X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=ca58b2ef4eeb06915ab19343d0987be77cbe1dc5;hb=a08e6f04aa3669cff094f94a9484c7c2bb1314d0;hp=f72a33754a92e050fe5263485be6762df47429c7;hpb=9c359401bbbb6cc870909cf556f87a375807bdb7;p=libfirm diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c index f72a33754..ca58b2ef4 100644 --- a/ir/ir/iropt.c +++ b/ir/ir/iropt.c @@ -131,6 +131,12 @@ static ir_tarval *computed_value_Add(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) return tarval_add(ta, tb); + /* x+~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } + return tarval_bad; } /* computed_value_Add */ @@ -197,7 +203,7 @@ 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); } @@ -264,10 +270,17 @@ static ir_tarval *computed_value_And(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_and (ta, tb); - } else { - if (tarval_is_null(ta)) return ta; - if (tarval_is_null(tb)) return tb; } + + if (tarval_is_null(ta)) return ta; + if (tarval_is_null(tb)) return tb; + + /* x&~x => 0 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_null(get_irn_mode(n)); + } + return tarval_bad; } /* computed_value_And */ @@ -284,9 +297,15 @@ static ir_tarval *computed_value_Or(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_or (ta, tb); - } else { - if (tarval_is_all_one(ta)) return ta; - if (tarval_is_all_one(tb)) return tb; + } + + if (tarval_is_all_one(ta)) return ta; + if (tarval_is_all_one(tb)) return tb; + + /* x|~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); } return tarval_bad; } /* computed_value_Or */ @@ -303,6 +322,11 @@ static ir_tarval *computed_value_Eor(const ir_node *n) if (a == b) return get_mode_null(get_irn_mode(n)); + /* x^~x => -1 */ + if ((is_Not(a) && get_Not_op(a) == b) + || (is_Not(b) && get_Not_op(b) == a)) { + return get_mode_all_one(get_irn_mode(n)); + } ta = value_of(a); tb = value_of(b); @@ -327,6 +351,28 @@ static ir_tarval *computed_value_Not(const ir_node *n) return tarval_bad; } /* computed_value_Not */ +/** + * Tests whether a shift shifts more bits than available in the mode + */ +static bool is_oversize_shift(const ir_node *n) +{ + ir_node *count = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_tarval *tv = value_of(count); + long modulo_shift; + long shiftval; + if (tv == tarval_bad) + return false; + if (!tarval_is_long(tv)) + return false; + shiftval = get_tarval_long(tv); + modulo_shift = get_mode_modulo_shift(mode); + if (shiftval < 0 || (modulo_shift > 0 && shiftval >= modulo_shift)) + return false; + + return shiftval >= (long)get_mode_size_bits(mode); +} + /** * Return the value of a Shl. */ @@ -341,6 +387,10 @@ static ir_tarval *computed_value_Shl(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shl(ta, tb); } + + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; } /* computed_value_Shl */ @@ -358,6 +408,9 @@ static ir_tarval *computed_value_Shr(const ir_node *n) if ((ta != tarval_bad) && (tb != tarval_bad)) { return tarval_shr(ta, tb); } + if (is_oversize_shift(n)) + return get_mode_null(get_irn_mode(n)); + return tarval_bad; } /* computed_value_Shr */ @@ -395,17 +448,52 @@ static ir_tarval *computed_value_Rotl(const ir_node *n) return tarval_bad; } /* computed_value_Rotl */ +bool ir_zero_when_converted(const ir_node *node, ir_mode *dest_mode) +{ + ir_mode *mode = get_irn_mode(node); + if (get_mode_arithmetic(mode) != irma_twos_complement + || get_mode_arithmetic(dest_mode) != irma_twos_complement) + return false; + + if (is_Shl(node)) { + ir_node *count = get_Shl_right(node); + if (is_Const(count)) { + ir_tarval *tv = get_Const_tarval(count); + if (tarval_is_long(tv)) { + long shiftval = get_tarval_long(tv); + long destbits = get_mode_size_bits(dest_mode); + if (shiftval >= destbits + && shiftval < (long)get_mode_modulo_shift(mode)) + return true; + } + } + } + if (is_And(node)) { + ir_node *right = get_And_right(node); + if (is_Const(right)) { + ir_tarval *tv = get_Const_tarval(right); + ir_tarval *conved = tarval_convert_to(tv, dest_mode); + return tarval_is_null(conved); + } + } + return false; +} + /** * Return the value of a Conv. */ static ir_tarval *computed_value_Conv(const ir_node *n) { - ir_node *a = get_Conv_op(n); - ir_tarval *ta = value_of(a); + ir_node *a = get_Conv_op(n); + ir_tarval *ta = value_of(a); + ir_mode *mode = get_irn_mode(n); if (ta != tarval_bad) return tarval_convert_to(ta, get_irn_mode(n)); + if (ir_zero_when_converted(a, mode)) + return get_mode_null(mode); + return tarval_bad; } /* computed_value_Conv */ @@ -435,7 +523,7 @@ 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; @@ -444,132 +532,91 @@ static ir_tarval *computed_value_Confirm(const ir_node *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); - - 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 */ + 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; +} /** - * Return the value of a floating point Quot. + * 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 *do_computed_value_Quot(const ir_node *a, const ir_node *b) +static ir_tarval *computed_value_Cmp(const ir_node *cmp) { - ir_tarval *ta = value_of(a); - ir_tarval *tb = value_of(b); + 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); - /* cannot optimize 0 / b = 0 because of NaN */ - if (ta != tarval_bad && tb != tarval_bad) - return tarval_quo(ta, tb); - return tarval_bad; -} /* do_computed_value_Quot */ + /* 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); +} /** - * Calculate the value of an integer Div of two nodes. + * Calculate the value of an integer Div. * Special case: 0 / b */ -static ir_tarval *do_computed_value_Div(const ir_node *a, const ir_node *b) +static ir_tarval *do_computed_value_Div(const ir_node *div) { - ir_tarval *ta = value_of(a); + const ir_node *a = get_Div_left(div); + const ir_node *b = get_Div_right(div); + const ir_mode *mode = get_Div_resmode(div); + ir_tarval *ta = value_of(a); ir_tarval *tb; const ir_node *dummy; - /* Compute c1 / c2 or 0 / a, a != 0 */ - if (tarval_is_null(ta) && value_not_zero(b, &dummy)) - return ta; /* 0 / b == 0 */ + /* cannot optimize 0 / b = 0 because of NaN */ + if (!mode_is_float(mode)) { + if (tarval_is_null(ta) && value_not_zero(b, &dummy)) + return ta; /* 0 / b == 0 if b != 0 */ + } tb = value_of(b); if (ta != tarval_bad && tb != tarval_bad) return tarval_div(ta, tb); @@ -593,36 +640,16 @@ static ir_tarval *do_computed_value_Mod(const ir_node *a, const ir_node *b) return tarval_bad; } /* do_computed_value_Mod */ -/** - * Return the value of a Proj(DivMod). - */ -static ir_tarval *computed_value_Proj_DivMod(const ir_node *n) -{ - long proj_nr = get_Proj_proj(n); - - /* compute either the Div or the Mod part */ - if (proj_nr == pn_DivMod_res_div) { - const ir_node *a = get_Proj_pred(n); - return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a)); - } else if (proj_nr == pn_DivMod_res_mod) { - const ir_node *a = get_Proj_pred(n); - return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a)); - } - return tarval_bad; -} /* computed_value_Proj_DivMod */ - /** * Return the value of a Proj(Div). */ static ir_tarval *computed_value_Proj_Div(const ir_node *n) { long proj_nr = get_Proj_proj(n); + if (proj_nr != pn_Div_res) + return tarval_bad; - if (proj_nr == pn_Div_res) { - const ir_node *a = get_Proj_pred(n); - return do_computed_value_Div(get_Div_left(a), get_Div_right(a)); - } - return tarval_bad; + return do_computed_value_Div(get_Proj_pred(n)); } /* computed_value_Proj_Div */ /** @@ -633,26 +660,12 @@ static ir_tarval *computed_value_Proj_Mod(const ir_node *n) long proj_nr = get_Proj_proj(n); if (proj_nr == pn_Mod_res) { - const ir_node *a = get_Proj_pred(n); - return do_computed_value_Mod(get_Mod_left(a), get_Mod_right(a)); + const ir_node *mod = get_Proj_pred(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(Quot). - */ -static ir_tarval *computed_value_Proj_Quot(const ir_node *n) -{ - long proj_nr = get_Proj_proj(n); - - if (proj_nr == pn_Quot_res) { - const ir_node *a = get_Proj_pred(n); - return do_computed_value_Quot(get_Quot_left(a), get_Quot_right(a)); - } - return tarval_bad; -} /* computed_value_Proj_Quot */ - /** * Return the value of a Proj. */ @@ -674,7 +687,7 @@ 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 && vrp->valid && tarval_cmp(vrp->bits_set, vrp->bits_not_set) == ir_relation_equal) { return vrp->bits_set; } if (n->op->ops.computed_value) @@ -703,31 +716,29 @@ 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_PROJ(DivMod); + CASE(Sub); + CASE(SymConst); CASE_PROJ(Div); CASE_PROJ(Mod); - CASE_PROJ(Quot); - CASE(Proj); default: /* leave NULL */ break; @@ -738,141 +749,6 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops #undef CASE } /* firm_set_default_computed_value */ -/** - * Returns a equivalent block for another block. - * If the block has only one predecessor, this is - * the equivalent one. If the only predecessor of a block is - * the block itself, this is a dead block. - * - * If both predecessors of a block are the branches of a binary - * Cond, the equivalent block is Cond's block. - * - * If all predecessors of a block are bad or lies in a dead - * block, the current block is dead as well. - * - * 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, * so a op 0 = 0 op a = a. @@ -1081,7 +957,7 @@ static ir_node *equivalent_node_Sub(ir_node *n) /** - * Optimize an "self-inverse unary op", ie op(op(n)) = n. + * Optimize an "self-inverse unary op", i.e. op(op(n)) = n. * * @todo * -(-a) == a, but might overflow two times. @@ -1152,11 +1028,11 @@ 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; @@ -1185,11 +1061,11 @@ 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; @@ -1376,7 +1252,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) int i, n_preds; ir_node *oldn = n; - ir_node *block; ir_node *first_val = NULL; /* to shutup gcc */ if (!get_opt_optimize() && @@ -1385,52 +1260,25 @@ 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; } } @@ -1443,49 +1291,6 @@ static ir_node *equivalent_node_Phi(ir_node *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). */ @@ -1533,72 +1338,6 @@ static ir_node *equivalent_node_Proj_Div(ir_node *proj) return proj; } /* equivalent_node_Proj_Div */ -/** - * Optimize a / 1.0 = a. - */ -static ir_node *equivalent_node_Proj_Quot(ir_node *proj) -{ - ir_node *oldn = proj; - ir_node *quot = get_Proj_pred(proj); - ir_node *b = get_Quot_right(quot); - ir_tarval *tb = value_of(b); - - /* Div is not commutative. */ - if (tarval_is_one(tb)) { /* Quot(x, 1) == x */ - switch (get_Proj_proj(proj)) { - case pn_Quot_M: - proj = get_Quot_mem(quot); - DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); - return proj; - - case pn_Quot_res: - proj = get_Quot_left(quot); - DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); - return proj; - - default: - /* we cannot replace the exception Proj's here, this is done in - transform_node_Proj_Quot() */ - return proj; - } - } - return proj; -} /* equivalent_node_Proj_Quot */ - -/** - * Optimize a / 1 = a. - */ -static ir_node *equivalent_node_Proj_DivMod(ir_node *proj) -{ - ir_node *oldn = proj; - ir_node *divmod = get_Proj_pred(proj); - ir_node *b = get_DivMod_right(divmod); - ir_tarval *tb = value_of(b); - - /* Div is not commutative. */ - if (tarval_is_one(tb)) { /* div(x, 1) == x */ - switch (get_Proj_proj(proj)) { - case pn_DivMod_M: - proj = get_DivMod_mem(divmod); - DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); - return proj; - - case pn_DivMod_res_div: - proj = get_DivMod_left(divmod); - DBG_OPT_ALGSIM0(oldn, proj, FS_OPT_NEUTRAL_1); - return proj; - - default: - /* we cannot replace the exception Proj's here, this is done in - transform_node_Proj_DivMod(). - Note further that the pn_DivMod_res_div case is handled in - computed_value_Proj(). */ - return proj; - } - } - return proj; -} /* equivalent_node_Proj_DivMod */ - /** * Optimize CopyB(mem, x, x) into a Nop. */ @@ -1616,13 +1355,6 @@ 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; @@ -1666,12 +1398,6 @@ 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); @@ -1685,70 +1411,12 @@ static ir_node *equivalent_node_Proj_Bound(ir_node *proj) } /* equivalent_node_Proj_Bound */ /** - * Optimize an Exception Proj(Load) with a non-null address. + * Does all optimizations on nodes that must be done on its Projs + * because of creating new nodes. */ -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. - */ -static ir_node *equivalent_node_Proj(ir_node *proj) +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; @@ -1812,65 +1480,65 @@ 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. @@ -1878,17 +1546,16 @@ 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; } @@ -1928,9 +1595,6 @@ 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); @@ -1945,15 +1609,10 @@ 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(Quot); - CASE_PROJ(DivMod); CASE_PROJ(CopyB); CASE_PROJ(Bound); - CASE_PROJ(Load); - CASE_PROJ(Store); CASE(Proj); CASE(Id); CASE(Mux); @@ -2381,22 +2040,6 @@ static ir_node *transform_node_Add(ir_node *n) DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1); return n; } - if (op == b) { - /* ~x + x = -1 */ - n = new_r_Const(irg, get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } - } - if (is_Not(b)) { - ir_node *op = get_Not_op(b); - - if (op == a) { - /* x + ~x = -1 */ - n = new_r_Const(irg, get_mode_minus_one(mode)); - DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_X_NOT_X); - return n; - } } } } @@ -2517,19 +2160,19 @@ restart: /* a - (b - c) -> a + (c - b) * -> (a - b) + c iff (b - c) is a pointer */ dbg_info *s_dbg = get_irn_dbg_info(b); - ir_node *s_block = get_nodes_block(b); ir_node *s_left = get_Sub_left(b); ir_node *s_right = get_Sub_right(b); ir_mode *s_mode = get_irn_mode(b); if (mode_is_reference(s_mode)) { - ir_node *sub = new_rd_Sub(s_dbg, s_block, a, s_left, mode); + ir_node *lowest_block = get_nodes_block(n); /* a and b are live here */ + ir_node *sub = new_rd_Sub(s_dbg, lowest_block, a, s_left, mode); dbg_info *a_dbg = get_irn_dbg_info(n); - ir_node *a_block = get_nodes_block(n); if (s_mode != mode) - s_right = new_r_Conv(a_block, s_right, mode); - n = new_rd_Add(a_dbg, a_block, sub, s_right, mode); + s_right = new_r_Conv(lowest_block, s_right, mode); + n = new_rd_Add(a_dbg, lowest_block, sub, s_right, mode); } else { + ir_node *s_block = get_nodes_block(b); ir_node *sub = new_rd_Sub(s_dbg, s_block, s_right, s_left, s_mode); dbg_info *a_dbg = get_irn_dbg_info(n); ir_node *a_block = get_nodes_block(n); @@ -2538,6 +2181,7 @@ restart: } DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD); return n; +#if 0 } else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */ ir_node *m_right = get_Mul_right(b); if (is_Const(m_right)) { @@ -2556,6 +2200,7 @@ restart: return n; } } +#endif } /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */ @@ -2717,8 +2362,8 @@ restart: } if (get_mode_arithmetic(mode) == irma_twos_complement) { + /* c - ~X = X + (c+1) */ if (is_Const(a) && is_Not(b)) { - /* c - ~X = X + (c+1) */ ir_tarval *tv = get_Const_tarval(a); tv = tarval_add(tv, get_mode_one(mode)); @@ -2731,6 +2376,24 @@ restart: return n; } } + /* x-(x&y) = x & ~y */ + if (is_And(b)) { + ir_node *and_left = get_And_left(b); + ir_node *and_right = get_And_right(b); + if (and_right == a) { + ir_node *tmp = and_left; + and_left = and_right; + and_right = tmp; + } + if (and_left == a) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *mode = get_irn_mode(n); + ir_node *notn = new_rd_Not(dbgi, block, and_right, mode); + ir_node *and = new_rd_And(dbgi, block, a, notn, mode); + return and; + } + } } return n; } /* transform_node_Sub */ @@ -2907,55 +2570,87 @@ static ir_node *transform_node_Div(ir_node *n) ir_mode *mode = get_Div_resmode(n); ir_node *a = get_Div_left(n); ir_node *b = get_Div_right(n); - ir_node *value; + ir_node *value = n; const ir_node *dummy; - if (is_Const(b) && is_const_Phi(a)) { - /* check for Div(Phi, Const) */ - value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0); - if (value) { - DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); - goto make_tuple; - } - } - else if (is_Const(a) && is_const_Phi(b)) { - /* check for Div(Const, Phi) */ - value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1); - if (value) { - DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); - goto make_tuple; + if (mode_is_int(mode)) { + if (is_Const(b) && is_const_Phi(a)) { + /* check for Div(Phi, Const) */ + value = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } else if (is_Const(a) && is_const_Phi(b)) { + /* check for Div(Const, Phi) */ + value = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } + } else if (is_const_Phi(a) && is_const_Phi(b)) { + /* check for Div(Phi, Phi) */ + value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode); + if (value) { + DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + goto make_tuple; + } } - } - else if (is_const_Phi(a) && is_const_Phi(b)) { - /* check for Div(Phi, Phi) */ - value = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode); - if (value) { - DBG_OPT_ALGSIM0(n, value, FS_OPT_CONST_PHI); + + if (a == b && value_not_zero(a, &dummy)) { + ir_graph *irg = get_irn_irg(n); + /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ + value = new_r_Const(irg, get_mode_one(mode)); + DBG_OPT_CSTEVAL(n, value); goto make_tuple; + } else { + if (mode_is_signed(mode) && is_Const(b)) { + ir_tarval *tv = get_Const_tarval(b); + + if (tv == get_mode_minus_one(mode)) { + /* a / -1 */ + value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); + DBG_OPT_CSTEVAL(n, value); + goto make_tuple; + } + } + /* Try architecture dependent optimization */ + value = arch_dep_replace_div_by_const(n); } - } + } else { + assert(mode_is_float(mode)); - value = n; + /* Optimize x/c to x*(1/c) */ + if (get_mode_arithmetic(mode) == irma_ieee754) { + ir_tarval *tv = value_of(b); - if (a == b && value_not_zero(a, &dummy)) { - ir_graph *irg = get_irn_irg(n); - /* BEWARE: we can optimize a/a to 1 only if this cannot cause a exception */ - value = new_r_Const(irg, get_mode_one(mode)); - DBG_OPT_CSTEVAL(n, value); - goto make_tuple; - } else { - if (mode_is_signed(mode) && is_Const(b)) { - ir_tarval *tv = get_Const_tarval(b); + if (tv != tarval_bad) { + int rem = tarval_fp_ops_enabled(); - if (tv == get_mode_minus_one(mode)) { - /* a / -1 */ - value = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); - DBG_OPT_CSTEVAL(n, value); - goto make_tuple; + /* + * Floating point constant folding might be disabled here to + * prevent rounding. + * However, as we check for exact result, doing it is safe. + * Switch it on. + */ + tarval_enable_fp_ops(1); + tv = tarval_div(get_mode_one(mode), tv); + tarval_enable_fp_ops(rem); + + /* Do the transformation if the result is either exact or we are + not using strict rules. */ + if (tv != tarval_bad && + (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) { + ir_node *block = get_nodes_block(n); + ir_graph *irg = get_irn_irg(block); + ir_node *c = new_r_Const(irg, tv); + dbg_info *dbgi = get_irn_dbg_info(n); + value = new_rd_Mul(dbgi, block, a, c, mode); + + goto make_tuple; + } } } - /* Try architecture dependent optimization */ - value = arch_dep_replace_div_by_const(n); } if (value != n) { @@ -2970,10 +2665,10 @@ 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; @@ -3062,206 +2757,15 @@ 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 */ -/** - * Transform a DivMod node. - */ -static ir_node *transform_node_DivMod(ir_node *n) -{ - const ir_node *dummy; - ir_node *a = get_DivMod_left(n); - ir_node *b = get_DivMod_right(n); - ir_mode *mode = get_DivMod_resmode(n); - ir_node *va, *vb; - ir_graph *irg = get_irn_irg(n); - ir_tarval *ta, *tb; - int evaluated = 0; - - if (is_Const(b) && is_const_Phi(a)) { - /* check for Div(Phi, Const) */ - va = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_div, mode, 0); - vb = apply_binop_on_phi(a, get_Const_tarval(b), (eval_func) tarval_mod, mode, 0); - if (va && vb) { - DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); - DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); - goto make_tuple; - } - } - else if (is_Const(a) && is_const_Phi(b)) { - /* check for Div(Const, Phi) */ - va = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_div, mode, 1); - vb = apply_binop_on_phi(b, get_Const_tarval(a), (eval_func) tarval_mod, mode, 1); - if (va && vb) { - DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); - DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); - goto make_tuple; - } - } - else if (is_const_Phi(a) && is_const_Phi(b)) { - /* check for Div(Phi, Phi) */ - va = apply_binop_on_2_phis(a, b, (eval_func) tarval_div, mode); - vb = apply_binop_on_2_phis(a, b, (eval_func) tarval_mod, mode); - if (va && vb) { - DBG_OPT_ALGSIM0(n, va, FS_OPT_CONST_PHI); - DBG_OPT_ALGSIM0(n, vb, FS_OPT_CONST_PHI); - goto make_tuple; - } - } - - ta = value_of(a); - tb = value_of(b); - if (tb != tarval_bad) { - if (tb == get_mode_one(get_tarval_mode(tb))) { - va = a; - vb = new_r_Const(irg, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else if (ta != tarval_bad) { - ir_tarval *resa, *resb; - resa = tarval_div(ta, tb); - if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through - Jmp for X result!? */ - resb = tarval_mod(ta, tb); - if (resb == tarval_bad) return n; /* Causes exception! */ - va = new_r_Const(irg, resa); - vb = new_r_Const(irg, resb); - DBG_OPT_CSTEVAL(n, va); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else if (mode_is_signed(mode) && tb == get_mode_minus_one(mode)) { - va = new_rd_Minus(get_irn_dbg_info(n), get_nodes_block(n), a, mode); - vb = new_r_Const(irg, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, va); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else { /* Try architecture dependent optimization */ - va = a; - vb = b; - arch_dep_replace_divmod_by_const(&va, &vb, n); - evaluated = va != NULL; - } - } else if (a == b) { - if (value_not_zero(a, &dummy)) { - /* a/a && a != 0 */ - va = new_r_Const(irg, get_mode_one(mode)); - vb = new_r_Const(irg, get_mode_null(mode)); - DBG_OPT_CSTEVAL(n, va); - DBG_OPT_CSTEVAL(n, vb); - goto make_tuple; - } else { - /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */ - return n; - } - } else if (ta == get_mode_null(mode) && value_not_zero(b, &dummy)) { - /* 0 / non-Const = 0 */ - vb = va = a; - goto make_tuple; - } - - if (evaluated) { /* replace by tuple */ - ir_node *mem, *blk; - -make_tuple: - mem = get_DivMod_mem(n); - /* skip a potential Pin */ - mem = skip_Pin(mem); - - blk = get_nodes_block(n); - turn_into_tuple(n, pn_DivMod_max); - set_Tuple_pred(n, pn_DivMod_M, mem); - set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(blk)); - set_Tuple_pred(n, pn_DivMod_X_except, get_irg_bad(irg)); /*no exception*/ - set_Tuple_pred(n, pn_DivMod_res_div, va); - set_Tuple_pred(n, pn_DivMod_res_mod, vb); - } - - return n; -} /* transform_node_DivMod */ - -/** - * Optimize x / c to x * (1/c) - */ -static ir_node *transform_node_Quot(ir_node *n) -{ - ir_mode *mode = get_Quot_resmode(n); - ir_node *oldn = n; - - if (get_mode_arithmetic(mode) == irma_ieee754) { - ir_node *b = get_Quot_right(n); - ir_tarval *tv = value_of(b); - - if (tv != tarval_bad) { - int rem = tarval_fp_ops_enabled(); - - /* - * Floating point constant folding might be disabled here to - * prevent rounding. - * However, as we check for exact result, doing it is safe. - * Switch it on. - */ - tarval_enable_fp_ops(1); - tv = tarval_quo(get_mode_one(mode), tv); - tarval_enable_fp_ops(rem); - - /* Do the transformation if the result is either exact or we are not - using strict rules. */ - if (tv != tarval_bad && - (tarval_ieee754_get_exact() || (get_irg_fp_model(get_irn_irg(n)) & fp_strict_algebraic) == 0)) { - ir_node *blk = get_nodes_block(n); - ir_graph *irg = get_irn_irg(blk); - ir_node *c = new_r_Const(irg, tv); - ir_node *a = get_Quot_left(n); - ir_node *m = new_rd_Mul(get_irn_dbg_info(n), blk, a, c, mode); - ir_node *mem = get_Quot_mem(n); - - /* skip a potential Pin */ - mem = skip_Pin(mem); - turn_into_tuple(n, pn_Quot_max); - set_Tuple_pred(n, pn_Quot_M, mem); - set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(blk)); - set_Tuple_pred(n, pn_Quot_X_except, get_irg_bad(irg)); - set_Tuple_pred(n, pn_Quot_res, m); - DBG_OPT_ALGSIM1(oldn, a, b, m, FS_OPT_FP_INV_MUL); - } - } - } - return n; -} /* transform_node_Quot */ - -/** - * 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. * @@ -3287,13 +2791,13 @@ static ir_node *transform_node_Cond(ir_node *n) Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */ ir_node *blk = get_nodes_block(n); jmp = new_r_Jmp(blk); - turn_into_tuple(n, pn_Cond_max); + turn_into_tuple(n, pn_Cond_max+1); if (ta == tarval_b_true) { - set_Tuple_pred(n, pn_Cond_false, 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); @@ -3404,12 +2908,203 @@ static ir_node *transform_bitwise_distributive(ir_node *n, add_identities(n); } - DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); - return n; - } + DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND); + return n; + } + } + + return n; +} + +/** + * Create a 0 constant of given mode. + */ +static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode) +{ + ir_tarval *tv = get_mode_null(mode); + ir_node *cnst = new_r_Const(irg, tv); + + return cnst; +} + +static bool is_shiftop(const ir_node *n) +{ + return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n); +} + +/** + * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is good for things like x-(x&y) esp. in 186.crafty. + */ +static ir_node *transform_node_shift_bitop(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *left; + ir_node *bitop_left; + ir_node *bitop_right; + ir_op *op_left; + ir_node *block; + dbg_info *dbgi; + ir_node *new_shift; + ir_node *new_bitop; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_shift; + + if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + return n; + + assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n)); + + if (!is_Const(right)) + return n; + + left = get_binop_left(n); + op_left = get_irn_op(left); + if (op_left != op_And && op_left != op_Or && op_left != op_Eor) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) { + /* TODO: test if sign bit is affectes */ + return n; + } + + bitop_right = get_binop_right(left); + if (!is_Const(bitop_right)) + return n; + + bitop_left = get_binop_left(left); + + block = get_nodes_block(n); + dbgi = get_irn_dbg_info(n); + tv1 = get_Const_tarval(bitop_right); + tv2 = get_Const_tarval(right); + + assert(get_tarval_mode(tv1) == mode); + + if (is_Shl(n)) { + new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shl(tv1, tv2); + } else if (is_Shr(n)) { + new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shr(tv1, tv2); + } else if (is_Shrs(n)) { + new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_shrs(tv1, tv2); + } else { + assert(is_Rotl(n)); + new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); + tv_shift = tarval_rotl(tv1, tv2); + } + + assert(get_tarval_mode(tv_shift) == mode); + irg = get_irn_irg(n); + new_const = new_r_Const(irg, tv_shift); + + if (op_left == op_And) { + new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode); + } else if (op_left == op_Or) { + new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode); + } else { + assert(op_left == op_Eor); + new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode); + } + + return new_bitop; +} + +/** + * normalisation: (x >> c1) & c2 to (x & (c2<> c1 + * (we can use: + * - and, or, xor instead of & + * - Shl, Shr, Shrs, rotl instead of >> + * (with a special case for Or/Xor + Shrs) + * + * This normalisation is usually good for the backend since << C can often be + * matched as address-mode. + */ +static ir_node *transform_node_bitop_shift(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + ir_node *left = get_binop_left(n); + ir_node *right = get_binop_right(n); + ir_mode *mode = get_irn_mode(n); + ir_node *shift_left; + ir_node *shift_right; + ir_node *block; + dbg_info *dbg_bitop; + dbg_info *dbg_shift; + ir_node *new_bitop; + ir_node *new_shift; + ir_node *new_const; + ir_tarval *tv1; + ir_tarval *tv2; + ir_tarval *tv_bitop; + + if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2)) + return n; + + assert(is_And(n) || is_Or(n) || is_Eor(n)); + if (!is_Const(right) || !is_shiftop(left)) + return n; + + shift_left = get_binop_left(left); + shift_right = get_binop_right(left); + if (!is_Const(shift_right)) + return n; + + /* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */ + if (is_Shrs(left)) { + /* TODO this could be improved */ + return n; + } + + irg = get_irn_irg(n); + block = get_nodes_block(n); + dbg_bitop = get_irn_dbg_info(n); + dbg_shift = get_irn_dbg_info(left); + tv1 = get_Const_tarval(shift_right); + tv2 = get_Const_tarval(right); + assert(get_tarval_mode(tv2) == mode); + + if (is_Shl(left)) { + tv_bitop = tarval_shr(tv2, tv1); + } else if (is_Shr(left)) { + tv_bitop = tarval_shl(tv2, tv1); + } else { + assert(is_Rotl(left)); + tv_bitop = tarval_rotl(tv2, tarval_neg(tv1)); + } + new_const = new_r_Const(irg, tv_bitop); + + if (is_And(n)) { + new_bitop = new_rd_And(dbg_bitop, block, shift_left, new_const, mode); + } else if (is_Or(n)) { + new_bitop = new_rd_Or(dbg_bitop, block, shift_left, new_const, mode); + } else { + assert(is_Eor(n)); + new_bitop = new_rd_Eor(dbg_bitop, block, shift_left, new_const, mode); } - return n; + if (is_Shl(left)) { + new_shift = new_rd_Shl(dbg_shift, block, new_bitop, shift_right, mode); + } else if (is_Shr(left)) { + new_shift = new_rd_Shr(dbg_shift, block, new_bitop, shift_right, mode); + } else if (is_Rotl(left)) { + assert(is_Rotl(left)); + new_shift = new_rd_Rotl(dbg_shift, block, new_bitop, shift_right, mode); + } + + return new_shift; } /** @@ -3423,23 +3118,43 @@ static ir_node *transform_node_And(ir_node *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)) + && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_graph *irg = get_irn_irg(n); + ir_node *zero = create_zero_const(irg, a_mode); + return new_rd_Cmp(dbgi, block, or, zero, ir_relation_equal); + } + } + 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)) { ir_node *op = get_Not_op(b); @@ -3543,7 +3258,7 @@ static ir_node *transform_node_And(ir_node *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)) { + b_vrp->bits_not_set), get_Const_tarval(a)) == ir_relation_equal)) { return b; @@ -3551,11 +3266,13 @@ static ir_node *transform_node_And(ir_node *n) 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)) { + a_vrp->bits_not_set), get_Const_tarval(b)) == ir_relation_equal)) { 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 */ @@ -3597,23 +3314,24 @@ static ir_node *transform_node_Eor(ir_node *n) ir_node *b = get_Eor_right(n); ir_mode *mode = get_irn_mode(n); - HANDLE_BINOP_PHI((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); + /* 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); @@ -3641,6 +3359,9 @@ static ir_node *transform_node_Eor(ir_node *n) } n = transform_bitwise_distributive(n, transform_node_Eor); + if (is_Eor(n)) + n = transform_node_bitop_shift(n); + return n; } /* transform_node_Eor */ @@ -3656,15 +3377,14 @@ 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 */ @@ -3698,7 +3418,7 @@ static ir_node *transform_node_Not(ir_node *n) } } return n; -} /* transform_node_Not */ +} /** * Transform a Minus. @@ -3817,7 +3537,7 @@ 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); @@ -3849,7 +3569,7 @@ 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); @@ -3891,7 +3611,7 @@ 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: { @@ -3946,7 +3666,7 @@ 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: { @@ -3977,238 +3697,201 @@ static ir_node *transform_node_Proj_Mod(ir_node *proj) return proj; } /* transform_node_Proj_Mod */ -/** - * Transform a Proj(DivMod) with a non-zero value. - * Removes the exceptions and routes the memory to the NoMem node. - */ -static ir_node *transform_node_Proj_DivMod(ir_node *proj) -{ - ir_node *divmod = get_Proj_pred(proj); - ir_node *b = get_DivMod_right(divmod); - ir_node *res, *new_mem; - const ir_node *confirm; - long proj_nr; - - if (value_not_zero(b, &confirm)) { - /* DivMod(x, y) && y != 0 */ - proj_nr = get_Proj_proj(proj); - - if (confirm == NULL) { - /* we are sure we have a Const != 0 */ - new_mem = get_DivMod_mem(divmod); - new_mem = skip_Pin(new_mem); - set_DivMod_mem(divmod, new_mem); - set_irn_pinned(divmod, op_pin_state_floats); - } - - switch (proj_nr) { - - case pn_DivMod_X_regular: - return new_r_Jmp(get_nodes_block(divmod)); - - case pn_DivMod_X_except: { - /* we found an exception handler, remove it */ - ir_graph *irg = get_irn_irg(proj); - DBG_OPT_EXC_REM(proj); - return get_irg_bad(irg); - } - - case pn_DivMod_M: { - ir_graph *irg = get_irn_irg(proj); - res = get_DivMod_mem(divmod); - new_mem = get_irg_no_mem(irg); - - if (confirm) { - /* This node can only float up to the Confirm block */ - new_mem = new_r_Pin(get_nodes_block(confirm), new_mem); - } - /* this is a DivMod without exception, we can remove the memory edge */ - set_DivMod_mem(divmod, new_mem); - return res; - } - - case pn_DivMod_res_mod: - if (get_DivMod_left(divmod) == b) { - /* a % a = 0 if a != 0 */ - ir_graph *irg = get_irn_irg(proj); - ir_mode *mode = get_irn_mode(proj); - ir_node *res = new_r_Const(irg, get_mode_null(mode)); - - DBG_OPT_CSTEVAL(divmod, res); - return res; - } - } - } - return proj; -} /* transform_node_Proj_DivMod */ - /** * 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); + 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 (!get_opt_unreachable_code()) + return n; - if (tb != tarval_bad) { - /* we have a constant switch */ - long num = get_Proj_proj(proj); + if (mode_is_int(get_irn_mode(b))) { + ir_tarval *tb = value_of(b); - 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 (tb != tarval_bad) { + /* we have a constant switch */ + long num = get_Proj_proj(proj); - if (!