stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
jdk-24/src/hotspot/share/opto/subnode.cpp
Roland Westrelin 3d460bd295 8256655: rework long counted loop handling 
Reviewed-by: kvn, thartmann, vlivanov
2020-12-01 07:30:51 +00:00

1765 lines
68 KiB
C++
Raw Blame History

/*
* Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "compiler/compileLog.hpp"
#include "gc/shared/barrierSet.hpp"
#include "gc/shared/c2/barrierSetC2.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/loopnode.hpp"
#include "opto/matcher.hpp"
#include "opto/movenode.hpp"
#include "opto/mulnode.hpp"
#include "opto/opcodes.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
#include "runtime/sharedRuntime.hpp"
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
#include "math.h"
//=============================================================================
//------------------------------Identity---------------------------------------
// If right input is a constant 0, return the left input.
Node* SubNode::Identity(PhaseGVN* phase) {
assert(in(1) != this, "Must already have called Value");
assert(in(2) != this, "Must already have called Value");
// Remove double negation
const Type *zero = add_id();
if( phase->type( in(1) )->higher_equal( zero ) &&
in(2)->Opcode() == Opcode() &&
phase->type( in(2)->in(1) )->higher_equal( zero ) ) {
return in(2)->in(2);
}
// Convert "(X+Y) - Y" into X and "(X+Y) - X" into Y
if (in(1)->Opcode() == Op_AddI) {
if (in(1)->in(2) == in(2)) {
return in(1)->in(1);
}
if (in(1)->in(1) == in(2)) {
return in(1)->in(2);
}
// Also catch: "(X + Opaque2(Y)) - Y". In this case, 'Y' is a loop-varying
// trip counter and X is likely to be loop-invariant (that's how O2 Nodes
// are originally used, although the optimizer sometimes jiggers things).
// This folding through an O2 removes a loop-exit use of a loop-varying
// value and generally lowers register pressure in and around the loop.
if (in(1)->in(2)->Opcode() == Op_Opaque2 && in(1)->in(2)->in(1) == in(2)) {
return in(1)->in(1);
}
}
return ( phase->type( in(2) )->higher_equal( zero ) ) ? in(1) : this;
}
//------------------------------Value------------------------------------------
// A subtract node differences it's two inputs.
const Type* SubNode::Value_common(PhaseTransform *phase) const {
const Node* in1 = in(1);
const Node* in2 = in(2);
// Either input is TOP ==> the result is TOP
const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
if( t1 == Type::TOP ) return Type::TOP;
const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
if( t2 == Type::TOP ) return Type::TOP;
// Not correct for SubFnode and AddFNode (must check for infinity)
// Equal? Subtract is zero
if (in1->eqv_uncast(in2)) return add_id();
// Either input is BOTTOM ==> the result is the local BOTTOM
if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
return bottom_type();
return NULL;
}
const Type* SubNode::Value(PhaseGVN* phase) const {
const Type* t = Value_common(phase);
if (t != NULL) {
return t;
}
const Type* t1 = phase->type(in(1));
const Type* t2 = phase->type(in(2));
return sub(t1,t2); // Local flavor of type subtraction
}
SubNode* SubNode::make(Node* in1, Node* in2, BasicType bt) {
switch (bt) {
case T_INT:
return new SubINode(in1, in2);
case T_LONG:
return new SubLNode(in1, in2);
default:
fatal("Not implemented for %s", type2name(bt));
}
return NULL;
}
//=============================================================================
//------------------------------Helper function--------------------------------
static bool is_cloop_increment(Node* inc) {
precond(inc->Opcode() == Op_AddI || inc->Opcode() == Op_AddL);
if (!inc->in(1)->is_Phi()) {
return false;
}
const PhiNode* phi = inc->in(1)->as_Phi();
if (!phi->region()->is_CountedLoop()) {
return false;
}
return inc == phi->region()->as_CountedLoop()->incr();
}
// Given the expression '(x + C) - v', or
// 'v - (x + C)', we examine nodes '+' and 'v':
//
// 1. Do not convert if '+' is a counted-loop increment, because the '-' is
// loop invariant and converting extends the live-range of 'x' to overlap
// with the '+', forcing another register to be used in the loop.
//
// 2. Do not convert if 'v' is a counted-loop induction variable, because
// 'x' might be invariant.
//
static bool ok_to_convert(Node* inc, Node* var) {
return !(is_cloop_increment(inc) || var->is_cloop_ind_var());
}
//------------------------------Ideal------------------------------------------
Node *SubINode::Ideal(PhaseGVN *phase, bool can_reshape){
Node *in1 = in(1);
Node *in2 = in(2);
uint op1 = in1->Opcode();
uint op2 = in2->Opcode();
#ifdef ASSERT
// Check for dead loop
if ((in1 == this) || (in2 == this) ||
((op1 == Op_AddI || op1 == Op_SubI) &&
((in1->in(1) == this) || (in1->in(2) == this) ||
(in1->in(1) == in1) || (in1->in(2) == in1)))) {
assert(false, "dead loop in SubINode::Ideal");
}
#endif
const Type *t2 = phase->type( in2 );
if( t2 == Type::TOP ) return NULL;
// Convert "x-c0" into "x+ -c0".
if( t2->base() == Type::Int ){ // Might be bottom or top...
const TypeInt *i = t2->is_int();
if( i->is_con() )
return new AddINode(in1, phase->intcon(-i->get_con()));
}
// Convert "(x+c0) - y" into (x-y) + c0"
// Do not collapse (x+c0)-y if "+" is a loop increment or
// if "y" is a loop induction variable.
if( op1 == Op_AddI && ok_to_convert(in1, in2) ) {
const Type *tadd = phase->type( in1->in(2) );
if( tadd->singleton() && tadd != Type::TOP ) {
Node *sub2 = phase->transform( new SubINode( in1->in(1), in2 ));
return new AddINode( sub2, in1->in(2) );
}
}
// Convert "x - (y+c0)" into "(x-y) - c0"
// Need the same check as in above optimization but reversed.
