X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=ec61a829c5a3427baff50c800dc6904ddd7ad292;hb=8a5f15326aeefb771a9999c87d963972f47b19b9;hp=679410da3f0fa60ac5a9fded3452f880dc68e96c;hpb=64a05bd2846f1dbe27318977303d411fc4149484;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index 679410da3..ec61a829c 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -21,7 +21,6 @@ * @file * @brief iropt --- optimizations intertwined with IR construction. * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck - * @version $Id$ */ #include "config.h" @@ -44,7 +43,6 @@ #include "irhooks.h" #include "irarch.h" #include "hashptr.h" -#include "opt_polymorphy.h" #include "irtools.h" #include "irhooks.h" #include "array_t.h" @@ -56,6 +54,22 @@ /* Make types visible to allow most efficient access */ #include "entity_t.h" +static bool is_Or_Eor_Add(const ir_node *node) +{ + if (is_Or(node) || is_Eor(node) || is_Add(node)) { + ir_node *left = get_binop_left(node); + ir_node *right = get_binop_right(node); + vrp_attr *vrp_left = vrp_get_info(left); + vrp_attr *vrp_right = vrp_get_info(right); + if (vrp_left != NULL && vrp_right != NULL) { + ir_tarval *vrp_val + = tarval_and(vrp_left->bits_not_set, vrp_right->bits_not_set); + return tarval_is_null(vrp_val); + } + } + return false; +} + /** * Returns the tarval of a Const node or tarval_bad for all other nodes. */ @@ -84,7 +98,7 @@ void set_value_of_func(value_of_func func) static ir_tarval *computed_value_Const(const ir_node *n) { return get_Const_tarval(n); -} /* computed_value_Const */ +} /** * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst. @@ -115,7 +129,7 @@ static ir_tarval *computed_value_SymConst(const ir_node *n) break; } return tarval_bad; -} /* computed_value_SymConst */ +} /** * Return the value of an Add. @@ -131,8 +145,14 @@ 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 */ +} /** * Return the value of a Sub. @@ -160,7 +180,7 @@ static ir_tarval *computed_value_Sub(const ir_node *n) return tarval_sub(ta, tb, mode); return tarval_bad; -} /* computed_value_Sub */ +} /** * Return the value of a Carry. @@ -182,7 +202,7 @@ static ir_tarval *computed_value_Carry(const ir_node *n) return get_mode_null(m); } return tarval_bad; -} /* computed_value_Carry */ +} /** * Return the value of a Borrow. @@ -202,7 +222,7 @@ static ir_tarval *computed_value_Borrow(const ir_node *n) return get_mode_null(m); } return tarval_bad; -} /* computed_value_Borrow */ +} /** * Return the value of an unary Minus. @@ -216,7 +236,7 @@ static ir_tarval *computed_value_Minus(const ir_node *n) return tarval_neg(ta); return tarval_bad; -} /* computed_value_Minus */ +} /** * Return the value of a Mul. @@ -249,7 +269,7 @@ static ir_tarval *computed_value_Mul(const ir_node *n) } } return tarval_bad; -} /* computed_value_Mul */ +} /** * Return the value of an And. @@ -264,12 +284,19 @@ 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 */ +} /** * Return the value of an Or. @@ -284,12 +311,18 @@ 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 */ +} /** * Return the value of an Eor. @@ -303,6 +336,11 @@ 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); @@ -311,7 +349,7 @@ static ir_tarval *computed_value_Eor(const ir_node *n) return tarval_eor(ta, tb); } return tarval_bad; -} /* computed_value_Eor */ +} /** * Return the value of a Not. @@ -325,7 +363,29 @@ static ir_tarval *computed_value_Not(const ir_node *n) return tarval_not(ta); 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. @@ -341,8 +401,12 @@ 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 */ +} /** * Return the value of a Shr. @@ -358,8 +422,11 @@ 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 */ +} /** * Return the value of a Shrs. @@ -376,7 +443,7 @@ static ir_tarval *computed_value_Shrs(const ir_node *n) return tarval_shrs(ta, tb); } return tarval_bad; -} /* computed_value_Shrs */ +} /** * Return the value of a Rotl. @@ -393,21 +460,56 @@ static ir_tarval *computed_value_Rotl(const ir_node *n) return tarval_rotl(ta, tb); } 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 */ +} /** * Calculate the value of a Mux: can be evaluated, if the @@ -427,7 +529,7 @@ static ir_tarval *computed_value_Mux(const ir_node *n) return value_of(v); } return tarval_bad; -} /* computed_value_Mux */ +} /** * Calculate the value of a Confirm: can be evaluated, @@ -441,7 +543,7 @@ static ir_tarval *computed_value_Confirm(const ir_node *n) return tv; } return value_of(get_Confirm_value(n)); -} /* computed_value_Confirm */ +} /** * gives a (conservative) estimation of possible relation when comparing @@ -484,17 +586,20 @@ ir_relation ir_get_possible_cmp_relations(const ir_node *left, /* Alloc nodes never return null (but throw an exception) */ if (is_Alloc(left) && tarval_is_null(tv_r)) possible &= ~ir_relation_equal; + /* stuff known through confirm nodes */ + if (is_Confirm(left) && get_Confirm_bound(left) == right) { + possible &= get_Confirm_relation(left); + } + if (is_Confirm(right) && get_Confirm_bound(right) == left) { + ir_relation relation = get_Confirm_relation(right); + relation = get_inversed_relation(relation); + possible &= relation; + } return possible; } -/** - * Return the value of a Cmp. - * - * The basic idea here is to determine which relations are possible and which - * one are definitely impossible. - */ -static ir_tarval *computed_value_Cmp(const ir_node *cmp) +static ir_tarval *compute_cmp(const ir_node *cmp) { ir_node *left = get_Cmp_left(cmp); ir_node *right = get_Cmp_right(cmp); @@ -511,6 +616,21 @@ static ir_tarval *computed_value_Cmp(const ir_node *cmp) return computed_value_Cmp_Confirm(cmp, left, right, relation); } +/** + * Return the value of a Cmp. + * + * The basic idea here is to determine which relations are possible and which + * one are definitely impossible. + */ +static ir_tarval *computed_value_Cmp(const ir_node *cmp) +{ + /* we can't construct Constb after lowering mode_b nodes */ + if (is_irg_state(get_irn_irg(cmp), IR_GRAPH_STATE_MODEB_LOWERED)) + return tarval_bad; + + return compute_cmp(cmp); +} + /** * Calculate the value of an integer Div. * Special case: 0 / b @@ -533,7 +653,7 @@ static ir_tarval *do_computed_value_Div(const ir_node *div) if (ta != tarval_bad && tb != tarval_bad) return tarval_div(ta, tb); return tarval_bad; -} /* do_computed_value_Div */ +} /** * Calculate the value of an integer Mod of two nodes. @@ -550,7 +670,7 @@ static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) if (ta != tarval_bad && tb != tarval_bad) return tarval_mod(ta, tb); return tarval_bad; -} /* do_computed_value_Mod */ +} /** * Return the value of a Proj(Div). @@ -562,7 +682,7 @@ static ir_tarval *computed_value_Proj_Div(const ir_node *n) return tarval_bad; return do_computed_value_Div(get_Proj_pred(n)); -} /* computed_value_Proj_Div */ +} /** * Return the value of a Proj(Mod). @@ -576,7 +696,7 @@ static ir_tarval *computed_value_Proj_Mod(const ir_node *n) return do_computed_value_Mod(get_Mod_left(mod), get_Mod_right(mod)); } return tarval_bad; -} /* computed_value_Proj_Mod */ +} /** * Return the value of a Proj. @@ -588,7 +708,7 @@ static ir_tarval *computed_value_Proj(const ir_node *proj) if (n->op->ops.computed_value_Proj != NULL) return n->op->ops.computed_value_Proj(proj); return tarval_bad; -} /* computed_value_Proj */ +} /** * If the parameter n can be computed, return its value, else tarval_bad. @@ -599,24 +719,15 @@ static ir_tarval *computed_value_Proj(const ir_node *proj) ir_tarval *computed_value(const ir_node *n) { vrp_attr *vrp = vrp_get_info(n); - if (vrp && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == ir_relation_equal) { + if (vrp != NULL && vrp->bits_set == vrp->bits_not_set) return vrp->bits_set; - } + if (n->op->ops.computed_value) return n->op->ops.computed_value(n); return tarval_bad; -} /* computed_value */ +} -/** - * Set the default computed_value evaluator in an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops) +void firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -655,77 +766,9 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops /* leave NULL */ break; } - - return ops; #undef CASE_PROJ #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, @@ -764,7 +807,7 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) } return n; -} /* equivalent_node_neutral_zero */ +} /** * Eor is commutative and has neutral 0. @@ -781,9 +824,9 @@ static ir_node *equivalent_node_Eor(ir_node *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 (is_Eor(a) || is_Or_Eor_Add(a)) { + ir_node *aa = get_binop_left(a); + ir_node *ab = get_binop_right(a); if (aa == b) { /* (a ^ b) ^ a -> b */ @@ -797,9 +840,9 @@ static ir_node *equivalent_node_Eor(ir_node *n) return n; } } - if (is_Eor(b)) { - ir_node *ba = get_Eor_left(b); - ir_node *bb = get_Eor_right(b); + if (is_Eor(b) || is_Or_Eor_Add(b)) { + ir_node *ba = get_binop_left(b); + ir_node *bb = get_binop_right(b); if (ba == a) { /* a ^ (a ^ b) -> b */ @@ -868,7 +911,7 @@ static ir_node *equivalent_node_Add(ir_node *n) } } return n; -} /* equivalent_node_Add */ +} /** * optimize operations that are not commutative but have neutral 0 on left, @@ -888,7 +931,7 @@ static ir_node *equivalent_node_left_zero(ir_node *n) DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0); } return n; -} /* equivalent_node_left_zero */ +} #define equivalent_node_Shl equivalent_node_left_zero #define equivalent_node_Shr equivalent_node_left_zero @@ -931,11 +974,11 @@ static ir_node *equivalent_node_Sub(ir_node *n) } } return n; -} /* equivalent_node_Sub */ +} /** - * 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. @@ -953,7 +996,7 @@ static ir_node *equivalent_node_idempotent_unop(ir_node *n) DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY); } return n; -} /* equivalent_node_idempotent_unop */ +} /** Optimize Not(Not(x)) == x. */ #define equivalent_node_Not equivalent_node_idempotent_unop @@ -992,7 +1035,7 @@ static ir_node *equivalent_node_Mul(ir_node *n) } } return n; -} /* equivalent_node_Mul */ +} /** * Use algebraic simplification a | a = a | 0 = 0 | a = a. @@ -1025,7 +1068,7 @@ static ir_node *equivalent_node_Or(ir_node *n) } return n; -} /* equivalent_node_Or */ +} /** * Optimize a & 0b1...1 = 0b1...1 & a = a & a = (a|X) & a = a. @@ -1056,8 +1099,13 @@ static ir_node *equivalent_node_And(ir_node *n) ir_node *convop = get_Conv_op(a); ir_mode *convopmode = get_irn_mode(convop); if (!mode_is_signed(convopmode)) { - if (tarval_is_all_one(tarval_convert_to(tv, convopmode))) { - /* Conv(X) & all_one(mode(X)) = Conv(X) */ + /* Check Conv(all_one) & Const = all_one */ + ir_tarval *one = get_mode_all_one(convopmode); + ir_tarval *conv = tarval_convert_to(one, mode); + ir_tarval *and = tarval_and(conv, tv); + + if (tarval_is_all_one(and)) { + /* Conv(X) & Const = X */ n = a; DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; @@ -1071,24 +1119,22 @@ static ir_node *equivalent_node_And(ir_node *n) DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); return n; } - if (is_Or(a)) { - if (b == get_Or_left(a) || b == get_Or_right(a)) { - /* (a|X) & a */ - n = b; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); - return n; - } + /* (a|X) & a => a*/ + if ((is_Or(a) || is_Or_Eor_Add(a)) + && (b == get_binop_left(a) || b == get_binop_right(a))) { + n = b; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; } - if (is_Or(b)) { - if (a == get_Or_left(b) || a == get_Or_right(b)) { - /* a & (a|X) */ - n = a; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); - return n; - } + /* a & (a|X) => a*/ + if ((is_Or(b) || is_Or_Eor_Add(b)) + && (a == get_binop_left(b) || a == get_binop_right(b))) { + n = a; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND); + return n; } return n; -} /* equivalent_node_And */ +} /** * Try to remove useless Conv's: @@ -1199,7 +1245,7 @@ restart: if (mode_is_int(n_mode) && get_mode_arithmetic(a_mode) == irma_ieee754) { /* ConvI(ConvF(I)) -> I, iff float mantissa >= int mode */ unsigned int_mantissa = get_mode_size_bits(n_mode) - (mode_is_signed(n_mode) ? 