_Mux(dbg, irg, block, sel, f, t, mode); + n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode); + return n; + } + } + + if (is_Proj(sel)) { ir_node *cmp = get_Proj_pred(sel); long pn = get_Proj_proj(sel); - ir_node *f = get_Mux_false(n); - ir_node *t = get_Mux_true(n); /* * Note: normalization puts the constant on the right side, @@ -4688,37 +5449,100 @@ static ir_node *transform_node_Mux(ir_node *n) { * * 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. + * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs + * transformations. */ if (is_Cmp(cmp)) { ir_node *cmp_r = get_Cmp_right(cmp); if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - ir_node *block = get_irn_n(n, -1); + ir_node *block = get_nodes_block(n); + ir_node *cmp_l = get_Cmp_left(cmp); - if(is_negated_value(f, t)) { - ir_node *cmp_left = get_Cmp_left(cmp); + if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) { + /* f = -t */ - /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */ - if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)) - || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))) + if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)) + || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))) { + /* Mux(a >/>= 0, a, -a) = Mux(a Abs(a) */ n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, - cmp_left, mode); + cmp_l, mode); DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; - /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */ - } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)) - || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))) + } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)) + || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))) { + /* Mux(a />= 0, -a, a) ==> -Abs(a) */ n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, - cmp_left, mode); + cmp_l, 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)) { + /* integer only */ + if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) { + /* Mux((a & b) != 0, c, 0) */ + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && f == cmp_r) { + if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) { + if (pn == pn_Cmp_Lg) { + /* Mux((a & 2^C) != 0, 2^C, 0) */ + n = cmp_l; + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } else { + /* Mux((a & 2^C) == 0, 2^C, 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } + return n; + } + } + if (is_Shl(and_r)) { + ir_node *shl_l = get_Shl_left(and_r); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_r == t && f == cmp_r) { + if (pn == pn_Cmp_Lg) { + /* (a & (1 << n)) != 0, (1 << n), 0) */ + n = cmp_l; + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } else { + /* (a & (1 << n)) == 0, (1 << n), 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } + return n; + } + } + } + and_l = get_And_left(cmp_l); + if (is_Shl(and_l)) { + ir_node *shl_l = get_Shl_left(and_l); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_l == t && f == cmp_r) { + if (pn == pn_Cmp_Lg) { + /* ((1 << n) & a) != 0, (1 << n), 0) */ + n = cmp_l; + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } else { + /* ((1 << n) & a) == 0, (1 << n), 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + } + return n; + } + } + } + } + } } } } @@ -4726,42 +5550,39 @@ static ir_node *transform_node_Mux(ir_node *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 */ - -/** - * optimize sync nodes that have other syncs as input we simply add the inputs + * optimize Sync nodes that have other syncs as input we simply add the inputs * of the other sync to our own inputs */ static ir_node *transform_node_Sync(ir_node *n) { - int i, arity; + int arity = get_Sync_n_preds(n); + int i; - arity = get_irn_arity(n); - for(i = 0; i < get_irn_arity(n); /*empty*/) { - int i2, arity2; - ir_node *in = get_irn_n(n, i); - if(!is_Sync(in)) { + for (i = 0; i < arity;) { + ir_node *pred = get_Sync_pred(n, i); + int pred_arity; + int j; + + if (!is_Sync(pred)) { ++i; continue; } - /* set sync input 0 instead of the sync */ - set_irn_n(n, i, get_irn_n(in, 0)); - /* so we check this input again for syncs */ + del_Sync_n(n, i); + --arity; + + pred_arity = get_Sync_n_preds(pred); + for (j = 0; j < pred_arity; ++j) { + ir_node *pred_pred = get_Sync_pred(pred, j); + int k; - /* append all other inputs of the sync to our sync */ - arity2 = get_irn_arity(in); - for(i2 = 1; i2 < arity2; ++i2) { - ir_node *in_in = get_irn_n(in, i2); - add_irn_n(n, in_in); - /* increase arity so we also check the new inputs for syncs */ - arity++; + for (k = 0;; ++k) { + if (k >= arity) { + add_irn_n(n, pred_pred); + ++arity; + break; + } + if (get_Sync_pred(n, k) == pred_pred) break; + } } } @@ -4805,44 +5626,59 @@ static ir_node *transform_node(ir_node *n) { */ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops) { -#define CASE(a) \ - case iro_##a: \ - ops->transform_node = transform_node_##a; \ +#define CASE(a) \ + case iro_##a: \ + ops->transform_node = transform_node_##a; \ + break +#define CASE_PROJ(a) \ + case iro_##a: \ + ops->transform_node_Proj = transform_node_Proj_##a; \ + break +#define CASE_PROJ_EX(a) \ + case iro_##a: \ + ops->transform_node = transform_node_##a; \ + ops->transform_node_Proj = transform_node_Proj_##a; \ break switch (code) { CASE(Add); CASE(Sub); CASE(Mul); - CASE(Div); - CASE(Mod); - CASE(DivMod); + CASE_PROJ_EX(Div); + CASE_PROJ_EX(Mod); + CASE_PROJ_EX(DivMod); CASE(Quot); CASE(Abs); - CASE(Cond); + CASE_PROJ_EX(Cmp); + CASE_PROJ_EX(Cond); CASE(And); - CASE(Or); CASE(Eor); - CASE(Minus); CASE(Not); + CASE(Minus); CASE(Cast); + CASE_PROJ(Load); + CASE_PROJ(Store); + CASE_PROJ(Bound); + CASE_PROJ(CopyB); CASE(Proj); CASE(Phi); + CASE(Or); CASE(Sel); CASE(Shr); CASE(Shrs); CASE(Shl); - CASE(Rot); + CASE(Rotl); CASE(Conv); CASE(End); CASE(Mux); - CASE(Psi); CASE(Sync); default: /* leave NULL */; } return ops; +#undef CASE_PROJ_EX +#undef CASE_PROJ #undef CASE } /* firm_set_default_transform_node */ @@ -4887,26 +5723,31 @@ static int node_cmp_attr_Free(ir_node *a, ir_node *b) { static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { const symconst_attr *pa = get_irn_symconst_attr(a); const symconst_attr *pb = get_irn_symconst_attr(b); - return (pa->num != pb->num) + return (pa->kind != pb->kind) || (pa->sym.type_p != pb->sym.type_p) || (pa->tp != pb->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)); + 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) { const ir_entity *a_ent = get_Sel_entity(a); const ir_entity *b_ent = get_Sel_entity(b); +#if 0 return (a_ent->kind != b_ent->kind) || (a_ent->name != b_ent->name) || (a_ent->owner != b_ent->owner) || (a_ent->ld_name != b_ent->ld_name) || (a_ent->type != b_ent->type); +#endif + /* Matze: inlining of functions can produce 2 entities with same type, + * name, etc. */ + return a_ent != b_ent; } /* node_cmp_attr_Sel */ /** Compares the attributes of two Phi nodes. */ @@ -4914,8 +5755,8 @@ static int node_cmp_attr_Phi(ir_node *a, ir_node *b) { /* we can only enter this function if both nodes have the same number of inputs, hence it is enough to check if one of them is a Phi0 */ if (is_Phi0(a)) { - /* check the Phi0 attribute */ - return get_irn_phi0_attr(a) != get_irn_phi0_attr(b); + /* check the Phi0 pos attribute */ + return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos; } return 0; } /* node_cmp_attr_Phi */ @@ -4954,6 +5795,49 @@ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { get_Store_volatility(b) == volatility_is_volatile); } /* node_cmp_attr_Store */ +/** Compares two exception attributes */ +static int node_cmp_exception(ir_node *a, ir_node *b) { + const except_attr *ea = get_irn_except_attr(a); + const except_attr *eb = get_irn_except_attr(b); + + return ea->pin_state != eb->pin_state; +} + +#define node_cmp_attr_Bound node_cmp_exception + +/** Compares the attributes of two Div nodes. */ +static int node_cmp_attr_Div(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode || + ma->no_remainder != mb->no_remainder; +} /* node_cmp_attr_Div */ + +/** Compares the attributes of two DivMod nodes. */ +static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_DivMod */ + +/** Compares the attributes of two Mod nodes. */ +static int node_cmp_attr_Mod(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_Mod */ + +/** Compares the attributes of two Quot nodes. */ +static int node_cmp_attr_Quot(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_Quot */ + /** 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)); @@ -5037,6 +5921,12 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) CASE(Store); CASE(Confirm); CASE(ASM); + CASE(Div); + CASE(DivMod); + CASE(Mod); + CASE(Quot); + CASE(Bound); + /* FIXME CopyB */ default: /* leave NULL */; } @@ -5046,30 +5936,32 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) } /* 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. + * Compare function for two nodes in the value table. Gets two + * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression. */ int identities_cmp(const void *elt, const void *key) { - ir_node *a, *b; + ir_node *a = (ir_node *)elt; + ir_node *b = (ir_node *)key; int i, irn_arity_a; - a = (void *)elt; - b = (void *)key; - if (a == b) return 0; if ((get_irn_op(a) != get_irn_op(b)) || (get_irn_mode(a) != get_irn_mode(b))) return 1; /* compare if a's in and b's in are of equal length */ - irn_arity_a = get_irn_intra_arity (a); + irn_arity_a = get_irn_intra_arity(a); if (irn_arity_a != get_irn_intra_arity(b)) return 1; - /* for block-local cse and op_pin_state_pinned nodes: */ - if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) { + if (get_irn_pinned(a) == op_pin_state_pinned) { + /* for pinned nodes, the block inputs must be equal */ if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1)) return 1; + } else if (! get_opt_global_cse()) { + /* for