if (op2 == Op_AddI && ok_to_convert(in2, in1)) {
Node* in21 = in2->in(1);
Node* in22 = in2->in(2);
const TypeInt* tcon = phase->type(in22)->isa_int();
if (tcon != NULL && tcon->is_con()) {
Node* sub2 = phase->transform( new SubINode(in1, in21) );
Node* neg_c0 = phase->intcon(- tcon->get_con());
return new AddINode(sub2, neg_c0);
}
}
const Type *t1 = phase->type( in1 );
if( t1 == Type::TOP ) return NULL;
#ifdef ASSERT
// Check for dead loop
if ((op2 == Op_AddI || op2 == Op_SubI) &&
((in2->in(1) == this) || (in2->in(2) == this) ||
(in2->in(1) == in2) || (in2->in(2) == in2))) {
assert(false, "dead loop in SubINode::Ideal");
}
#endif
// Convert "x - (x+y)" into "-y"
if (op2 == Op_AddI && in1 == in2->in(1)) {
return new SubINode(phase->intcon(0), in2->in(2));
}
// Convert "(x-y) - x" into "-y"
if (op1 == Op_SubI && in1->in(1) == in2) {
return new SubINode(phase->intcon(0), in1->in(2));
}
// Convert "x - (y+x)" into "-y"
if (op2 == Op_AddI && in1 == in2->in(2)) {
return new SubINode(phase->intcon(0), in2->in(1));
}
// Convert "0 - (x-y)" into "y-x"
if( t1 == TypeInt::ZERO && op2 == Op_SubI )
return new SubINode( in2->in(2), in2->in(1) );
// Convert "0 - (x+con)" into "-con-x"
jint con;
if( t1 == TypeInt::ZERO && op2 == Op_AddI &&
(con = in2->in(2)->find_int_con(0)) != 0 )
return new SubINode( phase->intcon(-con), in2->in(1) );
// Convert "(X+A) - (X+B)" into "A - B"
if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(1) )
return new SubINode( in1->in(2), in2->in(2) );
// Convert "(A+X) - (B+X)" into "A - B"
if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(2) )
return new SubINode( in1->in(1), in2->in(1) );
// Convert "(A+X) - (X+B)" into "A - B"
if( op1 == Op_AddI && op2 == Op_AddI && in1->in(2) == in2->in(1) )
return new SubINode( in1->in(1), in2->in(2) );
// Convert "(X+A) - (B+X)" into "A - B"
if( op1 == Op_AddI && op2 == Op_AddI && in1->in(1) == in2->in(2) )
return new SubINode( in1->in(2), in2->in(1) );
// Convert "A-(B-C)" into (A+C)-B", since add is commutative and generally
// nicer to optimize than subtract.
if( op2 == Op_SubI && in2->outcnt() == 1) {
Node *add1 = phase->transform( new AddINode( in1, in2->in(2) ) );
return new SubINode( add1, in2->in(1) );
}
// Convert "0-(A>>31)" into "(A>>>31)"
if ( op2 == Op_RShiftI ) {
Node *in21 = in2->in(1);
Node *in22 = in2->in(2);
const TypeInt *zero = phase->type(in1)->isa_int();
const TypeInt *t21 = phase->type(in21)->isa_int();
const TypeInt *t22 = phase->type(in22)->isa_int();
if ( t21 && t22 && zero == TypeInt::ZERO && t22->is_con(31) ) {
return new URShiftINode(in21, in22);
}
}
return NULL;
}
//------------------------------sub--------------------------------------------
// A subtract node differences it's two inputs.
const Type *SubINode::sub( const Type *t1, const Type *t2 ) const {
const TypeInt *r0 = t1->is_int(); // Handy access
const TypeInt *r1 = t2->is_int();
int32_t lo = java_subtract(r0->_lo, r1->_hi);
int32_t hi = java_subtract(r0->_hi, r1->_lo);
// We next check for 32-bit overflow.
// If that happens, we just assume all integers are possible.
if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR
((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND
(((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR
((r0->_hi ^ hi) >= 0)) ) // hi results have same signs
return TypeInt::make(lo,hi,MAX2(r0->_widen,r1->_widen));
else // Overflow; assume all integers
return TypeInt::INT;
}
//=============================================================================
//------------------------------Ideal------------------------------------------
Node *SubLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
Node *in1 = in(1);
Node *in2 = in(2);
uint op1 = in1->Opcode();
uint op2 = in2->Opcode();
#ifdef ASSERT
// Check for dead loop
if ((in1 == this) || (in2 == this) ||
((op1 == Op_AddL || op1 == Op_SubL) &&
((in1->in(1) == this) || (in1->in(2) == this) ||
(in1->in(1) == in1) || (in1->in(2) == in1)))) {
assert(false, "dead loop in SubLNode::Ideal");
}
#endif
if( phase->type( in2 ) == Type::TOP ) return NULL;
const TypeLong *i = phase->type( in2 )->isa_long();
// Convert "x-c0" into "x+ -c0".
if( i && // Might be bottom or top...
i->is_con() )
return new AddLNode(in1, phase->longcon(-i->get_con()));
// Convert "(x+c0) - y" into (x-y) + c0"
// Do not collapse (x+c0)-y if "+" is a loop increment or
// if "y" is a loop induction variable.
if( op1 == Op_AddL && ok_to_convert(in1, in2) ) {
Node *in11 = in1->in(1);
const Type *tadd = phase->type( in1->in(2) );
if( tadd->singleton() && tadd != Type::TOP ) {
Node *sub2 = phase->transform( new SubLNode( in11, in2 ));
return new AddLNode( sub2, in1->in(2) );
}
}
// Convert "x - (y+c0)" into "(x-y) - c0"
// Need the same check as in above optimization but reversed.
if (op2 == Op_AddL && ok_to_convert(in2, in1)) {
Node* in21 = in2->in(1);
Node* in22 = in2->in(2);
const TypeLong* tcon = phase->type(in22)->isa_long();
if (tcon != NULL && tcon->is_con()) {
Node* sub2 = phase->transform( new SubLNode(in1, in21) );
Node* neg_c0 = phase->longcon(- tcon->get_con());
return new AddLNode(sub2, neg_c0);
}
}
const Type *t1 = phase->type( in1 );
if( t1 == Type::TOP ) return NULL;
#ifdef ASSERT
// Check for dead loop
if ((op2 == Op_AddL || op2 == Op_SubL) &&
((in2->in(1) == this) || (in2->in(2) == this) ||
(in2->in(1) == in2) || (in2->in(2) == in2))) {
assert(false, "dead loop in SubLNode::Ideal");
}
#endif
// Convert "x - (x+y)" into "-y"
if (op2 == Op_AddL && in1 == in2->in(1)) {
return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(2));
}
// Convert "x - (y+x)" into "-y"
if (op2 == Op_AddL && in1 == in2->in(2)) {
return new SubLNode(phase->makecon(TypeLong::ZERO), in2->in(1));
}
// Convert "0 - (x-y)" into "y-x"
if( phase->type( in1 ) == TypeLong::ZERO && op2 == Op_SubL )
return new SubLNode( in2->in(2), in2->in(1) );
// Convert "(X+A) - (X+B)" into "A - B"
if( op1 == Op_AddL && op2 == Op_AddL && in1->in(1) == in2->in(1) )
return new SubLNode( in1->in(2), in2->in(2) );
// Convert "(A+X) - (B+X)" into "A - B"
if( op1 == Op_AddL && op2 == Op_AddL && in1->in(2) == in2->in(2) )
return new SubLNode( in1->in(1), in2->in(1) );
// Convert "A-(B-C)" into (A+C)-B"
if( op2 == Op_SubL && in2->outcnt() == 1) {
Node *add1 = phase->transform( new AddLNode( in1, in2->in(2) ) );
return new SubLNode( add1, in2->in(1) );
}
// Convert "0L-(A>>63)" into "(A>>>63)"
if ( op2 == Op_RShiftL ) {
Node *in21 = in2->in(1);
Node *in22 = in2->in(2);
const TypeLong *zero = phase->type(in1)->isa_long();
const TypeLong *t21 = phase->type(in21)->isa_long();
const TypeInt *t22 = phase->type(in22)->isa_int();
if ( t21 && t22 && zero == TypeLong::ZERO && t22->is_con(63) ) {
return new URShiftLNode(in21, in22);
}
}
return NULL;
}
//------------------------------sub--------------------------------------------
// A subtract node differences it's two inputs.
const Type *SubLNode::sub( const Type *t1, const Type *t2 ) const {
const TypeLong *r0 = t1->is_long(); // Handy access
const TypeLong *r1 = t2->is_long();
jlong lo = java_subtract(r0->_lo, r1->_hi);
jlong hi = java_subtract(r0->_hi, r1->_lo);
// We next check for 32-bit overflow.