1 : 0); - unsigned float_mantissa = tarval_ieee754_get_mantissa_size(a_mode); + unsigned float_mantissa = get_mode_mantissa_size(a_mode); if (float_mantissa >= int_mantissa) { n = b; @@ -1219,7 +1265,7 @@ restart: } } return n; -} /* equivalent_node_Conv */ +} /** * - fold Phi-nodes, iff they have only one predecessor except @@ -1230,7 +1276,6 @@ 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 +1284,6 @@ 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; @@ -1270,7 +1313,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) DBG_OPT_PHI(oldn, n); } return n; -} /* equivalent_node_Phi */ +} /** * Optimize Proj(Tuple). @@ -1285,7 +1328,7 @@ static ir_node *equivalent_node_Proj_Tuple(ir_node *proj) DBG_OPT_TUPLE(oldn, tuple, proj); return proj; -} /* equivalent_node_Proj_Tuple */ +} /** * Optimize a / 1 = a. @@ -1317,7 +1360,7 @@ static ir_node *equivalent_node_Proj_Div(ir_node *proj) } } return proj; -} /* equivalent_node_Proj_Div */ +} /** * Optimize CopyB(mem, x, x) into a Nop. @@ -1339,7 +1382,7 @@ static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) } } return proj; -} /* equivalent_node_Proj_CopyB */ +} /** * Optimize Bounds(idx, idx, upper) into idx. @@ -1389,7 +1432,7 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) } } return proj; -} /* equivalent_node_Proj_Bound */ +} /** * Does all optimizations on nodes that must be done on its Projs @@ -1401,7 +1444,7 @@ static ir_node *equivalent_node_Proj(ir_node *proj) if (n->op->ops.equivalent_node_Proj) return n->op->ops.equivalent_node_Proj(proj); return proj; -} /* equivalent_node_Proj */ +} /** * Remove Id's. @@ -1416,7 +1459,7 @@ static ir_node *equivalent_node_Id(ir_node *n) DBG_OPT_ID(oldn, n); return n; -} /* equivalent_node_Id */ +} /** * Optimize a Mux. @@ -1427,6 +1470,12 @@ static ir_node *equivalent_node_Mux(ir_node *n) ir_node *n_t, *n_f; ir_tarval *ts = value_of(sel); + if (ts == tarval_bad && is_Cmp(sel)) { + /* try again with a direct call to compute_cmp, as we don't care + * about the MODEB_LOWERED flag here */ + ts = compute_cmp(sel); + } + /* Mux(true, f, t) == t */ if (ts == tarval_b_true) { n = get_Mux_true(n); @@ -1553,18 +1602,9 @@ ir_node *equivalent_node(ir_node *n) if (n->op->ops.equivalent_node) return n->op->ops.equivalent_node(n); return n; -} /* equivalent_node */ +} -/** - * Sets the default equivalent node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops) +void firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -1576,7 +1616,6 @@ 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); @@ -1603,11 +1642,9 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op /* leave NULL */ break; } - - return ops; #undef CASE #undef CASE_PROJ -} /* firm_set_default_equivalent_node */ +} /** * Returns non-zero if a node is a Phi node @@ -1624,7 +1661,7 @@ static int is_const_Phi(ir_node *n) return 0; } return 1; -} /* is_const_Phi */ +} typedef ir_tarval *(*tarval_sub_type)(ir_tarval *a, ir_tarval *b, ir_mode *mode); typedef ir_tarval *(*tarval_binop_type)(ir_tarval *a, ir_tarval *b); @@ -1702,7 +1739,7 @@ static ir_node *apply_binop_on_phi(ir_node *phi, ir_tarval *other, eval_func eva res[i] = new_r_Const(irg, (ir_tarval*)res[i]); } return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_binop_on_phi */ +} /** * Apply an evaluator on a binop with two constant Phi. @@ -1747,7 +1784,7 @@ static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, eval_func eval, ir res[i] = new_r_Const(irg, (ir_tarval*)res[i]); } return new_r_Phi(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). @@ -1785,7 +1822,7 @@ static ir_node *apply_unop_on_phi(ir_node *phi, ir_tarval *(*eval)(ir_tarval *)) res[i] = new_r_Const(irg, (ir_tarval*)res[i]); } return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_unop_on_phi */ +} /** * Apply a conversion on a constant operator (a Phi). @@ -1820,7 +1857,7 @@ static ir_node *apply_conv_on_phi(ir_node *phi, ir_mode *mode) res[i] = new_r_Const(irg, (ir_tarval*)res[i]); } return new_r_Phi(get_nodes_block(phi), n, (ir_node **)res, mode); -} /* apply_conv_on_phi */ +} /** * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and @@ -1897,7 +1934,7 @@ static ir_node *transform_node_AddSub(ir_node *n) } return n; -} /* transform_node_AddSub */ +} #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \ do { \ @@ -1934,59 +1971,798 @@ static ir_node *transform_node_AddSub(ir_node *n) } while(0) /** - * Do the AddSub optimization, then Transform - * Constant folding on Phi - * Add(a,a) -> Mul(a, 2) - * Add(Mul(a, x), a) -> Mul(a, x+1) - * if the mode is integer or float. - * Transform Add(a,-b) into Sub(a,b). - * Reassociation might fold this further. + * Create a 0 constant of given mode. */ -static ir_node *transform_node_Add(ir_node *n) +static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode) { - ir_mode *mode; - ir_node *a, *b, *c, *oldn = n; - vrp_attr *a_vrp, *b_vrp; + ir_tarval *tv = get_mode_null(mode); + ir_node *cnst = new_r_Const(irg, tv); - n = transform_node_AddSub(n); + return cnst; +} - a = get_Add_left(n); - b = get_Add_right(n); +static bool is_shiftop(const ir_node *n) +{ + return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n); +} - mode = get_irn_mode(n); +/* the order of the values is important! */ +typedef enum const_class { + const_const = 0, + const_like = 1, + const_other = 2 +} const_class; - if (mode_is_reference(mode)) { - ir_mode *lmode = get_irn_mode(a); +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; +} - 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, get_nodes_block(n), a, mode); - } - } +/** + * Determines whether r is more constlike or has a larger index (in that order) + * than l. + */ +static bool 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)); +} - HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode); +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); - /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ - if (mode_is_float(mode)) { - ir_graph *irg = get_irn_irg(n); - if (get_irg_fp_model(irg) & fp_strict_algebraic) - return n; - } + if (relation == ir_relation_less_greater) + return true; - if (mode_is_num(mode)) { - ir_graph *irg = get_irn_irg(n); - /* the following code leads to endless recursion when Mul are replaced by a simple instruction chain */ - if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP) - && a == b && mode_is_int(mode)) { - ir_node *block = get_nodes_block(n); + if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right)) + return relation == ir_relation_greater; + return false; +} - n = new_rd_Mul( - get_irn_dbg_info(n), - block, - a, - new_r_Const_long(irg, mode, 2), - mode); +/** + * returns true for Cmp(x == 0) or Cmp(x != 0) + */ +static bool is_cmp_equality_zero(const ir_node *node) +{ + ir_relation relation; + ir_node *right = get_Cmp_right(node); + + if (!is_Const(right) || !is_Const_null(right)) + return false; + relation = get_Cmp_relation(node); + return relation == ir_relation_equal + || relation == ir_relation_less_greater + || (!mode_is_signed(get_irn_mode(right)) + && relation == ir_relation_greater); +} + +/** + * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. + * Such pattern may arise in bitfield stores. + * + * value c4 value c4 & c2 + * AND c3 AND c1 | c3 + * OR c2 ===> OR + * AND c1 + * OR + * + * + * value c2 value c1 + * AND c1 ===> OR if (c1 | c2) == 0x111..11 + * OR + */ +static ir_node *transform_node_Or_bf_store(ir_node *irn_or) +{ + ir_node *irn_and, *c1; + ir_node *or_l, *c2; + ir_node *and_l, *c3; + ir_node *value, *c4; + ir_node *new_and, *new_const, *block; + ir_mode *mode = get_irn_mode(irn_or); + + ir_tarval *tv1, *tv2, *tv3, *tv4, *tv; + + for (;;) { + ir_graph *irg; + irn_and = get_binop_left(irn_or); + c1 = get_binop_right(irn_or); + if (!is_Const(c1) || !is_And(irn_and)) + return irn_or; + + or_l = get_binop_left(irn_and); + c2 = get_binop_right(irn_and); + if (!is_Const(c2)) + return irn_or; + + tv1 = get_Const_tarval(c1); + tv2 = get_Const_tarval(c2); + + tv = tarval_or(tv1, tv2); + if (tarval_is_all_one(tv)) { + /* the AND does NOT clear a bit with isn't set by the OR */ + set_binop_left(irn_or, or_l); + set_binop_right(irn_or, c1); + + /* check for more */ + continue; + } + + if (!is_Or(or_l) && !is_Or_Eor_Add(or_l)) + return irn_or; + + and_l = get_binop_left(or_l); + c3 = get_binop_right(or_l); + if (!is_Const(c3) || !is_And(and_l)) + return irn_or; + + value = get_binop_left(and_l); + c4 = get_binop_right(and_l); + if (!is_Const(c4)) + return irn_or; + + /* ok, found the pattern, check for conditions */ + assert(mode == get_irn_mode(irn_and)); + assert(mode == get_irn_mode(or_l)); + assert(mode == get_irn_mode(and_l)); + + tv3 = get_Const_tarval(c3); + tv4 = get_Const_tarval(c4); + + tv = tarval_or(tv4, tv2); + if (!tarval_is_all_one(tv)) { + /* have at least one 0 at the same bit position */ + return irn_or; + } + + if (tv3 != tarval_andnot(tv3, tv4)) { + /* bit in the or_mask is outside the and_mask */ + return irn_or; + } + + if (tv1 != tarval_andnot(tv1, tv2)) { + /* bit in the or_mask is outside the and_mask */ + return irn_or; + } + + /* ok, all conditions met */ + block = get_irn_n(irn_or, -1); + irg = get_irn_irg(block); + + new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode); + + new_const = new_r_Const(irg, tarval_or(tv3, tv1)); + + set_binop_left(irn_or, new_and); + set_binop_right(irn_or, new_const); + + /* check for more */ + } +} + +/** + * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl + */ +static ir_node *transform_node_Or_Rotl(ir_node *irn_or) +{ + ir_mode *mode = get_irn_mode(irn_or); + ir_node *shl, *shr, *block; + ir_node *irn, *x, *c1, *c2, *n; + ir_tarval *tv1, *tv2; + + /* some backends can't handle rotl */ + if (!be_get_backend_param()->support_rotl) + return irn_or; + + if (! mode_is_int(mode)) + return irn_or; + + shl = get_binop_left(irn_or); + shr = get_binop_right(irn_or); + + if (is_Shr(shl)) { + if (!is_Shl(shr)) + return irn_or; + + irn = shl; + shl = shr; + shr = irn; + } else if (!is_Shl(shl)) { + return irn_or; + } else if (!is_Shr(shr)) { + return irn_or; + } + x = get_Shl_left(shl); + if (x != get_Shr_left(shr)) + return irn_or; + + c1 = get_Shl_right(shl); + c2 = get_Shr_right(shr); + if (is_Const(c1) && is_Const(c2)) { + tv1 = get_Const_tarval(c1); + if (! tarval_is_long(tv1)) + return irn_or; + + tv2 = get_Const_tarval(c2); + if (! tarval_is_long(tv2)) + return irn_or; + + if (get_tarval_long(tv1) + get_tarval_long(tv2) + != (int) get_mode_size_bits(mode)) + return irn_or; + + /* yet, condition met */ + block = get_nodes_block(irn_or); + + n = new_r_Rotl(block, x, c1, mode); + + DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); + return n; + } + + /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets + * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */ + if (!ir_is_negated_value(c1, c2)) { + return irn_or; + } + + /* yet, condition met */ + block = get_nodes_block(irn_or); + n = new_r_Rotl(block, x, c1, mode); + DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL); + return 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 + * and(a OP c, b OP c) -> and(a, b) OP c + * note, might return a different op than n + */ +static ir_node *transform_bitwise_distributive(ir_node *n, + recursive_transform trans_func) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_op *op = get_irn_op(a); + ir_op *op_root = get_irn_op(n); + + if (op != get_irn_op(b)) + return n; + + /* and(conv(a), conv(b)) -> conv(and(a,b)) */ + if (op == op_Conv) { + ir_node *a_op = get_Conv_op(a); + ir_node *b_op = get_Conv_op(b); + ir_mode *a_mode = get_irn_mode(a_op); + ir_mode *b_mode = get_irn_mode(b_op); + if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) { + ir_node *blk = get_nodes_block(n); + + n = exact_copy(n); + set_binop_left(n, a_op); + set_binop_right(n, b_op); + set_irn_mode(n, a_mode); + n = trans_func(n); + n = new_r_Conv(blk, n, get_irn_mode(oldn)); + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); + return n; + } + } + + if (op == op_Eor) { + /* nothing to gain here */ + return n; + } + + if (op == op_Shrs || op == op_Shr || op == op_Shl + || op == op_And || op == op_Or || op == op_Eor) { + ir_node *a_left = get_binop_left(a); + ir_node *a_right = get_binop_right(a); + ir_node *b_left = get_binop_left(b); + ir_node *b_right = get_binop_right(b); + ir_node *c = NULL; + ir_node *op1 = NULL; + ir_node *op2 = NULL; + + if (is_op_commutative(op)) { + if (a_left == b_left) { + c = a_left; + op1 = a_right; + op2 = b_right; + } else if (a_left == b_right) { + c = a_left; + op1 = a_right; + op2 = b_left; + } else if (a_right == b_left) { + c = a_right; + op1 = a_left; + op2 = b_right; + } + } + if (a_right == b_right) { + c = a_right; + op1 = a_left; + op2 = b_left; + } + + if (c != NULL) { + /* (a sop c) & (b sop c) => (a & b) sop c */ + ir_node *blk = get_nodes_block(n); + + ir_node *new_n = exact_copy(n); + set_binop_left(new_n, op1); + set_binop_right(new_n, op2); + new_n = trans_func(new_n); + + if (op_root == op_Eor && op == op_Or) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *mode = get_irn_mode(c); + + c = new_rd_Not(dbgi, blk, c, mode); + n = new_rd_And(dbgi, blk, new_n, c, mode); + } else { + n = exact_copy(a); + set_nodes_block(n, blk); + set_binop_left(n, new_n); + set_binop_right(n, c); + add_identities(n); + } + + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); + return n; + } + } + + return n; +} + +/** + * 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) || is_Or_Eor_Add(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); + + /* Check whether we have lost some bits during the right shift. */ + if (!is_And(n)) { + ir_tarval *tv_back_again = tarval_shl(tv_bitop, tv1); + + if (tarval_cmp(tv_back_again, tv2) != ir_relation_equal) + return n; + } + } else if (is_Shr(left)) { + if (!is_And(n)) { + /* + * TODO this can be improved by checking whether + * the left shift produces an overflow + */ + return n; + } + 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) || is_Or_Eor_Add(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 { + assert(is_Rotl(left)); + new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode); + } + + return new_shift; +} + +static bool complement_values(const ir_node *a, const ir_node *b) +{ + if (is_Not(a) && get_Not_op(a) == b) + return true; + if (is_Not(b) && get_Not_op(b) == a) + return true; + if (is_Const(a) && is_Const(b)) { + ir_tarval *tv_a = get_Const_tarval(a); + ir_tarval *tv_b = get_Const_tarval(b); + return tarval_not(tv_a) == tv_b; + } + return false; +} + +typedef ir_tarval *(tv_fold_binop_func)(ir_tarval *a, ir_tarval *b); + +/** + * for associative operations fold: + * op(op(x, c0), c1) to op(x, op(c0, c1)) with constants folded. + * This is a "light" version of the reassociation phase + */ +static ir_node *fold_constant_associativity(ir_node *node, + tv_fold_binop_func fold) +{ + ir_graph *irg; + ir_op *op; + ir_node *left; + ir_node *right = get_binop_right(node); + ir_node *left_right; + ir_node *left_left; + ir_tarval *c0; + ir_tarval *c1; + ir_tarval *new_c; + ir_node *new_const; + ir_node *new_node; + if (!is_Const(right)) + return node; + + op = get_irn_op(node); + left = get_binop_left(node); + if (get_irn_op(left) != op) + return node; + + left_right = get_binop_right(left); + if (!is_Const(left_right)) + return node; + + left_left = get_binop_left(left); + c0 = get_Const_tarval(left_right); + c1 = get_Const_tarval(right); + irg = get_irn_irg(node); + if (get_tarval_mode(c0) != get_tarval_mode(c1)) + return node; + new_c = fold(c0, c1); + if (new_c == tarval_bad) + return node; + new_const = new_r_Const(irg, new_c); + new_node = exact_copy(node); + set_binop_left(new_node, left_left); + set_binop_right(new_node, new_const); + return new_node; +} + +/** + * Transform an Or. + */ +static ir_node *transform_node_Or_(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_node *c; + ir_mode *mode; + + n = fold_constant_associativity(n, tarval_or); + if (n != oldn) + return n; + + if (is_Not(a) && is_Not(b)) { + /* ~a | ~b = ~(a&b) */ + ir_node *block = get_nodes_block(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), block, a, b, mode); + n = new_rd_Not(get_irn_dbg_info(n), block, n, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); + return n; + } + + /* we can combine the relations of two compares with the same operands */ + if (is_Cmp(a) && is_Cmp(b)) { + ir_node *a_left = get_Cmp_left(a); + ir_node *a_right = get_Cmp_right(a); + ir_node *b_left = get_Cmp_left(b); + ir_node *b_right = get_Cmp_right(b); + if (a_left == b_left && b_left == b_right) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + ir_relation new_relation = a_relation | b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + /* 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))) { + if (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, xora, b_mode); + ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_node *zero = create_zero_const(irg, b_mode); + return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); + } + if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_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); + } + } + } + + mode = get_irn_mode(n); + HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode); + + n = transform_node_Or_bf_store(n); + if (n != oldn) + return n; + n = transform_node_Or_Rotl(n); + if (n != oldn) + return n; + + n = transform_bitwise_distributive(n, transform_node_Or_); + if (n != oldn) + return n; + n = transform_node_bitop_shift(n); + if (n != oldn) + return n; + + return n; +} + +static ir_node *transform_node_Or(ir_node *n) +{ + if (is_Or_Eor_Add(n)) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *left = get_Or_left(n); + ir_node *right = get_Or_right(n); + ir_mode *mode = get_irn_mode(n); + return new_rd_Add(dbgi, block, left, right, mode); + } + return transform_node_Or_(n); +} + +/** + * Transform an Eor. + */ +static ir_node *transform_node_Eor_(ir_node *n) +{ + ir_node *oldn = n; + ir_node *a = get_binop_left(n); + ir_node *b = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *c; + + n = fold_constant_associativity(n, tarval_eor); + if (n != oldn) + return n; + + /* we can combine the relations of two compares with the same operands */ + if (is_Cmp(a) && is_Cmp(b)) { + ir_node *a_left = get_Cmp_left(a); + ir_node *a_right = get_Cmp_left(a); + ir_node *b_left = get_Cmp_left(b); + ir_node *b_right = get_Cmp_right(b); + if (a_left == b_left && b_left == b_right) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + ir_relation new_relation = a_relation ^ b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + } + + HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode); + + /* normalize not nodes... ~a ^ b <=> a ^ ~b */ + if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *new_not = new_rd_Not(dbg, block, b, mode); + ir_node *new_left = get_Not_op(a); + n = new_rd_Eor(dbg, block, new_left, new_not, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *new_not = new_rd_Not(dbg, block, a, mode); + ir_node *new_right = get_Not_op(b); + n = new_rd_Eor(dbg, block, new_not, new_right, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + + /* x ^ 1...1 -> ~1 */ + if (is_Const(b) && is_Const_all_one(b)) { + n = new_r_Not(get_nodes_block(n), a, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); + return n; + } + + n = transform_bitwise_distributive(n, transform_node_Eor_); + if (n != oldn) + return n; + n = transform_node_bitop_shift(n); + if (n != oldn) + return n; + + return n; +} + +static ir_node *transform_node_Eor(ir_node *n) +{ + if (is_Or_Eor_Add(n)) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *left = get_Eor_left(n); + ir_node *right = get_Eor_right(n); + ir_mode *mode = get_irn_mode(n); + return new_rd_Add(dbgi, block, left, right, mode); + } + return transform_node_Eor_(n); +} + +/** + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Add(a,a) -> Mul(a, 2) + * Add(Mul(a, x), a) -> Mul(a, x+1) + * if the mode is integer or float. + * Transform Add(a,-b) into Sub(a,b). + * Reassociation might fold this further. + */ +static ir_node *transform_node_Add(ir_node *n) +{ + ir_mode *mode; + ir_node *a; + ir_node *b; + ir_node *c; + ir_node *oldn = n; + + n = fold_constant_associativity(n, tarval_add); + if (n != oldn) + return n; + + n = transform_node_AddSub(n); + if (n != oldn) + return n; + + a = get_Add_left(n); + b = get_Add_right(n); + 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, get_nodes_block(n), a, mode); + } + } + + if (is_Const(b) && get_mode_arithmetic(mode) == irma_twos_complement) { + ir_tarval *tv = get_Const_tarval(b); + ir_tarval *min = get_mode_min(mode); + /* if all bits are set, then this has the same effect as a Not. + * Note that the following == gives false for different modes which + * is exactly what we want */ + if (tv == min) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + ir_node *block = get_nodes_block(n); + ir_node *cnst = new_r_Const(irg, min); + return new_rd_Eor(dbgi, block, a, cnst, mode); + } + } + + HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode); + + /* for FP the following optimizations are only allowed if + * fp_strict_algebraic is disabled */ + if (mode_is_float(mode)) { + ir_graph *irg = get_irn_irg(n); + if (get_irg_fp_model(irg) & fp_strict_algebraic) + return n; + } + + if (mode_is_num(mode)) { + ir_graph *irg = get_irn_irg(n); + /* the following code leads to endless recursion when Mul are replaced + * by a simple instruction chain */ + if (!is_irg_state(irg, IR_GRAPH_STATE_ARCH_DEP) + && a == b && mode_is_int(mode)) { + ir_node *block = get_nodes_block(n); + + n = new_rd_Mul( + get_irn_dbg_info(n), + block, + a, + new_r_Const_long(irg, mode, 2), + mode); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A); return n; } @@ -2022,39 +2798,21 @@ 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; - } } } } - a_vrp = vrp_get_info(a); - b_vrp = vrp_get_info(b); - - if (a_vrp && b_vrp) { - ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set); - - if (tarval_is_null(c)) { - dbg_info *dbgi = get_irn_dbg_info(n); - return new_rd_Or(dbgi, get_nodes_block(n), a, b, mode); - } + if (is_Or_Eor_Add(n)) { + n = transform_node_Or_(n); + if (n != oldn) + return n; + n = transform_node_Eor_(n); + if (n != oldn) + return n; } + return n; -} /* transform_node_Add */ +} /** * returns -cnst or NULL if impossible @@ -2179,6 +2937,7 @@ 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 +2956,7 @@ restart: return n; } } +#endif } /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ @@ -2209,58 +2969,54 @@ restart: DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A); return n; } - if (is_Add(a)) { - if (mode_wrap_around(mode)) { - ir_node *left = get_Add_left(a); - ir_node *right = get_Add_right(a); + if ((is_Add(a) || is_Or_Eor_Add(a)) && mode_wrap_around(mode)) { + ir_node *left = get_binop_left(a); + ir_node *right = get_binop_right(a); - /* FIXME: Does the Conv's work only for two complement or generally? */ - if (left == b) { - if (mode != get_irn_mode(right)) { - /* This Sub is an effective Cast */ - right = new_r_Conv(get_nodes_block(n), right, mode); - } - n = right; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; - } else if (right == b) { - if (mode != get_irn_mode(left)) { - /* This Sub is an effective Cast */ - left = new_r_Conv(get_nodes_block(n), left, mode); - } - n = left; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == b) { + if (mode != get_irn_mode(right)) { + /* This Sub is an effective Cast */ + right = new_r_Conv(get_nodes_block(n), right, mode); + } + n = right; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } else if (right == b) { + if (mode != get_irn_mode(left)) { + /* This Sub is an effective Cast */ + left = new_r_Conv(get_nodes_block(n), left, mode); } + n = left; + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; } } - if (is_Add(b)) { - if (mode_wrap_around(mode)) { - ir_node *left = get_Add_left(b); - ir_node *right = get_Add_right(b); + if ((is_Add(b) || is_Or_Eor_Add(b)) && mode_wrap_around(mode)) { + ir_node *left = get_binop_left(b); + ir_node *right = get_binop_right(b); - /* FIXME: Does the Conv's work only for two complement or generally? */ - if (left == a) { - ir_mode *r_mode = get_irn_mode(right); + /* FIXME: Does the Conv's work only for two complement or generally? */ + if (left == a) { + ir_mode *r_mode = get_irn_mode(right); - n = new_r_Minus(get_nodes_block(n), right, r_mode); - if (mode != r_mode) { - /* This Sub is an effective Cast */ - n = new_r_Conv(get_nodes_block(n), n, mode); - } - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; - } else if (right == a) { - ir_mode *l_mode = get_irn_mode(left); + n = new_r_Minus(get_nodes_block(n), right, r_mode); + if (mode != r_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(get_nodes_block(n), n, mode); + } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; + } else if (right == a) { + ir_mode *l_mode = get_irn_mode(left); - n = new_r_Minus(get_nodes_block(n), left, l_mode); - if (mode != l_mode) { - /* This Sub is an effective Cast */ - n = new_r_Conv(get_nodes_block(n), n, mode); - } - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); - return n; + n = new_r_Minus(get_nodes_block(n), left, l_mode); + if (mode != l_mode) { + /* This Sub is an effective Cast */ + n = new_r_Conv(get_nodes_block(n), n, mode); } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB); + return n; } } if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) { @@ -2358,8 +3114,8 @@ 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)); @@ -2372,9 +3128,27 @@ 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 */ +} /** * Several transformation done on n*n=2n bits mul. @@ -2434,7 +3208,8 @@ static ir_node *transform_node_Mul(ir_node *n) ir_node *a = get_Mul_left(n); ir_node *b = get_Mul_right(n); - if (is_Bad(a) || is_Bad(b)) + n = fold_constant_associativity(n, tarval_mul); + if (n != oldn) return n; if (mode != get_irn_mode(a)) @@ -2515,10 +3290,11 @@ static ir_node *transform_node_Mul(ir_node *n) return n; } } - if (get_mode_arithmetic(mode) == irma_ieee754) { + if (get_mode_arithmetic(mode) == irma_ieee754 + || get_mode_arithmetic(mode) == irma_x86_extended_float) { if (is_Const(a)) { ir_tarval *tv = get_Const_tarval(a); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* 2.0 * b = b + b */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), b, b, mode); @@ -2528,7 +3304,7 @@ static ir_node *transform_node_Mul(ir_node *n) } else if (is_Const(b)) { ir_tarval *tv = get_Const_tarval(b); - if (tarval_ieee754_get_exponent(tv) == 1 && tarval_ieee754_zero_mantissa(tv) + if (tarval_get_exponent(tv) == 1 && tarval_zero_mantissa(tv) && !tarval_is_negative(tv)) { /* a * 2.0 = a + a */ n = new_rd_Add(get_irn_dbg_info(n), get_nodes_block(n), a, a, mode); @@ -2538,7 +3314,7 @@ static ir_node *transform_node_Mul(ir_node *n) } } return arch_dep_replace_mul_with_shifts(n); -} /* transform_node_Mul */ +} /** * Transform a Div Node. @@ -2643,14 +3419,14 @@ 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)); set_Tuple_pred(n, pn_Div_res, value); } return n; -} /* transform_node_Div */ +} /** * Transform a Mod node. @@ -2735,14 +3511,14 @@ 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)); set_Tuple_pred(n, pn_Mod_res, value); } return n; -} /* transform_node_Mod */ +} /** * Transform a Cond node. @@ -2752,24 +3528,28 @@ make_tuple: */ static ir_node *transform_node_Cond(ir_node *n) { - ir_node *a = get_Cond_selector(n); - ir_tarval *ta = value_of(a); ir_graph *irg = get_irn_irg(n); + ir_tarval *ta; ir_node *jmp; /* we need block info which is not available in floating irgs */ if (get_irg_pinned(irg) == op_pin_state_floats) return n; - if ((ta != tarval_bad) && - (get_irn_mode(a) == mode_b) && - (get_opt_unreachable_code())) { + ta = value_of(a); + if (ta == tarval_bad && is_Cmp(a)) { + /* try again with a direct call to compute_cmp, as we don't care + * about the MODEB_LOWERED flag here */ + ta = compute_cmp(a); + } + + if (ta != tarval_bad && get_irn_mode(a) == mode_b) { /* It's a boolean Cond, branching on a boolean constant. 