X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=60ad50ea6f9e8a3c4f16b622e5ece73a81238716;hb=3e22a883f6ad7c41ab4972d966db7f12f8a21f89;hp=a696698be838d0038258801d6427b164368d4763;hpb=62361a42741ff68ef92e0e18ae2459145c3ede7f;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index a696698be..60ad50ea6 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -27,10 +27,17 @@ # include "irflag_t.h" # include "firmstat.h" # include "irarch.h" +# include "hashptr.h" /* Make types visible to allow most efficient access */ # include "entity_t.h" +# ifdef DO_HEAPANALYSIS +/* heapanal can't cope with NoMems */ +# else /* if defined DO_HEAPANALYSIS */ +# define USE_NOMEM +# endif /* defined DO_HEAPANALYSIS */ + /** * Trivial INLINEable routine for copy propagation. * Does follow Ids, needed to optimize INLINEd code. @@ -42,18 +49,6 @@ follow_Id (ir_node *n) return n; } -/** - * Returns the tarval of a Const node or tarval_bad for all other nodes. - */ -static INLINE tarval * -value_of (ir_node *n) -{ - if ((n != NULL) && (get_irn_op(n) == op_Const)) - return get_Const_tarval(n); /* might return tarval_bad */ - else - return tarval_bad; -} - /** * return the value of a Constant */ @@ -73,6 +68,9 @@ static tarval *computed_value_SymConst(ir_node *n) return tarval_bad; } +/** + * return the value of an Add + */ static tarval *computed_value_Add(ir_node *n) { ir_node *a = get_Add_left(n); @@ -81,14 +79,16 @@ static tarval *computed_value_Add(ir_node *n) tarval *ta = value_of(a); tarval *tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) - && (get_irn_mode(a) == get_irn_mode(b)) - && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) { + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) return tarval_add(ta, tb); - } + return tarval_bad; } +/** + * return the value of a Sub + * Special case: a - a + */ static tarval *computed_value_Sub(ir_node *n) { ir_node *a = get_Sub_left(n); @@ -103,14 +103,15 @@ static tarval *computed_value_Sub(ir_node *n) ta = value_of(a); tb = value_of(b); - if ((ta != tarval_bad) && (tb != tarval_bad) - && (get_irn_mode(a) == get_irn_mode(b)) - && !(get_mode_sort(get_irn_mode(a)) == irms_reference)) { + if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) return tarval_sub(ta, tb); - } + return tarval_bad; } +/** + * return the value of an unary Minus + */ static tarval *computed_value_Minus(ir_node *n) { ir_node *a = get_Minus_op(n); @@ -122,6 +123,9 @@ static tarval *computed_value_Minus(ir_node *n) return tarval_bad; } +/** + * return the value of a Mul + */ static tarval *computed_value_Mul(ir_node *n) { ir_node *a = get_Mul_left(n); @@ -136,18 +140,17 @@ static tarval *computed_value_Mul(ir_node *n) /* a*0 = 0 or 0*b = 0: calls computed_value recursive and returns the 0 with proper mode. */ - tarval *v; - - if ( ( ((v = ta) != tarval_bad) - && (v == get_mode_null(get_tarval_mode(v))) ) - || ( ((v = tb) != tarval_bad) - && (v == get_mode_null(get_tarval_mode(v))) )) { - return v; - } + if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta)))) + return ta; + if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb)))) + return tb; } return tarval_bad; } +/** + * return the value of a floating point Quot + */ static tarval *computed_value_Quot(ir_node *n) { ir_node *a = get_Quot_left(n); @@ -164,39 +167,64 @@ static tarval *computed_value_Quot(ir_node *n) return tarval_bad; } -static tarval *computed_value_Div(ir_node *n) +/** + * calculate the value of an integer Div of two nodes + * Special case: 0 / b + */ +static tarval *do_computed_value_Div(ir_node *a, ir_node *b) { - ir_node *a = get_Div_left(n); - ir_node *b = get_Div_right(n); - tarval *ta = value_of(a); tarval *tb = value_of(b); - /* This was missing in original implementation. Why? */ - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + /* Compute c1 / c2 or 0 / a, a != 0 */ + if (ta != tarval_bad) { + if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */ return tarval_div(ta, tb); + else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */ + return ta; } return tarval_bad; } -static tarval *computed_value_Mod(ir_node *n) +/** + * return the value of an integer Div + */ +static tarval *computed_value_Div(ir_node *n) { - ir_node *a = get_Mod_left(n); - ir_node *b = get_Mod_right(n); + return do_computed_value_Div(get_Div_left(n), get_Div_right(n)); +} +/** + * calculate