// If that happens, we just assume all integers are possible.
if( (((r0->_lo ^ r1->_hi) >= 0) || // lo ends have same signs OR
((r0->_lo ^ lo) >= 0)) && // lo results have same signs AND
(((r0->_hi ^ r1->_lo) >= 0) || // hi ends have same signs OR
((r0->_hi ^ hi) >= 0)) ) // hi results have same signs
return TypeLong::make(lo,hi,MAX2(r0->_widen,r1->_widen));
else // Overflow; assume all integers
return TypeLong::LONG;
}
//=============================================================================
//------------------------------Value------------------------------------------
// A subtract node differences its two inputs.
const Type* SubFPNode::Value(PhaseGVN* phase) const {
const Node* in1 = in(1);
const Node* in2 = in(2);
// Either input is TOP ==> the result is TOP
const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
if( t1 == Type::TOP ) return Type::TOP;
const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
if( t2 == Type::TOP ) return Type::TOP;
// if both operands are infinity of same sign, the result is NaN; do
// not replace with zero
if (t1->is_finite() && t2->is_finite() && in1 == in2) {
return add_id();
}
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
return sub(t1,t2); // Local flavor of type subtraction
}
//=============================================================================
//------------------------------Ideal------------------------------------------
Node *SubFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
const Type *t2 = phase->type( in(2) );
// Convert "x-c0" into "x+ -c0".
if( t2->base() == Type::FloatCon ) { // Might be bottom or top...
// return new (phase->C, 3) AddFNode(in(1), phase->makecon( TypeF::make(-t2->getf()) ) );
}
// Not associative because of boundary conditions (infinity)
if (IdealizedNumerics && !phase->C->method()->is_strict() &&
in(2)->is_Add() && in(1) == in(2)->in(1)) {
// Convert "x - (x+y)" into "-y"
return new SubFNode(phase->makecon(TypeF::ZERO), in(2)->in(2));
}
// Cannot replace 0.0-X with -X because a 'fsub' bytecode computes
// 0.0-0.0 as +0.0, while a 'fneg' bytecode computes -0.0.
//if( phase->type(in(1)) == TypeF::ZERO )
//return new (phase->C, 2) NegFNode(in(2));
return NULL;
}
//------------------------------sub--------------------------------------------
// A subtract node differences its two inputs.
const Type *SubFNode::sub( const Type *t1, const Type *t2 ) const {
// no folding if one of operands is infinity or NaN, do not do constant folding
if( g_isfinite(t1->getf()) && g_isfinite(t2->getf()) ) {
return TypeF::make( t1->getf() - t2->getf() );
}
else if( g_isnan(t1->getf()) ) {
return t1;
}
else if( g_isnan(t2->getf()) ) {
return t2;
}
else {
return Type::FLOAT;
}
}
//=============================================================================
//------------------------------Ideal------------------------------------------
Node *SubDNode::Ideal(PhaseGVN *phase, bool can_reshape){
const Type *t2 = phase->type( in(2) );
// Convert "x-c0" into "x+ -c0".
if( t2->base() == Type::DoubleCon ) { // Might be bottom or top...
// return new (phase->C, 3) AddDNode(in(1), phase->makecon( TypeD::make(-t2->getd()) ) );
}
// Not associative because of boundary conditions (infinity)
if (IdealizedNumerics && !phase->C->method()->is_strict() &&
in(2)->is_Add() && in(1) == in(2)->in(1)) {
// Convert "x - (x+y)" into "-y"
return new SubDNode(phase->makecon(TypeD::ZERO), in(2)->in(2));
}
// Cannot replace 0.0-X with -X because a 'dsub' bytecode computes
// 0.0-0.0 as +0.0, while a 'dneg' bytecode computes -0.0.
//if( phase->type(in(1)) == TypeD::ZERO )
//return new (phase->C, 2) NegDNode(in(2));
return NULL;
}
//------------------------------sub--------------------------------------------
// A subtract node differences its two inputs.
const Type *SubDNode::sub( const Type *t1, const Type *t2 ) const {
// no folding if one of operands is infinity or NaN, do not do constant folding
if( g_isfinite(t1->getd()) && g_isfinite(t2->getd()) ) {
return TypeD::make( t1->getd() - t2->getd() );
}
else if( g_isnan(t1->getd()) ) {
return t1;
}
else if( g_isnan(t2->getd()) ) {
return t2;
}
else {
return Type::DOUBLE;
}
}
//=============================================================================
//------------------------------Idealize---------------------------------------
// Unlike SubNodes, compare must still flatten return value to the
// range -1, 0, 1.
// And optimizations like those for (X + Y) - X fail if overflow happens.
Node* CmpNode::Identity(PhaseGVN* phase) {
return this;
}
#ifndef PRODUCT
//----------------------------related------------------------------------------
// Related nodes of comparison nodes include all data inputs (until hitting a
// control boundary) as well as all outputs until and including control nodes
// as well as their projections. In compact mode, data inputs till depth 1 and
// all outputs till depth 1 are considered.
void CmpNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
if (compact) {
this->collect_nodes(in_rel, 1, false, true);
this->collect_nodes(out_rel, -1, false, false);
} else {
this->collect_nodes_in_all_data(in_rel, false);
this->collect_nodes_out_all_ctrl_boundary(out_rel);
// Now, find all control nodes in out_rel, and include their projections
// and projection targets (if any) in the result.
GrowableArray<Node*> proj(Compile::current()->unique());
for (GrowableArrayIterator<Node*> it = out_rel->begin(); it != out_rel->end(); ++it) {
Node* n = *it;
if (n->is_CFG() && !n->is_Proj()) {
// Assume projections and projection targets are found at levels 1 and 2.
n->collect_nodes(&proj, -2, false, false);
for (GrowableArrayIterator<Node*> p = proj.begin(); p != proj.end(); ++p) {
out_rel->append_if_missing(*p);
}
proj.clear();
}
}
}
}
#endif
CmpNode *CmpNode::make(Node *in1, Node *in2, BasicType bt, bool unsigned_comp) {
switch (bt) {
case T_INT:
if (unsigned_comp) {
return new CmpUNode(in1, in2);
}
return new CmpINode(in1, in2);
case T_LONG:
if (unsigned_comp) {
return new CmpULNode(in1, in2);
}
return new CmpLNode(in1, in2);
default:
fatal("Not implemented for %s", type2name(bt));
}
return NULL;
}
//=============================================================================
//------------------------------cmp--------------------------------------------
// Simplify a CmpI (compare 2 integers) node, based on local information.
