X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=880db762a9933476173a45d3aa1b25ae543ddc2e;hb=c386867ba33d1ad72fbd62ea3c095be220c92e1e;hp=a7b75d5a46e1f2f1926cec44dadda8947f1e7b10;hpb=09913c6466bf204f32f0fe00961011e2ba952157;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index a7b75d5a4..880db762a 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -56,6 +56,22 @@ /* 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 +100,7 @@ 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 +131,7 @@ 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 +147,14 @@ 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 +182,7 @@ 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 +204,7 @@ 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. @@ -197,12 +219,12 @@ static ir_tarval *computed_value_Borrow(const ir_node *n) ir_tarval *tb = value_of(b); if ((ta != tarval_bad) && (tb != tarval_bad)) { - return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m); + return tarval_cmp(ta, tb) == ir_relation_less ? get_mode_one(m) : get_mode_null(m); } else if (tarval_is_null(ta)) { return get_mode_null(m); } return tarval_bad; -} /* computed_value_Borrow */ +} /** * Return the value of an unary Minus. @@ -216,7 +238,7 @@ 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 +271,7 @@ static ir_tarval *computed_value_Mul(const ir_node *n) } } return tarval_bad; -} /* computed_value_Mul */ +} /** * Return the value of an And. @@ -264,12 +286,19 @@ 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 +313,18 @@ 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 +338,11 @@ 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 +351,7 @@ 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 +365,29 @@ 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 +403,12 @@ 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 +424,11 @@ 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 +445,7 @@ 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 +462,56 @@ 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 +531,7 @@ 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, @@ -435,113 +539,90 @@ static ir_tarval *computed_value_Mux(const ir_node *n) */ static ir_tarval *computed_value_Confirm(const ir_node *n) { - if (get_Confirm_cmp(n) == pn_Cmp_Eq) { + if (get_Confirm_relation(n) == ir_relation_equal) { ir_tarval *tv = value_of(get_Confirm_bound(n)); if (tv != tarval_bad) return tv; } return value_of(get_Confirm_value(n)); -} /* computed_value_Confirm */ +} /** - * Return the value of a Proj(Cmp). - * - * This performs a first step of unreachable code elimination. - * Proj can not be computed, but folding a Cmp above the Proj here is - * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which - * only 1 is used. - * There are several case where we can evaluate a Cmp node, see later. + * gives a (conservative) estimation of possible relation when comparing + * left+right */ -static ir_tarval *computed_value_Proj_Cmp(const ir_node *n) +ir_relation ir_get_possible_cmp_relations(const ir_node *left, + const ir_node *right) { - ir_node *cmp = get_Proj_pred(n); - ir_node *left = get_Cmp_left(cmp); - ir_node *right = get_Cmp_right(cmp); - pn_Cmp pn_cmp = get_Proj_pn_cmp(n); - ir_mode *mode = get_irn_mode(left); - ir_tarval *tv_l, *tv_r; - - /* - * BEWARE: a == a is NOT always True for floating Point values, as - * NaN != NaN is defined, so we must check this here. - */ - if (left == right && (!mode_is_float(mode) || pn_cmp == pn_Cmp_Lt || pn_cmp == pn_Cmp_Gt)) { - /* This is a trick with the bits used for encoding the Cmp - Proj numbers, the following statement is not the same: - return new_tarval_from_long(pn_cmp == pn_Cmp_Eq, mode_b) */ - return new_tarval_from_long(pn_cmp & pn_Cmp_Eq, mode_b); - } - tv_l = value_of(left); - tv_r = value_of(right); + ir_relation possible = ir_relation_true; + ir_tarval *tv_l = value_of(left); + ir_tarval *tv_r = value_of(right); + ir_mode *mode = get_irn_mode(left); + ir_tarval *min = mode == mode_b ? tarval_b_false : get_mode_min(mode); + ir_tarval *max = mode == mode_b ? tarval_b_true : get_mode_max(mode); + /* both values known - evaluate them */ if ((tv_l != tarval_bad) && (tv_r != tarval_bad)) { - /* - * The predecessors of Cmp are target values. We can evaluate - * the Cmp. - */ - pn_Cmp flags = tarval_cmp(tv_l, tv_r); - if (flags != pn_Cmp_False) { - return new_tarval_from_long (pn_cmp & flags, mode_b); - } - } else if (mode_is_int(mode)) { - /* for integer values, we can check against MIN/MAX */ - pn_Cmp cmp_result; - - if (tv_l == get_mode_min(mode)) { - /* MIN <=/> x. This results in true/false. */ - if (pn_cmp == pn_Cmp_Le) - return tarval_b_true; - else if (pn_cmp == pn_Cmp_Gt) - return tarval_b_false; - } else if (tv_r == get_mode_min(mode)) { - /* x >=/< MIN. This results in true/false. */ - if (pn_cmp == pn_Cmp_Ge) - return tarval_b_true; - else if (pn_cmp == pn_Cmp_Lt) - return tarval_b_false; - } else if (tv_l == get_mode_max(mode)) { - /* MAX >=/< x. This results in true/false. */ - if (pn_cmp == pn_Cmp_Ge) - return tarval_b_true; - else if (pn_cmp == pn_Cmp_Lt) - return tarval_b_false; - } else if (tv_r == get_mode_max(mode)) { - /* x <=/> MAX. This results in true/false. */ - if (pn_cmp == pn_Cmp_Le) - return tarval_b_true; - else if (pn_cmp == pn_Cmp_Gt) - return tarval_b_false; - } - - cmp_result = vrp_cmp(left, right); - if (cmp_result != pn_Cmp_False) { - if (cmp_result == pn_Cmp_Lg) { - if (pn_cmp == pn_Cmp_Eq) { - return tarval_b_false; - } else if (pn_cmp == pn_Cmp_Lg) { - return tarval_b_true; - } - } else { - return new_tarval_from_long(cmp_result & pn_cmp, mode_b); - } - } - } else if (mode_is_reference(mode)) { - /* pointer compare */ - ir_node *s_l = skip_Proj(left); - ir_node *s_r = skip_Proj(right); + possible = tarval_cmp(tv_l, tv_r); + /* we can return now, won't get any better */ + return possible; + } + /* a == a is never less or greater (but might be equal or unordered) */ + if (left == right) + possible &= ~ir_relation_less_greater; + /* unordered results only happen for float compares */ + if (!mode_is_float(mode)) + possible &= ~ir_relation_unordered; + /* values can never be less than the least representable number or + * greater than the greatest representable number */ + if (tv_l == min) + possible &= ~ir_relation_greater; + if (tv_l == max) + possible &= ~ir_relation_less; + if (tv_r == max) + possible &= ~ir_relation_greater; + if (tv_r == min) + possible &= ~ir_relation_less; + /* maybe vrp can tell us more */ + possible &= vrp_cmp(left, right); + /* Alloc nodes never return null (but throw an exception) */ + if (is_Alloc(left) && tarval_is_null(tv_r)) + possible &= ~ir_relation_equal; + + return possible; +} - if ((is_Alloc(s_l) && tarval_is_null(tv_r)) || - (tarval_is_null(tv_l) && is_Alloc(s_r))) { - /* - * The predecessors are Allocs and (void*)(0) constants. In Firm Allocs never - * return NULL, they raise an exception. Therefore we can predict - * the Cmp result. - */ - return new_tarval_from_long(pn_cmp & pn_Cmp_Lg, mode_b); - } - } - return computed_value_Cmp_Confirm(cmp, left, right, pn_cmp); -} /* computed_value_Proj_Cmp */ +static ir_tarval *compute_cmp(const ir_node *cmp) +{ + ir_node *left = get_Cmp_left(cmp); + ir_node *right = get_Cmp_right(cmp); + ir_relation possible = ir_get_possible_cmp_relations(left, right); + ir_relation relation = get_Cmp_relation(cmp); + + /* if none of the requested relations is possible, return false */ + if ((possible & relation) == ir_relation_false) + return tarval_b_false; + /* if possible relations are a subset of the requested ones return true */ + if ((possible & ~relation) == ir_relation_false) + return tarval_b_true; + + 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. @@ -565,7 +646,7 @@ 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. @@ -582,7 +663,7 @@ 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). @@ -594,7 +675,7 @@ 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). @@ -608,7 +689,7 @@ 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. @@ -620,7 +701,7 @@ 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. @@ -631,13 +712,13 @@ 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) == pn_Cmp_Eq) { + 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. @@ -660,29 +741,29 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops break switch (code) { - CASE(Const); - CASE(SymConst); CASE(Add); - CASE(Sub); - CASE(Carry); + CASE(And); CASE(Borrow); + CASE(Carry); + CASE(Cmp); + CASE(Confirm); + CASE(Const); + CASE(Conv); + CASE(Eor); CASE(Minus); CASE(Mul); - CASE(And); - CASE(Or); - CASE(Eor); + CASE(Mux); CASE(Not); + CASE(Or); + CASE(Proj); + CASE(Rotl); CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Rotl); - CASE(Conv); - CASE(Mux); - CASE(Confirm); - CASE_PROJ(Cmp); + CASE(Sub); + CASE(SymConst); CASE_PROJ(Div); CASE_PROJ(Mod); - CASE(Proj); default: /* leave NULL */ break; @@ -691,142 +772,7 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops 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. - * - * Note, that blocks are NEVER turned into Bad's, instead - * the dead_block flag is set. So, never test for is_Bad(block), - * always use is_dead_Block(block). - */ -static ir_node *equivalent_node_Block(ir_node *n) -{ - ir_node *oldn = n; - int n_preds; - ir_graph *irg; - - /* don't optimize dead or labeled blocks */ - if (is_Block_dead(n) || has_Block_entity(n)) - return n; - - n_preds = get_Block_n_cfgpreds(n); - - /* The Block constructor does not call optimize, but mature_immBlock() - calls the optimization. */ - assert(get_Block_matured(n)); - - irg = get_irn_irg(n); - - /* Straightening: a single entry Block following a single exit Block - can be merged, if it is not the Start block. */ - /* !!! Beware, all Phi-nodes of n must have been optimized away. - This should be true, as the block is matured before optimize is called. - But what about Phi-cycles with the Phi0/Id that could not be resolved? - Remaining Phi nodes are just Ids. */ - if (n_preds == 1) { - ir_node *pred = skip_Proj(get_Block_cfgpred(n, 0)); - - if (is_Jmp(pred)) { - ir_node *predblock = get_nodes_block(pred); - if (predblock == oldn) { - /* Jmp jumps into the block it is in -- deal self cycle. */ - n = set_Block_dead(n); - DBG_OPT_DEAD_BLOCK(oldn, n); - } else { - n = predblock; - DBG_OPT_STG(oldn, n); - } - } else if (is_Cond(pred)) { - ir_node *predblock = get_nodes_block(pred); - if (predblock == oldn) { - /* Jmp jumps