the value of an integer Mod of two nodes + * Special case: a % 1 + */ +static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) +{ tarval *ta = value_of(a); tarval *tb = value_of(b); - /* This was missing in original implementation. Why? */ - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ + /* Compute c1 % c2 or a % 1 */ + if (tb != tarval_bad) { + if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */ return tarval_mod(ta, tb); + else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */ + return get_mode_null(get_irn_mode(a)); } return tarval_bad; } +/** + * return the value of an integer Mod + */ +static tarval *computed_value_Mod(ir_node *n) +{ + return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n)); +} + +/** + * return the value of an Abs + */ static tarval *computed_value_Abs(ir_node *n) { ir_node *a = get_Abs_op(n); @@ -208,6 +236,10 @@ static tarval *computed_value_Abs(ir_node *n) return tarval_bad; } +/** + * return the value of an And + * Special case: a & 0, 0 & b + */ static tarval *computed_value_And(ir_node *n) { ir_node *a = get_And_left(n); @@ -221,14 +253,18 @@ static tarval *computed_value_And(ir_node *n) } else { tarval *v; - if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_NULL) - || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_NULL)) { + if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL) + || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) { return v; } } return tarval_bad; } +/** + * return the value of an Or + * Special case: a | 1...1, 1...1 | b + */ static tarval *computed_value_Or(ir_node *n) { ir_node *a = get_Or_left(n); @@ -241,21 +277,29 @@ static tarval *computed_value_Or(ir_node *n) return tarval_or (ta, tb); } else { tarval *v; - if ( (classify_tarval ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) - || (classify_tarval ((v = computed_value (b))) == TV_CLASSIFY_ALL_ONE)) { + if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE) + || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) { return v; } } return tarval_bad; } +/** + * return the value of an Eor + */ static tarval *computed_value_Eor(ir_node *n) { ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + tarval *ta, *tb; + + if (a == b) + return get_tarval_null(get_irn_mode(n)); + + ta = value_of(a); + tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_eor (ta, tb); @@ -263,6 +307,9 @@ static tarval *computed_value_Eor(ir_node *n) return tarval_bad; } +/** + * return the value of a Not + */ static tarval *computed_value_Not(ir_node *n) { ir_node *a = get_Not_op(n); @@ -274,6 +321,9 @@ static tarval *computed_value_Not(ir_node *n) return tarval_bad; } +/** + * return the value of a Shl + */ static tarval *computed_value_Shl(ir_node *n) { ir_node *a = get_Shl_left(n); @@ -288,6 +338,9 @@ static tarval *computed_value_Shl(ir_node *n) return tarval_bad; } +/** + * return the value of a Shr + */ static tarval *computed_value_Shr(ir_node *n) { ir_node *a = get_Shr_left(n); @@ -302,6 +355,9 @@ static tarval *computed_value_Shr(ir_node *n) return tarval_bad; } +/** + * return the value of a Shrs + */ static tarval *computed_value_Shrs(ir_node *n) { ir_node *a = get_Shrs_left(n); @@ -316,6 +372,9 @@ static tarval *computed_value_Shrs(ir_node *n) return tarval_bad; } +/** + * return the value of a Rot + */ static tarval *computed_value_Rot(ir_node *n) { ir_node *a = get_Rot_left(n); @@ -330,6 +389,9 @@ static tarval *computed_value_Rot(ir_node *n) return tarval_bad; } +/** + * return the value of a Conv + */ static tarval *computed_value_Conv(ir_node *n) { ir_node *a = get_Conv_op(n); @@ -341,10 +403,14 @@ static tarval *computed_value_Conv(ir_node *n) return tarval_bad; } +/** + * return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod), Proj(DivMod) + */ static tarval *computed_value_Proj(ir_node *n) { - ir_node *a = get_Proj_pred(n), *b; + ir_node *a = get_Proj_pred(n); ir_node *aa, *ab; + long proj_nr; /* Optimize Cmp nodes. This performs a first step of unreachable code elimination. @@ -364,21 +430,23 @@ static tarval *computed_value_Proj(ir_node *n) case iro_Cmp: aa = get_Cmp_left(a); ab = get_Cmp_right(a); + proj_nr = get_Proj_proj(n); - if (aa == ab) { /* 1.: */ + if (aa == ab && !mode_is_float(get_irn_mode(aa))) { /* 1.: */ + /* BEWARE: a == a is NOT always True for floating