// If both inputs are constants, compare them.
const Type *CmpINode::sub( const Type *t1, const Type *t2 ) const {
const TypeInt *r0 = t1->is_int(); // Handy access
const TypeInt *r1 = t2->is_int();
if( r0->_hi < r1->_lo ) // Range is always low?
return TypeInt::CC_LT;
else if( r0->_lo > r1->_hi ) // Range is always high?
return TypeInt::CC_GT;
else if( r0->is_con() && r1->is_con() ) { // comparing constants?
assert(r0->get_con() == r1->get_con(), "must be equal");
return TypeInt::CC_EQ; // Equal results.
} else if( r0->_hi == r1->_lo ) // Range is never high?
return TypeInt::CC_LE;
else if( r0->_lo == r1->_hi ) // Range is never low?
return TypeInt::CC_GE;
return TypeInt::CC; // else use worst case results
}
// Simplify a CmpU (compare 2 integers) node, based on local information.
// If both inputs are constants, compare them.
const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const {
assert(!t1->isa_ptr(), "obsolete usage of CmpU");
// comparing two unsigned ints
const TypeInt *r0 = t1->is_int(); // Handy access
const TypeInt *r1 = t2->is_int();
// Current installed version
// Compare ranges for non-overlap
juint lo0 = r0->_lo;
juint hi0 = r0->_hi;
juint lo1 = r1->_lo;
juint hi1 = r1->_hi;
// If either one has both negative and positive values,
// it therefore contains both 0 and -1, and since [0..-1] is the
// full unsigned range, the type must act as an unsigned bottom.
bool bot0 = ((jint)(lo0 ^ hi0) < 0);
bool bot1 = ((jint)(lo1 ^ hi1) < 0);
if (bot0 || bot1) {
// All unsigned values are LE -1 and GE 0.
if (lo0 == 0 && hi0 == 0) {
return TypeInt::CC_LE; // 0 <= bot
} else if ((jint)lo0 == -1 && (jint)hi0 == -1) {
return TypeInt::CC_GE; // -1 >= bot
} else if (lo1 == 0 && hi1 == 0) {
return TypeInt::CC_GE; // bot >= 0
} else if ((jint)lo1 == -1 && (jint)hi1 == -1) {
return TypeInt::CC_LE; // bot <= -1
}
} else {
// We can use ranges of the form [lo..hi] if signs are the same.
assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
// results are reversed, '-' > '+' for unsigned compare
if (hi0 < lo1) {
return TypeInt::CC_LT; // smaller
} else if (lo0 > hi1) {
return TypeInt::CC_GT; // greater
} else if (hi0 == lo1 && lo0 == hi1) {
return TypeInt::CC_EQ; // Equal results
} else if (lo0 >= hi1) {
return TypeInt::CC_GE;
} else if (hi0 <= lo1) {
// Check for special case in Hashtable::get. (See below.)
if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
return TypeInt::CC_LT;
return TypeInt::CC_LE;
}
}
// Check for special case in Hashtable::get - the hash index is
// mod'ed to the table size so the following range check is useless.
// Check for: (X Mod Y) CmpU Y, where the mod result and Y both have
// to be positive.
// (This is a gross hack, since the sub method never
// looks at the structure of the node in any other case.)
if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
return TypeInt::CC_LT;
return TypeInt::CC; // else use worst case results
}
const Type* CmpUNode::Value(PhaseGVN* phase) const {
const Type* t = SubNode::Value_common(phase);
if (t != NULL) {
return t;
}
const Node* in1 = in(1);
const Node* in2 = in(2);
const Type* t1 = phase->type(in1);
const Type* t2 = phase->type(in2);
assert(t1->isa_int(), "CmpU has only Int type inputs");
if (t2 == TypeInt::INT) { // Compare to bottom?
return bottom_type();
}
uint in1_op = in1->Opcode();
if (in1_op == Op_AddI || in1_op == Op_SubI) {
// The problem rise when result of AddI(SubI) may overflow
// signed integer value. Let say the input type is
// [256, maxint] then +128 will create 2 ranges due to
// overflow: [minint, minint+127] and [384, maxint].
// But C2 type system keep only 1 type range and as result
// it use general [minint, maxint] for this case which we
// can't optimize.
//
// Make 2 separate type ranges based on types of AddI(SubI) inputs
// and compare results of their compare. If results are the same
// CmpU node can be optimized.
const Node* in11 = in1->in(1);
const Node* in12 = in1->in(2);
const Type* t11 = (in11 == in1) ? Type::TOP : phase->type(in11);
const Type* t12 = (in12 == in1) ? Type::TOP : phase->type(in12);
// Skip cases when input types are top or bottom.
if ((t11 != Type::TOP) && (t11 != TypeInt::INT) &&
(t12 != Type::TOP) && (t12 != TypeInt::INT)) {
const TypeInt *r0 = t11->is_int();
const TypeInt *r1 = t12->is_int();
jlong lo_r0 = r0->_lo;
jlong hi_r0 = r0->_hi;
jlong lo_r1 = r1->_lo;
jlong hi_r1 = r1->_hi;
if (in1_op == Op_SubI) {
jlong tmp = hi_r1;
hi_r1 = -lo_r1;
lo_r1 = -tmp;
// Note, for substructing [minint,x] type range
// long arithmetic provides correct overflow answer.
// The confusion come from the fact that in 32-bit
// -minint == minint but in 64-bit -minint == maxint+1.
}
jlong lo_long = lo_r0 + lo_r1;
jlong hi_long = hi_r0 + hi_r1;
int lo_tr1 = min_jint;
int hi_tr1 = (int)hi_long;
int lo_tr2 = (int)lo_long;
int hi_tr2 = max_jint;
bool underflow = lo_long != (jlong)lo_tr2;
bool overflow = hi_long != (jlong)hi_tr1;
// Use sub(t1, t2) when there is no overflow (one type range)
// or when both overflow and underflow (too complex).
if ((underflow != overflow) && (hi_tr1 < lo_tr2)) {
// Overflow only on one boundary, compare 2 separate type ranges.
int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
const TypeInt* tr1 = TypeInt::make(lo_tr1, hi_tr1, w);
const TypeInt* tr2 = TypeInt::make(lo_tr2, hi_tr2, w);
const Type* cmp1 = sub(tr1, t2);
const Type* cmp2 = sub(tr2, t2);
if (cmp1 == cmp2) {
return cmp1; // Hit!
}
}
}
}
return sub(t1, t2); // Local flavor of type subtraction
}
bool CmpUNode::is_index_range_check() const {
// Check for the "(X ModI Y) CmpU Y" shape
return (in(1)->Opcode() == Op_ModI &&
in(1)->in(2)->eqv_uncast(in(2)));
}
//------------------------------Idealize---------------------------------------
Node *CmpINode::Ideal( PhaseGVN *phase, bool can_reshape ) {
if (phase->type(in(2))->higher_equal(TypeInt::ZERO)) {
switch (in(1)->Opcode()) {
case Op_CmpL3: // Collapse a CmpL3/CmpI into a CmpL
return new CmpLNode(in(1)->in(1),in(1)->in(2));
case Op_CmpF3: // Collapse a CmpF3/CmpI into a CmpF
return new CmpFNode(in(1)->in(1),in(1)->in(2));
case Op_CmpD3: // Collapse a CmpD3/CmpI into a CmpD
return new CmpDNode(in(1)->in(1),in(1)->in(2));
//case Op_SubI:
// If (x - y) cannot overflow, then ((x - y) <?> 0)
// can be turned into (x <?> y).