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); @@ -2779,130 +3559,141 @@ static ir_node *transform_node_Cond(ir_node *n) } /* We might generate an endless loop, so keep it alive. */ add_End_keepalive(get_irg_end(irg), blk); + clear_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); } return n; -} /* transform_node_Cond */ - -/** - * 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 - * and(a OP c, b OP c) -> and(a, b) OP c - * note, might return a different op than n - */ -static ir_node *transform_bitwise_distributive(ir_node *n, - recursive_transform trans_func) -{ - ir_node *oldn = n; - ir_node *a = get_binop_left(n); - ir_node *b = get_binop_right(n); - ir_op *op = get_irn_op(a); - ir_op *op_root = get_irn_op(n); - - if (op != get_irn_op(b)) - return n; - - /* and(conv(a), conv(b)) -> conv(and(a,b)) */ - if (op == op_Conv) { - ir_node *a_op = get_Conv_op(a); - ir_node *b_op = get_Conv_op(b); - ir_mode *a_mode = get_irn_mode(a_op); - ir_mode *b_mode = get_irn_mode(b_op); - if (a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) { - ir_node *blk = get_nodes_block(n); - - n = exact_copy(n); - set_binop_left(n, a_op); - set_binop_right(n, b_op); - set_irn_mode(n, a_mode); - n = trans_func(n); - n = new_r_Conv(blk, n, get_irn_mode(oldn)); - - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - return n; - } - } - - if (op == op_Eor) { - /* nothing to gain here */ - return n; - } - - if (op == op_Shrs || op == op_Shr || op == op_Shl - || op == op_And || op == op_Or || op == op_Eor) { - ir_node *a_left = get_binop_left(a); - ir_node *a_right = get_binop_right(a); - ir_node *b_left = get_binop_left(b); - ir_node *b_right = get_binop_right(b); - ir_node *c = NULL; - ir_node *op1 = NULL; - ir_node *op2 = NULL; +} - if (is_op_commutative(op)) { - if (a_left == b_left) { - c = a_left; - op1 = a_right; - op2 = b_right; - } else if (a_left == b_right) { - c = a_left; - op1 = a_right; - op2 = b_left; - } else if (a_right == b_left) { - c = a_right; - op1 = a_left; - op2 = b_right; +static ir_node *transform_node_Switch(ir_node *n) +{ + ir_node *op = get_Switch_selector(n); + ir_tarval *val = value_of(op); + if (val != tarval_bad) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_graph *irg = get_irn_irg(n); + unsigned n_outs = get_Switch_n_outs(n); + ir_node *block = get_nodes_block(n); + ir_node *bad = new_r_Bad(irg, mode_X); + ir_node **in = XMALLOCN(ir_node*, n_outs); + const ir_switch_table *table = get_Switch_table(n); + size_t n_entries = ir_switch_table_get_n_entries(table); + long jmp_pn = 0; + size_t i; + unsigned o; + for (i = 0; i < n_entries; ++i) { + const ir_switch_table_entry *entry + = ir_switch_table_get_entry_const(table, i); + ir_tarval *min = entry->min; + ir_tarval *max = entry->max; + if (entry->pn == 0) + continue; + if ((min == max && min == val) + || (tarval_cmp(val, min) != ir_relation_less + && tarval_cmp(val, max) != ir_relation_greater)) { + jmp_pn = entry->pn; + break; } } - if (a_right == b_right) { - c = a_right; - op1 = a_left; - op2 = b_left; - } - - if (c != NULL) { - /* (a sop c) & (b sop c) => (a & b) sop c */ - ir_node *blk = get_nodes_block(n); - - ir_node *new_n = exact_copy(n); - set_binop_left(new_n, op1); - set_binop_right(new_n, op2); - new_n = trans_func(new_n); - - if (op_root == op_Eor && op == op_Or) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_mode *mode = get_irn_mode(c); - - c = new_rd_Not(dbgi, blk, c, mode); - n = new_rd_And(dbgi, blk, new_n, c, mode); + for (o = 0; o < n_outs; ++o) { + if (o == (unsigned)jmp_pn) { + in[o] = new_rd_Jmp(dbgi, block); } else { - n = exact_copy(a); - set_nodes_block(n, blk); - set_binop_left(n, new_n); - set_binop_right(n, c); - add_identities(n); + in[o] = bad; } - - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); - return n; } + return new_r_Tuple(block, (int)n_outs, in); } - return n; } /** - * Create a 0 constant of given mode. + * 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 *create_zero_const(ir_graph *irg, ir_mode *mode) +static ir_node *transform_node_shift_bitop(ir_node *n) { - ir_tarval *tv = get_mode_null(mode); - ir_node *cnst = new_r_Const(irg, tv); + 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; - return cnst; + 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; } /** @@ -2914,7 +3705,10 @@ static ir_node *transform_node_And(ir_node *n) ir_node *a = get_And_left(n); ir_node *b = get_And_right(n); ir_mode *mode; - vrp_attr *a_vrp, *b_vrp; + + n = fold_constant_associativity(n, tarval_and); + if (n != oldn) + return n; if (is_Cmp(a) && is_Cmp(b)) { ir_node *a_left = get_Cmp_left(a); @@ -2931,36 +3725,63 @@ 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); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); + /* 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))) { + if (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, xora, b_mode); + ir_node *or = new_rd_Or(dbgi, block, conv, xorb, b_mode); + ir_graph *irg = get_irn_irg(n); + ir_node *zero = create_zero_const(irg, b_mode); + return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); + } + if (values_in_mode(get_irn_mode(b_left), get_irn_mode(a_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); + } } } mode = get_irn_mode(n); HANDLE_BINOP_PHI((eval_func) tarval_and, a, b, c, mode); - if (is_Or(a)) { - if (is_Not(b)) { + if (is_Or(a) || is_Or_Eor_Add(a)) { + ir_node *or_left = get_binop_left(a); + ir_node *or_right = get_binop_right(a); + if (complement_values(or_left, b)) { + /* (a|b) & ~a => b & ~a */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_right, b, mode); + } else if (complement_values(or_right, b)) { + /* (a|b) & ~b => a & ~b */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_left, b, mode); + } else if (is_Not(b)) { ir_node *op = get_Not_op(b); if (is_And(op)) { ir_node *ba = get_And_left(op); ir_node *bb = get_And_right(op); /* it's enough to test the following cases due to normalization! */ - if (get_Or_left(a) == ba && get_Or_right(a) == bb) { + if (or_left == ba && or_right == bb) { /* (a|b) & ~(a&b) = a^b */ ir_node *block = get_nodes_block(n); @@ -2971,15 +3792,27 @@ static ir_node *transform_node_And(ir_node *n) } } } - if (is_Or(b)) { - if (is_Not(a)) { + if (is_Or(b) || is_Or_Eor_Add(b)) { + ir_node *or_left = get_binop_left(b); + ir_node *or_right = get_binop_right(b); + if (complement_values(or_left, a)) { + /* (a|b) & ~a => b & ~a */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_right, a, mode); + } else if (complement_values(or_right, a)) { + /* (a|b) & ~b => a & ~b */ + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_And(dbgi, block, or_left, a, mode); + } else if (is_Not(a)) { ir_node *op = get_Not_op(a); if (is_And(op)) { ir_node *aa = get_And_left(op); ir_node *ab = get_And_right(op); /* it's enough to test the following cases due to normalization! */ - if (get_Or_left(b) == aa && get_Or_right(b) == ab) { + if (or_left == aa && or_right == ab) { /* (a|b) & ~(a&b) = a^b */ ir_node *block = get_nodes_block(n); @@ -2990,9 +3823,9 @@ static ir_node *transform_node_And(ir_node *n) } } } - if (is_Eor(a)) { - ir_node *al = get_Eor_left(a); - ir_node *ar = get_Eor_right(a); + if (is_Eor(a) || is_Or_Eor_Add(a)) { + ir_node *al = get_binop_left(a); + ir_node *ar = get_binop_right(a); if (al == b) { /* (b ^ a) & b -> ~a & b */ @@ -3015,9 +3848,9 @@ static ir_node *transform_node_And(ir_node *n) return n; } } - if (is_Eor(b)) { - ir_node *bl = get_Eor_left(b); - ir_node *br = get_Eor_right(b); + if (is_Eor(b) || is_Or_Eor_Add(b)) { + ir_node *bl = get_binop_left(b); + ir_node *br = get_binop_right(b); if (bl == a) { /* a & (a ^ b) -> a & ~b */ @@ -3053,110 +3886,29 @@ static ir_node *transform_node_And(ir_node *n) return n; } - b_vrp = vrp_get_info(b); - if (is_Const(a) && b_vrp && (tarval_cmp(tarval_or(get_Const_tarval(a), - b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) { - - return b; - + if (is_Const(a)) { + vrp_attr *b_vrp = vrp_get_info(b); + ir_tarval *a_val = get_Const_tarval(a); + if (b_vrp != NULL && tarval_or(a_val, b_vrp->bits_not_set) == a_val) { + return b; + } } - a_vrp = vrp_get_info(a); - if (is_Const(b) && a_vrp && (tarval_cmp(tarval_or(get_Const_tarval(b), - a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) { - return a; + if (is_Const(b)) { + vrp_attr *a_vrp = vrp_get_info(a); + ir_tarval *b_val = get_Const_tarval(b); + if (a_vrp != NULL && tarval_or(b_val, a_vrp->bits_not_set) == b_val) { + return a; + } } n = transform_bitwise_distributive(n, transform_node_And); + if (is_And(n)) + n = transform_node_bitop_shift(n); return n; -} /* transform_node_And */ - -/* 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 bool 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)); } -/** - * Transform an Eor. - */ -static ir_node *transform_node_Eor(ir_node *n) -{ - ir_node *c, *oldn = n; - ir_node *a = get_Eor_left(n); - ir_node *b = get_Eor_right(n); - ir_mode *mode = get_irn_mode(n); - - /* we can combine the relations of two compares with the same operands */ - if (is_Cmp(a) && is_Cmp(b)) { - ir_node *a_left = get_Cmp_left(a); - ir_node *a_right = get_Cmp_left(a); - ir_node *b_left = get_Cmp_left(b); - ir_node *b_right = get_Cmp_right(b); - if (a_left == b_left && b_left == b_right) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_relation a_relation = get_Cmp_relation(a); - ir_relation b_relation = get_Cmp_relation(b); - ir_relation new_relation = a_relation ^ b_relation; - return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); - } - } - - HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode); - - /* normalize not nodes... ~a ^ b <=> a ^ ~b */ - if (is_Not(a) && operands_are_normalized(get_Not_op(a), b)) { - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_node *new_not = new_rd_Not(dbg, block, b, mode); - ir_node *new_left = get_Not_op(a); - n = new_rd_Eor(dbg, block, new_left, new_not, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); - return n; - } else if (is_Not(b) && !operands_are_normalized(a, get_Not_op(b))) { - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_node *new_not = new_rd_Not(dbg, block, a, mode); - ir_node *new_right = get_Not_op(b); - n = new_rd_Eor(dbg, block, new_not, new_right, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); - return n; - } - - /* x ^ 1...1 -> ~1 */ - if (is_Const(b) && is_Const_all_one(b)) { - n = new_r_Not(get_nodes_block(n), a, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT); - return n; - } - - n = transform_bitwise_distributive(n, transform_node_Eor); - return n; -} /* transform_node_Eor */ - /** * Transform a Not. */ @@ -3180,11 +3932,11 @@ static ir_node *transform_node_Not(ir_node *n) } /* normalize ~(a ^ b) => a ^ ~b */ - if (is_Eor(a)) { + if (is_Eor(a) || is_Or_Eor_Add(a)) { dbg_info *dbg = get_irn_dbg_info(n); ir_node *block = get_nodes_block(n); - ir_node *eor_right = get_Eor_right(a); - ir_node *eor_left = get_Eor_left(a); + ir_node *eor_right = get_binop_right(a); + ir_node *eor_left = get_binop_left(a); eor_right = new_rd_Not(dbg, block, eor_right, mode); n = new_rd_Eor(dbg, block, eor_left, eor_right, mode); return n; @@ -3198,11 +3950,11 @@ static ir_node *transform_node_Not(ir_node *n) ir_node *add_l = get_Minus_op(a); ir_node *add_r = new_rd_Const(dbg, irg, get_mode_minus_one(mode)); n = new_rd_Add(dbg, block, add_l, add_r, mode); - } else if (is_Add(a)) { - ir_node *add_r = get_Add_right(a); + } else if (is_Add(a) || is_Or_Eor_Add(a)) { + ir_node *add_r = get_binop_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 *op = get_binop_left(a); ir_node *blk = get_nodes_block(n); n = new_rd_Minus(get_irn_dbg_info(n), blk, op, get_irn_mode(n)); DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1); @@ -3306,71 +4058,67 @@ static ir_node *transform_node_Minus(ir_node *n) } return n; -} /* transform_node_Minus */ +} /** * Transform a Proj(Load) with a non-null address. */ static ir_node *transform_node_Proj_Load(ir_node *proj) { - if (get_opt_ldst_only_null_ptr_exceptions()) { - if (get_irn_mode(proj) == mode_X) { - ir_node *load = get_Proj_pred(proj); + if (get_irn_mode(proj) == mode_X) { + ir_node *load = get_Proj_pred(proj); - /* get the Load address */ - const ir_node *addr = get_Load_ptr(load); - const ir_node *confirm; + /* get the Load address */ + const ir_node *addr = get_Load_ptr(load); + const ir_node *confirm; - if (value_not_null(addr, &confirm)) { - if (confirm == NULL) { - /* this node may float if it did not depend on a Confirm */ - set_irn_pinned(load, op_pin_state_floats); - } - if (get_Proj_proj(proj) == pn_Load_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return new_r_Bad(irg, mode_X); - } else { - ir_node *blk = get_nodes_block(load); - return new_r_Jmp(blk); - } + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(load, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Load_X_except) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(load); + return new_r_Jmp(blk); } } } return proj; -} /* transform_node_Proj_Load */ +} /** * Transform a Proj(Store) with a non-null address. */ static ir_node *transform_node_Proj_Store(ir_node *proj) { - if (get_opt_ldst_only_null_ptr_exceptions()) { - if (get_irn_mode(proj) == mode_X) { - ir_node *store = get_Proj_pred(proj); + if (get_irn_mode(proj) == mode_X) { + ir_node *store = get_Proj_pred(proj); - /* get the load/store address */ - const ir_node *addr = get_Store_ptr(store); - const ir_node *confirm; + /* get the load/store address */ + const ir_node *addr = get_Store_ptr(store); + const ir_node *confirm; - if (value_not_null(addr, &confirm)) { - if (confirm == NULL) { - /* this node may float if it did not depend on a Confirm */ - set_irn_pinned(store, op_pin_state_floats); - } - if (get_Proj_proj(proj) == pn_Store_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return new_r_Bad(irg, mode_X); - } else { - ir_node *blk = get_nodes_block(store); - return new_r_Jmp(blk); - } + if (value_not_null(addr, &confirm)) { + if (confirm == NULL) { + /* this node may float if it did not depend on a Confirm */ + set_irn_pinned(store, op_pin_state_floats); + } + if (get_Proj_proj(proj) == pn_Store_X_except) { + ir_graph *irg = get_irn_irg(proj); + DBG_OPT_EXC_REM(proj); + return new_r_Bad(irg, mode_X); + } else { + ir_node *blk = get_nodes_block(store); + return new_r_Jmp(blk); } } } return proj; -} /* transform_node_Proj_Store */ +} /** * Transform a Proj(Div) with a non-zero value. @@ -3423,7 +4171,7 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) } } return proj; -} /* transform_node_Proj_Div */ +} /** * Transform a Proj(Mod) with a non-zero value. @@ -3487,85 +4235,6 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) } } return proj; -} /* transform_node_Proj_Mod */ - -/** - * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases. - */ -static ir_node *transform_node_Proj_Cond(ir_node *proj) -{ - ir_node *n = get_Proj_pred(proj); - ir_node *b = get_Cond_selector(n); - - if (!get_opt_unreachable_code()) - return n; - - if (mode_is_int(get_irn_mode(b))) { - ir_tarval *tb = value_of(b); - - if (tb != tarval_bad) { - /* we have a constant switch */ - long num = get_Proj_proj(proj); - - if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */ - if (get_tarval_long(tb) == num) { - /* Do NOT create a jump here, or we will have 2 control flow ops - * in a block. This case is optimized away in optimize_cf(). */ - return proj; - } else { - ir_graph *irg = get_irn_irg(proj); - /* this case will NEVER be taken, kill it */ - return new_r_Bad(irg, mode_X); - } - } - } else { - long num = get_Proj_proj(proj); - vrp_attr *b_vrp = vrp_get_info(b); - if (num != get_Cond_default_proj(n) && b_vrp) { - /* Try handling with vrp data. We only remove dead parts. */ - ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b)); - - if (b_vrp->range_type == VRP_RANGE) { - ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & ir_relation_greater) == cmp_result - && (cmp_result2 & ir_relation_less) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - return new_r_Bad(irg, mode_X); - } - } else if (b_vrp->range_type == VRP_ANTIRANGE) { - ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & ir_relation_less_equal) == cmp_result - && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - return new_r_Bad(irg, mode_X); - } - } - - if (!(tarval_cmp( - tarval_and( b_vrp->bits_set, tp), - b_vrp->bits_set - ) == ir_relation_equal)) { - ir_graph *irg = get_irn_irg(proj); - return new_r_Bad(irg, mode_X); - } - - if (!(tarval_cmp( - tarval_and( - tarval_not(tp), - tarval_not(b_vrp->bits_not_set)), - tarval_not(b_vrp->bits_not_set)) - == ir_relation_equal)) { - ir_graph *irg = get_irn_irg(proj); - return new_r_Bad(irg, mode_X); - } - } - } - } - return proj; } /** @@ -3590,6 +4259,26 @@ static bool is_single_bit(const ir_node *node) return false; } +/** + * checks if node just flips a bit in another node and returns that other node + * if so. @p tv should be a value having just 1 bit set + */ +static ir_node *flips_bit(const ir_node *node, ir_tarval *tv) +{ + if (is_Not(node)) + return get_Not_op(node); + if (is_Eor(node)) { + ir_node *right = get_Eor_right(node); + if (is_Const(right)) { + ir_tarval *right_tv = get_Const_tarval(right); + ir_mode *mode = get_irn_mode(node); + if (tarval_and(right_tv, tv) != get_mode_null(mode)) + return get_Eor_left(node); + } + } + return NULL; +} + /** * Normalizes and optimizes Cmp nodes. */ @@ -3612,30 +4301,51 @@ 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); } @@ -3745,11 +4455,11 @@ static ir_node *transform_node_Cmp(ir_node *n) } /* X+A == A, A+X == A, A-X == A -> X == 0 */ - if (is_Add(left) || is_Sub(left)) { + if (is_Add(left) || is_Sub(left) || is_Or_Eor_Add(left)) { ir_node *ll = get_binop_left(left); ir_node *lr = get_binop_right(left); - if (lr == right && is_Add(left)) { + if (lr == right && (is_Add(left) || is_Or_Eor_Add(left))) { ir_node *tmp = ll; ll = lr; lr = tmp; @@ -3762,11 +4472,11 @@ static ir_node *transform_node_Cmp(ir_node *n) DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } - if (is_Add(right) || is_Sub(right)) { + if (is_Add(right) || is_Sub(right) || is_Or_Eor_Add(right)) { ir_node *rl = get_binop_left(right); ir_node *rr = get_binop_right(right); - if (rr == left && is_Add(right)) { + if (rr == left && (is_Add(right) || is_Or_Eor_Add(right))) { ir_node *tmp = rl; rl = rr; rr = tmp; @@ -3807,33 +4517,62 @@ static ir_node *transform_node_Cmp(ir_node *n) } /* Cmp(Eor(x, y), 0) <=> Cmp(x, y) at least for the ==0,!=0 * cases */ - if (is_Const(right) && is_Const_null(right) && is_Eor(left)) { + if (is_Const(right) && is_Const_null(right) && + (is_Eor(left) || is_Or_Eor_Add(left))) { right = get_Eor_right(left); left = get_Eor_left(left); changed = true; } - } /* mode_is_int(...) */ + } } - /* Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced - * by the simpler Cmp(And(1bit), val), 0) negated pnc */ - if (mode_is_int(mode) && is_And(left) - && (relation == ir_relation_equal + if (mode_is_int(mode) && is_And(left)) { + /* a complicated Cmp(And(1bit, val), 1bit) "bit-testing" can be replaced + * by the simpler Cmp(And(1bit, val), 0) negated pnc */ + if (relation == ir_relation_equal || (mode_is_signed(mode) && relation == ir_relation_less_greater) - || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less))) { - ir_node *and0 = get_And_left(left); - ir_node *and1 = get_And_right(left); - if (and1 == right) { - ir_node *tmp = and0; - and0 = and1; - and1 = tmp; - } - if (and0 == right && is_single_bit(and0)) { - ir_graph *irg = get_irn_irg(n); - relation = - relation == ir_relation_equal ? ir_relation_less_greater : ir_relation_equal; - right = create_zero_const(irg, mode); - changed |= 1; + || (!mode_is_signed(mode) && (relation & ir_relation_less_equal) == ir_relation_less)) { + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (and1 == right) { + ir_node *tmp = and0; + and0 = and1; + and1 = tmp; + } + if (and0 == right && is_single_bit(and0)) { + ir_graph *irg = get_irn_irg(n); + relation = + relation == ir_relation_equal ? ir_relation_less_greater + : ir_relation_equal; + right = create_zero_const(irg, mode); + changed |= 1; + goto is_bittest; + } + } + + if (is_Const(right) && is_Const_null(right) && + (relation == ir_relation_equal + || (relation == ir_relation_less_greater) + || (!mode_is_signed(mode) && relation == ir_relation_greater))) { +is_bittest: { + /* instead of flipping the bit before the bit-test operation negate + * pnc */ + ir_node *and0 = get_And_left(left); + ir_node *and1 = get_And_right(left); + if (is_Const(and1)) { + ir_tarval *tv = get_Const_tarval(and1); + if (tarval_is_single_bit(tv)) { + ir_node *flipped = flips_bit(and0, tv); + if (flipped != NULL) { + dbg_info *dbgi = get_irn_dbg_info(left); + ir_node *block = get_nodes_block(left); + relation = get_negated_relation(relation); + left = new_rd_And(dbgi, block, flipped, and1, mode); + changed |= 1; + } + } + } + } } } @@ -3897,6 +4636,43 @@ static ir_node *transform_node_Cmp(ir_node *n) if (tv != tarval_bad) { ir_mode *mode = get_irn_mode(right); + /* cmp(mux(x, cf, ct), c2) can be eliminated: + * cmp(ct,c2) | cmp(cf,c2) | result + * -----------|------------|-------- + * true | true | True + * false | false | False + * true | false | x + * false | true | not(x) + */ + if (is_Mux(left)) { + ir_node *mux_true = get_Mux_true(left); + ir_node *mux_false = get_Mux_false(left); + if (is_Const(mux_true) && is_Const(mux_false)) { + /* we can fold true/false constant separately */ + ir_tarval *tv_true = get_Const_tarval(mux_true); + ir_tarval *tv_false = get_Const_tarval(mux_false); + ir_relation r_true = tarval_cmp(tv_true, tv); + ir_relation r_false = tarval_cmp(tv_false, tv); + if (r_true != ir_relation_false + || r_false != ir_relation_false) { + bool rel_true = (r_true & relation) != 0; + bool rel_false = (r_false & relation) != 0; + ir_node *cond = get_Mux_sel(left); + if (rel_true == rel_false) { + relation = rel_true ? ir_relation_true + : ir_relation_false; + } else if (rel_true) { + return cond; + } else { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *notn = new_rd_Not(dbgi, block, cond, mode_b); + return notn; + } + } + } + } + /* TODO extend to arbitrary constants */ if (is_Conv(left) && tarval_is_null(tv)) { ir_node *op = get_Conv_op(left); @@ -3951,7 +4727,7 @@ static ir_node *transform_node_Cmp(ir_node *n) } /* for integer modes, we have more */ - if (mode_is_int(mode)) { + if (mode_is_int(mode) && !is_Const(left)) { /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ if ((relation == ir_relation_less || relation == ir_relation_greater_equal) && tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) { @@ -3995,9 +4771,9 @@ static ir_node *transform_node_Cmp(ir_node *n) } } /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */ - else if (is_Add(left)) { - ir_node *a_l = get_Add_left(left); - ir_node *a_r = get_Add_right(left); + else if (is_Add(left) || is_Or_Eor_Add(left)) { + ir_node *a_l = get_binop_left(left); + ir_node *a_r = get_binop_right(left); ir_node *a; ir_tarval *tv2; @@ -4032,8 +4808,8 @@ static ir_node *transform_node_Cmp(ir_node *n) } } } - } /* == or != */ - } /* mode_is_int */ + } + } if (relation == ir_relation_equal || relation == ir_relation_less_greater) { switch (get_irn_opcode(left)) { @@ -4194,9 +4970,9 @@ static ir_node *transform_node_Cmp(ir_node *n) DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; - } /* switch */ + } } - } /* tarval != bad */ + } } if (changedc) { /* need a new Const */ @@ -4268,7 +5044,7 @@ static ir_node *transform_node_Proj_CopyB(ir_node *proj) } } return proj; -} /* transform_node_Proj_CopyB */ +} /** * Optimize Bounds(idx, idx, upper) into idx. @@ -4325,7 +5101,7 @@ static ir_node *transform_node_Proj_Bound(ir_node *proj) } } return proj; -} /* transform_node_Proj_Bound */ +} /** * Does all optimizations on nodes that must be done on its Projs @@ -4338,12 +5114,23 @@ static ir_node *transform_node_Proj(ir_node *proj) if (n->op->ops.transform_node_Proj) return n->op->ops.transform_node_Proj(proj); return proj; -} /* transform_node_Proj */ +} +/** + * Test whether a block is unreachable + * Note: That this only returns true when + * IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE is set. + * This is important, as you easily end up producing invalid constructs in the + * unreachable code when optimizing away edges into the unreachable code. + * So only set this flag when you iterate localopts to the fixpoint. + * When you reach the fixpoint then all unreachable code is dead + * (= can't be reached by firm edges) and you won't see the invalid constructs + * anymore. + */ static bool is_block_unreachable(const ir_node *block) { const ir_graph *irg = get_irn_irg(block); - if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) + if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) return false; return get_Block_dom_depth(block) < 0; } @@ -4355,13 +5142,12 @@ static ir_node *transform_node_Block(ir_node *block) ir_node *bad = NULL; int i; - if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) + if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) return block; for (i = 