Point!!! */ /* This is a trick with the bits used for encoding the Cmp Proj numbers, the following statement is not the same: - return new_tarval_from_long ((get_Proj_proj(n) == Eq), mode_b) */ - return new_tarval_from_long ((get_Proj_proj(n) & Eq), mode_b); + return new_tarval_from_long (proj_nr == Eq, mode_b) */ + return new_tarval_from_long (proj_nr & Eq, mode_b); } else { - tarval *taa = computed_value (aa); - tarval *tab = computed_value (ab); + tarval *taa = value_of(aa); + tarval *tab = value_of(ab); if ((taa != tarval_bad) && (tab != tarval_bad)) { /* 2.: */ /* strange checks... */ pnc_number flags = tarval_cmp (taa, tab); if (flags != False) { - return new_tarval_from_long (get_Proj_proj(n) & flags, mode_b); + return new_tarval_from_long (proj_nr & flags, mode_b); } } else { /* check for 3.: */ ir_node *aaa = skip_Id(skip_Proj(aa)); @@ -405,54 +473,29 @@ static tarval *computed_value_Proj(ir_node *n) && (mode_is_reference(get_irn_mode(ab))) && (get_irn_op(aba) == op_Alloc))) /* 3.: */ - return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b); + return new_tarval_from_long (proj_nr & Ne, mode_b); } } break; case iro_DivMod: - { - tarval *tb = value_of(b = get_DivMod_right(a)); - tarval *ta = value_of(a = get_DivMod_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_DivMod_res_div) - return tarval_div(ta, tb); - else if (get_Proj_proj(n)== pn_DivMod_res_mod) - return tarval_mod(ta, tb); - } + /* compute either the Div or the Mod part */ + proj_nr = get_Proj_proj(n); + if (proj_nr == pn_DivMod_res_div) + return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); + else if (proj_nr == pn_DivMod_res_mod) + return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); break; - } case iro_Div: - { - tarval *tb = value_of(b = get_Div_right(a)); - tarval *ta = value_of(a = get_Div_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_Div_res) - return tarval_div(ta, tb); - } + if (get_Proj_proj(n) == pn_Div_res) + return computed_value(a); break; - } case iro_Mod: - { - tarval *tb = value_of(b = get_Mod_right(a)); - tarval *ta = value_of(a = get_Mod_left(a)); - - if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) { - if (tb == get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */ - return tarval_bad; - if (get_Proj_proj(n)== pn_Mod_res) - return tarval_mod(ta, tb); - } + if (get_Proj_proj(n) == pn_Mod_res) + return computed_value(a); break; - } default: return tarval_bad; @@ -621,26 +664,6 @@ static ir_node *equivalent_node_Cond(ir_node *n) return n; } -/** - * Use algebraic simplification a v a = a. - */ -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); - - /* remove a v a */ - if (a == b) { - n = a; - - DBG_OPT_ALGSIM1(oldn, a, b, n); - } - - return n; -} - /** * optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. @@ -657,9 +680,9 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /* After running compute_node there is only one constant predecessor. Find this predecessors value and remember the other node: */ - if ((tv = computed_value(a)) != tarval_bad) { + if ((tv = value_of(a)) != tarval_bad) { on = b; - } else if ((tv = computed_value(b)) != tarval_bad) { + } else if ((tv = value_of(b)) != tarval_bad) { on = a; } else return n; @@ -689,7 +712,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) ir_node *a = get_binop_left(n); ir_node *b = get_binop_right(n); - if (classify_tarval(computed_value(b)) == TV_CLASSIFY_NULL) { + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); @@ -738,10 +761,10 @@ static ir_node *equivalent_node_Mul(ir_node *n) ir_node *b = get_Mul_right(n); /* Mul is commutative and has again an other neutral element. */ - if (classify_tarval (computed_value (a)) == TV_CLASSIFY_ONE) { + if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n); - } else if (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) { + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); } @@ -757,7 +780,7 @@ static ir_node *equivalent_node_Div(ir_node *n) ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ /* Turn Div into a tuple (mem, bad, a) */ ir_node *mem = get_Div_mem(n); turn_into_tuple(n, 3); @@ -768,6 +791,52 @@ static ir_node *equivalent_node_Div(ir_node *n) return n; } +/** + * Optimize a / 1 = a. + */ +static ir_node *equivalent_node_DivMod(ir_node *n) +{ + ir_node *a = get_DivMod_left(n); + ir_node *b = get_DivMod_right(n); + + /* Div is not commutative. */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */ + /* Turn DivMod into a tuple (mem, bad, a, 0) */ + ir_node *mem = get_Div_mem(n); + ir_mode *mode = get_irn_mode(b); + + turn_into_tuple(n, 4); + set_Tuple_pred(n, pn_DivMod_M, mem); + set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_DivMod_res_div, a); + set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode))); + } + return n; +} + +/** + * Use algebraic simplification a | a = a | 0 = 0 | a = a. + */ +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); + + if (a == b) { + n = a; /* Or has it's own neutral element */ + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n); + } + + return n; +} + /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. */ @@ -780,10 +849,10 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ - } else if (classify_tarval(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) { n = b; DBG_OPT_ALGSIM1(oldn, a, b, n); - } else if (classify_tarval(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; DBG_OPT_ALGSIM1(oldn, a, b, n); } @@ -815,12 +884,12 @@ static ir_node *equivalent_node_Conv(ir_node *n) if (n_mode == b_mode) { if (n_mode == mode_b) { n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n); } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { if (smaller_mode(b_mode, a_mode)){ n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n); + DBG_OPT_ALGSIM1(oldn, a, b, n); } } } @@ -828,14 +897,25 @@ static ir_node *equivalent_node_Conv(ir_node *n) return n; } +/** + * A Cast may be removed if the type of the previous node + * is already to type of the Cast. + */ +static ir_node *equivalent_node_Cast(ir_node *n) { + ir_node *pred = get_Cast_op(n); + if (get_irn_type(pred) == get_Cast_type(n)) + n = pred; + return n; +} + +/* Several optimizations: + - no Phi in start block. + - remove Id operators that are inputs to Phi + - fold Phi-nodes, iff they have only one predecessor except + themselves. +*/ static ir_node *equivalent_node_Phi(ir_node *n) { - /* Several optimizations: - - no Phi in start block. - - remove Id operators that are inputs to Phi - - fold Phi-nodes, iff they have only one predecessor except - themselves. - */ int i, n_preds; ir_node *oldn = n; @@ -1001,8 +1081,10 @@ static ir_op *firm_set_default_equivalent_node(ir_op *op) CASE(Minus); CASE(Mul); CASE(Div); + CASE(DivMod); CASE(And); CASE(Conv); + CASE(Cast); CASE(Phi); CASE(Proj); CASE(Id); @@ -1047,14 +1129,79 @@ optimize_preds(ir_node *n) { } /* end switch */ } -static ir_node *transform_node_Mul(ir_node *n) -{ +/** + * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and + * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)) if possible. + */ +static ir_node *transform_node_AddSub(ir_node *n) +{ + ir_mode *mode = get_irn_mode(n); + + if (mode_is_reference(mode)) { + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + int ref_bits = get_mode_size_bits(mode); + + if (get_irn_op(left) == op_Conv) { + ir_mode *mode = get_irn_mode(left); + int bits = get_mode_size_bits(mode); + + if (ref_bits == bits && + mode_is_int(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *pre = get_Conv_op(left); + ir_mode *pre_mode = get_irn_mode(pre); + + if (mode_is_int(pre_mode) && + get_mode_size_bits(pre_mode) == bits && + get_mode_arithmetic(pre_mode) == irma_twos_complement) { + /* ok, this conv just changes to sign, moreover the calculation + * is done with same number of bits as our address mode, so + * we can ignore the conv as address calculation can be viewed + * as either signed or unsigned + */ + set_binop_left(n, pre); + } + } + } + + if (get_irn_op(right) == op_Conv) { + ir_mode *mode = get_irn_mode(right); + int bits = get_mode_size_bits(mode); + + if (ref_bits == bits && + mode_is_int(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *pre = get_Conv_op(right); + ir_mode *pre_mode = get_irn_mode(pre); + + if (mode_is_int(pre_mode) && + get_mode_size_bits(pre_mode) == bits && + get_mode_arithmetic(pre_mode) == irma_twos_complement) { + /* ok, this conv just changes to sign, moreover the calculation + * is done with same number of bits as our address mode, so + * we can ignore the conv as address calculation can be viewed + * as either signed or unsigned + */ + set_binop_right(n, pre); + } + } + } + } + return