// This is handled (with more general cases) by Ideal_sub_algebra.
}
}
return NULL; // No change
}
Node *CmpLNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
const TypeLong *t2 = phase->type(in(2))->isa_long();
if (Opcode() == Op_CmpL && in(1)->Opcode() == Op_ConvI2L && t2 && t2->is_con()) {
const jlong con = t2->get_con();
if (con >= min_jint && con <= max_jint) {
return new CmpINode(in(1)->in(1), phase->intcon((jint)con));
}
}
return NULL;
}
//=============================================================================
// Simplify a CmpL (compare 2 longs ) node, based on local information.
// If both inputs are constants, compare them.
const Type *CmpLNode::sub( const Type *t1, const Type *t2 ) const {
const TypeLong *r0 = t1->is_long(); // Handy access
const TypeLong *r1 = t2->is_long();
if( r0->_hi < r1->_lo ) // Range is always low?
return TypeInt::CC_LT;
else if( r0->_lo > r1->_hi ) // Range is always high?
return TypeInt::CC_GT;
else if( r0->is_con() && r1->is_con() ) { // comparing constants?
assert(r0->get_con() == r1->get_con(), "must be equal");
return TypeInt::CC_EQ; // Equal results.
} else if( r0->_hi == r1->_lo ) // Range is never high?
return TypeInt::CC_LE;
else if( r0->_lo == r1->_hi ) // Range is never low?
return TypeInt::CC_GE;
return TypeInt::CC; // else use worst case results
}
// Simplify a CmpUL (compare 2 unsigned longs) node, based on local information.
// If both inputs are constants, compare them.
const Type* CmpULNode::sub(const Type* t1, const Type* t2) const {
assert(!t1->isa_ptr(), "obsolete usage of CmpUL");
// comparing two unsigned longs
const TypeLong* r0 = t1->is_long(); // Handy access
const TypeLong* r1 = t2->is_long();
// Current installed version
// Compare ranges for non-overlap
julong lo0 = r0->_lo;
julong hi0 = r0->_hi;
julong lo1 = r1->_lo;
julong hi1 = r1->_hi;
// If either one has both negative and positive values,
// it therefore contains both 0 and -1, and since [0..-1] is the
// full unsigned range, the type must act as an unsigned bottom.
bool bot0 = ((jlong)(lo0 ^ hi0) < 0);
bool bot1 = ((jlong)(lo1 ^ hi1) < 0);
if (bot0 || bot1) {
// All unsigned values are LE -1 and GE 0.
if (lo0 == 0 && hi0 == 0) {
return TypeInt::CC_LE; // 0 <= bot
} else if ((jlong)lo0 == -1 && (jlong)hi0 == -1) {
return TypeInt::CC_GE; // -1 >= bot
} else if (lo1 == 0 && hi1 == 0) {
return TypeInt::CC_GE; // bot >= 0
} else if ((jlong)lo1 == -1 && (jlong)hi1 == -1) {
return TypeInt::CC_LE; // bot <= -1
}
} else {
// We can use ranges of the form [lo..hi] if signs are the same.
assert(lo0 <= hi0 && lo1 <= hi1, "unsigned ranges are valid");
// results are reversed, '-' > '+' for unsigned compare
if (hi0 < lo1) {
return TypeInt::CC_LT; // smaller
} else if (lo0 > hi1) {
return TypeInt::CC_GT; // greater
} else if (hi0 == lo1 && lo0 == hi1) {
return TypeInt::CC_EQ; // Equal results
} else if (lo0 >= hi1) {
return TypeInt::CC_GE;
} else if (hi0 <= lo1) {
return TypeInt::CC_LE;
}
}
return TypeInt::CC; // else use worst case results
}
//=============================================================================
//------------------------------sub--------------------------------------------
// Simplify an CmpP (compare 2 pointers) node, based on local information.
// If both inputs are constants, compare them.
const Type *CmpPNode::sub( const Type *t1, const Type *t2 ) const {
const TypePtr *r0 = t1->is_ptr(); // Handy access
const TypePtr *r1 = t2->is_ptr();
// Undefined inputs makes for an undefined result
if( TypePtr::above_centerline(r0->_ptr) ||
TypePtr::above_centerline(r1->_ptr) )
return Type::TOP;
if (r0 == r1 && r0->singleton()) {
// Equal pointer constants (klasses, nulls, etc.)
return TypeInt::CC_EQ;
}
// See if it is 2 unrelated classes.
const TypeOopPtr* oop_p0 = r0->isa_oopptr();
const TypeOopPtr* oop_p1 = r1->isa_oopptr();
bool both_oop_ptr = oop_p0 && oop_p1;
if (both_oop_ptr) {
Node* in1 = in(1)->uncast();
Node* in2 = in(2)->uncast();
AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL);
AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL);
if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) {
return TypeInt::CC_GT; // different pointers
}
}
const TypeKlassPtr* klass_p0 = r0->isa_klassptr();
const TypeKlassPtr* klass_p1 = r1->isa_klassptr();
if (both_oop_ptr || (klass_p0 && klass_p1)) { // both or neither are klass pointers
ciKlass* klass0 = NULL;
bool xklass0 = false;
ciKlass* klass1 = NULL;
bool xklass1 = false;
if (oop_p0) {
klass0 = oop_p0->klass();
xklass0 = oop_p0->klass_is_exact();
} else {
assert(klass_p0, "must be non-null if oop_p0 is null");
klass0 = klass_p0->klass();
xklass0 = klass_p0->klass_is_exact();
}
if (oop_p1) {
klass1 = oop_p1->klass();
xklass1 = oop_p1->klass_is_exact();
} else {
assert(klass_p1, "must be non-null if oop_p1 is null");
klass1 = klass_p1->klass();
xklass1 = klass_p1->klass_is_exact();
}
if (klass0 && klass1 &&
klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces
klass1->is_loaded() && !klass1->is_interface() &&
(!klass0->is_obj_array_klass() ||
!klass0->as_obj_array_klass()->base_element_klass()->is_interface()) &&
(!klass1->is_obj_array_klass() ||
!klass1->as_obj_array_klass()->base_element_klass()->is_interface())) {
bool unrelated_classes = false;
// See if neither subclasses the other, or if the class on top
// is precise. In either of these cases, the compare is known
// to fail if at least one of the pointers is provably not null.
if (klass0->equals(klass1)) { // if types are unequal but klasses are equal
// Do nothing; we know nothing for imprecise types
} else if (klass0->is_subtype_of(klass1)) {
// If klass1's type is PRECISE, then classes are unrelated.
unrelated_classes = xklass1;
} else if (klass1->is_subtype_of(klass0)) {
// If klass0's type is PRECISE, then classes are unrelated.
unrelated_classes = xklass0;
} else { // Neither subtypes the other
unrelated_classes = true;
}
if (unrelated_classes) {
// The oops classes are known to be unrelated. If the joined PTRs of
// two oops is not Null and not Bottom, then we are sure that one
// of the two oops is non-null, and the comparison will always fail.
