X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=ca58b2ef4eeb06915ab19343d0987be77cbe1dc5;hb=a08e6f04aa3669cff094f94a9484c7c2bb1314d0;hp=679410da3f0fa60ac5a9fded3452f880dc68e96c;hpb=64a05bd2846f1dbe27318977303d411fc4149484;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 679410da3..ca58b2ef4 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -131,6 +131,12 @@ static ir_tarval *computed_value_Add(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) return tarval_add(ta, tb); + /* x+~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } + return tarval_bad; } /* computed_value_Add */ @@ -264,10 +270,17 @@ static ir_tarval *computed_value_And(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_and (ta, tb); - } else { - if (tarval_is_null(ta)) return ta; - if (tarval_is_null(tb)) return tb; } + + if (tarval_is_null(ta)) return ta; + if (tarval_is_null(tb)) return tb; + + /* x&~x => 0 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_null(get_irn_mode(n)); + } + return tarval_bad; } /* computed_value_And */ @@ -284,9 +297,15 @@ static ir_tarval *computed_value_Or(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_or (ta, tb); - } else { - if (tarval_is_all_one(ta)) return ta; - if (tarval_is_all_one(tb)) return tb; + } + + if (tarval_is_all_one(ta)) return ta; + if (tarval_is_all_one(tb)) return tb; + + /* x|~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); } return tarval_bad; } /* computed_value_Or */ @@ -303,6 +322,11 @@ static ir_tarval *computed_value_Eor(const ir_node *n) if (a == b) return get_mode_null(get_irn_mode(n)); + /* x^~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } ta = value_of(a); tb = value_of(b); @@ -327,6 +351,28 @@ static ir_tarval *computed_value_Not(const ir_node *n) return tarval_bad; } /* computed_value_Not */ +/** + * Tests whether a shift shifts more bits than available in the mode + */ +static bool is_oversize_shift(const ir_node *n) +{ + ir_node *count = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_tarval *tv = value_of(count); + long modulo_shift; + long shiftval; + if (tv == tarval_bad) + return false; + if (!tarval_is_long(tv)) + return false; + shiftval = get_tarval_long(tv); + modulo_shift = get_mode_modulo_shift(mode); + if (shiftval < 0 || (modulo_shift > 0 && shiftval >= modulo_shift)) + return false; + + return shiftval >= (long)get_mode_size_bits(mode); +} + /** * Return the value of a Shl. */ @@ -341,6 +387,10 @@ static ir_tarval *computed_value_Shl(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shl(ta, tb); } + + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; } /* computed_value_Shl */ @@ -358,6 +408,9 @@ static ir_tarval *computed_value_Shr(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shr(ta, tb); } + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; } /* computed_value_Shr */ @@ -395,17 +448,52 @@ static ir_tarval *computed_value_Rotl(const ir_node *n) return tarval_bad; } /* computed_value_Rotl */ +bool ir_zero_when_converted(const ir_node *node, ir_mode *dest_mode) +{ + ir_mode *mode = get_irn_mode(node); + if (get_mode_arithmetic(mode) != irma_twos_complement + || get_mode_arithmetic(dest_mode) != irma_twos_complement) + return false; + + if (is_Shl(node)) { + ir_node *count = get_Shl_right(node); + if (is_Const(count)) { + ir_tarval *tv = get_Const_tarval(count); + if (tarval_is_long(tv)) { + long shiftval = get_tarval_long(tv); + long destbits = get_mode_size_bits(dest_mode); + if (shiftval >= destbits + && shiftval < (long)get_mode_modulo_shift(mode)) + return true; + } + } + } + if (is_And(node)) { + ir_node *right = get_And_right(node); + if (is_Const(right)) { + ir_tarval *tv = get_Const_tarval(right); + ir_tarval *conved = tarval_convert_to(tv, dest_mode); + return tarval_is_null(conved); + } + } + return false; +} + /** * Return the value of a Conv. */ static ir_tarval *computed_value_Conv(const ir_node *n) { - ir_node *a = get_Conv_op(n); - ir_tarval *ta = value_of(a); + ir_node *a = get_Conv_op(n); + ir_tarval *ta = value_of(a); + ir_mode *mode = get_irn_mode(n); if (ta != tarval_bad) return tarval_convert_to(ta, get_irn_mode(n)); + if (ir_zero_when_converted(a, mode)) + return get_mode_null(mode); + return tarval_bad; } /* computed_value_Conv */ @@ -661,72 +749,6 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops #undef CASE } /* firm_set_default_computed_value */ -/** - * Returns a equivalent block for another block. - * If the block has only one predecessor, this is - * the equivalent one. If the only predecessor of a block is - * the block itself, this is a dead block. - * - * If both predecessors of a block are the branches of a binary - * Cond, the equivalent block is Cond's block. - * - * If all predecessors of a block are bad or lies in a dead - * block, the current block is dead as well. - */ -static ir_node *equivalent_node_Block(ir_node *n) -{ - ir_node *oldn = n; - int n_preds; - ir_graph *irg; - - /* don't optimize labeled blocks */ - if (has_Block_entity(n)) - return n; - if (!get_Block_matured(n)) - return n; - - n_preds = get_Block_n_cfgpreds(n); - - irg = get_irn_irg(n); - - /* Straightening: a single entry Block following a single exit Block - * can be merged. */ - if (n_preds == 1) { - ir_node *pred = get_Block_cfgpred(n, 0); - - if (is_Jmp(pred)) { - ir_node *pred_block = get_nodes_block(pred); - DBG_OPT_STG(n, pred_block); - return pred_block; - } - } else if (n_preds == 2) { - /* Test whether Cond jumps twice to this block - * The more general case which more than 2 predecessors is handles - * in optimize_cf(), we handle only this special case for speed here. - */ - ir_node *a = get_Block_cfgpred(n, 0); - ir_node *b = get_Block_cfgpred(n, 1); - - if (is_Proj(a) && is_Proj(b)) { - ir_node *cond = get_Proj_pred(a); - - if (cond == get_Proj_pred(b) && is_Cond(cond) && - get_irn_mode(get_Cond_selector(cond)) == mode_b) { - /* Also a single entry Block following a single exit Block. - * Phis have twice the same operand and will be optimized away. - */ - n = get_nodes_block(cond); - DBG_OPT_IFSIM1(oldn, a, b, n); - } - } - } - - return n; -} /* equivalent_node_Block */ - -/* We do not evaluate Cond here as we replace it by a new node, a Jmp. - See transform_node_Proj_Cond(). */ - /** * Optimize operations that are commutative and have neutral 0, * so a op 0 = 0 op a = a. @@ -935,7 +957,7 @@ static ir_node *equivalent_node_Sub(ir_node *n) /** - * Optimize an "self-inverse unary op", ie op(op(n)) = n. + * Optimize an "self-inverse unary op", i.e. op(op(n)) = n. * * @todo * -(-a) == a, but might overflow two times. @@ -1230,7 +1252,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) int i, n_preds; ir_node *oldn = n; - ir_node *block; ir_node *first_val = NULL; /* to shutup gcc */ if (!get_opt_optimize() && @@ -1239,8 +1260,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) n_preds = get_Phi_n_preds(n); - block = get_nodes_block(n); - /* Phi of dead Region without predecessors. */ if (n_preds == 0) return n; @@ -1576,7 +1595,6 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op break switch (code) { - CASE(Block); CASE(Eor); CASE(Add); CASE(Shl); @@ -2022,22 +2040,6 @@ static ir_node *transform_node_Add(ir_node *n) DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); return n; } - if (op == b) { - /* ~x + x = -1 */ - n = new_r_Const(irg, get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } - } - if (is_Not(b)) { - ir_node *op = get_Not_op(b); - - if (op == a) { - /* x + ~x = -1 */ - n = new_r_Const(irg, get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } } } } @@ -2179,6 +2181,7 @@ restart: } DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; +#if 0 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ ir_node *m_right = get_Mul_right(b); if (is_Const(m_right)) { @@ -2197,6 +2200,7 @@ restart: return n; } } +#endif } /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ @@ -2358,8 +2362,8 @@ restart: } if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* c - ~X = X + (c+1) */ if (is_Const(a) && is_Not(b)) { - /* c - ~X = X + (c+1) */ ir_tarval *tv = get_Const_tarval(a); tv = tarval_add(tv, get_mode_one(mode)); @@ -2372,6 +2376,24 @@ restart: return n; } } + /* x-(x&y) = x & ~y */ + if (is_And(b)) { + ir_node *and_left = get_And_left(b); + ir_node *and_right = get_And_right(b); + if (and_right == a) { + ir_node *tmp = and_left; + and_left = and_right; + and_right = tmp; + } + if (and_left == a) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + ir_node *notn = new_rd_Not(dbgi, block, and_right, mode); + ir_node *and = new_rd_And(dbgi, block, a, notn, mode); + return and; + } + } } return n; } /* transform_node_Sub */ @@ -2643,7 +2665,7 @@ make_tuple: /* skip a potential Pin */ mem = skip_Pin(mem); - turn_into_tuple(n, pn_Div_max); + turn_into_tuple(n, pn_Div_max+1); set_Tuple_pred(n, pn_Div_M, mem); set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk)); set_Tuple_pred(n, pn_Div_X_except, new_r_Bad(irg, mode_X)); @@ -2735,7 +2757,7 @@ make_tuple: /* skip a potential Pin */ mem = skip_Pin(mem); - turn_into_tuple(n, pn_Mod_max); + turn_into_tuple(n, pn_Mod_max+1); set_Tuple_pred(n, pn_Mod_M, mem); set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk)); set_Tuple_pred(n, pn_Mod_X_except, new_r_Bad(irg, mode_X)); @@ -2769,7 +2791,7 @@ static ir_node *transform_node_Cond(ir_node *n) Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ ir_node *blk = get_nodes_block(n); jmp = new_r_Jmp(blk); - turn_into_tuple(n, pn_Cond_max); + turn_into_tuple(n, pn_Cond_max+1); if (ta == tarval_b_true) { set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X)); set_Tuple_pred(n, pn_Cond_true, jmp); @@ -2905,6 +2927,186 @@ static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode) return cnst; } +static bool is_shiftop(const ir_node *n) +{ + return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n); +} + +/** + * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is good for things like x-(x&y) esp. in 186.crafty. + */ +static ir_node *transform_node_shift_bitop(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *left; + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_shift; + + if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + return n; + + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); + + if (!is_Const(right)) + return n; + + left = get_binop_left(n); + op_left = get_irn_op(left); + if (op_left != op_And && op_left != op_Or && op_left != op_Eor) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { + /* TODO: test if sign bit is affectes */ + return n; + } + + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; + + bitop_left = get_binop_left(left); + + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); + + assert(get_tarval_mode(tv1) == mode); + + if (is_Shl(n)) { + new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shl(tv1, tv2); + } else if (is_Shr(n)) { + new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shr(tv1, tv2); + } else if (is_Shrs(n)) { + new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shrs(tv1, tv2); + } else { + assert(is_Rotl(n)); + new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_rotl(tv1, tv2); + } + + assert(get_tarval_mode(tv_shift) == mode); + irg = get_irn_irg(n); + new_const = new_r_Const(irg, tv_shift); + + if (op_left == op_And) { + new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); + } else if (op_left == op_Or) { + new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); + } else { + assert(op_left == op_Eor); + new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); + } + + return new_bitop; +} + +/** + * normalisation: (x >> c1) & c2 to (x & (c2<> c1 + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is usually good for the backend since << C can often be + * matched as address-mode. + */ +static ir_node *transform_node_bitop_shift(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *shift_left; + ir_node *shift_right; + ir_node *block; + dbg_info *dbg_bitop; + dbg_info *dbg_shift; + ir_node *new_bitop; + ir_node *new_shift; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_bitop; + + if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + return n; + + assert(is_And(n) || is_Or(n) || is_Eor(n)); + if (!is_Const(right) || !is_shiftop(left)) + return n; + + shift_left = get_binop_left(left); + shift_right = get_binop_right(left); + if (!is_Const(shift_right)) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(left)) { + /* TODO this could be improved */ + return n; + } + + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbg_bitop = get_irn_dbg_info(n); + dbg_shift = get_irn_dbg_info(left); + tv1 = get_Const_tarval(shift_right); + tv2 = get_Const_tarval(right); + assert(get_tarval_mode(tv2) == mode); + + if (is_Shl(left)) { + tv_bitop = tarval_shr(tv2, tv1); + } else if (is_Shr(left)) { + tv_bitop = tarval_shl(tv2, tv1); + } else { + assert(is_Rotl(left)); + tv_bitop = tarval_rotl(tv2, tarval_neg(tv1)); + } + new_const = new_r_Const(irg, tv_bitop); + + if (is_And(n)) { + new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode); + } else if (is_Or(n)) { + new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode); + } else { + assert(is_Eor(n)); + new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode); + } + + if (is_Shl(left)) { + new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode); + } else if (is_Shr(left)) { + new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode); + } else if (is_Rotl(left)) { + assert(is_Rotl(left)); + new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode); + } + + return new_shift; +} + /** * Transform an And. */ @@ -2931,20 +3133,21 @@ static ir_node *transform_node_And(ir_node *n) ir_relation new_relation = a_relation & b_relation; return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); } - /* Cmp(a==0) and Cmp(b==0) can be optimized to Cmp(a|b==0) */ - if (is_Const(a_right) && is_Const_null(a_right) - && is_Const(b_right) && is_Const_null(b_right) - && a_relation == b_relation && a_relation == ir_relation_equal - && !mode_is_float(get_irn_mode(a_left)) - && !mode_is_float(get_irn_mode(b_left))) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_mode *mode = get_irn_mode(a_left); - ir_node *n_b_left = get_irn_mode(b_left) != mode ? - new_rd_Conv(dbgi, block, b_left, mode) : b_left; - ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode); - ir_graph *irg = get_irn_irg(n); - ir_node *zero = create_zero_const(irg, mode); + /* Cmp(a==b) and Cmp(c==d) can be optimized to Cmp((a^b)|(c^d)==0) */ + if (a_relation == b_relation && a_relation == ir_relation_equal + && !mode_is_float(get_irn_mode(a_left)) + && !mode_is_float(get_irn_mode(b_left)) + && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_graph *irg = get_irn_irg(n); + ir_node *zero = create_zero_const(irg, a_mode); return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); } } @@ -3068,6 +3271,8 @@ static ir_node *transform_node_And(ir_node *n) } n = transform_bitwise_distributive(n, transform_node_And); + if (is_And(n)) + n = transform_node_bitop_shift(n); return n; } /* transform_node_And */ @@ -3154,6 +3359,9 @@ static ir_node *transform_node_Eor(ir_node *n) } n = transform_bitwise_distributive(n, transform_node_Eor); + if (is_Eor(n)) + n = transform_node_bitop_shift(n); + return n; } /* transform_node_Eor */ @@ -3612,30 +3820,51 @@ static ir_node *transform_node_Cmp(ir_node *n) } /* Remove unnecessary conversions */ - /* TODO handle conv+constant */ if (is_Conv(left) && is_Conv(right)) { - ir_node *op_left = get_Conv_op(left); - ir_node *op_right = get_Conv_op(right); - ir_mode *mode_left = get_irn_mode(op_left); - ir_mode *mode_right = get_irn_mode(op_right); + ir_node *op_left = get_Conv_op(left); + ir_node *op_right = get_Conv_op(right); + ir_mode *mode_left = get_irn_mode(op_left); + ir_mode *mode_right = get_irn_mode(op_right); if (smaller_mode(mode_left, mode) && smaller_mode(mode_right, mode) && mode_left != mode_b && mode_right != mode_b) { - ir_node *block = get_nodes_block(n); + ir_node *block = get_nodes_block(n); if (mode_left == mode_right) { - left = op_left; - right = op_right; + left = op_left; + right = op_right; changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV_CONV); } else if (smaller_mode(mode_left, mode_right)) { - left = new_r_Conv(block, op_left, mode_right); - right = op_right; + left = new_r_Conv(block, op_left, mode_right); + right = op_right; changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } else if (smaller_mode(mode_right, mode_left)) { - left = op_left; - right = new_r_Conv(block, op_right, mode_left); + left = op_left; + right = new_r_Conv(block, op_right, mode_left); + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); + } + mode = get_irn_mode(left); + } + } + if (is_Conv(left) && is_Const(right)) { + ir_node *op_left = get_Conv_op(left); + ir_mode *mode_left = get_irn_mode(op_left); + if (smaller_mode(mode_left, mode) && mode_left != mode_b) { + ir_tarval *tv = get_Const_tarval(right); + tarval_int_overflow_mode_t last_mode + = tarval_get_integer_overflow_mode(); + ir_tarval *new_tv; + tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD); + new_tv = tarval_convert_to(tv, mode_left); + tarval_set_integer_overflow_mode(last_mode); + if (new_tv != tarval_bad) { + ir_graph *irg = get_irn_irg(n); + left = op_left; + right = new_r_Const(irg, new_tv); + mode = get_irn_mode(left); changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } @@ -4619,20 +4848,19 @@ static ir_node *transform_node_Or_Rotl(ir_node *irn_or) return n; } /* transform_node_Or_Rotl */ -static bool is_cmp_unequal_zero(const ir_node *node) +static bool is_cmp_unequal(const ir_node *node) { ir_relation relation = get_Cmp_relation(node); ir_node *left = get_Cmp_left(node); ir_node *right = get_Cmp_right(node); ir_mode *mode = get_irn_mode(left); - if (!is_Const(right) || !is_Const_null(right)) - return false; - if (mode_is_signed(mode)) { - return relation == ir_relation_less_greater; - } else { + if (relation == ir_relation_less_greater) + return true; + + if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right)) return relation == ir_relation_greater; - } + return false; } /** @@ -4672,18 +4900,21 @@ static ir_node *transform_node_Or(ir_node *n) ir_relation new_relation = a_relation | b_relation; return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); } - /* Cmp(a!=0) or Cmp(b!=0) => Cmp(a|b != 0) */ - if (is_cmp_unequal_zero(a) && is_cmp_unequal_zero(b) - && !mode_is_float(get_irn_mode(a_left)) - && !mode_is_float(get_irn_mode(b_left))) { - ir_graph *irg = get_irn_irg(n); - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_mode *mode = get_irn_mode(a_left); - ir_node *n_b_left = get_irn_mode(b_left) != mode ? - new_rd_Conv(dbgi, block, b_left, mode) : b_left; - ir_node *or = new_rd_Or(dbgi, block, a_left, n_b_left, mode); - ir_node *zero = create_zero_const(irg, mode); + /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */ + if (is_cmp_unequal(a) && is_cmp_unequal(b) + && !mode_is_float(get_irn_mode(a_left)) + && !mode_is_float(get_irn_mode(b_left)) + && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + ir_graph *irg = get_irn_irg(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_node *zero = create_zero_const(irg, a_mode); return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); } } @@ -4697,6 +4928,8 @@ static ir_node *transform_node_Or(ir_node *n) return n; n = transform_bitwise_distributive(n, transform_node_Or); + if (is_Or(n)) + n = transform_node_bitop_shift(n); return n; } /* transform_node_Or */ @@ -4714,9 +4947,11 @@ static ir_node *transform_node_shift(ir_node *n) { ir_node *left, *right; ir_mode *mode; + ir_mode *count_mode; ir_tarval *tv1, *tv2, *res; ir_node *in[2], *irn, *block; ir_graph *irg; + int modulo_shf; left = get_binop_left(n); @@ -4725,7 +4960,7 @@ static ir_node *transform_node_shift(ir_node *n) return n; right = get_binop_right(n); - tv1 = value_of(right); + tv1 = value_of(right); if (tv1 == tarval_bad) return n; @@ -4733,37 +4968,55 @@ static ir_node *transform_node_shift(ir_node *n) if (tv2 == tarval_bad) return n; - res = tarval_add(tv1, tv2); - mode = get_irn_mode(n); - irg = get_irn_irg(n); + count_mode = get_tarval_mode(tv1); + if (get_tarval_mode(tv2) != count_mode) { + /* TODO: search bigger mode or something and convert... */ + return n; + } - /* beware: a simple replacement works only, if res < modulo shift */ - if (!is_Rotl(n)) { - int modulo_shf = get_mode_modulo_shift(mode); - if (modulo_shf > 0) { - ir_tarval *modulo = new_tarval_from_long(modulo_shf, - get_tarval_mode(res)); + mode = get_irn_mode(n); + modulo_shf = get_mode_modulo_shift(mode); - assert(modulo_shf >= (int) get_mode_size_bits(mode)); + if (modulo_shf > 0) { + ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode); - /* shifting too much */ - if (!(tarval_cmp(res, modulo) & ir_relation_less)) { - if (is_Shrs(n)) { - ir_node *block = get_nodes_block(n); - dbg_info *dbgi = get_irn_dbg_info(n); - ir_mode *smode = get_irn_mode(right); - ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1); - return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode); - } + /* I'm not so sure what happens in one complement... */ + assert(get_mode_arithmetic(count_mode) == irma_twos_complement); + /* modulo shifts should always be a power of 2 (otherwise modulo_mask + * above will be invalid) */ + assert(modulo_shf<=0 || is_po2(modulo_shf)); + + tv1 = tarval_and(tv1, modulo_mask); + tv2 = tarval_and(tv2, modulo_mask); + } + res = tarval_add(tv1, tv2); + irg = get_irn_irg(n); - return new_r_Const(irg, get_mode_null(mode)); + /* beware: a simple replacement works only, if res < modulo shift */ + if (is_Rotl(n)) { + int bits = get_mode_size_bits(mode); + ir_tarval *modulo = new_tarval_from_long(bits, count_mode); + res = tarval_mod(res, modulo); + } else { + long bits = get_mode_size_bits(mode); + ir_tarval *mode_size = new_tarval_from_long(bits, count_mode); + + /* shifting too much */ + if (!(tarval_cmp(res, mode_size) & ir_relation_less)) { + if (is_Shrs(n)) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *smode = get_irn_mode(right); + ir_node *cnst = new_r_Const_long(irg, smode, get_mode_size_bits(mode) - 1); + return new_rd_Shrs(dbgi, block, get_binop_left(left), cnst, mode); } + + return new_r_Const(irg, get_mode_null(mode)); } - } else { - res = tarval_mod(res, new_tarval_from_long(get_mode_size_bits(mode), get_tarval_mode(res))); } /* ok, we can replace it */ + assert(modulo_shf >= (int) get_mode_size_bits(mode)); block = get_nodes_block(n); in[0] = get_binop_left(left); @@ -4774,91 +5027,6 @@ static ir_node *transform_node_shift(ir_node *n) DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); return transform_node(irn); -} /* transform_node_shift */ - -/** - * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) - * (we can use: - * - and, or, xor instead of & - * - Shl, Shr, Shrs, rotl instead of >> - * (with a special case for Or/Xor + Shrs) - */ -static ir_node *transform_node_bitop_shift(ir_node *n) -{ - ir_node *left; - ir_node *right = get_binop_right(n); - ir_mode *mode = get_irn_mode(n); - ir_node *bitop_left; - ir_node *bitop_right; - ir_op *op_left; - ir_node *block; - dbg_info *dbgi; - ir_graph *irg; - ir_node *new_shift; - ir_node *new_bitop; - ir_node *new_const; - ir_tarval *tv1; - ir_tarval *tv2; - ir_tarval *tv_shift; - - assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); - - if (!is_Const(right)) - return n; - - left = get_binop_left(n); - op_left = get_irn_op(left); - if (op_left != op_And && op_left != op_Or && op_left != op_Eor) - return n; - - /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ - if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { - /* TODO: test if sign bit is affectes */ - return n; - } - - bitop_right = get_binop_right(left); - if (!is_Const(bitop_right)) - return n; - - bitop_left = get_binop_left(left); - - block = get_nodes_block(n); - dbgi = get_irn_dbg_info(n); - tv1 = get_Const_tarval(bitop_right); - tv2 = get_Const_tarval(right); - - assert(get_tarval_mode(tv1) == mode); - - if (is_Shl(n)) { - new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shl(tv1, tv2); - } else if (is_Shr(n)) { - new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shr(tv1, tv2); - } else if (is_Shrs(n)) { - new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shrs(tv1, tv2); - } else { - assert(is_Rotl(n)); - new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_rotl(tv1, tv2); - } - - assert(get_tarval_mode(tv_shift) == mode); - irg = get_irn_irg(n); - new_const = new_r_Const(irg, tv_shift); - - if (op_left == op_And) { - new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); - } else if (op_left == op_Or) { - new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); - } else { - assert(op_left == op_Eor); - new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); - } - - return new_bitop; } /** @@ -4979,7 +5147,7 @@ typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block, * then we can use that to minimize the value of Add(x, const) or * Sub(Const, x). In particular this often avoids 1 instruction in some * backends for the Shift(x, Sub(Const, y)) case because it can be replaced - * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed. + * by Shift(x, Minus(y)) which does not need an explicit Const constructed. */ static ir_node *transform_node_shift_modulo(ir_node *n, new_shift_func new_shift) @@ -5067,7 +5235,7 @@ static ir_node *transform_node_Shr(ir_node *n) if (is_Shr(n)) n = transform_node_shl_shr(n); if (is_Shr(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shr */ @@ -5082,13 +5250,23 @@ static ir_node *transform_node_Shrs(ir_node *n) ir_node *b = get_Shrs_right(n); ir_mode *mode = get_irn_mode(n); + if (is_oversize_shift(n)) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *cmode = get_irn_mode(b); + long val = get_mode_size_bits(cmode)-1; + ir_graph *irg = get_irn_irg(n); + ir_node *cnst = new_r_Const_long(irg, cmode, val); + return new_rd_Shrs(dbgi, block, a, cnst, mode); + } + HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode); n = transform_node_shift(n); if (is_Shrs(n)) n = transform_node_shift_modulo(n, new_rd_Shrs); if (is_Shrs(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shrs */ @@ -5111,7 +5289,7 @@ static ir_node *transform_node_Shl(ir_node *n) if (is_Shl(n)) n = transform_node_shl_shr(n); if (is_Shl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shl */ @@ -5130,7 +5308,7 @@ static ir_node *transform_node_Rotl(ir_node *n) n = transform_node_shift(n); if (is_Rotl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Rotl */ @@ -5206,7 +5384,7 @@ static ir_node *transform_node_End(ir_node *n) /* no need to keep Bad */ if (is_Bad(ka)) continue; - /* dont keep unreachable code */ + /* do not keep unreachable code */ block = is_Block(ka) ? ka : get_nodes_block(ka); if (is_block_unreachable(block)) continue; @@ -5236,16 +5414,115 @@ int ir_is_negated_value(const ir_node *a, const ir_node *b) return false; } +static const ir_node *skip_upconv(const ir_node *node) +{ + while (is_Conv(node)) { + ir_mode *mode = get_irn_mode(node); + const ir_node *op = get_Conv_op(node); + ir_mode *op_mode = get_irn_mode(op); + if (!smaller_mode(op_mode, mode)) + break; + node = op; + } + return node; +} + +int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true, + const ir_node *mux_false) +{ + ir_node *cmp_left; + ir_node *cmp_right; + ir_mode *mode; + ir_relation relation; + + if (!is_Cmp(sel)) + return 0; + + /** + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs + * transformations. + */ + mode = get_irn_mode(mux_true); + if (mode_honor_signed_zeros(mode)) + return 0; + + /* must be <, <=, >=, > */ + relation = get_Cmp_relation(sel); + if ((relation & ir_relation_less_greater) == 0) + return 0; + + if (!ir_is_negated_value(mux_true, mux_false)) + return 0; + + mux_true = skip_upconv(mux_true); + mux_false = skip_upconv(mux_false); + + /* must be x cmp 0 */ + cmp_right = get_Cmp_right(sel); + if (!is_Const(cmp_right) || !is_Const_null(cmp_right)) + return 0; + + cmp_left = get_Cmp_left(sel); + if (cmp_left == mux_false) { + if (relation & ir_relation_less) { + return 1; + } else { + assert(relation & ir_relation_greater); + return -1; + } + } else if (cmp_left == mux_true) { + if (relation & ir_relation_less) { + return -1; + } else { + assert(relation & ir_relation_greater); + return 1; + } + } + + return 0; +} + +ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true, + ir_node *mux_false) +{ + ir_node *cmp_left = get_Cmp_left(sel); + return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true; +} + /** * Optimize a Mux into some simpler cases. */ static ir_node *transform_node_Mux(ir_node *n) { - ir_node *oldn = n, *sel = get_Mux_sel(n); - ir_mode *mode = get_irn_mode(n); - ir_node *t = get_Mux_true(n); - ir_node *f = get_Mux_false(n); - ir_graph *irg = get_irn_irg(n); + ir_node *oldn = n; + ir_node *sel = get_Mux_sel(n); + ir_mode *mode = get_irn_mode(n); + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_graph *irg = get_irn_irg(n); + + /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */ + if (get_mode_arithmetic(mode) == irma_twos_complement) { + int abs = ir_mux_is_abs(sel, t, f); + if (abs != 0) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *op = ir_get_abs_op(sel, t, f); + int bits = get_mode_size_bits(mode); + ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1); + ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode); + ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode); + ir_node *res; + if (abs > 0) { + res = new_rd_Sub(dbgi, block, xorn, sext, mode); + } else { + res = new_rd_Sub(dbgi, block, sext, xorn, mode); + } + return res; + } + } if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) return n; @@ -5556,7 +5833,7 @@ static ir_node *transform_node_Load(ir_node *n) ir_node *bad = new_r_Bad(irg, mode_X); ir_mode *mode = get_Load_mode(n); ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res); - ir_node *in[pn_Load_max] = { mem, jmp, bad, res }; + ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad }; ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in); return tuple; } @@ -5576,7 +5853,7 @@ static ir_node *transform_node_Load(ir_node *n) ir_graph *irg = get_irn_irg(n); ir_node *bad = new_r_Bad(irg, mode_X); ir_node *res = value; - ir_node *in[pn_Load_max] = { mem, jmp, bad, res }; + ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad }; ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in); return tuple; } @@ -6039,10 +6316,23 @@ int identities_cmp(const void *elt, const void *key) /* for pinned nodes, the block inputs must be equal */ if (get_irn_n(a, -1) != get_irn_n(b, -1)) return 1; - } else if (! get_opt_global_cse()) { - /* for block-local CSE both nodes must be in the same Block */ - if (get_nodes_block(a) != get_nodes_block(b)) - return 1; + } else { + ir_node *block_a = get_nodes_block(a); + ir_node *block_b = get_nodes_block(b); + if (! get_opt_global_cse()) { + /* for block-local CSE both nodes must be in the same Block */ + if (block_a != block_b) + return 1; + } else { + /* The optimistic approach would be to do nothing here. + * However doing GCSE optimistically produces a lot of partially dead code which appears + * to be worse in practice than the missed opportunities. + * So we use a very conservative variant here and only CSE if 1 value dominates the + * other. */ + if (!block_dominates(block_a, block_b) + && !block_dominates(block_b, block_a)) + return 1; + } } /* compare a->in[0..ins] with b->in[0..ins] */