X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=69ce9d0f25bcb339030c92f1e1eb68320291b118;hb=abb11c60e50826d82d0efcee52842c0ccdde95e0;hp=a31b138e9cda5f00911c40f50d8727e5128348e3;hpb=eb69bf6043d29071581499f9be23a4b59fb61214;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index a31b138e9..69ce9d0f2 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -3,19 +3,30 @@ * File name: ir/ir/iropt.c * Purpose: iropt --- optimizations intertwined with IR construction. * Author: Christian Schaefer - * Modified by: Goetz Lindenmaier + * Modified by: Goetz Lindenmaier, Michael Beck * Created: * CVS-ID: $Id$ - * Copyright: (c) 1998-2003 Universität Karlsruhe + * Copyright: (c) 1998-2005 Universität Karlsruhe * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. */ #ifdef HAVE_CONFIG_H -# include +# include "config.h" +#endif + +#ifdef HAVE_ALLOCA_H +#include +#endif +#ifdef HAVE_MALLOC_H +#include +#endif +#ifdef HAVE_STRING_H +#include #endif # include "irnode_t.h" # include "irgraph_t.h" +# include "iredges_t.h" # include "irmode_t.h" # include "iropt_t.h" # include "ircons_t.h" @@ -25,47 +36,38 @@ # include "dbginfo_t.h" # include "iropt_dbg.h" # include "irflag_t.h" -# include "firmstat.h" +# include "irhooks.h" +# include "irarch.h" +# include "hashptr.h" +# include "archop.h" +# include "opt_polymorphy.h" +# include "opt_confirms.h" /* Make types visible to allow most efficient access */ # include "entity_t.h" /** - * Trivial INLINEable routine for copy propagation. - * Does follow Ids, needed to optimize INLINEd code. + * return the value of a Constant */ -static INLINE ir_node * -follow_Id (ir_node *n) +static tarval *computed_value_Const(ir_node *n) { - while (get_irn_op (n) == op_Id) n = get_Id_pred (n); - return n; + return get_Const_tarval(n); } /** - * Returns the tarval of a Const node or tarval_bad for all other nodes. + * return the value of a 'sizeof' SymConst */ -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; -} - -static tarval *computed_value_Const(ir_node *n) -{ - return get_Const_tarval(n); -} - static tarval *computed_value_SymConst(ir_node *n) { - if ((get_SymConst_kind(n) ==symconst_size) && + if ((get_SymConst_kind(n) == symconst_size) && (get_type_state(get_SymConst_type(n))) == layout_fixed) - return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), mode_Is); + return new_tarval_from_long(get_type_size_bytes(get_SymConst_type(n)), get_irn_mode(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); @@ -74,30 +76,39 @@ 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); ir_node *b = get_Sub_right(n); + tarval *ta; + tarval *tb; - tarval *ta = value_of(a); - tarval *tb = value_of(b); + /* a - a */ + if (a == b && !is_Bad(a)) + return get_mode_null(get_irn_mode(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); @@ -109,6 +120,9 @@ 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); @@ -123,18 +137,17 @@ 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); @@ -151,38 +164,64 @@ 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); @@ -194,6 +233,10 @@ 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); @@ -207,14 +250,18 @@ static tarval *computed_value_And(ir_node *n) } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_NULL) - || (tarval_classify ((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); @@ -227,21 +274,29 @@ static tarval *computed_value_Or(ir_node *n) return tarval_or (ta, tb); } else { tarval *v; - if ( (tarval_classify ((v = computed_value (a))) == TV_CLASSIFY_ALL_ONE) - || (tarval_classify ((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_mode_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); @@ -249,6 +304,9 @@ 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); @@ -260,6 +318,9 @@ 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); @@ -274,6 +335,9 @@ 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); @@ -288,6 +352,9 @@ 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); @@ -302,6 +369,9 @@ 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); @@ -311,11 +381,14 @@ static tarval *computed_value_Rot(ir_node *n) tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - /* return tarval_rot (ta, tb); */ + return tarval_rot (ta, tb); } return tarval_bad; } +/** + * return the value of a Conv + */ static tarval *computed_value_Conv(ir_node *n) { ir_node *a = get_Conv_op(n); @@ -327,91 +400,194 @@ static tarval *computed_value_Conv(ir_node *n) return tarval_bad; } -static tarval *computed_value_Proj(ir_node *n) +/** + * return the value of a Proj(Cmp) + * + * This performs a first step of unreachable code elimination. + * Proj can not be computed, but folding a Cmp above the Proj here is + * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which + * only 1 is used. + * There are several case where we can evaluate a Cmp node, see later. + */ +static tarval *computed_value_Proj_Cmp(ir_node *n) { - ir_node *a = get_Proj_pred(n), *b; - ir_node *aa, *ab; - - /* Optimize Cmp nodes. - This performs a first step of unreachable code elimination. - Proj can not be computed, but folding a Cmp above the Proj here is - not as wasteful as folding a Cmp into a Tuple of 16 Consts of which - only 1 is used. - There are several case where we can evaluate a Cmp node: - 1. The nodes compared are both the same. If we compare for - equal, greater equal, ... this will return true, else it - will return false. This step relies on cse. - 2. The predecessors of Cmp are target values. We can evaluate - the Cmp. - 3. The predecessors are Allocs or void* constants. Allocs never - return NULL, they raise an exception. Therefore we can predict - the Cmp result. */ - if (get_irn_op(a) == op_Cmp) { - aa = get_Cmp_left(a); - ab = get_Cmp_right(a); - - if (aa == ab) { /* 1.: */ - /* This is a tric 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); - } else { - tarval *taa = computed_value (aa); - tarval *tab = computed_value (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); - } - } else { /* check for 3.: */ - ir_node *aaa = skip_nop(skip_Proj(aa)); - ir_node *aba = skip_nop(skip_Proj(ab)); - - if ( ( (/* aa is ProjP and aaa is Alloc */ - (get_irn_op(aa) == op_Proj) - && (mode_is_reference(get_irn_mode(aa))) - && (get_irn_op(aaa) == op_Alloc)) - && ( (/* ab is constant void */ - (get_irn_op(ab) == op_Const) - && (mode_is_reference(get_irn_mode(ab))) - && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) - || (/* ab is other Alloc */ - (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc) - && (aaa != aba)))) - || (/* aa is void and aba is Alloc */ - (get_irn_op(aa) == op_Const) - && (mode_is_reference(get_irn_mode(aa))) - && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) - && (get_irn_op(ab) == op_Proj) - && (mode_is_reference(get_irn_mode(ab))) - && (get_irn_op(aba) == op_Alloc))) - /* 3.: */ - return new_tarval_from_long (get_Proj_proj(n) & Ne, mode_b); + ir_node *a = get_Proj_pred(n); + ir_node *aa = get_Cmp_left(a); + ir_node *ab = get_Cmp_right(a); + long proj_nr = get_Proj_proj(n); + + /* + * BEWARE: a == a is NOT always True for floating Point values, as + * NaN != NaN is defined, so we must check this here. + */ + if (aa == ab && ( + !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt) + ) { /* 1.: */ + + /* This is a trick with the bits used for encoding the Cmp + Proj numbers, the following statement is not the same: + return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */ + return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b); + } + else { + tarval *taa = value_of(aa); + tarval *tab = value_of(ab); + ir_mode *mode = get_irn_mode(aa); + + /* + * The predecessors of Cmp are target values. We can evaluate + * the Cmp. + */ + if ((taa != tarval_bad) && (tab != tarval_bad)) { + /* strange checks... */ + pn_Cmp flags = tarval_cmp(taa, tab); + if (flags != pn_Cmp_False) { + return new_tarval_from_long (proj_nr & flags, mode_b); + } + } + /* for integer values, we can check against MIN/MAX */ + else if (mode_is_int(mode)) { + /* MIN <=/> x. This results in true/false. */ + if (taa == get_mode_min(mode)) { + /* a compare with the MIN value */ + if (proj_nr == pn_Cmp_Le) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Gt) + return get_tarval_b_false(); + } + /* x >=/< MIN. This results in true/false. */ + else + if (tab == get_mode_min(mode)) { + /* a compare with the MIN value */ + if (proj_nr == pn_Cmp_Ge) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Lt) + return get_tarval_b_false(); + } + /* MAX >=/< x. This results in true/false. */ + else if (taa == get_mode_max(mode)) { + if (proj_nr == pn_Cmp_Ge) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Lt) + return get_tarval_b_false(); + } + /* x <=/> MAX. This results in true/false. */ + else if (tab == get_mode_max(mode)) { + if (proj_nr == pn_Cmp_Le) + return get_tarval_b_true(); + else if (proj_nr == pn_Cmp_Gt) + return get_tarval_b_false(); } } - } else if (get_irn_op(a) == op_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)== 0) /* Div */ - return tarval_div(ta, tb); - else /* Mod */ - return tarval_mod(ta, tb); + /* + * The predecessors are Allocs or (void*)(0) constants. Allocs never + * return NULL, they raise an exception. Therefore we can predict + * the Cmp result. + */ + else { + ir_node *aaa = skip_Id(skip_Proj(aa)); + ir_node *aba = skip_Id(skip_Proj(ab)); + + if ( ( (/* aa is ProjP and aaa is Alloc */ + (get_irn_op(aa) == op_Proj) + && (mode_is_reference(get_irn_mode(aa))) + && (get_irn_op(aaa) == op_Alloc)) + && ( (/* ab is NULL */ + (get_irn_op(ab) == op_Const) + && (mode_is_reference(get_irn_mode(ab))) + && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) + || (/* ab is other Alloc */ + (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc) + && (aaa != aba)))) + || (/* aa is NULL and aba is Alloc */ + (get_irn_op(aa) == op_Const) + && (mode_is_reference(get_irn_mode(aa))) + && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) + && (get_irn_op(ab) == op_Proj) + && (mode_is_reference(get_irn_mode(ab))) + && (get_irn_op(aba) == op_Alloc))) + /* 3.: */ + return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b); } } + + return computed_value_Cmp_Confirm(a, aa, ab, proj_nr); +} + +/** + * 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); + long proj_nr; + + switch (get_irn_opcode(a)) { + case iro_Cmp: + return computed_value_Proj_Cmp(n); + + case iro_DivMod: + /* compute either the Div or the Mod part */ + proj_nr = get_Proj_proj(n); + if (proj_nr == pn_DivMod_res_div) + return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); + else if (proj_nr == pn_DivMod_res_mod) + return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); + break; + + case iro_Div: + if (get_Proj_proj(n) == pn_Div_res) + return computed_value(a); + break; + + case iro_Mod: + if (get_Proj_proj(n) == pn_Mod_res) + return computed_value(a); + break; + + default: + return tarval_bad; + } return tarval_bad; } +/** + * calculate the value of a Mux: can be evaluated, if the + * sel and the right input are known + */ +static tarval *computed_value_Mux(ir_node *n) +{ + ir_node *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); + + if (ts == get_tarval_b_true()) { + ir_node *v = get_Mux_true(n); + return value_of(v); + } + else if (ts == get_tarval_b_false()) { + ir_node *v = get_Mux_false(n); + return value_of(v); + } + return tarval_bad; +} + +/** + * calculate the value of a Confirm: can be evaluated, + * if it has the form Confirm(x, '=', Const). + */ +static tarval *computed_value_Confirm(ir_node *n) +{ + return get_Confirm_cmp(n) == pn_Cmp_Eq ? + value_of(get_Confirm_bound(n)) : tarval_bad; +} + /** * If the parameter n can be computed, return its value, else tarval_bad. * Performs constant folding. + * + * @param n The node this should be evaluated */ tarval *computed_value(ir_node *n) { @@ -425,9 +601,9 @@ tarval *computed_value(ir_node *n) */ static ir_op *firm_set_default_computed_value(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->computed_value = computed_value_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->computed_value = computed_value_##a; \ break switch (op->code) { @@ -451,6 +627,8 @@ static ir_op *firm_set_default_computed_value(ir_op *op) CASE(Rot); CASE(Conv); CASE(Proj); + CASE(Mux); + CASE(Confirm); default: op->computed_value = NULL; } @@ -471,18 +649,35 @@ different_identity (ir_node *a, ir_node *b) ir_node *a1 = get_Proj_pred (a); ir_node *b1 = get_Proj_pred (b); if (a1 != b1 && get_irn_op (a1) == op_Alloc - && get_irn_op (b1) == op_Alloc) + && get_irn_op (b1) == op_Alloc) return 1; } return 0; } #endif +/** + * Returns a equivalent block for another block. + * If the block has only one predecessor, this is + * the equivalent one. If the only predecessor of a block is + * the block itself, this is a dead block. + * + * If both predecessors of a block are the branches of a binary + * Cond, the equivalent block is Cond's block. + * + * If all predecessors of a block are bad or lies in a dead + * block, the current block is dead as well. + * + * Note, that blocks are NEVER turned into Bad's, instead + * the dead_block flag is set. So, never test for is_Bad(block), + * always use is_dead_Block(block). + */ static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; + int n_preds = get_Block_n_cfgpreds(n); - /* The Block constructor does not call optimize, but mature_block + /* The Block constructor does not call optimize, but mature_immBlock calls the optimization. */ assert(get_Block_matured(n)); @@ -492,52 +687,70 @@ static ir_node *equivalent_node_Block(ir_node *n) This should be true, as the block is matured before optimize is called. But what about Phi-cycles with the Phi0/Id that could not be resolved? Remaining Phi nodes are just Ids. */ - if ((get_Block_n_cfgpreds(n) == 1) && - (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { - ir_node *predblock = get_nodes_Block(get_Block_cfgpred(n, 0)); + if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) { + ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0)); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ - n = new_Bad(); DBG_OPT_DEAD; + n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); } else if (get_opt_control_flow_straightening()) { - n = predblock; DBG_OPT_STG; + n = predblock; + DBG_OPT_STG(oldn, n); } } - else if ((get_Block_n_cfgpreds(n) == 1) && - (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { - ir_node *predblock = get_nodes_Block(get_Block_cfgpred(n, 0)); + else if ((n_preds == 1) && + (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) { + ir_node *predblock = get_Block_cfgpred_block(n, 0); if (predblock == oldn) { /* Jmp jumps into the block it is in -- deal self cycle. */ - n = new_Bad(); DBG_OPT_DEAD; + n = set_Block_dead(n); + DBG_OPT_DEAD_BLOCK(oldn, n); } } - else if ((get_Block_n_cfgpreds(n) == 2) && - (get_opt_control_flow_weak_simplification())) { + else if ((n_preds == 2) && + (get_opt_control_flow_weak_simplification())) { /* Test whether Cond jumps twice to this block - @@@ we could do this also with two loops finding two preds from several ones. */ + * The more general case which more than 2 predecessors is handles + * in optimize_cf(), we handle only this special case for speed here. + */ ir_node *a = get_Block_cfgpred(n, 0); ir_node *b = get_Block_cfgpred(n, 1); if ((get_irn_op(a) == op_Proj) && - (get_irn_op(b) == op_Proj) && - (get_Proj_pred(a) == get_Proj_pred(b)) && - (get_irn_op(get_Proj_pred(a)) == op_Cond) && - (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { + (get_irn_op(b) == op_Proj) && + (get_Proj_pred(a) == get_Proj_pred(b)) && + (get_irn_op(get_Proj_pred(a)) == op_Cond) && + (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) { /* Also a single entry Block following a single exit Block. Phis have - twice the same operand and will be optimized away. */ - n = get_nodes_Block(a); DBG_OPT_IFSIM; + twice the same operand and will be optimized away. */ + n = get_nodes_block(a); + DBG_OPT_IFSIM1(oldn, a, b, n); } - } else if (get_opt_unreachable_code() && - (n != current_ir_graph->start_block) && - (n != current_ir_graph->end_block) ) { - int i; + } + else if (get_opt_unreachable_code() && + (n != current_ir_graph->start_block) && + (n != current_ir_graph->end_block) ) { + int i, n_cfg = get_Block_n_cfgpreds(n); + /* If all inputs are dead, this block is dead too, except if it is the start or end block. This is a step of unreachable code elimination */ - for (i = 0; i < get_Block_n_cfgpreds(n); i++) { - if (!is_Bad(get_Block_cfgpred(n, i))) break; + for (i = 0; i < n_cfg; 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 == get_Block_n_cfgpreds(n)) - n = new_Bad(); + if (i == n_cfg) + n = set_Block_dead(n); } return n; @@ -551,36 +764,14 @@ static ir_node *equivalent_node_Jmp(ir_node *n) { /* GL: Why not same for op_Raise?? */ /* unreachable code elimination */ - if (is_Bad(get_nodes_Block(n))) + if (is_Block_dead(get_nodes_block(n))) n = new_Bad(); return n; } -static ir_node *equivalent_node_Cond(ir_node *n) -{ - /* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See cases for iro_Cond and iro_Proj in transform_node. */ - return n; -} - -/** - * 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; - } - - return n; -} +/* 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, @@ -598,25 +789,79 @@ 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; /* If this predecessors constant value is zero, the operation is - unnecessary. Remove it: */ - if (tarval_classify (tv) == TV_CLASSIFY_NULL) { - n = on; DBG_OPT_ALGSIM1; + * unnecessary. Remove it. + * + * Beware: If n is a Add, the mode of on and n might be different + * which happens in this rare construction: NULL + 3. + * Then, a Conv would be needed which we cannot include here. + */ + if (classify_tarval (tv) == TV_CLASSIFY_NULL) { + if (get_irn_mode(on) == get_irn_mode(n)) { + n = on; + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); + } } return n; } -#define equivalent_node_Add equivalent_node_neutral_zero #define equivalent_node_Eor equivalent_node_neutral_zero +/* + * Optimize a - 0 and (a - x) + x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of an Add may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. + */ +static ir_node *equivalent_node_Add(ir_node *n) +{ + ir_node *oldn = n; + ir_node *left, *right; + + n = equivalent_node_neutral_zero(n); + if (n != oldn) + return n; + + left = get_Add_left(n); + right = get_Add_right(n); + + if (get_irn_op(left) == op_Sub) { + if (get_Sub_right(left) == right) { + /* (a - x) + x */ + + n = get_Sub_left(left); + if (get_irn_mode(oldn) == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + if (get_irn_op(right) == op_Sub) { + if (get_Sub_right(right) == left) { + /* x + (a - x) */ + + n = get_Sub_left(right); + if (get_irn_mode(oldn) == get_irn_mode(n)) { + DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB); + return n; + } + } + } + return n; +} + /** * optimize operations that are not commutative but have neutral 0 on left, * so a op 0 = a. @@ -628,39 +873,96 @@ 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 (tarval_classify(computed_value(b)) == TV_CLASSIFY_NULL) { - n = a; DBG_OPT_ALGSIM1; + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + n = a; + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } return n; } -#define equivalent_node_Sub equivalent_node_left_zero #define equivalent_node_Shl equivalent_node_left_zero #define equivalent_node_Shr equivalent_node_left_zero #define equivalent_node_Shrs equivalent_node_left_zero #define equivalent_node_Rot equivalent_node_left_zero /** - * Er, a "symmetic unop", ie op(op(n)) = n. + * Optimize a - 0 and (a + x) - x (for modes with wrap-around). + * + * The second one looks strange, but this construct + * is used heavily in the LCC sources :-). + * + * Beware: The Mode of a Sub may be different than the mode of its + * predecessors, so we could not return a predecessors in all cases. */ -static ir_node *equivalent_node_symmetric_unop(ir_node *n) +static ir_node *equivalent_node_Sub(ir_node *n) { ir_node *oldn = n; + ir_node *a = get_Sub_left(n); + ir_node *b = get_Sub_right(n); + + /* Beware: modes might be different */ + if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + if (get_irn_mode(n) == get_irn_mode(a)) { + n = a; + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); + } + } + else if (get_irn_op(a) == op_Add) { + ir_mode *mode = get_irn_mode(n); + + if (mode_wrap_around(mode)) { + ir_node *left = get_Add_left(a); + ir_node *right = get_Add_right(a); + + if (left == b) { + if (get_irn_mode(n) == get_irn_mode(right)) { + n = right; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + } + } + else if (right == b) { + if (get_irn_mode(n) == get_irn_mode(left)) { + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + } + } + } + } + + return n; +} + + +/** + * Optimize an "idempotent unary op", ie op(op(n)) = n. + * + * @fixme -(-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) +{ + ir_node *oldn = n; + ir_node *pred = get_unop_op(n); + /* optimize symmetric unop */ - if (get_irn_op(get_unop_op(n)) == get_irn_op(n)) { - n = get_unop_op(get_unop_op(n)); DBG_OPT_ALGSIM2; + if (get_irn_op(pred) == get_irn_op(n)) { + n = get_unop_op(pred); + DBG_OPT_ALGSIM2(oldn, pred, n); } return n; } -/* NotNot x == x */ -#define equivalent_node_Not equivalent_node_symmetric_unop +/* 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_symmetric_unop + out of bounds exception if min =! max? */ +#define equivalent_node_Minus equivalent_node_idempotent_unop /** * Optimize a * 1 = 1 * a = a. @@ -673,10 +975,12 @@ 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 (tarval_classify (computed_value (a)) == TV_CLASSIFY_ONE) { - n = b; DBG_OPT_ALGSIM1; - } else if (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) { - n = a; DBG_OPT_ALGSIM1; + if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); } return n; } @@ -690,17 +994,64 @@ static ir_node *equivalent_node_Div(ir_node *n) ir_node *b = get_Div_right(n); /* Div is not commutative. */ - if (tarval_classify(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); set_Tuple_pred(n, pn_Div_M, mem); - set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ + set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_Div_res, a); } 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 */ + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR); + } + + return n; +} + /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a. */ @@ -713,17 +1064,19 @@ static ir_node *equivalent_node_And(ir_node *n) if (a == b) { n = a; /* And has it's own neutral element */ - } else if (tarval_classify(computed_value(a)) == TV_CLASSIFY_ALL_ONE) { + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); + } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) { n = b; - } else if (tarval_classify(computed_value(b)) == TV_CLASSIFY_ALL_ONE) { + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) { n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); } - if (n != oldn) DBG_OPT_ALGSIM1; return n; } /** - * Try to remove useless conv's: + * Try to remove useless Conv's: */ static ir_node *equivalent_node_Conv(ir_node *n) { @@ -735,7 +1088,8 @@ static ir_node *equivalent_node_Conv(ir_node *n) ir_mode *a_mode = get_irn_mode(a); if (n_mode == a_mode) { /* No Conv necessary */ - n = a; DBG_OPT_ALGSIM3; + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */ ir_mode *b_mode; @@ -745,26 +1099,43 @@ 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; + n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) { - if (smaller_mode(b_mode, a_mode)){ - n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ DBG_OPT_ALGSIM1; - } + if (smaller_mode(b_mode, a_mode)){ + n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + } } } } return n; } +/** + * 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; +} + +/* 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; @@ -779,49 +1150,33 @@ static ir_node *equivalent_node_Phi(ir_node *n) block = get_nodes_block(n); /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!! assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */ - if ((is_Bad(block)) || /* Control dead */ + if ((is_Block_dead(block)) || /* Control dead */ (block == current_ir_graph->start_block)) /* There should be no Phi nodes */ - return new_Bad(); /* in the Start Block. */ + return new_Bad(); /* in the Start Block. */ if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ -#if 0 - /* first we test for a special case: */ - /* Confirm is a special node fixing additional information for a - value that is known at a certain point. This is useful for - dataflow analysis. */ - if (n_preds == 2) { - ir_node *a = get_Phi_pred(n, 0); - ir_node *b = get_Phi_pred(n, 1); - if ( (get_irn_op(a) == op_Confirm) - && (get_irn_op(b) == op_Confirm) - && follow_Id (get_irn_n(a, 0) == get_irn_n(b, 0)) - && (get_irn_n(a, 1) == get_irn_n (b, 1)) - && (a->data.num == (~b->data.num & irpn_True) )) { - return get_irn_n(a, 0); - } - } -#endif - /* If the Block has a Bad pred, we also have one. */ for (i = 0; i < n_preds; ++i) - if (is_Bad (get_Block_cfgpred(block, i))) + if (is_Bad(get_Block_cfgpred(block, i))) set_Phi_pred(n, i, new_Bad()); /* Find first non-self-referencing input */ - for (i = 0; i < n_preds; ++i) { + for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ #if 1 - && (get_irn_op(first_val) != op_Bad) + && (! is_Bad(first_val)) #endif - ) { /* value not dead */ + ) { /* value not dead */ break; /* then found first value. */ } } - /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - if (i >= n_preds) { return new_Bad(); } + if (i >= n_preds) { + /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ + return new_Bad(); + } scnd_val = NULL; @@ -830,18 +1185,19 @@ static ir_node *equivalent_node_Phi(ir_node *n) while (++i < n_preds) { scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) - && (scnd_val != first_val) + && (scnd_val != first_val) #if 1 - && (get_irn_op(scnd_val) != op_Bad) + && (! is_Bad(scnd_val)) #endif - ) { + ) { break; } } - /* Fold, if no multiple distinct non-self-referencing inputs */ if (i >= n_preds) { - n = first_val; DBG_OPT_PHI; + /* Fold, if no multiple distinct non-self-referencing inputs */ + n = first_val; + DBG_OPT_PHI(oldn, n); } else { /* skip the remaining Ids (done in get_Phi_pred). */ /* superfluous, since we walk all to propagate Block's Bads. @@ -851,103 +1207,159 @@ static ir_node *equivalent_node_Phi(ir_node *n) } /** - * Optimize Loads after Store. - * - * @todo FAILS for volatile entities + * optimize Proj(Tuple) and gigo() for ProjX in Bad block */ -static ir_node *equivalent_node_Load(ir_node *n) +static ir_node *equivalent_node_Proj(ir_node *n) { ir_node *oldn = n; - /* remove unnecessary Load. */ - ir_node *a = skip_Proj(get_Load_mem(n)); - ir_node *b = get_Load_ptr(n); - ir_node *c; - - /* TODO: check for volatile */ - if (get_irn_op(a) == op_Store && get_Store_ptr(a) == b) { - /* We load immediately after a store -- a read after write. */ - ir_node *mem = get_Load_mem(n); + ir_node *a = get_Proj_pred(n); - c = get_Store_value(a); - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Load_M, mem); - set_Tuple_pred(n, pn_Load_res, c); - set_Tuple_pred(n, pn_Load_X_except, new_Bad()); DBG_OPT_RAW; + if ( get_irn_op(a) == op_Tuple) { + /* Remove the Tuple/Proj combination. */ + if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) { + n = get_Tuple_pred(a, get_Proj_proj(n)); + DBG_OPT_TUPLE(oldn, a, n); + } else { + assert(0); /* This should not happen! */ + n = new_Bad(); + } + } else if (get_irn_mode(n) == mode_X) { + if (is_Block_dead(get_nodes_block(skip_Proj(n)))) { + /* Remove dead control flow -- early gigo(). */ + n = new_Bad(); + } } return n; } /** - * Optimize store after store and load atfter store. - * - * @todo FAILS for volatile entities + * Remove Id's. */ -static ir_node *equivalent_node_Store(ir_node *n) +static ir_node *equivalent_node_Id(ir_node *n) { ir_node *oldn = n; - /* remove unnecessary store. */ - ir_node *a = skip_Proj(get_Store_mem(n)); - ir_node *b = get_Store_ptr(n); - ir_node *c = skip_Proj(get_Store_value(n)); - - if (get_irn_op(a) == op_Store - && get_Store_ptr(a) == b - && skip_Proj(get_Store_value(a)) == c) { - /* We have twice exactly the same store -- a write after write. */ - n = a; DBG_OPT_WAW; - } else if (get_irn_op(c) == op_Load - && (a == c || skip_Proj(get_Load_mem(c)) == a) - && get_Load_ptr(c) == b ) { - /* We just loaded the value from the same memory, i.e., the store - doesn't change the memory -- a write after read. */ - a = get_Store_mem(n); - turn_into_tuple(n, 2); - set_Tuple_pred(n, pn_Store_M, a); - set_Tuple_pred(n, pn_Store_X_except, new_Bad()); DBG_OPT_WAR; - } + do { + n = get_Id_pred(n); + } while (get_irn_op(n) == op_Id); + + DBG_OPT_ID(oldn, n); return n; } -static ir_node *equivalent_node_Proj(ir_node *n) +/** + * optimize a Mux + */ +static ir_node *equivalent_node_Mux(ir_node *n) { - ir_node *oldn = n; - - ir_node *a = get_Proj_pred(n); + ir_node *oldn = n, *sel = get_Mux_sel(n); + tarval *ts = value_of(sel); - if ( get_irn_op(a) == op_Tuple) { - /* Remove the Tuple/Proj combination. */ - if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) { - n = get_Tuple_pred(a, get_Proj_proj(n)); DBG_OPT_TUPLE; - } else { - assert(0); /* This should not happen! */ - n = new_Bad(); + /* Mux(true, f, t) == t */ + if (ts == tarval_b_true) { + n = get_Mux_true(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); + } + /* Mux(false, f, t) == f */ + else if (ts == tarval_b_false) { + n = get_Mux_false(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C); + } + /* Mux(v, x, x) == x */ + else if (get_Mux_false(n) == get_Mux_true(n)) { + n = get_Mux_true(n); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ); + } + else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) { + ir_node *cmp = get_Proj_pred(sel); + long proj_nr = get_Proj_proj(sel); + ir_node *b = get_Mux_false(n); + ir_node *a = get_Mux_true(n); + + /* + * Note: normalization puts the constant on the right site, + * so we check only one case. + * + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the first one. + */ + if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) { + if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { + /* Mux(a CMP 0, X, a) */ + if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) { + /* Mux(a CMP 0, -a, a) */ + if (proj_nr == pn_Cmp_Eq) { + /* Mux(a == 0, -a, a) ==> -a */ + n = b; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(a != 0, -a, a) ==> a */ + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + } + else if (classify_Const(b) == CNST_NULL) { + /* Mux(a CMP 0, 0, a) */ + if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(a != 0, 0, a) ==> a */ + n = a; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + else if (proj_nr == pn_Cmp_Eq) { + /* Mux(a == 0, 0, a) ==> 0 */ + n = b; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } + } + } } - } else if (get_irn_mode(n) == mode_X && - is_Bad(get_nodes_Block(n))) { - /* Remove dead control flow -- early gigo. */ - n = new_Bad(); } return n; } /** - * Remove Id's. + * Optimize -a CMP -b into b CMP a. + * This works only for for modes where unary Minus + * cannot Overflow. + * Note that two-complement integers can Overflow + * so it will NOT work. */ -static ir_node *equivalent_node_Id(ir_node *n) +static ir_node *equivalent_node_Cmp(ir_node *n) { - ir_node *oldn = n; + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); - n = follow_Id(n); DBG_OPT_ID; + if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus && + !mode_overflow_on_unary_Minus(get_irn_mode(left))) { + left = get_Minus_op(left); + right = get_Minus_op(right); + set_Cmp_left(n, right); + set_Cmp_right(n, left); + } return n; } -/* -case iro_Mod, Quot, DivMod - DivMod allocates new nodes --> it's treated in transform node. - What about Quot, DivMod? -*/ +/** + * Remove Confirm nodes if setting is on. + */ +static ir_node *equivalent_node_Confirm(ir_node *n) +{ + if (get_Confirm_cmp(n) == pn_Cmp_Eq) { + ir_node *bound = get_Confirm_bound(n); + ir_op *op = get_irn_op(bound); + + /* + * Optimize a rare case: + * Confirm(x, '=', Const) ==> Const + */ + if (op == op_Const || op == op_SymConst) + return bound; + } + return get_opt_remove_Confirm() ? get_Confirm_value(n) : n; +} /** * equivalent_node() returns a node equivalent to input n. It skips all nodes that @@ -969,15 +1381,14 @@ equivalent_node(ir_node *n) */ static ir_op *firm_set_default_equivalent_node(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->equivalent_node = equivalent_node_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->equivalent_node = equivalent_node_##a; \ break switch (op->code) { CASE(Block); CASE(Jmp); - CASE(Cond); CASE(Or); CASE(Add); CASE(Eor); @@ -990,13 +1401,16 @@ 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(Load); /* dangerous */ - CASE(Store); /* dangerous, see todo */ CASE(Proj); CASE(Id); + CASE(Mux); + CASE(Cmp); + CASE(Confirm); default: op->equivalent_node = NULL; } @@ -1038,41 +1452,239 @@ optimize_preds(ir_node *n) { } /* end switch */ } +/** + * 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) +{ + 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; +} + +/** + * Do the AddSub optimization, then Transform Add(a,a) into Mul(a, 2) + * 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 *oldn = n; + + n = transform_node_AddSub(n); + + mode = get_irn_mode(n); + if (mode_is_num(mode)) { + ir_node *a = get_Add_left(n); + + if (a == get_Add_right(n)) { + ir_node *block = get_irn_n(n, -1); + + n = new_rd_Mul( + get_irn_dbg_info(n), + current_ir_graph, + block, + a, + new_r_Const_long(current_ir_graph, block, mode, 2), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); + } + else { + ir_node *b = get_Add_right(n); + + if (get_irn_op(a) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + b, + get_Minus_op(a), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + else if (get_irn_op(b) == op_Minus) { + n = new_rd_Sub( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + a, + get_Minus_op(b), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + } + } + } + return n; +} + +/** + * Do the AddSub optimization, then Transform Sub(0,a) into Minus(a). + */ +static ir_node *transform_node_Sub(ir_node *n) +{ + ir_mode *mode; + ir_node *oldn = n; + + n = transform_node_AddSub(n); + + mode = get_irn_mode(n); + if (mode_is_num(mode) && (classify_Const(get_Sub_left(n)) == CNST_NULL)) { + n = new_rd_Minus( + get_irn_dbg_info(n), + current_ir_graph, + get_irn_n(n, -1), + get_Sub_right(n), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); + } + + return n; +} + +/** + * Transform Mul(a,-1) into -a. + * Do architecture dependent optimizations on Mul nodes + */ +static ir_node *transform_node_Mul(ir_node *n) { + ir_node *oldn = n; + ir_mode *mode = get_irn_mode(n); + + if (mode_is_signed(mode)) { + ir_node *r = NULL; + ir_node *a = get_Mul_left(n); + ir_node *b = get_Mul_right(n); + + if (value_of(a) == get_mode_minus_one(mode)) + r = b; + else if (value_of(b) == get_mode_minus_one(mode)) + r = a; + if (r) { + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode); + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1); + return n; + } + } + return arch_dep_replace_mul_with_shifts(n); +} + +/** + * transform a Div Node + */ 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 */ if (tv != tarval_bad) { + value = new_Const(get_tarval_mode(tv), tv); + + DBG_OPT_CSTEVAL(n, value); + } + else /* Try architecture dependent optimization */ + value = arch_dep_replace_div_by_const(n); + + if (value != n) { /* Turn Div into a tuple (mem, bad, value) */ ir_node *mem = get_Div_mem(n); turn_into_tuple(n, 3); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_except, new_Bad()); - set_Tuple_pred(n, pn_Div_res, new_Const(get_tarval_mode(tv), tv)); + set_Tuple_pred(n, pn_Div_res, value); } return n; } +/** + * transform a Mod node + */ static ir_node *transform_node_Mod(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 0, as this may cause a exception */ if (tv != tarval_bad) { + value = new_Const(get_tarval_mode(tv), tv); + + DBG_OPT_CSTEVAL(n, value); + } + else /* Try architecture dependent optimization */ + value = arch_dep_replace_mod_by_const(n); + + if (value != n) { /* Turn Mod into a tuple (mem, bad, value) */ ir_node *mem = get_Mod_mem(n); + turn_into_tuple(n, 3); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, new_Const(get_tarval_mode(tv), tv)); + set_Tuple_pred(n, pn_Mod_res, value); } return n; } +/** + * transform a DivMod node + */ static ir_node *transform_node_DivMod(ir_node *n) { int evaluated = 0; @@ -1092,21 +1704,33 @@ static ir_node *transform_node_DivMod(ir_node *n) if (tb == get_mode_one(get_tarval_mode(tb))) { b = new_Const (mode, get_mode_null(mode)); evaluated = 1; - } else if (ta != tarval_bad) { + + DBG_OPT_CSTEVAL(n, b); + } + else if (ta != tarval_bad) { tarval *resa, *resb; resa = tarval_div (ta, tb); if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through - Jmp for X result!? */ + Jmp for X result!? */ resb = tarval_mod (ta, tb); if (resb == tarval_bad) return n; /* Causes exception! */ a = new_Const (mode, resa); b = new_Const (mode, resb); evaluated = 1; + + DBG_OPT_CSTEVAL(n, a); + DBG_OPT_CSTEVAL(n, b); + } + else { /* Try architecture dependent 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); @@ -1114,12 +1738,49 @@ static ir_node *transform_node_DivMod(ir_node *n) set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ set_Tuple_pred(n, pn_DivMod_res_div, a); set_Tuple_pred(n, pn_DivMod_res_mod, b); - assert(get_nodes_Block(n)); } return n; } +/** + * Optimize Abs(x) into x if x is Confirmed >= 0 + * Optimize Abs(x) into -x if x is Confirmed <= 0 + */ +static ir_node *transform_node_Abs(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_Abs_op(n); + value_classify sign = classify_value_sign(a); + + if (sign == VALUE_NEGATIVE) { + ir_mode *mode = get_irn_mode(n); + + /* + * We can replace the Abs by -x here. + * We even could add a new Confirm here. + * + * Note that -x would create a new node, so we could + * not run it in the equivalent_node() context. + */ + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, + get_irn_n(n, -1), a, mode); + + DBG_OPT_CONFIRM(oldn, n); + } + else if (sign == VALUE_POSITIVE) { + /* n is positive, Abs is not needed */ + n = a; + + DBG_OPT_CONFIRM(oldn, n); + } + + return n; +} + +/** + * transform a Cond node + */ static ir_node *transform_node_Cond(ir_node *n) { /* Replace the Cond by a Jmp if it branches on a constant @@ -1128,12 +1789,16 @@ static ir_node *transform_node_Cond(ir_node *n) ir_node *a = get_Cond_selector(n); tarval *ta = value_of(a); + /* we need block info which is not available in floating irgs */ + if (get_irg_pinned(current_ir_graph) == op_pin_state_floats) + return n; + if ((ta != tarval_bad) && (get_irn_mode(a) == mode_b) && (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. - Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ - jmp = new_r_Jmp(current_ir_graph, get_nodes_Block(n)); + Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ + jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); turn_into_tuple(n, 2); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); @@ -1143,195 +1808,484 @@ static ir_node *transform_node_Cond(ir_node *n) set_Tuple_pred(n, pn_Cond_true, new_Bad()); } /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n)); + add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); } else if ((ta != tarval_bad) && - (get_irn_mode(a) == mode_Iu) && - (get_Cond_kind(n) == dense) && - (get_opt_unreachable_code())) { + (get_irn_mode(a) == mode_Iu) && + (get_Cond_kind(n) == dense) && + (get_opt_unreachable_code())) { /* I don't want to allow Tuples smaller than the biggest Proj. Also this tuple might get really big... I generate the Jmp here, and remember it in link. Link is used when optimizing Proj. */ - set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_Block(n))); + set_irn_link(n, new_r_Jmp(current_ir_graph, get_nodes_block(n))); /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_Block(n)); + 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) - && (tarval_classify(computed_value(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { + && (get_irn_mode(a) == mode_b) + && (classify_tarval(value_of(get_Eor_right(a))) == TV_CLASSIFY_ONE)) { /* The Eor is a negate. Generate a new Cond without the negate, simulate the negate by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n), - get_Eor_left(a))); + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), + get_Eor_left(a))); } else if ((get_irn_op(a) == op_Not) - && (get_irn_mode(a) == mode_b)) { + && (get_irn_mode(a) == mode_b)) { /* A Not before the Cond. Generate a new Cond without the Not, simulate the Not by exchanging the results. */ - set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_Block(n), - get_Not_op(a))); + set_irn_link(n, new_r_Cond(current_ir_graph, get_nodes_block(n), + get_Not_op(a))); } return n; } +/** + * Transform an Eor. + */ static ir_node *transform_node_Eor(ir_node *n) { + ir_node *oldn = n; ir_node *a = get_Eor_left(n); ir_node *b = get_Eor_right(n); + ir_mode *mode = get_irn_mode(n); - if ((get_irn_mode(n) == mode_b) + if (a == b) { + /* a ^ a = 0 */ + n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), + mode, get_mode_null(mode)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A); + } + else if ((mode == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) + && (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) - && (tarval_classify (computed_value (b)) == TV_CLASSIFY_ONE)) + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode)); + + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL); + } + else if ((mode == mode_b) + && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) { /* The Eor is a Not. Replace it by a Not. */ /* ????!!!Extend to bitfield 1111111. */ - n = new_r_Not(current_ir_graph, get_nodes_Block(n), a, mode_b); + n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b); + + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + } return n; } /** - * Transfor a boolean Not. + * Transform a boolean Not. */ static ir_node *transform_node_Not(ir_node *n) { + ir_node *oldn = n; ir_node *a = get_Not_op(n); if ( (get_irn_mode(n) == mode_b) && (get_irn_op(a) == op_Proj) && (get_irn_mode(a) == mode_b) - && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) + && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) { /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(current_ir_graph, get_nodes_Block(n), get_Proj_pred(a), - mode_b, get_negated_pnc(get_Proj_proj(a))); + n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), + mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); + } return n; } /** - * Transform a Div/Mod/DivMod with a non-zero constant. Must be - * done here to avoid that this optimization runs more than once... + * Transform a Cast_type(Const) into a new Const_type */ -static ir_node *transform_node_Proj(ir_node *proj) +static ir_node *transform_node_Cast(ir_node *n) { + ir_node *oldn = n; + ir_node *pred = get_Cast_op(n); + type *tp = get_irn_type(n); + + if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) { + n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), + get_Const_tarval(pred), tp); + DBG_OPT_CSTEVAL(oldn, n); + } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) { + n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred), + get_SymConst_kind(pred), tp); + DBG_OPT_CSTEVAL(oldn, n); + } + + return n; +} + +/** + * 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) { ir_node *n = get_Proj_pred(proj); - ir_node *b; - tarval *tb; + ir_node *b = get_Div_right(n); + long proj_nr; - switch (get_irn_opcode(n)) { - case iro_Div: - b = get_Div_right(n); - tb = computed_value(b); + if (value_not_zero(b)) { + /* div(x, y) && y != 0 */ + proj_nr = get_Proj_proj(proj); - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* div(x, c) && c != 0 */ - ir_node *div, *proj; - ir_node *a = get_Div_left(n); - ir_node *mem = get_Div_mem(n); - int rem = get_optimize(); + /* this node may float */ + set_irn_pinned(n, op_pin_state_floats); - set_optimize(0); - { - div = new_rd_Div(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); + 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 */ + ir_node *res = get_Div_mem(n); + set_Div_mem(n, get_irg_no_mem(current_ir_graph)); - proj = new_r_Proj(current_ir_graph, get_nodes_Block(n), div, get_irn_mode(a), pn_Div_res); - } - set_optimize(rem); + return res; + } + } + return proj; +} + +/** + * Transform a Proj(Mod) with a non-zero value. + * Removes the exceptions and routes the memory to the NoMem node. + */ +static ir_node *transform_node_Proj_Mod(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + ir_node *b = get_Mod_right(n); + long proj_nr; + + if (value_not_zero(b)) { + /* mod(x, y) && y != 0 */ + proj_nr = get_Proj_proj(proj); - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Mod_M, mem); - set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, 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 */ + ir_node *res = get_Mod_mem(n); + set_Mod_mem(n, get_irg_no_mem(current_ir_graph)); + + return res; } - break; - case iro_Mod: - b = get_Mod_right(n); - tb = computed_value(b); + else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* mod(x, c) && c != 0 */ - ir_node *mod, *proj; - ir_node *a = get_Mod_left(n); - ir_node *mem = get_Mod_mem(n); - int rem = get_optimize(); + DBG_OPT_CSTEVAL(n, res); + return res; + } + } + return proj; +} + +/** + * Transform a Proj(DivMod) with a non-zero value. + * Removes the exceptions and routes the memory to the NoMem node. + */ +static ir_node *transform_node_Proj_DivMod(ir_node *proj) +{ + ir_node *n = get_Proj_pred(proj); + ir_node *b = get_DivMod_right(n); + long proj_nr; + + if (value_not_zero(b)) { + /* DivMod(x, y) && y != 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 */ + ir_node *res = get_DivMod_mem(n); + set_DivMod_mem(n, get_irg_no_mem(current_ir_graph)); + + return res; + } + else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) { + /* a % a = 0 if a != 0 */ + ir_mode *mode = get_irn_mode(proj); + ir_node *res = new_Const(mode, get_mode_null(mode)); - set_optimize(0); - { - mod = new_rd_Mod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); + DBG_OPT_CSTEVAL(n, res); + return res; + } + } + return proj; +} - proj = new_r_Proj(current_ir_graph, get_nodes_Block(n), mod, get_irn_mode(a), pn_Mod_res); +/** + * 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); + tarval *tb = value_of(b); + + if (tb != tarval_bad && mode_is_int(get_tarval_mode(tb))) { + /* we have a constant switch */ + long num = get_Proj_proj(proj); + + if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */ + if (get_tarval_long(tb) == num) { + /* Do NOT create a jump here, or we will have 2 control flow ops + * in a block. This case is optimized away in optimize_cf(). */ + return proj; + } + else { + /* this case will NEVER be taken, kill it */ + return new_Bad(); + } } - set_optimize(rem); + } + } + return proj; +} + +/** + * Normalizes and optimizes Cmp nodes. + */ +static ir_node *transform_node_Proj_Cmp(ir_node *proj) +{ + if (get_opt_reassociation()) { + ir_node *n = get_Proj_pred(proj); + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + ir_node *c = NULL; + tarval *tv = NULL; + int changed = 0; + ir_mode *mode = NULL; + long proj_nr = get_Proj_proj(proj); + + /* + * First step: normalize the compare op + * by placing the constant on the right site + * or moving the lower address node to the left. + * We ignore the case that both are constants + * this case should be optimized away. + */ + if (get_irn_op(right) == op_Const) + c = right; + else if (get_irn_op(left) == op_Const) { + c = left; + left = right; + right = c; + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; + } + else if (left > right) { + ir_node *t = left; + + left = right; + right = t; - turn_into_tuple(n, 3); - set_Tuple_pred(n, pn_Mod_M, mem); - set_Tuple_pred(n, pn_Mod_X_except, new_Bad()); - set_Tuple_pred(n, pn_Mod_res, proj); + proj_nr = get_inversed_pnc(proj_nr); + changed |= 1; } - break; - case iro_DivMod: - b = get_DivMod_right(n); - tb = computed_value(b); - - if (tb != tarval_bad && tarval_classify(tb) != TV_CLASSIFY_NULL) { /* DivMod(x, c) && c != 0 */ - ir_node *div_mod, *proj_div, *proj_mod; - ir_node *a = get_Mod_left(n); - ir_node *mem = get_Mod_mem(n); - int rem = get_optimize(); - - set_optimize(0); - { - div_mod = new_rd_DivMod(get_irn_dbg_info(n), current_ir_graph, - get_nodes_Block(n), get_irg_initial_mem(current_ir_graph), a, b); - - proj_div = new_r_Proj(current_ir_graph, get_nodes_Block(n), div_mod, get_irn_mode(a), pn_DivMod_res_div); - proj_mod = new_r_Proj(current_ir_graph, get_nodes_Block(n), div_mod, get_irn_mode(a), pn_DivMod_res_mod); + + /* + * Second step: Try to reduce the magnitude + * of a constant. This may help to generate better code + * later and may help to normalize more compares. + * Of course this is only possible for integer values. + */ + if (c) { + mode = get_irn_mode(c); + tv = get_Const_tarval(c); + + if (tv != tarval_bad) { + /* the following optimization is possible on modes without Overflow + * on Unary Minus or on == and !=: + * -a CMP c ==> a swap(CMP) -c + * + * Beware: for two-complement Overflow may occur, so only == and != can + * be optimized, see this: + * -MININT < 0 =/=> MININT > 0 !!! + */ + if (get_opt_constant_folding() && get_irn_op(left) == op_Minus && + (!mode_overflow_on_unary_Minus(mode) || + (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { + left = get_Minus_op(left); + tv = tarval_sub(get_mode_null(mode), tv); + + proj_nr = get_inversed_pnc(proj_nr); + changed |= 2; + } + + /* for integer modes, we have more */ + if (mode_is_int(mode)) { + /* Ne includes Unordered which is not possible on integers. + * However, frontends often use this wrong, so fix it here */ + if (proj_nr & pn_Cmp_Uo) { + proj_nr &= ~pn_Cmp_Uo; + set_Proj_proj(proj, proj_nr); + } + + /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ + if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { + tv = tarval_sub(tv, get_mode_one(mode)); + + proj_nr ^= pn_Cmp_Eq; + changed |= 2; + } + /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ + else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && + tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { + tv = tarval_add(tv, get_mode_one(mode)); + + proj_nr ^= pn_Cmp_Eq; + changed |= 2; + } + + /* the following reassociations work only for == and != */ + if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + + /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ + if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) { + right = get_Sub_right(left); + left = get_Sub_left(left); + + tv = value_of(right); + changed = 1; + } + + if (tv != tarval_bad) { + ir_op *op = get_irn_op(left); + + /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ + if (op == op_Sub) { + ir_node *c1 = get_Sub_right(left); + tarval *tv2 = value_of(c1); + + if (tv2 != tarval_bad) { + tv2 = tarval_add(tv, value_of(c1)); + + if (tv2 != tarval_bad) { + left = get_Sub_left(left); + tv = tv2; + changed |= 2; + } + } + } + /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ + else if (op == op_Add) { + ir_node *a_l = get_Add_left(left); + ir_node *a_r = get_Add_right(left); + ir_node *a; + tarval *tv2; + + if (get_irn_op(a_l) == op_Const) { + a = a_r; + tv2 = value_of(a_l); + } + else { + a = a_l; + tv2 = value_of(a_r); + } + + if (tv2 != tarval_bad) { + tv2 = tarval_sub(tv, tv2); + + if (tv2 != tarval_bad) { + left = a; + tv = tv2; + changed |= 2; + } + } + } + /* -a == c ==> a == -c, -a != c ==> a != -c */ + else if (op == op_Minus) { + tarval *tv2 = tarval_sub(get_mode_null(mode), tv); + + if (tv2 != tarval_bad) { + left = get_Minus_op(left); + tv = tv2; + changed |= 2; + } + } + } + } /* == or != */ + /* the following reassociations work only for <= */ + else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + if (tv != tarval_bad) { + ir_op *op = get_irn_op(left); + + /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */ + if (op == op_Abs) { + } + } + } + } /* mode_is_int */ } - set_optimize(rem); + } + + if (changed) { + ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ + + if (changed & 2) /* need a new Const */ + right = new_Const(mode, tv); - turn_into_tuple(n, 4); - set_Tuple_pred(n, pn_DivMod_M, mem); - set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); - set_Tuple_pred(n, pn_DivMod_res_div, proj_div); - set_Tuple_pred(n, pn_DivMod_res_mod, proj_mod); + /* create a new compare */ + n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, + left, right); + + set_Proj_pred(proj, n); + set_Proj_proj(proj, proj_nr); } - break; - default: - /* do nothing */ - return proj; } - - /* we have added a Tuple, optimize it for the current Proj away */ - return equivalent_node_Proj(proj); + return proj; } /** - * Transform a Store before a Store to the same address... - * Both nodes must be in the same block. - * - * @todo Check for volatile! Moreover, what if the first store - * has a exception handler while the other has not? + * Does all optimizations on nodes that must be done on it's Proj's + * because of creating new nodes. */ -static ir_node *transform_node_Store(ir_node *store) +static ir_node *transform_node_Proj(ir_node *proj) { - ir_node *pred = skip_Proj(get_Store_mem(store)); - ir_node *ptr = get_Store_ptr(store); + ir_node *n = get_Proj_pred(proj); + + switch (get_irn_opcode(n)) { + case iro_Div: + return transform_node_Proj_Div(proj); - if (get_irn_op(pred) == op_Store && - get_Store_ptr(pred) == ptr && - get_nodes_block(pred) == get_nodes_block(store)) { - /* the Store n is useless, as it is overwritten by the store store */ - ir_node *mem = get_Store_mem(pred); + case iro_Mod: + return transform_node_Proj_Mod(proj); + + case iro_DivMod: + return transform_node_Proj_DivMod(proj); + + case iro_Cond: + return transform_node_Proj_Cond(proj); - turn_into_tuple(pred, 2); - set_Tuple_pred(pred, pn_Store_M, mem); - set_Tuple_pred(pred, pn_Store_X_except, new_Bad()); + case iro_Cmp: + return transform_node_Proj_Cmp(proj); + + case iro_Tuple: + /* should not happen, but if it does will be optimized away */ + return equivalent_node_Proj(proj); + + default: + /* do nothing */ + return proj; } - return store; } /** @@ -1365,7 +2319,7 @@ static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) * AND c1 * OR */ -static ir_node *transform_node_Or(ir_node *or) +static ir_node *transform_node_Or_bf_store(ir_node *or) { ir_node *and, *c1; ir_node *or_l, *c2; @@ -1403,7 +2357,7 @@ static ir_node *transform_node_Or(ir_node *or) tv4 = get_Const_tarval(c4); tv = tarval_or(tv4, tv2); - if (tarval_classify(tv) != TV_CLASSIFY_ALL_ONE) { + if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) { /* have at least one 0 at the same bit position */ return or; } @@ -1421,7 +2375,7 @@ static ir_node *transform_node_Or(ir_node *or) } /* ok, all conditions met */ - block = get_nodes_block(or); + block = get_irn_n(or, -1); new_and = new_r_And(current_ir_graph, block, value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode); @@ -1432,7 +2386,350 @@ static ir_node *transform_node_Or(ir_node *or) set_Or_right(or, new_const); /* check for more */ - return transform_node_Or(or); + return transform_node_Or_bf_store(or); +} + +/** + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot + */ +static ir_node *transform_node_Or_Rot(ir_node *or) +{ + ir_mode *mode = get_irn_mode(or); + ir_node *shl, *shr, *block; + ir_node *irn, *x, *c1, *c2, *v, *sub, *n; + tarval *tv1, *tv2; + + if (! mode_is_int(mode)) + return or; + + shl = get_binop_left(or); + shr = get_binop_right(or); + + if (get_irn_op(shl) == op_Shr) { + if (get_irn_op(shr) != op_Shl) + return or; + + irn = shl; + shl = shr; + shr = irn; + } + else if (get_irn_op(shl) != op_Shl) + return or; + else if (get_irn_op(shr) != op_Shr) + return or; + + x = get_Shl_left(shl); + if (x != get_Shr_left(shr)) + return or; + + c1 = get_Shl_right(shl); + c2 = get_Shr_right(shr); + if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) { + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + tv2 = get_Const_tarval(c2); + if (! tarval_is_long(tv2)) + return or; + + if (get_tarval_long(tv1) + get_tarval_long(tv2) + != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_irn_n(or, -1); + + n = new_r_Rot(current_ir_graph, block, x, c1, mode); + + DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + else if (get_irn_op(c1) == op_Sub) { + v = c2; + sub = c1; + + if (get_Sub_right(sub) != v) + return or; + + c1 = get_Sub_left(sub); + if (get_irn_op(c1) != op_Const) + return or; + + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_nodes_block(or); + + /* a Rot right is not supported, so use a rot left */ + n = new_r_Rot(current_ir_graph, block, x, sub, mode); + + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + else if (get_irn_op(c2) == op_Sub) { + v = c1; + sub = c2; + + c1 = get_Sub_left(sub); + if (get_irn_op(c1) != op_Const) + return or; + + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return or; + + if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + return or; + + /* yet, condition met */ + block = get_irn_n(or, -1); + + /* a Rot Left */ + n = new_r_Rot(current_ir_graph, block, x, v, mode); + + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + return n; + } + + return or; +} + +/** + * Optimize an Or + */ +static ir_node *transform_node_Or(ir_node *or) +{ + or = transform_node_Or_bf_store(or); + or = transform_node_Or_Rot(or); + + return or; +} + +/* forward */ +static ir_node *transform_node(ir_node *n); + +/** + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl. + * + * Should be moved to reassociation? + */ +static ir_node *transform_node_shift(ir_node *n) +{ + ir_node *left, *right; + tarval *tv1, *tv2, *res; + ir_mode *mode; + int modulo_shf, flag; + + left = get_binop_left(n); + + /* different operations */ + if (get_irn_op(left) != get_irn_op(n)) + return n; + + right = get_binop_right(n); + tv1 = value_of(right); + if (tv1 == tarval_bad) + return n; + + tv2 = value_of(get_binop_right(left)); + if (tv2 == tarval_bad) + return n; + + res = tarval_add(tv1, tv2); + + /* 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) & pn_Cmp_Lt) + flag = 1; + } + else + flag = 1; + + if (flag) { + /* ok, we can replace it */ + ir_node *in[2], *irn, *block = get_irn_n(n, -1); + + in[0] = get_binop_left(left); + in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + + irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + + DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + + return transform_node(irn); + } + return n; +} + +#define transform_node_Shr transform_node_shift +#define transform_node_Shrs transform_node_shift +#define transform_node_Shl transform_node_shift + +/** + * Remove dead blocks and nodes in dead blocks + * in keep alive list. We do not generate a new End node. + */ +static ir_node *transform_node_End(ir_node *n) { + int i, n_keepalives = get_End_n_keepalives(n); + + for (i = 0; i < n_keepalives; ++i) { + ir_node *ka = get_End_keepalive(n, i); + if (is_Block(ka)) { + if (is_Block_dead(ka)) { + set_End_keepalive(n, i, new_Bad()); + } + } + else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) + set_End_keepalive(n, i, new_Bad()); + } + return n; +} + +/** + * Optimize a Mux into some simpler cases. + */ +static ir_node *transform_node_Mux(ir_node *n) +{ + ir_node *oldn = n, *sel = get_Mux_sel(n); + ir_mode *mode = get_irn_mode(n); + + if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) { + ir_node *cmp = get_Proj_pred(sel); + long proj_nr = get_Proj_proj(sel); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(n); + + if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) { + ir_node *block = get_irn_n(n, -1); + + /* + * Note: normalization puts the constant on the right site, + * so we check only one case. + * + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the first one. + */ + if (get_irn_op(f) == op_Minus && + get_Minus_op(f) == t && + get_Cmp_left(cmp) == t) { + + if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { + /* Mux(a >=/> 0, -a, a) ==> Abs(a) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + t, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + t, mode); + n = new_rd_Minus(get_irn_dbg_info(n), + current_ir_graph, + block, + n, mode); + + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + } + else if (get_irn_op(t) == op_Minus && + get_Minus_op(t) == f && + get_Cmp_left(cmp) == f) { + + if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) { + /* Mux(a <=/< 0, a, -a) ==> Abs(a) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + f, mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) { + /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */ + n = new_rd_Abs(get_irn_dbg_info(n), + current_ir_graph, + block, + f, mode); + n = new_rd_Minus(get_irn_dbg_info(n), + current_ir_graph, + block, + n, mode); + + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); + return n; + } + } + + if (mode_is_int(mode) && mode_is_signed(mode) && + get_mode_arithmetic(mode) == irma_twos_complement) { + ir_node *x = get_Cmp_left(cmp); + + /* the following optimization works only with signed integer two-complement mode */ + + if (mode == get_irn_mode(x)) { + /* + * FIXME: this restriction is two rigid, as it would still + * work if mode(x) = Hs and mode == Is, but at least it removes + * all wrong cases. + */ + if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) && + classify_Const(t) == CNST_ALL_ONE && + classify_Const(f) == CNST_NULL) { + /* + * Mux(x:T Shrs(x, sizeof_bits(T) - 1) + * Conditions: + * T must be signed. + */ + n = new_rd_Shrs(get_irn_dbg_info(n), + current_ir_graph, block, x, + new_r_Const_long(current_ir_graph, block, mode_Iu, + get_mode_size_bits(mode) - 1), + mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + return n; + } + else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) && + classify_Const(t) == CNST_ONE && + classify_Const(f) == CNST_NULL) { + /* + * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1) + * Conditions: + * T must be signed. + */ + n = new_rd_Shr(get_irn_dbg_info(n), + current_ir_graph, block, + new_r_Minus(current_ir_graph, block, x, mode), + new_r_Const_long(current_ir_graph, block, mode_Iu, + get_mode_size_bits(mode) - 1), + mode); + DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR); + return n; + } + } + } + } + } + return arch_transform_node_Mux(n); } /** @@ -1453,23 +2750,33 @@ static ir_node *transform_node(ir_node *n) */ static ir_op *firm_set_default_transform_node(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->transform_node = transform_node_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->transform_node = transform_node_##a; \ break switch (op->code) { + CASE(Add); + CASE(Sub); + CASE(Mul); CASE(Div); CASE(Mod); CASE(DivMod); + CASE(Abs); CASE(Cond); CASE(Eor); CASE(Not); + CASE(Cast); CASE(Proj); - CASE(Store); + CASE(Sel); CASE(Or); + CASE(Shr); + CASE(Shrs); + CASE(Shl); + CASE(End); + CASE(Mux); default: - op->transform_node = NULL; + op->transform_node = NULL; } return op; @@ -1483,66 +2790,98 @@ 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); + 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); + 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) + 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); } +/** 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)); + return (get_irn_free_attr(a).where != get_irn_free_attr(b).where) + || (get_irn_free_attr(a).type != get_irn_free_attr(b).type); } +/** 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); + return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num) + || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p) + || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp); } +/** 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)); + return (get_irn_call_attr(a) != get_irn_call_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); + 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); + 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 || + get_Load_volatility(b) == volatility_is_volatile) + /* NEVER do CSE on volatile Loads */ + return 1; + + 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_Store_volatility(b) == volatility_is_volatile); +} + +/** Compares the attributes of two Confirm nodes. */ +static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) +{ + return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); } /** @@ -1550,9 +2889,9 @@ static int node_cmp_attr_Cast(ir_node *a, ir_node *b) */ static ir_op *firm_set_default_node_cmp_attr(ir_op *op) { -#define CASE(a) \ - case iro_##a: \ - op->node_cmp_attr = node_cmp_attr_##a; \ +#define CASE(a) \ + case iro_##a: \ + op->node_cmp_attr = node_cmp_attr_##a; \ break switch (op->code) { @@ -1563,10 +2902,12 @@ 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); + CASE(Load); + CASE(Store); + CASE(Confirm); default: op->node_cmp_attr = NULL; } @@ -1593,20 +2934,20 @@ 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 pinned nodes: */ - if (!get_opt_global_cse() || (get_op_pinned(get_irn_op(a)) == pinned)) { - if (get_irn_n(a, -1) != get_irn_n(b, -1)) + /* for block-local cse and op_pin_state_pinned nodes: */ + if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) { + if (get_irn_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; /* @@ -1630,41 +2971,39 @@ 