TypePtr::PTR jp = r0->join_ptr(r1->_ptr);
if (jp != TypePtr::Null && jp != TypePtr::BotPTR) {
return TypeInt::CC_GT;
}
}
}
}
// Known constants can be compared exactly
// Null can be distinguished from any NotNull pointers
// Unknown inputs makes an unknown result
if( r0->singleton() ) {
intptr_t bits0 = r0->get_con();
if( r1->singleton() )
return bits0 == r1->get_con() ? TypeInt::CC_EQ : TypeInt::CC_GT;
return ( r1->_ptr == TypePtr::NotNull && bits0==0 ) ? TypeInt::CC_GT : TypeInt::CC;
} else if( r1->singleton() ) {
intptr_t bits1 = r1->get_con();
return ( r0->_ptr == TypePtr::NotNull && bits1==0 ) ? TypeInt::CC_GT : TypeInt::CC;
} else
return TypeInt::CC;
}
static inline Node* isa_java_mirror_load(PhaseGVN* phase, Node* n) {
// Return the klass node for (indirect load from OopHandle)
// LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror))))
// or NULL if not matching.
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
n = bs->step_over_gc_barrier(n);
if (n->Opcode() != Op_LoadP) return NULL;
const TypeInstPtr* tp = phase->type(n)->isa_instptr();
if (!tp || tp->klass() != phase->C->env()->Class_klass()) return NULL;
Node* adr = n->in(MemNode::Address);
// First load from OopHandle: ((OopHandle)mirror)->resolve(); may need barrier.
if (adr->Opcode() != Op_LoadP || !phase->type(adr)->isa_rawptr()) return NULL;
adr = adr->in(MemNode::Address);
intptr_t off = 0;
Node* k = AddPNode::Ideal_base_and_offset(adr, phase, off);
if (k == NULL) return NULL;
const TypeKlassPtr* tkp = phase->type(k)->isa_klassptr();
if (!tkp || off != in_bytes(Klass::java_mirror_offset())) return NULL;
// We've found the klass node of a Java mirror load.
return k;
}
static inline Node* isa_const_java_mirror(PhaseGVN* phase, Node* n) {
// for ConP(Foo.class) return ConP(Foo.klass)
// otherwise return NULL
if (!n->is_Con()) return NULL;
const TypeInstPtr* tp = phase->type(n)->isa_instptr();
if (!tp) return NULL;
ciType* mirror_type = tp->java_mirror_type();
// TypeInstPtr::java_mirror_type() returns non-NULL for compile-
// time Class constants only.
if (!mirror_type) return NULL;
// x.getClass() == int.class can never be true (for all primitive types)
// Return a ConP(NULL) node for this case.
if (mirror_type->is_classless()) {
return phase->makecon(TypePtr::NULL_PTR);
}
// return the ConP(Foo.klass)
assert(mirror_type->is_klass(), "mirror_type should represent a Klass*");
return phase->makecon(TypeKlassPtr::make(mirror_type->as_klass()));
}
//------------------------------Ideal------------------------------------------
// Normalize comparisons between Java mirror loads to compare the klass instead.
//
// Also check for the case of comparing an unknown klass loaded from the primary
// super-type array vs a known klass with no subtypes. This amounts to
// checking to see an unknown klass subtypes a known klass with no subtypes;
// this only happens on an exact match. We can shorten this test by 1 load.
Node *CmpPNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
// Normalize comparisons between Java mirrors into comparisons of the low-
// level klass, where a dependent load could be shortened.
//
// The new pattern has a nice effect of matching the same pattern used in the
// fast path of instanceof/checkcast/Class.isInstance(), which allows
// redundant exact type check be optimized away by GVN.
// For example, in
// if (x.getClass() == Foo.class) {
// Foo foo = (Foo) x;
// // ... use a ...
// }
// a CmpPNode could be shared between if_acmpne and checkcast
{
Node* k1 = isa_java_mirror_load(phase, in(1));
Node* k2 = isa_java_mirror_load(phase, in(2));
Node* conk2 = isa_const_java_mirror(phase, in(2));
if (k1 && (k2 || conk2)) {
Node* lhs = k1;
Node* rhs = (k2 != NULL) ? k2 : conk2;
PhaseIterGVN* igvn = phase->is_IterGVN();
if (igvn != NULL) {
set_req_X(1, lhs, igvn);
set_req_X(2, rhs, igvn);
} else {
set_req(1, lhs);
set_req(2, rhs);
}
return this;
}
}
// Constant pointer on right?
const TypeKlassPtr* t2 = phase->type(in(2))->isa_klassptr();
if (t2 == NULL || !t2->klass_is_exact())
return NULL;
// Get the constant klass we are comparing to.
ciKlass* superklass = t2->klass();
// Now check for LoadKlass on left.
Node* ldk1 = in(1);
if (ldk1->is_DecodeNKlass()) {
ldk1 = ldk1->in(1);
if (ldk1->Opcode() != Op_LoadNKlass )
return NULL;
} else if (ldk1->Opcode() != Op_LoadKlass )
return NULL;
// Take apart the address of the LoadKlass:
Node* adr1 = ldk1->in(MemNode::Address);
intptr_t con2 = 0;
Node* ldk2 = AddPNode::Ideal_base_and_offset(adr1, phase, con2);
if (ldk2 == NULL)
return NULL;
if (con2 == oopDesc::klass_offset_in_bytes()) {
// We are inspecting an object's concrete class.
// Short-circuit the check if the query is abstract.
if (superklass->is_interface() ||
superklass->is_abstract()) {
// Make it come out always false:
this->set_req(2, phase->makecon(TypePtr::NULL_PTR));
return this;
}
}
// Check for a LoadKlass from primary supertype array.
// Any nested loadklass from loadklass+con must be from the p.s. array.
if (ldk2->is_DecodeNKlass()) {
// Keep ldk2 as DecodeN since it could be used in CmpP below.
if (ldk2->in(1)->Opcode() != Op_LoadNKlass )
return NULL;
} else if (ldk2->Opcode() != Op_LoadKlass)
return NULL;
// Verify that we understand the situation
if (con2 != (intptr_t) superklass->super_check_offset())
return NULL; // Might be element-klass loading from array klass
// If 'superklass' has no subklasses and is not an interface, then we are
// assured that the only input which will pass the type check is
// 'superklass' itself.
//
// We could be more liberal here, and allow the optimization on interfaces
// which have a single implementor. This would require us to increase the
// expressiveness of the add_dependency() mechanism.
// %%% Do this after we fix TypeOopPtr: Deps are expressive enough now.