(tarval_cmp( - tarval_and( b_vrp->bits_set, tp), - b_vrp->bits_set - ) == pn_Cmp_Eq)) { + if (num != get_Cond_default_proj(n)) { /* we cannot optimize default Proj's yet */ + if (get_tarval_long(tb) == num) { + /* Do NOT create a jump here, or we will have 2 control flow ops + * in a block. This case is optimized away in optimize_cf(). */ + return proj; + } else { + ir_graph *irg = get_irn_irg(proj); + /* this case will NEVER be taken, kill it */ + return new_r_Bad(irg, mode_X); + } + } + } else { + long num = get_Proj_proj(proj); + vrp_attr *b_vrp = vrp_get_info(b); + if (num != get_Cond_default_proj(n) && b_vrp) { + /* Try handling with vrp data. We only remove dead parts. */ + ir_tarval *tp = new_tarval_from_long(num, get_irn_mode(b)); + + if (b_vrp->range_type == VRP_RANGE) { + ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); + ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); + + if ((cmp_result & ir_relation_greater) == cmp_result + && (cmp_result2 & ir_relation_less) == cmp_result2) { ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } + } else if (b_vrp->range_type == VRP_ANTIRANGE) { + ir_relation cmp_result = tarval_cmp(b_vrp->range_bottom, tp); + ir_relation cmp_result2 = tarval_cmp(b_vrp->range_top, tp); - if (!(tarval_cmp( - tarval_and( - tarval_not(tp), - tarval_not(b_vrp->bits_not_set)), - tarval_not(b_vrp->bits_not_set)) - == pn_Cmp_Eq)) { + if ((cmp_result & ir_relation_less_equal) == cmp_result + && (cmp_result2 & ir_relation_greater_equal) == cmp_result2) { ir_graph *irg = get_irn_irg(proj); - return get_irg_bad(irg); + return new_r_Bad(irg, mode_X); } + } + if (!(tarval_cmp( + tarval_and( b_vrp->bits_set, tp), + b_vrp->bits_set + ) == ir_relation_equal)) { + ir_graph *irg = get_irn_irg(proj); + return new_r_Bad(irg, mode_X); + } + if (!(tarval_cmp( + tarval_and( + tarval_not(tp), + tarval_not(b_vrp->bits_not_set)), + tarval_not(b_vrp->bits_not_set)) + == ir_relation_equal)) { + ir_graph *irg = get_irn_irg(proj); + return new_r_Bad(irg, mode_X); } } } } return proj; -} /* transform_node_Proj_Cond */ +} /** - * Create a 0 constant of given mode. - */ -static ir_node *create_zero_const(ir_graph *irg, ir_mode *mode) -{ - ir_tarval *tv = get_mode_null(mode); - ir_node *cnst = new_r_Const(irg, tv); - - return cnst; + * return true if the operation returns a value with exactly 1 bit set + */ +static bool is_single_bit(const ir_node *node) +{ + /* a first implementation, could be extended with vrp and others... */ + if (is_Shl(node)) { + ir_node *shl_l = get_Shl_left(node); + ir_mode *mode = get_irn_mode(node); + int modulo = get_mode_modulo_shift(mode); + /* this works if we shift a 1 and we have modulo shift */ + if (is_Const(shl_l) && is_Const_one(shl_l) + && 0 < modulo && modulo <= (int)get_mode_size_bits(mode)) { + return true; + } + } else if (is_Const(node)) { + ir_tarval *tv = get_Const_tarval(node); + return tarval_is_single_bit(tv); + } + return false; } /** * 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 = NULL; - 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(get_irn_mode(left))) { - 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_mode *mode = get_irn_mode(left); - 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 */ - if (mode_is_int(get_irn_mode(left))) { + if (mode_is_int(mode)) { unsigned lop = get_irn_opcode(left); if (lop == get_irn_opcode(right)) { @@ -4221,7 +3904,7 @@ 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: @@ -4234,25 +3917,25 @@ 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; @@ -4266,13 +3949,13 @@ 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; @@ -4281,7 +3964,7 @@ 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; @@ -4303,8 +3986,8 @@ 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, get_irn_mode(left)); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } @@ -4320,11 +4003,12 @@ 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, get_irn_mode(left)); - changed |= 1; + right = create_zero_const(irg, mode); + changed = true; DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_OP_OP); } } + if (is_And(left) && is_Const(right)) { ir_node *ll = get_binop_left(left); ir_node *lr = get_binop_right(left); @@ -4346,7 +4030,7 @@ 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; } } } @@ -4355,26 +4039,64 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) if (is_Const(right) && is_Const_null(right) && is_Eor(left)) { right = get_Eor_right(left); left = get_Eor_left(left); - changed |= 1; + changed = true; } } /* 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) + && (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; + } + } /* 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; } @@ -4387,12 +4109,11 @@ 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; } /* @@ -4403,7 +4124,7 @@ 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)) { @@ -4416,13 +4137,13 @@ 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); } } @@ -4438,73 +4159,53 @@ 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); - } - /* 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)) { @@ -4517,7 +4218,7 @@ 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); } } @@ -4543,7 +4244,7 @@ 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); } } @@ -4555,7 +4256,7 @@ 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); } } @@ -4563,7 +4264,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) } /* == 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; @@ -4578,9 +4279,9 @@ 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; } @@ -4599,8 +4300,8 @@ 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); } } @@ -4616,9 +4317,9 @@ 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; } } @@ -4642,16 +4343,16 @@ 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; @@ -4674,16 +4375,16 @@ 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; @@ -4709,16 +4410,16 @@ 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; @@ -4727,16 +4428,16 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) } /* 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) || - (is_DivMod(op) && get_Proj_proj(left) == pn_DivMod_res_mod)) { + if (is_Mod(op) && get_Proj_proj(left) == pn_Mod_res) { ir_node *c = get_binop_right(op); if (is_Const(c)) { @@ -4751,7 +4452,7 @@ 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); } } @@ -4759,15 +4460,15 @@ 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. @@ -4785,10 +4486,12 @@ 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; } @@ -4836,7 +4539,7 @@ 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; @@ -4854,7 +4557,7 @@ static ir_node *transform_node_Proj_Bound(ir_node *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) @@ -4866,28 +4569,71 @@ static ir_node *transform_node_Proj(ir_node *proj) return proj; } /* transform_node_Proj */ -/** - * Move Confirms down through Phi nodes. - */ +static bool is_block_unreachable(const ir_node *block) +{ + const ir_graph *irg = get_irn_irg(block); + if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) + return false; + return get_Block_dom_depth(block) < 0; +} + +static ir_node *transform_node_Block(ir_node *block) +{ + ir_graph *irg = get_irn_irg(block); + int arity = get_irn_arity(block); + ir_node *bad = NULL; + int i; + + if (!is_irg_state(irg, IR_GRAPH_STATE_BAD_BLOCK)) + return block; + + for (i = 0; i < arity; ++i) { + ir_node *pred = get_Block_cfgpred(block, i); + ir_node *pred_block = get_nodes_block(pred); + if (!is_Bad(pred) && !is_block_unreachable(pred_block)) + continue; + if (bad == NULL) + bad = new_r_Bad(irg, mode_X); + set_irn_n(block, i, bad); + } + + return block; +} + static ir_node *transform_node_Phi(ir_node *phi) { - int i, n; - ir_mode *mode = get_irn_mode(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; + + /* Set phi-operands for bad-block inputs to bad */ + for (i = 0; i < n; ++i) { + ir_node *pred = get_Block_cfgpred(block, i); + if (is_Bad(pred) || is_block_unreachable(get_nodes_block(pred))) { + if (bad == NULL) + bad = new_r_Bad(irg, mode); + set_irn_n(phi, i, bad); + } + } + /* Move Confirms down through Phi nodes. */ if (mode_is_reference(mode)) { n = get_irn_arity(phi); /* Beware of Phi0 */ if (n > 0) { - ir_node *pred = get_irn_n(phi, 0); - ir_node *bound, *new_phi, *block, **in; - pn_Cmp pnc; + ir_node *pred = get_irn_n(phi, 0); + ir_node *bound, *new_phi, *block, **in; + ir_relation relation; if (! is_Confirm(pred)) return phi; - bound = get_Confirm_bound(pred); - pnc = get_Confirm_cmp(pred); + bound = get_Confirm_bound(pred); + relation = get_Confirm_relation(pred); NEW_ARR_A(ir_node *, in, n); in[0] = get_Confirm_value(pred); @@ -4897,18 +4643,18 @@ static ir_node *transform_node_Phi(ir_node *phi) if (! is_Confirm(pred) || get_Confirm_bound(pred) != bound || - get_Confirm_cmp(pred) != pnc) + get_Confirm_relation(pred) != relation) return phi; in[i] = get_Confirm_value(pred); } /* move the Confirm nodes "behind" the Phi */ block = get_irn_n(phi, -1); new_phi = new_r_Phi(block, n, in, get_irn_mode(phi)); - return new_r_Confirm(block, new_phi, bound, pnc); + return new_r_Confirm(block, new_phi, bound, relation); } } return phi; -} /* transform_node_Phi */ +} /** * Returns the operands of a commutative bin-op, if one operand is @@ -5091,7 +4837,7 @@ static ir_node *transform_node_Or_Rotl(ir_node *irn_or) /* 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)) { + if (!ir_is_negated_value(c1, c2)) { return irn_or; } @@ -5102,6 +4848,21 @@ static ir_node *transform_node_Or_Rotl(ir_node *irn_or) return n; } /* transform_node_Or_Rotl */ +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 (relation == ir_relation_less_greater) + return true; + + if (!mode_is_signed(mode) && is_Const(right) && is_Const_null(right)) + return relation == ir_relation_greater; + return false; +} + /** * Transform an Or. */ @@ -5125,18 +4886,36 @@ static ir_node *transform_node_Or(ir_node *n) return n; } - /* 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; - - return new_rd_Proj(dbgi, pred_a, mode_b, new_pnc); + /* 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); + } + /* Cmp(a!