X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=c2cecb135541755428bf95090f081897abbfa566;hb=7498ed3766f3f98604d96d1b75978a3a9e6768ed;hp=b034544ea73ae9f63ce3008b84d6e74c5305347b;hpb=ef68952c161680fb53ba0bd3edbc51f02d79e4d6;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index b034544ea..c2cecb135 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. * @@ -722,10 +722,16 @@ 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; - /* The Block constructor does not call optimize, but mature_immBlock - calls the optimization. */ + /* 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. */ assert(get_Block_matured(n)); /* Straightening: a single entry Block following a single exit Block @@ -971,7 +977,7 @@ 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. @@ -1159,7 +1165,6 @@ 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); @@ -1189,11 +1194,8 @@ static ir_node *equivalent_node_Conv(ir_node *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 (n_mode == b_mode) { if (n_mode == mode_b) { @@ -1236,7 +1238,7 @@ 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; @@ -1244,8 +1246,6 @@ 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. */ @@ -1303,45 +1303,36 @@ 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 */ @@ -1449,41 +1440,54 @@ 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); } } @@ -1581,34 +1585,27 @@ 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) { @@ -1760,6 +1757,50 @@ 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). * @@ -1842,17 +1883,17 @@ 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); @@ -1870,12 +1911,12 @@ 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); @@ -1891,6 +1932,19 @@ 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 */ @@ -1905,6 +1959,10 @@ static ir_node *transform_node_AddSub(ir_node *n) { /* check for Op(Const, Phi) */ \ c = apply_binop_on_phi(b, get_Const_tarval(a), eval, mode, 1);\ } \ + else if (is_const_Phi(a) && is_const_Phi(b)) { \ + /* check for Op(Phi, Phi) */ \ + c = apply_binop_on_2_phis(a, b, eval, mode); \ + } \ if (c) { \ DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \ return c; \ @@ -1983,74 +2041,6 @@ static ir_node *transform_node_Add(ir_node *n) { DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); return n; } - if (! is_reassoc_running()) { - /* do NOT execute this code if reassociation is enabled, it does the inverse! */ - if (is_Mul(a)) { - ir_node *ma = get_Mul_left(a); - ir_node *mb = get_Mul_right(a); - - if (b == ma) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } else if (b == mb) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - } - if (is_Mul(b)) { - ir_node *ma = get_Mul_left(b); - ir_node *mb = get_Mul_right(b); - - if (a == ma) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - if (a == mb) { - ir_node *blk = get_irn_n(n, -1); - n = new_rd_Mul( - get_irn_dbg_info(n), current_ir_graph, blk, - mb, - new_rd_Add( - get_irn_dbg_info(n), current_ir_graph, blk, - ma, - new_r_Const_long(current_ir_graph, blk, mode, 1), - mode), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A); - return n; - } - } - } if (get_mode_arithmetic(mode) == irma_twos_complement) { /* Here we rely on constants be on the RIGHT side */ if (is_Not(a)) { @@ -2267,16 +2257,16 @@ 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); @@ -2385,12 +2375,14 @@ 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); @@ -2398,12 +2390,14 @@ 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); @@ -2489,7 +2483,8 @@ 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; } @@ -2497,7 +2492,8 @@ 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; } @@ -2510,10 +2506,39 @@ 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); @@ -2569,10 +2594,39 @@ 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); @@ -2632,14 +2686,48 @@ 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; @@ -2648,28 +2736,30 @@ 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 */ @@ -2677,7 +2767,7 @@ 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; } @@ -2695,8 +2785,8 @@ 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; @@ -2745,6 +2835,7 @@ static ir_node *transform_node_Quot(ir_node *n) { /** * Optimize Abs(x) into x if x is Confirmed >= 0 * Optimize Abs(x) into -x if x is Confirmed <= 0 + * Optimize Abs(-x) int Abs(x) */ static ir_node *transform_node_Abs(ir_node *n) { ir_node *c, *oldn = n; @@ -2759,13 +2850,14 @@ static ir_node *transform_node_Abs(ir_node *n) { /* * We can replace the Abs by -x here. - * We even could add a new Confirm here. + * We even could add a new Confirm here + * (if not twos complement) * * Note that -x would create a new node, so we could * not run it in the equivalent_node() context. */ n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, - get_irn_n(n, -1), a, mode); + get_nodes_block(n), a, mode); DBG_OPT_CONFIRM(oldn, n); return n; @@ -2776,8 +2868,17 @@ static ir_node *transform_node_Abs(ir_node *n) { DBG_OPT_CONFIRM(oldn, n); return n; default: + break; + } + if (is_Minus(a)) { + /* Abs(-x) = Abs(x) */ + mode = get_irn_mode(n); + n = new_rd_Abs(get_irn_dbg_info(n), current_ir_graph, + get_nodes_block(n), get_Minus_op(a), mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ABS_MINUS_X); return n; } + return n; } /* transform_node_Abs */ /** @@ -2801,7 +2902,8 @@ 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()); @@ -2811,12 +2913,16 @@ 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 @@ -2867,7 +2973,8 @@ 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) { @@ -2908,7 +3015,7 @@ static ir_node *transform_bitwise_distributive(ir_node *n, n = new_rd_And(dbgi, irg, blk, new_n, c, mode); } else { n = exact_copy(a); - set_irn_n(n, -1, blk); + set_nodes_block(n, blk); set_binop_left(n, new_n); set_binop_right(n, c); add_identities(current_ir_graph->value_table, n); @@ -3208,7 +3315,7 @@ 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); @@ -3224,7 +3331,7 @@ 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); @@ -3251,13 +3358,15 @@ static ir_node *transform_node_Minus(ir_node *n) { ir_node *mul_r = get_Mul_right(a); if (is_Const(mul_r)) { tarval *tv = tarval_neg(get_Const_tarval(mul_r)); - ir_node *cnst = new_Const(mode, tv); - dbg_info *dbg = get_irn_dbg_info(a); - ir_graph *irg = current_ir_graph; - ir_node *block = get_nodes_block(a); - n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode); - DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C); - return n; + 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; + } } } @@ -3277,8 +3386,8 @@ 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); } @@ -3481,38 +3590,72 @@ 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 */ + 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 */ - if (is_Cast(left)) - left = get_Cast_op(left); - if (is_Cast(right)) - right = get_Cast_op(right); + /* remove Casts of both sides */ + left = skip_Cast(left); + right = skip_Cast(right); /* Remove unnecessary conversions */ /* TODO handle constants */ @@ -3523,7 +3666,8 @@ 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); @@ -3546,90 +3690,131 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } } - /* remove operation of both sides if possible */ + /* remove operation on both sides if possible */ if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { - 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; - 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; + /* + * 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_Rot: + if (get_Rot_right(left) == get_Rot_right(right)) { + /* a ROT X CMP b ROT X ==> a CMP b */ + left = get_Rot_left(left); + right = get_Rot_left(right); + 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; - 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; + } + + /* 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); @@ -3650,28 +3835,12 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { } } - if (!get_opt_reassociation()) - return proj; - /* * First step: normalize the compare op * by placing the constant on the right side * or moving the lower address node to the left. - * We ignore the case that both are constants - * this case should be optimized away. */ - if (is_Const(right)) { - c = right; - } else if (is_irn_constlike(left)) { - c = left; - left = right; - right = c; - if(!is_Const(c)) - c = NULL; - - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; - } else if (get_irn_idx(left) > get_irn_idx(right)) { + if (!operands_are_normalized(left, right)) { ir_node *t = left; left = right; @@ -3687,9 +3856,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { * later and may help to normalize more compares. * Of course this is only possible for integer values. */ - if (c) { - mode = get_irn_mode(c); - tv = get_Const_tarval(c); + 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)) { @@ -4046,9 +4215,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) { /* 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; @@ -4291,7 +4458,7 @@ 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 */ @@ -4316,7 +4483,7 @@ 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 */ @@ -4339,7 +4506,7 @@ 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 */ @@ -4535,6 +4702,12 @@ static ir_node *transform_node_Conv(ir_node *n) { return c; } } + + if (is_Unknown(a)) { /* Conv_A(Unknown_B) -> Unknown_A */ + ir_mode *mode = get_irn_mode(n); + return new_r_Unknown(current_ir_graph, mode); + } + return n; } /* transform_node_Conv */ @@ -4674,7 +4847,6 @@ static ir_node *transform_node_Mux(ir_node *n) { return n; } } - } } } @@ -4696,28 +4868,35 @@ static ir_node *transform_node_Psi(ir_node *n) { * of the other sync to our own inputs */ static ir_node *transform_node_Sync(ir_node *n) { - int i, arity; + int arity = get_Sync_n_preds(n); + int i; + + for (i = 0; i < arity;) { + ir_node *pred = get_Sync_pred(n, i); + int pred_arity; + int j; - arity = get_irn_arity(n); - for(i = 0; i < get_irn_arity(n); /*empty*/) { - int i2, arity2; - ir_node *in = get_irn_n(n, i); - if(!is_Sync(in)) { + if (!is_Sync(pred)) { ++i; continue; } - /* set sync input 0 instead of the sync */ - set_irn_n(n, i, get_irn_n(in, 0)); - /* so we check this input again for syncs */ - - /* append all other inputs of the sync to our sync */ - arity2 = get_irn_arity(in); - for(i2 = 1; i2 < arity2; ++i2) { - ir_node *in_in = get_irn_n(in, i2); - add_irn_n(n, in_in); - /* increase arity so we also check the new inputs for syncs */ - arity++; + 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; + } } } @@ -4843,14 +5022,14 @@ 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. */ @@ -4870,8 +5049,8 @@ 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 */ @@ -4910,6 +5089,49 @@ 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)); @@ -4993,6 +5215,12 @@ 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 */; } @@ -5002,30 +5230,32 @@ 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] */ @@ -5086,34 +5316,74 @@ 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. * @@ -5134,8 +5404,10 @@ static INLINE ir_node *identify(pset *value_table, ir_node *n) { normalize_node(n); o = pset_find(value_table, n, ir_node_hash(n)); - if (!o) return n; + if (o == NULL) + return n; + update_known_irn(o, n); DBG_OPT_CSE(n, o); return o; @@ -5145,12 +5417,15 @@ static INLINE ir_node *identify(pset *value_table, ir_node *n) { * 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)) + 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 */ @@ -5159,6 +5434,13 @@ static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) { * 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 HashSet containing all nodes in the + * current IR graph + * @param n the node to look up + * + * @return a node that computes the same value as n or n if no such + * node could be found */ ir_node *identify_remember(pset *value_table, ir_node *n) { ir_node *o = NULL; @@ -5170,6 +5452,7 @@ ir_node *identify_remember(pset *value_table, ir_node *n) { o = pset_insert(value_table, n, ir_node_hash(n)); if (o != n) { + update_known_irn(o, n); DBG_OPT_CSE(n, o); } @@ -5207,22 +5490,35 @@ 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(); + } } } } @@ -5236,7 +5532,7 @@ 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++) {