// Object arrays must have their base element have no subtypes
while (superklass->is_obj_array_klass()) {
ciType* elem = superklass->as_obj_array_klass()->element_type();
superklass = elem->as_klass();
}
if (superklass->is_instance_klass()) {
ciInstanceKlass* ik = superklass->as_instance_klass();
if (ik->has_subklass() || ik->is_interface()) return NULL;
// Add a dependency if there is a chance that a subclass will be added later.
if (!ik->is_final()) {
phase->C->dependencies()->assert_leaf_type(ik);
}
}
// Bypass the dependent load, and compare directly
this->set_req(1,ldk2);
return this;
}
//=============================================================================
//------------------------------sub--------------------------------------------
// Simplify an CmpN (compare 2 pointers) node, based on local information.
// If both inputs are constants, compare them.
const Type *CmpNNode::sub( const Type *t1, const Type *t2 ) const {
ShouldNotReachHere();
return bottom_type();
}
//------------------------------Ideal------------------------------------------
Node *CmpNNode::Ideal( PhaseGVN *phase, bool can_reshape ) {
return NULL;
}
//=============================================================================
//------------------------------Value------------------------------------------
// Simplify an CmpF (compare 2 floats ) node, based on local information.
// If both inputs are constants, compare them.
const Type* CmpFNode::Value(PhaseGVN* phase) const {
const Node* in1 = in(1);
const Node* in2 = in(2);
// Either input is TOP ==> the result is TOP
const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
if( t1 == Type::TOP ) return Type::TOP;
const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
if( t2 == Type::TOP ) return Type::TOP;
// Not constants? Don't know squat - even if they are the same
// value! If they are NaN's they compare to LT instead of EQ.
const TypeF *tf1 = t1->isa_float_constant();
const TypeF *tf2 = t2->isa_float_constant();
if( !tf1 || !tf2 ) return TypeInt::CC;
// This implements the Java bytecode fcmpl, so unordered returns -1.
if( tf1->is_nan() || tf2->is_nan() )
return TypeInt::CC_LT;
if( tf1->_f < tf2->_f ) return TypeInt::CC_LT;
if( tf1->_f > tf2->_f ) return TypeInt::CC_GT;
assert( tf1->_f == tf2->_f, "do not understand FP behavior" );
return TypeInt::CC_EQ;
}
//=============================================================================
//------------------------------Value------------------------------------------
// Simplify an CmpD (compare 2 doubles ) node, based on local information.
// If both inputs are constants, compare them.
const Type* CmpDNode::Value(PhaseGVN* phase) const {
const Node* in1 = in(1);
const Node* in2 = in(2);
// Either input is TOP ==> the result is TOP
const Type* t1 = (in1 == this) ? Type::TOP : phase->type(in1);
if( t1 == Type::TOP ) return Type::TOP;
const Type* t2 = (in2 == this) ? Type::TOP : phase->type(in2);
if( t2 == Type::TOP ) return Type::TOP;
// Not constants? Don't know squat - even if they are the same
// value! If they are NaN's they compare to LT instead of EQ.
const TypeD *td1 = t1->isa_double_constant();
const TypeD *td2 = t2->isa_double_constant();
if( !td1 || !td2 ) return TypeInt::CC;
// This implements the Java bytecode dcmpl, so unordered returns -1.
if( td1->is_nan() || td2->is_nan() )
return TypeInt::CC_LT;
if( td1->_d < td2->_d ) return TypeInt::CC_LT;
if( td1->_d > td2->_d ) return TypeInt::CC_GT;
assert( td1->_d == td2->_d, "do not understand FP behavior" );
return TypeInt::CC_EQ;
}
//------------------------------Ideal------------------------------------------
Node *CmpDNode::Ideal(PhaseGVN *phase, bool can_reshape){
// Check if we can change this to a CmpF and remove a ConvD2F operation.
// Change (CMPD (F2D (float)) (ConD value))
// To (CMPF (float) (ConF value))
// Valid when 'value' does not lose precision as a float.
// Benefits: eliminates conversion, does not require 24-bit mode
// NaNs prevent commuting operands. This transform works regardless of the
// order of ConD and ConvF2D inputs by preserving the original order.
int idx_f2d = 1; // ConvF2D on left side?
if( in(idx_f2d)->Opcode() != Op_ConvF2D )
idx_f2d = 2; // No, swap to check for reversed args
int idx_con = 3-idx_f2d; // Check for the constant on other input
if( ConvertCmpD2CmpF &&
in(idx_f2d)->Opcode() == Op_ConvF2D &&
in(idx_con)->Opcode() == Op_ConD ) {
const TypeD *t2 = in(idx_con)->bottom_type()->is_double_constant();
double t2_value_as_double = t2->_d;
float t2_value_as_float = (float)t2_value_as_double;
if( t2_value_as_double == (double)t2_value_as_float ) {
// Test value can be represented as a float
// Eliminate the conversion to double and create new comparison
Node *new_in1 = in(idx_f2d)->in(1);
Node *new_in2 = phase->makecon( TypeF::make(t2_value_as_float) );
if( idx_f2d != 1 ) { // Must flip args to match original order
Node *tmp = new_in1;
new_in1 = new_in2;
new_in2 = tmp;
}
CmpFNode *new_cmp = (Opcode() == Op_CmpD3)
? new CmpF3Node( new_in1, new_in2 )
: new CmpFNode ( new_in1, new_in2 ) ;
return new_cmp; // Changed to CmpFNode
}
// Testing value required the precision of a double
}
return NULL; // No change
}
//=============================================================================
//------------------------------cc2logical-------------------------------------
// Convert a condition code type to a logical type
const Type *BoolTest::cc2logical( const Type *CC ) const {
if( CC == Type::TOP ) return Type::TOP;
if( CC->base() != Type::Int ) return TypeInt::BOOL; // Bottom or worse
const TypeInt *ti = CC->is_int();
if( ti->is_con() ) { // Only 1 kind of condition codes set?
// Match low order 2 bits
int tmp = ((ti->get_con()&3) == (_test&3)) ? 1 : 0;
if( _test & 4 ) tmp = 1-tmp; // Optionally complement result
return TypeInt::make(tmp); // Boolean result
}
if( CC == TypeInt::CC_GE ) {
if( _test == ge ) return TypeInt::ONE;
if( _test == lt ) return TypeInt::ZERO;
}
if( CC == TypeInt::CC_LE ) {
if( _test == le ) return TypeInt::ONE;
if( _test == gt ) return TypeInt::ZERO;
}
return TypeInt::BOOL;
}
//------------------------------dump_spec-------------------------------------
// Print special per-node info
void BoolTest::dump_on(outputStream *st) const {
const char *msg[] = {"eq","gt","of","lt","ne","le","nof","ge"};
st->print("%s", msg[_test]);
}
// Returns the logical AND of two tests (or 'never' if both tests can never be true).
// For example, a test for 'le' followed by a test for 'lt' is equivalent with 'lt'.