0; i < arity; ++i) { - ir_node *pred = get_Block_cfgpred(block, i); - ir_node *pred_block = get_nodes_block(pred); - if (!is_Bad(pred) && !is_block_unreachable(pred_block)) + ir_node *const pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred))) continue; if (bad == NULL) bad = new_r_Bad(irg, mode_X); @@ -4382,328 +5168,96 @@ static ir_node *transform_node_Phi(ir_node *phi) /* Set phi-operands for bad-block inputs to bad */ for (i = 0; i < n; ++i) { - ir_node *pred = get_Block_cfgpred(block, i); - if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) { - if (bad == NULL) - bad = new_r_Bad(irg, mode); - set_irn_n(phi, i, bad); - } - } - - /* Move Confirms down through Phi nodes. */ - if (mode_is_reference(mode)) { - n = get_irn_arity(phi); - - /* Beware of Phi0 */ - if (n > 0) { - ir_node *pred = get_irn_n(phi, 0); - ir_node *bound, *new_phi, *block, **in; - ir_relation relation; - - if (! is_Confirm(pred)) - return phi; - - bound = get_Confirm_bound(pred); - relation = get_Confirm_relation(pred); - - NEW_ARR_A(ir_node *, in, n); - in[0] = get_Confirm_value(pred); - - for (i = 1; i < n; ++i) { - pred = get_irn_n(phi, i); - - if (! is_Confirm(pred) || - get_Confirm_bound(pred) != bound || - get_Confirm_relation(pred) != relation) - return phi; - in[i] = get_Confirm_value(pred); - } - /* move the Confirm nodes "behind" the Phi */ - block = get_irn_n(phi, -1); - new_phi = new_r_Phi(block, n, in, get_irn_mode(phi)); - return new_r_Confirm(block, new_phi, bound, relation); - } - } - return phi; -} - -/** - * Returns the operands of a commutative bin-op, if one operand is - * a const, it is returned as the second one. - */ -static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) -{ - ir_node *op_a = get_binop_left(binop); - ir_node *op_b = get_binop_right(binop); - - assert(is_op_commutative(get_irn_op(binop))); - - if (is_Const(op_a)) { - *a = op_b; - *c = op_a; - } else { - *a = op_a; - *c = op_b; - } -} /* get_comm_Binop_Ops */ - -/** - * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible. - * Such pattern may arise in bitfield stores. - * - * value c4 value c4 & c2 - * AND c3 AND c1 | c3 - * OR c2 ===> OR - * AND c1 - * OR - * - * - * value c2 value c1 - * AND c1 ===> OR if (c1 | c2) == 0x111..11 - * OR - */ -static ir_node *transform_node_Or_bf_store(ir_node *irn_or) -{ - ir_node *irn_and, *c1; - ir_node *or_l, *c2; - ir_node *and_l, *c3; - ir_node *value, *c4; - ir_node *new_and, *new_const, *block; - ir_mode *mode = get_irn_mode(irn_or); - - ir_tarval *tv1, *tv2, *tv3, *tv4, *tv; - - for (;;) { - ir_graph *irg; - get_comm_Binop_Ops(irn_or, &irn_and, &c1); - if (!is_Const(c1) || !is_And(irn_and)) - return irn_or; - - get_comm_Binop_Ops(irn_and, &or_l, &c2); - if (!is_Const(c2)) - return irn_or; - - tv1 = get_Const_tarval(c1); - tv2 = get_Const_tarval(c2); - - tv = tarval_or(tv1, tv2); - if (tarval_is_all_one(tv)) { - /* the AND does NOT clear a bit with isn't set by the OR */ - set_Or_left(irn_or, or_l); - set_Or_right(irn_or, c1); - - /* check for more */ - continue; - } - - if (!is_Or(or_l)) - return irn_or; - - get_comm_Binop_Ops(or_l, &and_l, &c3); - if (!is_Const(c3) || !is_And(and_l)) - return irn_or; - - get_comm_Binop_Ops(and_l, &value, &c4); - if (!is_Const(c4)) - return irn_or; - - /* ok, found the pattern, check for conditions */ - assert(mode == get_irn_mode(irn_and)); - assert(mode == get_irn_mode(or_l)); - assert(mode == get_irn_mode(and_l)); - - tv3 = get_Const_tarval(c3); - tv4 = get_Const_tarval(c4); - - tv = tarval_or(tv4, tv2); - if (!tarval_is_all_one(tv)) { - /* have at least one 0 at the same bit position */ - return irn_or; - } - - if (tv3 != tarval_andnot(tv3, tv4)) { - /* bit in the or_mask is outside the and_mask */ - return irn_or; - } - - if (tv1 != tarval_andnot(tv1, tv2)) { - /* bit in the or_mask is outside the and_mask */ - return irn_or; - } - - /* ok, all conditions met */ - block = get_irn_n(irn_or, -1); - irg = get_irn_irg(block); - - new_and = new_r_And(block, value, new_r_Const(irg, tarval_and(tv4, tv2)), mode); - - new_const = new_r_Const(irg, tarval_or(tv3, tv1)); - - set_Or_left(irn_or, new_and); - set_Or_right(irn_or, new_const); - - /* check for more */ - } -} /* transform_node_Or_bf_store */ - -/** - * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rotl - */ -static ir_node *transform_node_Or_Rotl(ir_node *irn_or) -{ - ir_mode *mode = get_irn_mode(irn_or); - ir_node *shl, *shr, *block; - ir_node *irn, *x, *c1, *c2, *n; - ir_tarval *tv1, *tv2; - - /* some backends can't handle rotl */ - if (!be_get_backend_param()->support_rotl) - return irn_or; - - if (! mode_is_int(mode)) - return irn_or; - - shl = get_binop_left(irn_or); - shr = get_binop_right(irn_or); - - if (is_Shr(shl)) { - if (!is_Shl(shr)) - return irn_or; - - irn = shl; - shl = shr; - shr = irn; - } else if (!is_Shl(shl)) { - return irn_or; - } else if (!is_Shr(shr)) { - return irn_or; - } - x = get_Shl_left(shl); - if (x != get_Shr_left(shr)) - return irn_or; - - c1 = get_Shl_right(shl); - c2 = get_Shr_right(shr); - if (is_Const(c1) && is_Const(c2)) { - tv1 = get_Const_tarval(c1); - if (! tarval_is_long(tv1)) - return irn_or; - - tv2 = get_Const_tarval(c2); - if (! tarval_is_long(tv2)) - return irn_or; - - if (get_tarval_long(tv1) + get_tarval_long(tv2) - != (int) get_mode_size_bits(mode)) - return irn_or; - - /* yet, condition met */ - block = get_nodes_block(irn_or); - - n = new_r_Rotl(block, x, c1, mode); - - DBG_OPT_ALGSIM1(irn_or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROTL); - return n; - } - - /* Note: the obvious rot formulation (a << x) | (a >> (32-x)) gets - * transformed to (a << x) | (a >> -x) by transform_node_shift_modulo() */ - if (!ir_is_negated_value(c1, c2)) { - return irn_or; - } - - /* yet, condition met */ - block = get_nodes_block(irn_or); - n = new_r_Rotl(block, x, c1, mode); - DBG_OPT_ALGSIM0(irn_or, n, FS_OPT_OR_SHFT_TO_ROTL); - return n; -} /* transform_node_Or_Rotl */ - -static bool is_cmp_unequal_zero(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 { - return relation == ir_relation_greater; - } -} - -/** - * Transform an Or. - */ -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); - - mode = get_irn_mode(n); - a = get_Not_op(a); - b = get_Not_op(b); - n = new_rd_And(get_irn_dbg_info(n), block, a, b, mode); - n = new_rd_Not(get_irn_dbg_info(n), block, n, mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_DEMORGAN); - return n; - } - - /* we can combine the relations of two compares with the same operands */ - if (is_Cmp(a) && is_Cmp(b)) { - ir_node *a_left = get_Cmp_left(a); - ir_node *a_right = get_Cmp_left(a); - ir_node *b_left = get_Cmp_left(b); - ir_node *b_right = get_Cmp_right(b); - if (a_left == b_left && b_left == b_right) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); - ir_relation a_relation = get_Cmp_relation(a); - ir_relation b_relation = get_Cmp_relation(b); - ir_relation new_relation = a_relation | b_relation; - return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + if (!is_Bad(get_Phi_pred(phi, i))) { + ir_node *pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) { + if (bad == NULL) + bad = new_r_Bad(irg, mode); + set_irn_n(phi, i, bad); + } } - /* 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); - return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); + } + + /* Move Pin nodes down through Phi nodes. */ + if (mode == mode_M) { + n = get_irn_arity(phi); + + /* Beware of Phi0 */ + if (n > 0) { + ir_node **in; + ir_node *new_phi; + bool has_pin = false; + + NEW_ARR_A(ir_node *, in, n); + + for (i = 0; i < n; ++i) { + ir_node *pred = get_irn_n(phi, i); + + if (is_Pin(pred)) { + in[i] = get_Pin_op(pred); + has_pin = true; + } else if (is_Bad(pred)) { + in[i] = pred; + } else { + return phi; + } + } + + if (!has_pin) + return phi; + + /* Move the Pin nodes "behind" the Phi. */ + block = get_irn_n(phi, -1); + new_phi = new_r_Phi(block, n, in, mode_M); + return new_r_Pin(block, new_phi); } } + /* Move Confirms down through Phi nodes. */ + else if (mode_is_reference(mode)) { + n = get_irn_arity(phi); - mode = get_irn_mode(n); - HANDLE_BINOP_PHI((eval_func) tarval_or, a, b, c, mode); + /* Beware of Phi0 */ + if (n > 0) { + ir_node *pred = get_irn_n(phi, 0); + ir_node *bound, *new_phi, *block, **in; + ir_relation relation; + bool has_confirm = false; - n = transform_node_Or_bf_store(n); - n = transform_node_Or_Rotl(n); - if (n != oldn) - return n; + if (! is_Confirm(pred)) + return phi; - n = transform_bitwise_distributive(n, transform_node_Or); + bound = get_Confirm_bound(pred); + relation = get_Confirm_relation(pred); - return n; -} /* transform_node_Or */ + NEW_ARR_A(ir_node *, in, n); + in[0] = get_Confirm_value(pred); + + for (i = 1; i < n; ++i) { + pred = get_irn_n(phi, i); + + if (is_Confirm(pred) && + get_Confirm_bound(pred) == bound && + get_Confirm_relation(pred) == relation) { + in[i] = get_Confirm_value(pred); + has_confirm = true; + } else if (is_Bad(pred)) { + in[i] = pred; + } else { + return phi; + } + } + if (!has_confirm) + return phi; -/* forward */ -static ir_node *transform_node(ir_node *n); + /* move the Confirm nodes "behind" the Phi */ + block = get_irn_n(phi, -1); + new_phi = new_r_Phi(block, n, in, get_irn_mode(phi)); + return new_r_Confirm(block, new_phi, bound, relation); + } + } + return phi; +} /** * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl, Rotl. @@ -4714,9 +5268,11 @@ 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 +5281,7 @@ 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 +5289,55 @@ 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); + + /* 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); - return new_r_Const(irg, get_mode_null(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); @@ -4773,92 +5347,7 @@ 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; + return irn; } /** @@ -4979,7 +5468,7 @@ 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) @@ -5011,8 +5500,8 @@ static ir_node *transform_node_shift_modulo(ir_node *n, return n; newop = new_r_Const(irg, tv_mod); - } else if (is_Add(right)) { - ir_node *add_right = get_Add_right(right); + } else if (is_Add(right) || is_Or_Eor_Add(right)) { + ir_node *add_right = get_binop_right(right); if (is_Const(add_right)) { ir_tarval *tv = get_Const_tarval(add_right); ir_tarval *tv_mod = get_modulo_tv_value(tv, modulo); @@ -5021,7 +5510,7 @@ static ir_node *transform_node_shift_modulo(ir_node *n, return n; newconst = new_r_Const(irg, tv_mod); - newop = new_r_Add(block, get_Add_left(right), newconst, + newop = new_r_Add(block, get_binop_left(right), newconst, mode_right); } } else if (is_Sub(right)) { @@ -5067,31 +5556,60 @@ 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 */ +} /** * Transform a Shrs. */ 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_mode *mode = get_irn_mode(n); + ir_node *oldn = n; + ir_node *a = get_Shrs_left(n); + ir_node *b = get_Shrs_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *c; + vrp_attr *attr; + + 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 (n != oldn) + return 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_modulo(n, new_rd_Shrs); + if (n != oldn) + return n; + n = transform_node_shift_bitop(n); + if (n != oldn) + return n; + + /* normalisation: use Shr when sign bit is guaranteed to be cleared */ + attr = vrp_get_info(a); + if (attr != NULL) { + unsigned bits = get_mode_size_bits(mode); + ir_tarval *scount = new_tarval_from_long(bits-1, mode_Iu); + ir_tarval *sign = tarval_shl(get_mode_one(mode), scount); + if (tarval_is_null(tarval_and(attr->bits_not_set, sign))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + return new_rd_Shr(dbgi, block, a, b, mode); + } + } return n; -} /* transform_node_Shrs */ +} /** * Transform a Shl. @@ -5111,10 +5629,10 @@ 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 */ +} /** * Transform a Rotl. @@ -5130,10 +5648,10 @@ 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 */ +} /** * Transform a Conv. @@ -5187,7 +5705,7 @@ static ir_node *transform_node_Conv(ir_node *n) } return n; -} /* transform_node_Conv */ +} /** * Remove dead blocks and nodes in dead blocks @@ -5206,7 +5724,7 @@ 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; @@ -5215,7 +5733,7 @@ static ir_node *transform_node_End(ir_node *n) if (j != n_keepalives) set_End_keepalives(n, j, in); return n; -} /* transform_node_End */ +} int ir_is_negated_value(const ir_node *a, const ir_node *b) { @@ -5236,71 +5754,236 @@ 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_false, + const ir_node *mux_true) +{ + 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_false, + ir_node *mux_true) +{ + ir_node *cmp_left = get_Cmp_left(sel); + return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true; +} + +bool ir_is_optimizable_mux(const ir_node *sel, const ir_node *mux_false, + const ir_node *mux_true) +{ + /* this code should return true each time transform_node_Mux would + * optimize the Mux completely away */ + + ir_mode *mode = get_irn_mode(mux_false); + if (get_mode_arithmetic(mode) == irma_twos_complement + && ir_mux_is_abs(sel, mux_false, mux_true)) + return true; + + if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) { + const ir_node *cmp_r = get_Cmp_right(sel); + const ir_node *cmp_l = get_Cmp_left(sel); + const ir_node *f = mux_false; + const ir_node *t = mux_true; + + if (is_Const(t) && is_Const_null(t)) { + t = mux_false; + f = mux_true; + } + + if (is_And(cmp_l) && f == cmp_r) { + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && is_single_bit(and_r)) + return true; + and_l = get_And_left(cmp_l); + if (and_l == t && is_single_bit(and_l)) + return true; + } + } + + return false; +} + +/** + * Optimize a Mux(c, 0, 1) node (sometimes called a "set" instruction) + */ +static ir_node *transform_Mux_set(ir_node *n) +{ + ir_node *cond = get_Mux_sel(n); + ir_mode *dest_mode; + ir_mode *mode; + ir_node *left; + ir_node *right; + ir_relation relation; + bool need_not; + dbg_info *dbgi; + ir_node *block; + ir_graph *irg; + ir_node *a; + ir_node *b; + unsigned bits; + ir_tarval *tv; + ir_node *shift_cnt; + ir_node *res; + + if (!is_Cmp(cond)) + return n; + left = get_Cmp_left(cond); + mode = get_irn_mode(left); + if (!mode_is_int(mode) && !mode_is_reference(mode)) + return n; + dest_mode = get_irn_mode(n); + if (!mode_is_int(dest_mode) && !mode_is_reference(dest_mode)) + return n; + right = get_Cmp_right(cond); + relation = get_Cmp_relation(cond) & ~ir_relation_unordered; + if (get_mode_size_bits(mode) >= get_mode_size_bits(dest_mode) + && !