=b) or Cmp(c!=d) => Cmp((a^b)|(c^d) != 0) */ + if (is_cmp_unequal(a) && is_cmp_unequal(b) + && !mode_is_float(get_irn_mode(a_left)) + && !mode_is_float(get_irn_mode(b_left)) + && values_in_mode(get_irn_mode(a_left), get_irn_mode(b_left))) { + ir_graph *irg = get_irn_irg(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_mode *a_mode = get_irn_mode(a_left); + ir_mode *b_mode = get_irn_mode(b_left); + ir_node *xora = new_rd_Eor(dbgi, block, a_left, a_right, a_mode); + ir_node *xorb = new_rd_Eor(dbgi, block, b_left, b_right, b_mode); + ir_node *conv = new_rd_Conv(dbgi, block, xorb, a_mode); + ir_node *or = new_rd_Or(dbgi, block, xora, conv, a_mode); + ir_node *zero = create_zero_const(irg, a_mode); + return new_rd_Cmp(dbgi, block, or, zero, ir_relation_less_greater); } } @@ -5149,6 +4928,8 @@ static ir_node *transform_node_Or(ir_node *n) return n; n = transform_bitwise_distributive(n, transform_node_Or); + if (is_Or(n)) + n = transform_node_bitop_shift(n); return n; } /* transform_node_Or */ @@ -5166,9 +4947,11 @@ static ir_node *transform_node_shift(ir_node *n) { ir_node *left, *right; ir_mode *mode; + ir_mode *count_mode; ir_tarval *tv1, *tv2, *res; ir_node *in[2], *irn, *block; ir_graph *irg; + int modulo_shf; left = get_binop_left(n); @@ -5177,7 +4960,7 @@ static ir_node *transform_node_shift(ir_node *n) return n; right = get_binop_right(n); - tv1 = value_of(right); + tv1 = value_of(right); if (tv1 == tarval_bad) return n; @@ -5185,132 +4968,65 @@ 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); - - /* 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)); - - assert(modulo_shf >= (int) get_mode_size_bits(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); - } - - 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 */ - block = get_nodes_block(n); - - in[0] = get_binop_left(left); - in[1] = new_r_Const(irg, res); - - irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in); - - DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); - - return transform_node(irn); -} /* transform_node_shift */ - -/** - * normalisation: (x & c1) >> c2 to (x >> c2) & (c1 >> c2) - * (we can use: - * - and, or, xor instead of & - * - Shl, Shr, Shrs, rotl instead of >> - * (with a special case for Or/Xor + Shrs) - */ -static ir_node *transform_node_bitop_shift(ir_node *n) -{ - ir_node *left; - ir_node *right = get_binop_right(n); - ir_mode *mode = get_irn_mode(n); - ir_node *bitop_left; - ir_node *bitop_right; - ir_op *op_left; - ir_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 */ + count_mode = get_tarval_mode(tv1); + if (get_tarval_mode(tv2) != count_mode) { + /* TODO: search bigger mode or something and convert... */ return n; } - bitop_right = get_binop_right(left); - if (!is_Const(bitop_right)) - return n; + mode = get_irn_mode(n); + modulo_shf = get_mode_modulo_shift(mode); - bitop_left = get_binop_left(left); + if (modulo_shf > 0) { + ir_tarval *modulo_mask = new_tarval_from_long(modulo_shf-1, count_mode); - block = get_nodes_block(n); - dbgi = get_irn_dbg_info(n); - tv1 = get_Const_tarval(bitop_right); - tv2 = get_Const_tarval(right); + /* 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)); - assert(get_tarval_mode(tv1) == mode); + tv1 = tarval_and(tv1, modulo_mask); + tv2 = tarval_and(tv2, modulo_mask); + } + res = tarval_add(tv1, tv2); + irg = get_irn_irg(n); - 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); + /* 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 { - assert(is_Rotl(n)); - new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode); - tv_shift = tarval_rotl(tv1, tv2); - } + long bits = get_mode_size_bits(mode); + ir_tarval *mode_size = new_tarval_from_long(bits, count_mode); - assert(get_tarval_mode(tv_shift) == mode); - irg = get_irn_irg(n); - new_const = new_r_Const(irg, tv_shift); + /* 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); + } - 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_r_Const(irg, get_mode_null(mode)); + } } - return new_bitop; + /* 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); + in[1] = new_r_Const(irg, res); + + irn = new_ir_node(NULL, get_Block_irg(block), block, get_irn_op(n), mode, 2, in); + + DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT); + + return transform_node(irn); } /** @@ -5336,7 +5052,7 @@ 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)); @@ -5358,7 +5074,7 @@ 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); @@ -5394,8 +5110,8 @@ 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) { @@ -5404,7 +5120,7 @@ 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); @@ -5431,7 +5147,7 @@ typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block, * then we can use that to minimize the value of Add(x, const) or * Sub(Const, x). In particular this often avoids 1 instruction in some * backends for the Shift(x, Sub(Const, y)) case because it can be replaced - * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed. + * by Shift(x, Minus(y)) which does not need an explicit Const constructed. */ static ir_node *transform_node_shift_modulo(ir_node *n, new_shift_func new_shift) @@ -5519,7 +5235,7 @@ static ir_node *transform_node_Shr(ir_node *n) if (is_Shr(n)) n = transform_node_shl_shr(n); if (is_Shr(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shr */ @@ -5534,13 +5250,23 @@ static ir_node *transform_node_Shrs(ir_node *n) ir_node *b = get_Shrs_right(n); ir_mode *mode = get_irn_mode(n); + if (is_oversize_shift(n)) { + ir_node *block = get_nodes_block(n); + dbg_info *dbgi = get_irn_dbg_info(n); + ir_mode *cmode = get_irn_mode(b); + long val = get_mode_size_bits(cmode)-1; + ir_graph *irg = get_irn_irg(n); + ir_node *cnst = new_r_Const_long(irg, cmode, val); + return new_rd_Shrs(dbgi, block, a, cnst, mode); + } + HANDLE_BINOP_PHI((eval_func) tarval_shrs, a, b, c, mode); n = transform_node_shift(n); if (is_Shrs(n)) n = transform_node_shift_modulo(n, new_rd_Shrs); if (is_Shrs(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shrs */ @@ -5563,7 +5289,7 @@ static ir_node *transform_node_Shl(ir_node *n) if (is_Shl(n)) n = transform_node_shl_shr(n); if (is_Shl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Shl */ @@ -5582,7 +5308,7 @@ static ir_node *transform_node_Rotl(ir_node *n) n = transform_node_shift(n); if (is_Rotl(n)) - n = transform_node_bitop_shift(n); + n = transform_node_shift_bitop(n); return n; } /* transform_node_Rotl */ @@ -5654,17 +5380,14 @@ 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; - } - continue; - } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) { + ir_node *block; + /* no need to keep Bad */ + if (is_Bad(ka)) continue; - } else if (is_Bad(ka)) { - /* no need to keep Bad */ + /* do not keep unreachable code */ + block = is_Block(ka) ? ka : get_nodes_block(ka); + if (is_block_unreachable(block)) continue; - } in[j++] = ka; } if (j != n_keepalives) @@ -5672,7 +5395,7 @@ static ir_node *transform_node_End(ir_node *n) 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; @@ -5691,16 +5414,115 @@ bool is_negated_value(ir_node *a, ir_node *b) return false; } +static const ir_node *skip_upconv(const ir_node *node) +{ + while (is_Conv(node)) { + ir_mode *mode = get_irn_mode(node); + const ir_node *op = get_Conv_op(node); + ir_mode *op_mode = get_irn_mode(op); + if (!smaller_mode(op_mode, mode)) + break; + node = op; + } + return node; +} + +int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true, + const ir_node *mux_false) +{ + ir_node *cmp_left; + ir_node *cmp_right; + ir_mode *mode; + ir_relation relation; + + if (!is_Cmp(sel)) + return 0; + + /** + * Note further that these optimization work even for floating point + * with NaN's because -NaN == NaN. + * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs + * transformations. + */ + mode = get_irn_mode(mux_true); + if (mode_honor_signed_zeros(mode)) + return 0; + + /* must be <, <=, >=, > */ + relation = get_Cmp_relation(sel); + if ((relation & ir_relation_less_greater) == 0) + return 0; + + if (!ir_is_negated_value(mux_true, mux_false)) + return 0; + + mux_true = skip_upconv(mux_true); + mux_false = skip_upconv(mux_false); + + /* must be x cmp 0 */ + cmp_right = get_Cmp_right(sel); + if (!is_Const(cmp_right) || !is_Const_null(cmp_right)) + return 0; + + cmp_left = get_Cmp_left(sel); + if (cmp_left == mux_false) { + if (relation & ir_relation_less) { + return 1; + } else { + assert(relation & ir_relation_greater); + return -1; + } + } else if (cmp_left == mux_true) { + if (relation & ir_relation_less) { + return -1; + } else { + assert(relation & ir_relation_greater); + return 1; + } + } + + return 0; +} + +ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true, + ir_node *mux_false) +{ + ir_node *cmp_left = get_Cmp_left(sel); + return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true; +} + /** * Optimize a Mux into some simpler cases. */ static ir_node *transform_node_Mux(ir_node *n) { - ir_node *oldn = n, *sel = get_Mux_sel(n); - ir_mode *mode = get_irn_mode(n); - ir_node *t = get_Mux_true(n); - ir_node *f = get_Mux_false(n); - ir_graph *irg = get_irn_irg(n); + ir_node *oldn = n; + ir_node *sel = get_Mux_sel(n); + ir_mode *mode = get_irn_mode(n); + ir_node *t = get_Mux_true(n); + ir_node *f = get_Mux_false(n); + ir_graph *irg = get_irn_irg(n); + + /* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */ + if (get_mode_arithmetic(mode) == irma_twos_complement) { + int abs = ir_mux_is_abs(sel, t, f); + if (abs != 0) { + dbg_info *dbgi = get_irn_dbg_info(n); + ir_node *block = get_nodes_block(n); + ir_node *op = ir_get_abs_op(sel, t, f); + int bits = get_mode_size_bits(mode); + ir_node *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1); + ir_node *sext = new_rd_Shrs(dbgi, block, op, shiftconst, mode); + ir_node *xorn = new_rd_Eor(dbgi, block, op, sext, mode); + ir_node *res; + if (abs > 0) { + res = new_rd_Sub(dbgi, block, xorn, sext, mode); + } else { + res = new_rd_Sub(dbgi, block, sext, xorn, mode); + } + return res; + } + } if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX)) return n; @@ -5761,21 +5583,20 @@ static ir_node *transform_node_Mux(ir_node *n) /* 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); - - 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; - - /* 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_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); } /* note: after normalization, false can only happen on default */ @@ -5838,77 +5659,69 @@ static ir_node *transform_node_Mux(ir_node *n) } } - if (is_Proj(sel)) { - ir_node *cmp = get_Proj_pred(sel); - long pn = get_Proj_proj(sel); + if (is_Cmp(sel)) { + ir_node *cmp_r = get_Cmp_right(sel); + if (is_Const(cmp_r) && is_Const_null(cmp_r)) { + ir_node *block = get_nodes_block(n); + ir_node *cmp_l = get_Cmp_left(sel); - /* - * 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) */ + if (mode_is_int(mode)) { + ir_relation relation = get_Cmp_relation(sel); + /* integer only */ + if ((relation == ir_relation_less_greater || relation == ir_relation_equal) && is_And(cmp_l)) { + /* Mux((a & b) != 0, c, 0) */ + ir_node *and_r = get_And_right(cmp_l); + ir_node *and_l; + + if (and_r == t && f == cmp_r) { + if (is_Const(t) && tarval_is_single_bit(get_Const_tarval(t))) { + if (relation == ir_relation_less_greater) { + /* Mux((a & 2^C) != 0, 2^C, 0) */ + n = cmp_l; + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); + } else { + /* Mux((a & 2^C) == 0, 2^C, 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), + block, cmp_l, t, mode); + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); + } + return n; + } + } + if (is_Shl(and_r)) { + ir_node *shl_l = get_Shl_left(and_r); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_r == t && f == cmp_r) { + if (relation == ir_relation_less_greater) { + /* (a & (1 << n)) != 0, (1 << n), 0) */ n = cmp_l; - DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_BITOP); + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); } else { - /* Mux((a & 2^C) == 0, 2^C, 0) */ + /* (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); + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); } return n; } } - if (is_Shl(and_r)) { - ir_node *shl_l = get_Shl_left(and_r); - if (is_Const(shl_l) && is_Const_one(shl_l)) { - if (and_r == t && f == cmp_r) { - if (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; + } + and_l = get_And_left(cmp_l); + if (is_Shl(and_l)) { + ir_node *shl_l = get_Shl_left(and_l); + if (is_Const(shl_l) && is_Const_one(shl_l)) { + if (and_l == t && f == cmp_r) { + if (relation == ir_relation_less_greater) { + /* ((1 << n) & a) != 0, (1 << n), 0) */ + n = cmp_l; + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); + } else { + /* ((1 << n) & a) == 0, (1 << n), 0) */ + n = new_rd_Eor(get_irn_dbg_info(n), + block, cmp_l, t, mode); + DBG_OPT_ALGSIM1(oldn, sel, sel, n, FS_OPT_MUX_TO_BITOP); } + return n; } } } @@ -5918,7 +5731,7 @@ static ir_node *transform_node_Mux(ir_node *n) } return n; -} /* transform_node_Mux */ +} /** * optimize Sync nodes that have other syncs as input we simply add the inputs @@ -5934,6 +5747,24 @@ static ir_node *transform_node_Sync(ir_node *n) int pred_arity; int j; + /* Remove Bad predecessors */ + if (is_Bad(pred)) { + del_Sync_n(n, i); + --arity; + continue; + } + + /* Remove duplicate predecessors */ + for (j = 0; j < i; ++j) { + if (get_Sync_pred(n, j) == pred) { + del_Sync_n(n, i); + --arity; + break; + } + } + if (j < i) + continue; + if (!is_Sync(pred)) { ++i; continue; @@ -5958,11 +5789,78 @@ static ir_node *transform_node_Sync(ir_node *n) } } + if (arity == 0) { + ir_graph *irg = get_irn_irg(n); + return new_r_Bad(irg, mode_M); + } + if (arity == 1) { + return get_Sync_pred(n, 0); + } + /* rehash the sync node */ add_identities(n); + return n; +} + +static ir_node *transform_node_Load(ir_node *n) +{ + /* if our memory predecessor is a load from the same address, then reuse the + * previous result */ + ir_node *mem = get_Load_mem(n); + ir_node *mem_pred; + + if (!is_Proj(mem)) + return n; + /* don't touch volatile loads */ + if (get_Load_volatility(n) == volatility_is_volatile) + return n; + mem_pred = get_Proj_pred(mem); + if (is_Load(mem_pred)) { + ir_node *pred_load = mem_pred; + + /* conservatively compare the 2 loads. TODO: This could be less strict + * with fixup code in some situations (like smaller/bigger modes) */ + if (get_Load_ptr(pred_load) != get_Load_ptr(n)) + return n; + if (get_Load_mode(pred_load) != get_Load_mode(n)) + return n; + /* all combinations of aligned/unaligned pred/n should be fine so we do + * not compare the unaligned attribute */ + { + ir_node *block = get_nodes_block(n); + ir_node *jmp = new_r_Jmp(block); + ir_graph *irg = get_irn_irg(n); + ir_node *bad = new_r_Bad(irg, mode_X); + ir_mode *mode = get_Load_mode(n); + ir_node *res = new_r_Proj(pred_load, mode, pn_Load_res); + ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad }; + ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in); + return tuple; + } + } else if (is_Store(mem_pred)) { + ir_node *pred_store = mem_pred; + ir_node *value = get_Store_value(pred_store); + + if (get_Store_ptr(pred_store) != get_Load_ptr(n)) + return n; + if (get_irn_mode(value) != get_Load_mode(n)) + return n; + /* all combinations of aligned/unaligned pred/n should be fine so we do + * not compare the unaligned attribute */ + { + ir_node *block = get_nodes_block(n); + ir_node *jmp = new_r_Jmp(block); + ir_graph *irg = get_irn_irg(n); + ir_node *bad = new_r_Bad(irg, mode_X); + ir_node *res = value; + ir_node *in[pn_Load_max+1] = { mem, res, jmp, bad }; + ir_node *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in); + return tuple; + } + } return n; -} /* transform_node_Sync */ +} /** * optimize a trampoline Call into a direct Call @@ -5975,7 +5873,7 @@ static ir_node *transform_node_Call(ir_node *call) ir_graph *irg; type_dbg_info *tdb; dbg_info *db; - int i, n_res, n_param; + size_t i, n_res, n_param; ir_variadicity var; if (! is_Proj(callee)) @@ -6019,9 +5917,6 @@ static ir_node *transform_node_Call(ir_node *call) } var = get_method_variadicity(mtp); set_method_variadicity(ctp, var); - if (var == variadicity_variadic) { - set_method_first_variadic_param_index(ctp, get_method_first_variadic_param_index(mtp) + 1); - } /* When we resolve a trampoline, the function must be called by a this-call */ set_method_calling_convention(ctp, get_method_calling_convention(mtp) | cc_this_call); set_method_additional_properties(ctp, get_method_additional_properties(mtp)); @@ -6088,35 +5983,34 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops switch (code) { CASE(Add); - CASE(Sub); - CASE(Mul); - CASE_PROJ_EX(Div); - CASE_PROJ_EX(Mod); - CASE_PROJ_EX(DivMod); - CASE(Quot); - CASE_PROJ_EX(Cmp); - CASE_PROJ_EX(Cond); CASE(And); + CASE(Block); + CASE(Call); + CASE(Cmp); + CASE(Conv); + CASE(End); CASE(Eor); - CASE(Not); CASE(Minus); - CASE_PROJ(Load); - CASE_PROJ(Store); - CASE_PROJ(Bound); - CASE_PROJ(CopyB); - CASE(Proj); - CASE(Phi); + CASE(Mul); + CASE(Mux); + CASE(Not); CASE(Or); + CASE(Phi); + CASE(Proj); + CASE(Rotl); CASE(Sel); + CASE(Shl); CASE(Shr); CASE(Shrs); - CASE(Shl); - CASE(Rotl); - CASE(Conv); - CASE(End); - CASE(Mux); + CASE(Sub); CASE(Sync); - CASE(Call); + CASE_PROJ(Bound); + CASE_PROJ(CopyB); + CASE_PROJ(Store); + CASE_PROJ_EX(Cond); + CASE_PROJ_EX(Div); + CASE_PROJ_EX(Load); + CASE_PROJ_EX(Mod); default: /* leave NULL */; } @@ -6135,35 +6029,35 @@ static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops #define N_IR_NODES 512 /** Compares the attributes of two Const nodes. */ -static int node_cmp_attr_Const(ir_node *a, ir_node *b) +static int node_cmp_attr_Const(const ir_node *a, const ir_node *b) { return get_Const_tarval(a) != get_Const_tarval(b); } /** Compares the attributes of two Proj nodes. */ -static int node_cmp_attr_Proj(ir_node *a, 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(ir_node *a, ir_node *b) +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 */ +} /** Compares the attributes of two Free nodes. */ -static int node_cmp_attr_Free(ir_node *a, ir_node *b) +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(ir_node *a, ir_node *b) +static int node_cmp_attr_SymConst(const ir_node *a, const ir_node *b) { const symconst_attr *pa = &a->attr.symc; const symconst_attr *pb = &b->attr.symc; @@ -6172,24 +6066,24 @@ static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) } /** Compares the attributes of two Call nodes. */ -static int node_cmp_attr_Call(ir_node *a, ir_node *b) +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 */ +} /** Compares the attributes of two Sel nodes. */ -static int node_cmp_attr_Sel(ir_node *a, ir_node *b) +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(ir_node *a, ir_node *b) +static int node_cmp_attr_Phi(const ir_node *a, const ir_node *b) { /* we can only enter this function if both nodes have the same number of inputs, hence it is enough to check if one of them is a Phi0 */ @@ -6198,48 +6092,48 @@ static int node_cmp_attr_Phi(ir_node *a, 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(ir_node *a, ir_node *b) +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(ir_node *a, ir_node *b) +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(ir_node *a, ir_node *b) +static int node_cmp_attr_Load(const ir_node *a, const ir_node *b) { if (get_Load_volatility(a) == volatility_is_volatile || get_Load_volatility(b) == volatility_is_volatile) /* 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 */ +} /** Compares the attributes of two Store nodes. */ -static int node_cmp_attr_Store(ir_node *a, ir_node *b) +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 */ +} /** Compares two exception attributes */ -static int node_cmp_exception(ir_node *a, ir_node *b) +static int node_cmp_exception(const ir_node *a, const ir_node *b) { const except_attr *ea = &a->attr.except; const except_attr *eb = &b->attr.except; @@ -6250,52 +6144,48 @@ static int node_cmp_exception(ir_node *a, ir_node *b) #define node_cmp_attr_Bound node_cmp_exception /** Compares the attributes of two Div nodes. */ -static int node_cmp_attr_Div(ir_node *a, ir_node *b) +static int node_cmp_attr_Div(const ir_node *a, const ir_node *b) { - const divmod_attr *ma = &a->attr.divmod; - const divmod_attr *mb = &b->attr.divmod; + 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 */ +} -/** Compares the attributes of two DivMod nodes. */ -static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) +/** Compares the attributes of two Mod nodes. */ +static int node_cmp_attr_Mod(const ir_node *a, const ir_node *b) { - const divmod_attr *ma = &a->attr.divmod; - const divmod_attr *mb = &b->attr.divmod; + 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_DivMod */ - -/** Compares the attributes of two Mod nodes. */ -static int node_cmp_attr_Mod(ir_node *a, ir_node *b) -{ - return node_cmp_attr_DivMod(a, b); -} /* node_cmp_attr_Mod */ +} -/** Compares the attributes of two Quot nodes. */ -static int node_cmp_attr_Quot(ir_node *a, ir_node *b) +static int node_cmp_attr_Cmp(const ir_node *a, const ir_node *b) { - return node_cmp_attr_DivMod(a, b); -} /* node_cmp_attr_Quot */ + 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(ir_node *a, ir_node *b) +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(ir_node *a, ir_node *b) +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 */ +} /** Compares the attributes of two ASM nodes. */ -static int node_cmp_attr_ASM(ir_node *a, ir_node *b) +static int node_cmp_attr_ASM(const ir_node *a, const ir_node *b) { int i, n; const ir_asm_constraint *ca; @@ -6308,29 +6198,31 @@ static int node_cmp_attr_ASM(ir_node *a, 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); @@ -6339,10 +6231,10 @@ static int node_cmp_attr_ASM(ir_node *a, ir_node *b) return 1; } return 0; -} /* node_cmp_attr_ASM */ +} /** Compares the inexistent attributes of two Dummy nodes. */ -static int node_cmp_attr_Dummy(ir_node *a, ir_node *b) +static int node_cmp_attr_Dummy(const ir_node *a, const ir_node *b) { (void) a; (void) b; @@ -6374,6 +6266,7 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) CASE(Call); CASE(Sel); CASE(Phi); + CASE(Cmp); CASE(Conv); CASE(Cast); CASE(Load); @@ -6381,9 +6274,7 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops) CASE(Confirm); CASE(ASM); CASE(Div); - CASE(DivMod); CASE(Mod); - CASE(Quot); CASE(Bound); CASE(Builtin); CASE(Dummy); @@ -6425,10 +6316,23 @@ int identities_cmp(const void *elt, const void *key) /* for pinned nodes, the block inputs must be equal */ if (get_irn_n(a, -1) != get_irn_n(b, -1)) return 1; - } else if (! get_opt_global_cse()) { - /* for block-local CSE both nodes must be in the same Block */ - if (get_nodes_block(a) != get_nodes_block(b)) - return 1; + } else { + ir_node *block_a = get_nodes_block(a); + ir_node *block_b = get_nodes_block(b); + if (! get_opt_global_cse()) { + /* for block-local CSE both nodes must be in the same Block */ + if (block_a != block_b) + return 1; + } else { + /* The optimistic approach would be to do nothing here. + * However doing GCSE optimistically produces a lot of partially dead code which appears + * to be worse in practice than the missed opportunities. + * So we use a very conservative variant here and only CSE if 1 value dominates the + * other. */ + if (!block_dominates(block_a, block_b) + && !block_dominates(block_b, block_a)) + return 1; + } } /* compare a->in[0..ins] with b->in[0..ins] */ @@ -6521,7 +6425,7 @@ 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; @@ -6570,100 +6474,6 @@ void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *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. * It can only be called if it is guaranteed that no other nodes @@ -6692,7 +6502,7 @@ 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; @@ -6705,7 +6515,7 @@ 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); @@ -6736,7 +6546,7 @@ 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); @@ -6751,10 +6561,6 @@ 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; @@ -6828,10 +6634,6 @@ ir_node *optimize_in_place_2(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 can verify the node, as it has no dead inputs any more. */ irn_verify(n); @@ -6859,8 +6661,6 @@ 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. */ @@ -6903,7 +6703,7 @@ 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: \ @@ -6929,7 +6729,7 @@ 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);