n; +} + +#define transform_node_Add transform_node_AddSub +#define transform_node_Sub transform_node_AddSub + +/** Do architecture dependend optimizations on Mul nodes */ +static ir_node *transform_node_Mul(ir_node *n) { return arch_dep_replace_mul_with_shifts(n); } static ir_node *transform_node_Div(ir_node *n) { - tarval *tv = computed_value(n); + tarval *tv = value_of(n); ir_node *value = n; /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ @@ -1062,7 +1209,7 @@ static ir_node *transform_node_Div(ir_node *n) if (tv != tarval_bad) value = new_Const(get_tarval_mode(tv), tv); else /* Try architecture dependand optimization */ - value = arch_dep_replace_div_with_shifts(n); + value = arch_dep_replace_div_by_const(n); if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ @@ -1078,18 +1225,15 @@ static ir_node *transform_node_Div(ir_node *n) static ir_node *transform_node_Mod(ir_node *n) { - tarval *tv = computed_value(n); + tarval *tv = value_of(n); ir_node *value = n; - tarval *tb = value_of(get_Mod_right(n)); /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */ if (tv != tarval_bad) value = new_Const(get_tarval_mode(tv), tv); - else if (tb != tarval_bad && tb == get_mode_one(get_tarval_mode(tb))) { /* x mod 1 == 0 */ - ir_mode *mode = get_irn_mode(get_DivMod_left(n)); - value = new_Const(mode, get_mode_null(mode)); - } + else /* Try architecture dependand optimization */ + value = arch_dep_replace_mod_by_const(n); if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ @@ -1133,10 +1277,16 @@ static ir_node *transform_node_DivMod(ir_node *n) b = new_Const (mode, resb); evaluated = 1; } + else { /* Try architecture dependand optimization */ + arch_dep_replace_divmod_by_const(&a, &b, n); + evaluated = a != NULL; + } } else if (ta == get_mode_null(mode)) { + /* 0 / non-Const = 0 */ b = a; evaluated = 1; } + if (evaluated) { /* replace by tuple */ ir_node *mem = get_DivMod_mem(n); turn_into_tuple(n, 4); @@ -1187,7 +1337,7 @@ static ir_node *transform_node_Cond(ir_node *n) add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); } else if ((get_irn_op(a) == op_Eor) && (get_irn_mode(a) == mode_b) - && (classify_tarval(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { + && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { /* The Eor is a negate. Generate a new Cond without the negate, simulate the negate by exchanging the results. */ set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), @@ -1202,6 +1352,9 @@ static ir_node *transform_node_Cond(ir_node *n) return n; } +/** + * Transform an Eor. + */ static ir_node *transform_node_Eor(ir_node *n) { ir_node *a = get_Eor_left(n); @@ -1210,13 +1363,13 @@ static ir_node *transform_node_Eor(ir_node *n) if ((get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE) && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) /* The Eor negates a Cmp. The Cmp has the negated result anyways! */ n = new_r_Proj(current_ir_graph, get_nodes_block(n), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a))); else if ((get_irn_mode(n) == mode_b) - && (classify_tarval (computed_value (b)) == TV_CLASSIFY_ONE)) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode_b); @@ -1225,7 +1378,7 @@ static ir_node *transform_node_Eor(ir_node *n) } /** - * Transfor a boolean Not. + * Transform a boolean Not. */ static ir_node *transform_node_Not(ir_node *n) { @@ -1242,10 +1395,30 @@ static ir_node *transform_node_Not(ir_node *n) return n; } +/** + * Transform a Cast of a Const into a new Const + */ +static ir_node *transform_node_Cast(ir_node *n) { + ir_node *pred = get_Cast_op(n); + type *tp = get_irn_type(pred); + + if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { + n = new_rd_Const_type(NULL, current_ir_graph, get_nodes_block(pred), get_irn_mode(pred), + get_Const_tarval(pred), tp); + } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) { + n = new_rd_SymConst_type(NULL, current_ir_graph, get_nodes_block(pred), get_SymConst_symbol(pred), + get_SymConst_kind(pred), tp); + } + return n; +} + /** * Transform a Div/Mod/DivMod with a non-zero constant. Must be - * done here instead of equivalent node because in creates new + * done here instead of equivalent node because it creates new * nodes. + * Removes the exceptions and routes the memory to the NoMem node. + * + * Further, it optimizes jump tables by