(mode_is_signed(mode) && is_Const(right) && is_Const_null(right) + && relation != ir_relation_greater)) + return n; + + need_not = false; + switch (relation) { + case ir_relation_less: + /* a < b -> (a - b) >> 31 */ + a = left; + b = right; + break; + case ir_relation_less_equal: + /* a <= b -> ~(a - b) >> 31 */ + a = right; + b = left; + need_not = true; + break; + case ir_relation_greater: + /* a > b -> (b - a) >> 31 */ + a = right; + b = left; + break; + case ir_relation_greater_equal: + /* a >= b -> ~(a - b) >> 31 */ + a = left; + b = right; + need_not = true; + break; + default: + return n; + } + + dbgi = get_irn_dbg_info(n); + block = get_nodes_block(n); + irg = get_irn_irg(block); + bits = get_mode_size_bits(dest_mode); + tv = new_tarval_from_long(bits-1, mode_Iu); + shift_cnt = new_rd_Const(dbgi, irg, tv); + + if (mode != dest_mode) { + a = new_rd_Conv(dbgi, block, a, dest_mode); + b = new_rd_Conv(dbgi, block, b, dest_mode); + } + + res = new_rd_Sub(dbgi, block, a, b, dest_mode); + if (need_not) { + res = new_rd_Not(dbgi, block, res, dest_mode); + } + res = new_rd_Shr(dbgi, block, res, shift_cnt, dest_mode); + return res; +} + /** * 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); - - if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) - return 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); - if (is_Mux(t)) { - ir_node* block = get_nodes_block(n); - ir_node* c0 = sel; - ir_node* c1 = get_Mux_sel(t); - ir_node* t1 = get_Mux_true(t); - ir_node* f1 = get_Mux_false(t); - if (f == f1) { - /* Mux(cond0, Mux(cond1, x, y), y) -> typical if (cond0 && cond1) x else y */ - ir_node* and_ = new_r_And(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, f1, t1, mode); - n = new_mux; - sel = and_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); - } else if (f == t1) { - /* Mux(cond0, Mux(cond1, x, y), x) */ - ir_node* not_c1 = new_r_Not(block, c1, mode_b); - ir_node* and_ = new_r_And(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, and_, t1, f1, mode); - n = new_mux; - sel = and_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, t, FS_OPT_MUX_COMBINE); - } - } else if (is_Mux(f)) { - ir_node* block = get_nodes_block(n); - ir_node* c0 = sel; - ir_node* c1 = get_Mux_sel(f); - ir_node* t1 = get_Mux_true(f); - ir_node* f1 = get_Mux_false(f); - if (t == t1) { - /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */ - ir_node* or_ = new_r_Or(block, c0, c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, f1, t1, mode); - n = new_mux; - sel = or_; - f = f1; - t = t1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); - } else if (t == f1) { - /* Mux(cond0, x, Mux(cond1, y, x)) */ - ir_node* not_c1 = new_r_Not(block, c1, mode_b); - ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b); - ir_node* new_mux = new_r_Mux(block, or_, t1, f1, mode); - n = new_mux; - sel = or_; - f = t1; - t = f1; - DBG_OPT_ALGSIM0(oldn, f, FS_OPT_MUX_COMBINE); + /* 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, f, t); + 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, f, t); + 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; } } @@ -5319,136 +6002,131 @@ static ir_node *transform_node_Mux(ir_node *n) relation = get_negated_relation(relation); sel = new_rd_Cmp(seldbgi, block, get_Cmp_left(sel), get_Cmp_right(sel), relation); - n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); + return new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); } - /* note: after normalization, false can only happen on default */ - if (mode == mode_b) { - dbg_info *dbg = get_irn_dbg_info(n); - ir_node *block = get_nodes_block(n); + if (is_Const(f) && is_Const_null(f) && is_Const(t) && is_Const_one(t)) { + n = transform_Mux_set(n); + if (n != oldn) + return n; + } - if (is_Const(t)) { - ir_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 { - /* Muxb(sel, true, x) = Or(sel, x) */ - n = new_rd_Or(dbg, block, sel, f, mode_b); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL); - return n; - } + /* the following optimisations create new mode_b nodes, so only do them + * before mode_b lowering */ + if (!is_irg_state(irg, IR_GRAPH_STATE_MODEB_LOWERED)) { + if (is_Mux(t)) { + ir_node* block = get_nodes_block(n); + ir_node* c0 = sel; + ir_node* c1 = get_Mux_sel(t); + ir_node* t1 = get_Mux_true(t); + ir_node* f1 = get_Mux_false(t); + if (f == f1) { + /* Mux(cond0, Mux(cond1, x, y), y) => Mux(cond0 && cond1, x, y) */ + ir_node* and_ = new_r_And(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, f1, t1, mode); + } else if (f == t1) { + /* Mux(cond0, Mux(cond1, x, y), x) */ + ir_node* not_c1 = new_r_Not(block, c1, mode_b); + ir_node* and_ = new_r_And(block, c0, not_c1, mode_b); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, and_, t1, f1, mode); } - } else if (is_Const(f)) { - ir_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, block, sel, mode_b); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); - n = new_rd_Or(dbg, block, not_sel, t, mode_b); - return n; - } else { - /* Muxb(sel, x, false) = And(sel, x) */ - assert(tv_f == tarval_b_false); - n = new_rd_And(dbg, block, sel, t, mode_b); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); - return n; + } else if (is_Mux(f)) { + ir_node* block = get_nodes_block(n); + ir_node* c0 = sel; + ir_node* c1 = get_Mux_sel(f); + ir_node* t1 = get_Mux_true(f); + ir_node* f1 = get_Mux_false(f); + if (t == t1) { + /* Mux(cond0, x, Mux(cond1, x, y)) -> typical if (cond0 || cond1) x else y */ + ir_node* or_ = new_r_Or(block, c0, c1, mode_b); + DBG_OPT_ALGSIM0(oldn, f1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, f1, t1, mode); + } else if (t == f1) { + /* Mux(cond0, x, Mux(cond1, y, x)) */ + ir_node* not_c1 = new_r_Not(block, c1, mode_b); + ir_node* or_ = new_r_Or(block, c0, not_c1, mode_b); + DBG_OPT_ALGSIM0(oldn, t1, FS_OPT_MUX_COMBINE); + return new_r_Mux(block, or_, t1, f1, mode); } } - } - /* more normalization: Mux(sel, 0, 1) is simply a conv from the mode_b - * value to integer. */ - if (is_Const(t) && is_Const(f) && mode_is_int(mode)) { - ir_tarval *a = get_Const_tarval(t); - ir_tarval *b = get_Const_tarval(f); + /* note: after normalization, false can only happen on default */ + if (mode == mode_b) { + dbg_info *dbg = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); - if (tarval_is_one(a) && tarval_is_null(b)) { - ir_node *block = get_nodes_block(n); - ir_node *conv = new_r_Conv(block, sel, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); - return n; - } else if (tarval_is_null(a) && tarval_is_one(b)) { - ir_node *block = get_nodes_block(n); - ir_node *not_ = new_r_Not(block, sel, mode_b); - ir_node *conv = new_r_Conv(block, not_, mode); - n = conv; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_CONV); - return n; + if (is_Const(t)) { + ir_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 { + /* Muxb(sel, true, x) = Or(sel, x) */ + n = new_rd_Or(dbg, block, sel, f, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_OR_BOOL); + return n; + } + } + } else if (is_Const(f)) { + ir_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, block, sel, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_ORNOT_BOOL); + n = new_rd_Or(dbg, block, not_sel, t, mode_b); + return n; + } else { + /* Muxb(sel, x, false) = And(sel, x) */ + assert(tv_f == tarval_b_false); + n = new_rd_And(dbg, block, sel, t, mode_b); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_AND_BOOL); + return n; + } + } } } - if (is_Cmp(sel)) { + if (is_Cmp(sel) && mode_is_int(mode) && is_cmp_equality_zero(sel)) { + ir_relation relation = get_Cmp_relation(sel); ir_node *cmp_r = get_Cmp_right(sel); - if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - ir_node *block = get_nodes_block(n); - ir_node *cmp_l = get_Cmp_left(sel); - - if (mode_is_int(mode)) { - ir_relation relation = get_Cmp_relation(sel); - /* integer only */ - if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) { - /* Mux((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 (relation == ir_relation_less_greater) { - /* Mux((a & 2^C) != 0, 2^C, 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } else { - /* Mux((a & 2^C) == 0, 2^C, 0) */ - n = new_rd_Eor(get_irn_dbg_info(n), - block, cmp_l, t, mode); - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } - 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 (relation == ir_relation_less_greater) { - /* (a & (1 << n)) != 0, (1 << n), 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } else { - /* (a & (1 << n)) == 0, (1 << n), 0) */ - n = new_rd_Eor(get_irn_dbg_info(n), - block, cmp_l, t, mode); - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } - 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 (relation == ir_relation_less_greater) { - /* ((1 << n) & a) != 0, (1 << n), 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } else { - /* ((1 << n) & a) == 0, (1 << n), 0) */ - n = new_rd_Eor(get_irn_dbg_info(n), - block, cmp_l, t, mode); - DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); - } - return n; - } - } - } + ir_node *cmp_l = get_Cmp_left(sel); + ir_node *block = get_nodes_block(n); + + if (is_And(cmp_l) && f == cmp_r) { + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && is_single_bit(and_r)) { + if (relation == ir_relation_equal) { + /* Mux((a & (1<op->ops.transform_node != NULL) - n = n->op->ops.transform_node(n); - } while (oldn != n); - - return n; -} /* transform_node */ +} -/** - * Sets the default transform node operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops) +void firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -5710,6 +6355,7 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops CASE(Block); CASE(Call); CASE(Cmp); + CASE(Cond); CASE(Conv); CASE(End); CASE(Eor); @@ -5721,29 +6367,82 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops CASE(Phi); CASE(Proj); CASE(Rotl); - CASE(Sel); CASE(Shl); CASE(Shr); CASE(Shrs); CASE(Sub); + CASE(Switch); CASE(Sync); CASE_PROJ(Bound); CASE_PROJ(CopyB); CASE_PROJ(Store); - CASE_PROJ_EX(Cond); CASE_PROJ_EX(Div); CASE_PROJ_EX(Load); CASE_PROJ_EX(Mod); default: - /* leave NULL */; + break; } - - return ops; #undef CASE_PROJ_EX #undef CASE_PROJ #undef CASE -} /* firm_set_default_transform_node */ +} + +/** + * Tries several [inplace] [optimizing] transformations and returns an + * equivalent node. The difference to equivalent_node() is that these + * transformations _do_ generate new nodes, and thus the old node must + * not be freed even if the equivalent node isn't the old one. + */ +static ir_node *transform_node(ir_node *n) +{ + ir_node *old_n; + unsigned iro; +restart: + old_n = n; + iro = get_irn_opcode_(n); + /* constant expression evaluation / constant folding */ + if (get_opt_constant_folding()) { + /* neither constants nor Tuple values can be evaluated */ + if (iro != iro_Const && get_irn_mode(n) != mode_T) { + /* try to evaluate */ + ir_tarval *tv = computed_value(n); + if (tv != tarval_bad) { + /* evaluation was successful -- replace the node. */ + ir_graph *irg = get_irn_irg(n); + + n = new_r_Const(irg, tv); + + DBG_OPT_CSTEVAL(old_n, n); + return n; + } + } + } + + /* remove unnecessary nodes */ + if (get_opt_constant_folding() || + (iro == iro_Phi) || /* always optimize these nodes. */ + (iro == iro_Id) || /* ... */ + (iro == iro_Proj) || /* ... */ + (iro == iro_Block)) { /* Flags tested local. */ + n = equivalent_node(n); + if (n != old_n) + goto restart; + } + + /* Some more constant expression evaluation. */ + if (get_opt_algebraic_simplification() || + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ + if (n->op->ops.transform_node != NULL) { + n = n->op->ops.transform_node(n); + if (n != old_n) { + goto restart; + } + } + } + return n; +} /* **************** Common Subexpression Elimination **************** */ @@ -5751,6 +6450,14 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops in a graph. */ #define N_IR_NODES 512 +/** Compares two exception attributes */ +static int node_cmp_exception(const ir_node *a, const ir_node *b) +{ + const except_attr *ea = &a->attr.except; + const except_attr *eb = &b->attr.except; + return ea->pin_state != eb->pin_state; +} + /** Compares the attributes of two Const nodes. */ static int node_cmp_attr_Const(const ir_node *a, const ir_node *b) { @@ -5768,7 +6475,9 @@ static int node_cmp_attr_Alloc(const ir_node *a, const ir_node *b) { const alloc_attr *pa = &a->attr.alloc; const alloc_attr *pb = &b->attr.alloc; - return (pa->where != pb->where) || (pa->type != pb->type); + if (pa->where != pb->where || pa->type != pb->type) + return 1; + return node_cmp_exception(a, b); } /** Compares the attributes of two Free nodes. */ @@ -5793,8 +6502,9 @@ static int node_cmp_attr_Call(const ir_node *a, const ir_node *b) { const call_attr *pa = &a->attr.call; const call_attr *pb = &b->attr.call; - return (pa->type != pb->type) - || (pa->tail_call != pb->tail_call); + if (pa->type != pb->type) + return 1; + return node_cmp_exception(a, b); } /** Compares the attributes of two Sel nodes. */ @@ -5839,8 +6549,9 @@ static int node_cmp_attr_Load(const ir_node *a, const ir_node *b) /* do not CSE Loads with different alignment. Be conservative. */ if (get_Load_unaligned(a) != get_Load_unaligned(b)) return 1; - - return get_Load_mode(a) != get_Load_mode(b); + if (get_Load_mode(a) != get_Load_mode(b)) + return 1; + return node_cmp_exception(a, b); } /** Compares the attributes of two Store nodes. */ @@ -5849,31 +6560,34 @@ static int node_cmp_attr_Store(const ir_node *a, const ir_node *b) /* do not CSE Stores with different alignment. Be conservative. */ if (get_Store_unaligned(a) != get_Store_unaligned(b)) return 1; - /* NEVER do CSE on volatile Stores */ - return (get_Store_volatility(a) == volatility_is_volatile || - get_Store_volatility(b) == volatility_is_volatile); + if (get_Store_volatility(a) == volatility_is_volatile || + get_Store_volatility(b) == volatility_is_volatile) + return 1; + return node_cmp_exception(a, b); } -/** Compares two exception attributes */ -static int node_cmp_exception(const ir_node *a, const ir_node *b) +static int node_cmp_attr_CopyB(const ir_node *a, const ir_node *b) { - const except_attr *ea = &a->attr.except; - const except_attr *eb = &b->attr.except; + if (get_CopyB_type(a) != get_CopyB_type(b)) + return 1; - return ea->pin_state != eb->pin_state; + return node_cmp_exception(a, b); } -#define node_cmp_attr_Bound node_cmp_exception +static int node_cmp_attr_Bound(const ir_node *a, const ir_node *b) +{ + return node_cmp_exception(a, b); +} /** Compares the attributes of two Div nodes. */ static int node_cmp_attr_Div(const ir_node *a, const ir_node *b) { const div_attr *ma = &a->attr.div; const div_attr *mb = &b->attr.div; - return ma->exc.pin_state != mb->exc.pin_state || - ma->resmode != mb->resmode || - ma->no_remainder != mb->no_remainder; + if (ma->resmode != mb->resmode || ma->no_remainder != mb->no_remainder) + return 1; + return node_cmp_exception(a, b); } /** Compares the attributes of two Mod nodes. */ @@ -5881,8 +6595,9 @@ static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b) { const mod_attr *ma = &a->attr.mod; const mod_attr *mb = &b->attr.mod; - return ma->exc.pin_state != mb->exc.pin_state || - ma->resmode != mb->resmode; + if (ma->resmode != mb->resmode) + return 1; + return node_cmp_exception(a, b); } static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b) @@ -5903,14 +6618,18 @@ static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b) /** Compares the attributes of two Builtin nodes. */ static int node_cmp_attr_Builtin(const ir_node *a, const ir_node *b) { - /* no need to compare the type, equal kind means equal type */ - return get_Builtin_kind(a) != get_Builtin_kind(b); + if (get_Builtin_kind(a) != get_Builtin_kind(b)) + return 1; + if (get_Builtin_type(a) != get_Builtin_type(b)) + return 1; + return node_cmp_exception(a, b); } /** Compares the attributes of two ASM nodes. */ static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) { - int i, n; + size_t n; + size_t i; const ir_asm_constraint *ca; const ir_asm_constraint *cb; ident **cla, **clb; @@ -5953,7 +6672,8 @@ static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) if (cla[i] != clb[i]) return 1; } - return 0; + + return node_cmp_exception(a, b); } /** Compares the inexistent attributes of two Dummy nodes. */ @@ -5961,19 +6681,18 @@ static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b) { (void) a; (void) b; + /* Dummy nodes never equal by definition */ return 1; } -/** - * Set the default node attribute compare operation for an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) +static int node_cmp_attr_InstOf(const ir_node *a, const ir_node *b) +{ + if (get_InstOf_type(a) != get_InstOf_type(b)) + return 1; + return node_cmp_exception(a, b); +} + +void firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -5981,35 +6700,34 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) break switch (code) { - CASE(Const); - CASE(Proj); + CASE(ASM); CASE(Alloc); - CASE(Free); - CASE(SymConst); + CASE(Bound); + CASE(Builtin); CASE(Call); - CASE(Sel); - CASE(Phi); - CASE(Cmp); - CASE(Conv); CASE(Cast); - CASE(Load); - CASE(Store); + CASE(Cmp); CASE(Confirm); - CASE(ASM); + CASE(Const); + CASE(Conv); + CASE(CopyB); CASE(Div); - CASE(Mod); - CASE(Bound); - CASE(Builtin); CASE(Dummy); - /* FIXME CopyB */ + CASE(Free); + CASE(InstOf); + CASE(Load); + CASE(Mod); + CASE(Phi); + CASE(Proj); + CASE(Sel); + CASE(Store); + CASE(SymConst); default: /* leave NULL */ break; } - - return ops; #undef CASE -} /* firm_set_default_node_cmp_attr */ +} /* * Compare function for two nodes in the value table. Gets two @@ -6039,10 +6757,23 @@ 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] */ @@ -6064,7 +6795,7 @@ int identities_cmp(const void *elt, const void *key) return a->op->ops.node_cmp_attr(a, b); return 0; -} /* identities_cmp */ +} /* * Calculate a hash value of a node. @@ -6074,7 +6805,7 @@ int identities_cmp(const void *elt, const void *key) unsigned ir_node_hash(const ir_node *node) { return node->op->ops.hash(node); -} /* ir_node_hash */ +} void new_identities(ir_graph *irg) @@ -6082,13 +6813,13 @@ void new_identities(ir_graph *irg) if (irg->value_table != NULL) del_pset(irg->value_table); irg->value_table = new_pset(identities_cmp, N_IR_NODES); -} /* new_identities */ +} void del_identities(ir_graph *irg) { if (irg->value_table != NULL) del_pset(irg->value_table); -} /* del_identities */ +} /* Normalize a node by putting constants (and operands with larger * node index) on the right (operator side). */ @@ -6108,7 +6839,7 @@ void ir_normalize_node(ir_node *n) hook_normalize(n); } } -} /* ir_normalize_node */ +} /* * Return the canonical node computing the same value as n. @@ -6139,7 +6870,7 @@ ir_node *identify_remember(ir_node *n) } return nn; -} /* identify_remember */ +} /** * During construction we set the op_pin_state_pinned flag in the graph right @@ -6158,7 +6889,7 @@ static inline ir_node *identify_cons(ir_node *n) set_irg_pinned(irg, op_pin_state_floats); } return n; -} /* identify_cons */ +} /* Add a node to the identities value table. */ void add_identities(ir_node *node) @@ -6182,7 +6913,7 @@ void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) visit(node, env); } current_ir_graph = rem; -} /* visit_all_identities */ +} /** * These optimizations deallocate nodes from the obstack. @@ -6267,12 +6998,13 @@ ir_node *optimize_node(ir_node *n) free the node. */ iro = get_irn_opcode(n); if (get_opt_algebraic_simplification() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ + (iro == iro_Cond) || + (iro == iro_Proj)) { /* Flags tested local. */ n = transform_node(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)) { + if (get_opt_cse()) { ir_node *o = n; n = identify_remember(o); if (o != n) @@ -6280,7 +7012,7 @@ ir_node *optimize_node(ir_node *n) } return n; -} /* optimize_node */ +} /** @@ -6290,67 +7022,40 @@ ir_node *optimize_node(ir_node *n) */ ir_node *optimize_in_place_2(ir_node *n) { - ir_tarval *tv; - ir_node *oldn = n; - unsigned iro = get_irn_opcode(n); - if (!get_opt_optimize() && !is_Phi(n)) return n; - if (iro == iro_Deleted) + if (is_Deleted(n)) return n; - /* constant expression evaluation / constant folding */ - if (get_opt_constant_folding()) { - /* neither constants nor Tuple values can be evaluated */ - if (iro != iro_Const && get_irn_mode(n) != mode_T) { - /* try to evaluate */ - tv = computed_value(n); - if (tv != tarval_bad) { - /* evaluation was successful -- replace the node. */ - ir_graph *irg = get_irn_irg(n); - - n = new_r_Const(irg, tv); - - DBG_OPT_CSTEVAL(oldn, n); - return n; - } - } - } - - /* remove unnecessary nodes */ - if (get_opt_constant_folding() || - (iro == iro_Phi) || /* always optimize these nodes. */ - (iro == iro_Id) || /* ... */ - (iro == iro_Proj) || /* ... */ - (iro == iro_Block) ) /* Flags tested local. */ - n = equivalent_node(n); - /** common subexpression elimination **/ /* Checks whether n is already available. */ - /* The block input is used to distinguish different subexpressions. Right - now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common - subexpressions within a block. */ + /* The block input is used to distinguish different subexpressions. + * Right now all nodes are op_pin_state_pinned to blocks, i.e., the cse + * only finds common subexpressions within a block. */ if (get_opt_cse()) { ir_node *o = n; - n = identify_remember(o); - if (o != n) + n = identify_remember(n); + if (n != o) { DBG_OPT_CSE(o, n); + /* we have another existing node now, we do not optimize it here */ + return n; + } } - /* Some more constant expression evaluation. */ - iro = get_irn_opcode(n); - if (get_opt_constant_folding() || - (iro == iro_Cond) || - (iro == iro_Proj)) /* Flags tested local. */ - n = transform_node(n); + n = transform_node(n); /* Now we can verify the node, as it has no dead inputs any more. */ irn_verify(n); /* Now we have a legal, useful node. Enter it in hash table for cse. - Blocks should be unique anyways. (Except the successor of start: - is cse with the start block!) */ - if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { + * Blocks should be unique anyways. (Except the successor of start: + * is cse with the start block!) + * + * Note: This is only necessary because some of the optimisations + * operate in-place (set_XXX_bla, turn_into_tuple, ...) which is considered + * bad practice and should be fixed sometime. + */ + if (get_opt_cse()) { ir_node *o = n; n = identify_remember(o); if (o != n) @@ -6358,7 +7063,7 @@ ir_node *optimize_in_place_2(ir_node *n) } return n; -} /* optimize_in_place_2 */ +} /** * Wrapper for external use, set proper status bits after optimization. @@ -6374,9 +7079,9 @@ ir_node *optimize_in_place(ir_node *n) /* FIXME: Maybe we could also test whether optimizing the node can change the control graph. */ - set_irg_doms_inconsistent(irg); + clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE); return optimize_in_place_2(n); -} /* optimize_in_place */ +} /** * Calculate a hash value of a Const node. @@ -6389,7 +7094,7 @@ static unsigned hash_Const(const ir_node *node) h = HASH_PTR(node->attr.con.tarval); return h; -} /* hash_Const */ +} /** * Calculate a hash value of a SymConst node. @@ -6402,18 +7107,9 @@ static unsigned hash_SymConst(const ir_node *node) h = HASH_PTR(node->attr.symc.sym.type_p); return h; -} /* hash_SymConst */ +} -/** - * Set the default hash operation in an ir_op_ops. - * - * @param code the opcode for the default operation - * @param ops the operations initialized - * - * @return - * The operations. - */ -static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops) +void firm_set_default_hash(unsigned code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -6422,7 +7118,7 @@ static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops) /* hash function already set */ if (ops->hash != NULL) - return ops; + return; switch (code) { CASE(Const); @@ -6431,23 +7127,5 @@ static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops) /* use input/mode default hash if no function was given */ ops->hash = firm_default_hash; } - - return ops; #undef CASE } - -/* - * Sets the default operation for an ir_ops. - */ -ir_op_ops *firm_set_default_operations(unsigned code, ir_op_ops *ops) -{ - ops = firm_set_default_hash(code, ops); - ops = firm_set_default_computed_value(code, ops); - ops = firm_set_default_equivalent_node(code, ops); - ops = firm_set_default_transform_node(code, ops); - ops = firm_set_default_node_cmp_attr(code, ops); - ops = firm_set_default_get_type_attr(code, ops); - ops = firm_set_default_get_entity_attr(code, ops); - - return ops; -} /* firm_set_default_operations */