into the block it is in -- deal self cycle. */ - n = set_Block_dead(n); - DBG_OPT_DEAD_BLOCK(oldn, n); - } - } - } 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); - } - } - } else if (get_opt_unreachable_code() && - (n != get_irg_start_block(irg)) && - (n != get_irg_end_block(irg))) { - int i; - - /* If all inputs are dead, this block is dead too, except if it is - the start or end block. This is one step of unreachable code - elimination */ - for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) { - ir_node *pred = get_Block_cfgpred(n, i); - ir_node *pred_blk; - - if (is_Bad(pred)) continue; - pred_blk = get_nodes_block(skip_Proj(pred)); - - if (is_Block_dead(pred_blk)) continue; - - if (pred_blk != n) { - /* really found a living input */ - break; - } - } - if (i < 0) { - n = set_Block_dead(n); - DBG_OPT_DEAD_BLOCK(oldn, n); - } - } - - return n; -} /* equivalent_node_Block */ - -/** - * Returns a equivalent node for a Jmp, a Bad :-) - * Of course this only happens if the Block of the Jmp is dead. - */ -static ir_node *equivalent_node_Jmp(ir_node *n) -{ - ir_node *oldn = n; - - /* unreachable code elimination */ - if (is_Block_dead(get_nodes_block(n))) { - ir_graph *irg = get_irn_irg(n); - n = get_irg_bad(irg); - DBG_OPT_DEAD_BLOCK(oldn, n); - } - return n; -} /* equivalent_node_Jmp */ - -/** Raise is handled in the same way as Jmp. */ -#define equivalent_node_Raise equivalent_node_Jmp - - -/* 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, @@ -865,7 +811,7 @@ static ir_node *equivalent_node_neutral_zero(ir_node *n) } return n; -} /* equivalent_node_neutral_zero */ +} /** * Eor is commutative and has neutral 0. @@ -882,9 +828,9 @@ 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 */ @@ -898,9 +844,9 @@ 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 */ @@ -969,7 +915,7 @@ 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, @@ -989,7 +935,7 @@ 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 @@ -1032,11 +978,11 @@ 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. @@ -1054,7 +1000,7 @@ 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 @@ -1093,7 +1039,7 @@ static ir_node *equivalent_node_Mul(ir_node *n) } } return n; -} /* equivalent_node_Mul */ +} /** * Use algebraic simplification a | a = a | 0 = 0 | a = a. @@ -1107,11 +1053,11 @@ static ir_node *equivalent_node_Or(ir_node *n) ir_tarval *tv; if (a == b) { - n = a; /* Or has it's own neutral element */ + n = a; /* idempotence */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR); return n; } - /* constants are cormalized to right, check this site first */ + /* constants are normalized to right, check this side first */ tv = value_of(b); if (tarval_is_null(tv)) { n = a; @@ -1126,7 +1072,7 @@ 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. @@ -1140,11 +1086,11 @@ static ir_node *equivalent_node_And(ir_node *n) ir_tarval *tv; if (a == b) { - n = a; /* And has it's own neutral element */ + n = a; /* idempotence */ DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND); return n; } - /* constants are normalized to right, check this site first */ + /* constants are normalized to right, check this side first */ tv = value_of(b); if (tarval_is_all_one(tv)) { n = a; @@ -1157,8 +1103,13 @@ 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; @@ -1172,24 +1123,22 @@ 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: @@ -1300,7 +1249,7 @@ 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; @@ -1320,7 +1269,7 @@ restart: } } return n; -} /* equivalent_node_Conv */ +} /** * - fold Phi-nodes, iff they have only one predecessor except @@ -1331,7 +1280,6 @@ 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() && @@ -1340,52 +1288,25 @@ static ir_node *equivalent_node_Phi(ir_node *n) n_preds = get_Phi_n_preds(n); - block = get_nodes_block(n); - /* Control dead */ - if (is_Block_dead(block)) { - ir_graph *irg = get_irn_irg(n); - return get_irg_bad(irg); - } - - if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */ + /* Phi of dead Region without predecessors. */ + if (n_preds == 0) + return n; /* Find first non-self-referencing input */ for (i = 0; i < n_preds; ++i) { first_val = get_Phi_pred(n, i); - if ( (first_val != n) /* not self pointer */ -#if 0 - /* BEWARE: when the if is changed to 1, Phi's will ignore it's Bad - * predecessors. Then, Phi nodes in dead code might be removed, causing - * nodes pointing to themself (Add's for instance). - * This is really bad and causes endless recursions in several - * code pathes, so we do NOT optimize such a code. - * This is not that bad as it sounds, optimize_cf() removes bad control flow - * (and bad Phi predecessors), so live code is optimized later. - */ - && (! is_Bad(get_Block_cfgpred(block, i))) -#endif - ) { /* value not dead */ - break; /* then found first value. */ + /* not self pointer */ + if (first_val != n) { + /* then found first value. */ + break; } } - if (i >= n_preds) { - ir_graph *irg = get_irn_irg(n); - /* A totally Bad or self-referencing Phi (we didn't break the above loop) */ - return get_irg_bad(irg); - } - /* search for rest of inputs, determine if any of these are non-self-referencing */ while (++i < n_preds) { ir_node *scnd_val = get_Phi_pred(n, i); - if ( (scnd_val != n) - && (scnd_val != first_val) -#if 0 - /* see above */ - && (! is_Bad(get_Block_cfgpred(block, i))) -#endif - ) { + if (scnd_val != n && scnd_val != first_val) { break; } } @@ -1396,50 +1317,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) DBG_OPT_PHI(oldn, n); } return n; -} /* equivalent_node_Phi */ - -/** - * Several optimizations: - * - fold Sync-nodes, iff they have only one predecessor except - * themselves. - */ -static ir_node *equivalent_node_Sync(ir_node *n) -{ - int arity = get_Sync_n_preds(n); - int i; - - for (i = 0; i < arity;) { - ir_node *pred = get_Sync_pred(n, i); - int j; - - /* Remove Bad predecessors */ - if (is_Bad(pred)) { - del_Sync_n(n, i); - --arity; - continue; - } - - /* Remove duplicate predecessors */ - for (j = 0;; ++j) { - if (j >= i) { - ++i; - break; - } - if (get_Sync_pred(n, j) == pred) { - del_Sync_n(n, i); - --arity; - break; - } - } - } - - if (arity == 0) { - ir_graph *irg = get_irn_irg(n); - return get_irg_bad(irg); - } - if (arity == 1) return get_Sync_pred(n, 0); - return n; -} /* equivalent_node_Sync */ +} /** * Optimize Proj(Tuple). @@ -1454,7 +1332,7 @@ 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. @@ -1486,7 +1364,7 @@ static ir_node *equivalent_node_Proj_Div(ir_node *proj) } } return proj; -} /* equivalent_node_Proj_Div */ +} /** * Optimize CopyB(mem, x, x) into a Nop. @@ -1505,17 +1383,10 @@ static ir_node *equivalent_node_Proj_CopyB(ir_node *proj) proj = get_CopyB_mem(copyb); DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); break; - - case pn_CopyB_X_except: { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(irg); - break; - } } } return proj; -} /* equivalent_node_Proj_CopyB */ +} /** * Optimize Bounds(idx, idx, upper) into idx. @@ -1555,12 +1426,6 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) DBG_OPT_EXC_REM(proj); proj = get_Bound_mem(bound); break; - case pn_Bound_X_except: { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(irg); - break; - } case pn_Bound_res: proj = idx; DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NOP); @@ -1571,77 +1436,19 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) } } return proj; -} /* equivalent_node_Proj_Bound */ +} /** - * Optimize an Exception Proj(Load) with a non-null address. - */ -static ir_node *equivalent_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); - - /* get the Load address */ - const ir_node *addr = get_Load_ptr(load); - const ir_node *confirm; - - if (value_not_null(addr, &confirm)) { - if (get_Proj_proj(proj) == pn_Load_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); - } - } - } - } - return proj; -} /* equivalent_node_Proj_Load */ - -/** - * Optimize an Exception Proj(Store) with a non-null address. - */ -static ir_node *equivalent_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); - - /* get the load/store address */ - const ir_node *addr = get_Store_ptr(store); - const ir_node *confirm; - - if (value_not_null(addr, &confirm)) { - if (get_Proj_proj(proj) == pn_Store_X_except) { - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); - } - } - } - } - return proj; -} /* equivalent_node_Proj_Store */ - -/** - * Does all optimizations on nodes that must be done on it's Proj's - * because of creating new nodes. + * Does all optimizations on nodes that must be done on its Projs + * because of creating new nodes. */ static ir_node *equivalent_node_Proj(ir_node *proj) { ir_node *n = get_Proj_pred(proj); - - if (get_irn_mode(proj) == mode_X) { - if (is_Block_dead(get_nodes_block(n))) { - /* Remove dead control flow -- early gigo(). */ - ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); - } - } if (n->op->ops.equivalent_node_Proj) return n->op->ops.equivalent_node_Proj(proj); return proj; -} /* equivalent_node_Proj */ +} /** * Remove Id's. @@ -1656,7 +1463,7 @@ static ir_node *equivalent_node_Id(ir_node *n) DBG_OPT_ID(oldn, n); return n; -} /* equivalent_node_Id */ +} /** * Optimize a Mux. @@ -1667,6 +1474,12 @@ 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); @@ -1701,65 +1514,65 @@ static ir_node *equivalent_node_Mux(ir_node *n) DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ); return n; } - if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) { - ir_node *cmp = get_Proj_pred(sel); - long proj_nr = get_Proj_proj(sel); - ir_node *f = get_Mux_false(n); - ir_node *t = get_Mux_true(n); + if (is_Cmp(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) { + ir_relation relation = get_Cmp_relation(sel); + ir_node *f = get_Mux_false(n); + ir_node *t = get_Mux_true(n); /* * Note further that these optimization work even for floating point * with NaN's because -NaN == NaN. * However, if +0 and -0 is handled differently, we cannot use the first one. */ - if (is_Cmp(cmp)) { - ir_node *const cmp_l = get_Cmp_left(cmp); - ir_node *const cmp_r = get_Cmp_right(cmp); - - switch (proj_nr) { - case pn_Cmp_Eq: - if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */ - (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */ - n = f; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - return n; - } - break; + ir_node *const cmp_l = get_Cmp_left(sel); + ir_node *const cmp_r = get_Cmp_right(sel); + + switch (relation) { + case ir_relation_equal: + if ((cmp_l == t && cmp_r == f) || /* Mux(t == f, t, f) -> f */ + (cmp_l == f && cmp_r == t)) { /* Mux(f == t, t, f) -> f */ + n = f; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + return n; + } + break; - case pn_Cmp_Lg: - case pn_Cmp_Ne: - if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */ - (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */ - n = t; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - return n; - } - break; + case ir_relation_less_greater: + case ir_relation_unordered_less_greater: + if ((cmp_l == t && cmp_r == f) || /* Mux(t != f, t, f) -> t */ + (cmp_l == f && cmp_r == t)) { /* Mux(f != t, t, f) -> t */ + n = t; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + return n; } + break; + default: + break; + } - /* - * Note: normalization puts the constant on the right side, - * so we check only one case. - */ - if (cmp_l == t && tarval_is_null(value_of(cmp_r))) { - /* Mux(t CMP 0, X, t) */ - if (is_Minus(f) && get_Minus_op(f) == t) { - /* Mux(t CMP 0, -t, t) */ - if (proj_nr == pn_Cmp_Eq) { - /* Mux(t == 0, -t, t) ==> -t */ - n = f; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) { - /* Mux(t != 0, -t, t) ==> t */ - n = t; - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); - } + /* + * Note: normalization puts the constant on the right side, + * so we check only one case. + */ + if (cmp_l == t && tarval_is_null(value_of(cmp_r))) { + /* Mux(t CMP 0, X, t) */ + if (is_Minus(f) && get_Minus_op(f) == t) { + /* Mux(t CMP 0, -t, t) */ + if (relation == ir_relation_equal) { + /* Mux(t == 0, -t, t) ==> -t */ + n = f; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); + } else if (relation == ir_relation_less_greater || relation == ir_relation_unordered_less_greater) { + /* Mux(t != 0, -t, t) ==> t */ + n = t; + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM); } } } } + return n; -} /* equivalent_node_Mux */ +} /** * Remove Confirm nodes if setting is on. @@ -1767,17 +1580,16 @@ static ir_node *equivalent_node_Mux(ir_node *n) */ static ir_node *equivalent_node_Confirm(ir_node *n) { - ir_node *pred = get_Confirm_value(n); - pn_Cmp pnc = get_Confirm_cmp(n); + ir_node *pred = get_Confirm_value(n); + ir_relation relation = get_Confirm_relation(n); - while (is_Confirm(pred) && pnc == get_Confirm_cmp(pred)) { + while (is_Confirm(pred) && relation == get_Confirm_relation(pred)) { /* * rare case: two identical Confirms one after another, * replace the second one with the first. */ n = pred; pred = get_Confirm_value(n); - pnc = get_Confirm_cmp(n); } return n; } @@ -1794,7 +1606,7 @@ 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. @@ -1817,9 +1629,6 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op break switch (code) { - CASE(Block); - CASE(Jmp); - CASE(Raise); CASE(Eor); CASE(Add); CASE(Shl); @@ -1834,13 +1643,10 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op CASE(And); CASE(Conv); CASE(Phi); - CASE(Sync); CASE_PROJ(Tuple); CASE_PROJ(Div); CASE_PROJ(CopyB); CASE_PROJ(Bound); - CASE_PROJ(Load); - CASE_PROJ(Store); CASE(Proj); CASE(Id); CASE(Mux); @@ -1853,7 +1659,7 @@ static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *op return ops; #undef CASE #undef CASE_PROJ -} /* firm_set_default_equivalent_node */ +} /** * Returns non-zero if a node is a Phi node @@ -1870,7 +1676,7 @@ 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); @@ -1948,7 +1754,7 @@ 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. @@ -1993,7 +1799,7 @@ 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). @@ -2031,7 +1837,7 @@ 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). @@ -2066,7 +1872,7 @@ 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 @@ -2143,7 +1949,7 @@ static ir_node *transform_node_AddSub(ir_node *n) } return n; -} /* transform_node_AddSub */ +} #define HANDLE_BINOP_PHI(eval, a, b, c, mode) \ do { \ @@ -2180,162 +1986,809 @@ 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; - - n = transform_node_AddSub(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); - } - } - - HANDLE_BINOP_PHI((eval_func) tarval_add, a, b, c, mode); + ir_tarval *tv = get_mode_null(mode); + ir_node *cnst = new_r_Const(irg, tv); - /* 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; - } + return cnst; +} - 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); +static bool is_shiftop(const ir_node *n) +{ + return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(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; - } - if (is_Minus(a)) { - n = new_rd_Sub( - get_irn_dbg_info(n), - get_nodes_block(n), - b, - get_Minus_op(a), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); - return n; - } - if (is_Minus(b)) { - n = new_rd_Sub( - get_irn_dbg_info(n), - get_nodes_block(n), - a, - get_Minus_op(b), - mode); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); - return n; - } - if (get_mode_arithmetic(mode) == irma_twos_complement) { - /* Here we rely on constants be on the RIGHT side */ - if (is_Not(a)) { - ir_node *op = get_Not_op(a); +/* the order of the values is important! */ +typedef enum const_class { + const_const = 0, + const_like = 1, + const_other = 2 +} const_class; - if (is_Const(b) && is_Const_one(b)) { - /* ~x + 1 = -x */ - ir_node *blk = get_nodes_block(n); - n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode); - 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); +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 (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; - } - } - } - } +/** + * 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)); +} - a_vrp = vrp_get_info(a); - b_vrp = vrp_get_info(b); +static bool is_cmp_unequal(const ir_node *node) +{ + ir_relation relation = get_Cmp_relation(node); + ir_node *left = get_Cmp_left(node); + ir_node *right = get_Cmp_right(node); + ir_mode *mode = get_irn_mode(left); - if (a_vrp && b_vrp) { - ir_tarval *c = tarval_and(a_vrp->bits_not_set, b_vrp->bits_not_set); + if (relation == ir_relation_less_greater) + return true; - 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); - } - } - return n; -} /* transform_node_Add */ + if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right)) + return relation == ir_relation_greater; + return false; +} /** - * returns -cnst or NULL if impossible + * returns true for Cmp(x == 0) or Cmp(x != 0) */ -static ir_node *const_negate(ir_node *cnst) -{ - ir_tarval *tv = tarval_neg(get_Const_tarval(cnst)); - dbg_info *dbgi = get_irn_dbg_info(cnst); - ir_graph *irg = get_irn_irg(cnst); - if (tv == tarval_bad) return NULL; - return new_rd_Const(dbgi, irg, tv); +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); } /** - * Do the AddSub optimization, then Transform - * Constant folding on Phi - * Sub(0,a) -> Minus(a) - * Sub(Mul(a, x), a) -> Mul(a, x-1) - * Sub(Sub(x, y), b) -> Sub(x, Add(y,b)) - * Sub(Add(a, x), x) -> a - * Sub(x, Add(x, a)) -> -a - * Sub(x, Const) -> Add(x, -Const) + * 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_Sub(ir_node *n) +static ir_node *transform_node_Or_bf_store(ir_node *irn_or) { - ir_mode *mode; - ir_node *oldn = n; - ir_node *a, *b, *c; - - n = transform_node_AddSub(n); + 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); - a = get_Sub_left(n); - b = get_Sub_right(n); + ir_tarval *tv1, *tv2, *tv3, *tv4, *tv; - mode = get_irn_mode(n); + 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; +} + +/** + * 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; + + 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; + + /* 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 = 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); + } + } + + 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; + } + if (is_Minus(a)) { + n = new_rd_Sub( + get_irn_dbg_info(n), + get_nodes_block(n), + b, + get_Minus_op(a), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + return n; + } + if (is_Minus(b)) { + n = new_rd_Sub( + get_irn_dbg_info(n), + get_nodes_block(n), + a, + get_Minus_op(b), + mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B); + return n; + } + if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* Here we rely on constants be on the RIGHT side */ + if (is_Not(a)) { + ir_node *op = get_Not_op(a); + + if (is_Const(b) && is_Const_one(b)) { + /* ~x + 1 = -x */ + ir_node *blk = get_nodes_block(n); + n = new_rd_Minus(get_irn_dbg_info(n), blk, op, mode); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); + return n; + } + } + } + } + + 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; +} + +/** + * returns -cnst or NULL if impossible + */ +static ir_node *const_negate(ir_node *cnst) +{ + ir_tarval *tv = tarval_neg(get_Const_tarval(cnst)); + dbg_info *dbgi = get_irn_dbg_info(cnst); + ir_graph *irg = get_irn_irg(cnst); + if (tv == tarval_bad) return NULL; + return new_rd_Const(dbgi, irg, tv); +} + +/** + * Do the AddSub optimization, then Transform + * Constant folding on Phi + * Sub(0,a) -> Minus(a) + * Sub(Mul(a, x), a) -> Mul(a, x-1) + * Sub(Sub(x, y), b) -> Sub(x, Add(y,b)) + * Sub(Add(a, x), x) -> a + * Sub(x, Add(x, a)) -> -a + * Sub(x, Const) -> Add(x, -Const) + */ +static ir_node *transform_node_Sub(ir_node *n) +{ + ir_mode *mode; + ir_node *oldn = n; + ir_node *a, *b, *c; + + n = transform_node_AddSub(n); + + a = get_Sub_left(n); + b = get_Sub_right(n); + + mode = get_irn_mode(n); if (mode_is_int(mode)) { ir_mode *lmode = get_irn_mode(a); @@ -2425,6 +2878,7 @@ 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)) { @@ -2443,6 +2897,7 @@ restart: return n; } } +#endif } /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ @@ -2455,58 +2910,54 @@ 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)) { @@ -2604,8 +3055,8 @@ 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)); @@ -2618,9 +3069,27 @@ 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. @@ -2761,10 +3230,11 @@ 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); @@ -2774,7 +3244,7 @@ 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); @@ -2784,7 +3254,7 @@ static ir_node *transform_node_Mul(ir_node *n) } } return arch_dep_replace_mul_with_shifts(n); -} /* transform_node_Mul */ +} /** * Transform a Div Node. @@ -2889,14 +3359,14 @@ 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, get_irg_bad(irg)); + 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. @@ -2981,40 +3451,14 @@ 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, get_irg_bad(irg)); + 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 */ - -/** - * Optimize -a CMP -b into b CMP