removing all impossible cases. */ static ir_node *transform_node_Proj(ir_node *proj) { @@ -1257,57 +1430,74 @@ static ir_node *transform_node_Proj(ir_node *proj) switch (get_irn_opcode(n)) { case iro_Div: b = get_Div_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_Div_X_except) { /* we found an exception handler, remove it */ return new_Bad(); - } - else if (proj_nr == pn_Div_M) { - /* the memory Proj can be removed */ + } else { + /* the memory Proj can be removed */ ir_node *res = get_Div_mem(n); - set_Div_mem(n, get_irg_initial_mem(current_ir_graph)); - return res; +# ifdef USE_NOMEM + set_Div_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ + if (proj_nr == pn_Div_M) + return res; } } break; case iro_Mod: b = get_Mod_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_Mod_X_except) { /* we found an exception handler, remove it */ return new_Bad(); - } - else if (proj_nr == pn_Mod_M) { - /* the memory Proj can be removed */ + } else { + /* the memory Proj can be removed */ ir_node *res = get_Mod_mem(n); - set_Mod_mem(n, get_irg_initial_mem(current_ir_graph)); - return res; +# ifdef USE_NOMEM + set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ + if (proj_nr == pn_Mod_M) + return res; } } break; case iro_DivMod: b = get_DivMod_right(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && classify_tarval(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ proj_nr = get_Proj_proj(proj); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); + if (proj_nr == pn_DivMod_X_except) { + /* we found an exception handler, remove it */ return new_Bad(); } - else if (proj_nr == pn_DivMod_M) { - /* the memory Proj can be removed */ + else { + /* the memory Proj can be removed */ ir_node *res = get_DivMod_mem(n); - set_DivMod_mem(n, get_irg_initial_mem(current_ir_graph)); - return res; +# ifdef USE_NOMEM + set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); +# endif /* defined USE_NOMEM */ + if (proj_nr == pn_DivMod_M) + return res; } } break; @@ -1315,7 +1505,7 @@ static ir_node *transform_node_Proj(ir_node *proj) case iro_Cond: if (get_opt_unreachable_code()) { b = get_Cond_selector(n); - tb = computed_value(b); + tb = value_of(b); if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) { /* we have a constant switch */ @@ -1448,6 +1638,65 @@ static ir_node *transform_node_Or(ir_node *or) return transform_node_Or(or); } +/* forward */ +static ir_node *transform_node(ir_node *n); + +/** + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl + */ +static ir_node * transform_node_shift(ir_node *n) +{ + ir_node *left; + tarval *tv1, *tv2, *res; + ir_mode *mode; + int modulo_shf, flag; + + left = get_binop_left(n); + + /* different operations */ + if (get_irn_op(left) != get_irn_op(n)) + return n; + + tv1 = value_of(get_binop_right(n)); + if (tv1 == tarval_bad) + return n; + + tv2 = value_of(get_binop_right(left)); + if (tv2 == tarval_bad) + return n; + + res = tarval_add(tv1, tv2); + + /* beware: a simple replacement works only, if res < modulo shift */ + mode = get_irn_mode(n); + + flag = 0; + + modulo_shf = get_mode_modulo_shift(mode); + if (modulo_shf > 0) { + tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res)); + + if (tarval_cmp(res, modulo) & Lt) + flag = 1; + } + else + flag = 1; + + if (flag) { + /* ok, we can replace it */ + ir_node *in[2], *irn, *block = get_nodes_block(n); + + in[0] = get_binop_left(left); + in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + + irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + + return transform_node(irn); + } + return n; +} + + /** * Tries several [inplace] [optimizing] transformations and returns an * equivalent node. The difference to equivalent_node() is that these @@ -1472,6 +1721,8 @@ static ir_op *firm_set_default_transform_node(ir_op *op) break switch (op->code) { + CASE(Add); + CASE(Sub); CASE(Mul); CASE(Div); CASE(Mod); @@ -1479,8 +1730,14 @@ static ir_op *firm_set_default_transform_node(ir_op *op) CASE(Cond); CASE(Eor); CASE(Not); + CASE(Cast); CASE(Proj); CASE(Or); + case iro_Shr: + case iro_Shrs: + case iro_Shl: + op->transform_node = transform_node_shift; + break; default: op->transform_node = NULL; } @@ -1496,68 +1753,75 @@ static ir_op *firm_set_default_transform_node(ir_op *op) in a graph. */ #define N_IR_NODES 512 +/** 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)); } +/** 