X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=631570b4ba1d124a0889aeb26dbc95f955b77be3;hb=8afe079f9adaa453d795b711f7a0dab7e3c27156;hp=8b5719abb6d82790998c9fee9561c86d8a2a37cd;hpb=6b91f5d25b8eabae39539fa9cac85d6be0dc62dc;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 8b5719abb..631570b4b 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -53,6 +53,25 @@ /* Make types visible to allow most efficient access */ #include "entity_t.h" +/** + * Returns the tarval of a Const node or tarval_bad for all other nodes. + */ +static tarval *default_value_of(const ir_node *n) { + if (is_Const(n)) + return get_Const_tarval(n); /* might return tarval_bad */ + else + return tarval_bad; +} + +value_of_func value_of_ptr = default_value_of; + +void set_value_of_func(value_of_func func) { + if (func != NULL) + value_of_ptr = func; + else + value_of_ptr = default_value_of; +} + /** * Return the value of a Constant. */ @@ -178,7 +197,7 @@ static tarval *computed_value_Minus(ir_node *n) { ir_node *a = get_Minus_op(n); tarval *ta = value_of(a); - if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a))) + if (ta != tarval_bad) return tarval_neg(ta); return tarval_bad; @@ -418,20 +437,20 @@ static tarval *computed_value_Shrs(ir_node *n) { } /* computed_value_Shrs */ /** - * Return the value of a Rot. + * Return the value of a Rotl. */ -static tarval *computed_value_Rot(ir_node *n) { - ir_node *a = get_Rot_left(n); - ir_node *b = get_Rot_right(n); +static tarval *computed_value_Rotl(ir_node *n) { + ir_node *a = get_Rotl_left(n); + ir_node *b = get_Rotl_right(n); tarval *ta = value_of(a); tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - return tarval_rot (ta, tb); + return tarval_rotl(ta, tb); } return tarval_bad; -} /* computed_value_Rot */ +} /* computed_value_Rotl */ /** * Return the value of a Conv. @@ -540,7 +559,7 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) { && ( (/* ab is NULL */ is_Const(ab) && mode_is_reference(get_irn_mode(ab)) - && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab)))) + && is_Const_null(ab)) || (/* ab is other Alloc */ is_Proj(ab) && mode_is_reference(get_irn_mode(ab)) @@ -549,7 +568,7 @@ static tarval *computed_value_Proj_Cmp(ir_node *n) { || (/* aa is NULL and aba is Alloc */ is_Const(aa) && mode_is_reference(get_irn_mode(aa)) - && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa))) + && is_Const_null(aa) && is_Proj(ab) && mode_is_reference(get_irn_mode(ab)) && is_Alloc(aba))) @@ -637,8 +656,16 @@ static tarval *computed_value_Psi(ir_node *n) { * 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; + /* + * Beware: we might produce Phi(Confirm(x == true), Confirm(x == false)). + * Do NOT optimize them away (CondEval wants them), so wait until + * remove_confirm is activated. + */ + if (get_opt_remove_confirm()) { + return get_Confirm_cmp(n) == pn_Cmp_Eq ? + value_of(get_Confirm_bound(n)) : tarval_bad; + } + return tarval_bad; } /* computed_value_Confirm */ /** @@ -687,7 +714,7 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Rot); + CASE(Rotl); CASE(Carry); CASE(Borrow); CASE(Conv); @@ -722,10 +749,16 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops static ir_node *equivalent_node_Block(ir_node *n) { ir_node *oldn = n; - int n_preds = get_Block_n_cfgpreds(n); + int n_preds; + + /* don't optimize dead blocks */ + if (is_Block_dead(n)) + return n; + + n_preds = get_Block_n_cfgpreds(n); - /* The Block constructor does not call optimize, but mature_immBlock - calls the optimization. */ + /* The Block constructor does not call optimize, but mature_immBlock() + calls the optimization. */ assert(get_Block_matured(n)); /* Straightening: a single entry Block following a single exit Block @@ -862,7 +895,53 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) /** * Eor is commutative and has neutral 0. */ -#define equivalent_node_Eor equivalent_node_neutral_zero +static ir_node *equivalent_node_Eor(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a; + ir_node *b; + + n = equivalent_node_neutral_zero(n); + if (n != oldn) return n; + + a = get_Eor_left(n); + b = get_Eor_right(n); + + if (is_Eor(a)) { + ir_node *aa = get_Eor_left(a); + ir_node *ab = get_Eor_right(a); + + if (aa == b) { + /* (a ^ b) ^ a -> b */ + n = ab; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } else if (ab == b) { + /* (a ^ b) ^ b -> a */ + n = aa; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } + } + if (is_Eor(b)) { + ir_node *ba = get_Eor_left(b); + ir_node *bb = get_Eor_right(b); + + if (ba == a) { + /* a ^ (a ^ b) -> b */ + n = bb; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } else if (bb == a) { + /* a ^ (b ^ a) -> b */ + n = ba; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_EOR_A_B_A); + return n; + } + } + + return n; +} /* * Optimize a - 0 and (a - x) + x (for modes with wrap-around). @@ -878,14 +957,14 @@ static ir_node *equivalent_node_Add(ir_node *n) { ir_node *left, *right; ir_mode *mode = get_irn_mode(n); - /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ - if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) - return n; - n = equivalent_node_neutral_zero(n); if (n != oldn) return n; + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return n; + left = get_Add_left(n); right = get_Add_right(n); @@ -935,7 +1014,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) { #define equivalent_node_Shl equivalent_node_left_zero #define equivalent_node_Shr equivalent_node_left_zero #define equivalent_node_Shrs equivalent_node_left_zero -#define equivalent_node_Rot equivalent_node_left_zero +#define equivalent_node_Rotl equivalent_node_left_zero /** * Optimize a - 0 and (a + x) - x (for modes with wrap-around). @@ -971,7 +1050,7 @@ static ir_node *equivalent_node_Sub(ir_node *n) { /** - * Optimize an "idempotent unary op", ie op(op(n)) = n. + * Optimize an "self-inverse unary op", ie op(op(n)) = n. * * @todo * -(-a) == a, but might overflow two times. @@ -1159,42 +1238,67 @@ static ir_node *equivalent_node_And(ir_node *n) { static ir_node *equivalent_node_Conv(ir_node *n) { ir_node *oldn = n; ir_node *a = get_Conv_op(n); - ir_node *b; ir_mode *n_mode = get_irn_mode(n); ir_mode *a_mode = get_irn_mode(a); +restart: if (n_mode == a_mode) { /* No Conv necessary */ if (get_Conv_strict(n)) { /* special case: the predecessor might be a also a Conv */ if (is_Conv(a)) { if (! get_Conv_strict(a)) { /* first one is not strict, kick it */ - set_Conv_op(n, get_Conv_op(a)); - return n; + a = get_Conv_op(a); + a_mode = get_irn_mode(a); + set_Conv_op(n, a); + goto restart; + } + /* else both are strict conv, second is superfluous */ + } else if (is_Proj(a)) { + ir_node *pred = get_Proj_pred(a); + if (is_Load(pred)) { + /* loads always return with the exact precision of n_mode */ + assert(get_Load_mode(pred) == n_mode); + return a; } - /* else both are strict conv, second is superflous */ - } else { - /* leave strict floating point Conv's */ - return n; } + + /* leave strict floating point Conv's */ + return n; } n = a; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV); } else if (is_Conv(a)) { /* Conv(Conv(b)) */ - ir_mode *b_mode; - - b = get_Conv_op(a); - n_mode = get_irn_mode(n); - b_mode = get_irn_mode(b); - + ir_node *b = get_Conv_op(a); + ir_mode *b_mode = get_irn_mode(b); + + if (get_Conv_strict(n) && get_Conv_strict(a)) { + /* both are strict conv */ + if (smaller_mode(a_mode, n_mode)) { + /* both are strict, but the first is smaller, so + the second cannot remove more precision, remove the + strict bit */ + set_Conv_strict(n, 0); + } + } if (n_mode == b_mode) { - if (n_mode == mode_b) { - n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - } else if (mode_is_int(n_mode)) { - if (smaller_mode(b_mode, a_mode)){ - n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + if (! get_Conv_strict(n) && ! get_Conv_strict(a)) { + if (n_mode == mode_b) { + n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + } else if (get_mode_arithmetic(n_mode) == get_mode_arithmetic(a_mode)) { + if (smaller_mode(b_mode, a_mode)) { + n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */ + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + } + } + } + if (is_Conv(b)) { + if (smaller_mode(b_mode, a_mode)) { + if (get_Conv_strict(n)) + set_Conv_strict(b, 1); + n = b; /* ConvA(ConvB(ConvA(...))) == ConvA(...) */ DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); } } @@ -1229,7 +1333,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) { int i, n_preds; ir_node *oldn = n; - ir_node *block = NULL; /* to shutup gcc */ + ir_node *block; ir_node *first_val = NULL; /* to shutup gcc */ if (!get_opt_normalize()) return n; @@ -1237,8 +1341,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) { n_preds = get_Phi_n_preds(n); block = get_nodes_block(n); - /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!! - assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */ if ((is_Block_dead(block)) || /* Control dead */ (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */ return new_Bad(); /* in the Start Block. */ @@ -1254,7 +1356,15 @@ static ir_node *equivalent_node_Phi(ir_node *n) { for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); if ( (first_val != n) /* not self pointer */ -#if 1 +#if 0 + /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad + * predecessors. Then, Phi nodes in dead code might be removed, causing + * nodes pointing to themself (Add's for instance). + * This is really bad and causes endless recursions in several + * code pathes, so we do NOT optimize such a code. + * This is not that bad as it sounds, optimize_cf() removes bad control flow + * (and bad Phi predecessors), so live code is optimized later. + */ && (! is_Bad(first_val)) #endif ) { /* value not dead */ @@ -1273,7 +1383,8 @@ static ir_node *equivalent_node_Phi(ir_node *n) { ir_node *scnd_val = get_Phi_pred(n, i); if ( (scnd_val != n) && (scnd_val != first_val) -#if 1 +#if 0 + /* see above */ && (! is_Bad(scnd_val)) #endif ) { @@ -1296,45 +1407,36 @@ static ir_node *equivalent_node_Phi(ir_node *n) { * themselves. */ static ir_node *equivalent_node_Sync(ir_node *n) { - int i, n_preds; - - ir_node *oldn = n; - ir_node *first_val = NULL; /* to shutup gcc */ - - if (!get_opt_normalize()) return n; + int arity = get_Sync_n_preds(n); + int i; - n_preds = get_Sync_n_preds(n); + for (i = 0; i < arity;) { + ir_node *pred = get_Sync_pred(n, i); + int j; - /* Find first non-self-referencing input */ - for (i = 0; i < n_preds; ++i) { - first_val = get_Sync_pred(n, i); - if ((first_val != n) /* not self pointer */ && - (! is_Bad(first_val)) - ) { /* value not dead */ - break; /* then found first value. */ + /* Remove Bad predecessors */ + if (is_Bad(pred)) { + del_Sync_n(n, i); + --arity; + continue; } - } - - if (i >= n_preds) - /* A totally Bad or self-referencing Sync (we didn't break the above loop) */ - return new_Bad(); - /* search the rest of inputs, determine if any of these - are non-self-referencing */ - while (++i < n_preds) { - ir_node *scnd_val = get_Sync_pred(n, i); - if ((scnd_val != n) && - (scnd_val != first_val) && - (! is_Bad(scnd_val)) - ) - break; + /* Remove duplicate predecessors */ + for (j = 0;; ++j) { + if (j >= i) { + ++i; + break; + } + if (get_Sync_pred(n, j) == pred) { + del_Sync_n(n, i); + --arity; + break; + } + } } - if (i >= n_preds) { - /* Fold, if no multiple distinct non-self-referencing inputs */ - n = first_val; - DBG_OPT_SYNC(oldn, n); - } + if (arity == 0) return new_Bad(); + if (arity == 1) return get_Sync_pred(n, 0); return n; } /* equivalent_node_Sync */ @@ -1442,41 +1544,54 @@ static ir_node *equivalent_node_Mux(ir_node *n) else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) { ir_node *cmp = get_Proj_pred(sel); long proj_nr = get_Proj_proj(sel); - ir_node *b = get_Mux_false(n); - ir_node *a = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(n); /* - * Note: normalization puts the constant on the right site, - * so we check only one case. - * * Note further that these optimization work even for floating point * with NaN's because -NaN == NaN. * However, if +0 and -0 is handled differently, we cannot use the first one. */ - if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) { - ir_node *cmp_r = get_Cmp_right(cmp); - if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - /* Mux(a CMP 0, X, a) */ - if (is_Minus(b) && get_Minus_op(b) == a) { - /* Mux(a CMP 0, -a, a) */ - if (proj_nr == pn_Cmp_Eq) { - /* Mux(a == 0, -a, a) ==> -a */ - n = b; + if (is_Cmp(cmp)) { + ir_node *const cmp_l = get_Cmp_left(cmp); + ir_node *const cmp_r = get_Cmp_right(cmp); + + switch (proj_nr) { + case pn_Cmp_Eq: + if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */ + (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */ + n = f; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { - /* Mux(a != 0, -a, a) ==> a */ - n = a; + return n; + } + break; + + case pn_Cmp_Lg: + case pn_Cmp_Ne: + if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */ + (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */ + n = t; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + return n; } - } else if (is_Const(b) && is_Const_null(b)) { - /* Mux(a CMP 0, 0, a) */ - if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { - /* Mux(a != 0, 0, a) ==> a */ - n = a; + break; + } + + /* + * Note: normalization puts the constant on the right side, + * so we check only one case. + */ + if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) { + /* Mux(t CMP 0, X, t) */ + if (is_Minus(f) && get_Minus_op(f) == t) { + /* Mux(t CMP 0, -t, t) */ + if (proj_nr == pn_Cmp_Eq) { + /* Mux(t == 0, -t, t) ==> -t */ + n = f; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } else if (proj_nr == pn_Cmp_Eq) { - /* Mux(a == 0, 0, a) ==> 0 */ - n = b; + } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { + /* Mux(t != 0, -t, t) ==> t */ + n = t; DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } @@ -1535,19 +1650,9 @@ static ir_node *equivalent_node_Confirm(ir_node *n) { */ n = pred; } - if (pnc == pn_Cmp_Eq) { - ir_node *bound = get_Confirm_bound(n); - - /* - * Optimize a rare case: - * Confirm(x, '=', Constlike) ==> Constlike - */ - if (is_irn_constlike(bound)) { - DBG_OPT_CONFIRM(n, bound); - return bound; - } - } - return get_opt_remove_confirm() ? get_Confirm_value(n) : n; + if (get_opt_remove_confirm()) + return get_Confirm_value(n); + return n; } /** @@ -1574,34 +1679,27 @@ static ir_node *equivalent_node_CopyB(ir_node *n) { * Optimize Bounds(idx, idx, upper) into idx. */ static ir_node *equivalent_node_Bound(ir_node *n) { - ir_node *idx = get_Bound_index(n); - ir_node *lower = get_Bound_lower(n); + ir_node *idx = get_Bound_index(n); + ir_node *pred = skip_Proj(idx); int ret_tuple = 0; - /* By definition lower < upper, so if idx == lower --> - lower <= idx && idx < upper */ - if (idx == lower) { - /* Turn Bound into a tuple (mem, jmp, bad, idx) */ - ret_tuple = 1; - } else { - ir_node *pred = skip_Proj(idx); - - if (get_irn_op(pred) == op_Bound) { + if (is_Bound(pred)) { + /* + * idx was Bounds checked in the same MacroBlock previously, + * it is still valid if lower <= pred_lower && pred_upper <= upper. + */ + ir_node *lower = get_Bound_lower(n); + ir_node *upper = get_Bound_upper(n); + if (get_Bound_lower(pred) == lower && + get_Bound_upper(pred) == upper && + get_irn_MacroBlock(n) == get_irn_MacroBlock(pred)) { /* - * idx was Bounds_check previously, it is still valid if - * lower <= pred_lower && pred_upper <= upper. + * One could expect that we simply return the previous + * Bound here. However, this would be wrong, as we could + * add an exception Proj to a new location then. + * So, we must turn in into a tuple. */ - ir_node *upper = get_Bound_upper(n); - if (get_Bound_lower(pred) == lower && - get_Bound_upper(pred) == upper) { - /* - * One could expect that we simply return the previous - * Bound here. However, this would be wrong, as we could - * add an exception Proj to a new location then. - * So, we must turn in into a tuple. - */ - ret_tuple = 1; - } + ret_tuple = 1; } } if (ret_tuple) { @@ -1657,7 +1755,7 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Rot); + CASE(Rotl); CASE(Not); CASE(Minus); CASE(Mul); @@ -1692,7 +1790,7 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op static int is_const_Phi(ir_node *n) { int i; - if (! is_Phi(n)) + if (! is_Phi(n) || get_irn_arity(n) == 0) return 0; for (i = get_irn_arity(n) - 1; i >= 0; --i) if (! is_Const(get_irn_n(n, i))) @@ -1706,15 +1804,15 @@ static int is_const_Phi(ir_node *n) { * @param phi the Phi node * @param other the other operand * @param eval an evaluator function + * @param mode the mode of the result, may be different from the mode of the Phi! * @param left if non-zero, other is the left operand, else the right * * @return a new Phi node if the conversion was successful, NULL else */ -static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) { +static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), ir_mode *mode, int left) { tarval *tv; void **res; ir_node *pred; - ir_mode *mode; ir_graph *irg; int i, n = get_irn_arity(phi); @@ -1744,7 +1842,6 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)( res[i] = tv; } } - mode = get_irn_mode(phi); irg = current_ir_graph; for (i = 0; i < n; ++i) { pred = get_irn_n(phi, i); @@ -1754,6 +1851,50 @@ static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)( return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); } /* apply_binop_on_phi */ +/** + * Apply an evaluator on a binop with two constant Phi. + * + * @param a the left Phi node + * @param b the right Phi node + * @param eval an evaluator function + * @param mode the mode of the result, may be different from the mode of the Phi! + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(tarval *, tarval *), ir_mode *mode) { + tarval *tv_l, *tv_r, *tv; + void **res; + ir_node *pred; + ir_graph *irg; + int i, n; + + if (get_nodes_block(a) != get_nodes_block(b)) + return NULL; + + n = get_irn_arity(a); + NEW_ARR_A(void *, res, n); + + for (i = 0; i < n; ++i) { + pred = get_irn_n(a, i); + tv_l = get_Const_tarval(pred); + pred = get_irn_n(b, i); + tv_r = get_Const_tarval(pred); + tv = eval(tv_l, tv_r); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return NULL; + } + res[i] = tv; + } + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(a, i); + res[i] = new_r_Const_type(irg, get_irg_start_block(irg), mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(a), n, (ir_node **)res, mode); +} /* apply_binop_on_2_phis */ + /** * Apply an evaluator on a unop with a constant operator (a Phi). * @@ -1792,6 +1933,41 @@ static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) { return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); } /* apply_unop_on_phi */ +/** + * Apply a conversion on a constant operator (a Phi). + * + * @param phi the Phi node + * + * @return a new Phi node if the conversion was successful, NULL else + */ +static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) { + tarval *tv; + void **res; + ir_node *pred; + ir_graph *irg; + int i, n = get_irn_arity(phi); + + NEW_ARR_A(void *, res, n); + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + tv = get_Const_tarval(pred); + tv = tarval_convert_to(tv, mode); + + if (tv == tarval_bad) { + /* folding failed, bad */ + return 0; + } + res[i] = tv; + } + irg = current_ir_graph; + for (i = 0; i < n; ++i) { + pred = get_irn_n(phi, i); + res[i] = new_r_Const_type(irg, get_irg_start_block(irg), + mode, res[i], get_Const_type(pred)); + } + return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode); +} /* apply_conv_on_phi */ + /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)). @@ -1801,17 +1977,17 @@ static ir_node *transform_node_AddSub(ir_node *n) { ir_mode *mode = get_irn_mode(n); if (mode_is_reference(mode)) { - ir_node *left = get_binop_left(n); - ir_node *right = get_binop_right(n); - int ref_bits = get_mode_size_bits(mode); + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + unsigned ref_bits = get_mode_size_bits(mode); if (is_Conv(left)) { - ir_mode *mode = get_irn_mode(left); - int bits = get_mode_size_bits(mode); + ir_mode *lmode = get_irn_mode(left); + unsigned bits = get_mode_size_bits(lmode); if (ref_bits == bits && - mode_is_int(mode) && - get_mode_arithmetic(mode) == irma_twos_complement) { + mode_is_int(lmode) && + get_mode_arithmetic(lmode) == irma_twos_complement) { ir_node *pre = get_Conv_op(left); ir_mode *pre_mode = get_irn_mode(pre); @@ -1829,12 +2005,12 @@ static ir_node *transform_node_AddSub(ir_node *n) { } if (is_Conv(right)) { - ir_mode *mode = get_irn_mode(right); - int bits = get_mode_size_bits(mode); + ir_mode *rmode = get_irn_mode(right); + unsigned bits = get_mode_size_bits(rmode); if (ref_bits == bits && - mode_is_int(mode) && - get_mode_arithmetic(mode) == irma_twos_complement) { + mode_is_int(rmode) && + get_mode_arithmetic(rmode) == irma_twos_complement) { ir_node *pre = get_Conv_op(right); ir_mode *pre_mode = get_irn_mode(pre); @@ -1850,30 +2026,48 @@ static ir_node *transform_node_AddSub(ir_node *n) { } } } + + /* let address arithmetic use unsigned modes */ + if (is_Const(right)) { + ir_mode *rmode = get_irn_mode(right); + + if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) { + /* convert a AddP(P, *s) into AddP(P, *u) */ + ir_mode *nm = get_reference_mode_unsigned_eq(mode); + + ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm); + set_binop_right(n, pre); + } + } } + return n; } /* transform_node_AddSub */ -#define HANDLE_BINOP_PHI(op,a,b,c) \ - c = NULL; \ - if (is_Const(b) && is_const_Phi(a)) { \ - /* check for Op(Phi, Const) */ \ - c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \ - } \ - else if (is_Const(a) && is_const_Phi(b)) { \ - /* check for Op(Const, Phi) */ \ - c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \ - } \ - if (c) { \ - DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ - return c; \ +#define HANDLE_BINOP_PHI(eval, a, b, c, mode) \ + c = NULL; \ + if (is_Const(b) && is_const_Phi(a)) { \ + /* check for Op(Phi, Const) */ \ + c = apply_binop_on_phi(a, get_Const_tarval(b), eval, mode, 0);\ + } \ + else if (is_Const(a) && is_const_Phi(b)) { \ + /* check for Op(Const, Phi) */ \ + c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\ + } \ + else if (is_const_Phi(a) && is_const_Phi(b)) { \ + /* check for Op(Phi, Phi) */ \ + c = apply_binop_on_2_phis(a, b, eval, mode); \ + } \ + if (c) { \ + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ + return c; \ } -#define HANDLE_UNOP_PHI(op,a,c) \ +#define HANDLE_UNOP_PHI(eval, a, c) \ c = NULL; \ if (is_const_Phi(a)) { \ /* check for Op(Phi) */ \ - c = apply_unop_on_phi(a, op); \ + c = apply_unop_on_phi(a, eval); \ if (c) { \ DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ return c; \ @@ -1898,18 +2092,28 @@ static ir_node *transform_node_Add(ir_node *n) { a = get_Add_left(n); b = get_Add_right(n); - HANDLE_BINOP_PHI(tarval_add, a,b,c); - mode = get_irn_mode(n); + if (mode_is_reference(mode)) { + ir_mode *lmode = get_irn_mode(a); + + if (is_Const(b) && is_Const_null(b) && mode_is_int(lmode)) { + /* an Add(a, NULL) is a hidden Conv */ + dbg_info *dbg = get_irn_dbg_info(n); + return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode); + } + } + + HANDLE_BINOP_PHI(tarval_add, a, b, c, mode); + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) return n; if (mode_is_num(mode)) { /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */ - if (!get_opt_arch_dep_running() && a == b && mode_is_int(mode)) { - ir_node *block = get_irn_n(n, -1); + if (!is_arch_dep_running() && a == b && mode_is_int(mode)) { + ir_node *block = get_nodes_block(n); n = new_rd_Mul( get_irn_dbg_info(n), @@ -1943,74 +2147,6 @@ static ir_node *transform_node_Add(ir_node *n) { DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); return n; } - if (! get_opt_reassociation()) { - /* do NOT execute this code if reassociation is enabled, it does the inverse! */ - if (is_Mul(a)) { - ir_node *ma = get_Mul_left(a); - ir_node *mb = get_Mul_right(a); - - if (b == ma) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } else if (b == mb) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - } - if (is_Mul(b)) { - ir_node *ma = get_Mul_left(b); - ir_node *mb = get_Mul_right(b); - - if (a == ma) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - if (a == mb) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - } - } if (get_mode_arithmetic(mode) == irma_twos_complement) { /* Here we rely on constants be on the RIGHT side */ if (is_Not(a)) { @@ -2035,7 +2171,7 @@ static ir_node *transform_node_Add(ir_node *n) { ir_node *op = get_Not_op(b); if (op == a) { - /* ~x + x = -1 */ + /* x + ~x = -1 */ ir_node *blk = get_irn_n(n, -1); n = new_r_Const(current_ir_graph, blk, mode, get_mode_minus_one(mode)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); @@ -2047,7 +2183,9 @@ static ir_node *transform_node_Add(ir_node *n) { return n; } /* transform_node_Add */ -/* returns -cnst */ +/** + * returns -cnst or NULL if impossible + */ static ir_node *const_negate(ir_node *cnst) { tarval *tv = tarval_neg(get_Const_tarval(cnst)); dbg_info *dbgi = get_irn_dbg_info(cnst); @@ -2080,8 +2218,18 @@ static ir_node *transform_node_Sub(ir_node *n) { mode = get_irn_mode(n); + if (mode_is_int(mode)) { + ir_mode *lmode = get_irn_mode(a); + + if (is_Const(b) && is_Const_null(b) && mode_is_reference(lmode)) { + /* a Sub(a, NULL) is a hidden Conv */ + dbg_info *dbg = get_irn_dbg_info(n); + return new_rd_Conv(dbg, current_ir_graph, get_nodes_block(n), a, mode); + } + } + restart: - HANDLE_BINOP_PHI(tarval_sub, a,b,c); + HANDLE_BINOP_PHI(tarval_sub, a, b, c, mode); /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) @@ -2106,7 +2254,6 @@ restart: dbg_info *dbg = get_irn_dbg_info(n); ir_node *block = get_nodes_block(n); ir_node *left = get_Minus_op(a); - ir_mode *mode = get_irn_mode(n); ir_node *add = new_rd_Add(dbg, irg, block, left, b, mode); n = new_rd_Minus(dbg, irg, block, add, mode); @@ -2117,7 +2264,6 @@ restart: dbg_info *dbg = get_irn_dbg_info(n); ir_node *block = get_nodes_block(n); ir_node *right = get_Minus_op(b); - ir_mode *mode = get_irn_mode(n); n = new_rd_Add(dbg, irg, block, a, right, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MINUS); @@ -2132,9 +2278,8 @@ restart: ir_node *sub = new_rd_Sub(s_dbg, irg, s_block, s_left, s_right, s_mode); dbg_info *a_dbg = get_irn_dbg_info(n); ir_node *a_block = get_nodes_block(n); - ir_mode *a_mode = get_irn_mode(n); - n = new_rd_Add(a_dbg, irg, a_block, a, sub, a_mode); + n = new_rd_Add(a_dbg, irg, a_block, a, sub, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ @@ -2150,9 +2295,8 @@ restart: ir_node *mul = new_rd_Mul(m_dbg, irg, m_block, m_left, cnst2, m_mode); dbg_info *a_dbg = get_irn_dbg_info(n); ir_node *a_block = get_nodes_block(n); - ir_mode *a_mode = get_irn_mode(n); - n = new_rd_Add(a_dbg, irg, a_block, a, mul, a_mode); + n = new_rd_Add(a_dbg, irg, a_block, a, mul, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; } @@ -2229,16 +2373,16 @@ restart: if (mode == get_irn_mode(b)) { ir_mode *ma, *mb; - - a = get_Conv_op(a); - b = get_Conv_op(b); + ir_node *op_a = get_Conv_op(a); + ir_node *op_b = get_Conv_op(b); /* check if it's allowed to skip the conv */ - ma = get_irn_mode(a); - mb = get_irn_mode(b); + ma = get_irn_mode(op_a); + mb = get_irn_mode(op_b); if (mode_is_reference(ma) && mode_is_reference(mb)) { /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */ + a = op_a; b = op_b; set_Sub_left(n, a); set_Sub_right(n, b); @@ -2247,7 +2391,7 @@ restart: } } /* do NOT execute this code if reassociation is enabled, it does the inverse! */ - if (get_opt_reassociation() && is_Mul(a)) { + if (!is_reassoc_running() && is_Mul(a)) { ir_node *ma = get_Mul_left(a); ir_node *mb = get_Mul_right(a); @@ -2283,21 +2427,22 @@ restart: return n; } } - if (is_Sub(a)) { - ir_node *x = get_Sub_left(a); - ir_node *y = get_Sub_right(a); - ir_node *blk = get_irn_n(n, -1); - ir_mode *m_b = get_irn_mode(b); - ir_mode *m_y = get_irn_mode(y); + if (is_Sub(a)) { /* (x - y) - b -> x - (y + b) */ + ir_node *x = get_Sub_left(a); + ir_node *y = get_Sub_right(a); + ir_node *blk = get_irn_n(n, -1); + ir_mode *m_b = get_irn_mode(b); + ir_mode *m_y = get_irn_mode(y); + ir_mode *add_mode; ir_node *add; /* Determine the right mode for the Add. */ if (m_b == m_y) - mode = m_b; + add_mode = m_b; else if (mode_is_reference(m_b)) - mode = m_b; + add_mode = m_b; else if (mode_is_reference(m_y)) - mode = m_y; + add_mode = m_y; else { /* * Both modes are different but none is reference, @@ -2308,7 +2453,7 @@ restart: return n; } - add = new_r_Add(current_ir_graph, blk, y, b, mode); + add = new_r_Add(current_ir_graph, blk, y, b, add_mode); n = new_rd_Sub(get_irn_dbg_info(n), current_ir_graph, blk, x, add, mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_SUB_X_Y_Z); @@ -2346,12 +2491,14 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { ir_mode *smode = get_irn_mode(a); if (ta == get_mode_one(smode)) { + /* (L)1 * (L)b = (L)b */ ir_node *blk = get_irn_n(n, -1); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, b, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; } else if (ta == get_mode_minus_one(smode)) { + /* (L)-1 * (L)b = (L)b */ ir_node *blk = get_irn_n(n, -1); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, b, smode); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); @@ -2359,12 +2506,14 @@ static ir_node *transform_node_Mul2n(ir_node *n, ir_mode *mode) { return n; } if (tb == get_mode_one(smode)) { + /* (L)a * (L)1 = (L)a */ ir_node *blk = get_irn_n(a, -1); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1); return n; } else if (tb == get_mode_minus_one(smode)) { + /* (L)a * (L)-1 = (L)-a */ ir_node *blk = get_irn_n(n, -1); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, a, smode); n = new_rd_Conv(get_irn_dbg_info(n), current_ir_graph, blk, n, mode); @@ -2391,7 +2540,7 @@ static ir_node *transform_node_Mul(ir_node *n) { if (mode != get_irn_mode(a)) return transform_node_Mul2n(n, mode); - HANDLE_BINOP_PHI(tarval_mul, a,b,c); + HANDLE_BINOP_PHI(tarval_mul, a, b, c, mode); if (mode_is_signed(mode)) { ir_node *r = NULL; @@ -2450,7 +2599,8 @@ static ir_node *transform_node_Mul(ir_node *n) { if (is_Const(a)) { tarval *tv = get_Const_tarval(a); if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { - n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), b, b, mode); + /* 2.0 * b = b + b */ + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), b, b, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); return n; } @@ -2458,7 +2608,8 @@ static ir_node *transform_node_Mul(ir_node *n) { else if (is_Const(b)) { tarval *tv = get_Const_tarval(b); if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv)) { - n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, a, mode); + /* a * 2.0 = a + a */ + n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, get_nodes_block(n), a, a, mode); DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_A_A); return n; } @@ -2471,10 +2622,39 @@ static ir_node *transform_node_Mul(ir_node *n) { * Transform a Div Node. */ static ir_node *transform_node_Div(ir_node *n) { - tarval *tv = value_of(n); ir_mode *mode = get_Div_resmode(n); - ir_node *value = n; + ir_node *a = get_Div_left(n); + ir_node *b = get_Div_right(n); + ir_node *value; + tarval *tv; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + value = apply_binop_on_2_phis(a, b, tarval_div, mode); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + value = n; + tv = value_of(n); if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); @@ -2530,10 +2710,39 @@ make_tuple: * Transform a Mod node. */ static ir_node *transform_node_Mod(ir_node *n) { - tarval *tv = value_of(n); ir_mode *mode = get_Mod_resmode(n); - ir_node *value = n; + ir_node *a = get_Mod_left(n); + ir_node *b = get_Mod_right(n); + ir_node *value; + tarval *tv; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + value = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + value = apply_binop_on_2_phis(a, b, tarval_mod, mode); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + value = n; + tv = value_of(n); if (tv != tarval_bad) { value = new_Const(get_tarval_mode(tv), tv); @@ -2593,14 +2802,48 @@ static ir_node *transform_node_DivMod(ir_node *n) { ir_node *a = get_DivMod_left(n); ir_node *b = get_DivMod_right(n); ir_mode *mode = get_DivMod_resmode(n); - tarval *ta = value_of(a); - tarval *tb = value_of(b); + tarval *ta, *tb; int evaluated = 0; + ir_node *va, *vb; + + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + va = apply_binop_on_phi(a, get_Const_tarval(b), tarval_div, mode, 0); + vb = apply_binop_on_phi(a, get_Const_tarval(b), tarval_mod, mode, 0); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + va = apply_binop_on_phi(b, get_Const_tarval(a), tarval_div, mode, 1); + vb = apply_binop_on_phi(b, get_Const_tarval(a), tarval_mod, mode, 1); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + va = apply_binop_on_2_phis(a, b, tarval_div, mode); + vb = apply_binop_on_2_phis(a, b, tarval_mod, mode); + if (va && vb) { + DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); + DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); + goto make_tuple; + } + } + ta = value_of(a); + tb = value_of(b); if (tb != tarval_bad) { if (tb == get_mode_one(get_tarval_mode(tb))) { - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, b); + va = a; + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else if (ta != tarval_bad) { tarval *resa, *resb; @@ -2609,28 +2852,30 @@ static ir_node *transform_node_DivMod(ir_node *n) { Jmp for X result!? */ resb = tarval_mod(ta, tb); if (resb == tarval_bad) return n; /* Causes exception! */ - a = new_Const(mode, resa); - b = new_Const(mode, resb); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_Const(mode, resa); + vb = new_Const(mode, resb); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) { - a = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), a, mode); + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else { /* Try architecture dependent optimization */ - arch_dep_replace_divmod_by_const(&a, &b, n); - evaluated = a != NULL; + va = a; + vb = b; + arch_dep_replace_divmod_by_const(&va, &vb, n); + evaluated = va != NULL; } } else if (a == b) { if (value_not_zero(a, &dummy)) { /* a/a && a != 0 */ - a = new_Const(mode, get_mode_one(mode)); - b = new_Const(mode, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, a); - DBG_OPT_CSTEVAL(n, b); + va = new_Const(mode, get_mode_one(mode)); + vb = new_Const(mode, get_mode_null(mode)); + DBG_OPT_CSTEVAL(n, va); + DBG_OPT_CSTEVAL(n, vb); goto make_tuple; } else { /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ @@ -2638,7 +2883,7 @@ static ir_node *transform_node_DivMod(ir_node *n) { } } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) { /* 0 / non-Const = 0 */ - b = a; + vb = va = a; goto make_tuple; } @@ -2656,8 +2901,8 @@ make_tuple: set_Tuple_pred(n, pn_DivMod_M, mem); set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk)); set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */ - set_Tuple_pred(n, pn_DivMod_res_div, a); - set_Tuple_pred(n, pn_DivMod_res_mod, b); + set_Tuple_pred(n, pn_DivMod_res_div, va); + set_Tuple_pred(n, pn_DivMod_res_mod, vb); } return n; @@ -2675,8 +2920,17 @@ static ir_node *transform_node_Quot(ir_node *n) { if (is_Const(b)) { tarval *tv = get_Const_tarval(b); + int rem; - tv = tarval_quo(get_mode_one(mode), tv); + /* + * Floating point constant folding might be disabled here to + * prevent rounding. + * However, as we check for exact result, doing it is safe. + * Switch it on. + */ + rem = tarval_enable_fp_ops(1); + tv = tarval_quo(get_mode_one(mode), tv); + (void)tarval_enable_fp_ops(rem); /* Do the transformation if the result is either exact or we are not using strict rules. */ @@ -2706,33 +2960,49 @@ static ir_node *transform_node_Quot(ir_node *n) { /** * Optimize Abs(x) into x if x is Confirmed >= 0 * Optimize Abs(x) into -x if x is Confirmed <= 0 + * Optimize Abs(-x) int Abs(x) */ static ir_node *transform_node_Abs(ir_node *n) { - ir_node *oldn = n; - ir_node *a = get_Abs_op(n); - value_classify_sign sign = classify_value_sign(a); + ir_node *c, *oldn = n; + ir_node *a = get_Abs_op(n); + ir_mode *mode; - if (sign == value_classified_negative) { - ir_mode *mode = get_irn_mode(n); + HANDLE_UNOP_PHI(tarval_abs, a, c); + + switch (classify_value_sign(a)) { + case value_classified_negative: + mode = get_irn_mode(n); /* * We can replace the Abs by -x here. - * We even could add a new Confirm here. + * We even could add a new Confirm here + * (if not twos complement) * * Note that -x would create a new node, so we could * not run it in the equivalent_node() context. */ n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, - get_irn_n(n, -1), a, mode); + get_nodes_block(n), a, mode); DBG_OPT_CONFIRM(oldn, n); - } else if (sign == value_classified_positive) { + return n; + case value_classified_positive: /* n is positive, Abs is not needed */ n = a; DBG_OPT_CONFIRM(oldn, n); + return n; + default: + break; + } + if (is_Minus(a)) { + /* Abs(-x) = Abs(x) */ + mode = get_irn_mode(n); + n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, + get_nodes_block(n), get_Minus_op(a), mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X); + return n; } - return n; } /* transform_node_Abs */ @@ -2757,7 +3027,8 @@ static ir_node *transform_node_Cond(ir_node *n) { (get_opt_unreachable_code())) { /* It's a boolean Cond, branching on a boolean constant. Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ - jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n)); + ir_node *blk = get_nodes_block(n); + jmp = new_r_Jmp(current_ir_graph, blk); turn_into_tuple(n, pn_Cond_max); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_Bad()); @@ -2767,12 +3038,16 @@ static ir_node *transform_node_Cond(ir_node *n) { set_Tuple_pred(n, pn_Cond_true, new_Bad()); } /* We might generate an endless loop, so keep it alive. */ - add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n)); + add_End_keepalive(get_irg_end(current_ir_graph), blk); } return n; } /* transform_node_Cond */ -typedef ir_node* (*recursive_transform) (ir_node *n); +/** + * Prototype of a recursive transform function + * for bitwise distributive transformations. + */ +typedef ir_node* (*recursive_transform)(ir_node *n); /** * makes use of distributive laws for and, or, eor @@ -2823,7 +3098,8 @@ static ir_node *transform_bitwise_distributive(ir_node *n, ir_node *b_left = get_binop_left(b); ir_node *b_right = get_binop_right(b); ir_node *c = NULL; - ir_node *op1, *op2; + ir_node *op1 = NULL; + ir_node *op2 = NULL; if (is_op_commutative(op)) { if (a_left == b_left) { @@ -2864,9 +3140,10 @@ static ir_node *transform_bitwise_distributive(ir_node *n, n = new_rd_And(dbgi, irg, blk, new_n, c, mode); } else { n = exact_copy(a); - set_irn_n(n, -1, blk); + set_nodes_block(n, blk); set_binop_left(n, new_n); set_binop_right(n, c); + add_identities(current_ir_graph->value_table, n); } DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); @@ -2886,9 +3163,8 @@ static ir_node *transform_node_And(ir_node *n) { ir_node *b = get_And_right(n); ir_mode *mode; - HANDLE_BINOP_PHI(tarval_and, a,b,c); - mode = get_irn_mode(n); + HANDLE_BINOP_PHI(tarval_and, a, b, c, mode); /* we can evaluate 2 Projs of the same Cmp */ if (mode == mode_b && is_Proj(a) && is_Proj(b)) { @@ -3020,10 +3296,10 @@ static ir_node *transform_node_Eor(ir_node *n) { ir_node *b = get_Eor_right(n); ir_mode *mode = get_irn_mode(n); - HANDLE_BINOP_PHI(tarval_eor, a,b,c); + HANDLE_BINOP_PHI(tarval_eor, a, b, c, mode); /* we can evaluate 2 Projs of the same Cmp */ - if(get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) { + if (mode == mode_b && is_Proj(a) && is_Proj(b)) { ir_node *pred_a = get_Proj_pred(a); ir_node *pred_b = get_Proj_pred(b); if(pred_a == pred_b) { @@ -3053,12 +3329,20 @@ static ir_node *transform_node_Eor(ir_node *n) { 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 && is_Const(b) && is_Const_one(b)) { - /* The Eor is a Not. Replace it by a Not. */ - /* ????!!!Extend to bitfield 1111111. */ - 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); + } else if (is_Const(b)) { + if (is_Not(a)) { /* ~x ^ const -> x ^ ~const */ + ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(b))); + ir_node *not_op = get_Not_op(a); + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + n = new_rd_Eor(dbg, irg, block, not_op, cnst, mode); + return n; + } else if (is_Const_all_one(b)) { /* x ^ 1...1 -> ~1 */ + n = new_r_Not(current_ir_graph, get_nodes_block(n), a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + } } else { n = transform_bitwise_distributive(n, transform_node_Eor); } @@ -3071,28 +3355,51 @@ static ir_node *transform_node_Eor(ir_node *n) { */ static ir_node *transform_node_Not(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Not_op(n); + ir_node *a = get_Not_op(n); + ir_mode *mode = get_irn_mode(n); HANDLE_UNOP_PHI(tarval_not,a,c); /* check for a boolean Not */ - if (get_irn_mode(n) == mode_b && + if (mode == mode_b && is_Proj(a) && is_Cmp(get_Proj_pred(a))) { /* We negate a Cmp. The Cmp has the negated result anyways! */ n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a), mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); - return n; + return n; } - if (is_Sub(a)) { - ir_node *sub_r = get_Sub_right(a); - if (is_Const(sub_r) && is_Const_one(sub_r)) { - /* ~(x-1) = -x */ - ir_node *op = get_Sub_left(a); - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); + if (is_Eor(a)) { + ir_node *eor_b = get_Eor_right(a); + if (is_Const(eor_b)) { /* ~(x ^ const) -> x ^ ~const */ + ir_node *cnst = new_Const(mode, tarval_not(get_Const_tarval(eor_b))); + ir_node *eor_a = get_Eor_left(a); + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + n = new_rd_Eor(dbg, irg, block, eor_a, cnst, mode); + return n; + } + } + if (get_mode_arithmetic(mode) == irma_twos_complement) { + if (is_Minus(a)) { /* ~-x -> x + -1 */ + dbg_info *dbg = get_irn_dbg_info(n); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(n); + ir_node *add_l = get_Minus_op(a); + ir_node *add_r = new_rd_Const(dbg, irg, block, mode, get_mode_minus_one(mode)); + n = new_rd_Add(dbg, irg, block, add_l, add_r, mode); + } else if (is_Add(a)) { + ir_node *add_r = get_Add_right(a); + if (is_Const(add_r) && is_Const_all_one(add_r)) { + /* ~(x + -1) = -x */ + ir_node *op = get_Add_left(a); + ir_node *blk = get_irn_n(n, -1); + n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n)); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); + } } } return n; @@ -3103,6 +3410,9 @@ static ir_node *transform_node_Not(ir_node *n) { * Optimize: * -(~x) = x + 1 * -(a-b) = b - a + * -(a >>u (size-1)) = a >>s (size-1) + * -(a >>s (size-1)) = a >>u (size-1) + * -(a * const) -> a * -const */ static ir_node *transform_node_Minus(ir_node *n) { ir_node *c, *oldn = n; @@ -3130,7 +3440,7 @@ static ir_node *transform_node_Minus(ir_node *n) { if (is_Const(c)) { tarval *tv = get_Const_tarval(c); - if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>u (size-1)) = a >>s (size-1) */ ir_node *v = get_Shr_left(a); @@ -3146,7 +3456,7 @@ static ir_node *transform_node_Minus(ir_node *n) { if (is_Const(c)) { tarval *tv = get_Const_tarval(c); - if (tarval_is_long(tv) && get_tarval_long(tv) == get_mode_size_bits(mode) - 1) { + if (tarval_is_long(tv) && get_tarval_long(tv) == (int) get_mode_size_bits(mode) - 1) { /* -(a >>s (size-1)) = a >>u (size-1) */ ir_node *v = get_Shrs_left(a); @@ -3168,6 +3478,23 @@ static ir_node *transform_node_Minus(ir_node *n) { return n; } + if (is_Mul(a)) { /* -(a * const) -> a * -const */ + ir_node *mul_l = get_Mul_left(a); + ir_node *mul_r = get_Mul_right(a); + if (is_Const(mul_r)) { + tarval *tv = tarval_neg(get_Const_tarval(mul_r)); + if(tv != tarval_bad) { + ir_node *cnst = new_Const(mode, tv); + dbg_info *dbg = get_irn_dbg_info(a); + ir_graph *irg = current_ir_graph; + ir_node *block = get_nodes_block(a); + n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode); + DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C); + return n; + } + } + } + return n; } /* transform_node_Minus */ @@ -3184,8 +3511,8 @@ static ir_node *transform_node_Cast(ir_node *n) { get_Const_tarval(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } else if (is_SymConst(pred) && get_SymConst_value_type(pred) != tp) { - n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred), - get_SymConst_kind(pred), tp); + n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred), + get_SymConst_symbol(pred), get_SymConst_kind(pred), tp); DBG_OPT_CSTEVAL(oldn, n); } @@ -3204,6 +3531,15 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) { if (value_not_zero(b, &confirm)) { /* div(x, y) && y != 0 */ + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_Div_mem(div); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_Div_mem(div, new_mem); + set_irn_pinned(div, op_pin_state_floats); + } + proj_nr = get_Proj_proj(proj); switch (proj_nr) { case pn_Div_X_regular: @@ -3245,6 +3581,15 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { /* mod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_Mod_mem(mod); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_Mod_mem(mod, new_mem); + set_irn_pinned(mod, op_pin_state_floats); + } + switch (proj_nr) { case pn_Mod_X_regular: @@ -3263,9 +3608,8 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } - set_irn_pinned(mod, op_pin_state_floats); /* this is a Mod without exception, we can remove the memory edge */ - set_Mod_mem(mod, get_irg_no_mem(current_ir_graph)); + set_Mod_mem(mod, new_mem); return res; case pn_Mod_res: if (get_Mod_left(mod) == b) { @@ -3295,6 +3639,15 @@ static ir_node *transform_node_Proj_DivMod(ir_node *proj) { /* DivMod(x, y) && y != 0 */ proj_nr = get_Proj_proj(proj); + if (confirm == NULL) { + /* we are sure we have a Const != 0 */ + new_mem = get_DivMod_mem(divmod); + if (is_Pin(new_mem)) + new_mem = get_Pin_op(new_mem); + set_DivMod_mem(divmod, new_mem); + set_irn_pinned(divmod, op_pin_state_floats); + } + switch (proj_nr) { case pn_DivMod_X_regular: @@ -3313,9 +3666,8 @@ static ir_node *transform_node_Proj_DivMod(ir_node *proj) { /* This node can only float up to the Confirm block */ new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem); } - set_irn_pinned(divmod, op_pin_state_floats); /* this is a DivMod without exception, we can remove the memory edge */ - set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph)); + set_DivMod_mem(divmod, new_mem); return res; case pn_DivMod_res_mod: @@ -3363,33 +3715,73 @@ static ir_node *transform_node_Proj_Cond(ir_node *proj) { return proj; } /* transform_node_Proj_Cond */ +/** + * Create a 0 constant of given mode. + */ +static ir_node *create_zero_const(ir_mode *mode) { + tarval *tv = get_mode_null(mode); + ir_node *cnst = new_Const(mode, tv); + + return cnst; +} + +/* the order of the values is important! */ +typedef enum const_class { + const_const = 0, + const_like = 1, + const_other = 2 +} const_class; + +static const_class classify_const(const ir_node* n) +{ + if (is_Const(n)) return const_const; + if (is_irn_constlike(n)) return const_like; + return const_other; +} + +/** + * Determines whether r is more constlike or has a larger index (in that order) + * than l. + */ +static int operands_are_normalized(const ir_node *l, const ir_node *r) +{ + const const_class l_order = classify_const(l); + const const_class r_order = classify_const(r); + return + l_order > r_order || + (l_order == r_order && get_irn_idx(l) <= get_irn_idx(r)); +} + /** * Normalizes and optimizes Cmp nodes. */ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { - ir_node *n = get_Proj_pred(proj); - ir_node *left = get_Cmp_left(n); - ir_node *right = get_Cmp_right(n); - ir_node *c = NULL; - tarval *tv = NULL; - int changed = 0; - ir_mode *mode = NULL; - long proj_nr = get_Proj_proj(proj); - - /* we can evaluate this direct */ - switch(proj_nr) { + ir_node *n = get_Proj_pred(proj); + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + tarval *tv = NULL; + int changed = 0; + ir_mode *mode = NULL; + long proj_nr = get_Proj_proj(proj); + + /* we can evaluate some cases directly */ + switch (proj_nr) { case pn_Cmp_False: return new_Const(mode_b, get_tarval_b_false()); case pn_Cmp_True: return new_Const(mode_b, get_tarval_b_true()); case pn_Cmp_Leg: - if(!mode_is_float(get_irn_mode(left))) + if (!mode_is_float(get_irn_mode(left))) return new_Const(mode_b, get_tarval_b_true()); break; default: break; } + /* remove Casts of both sides */ + left = skip_Cast(left); + right = skip_Cast(right); + /* Remove unnecessary conversions */ /* TODO handle constants */ if (is_Conv(left) && is_Conv(right)) { @@ -3399,7 +3791,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { ir_mode *mode_left = get_irn_mode(op_left); ir_mode *mode_right = get_irn_mode(op_right); - if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode)) { + if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode) + && mode_left != mode_b && mode_right != mode_b) { ir_graph *irg = current_ir_graph; ir_node *block = get_nodes_block(n); @@ -3407,149 +3800,172 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = op_left; right = op_right; changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV); } else if (smaller_mode(mode_left, mode_right)) { left = new_r_Conv(irg, block, op_left, mode_right); right = op_right; changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } else if (smaller_mode(mode_right, mode_left)) { left = op_left; right = new_r_Conv(irg, block, op_right, mode_left); changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } } } - /* TODO extend to arbitrary constants */ - if (is_Conv(left) && is_Const(right) && is_Const_null(right)) { - ir_mode* mode = get_irn_mode(left); - ir_node* op = get_Conv_op(left); - ir_mode* op_mode = get_irn_mode(op); - - if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) && - (mode_is_signed(mode) || !mode_is_signed(op_mode))) { - ir_node *null = new_Const(op_mode, get_mode_null(op_mode)); - set_Cmp_left( n, op); - set_Cmp_right(n, null); - return proj; - } - } - - /* remove Casts */ - if (is_Cast(left)) - left = get_Cast_op(left); - if (is_Cast(right)) - right = get_Cast_op(right); - - /* remove operation of both sides if possible */ + /* remove operation on both sides if possible */ if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { - ir_opcode lop = get_irn_opcode(left); - - if (lop == get_irn_opcode(right)) { - ir_node *ll, *lr, *rl, *rr; - - /* same operation on both sides, try to remove */ - switch (lop) { - case iro_Not: - case iro_Minus: - /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */ - left = get_unop_op(left); - right = get_unop_op(right); - changed |= 1; - break; - case iro_Add: - ll = get_Add_left(left); - lr = get_Add_right(left); - rl = get_Add_left(right); - rr = get_Add_right(right); - - if (ll == rl) { - /* X + a CMP X + b ==> a CMP b */ - left = lr; - right = rr; - changed |= 1; - } else if (ll == rr) { - /* X + a CMP b + X ==> a CMP b */ - left = lr; - right = rl; - changed |= 1; - } else if (lr == rl) { - /* a + X CMP X + b ==> a CMP b */ - left = ll; - right = rr; - changed |= 1; - } else if (lr == rr) { - /* a + X CMP b + X ==> a CMP b */ - left = ll; - right = rl; + /* + * The following operations are NOT safe for floating point operations, for instance + * 1.0 + inf == 2.0 + inf, =/=> x == y + */ + if (mode_is_int(get_irn_mode(left))) { + unsigned lop = get_irn_opcode(left); + + if (lop == get_irn_opcode(right)) { + ir_node *ll, *lr, *rl, *rr; + + /* same operation on both sides, try to remove */ + switch (lop) { + case iro_Not: + case iro_Minus: + /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */ + left = get_unop_op(left); + right = get_unop_op(right); changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + break; + case iro_Add: + ll = get_Add_left(left); + lr = get_Add_right(left); + rl = get_Add_left(right); + rr = get_Add_right(right); + + if (ll == rl) { + /* X + a CMP X + b ==> a CMP b */ + left = lr; + right = rr; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } else if (ll == rr) { + /* X + a CMP b + X ==> a CMP b */ + left = lr; + right = rl; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } else if (lr == rl) { + /* a + X CMP X + b ==> a CMP b */ + left = ll; + right = rr; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } else if (lr == rr) { + /* a + X CMP b + X ==> a CMP b */ + left = ll; + right = rl; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } + break; + case iro_Sub: + ll = get_Sub_left(left); + lr = get_Sub_right(left); + rl = get_Sub_left(right); + rr = get_Sub_right(right); + + if (ll == rl) { + /* X - a CMP X - b ==> a CMP b */ + left = lr; + right = rr; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } else if (lr == rr) { + /* a - X CMP b - X ==> a CMP b */ + left = ll; + right = rl; + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } + break; + case iro_Rotl: + if (get_Rotl_right(left) == get_Rotl_right(right)) { + /* a ROTL X CMP b ROTL X ==> a CMP b */ + left = get_Rotl_left(left); + right = get_Rotl_left(right); + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } + break; + default: + break; } - break; - case iro_Sub: - ll = get_Sub_left(left); - lr = get_Sub_right(left); - rl = get_Sub_left(right); - rr = get_Sub_right(right); - - if (ll == rl) { - /* X - a CMP X - b ==> a CMP b */ - left = lr; - right = rr; - changed |= 1; - } else if (lr == rr) { - /* a - X CMP b - X ==> a CMP b */ - left = ll; - right = rl; + } + + /* X+A == A, A+X == A, A-X == A -> X == 0 */ + if (is_Add(left) || is_Sub(left)) { + ir_node *ll = get_binop_left(left); + ir_node *lr = get_binop_right(left); + + if (lr == right && is_Add(left)) { + ir_node *tmp = ll; + ll = lr; + lr = tmp; + } + if (ll == right) { + left = lr; + right = create_zero_const(get_irn_mode(left)); changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } - break; - case iro_Rot: - if (get_Rot_right(left) == get_Rot_right(right)) { - /* a ROT X CMP b ROT X */ - left = get_Rot_left(left); - right = get_Rot_left(right); + } + if (is_Add(right) || is_Sub(right)) { + ir_node *rl = get_binop_left(right); + ir_node *rr = get_binop_right(right); + + if (rr == left && is_Add(right)) { + ir_node *tmp = rl; + rl = rr; + rr = tmp; + } + if (rl == left) { + left = rr; + right = create_zero_const(get_irn_mode(left)); changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } - break; - default: - break; } - } - } + } /* mode_is_int(...) */ + } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */ + /* replace mode_b compares with ands/ors */ if (get_irn_mode(left) == mode_b) { ir_graph *irg = current_ir_graph; ir_node *block = get_nodes_block(n); + ir_node *bres; switch (proj_nr) { - case pn_Cmp_Le: return new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); - case pn_Cmp_Lt: return new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); - case pn_Cmp_Ge: return new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); - case pn_Cmp_Gt: return new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); - case pn_Cmp_Lg: return new_r_Eor(irg, block, left, right, mode_b); - case pn_Cmp_Eq: return new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); + case pn_Cmp_Le: bres = new_r_Or( irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break; + case pn_Cmp_Lt: bres = new_r_And(irg, block, new_r_Not(irg, block, left, mode_b), right, mode_b); break; + case pn_Cmp_Ge: bres = new_r_Or( irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break; + case pn_Cmp_Gt: bres = new_r_And(irg, block, left, new_r_Not(irg, block, right, mode_b), mode_b); break; + case pn_Cmp_Lg: bres = new_r_Eor(irg, block, left, right, mode_b); break; + case pn_Cmp_Eq: bres = new_r_Not(irg, block, new_r_Eor(irg, block, left, right, mode_b), mode_b); break; + default: bres = NULL; + } + if (bres) { + DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL); + return bres; } } - if (!get_opt_reassociation()) - return proj; - /* * First step: normalize the compare op * by placing the constant on the right side * or moving the lower address node to the left. - * We ignore the case that both are constants - * this case should be optimized away. */ - if (is_Const(right)) { - c = right; - } else if (is_Const(left)) { - c = left; - left = right; - right = c; - - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; - } else if (get_irn_idx(left) > get_irn_idx(right)) { + if (!operands_are_normalized(left, right)) { ir_node *t = left; left = right; @@ -3565,9 +3981,28 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { * later and may help to normalize more compares. * Of course this is only possible for integer values. */ - if (c) { - mode = get_irn_mode(c); - tv = get_Const_tarval(c); + if (is_Const(right)) { + mode = get_irn_mode(right); + tv = get_Const_tarval(right); + + /* TODO extend to arbitrary constants */ + if (is_Conv(left) && tarval_is_null(tv)) { + ir_node *op = get_Conv_op(left); + ir_mode *op_mode = get_irn_mode(op); + + /* + * UpConv(x) REL 0 ==> x REL 0 + */ + if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) && + ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) || + mode_is_signed(mode) || !mode_is_signed(op_mode))) { + tv = get_mode_null(op_mode); + left = op; + mode = op_mode; + changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + } + } if (tv != tarval_bad) { /* the following optimization is possible on modes without Overflow @@ -3587,13 +4022,16 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = get_Minus_op(left); proj_nr = get_inversed_pnc(proj_nr); changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) { + /* Not(a) ==/!= c ==> a ==/!= Not(c) */ tv = tarval_not(tv); if (tv != tarval_bad) { left = get_Not_op(left); changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } @@ -3614,6 +4052,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (tv != tarval_bad) { proj_nr ^= pn_Cmp_Eq; changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ @@ -3624,20 +4063,24 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { if (tv != tarval_bad) { proj_nr ^= pn_Cmp_Eq; changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* the following reassociations work only for == and != */ if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { +#if 0 /* Might be not that good in general */ /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ if (tarval_is_null(tv) && is_Sub(left)) { - right =get_Sub_right(left); + right = get_Sub_right(left); left = get_Sub_left(left); tv = value_of(right); changed = 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } +#endif if (tv != tarval_bad) { /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ @@ -3652,6 +4095,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = get_Sub_left(left); tv = tv2; changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } @@ -3677,6 +4121,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = a; tv = tv2; changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } @@ -3688,6 +4133,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { left = get_Minus_op(left); tv = tv2; changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } @@ -3702,43 +4148,199 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } } /* mode_is_int */ - /* - * optimization for AND: - * Optimize: - * And(x, C) == C ==> And(x, C) != 0 - * And(x, C) != C ==> And(X, C) == 0 - * - * if C is a single Bit constant. - */ - if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_And(left)) { - if (tarval_is_single_bit(tv)) { - /* check for Constant's match. We have check hare the tarvals, - because our const might be changed */ - ir_node *la = get_And_left(left); - ir_node *ra = get_And_right(left); - if ((is_Const(la) && get_Const_tarval(la) == tv) || - (is_Const(ra) && get_Const_tarval(ra) == tv)) { - /* fine: do the transformation */ - tv = get_mode_null(get_tarval_mode(tv)); - proj_nr ^= pn_Cmp_Leg; - changed |= 2; + if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + switch (get_irn_opcode(left)) { + ir_node *c1; + + case iro_And: + c1 = get_And_right(left); + if (is_Const(c1)) { + /* + * And(x, C1) == C2 ==> FALSE if C2 & C1 != C2 + * And(x, C1) != C2 ==> TRUE if C2 & C1 != C2 + */ + tarval *mask = tarval_and(get_Const_tarval(c1), tv); + if (mask != tv) { + /* TODO: move to constant evaluation */ + tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_Const(mode_b, tv); + DBG_OPT_CSTEVAL(proj, c1); + return c1; + } + + if (tarval_is_single_bit(tv)) { + /* + * optimization for AND: + * Optimize: + * And(x, C) == C ==> And(x, C) != 0 + * And(x, C) != C ==> And(X, C) == 0 + * + * if C is a single Bit constant. + */ + + /* check for Constant's match. We have check hare the tarvals, + because our const might be changed */ + if (get_Const_tarval(c1) == tv) { + /* fine: do the transformation */ + tv = get_mode_null(get_tarval_mode(tv)); + proj_nr ^= pn_Cmp_Leg; + changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); + } + } } - } + break; + case iro_Or: + c1 = get_Or_right(left); + if (is_Const(c1) && tarval_is_null(tv)) { + /* + * Or(x, C) == 0 && C != 0 ==> FALSE + * Or(x, C) != 0 && C != 0 ==> TRUE + */ + if (! tarval_is_null(get_Const_tarval(c1))) { + /* TODO: move to constant evaluation */ + tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_Const(mode_b, tv); + DBG_OPT_CSTEVAL(proj, c1); + return c1; + } + } + break; + case iro_Shl: + /* + * optimize x << c1 == c into x & (-1 >>u c1) == c >> c1 if c & (-1 << c1) == c + * FALSE else + * optimize x << c1 != c into x & (-1 >>u c1) != c >> c1 if c & (-1 << c1) == c + * TRUE else + */ + c1 = get_Shl_right(left); + if (is_Const(c1)) { + tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + tarval *minus1 = get_mode_all_one(mode); + tarval *amask = tarval_shr(minus1, tv1); + tarval *cmask = tarval_shl(minus1, tv1); + ir_node *sl, *blk; + + if (tarval_and(tv, cmask) != tv) { + /* condition not met */ + tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_Const(mode_b, tv); + DBG_OPT_CSTEVAL(proj, c1); + return c1; + } + sl = get_Shl_left(left); + blk = get_nodes_block(n); + left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode); + tv = tarval_shr(tv, tv1); + changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); + } + break; + case iro_Shr: + /* + * optimize x >>u c1 == c into x & (-1 << c1) == c << c1 if c & (-1 >>u c1) == c + * FALSE else + * optimize x >>u c1 != c into x & (-1 << c1) != c << c1 if c & (-1 >>u c1) == c + * TRUE else + */ + c1 = get_Shr_right(left); + if (is_Const(c1)) { + tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + tarval *minus1 = get_mode_all_one(mode); + tarval *amask = tarval_shl(minus1, tv1); + tarval *cmask = tarval_shr(minus1, tv1); + ir_node *sl, *blk; + + if (tarval_and(tv, cmask) != tv) { + /* condition not met */ + tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_Const(mode_b, tv); + DBG_OPT_CSTEVAL(proj, c1); + return c1; + } + sl = get_Shr_left(left); + blk = get_nodes_block(n); + left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode); + tv = tarval_shl(tv, tv1); + changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); + } + break; + case iro_Shrs: + /* + * optimize x >>s c1 == c into x & (-1 << c1) == c << c1 if (c >>s (BITS - c1)) \in {0,-1} + * FALSE else + * optimize x >>s c1 != c into x & (-1 << c1) != c << c1 if (c >>s (BITS - c1)) \in {0,-1} + * TRUE else + */ + c1 = get_Shrs_right(left); + if (is_Const(c1)) { + tarval *tv1 = get_Const_tarval(c1); + ir_mode *mode = get_irn_mode(left); + tarval *minus1 = get_mode_all_one(mode); + tarval *amask = tarval_shl(minus1, tv1); + tarval *cond = new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(tv1)); + ir_node *sl, *blk; + + cond = tarval_sub(cond, tv1); + cond = tarval_shrs(tv, cond); + + if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) { + /* condition not met */ + tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + c1 = new_Const(mode_b, tv); + DBG_OPT_CSTEVAL(proj, c1); + return c1; + } + sl = get_Shrs_left(left); + blk = get_nodes_block(n); + left = new_rd_And(get_irn_dbg_info(left), current_ir_graph, blk, sl, new_Const(mode, amask), mode); + tv = tarval_shl(tv, tv1); + changed |= 2; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); + } + break; + } /* switch */ } } /* tarval != bad */ } + if (changed & 2) /* need a new Const */ + right = new_Const(mode, tv); + + if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { + ir_node *op = get_Proj_pred(left); + + if ((is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) || + (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) { + ir_node *c = get_binop_right(op); + + if (is_Const(c)) { + tarval *tv = get_Const_tarval(c); + + if (tarval_is_single_bit(tv)) { + /* special case: (x % 2^n) CMP 0 ==> x & (2^n-1) CMP 0 */ + ir_node *v = get_binop_left(op); + ir_node *blk = get_irn_n(op, -1); + ir_mode *mode = get_irn_mode(v); + + tv = tarval_sub(tv, get_mode_one(mode)); + left = new_rd_And(get_irn_dbg_info(op), current_ir_graph, blk, v, new_Const(mode, tv), mode); + changed |= 1; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND); + } + } + } + } + if (changed) { ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */ - if (changed & 2) /* need a new Const */ - right = new_Const(mode, tv); - /* create a new compare */ n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block, left, right); - - set_Proj_pred(proj, n); - set_Proj_proj(proj, proj_nr); + proj = new_rd_Proj(get_irn_dbg_info(proj), current_ir_graph, block, n, get_irn_mode(proj), proj_nr); } return proj; @@ -3939,9 +4541,9 @@ static ir_node *transform_node_Or_bf_store(ir_node *or) { } /* transform_node_Or_bf_store */ /** - * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl */ -static ir_node *transform_node_Or_Rot(ir_node *or) { +static ir_node *transform_node_Or_Rotl(ir_node *or) { ir_mode *mode = get_irn_mode(or); ir_node *shl, *shr, *block; ir_node *irn, *x, *c1, *c2, *v, *sub, *n; @@ -3981,15 +4583,15 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { return or; if (get_tarval_long(tv1) + get_tarval_long(tv2) - != get_mode_size_bits(mode)) + != (int) 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); + n = new_r_Rotl(current_ir_graph, block, x, c1, mode); - DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT); + DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); return n; } else if (is_Sub(c1)) { v = c2; @@ -4006,16 +4608,16 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { if (! tarval_is_long(tv1)) return or; - if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + if (get_tarval_long(tv1) != (int) get_mode_size_bits(mode)) return or; /* yet, condition met */ block = get_nodes_block(or); /* a Rot right is not supported, so use a rot left */ - n = new_r_Rot(current_ir_graph, block, x, sub, mode); + n = new_r_Rotl(current_ir_graph, block, x, sub, mode); - DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL); return n; } else if (is_Sub(c2)) { v = c1; @@ -4029,21 +4631,21 @@ static ir_node *transform_node_Or_Rot(ir_node *or) { if (! tarval_is_long(tv1)) return or; - if (get_tarval_long(tv1) != get_mode_size_bits(mode)) + if (get_tarval_long(tv1) != (int) 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); + n = new_r_Rotl(current_ir_graph, block, x, v, mode); - DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT); + DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROTL); return n; } return or; -} /* transform_node_Or_Rot */ +} /* transform_node_Or_Rotl */ /** * Transform an Or. @@ -4052,12 +4654,13 @@ static ir_node *transform_node_Or(ir_node *n) { ir_node *c, *oldn = n; ir_node *a = get_Or_left(n); ir_node *b = get_Or_right(n); + ir_mode *mode; if (is_Not(a) && is_Not(b)) { /* ~a | ~b = ~(a&b) */ ir_node *block = get_nodes_block(n); - ir_mode *mode = get_irn_mode(n); + mode = get_irn_mode(n); a = get_Not_op(a); b = get_Not_op(b); n = new_rd_And(get_irn_dbg_info(n), current_ir_graph, block, a, b, mode); @@ -4083,10 +4686,11 @@ static ir_node *transform_node_Or(ir_node *n) { } } - HANDLE_BINOP_PHI(tarval_or, a,b,c); + mode = get_irn_mode(n); + HANDLE_BINOP_PHI(tarval_or, a, b, c, mode); n = transform_node_Or_bf_store(n); - n = transform_node_Or_Rot(n); + n = transform_node_Or_Rotl(n); if (n != oldn) return n; @@ -4100,15 +4704,15 @@ static ir_node *transform_node_Or(ir_node *n) { static ir_node *transform_node(ir_node *n); /** - * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl. + * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl. * * 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; + tarval *tv1, *tv2, *res; + ir_node *in[2], *irn, *block; left = get_binop_left(n); @@ -4125,48 +4729,249 @@ static ir_node *transform_node_shift(ir_node *n) { if (tv2 == tarval_bad) return n; - res = tarval_add(tv1, tv2); + res = tarval_add(tv1, tv2); + mode = get_irn_mode(n); /* beware: a simple replacement works only, if res < modulo shift */ + if (!is_Rotl(n)) { + int modulo_shf = get_mode_modulo_shift(mode); + assert(modulo_shf >= (int) get_mode_size_bits(mode)); + if (modulo_shf > 0) { + tarval *modulo = new_tarval_from_long(modulo_shf, + get_tarval_mode(res)); + + /* shifting too much */ + if (!(tarval_cmp(res, modulo) & pn_Cmp_Lt)) { + if (is_Shrs(n)) { + ir_graph *irg = get_irn_irg(n); + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *cnst = new_Const(mode_Iu, new_tarval_from_long(get_mode_size_bits(mode)-1, mode_Iu)); + return new_rd_Shrs(dbgi, irg, block, get_binop_left(left), + cnst, mode); + } + + return new_Const(mode, get_mode_null(mode)); + } + } + } else { + res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res))); + } + + /* ok, we can replace it */ + block = get_nodes_block(n); + + in[0] = get_binop_left(left); + in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + + irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + + DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + + return transform_node(irn); +} /* transform_node_shift */ + +/** + * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + */ +static ir_node *transform_node_bitop_shift(ir_node *n) { + ir_node *left; + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_graph *irg; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + tarval *tv1; + tarval *tv2; + tarval *tv_shift; + + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); + + if (!is_Const(right)) + return n; + + left = get_binop_left(n); + op_left = get_irn_op(left); + if (op_left != op_And && op_left != op_Or && op_left != op_Eor) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { + /* TODO: test if sign bit is affectes */ + return n; + } + + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; + + bitop_left = get_binop_left(left); + + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); + + assert(get_tarval_mode(tv1) == mode); + + if (is_Shl(n)) { + new_shift = new_rd_Shl(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shl(tv1, tv2); + } else if(is_Shr(n)) { + new_shift = new_rd_Shr(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shr(tv1, tv2); + } else if(is_Shrs(n)) { + new_shift = new_rd_Shrs(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_shrs(tv1, tv2); + } else { + assert(is_Rotl(n)); + new_shift = new_rd_Rotl(dbgi, irg, block, bitop_left, right, mode); + tv_shift = tarval_rotl(tv1, tv2); + } + + assert(get_tarval_mode(tv_shift) == mode); + new_const = new_Const(mode, tv_shift); + + if (op_left == op_And) { + new_bitop = new_rd_And(dbgi, irg, block, new_shift, new_const, mode); + } else if(op_left == op_Or) { + new_bitop = new_rd_Or(dbgi, irg, block, new_shift, new_const, mode); + } else { + assert(op_left == op_Eor); + new_bitop = new_rd_Eor(dbgi, irg, block, new_shift, new_const, mode); + } + + return new_bitop; +} + +/** + * normalisation: + * (x << c1) >> c2 <=> x>>(c2-c1) & (-1>>c2) + * also: + * if c2 > c1: x << (c2-c1) + * (x >>s c2) << c1 <=> x>>s(c2-c1) & (-1>>c2) + */ +static ir_node *transform_node_shl_shr(ir_node *n) { + ir_node *left; + ir_node *right = get_binop_right(n); + ir_node *x; + ir_graph *irg; + ir_node *block; + ir_mode *mode; + dbg_info *dbgi; + ir_node *new_const; + ir_node *new_shift; + ir_node *new_and; + tarval *tv_shl; + tarval *tv_shr; + tarval *tv_shift; + tarval *tv_mask; + pn_Cmp pnc; + int need_shrs = 0; + + assert(is_Shl(n) || is_Shr(n) || is_Shrs(n)); + + if (!is_Const(right)) + return n; + + left = get_binop_left(n); mode = get_irn_mode(n); + if (is_Shl(n) && (is_Shr(left) || is_Shrs(left))) { + ir_node *shr_right = get_binop_right(left); - flag = 0; + if (!is_Const(shr_right)) + return n; - modulo_shf = get_mode_modulo_shift(mode); - if (modulo_shf > 0) { - tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res)); + x = get_binop_left(left); + tv_shr = get_Const_tarval(shr_right); + tv_shl = get_Const_tarval(right); - if (tarval_cmp(res, modulo) & pn_Cmp_Lt) - flag = 1; - } else - flag = 1; + if (is_Shrs(left)) { + /* shrs variant only allowed if c1 >= c2 */ + if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge)) + return n; - if (flag) { - /* ok, we can replace it */ - ir_node *in[2], *irn, *block = get_irn_n(n, -1); + tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr); + need_shrs = 1; + } else { + tv_mask = tarval_shr(get_mode_all_one(mode), tv_shr); + } + tv_mask = tarval_shl(tv_mask, tv_shl); + } else if(is_Shr(n) && is_Shl(left)) { + ir_node *shl_right = get_Shl_right(left); - in[0] = get_binop_left(left); - in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res); + if (!is_Const(shl_right)) + return n; - irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in); + x = get_Shl_left(left); + tv_shr = get_Const_tarval(right); + tv_shl = get_Const_tarval(shl_right); - DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + tv_mask = tarval_shl(get_mode_all_one(mode), tv_shl); + tv_mask = tarval_shr(tv_mask, tv_shr); + } else { + return n; + } - return transform_node(irn); + assert(get_tarval_mode(tv_shl) == get_tarval_mode(tv_shr)); + assert(tv_mask != tarval_bad); + assert(get_tarval_mode(tv_mask) == mode); + + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + + pnc = tarval_cmp(tv_shl, tv_shr); + if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) { + tv_shift = tarval_sub(tv_shr, tv_shl); + new_const = new_Const(get_tarval_mode(tv_shift), tv_shift); + if (need_shrs) { + new_shift = new_rd_Shrs(dbgi, irg, block, x, new_const, mode); + } else { + new_shift = new_rd_Shr(dbgi, irg, block, x, new_const, mode); + } + } else { + assert(pnc == pn_Cmp_Gt); + tv_shift = tarval_sub(tv_shl, tv_shr); + new_const = new_Const(get_tarval_mode(tv_shift), tv_shift); + new_shift = new_rd_Shl(dbgi, irg, block, x, new_const, mode); } - return n; -} /* transform_node_shift */ + + new_const = new_Const(mode, tv_mask); + new_and = new_rd_And(dbgi, irg, block, new_shift, new_const, mode); + + return new_and; +} /** * Transform a Shr. */ static ir_node *transform_node_Shr(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Shr_left(n); - ir_node *b = get_Shr_right(n); + ir_node *left = get_Shr_left(n); + ir_node *right = get_Shr_right(n); + ir_mode *mode = get_irn_mode(n); + + HANDLE_BINOP_PHI(tarval_shr, left, right, c, mode); + n = transform_node_shift(n); - HANDLE_BINOP_PHI(tarval_shr, a, b, c); - return transform_node_shift(n); + if (is_Shr(n)) + n = transform_node_shl_shr(n); + if (is_Shr(n)) + n = transform_node_bitop_shift(n); + + return n; } /* transform_node_Shr */ /** @@ -4174,11 +4979,17 @@ static ir_node *transform_node_Shr(ir_node *n) { */ static ir_node *transform_node_Shrs(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Shrs_left(n); - ir_node *b = get_Shrs_right(n); + ir_node *a = get_Shrs_left(n); + ir_node *b = get_Shrs_right(n); + ir_mode *mode = get_irn_mode(n); + + HANDLE_BINOP_PHI(tarval_shrs, a, b, c, mode); + n = transform_node_shift(n); + + if (is_Shrs(n)) + n = transform_node_bitop_shift(n); - HANDLE_BINOP_PHI(tarval_shrs, a, b, c); - return transform_node_shift(n); + return n; } /* transform_node_Shrs */ /** @@ -4186,13 +4997,62 @@ static ir_node *transform_node_Shrs(ir_node *n) { */ static ir_node *transform_node_Shl(ir_node *n) { ir_node *c, *oldn = n; - ir_node *a = get_Shl_left(n); - ir_node *b = get_Shl_right(n); + ir_node *a = get_Shl_left(n); + ir_node *b = get_Shl_right(n); + ir_mode *mode = get_irn_mode(n); + + HANDLE_BINOP_PHI(tarval_shl, a, b, c, mode); + n = transform_node_shift(n); + + if (is_Shl(n)) + n = transform_node_shl_shr(n); + if (is_Shl(n)) + n = transform_node_bitop_shift(n); - HANDLE_BINOP_PHI(tarval_shl, a, b, c); - return transform_node_shift(n); + return n; } /* transform_node_Shl */ +/** + * Transform a Rotl. + */ +static ir_node *transform_node_Rotl(ir_node *n) { + ir_node *c, *oldn = n; + ir_node *a = get_Rotl_left(n); + ir_node *b = get_Rotl_right(n); + ir_mode *mode = get_irn_mode(n); + + HANDLE_BINOP_PHI(tarval_rotl, a, b, c, mode); + n = transform_node_shift(n); + + if (is_Rotl(n)) + n = transform_node_bitop_shift(n); + + return n; +} /* transform_node_Rotl */ + +/** + * Transform a Conv. + */ +static ir_node *transform_node_Conv(ir_node *n) { + ir_node *c, *oldn = n; + ir_node *a = get_Conv_op(n); + + if (is_const_Phi(a)) { + c = apply_conv_on_phi(a, get_irn_mode(n)); + if (c) { + DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); + return c; + } + } + + if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */ + ir_mode *mode = get_irn_mode(n); + return new_r_Unknown(current_ir_graph, mode); + } + + return n; +} /* transform_node_Conv */ + /** * Remove dead blocks and nodes in dead blocks * in keep alive list. We do not generate a new End node. @@ -4224,17 +5084,17 @@ static ir_node *transform_node_End(ir_node *n) { /** returns 1 if a == -b */ static int is_negated_value(ir_node *a, ir_node *b) { - if(is_Minus(a) && get_Minus_op(a) == b) + if (is_Minus(a) && get_Minus_op(a) == b) return 1; - if(is_Minus(b) && get_Minus_op(b) == a) + if (is_Minus(b) && get_Minus_op(b) == a) return 1; - if(is_Sub(a) && is_Sub(b)) { + if (is_Sub(a) && is_Sub(b)) { ir_node *a_left = get_Sub_left(a); ir_node *a_right = get_Sub_right(a); ir_node *b_left = get_Sub_left(b); ir_node *b_right = get_Sub_right(b); - if(a_left == b_right && a_right == b_left) + if (a_left == b_right && a_right == b_left) return 1; } @@ -4247,50 +5107,103 @@ static int is_negated_value(ir_node *a, ir_node *b) { static ir_node *transform_node_Mux(ir_node *n) { ir_node *oldn = n, *sel = get_Mux_sel(n); ir_mode *mode = get_irn_mode(n); + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_graph *irg = current_ir_graph; + ir_node *conds[1], *vals[2]; + + /* first normalization step: move a possible zero to the false case */ + if (is_Proj(sel)) { + ir_node *cmp = get_Proj_pred(sel); + + if (is_Cmp(cmp)) { + if (is_Const(t) && is_Const_null(t)) { + /* Psi(x, 0, y) => Psi(x, y, 0) */ + pn_Cmp pnc = get_Proj_proj(sel); + sel = new_r_Proj(irg, get_nodes_block(cmp), cmp, mode_b, + get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp)))); + conds[0] = sel; + vals[0] = f; + vals[1] = t; + n = new_rd_Psi(get_irn_dbg_info(n), irg, get_nodes_block(n), 1, conds, vals, mode); + t = vals[0]; + f = vals[1]; + } + } + } + /* note: after normalization, false can only happen on default */ if (mode == mode_b) { - ir_node *t = get_Mux_true(n); - ir_node *f = get_Mux_false(n); dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_irn_n(n, -1); + ir_node *block = get_nodes_block(n); ir_graph *irg = current_ir_graph; if (is_Const(t)) { tarval *tv_t = get_Const_tarval(t); if (tv_t == tarval_b_true) { if (is_Const(f)) { + /* Muxb(sel, true, false) = sel */ assert(get_Const_tarval(f) == tarval_b_false); + DBG_OPT_ALGSIM0(oldn, sel, FS_OPT_MUX_BOOL); return sel; } else { - return new_rd_Or(dbg, irg, block, sel, f, mode_b); - } - } else { - ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); - assert(tv_t == tarval_b_false); - if (is_Const(f)) { - assert(get_Const_tarval(f) == tarval_b_true); - return not_sel; - } else { - return new_rd_And(dbg, irg, block, not_sel, f, mode_b); + /* Muxb(sel, true, x) = Or(sel, x) */ + n = new_rd_Or(dbg, irg, block, sel, f, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL); + return n; } } } else if (is_Const(f)) { tarval *tv_f = get_Const_tarval(f); if (tv_f == tarval_b_true) { + /* Muxb(sel, x, true) = Or(Not(sel), x) */ ir_node* not_sel = new_rd_Not(dbg, irg, block, sel, mode_b); - return new_rd_Or(dbg, irg, block, not_sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); + n = new_rd_Or(dbg, irg, block, not_sel, t, mode_b); + return n; } else { + /* Muxb(sel, x, false) = And(sel, x) */ assert(tv_f == tarval_b_false); - return new_rd_And(dbg, irg, block, sel, t, mode_b); + n = new_rd_And(dbg, irg, block, sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); + return n; } } } - if (is_Proj(sel) && !mode_honor_signed_zeros(mode)) { + /* more normalization: try to normalize Mux(x, C1, C2) into Mux(x, +1/-1, 0) op C2 */ + if (is_Const(t) && is_Const(f) && mode_is_int(mode)) { + tarval *a = get_Const_tarval(t); + tarval *b = get_Const_tarval(f); + tarval *null = get_tarval_null(mode); + tarval *diff, *min; + + if (tarval_cmp(a, b) & pn_Cmp_Gt) { + diff = tarval_sub(a, b); + min = b; + } else { + diff = tarval_sub(b, a); + min = a; + } + + if (diff == get_tarval_one(mode) && min != null) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_graph *irg = current_ir_graph; + + + conds[0] = sel; + vals[0] = new_Const(mode, tarval_sub(a, min)); + vals[1] = new_Const(mode, tarval_sub(b, min)); + n = new_rd_Psi(dbg, irg, block, 1, conds, vals, mode); + n = new_rd_Add(dbg, irg, block, n, new_Const(mode, min), mode); + return n; + } + } + + if (is_Proj(sel)) { ir_node *cmp = get_Proj_pred(sel); long pn = get_Proj_proj(sel); - ir_node *f = get_Mux_false(n); - ir_node *t = get_Mux_true(n); /* * Note: normalization puts the constant on the right side, @@ -4298,31 +5211,32 @@ static ir_node *transform_node_Mux(ir_node *n) { * * Note further that these optimization work even for floating point * with NaN's because -NaN == NaN. - * However, if +0 and -0 is handled differently, we cannot use the first - * one. + * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs + * transformations. */ if (is_Cmp(cmp)) { ir_node *cmp_r = get_Cmp_right(cmp); if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - ir_node *block = get_irn_n(n, -1); + ir_node *block = get_nodes_block(n); + ir_node *cmp_l = get_Cmp_left(cmp); - if(is_negated_value(f, t)) { - ir_node *cmp_left = get_Cmp_left(cmp); + if (!mode_honor_signed_zeros(mode) && is_negated_value(f, t)) { + /* f = -t */ - /* Psi(a >= 0, a, -a) = Psi(a <= 0, -a, a) ==> Abs(a) */ - if ( (cmp_left == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)) - || (cmp_left == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))) + if ( (cmp_l == t && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt)) + || (cmp_l == f && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt))) { + /* Psi(a >/>= 0, a, -a) = Psi(a Abs(a) */ n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, - cmp_left, mode); + cmp_l, mode); DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); return n; - /* Psi(a <= 0, a, -a) = Psi(a >= 0, -a, a) ==> -Abs(a) */ - } else if ((cmp_left == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)) - || (cmp_left == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))) + } else if ((cmp_l == t && (pn == pn_Cmp_Le || pn == pn_Cmp_Lt)) + || (cmp_l == f && (pn == pn_Cmp_Ge || pn == pn_Cmp_Gt))) { + /* Psi(a />= 0, -a, a) ==> -Abs(a) */ n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, block, - cmp_left, mode); + cmp_l, mode); n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, block, n, mode); DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS); @@ -4330,49 +5244,58 @@ static ir_node *transform_node_Mux(ir_node *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 ((pn == pn_Cmp_Lt || pn == pn_Cmp_Le) && - is_Const(t) && is_Const_all_one(t) && - is_Const(f) && is_Const_null(f)) { - /* - * 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 ((pn == pn_Cmp_Gt || pn == pn_Cmp_Ge) && - is_Const(t) && is_Const_one(t) && - is_Const(f) && is_Const_null(f)) { - /* - * 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; + if (mode_is_int(mode)) { + /* integer only */ + if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && is_And(cmp_l)) { + /* Psi((a & b) != 0, c, 0) */ + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && f == cmp_r) { + if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) { + if (pn == pn_Cmp_Lg) { + /* Psi((a & 2^C) != 0, 2^C, 0) */ + n = cmp_l; + } else { + /* Psi((a & 2^C) == 0, 2^C, 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + } + return n; + } + } + if (is_Shl(and_r)) { + ir_node *shl_l = get_Shl_left(and_r); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_r == t && f == cmp_r) { + if (pn == pn_Cmp_Lg) { + /* (a & (1 << n)) != 0, (1 << n), 0) */ + n = cmp_l; + } else { + /* (a & (1 << n)) == 0, (1 << n), 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + } + return n; + } + } + } + and_l = get_And_left(cmp_l); + if (is_Shl(and_l)) { + ir_node *shl_l = get_Shl_left(and_l); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_l == t && f == cmp_r) { + if (pn == pn_Cmp_Lg) { + /* ((1 << n) & a) != 0, (1 << n), 0) */ + n = cmp_l; + } else { + /* ((1 << n) & a) == 0, (1 << n), 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), current_ir_graph, + block, cmp_l, t, mode); + } + return n; + } + } } } } @@ -4392,6 +5315,49 @@ static ir_node *transform_node_Psi(ir_node *n) { return n; } /* transform_node_Psi */ +/** + * optimize sync nodes that have other syncs as input we simply add the inputs + * of the other sync to our own inputs + */ +static ir_node *transform_node_Sync(ir_node *n) { + int arity = get_Sync_n_preds(n); + int i; + + for (i = 0; i < arity;) { + ir_node *pred = get_Sync_pred(n, i); + int pred_arity; + int j; + + if (!is_Sync(pred)) { + ++i; + continue; + } + + del_Sync_n(n, i); + --arity; + + pred_arity = get_Sync_n_preds(pred); + for (j = 0; j < pred_arity; ++j) { + ir_node *pred_pred = get_Sync_pred(pred, j); + int k; + + for (k = 0;; ++k) { + if (k >= arity) { + add_irn_n(n, pred_pred); + ++arity; + break; + } + if (get_Sync_pred(n, k) == pred_pred) break; + } + } + } + + /* rehash the sync node */ + add_identities(current_ir_graph->value_table, n); + + return n; +} + /** * Tries several [inplace] [optimizing] transformations and returns an * equivalent node. The difference to equivalent_node() is that these @@ -4453,9 +5419,12 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops CASE(Shr); CASE(Shrs); CASE(Shl); + CASE(Rotl); + CASE(Conv); CASE(End); CASE(Mux); CASE(Psi); + CASE(Sync); default: /* leave NULL */; } @@ -4505,14 +5474,14 @@ static int