a. - * This works only for for modes where unary Minus - * cannot Overflow. - * Note that two-complement integers can Overflow - * so it will NOT work. - * - * For == and != can be handled in Proj(Cmp) - */ -static ir_node *transform_node_Cmp(ir_node *n) -{ - ir_node *oldn = n; - ir_node *left = get_Cmp_left(n); - ir_node *right = get_Cmp_right(n); - - if (is_Minus(left) && is_Minus(right) && - !mode_overflow_on_unary_Minus(get_irn_mode(left))) { - ir_node *const new_left = get_Minus_op(right); - ir_node *const new_right = get_Minus_op(left); - n = new_rd_Cmp(get_irn_dbg_info(n), get_nodes_block(n), new_left, new_right); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CMP_OP_OP); - } - return n; -} /* transform_node_Cmp */ - +} /** * Transform a Cond node. @@ -3024,146 +3468,172 @@ static ir_node *transform_node_Cmp(ir_node *n) */ 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, get_irg_bad(irg)); + set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X)); set_Tuple_pred(n, pn_Cond_true, jmp); } else { set_Tuple_pred(n, pn_Cond_false, jmp); - set_Tuple_pred(n, pn_Cond_true, get_irg_bad(irg)); + set_Tuple_pred(n, pn_Cond_true, new_r_Bad(irg, mode_X)); } /* 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); +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; + } + } + for (o = 0; o < n_outs; ++o) { + if (o == (unsigned)jmp_pn) { + in[o] = new_rd_Jmp(dbgi, block); + } else { + in[o] = bad; + } + } + return new_r_Tuple(block, (int)n_outs, in); + } + return 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 + * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is good for things like x-(x&y) esp. in 186.crafty. */ -static ir_node *transform_bitwise_distributive(ir_node *n, - recursive_transform trans_func) +static ir_node *transform_node_shift_bitop(ir_node *n) { - 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); + ir_graph *irg = get_irn_irg(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *left; + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_shift; - if (op != get_irn_op(b)) + if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) 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); + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(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)); + if (!is_Const(right)) + return n; - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV); - 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; - if (op == op_Eor) { - /* nothing to gain here */ + /* 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; } - 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; + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; - 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; - } + bitop_left = get_binop_left(left); - if (c != NULL) { - /* (a sop c) & (b sop c) => (a & b) sop c */ - ir_node *blk = get_nodes_block(n); + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); - 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); + assert(get_tarval_mode(tv1) == mode); - if (op_root == op_Eor && op == op_Or) { - dbg_info *dbgi = get_irn_dbg_info(n); - ir_mode *mode = get_irn_mode(c); + 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); + } - 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); - } + assert(get_tarval_mode(tv_shift) == mode); + irg = get_irn_irg(n); + new_const = new_r_Const(irg, tv_shift); - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); - return n; - } + 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 n; + return new_bitop; } /** @@ -3175,34 +3645,79 @@ 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; + + 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); + ir_relation a_relation = get_Cmp_relation(a); + ir_relation b_relation = get_Cmp_relation(b); + /* we can combine the relations of two compares with the same + * operands */ + 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 new_relation = a_relation & b_relation; + return new_rd_Cmp(dbgi, block, a_left, a_right, new_relation); + } + /* 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); - /* we can evaluate 2 Projs of the same Cmp */ - if (mode == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if (pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - pn_Cmp pn_a = get_Proj_pn_cmp(a); - pn_Cmp pn_b = get_Proj_pn_cmp(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a & pn_b; - - return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc); - } - } - 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); @@ -3213,15 +3728,27 @@ 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); @@ -3232,9 +3759,9 @@ 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 */ @@ -3257,9 +3784,9 @@ 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 */ @@ -3295,109 +3822,29 @@ 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)) == pn_Cmp_Eq)) { - - 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)) == pn_Cmp_Eq)) { - 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); - - HANDLE_BINOP_PHI((eval_func) tarval_eor, a, b, c, mode); - - /* we can evaluate 2 Projs of the same Cmp */ - if (mode == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if (pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - pn_Cmp pn_a = get_Proj_pn_cmp(a); - pn_Cmp pn_b = get_Proj_pn_cmp(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a ^ pn_b; - - return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc); - } - } - - /* 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. */ @@ -3410,23 +3857,22 @@ static ir_node *transform_node_Not(ir_node *n) HANDLE_UNOP_PHI(tarval_not,a,c); /* check for a boolean Not */ - if (mode == mode_b && is_Proj(a)) { - ir_node *a_pred = get_Proj_pred(a); - if (is_Cmp(a_pred)) { - /* We negate a Cmp. The Cmp has the negated result anyways! */ - n = new_r_Proj(get_Proj_pred(a), - mode_b, get_negated_pnc(get_Proj_proj(a), mode_b)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); - return n; - } + if (is_Cmp(a)) { + dbg_info *dbgi = get_irn_dbg_info(a); + ir_node *block = get_nodes_block(a); + ir_relation relation = get_Cmp_relation(a); + relation = get_negated_relation(relation); + n = new_rd_Cmp(dbgi, block, get_Cmp_left(a), get_Cmp_right(a), relation); + DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP); + return 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; @@ -3440,11 +3886,11 @@ 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); @@ -3452,7 +3898,7 @@ static ir_node *transform_node_Not(ir_node *n) } } return n; -} /* transform_node_Not */ +} /** * Transform a Minus. @@ -3548,7 +3994,7 @@ static ir_node *transform_node_Minus(ir_node *n) } return n; -} /* transform_node_Minus */ +} /** * Transform a Proj(Load) with a non-null address. @@ -3571,7 +4017,7 @@ static ir_node *transform_node_Proj_Load(ir_node *proj) if (get_Proj_proj(proj) == pn_Load_X_except) { ir_graph *irg = get_irn_irg(proj); DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } else { ir_node *blk = get_nodes_block(load); return new_r_Jmp(blk); @@ -3580,7 +4026,7 @@ static ir_node *transform_node_Proj_Load(ir_node *proj) } } return proj; -} /* transform_node_Proj_Load */ +} /** * Transform a Proj(Store) with a non-null address. @@ -3603,7 +4049,7 @@ static ir_node *transform_node_Proj_Store(ir_node *proj) if (get_Proj_proj(proj) == pn_Store_X_except) { ir_graph *irg = get_irn_irg(proj); DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } else { ir_node *blk = get_nodes_block(store); return new_r_Jmp(blk); @@ -3612,7 +4058,7 @@ static ir_node *transform_node_Proj_Store(ir_node *proj) } } return proj; -} /* transform_node_Proj_Store */ +} /** * Transform a Proj(Div) with a non-zero value. @@ -3645,7 +4091,7 @@ static ir_node *transform_node_Proj_Div(ir_node *proj) ir_graph *irg = get_irn_irg(proj); /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } case pn_Div_M: { @@ -3665,7 +4111,7 @@ 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. @@ -3700,7 +4146,7 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) ir_graph *irg = get_irn_irg(proj); /* we found an exception handler, remove it */ DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } case pn_Mod_M: { @@ -3729,87 +4175,7 @@ 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) -{ - if (get_opt_unreachable_code()) { - ir_node *n = get_Proj_pred(proj); - ir_node *b = get_Cond_selector(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 get_irg_bad(irg); - } - } - } 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) { - pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & pn_Cmp_Gt) == cmp_result && (cmp_result2 - & pn_Cmp_Lt) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); - } - } else if (b_vrp->range_type == VRP_ANTIRANGE) { - pn_Cmp cmp_result = tarval_cmp(b_vrp->range_bottom, tp); - pn_Cmp cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - - if ((cmp_result & pn_Cmp_Le) == cmp_result && (cmp_result2 - & pn_Cmp_Ge) == cmp_result2) { - ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); - } - } - - if (!(tarval_cmp( - tarval_and( b_vrp->bits_set, tp), - b_vrp->bits_set - ) == pn_Cmp_Eq)) { - ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); - } - - if (!