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); } +/** 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); } +/** Compares the attributes of two Alloc nodes. */ static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) { return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where) - || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); + || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type); } +/** Compares the attributes of two Free nodes. */ static int node_cmp_attr_Free(ir_node *a, ir_node *b) { return (get_irn_free_attr(a) != get_irn_free_attr(b)); } +/** Compares the attributes of two SymConst nodes. */ static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) - || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p); + || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p) + || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp); } +/** Compares the attributes of two Call nodes. */ static int node_cmp_attr_Call(ir_node *a, ir_node *b) { return (get_irn_call_attr(a) != get_irn_call_attr(b)); } -static int node_cmp_attr_FuncCall(ir_node *a, ir_node *b) -{ - return (get_irn_funccall_attr(a) != get_irn_funccall_attr(b)); -} - +/** Compares the attributes of two Sel nodes. */ static int node_cmp_attr_Sel(ir_node *a, ir_node *b) { - return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) - || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) - || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) + return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind) + || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name) + || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner) || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name) - || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); + || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type); } +/** Compares the attributes of two Phi nodes. */ static int node_cmp_attr_Phi(ir_node *a, ir_node *b) { return get_irn_phi_attr (a) != get_irn_phi_attr (b); } +/** 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); } +/** 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 || @@ -1568,11 +1832,12 @@ static int node_cmp_attr_Load(ir_node *a, ir_node *b) return get_Load_mode(a) != get_Load_mode(b); } +/** Compares the attributes of two Store nodes. */ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { /* NEVER do CSE on volatile Stores */ return (get_Store_volatility(a) == volatility_is_volatile || - get_Load_volatility(b) == volatility_is_volatile); + get_Store_volatility(b) == volatility_is_volatile); } /** @@ -1593,7 +1858,6 @@ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) CASE(Free); CASE(SymConst); CASE(Call); - CASE(FuncCall); CASE(Sel); CASE(Phi); CASE(Cast); @@ -1625,20 +1889,20 @@ vt_cmp (const void *elt, const void *key) if ((get_irn_op(a) != get_irn_op(b)) || (get_irn_mode(a) != get_irn_mode(b))) return 1; - /* compare if a's in and b's in are equal */ - irn_arity_a = get_irn_arity (a); - if (irn_arity_a != get_irn_arity(b)) + /* 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)) return 1; /* for block-local cse and op_pin_state_pinned nodes: */ - if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == op_pin_state_pinned)) { - if (get_irn_n(a, -1) != get_irn_n(b, -1)) + if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) { + if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1)) return 1; } /* compare a->in[0..ins] with b->in[0..ins] */ for (i = 0; i < irn_arity_a; i++) - if (get_irn_n(a, i) != get_irn_n(b, i)) + if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i)) return 1; /* @@ -1662,41 +1926,39 @@ ir_node_hash (ir_node *node) if (node->op == op_Const) { /* special value for const, as they only differ in their tarval. */ - h = ((unsigned) node->attr.con.tv)>>3 ; - h = 9*h + (unsigned)get_irn_mode(node); + h = HASH_PTR(node->attr.con.tv); + h = 9*h + HASH_PTR(get_irn_mode(node)); } else if (node->op == op_SymConst) { /* special value for const, as they only differ in their symbol. */ - h = ((unsigned) node->attr.i.sym.type_p)>>3 ; - h = 9*h + (unsigned)get_irn_mode(node); + h = HASH_PTR(node->attr.i.sym.type_p); + h = 9*h + HASH_PTR(get_irn_mode(node)); } else { /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */ - h = irn_arity = get_irn_arity(node); + h = irn_arity = get_irn_intra_arity(node); - /* consider all in nodes... except the block. */ - for (i = 0; i < irn_arity; i++) { - h = 9*h + (unsigned)get_irn_n(node, i); + /* consider all in nodes... except the block if not a control flow. */ + for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) { + h = 9*h + HASH_PTR(get_irn_intra_n(node, i)); } /* ...mode,... */ - h = 9*h + (unsigned) get_irn_mode (node); + h = 9*h + HASH_PTR(get_irn_mode(node)); /* ...and code */ - h = 9*h + (unsigned) get_irn_op (node); + h = 9*h + HASH_PTR(get_irn_op(node)); } return h; } pset * -new_identities (void) -{ - return new_pset (vt_cmp, N_IR_NODES); +new_identities(void) { + return new_pset(vt_cmp, N_IR_NODES); } void -del_identities (pset *value_table) -{ - del_pset (value_table); +del_identities(pset *value_table) { + del_pset(value_table); } /** @@ -1714,7 +1976,6 @@ identify (pset *value_table, ir_node *n) if (!value_table) return n; - /* TODO: use a generic commutative attribute */ if (get_opt_reassociation()) { if (is_op_commutative(get_irn_op(n))) { ir_node *l = get_binop_left(n); @@ -1757,23 +2018,39 @@ identify_cons (pset *value_table, ir_node *n) { * if it isn't there yet. */ static ir_node * -identify_remember (pset *value_table, ir_node *node) +identify_remember (pset *value_table, ir_node *n) { ir_node *o = NULL; - if (!value_table) return node; + if (!value_table) return n; + + if (get_opt_reassociation()) { + if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + + /* for commutative operators perform a OP b == b OP a */ + if (l > r) { + set_binop_left(n, r); + set_binop_right(n, l); + } + } + } /* lookup or insert in hash table with given hash key. */ - o = pset_insert (value_table, node, ir_node_hash (node)); + o = pset_insert (value_table, n, ir_node_hash (n)); - if (o == node) return node; + if (o != n) { + DBG_OPT_CSE(n, o); + } return o; } void add_identities (pset *value_table, ir_node *node) { - identify_remember (value_table, node); + if (get_opt_cse() && (get_irn_opcode(node) != iro_Block)) + identify_remember (value_table, node); } /** @@ -1839,6 +2116,13 @@ optimize_node (ir_node *n) ir_node *oldn = n; opcode iro = get_irn_opcode(n); + type *old_tp = get_irn_type(n); + { + int i, arity = get_irn_arity(n); + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + } + /* Allways optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; @@ -1847,26 +2131,27 @@ optimize_node (ir_node *n) /* constants can not be evaluated */ if (iro != iro_Const) { /* try to evaluate */ - tv = computed_value (n); + tv = computed_value(n); if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { /* * we MUST copy the node here temporary, because it's still needed - * for DBG_OPT_ALGSIM0 + * for DBG_OPT_CSTEVAL */ int node_size = offsetof(ir_node, attr) + n->op->attr_size; oldn = alloca(node_size); memcpy(oldn, n, node_size); - CLONE_ARR_A(ir_node *, oldn->in, n->in); + CLONE_ARR_A(ir_node *, oldn->in, n->in); - /* ARG, copy the in array, we need it for statistics */ - memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); + /* ARG, copy the in array, we need it for statistics */ + memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); /* evaluation was successful -- replace the node. */ obstack_free (current_ir_graph->obst, n); n = new_Const (get_tarval_mode (tv), tv); - - DBG_OPT_ALGSIM0(oldn, n); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(n, old_tp); + DBG_OPT_CSTEVAL(oldn, n); return n; } } @@ -1930,6 +2215,13 @@ optimize_in_place_2 (ir_node *n) ir_node *oldn = n; opcode iro = get_irn_opcode(n); + type *old_tp = get_irn_type(n); + { + int i, arity = get_irn_arity(n); + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + } + if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; /* if not optimize return n */ @@ -1939,18 +2231,20 @@ optimize_in_place_2 (ir_node *n) return n; } - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { /* constants can not be evaluated */ if (iro != iro_Const) { /* try to evaluate */ - tv = computed_value (n); + tv = computed_value(n); if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { /* evaluation was successful -- replace the node. */ n = new_Const (get_tarval_mode (tv), tv); - DBG_OPT_ALGSIM0(oldn, n); + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(n, old_tp); + + DBG_OPT_CSTEVAL(oldn, n); return n; } }