node_cmp_attr_Free(ir_node *a, ir_node *b) { static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) { const symconst_attr *pa = get_irn_symconst_attr(a); const symconst_attr *pb = get_irn_symconst_attr(b); - return (pa->num != pb->num) + return (pa->kind != pb->kind) || (pa->sym.type_p != pb->sym.type_p) || (pa->tp != pb->tp); } /* node_cmp_attr_SymConst */ /** Compares the attributes of two Call nodes. */ static int node_cmp_attr_Call(ir_node *a, ir_node *b) { - return (get_irn_call_attr(a) != get_irn_call_attr(b)); + return get_irn_call_attr(a) != get_irn_call_attr(b); } /* node_cmp_attr_Call */ /** Compares the attributes of two Sel nodes. */ @@ -4532,8 +5501,8 @@ static int node_cmp_attr_Phi(ir_node *a, ir_node *b) { /* we can only enter this function if both nodes have the same number of inputs, hence it is enough to check if one of them is a Phi0 */ if (is_Phi0(a)) { - /* check the Phi0 attribute */ - return get_irn_phi0_attr(a) != get_irn_phi0_attr(b); + /* check the Phi0 pos attribute */ + return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos; } return 0; } /* node_cmp_attr_Phi */ @@ -4572,6 +5541,49 @@ static int node_cmp_attr_Store(ir_node *a, ir_node *b) { get_Store_volatility(b) == volatility_is_volatile); } /* node_cmp_attr_Store */ +/** Compares two exception attributes */ +static int node_cmp_exception(ir_node *a, ir_node *b) { + const except_attr *ea = get_irn_except_attr(a); + const except_attr *eb = get_irn_except_attr(b); + + return ea->pin_state != eb->pin_state; +} + +#define node_cmp_attr_Bound node_cmp_exception + +/** Compares the attributes of two Div nodes. */ +static int node_cmp_attr_Div(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode || + ma->no_remainder != mb->no_remainder; +} /* node_cmp_attr_Div */ + +/** Compares the attributes of two DivMod nodes. */ +static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_DivMod */ + +/** Compares the attributes of two Mod nodes. */ +static int node_cmp_attr_Mod(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_Mod */ + +/** Compares the attributes of two Quot nodes. */ +static int node_cmp_attr_Quot(ir_node *a, ir_node *b) { + const divmod_attr *ma = get_irn_divmod_attr(a); + const divmod_attr *mb = get_irn_divmod_attr(b); + return ma->exc.pin_state != mb->exc.pin_state || + ma->res_mode != mb->res_mode; +} /* node_cmp_attr_Quot */ + /** Compares the attributes of two Confirm nodes. */ static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) { return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); @@ -4655,6 +5667,12 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) CASE(Store); CASE(Confirm); CASE(ASM); + CASE(Div); + CASE(DivMod); + CASE(Mod); + CASE(Quot); + CASE(Bound); + /* FIXME CopyB */ default: /* leave NULL */; } @@ -4664,30 +5682,32 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) } /* firm_set_default_node_cmp_attr */ /* - * Compare function for two nodes in the hash table. Gets two - * nodes as parameters. Returns 0 if the nodes are a cse. + * Compare function for two nodes in the value table. Gets two + * nodes as parameters. Returns 0 if the nodes are a Common Sub Expression. */ int identities_cmp(const void *elt, const void *key) { - ir_node *a, *b; + ir_node *a = (ir_node *)elt; + ir_node *b = (ir_node *)key; int i, irn_arity_a; - a = (void *)elt; - b = (void *)key; - if (a == b) return 0; if ((get_irn_op(a) != get_irn_op(b)) || (get_irn_mode(a) != get_irn_mode(b))) return 1; /* compare if a's in and b's in are of equal length */ - irn_arity_a = get_irn_intra_arity (a); + irn_arity_a = get_irn_intra_arity(a); if (irn_arity_a != get_irn_intra_arity(b)) return 1; - /* for block-local cse and op_pin_state_pinned nodes: */ - if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) { + if (get_irn_pinned(a) == op_pin_state_pinned) { + /* for pinned nodes, the block inputs must be equal */ if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1)) return 1; + } else if (! get_opt_global_cse()) { + /* for block-local CSE both nodes must be in the same MacroBlock */ + if (get_irn_MacroBlock(a) != get_irn_MacroBlock(b)) + return 1; } /* compare a->in[0..ins] with b->in[0..ins] */ @@ -4708,7 +5728,7 @@ int identities_cmp(const void *elt, const void *key) { /* * Calculate a hash value of a node. */ -unsigned ir_node_hash(ir_node *node) { +unsigned ir_node_hash(const ir_node *node) { unsigned h; int i, irn_arity; @@ -4748,96 +5768,120 @@ void del_identities(pset *value_table) { } /* del_identities */ /** - * Normalize a node by putting constants (and operands with smaller - * node index) on the right + * Normalize a node by putting constants (and operands with larger + * node index) on the right (operator side). * * @param n The node to normalize */ static void normalize_node(ir_node *n) { - if (get_opt_reassociation()) { - if (is_op_commutative(get_irn_op(n))) { - ir_node *l = get_binop_left(n); - ir_node *r = get_binop_right(n); - int l_idx = get_irn_idx(l); - int r_idx = get_irn_idx(r); - - /* For commutative operators perform a OP b == b OP a but keep - constants on the RIGHT side. This helps greatly in some optimizations. - Moreover we use the idx number to make the form deterministic. */ - if (is_irn_constlike(l)) - l_idx = -l_idx; - if (is_irn_constlike(r)) - r_idx = -r_idx; - if (l_idx < r_idx) { - set_binop_left(n, r); - set_binop_right(n, l); - } + if (is_op_commutative(get_irn_op(n))) { + ir_node *l = get_binop_left(n); + ir_node *r = get_binop_right(n); + + /* For commutative operators perform a OP b == b OP a but keep + * constants on the RIGHT side. This helps greatly in some + * optimizations. Moreover we use the idx number to make the form + * deterministic. */ + if (!operands_are_normalized(l, r)) { + set_binop_left(n, r); + set_binop_right(n, l); } } } /* normalize_node */ /** + * Update the nodes after a match in the value table. If both nodes have + * the same MacroBlock but different Blocks, we must ensure that the node + * with the dominating Block (the node that is near to the MacroBlock header + * is stored in the table. + * Because a MacroBlock has only one "non-exception" flow, we don't need + * dominance info here: We known, that one block must dominate the other and + * following the only block input will allow to find it. + */ +static void update_known_irn(ir_node *known_irn, const ir_node *new_ir_node) { + ir_node *known_blk, *new_block, *block, *mbh; + + if (get_opt_global_cse()) { + /* Block inputs are meaning less */ + return; + } + known_blk = get_irn_n(known_irn, -1); + new_block = get_irn_n(new_ir_node, -1); + if (known_blk == new_block) { + /* already in the same block */ + return; + } + /* + * We expect the typical case when we built the graph. In that case, the + * known_irn is already the upper one, so checking this should be faster. + */ + block = new_block; + mbh = get_Block_MacroBlock(new_block); + for (;;) { + if (block == known_blk) { + /* ok, we have found it: known_block dominates new_block as expected */ + return; + } + if (block == mbh) { + /* + * We have reached the MacroBlock header NOT founding + * the known_block. new_block must dominate known_block. + * Update known_irn. + */ + set_irn_n(known_irn, -1, new_block); + return; + } + assert(get_Block_n_cfgpreds(block) == 1); + block = get_Block_cfgpred_block(block, 0); + } +} /* update_value_table */ + +/* * Return the canonical node computing the same value as n. + * Looks up the node in a hash table, enters it in the table + * if it isn't there yet. * - * @param value_table The value table - * @param n The node to lookup - * - * Looks up the node in a hash table. + * @param value_table the HashSet containing all nodes in the + * current IR graph + * @param n the node to look up * - * For Const nodes this is performed in the constructor, too. Const - * nodes are extremely time critical because of their frequent use in - * constant string arrays. + * @return a node that computes the same value as n or n if no such + * node could be found */ -static INLINE ir_node *identify(pset *value_table, ir_node *n) { +ir_node *identify_remember(pset *value_table, ir_node *n) { ir_node *o = NULL; if (!value_table) return n; normalize_node(n); + /* lookup or insert in hash table with given hash key. */ + o = pset_insert(value_table, n, ir_node_hash(n)); - o = pset_find(value_table, n, ir_node_hash(n)); - if (!o) return n; - - DBG_OPT_CSE(n, o); + if (o != n) { + update_known_irn(o, n); + DBG_OPT_CSE(n, o); + } return o; -} /* identify */ +} /* identify_remember */ /** * During construction we set the op_pin_state_pinned flag in the graph right when the * optimization is performed. The flag turning on procedure global cse could * be changed between two allocations. This way we are safe. + * + * @param value_table The value table + * @param n The node to lookup */ static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) { ir_node *old = n; - n = identify(value_table, n); - if (get_irn_n(old, -1) != get_irn_n(n, -1)) + n = identify_remember(value_table, n); + if (n != old && get_irn_MacroBlock(old) != get_irn_MacroBlock(n)) set_irg_pinned(current_ir_graph, op_pin_state_floats); return n; } /* identify_cons */ -/* - * Return the canonical node computing the same value as n. - * Looks up the node in a hash table, enters it in the table - * if it isn't there yet. - */ -ir_node *identify_remember(pset *value_table, ir_node *n) { - ir_node *o = NULL; - - if (!value_table) return n; - - normalize_node(n); - /* lookup or insert in hash table with given hash key. */ - o = pset_insert(value_table, n, ir_node_hash(n)); - - if (o != n) { - DBG_OPT_CSE(n, o); - } - - return o; -} /* identify_remember */ - /* Add a node to the identities value table. */ void add_identities(pset *value_table, ir_node *node) { if (get_opt_cse() && is_no_Block(node)) @@ -4869,22 +5913,35 @@ static ir_node *gigo(ir_node *node) { ir_node *block = get_nodes_block(skip_Proj(node)); /* Don't optimize nodes in immature blocks. */ - if (!get_Block_matured(block)) return node; + if (!get_Block_matured(block)) + return node; /* Don't optimize End, may have Bads. */ if (op == op_End) return node; if (is_Block(block)) { - irn_arity = get_irn_arity(block); - for (i = 0; i < irn_arity; i++) { + if (is_Block_dead(block)) { + /* control flow from dead block is dead */ + return new_Bad(); + } + + for (i = get_irn_arity(block) - 1; i >= 0; --i) { if (!is_Bad(get_irn_n(block, i))) break; } - if (i == irn_arity) { + if (i < 0) { ir_graph *irg = get_irn_irg(block); /* the start block is never dead */ if (block != get_irg_start_block(irg) - && block != get_irg_end_block(irg)) + && block != get_irg_end_block(irg)) { + /* + * Do NOT kill control flow without setting + * the block to dead of bad things can happen: + * We get a Block that is not reachable be irg_block_walk() + * but can be found by irg_walk()! + */ + set_Block_dead(block); return new_Bad(); + } } } } @@ -4898,7 +5955,7 @@ static ir_node *gigo(ir_node *node) { * Beware: we can only read the block of a non-floating node. */ if (is_irn_pinned_in_irg(node) && - is_Block_dead(get_nodes_block(node))) + is_Block_dead(get_nodes_block(skip_Proj(node)))) return new_Bad(); for (i = 0; i < irn_arity; i++) { @@ -5001,7 +6058,7 @@ ir_node *optimize_node(ir_node *n) { } /* remove unnecessary nodes */ - if (get_opt_constant_folding() || + if (get_opt_algebraic_simplification() || (iro == iro_Phi) || /* always optimize these nodes. */ (iro == iro_Id) || (iro == iro_Proj) || @@ -5029,14 +6086,14 @@ ir_node *optimize_node(ir_node *n) { /* Some more constant expression evaluation that does not allow to free the node. */ iro = get_irn_opcode(n); - if (get_opt_constant_folding() || + if (get_opt_algebraic_simplification() || (iro == iro_Cond) || (iro == iro_Proj)) /* Flags tested local. */ n = transform_node(n); /* Remove nodes with dead (Bad) input. Run always for transformation induced Bads. */ - n = gigo (n); + n = gigo(n); /* Now we have a legal, useful node. Enter it in hash table for CSE */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { @@ -5105,7 +6162,7 @@ ir_node *optimize_in_place_2(ir_node *n) { now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common subexpressions within a block. */ if (get_opt_cse()) { - n = identify(current_ir_graph->value_table, n); + n = identify_remember(current_ir_graph->value_table, n); } /* Some more constant expression evaluation. */