(tarval_cmp( - tarval_and( - tarval_not(tp), - tarval_not(b_vrp->bits_not_set)), - tarval_not(b_vrp->bits_not_set)) - == pn_Cmp_Eq)) { - ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); - } - - - } - } - } - } - return proj; -} /* transform_node_Proj_Cond */ +} /** * return true if the operation returns a value with exactly 1 bit set @@ -3834,86 +4200,114 @@ static bool is_single_bit(const ir_node *node) } /** - * Create a 0 constant of given mode. + * 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 *create_zero_const(ir_graph *irg, ir_mode *mode) +static ir_node *flips_bit(const ir_node *node, ir_tarval *tv) { - ir_tarval *tv = get_mode_null(mode); - ir_node *cnst = new_r_Const(irg, tv); - - return cnst; + 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. */ -static ir_node *transform_node_Proj_Cmp(ir_node *proj) -{ - ir_node *n = get_Proj_pred(proj); - ir_node *left = get_Cmp_left(n); - ir_node *right = get_Cmp_right(n); - ir_tarval *tv = NULL; - int changed = 0; - ir_mode *mode = get_irn_mode(left); - long proj_nr = get_Proj_proj(proj); - - /* we can evaluate some cases directly */ - switch (proj_nr) { - case pn_Cmp_False: { - ir_graph *irg = get_irn_irg(proj); - return new_r_Const(irg, get_tarval_b_false()); - } - case pn_Cmp_True: { - ir_graph *irg = get_irn_irg(proj); - return new_r_Const(irg, get_tarval_b_true()); - } - case pn_Cmp_Leg: - if (!mode_is_float(mode)) { - ir_graph *irg = get_irn_irg(proj); - return new_r_Const(irg, get_tarval_b_true()); - } - break; - default: - break; - } +static ir_node *transform_node_Cmp(ir_node *n) +{ + ir_node *left = get_Cmp_left(n); + ir_node *right = get_Cmp_right(n); + ir_mode *mode = get_irn_mode(left); + ir_tarval *tv = NULL; + bool changed = false; + bool changedc = false; + ir_relation relation = get_Cmp_relation(n); + ir_relation possible = ir_get_possible_cmp_relations(left, right); - /* remove Casts of both sides */ - left = skip_Cast(left); - right = skip_Cast(right); + /* mask out impossible relations */ + ir_relation new_relation = relation & possible; + if (new_relation != relation) { + relation = new_relation; + changed = true; + } /* Remove unnecessary conversions */ - /* TODO handle constants */ 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; - changed |= 1; + 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; - changed |= 1; + 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); - changed |= 1; + 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); } } } + /* + * Optimize -a CMP -b into b CMP a. + * This works only for modes where unary Minus cannot Overflow. + * Note that two-complement integers can Overflow so it will NOT work. + */ + if (!mode_overflow_on_unary_Minus(mode) && + is_Minus(left) && is_Minus(right)) { + left = get_Minus_op(left); + right = get_Minus_op(right); + relation = get_inversed_relation(relation); + changed = true; + DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); + } + /* remove operation on both sides if possible */ - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { /* * The following operations are NOT safe for floating point operations, for instance * 1.0 + inf == 2.0 + inf, =/=> x == y @@ -3931,7 +4325,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) /* ~a CMP ~b => a CMP b, -a CMP -b ==> a CMP b */ left = get_unop_op(left); right = get_unop_op(right); - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); break; case iro_Add: @@ -3944,25 +4338,25 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) /* X + a CMP X + b ==> a CMP b */ left = lr; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (ll == rr) { /* X + a CMP b + X ==> a CMP b */ left = lr; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rl) { /* a + X CMP X + b ==> a CMP b */ left = ll; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rr) { /* a + X CMP b + X ==> a CMP b */ left = ll; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -3976,13 +4370,13 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) /* X - a CMP X - b ==> a CMP b */ left = lr; right = rr; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } else if (lr == rr) { /* a - X CMP b - X ==> a CMP b */ left = ll; right = rl; - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -3991,7 +4385,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) /* a ROTL X CMP b ROTL X ==> a CMP b */ left = get_Rotl_left(left); right = get_Rotl_left(right); - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } break; @@ -4001,11 +4395,11 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) } /* 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; @@ -4013,16 +4407,16 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (ll == right) { ir_graph *irg = get_irn_irg(n); left = lr; - right = create_zero_const(irg, mode); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; 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; @@ -4030,8 +4424,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (rl == left) { ir_graph *irg = get_irn_irg(n); left = rr; - right = create_zero_const(irg, mode); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } @@ -4057,55 +4451,102 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) left = new_rd_And(dbg, block, get_Shr_left(ll), new_r_Const(irg, mask), mode); right = new_r_Const(irg, value); - changed |= 1; + changed = true; } } } /* 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; + } + } + } + + 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; + 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; + } + } + } } - } /* mode_is_int(...) */ - } /* proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg */ - - /* 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) && (proj_nr == pn_Cmp_Eq - || (mode_is_signed(mode) && proj_nr == pn_Cmp_Lg) - || (!mode_is_signed(mode) && (proj_nr & pn_Cmp_Le) == pn_Cmp_Lt))) { - 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); - proj_nr = proj_nr == pn_Cmp_Eq ? pn_Cmp_Lg : pn_Cmp_Eq; - right = create_zero_const(irg, mode); - changed |= 1; } } /* replace mode_b compares with ands/ors */ - if (get_irn_mode(left) == mode_b) { + if (mode == mode_b) { ir_node *block = get_nodes_block(n); ir_node *bres; - switch (proj_nr) { - case pn_Cmp_Le: bres = new_r_Or( block, new_r_Not(block, left, mode_b), right, mode_b); break; - case pn_Cmp_Lt: bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); break; - case pn_Cmp_Ge: bres = new_r_Or( block, left, new_r_Not(block, right, mode_b), mode_b); break; - case pn_Cmp_Gt: bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); break; - case pn_Cmp_Lg: bres = new_r_Eor(block, left, right, mode_b); break; - case pn_Cmp_Eq: bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); break; - default: bres = NULL; - } - if (bres) { + switch (relation) { + case ir_relation_less_equal: + bres = new_r_Or(block, new_r_Not(block, left, mode_b), right, mode_b); + break; + case ir_relation_less: + bres = new_r_And(block, new_r_Not(block, left, mode_b), right, mode_b); + break; + case ir_relation_greater_equal: + bres = new_r_Or(block, left, new_r_Not(block, right, mode_b), mode_b); + break; + case ir_relation_greater: + bres = new_r_And(block, left, new_r_Not(block, right, mode_b), mode_b); + break; + case ir_relation_less_greater: + bres = new_r_Eor(block, left, right, mode_b); + break; + case ir_relation_equal: + bres = new_r_Not(block, new_r_Eor(block, left, right, mode_b), mode_b); + break; + default: +#ifdef DEBUG_libfirm + ir_fprintf(stderr, "Optimisation warning, unexpected mode_b Cmp %+F\n", n); +#endif + bres = NULL; + } + if (bres != NULL) { DBG_OPT_ALGSIM0(n, bres, FS_OPT_CMP_TO_BOOL); return bres; } @@ -4118,12 +4559,11 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) */ if (!operands_are_normalized(left, right)) { ir_node *t = left; - left = right; right = t; - proj_nr = get_inversed_pnc(proj_nr); - changed |= 1; + relation = get_inversed_relation(relation); + changed = true; } /* @@ -4134,7 +4574,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) */ tv = value_of(right); if (tv != tarval_bad) { - mode = get_irn_mode(right); + ir_mode *mode = get_irn_mode(right); /* TODO extend to arbitrary constants */ if (is_Conv(left) && tarval_is_null(tv)) { @@ -4147,13 +4587,13 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) * win. (on the other side it makes detection/creation of fabs hard) */ if (get_mode_size_bits(mode) > get_mode_size_bits(op_mode) && - ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) || + ((relation == ir_relation_equal || relation == ir_relation_less_greater) || mode_is_signed(mode) || !mode_is_signed(op_mode)) && !mode_is_float(mode)) { tv = get_mode_null(op_mode); left = op; mode = op_mode; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV); } } @@ -4169,73 +4609,53 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) */ if (is_Minus(left) && (!mode_overflow_on_unary_Minus(mode) || - (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) { + (mode_is_int(mode) && (relation == ir_relation_equal || relation == ir_relation_less_greater)))) { tv = tarval_neg(tv); if (tv != tarval_bad) { left = get_Minus_op(left); - proj_nr = get_inversed_pnc(proj_nr); - changed |= 2; + relation = get_inversed_relation(relation); + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } - } else if (is_Not(left) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)) { + } else if (is_Not(left) && (relation == ir_relation_equal || relation == ir_relation_less_greater)) { /* Not(a) ==/!= c ==> a ==/!= Not(c) */ tv = tarval_not(tv); if (tv != tarval_bad) { left = get_Not_op(left); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } /* for integer modes, we have more */ - if (mode_is_int(mode)) { - /* Ne includes Unordered which is not possible on integers. - * However, frontends often use this wrong, so fix it here */ - if (proj_nr & pn_Cmp_Uo) { - proj_nr &= ~pn_Cmp_Uo; - set_Proj_proj(proj, proj_nr); - } - + if (mode_is_int(mode) && !is_Const(left)) { /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */ - if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) && - tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) { + if ((relation == ir_relation_less || relation == ir_relation_greater_equal) && + tarval_cmp(tv, get_mode_null(mode)) == ir_relation_greater) { tv = tarval_sub(tv, get_mode_one(mode), NULL); if (tv != tarval_bad) { - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + relation ^= ir_relation_equal; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */ - else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) && - tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) { + else if ((relation == ir_relation_greater || relation == ir_relation_less_equal) && + tarval_cmp(tv, get_mode_null(mode)) == ir_relation_less) { tv = tarval_add(tv, get_mode_one(mode)); if (tv != tarval_bad) { - proj_nr ^= pn_Cmp_Eq; - changed |= 2; + relation ^= ir_relation_equal; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } /* the following reassociations work only for == and != */ - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { - -#if 0 /* Might be not that good in general */ - /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */ - if (tarval_is_null(tv) && is_Sub(left)) { - right = get_Sub_right(left); - left = get_Sub_left(left); - - tv = value_of(right); - changed = 1; - DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); - } -#endif - + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { if (tv != tarval_bad) { /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */ if (is_Sub(left)) { @@ -4248,15 +4668,15 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (tv2 != tarval_bad) { left = get_Sub_left(left); tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } /* 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; @@ -4274,7 +4694,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (tv2 != tarval_bad) { left = a; tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } @@ -4286,15 +4706,15 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (tv2 != tarval_bad) { left = get_Minus_op(left); tv = tv2; - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_C); } } } - } /* == or != */ - } /* mode_is_int */ + } + } - if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) { + if (relation == ir_relation_equal || relation == ir_relation_less_greater) { switch (get_irn_opcode(left)) { ir_node *c1; @@ -4309,9 +4729,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (mask != tv) { /* TODO: move to constant evaluation */ ir_graph *irg = get_irn_irg(n); - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); c1 = new_r_Const(irg, tv); - DBG_OPT_CSTEVAL(proj, c1); + DBG_OPT_CSTEVAL(n, c1); return c1; } @@ -4330,8 +4750,8 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (get_Const_tarval(c1) == tv) { /* fine: do the transformation */ tv = get_mode_null(get_tarval_mode(tv)); - proj_nr ^= pn_Cmp_Leg; - changed |= 2; + relation ^= ir_relation_less_equal_greater; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CNST_MAGN); } } @@ -4347,9 +4767,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (! tarval_is_null(get_Const_tarval(c1))) { /* TODO: move to constant evaluation */ ir_graph *irg = get_irn_irg(n); - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); c1 = new_r_Const(irg, tv); - DBG_OPT_CSTEVAL(proj, c1); + DBG_OPT_CSTEVAL(n, c1); return c1; } } @@ -4373,16 +4793,16 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (tarval_and(tv, cmask) != tv) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); c1 = new_r_Const(irg, tv); - DBG_OPT_CSTEVAL(proj, c1); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shl_left(left); blk = get_nodes_block(n); left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shr(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; @@ -4405,16 +4825,16 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (tarval_and(tv, cmask) != tv) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); c1 = new_r_Const(irg, tv); - DBG_OPT_CSTEVAL(proj, c1); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shr_left(left); blk = get_nodes_block(n); left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shl(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; @@ -4440,30 +4860,31 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (!tarval_is_all_one(cond) && !tarval_is_null(cond)) { /* condition not met */ - tv = proj_nr == pn_Cmp_Eq ? get_tarval_b_false() : get_tarval_b_true(); + tv = relation == ir_relation_equal ? get_tarval_b_false() : get_tarval_b_true(); c1 = new_r_Const(irg, tv); - DBG_OPT_CSTEVAL(proj, c1); + DBG_OPT_CSTEVAL(n, c1); return c1; } sl = get_Shrs_left(left); blk = get_nodes_block(n); left = new_rd_And(get_irn_dbg_info(left), blk, sl, new_r_Const(irg, amask), mode); tv = tarval_shl(tv, tv1); - changed |= 2; + changedc = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_SHF_TO_AND); } break; - } /* switch */ + } } - } /* tarval != bad */ + } } - if (changed & 2) { /* need a new Const */ + if (changedc) { /* need a new Const */ ir_graph *irg = get_irn_irg(n); right = new_r_Const(irg, tv); + changed = true; } - if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { + if ((relation == ir_relation_equal || relation == ir_relation_less_greater) && is_Const(right) && is_Const_null(right) && is_Proj(left)) { ir_node *op = get_Proj_pred(left); if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) { @@ -4481,7 +4902,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) tv = tarval_sub(tv, get_mode_one(mode), NULL); left = new_rd_And(get_irn_dbg_info(op), blk, v, new_r_Const(irg, tv), mode); - changed |= 1; + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_MOD_TO_AND); } } @@ -4489,15 +4910,15 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) } if (changed) { - ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); /* create a new compare */ - n = new_rd_Cmp(get_irn_dbg_info(n), block, left, right); - proj = new_rd_Proj(get_irn_dbg_info(proj), n, get_irn_mode(proj), proj_nr); + n = new_rd_Cmp(dbgi, block, left, right, relation); } - return proj; -} /* transform_node_Proj_Cmp */ + return n; +} /** * Optimize CopyB(mem, x, x) into a Nop. @@ -4515,16 +4936,18 @@ static ir_node *transform_node_Proj_CopyB(ir_node *proj) DBG_OPT_EXC_REM(proj); proj = new_r_Jmp(get_nodes_block(copyb)); break; - case pn_CopyB_X_except: + case pn_CopyB_X_except: { + ir_graph *irg = get_irn_irg(proj); DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(get_irn_irg(proj)); + proj = new_r_Bad(irg, mode_X); break; + } default: break; } } return proj; -} /* transform_node_Proj_CopyB */ +} /** * Optimize Bounds(idx, idx, upper) into idx. @@ -4566,7 +4989,7 @@ static ir_node *transform_node_Proj_Bound(ir_node *proj) break; case pn_Bound_X_except: DBG_OPT_EXC_REM(proj); - proj = get_irg_bad(get_irn_irg(proj)); + proj = new_r_Bad(get_irn_irg(proj), mode_X); break; case pn_Bound_res: proj = idx; @@ -4581,10 +5004,10 @@ 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 it's Proj's + * Does all optimizations on nodes that must be done on its Projs * because of creating new nodes. */ static ir_node *transform_node_Proj(ir_node *proj) @@ -4594,295 +5017,150 @@ 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 */ - -/** - * Move Confirms down through Phi nodes. - */ -static ir_node *transform_node_Phi(ir_node *phi) -{ - int i, n; - ir_mode *mode = get_irn_mode(phi); - - 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; - pn_Cmp pnc; - - if (! is_Confirm(pred)) - return phi; - - bound = get_Confirm_bound(pred); - pnc = get_Confirm_cmp(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_cmp(pred) != pnc) - 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, pnc); - } - } - return phi; -} /* transform_node_Phi */ +} /** - * Returns the operands of a commutative bin-op, if one operand is - * a const, it is returned as the second one. + * 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 void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) +static bool is_block_unreachable(const ir_node *block) { - 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 */ + const ir_graph *irg = get_irn_irg(block); + if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) + return false; + return get_Block_dom_depth(block) < 0; +} -/** - * 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) +static ir_node *transform_node_Block(ir_node *block) { - 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); + ir_graph *irg = get_irn_irg(block); + int arity = get_irn_arity(block); + ir_node *bad = NULL; + int i; - 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); + if (!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)) + return block; - /* check for more */ + for (i = 0; i < arity; ++i) { + ir_node *const pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || !is_block_unreachable(get_nodes_block(pred))) 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 (!is_negated_value(c1, c2)) { - return irn_or; + if (bad == NULL) + bad = new_r_Bad(irg, mode_X); + set_irn_n(block, i, bad); } - /* 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 */ - -/** - * 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; + return block; +} - if (is_Not(a) && is_Not(b)) { - /* ~a | ~b = ~(a&b) */ - ir_node *block = get_nodes_block(n); +static ir_node *transform_node_Phi(ir_node *phi) +{ + int n = get_irn_arity(phi); + ir_mode *mode = get_irn_mode(phi); + ir_node *block = get_nodes_block(phi); + ir_graph *irg = get_irn_irg(phi); + ir_node *bad = NULL; + int i; - 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; + /* Set phi-operands for bad-block inputs to bad */ + for (i = 0; i < n; ++i) { + 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); + } + } } - /* we can evaluate 2 Projs of the same Cmp */ - if (get_irn_mode(n) == mode_b && is_Proj(a) && is_Proj(b)) { - ir_node *pred_a = get_Proj_pred(a); - ir_node *pred_b = get_Proj_pred(b); - if (pred_a == pred_b) { - dbg_info *dbgi = get_irn_dbg_info(n); - pn_Cmp pn_a = get_Proj_pn_cmp(a); - pn_Cmp pn_b = get_Proj_pn_cmp(b); - /* yes, we can simply calculate with pncs */ - pn_Cmp new_pnc = pn_a | pn_b; + /* 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; + } + } - return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc); + 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; + + /* 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; +} /* forward */ static ir_node *transform_node(ir_node *n); @@ -4896,9 +5174,11 @@ 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); @@ -4907,7 +5187,7 @@ 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; @@ -4915,37 +5195,55 @@ 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) & pn_Cmp_Lt)) { - 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); @@ -4956,91 +5254,6 @@ static ir_node *transform_node_shift(ir_node *n) DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); return transform_node(irn); -} /* transform_node_shift */ - -/** - * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) - * (we can use: - * - and, or, xor instead of & - * - Shl, Shr, Shrs, rotl instead of >> - * (with a special case for Or/Xor + Shrs) - */ -static ir_node *transform_node_bitop_shift(ir_node *n) -{ - ir_node *left; - ir_node *right = get_binop_right(n); - ir_mode *mode = get_irn_mode(n); - ir_node *bitop_left; - ir_node *bitop_right; - ir_op *op_left; - ir_node *block; - dbg_info *dbgi; - ir_graph *irg; - ir_node *new_shift; - ir_node *new_bitop; - ir_node *new_const; - ir_tarval *tv1; - ir_tarval *tv2; - ir_tarval *tv_shift; - - assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); - - if (!is_Const(right)) - return n; - - left = get_binop_left(n); - op_left = get_irn_op(left); - if (op_left != op_And && op_left != op_Or && op_left != op_Eor) - return n; - - /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ - if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { - /* TODO: test if sign bit is affectes */ - return n; - } - - bitop_right = get_binop_right(left); - if (!is_Const(bitop_right)) - return n; - - bitop_left = get_binop_left(left); - - block = get_nodes_block(n); - dbgi = get_irn_dbg_info(n); - tv1 = get_Const_tarval(bitop_right); - tv2 = get_Const_tarval(right); - - assert(get_tarval_mode(tv1) == mode); - - if (is_Shl(n)) { - new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shl(tv1, tv2); - } else if (is_Shr(n)) { - new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shr(tv1, tv2); - } else if (is_Shrs(n)) { - new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_shrs(tv1, tv2); - } else { - assert(is_Rotl(n)); - new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_rotl(tv1, tv2); - } - - assert(get_tarval_mode(tv_shift) == mode); - irg = get_irn_irg(n); - new_const = new_r_Const(irg, tv_shift); - - if (op_left == op_And) { - new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); - } else if (op_left == op_Or) { - new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); - } else { - assert(op_left == op_Eor); - new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); - } - - return new_bitop; } /** @@ -5066,7 +5279,7 @@ static ir_node *transform_node_shl_shr(ir_node *n) ir_tarval *tv_shift; ir_tarval *tv_mask; ir_graph *irg; - pn_Cmp pnc; + ir_relation relation; int need_shrs = 0; assert(is_Shl(n) || is_Shr(n) || is_Shrs(n)); @@ -5088,7 +5301,7 @@ static ir_node *transform_node_shl_shr(ir_node *n) if (is_Shrs(left)) { /* shrs variant only allowed if c1 >= c2 */ - if (! (tarval_cmp(tv_shl, tv_shr) & pn_Cmp_Ge)) + if (! (tarval_cmp(tv_shl, tv_shr) & ir_relation_greater_equal)) return n; tv_mask = tarval_shrs(get_mode_all_one(mode), tv_shr); @@ -5124,8 +5337,8 @@ static ir_node *transform_node_shl_shr(ir_node *n) irg = get_irn_irg(block); dbgi = get_irn_dbg_info(n); - pnc = tarval_cmp(tv_shl, tv_shr); - if (pnc == pn_Cmp_Lt || pnc == pn_Cmp_Eq) { + relation = tarval_cmp(tv_shl, tv_shr); + if (relation == ir_relation_less || relation == ir_relation_equal) { tv_shift = tarval_sub(tv_shr, tv_shl, NULL); new_const = new_r_Const(irg, tv_shift); if (need_shrs) { @@ -5134,7 +5347,7 @@ static ir_node *transform_node_shl_shr(ir_node *n) new_shift = new_rd_Shr(dbgi, block, x, new_const, mode); } } else { - assert(pnc == pn_Cmp_Gt); + assert(relation == ir_relation_greater); tv_shift = tarval_sub(tv_shl, tv_shr, NULL); new_const = new_r_Const(irg, tv_shift); new_shift = new_rd_Shl(dbgi, block, x, new_const, mode); @@ -5161,7 +5374,7 @@ 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) @@ -5193,8 +5406,8 @@ 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); @@ -5203,7 +5416,7 @@ 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)) { @@ -5249,31 +5462,60 @@ 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. @@ -5293,10 +5535,10 @@ 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. @@ -5312,10 +5554,10 @@ 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. @@ -5369,7 +5611,7 @@ static ir_node *transform_node_Conv(ir_node *n) } return n; -} /* transform_node_Conv */ +} /** * Remove dead blocks and nodes in dead blocks @@ -5384,25 +5626,22 @@ static ir_node *transform_node_End(ir_node *n) for (i = j = 0; i < n_keepalives; ++i) { ir_node *ka = get_End_keepalive(n, i); - if (is_Block(ka)) { - if (! is_Block_dead(ka)) { - in[j++] = ka; - } + ir_node *block; + /* no need to keep Bad */ + if (is_Bad(ka)) continue; - } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) { + /* do not keep unreachable code */ + block = is_Block(ka) ? ka : get_nodes_block(ka); + if (is_block_unreachable(block)) continue; - } else if (is_Bad(ka)) { - /* no need to keep Bad */ - continue; - } in[j++] = ka; } if (j != n_keepalives) set_End_keepalives(n, j, in); return n; -} /* transform_node_End */ +} -bool is_negated_value(ir_node *a, ir_node *b) +int ir_is_negated_value(const ir_node *a, const ir_node *b) { if (is_Minus(a) && get_Minus_op(a) == b) return true; @@ -5421,234 +5660,331 @@ bool is_negated_value(ir_node *a, ir_node *b) return false; } -/** - * Optimize a Mux into some simpler cases. - */ -static ir_node *transform_node_Mux(ir_node *n) +static const ir_node *skip_upconv(const ir_node *node) { - 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); + 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; +} - if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) - return n; +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; - 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); + /* 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; } } - /* first normalization step: try to move a constant to the false side, - * 0 preferred on false side too */ - if (is_Proj(sel)) { - ir_node *cmp = get_Proj_pred(sel); + return 0; +} - if (is_Cmp(cmp) && is_Const(t) && - (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) { - pn_Cmp pnc = get_Proj_pn_cmp(sel); - ir_node *tmp = t; - t = f; - f = tmp; +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; + } - /* Mux(x, a, b) => Mux(not(x), b, a) */ - sel = new_r_Proj(cmp, mode_b, - get_negated_pnc(pnc, get_irn_mode(get_Cmp_left(cmp)))); - n = new_rd_Mux(get_irn_dbg_info(n), get_nodes_block(n), sel, f, t, mode); + 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; } } - /* 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); + return false; +} + +/** + * Optimize a Mux into some simpler cases. + */ +static ir_node *transform_node_Mux(ir_node *n) +{ + ir_node *oldn = n; + ir_node *sel = get_Mux_sel(n); + ir_mode *mode = get_irn_mode(n); + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_graph *irg = get_irn_irg(n); + + /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */ + if (get_mode_arithmetic(mode) == irma_twos_complement) { + int abs = ir_mux_is_abs(sel, 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; + } + } + + /* first normalization step: try to move a constant to the false side, + * 0 preferred on false side too */ + if (is_Cmp(sel) && is_Const(t) && + (!is_Const(f) || (is_Const_null(t) && !is_Const_null(f)))) { + dbg_info *seldbgi = get_irn_dbg_info(sel); + ir_node *block = get_nodes_block(sel); + ir_relation relation = get_Cmp_relation(sel); + ir_node *tmp = t; + t = f; + f = tmp; + + /* Mux(x, a, b) => Mux(not(x), b, a) */ + 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); + } + + /* 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); + 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); + } + } + + /* 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(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; + 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, 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); + /* 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_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); + } + + /* 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); + + 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 { - /* 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); + } 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; } } } - /* 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); - - 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_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); + ir_node *cmp_l = get_Cmp_left(sel); + ir_node *block = get_nodes_block(n); - if (is_Proj(sel)) { - ir_node *cmp = get_Proj_pred(sel); - long pn = get_Proj_proj(sel); + if (is_And(cmp_l) && f == cmp_r) { + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; - /* - * Note: normalization puts the constant on the right side, - * so we check only one case. - */ - if (is_Cmp(cmp)) { - ir_node *cmp_r = get_Cmp_right(cmp); - if (is_Const(cmp_r) && is_Const_null(cmp_r)) { - ir_node *block = get_nodes_block(n); - ir_node *cmp_l = get_Cmp_left(cmp); - - if (mode_is_int(mode)) { - /* integer only */ - if ((pn == pn_Cmp_Lg || pn == pn_Cmp_Eq) && 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 (pn == pn_Cmp_Lg) { - /* Mux((a & 2^C) != 0, 2^C, 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, cmp, 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, cmp, 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 (pn == pn_Cmp_Lg) { - /* (a & (1 << n)) != 0, (1 << n), 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, cmp, 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, cmp, 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 (pn == pn_Cmp_Lg) { - /* ((1 << n) & a) != 0, (1 << n), 0) */ - n = cmp_l; - DBG_OPT_ALGSIM1(oldn, cmp, 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, cmp, sel, n, FS_OPT_MUX_TO_BITOP); - } - return n; - } - } - } - } + if (and_r == t && is_single_bit(and_r)) { + if (relation == ir_relation_equal) { + /* Mux((a & (1<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) { @@ -5872,15 +6300,17 @@ static int node_cmp_attr_Const(const ir_node *a, const ir_node *b) static int node_cmp_attr_Proj(const ir_node *a, const ir_node *b) { return a->attr.proj.proj != b->attr.proj.proj; -} /* node_cmp_attr_Proj */ +} /** Compares the attributes of two Alloc nodes. */ 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); -} /* node_cmp_attr_Alloc */ + if (pa->where != pb->where || pa->type != pb->type) + return 1; + return node_cmp_exception(a, b); +} /** Compares the attributes of two Free nodes. */ static int node_cmp_attr_Free(const ir_node *a, const ir_node *b) @@ -5888,7 +6318,7 @@ static int node_cmp_attr_Free(const ir_node *a, const ir_node *b) const free_attr *pa = &a->attr.free; const free_attr *pb = &b->attr.free; return (pa->where != pb->where) || (pa->type != pb->type); -} /* node_cmp_attr_Free */ +} /** Compares the attributes of two SymConst nodes. */ static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b) @@ -5904,9 +6334,10 @@ 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); -} /* node_cmp_attr_Call */ + if (pa->type != pb->type || pa->tail_call != pb->tail_call) + return 1; + return node_cmp_exception(a, b); +} /** Compares the attributes of two Sel nodes. */ static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b) @@ -5914,7 +6345,7 @@ static int node_cmp_attr_Sel(const ir_node *a, const ir_node *b) const ir_entity *a_ent = get_Sel_entity(a); const ir_entity *b_ent = get_Sel_entity(b); return a_ent != b_ent; -} /* node_cmp_attr_Sel */ +} /** Compares the attributes of two Phi nodes. */ static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b) @@ -5926,19 +6357,19 @@ static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b) return a->attr.phi.u.pos != b->attr.phi.u.pos; } return 0; -} /* node_cmp_attr_Phi */ +} /** Compares the attributes of two Conv nodes. */ static int node_cmp_attr_Conv(const ir_node *a, const ir_node *b) { return get_Conv_strict(a) != get_Conv_strict(b); -} /* node_cmp_attr_Conv */ +} /** Compares the attributes of two Cast nodes. */ static int node_cmp_attr_Cast(const