X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=3164c615b318ffc1480aa8b6d720ce2d774c84e2;hb=2b84102b8f003a13d083330fbc6aaff92b6563e9;hp=6542ecfc9bbb8b0f9b3f0f85097868b0c507910b;hpb=884339329314777cf0094cedcd11ebae52656ab2;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 6542ecfc9..3164c615b 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -21,11 +21,11 @@ * @file * @brief iropt --- optimizations intertwined with IR construction. * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck - * @version $Id$ */ #include "config.h" #include +#include #include "irnode_t.h" #include "irgraph_t.h" @@ -34,7 +34,8 @@ #include "iropt_t.h" #include "ircons_t.h" #include "irgmod.h" -#include "irvrfy.h" +#include "irverify.h" +#include "iroptimize.h" #include "tv_t.h" #include "dbginfo_t.h" #include "iropt_dbg.h" @@ -42,19 +43,39 @@ #include "irhooks.h" #include "irarch.h" #include "hashptr.h" -#include "opt_confirms.h" -#include "opt_polymorphy.h" #include "irtools.h" #include "irhooks.h" #include "array_t.h" +#include "vrp.h" +#include "firm_types.h" +#include "bitfiddle.h" +#include "be.h" +#include "error.h" +#include "firmstat_t.h" -/* Make types visible to allow most efficient access */ #include "entity_t.h" +static bool is_Or_Eor_Add(const ir_node *node) +{ + if (is_Or(node) || is_Eor(node) || is_Add(node)) { + ir_node *left = get_binop_left(node); + ir_node *right = get_binop_right(node); + vrp_attr *vrp_left = vrp_get_info(left); + vrp_attr *vrp_right = vrp_get_info(right); + if (vrp_left != NULL && vrp_right != NULL) { + ir_tarval *vrp_val + = tarval_and(vrp_left->bits_not_set, vrp_right->bits_not_set); + return tarval_is_null(vrp_val); + } + } + return false; +} + /** * Returns the tarval of a Const node or tarval_bad for all other nodes. */ -static tarval *default_value_of(const ir_node *n) { +static ir_tarval *default_value_of(const ir_node *n) +{ if (is_Const(n)) return get_Const_tarval(n); /* might return tarval_bad */ else @@ -63,8 +84,8 @@ static tarval *default_value_of(const ir_node *n) { value_of_func value_of_ptr = default_value_of; -/* * Set a new value_of function. */ -void set_value_of_func(value_of_func func) { +void set_value_of_func(value_of_func func) +{ if (func != NULL) value_of_ptr = func; else @@ -74,14 +95,16 @@ void set_value_of_func(value_of_func func) { /** * Return the value of a Constant. */ -static tarval *computed_value_Const(const ir_node *n) { +static ir_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(const ir_node *n) { +static ir_tarval *computed_value_SymConst(const ir_node *n) +{ ir_type *type; ir_entity *ent; @@ -106,34 +129,42 @@ static tarval *computed_value_SymConst(const ir_node *n) { break; } return tarval_bad; -} /* computed_value_SymConst */ +} /** * Return the value of an Add. */ -static tarval *computed_value_Add(const ir_node *n) { +static ir_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); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) return tarval_add(ta, tb); + /* x+~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } + return tarval_bad; -} /* computed_value_Add */ +} /** * Return the value of a Sub. * Special case: a - a */ -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; +static ir_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); + ir_tarval *ta; + ir_tarval *tb; /* NaN - NaN != 0 */ if (! mode_is_float(mode)) { @@ -149,19 +180,19 @@ static tarval *computed_value_Sub(const ir_node *n) { return tarval_sub(ta, tb, mode); return tarval_bad; -} /* computed_value_Sub */ +} /** * Return the value of a Carry. * Special : a op 0, 0 op b */ -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); - - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_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); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { tarval_add(ta, tb); @@ -171,51 +202,52 @@ static tarval *computed_value_Carry(const ir_node *n) { return get_mode_null(m); } return tarval_bad; -} /* computed_value_Carry */ +} /** * Return the value of a Borrow. * Special : a op 0 */ -static tarval *computed_value_Borrow(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); - - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_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); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m); + return tarval_cmp(ta, tb) == ir_relation_less ? get_mode_one(m) : get_mode_null(m); } else if (tarval_is_null(ta)) { return get_mode_null(m); } return tarval_bad; -} /* computed_value_Borrow */ +} /** * Return the value of an unary Minus. */ -static tarval *computed_value_Minus(const ir_node *n) { - ir_node *a = get_Minus_op(n); - tarval *ta = value_of(a); +static ir_tarval *computed_value_Minus(const ir_node *n) +{ + ir_node *a = get_Minus_op(n); + ir_tarval *ta = value_of(a); if (ta != tarval_bad) return tarval_neg(ta); return tarval_bad; -} /* computed_value_Minus */ +} /** * Return the value of a Mul. */ -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; - - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_tarval *computed_value_Mul(const ir_node *n) +{ + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); + ir_mode *mode; mode = get_irn_mode(n); if (mode != get_irn_mode(a)) { @@ -237,72 +269,78 @@ static tarval *computed_value_Mul(const ir_node *n) { } } return tarval_bad; -} /* computed_value_Mul */ - -/** - * Return the value of an Abs. - */ -static tarval *computed_value_Abs(const ir_node *n) { - ir_node *a = get_Abs_op(n); - tarval *ta = value_of(a); - - if (ta != tarval_bad) - return tarval_abs(ta); - - return tarval_bad; -} /* computed_value_Abs */ +} /** * Return the value of an And. * Special case: a & 0, 0 & b */ -static tarval *computed_value_And(const ir_node *n) { - ir_node *a = get_And_left(n); - ir_node *b = get_And_right(n); - - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_tarval *computed_value_And(const ir_node *n) +{ + ir_node *a = get_And_left(n); + ir_node *b = get_And_right(n); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_and (ta, tb); - } else { - if (tarval_is_null(ta)) return ta; - if (tarval_is_null(tb)) return tb; } + + if (tarval_is_null(ta)) return ta; + if (tarval_is_null(tb)) return tb; + + /* x&~x => 0 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_null(get_irn_mode(n)); + } + return tarval_bad; -} /* computed_value_And */ +} /** * Return the value of an Or. * Special case: a | 1...1, 1...1 | b */ -static tarval *computed_value_Or(const ir_node *n) { - ir_node *a = get_Or_left(n); - ir_node *b = get_Or_right(n); - - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_tarval *computed_value_Or(const ir_node *n) +{ + ir_node *a = get_Or_left(n); + ir_node *b = get_Or_right(n); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_or (ta, tb); - } else { - if (tarval_is_all_one(ta)) return ta; - if (tarval_is_all_one(tb)) return tb; + } + + if (tarval_is_all_one(ta)) return ta; + if (tarval_is_all_one(tb)) return tb; + + /* x|~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); } return tarval_bad; -} /* computed_value_Or */ +} /** * Return the value of an Eor. */ -static tarval *computed_value_Eor(const ir_node *n) { +static ir_tarval *computed_value_Eor(const ir_node *n) +{ ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); - tarval *ta, *tb; + ir_tarval *ta, *tb; if (a == b) return get_mode_null(get_irn_mode(n)); + /* x^~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } ta = value_of(a); tb = value_of(b); @@ -311,105 +349,176 @@ static tarval *computed_value_Eor(const ir_node *n) { return tarval_eor(ta, tb); } return tarval_bad; -} /* computed_value_Eor */ +} /** * Return the value of a Not. */ -static tarval *computed_value_Not(const ir_node *n) { - ir_node *a = get_Not_op(n); - tarval *ta = value_of(a); +static ir_tarval *computed_value_Not(const ir_node *n) +{ + ir_node *a = get_Not_op(n); + ir_tarval *ta = value_of(a); if (ta != tarval_bad) return tarval_not(ta); return tarval_bad; -} /* computed_value_Not */ +} + +/** + * Tests whether a shift shifts more bits than available in the mode + */ +static bool is_oversize_shift(const ir_node *n) +{ + ir_node *count = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_tarval *tv = value_of(count); + long modulo_shift; + long shiftval; + if (tv == tarval_bad) + return false; + if (!tarval_is_long(tv)) + return false; + shiftval = get_tarval_long(tv); + modulo_shift = get_mode_modulo_shift(mode); + if (shiftval < 0 || (modulo_shift > 0 && shiftval >= modulo_shift)) + return false; + + return shiftval >= (long)get_mode_size_bits(mode); +} /** * Return the value of a Shl. */ -static tarval *computed_value_Shl(const ir_node *n) { +static ir_tarval *computed_value_Shl(const ir_node *n) +{ ir_node *a = get_Shl_left(n); ir_node *b = get_Shl_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shl(ta, tb); } + + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; -} /* computed_value_Shl */ +} /** * Return the value of a Shr. */ -static tarval *computed_value_Shr(const ir_node *n) { +static ir_tarval *computed_value_Shr(const ir_node *n) +{ ir_node *a = get_Shr_left(n); ir_node *b = get_Shr_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shr(ta, tb); } + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; -} /* computed_value_Shr */ +} /** * Return the value of a Shrs. */ -static tarval *computed_value_Shrs(const ir_node *n) { +static ir_tarval *computed_value_Shrs(const ir_node *n) +{ ir_node *a = get_Shrs_left(n); ir_node *b = get_Shrs_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shrs(ta, tb); } return tarval_bad; -} /* computed_value_Shrs */ +} /** * Return the value of a Rotl. */ -static tarval *computed_value_Rotl(const ir_node *n) { +static ir_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); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_rotl(ta, tb); } return tarval_bad; -} /* computed_value_Rotl */ +} + +bool ir_zero_when_converted(const ir_node *node, ir_mode *dest_mode) +{ + ir_mode *mode = get_irn_mode(node); + if (get_mode_arithmetic(mode) != irma_twos_complement + || get_mode_arithmetic(dest_mode) != irma_twos_complement) + return false; + + if (is_Shl(node)) { + ir_node *count = get_Shl_right(node); + if (is_Const(count)) { + ir_tarval *tv = get_Const_tarval(count); + if (tarval_is_long(tv)) { + long shiftval = get_tarval_long(tv); + long destbits = get_mode_size_bits(dest_mode); + if (shiftval >= destbits + && shiftval < (long)get_mode_modulo_shift(mode)) + return true; + } + } + } + if (is_And(node)) { + ir_node *right = get_And_right(node); + if (is_Const(right)) { + ir_tarval *tv = get_Const_tarval(right); + ir_tarval *conved = tarval_convert_to(tv, dest_mode); + return tarval_is_null(conved); + } + } + return false; +} /** * Return the value of a Conv. */ -static tarval *computed_value_Conv(const ir_node *n) { - ir_node *a = get_Conv_op(n); - tarval *ta = value_of(a); +static ir_tarval *computed_value_Conv(const ir_node *n) +{ + ir_node *a = get_Conv_op(n); + ir_tarval *ta = value_of(a); + ir_mode *mode = get_irn_mode(n); if (ta != tarval_bad) return tarval_convert_to(ta, get_irn_mode(n)); + if (ir_zero_when_converted(a, mode)) + return get_mode_null(mode); + 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) { +static ir_tarval *computed_value_Mux(const ir_node *n) +{ ir_node *sel = get_Mux_sel(n); - tarval *ts = value_of(sel); + ir_tarval *ts = value_of(sel); if (ts == get_tarval_b_true()) { ir_node *v = get_Mux_true(n); @@ -420,178 +529,151 @@ static tarval *computed_value_Mux(const ir_node *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 (jump threading 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; - } +static ir_tarval *computed_value_Confirm(const ir_node *n) +{ + if (get_Confirm_relation(n) == ir_relation_equal) { + ir_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). - * - * This performs a first step of unreachable code elimination. - * Proj can not be computed, but folding a Cmp above the Proj here is - * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which - * only 1 is used. - * There are several case where we can evaluate a Cmp node, see later. + * gives a (conservative) estimation of possible relation when comparing + * left+right */ -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); - long proj_nr = get_Proj_proj(n); - - /* - * BEWARE: a == a is NOT always True for floating Point values, as - * NaN != NaN is defined, so we must check this here. - */ - if (aa == ab && ( - !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt) - ) { /* 1.: */ +ir_relation ir_get_possible_cmp_relations(const ir_node *left, + const ir_node *right) +{ + ir_relation possible = ir_relation_true; + ir_tarval *tv_l = value_of(left); + ir_tarval *tv_r = value_of(right); + ir_mode *mode = get_irn_mode(left); + ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode); + ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode); + + /* both values known - evaluate them */ + if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) { + possible = tarval_cmp(tv_l, tv_r); + /* we can return now, won't get any better */ + return possible; + } + /* a == a is never less or greater (but might be equal or unordered) */ + if (left == right) + possible &= ~ir_relation_less_greater; + /* unordered results only happen for float compares */ + if (!mode_is_float(mode)) + possible &= ~ir_relation_unordered; + /* values can never be less than the least representable number or + * greater than the greatest representable number */ + if (tv_l == min) + possible &= ~ir_relation_greater; + if (tv_l == max) + possible &= ~ir_relation_less; + if (tv_r == max) + possible &= ~ir_relation_greater; + if (tv_r == min) + possible &= ~ir_relation_less; + /* maybe vrp can tell us more */ + possible &= vrp_cmp(left, right); + /* Alloc nodes never return null (but throw an exception) */ + if (is_Alloc(left) && tarval_is_null(tv_r)) + possible &= ~ir_relation_equal; + /* stuff known through confirm nodes */ + if (is_Confirm(left) && get_Confirm_bound(left) == right) { + possible &= get_Confirm_relation(left); + } + if (is_Confirm(right) && get_Confirm_bound(right) == left) { + ir_relation relation = get_Confirm_relation(right); + relation = get_inversed_relation(relation); + possible &= relation; + } + + return possible; +} - /* This is a trick with the bits used for encoding the Cmp - Proj numbers, the following statement is not the same: - return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */ - return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b); - } - else { - tarval *taa = value_of(aa); - tarval *tab = value_of(ab); - ir_mode *mode = get_irn_mode(aa); +static ir_tarval *compute_cmp(const ir_node *cmp) +{ + ir_node *left = get_Cmp_left(cmp); + ir_node *right = get_Cmp_right(cmp); + ir_relation possible = ir_get_possible_cmp_relations(left, right); + ir_relation relation = get_Cmp_relation(cmp); + + /* if none of the requested relations is possible, return false */ + if ((possible & relation) == ir_relation_false) + return tarval_b_false; + /* if possible relations are a subset of the requested ones return true */ + if ((possible & ~relation) == ir_relation_false) + return tarval_b_true; + + return computed_value_Cmp_Confirm(cmp, left, right, relation); +} - /* - * The predecessors of Cmp are target values. We can evaluate - * the Cmp. - */ - if ((taa != tarval_bad) && (tab != tarval_bad)) { - /* strange checks... */ - pn_Cmp flags = tarval_cmp(taa, tab); - if (flags != pn_Cmp_False) { - return new_tarval_from_long (proj_nr & flags, mode_b); - } - } - /* for integer values, we can check against MIN/MAX */ - else if (mode_is_int(mode)) { - /* MIN <=/> x. This results in true/false. */ - if (taa == get_mode_min(mode)) { - /* a compare with the MIN value */ - if (proj_nr == pn_Cmp_Le) - return get_tarval_b_true(); - else if (proj_nr == pn_Cmp_Gt) - return get_tarval_b_false(); - } - /* x >=/< MIN. This results in true/false. */ - else - if (tab == get_mode_min(mode)) { - /* a compare with the MIN value */ - if (proj_nr == pn_Cmp_Ge) - return get_tarval_b_true(); - else if (proj_nr == pn_Cmp_Lt) - return get_tarval_b_false(); - } - /* MAX >=/< x. This results in true/false. */ - else if (taa == get_mode_max(mode)) { - if (proj_nr == pn_Cmp_Ge) - return get_tarval_b_true(); - else if (proj_nr == pn_Cmp_Lt) - return get_tarval_b_false(); - } - /* x <=/> MAX. This results in true/false. */ - else if (tab == get_mode_max(mode)) { - if (proj_nr == pn_Cmp_Le) - return get_tarval_b_true(); - else if (proj_nr == pn_Cmp_Gt) - return get_tarval_b_false(); - } - } - /* - * The predecessors are Allocs or (void*)(0) constants. Allocs never - * return NULL, they raise an exception. Therefore we can predict - * the Cmp result. - */ - else { - ir_node *aaa = skip_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 */ - 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 */ - 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_Lg, mode_b); - } - } - return computed_value_Cmp_Confirm(a, aa, ab, proj_nr); -} /* computed_value_Proj_Cmp */ +/** + * some people want to call compute_cmp directly, in this case we have to + * test the constant folding flag again + */ +static ir_tarval *compute_cmp_ext(const ir_node *cmp) +{ + if (!get_opt_constant_folding()) + return tarval_bad; + return compute_cmp(cmp); +} /** - * Return the value of a floating point Quot. + * Return the value of a Cmp. + * + * The basic idea here is to determine which relations are possible and which + * one are definitely impossible. */ -static tarval *do_computed_value_Quot(const ir_node *a, const ir_node *b) { - tarval *ta = value_of(a); - tarval *tb = value_of(b); +static ir_tarval *computed_value_Cmp(const ir_node *cmp) +{ + /* we can't construct Constb after lowering mode_b nodes */ + if (irg_is_constrained(get_irn_irg(cmp), IR_GRAPH_CONSTRAINT_MODEB_LOWERED)) + return tarval_bad; - /* 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 */ + return compute_cmp(cmp); +} /** - * Calculate the value of an integer Div of two nodes. + * Calculate the value of an integer Div. * 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; +static ir_tarval *do_computed_value_Div(const ir_node *div) +{ + const ir_node *a = get_Div_left(div); + const ir_node *b = get_Div_right(div); + const ir_mode *mode = get_Div_resmode(div); + ir_tarval *ta = value_of(a); + ir_tarval *tb; const ir_node *dummy; - /* Compute c1 / c2 or 0 / a, a != 0 */ - if (tarval_is_null(ta) && value_not_zero(b, &dummy)) - return ta; /* 0 / b == 0 */ + /* cannot optimize 0 / b = 0 because of NaN */ + if (!mode_is_float(mode)) { + if (tarval_is_null(ta) && value_not_zero(b, &dummy)) + return ta; /* 0 / b == 0 if 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 */ +} /** * 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); +static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) +{ + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); /* Compute a % 1 or c1 % c2 */ if (tarval_is_one(tb)) @@ -599,74 +681,45 @@ static tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) { 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_DivMod */ +} /** * Return the value of a Proj(Div). */ -static tarval *computed_value_Proj_Div(const ir_node *n) { +static ir_tarval *computed_value_Proj_Div(const ir_node *n) +{ long proj_nr = get_Proj_proj(n); + if (proj_nr != pn_Div_res) + return tarval_bad; - 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)); - } - return tarval_bad; -} /* computed_value_Proj_Div */ + return do_computed_value_Div(get_Proj_pred(n)); +} /** * Return the value of a Proj(Mod). */ -static tarval *computed_value_Proj_Mod(const ir_node *n) { +static ir_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_Proj_Mod */ - -/** - * Return the value of a Proj(Quot). - */ -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)); + const ir_node *mod = get_Proj_pred(n); + return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod)); } return tarval_bad; -} /* computed_value_Proj_Quot */ +} /** * Return the value of a Proj. */ -static tarval *computed_value_Proj(const ir_node *proj) { +static ir_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. @@ -674,210 +727,29 @@ static tarval *computed_value_Proj(const ir_node *proj) { * * @param n The node this should be evaluated */ -tarval *computed_value(const ir_node *n) { +ir_tarval *computed_value(const ir_node *n) +{ + vrp_attr *vrp = vrp_get_info(n); + if (vrp != NULL && vrp->bits_set == vrp->bits_not_set) + return vrp->bits_set; + if (n->op->ops.computed_value) return n->op->ops.computed_value(n); return tarval_bad; -} /* computed_value */ - -/** - * Set the default computed_value evaluator in an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_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; \ - break -#define CASE_PROJ(a) \ - case iro_##a: \ - ops->computed_value_Proj = computed_value_Proj_##a; \ - break - - switch (code) { - CASE(Const); - CASE(SymConst); - CASE(Add); - CASE(Sub); - CASE(Carry); - CASE(Borrow); - CASE(Minus); - CASE(Mul); - CASE(Abs); - CASE(And); - CASE(Or); - CASE(Eor); - CASE(Not); - CASE(Shl); - CASE(Shr); - CASE(Shrs); - CASE(Rotl); - CASE(Conv); - CASE(Mux); - 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 */ - -/** - * Returns a equivalent block for another block. - * If the block has only one predecessor, this is - * the equivalent one. If the only predecessor of a block is - * the block itself, this is a dead block. - * - * If both predecessors of a block are the branches of a binary - * Cond, the equivalent block is Cond's block. - * - * If all predecessors of a block are bad or lies in a dead - * block, the current block is dead as well. - * - * Note, that blocks are NEVER turned into Bad's, instead - * the dead_block flag is set. So, never test for is_Bad(block), - * always use is_dead_Block(block). - */ -static ir_node *equivalent_node_Block(ir_node *n) -{ - ir_node *oldn = n; - int n_preds; - - /* don't optimize dead or labeled blocks */ - if (is_Block_dead(n) || has_Block_entity(n)) - return n; - - n_preds = get_Block_n_cfgpreds(n); - - /* 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 - can be merged, if it is not the Start block. */ - /* !!! Beware, all Phi-nodes of n must have been optimized away. - 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) { - 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())) { - /* Test whether Cond jumps twice to this block - * The more general case which more than 2 predecessors is handles - * in optimize_cf(), we handle only this special case for speed here. - */ - ir_node *a = get_Block_cfgpred(n, 0); - ir_node *b = get_Block_cfgpred(n, 1); - - 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)) && - (n != get_irg_end_block(current_ir_graph)) ) { - int i; - - /* If all inputs are dead, this block is dead too, except if it is - the start or end block. This is one step of unreachable code - elimination */ - for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) { - ir_node *pred = get_Block_cfgpred(n, i); - ir_node *pred_blk; - - if (is_Bad(pred)) continue; - pred_blk = get_nodes_block(skip_Proj(pred)); - - if (is_Block_dead(pred_blk)) continue; - - if (pred_blk != n) { - /* really found a living input */ - break; - } - } - if (i < 0) { - n = set_Block_dead(n); - DBG_OPT_DEAD_BLOCK(oldn, n); - } - } - - return n; -} /* equivalent_node_Block */ - -/** - * Returns a equivalent node for a Jmp, a Bad :-) - * Of course this only happens if the Block of the Jmp is dead. - */ -static ir_node *equivalent_node_Jmp(ir_node *n) { - 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 */ - -/** Raise is handled in the same way as Jmp. */ -#define equivalent_node_Raise equivalent_node_Jmp - - -/* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See transform_node_Proj_Cond(). */ +} /** * Optimize operations that are commutative and have neutral 0, * 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); ir_node *b = get_binop_right(n); - tarval *tv; + ir_tarval *tv; ir_node *on; /* After running compute_node there is only one constant predecessor. @@ -903,12 +775,13 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) { } return n; -} /* equivalent_node_neutral_zero */ +} /** * Eor is commutative and has neutral 0. */ -static ir_node *equivalent_node_Eor(ir_node *n) { +static ir_node *equivalent_node_Eor(ir_node *n) +{ ir_node *oldn = n; ir_node *a; ir_node *b; @@ -919,9 +792,9 @@ static ir_node *equivalent_node_Eor(ir_node *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 (is_Eor(a) || is_Or_Eor_Add(a)) { + ir_node *aa = get_binop_left(a); + ir_node *ab = get_binop_right(a); if (aa == b) { /* (a ^ b) ^ a -> b */ @@ -935,9 +808,9 @@ static ir_node *equivalent_node_Eor(ir_node *n) { return n; } } - if (is_Eor(b)) { - ir_node *ba = get_Eor_left(b); - ir_node *bb = get_Eor_right(b); + if (is_Eor(b) || is_Or_Eor_Add(b)) { + ir_node *ba = get_binop_left(b); + ir_node *bb = get_binop_right(b); if (ba == a) { /* a ^ (a ^ b) -> b */ @@ -963,7 +836,8 @@ static ir_node *equivalent_node_Eor(ir_node *n) { * Beware: The Mode of an Add may be different than the mode of its * predecessors, so we could not return a predecessors in all cases. */ -static ir_node *equivalent_node_Add(ir_node *n) { +static ir_node *equivalent_node_Add(ir_node *n) +{ ir_node *oldn = n; ir_node *left, *right; ir_mode *mode = get_irn_mode(n); @@ -973,8 +847,11 @@ static ir_node *equivalent_node_Add(ir_node *n) { 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; + if (mode_is_float(mode)) { + ir_graph *irg = get_irn_irg(n); + if (get_irg_fp_model(irg) & fp_strict_algebraic) + return n; + } left = get_Add_left(n); right = get_Add_right(n); @@ -1002,18 +879,19 @@ static ir_node *equivalent_node_Add(ir_node *n) { } } return n; -} /* equivalent_node_Add */ +} /** * optimize operations that are not commutative but have neutral 0 on left, * so a op 0 = a. */ -static ir_node *equivalent_node_left_zero(ir_node *n) { +static ir_node *equivalent_node_left_zero(ir_node *n) +{ ir_node *oldn = n; - ir_node *a = get_binop_left(n); - ir_node *b = get_binop_right(n); - tarval *tb = value_of(b); + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_tarval *tb = value_of(b); if (tarval_is_null(tb)) { n = a; @@ -1021,12 +899,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) { DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } return n; -} /* equivalent_node_left_zero */ - -#define equivalent_node_Shl equivalent_node_left_zero -#define equivalent_node_Shr equivalent_node_left_zero -#define equivalent_node_Shrs equivalent_node_left_zero -#define equivalent_node_Rotl equivalent_node_left_zero +} /** * Optimize a - 0 and (a + x) - x (for modes with wrap-around). @@ -1037,15 +910,19 @@ static ir_node *equivalent_node_left_zero(ir_node *n) { * Beware: The Mode of a Sub may be different than the mode of its * predecessors, so we could not return a predecessors in all cases. */ -static ir_node *equivalent_node_Sub(ir_node *n) { - ir_node *oldn = n; - ir_node *b; - ir_mode *mode = get_irn_mode(n); - tarval *tb; +static ir_node *equivalent_node_Sub(ir_node *n) +{ + ir_node *oldn = n; + ir_node *b; + ir_mode *mode = get_irn_mode(n); + ir_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; + if (mode_is_float(mode)) { + ir_graph *irg = get_irn_irg(n); + if (get_irg_fp_model(irg) & fp_strict_algebraic) + return n; + } b = get_Sub_right(n); tb = value_of(b); @@ -1060,47 +937,40 @@ static ir_node *equivalent_node_Sub(ir_node *n) { } } return n; -} /* equivalent_node_Sub */ +} /** - * Optimize an "self-inverse unary op", ie op(op(n)) = n. + * Optimize an "self-inverse unary op", i.e. op(op(n)) = n. * * @todo * -(-a) == a, but might overflow two times. * We handle it anyway here but the better way would be a * flag. This would be needed for Pascal for instance. */ -static ir_node *equivalent_node_idempotent_unop(ir_node *n) { +static ir_node *equivalent_node_involution(ir_node *n) +{ ir_node *oldn = n; ir_node *pred = get_unop_op(n); - - /* optimize symmetric unop */ if (get_irn_op(pred) == get_irn_op(n)) { n = get_unop_op(pred); - DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY); + DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_INVOLUTION); } return n; -} /* equivalent_node_idempotent_unop */ - -/** Optimize Not(Not(x)) == x. */ -#define equivalent_node_Not equivalent_node_idempotent_unop - -/** -(-x) == x ??? Is this possible or can --x raise an - out of bounds exception if min =! max? */ -#define equivalent_node_Minus equivalent_node_idempotent_unop +} /** * Optimize a * 1 = 1 * a = a. */ -static ir_node *equivalent_node_Mul(ir_node *n) { +static ir_node *equivalent_node_Mul(ir_node *n) +{ ir_node *oldn = n; ir_node *a = get_Mul_left(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; + ir_node *b = get_Mul_right(n); + ir_tarval *tv; /* * Mul is commutative and has again an other neutral element. @@ -1119,24 +989,25 @@ static ir_node *equivalent_node_Mul(ir_node *n) { } } return n; -} /* equivalent_node_Mul */ +} /** * Use algebraic simplification a | a = a | 0 = 0 | a = a. */ -static ir_node *equivalent_node_Or(ir_node *n) { +static ir_node *equivalent_node_Or(ir_node *n) +{ ir_node *oldn = n; - ir_node *a = get_Or_left(n); - ir_node *b = get_Or_right(n); - tarval *tv; + ir_node *a = get_Or_left(n); + ir_node *b = get_Or_right(n); + ir_tarval *tv; if (a == b) { - n = a; /* Or has it's own neutral element */ + n = a; /* idempotence */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); return n; } - /* constants are cormalized to right, check this site first */ + /* constants are normalized to right, check this side first */ tv = value_of(b); if (tarval_is_null(tv)) { n = a; @@ -1151,24 +1022,25 @@ static ir_node *equivalent_node_Or(ir_node *n) { } return n; -} /* equivalent_node_Or */ +} /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a. */ -static ir_node *equivalent_node_And(ir_node *n) { +static ir_node *equivalent_node_And(ir_node *n) +{ ir_node *oldn = n; - ir_node *a = get_And_left(n); - ir_node *b = get_And_right(n); - tarval *tv; + ir_node *a = get_And_left(n); + ir_node *b = get_And_right(n); + ir_tarval *tv; if (a == b) { - n = a; /* And has it's own neutral element */ + n = a; /* idempotence */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); return n; } - /* constants are normalized to right, check this site first */ + /* constants are normalized to right, check this side first */ tv = value_of(b); if (tarval_is_all_one(tv)) { n = a; @@ -1181,8 +1053,13 @@ static ir_node *equivalent_node_And(ir_node *n) { ir_node *convop = get_Conv_op(a); ir_mode *convopmode = get_irn_mode(convop); if (!mode_is_signed(convopmode)) { - if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) { - /* Conv(X) & all_one(mode(X)) = Conv(X) */ + /* Check Conv(all_one) & Const = all_one */ + ir_tarval *one = get_mode_all_one(convopmode); + ir_tarval *conv = tarval_convert_to(one, mode); + ir_tarval *tand = tarval_and(conv, tv); + + if (tarval_is_all_one(tand)) { + /* Conv(X) & Const = X */ n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; @@ -1196,284 +1073,104 @@ static ir_node *equivalent_node_And(ir_node *n) { DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; } - if (is_Or(a)) { - if (b == get_Or_left(a) || b == get_Or_right(a)) { - /* (a|X) & a */ - n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); - return n; - } + /* (a|X) & a => a*/ + if ((is_Or(a) || is_Or_Eor_Add(a)) + && (b == get_binop_left(a) || b == get_binop_right(a))) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; } - if (is_Or(b)) { - if (a == get_Or_left(b) || a == get_Or_right(b)) { - /* a & (a|X) */ - n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); - return n; - } + /* a & (a|X) => a*/ + if ((is_Or(b) || is_Or_Eor_Add(b)) + && (a == get_binop_left(b) || a == get_binop_right(b))) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; } return n; -} /* equivalent_node_And */ +} /** * Try to remove useless Conv's: */ -static ir_node *equivalent_node_Conv(ir_node *n) { +static ir_node *equivalent_node_Conv(ir_node *n) +{ ir_node *oldn = n; ir_node *a = get_Conv_op(n); ir_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)) { - ir_node *p = a; - - /* neither Minus nor Abs nor Confirm change the precision, - so we can "look-through" */ - for (;;) { - if (is_Minus(p)) { - p = get_Minus_op(p); - } else if (is_Abs(p)) { - p = get_Abs_op(p); - } else if (is_Confirm(p)) { - p = get_Confirm_value(p); - } else { - /* stop here */ - break; - } - } - if (is_Conv(p) && get_Conv_strict(p)) { - /* we known already, that a_mode == n_mode, and neither - Abs nor Minus change the mode, so the second Conv - can be kicked */ - assert(get_irn_mode(p) == n_mode); - n = a; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - return n; - } - if (is_Proj(p)) { - ir_node *pred = get_Proj_pred(p); - if (is_Load(pred)) { - /* Loads always return with the exact precision of n_mode */ - assert(get_Load_mode(pred) == n_mode); - n = a; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - return n; - } - if (is_Proj(pred) && get_Proj_proj(pred) == pn_Start_T_args) { - pred = get_Proj_pred(pred); - if (is_Start(pred)) { - /* Arguments always return with the exact precision, - as strictConv's are place before Call -- if the - caller was compiled with the same setting. - Otherwise, the semantics is probably still right. */ - assert(get_irn_mode(p) == n_mode); - n = a; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - return n; - } - } - } - if (is_Conv(a)) { - /* special case: the immediate predecessor is also a Conv */ - if (! get_Conv_strict(a)) { - /* first one is not strict, kick it */ - 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 superfluous */ - n = a; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - return n; - } - } else { - n = a; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); - return n; - } + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); + return n; } else if (is_Conv(a)) { /* Conv(Conv(b)) */ 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 (! 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); - return n; - } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) { - if (values_in_mode(b_mode, a_mode)) { - n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - return n; - } - } - } - if (mode_is_int(n_mode) && 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); - return n; - } - } + if (n_mode == b_mode && values_in_mode(b_mode, a_mode)) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + return n; } } return n; -} /* equivalent_node_Conv */ - -/** - * A Cast may be removed if the type of the previous node - * is already the type of the Cast. - */ -static ir_node *equivalent_node_Cast(ir_node *n) { - ir_node *oldn = n; - ir_node *pred = get_Cast_op(n); - - if (get_irn_type(pred) == get_Cast_type(n)) { - n = pred; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST); - } - return n; -} /* equivalent_node_Cast */ +} /** * - fold Phi-nodes, iff they have only one predecessor except * themselves. */ -static ir_node *equivalent_node_Phi(ir_node *n) { +static ir_node *equivalent_node_Phi(ir_node *n) +{ int i, n_preds; ir_node *oldn = n; - ir_node *block; ir_node *first_val = NULL; /* to shutup gcc */ - if (!get_opt_normalize()) return n; + if (!get_opt_optimize() && + !irg_is_constrained(get_irn_irg(n), IR_GRAPH_CONSTRAINT_CONSTRUCTION)) + return n; n_preds = get_Phi_n_preds(n); - block = get_nodes_block(n); - 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. */ + /* Phi of dead Region without predecessors. */ + if (n_preds == 0) + return n; /* 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 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. */ + /* not self pointer */ + if (first_val != n) { + /* then found first value. */ + break; } } - if (i >= n_preds) { - /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - return get_irg_bad(current_ir_graph); - } - /* search for rest of inputs, determine if any of these are non-self-referencing */ while (++i < n_preds) { ir_node *scnd_val = get_Phi_pred(n, i); - if ( (scnd_val != n) - && (scnd_val != first_val) -#if 0 - /* see above */ - && (! is_Bad(get_Block_cfgpred(block, i))) -#endif - ) { + if (scnd_val != n && scnd_val != first_val) { break; } } - if (i >= n_preds) { + if (i >= n_preds && !is_Dummy(first_val)) { /* Fold, if no multiple distinct non-self-referencing inputs */ n = first_val; DBG_OPT_PHI(oldn, n); } return n; -} /* equivalent_node_Phi */ - -/** - * Several optimizations: - * - fold Sync-nodes, iff they have only one predecessor except - * themselves. - */ -static ir_node *equivalent_node_Sync(ir_node *n) { - int arity = get_Sync_n_preds(n); - int i; - - for (i = 0; i < arity;) { - ir_node *pred = get_Sync_pred(n, i); - int j; - - /* 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; - } - } - } - - if (arity == 0) return get_irg_bad(current_ir_graph); - if (arity == 1) return get_Sync_pred(n, 0); - return n; -} /* equivalent_node_Sync */ +} /** * Optimize Proj(Tuple). */ -static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) { +static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) +{ ir_node *oldn = proj; ir_node *tuple = get_Proj_pred(proj); @@ -1482,16 +1179,17 @@ static ir_node *equivalent_node_Proj_Tuple(ir_node *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); +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); + ir_tarval *tb = value_of(b); /* Div is not commutative. */ if (tarval_is_one(tb)) { /* div(x, 1) == x */ @@ -1513,76 +1211,13 @@ static ir_node *equivalent_node_Proj_Div(ir_node *proj) { } } return proj; -} /* equivalent_node_Proj_Div */ - -/** - * 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); - - /* 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 */ - -/** - * 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 proj; -} /* equivalent_node_Proj_DivMod */ +} /** * Optimize CopyB(mem, x, x) into a Nop. */ -static ir_node *equivalent_node_Proj_CopyB(ir_node *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); @@ -1591,24 +1226,20 @@ static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) { if (a == b) { /* Turn CopyB into a tuple (mem, jmp, bad, bad) */ switch (get_Proj_proj(proj)) { - case pn_CopyB_M_regular: + case pn_CopyB_M: proj = get_CopyB_mem(copyb); DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); break; - - case pn_CopyB_X_except: - DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(current_ir_graph); - break; } } return proj; -} /* equivalent_node_Proj_CopyB */ +} /** * Optimize Bounds(idx, idx, upper) into idx. */ -static ir_node *equivalent_node_Proj_Bound(ir_node *proj) { +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); @@ -1619,14 +1250,13 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) { 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. + * idx was Bounds checked 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)) { + get_Bound_upper(pred) == upper) { /* * One could expect that we simply return the previous * Bound here. However, this would be wrong, as we could @@ -1643,90 +1273,35 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) { 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); - } - } - } - } - 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); - } - } + break; } } return proj; -} /* equivalent_node_Proj_Store */ +} /** - * Does all optimizations on nodes that must be done on it's Proj's + * Does all optimizations on nodes that must be done on its Projs * because of creating new nodes. */ -static ir_node *equivalent_node_Proj(ir_node *proj) { +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 */ +} /** * Remove Id's. */ -static ir_node *equivalent_node_Id(ir_node *n) { +static ir_node *equivalent_node_Id(ir_node *n) +{ ir_node *oldn = n; do { @@ -1735,16 +1310,22 @@ static ir_node *equivalent_node_Id(ir_node *n) { DBG_OPT_ID(oldn, n); return n; -} /* equivalent_node_Id */ +} /** * Optimize a Mux. */ static ir_node *equivalent_node_Mux(ir_node *n) { - ir_node *oldn = n, *sel = get_Mux_sel(n); - ir_node *n_t, *n_f; - tarval *ts = value_of(sel); + ir_node *oldn = n, *sel = get_Mux_sel(n); + ir_node *n_t, *n_f; + ir_tarval *ts = value_of(sel); + + if (ts == tarval_bad && is_Cmp(sel)) { + /* try again with a direct call to compute_cmp, as we don't care + * about the MODEB_LOWERED flag here */ + ts = compute_cmp_ext(sel); + } /* Mux(true, f, t) == t */ if (ts == tarval_b_true) { @@ -1780,85 +1361,83 @@ static ir_node *equivalent_node_Mux(ir_node *n) DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ); return n; } - 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 *f = get_Mux_false(n); - ir_node *t = get_Mux_true(n); + if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) { + ir_relation relation = get_Cmp_relation(sel); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(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. */ - 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); - return n; - } - break; + ir_node *const cmp_l = get_Cmp_left(sel); + ir_node *const cmp_r = get_Cmp_right(sel); + + switch (relation) { + case ir_relation_equal: + 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); + 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; - } - break; + case ir_relation_less_greater: + case ir_relation_unordered_less_greater: + 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; } + break; + default: + 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_Lg || proj_nr == pn_Cmp_Ne) { - /* Mux(t != 0, -t, t) ==> t */ - n = t; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } + /* + * 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 (relation == ir_relation_equal) { + /* Mux(t == 0, -t, t) ==> -t */ + n = f; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) { + /* Mux(t != 0, -t, t) ==> t */ + n = t; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } } } + return n; -} /* equivalent_node_Mux */ +} /** * Remove Confirm nodes if setting is on. * Replace Confirms(x, '=', Constlike) by Constlike. */ -static ir_node *equivalent_node_Confirm(ir_node *n) { - ir_node *pred = get_Confirm_value(n); - pn_Cmp pnc = get_Confirm_cmp(n); +static ir_node *equivalent_node_Confirm(ir_node *n) +{ + ir_node *pred = get_Confirm_value(n); + ir_relation relation = get_Confirm_relation(n); - while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) { + while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) { /* * rare case: two identical Confirms one after another, * replace the second one with the first. */ n = pred; pred = get_Confirm_value(n); - pnc = get_Confirm_cmp(n); } - if (get_opt_remove_confirm()) - return get_Confirm_value(n); return n; } @@ -1869,78 +1448,19 @@ static ir_node *equivalent_node_Confirm(ir_node *n) { * If a node returns a Tuple we can not just skip it. If the size of the * in array fits, we transform n into a tuple (e.g., Div). */ -ir_node *equivalent_node(ir_node *n) { +ir_node *equivalent_node(ir_node *n) +{ if (n->op->ops.equivalent_node) return n->op->ops.equivalent_node(n); return n; -} /* equivalent_node */ - -/** - * Sets the default equivalent node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - 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(Eor); - CASE(Add); - CASE(Shl); - CASE(Shr); - CASE(Shrs); - CASE(Rotl); - CASE(Sub); - CASE(Not); - CASE(Minus); - CASE(Mul); - 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(Confirm); - default: - /* leave NULL */; - } - - return ops; -#undef CASE -#undef CASE_PROJ -} /* firm_set_default_equivalent_node */ +} /** * Returns non-zero if a node is a Phi node * with all predecessors constant. */ -static int is_const_Phi(ir_node *n) { +static int is_const_Phi(ir_node *n) +{ int i; if (! is_Phi(n) || get_irn_arity(n) == 0) @@ -1950,10 +1470,10 @@ static int is_const_Phi(ir_node *n) { return 0; } 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); +typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode); +typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b); /** * in reality eval_func should be tarval (*eval_func)() but incomplete @@ -1964,7 +1484,7 @@ typedef void (*eval_func)(void); /** * Wrapper for the tarval binop evaluation, tarval_sub has one more parameter. */ -static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode) +static ir_tarval *do_eval(eval_func eval, ir_tarval *a, ir_tarval *b, ir_mode *mode) { if (eval == (eval_func) tarval_sub) { tarval_sub_type func = (tarval_sub_type)eval; @@ -1988,46 +1508,43 @@ static tarval *do_eval(eval_func eval, tarval *a, tarval *b, ir_mode *mode) * * @return a new Phi node if the conversion was successful, NULL else */ -static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, ir_mode *mode, int left) { - tarval *tv; - void **res; - ir_node *pred; - ir_graph *irg; - int i, n = get_irn_arity(phi); - - NEW_ARR_A(void *, res, n); +static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eval, ir_mode *mode, int left) +{ + int n = get_irn_arity(phi); + ir_tarval **tvs = ALLOCAN(ir_tarval*, n); if (left) { - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - tv = get_Const_tarval(pred); - tv = do_eval(eval, other, tv, mode); + for (int i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); + ir_tarval *tv = get_Const_tarval(pred); + tv = do_eval(eval, other, tv, mode); if (tv == tarval_bad) { /* folding failed, bad */ return NULL; } - res[i] = tv; + tvs[i] = tv; } } else { - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - tv = get_Const_tarval(pred); - tv = do_eval(eval, tv, other, mode); + for (int i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); + ir_tarval *tv = get_Const_tarval(pred); + tv = do_eval(eval, tv, other, mode); if (tv == tarval_bad) { /* folding failed, bad */ return 0; } - res[i] = tv; + tvs[i] = tv; } } - irg = current_ir_graph; - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred)); + ir_graph *irg = get_irn_irg(phi); + ir_node **res = ALLOCAN(ir_node*, n); + for (int i = 0; i < n; ++i) { + res[i] = new_r_Const(irg, tvs[i]); } - return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_binop_on_phi */ + ir_node *block = get_nodes_block(phi); + return new_r_Phi(block, n, res, mode); +} /** * Apply an evaluator on a binop with two constant Phi. @@ -2039,39 +1556,34 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, eval_func eval, * * @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, eval_func eval, ir_mode *mode) { - tarval *tv_l, *tv_r, *tv; - void **res; - ir_node *pred; - ir_graph *irg; - int i, n; - +static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir_mode *mode) +{ 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); + int n = get_irn_arity(a); + ir_tarval **tvs = ALLOCAN(ir_tarval*, n); + for (int i = 0; i < n; ++i) { + ir_node *pred_a = get_irn_n(a, i); + ir_tarval *tv_l = get_Const_tarval(pred_a); + ir_node *pred_b = get_irn_n(b, i); + ir_tarval *tv_r = get_Const_tarval(pred_b); + ir_tarval *tv = do_eval(eval, tv_l, tv_r, mode); if (tv == tarval_bad) { /* folding failed, bad */ return NULL; } - res[i] = tv; + tvs[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, res[i], get_Const_type(pred)); + ir_graph *irg = get_irn_irg(a); + ir_node **res = ALLOCAN(ir_node*, n); + for (int i = 0; i < n; ++i) { + res[i] = new_r_Const(irg, tvs[i]); } - return new_r_Phi(get_nodes_block(a), n, (ir_node **)res, mode); -} /* apply_binop_on_2_phis */ + ir_node *block = get_nodes_block(a); + return new_r_Phi(block, n, res, mode); +} /** * Apply an evaluator on a unop with a constant operator (a Phi). @@ -2081,34 +1593,30 @@ static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir * * @return a new Phi node if the conversion was successful, NULL else */ -static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) { - tarval *tv; - void **res; - ir_node *pred; - ir_mode *mode; - ir_graph *irg; - int i, n = get_irn_arity(phi); - - NEW_ARR_A(void *, res, n); - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - tv = get_Const_tarval(pred); - tv = eval(tv); +static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *)) +{ + int n = get_irn_arity(phi); + ir_tarval **tvs = ALLOCAN(ir_tarval*, n); + for (int i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); + ir_tarval *tv = get_Const_tarval(pred); + tv = eval(tv); if (tv == tarval_bad) { /* folding failed, bad */ return 0; } - res[i] = tv; + tvs[i] = tv; } - mode = get_irn_mode(phi); - irg = current_ir_graph; - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred)); + ir_graph *irg = get_irn_irg(phi); + ir_node **res = ALLOCAN(ir_node*, n); + for (int i = 0; i < n; ++i) { + res[i] = new_r_Const(irg, tvs[i]); } - return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_unop_on_phi */ + ir_node *block = get_nodes_block(phi); + ir_mode *mode = get_irn_mode(phi); + return new_r_Phi(block, n, res, mode); +} /** * Apply a conversion on a constant operator (a Phi). @@ -2117,39 +1625,37 @@ static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) { * * @return a new Phi node if the conversion was successful, NULL else */ -static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) { - tarval *tv; - void **res; - ir_node *pred; - ir_graph *irg; - int i, n = get_irn_arity(phi); - - NEW_ARR_A(void *, res, n); - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - tv = get_Const_tarval(pred); - tv = tarval_convert_to(tv, mode); +static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) +{ + int n = get_irn_arity(phi); + ir_tarval **tvs = ALLOCAN(ir_tarval*, n); + for (int i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); + ir_tarval *tv = get_Const_tarval(pred); + tv = tarval_convert_to(tv, mode); if (tv == tarval_bad) { /* folding failed, bad */ return 0; } - res[i] = tv; + tvs[i] = tv; } - irg = current_ir_graph; - for (i = 0; i < n; ++i) { - pred = get_irn_n(phi, i); - res[i] = new_r_Const_type(irg, res[i], get_Const_type(pred)); + ir_graph *irg = get_irn_irg(phi); + ir_node **res = ALLOCAN(ir_node*, n); + for (int i = 0; i < n; ++i) { + res[i] = new_r_Const(irg, tvs[i]); } - return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_conv_on_phi */ + ir_node *block = get_nodes_block(phi); + return new_r_Phi(block, n, res, mode); +} /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)). * If possible, remove the Conv's. */ -static ir_node *transform_node_AddSub(ir_node *n) { +static ir_node *transform_node_AddSub(ir_node *n) +{ ir_mode *mode = get_irn_mode(n); if (mode_is_reference(mode)) { @@ -2218,9 +1724,10 @@ static ir_node *transform_node_AddSub(ir_node *n) { } return n; -} /* transform_node_AddSub */ +} #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \ + do { \ c = NULL; \ if (is_Const(b) && is_const_Phi(a)) { \ /* check for Op(Phi, Const) */ \ @@ -2237,9 +1744,11 @@ static ir_node *transform_node_AddSub(ir_node *n) { if (c) { \ DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ return c; \ - } + } \ + } while(0) #define HANDLE_UNOP_PHI(eval, a, c) \ + do { \ c = NULL; \ if (is_const_Phi(a)) { \ /* check for Op(Phi) */ \ @@ -2248,47 +1757,793 @@ static ir_node *transform_node_AddSub(ir_node *n) { DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ return c; \ } \ - } + } \ + } while(0) /** - * Do the AddSub optimization, then Transform - * Constant folding on Phi - * Add(a,a) -> Mul(a, 2) - * Add(Mul(a, x), a) -> Mul(a, x+1) - * if the mode is integer or float. - * Transform Add(a,-b) into Sub(a,b). - * Reassociation might fold this further. + * Create a 0 constant of given mode. */ -static ir_node *transform_node_Add(ir_node *n) { - ir_mode *mode; - ir_node *a, *b, *c, *oldn = n; - - n = transform_node_AddSub(n); +static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode) +{ + ir_tarval *tv = get_mode_null(mode); + ir_node *cnst = new_r_Const(irg, tv); - a = get_Add_left(n); - b = get_Add_right(n); + return cnst; +} - mode = get_irn_mode(n); +static bool is_shiftop(const ir_node *n) +{ + return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n); +} - if (mode_is_reference(mode)) { - ir_mode *lmode = get_irn_mode(a); +/* 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 bool 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)); +} + +static bool is_cmp_unequal(const ir_node *node) +{ + ir_relation relation = get_Cmp_relation(node); + ir_node *left = get_Cmp_left(node); + ir_node *right = get_Cmp_right(node); + ir_mode *mode = get_irn_mode(left); + + if (relation == ir_relation_less_greater) + return true; + + if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right)) + return relation == ir_relation_greater; + return false; +} + +/** + * returns true for Cmp(x == 0) or Cmp(x != 0) + */ +static bool is_cmp_equality_zero(const ir_node *node) +{ + ir_relation relation; + ir_node *right = get_Cmp_right(node); + + if (!is_Const(right) || !is_Const_null(right)) + return false; + relation = get_Cmp_relation(node); + return relation == ir_relation_equal + || relation == ir_relation_less_greater + || (!mode_is_signed(get_irn_mode(right)) + && relation == ir_relation_greater); +} + +/** + * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. + * Such pattern may arise in bitfield stores. + * + * value c4 value c4 & c2 + * AND c3 AND c1 | c3 + * OR c2 ===> OR + * AND c1 + * OR + * + * + * value c2 value c1 + * AND c1 ===> OR if (c1 | c2) == 0x111..11 + * OR + */ +static ir_node *transform_node_Or_bf_store(ir_node *irn_or) +{ + ir_node *irn_and, *c1; + ir_node *or_l, *c2; + ir_node *and_l, *c3; + ir_node *value, *c4; + ir_node *new_and, *new_const, *block; + ir_mode *mode = get_irn_mode(irn_or); + + ir_tarval *tv1, *tv2, *tv3, *tv4, *tv; + + for (;;) { + ir_graph *irg; + irn_and = get_binop_left(irn_or); + c1 = get_binop_right(irn_or); + if (!is_Const(c1) || !is_And(irn_and)) + return irn_or; + + or_l = get_binop_left(irn_and); + c2 = get_binop_right(irn_and); + if (!is_Const(c2)) + return irn_or; + + tv1 = get_Const_tarval(c1); + tv2 = get_Const_tarval(c2); + + tv = tarval_or(tv1, tv2); + if (tarval_is_all_one(tv)) { + /* the AND does NOT clear a bit with isn't set by the OR */ + set_binop_left(irn_or, or_l); + set_binop_right(irn_or, c1); + + /* check for more */ + continue; + } + + if (!is_Or(or_l) && !is_Or_Eor_Add(or_l)) + return irn_or; + + and_l = get_binop_left(or_l); + c3 = get_binop_right(or_l); + if (!is_Const(c3) || !is_And(and_l)) + return irn_or; + + value = get_binop_left(and_l); + c4 = get_binop_right(and_l); + if (!is_Const(c4)) + return irn_or; + + /* ok, found the pattern, check for conditions */ + assert(mode == get_irn_mode(irn_and)); + assert(mode == get_irn_mode(or_l)); + assert(mode == get_irn_mode(and_l)); + + tv3 = get_Const_tarval(c3); + tv4 = get_Const_tarval(c4); + + tv = tarval_or(tv4, tv2); + if (!tarval_is_all_one(tv)) { + /* have at least one 0 at the same bit position */ + return irn_or; + } + + if (tv3 != tarval_andnot(tv3, tv4)) { + /* bit in the or_mask is outside the and_mask */ + return irn_or; + } + + if (tv1 != tarval_andnot(tv1, tv2)) { + /* bit in the or_mask is outside the and_mask */ + return irn_or; + } + + /* ok, all conditions met */ + block = get_nodes_block(irn_or); + irg = get_irn_irg(block); + + new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode); + + new_const = new_r_Const(irg, tarval_or(tv3, tv1)); + + set_binop_left(irn_or, new_and); + set_binop_right(irn_or, new_const); + + /* check for more */ + } +} + +/** + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl + */ +static ir_node *transform_node_Or_Rotl(ir_node *irn_or) +{ + ir_mode *mode = get_irn_mode(irn_or); + ir_node *shl, *shr, *block; + ir_node *irn, *x, *c1, *c2, *n; + ir_tarval *tv1, *tv2; + + /* some backends can't handle rotl */ + if (!be_get_backend_param()->support_rotl) + return irn_or; + + if (! mode_is_int(mode)) + return irn_or; + + shl = get_binop_left(irn_or); + shr = get_binop_right(irn_or); + + if (is_Shr(shl)) { + if (!is_Shl(shr)) + return irn_or; + + irn = shl; + shl = shr; + shr = irn; + } else if (!is_Shl(shl)) { + return irn_or; + } else if (!is_Shr(shr)) { + return irn_or; + } + x = get_Shl_left(shl); + if (x != get_Shr_left(shr)) + return irn_or; + + c1 = get_Shl_right(shl); + c2 = get_Shr_right(shr); + if (is_Const(c1) && is_Const(c2)) { + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return irn_or; + + tv2 = get_Const_tarval(c2); + if (! tarval_is_long(tv2)) + return irn_or; + + if (get_tarval_long(tv1) + get_tarval_long(tv2) + != (int) get_mode_size_bits(mode)) + return irn_or; + + /* yet, condition met */ + block = get_nodes_block(irn_or); + + n = new_r_Rotl(block, x, c1, mode); + + DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); + return n; + } + + /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets + * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */ + if (!ir_is_negated_value(c1, c2)) { + return irn_or; + } + + /* yet, condition met */ + block = get_nodes_block(irn_or); + n = new_r_Rotl(block, x, c1, mode); + DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL); + return 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 + * and(a OP c, b OP c) -> and(a, b) OP c + * note, might return a different op than n + */ +static ir_node *transform_bitwise_distributive(ir_node *n, + recursive_transform trans_func) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_op *op = get_irn_op(a); + ir_op *op_root = get_irn_op(n); + + if (op != get_irn_op(b)) + return n; + + /* and(conv(a), conv(b)) -> conv(and(a,b)) */ + if (op == op_Conv) { + ir_node *a_op = get_Conv_op(a); + ir_node *b_op = get_Conv_op(b); + 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_nodes_block(n); + + n = exact_copy(n); + set_binop_left(n, a_op); + set_binop_right(n, b_op); + set_irn_mode(n, a_mode); + n = trans_func(n); + n = new_r_Conv(blk, n, get_irn_mode(oldn)); + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + return n; + } + } + + if (op == op_Eor) { + /* nothing to gain here */ + return n; + } + + if (op == op_Shrs || op == op_Shr || op == op_Shl + || op == op_And || op == op_Or || op == op_Eor) { + ir_node *a_left = get_binop_left(a); + ir_node *a_right = get_binop_right(a); + ir_node *b_left = get_binop_left(b); + ir_node *b_right = get_binop_right(b); + ir_node *c = NULL; + ir_node *op1 = NULL; + ir_node *op2 = NULL; + + if (is_op_commutative(op)) { + if (a_left == b_left) { + c = a_left; + op1 = a_right; + op2 = b_right; + } else if (a_left == b_right) { + c = a_left; + op1 = a_right; + op2 = b_left; + } else if (a_right == b_left) { + c = a_right; + op1 = a_left; + op2 = b_right; + } + } + if (a_right == b_right) { + c = a_right; + op1 = a_left; + op2 = b_left; + } + + if (c != NULL) { + /* (a sop c) & (b sop c) => (a & b) sop c */ + ir_node *blk = get_nodes_block(n); + + ir_node *new_n = exact_copy(n); + set_binop_left(new_n, op1); + set_binop_right(new_n, op2); + new_n = trans_func(new_n); + + if (op_root == op_Eor && op == op_Or) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *mode = get_irn_mode(c); + + c = new_rd_Not(dbgi, blk, c, mode); + n = new_rd_And(dbgi, blk, new_n, c, mode); + } else { + n = exact_copy(a); + set_nodes_block(n, blk); + set_binop_left(n, new_n); + set_binop_right(n, c); + add_identities(n); + } + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); + return n; + } + } + + return n; +} + +/** + * normalisation: (x >> c1) & c2 to (x & (c2<> c1 + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is usually good for the backend since << C can often be + * matched as address-mode. + */ +static ir_node *transform_node_bitop_shift(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *shift_left; + ir_node *shift_right; + ir_node *block; + dbg_info *dbg_bitop; + dbg_info *dbg_shift; + ir_node *new_bitop; + ir_node *new_shift; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_bitop; + + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2)) + return n; + + assert(is_And(n) || is_Or(n) || is_Eor(n) || is_Or_Eor_Add(n)); + if (!is_Const(right) || !is_shiftop(left)) + return n; + + shift_left = get_binop_left(left); + shift_right = get_binop_right(left); + if (!is_Const(shift_right)) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(left)) { + /* TODO this could be improved */ + return n; + } + + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbg_bitop = get_irn_dbg_info(n); + dbg_shift = get_irn_dbg_info(left); + tv1 = get_Const_tarval(shift_right); + tv2 = get_Const_tarval(right); + assert(get_tarval_mode(tv2) == mode); + + if (is_Shl(left)) { + tv_bitop = tarval_shr(tv2, tv1); + + /* Check whether we have lost some bits during the right shift. */ + if (!is_And(n)) { + ir_tarval *tv_back_again = tarval_shl(tv_bitop, tv1); + + if (tarval_cmp(tv_back_again, tv2) != ir_relation_equal) + return n; + } + } else if (is_Shr(left)) { + if (!is_And(n)) { + /* + * TODO this can be improved by checking whether + * the left shift produces an overflow + */ + return n; + } + tv_bitop = tarval_shl(tv2, tv1); + } else { + assert(is_Rotl(left)); + tv_bitop = tarval_rotl(tv2, tarval_neg(tv1)); + } + new_const = new_r_Const(irg, tv_bitop); + + if (is_And(n)) { + new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode); + } else if (is_Or(n) || is_Or_Eor_Add(n)) { + new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode); + } else { + assert(is_Eor(n)); + new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode); + } + + if (is_Shl(left)) { + new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode); + } else if (is_Shr(left)) { + new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode); + } else { + assert(is_Rotl(left)); + new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode); + } + + return new_shift; +} + +static bool complement_values(const ir_node *a, const ir_node *b) +{ + if (is_Not(a) && get_Not_op(a) == b) + return true; + if (is_Not(b) && get_Not_op(b) == a) + return true; + if (is_Const(a) && is_Const(b)) { + ir_tarval *tv_a = get_Const_tarval(a); + ir_tarval *tv_b = get_Const_tarval(b); + return tarval_not(tv_a) == tv_b; + } + return false; +} + +typedef ir_tarval *(tv_fold_binop_func)(ir_tarval *a, ir_tarval *b); + +/** + * for associative operations fold: + * op(op(x, c0), c1) to op(x, op(c0, c1)) with constants folded. + * This is a "light" version of the reassociation phase + */ +static ir_node *fold_constant_associativity(ir_node *node, + tv_fold_binop_func fold) +{ + ir_graph *irg; + ir_op *op; + ir_node *left; + ir_node *right = get_binop_right(node); + ir_node *left_right; + ir_node *left_left; + ir_tarval *c0; + ir_tarval *c1; + ir_tarval *new_c; + ir_node *new_const; + ir_node *new_node; + if (!is_Const(right)) + return node; + + op = get_irn_op(node); + left = get_binop_left(node); + if (get_irn_op(left) != op) + return node; + + left_right = get_binop_right(left); + if (!is_Const(left_right)) + return node; + + left_left = get_binop_left(left); + c0 = get_Const_tarval(left_right); + c1 = get_Const_tarval(right); + irg = get_irn_irg(node); + if (get_tarval_mode(c0) != get_tarval_mode(c1)) + return node; + new_c = fold(c0, c1); + if (new_c == tarval_bad) + return node; + new_const = new_r_Const(irg, new_c); + new_node = exact_copy(node); + set_binop_left(new_node, left_left); + set_binop_right(new_node, new_const); + return new_node; +} + +/** + * Transform an Or. + */ +static ir_node *transform_node_Or_(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_node *c; + ir_mode *mode; + + n = fold_constant_associativity(n, tarval_or); + if (n != oldn) + return n; + + if (is_Not(a) && is_Not(b)) { + /* ~a | ~b = ~(a&b) */ + ir_node *block = get_nodes_block(n); + + mode = get_irn_mode(n); + a = get_Not_op(a); + b = get_Not_op(b); + n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode); + n = new_rd_Not(get_irn_dbg_info(n), block, n, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); + return n; + } + + /* we can combine the relations of two compares with the same operands */ + if (is_Cmp(a) && is_Cmp(b)) { + ir_node *a_left = get_Cmp_left(a); + ir_node *a_right = get_Cmp_right(a); + ir_node *b_left = get_Cmp_left(b); + ir_node *b_right = get_Cmp_right(b); + if (a_left == b_left && b_left == b_right) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + ir_relation new_relation = a_relation | b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */ + if (is_cmp_unequal(a) && is_cmp_unequal(b) + && !mode_is_float(get_irn_mode(a_left)) + && !mode_is_float(get_irn_mode(b_left))) { + if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + ir_graph *irg = get_irn_irg(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode); + ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_node *zero = create_zero_const(irg, b_mode); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater); + } + if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) { + ir_graph *irg = get_irn_irg(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_node *zero = create_zero_const(irg, a_mode); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_less_greater); + } + } + } + + mode = get_irn_mode(n); + HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode); + + n = transform_node_Or_bf_store(n); + if (n != oldn) + return n; + n = transform_node_Or_Rotl(n); + if (n != oldn) + return n; + + n = transform_bitwise_distributive(n, transform_node_Or_); + if (n != oldn) + return n; + n = transform_node_bitop_shift(n); + if (n != oldn) + return n; + + return n; +} + +static ir_node *transform_node_Or(ir_node *n) +{ + if (is_Or_Eor_Add(n)) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *left = get_Or_left(n); + ir_node *right = get_Or_right(n); + ir_mode *mode = get_irn_mode(n); + return new_rd_Add(dbgi, block, left, right, mode); + } + return transform_node_Or_(n); +} + +/** + * Transform an Eor. + */ +static ir_node *transform_node_Eor_(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *c; + + n = fold_constant_associativity(n, tarval_eor); + if (n != oldn) + return n; + + /* we can combine the relations of two compares with the same operands */ + if (is_Cmp(a) && is_Cmp(b)) { + ir_node *a_left = get_Cmp_left(a); + ir_node *a_right = get_Cmp_left(a); + ir_node *b_left = get_Cmp_left(b); + ir_node *b_right = get_Cmp_right(b); + if (a_left == b_left && b_left == b_right) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + ir_relation new_relation = a_relation ^ b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + } + + HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode); + + /* normalize not nodes... ~a ^ b <=> a ^ ~b */ + if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *new_not = new_rd_Not(dbg, block, b, mode); + ir_node *new_left = get_Not_op(a); + n = new_rd_Eor(dbg, block, new_left, new_not, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *new_not = new_rd_Not(dbg, block, a, mode); + ir_node *new_right = get_Not_op(b); + n = new_rd_Eor(dbg, block, new_not, new_right, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + + /* x ^ 1...1 -> ~1 */ + if (is_Const(b) && is_Const_all_one(b)) { + n = new_r_Not(get_nodes_block(n), a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + + n = transform_bitwise_distributive(n, transform_node_Eor_); + if (n != oldn) + return n; + n = transform_node_bitop_shift(n); + if (n != oldn) + return n; + + return n; +} + +static ir_node *transform_node_Eor(ir_node *n) +{ + if (is_Or_Eor_Add(n)) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *left = get_Eor_left(n); + ir_node *right = get_Eor_right(n); + ir_mode *mode = get_irn_mode(n); + return new_rd_Add(dbgi, block, left, right, mode); + } + return transform_node_Eor_(n); +} + +/** + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Add(a,a) -> Mul(a, 2) + * Add(Mul(a, x), a) -> Mul(a, x+1) + * if the mode is integer or float. + * Transform Add(a,-b) into Sub(a,b). + * Reassociation might fold this further. + */ +static ir_node *transform_node_Add(ir_node *n) +{ + ir_mode *mode; + ir_node *a; + ir_node *b; + ir_node *c; + ir_node *oldn = n; + + n = fold_constant_associativity(n, tarval_add); + if (n != oldn) + return n; + + n = transform_node_AddSub(n); + if (n != oldn) + return n; + + a = get_Add_left(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)) { + 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, get_nodes_block(n), a, mode); } } + if (is_Const(b) && get_mode_arithmetic(mode) == irma_twos_complement) { + ir_tarval *tv = get_Const_tarval(b); + ir_tarval *min = get_mode_min(mode); + /* if all bits are set, then this has the same effect as a Not. + * Note that the following == gives false for different modes which + * is exactly what we want */ + if (tv == min) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + ir_node *block = get_nodes_block(n); + ir_node *cnst = new_r_Const(irg, min); + return new_rd_Eor(dbgi, block, a, cnst, mode); + } + } + HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode); - /* 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; + /* for FP the following optimizations are only allowed if + * fp_strict_algebraic is disabled */ + if (mode_is_float(mode)) { + ir_graph *irg = get_irn_irg(n); + if (get_irg_fp_model(irg) & fp_strict_algebraic) + return n; + } if (mode_is_num(mode)) { - /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */ - if (!is_irg_state(current_ir_graph, IR_GRAPH_STATE_ARCH_DEP) + ir_graph *irg = get_irn_irg(n); + /* the following code leads to endless recursion when Mul are replaced + * by a simple instruction chain */ + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_ARCH_DEP) && a == b && mode_is_int(mode)) { ir_node *block = get_nodes_block(n); @@ -2296,7 +2551,7 @@ static ir_node *transform_node_Add(ir_node *n) { get_irn_dbg_info(n), block, a, - new_Const_long(mode, 2), + new_r_Const_long(irg, mode, 2), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); return n; @@ -2333,35 +2588,30 @@ static ir_node *transform_node_Add(ir_node *n) { DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); return n; } - if (op == b) { - /* ~x + x = -1 */ - n = new_Const(get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } - } - if (is_Not(b)) { - ir_node *op = get_Not_op(b); - - if (op == a) { - /* x + ~x = -1 */ - n = new_Const(get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } } } } + + if (is_Or_Eor_Add(n)) { + n = transform_node_Or_(n); + if (n != oldn) + return n; + n = transform_node_Eor_(n); + if (n != oldn) + return n; + } + return n; -} /* transform_node_Add */ +} /** * returns -cnst or NULL if impossible */ -static ir_node *const_negate(ir_node *cnst) { - tarval *tv = tarval_neg(get_Const_tarval(cnst)); - dbg_info *dbgi = get_irn_dbg_info(cnst); - ir_graph *irg = get_irn_irg(cnst); +static ir_node *const_negate(ir_node *cnst) +{ + ir_tarval *tv = tarval_neg(get_Const_tarval(cnst)); + dbg_info *dbgi = get_irn_dbg_info(cnst); + ir_graph *irg = get_irn_irg(cnst); if (tv == tarval_bad) return NULL; return new_rd_Const(dbgi, irg, tv); } @@ -2376,7 +2626,8 @@ static ir_node *const_negate(ir_node *cnst) { * Sub(x, Add(x, a)) -> -a * Sub(x, Const) -> Add(x, -Const) */ -static ir_node *transform_node_Sub(ir_node *n) { +static ir_node *transform_node_Sub(ir_node *n) +{ ir_mode *mode; ir_node *oldn = n; ir_node *a, *b, *c; @@ -2415,8 +2666,11 @@ restart: HANDLE_BINOP_PHI((eval_func) tarval_sub, a, b, c, mode); /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ - if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) - return n; + if (mode_is_float(mode)) { + ir_graph *irg = get_irn_irg(n); + if (get_irg_fp_model(irg) & fp_strict_algebraic) + return n; + } if (is_Const(b) && !mode_is_reference(get_irn_mode(b))) { /* a - C -> a + (-C) */ @@ -2452,19 +2706,19 @@ restart: /* a - (b - c) -> a + (c - b) * -> (a - b) + c iff (b - c) is a pointer */ dbg_info *s_dbg = get_irn_dbg_info(b); - ir_node *s_block = get_nodes_block(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); if (mode_is_reference(s_mode)) { - ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode); + ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */ + ir_node *sub = new_rd_Sub(s_dbg, lowest_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(a_block, s_right, mode); - n = new_rd_Add(a_dbg, a_block, sub, s_right, mode); + s_right = new_r_Conv(lowest_block, s_right, mode); + n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode); } else { + ir_node *s_block = get_nodes_block(b); ir_node *sub = new_rd_Sub(s_dbg, 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); @@ -2473,6 +2727,7 @@ restart: } DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; +#if 0 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ ir_node *m_right = get_Mul_right(b); if (is_Const(m_right)) { @@ -2491,6 +2746,7 @@ restart: return n; } } +#endif } /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ @@ -2503,58 +2759,54 @@ restart: DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); return n; } - if (is_Add(a)) { - if (mode_wrap_around(mode)) { - ir_node *left = get_Add_left(a); - ir_node *right = get_Add_right(a); + if ((is_Add(a) || is_Or_Eor_Add(a)) && mode_wrap_around(mode)) { + ir_node *left = get_binop_left(a); + ir_node *right = get_binop_right(a); - /* FIXME: Does the Conv's work only for two complement or generally? */ - if (left == b) { - if (mode != get_irn_mode(right)) { - /* This Sub is an effective Cast */ - right = new_r_Conv(get_nodes_block(n), right, mode); - } - n = right; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; - } else if (right == b) { - if (mode != get_irn_mode(left)) { - /* This Sub is an effective Cast */ - left = new_r_Conv(get_nodes_block(n), left, mode); - } - n = left; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == b) { + if (mode != get_irn_mode(right)) { + /* This Sub is an effective Cast */ + right = new_r_Conv(get_nodes_block(n), right, mode); + } + n = right; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } else if (right == b) { + if (mode != get_irn_mode(left)) { + /* This Sub is an effective Cast */ + left = new_r_Conv(get_nodes_block(n), left, mode); } + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; } } - if (is_Add(b)) { - if (mode_wrap_around(mode)) { - ir_node *left = get_Add_left(b); - ir_node *right = get_Add_right(b); + if ((is_Add(b) || is_Or_Eor_Add(b)) && mode_wrap_around(mode)) { + ir_node *left = get_binop_left(b); + ir_node *right = get_binop_right(b); - /* FIXME: Does the Conv's work only for two complement or generally? */ - if (left == a) { - ir_mode *r_mode = get_irn_mode(right); + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == a) { + ir_mode *r_mode = get_irn_mode(right); - n = new_r_Minus(get_nodes_block(n), right, r_mode); - if (mode != r_mode) { - /* This Sub is an effective Cast */ - n = new_r_Conv(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_nodes_block(n), right, r_mode); + if (mode != r_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(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_nodes_block(n), left, l_mode); - if (mode != l_mode) { - /* This Sub is an effective Cast */ - n = new_r_Conv(get_nodes_block(n), n, mode); - } - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; + n = new_r_Minus(get_nodes_block(n), left, l_mode); + if (mode != l_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(get_nodes_block(n), n, mode); } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; } } if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) { @@ -2585,7 +2837,8 @@ restart: ir_node *mb = get_Mul_right(a); if (ma == b) { - ir_node *blk = get_nodes_block(n); + ir_node *blk = get_nodes_block(n); + ir_graph *irg = get_irn_irg(n); n = new_rd_Mul( get_irn_dbg_info(n), blk, @@ -2594,13 +2847,14 @@ restart: get_irn_dbg_info(n), blk, mb, - new_Const_long(mode, 1), + new_r_Const(irg, get_mode_one(mode)), mode), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); return n; } else if (mb == b) { - ir_node *blk = get_nodes_block(n); + ir_node *blk = get_nodes_block(n); + ir_graph *irg = get_irn_irg(n); n = new_rd_Mul( get_irn_dbg_info(n), blk, @@ -2609,7 +2863,7 @@ restart: get_irn_dbg_info(n), blk, ma, - new_Const_long(mode, 1), + new_r_Const(irg, get_mode_one(mode)), mode), mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A); @@ -2650,34 +2904,54 @@ restart: } if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* c - ~X = X + (c+1) */ if (is_Const(a) && is_Not(b)) { - /* c - ~X = X + (c+1) */ - tarval *tv = get_Const_tarval(a); + ir_tarval *tv = get_Const_tarval(a); tv = tarval_add(tv, get_mode_one(mode)); if (tv != tarval_bad) { - ir_node *blk = get_nodes_block(n); - ir_node *c = new_Const(tv); + ir_node *blk = get_nodes_block(n); + ir_graph *irg = get_irn_irg(n); + ir_node *c = new_r_Const(irg, tv); n = new_rd_Add(get_irn_dbg_info(n), blk, get_Not_op(b), c, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_C_NOT_X); return n; } } + /* x-(x&y) = x & ~y */ + if (is_And(b)) { + ir_node *and_left = get_And_left(b); + ir_node *and_right = get_And_right(b); + if (and_right == a) { + ir_node *tmp = and_left; + and_left = and_right; + and_right = tmp; + } + if (and_left == a) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + ir_node *notn = new_rd_Not(dbgi, block, and_right, mode); + ir_node *andn = new_rd_And(dbgi, block, a, notn, mode); + return andn; + } + } } return n; -} /* transform_node_Sub */ +} /** * Several transformation done on n*n=2n bits mul. * These transformations must be done here because new nodes may be produced. */ -static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { - ir_node *oldn = n; - ir_node *a = get_Mul_left(n); - ir_node *b = get_Mul_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); - ir_mode *smode = get_irn_mode(a); +static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) +{ + ir_node *oldn = n; + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + ir_tarval *ta = value_of(a); + ir_tarval *tb = value_of(b); + ir_mode *smode = get_irn_mode(a); if (ta == get_mode_one(smode)) { /* (L)1 * (L)b = (L)b */ @@ -2696,7 +2970,7 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { } if (tb == get_mode_one(smode)) { /* (L)a * (L)1 = (L)a */ - ir_node *blk = get_irn_n(a, -1); + ir_node *blk = get_nodes_block(a); n = new_rd_Conv(get_irn_dbg_info(n), blk, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; @@ -2717,13 +2991,15 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { * Do constant evaluation of Phi nodes. * Do architecture dependent optimizations on Mul nodes */ -static ir_node *transform_node_Mul(ir_node *n) { +static ir_node *transform_node_Mul(ir_node *n) +{ ir_node *c, *oldn = n; ir_mode *mode = get_irn_mode(n); ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); - if (is_Bad(a) || is_Bad(b)) + n = fold_constant_associativity(n, tarval_mul); + if (n != oldn) return n; if (mode != get_irn_mode(a)) @@ -2804,10 +3080,11 @@ static ir_node *transform_node_Mul(ir_node *n) { return n; } } - if (get_mode_arithmetic(mode) == irma_ieee754) { + if (get_mode_arithmetic(mode) == irma_ieee754 + || get_mode_arithmetic(mode) == irma_x86_extended_float) { if (is_Const(a)) { - tarval *tv = get_Const_tarval(a); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + ir_tarval *tv = get_Const_tarval(a); + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* 2.0 * b = b + b */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode); @@ -2816,8 +3093,8 @@ static ir_node *transform_node_Mul(ir_node *n) { } } else if (is_Const(b)) { - tarval *tv = get_Const_tarval(b); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + ir_tarval *tv = get_Const_tarval(b); + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* a * 2.0 = a + a */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode); @@ -2827,93 +3104,131 @@ static ir_node *transform_node_Mul(ir_node *n) { } } return arch_dep_replace_mul_with_shifts(n); -} /* transform_node_Mul */ +} /** * Transform a Div Node. */ -static ir_node *transform_node_Div(ir_node *n) { +static ir_node *transform_node_Div(ir_node *n) +{ ir_mode *mode = get_Div_resmode(n); ir_node *a = get_Div_left(n); ir_node *b = get_Div_right(n); - ir_node *value; + ir_node *value = n; const ir_node *dummy; - if (is_Const(b) && is_const_Phi(a)) { - /* check for Div(Phi, Const) */ - value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) 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), (eval_func) tarval_div, mode, 1); - if (value) { - DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); - goto make_tuple; + if (mode_is_int(mode)) { + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) 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), (eval_func) 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, (eval_func) tarval_div, mode); + 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, (eval_func) tarval_div, mode); - if (value) { - DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + + if (a == b && value_not_zero(a, &dummy)) { + ir_graph *irg = get_irn_irg(n); + /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ + value = new_r_Const(irg, get_mode_one(mode)); + DBG_OPT_CSTEVAL(n, value); goto make_tuple; + } else { + if (mode_is_signed(mode) && is_Const(b)) { + ir_tarval *tv = get_Const_tarval(b); + + if (tv == get_mode_minus_one(mode)) { + /* a / -1 */ + value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } + } + /* Try architecture dependent optimization */ + value = arch_dep_replace_div_by_const(n); } - } + } else { + assert(mode_is_float(mode)); - value = n; + /* Optimize x/c to x*(1/c) */ + if (get_mode_arithmetic(mode) == irma_ieee754) { + ir_tarval *tv = value_of(b); - if (a == b && value_not_zero(a, &dummy)) { - /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ - value = new_Const(get_mode_one(mode)); - DBG_OPT_CSTEVAL(n, value); - goto make_tuple; - } else { - if (mode_is_signed(mode) && is_Const(b)) { - tarval *tv = get_Const_tarval(b); + if (tv != tarval_bad) { + int rem = tarval_fp_ops_enabled(); - if (tv == get_mode_minus_one(mode)) { - /* a / -1 */ - value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); - DBG_OPT_CSTEVAL(n, value); - goto make_tuple; + /* + * 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. + */ + tarval_enable_fp_ops(1); + tv = tarval_div(get_mode_one(mode), tv); + 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(get_irn_irg(n)) & fp_strict_algebraic) == 0)) { + ir_node *block = get_nodes_block(n); + ir_graph *irg = get_irn_irg(block); + ir_node *c = new_r_Const(irg, tv); + dbg_info *dbgi = get_irn_dbg_info(n); + value = new_rd_Mul(dbgi, block, a, c, mode); + + goto make_tuple; + } } } - /* Try architecture dependent optimization */ - value = arch_dep_replace_div_by_const(n); } if (value != n) { ir_node *mem, *blk; + ir_graph *irg; make_tuple: /* Turn Div into a tuple (mem, jmp, bad, value) */ mem = get_Div_mem(n); blk = get_nodes_block(n); + irg = get_irn_irg(blk); /* skip a potential Pin */ mem = skip_Pin(mem); - turn_into_tuple(n, pn_Div_max); + turn_into_tuple(n, pn_Div_max+1); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk)); - set_Tuple_pred(n, pn_Div_X_except, new_Bad()); + set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X)); set_Tuple_pred(n, pn_Div_res, value); } return n; -} /* transform_node_Div */ +} /** * Transform a Mod node. */ -static ir_node *transform_node_Mod(ir_node *n) { - ir_mode *mode = get_Mod_resmode(n); - ir_node *a = get_Mod_left(n); - ir_node *b = get_Mod_right(n); - ir_node *value; - tarval *tv; +static ir_node *transform_node_Mod(ir_node *n) +{ + ir_mode *mode = get_Mod_resmode(n); + ir_node *a = get_Mod_left(n); + ir_node *b = get_Mod_right(n); + ir_graph *irg; + ir_node *value; + ir_tarval *tv; if (is_Const(b) && is_const_Phi(a)) { /* check for Div(Phi, Const) */ @@ -2942,8 +3257,9 @@ static ir_node *transform_node_Mod(ir_node *n) { value = n; tv = value_of(n); + irg = get_irn_irg(n); if (tv != tarval_bad) { - value = new_Const(tv); + value = new_r_Const(irg, tv); DBG_OPT_CSTEVAL(n, value); goto make_tuple; @@ -2954,16 +3270,16 @@ static ir_node *transform_node_Mod(ir_node *n) { if (a == b && value_not_zero(a, &dummy)) { /* BEWARE: we can optimize a%a to 0 only if this cannot cause a exception */ - value = new_Const(get_mode_null(mode)); + value = new_r_Const(irg, get_mode_null(mode)); DBG_OPT_CSTEVAL(n, value); goto make_tuple; } else { if (mode_is_signed(mode) && is_Const(b)) { - tarval *tv = get_Const_tarval(b); + ir_tarval *tv = get_Const_tarval(b); if (tv == get_mode_minus_one(mode)) { /* a % -1 = 0 */ - value = new_Const(get_mode_null(mode)); + value = new_r_Const(irg, get_mode_null(mode)); DBG_OPT_CSTEVAL(n, value); goto make_tuple; } @@ -2975,255 +3291,24 @@ static ir_node *transform_node_Mod(ir_node *n) { if (value != n) { ir_node *mem, *blk; + ir_graph *irg; make_tuple: /* Turn Mod into a tuple (mem, jmp, bad, value) */ mem = get_Mod_mem(n); blk = get_nodes_block(n); + irg = get_irn_irg(blk); /* skip a potential Pin */ mem = skip_Pin(mem); - turn_into_tuple(n, pn_Mod_max); + turn_into_tuple(n, pn_Mod_max+1); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk)); - set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); + set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X)); set_Tuple_pred(n, pn_Mod_res, value); } return n; -} /* transform_node_Mod */ - -/** - * Transform a DivMod node. - */ -static ir_node *transform_node_DivMod(ir_node *n) { - 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), (eval_func) tarval_div, mode, 0); - vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) 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), (eval_func) tarval_div, mode, 1); - vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) 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, (eval_func) tarval_div, mode); - vb = apply_binop_on_2_phis(a, b, (eval_func) 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))) { - va = a; - vb = new_Const(get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else if (ta != tarval_bad) { - tarval *resa, *resb; - resa = tarval_div(ta, tb); - if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through - Jmp for X result!? */ - resb = tarval_mod(ta, tb); - if (resb == tarval_bad) return n; /* Causes exception! */ - va = new_Const(resa); - vb = new_Const(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)) { - va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); - vb = new_Const(get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, va); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else { /* Try architecture dependent optimization */ - 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 */ - va = new_Const(get_mode_one(mode)); - vb = new_Const(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 */ - return n; - } - } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) { - /* 0 / non-Const = 0 */ - vb = va = a; - goto make_tuple; - } - - if (evaluated) { /* replace by tuple */ - ir_node *mem, *blk; - -make_tuple: - mem = get_DivMod_mem(n); - /* skip a potential Pin */ - mem = skip_Pin(mem); - - 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(blk)); - set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ - set_Tuple_pred(n, pn_DivMod_res_div, va); - set_Tuple_pred(n, pn_DivMod_res_mod, vb); - } - - return n; -} /* transform_node_DivMod */ - -/** - * Optimize x / c to x * (1/c) - */ -static ir_node *transform_node_Quot(ir_node *n) { - ir_mode *mode = get_Quot_resmode(n); - ir_node *oldn = n; - - if (get_mode_arithmetic(mode) == irma_ieee754) { - ir_node *b = get_Quot_right(n); - tarval *tv = value_of(b); - - if (tv != tarval_bad) { - int rem = tarval_fp_ops_enabled(); - - /* - * 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. - */ - tarval_enable_fp_ops(1); - tv = tarval_quo(get_mode_one(mode), tv); - 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_nodes_block(n); - ir_node *c = new_Const(tv); - ir_node *a = get_Quot_left(n); - ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode); - ir_node *mem = get_Quot_mem(n); - - /* skip a potential Pin */ - 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(blk)); - set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); - set_Tuple_pred(n, pn_Quot_res, m); - DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL); - } - } - } - return n; -} /* transform_node_Quot */ - -/** - * 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; - ir_node *a = get_Abs_op(n); - ir_mode *mode; - - HANDLE_UNOP_PHI(tarval_abs, a, c); - - switch (classify_value_sign(a)) { - case value_classified_negative: - mode = get_irn_mode(n); - - /* - * We can replace the Abs by -x 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), get_nodes_block(n), a, mode); - - DBG_OPT_CONFIRM(oldn, n); - return n; - case value_classified_positive: - /* n is positive, Abs is not needed */ - n = a; - - 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), 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), 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. @@ -3231,184 +3316,260 @@ static ir_node *transform_node_Cmp(ir_node *n) { * Replace the Cond by a Jmp if it branches on a constant * condition. */ -static ir_node *transform_node_Cond(ir_node *n) { - - ir_node *jmp; - ir_node *a = get_Cond_selector(n); - tarval *ta = value_of(a); +static ir_node *transform_node_Cond(ir_node *n) +{ + ir_node *a = get_Cond_selector(n); + ir_graph *irg = get_irn_irg(n); + ir_tarval *ta; + ir_node *jmp; /* we need block info which is not available in floating irgs */ - if (get_irg_pinned(current_ir_graph) == op_pin_state_floats) + if (get_irg_pinned(irg) == op_pin_state_floats) return n; - if ((ta != tarval_bad) && - (get_irn_mode(a) == mode_b) && - (get_opt_unreachable_code())) { + ta = value_of(a); + if (ta == tarval_bad && is_Cmp(a)) { + /* try again with a direct call to compute_cmp, as we don't care + * about the MODEB_LOWERED flag here */ + ta = compute_cmp_ext(a); + } + + if (ta != tarval_bad && get_irn_mode(a) == mode_b) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ ir_node *blk = get_nodes_block(n); jmp = new_r_Jmp(blk); - turn_into_tuple(n, pn_Cond_max); + turn_into_tuple(n, pn_Cond_max+1); if (ta == tarval_b_true) { - set_Tuple_pred(n, pn_Cond_false, new_Bad()); + set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X)); set_Tuple_pred(n, pn_Cond_true, jmp); } else { set_Tuple_pred(n, pn_Cond_false, jmp); - set_Tuple_pred(n, pn_Cond_true, new_Bad()); + set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X)); + } + clear_irg_properties(irg, IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE); + } + return n; +} + +static ir_node *transform_node_Switch(ir_node *n) +{ + ir_node *op = get_Switch_selector(n); + ir_tarval *val = value_of(op); + if (val != tarval_bad) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + unsigned n_outs = get_Switch_n_outs(n); + ir_node *block = get_nodes_block(n); + ir_node *bad = new_r_Bad(irg, mode_X); + ir_node **in = XMALLOCN(ir_node*, n_outs); + const ir_switch_table *table = get_Switch_table(n); + size_t n_entries = ir_switch_table_get_n_entries(table); + long jmp_pn = 0; + size_t i; + unsigned o; + for (i = 0; i < n_entries; ++i) { + const ir_switch_table_entry *entry + = ir_switch_table_get_entry_const(table, i); + ir_tarval *min = entry->min; + ir_tarval *max = entry->max; + if (entry->pn == 0) + continue; + if ((min == max && min == val) + || (tarval_cmp(val, min) != ir_relation_less + && tarval_cmp(val, max) != ir_relation_greater)) { + jmp_pn = entry->pn; + break; + } + } + for (o = 0; o < n_outs; ++o) { + if (o == (unsigned)jmp_pn) { + in[o] = new_rd_Jmp(dbgi, block); + } else { + in[o] = bad; + } } - /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), blk); + return new_r_Tuple(block, (int)n_outs, in); } return n; -} /* transform_node_Cond */ - -/** - * 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 - * and(a OP c, b OP c) -> and(a, b) OP c - * note, might return a different op than n + * 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) + * + * This normalisation is good for things like x-(x&y) esp. in 186.crafty. */ -static ir_node *transform_bitwise_distributive(ir_node *n, - recursive_transform trans_func) +static ir_node *transform_node_shift_bitop(ir_node *n) { - ir_node *oldn = n; - ir_node *a = get_binop_left(n); - ir_node *b = get_binop_right(n); - ir_op *op = get_irn_op(a); - ir_op *op_root = get_irn_op(n); - - if(op != get_irn_op(b)) + ir_graph *irg = get_irn_irg(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *left; + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_shift; + + if (irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_NORMALISATION2)) return n; - /* and(conv(a), conv(b)) -> conv(and(a,b)) */ - if (op == op_Conv) { - ir_node *a_op = get_Conv_op(a); - ir_node *b_op = get_Conv_op(b); - 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_nodes_block(n); + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); - n = exact_copy(n); - set_binop_left(n, a_op); - set_binop_right(n, b_op); - set_irn_mode(n, a_mode); - n = trans_func(n); - n = new_r_Conv(blk, n, get_irn_mode(oldn)); + if (!is_Const(right)) + return n; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - 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; - if (op == op_Eor) { - /* nothing to gain here */ + /* 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; } - if (op == op_Shrs || op == op_Shr || op == op_Shl - || op == op_And || op == op_Or || op == op_Eor) { - ir_node *a_left = get_binop_left(a); - ir_node *a_right = get_binop_right(a); - ir_node *b_left = get_binop_left(b); - ir_node *b_right = get_binop_right(b); - ir_node *c = NULL; - ir_node *op1 = NULL; - ir_node *op2 = NULL; + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; - if (is_op_commutative(op)) { - if (a_left == b_left) { - c = a_left; - op1 = a_right; - op2 = b_right; - } else if(a_left == b_right) { - c = a_left; - op1 = a_right; - op2 = b_left; - } else if(a_right == b_left) { - c = a_right; - op1 = a_left; - op2 = b_right; - } - } - if(a_right == b_right) { - c = a_right; - op1 = a_left; - op2 = b_left; - } + bitop_left = get_binop_left(left); - if (c != NULL) { - /* (a sop c) & (b sop c) => (a & b) sop c */ - ir_node *blk = get_nodes_block(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); - ir_node *new_n = exact_copy(n); - set_binop_left(new_n, op1); - set_binop_right(new_n, op2); - new_n = trans_func(new_n); + assert(get_tarval_mode(tv1) == mode); - if(op_root == op_Eor && op == op_Or) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_mode *mode = get_irn_mode(c); + if (is_Shl(n)) { + new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shl(tv1, tv2); + } else if (is_Shr(n)) { + new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shr(tv1, tv2); + } else if (is_Shrs(n)) { + new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shrs(tv1, tv2); + } else { + assert(is_Rotl(n)); + new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_rotl(tv1, tv2); + } - c = new_rd_Not(dbgi, blk, c, mode); - n = new_rd_And(dbgi, blk, new_n, c, mode); - } else { - n = exact_copy(a); - set_nodes_block(n, blk); - set_binop_left(n, new_n); - set_binop_right(n, c); - add_identities(current_ir_graph->value_table, n); - } + assert(get_tarval_mode(tv_shift) == mode); + irg = get_irn_irg(n); + new_const = new_r_Const(irg, tv_shift); - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); - return n; - } + if (op_left == op_And) { + new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); + } else if (op_left == op_Or) { + new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); + } else { + assert(op_left == op_Eor); + new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); } - return n; + return new_bitop; } /** * Transform an And. */ -static ir_node *transform_node_And(ir_node *n) { +static ir_node *transform_node_And(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); ir_mode *mode; - mode = get_irn_mode(n); - HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode); - - /* we can evaluate 2 Projs of the same Cmp */ - if (mode == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if (pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(pred_a); - pn_Cmp pn_a = get_Proj_proj(a); - pn_Cmp pn_b = get_Proj_proj(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a & pn_b; + n = fold_constant_associativity(n, tarval_and); + if (n != oldn) + return n; - return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc); + if (is_Cmp(a) && is_Cmp(b)) { + ir_node *a_left = get_Cmp_left(a); + ir_node *a_right = get_Cmp_right(a); + ir_node *b_left = get_Cmp_left(b); + ir_node *b_right = get_Cmp_right(b); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + /* we can combine the relations of two compares with the same + * operands */ + if (a_left == b_left && b_left == b_right) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_relation new_relation = a_relation & b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */ + if (a_relation == b_relation && a_relation == ir_relation_equal + && !mode_is_float(get_irn_mode(a_left)) + && !mode_is_float(get_irn_mode(b_left))) { + if (values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xora, b_mode); + ir_node *orn = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_graph *irg = get_irn_irg(n); + ir_node *zero = create_zero_const(irg, b_mode); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal); + } + if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_left))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *orn = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_graph *irg = get_irn_irg(n); + ir_node *zero = create_zero_const(irg, a_mode); + return new_rd_Cmp(dbgi, block, orn, zero, ir_relation_equal); + } } } - if (is_Or(a)) { - if (is_Not(b)) { + + mode = get_irn_mode(n); + HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode); + + if (is_Or(a) || is_Or_Eor_Add(a)) { + ir_node *or_left = get_binop_left(a); + ir_node *or_right = get_binop_right(a); + if (complement_values(or_left, b)) { + /* (a|b) & ~a => b & ~a */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_right, b, mode); + } else if (complement_values(or_right, b)) { + /* (a|b) & ~b => a & ~b */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_left, b, mode); + } else if (is_Not(b)) { ir_node *op = get_Not_op(b); if (is_And(op)) { ir_node *ba = get_And_left(op); ir_node *bb = get_And_right(op); /* it's enough to test the following cases due to normalization! */ - if (get_Or_left(a) == ba && get_Or_right(a) == bb) { + if (or_left == ba && or_right == bb) { /* (a|b) & ~(a&b) = a^b */ ir_node *block = get_nodes_block(n); @@ -3419,15 +3580,27 @@ static ir_node *transform_node_And(ir_node *n) { } } } - if (is_Or(b)) { - if (is_Not(a)) { + if (is_Or(b) || is_Or_Eor_Add(b)) { + ir_node *or_left = get_binop_left(b); + ir_node *or_right = get_binop_right(b); + if (complement_values(or_left, a)) { + /* (a|b) & ~a => b & ~a */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_right, a, mode); + } else if (complement_values(or_right, a)) { + /* (a|b) & ~b => a & ~b */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_left, a, mode); + } else if (is_Not(a)) { ir_node *op = get_Not_op(a); if (is_And(op)) { ir_node *aa = get_And_left(op); ir_node *ab = get_And_right(op); /* it's enough to test the following cases due to normalization! */ - if (get_Or_left(b) == aa && get_Or_right(b) == ab) { + if (or_left == aa && or_right == ab) { /* (a|b) & ~(a&b) = a^b */ ir_node *block = get_nodes_block(n); @@ -3438,9 +3611,9 @@ static ir_node *transform_node_And(ir_node *n) { } } } - if (is_Eor(a)) { - ir_node *al = get_Eor_left(a); - ir_node *ar = get_Eor_right(a); + if (is_Eor(a) || is_Or_Eor_Add(a)) { + ir_node *al = get_binop_left(a); + ir_node *ar = get_binop_right(a); if (al == b) { /* (b ^ a) & b -> ~a & b */ @@ -3463,9 +3636,9 @@ static ir_node *transform_node_And(ir_node *n) { return n; } } - if (is_Eor(b)) { - ir_node *bl = get_Eor_left(b); - ir_node *br = get_Eor_right(b); + if (is_Eor(b) || is_Or_Eor_Add(b)) { + ir_node *bl = get_binop_left(b); + ir_node *br = get_binop_right(b); if (bl == a) { /* a & (a ^ b) -> a & ~b */ @@ -3501,67 +3674,34 @@ static ir_node *transform_node_And(ir_node *n) { return n; } - n = transform_bitwise_distributive(n, transform_node_And); - - return n; -} /* transform_node_And */ - -/** - * Transform an Eor. - */ -static ir_node *transform_node_Eor(ir_node *n) { - ir_node *c, *oldn = n; - ir_node *a = get_Eor_left(n); - ir_node *b = get_Eor_right(n); - ir_mode *mode = get_irn_mode(n); - - HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode); - - /* we can evaluate 2 Projs of the same Cmp */ - if (mode == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if(pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(pred_a); - pn_Cmp pn_a = get_Proj_proj(a); - pn_Cmp pn_b = get_Proj_proj(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a ^ pn_b; - - return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc); + if (is_Const(a)) { + vrp_attr *b_vrp = vrp_get_info(b); + ir_tarval *a_val = get_Const_tarval(a); + if (b_vrp != NULL && tarval_or(a_val, b_vrp->bits_not_set) == a_val) { + return b; } } - if (a == b) { - /* a ^ a = 0 */ - n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, - get_mode_null(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A); - } else if (is_Const(b)) { - if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */ - ir_node *cnst = new_Const(tarval_not(get_Const_tarval(b))); - ir_node *not_op = get_Not_op(a); - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_mode *mode = get_irn_mode(n); - n = new_rd_Eor(dbg, block, not_op, cnst, mode); - return n; - } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */ - n = new_r_Not(get_nodes_block(n), a, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + if (is_Const(b)) { + vrp_attr *a_vrp = vrp_get_info(a); + ir_tarval *b_val = get_Const_tarval(b); + if (a_vrp != NULL && tarval_or(b_val, a_vrp->bits_not_set) == b_val) { + return a; } - } else { - n = transform_bitwise_distributive(n, transform_node_Eor); } + n = transform_bitwise_distributive(n, transform_node_And); + if (is_And(n)) + n = transform_node_bitop_shift(n); + return n; -} /* transform_node_Eor */ +} /** * Transform a Not. */ -static ir_node *transform_node_Not(ir_node *n) { +static ir_node *transform_node_Not(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *a = get_Not_op(n); ir_mode *mode = get_irn_mode(n); @@ -3569,42 +3709,40 @@ static ir_node *transform_node_Not(ir_node *n) { HANDLE_UNOP_PHI(tarval_not,a,c); /* check for a boolean Not */ - if (mode == mode_b && is_Proj(a)) { - ir_node *a_pred = get_Proj_pred(a); - if (is_Cmp(a_pred)) { - ir_node *cmp_block = get_nodes_block(a_pred); - /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(cmp_block, 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; - } + if (is_Cmp(a)) { + dbg_info *dbgi = get_irn_dbg_info(a); + ir_node *block = get_nodes_block(a); + ir_relation relation = get_Cmp_relation(a); + relation = get_negated_relation(relation); + n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); + return n; } - if (is_Eor(a)) { - ir_node *eor_b = get_Eor_right(a); - if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */ - ir_node *cnst = new_Const(tarval_not(get_Const_tarval(eor_b))); - ir_node *eor_a = get_Eor_left(a); - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_mode *mode = get_irn_mode(n); - n = new_rd_Eor(dbg, block, eor_a, cnst, mode); - return n; - } + + /* normalize ~(a ^ b) => a ^ ~b */ + if (is_Eor(a) || is_Or_Eor_Add(a)) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *eor_right = get_binop_right(a); + ir_node *eor_left = get_binop_left(a); + eor_right = new_rd_Not(dbg, block, eor_right, mode); + n = new_rd_Eor(dbg, block, eor_left, eor_right, mode); + return n; } + if (get_mode_arithmetic(mode) == irma_twos_complement) { if (is_Minus(a)) { /* ~-x -> x + -1 */ dbg_info *dbg = get_irn_dbg_info(n); - ir_graph *irg = current_ir_graph; + ir_graph *irg = get_irn_irg(n); ir_node *block = get_nodes_block(n); ir_node *add_l = get_Minus_op(a); ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode)); n = new_rd_Add(dbg, block, add_l, add_r, mode); - } else if (is_Add(a)) { - ir_node *add_r = get_Add_right(a); + } else if (is_Add(a) || is_Or_Eor_Add(a)) { + ir_node *add_r = get_binop_right(a); if (is_Const(add_r) && is_Const_all_one(add_r)) { /* ~(x + -1) = -x */ - ir_node *op = get_Add_left(a); + ir_node *op = get_binop_left(a); ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); @@ -3612,7 +3750,7 @@ static ir_node *transform_node_Not(ir_node *n) { } } return n; -} /* transform_node_Not */ +} /** * Transform a Minus. @@ -3623,7 +3761,8 @@ static ir_node *transform_node_Not(ir_node *n) { * -(a >>s (size-1)) = a >>u (size-1) * -(a * const) -> a * -const */ -static ir_node *transform_node_Minus(ir_node *n) { +static ir_node *transform_node_Minus(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *a = get_Minus_op(n); ir_mode *mode; @@ -3635,10 +3774,11 @@ static ir_node *transform_node_Minus(ir_node *n) { /* the following rules are only to twos-complement */ if (is_Not(a)) { /* -(~x) = x + 1 */ - ir_node *op = get_Not_op(a); - tarval *tv = get_mode_one(mode); - ir_node *blk = get_nodes_block(n); - ir_node *c = new_Const(tv); + ir_node *op = get_Not_op(a); + ir_tarval *tv = get_mode_one(mode); + ir_node *blk = get_nodes_block(n); + ir_graph *irg = get_irn_irg(blk); + ir_node *c = new_r_Const(irg, tv); n = new_rd_Add(get_irn_dbg_info(n), blk, op, c, mode); DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT); return n; @@ -3647,7 +3787,7 @@ static ir_node *transform_node_Minus(ir_node *n) { ir_node *c = get_Shr_right(a); if (is_Const(c)) { - tarval *tv = get_Const_tarval(c); + ir_tarval *tv = get_Const_tarval(c); if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>u (size-1)) = a >>s (size-1) */ @@ -3663,7 +3803,7 @@ static ir_node *transform_node_Minus(ir_node *n) { ir_node *c = get_Shrs_right(a); if (is_Const(c)) { - tarval *tv = get_Const_tarval(c); + ir_tarval *tv = get_Const_tarval(c); if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>s (size-1)) = a >>u (size-1) */ @@ -3688,13 +3828,14 @@ 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); - tarval *tv = value_of(mul_r); + ir_node *mul_l = get_Mul_left(a); + ir_node *mul_r = get_Mul_right(a); + ir_tarval *tv = value_of(mul_r); if (tv != tarval_bad) { tv = tarval_neg(tv); if (tv != tarval_bad) { - ir_node *cnst = new_Const(tv); + ir_graph *irg = get_irn_irg(n); + ir_node *cnst = new_r_Const(irg, tv); dbg_info *dbg = get_irn_dbg_info(a); ir_node *block = get_nodes_block(a); n = new_rd_Mul(dbg, block, mul_l, cnst, mode); @@ -3705,93 +3846,74 @@ static ir_node *transform_node_Minus(ir_node *n) { } return n; -} /* transform_node_Minus */ +} /** - * Transform a Cast_type(Const) into a new Const_type + * Transform a Proj(Load) with a non-null address. */ -static ir_node *transform_node_Cast(ir_node *n) { - ir_node *oldn = n; - ir_node *pred = get_Cast_op(n); - ir_type *tp = get_irn_type(n); - - if (is_Const(pred) && get_Const_type(pred) != tp) { - n = new_rd_Const_type(NULL, current_ir_graph, 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_mode(pred), - get_SymConst_symbol(pred), get_SymConst_kind(pred), tp); - DBG_OPT_CSTEVAL(oldn, n); - } +static ir_node *transform_node_Proj_Load(ir_node *proj) +{ + if (get_irn_mode(proj) == mode_X) { + ir_node *load = get_Proj_pred(proj); - return n; -} /* transform_node_Cast */ + /* get the Load address */ + const ir_node *addr = get_Load_ptr(load); + const ir_node *confirm; -/** - * 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(blk); - } + 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) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(load); + return new_r_Jmp(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(blk); - } +static ir_node *transform_node_Proj_Store(ir_node *proj) +{ + 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) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(store); + return new_r_Jmp(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. */ -static ir_node *transform_node_Proj_Div(ir_node *proj) { +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 *res, *new_mem; @@ -3813,14 +3935,17 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { case pn_Div_X_regular: return new_r_Jmp(get_nodes_block(div)); - case pn_Div_X_except: + case pn_Div_X_except: { + ir_graph *irg = get_irn_irg(proj); /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); - return new_Bad(); + return new_r_Bad(irg, mode_X); + } - case pn_Div_M: + case pn_Div_M: { + ir_graph *irg = get_irn_irg(proj); res = get_Div_mem(div); - new_mem = get_irg_no_mem(current_ir_graph); + new_mem = get_irg_no_mem(irg); if (confirm) { /* This node can only float up to the Confirm block */ @@ -3831,15 +3956,17 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { set_Div_mem(div, new_mem); return res; } + } } return proj; -} /* transform_node_Proj_Div */ +} /** * Transform a Proj(Mod) with a non-zero value. * Removes the exceptions and routes the memory to the NoMem node. */ -static ir_node *transform_node_Proj_Mod(ir_node *proj) { +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 *res, *new_mem; @@ -3861,234 +3988,182 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { switch (proj_nr) { case pn_Mod_X_regular: - return new_r_Jmp(get_irn_n(mod, -1)); + return new_r_Jmp(get_nodes_block(mod)); - case pn_Mod_X_except: + case pn_Mod_X_except: { + ir_graph *irg = get_irn_irg(proj); /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); - return new_Bad(); - - case pn_Mod_M: - res = get_Mod_mem(mod); - new_mem = get_irg_no_mem(current_ir_graph); - - if (confirm) { - /* This node can only float up to the Confirm block */ - new_mem = new_r_Pin(get_nodes_block(confirm), new_mem); - } - /* this is a Mod without exception, we can remove the memory edge */ - set_Mod_mem(mod, new_mem); - return res; - case pn_Mod_res: - if (get_Mod_left(mod) == b) { - /* a % a = 0 if a != 0 */ - ir_mode *mode = get_irn_mode(proj); - ir_node *res = new_Const(get_mode_null(mode)); - - DBG_OPT_CSTEVAL(mod, res); - return res; - } - } - } - return proj; -} /* transform_node_Proj_Mod */ - -/** - * Transform a Proj(DivMod) with a non-zero value. - * Removes the exceptions and routes the memory to the NoMem node. - */ -static ir_node *transform_node_Proj_DivMod(ir_node *proj) { - ir_node *divmod = get_Proj_pred(proj); - ir_node *b = get_DivMod_right(divmod); - ir_node *res, *new_mem; - const ir_node *confirm; - long proj_nr; - - if (value_not_zero(b, &confirm)) { - /* DivMod(x, y) && y != 0 */ - proj_nr = get_Proj_proj(proj); - - if (confirm == NULL) { - /* we are sure we have a Const != 0 */ - new_mem = get_DivMod_mem(divmod); - new_mem = skip_Pin(new_mem); - set_DivMod_mem(divmod, new_mem); - set_irn_pinned(divmod, op_pin_state_floats); + return new_r_Bad(irg, mode_X); } - switch (proj_nr) { - - case pn_DivMod_X_regular: - return new_r_Jmp(get_nodes_block(divmod)); - - case pn_DivMod_X_except: - /* we found an exception handler, remove it */ - DBG_OPT_EXC_REM(proj); - return new_Bad(); - - case pn_DivMod_M: - res = get_DivMod_mem(divmod); - new_mem = get_irg_no_mem(current_ir_graph); + case pn_Mod_M: { + ir_graph *irg = get_irn_irg(proj); + res = get_Mod_mem(mod); + new_mem = get_irg_no_mem(irg); if (confirm) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(get_nodes_block(confirm), new_mem); } - /* this is a DivMod without exception, we can remove the memory edge */ - set_DivMod_mem(divmod, new_mem); + /* this is a Mod without exception, we can remove the memory edge */ + set_Mod_mem(mod, new_mem); return res; - - case pn_DivMod_res_mod: - if (get_DivMod_left(divmod) == b) { + } + case pn_Mod_res: + if (get_Mod_left(mod) == b) { /* a % a = 0 if a != 0 */ - ir_mode *mode = get_irn_mode(proj); - ir_node *res = new_Const(get_mode_null(mode)); + ir_graph *irg = get_irn_irg(proj); + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_r_Const(irg, get_mode_null(mode)); - DBG_OPT_CSTEVAL(divmod, res); + DBG_OPT_CSTEVAL(mod, res); return res; } } } return proj; -} /* transform_node_Proj_DivMod */ - -/** - * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases. - */ -static ir_node *transform_node_Proj_Cond(ir_node *proj) { - if (get_opt_unreachable_code()) { - ir_node *n = get_Proj_pred(proj); - ir_node *b = get_Cond_selector(n); - - if (mode_is_int(get_irn_mode(b))) { - tarval *tb = value_of(b); - - if (tb != tarval_bad) { - /* we have a constant switch */ - long num = get_Proj_proj(proj); - - if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */ - if (get_tarval_long(tb) == num) { - /* Do NOT create a jump here, or we will have 2 control flow ops - * in a block. This case is optimized away in optimize_cf(). */ - return proj; - } else { - /* this case will NEVER be taken, kill it */ - return get_irg_bad(current_ir_graph); - } - } - } - } - } - return proj; -} /* transform_node_Proj_Cond */ +} /** - * Create a 0 constant of given mode. + * return true if the operation returns a value with exactly 1 bit set */ -static ir_node *create_zero_const(ir_mode *mode) { - tarval *tv = get_mode_null(mode); - ir_node *cnst = new_Const(tv); - - 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) +static bool is_single_bit(const ir_node *node) { - if (is_Const(n)) return const_const; - if (is_irn_constlike(n)) return const_like; - return const_other; + /* a first implementation, could be extended with vrp and others... */ + if (is_Shl(node)) { + ir_node *shl_l = get_Shl_left(node); + ir_mode *mode = get_irn_mode(node); + int modulo = get_mode_modulo_shift(mode); + /* this works if we shift a 1 and we have modulo shift */ + if (is_Const(shl_l) && is_Const_one(shl_l) + && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) { + return true; + } + } else if (is_Const(node)) { + ir_tarval *tv = get_Const_tarval(node); + return tarval_is_single_bit(tv); + } + return false; } /** - * Determines whether r is more constlike or has a larger index (in that order) - * than l. + * checks if node just flips a bit in another node and returns that other node + * if so. @p tv should be a value having just 1 bit set */ -static int operands_are_normalized(const ir_node *l, const ir_node *r) +static ir_node *flips_bit(const ir_node *node, ir_tarval *tv) { - 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)); + if (is_Not(node)) + return get_Not_op(node); + if (is_Eor(node)) { + ir_node *right = get_Eor_right(node); + if (is_Const(right)) { + ir_tarval *right_tv = get_Const_tarval(right); + ir_mode *mode = get_irn_mode(node); + if (tarval_and(right_tv, tv) != get_mode_null(mode)) + return get_Eor_left(node); + } + } + return NULL; } /** * 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); - 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) { - case pn_Cmp_False: - return new_Const(get_tarval_b_false()); - case pn_Cmp_True: - return new_Const(get_tarval_b_true()); - case pn_Cmp_Leg: - if (!mode_is_float(get_irn_mode(left))) - return new_Const(get_tarval_b_true()); - break; - default: - break; +static ir_node *transform_node_Cmp(ir_node *n) +{ + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + ir_mode *mode = get_irn_mode(left); + ir_tarval *tv = NULL; + bool changed = false; + bool changedc = false; + ir_relation relation = get_Cmp_relation(n); + ir_relation possible = ir_get_possible_cmp_relations(left, right); + + /* mask out impossible relations */ + ir_relation new_relation = relation & possible; + if (new_relation != relation) { + relation = new_relation; + changed = true; } - /* remove Casts of both sides */ - left = skip_Cast(left); - right = skip_Cast(right); - /* Remove unnecessary conversions */ - /* TODO handle constants */ - if (is_Conv(left) && is_Conv(right)) { - ir_mode *mode = get_irn_mode(left); - ir_node *op_left = get_Conv_op(left); - ir_node *op_right = get_Conv_op(right); - 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) - && mode_left != mode_b && mode_right != mode_b) { - ir_node *block = get_nodes_block(n); + if (!mode_is_float(mode) + || be_get_backend_param()->mode_float_arithmetic == NULL) { + if (is_Conv(left) && is_Conv(right)) { + ir_node *op_left = get_Conv_op(left); + ir_node *op_right = get_Conv_op(right); + 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) + && mode_left != mode_b && mode_right != mode_b) { + ir_node *block = get_nodes_block(n); - if (mode_left == mode_right) { - left = op_left; - right = op_right; - changed |= 1; - DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV); - } else if (smaller_mode(mode_left, mode_right)) { - left = new_r_Conv(block, op_left, mode_right); - right = op_right; - changed |= 1; - DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); - } else if (smaller_mode(mode_right, mode_left)) { - left = op_left; - right = new_r_Conv(block, op_right, mode_left); - changed |= 1; - DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + if (mode_left == mode_right) { + left = op_left; + right = op_right; + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV); + } else if (smaller_mode(mode_left, mode_right)) { + left = new_r_Conv(block, op_left, mode_right); + right = op_right; + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + } else if (smaller_mode(mode_right, mode_left)) { + left = op_left; + right = new_r_Conv(block, op_right, mode_left); + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + } + mode = get_irn_mode(left); + } + } + if (is_Conv(left) && is_Const(right)) { + ir_node *op_left = get_Conv_op(left); + ir_mode *mode_left = get_irn_mode(op_left); + if (smaller_mode(mode_left, mode) && mode_left != mode_b) { + ir_tarval *tv = get_Const_tarval(right); + tarval_int_overflow_mode_t last_mode + = tarval_get_integer_overflow_mode(); + ir_tarval *new_tv; + tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD); + new_tv = tarval_convert_to(tv, mode_left); + tarval_set_integer_overflow_mode(last_mode); + if (new_tv != tarval_bad) { + ir_graph *irg = get_irn_irg(n); + left = op_left; + right = new_r_Const(irg, new_tv); + mode = get_irn_mode(left); + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + } } } } + /* + * Optimize -a CMP -b into b CMP a. + * This works only for modes where unary Minus cannot Overflow. + * Note that two-complement integers can Overflow so it will NOT work. + */ + if (!mode_overflow_on_unary_Minus(mode) && + is_Minus(left) && is_Minus(right)) { + left = get_Minus_op(left); + right = get_Minus_op(right); + relation = get_inversed_relation(relation); + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } + /* remove operation on both sides if possible */ - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { /* * The following operations are NOT safe for floating point operations, for instance * 1.0 + inf == 2.0 + inf, =/=> x == y */ - if (mode_is_int(get_irn_mode(left))) { + if (mode_is_int(mode)) { unsigned lop = get_irn_opcode(left); if (lop == get_irn_opcode(right)) { @@ -4101,7 +4176,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */ left = get_unop_op(left); right = get_unop_op(right); - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); break; case iro_Add: @@ -4114,25 +4189,25 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* X + a CMP X + b ==> a CMP b */ left = lr; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (ll == rr) { /* X + a CMP b + X ==> a CMP b */ left = lr; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rl) { /* a + X CMP X + b ==> a CMP b */ left = ll; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rr) { /* a + X CMP b + X ==> a CMP b */ left = ll; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -4146,13 +4221,13 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* X - a CMP X - b ==> a CMP b */ left = lr; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rr) { /* a - X CMP b - X ==> a CMP b */ left = ll; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -4161,7 +4236,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* a ROTL X CMP b ROTL X ==> a CMP b */ left = get_Rotl_left(left); right = get_Rotl_left(right); - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -4171,38 +4246,41 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } /* X+A == A, A+X == A, A-X == A -> X == 0 */ - if (is_Add(left) || is_Sub(left)) { + if (is_Add(left) || is_Sub(left) || is_Or_Eor_Add(left)) { ir_node *ll = get_binop_left(left); ir_node *lr = get_binop_right(left); - if (lr == right && is_Add(left)) { + if (lr == right && (is_Add(left) || is_Or_Eor_Add(left))) { ir_node *tmp = ll; ll = lr; lr = tmp; } if (ll == right) { + ir_graph *irg = get_irn_irg(n); left = lr; - right = create_zero_const(get_irn_mode(left)); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } - if (is_Add(right) || is_Sub(right)) { + if (is_Add(right) || is_Sub(right) || is_Or_Eor_Add(right)) { ir_node *rl = get_binop_left(right); ir_node *rr = get_binop_right(right); - if (rr == left && is_Add(right)) { + if (rr == left && (is_Add(right) || is_Or_Eor_Add(right))) { ir_node *tmp = rl; rl = rr; rr = tmp; } if (rl == left) { + ir_graph *irg = get_irn_irg(n); left = rr; - right = create_zero_const(get_irn_mode(left)); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } + if (is_And(left) && is_Const(right)) { ir_node *ll = get_binop_left(left); ir_node *lr = get_binop_right(left); @@ -4214,37 +4292,112 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { ir_node *llr = get_Shr_right(ll); if (is_Const(llr)) { dbg_info *dbg = get_irn_dbg_info(left); + ir_graph *irg = get_irn_irg(left); + + ir_tarval *c1 = get_Const_tarval(llr); + ir_tarval *c2 = get_Const_tarval(lr); + ir_tarval *c3 = get_Const_tarval(right); + ir_tarval *mask = tarval_shl(c2, c1); + ir_tarval *value = tarval_shl(c3, c1); + + left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode); + right = new_r_Const(irg, value); + changed = true; + } + } + } + /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0 + * cases */ + if (is_Const(right) && is_Const_null(right) && + (is_Eor(left) || is_Or_Eor_Add(left))) { + right = get_Eor_right(left); + left = get_Eor_left(left); + changed = true; + } + } + } - tarval *c1 = get_Const_tarval(llr); - tarval *c2 = get_Const_tarval(lr); - tarval *c3 = get_Const_tarval(right); - tarval *mask = tarval_shl(c2, c1); - tarval *value = tarval_shl(c3, c1); + if (mode_is_int(mode) && is_And(left)) { + /* a complicated Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced + * by the simpler Cmp(And(1bit, val), 0) negated pnc */ + if (relation == ir_relation_equal + || (mode_is_signed(mode) && relation == ir_relation_less_greater) + || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less)) { + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (and1 == right) { + ir_node *tmp = and0; + and0 = and1; + and1 = tmp; + } + if (and0 == right && is_single_bit(and0)) { + ir_graph *irg = get_irn_irg(n); + relation = + relation == ir_relation_equal ? ir_relation_less_greater + : ir_relation_equal; + right = create_zero_const(irg, mode); + changed |= 1; + goto is_bittest; + } + } - left = new_rd_And(dbg, block, get_Shr_left(ll), new_Const(mask), mode); - right = new_Const(value); + if (is_Const(right) && is_Const_null(right) && + (relation == ir_relation_equal + || (relation == ir_relation_less_greater) + || (!mode_is_signed(mode) && relation == ir_relation_greater))) { +is_bittest: { + /* instead of flipping the bit before the bit-test operation negate + * pnc */ + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (is_Const(and1)) { + ir_tarval *tv = get_Const_tarval(and1); + if (tarval_is_single_bit(tv)) { + ir_node *flipped = flips_bit(and0, tv); + if (flipped != NULL) { + dbg_info *dbgi = get_irn_dbg_info(left); + ir_node *block = get_nodes_block(left); + relation = get_negated_relation(relation); + left = new_rd_And(dbgi, block, flipped, and1, mode); changed |= 1; } } } - } /* mode_is_int(...) */ - } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */ + } + } + } /* replace mode_b compares with ands/ors */ - if (get_irn_mode(left) == mode_b) { + if (mode == mode_b) { ir_node *block = get_nodes_block(n); ir_node *bres; - switch (proj_nr) { - case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break; - case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break; - case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break; - case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break; - case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break; - case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break; - default: bres = NULL; - } - if (bres) { + switch (relation) { + case ir_relation_less_equal: + bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b); + break; + case ir_relation_less: + bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); + break; + case ir_relation_greater_equal: + bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b); + break; + case ir_relation_greater: + bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); + break; + case ir_relation_less_greater: + bres = new_r_Eor(block, left, right, mode_b); + break; + case ir_relation_equal: + bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); + break; + default: +#ifdef DEBUG_libfirm + ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n); +#endif + bres = NULL; + } + if (bres != NULL) { DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL); return bres; } @@ -4257,12 +4410,11 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ if (!operands_are_normalized(left, right)) { ir_node *t = left; - left = right; right = t; - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; + relation = get_inversed_relation(relation); + changed = true; } /* @@ -4273,7 +4425,44 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ tv = value_of(right); if (tv != tarval_bad) { - mode = get_irn_mode(right); + ir_mode *mode = get_irn_mode(right); + + /* cmp(mux(x, cf, ct), c2) can be eliminated: + * cmp(ct,c2) | cmp(cf,c2) | result + * -----------|------------|-------- + * true | true | True + * false | false | False + * true | false | x + * false | true | not(x) + */ + if (is_Mux(left)) { + ir_node *mux_true = get_Mux_true(left); + ir_node *mux_false = get_Mux_false(left); + if (is_Const(mux_true) && is_Const(mux_false)) { + /* we can fold true/false constant separately */ + ir_tarval *tv_true = get_Const_tarval(mux_true); + ir_tarval *tv_false = get_Const_tarval(mux_false); + ir_relation r_true = tarval_cmp(tv_true, tv); + ir_relation r_false = tarval_cmp(tv_false, tv); + if (r_true != ir_relation_false + || r_false != ir_relation_false) { + bool rel_true = (r_true & relation) != 0; + bool rel_false = (r_false & relation) != 0; + ir_node *cond = get_Mux_sel(left); + if (rel_true == rel_false) { + relation = rel_true ? ir_relation_true + : ir_relation_false; + } else if (rel_true) { + return cond; + } else { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *notn = new_rd_Not(dbgi, block, cond, mode_b); + return notn; + } + } + } + } /* TODO extend to arbitrary constants */ if (is_Conv(left) && tarval_is_null(tv)) { @@ -4282,14 +4471,17 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* * UpConv(x) REL 0 ==> x REL 0 + * Don't do this for float values as it's unclear whether it is a + * win. (on the other side it makes detection/creation of fabs hard) */ if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) && - ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) || - mode_is_signed(mode) || !mode_is_signed(op_mode))) { + ((relation == ir_relation_equal || relation == ir_relation_less_greater) || + mode_is_signed(mode) || !mode_is_signed(op_mode)) && + !mode_is_float(mode)) { tv = get_mode_null(op_mode); left = op; mode = op_mode; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } } @@ -4305,78 +4497,58 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ if (is_Minus(left) && (!mode_overflow_on_unary_Minus(mode) || - (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { + (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) { tv = tarval_neg(tv); if (tv != tarval_bad) { left = get_Minus_op(left); - proj_nr = get_inversed_pnc(proj_nr); - changed |= 2; + relation = get_inversed_relation(relation); + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } - } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) { + } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) { /* Not(a) ==/!= c ==> a ==/!= Not(c) */ tv = tarval_not(tv); if (tv != tarval_bad) { left = get_Not_op(left); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } /* for integer modes, we have more */ - if (mode_is_int(mode)) { - /* Ne includes Unordered which is not possible on integers. - * However, frontends often use this wrong, so fix it here */ - if (proj_nr & pn_Cmp_Uo) { - proj_nr &= ~pn_Cmp_Uo; - set_Proj_proj(proj, proj_nr); - } - + if (mode_is_int(mode) && !is_Const(left)) { /* 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) { + if ((relation == ir_relation_less || relation == ir_relation_greater_equal) && + tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) { tv = tarval_sub(tv, get_mode_one(mode), NULL); if (tv != tarval_bad) { - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + relation ^= ir_relation_equal; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ - else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && - tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { + else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) && + tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) { tv = tarval_add(tv, get_mode_one(mode)); if (tv != tarval_bad) { - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + relation ^= ir_relation_equal; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* the following reassociations work only for == and != */ - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { - -#if 0 /* Might be not that good in general */ - /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ - if (tarval_is_null(tv) && is_Sub(left)) { - right = get_Sub_right(left); - left = get_Sub_left(left); - - tv = value_of(right); - changed = 1; - DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); - } -#endif - + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { if (tv != tarval_bad) { /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ if (is_Sub(left)) { ir_node *c1 = get_Sub_right(left); - tarval *tv2 = value_of(c1); + ir_tarval *tv2 = value_of(c1); if (tv2 != tarval_bad) { tv2 = tarval_add(tv, value_of(c1)); @@ -4384,17 +4556,17 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (tv2 != tarval_bad) { left = get_Sub_left(left); tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ - else if (is_Add(left)) { - ir_node *a_l = get_Add_left(left); - ir_node *a_r = get_Add_right(left); + else if (is_Add(left) || is_Or_Eor_Add(left)) { + ir_node *a_l = get_binop_left(left); + ir_node *a_r = get_binop_right(left); ir_node *a; - tarval *tv2; + ir_tarval *tv2; if (is_Const(a_l)) { a = a_r; @@ -4410,35 +4582,27 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (tv2 != tarval_bad) { left = a; tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } /* -a == c ==> a == -c, -a != c ==> a != -c */ else if (is_Minus(left)) { - tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL); + ir_tarval *tv2 = tarval_sub(get_mode_null(mode), tv, NULL); if (tv2 != tarval_bad) { left = get_Minus_op(left); tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } - } /* == or != */ - /* the following reassociations work only for <= */ - 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 (is_Abs(left)) { // TODO something is missing here - } - } } - } /* mode_is_int */ + } - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { switch (get_irn_opcode(left)) { ir_node *c1; @@ -4449,12 +4613,13 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2 * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2 */ - tarval *mask = tarval_and(get_Const_tarval(c1), tv); + ir_tarval *mask = tarval_and(get_Const_tarval(c1), tv); if (mask != tv) { /* TODO: move to constant evaluation */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); - c1 = new_Const(tv); - DBG_OPT_CSTEVAL(proj, c1); + ir_graph *irg = get_irn_irg(n); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_r_Const(irg, tv); + DBG_OPT_CSTEVAL(n, c1); return c1; } @@ -4473,8 +4638,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (get_Const_tarval(c1) == tv) { /* fine: do the transformation */ tv = get_mode_null(get_tarval_mode(tv)); - proj_nr ^= pn_Cmp_Leg; - changed |= 2; + relation ^= ir_relation_less_equal_greater; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } @@ -4489,9 +4654,10 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ if (! tarval_is_null(get_Const_tarval(c1))) { /* TODO: move to constant evaluation */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); - c1 = new_Const(tv); - DBG_OPT_CSTEVAL(proj, c1); + ir_graph *irg = get_irn_irg(n); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_r_Const(irg, tv); + DBG_OPT_CSTEVAL(n, c1); return c1; } } @@ -4505,25 +4671,26 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ c1 = get_Shl_right(left); if (is_Const(c1)) { - tarval *tv1 = get_Const_tarval(c1); - ir_mode *mode = get_irn_mode(left); - tarval *minus1 = get_mode_all_one(mode); - tarval *amask = tarval_shr(minus1, tv1); - tarval *cmask = tarval_shl(minus1, tv1); - ir_node *sl, *blk; + ir_graph *irg = get_irn_irg(c1); + ir_tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + ir_tarval *minus1 = get_mode_all_one(mode); + ir_tarval *amask = tarval_shr(minus1, tv1); + ir_tarval *cmask = tarval_shl(minus1, tv1); + ir_node *sl, *blk; if (tarval_and(tv, cmask) != tv) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); - c1 = new_Const(tv); - DBG_OPT_CSTEVAL(proj, c1); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_r_Const(irg, tv); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shl_left(left); blk = get_nodes_block(n); - left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode); + left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shr(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; @@ -4536,25 +4703,26 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ c1 = get_Shr_right(left); if (is_Const(c1)) { - tarval *tv1 = get_Const_tarval(c1); - ir_mode *mode = get_irn_mode(left); - tarval *minus1 = get_mode_all_one(mode); - tarval *amask = tarval_shl(minus1, tv1); - tarval *cmask = tarval_shr(minus1, tv1); - ir_node *sl, *blk; + ir_graph *irg = get_irn_irg(c1); + ir_tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + ir_tarval *minus1 = get_mode_all_one(mode); + ir_tarval *amask = tarval_shl(minus1, tv1); + ir_tarval *cmask = tarval_shr(minus1, tv1); + ir_node *sl, *blk; if (tarval_and(tv, cmask) != tv) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); - c1 = new_Const(tv); - DBG_OPT_CSTEVAL(proj, c1); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_r_Const(irg, tv); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shr_left(left); blk = get_nodes_block(n); - left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode); + left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shl(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; @@ -4567,11 +4735,12 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { */ c1 = get_Shrs_right(left); if (is_Const(c1)) { - tarval *tv1 = get_Const_tarval(c1); - ir_mode *mode = get_irn_mode(left); - tarval *minus1 = get_mode_all_one(mode); - tarval *amask = tarval_shl(minus1, tv1); - tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1)); + ir_graph *irg = get_irn_irg(c1); + ir_tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + ir_tarval *minus1 = get_mode_all_one(mode); + ir_tarval *amask = tarval_shl(minus1, tv1); + ir_tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1)); ir_node *sl, *blk; cond = tarval_sub(cond, tv1, NULL); @@ -4579,46 +4748,49 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); - c1 = new_Const(tv); - DBG_OPT_CSTEVAL(proj, c1); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_r_Const(irg, tv); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shrs_left(left); blk = get_nodes_block(n); - left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_Const(amask), mode); + left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shl(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; - } /* switch */ + } } - } /* tarval != bad */ + } } - if (changed & 2) /* need a new Const */ - right = new_Const(tv); + if (changedc) { /* need a new Const */ + ir_graph *irg = get_irn_irg(n); + right = new_r_Const(irg, tv); + changed = true; + } - if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { + if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { ir_node *op = get_Proj_pred(left); - if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) || - (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) { + if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) { ir_node *c = get_binop_right(op); if (is_Const(c)) { - tarval *tv = get_Const_tarval(c); + ir_tarval *tv = get_Const_tarval(c); if (tarval_is_single_bit(tv)) { /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */ ir_node *v = get_binop_left(op); - ir_node *blk = get_irn_n(op, -1); + ir_node *blk = get_nodes_block(op); + ir_graph *irg = get_irn_irg(op); ir_mode *mode = get_irn_mode(v); tv = tarval_sub(tv, get_mode_one(mode), NULL); - left = new_rd_And(get_irn_dbg_info(op), blk, v, new_Const(tv), mode); - changed |= 1; + left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND); } } @@ -4626,20 +4798,21 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } if (changed) { - ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); /* create a new compare */ - n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right); - proj = new_rd_Proj(get_irn_dbg_info(proj), block, n, get_irn_mode(proj), proj_nr); + n = new_rd_Cmp(dbgi, block, left, right, relation); } - return proj; -} /* transform_node_Proj_Cmp */ + return n; +} /** * Optimize CopyB(mem, x, x) into a Nop. */ -static ir_node *transform_node_Proj_CopyB(ir_node *proj) { +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); @@ -4651,21 +4824,24 @@ static ir_node *transform_node_Proj_CopyB(ir_node *proj) { DBG_OPT_EXC_REM(proj); proj = new_r_Jmp(get_nodes_block(copyb)); break; - case pn_CopyB_X_except: + case pn_CopyB_X_except: { + ir_graph *irg = get_irn_irg(proj); DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(get_irn_irg(proj)); + proj = new_r_Bad(irg, mode_X); 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) { +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); @@ -4676,374 +4852,216 @@ static ir_node *transform_node_Proj_Bound(ir_node *proj) { 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. + * idx was Bounds checked 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)) { + 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) */ - 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(get_irn_irg(proj)); - 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(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. - */ -static ir_node *transform_node_Proj(ir_node *proj) { - ir_node *n = get_Proj_pred(proj); - - if (n->op->ops.transform_node_Proj) - return n->op->ops.transform_node_Proj(proj); - return proj; -} /* transform_node_Proj */ - -/** - * Move Confirms down through Phi nodes. - */ -static ir_node *transform_node_Phi(ir_node *phi) { - int i, n; - ir_mode *mode = get_irn_mode(phi); - - if (mode_is_reference(mode)) { - n = get_irn_arity(phi); - - /* Beware of Phi0 */ - if (n > 0) { - ir_node *pred = get_irn_n(phi, 0); - ir_node *bound, *new_Phi, *block, **in; - pn_Cmp pnc; - - if (! is_Confirm(pred)) - return phi; - - bound = get_Confirm_bound(pred); - pnc = get_Confirm_cmp(pred); - - NEW_ARR_A(ir_node *, in, n); - in[0] = get_Confirm_value(pred); - - for (i = 1; i < n; ++i) { - pred = get_irn_n(phi, i); - - if (! is_Confirm(pred) || - get_Confirm_bound(pred) != bound || - get_Confirm_cmp(pred) != pnc) - return phi; - in[i] = get_Confirm_value(pred); - } - /* move the Confirm nodes "behind" the Phi */ - block = get_irn_n(phi, -1); - new_Phi = new_r_Phi(block, n, in, get_irn_mode(phi)); - return new_r_Confirm(block, new_Phi, bound, pnc); - } - } - return phi; -} /* transform_node_Phi */ - -/** - * Returns the operands of a commutative bin-op, if one operand is - * a const, it is returned as the second one. - */ -static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) { - ir_node *op_a = get_binop_left(binop); - ir_node *op_b = get_binop_right(binop); - - assert(is_op_commutative(get_irn_op(binop))); - - if (is_Const(op_a)) { - *a = op_b; - *c = op_a; - } else { - *a = op_a; - *c = op_b; - } -} /* get_comm_Binop_Ops */ - -/** - * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. - * Such pattern may arise in bitfield stores. - * - * value c4 value c4 & c2 - * AND c3 AND c1 | c3 - * OR c2 ===> OR - * AND c1 - * OR - * - * - * value c2 value c1 - * AND c1 ===> OR if (c1 | c2) == 0x111..11 - * OR - */ -static ir_node *transform_node_Or_bf_store(ir_node *or) { - ir_node *and, *c1; - ir_node *or_l, *c2; - ir_node *and_l, *c3; - ir_node *value, *c4; - ir_node *new_and, *new_const, *block; - ir_mode *mode = get_irn_mode(or); - - tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2; - - while (1) { - get_comm_Binop_Ops(or, &and, &c1); - if (!is_Const(c1) || !is_And(and)) - return or; - - get_comm_Binop_Ops(and, &or_l, &c2); - if (!is_Const(c2)) - return or; - - tv1 = get_Const_tarval(c1); - tv2 = get_Const_tarval(c2); - - tv = tarval_or(tv1, tv2); - if (tarval_is_all_one(tv)) { - /* the AND does NOT clear a bit with isn't set by the OR */ - set_Or_left(or, or_l); - set_Or_right(or, c1); - - /* check for more */ - continue; - } - - if (!is_Or(or_l)) - return or; - - get_comm_Binop_Ops(or_l, &and_l, &c3); - if (!is_Const(c3) || !is_And(and_l)) - return or; - - get_comm_Binop_Ops(and_l, &value, &c4); - if (!is_Const(c4)) - return or; - - /* ok, found the pattern, check for conditions */ - assert(mode == get_irn_mode(and)); - assert(mode == get_irn_mode(or_l)); - assert(mode == get_irn_mode(and_l)); - - tv3 = get_Const_tarval(c3); - tv4 = get_Const_tarval(c4); - - tv = tarval_or(tv4, tv2); - if (!tarval_is_all_one(tv)) { - /* have at least one 0 at the same bit position */ - return or; - } - - n_tv4 = tarval_not(tv4); - if (tv3 != tarval_and(tv3, n_tv4)) { - /* bit in the or_mask is outside the and_mask */ - return or; + * 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; } - - n_tv2 = tarval_not(tv2); - if (tv1 != tarval_and(tv1, n_tv2)) { - /* bit in the or_mask is outside the and_mask */ - return or; + } + 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 = new_r_Bad(get_irn_irg(proj), mode_X); + 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(get_nodes_block(bound)); + break; + default: + break; } - - /* ok, all conditions met */ - block = get_irn_n(or, -1); - - new_and = new_r_And(block, value, new_Const(tarval_and(tv4, tv2)), mode); - - new_const = new_Const(tarval_or(tv3, tv1)); - - set_Or_left(or, new_and); - set_Or_right(or, new_const); - - /* check for more */ } -} /* transform_node_Or_bf_store */ + return proj; +} /** - * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl + * Does all optimizations on nodes that must be done on its Projs + * because of creating new nodes. */ -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, *rotval; - tarval *tv1, *tv2; - - if (! mode_is_int(mode)) - return or; - - shl = get_binop_left(or); - shr = get_binop_right(or); - - if (is_Shr(shl)) { - if (!is_Shl(shr)) - return or; - - irn = shl; - shl = shr; - shr = irn; - } else if (!is_Shl(shl)) { - return or; - } else if (!is_Shr(shr)) { - return or; - } - x = get_Shl_left(shl); - if (x != get_Shr_left(shr)) - return or; - - c1 = get_Shl_right(shl); - c2 = get_Shr_right(shr); - if (is_Const(c1) && is_Const(c2)) { - tv1 = get_Const_tarval(c1); - if (! tarval_is_long(tv1)) - return or; +static ir_node *transform_node_Proj(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); - tv2 = get_Const_tarval(c2); - if (! tarval_is_long(tv2)) - return or; + if (n->op->ops.transform_node_Proj) + return n->op->ops.transform_node_Proj(proj); + return proj; +} - if (get_tarval_long(tv1) + get_tarval_long(tv2) - != (int) get_mode_size_bits(mode)) - return or; +/** + * Test whether a block is unreachable + * Note: That this only returns true when + * IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE is set. + * This is important, as you easily end up producing invalid constructs in the + * unreachable code when optimizing away edges into the unreachable code. + * So only set this flag when you iterate localopts to the fixpoint. + * When you reach the fixpoint then all unreachable code is dead + * (= can't be reached by firm edges) and you won't see the invalid constructs + * anymore. + */ +static bool is_block_unreachable(const ir_node *block) +{ + const ir_graph *irg = get_irn_irg(block); + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE)) + return false; + return get_Block_dom_depth(block) < 0; +} - /* yet, condition met */ - block = get_nodes_block(or); +static ir_node *transform_node_Block(ir_node *block) +{ + ir_graph *irg = get_irn_irg(block); + int arity = get_irn_arity(block); + ir_node *bad = NULL; + int i; - n = new_r_Rotl(block, x, c1, mode); + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_OPTIMIZE_UNREACHABLE_CODE)) + return block; - DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); - return n; + for (i = 0; i < arity; ++i) { + ir_node *const pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred))) + continue; + if (bad == NULL) + bad = new_r_Bad(irg, mode_X); + set_irn_n(block, i, bad); } - 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; - - if (get_Sub_right(sub) != v) - return or; - - c1 = get_Sub_left(sub); - if (!is_Const(c1)) - return or; + return block; +} - tv1 = get_Const_tarval(c1); - if (! tarval_is_long(tv1)) - return or; +static ir_node *transform_node_Phi(ir_node *phi) +{ + int n = get_irn_arity(phi); + ir_mode *mode = get_irn_mode(phi); + ir_node *block = get_nodes_block(phi); + ir_graph *irg = get_irn_irg(phi); + ir_node *bad = NULL; + int i; + + /* Set phi-operands for bad-block inputs to bad */ + for (i = 0; i < n; ++i) { + if (!is_Bad(get_Phi_pred(phi, i))) { + ir_node *pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) { + if (bad == NULL) + bad = new_r_Bad(irg, mode); + set_irn_n(phi, i, bad); + } + } + } - if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode)) - return or; + /* Move Pin nodes down through Phi nodes. */ + if (mode == mode_M) { + n = get_irn_arity(phi); - /* yet, condition met */ - block = get_nodes_block(or); + /* Beware of Phi0 */ + if (n > 0) { + ir_node **in; + ir_node *new_phi; + bool has_pin = false; - n = new_r_Rotl(block, x, rotval, mode); + NEW_ARR_A(ir_node *, in, n); - DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL); - return n; -} /* transform_node_Or_Rotl */ + for (i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); -/** - * Transform an Or. - */ -static ir_node *transform_node_Or(ir_node *n) { - ir_node *c, *oldn = n; - ir_node *a = get_Or_left(n); - ir_node *b = get_Or_right(n); - ir_mode *mode; + if (is_Pin(pred)) { + in[i] = get_Pin_op(pred); + has_pin = true; + } else if (is_Bad(pred)) { + in[i] = pred; + } else { + return phi; + } + } - if (is_Not(a) && is_Not(b)) { - /* ~a | ~b = ~(a&b) */ - ir_node *block = get_nodes_block(n); + if (!has_pin) + return phi; - mode = get_irn_mode(n); - a = get_Not_op(a); - b = get_Not_op(b); - n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode); - n = new_rd_Not(get_irn_dbg_info(n), block, n, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); - return n; + /* Move the Pin nodes "behind" the Phi. */ + new_phi = new_r_Phi(block, n, in, mode_M); + return new_r_Pin(block, new_phi); + } } + /* Move Confirms down through Phi nodes. */ + else if (mode_is_reference(mode)) { + n = get_irn_arity(phi); - /* we can evaluate 2 Projs of the same Cmp */ - if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if (pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(pred_a); - pn_Cmp pn_a = get_Proj_proj(a); - pn_Cmp pn_b = get_Proj_proj(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a | pn_b; + /* Beware of Phi0 */ + if (n > 0) { + ir_node *pred = get_irn_n(phi, 0); + ir_node *bound, *new_phi, **in; + ir_relation relation; + bool has_confirm = false; - return new_rd_Proj(dbgi, block, pred_a, mode_b, new_pnc); - } - } + if (! is_Confirm(pred)) + return phi; - mode = get_irn_mode(n); - HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode); + bound = get_Confirm_bound(pred); + relation = get_Confirm_relation(pred); - n = transform_node_Or_bf_store(n); - n = transform_node_Or_Rotl(n); - if (n != oldn) - return n; + NEW_ARR_A(ir_node *, in, n); + in[0] = get_Confirm_value(pred); - n = transform_bitwise_distributive(n, transform_node_Or); + for (i = 1; i < n; ++i) { + pred = get_irn_n(phi, i); - return n; -} /* transform_node_Or */ + if (is_Confirm(pred) && + get_Confirm_bound(pred) == bound && + get_Confirm_relation(pred) == relation) { + in[i] = get_Confirm_value(pred); + has_confirm = true; + } else if (is_Bad(pred)) { + in[i] = pred; + } else { + return phi; + } + } + if (!has_confirm) + return phi; -/* forward */ -static ir_node *transform_node(ir_node *n); + /* move the Confirm nodes "behind" the Phi */ + new_phi = new_r_Phi(block, n, in, get_irn_mode(phi)); + return new_r_Confirm(block, new_phi, bound, relation); + } + } + return phi; +} /** * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl. * * Should be moved to reassociation? */ -static ir_node *transform_node_shift(ir_node *n) { +static ir_node *transform_node_shift(ir_node *n) +{ ir_node *left, *right; ir_mode *mode; - tarval *tv1, *tv2, *res; + ir_mode *count_mode; + ir_tarval *tv1, *tv2, *res; ir_node *in[2], *irn, *block; + ir_graph *irg; + int modulo_shf; left = get_binop_left(n); @@ -5052,7 +5070,7 @@ static ir_node *transform_node_shift(ir_node *n) { return n; right = get_binop_right(n); - tv1 = value_of(right); + tv1 = value_of(right); if (tv1 == tarval_bad) return n; @@ -5060,128 +5078,65 @@ static ir_node *transform_node_shift(ir_node *n) { if (tv2 == tarval_bad) return n; - res = tarval_add(tv1, tv2); - mode = get_irn_mode(n); + count_mode = get_tarval_mode(tv1); + if (get_tarval_mode(tv2) != count_mode) { + /* TODO: search bigger mode or something and convert... */ + return n; + } - /* beware: a simple replacement works only, if res < modulo shift */ - if (!is_Rotl(n)) { - int modulo_shf = get_mode_modulo_shift(mode); - if (modulo_shf > 0) { - tarval *modulo = new_tarval_from_long(modulo_shf, - get_tarval_mode(res)); + mode = get_irn_mode(n); + modulo_shf = get_mode_modulo_shift(mode); - assert(modulo_shf >= (int) get_mode_size_bits(mode)); + if (modulo_shf > 0) { + ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode); - /* shifting too much */ - if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) { - if (is_Shrs(n)) { - ir_node *block = get_nodes_block(n); - dbg_info *dbgi = get_irn_dbg_info(n); - ir_mode *smode = get_irn_mode(right); - ir_node *cnst = new_Const_long(smode, get_mode_size_bits(mode) - 1); - return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode); - } + /* I'm not so sure what happens in one complement... */ + assert(get_mode_arithmetic(count_mode) == irma_twos_complement); + /* modulo shifts should always be a power of 2 (otherwise modulo_mask + * above will be invalid) */ + assert(modulo_shf<=0 || is_po2(modulo_shf)); + + tv1 = tarval_and(tv1, modulo_mask); + tv2 = tarval_and(tv2, modulo_mask); + } + res = tarval_add(tv1, tv2); + irg = get_irn_irg(n); + + /* beware: a simple replacement works only, if res < modulo shift */ + if (is_Rotl(n)) { + int bits = get_mode_size_bits(mode); + ir_tarval *modulo = new_tarval_from_long(bits, count_mode); + res = tarval_mod(res, modulo); + } else { + long bits = get_mode_size_bits(mode); + ir_tarval *mode_size = new_tarval_from_long(bits, count_mode); - return new_Const(get_mode_null(mode)); + /* shifting too much */ + if (!(tarval_cmp(res, mode_size) & ir_relation_less)) { + if (is_Shrs(n)) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *smode = get_irn_mode(right); + ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1); + return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode); } + + return new_r_Const(irg, 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 */ + assert(modulo_shf >= (int) get_mode_size_bits(mode)); block = get_nodes_block(n); in[0] = get_binop_left(left); - in[1] = new_Const(res); + in[1] = new_r_Const(irg, res); irn = new_ir_node(NULL, get_Block_irg(block), 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_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); - - 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, block, bitop_left, right, mode); - tv_shift = tarval_shl(tv1, tv2); - } else if(is_Shr(n)) { - new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shr(tv1, tv2); - } else if(is_Shrs(n)) { - new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shrs(tv1, tv2); - } else { - assert(is_Rotl(n)); - new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_rotl(tv1, tv2); - } - - assert(get_tarval_mode(tv_shift) == mode); - new_const = new_Const(tv_shift); - - if (op_left == op_And) { - new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); - } else if(op_left == op_Or) { - new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); - } else { - assert(op_left == op_Eor); - new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); - } - - return new_bitop; + return irn; } /** @@ -5191,23 +5146,24 @@ static ir_node *transform_node_bitop_shift(ir_node *n) { * (x >> c1) << c2 <=> x OP (c2-c1) & ((-1 >> c1) << c2) * (also with x >>s c1 when c1>=c2) */ -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; +static ir_node *transform_node_shl_shr(ir_node *n) +{ + ir_node *left; + ir_node *right = get_binop_right(n); + ir_node *x; + ir_node *block; + ir_mode *mode; + dbg_info *dbgi; + ir_node *new_const; + ir_node *new_shift; + ir_node *new_and; + ir_tarval *tv_shl; + ir_tarval *tv_shr; + ir_tarval *tv_shift; + ir_tarval *tv_mask; + ir_graph *irg; + ir_relation relation; + int need_shrs = 0; assert(is_Shl(n) || is_Shr(n) || is_Shrs(n)); @@ -5228,7 +5184,7 @@ static ir_node *transform_node_shl_shr(ir_node *n) { if (is_Shrs(left)) { /* shrs variant only allowed if c1 >= c2 */ - if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge)) + if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal)) return n; tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr); @@ -5237,7 +5193,7 @@ static ir_node *transform_node_shl_shr(ir_node *n) { 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)) { + } else if (is_Shr(n) && is_Shl(left)) { ir_node *shl_right = get_Shl_right(left); if (!is_Const(shl_right)) @@ -5260,36 +5216,122 @@ static ir_node *transform_node_shl_shr(ir_node *n) { assert(tv_mask != tarval_bad); assert(get_tarval_mode(tv_mask) == mode); - irg = get_irn_irg(n); block = get_nodes_block(n); + irg = get_irn_irg(block); dbgi = get_irn_dbg_info(n); - pnc = tarval_cmp(tv_shl, tv_shr); - if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) { + relation = tarval_cmp(tv_shl, tv_shr); + if (relation == ir_relation_less || relation == ir_relation_equal) { tv_shift = tarval_sub(tv_shr, tv_shl, NULL); - new_const = new_Const(tv_shift); + new_const = new_r_Const(irg, tv_shift); if (need_shrs) { new_shift = new_rd_Shrs(dbgi, block, x, new_const, mode); } else { new_shift = new_rd_Shr(dbgi, block, x, new_const, mode); } } else { - assert(pnc == pn_Cmp_Gt); + assert(relation == ir_relation_greater); tv_shift = tarval_sub(tv_shl, tv_shr, NULL); - new_const = new_Const(tv_shift); + new_const = new_r_Const(irg, tv_shift); new_shift = new_rd_Shl(dbgi, block, x, new_const, mode); } - new_const = new_Const(tv_mask); + new_const = new_r_Const(irg, tv_mask); new_and = new_rd_And(dbgi, block, new_shift, new_const, mode); - return new_and; + return new_and; +} + +static ir_tarval *get_modulo_tv_value(ir_tarval *tv, int modulo_val) +{ + ir_mode *mode = get_tarval_mode(tv); + ir_tarval *modulo_tv = new_tarval_from_long(modulo_val, mode); + return tarval_mod(tv, modulo_tv); +} + +typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block, + ir_node *left, ir_node *right, ir_mode *mode); + +/** + * Normalisation: if we have a shl/shr with modulo_shift behaviour + * then we can use that to minimize the value of Add(x, const) or + * Sub(Const, x). In particular this often avoids 1 instruction in some + * backends for the Shift(x, Sub(Const, y)) case because it can be replaced + * by Shift(x, Minus(y)) which does not need an explicit Const constructed. + */ +static ir_node *transform_node_shift_modulo(ir_node *n, + new_shift_func new_shift) +{ + ir_mode *mode = get_irn_mode(n); + int modulo = get_mode_modulo_shift(mode); + ir_node *newop = NULL; + ir_mode *mode_right; + ir_node *block; + ir_node *right; + ir_graph *irg; + + if (modulo == 0) + return n; + if (get_mode_arithmetic(mode) != irma_twos_complement) + return n; + if (!is_po2(modulo)) + return n; + + irg = get_irn_irg(n); + block = get_nodes_block(n); + right = get_binop_right(n); + mode_right = get_irn_mode(right); + if (is_Const(right)) { + ir_tarval *tv = get_Const_tarval(right); + ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo); + + if (tv_mod == tv) + return n; + + newop = new_r_Const(irg, tv_mod); + } else if (is_Add(right) || is_Or_Eor_Add(right)) { + ir_node *add_right = get_binop_right(right); + if (is_Const(add_right)) { + ir_tarval *tv = get_Const_tarval(add_right); + ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo); + ir_node *newconst; + if (tv_mod == tv) + return n; + + newconst = new_r_Const(irg, tv_mod); + newop = new_r_Add(block, get_binop_left(right), newconst, + mode_right); + } + } else if (is_Sub(right)) { + ir_node *sub_left = get_Sub_left(right); + if (is_Const(sub_left)) { + ir_tarval *tv = get_Const_tarval(sub_left); + ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo); + ir_node *newconst; + if (tv_mod == tv) + return n; + + newconst = new_r_Const(irg, tv_mod); + newop = new_r_Sub(block, newconst, get_Sub_right(right), + mode_right); + } + } else { + return n; + } + + if (newop != NULL) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *left = get_binop_left(n); + return new_shift(dbgi, block, left, newop, mode); + } + return n; } /** * Transform a Shr. */ -static ir_node *transform_node_Shr(ir_node *n) { +static ir_node *transform_node_Shr(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *left = get_Shr_left(n); ir_node *right = get_Shr_right(n); @@ -5298,36 +5340,71 @@ static ir_node *transform_node_Shr(ir_node *n) { HANDLE_BINOP_PHI((eval_func) tarval_shr, left, right, c, mode); n = transform_node_shift(n); + if (is_Shr(n)) + n = transform_node_shift_modulo(n, new_rd_Shr); if (is_Shr(n)) n = transform_node_shl_shr(n); if (is_Shr(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; -} /* transform_node_Shr */ +} /** * Transform a Shrs. */ -static ir_node *transform_node_Shrs(ir_node *n) { - ir_node *c, *oldn = n; - ir_node *a = get_Shrs_left(n); - ir_node *b = get_Shrs_right(n); - ir_mode *mode = get_irn_mode(n); +static ir_node *transform_node_Shrs(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_Shrs_left(n); + ir_node *b = get_Shrs_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *c; + vrp_attr *attr; + + if (is_oversize_shift(n)) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *cmode = get_irn_mode(b); + long val = get_mode_size_bits(cmode)-1; + ir_graph *irg = get_irn_irg(n); + ir_node *cnst = new_r_Const_long(irg, cmode, val); + return new_rd_Shrs(dbgi, block, a, cnst, mode); + } HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode); n = transform_node_shift(n); + if (n != oldn) + return n; - if (is_Shrs(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_modulo(n, new_rd_Shrs); + if (n != oldn) + return n; + n = transform_node_shift_bitop(n); + if (n != oldn) + return n; + + /* normalisation: use Shr when sign bit is guaranteed to be cleared */ + attr = vrp_get_info(a); + if (attr != NULL) { + unsigned bits = get_mode_size_bits(mode); + ir_tarval *scount = new_tarval_from_long(bits-1, mode_Iu); + ir_tarval *sign = tarval_shl(get_mode_one(mode), scount); + if (tarval_is_null(tarval_and(attr->bits_not_set, sign))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_Shr(dbgi, block, a, b, mode); + } + } return n; -} /* transform_node_Shrs */ +} /** * Transform a Shl. */ -static ir_node *transform_node_Shl(ir_node *n) { +static ir_node *transform_node_Shl(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *a = get_Shl_left(n); ir_node *b = get_Shl_right(n); @@ -5336,18 +5413,21 @@ static ir_node *transform_node_Shl(ir_node *n) { HANDLE_BINOP_PHI((eval_func) tarval_shl, a, b, c, mode); n = transform_node_shift(n); + if (is_Shl(n)) + n = transform_node_shift_modulo(n, new_rd_Shl); if (is_Shl(n)) n = transform_node_shl_shr(n); if (is_Shl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; -} /* transform_node_Shl */ +} /** * Transform a Rotl. */ -static ir_node *transform_node_Rotl(ir_node *n) { +static ir_node *transform_node_Rotl(ir_node *n) +{ ir_node *c, *oldn = n; ir_node *a = get_Rotl_left(n); ir_node *b = get_Rotl_right(n); @@ -5357,19 +5437,102 @@ static ir_node *transform_node_Rotl(ir_node *n) { n = transform_node_shift(n); if (is_Rotl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; -} /* transform_node_Rotl */ +} + +/** + * returns mode size for may_leave_out_middle_mode + */ +static unsigned get_significand_size(ir_mode *mode) +{ + const ir_mode_arithmetic arithmetic = get_mode_arithmetic(mode); + switch (arithmetic) { + case irma_ieee754: + case irma_x86_extended_float: + return get_mode_mantissa_size(mode) + 1; + case irma_twos_complement: + return get_mode_size_bits(mode); + case irma_none: + panic("Conv node with irma_none mode?"); + } + panic("unexpected mode_arithmetic in get_significand_size"); +} + +/** + * Returns true if a conversion from mode @p m0 to @p m1 has the same effect + * as converting from @p m0 to @p m1 and then to @p m2. + * Classifying the 3 modes as the big(b), middle(m) and small(s) mode this + * gives the following truth table: + * s -> b -> m : true + * s -> m -> b : !signed(s) || signed(m) + * m -> b -> s : true + * m -> s -> b : false + * b -> s -> m : false + * b -> m -> s : true + * + * s -> b -> b : true + * s -> s -> b : false + * + * additional float constraints: + * F -> F -> F: fine + * F -> I -> I: signedness of Is must match + * I -> F -> I: signedness of Is must match + * I -> I -> F: signedness of Is must match + * F -> I -> F: bad + * I -> F -> F: fine + * F -> F -> I: fine + * at least 1 float involved: signedness must match + */ +bool may_leave_out_middle_conv(ir_mode *m0, ir_mode *m1, ir_mode *m2) +{ + int n_floats = mode_is_float(m0) + mode_is_float(m1) + mode_is_float(m2); + if (n_floats == 1) { +#if 0 + int n_signed = mode_is_signed(m0) + mode_is_signed(m1) + + mode_is_signed(m2); + /* we assume that float modes are always signed */ + if ((n_signed & 1) != 1) + return false; +#else + /* because overflow gives strange results we don't touch this case */ + return false; +#endif + } else if (n_floats == 2 && !mode_is_float(m1)) { + return false; + } + + unsigned size0 = get_significand_size(m0); + unsigned size1 = get_significand_size(m1); + unsigned size2 = get_significand_size(m2); + if (size1 < size2 && size0 >= size1) + return false; + if (size1 >= size2) + return true; + return !mode_is_signed(m0) || mode_is_signed(m1); +} /** * Transform a Conv. */ -static ir_node *transform_node_Conv(ir_node *n) { +static ir_node *transform_node_Conv(ir_node *n) +{ ir_node *c, *oldn = n; ir_mode *mode = get_irn_mode(n); ir_node *a = get_Conv_op(n); + if (is_Conv(a)) { + ir_mode *a_mode = get_irn_mode(a); + ir_node *b = get_Conv_op(a); + ir_mode *b_mode = get_irn_mode(b); + if (may_leave_out_middle_conv(b_mode, a_mode, mode)) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_Conv(dbgi, block, b, mode); + } + } + if (mode != mode_b && is_const_Phi(a)) { /* Do NOT optimize mode_b Conv's, this leads to remaining * Phib nodes later, because the conv_b_lower operation @@ -5383,7 +5546,8 @@ static ir_node *transform_node_Conv(ir_node *n) { } if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */ - return new_r_Unknown(current_ir_graph, mode); + ir_graph *irg = get_irn_irg(n); + return new_r_Unknown(irg, mode); } if (mode_is_reference(mode) && @@ -5393,17 +5557,17 @@ static ir_node *transform_node_Conv(ir_node *n) { ir_node *r = get_Add_right(a); dbg_info *dbgi = get_irn_dbg_info(a); ir_node *block = get_nodes_block(n); - if(is_Conv(l)) { + if (is_Conv(l)) { ir_node *lop = get_Conv_op(l); - if(get_irn_mode(lop) == mode) { + if (get_irn_mode(lop) == mode) { /* ConvP(AddI(ConvI(P), x)) -> AddP(P, x) */ n = new_rd_Add(dbgi, block, lop, r, mode); return n; } } - if(is_Conv(r)) { + if (is_Conv(r)) { ir_node *rop = get_Conv_op(r); - if(get_irn_mode(rop) == mode) { + if (get_irn_mode(rop) == mode) { /* ConvP(AddI(x, ConvI(P))) -> AddP(x, P) */ n = new_rd_Add(dbgi, block, l, rop, mode); return n; @@ -5412,13 +5576,14 @@ static ir_node *transform_node_Conv(ir_node *n) { } return n; -} /* transform_node_Conv */ +} /** * Remove dead blocks and nodes in dead blocks * in keep alive list. We do not generate a new End node. */ -static ir_node *transform_node_End(ir_node *n) { +static ir_node *transform_node_End(ir_node *n) +{ int i, j, n_keepalives = get_End_n_keepalives(n); ir_node **in; @@ -5426,30 +5591,27 @@ static ir_node *transform_node_End(ir_node *n) { for (i = j = 0; i < n_keepalives; ++i) { ir_node *ka = get_End_keepalive(n, i); - if (is_Block(ka)) { - if (! is_Block_dead(ka)) { - in[j++] = ka; - } - continue; - } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) { + ir_node *block; + /* no need to keep Bad */ + if (is_Bad(ka)) continue; - } else if (is_Bad(ka)) { - /* no need to keep Bad */ + /* do not keep unreachable code */ + block = is_Block(ka) ? ka : get_nodes_block(ka); + if (is_block_unreachable(block)) continue; - } in[j++] = ka; } if (j != n_keepalives) set_End_keepalives(n, j, in); return n; -} /* transform_node_End */ +} -/** returns 1 if a == -b */ -static int is_negated_value(ir_node *a, ir_node *b) { +int ir_is_negated_value(const ir_node *a, const ir_node *b) +{ if (is_Minus(a) && get_Minus_op(a) == b) - return 1; + return true; if (is_Minus(b) && get_Minus_op(b) == a) - return 1; + return true; if (is_Sub(a) && is_Sub(b)) { ir_node *a_left = get_Sub_left(a); ir_node *a_right = get_Sub_right(a); @@ -5457,291 +5619,398 @@ static int is_negated_value(ir_node *a, ir_node *b) { ir_node *b_right = get_Sub_right(b); if (a_left == b_right && a_right == b_left) + return true; + } + + return false; +} + +static const ir_node *skip_upconv(const ir_node *node) +{ + while (is_Conv(node)) { + ir_mode *mode = get_irn_mode(node); + const ir_node *op = get_Conv_op(node); + ir_mode *op_mode = get_irn_mode(op); + if (!smaller_mode(op_mode, mode)) + break; + node = op; + } + return node; +} + +int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_false, + const ir_node *mux_true) +{ + ir_node *cmp_left; + ir_node *cmp_right; + ir_mode *mode; + ir_relation relation; + + if (!is_Cmp(sel)) + return 0; + + /** + * 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 Abs/-Abs + * transformations. + */ + mode = get_irn_mode(mux_true); + if (mode_honor_signed_zeros(mode)) + return 0; + + /* must be <, <=, >=, > */ + relation = get_Cmp_relation(sel); + if ((relation & ir_relation_less_greater) == 0) + return 0; + + if (!ir_is_negated_value(mux_true, mux_false)) + return 0; + + mux_true = skip_upconv(mux_true); + mux_false = skip_upconv(mux_false); + + /* must be x cmp 0 */ + cmp_right = get_Cmp_right(sel); + if (!is_Const(cmp_right) || !is_Const_null(cmp_right)) + return 0; + + cmp_left = get_Cmp_left(sel); + if (cmp_left == mux_false) { + if (relation & ir_relation_less) { + return 1; + } else { + assert(relation & ir_relation_greater); + return -1; + } + } else if (cmp_left == mux_true) { + if (relation & ir_relation_less) { + return -1; + } else { + assert(relation & ir_relation_greater); return 1; + } } return 0; } +ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_false, + ir_node *mux_true) +{ + ir_node *cmp_left = get_Cmp_left(sel); + return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true; +} + +bool ir_is_optimizable_mux(const ir_node *sel, const ir_node *mux_false, + const ir_node *mux_true) +{ + /* this code should return true each time transform_node_Mux would + * optimize the Mux completely away */ + + ir_mode *mode = get_irn_mode(mux_false); + if (get_mode_arithmetic(mode) == irma_twos_complement + && ir_mux_is_abs(sel, mux_false, mux_true)) + return true; + + if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) { + const ir_node *cmp_r = get_Cmp_right(sel); + const ir_node *cmp_l = get_Cmp_left(sel); + const ir_node *f = mux_false; + const ir_node *t = mux_true; + + if (is_Const(t) && is_Const_null(t)) { + t = mux_false; + f = mux_true; + } + + if (is_And(cmp_l) && f == cmp_r) { + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && is_single_bit(and_r)) + return true; + and_l = get_And_left(cmp_l); + if (and_l == t && is_single_bit(and_l)) + return true; + } + } + + return false; +} + /** - * Optimize a Mux into some simpler cases. + * Optimize a Mux(c, 0, 1) node (sometimes called a "set" instruction) */ -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; - - if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) +static ir_node *transform_Mux_set(ir_node *n) +{ + ir_node *cond = get_Mux_sel(n); + ir_mode *dest_mode; + ir_mode *mode; + ir_node *left; + ir_node *right; + ir_relation relation; + bool need_not; + dbg_info *dbgi; + ir_node *block; + ir_graph *irg; + ir_node *a; + ir_node *b; + unsigned bits; + ir_tarval *tv; + ir_node *shift_cnt; + ir_node *res; + + if (!is_Cmp(cond)) + return n; + left = get_Cmp_left(cond); + mode = get_irn_mode(left); + if (!mode_is_int(mode) && !mode_is_reference(mode)) return n; + dest_mode = get_irn_mode(n); + if (!mode_is_int(dest_mode) && !mode_is_reference(dest_mode)) + return n; + right = get_Cmp_right(cond); + relation = get_Cmp_relation(cond) & ~ir_relation_unordered; + if (get_mode_size_bits(mode) >= get_mode_size_bits(dest_mode) + && !(mode_is_signed(mode) && is_Const(right) && is_Const_null(right) + && relation != ir_relation_greater)) + return n; + + need_not = false; + switch (relation) { + case ir_relation_less: + /* a < b -> (a - b) >> 31 */ + a = left; + b = right; + break; + case ir_relation_less_equal: + /* a <= b -> ~(a - b) >> 31 */ + a = right; + b = left; + need_not = true; + break; + case ir_relation_greater: + /* a > b -> (b - a) >> 31 */ + a = right; + b = left; + break; + case ir_relation_greater_equal: + /* a >= b -> ~(a - b) >> 31 */ + a = left; + b = right; + need_not = true; + break; + default: + return n; + } - if (is_Mux(t)) { - ir_node* block = get_nodes_block(n); - ir_node* c0 = sel; - ir_node* c1 = get_Mux_sel(t); - ir_node* t1 = get_Mux_true(t); - ir_node* f1 = get_Mux_false(t); - if (f == f1) { - /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */ - ir_node* and_ = new_r_And(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode); - n = new_mux; - sel = and_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); - } else if (f == t1) { - /* Mux(cond0, Mux(cond1, x, y), x) */ - ir_node* not_c1 = new_r_Not(block, c1, mode_b); - ir_node* and_ = new_r_And(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode); - n = new_mux; - sel = and_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); - } - } else if (is_Mux(f)) { - ir_node* block = get_nodes_block(n); - ir_node* c0 = sel; - ir_node* c1 = get_Mux_sel(f); - ir_node* t1 = get_Mux_true(f); - ir_node* f1 = get_Mux_false(f); - if (t == t1) { - /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */ - ir_node* or_ = new_r_Or(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode); - n = new_mux; - sel = or_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); - } else if (t == f1) { - /* Mux(cond0, x, Mux(cond1, y, x)) */ - ir_node* not_c1 = new_r_Not(block, c1, mode_b); - ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode); - n = new_mux; - sel = or_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); - } - } - - /* 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(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), get_nodes_block(n), sel, t, f, mode); - tmp = t; - t = f; - f = tmp; - } - } + dbgi = get_irn_dbg_info(n); + block = get_nodes_block(n); + irg = get_irn_irg(block); + bits = get_mode_size_bits(dest_mode); + tv = new_tarval_from_long(bits-1, mode_Iu); + shift_cnt = new_rd_Const(dbgi, irg, tv); + + if (mode != dest_mode) { + a = new_rd_Conv(dbgi, block, a, dest_mode); + b = new_rd_Conv(dbgi, block, b, dest_mode); } - /* note: after normalization, false can only happen on default */ - if (mode == mode_b) { - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); + res = new_rd_Sub(dbgi, block, a, b, dest_mode); + if (need_not) { + res = new_rd_Not(dbgi, block, res, dest_mode); + } + res = new_rd_Shr(dbgi, block, res, shift_cnt, dest_mode); + return res; +} - 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 { - /* Muxb(sel, true, x) = Or(sel, x) */ - n = new_rd_Or(dbg, 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, block, sel, mode_b); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); - n = new_rd_Or(dbg, block, not_sel, t, mode_b); - return n; +/** + * Optimize a Mux into some simpler cases. + */ +static ir_node *transform_node_Mux(ir_node *n) +{ + ir_node *oldn = n; + ir_node *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 = get_irn_irg(n); + + /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */ + if (get_mode_arithmetic(mode) == irma_twos_complement) { + int abs = ir_mux_is_abs(sel, f, t); + if (abs != 0) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *op = ir_get_abs_op(sel, f, t); + int bits = get_mode_size_bits(mode); + ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1); + ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode); + ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode); + ir_node *res; + if (abs > 0) { + res = new_rd_Sub(dbgi, block, xorn, sext, mode); } else { - /* Muxb(sel, x, false) = And(sel, x) */ - assert(tv_f == tarval_b_false); - n = new_rd_And(dbg, block, sel, t, mode_b); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); - return n; + res = new_rd_Sub(dbgi, block, sext, xorn, mode); } + return res; } } - /* 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 *diff, *min; + /* first normalization step: try to move a constant to the false side, + * 0 preferred on false side too */ + if (is_Cmp(sel) && is_Const(t) && + (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) { + dbg_info *seldbgi = get_irn_dbg_info(sel); + ir_node *block = get_nodes_block(sel); + ir_relation relation = get_Cmp_relation(sel); + ir_node *tmp = t; + t = f; + f = tmp; - if (tarval_is_one(a) && tarval_is_null(b)) { - ir_node *block = get_nodes_block(n); - ir_node *conv = new_r_Conv(block, sel, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); - return n; - } else if (tarval_is_null(a) && tarval_is_one(b)) { - ir_node *block = get_nodes_block(n); - ir_node *not_ = new_r_Not(block, sel, mode_b); - ir_node *conv = new_r_Conv(block, not_, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); + /* Mux(x, a, b) => Mux(not(x), b, a) */ + relation = get_negated_relation(relation); + sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel), + get_Cmp_right(sel), relation); + return new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); + } + + if (is_Const(f) && is_Const_null(f) && is_Const(t) && is_Const_one(t)) { + n = transform_Mux_set(n); + if (n != oldn) return n; - } - /* TODO: it's not really clear if that helps in general or should be moved - * to backend, especially with the MUX->Conv transformation above */ - 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; + } + + /* the following optimisations create new mode_b nodes, so only do them + * before mode_b lowering */ + if (!irg_is_constrained(irg, IR_GRAPH_CONSTRAINT_MODEB_LOWERED)) { + if (is_Mux(t)) { + ir_node* block = get_nodes_block(n); + ir_node* c0 = sel; + ir_node* c1 = get_Mux_sel(t); + ir_node* t1 = get_Mux_true(t); + ir_node* f1 = get_Mux_false(t); + if (f == f1) { + /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */ + ir_node* and_ = new_r_And(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, f1, t1, mode); + } else if (f == t1) { + /* Mux(cond0, Mux(cond1, x, y), x) */ + ir_node* not_c1 = new_r_Not(block, c1, mode_b); + ir_node* and_ = new_r_And(block, c0, not_c1, mode_b); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, t1, f1, mode); + } + } else if (is_Mux(f)) { + ir_node* block = get_nodes_block(n); + ir_node* c0 = sel; + ir_node* c1 = get_Mux_sel(f); + ir_node* t1 = get_Mux_true(f); + ir_node* f1 = get_Mux_false(f); + if (t == t1) { + /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */ + ir_node* or_ = new_r_Or(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, f1, t1, mode); + } else if (t == f1) { + /* Mux(cond0, x, Mux(cond1, y, x)) */ + ir_node* not_c1 = new_r_Not(block, c1, mode_b); + ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, t1, f1, mode); + } } - if (diff == get_tarval_one(mode)) { + /* note: after normalization, false can only happen on default */ + if (mode == mode_b) { dbg_info *dbg = get_irn_dbg_info(n); ir_node *block = get_nodes_block(n); - ir_node *t = new_Const(tarval_sub(a, min, NULL)); - ir_node *f = new_Const(tarval_sub(b, min, NULL)); - n = new_rd_Mux(dbg, block, sel, f, t, mode); - n = new_rd_Add(dbg, block, n, new_Const(min), mode); - return n; + + if (is_Const(t)) { + ir_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 { + /* Muxb(sel, true, x) = Or(sel, x) */ + n = new_rd_Or(dbg, block, sel, f, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL); + return n; + } + } + } else if (is_Const(f)) { + ir_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, block, sel, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); + n = new_rd_Or(dbg, block, not_sel, t, mode_b); + return n; + } else { + /* Muxb(sel, x, false) = And(sel, x) */ + assert(tv_f == tarval_b_false); + n = new_rd_And(dbg, block, sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); + return n; + } + } } } - if (is_Proj(sel)) { - ir_node *cmp = get_Proj_pred(sel); - long pn = get_Proj_proj(sel); - - /* - * Note: normalization puts the constant on the right side, - * 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 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_nodes_block(n); - ir_node *cmp_l = get_Cmp_left(cmp); + if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) { + ir_relation relation = get_Cmp_relation(sel); + ir_node *cmp_r = get_Cmp_right(sel); + ir_node *cmp_l = get_Cmp_left(sel); + ir_node *block = get_nodes_block(n); - if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) { - /* f = -t */ + if (is_And(cmp_l) && f == cmp_r) { + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; - 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), block, cmp_l, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); - return n; - } 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), block, cmp_l, mode); - n = new_rd_Minus(get_irn_dbg_info(n), block, n, mode); - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); - return n; - } + if (and_r == t && is_single_bit(and_r)) { + if (relation == ir_relation_equal) { + /* Mux((a & (1<value_table, n); + add_identities(n); + return n; +} + +static ir_node *create_load_replacement_tuple(ir_node *n, ir_node *mem, + ir_node *res) +{ + ir_node *block = get_nodes_block(n); + ir_graph *irg = get_irn_irg(n); + ir_node *in[pn_Load_max+1]; + size_t n_in = 2; + in[pn_Load_M] = mem; + in[pn_Load_res] = res; + if (ir_throws_exception(n)) { + in[pn_Load_X_regular] = new_r_Jmp(block); + in[pn_Load_X_except] = new_r_Bad(irg, mode_X); + n_in = 4; + assert(pn_Load_max == 4); + } + ir_node *tuple = new_r_Tuple(block, n_in, in); + return tuple; +} + +static ir_node *transform_node_Load(ir_node *n) +{ + /* don't touch volatile loads */ + if (get_Load_volatility(n) == volatility_is_volatile) + return n; + + ir_node *ptr = get_Load_ptr(n); + const ir_node *confirm; + if (value_not_zero(ptr, &confirm) && confirm == NULL) { + set_irn_pinned(n, op_pin_state_floats); + } + + /* if our memory predecessor is a load from the same address, then reuse the + * previous result */ + ir_node *mem = get_Load_mem(n); + if (!is_Proj(mem)) + return n; + ir_node *mem_pred = get_Proj_pred(mem); + if (is_Load(mem_pred)) { + ir_node *pred_load = mem_pred; + + /* conservatively compare the 2 loads. TODO: This could be less strict + * with fixup code in some situations (like smaller/bigger modes) */ + if (get_Load_ptr(pred_load) != ptr) + return n; + if (get_Load_mode(pred_load) != get_Load_mode(n)) + return n; + /* all combinations of aligned/unaligned pred/n should be fine so we do + * not compare the unaligned attribute */ + ir_mode *mode = get_Load_mode(n); + ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res); + return create_load_replacement_tuple(n, mem, res); + } else if (is_Store(mem_pred)) { + ir_node *pred_store = mem_pred; + ir_node *value = get_Store_value(pred_store); + + if (get_Store_ptr(pred_store) != ptr) + return n; + if (get_irn_mode(value) != get_Load_mode(n)) + return n; + /* all combinations of aligned/unaligned pred/n should be fine so we do + * not compare the unaligned attribute */ + return create_load_replacement_tuple(n, mem, value); + } + + return n; +} + +static ir_node *transform_node_Store(ir_node *n) +{ + /* don't touch volatile stores */ + if (get_Store_volatility(n) == volatility_is_volatile) + return n; + ir_node *ptr = get_Store_ptr(n); + const ir_node *confirm; + if (value_not_zero(ptr, &confirm) && confirm == NULL) { + set_irn_pinned(n, op_pin_state_floats); + } return n; -} /* transform_node_Sync */ +} /** * optimize a trampoline Call into a direct Call */ -static ir_node *transform_node_Call(ir_node *call) { +static ir_node *transform_node_Call(ir_node *call) +{ ir_node *callee = get_Call_ptr(call); ir_node *adr, *mem, *res, *bl, **in; ir_type *ctp, *mtp, *tp; + ir_graph *irg; type_dbg_info *tdb; dbg_info *db; - int i, n_res, n_param; + size_t i, n_res, n_param; ir_variadicity var; if (! is_Proj(callee)) @@ -5821,7 +6199,8 @@ static ir_node *transform_node_Call(ir_node *call) { NEW_ARR_A(ir_node *, in, n_param + 1); /* FIXME: we don't need a new pointer type in every step */ - tp = get_irg_frame_type(current_ir_graph); + irg = get_irn_irg(call); + tp = get_irg_frame_type(irg); tp = new_type_pointer(tp); set_method_param_type(ctp, 0, tp); @@ -5832,9 +6211,6 @@ static ir_node *transform_node_Call(ir_node *call) { } var = get_method_variadicity(mtp); set_method_variadicity(ctp, var); - if (var == variadicity_variadic) { - set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1); - } /* When we resolve a trampoline, the function must be called by a this-call */ set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call); set_method_additional_properties(ctp, get_method_additional_properties(mtp)); @@ -5848,7 +6224,7 @@ static ir_node *transform_node_Call(ir_node *call) { if (get_irn_pinned(call) == op_pin_state_floats) set_irn_pinned(res, op_pin_state_floats); return res; -} /* transform_node_Call */ +} /** * Tries several [inplace] [optimizing] transformations and returns an @@ -5856,372 +6232,191 @@ static ir_node *transform_node_Call(ir_node *call) { * transformations _do_ generate new nodes, and thus the old node must * not be freed even if the equivalent node isn't the old one. */ -static ir_node *transform_node(ir_node *n) { - ir_node *oldn; - - /* - * Transform_node is the only "optimizing transformation" that might - * return a node with a different opcode. We iterate HERE until fixpoint - * to get the final result. - */ - do { - oldn = n; - if (n->op->ops.transform_node != NULL) - n = n->op->ops.transform_node(n); - } while (oldn != n); - - return n; -} /* transform_node */ - -/** - * Sets the default transform node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_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; \ - 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_PROJ_EX(Div); - CASE_PROJ_EX(Mod); - CASE_PROJ_EX(DivMod); - CASE(Quot); - CASE(Abs); - CASE_PROJ_EX(Cmp); - CASE_PROJ_EX(Cond); - CASE(And); - CASE(Eor); - 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(Rotl); - CASE(Conv); - CASE(End); - CASE(Mux); - CASE(Sync); - CASE(Call); - default: - /* leave NULL */; - } - - return ops; -#undef CASE_PROJ_EX -#undef CASE_PROJ -#undef CASE -} /* firm_set_default_transform_node */ - - -/* **************** Common Subexpression Elimination **************** */ +static ir_node *transform_node(ir_node *n) +{ + ir_node *old_n; + unsigned iro; +restart: + old_n = n; + iro = get_irn_opcode_(n); + /* constant expression evaluation / constant folding */ + if (get_opt_constant_folding()) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { + /* try to evaluate */ + ir_tarval *tv = computed_value(n); + if (tv != tarval_bad) { + /* evaluation was successful -- replace the node. */ + ir_graph *irg = get_irn_irg(n); -/** The size of the hash table used, should estimate the number of nodes - in a graph. */ -#define N_IR_NODES 512 + n = new_r_Const(irg, tv); -/** Compares the attributes of two Const nodes. */ -static int node_cmp_attr_Const(ir_node *a, ir_node *b) { - return (get_Const_tarval(a) != get_Const_tarval(b)) - || (get_Const_type(a) != get_Const_type(b)); -} /* node_cmp_attr_Const */ - -/** Compares the attributes of two Proj nodes. */ -static int node_cmp_attr_Proj(ir_node *a, ir_node *b) { - return get_irn_proj_attr(a) != get_irn_proj_attr(b); -} /* node_cmp_attr_Proj */ - -/** Compares the attributes of two Filter nodes. */ -static int node_cmp_attr_Filter(ir_node *a, ir_node *b) { - return get_Filter_proj(a) != get_Filter_proj(b); -} /* node_cmp_attr_Filter */ - -/** Compares the attributes of two Alloc nodes. */ -static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) { - const alloc_attr *pa = get_irn_alloc_attr(a); - const alloc_attr *pb = get_irn_alloc_attr(b); - return (pa->where != pb->where) || (pa->type != pb->type); -} /* node_cmp_attr_Alloc */ - -/** Compares the attributes of two Free nodes. */ -static int node_cmp_attr_Free(ir_node *a, ir_node *b) { - const free_attr *pa = get_irn_free_attr(a); - const free_attr *pb = get_irn_free_attr(b); - return (pa->where != pb->where) || (pa->type != pb->type); -} /* node_cmp_attr_Free */ - -/** Compares the attributes of two SymConst nodes. */ -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->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) { - const call_attr *pa = get_irn_call_attr(a); - const call_attr *pb = get_irn_call_attr(b); - return (pa->type != pb->type) - || (pa->tail_call != pb->tail_call); -} /* 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. */ -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 pos attribute */ - return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos; + DBG_OPT_CSTEVAL(old_n, n); + return n; + } + } } - return 0; -} /* node_cmp_attr_Phi */ - -/** Compares the attributes of two Conv nodes. */ -static int node_cmp_attr_Conv(ir_node *a, ir_node *b) { - return get_Conv_strict(a) != get_Conv_strict(b); -} /* node_cmp_attr_Conv */ - -/** Compares the attributes of two Cast nodes. */ -static int node_cmp_attr_Cast(ir_node *a, ir_node *b) { - return get_Cast_type(a) != get_Cast_type(b); -} /* node_cmp_attr_Cast */ - -/** Compares the attributes of two Load nodes. */ -static int node_cmp_attr_Load(ir_node *a, ir_node *b) { - if (get_Load_volatility(a) == volatility_is_volatile || - get_Load_volatility(b) == volatility_is_volatile) - /* NEVER do CSE on volatile Loads */ - return 1; - /* do not CSE Loads with different alignment. Be conservative. */ - if (get_Load_align(a) != get_Load_align(b)) - return 1; - return get_Load_mode(a) != get_Load_mode(b); -} /* node_cmp_attr_Load */ + /* remove unnecessary nodes */ + if (get_opt_constant_folding() || + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || /* ... */ + (iro == iro_Proj) || /* ... */ + (iro == iro_Block)) { /* Flags tested local. */ + n = equivalent_node(n); + if (n != old_n) + goto restart; + } -/** Compares the attributes of two Store nodes. */ -static int node_cmp_attr_Store(ir_node *a, ir_node *b) { - /* do not CSE Stores with different alignment. Be conservative. */ - if (get_Store_align(a) != get_Store_align(b)) - return 1; + /* Some more constant expression evaluation. */ + if (get_opt_algebraic_simplification() || + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ + if (n->op->ops.transform_node != NULL) { + n = n->op->ops.transform_node(n); + if (n != old_n) { + goto restart; + } + } + } - /* NEVER do CSE on volatile Stores */ - return (get_Store_volatility(a) == volatility_is_volatile || - 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->resmode != mb->resmode || - 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->resmode != mb->resmode; -} /* 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->resmode != mb->resmode; -} /* 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->resmode != mb->resmode; -} /* node_cmp_attr_Quot */ - -/** Compares the attributes of two Confirm nodes. */ -static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) { - /* no need to compare the bound, as this is a input */ - return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); -} /* node_cmp_attr_Confirm */ - -/** Compares the attributes of two Builtin nodes. */ -static int node_cmp_attr_Builtin(ir_node *a, ir_node *b) { - const builtin_attr *ma = get_irn_builtin_attr(a); - const builtin_attr *mb = get_irn_builtin_attr(b); - - /* no need to compare the type, equal kind means equal type */ - return ma->kind != mb->kind; -} /* node_cmp_attr_Builtin */ - -/** Compares the attributes of two ASM nodes. */ -static int node_cmp_attr_ASM(ir_node *a, ir_node *b) { - int i, n; - const ir_asm_constraint *ca; - const ir_asm_constraint *cb; - ident **cla, **clb; - - if (get_ASM_text(a) != get_ASM_text(b)) - return 1; + return n; +} - /* Should we really check the constraints here? Should be better, but is strange. */ - n = get_ASM_n_input_constraints(a); - if (n != get_ASM_n_input_constraints(b)) - return 0; +static void register_computed_value_func(ir_op *op, computed_value_func func) +{ + assert(op->ops.computed_value == NULL || op->ops.computed_value == func); + op->ops.computed_value = func; +} - ca = get_ASM_input_constraints(a); - cb = get_ASM_input_constraints(b); - for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) - return 1; - } +static void register_computed_value_func_proj(ir_op *op, + computed_value_func func) +{ + assert(op->ops.computed_value_Proj == NULL + || op->ops.computed_value_Proj == func); + op->ops.computed_value_Proj = func; +} - n = get_ASM_n_output_constraints(a); - if (n != get_ASM_n_output_constraints(b)) - return 0; +static void register_equivalent_node_func(ir_op *op, equivalent_node_func func) +{ + assert(op->ops.equivalent_node == NULL || op->ops.equivalent_node == func); + op->ops.equivalent_node = func; +} - ca = get_ASM_output_constraints(a); - cb = get_ASM_output_constraints(b); - for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) - return 1; - } +static void register_equivalent_node_func_proj(ir_op *op, + equivalent_node_func func) +{ + assert(op->ops.equivalent_node_Proj == NULL + || op->ops.equivalent_node_Proj == func); + op->ops.equivalent_node_Proj = func; +} - n = get_ASM_n_clobbers(a); - if (n != get_ASM_n_clobbers(b)) - return 0; +static void register_transform_node_func(ir_op *op, transform_node_func func) +{ + assert(op->ops.transform_node == NULL || op->ops.transform_node == func); + op->ops.transform_node = func; +} - cla = get_ASM_clobbers(a); - clb = get_ASM_clobbers(b); - for (i = 0; i < n; ++i) { - if (cla[i] != clb[i]) - return 1; - } - return 0; -} /* node_cmp_attr_ASM */ +static void register_transform_node_func_proj(ir_op *op, + transform_node_func func) +{ + assert(op->ops.transform_node_Proj == NULL + || op->ops.transform_node_Proj == func); + op->ops.transform_node_Proj = func; +} -/** Compares the inexistent attributes of two Dummy nodes. */ -static int node_cmp_attr_Dummy(ir_node *a, ir_node *b) +void ir_register_opt_node_ops(void) { - (void) a; - (void) b; - return 1; + register_computed_value_func(op_Add, computed_value_Add); + register_computed_value_func(op_And, computed_value_And); + register_computed_value_func(op_Borrow, computed_value_Borrow); + register_computed_value_func(op_Carry, computed_value_Carry); + register_computed_value_func(op_Cmp, computed_value_Cmp); + register_computed_value_func(op_Confirm, computed_value_Confirm); + register_computed_value_func(op_Const, computed_value_Const); + register_computed_value_func(op_Conv, computed_value_Conv); + register_computed_value_func(op_Eor, computed_value_Eor); + register_computed_value_func(op_Minus, computed_value_Minus); + register_computed_value_func(op_Mul, computed_value_Mul); + register_computed_value_func(op_Mux, computed_value_Mux); + register_computed_value_func(op_Not, computed_value_Not); + register_computed_value_func(op_Or, computed_value_Or); + register_computed_value_func(op_Proj, computed_value_Proj); + register_computed_value_func(op_Rotl, computed_value_Rotl); + register_computed_value_func(op_Shl, computed_value_Shl); + register_computed_value_func(op_Shr, computed_value_Shr); + register_computed_value_func(op_Shrs, computed_value_Shrs); + register_computed_value_func(op_Sub, computed_value_Sub); + register_computed_value_func(op_SymConst, computed_value_SymConst); + register_computed_value_func_proj(op_Div, computed_value_Proj_Div); + register_computed_value_func_proj(op_Mod, computed_value_Proj_Mod); + + register_equivalent_node_func(op_Add, equivalent_node_Add); + register_equivalent_node_func(op_And, equivalent_node_And); + register_equivalent_node_func(op_Confirm, equivalent_node_Confirm); + register_equivalent_node_func(op_Conv, equivalent_node_Conv); + register_equivalent_node_func(op_Eor, equivalent_node_Eor); + register_equivalent_node_func(op_Id, equivalent_node_Id); + register_equivalent_node_func(op_Minus, equivalent_node_involution); + register_equivalent_node_func(op_Mul, equivalent_node_Mul); + register_equivalent_node_func(op_Mux, equivalent_node_Mux); + register_equivalent_node_func(op_Not, equivalent_node_involution); + register_equivalent_node_func(op_Or, equivalent_node_Or); + register_equivalent_node_func(op_Phi, equivalent_node_Phi); + register_equivalent_node_func(op_Proj, equivalent_node_Proj); + register_equivalent_node_func(op_Rotl, equivalent_node_left_zero); + register_equivalent_node_func(op_Shl, equivalent_node_left_zero); + register_equivalent_node_func(op_Shr, equivalent_node_left_zero); + register_equivalent_node_func(op_Shrs, equivalent_node_left_zero); + register_equivalent_node_func(op_Sub, equivalent_node_Sub); + register_equivalent_node_func_proj(op_Bound, equivalent_node_Proj_Bound); + register_equivalent_node_func_proj(op_CopyB, equivalent_node_Proj_CopyB); + register_equivalent_node_func_proj(op_Div, equivalent_node_Proj_Div); + register_equivalent_node_func_proj(op_Tuple, equivalent_node_Proj_Tuple); + + register_transform_node_func(op_Add, transform_node_Add); + register_transform_node_func(op_And, transform_node_And); + register_transform_node_func(op_Block, transform_node_Block); + register_transform_node_func(op_Call, transform_node_Call); + register_transform_node_func(op_Cmp, transform_node_Cmp); + register_transform_node_func(op_Cond, transform_node_Cond); + register_transform_node_func(op_Conv, transform_node_Conv); + register_transform_node_func(op_Div, transform_node_Div); + register_transform_node_func(op_End, transform_node_End); + register_transform_node_func(op_Eor, transform_node_Eor); + register_transform_node_func(op_Load, transform_node_Load); + register_transform_node_func(op_Minus, transform_node_Minus); + register_transform_node_func(op_Mod, transform_node_Mod); + register_transform_node_func(op_Mul, transform_node_Mul); + register_transform_node_func(op_Mux, transform_node_Mux); + register_transform_node_func(op_Not, transform_node_Not); + register_transform_node_func(op_Or, transform_node_Or); + register_transform_node_func(op_Phi, transform_node_Phi); + register_transform_node_func(op_Proj, transform_node_Proj); + register_transform_node_func(op_Rotl, transform_node_Rotl); + register_transform_node_func(op_Shl, transform_node_Shl); + register_transform_node_func(op_Shrs, transform_node_Shrs); + register_transform_node_func(op_Shr, transform_node_Shr); + register_transform_node_func(op_Store, transform_node_Store); + register_transform_node_func(op_Sub, transform_node_Sub); + register_transform_node_func(op_Switch, transform_node_Switch); + register_transform_node_func(op_Sync, transform_node_Sync); + register_transform_node_func_proj(op_Bound, transform_node_Proj_Bound); + register_transform_node_func_proj(op_CopyB, transform_node_Proj_CopyB); + register_transform_node_func_proj(op_Div, transform_node_Proj_Div); + register_transform_node_func_proj(op_Load, transform_node_Proj_Load); + register_transform_node_func_proj(op_Mod, transform_node_Proj_Mod); + register_transform_node_func_proj(op_Store, transform_node_Proj_Store); } -/** - * Set the default node attribute compare operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) -{ -#define CASE(a) \ - case iro_##a: \ - ops->node_cmp_attr = node_cmp_attr_##a; \ - break - - switch (code) { - CASE(Const); - CASE(Proj); - CASE(Filter); - CASE(Alloc); - CASE(Free); - CASE(SymConst); - CASE(Call); - CASE(Sel); - CASE(Phi); - CASE(Conv); - CASE(Cast); - CASE(Load); - CASE(Store); - CASE(Confirm); - CASE(ASM); - CASE(Div); - CASE(DivMod); - CASE(Mod); - CASE(Quot); - CASE(Bound); - CASE(Builtin); - CASE(Dummy); - /* FIXME CopyB */ - default: - /* leave NULL */; - } +/* **************** Common Subexpression Elimination **************** */ - return ops; -#undef CASE -} /* firm_set_default_node_cmp_attr */ +/** The size of the hash table used, should estimate the number of nodes + in a graph. */ +#define N_IR_NODES 512 -/* - * 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) { +int identities_cmp(const void *elt, const void *key) +{ ir_node *a = (ir_node *)elt; ir_node *b = (ir_node *)key; int i, irn_arity_a; @@ -6232,24 +6427,51 @@ int identities_cmp(const void *elt, const void *key) { (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); - if (irn_arity_a != get_irn_intra_arity(b)) + irn_arity_a = get_irn_arity(a); + if (irn_arity_a != get_irn_arity(b)) + return 1; + + /* blocks are never the same */ + if (is_Block(a)) return 1; 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)) + if (get_nodes_block(a) != get_nodes_block(b)) return 1; + } else { + ir_node *block_a = get_nodes_block(a); + ir_node *block_b = get_nodes_block(b); + if (! get_opt_global_cse()) { + /* for block-local CSE both nodes must be in the same Block */ + if (block_a != block_b) + return 1; + } else { + /* The optimistic approach would be to do nothing here. + * However doing GCSE optimistically produces a lot of partially dead code which appears + * to be worse in practice than the missed opportunities. + * So we use a very conservative variant here and only CSE if 1 value dominates the + * other. */ + if (!block_dominates(block_a, block_b) + && !block_dominates(block_b, block_a)) + return 1; + /* respect the workaround rule: do not move nodes which are only + * held by keepalive edges */ + if (only_used_by_keepalive(a) || only_used_by_keepalive(b)) + return 1; + } } /* compare a->in[0..ins] with b->in[0..ins] */ - for (i = 0; i < irn_arity_a; i++) - if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i)) - return 1; + for (i = 0; i < irn_arity_a; ++i) { + ir_node *pred_a = get_irn_n(a, i); + ir_node *pred_b = get_irn_n(b, i); + if (pred_a != pred_b) { + /* if both predecessors are CSE neutral they might be different */ + if (!is_irn_cse_neutral(pred_a) || !is_irn_cse_neutral(pred_b)) + return 1; + } + } /* * here, we already now that the nodes are identical except their @@ -6259,29 +6481,37 @@ int identities_cmp(const void *elt, const void *key) { return a->op->ops.node_cmp_attr(a, b); return 0; -} /* identities_cmp */ +} -/* - * Calculate a hash value of a node. - * - * @param node The IR-node - */ -unsigned ir_node_hash(const ir_node *node) { +unsigned ir_node_hash(const ir_node *node) +{ return node->op->ops.hash(node); -} /* ir_node_hash */ +} +void new_identities(ir_graph *irg) +{ + if (irg->value_table != NULL) + del_pset(irg->value_table); + irg->value_table = new_pset(identities_cmp, N_IR_NODES); +} -pset *new_identities(void) { - return new_pset(identities_cmp, N_IR_NODES); -} /* new_identities */ +void del_identities(ir_graph *irg) +{ + if (irg->value_table != NULL) + del_pset(irg->value_table); +} -void del_identities(pset *value_table) { - del_pset(value_table); -} /* del_identities */ +static int cmp_node_nr(const void *a, const void *b) +{ + ir_node **p1 = (ir_node**)a; + ir_node **p2 = (ir_node**)b; + long n1 = get_irn_node_nr(*p1); + long n2 = get_irn_node_nr(*p2); + return (n1>n2) - (n1value_table; + ir_node *nn; - if (!value_table) return n; + if (value_table == NULL) + return n; ir_normalize_node(n); /* lookup or insert in hash table with given hash key. */ - o = pset_insert(value_table, n, ir_node_hash(n)); + nn = (ir_node*)pset_insert(value_table, n, ir_node_hash(n)); - if (o != n) { - update_known_irn(o, n); + if (nn != n) { + /* n is reachable again */ + edges_node_revival(nn); } - return o; -} /* identify_remember */ + return nn; +} /** - * 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. + * 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) { +static inline ir_node *identify_cons(ir_node *n) +{ ir_node *old = n; - 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); + n = identify_remember(n); + if (n != old && get_nodes_block(old) != get_nodes_block(n)) { + ir_graph *irg = get_irn_irg(n); + set_irg_pinned(irg, op_pin_state_floats); + } return n; -} /* identify_cons */ +} + +void add_identities(ir_node *node) +{ + if (!get_opt_cse()) + return; + if (is_Block(node)) + return; -/* Add a node to the identities value table. */ -void add_identities(pset *value_table, ir_node *node) { - if (get_opt_cse() && is_no_Block(node)) - identify_remember(value_table, node); -} /* add_identities */ + identify_remember(node); +} -/* Visit each node in the value table of a graph. */ -void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) { - ir_node *node; +void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) +{ ir_graph *rem = current_ir_graph; current_ir_graph = irg; - foreach_pset(irg->value_table, node) + foreach_pset(irg->value_table, ir_node, node) { visit(node, env); - current_ir_graph = rem; -} /* visit_all_identities */ - -/** - * Garbage in, garbage out. If a node has a dead input, i.e., the - * Bad node is input to the node, return the Bad node. - */ -static ir_node *gigo(ir_node *node) { - int i, irn_arity; - ir_op *op = get_irn_op(node); - - /* remove garbage blocks by looking at control flow that leaves the block - and replacing the control flow by Bad. */ - if (get_irn_mode(node) == mode_X) { - ir_node *block = get_nodes_block(skip_Proj(node)); - - /* Don't optimize nodes in immature blocks. */ - if (!get_Block_matured(block)) - return node; - /* Don't optimize End, may have Bads. */ - if (op == op_End) return node; - - if (is_Block(block)) { - 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 < 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)) { - /* - * 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(); - } - } - } - } - - /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the - blocks predecessors is dead. */ - if (op != op_Block && op != op_Phi && op != op_Tuple) { - irn_arity = get_irn_arity(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(skip_Proj(node)))) - return new_Bad(); - - for (i = 0; i < irn_arity; i++) { - ir_node *pred = get_irn_n(node, i); - - if (is_Bad(pred)) - return new_Bad(); -#if 0 - /* Propagating Unknowns here seems to be a bad idea, because - sometimes we need a node as a input and did not want that - it kills it's user. - However, it might be useful to move this into a later phase - (if you think that optimizing such code is useful). */ - if (is_Unknown(pred) && mode_is_data(get_irn_mode(node))) - return new_Unknown(get_irn_mode(node)); -#endif - } - } -#if 0 - /* With this code we violate the agreement that local_optimize - only leaves Bads in Block, Phi and Tuple nodes. */ - /* If Block has only Bads as predecessors it's garbage. */ - /* If Phi has only Bads as predecessors it's garbage. */ - if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) { - irn_arity = get_irn_arity(node); - for (i = 0; i < irn_arity; i++) { - if (!is_Bad(get_irn_n(node, i))) break; - } - if (i == irn_arity) node = new_Bad(); } -#endif - return node; -} /* gigo */ + current_ir_graph = rem; +} -/** - * These optimizations deallocate nodes from the obstack. - * It can only be called if it is guaranteed that no other nodes - * reference this one, i.e., right after construction of a node. - * - * @param n The node to optimize - * - * current_ir_graph must be set to the graph of the node! - */ -ir_node *optimize_node(ir_node *n) { - tarval *tv; - ir_node *oldn = n; - ir_opcode iro = get_irn_opcode(n); +ir_node *optimize_node(ir_node *n) +{ + ir_node *oldn = n; + ir_graph *irg = get_irn_irg(n); + unsigned iro = get_irn_opcode(n); + ir_tarval *tv; /* Always optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; @@ -6519,31 +6627,18 @@ ir_node *optimize_node(ir_node *n) { if (get_opt_constant_folding()) { /* neither constants nor Tuple values can be evaluated */ if (iro != iro_Const && (get_irn_mode(n) != mode_T)) { - unsigned fp_model = get_irg_fp_model(current_ir_graph); - int old_fp_mode = tarval_fp_ops_enabled(); - - tarval_enable_fp_ops(! (fp_model & fp_no_float_fold)); - /* try to evaluate */ tv = computed_value(n); if (tv != tarval_bad) { ir_node *nw; - ir_type *old_tp = get_irn_type(n); - int i, arity = get_irn_arity(n); - int node_size; + size_t node_size; /* - * Try to recover the type of the new expression. - */ - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - - /* - * we MUST copy the node here temporary, because it's still needed - * for DBG_OPT_CSTEVAL + * we MUST copy the node here temporarily, because it's still + * needed for DBG_OPT_CSTEVAL */ node_size = offsetof(ir_node, attr) + n->op->attr_size; - oldn = alloca(node_size); + oldn = (ir_node*)alloca(node_size); memcpy(oldn, n, node_size); CLONE_ARR_A(ir_node *, oldn->in, n->in); @@ -6552,19 +6647,15 @@ ir_node *optimize_node(ir_node *n) { memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); /* note the inplace edges module */ - edges_node_deleted(n, current_ir_graph); + edges_node_deleted(n); /* evaluation was successful -- replace the node. */ - irg_kill_node(current_ir_graph, n); - nw = new_Const(tv); + irg_kill_node(irg, n); + nw = new_r_Const(irg, tv); - if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) - set_Const_type(nw, old_tp); DBG_OPT_CSTEVAL(oldn, nw); - tarval_enable_fp_ops(old_fp_mode); return nw; } - tarval_enable_fp_ops(old_fp_mode); } } @@ -6584,13 +6675,13 @@ ir_node *optimize_node(ir_node *n) { * subexpressions within a block. */ if (get_opt_cse()) - n = identify_cons(current_ir_graph->value_table, n); + n = identify_cons(n); if (n != oldn) { - edges_node_deleted(oldn, current_ir_graph); + edges_node_deleted(oldn); /* We found an existing, better node, so we can deallocate the old node. */ - irg_kill_node(current_ir_graph, oldn); + irg_kill_node(irg, oldn); return n; } @@ -6598,205 +6689,74 @@ ir_node *optimize_node(ir_node *n) { free the node. */ iro = get_irn_opcode(n); if (get_opt_algebraic_simplification() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ n = transform_node(n); - - /* Remove nodes with dead (Bad) input. - Run always for transformation induced Bads. */ - n = gigo(n); + } /* Now we have a legal, useful node. Enter it in hash table for CSE */ - if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + if (get_opt_cse()) { ir_node *o = n; - n = identify_remember(current_ir_graph->value_table, o); + n = identify_remember(o); if (o != n) DBG_OPT_CSE(o, n); } return n; -} /* optimize_node */ - - -/** - * These optimizations never deallocate nodes (in place). This can cause dead - * nodes lying on the obstack. Remove these by a dead node elimination, - * i.e., a copying garbage collection. - */ -ir_node *optimize_in_place_2(ir_node *n) { - tarval *tv; - ir_node *oldn = n; - ir_opcode iro = get_irn_opcode(n); +} +ir_node *optimize_in_place_2(ir_node *n) +{ if (!get_opt_optimize() && !is_Phi(n)) return n; - /* constant expression evaluation / constant folding */ - if (get_opt_constant_folding()) { - /* neither constants nor Tuple values can be evaluated */ - if (iro != iro_Const && get_irn_mode(n) != mode_T) { - unsigned fp_model = get_irg_fp_model(current_ir_graph); - int old_fp_mode = tarval_fp_ops_enabled(); - - tarval_enable_fp_ops((fp_model & fp_strict_algebraic) == 0); - /* try to evaluate */ - tv = computed_value(n); - if (tv != tarval_bad) { - /* evaluation was successful -- replace the node. */ - ir_type *old_tp = get_irn_type(n); - int i, arity = get_irn_arity(n); - - /* - * Try to recover the type of the new expression. - */ - for (i = 0; i < arity && !old_tp; ++i) - old_tp = get_irn_type(get_irn_n(n, i)); - - n = new_Const(tv); - - if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) - set_Const_type(n, old_tp); - - DBG_OPT_CSTEVAL(oldn, n); - tarval_enable_fp_ops(old_fp_mode); - return n; - } - tarval_enable_fp_ops(old_fp_mode); - } - } - - /* remove unnecessary nodes */ - if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || /* ... */ - (iro == iro_Proj) || /* ... */ - (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node(n); + if (is_Deleted(n)) + return n; /** common subexpression elimination **/ /* Checks whether n is already available. */ - /* The block input is used to distinguish different subexpressions. Right - now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common - subexpressions within a block. */ + /* The block input is used to distinguish different subexpressions. + * Right now all nodes are op_pin_state_pinned to blocks, i.e., the cse + * only finds common subexpressions within a block. */ if (get_opt_cse()) { ir_node *o = n; - n = identify_remember(current_ir_graph->value_table, o); - if (o != n) + n = identify_remember(n); + if (n != o) { DBG_OPT_CSE(o, n); + /* we have another existing node now, we do not optimize it here */ + return n; + } } - /* Some more constant expression evaluation. */ - iro = get_irn_opcode(n); - if (get_opt_constant_folding() || - (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 = transform_node(n); /* Now we can verify the node, as it has no dead inputs any more. */ - irn_vrfy(n); + irn_verify(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)) { + * + * Note: This is only necessary because some of the optimisations + * operate in-place (set_XXX_bla, turn_into_tuple, ...) which is considered + * bad practice and should be fixed sometime. + */ + if (get_opt_cse()) { ir_node *o = n; - n = identify_remember(current_ir_graph->value_table, o); + n = identify_remember(o); if (o != n) DBG_OPT_CSE(o, n); } return n; -} /* optimize_in_place_2 */ +} -/** - * Wrapper for external use, set proper status bits after optimization. - */ -ir_node *optimize_in_place(ir_node *n) { - /* Handle graph state */ - assert(get_irg_phase_state(current_ir_graph) != phase_building); +ir_node *optimize_in_place(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, op_pin_state_floats); - if (get_irg_outs_state(current_ir_graph) == outs_consistent) - set_irg_outs_inconsistent(current_ir_graph); + set_irg_pinned(irg, op_pin_state_floats); /* FIXME: Maybe we could also test whether optimizing the node can change the control graph. */ - set_irg_doms_inconsistent(current_ir_graph); + clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE); 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); - - return h; -} /* hash_Const */ - -/** - * Calculate a hash value of a SymConst node. - */ -static unsigned hash_SymConst(const ir_node *node) { - unsigned h; - - /* all others are pointers */ - h = HASH_PTR(node->attr.symc.sym.type_p); - - 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); - ops = firm_set_default_node_cmp_attr(code, ops); - ops = firm_set_default_get_type(code, ops); - ops = firm_set_default_get_type_attr(code, ops); - ops = firm_set_default_get_entity_attr(code, ops); - - return ops; -} /* firm_set_default_operations */