block-local CSE both nodes must be in the same MacroBlock */ + if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b)) + return 1; } /* compare a->in[0..ins] with b->in[0..ins] */ @@ -5089,38 +5981,14 @@ int identities_cmp(const void *elt, const void *key) { /* * Calculate a hash value of a node. + * + * @param node The IR-node */ -unsigned ir_node_hash(ir_node *node) { - unsigned h; - int i, irn_arity; - - if (node->op == op_Const) { - /* special value for const, as they only differ in their tarval. */ - h = HASH_PTR(node->attr.con.tv); - 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.symc.sym.type_p); - h = 9*h + HASH_PTR(get_irn_mode(node)); - } else { - - /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */ - 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++) { - h = 9*h + HASH_PTR(get_irn_intra_n(node, i)); - } - - /* ...mode,... */ - h = 9*h + HASH_PTR(get_irn_mode(node)); - /* ...and code */ - h = 9*h + HASH_PTR(get_irn_op(node)); - } - - return h; +unsigned ir_node_hash(const ir_node *node) { + return node->op->ops.hash(node); } /* ir_node_hash */ + pset *new_identities(void) { return new_pset(identities_cmp, N_IR_NODES); } /* new_identities */ @@ -5130,96 +5998,119 @@ void del_identities(pset *value_table) { } /* del_identities */ /** - * Normalize a node by putting constants (and operands with smaller - * node index) on the right + * Normalize a node by putting constants (and operands with larger + * node index) on the right (operator side). * * @param n The node to normalize */ static void normalize_node(ir_node *n) { - if (get_opt_reassociation()) { - if (is_op_commutative(get_irn_op(n))) { - ir_node *l = get_binop_left(n); - ir_node *r = get_binop_right(n); - int l_idx = get_irn_idx(l); - int r_idx = get_irn_idx(r); - - /* For commutative operators perform a OP b == b OP a but keep - constants on the RIGHT side. This helps greatly in some optimizations. - Moreover we use the idx number to make the form deterministic. */ - if (is_irn_constlike(l)) - l_idx = -l_idx; - if (is_irn_constlike(r)) - r_idx = -r_idx; - if (l_idx < r_idx) { - set_binop_left(n, r); - set_binop_right(n, l); - } + if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + + /* For commutative operators perform a OP b == b OP a but keep + * constants on the RIGHT side. This helps greatly in some + * optimizations. Moreover we use the idx number to make the form + * deterministic. */ + if (!operands_are_normalized(l, r)) { + set_binop_left(n, r); + set_binop_right(n, l); } } } /* normalize_node */ /** + * Update the nodes after a match in the value table. If both nodes have + * the same MacroBlock but different Blocks, we must ensure that the node + * with the dominating Block (the node that is near to the MacroBlock header + * is stored in the table. + * Because a MacroBlock has only one "non-exception" flow, we don't need + * dominance info here: We known, that one block must dominate the other and + * following the only block input will allow to find it. + */ +static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) { + ir_node *known_blk, *new_block, *block, *mbh; + + if (get_opt_global_cse()) { + /* Block inputs are meaning less */ + return; + } + known_blk = get_irn_n(known_irn, -1); + new_block = get_irn_n(new_ir_node, -1); + if (known_blk == new_block) { + /* already in the same block */ + return; + } + /* + * We expect the typical case when we built the graph. In that case, the + * known_irn is already the upper one, so checking this should be faster. + */ + block = new_block; + mbh = get_Block_MacroBlock(new_block); + for (;;) { + if (block == known_blk) { + /* ok, we have found it: known_block dominates new_block as expected */ + return; + } + if (block == mbh) { + /* + * We have reached the MacroBlock header NOT founding + * the known_block. new_block must dominate known_block. + * Update known_irn. + */ + set_irn_n(known_irn, -1, new_block); + return; + } + assert(get_Block_n_cfgpreds(block) == 1); + block = get_Block_cfgpred_block(block, 0); + } +} /* update_value_table */ + +/* * 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. * - * @param value_table The value table - * @param n The node to lookup - * - * Looks up the node in a hash table. + * @param value_table the HashSet containing all nodes in the + * current IR graph + * @param n the node to look up * - * 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. + * @return a node that computes the same value as n or n if no such + * node could be found */ -static INLINE ir_node *identify(pset *value_table, ir_node *n) { +ir_node *identify_remember(pset *value_table, ir_node *n) { ir_node *o = NULL; if (!value_table) return n; normalize_node(n); + /* lookup or insert in hash table with given hash key. */ + o = pset_insert(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); + if (o != n) { + update_known_irn(o, n); + } return o; -} /* identify */ +} /* identify_remember */ /** * 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. + * + * @param value_table The value table + * @param n The node to lookup */ 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)) + n = identify_remember(value_table, n); + if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n)) 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. - */ -ir_node *identify_remember(pset *value_table, ir_node *n) { - ir_node *o = NULL; - - if (!value_table) return n; - - normalize_node(n); - /* lookup or insert in hash table with given hash key. */ - o = pset_insert(value_table, n, ir_node_hash(n)); - - if (o != n) { - DBG_OPT_CSE(n, o); - } - - return o; -} /* identify_remember */ - /* 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)) @@ -5251,22 +6142,35 @@ static ir_node *gigo(ir_node *node) { ir_node *block = get_nodes_block(skip_Proj(node)); /* Don't optimize nodes in immature blocks. */ - if (!get_Block_matured(block)) return node; + 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_Block_dead(block)) { + /* control flow from dead block is dead */ + return new_Bad(); + } + + for (i = get_irn_arity(block) - 1; i >= 0; --i) { if (!is_Bad(get_irn_n(block, i))) break; } - if (i == irn_arity) { + if (i < 0) { ir_graph *irg = get_irn_irg(block); /* the start block is never dead */ if (block != get_irg_start_block(irg) - && block != get_irg_end_block(irg)) + && block != get_irg_end_block(irg)) { + /* + * Do NOT kill control flow without setting + * the block to dead of bad things can happen: + * We get a Block that is not reachable be irg_block_walk() + * but can be found by irg_walk()! + */ + set_Block_dead(block); return new_Bad(); + } } } } @@ -5280,7 +6184,7 @@ static ir_node *gigo(ir_node *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))) + is_Block_dead(get_nodes_block(skip_Proj(node)))) return new_Bad(); for (i = 0; i < irn_arity; i++) { @@ -5383,7 +6287,7 @@ ir_node *optimize_node(ir_node *n) { } /* remove unnecessary nodes */ - if (get_opt_constant_folding() || + if (get_opt_algebraic_simplification() || (iro == iro_Phi) || /* always optimize these nodes. */ (iro == iro_Id) || (iro == iro_Proj) || @@ -5411,18 +6315,21 @@ ir_node *optimize_node(ir_node *n) { /* Some more constant expression evaluation that does not allow to free the node. */ iro = get_irn_opcode(n); - if (get_opt_constant_folding() || + if (get_opt_algebraic_simplification() || (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); + n = gigo(n); /* 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); + ir_node *o = n; + n = identify_remember(current_ir_graph->value_table, o); + if (o != n) + DBG_OPT_CSE(o, n); } return n; @@ -5487,7 +6394,10 @@ ir_node *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); + ir_node *o = n; + n = identify_remember(current_ir_graph->value_table, o); + if (o != n) + DBG_OPT_CSE(o, n); } /* Some more constant expression evaluation. */ @@ -5507,8 +6417,12 @@ ir_node *optimize_in_place_2(ir_node *n) { /* Now we have a legal, useful node. Enter it in hash table for cse. 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); + if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + ir_node *o = n; + n = identify_remember(current_ir_graph->value_table, o); + if (o != n) + DBG_OPT_CSE(o, n); + } return n; } /* optimize_in_place_2 */ @@ -5531,10 +6445,69 @@ ir_node *optimize_in_place(ir_node *n) { return optimize_in_place_2(n); } /* optimize_in_place */ +/** + * Calculate a hash value of a Const node. + */ +static unsigned hash_Const(const ir_node *node) { + unsigned h; + + /* special value for const, as they only differ in their tarval. */ + h = HASH_PTR(node->attr.con.tv); + h = 9*h + HASH_PTR(get_irn_mode(node)); + + return h; +} /* hash_Const */ + +/** + * Calculate a hash value of a SymConst node. + */ +static unsigned hash_SymConst(const ir_node *node) { + unsigned h; + + /* special value for const, as they only differ in their symbol. */ + h = HASH_PTR(node->attr.symc.sym.type_p); + h = 9*h + HASH_PTR(get_irn_mode(node)); + + return h; +} /* hash_SymConst */ + +/** + * Set the default hash operation in an ir_op_ops. + * + * @param code the opcode for the default operation + * @param ops the operations initialized + * + * @return + * The operations. + */ +static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops) +{ +#define CASE(a) \ + case iro_##a: \ + ops->hash = hash_##a; \ + break + + /* hash function already set */ + if (ops->hash != NULL) + return ops; + + switch (code) { + CASE(Const); + CASE(SymConst); + default: + /* use input/mode default hash if no function was given */ + ops->hash = firm_default_hash; + } + + return ops; +#undef CASE +} + /* * Sets the default operation for an ir_ops. */ ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) { + ops = firm_set_default_hash(code, 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);