ir_node *a, const ir_node *b) { return get_Cast_type(a) != get_Cast_type(b); -} /* node_cmp_attr_Cast */ +} /** Compares the attributes of two Load nodes. */ static int node_cmp_attr_Load(const ir_node *a, const ir_node *b) @@ -5948,67 +6379,83 @@ static int node_cmp_attr_Load(const ir_node *a, const ir_node *b) /* NEVER do CSE on volatile Loads */ return 1; /* do not CSE Loads with different alignment. Be conservative. */ - if (get_Load_align(a) != get_Load_align(b)) + if (get_Load_unaligned(a) != get_Load_unaligned(b)) return 1; - - return get_Load_mode(a) != get_Load_mode(b); -} /* node_cmp_attr_Load */ + if (get_Load_mode(a) != get_Load_mode(b)) + return 1; + return node_cmp_exception(a, b); +} /** Compares the attributes of two Store nodes. */ 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_align(a) != get_Store_align(b)) + 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); -} /* node_cmp_attr_Store */ + 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; -} /* node_cmp_attr_Div */ + 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. */ 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; -} /* node_cmp_attr_Mod */ + 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) +{ + const cmp_attr *ma = &a->attr.cmp; + const cmp_attr *mb = &b->attr.cmp; + return ma->relation != mb->relation; +} /** Compares the attributes of two Confirm nodes. */ static int node_cmp_attr_Confirm(const ir_node *a, const ir_node *b) { - /* no need to compare the bound, as this is a input */ - return (get_Confirm_cmp(a) != get_Confirm_cmp(b)); -} /* node_cmp_attr_Confirm */ + const confirm_attr *ma = &a->attr.confirm; + const confirm_attr *mb = &b->attr.confirm; + return ma->relation != mb->relation; +} /** 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); -} /* node_cmp_attr_Builtin */ + 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) @@ -6024,29 +6471,31 @@ static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) /* Should we really check the constraints here? Should be better, but is strange. */ n = get_ASM_n_input_constraints(a); if (n != get_ASM_n_input_constraints(b)) - return 0; + return 1; ca = get_ASM_input_constraints(a); cb = get_ASM_input_constraints(b); for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) + if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint + || ca[i].mode != cb[i].mode) return 1; } n = get_ASM_n_output_constraints(a); if (n != get_ASM_n_output_constraints(b)) - return 0; + return 1; ca = get_ASM_output_constraints(a); cb = get_ASM_output_constraints(b); for (i = 0; i < n; ++i) { - if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint) + if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint + || ca[i].mode != cb[i].mode) return 1; } n = get_ASM_n_clobbers(a); if (n != get_ASM_n_clobbers(b)) - return 0; + return 1; cla = get_ASM_clobbers(a); clb = get_ASM_clobbers(b); @@ -6054,17 +6503,26 @@ static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) if (cla[i] != clb[i]) return 1; } - return 0; -} /* node_cmp_attr_ASM */ + + return node_cmp_exception(a, b); +} /** Compares the inexistent attributes of two Dummy nodes. */ 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; } +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); +} + /** * Set the default node attribute compare operation for an ir_op_ops. * @@ -6082,26 +6540,28 @@ 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(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; @@ -6109,7 +6569,7 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) return ops; #undef CASE -} /* firm_set_default_node_cmp_attr */ +} /* * Compare function for two nodes in the value table. Gets two @@ -6139,10 +6599,23 @@ 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] */ @@ -6164,7 +6637,7 @@ 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. @@ -6174,7 +6647,7 @@ 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) @@ -6182,13 +6655,13 @@ 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). */ @@ -6208,7 +6681,7 @@ void ir_normalize_node(ir_node *n) hook_normalize(n); } } -} /* ir_normalize_node */ +} /* * Return the canonical node computing the same value as n. @@ -6235,11 +6708,11 @@ ir_node *identify_remember(ir_node *n) if (nn != n) { /* n is reachable again */ - edges_node_revival(nn, get_irn_irg(nn)); + edges_node_revival(nn); } return nn; -} /* identify_remember */ +} /** * During construction we set the op_pin_state_pinned flag in the graph right @@ -6258,7 +6731,7 @@ 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) @@ -6282,101 +6755,7 @@ void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) visit(node, env); } current_ir_graph = rem; -} /* visit_all_identities */ - -/** - * Garbage in, garbage out. If a node has a dead input, i.e., the - * Bad node is input to the node, return the Bad node. - */ -static ir_node *gigo(ir_node *node) -{ - int i, irn_arity; - ir_op *op = get_irn_op(node); - - /* remove garbage blocks by looking at control flow that leaves the block - and replacing the control flow by Bad. */ - if (get_irn_mode(node) == mode_X) { - ir_node *block = get_nodes_block(skip_Proj(node)); - ir_graph *irg = get_irn_irg(block); - - /* Don't optimize nodes in immature blocks. */ - if (!get_Block_matured(block)) - return node; - /* Don't optimize End, may have Bads. */ - if (op == op_End) return node; - - if (is_Block(block)) { - if (is_Block_dead(block)) { - /* control flow from dead block is dead */ - return get_irg_bad(irg); - } - - for (i = get_irn_arity(block) - 1; i >= 0; --i) { - if (!is_Bad(get_irn_n(block, i))) - break; - } - if (i < 0) { - ir_graph *irg = get_irn_irg(block); - /* the start block is never dead */ - if (block != get_irg_start_block(irg) - && block != get_irg_end_block(irg)) { - /* - * Do NOT kill control flow without setting - * the block to dead of bad things can happen: - * We get a Block that is not reachable be irg_block_walk() - * but can be found by irg_walk()! - */ - set_Block_dead(block); - return get_irg_bad(irg); - } - } - } - } - - /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the - blocks predecessors is dead. */ - if (op != op_Block && op != op_Phi && op != op_Tuple && op != op_Anchor) { - ir_graph *irg = get_irn_irg(node); - irn_arity = get_irn_arity(node); - - /* - * Beware: we can only read the block of a non-floating node. - */ - if (is_irn_pinned_in_irg(node) && - is_Block_dead(get_nodes_block(skip_Proj(node)))) - return get_irg_bad(irg); - - for (i = 0; i < irn_arity; i++) { - ir_node *pred = get_irn_n(node, i); - - if (is_Bad(pred)) - return get_irg_bad(irg); -#if 0 - /* Propagating Unknowns here seems to be a bad idea, because - sometimes we need a node as a input and did not want that - it kills it's user. - However, it might be useful to move this into a later phase - (if you think that optimizing such code is useful). */ - if (is_Unknown(pred) && mode_is_data(get_irn_mode(node))) - return new_r_Unknown(irg, get_irn_mode(node)); -#endif - } - } -#if 0 - /* With this code we violate the agreement that local_optimize - only leaves Bads in Block, Phi and Tuple nodes. */ - /* If Block has only Bads as predecessors it's garbage. */ - /* If Phi has only Bads as predecessors it's garbage. */ - if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) { - irn_arity = get_irn_arity(node); - for (i = 0; i < irn_arity; i++) { - if (!is_Bad(get_irn_n(node, i))) break; - } - if (i == irn_arity) node = get_irg_bad(irg); - } -#endif - return node; -} /* gigo */ +} /** * These optimizations deallocate nodes from the obstack. @@ -6406,7 +6785,7 @@ ir_node *optimize_node(ir_node *n) size_t node_size; /* - * we MUST copy the node here temporary, because it's still + * we MUST copy the node here temporarily, because it's still * needed for DBG_OPT_CSTEVAL */ node_size = offsetof(ir_node, attr) + n->op->attr_size; @@ -6419,7 +6798,7 @@ ir_node *optimize_node(ir_node *n) memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0])); /* note the inplace edges module */ - edges_node_deleted(n, irg); + edges_node_deleted(n); /* evaluation was successful -- replace the node. */ irg_kill_node(irg, n); @@ -6450,7 +6829,7 @@ ir_node *optimize_node(ir_node *n) n = identify_cons(n); if (n != oldn) { - edges_node_deleted(oldn, irg); + edges_node_deleted(oldn); /* We found an existing, better node, so we can deallocate the old node. */ irg_kill_node(irg, oldn); @@ -6465,10 +6844,6 @@ ir_node *optimize_node(ir_node *n) (iro == iro_Proj)) /* Flags tested local. */ n = transform_node(n); - /* Remove nodes with dead (Bad) input. - Run always for transformation induced Bads. */ - n = gigo(n); - /* Now we have a legal, useful node. Enter it in hash table for CSE */ if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) { ir_node *o = n; @@ -6478,7 +6853,7 @@ ir_node *optimize_node(ir_node *n) } return n; -} /* optimize_node */ +} /** @@ -6497,12 +6872,6 @@ ir_node *optimize_in_place_2(ir_node *n) if (iro == iro_Deleted) return n; - /* Remove nodes with dead (Bad) input. - Run always for transformation induced Bads. */ - n = gigo(n); - if (is_Bad(n)) - return n; - /* constant expression evaluation / constant folding */ if (get_opt_constant_folding()) { /* neither constants nor Tuple values can be evaluated */ @@ -6562,7 +6931,7 @@ 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. @@ -6575,14 +6944,12 @@ ir_node *optimize_in_place(ir_node *n) if (get_opt_global_cse()) set_irg_pinned(irg, op_pin_state_floats); - if (get_irg_outs_state(irg) == outs_consistent) - set_irg_outs_inconsistent(irg); /* 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. @@ -6595,7 +6962,7 @@ 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. @@ -6608,7 +6975,7 @@ 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. @@ -6619,7 +6986,7 @@ static unsigned hash_SymConst(const ir_node *node) * @return * The operations. */ -static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops) +static ir_op_ops *firm_set_default_hash(unsigned code, ir_op_ops *ops) { #define CASE(a) \ case iro_##a: \ @@ -6645,7 +7012,7 @@ static ir_op_ops *firm_set_default_hash(ir_opcode code, ir_op_ops *ops) /* * Sets the default operation for an ir_ops. */ -ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *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); @@ -6656,4 +7023,4 @@ ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) ops = firm_set_default_get_entity_attr(code, ops); return ops; -} /* firm_set_default_operations */ +}