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); } /** @@ -1682,14 +3021,15 @@ 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); + ir_node *r = get_binop_right(n); + /* for commutative operators perform a OP b == b OP a */ - if (get_binop_left(n) > get_binop_right(n)) { - ir_node *h = get_binop_left(n); - set_binop_left(n, get_binop_right(n)); - set_binop_right(n, h); + if (l > r) { + set_binop_left(n, r); + set_binop_right(n, l); } } } @@ -1697,20 +3037,23 @@ identify (pset *value_table, ir_node *n) o = pset_find (value_table, n, ir_node_hash (n)); if (!o) return n; + DBG_OPT_CSE(n, o); + return o; } /** - * During construction we set the pinned flag in the graph right when the + * During construction we set the op_pin_state_pinned flag in the graph right when the * optimization is performed. The flag turning on procedure global cse could * be changed between two allocations. This way we are safe. */ static INLINE ir_node * identify_cons (pset *value_table, ir_node *n) { ir_node *old = n; + n = identify(value_table, n); if (get_irn_n(old, -1) != get_irn_n(n, -1)) - set_irg_pinned(current_ir_graph, floats); + set_irg_pinned(current_ir_graph, op_pin_state_floats); return n; } @@ -1720,23 +3063,39 @@ 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); } /** @@ -1747,17 +3106,22 @@ static INLINE ir_node * gigo (ir_node *node) { int i, irn_arity; - ir_op* op = get_irn_op(node); + ir_op *op = get_irn_op(node); /* remove garbage blocks by looking at control flow that leaves the block and replacing the control flow by Bad. */ if (get_irn_mode(node) == mode_X) { - ir_node *block = get_nodes_block(node); - if (op == op_End) return node; /* Don't optimize End, may have Bads. */ - if (get_irn_op(block) == op_Block && get_Block_matured(block)) { + 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)) { irn_arity = get_irn_arity(block); for (i = 0; i < irn_arity; i++) { - if (!is_Bad(get_irn_n(block, i))) break; + if (!is_Bad(get_irn_n(block, i))) break; } if (i == irn_arity) return new_Bad(); } @@ -1767,7 +3131,15 @@ gigo (ir_node *node) blocks predecessors is dead. */ if ( op != op_Block && op != op_Phi && op != op_Tuple) { irn_arity = get_irn_arity(node); - for (i = -1; i < irn_arity; i++) { + + /* + * Beware: we can only read the block of a non-floating node. + */ + if (is_irn_pinned_in_irg(node) && + is_Block_dead(get_nodes_block(node))) + return new_Bad(); + + for (i = 0; i < irn_arity; i++) { if (is_Bad(get_irn_n(node, i))) { return new_Bad(); } @@ -1796,60 +3168,84 @@ gigo (ir_node *node) * reference this one, i.e., right after construction of a node. */ ir_node * -optimize_node (ir_node *n) +optimize_node(ir_node *n) { tarval *tv; ir_node *oldn = n; opcode iro = get_irn_opcode(n); - /* Allways optimize Phi nodes: part of the construction. */ + /* Always optimize Phi nodes: part of the construction. */ if ((!get_opt_optimize()) && (iro != iro_Phi)) return n; /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && (get_irn_mode(n) != mode_T)) { /* try to evaluate */ - tv = computed_value (n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { - /* - * we MUST copy the node here temparary, because it's still needed - * for DBG_OPT_ALGSIM0 - */ - int node_size = offsetof(ir_node, attr) + n->op->attr_size; - ir_node *x = alloca(node_size); - - memcpy(x, n, node_size); - oldn = x; + tv = computed_value(n); + if (tv != tarval_bad) { + ir_node *nw; + type *old_tp = get_irn_type(n); + int i, arity = get_irn_arity(n); + int node_size; + + /* + * Try to recover the type of the new expression. + */ + for (i = 0; i < arity && !old_tp; ++i) + old_tp = get_irn_type(get_irn_n(n, i)); + + /* + * we MUST copy the node here temporary, because it's still needed + * for DBG_OPT_CSTEVAL + */ + 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); + + /* ARG, copy the in array, we need it for statistics */ + memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); + + /* note the inplace edges module */ + edges_node_deleted(n, current_ir_graph); /* evaluation was successful -- replace the node. */ - obstack_free (current_ir_graph->obst, n); - n = new_Const (get_tarval_mode (tv), tv); - DBG_OPT_ALGSIM0; - return n; + obstack_free(current_ir_graph->obst, n); + nw = new_Const(get_tarval_mode (tv), tv); + + if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv)) + set_Const_type(nw, old_tp); + DBG_OPT_CSTEVAL(oldn, nw); + return nw; } } } /* remove unnecessary nodes */ if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || - (iro == iro_Proj) || - (iro == iro_Block) ) /* Flags tested local. */ + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || + (iro == iro_Proj) || + (iro == iro_Block) ) /* Flags tested local. */ n = equivalent_node (n); optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ - /** common subexpression elimination **/ - /* Checks whether n is already available. */ - /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common - subexpressions within a block. */ + /* Common Subexpression Elimination. + * + * Checks whether n is already available. + * The block input is used to distinguish different subexpressions. Right + * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common + * subexpressions within a block. + */ if (get_opt_cse()) n = identify_cons (current_ir_graph->value_table, n); if (n != oldn) { + edges_node_deleted(oldn, current_ir_graph); + /* We found an existing, better node, so we can deallocate the old node. */ obstack_free (current_ir_graph->obst, oldn); @@ -1860,15 +3256,16 @@ optimize_node (ir_node *n) free the node. */ iro = get_irn_opcode(n); if (get_opt_constant_folding() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ + (iro == iro_Cond) || + (iro == iro_Proj) || + (iro == iro_Sel)) /* Flags tested local. */ n = transform_node (n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ n = gigo (n); - /* Now we have a legal, useful node. Enter it in hash table for cse */ + /* Now we have a legal, useful node. Enter it in hash table for CSE */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { n = identify_remember (current_ir_graph->value_table, n); } @@ -1878,7 +3275,7 @@ optimize_node (ir_node *n) /** - * These optimizations never deallocate nodes. This can cause dead + * These optimizations never deallocate nodes (in place). This can cause dead * nodes lying on the obstack. Remove these by a dead node elimination, * i.e., a copying garbage collection. */ @@ -1891,59 +3288,64 @@ optimize_in_place_2 (ir_node *n) if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n; - /* if not optimize return n */ - if (n == NULL) { - assert(0); - /* Here this is possible. Why? */ - return n; - } - - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { - /* constants can not be evaluated */ - if (iro != iro_Const) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { /* try to evaluate */ - tv = computed_value (n); - if ((get_irn_mode(n) != mode_T) && (tv != tarval_bad)) { + tv = computed_value(n); + if (tv != tarval_bad) { /* evaluation was successful -- replace the node. */ - n = new_Const (get_tarval_mode (tv), tv); - DBG_OPT_ALGSIM0; - return n; + 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(get_tarval_mode(tv), tv); + + if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv)) + set_Const_type(n, old_tp); + + DBG_OPT_CSTEVAL(oldn, n); + return n; } } } /* remove unnecessary nodes */ - /*if (get_opt_constant_folding()) */ 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); + n = equivalent_node(n); optimize_preds(n); /* do node specific optimizations of nodes predecessors. */ /** common subexpression elimination **/ /* Checks whether n is already available. */ /* The block input is used to distinguish different subexpressions. Right - now all nodes are pinned to blocks, i.e., the cse only finds common + now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) { - n = identify (current_ir_graph->value_table, n); + n = identify(current_ir_graph->value_table, n); } /* Some more constant expression evaluation. */ iro = get_irn_opcode(n); if (get_opt_constant_folding() || (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ - n = transform_node (n); + (iro == iro_Proj) || + (iro == iro_Sel)) /* Flags tested local. */ + n = transform_node(n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ - n = gigo (n); + n = gigo(n); /* Now we can verify the node, as it has no dead inputs any more. */ irn_vrfy(n); @@ -1952,7 +3354,7 @@ optimize_in_place_2 (ir_node *n) Blocks should be unique anyways. (Except the successor of start: is cse with the start block!) */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) - n = identify_remember (current_ir_graph->value_table, n); + n = identify_remember(current_ir_graph->value_table, n); return n; } @@ -1967,9 +3369,10 @@ optimize_in_place (ir_node *n) assert(get_irg_phase_state(current_ir_graph) != phase_building); if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, floats); + 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); + /* Maybe we could also test whether optimizing the node can change the control graph. */ if (get_irg_dom_state(current_ir_graph) == dom_consistent) @@ -1986,6 +3389,7 @@ ir_op *firm_set_default_operations(ir_op *op) op = firm_set_default_equivalent_node(op); op = firm_set_default_transform_node(op); op = firm_set_default_node_cmp_attr(op); + op = firm_set_default_get_type(op); return op; }