BoolTest::mask BoolTest::merge(BoolTest other) const {
const mask res[illegal+1][illegal+1] = {
// eq, gt, of, lt, ne, le, nof, ge, never, illegal
{eq, never, illegal, never, never, eq, illegal, eq, never, illegal}, // eq
{never, gt, illegal, never, gt, never, illegal, gt, never, illegal}, // gt
{illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // of
{never, never, illegal, lt, lt, lt, illegal, never, never, illegal}, // lt
{never, gt, illegal, lt, ne, lt, illegal, gt, never, illegal}, // ne
{eq, never, illegal, lt, lt, le, illegal, eq, never, illegal}, // le
{illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, never, illegal}, // nof
{eq, gt, illegal, never, gt, eq, illegal, ge, never, illegal}, // ge
{never, never, never, never, never, never, never, never, never, illegal}, // never
{illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal, illegal}}; // illegal
return res[_test][other._test];
}
//=============================================================================
uint BoolNode::hash() const { return (Node::hash() << 3)|(_test._test+1); }
uint BoolNode::size_of() const { return sizeof(BoolNode); }
//------------------------------operator==-------------------------------------
bool BoolNode::cmp( const Node &n ) const {
const BoolNode *b = (const BoolNode *)&n; // Cast up
return (_test._test == b->_test._test);
}
//-------------------------------make_predicate--------------------------------
Node* BoolNode::make_predicate(Node* test_value, PhaseGVN* phase) {
if (test_value->is_Con()) return test_value;
if (test_value->is_Bool()) return test_value;
if (test_value->is_CMove() &&
test_value->in(CMoveNode::Condition)->is_Bool()) {
BoolNode* bol = test_value->in(CMoveNode::Condition)->as_Bool();
const Type* ftype = phase->type(test_value->in(CMoveNode::IfFalse));
const Type* ttype = phase->type(test_value->in(CMoveNode::IfTrue));
if (ftype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ttype)) {
return bol;
} else if (ttype == TypeInt::ZERO && !TypeInt::ZERO->higher_equal(ftype)) {
return phase->transform( bol->negate(phase) );
}
// Else fall through. The CMove gets in the way of the test.
// It should be the case that make_predicate(bol->as_int_value()) == bol.
}
Node* cmp = new CmpINode(test_value, phase->intcon(0));
cmp = phase->transform(cmp);
Node* bol = new BoolNode(cmp, BoolTest::ne);
return phase->transform(bol);
}
//--------------------------------as_int_value---------------------------------
Node* BoolNode::as_int_value(PhaseGVN* phase) {
// Inverse to make_predicate. The CMove probably boils down to a Conv2B.
Node* cmov = CMoveNode::make(NULL, this,
phase->intcon(0), phase->intcon(1),
TypeInt::BOOL);
return phase->transform(cmov);
}
//----------------------------------negate-------------------------------------
BoolNode* BoolNode::negate(PhaseGVN* phase) {
return new BoolNode(in(1), _test.negate());
}
// Change "bool eq/ne (cmp (add/sub A B) C)" into false/true if add/sub
// overflows and we can prove that C is not in the two resulting ranges.
// This optimization is similar to the one performed by CmpUNode::Value().
Node* BoolNode::fold_cmpI(PhaseGVN* phase, SubNode* cmp, Node* cmp1, int cmp_op,
int cmp1_op, const TypeInt* cmp2_type) {
// Only optimize eq/ne integer comparison of add/sub
if((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
(cmp_op == Op_CmpI) && (cmp1_op == Op_AddI || cmp1_op == Op_SubI)) {
// Skip cases were inputs of add/sub are not integers or of bottom type
const TypeInt* r0 = phase->type(cmp1->in(1))->isa_int();
const TypeInt* r1 = phase->type(cmp1->in(2))->isa_int();
if ((r0 != NULL) && (r0 != TypeInt::INT) &&
(r1 != NULL) && (r1 != TypeInt::INT) &&
(cmp2_type != TypeInt::INT)) {
// Compute exact (long) type range of add/sub result
jlong lo_long = r0->_lo;
jlong hi_long = r0->_hi;
if (cmp1_op == Op_AddI) {
lo_long += r1->_lo;
hi_long += r1->_hi;
} else {
lo_long -= r1->_hi;
hi_long -= r1->_lo;
}
// Check for over-/underflow by casting to integer
int lo_int = (int)lo_long;
int hi_int = (int)hi_long;
bool underflow = lo_long != (jlong)lo_int;
bool overflow = hi_long != (jlong)hi_int;
if ((underflow != overflow) && (hi_int < lo_int)) {
// Overflow on one boundary, compute resulting type ranges:
// tr1 [MIN_INT, hi_int] and tr2 [lo_int, MAX_INT]
int w = MAX2(r0->_widen, r1->_widen); // _widen does not matter here
const TypeInt* tr1 = TypeInt::make(min_jint, hi_int, w);
const TypeInt* tr2 = TypeInt::make(lo_int, max_jint, w);
// Compare second input of cmp to both type ranges
const Type* sub_tr1 = cmp->sub(tr1, cmp2_type);
const Type* sub_tr2 = cmp->sub(tr2, cmp2_type);
if (sub_tr1 == TypeInt::CC_LT && sub_tr2 == TypeInt::CC_GT) {
// The result of the add/sub will never equal cmp2. Replace BoolNode
// by false (0) if it tests for equality and by true (1) otherwise.
return ConINode::make((_test._test == BoolTest::eq) ? 0 : 1);
}
}
}
}
return NULL;
}
static bool is_counted_loop_cmp(Node *cmp) {
Node *n = cmp->in(1)->in(1);
return n != NULL &&
n->is_Phi() &&
n->in(0) != NULL &&
n->in(0)->is_CountedLoop() &&
n->in(0)->as_CountedLoop()->phi() == n;
}
//------------------------------Ideal------------------------------------------
Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Change "bool tst (cmp con x)" into "bool ~tst (cmp x con)".
// This moves the constant to the right. Helps value-numbering.
Node *cmp = in(1);
if( !cmp->is_Sub() ) return NULL;
int cop = cmp->Opcode();
if( cop == Op_FastLock || cop == Op_FastUnlock || cmp->is_SubTypeCheck()) return NULL;
Node *cmp1 = cmp->in(1);
Node *cmp2 = cmp->in(2);
if( !cmp1 ) return NULL;
if (_test._test == BoolTest::overflow || _test._test == BoolTest::no_overflow) {
return NULL;
}
// Constant on left?
Node *con = cmp1;
uint op2 = cmp2->Opcode();
// Move constants to the right of compare's to canonicalize.
// Do not muck with Opaque1 nodes, as this indicates a loop
// guard that cannot change shape.
if( con->is_Con() && !cmp2->is_Con() && op2 != Op_Opaque1 &&
// Because of NaN's, CmpD and CmpF are not commutative
cop != Op_CmpD && cop != Op_CmpF &&
// Protect against swapping inputs to a compare when it is used by a
// counted loop exit, which requires maintaining the loop-limit as in(2)
!is_counted_loop_exit_test() ) {
// Ok, commute the constant to the right of the cmp node.
// Clone the Node, getting a new Node of the same class
cmp = cmp->clone();
// Swap inputs to the clone
cmp->swap_edges(1, 2);
cmp = phase->transform( cmp );
return new BoolNode( cmp, _test.commute() );
}
// Change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)".
if (cop == Op_CmpI &&
(_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
cmp1->Opcode() == Op_AndI && cmp2->Opcode() == Op_ConI &&
cmp1->in(2)->Opcode() == Op_ConI) {
const TypeInt *t12 = phase->type(cmp2)->isa_int();
const TypeInt *t112 = phase->type(cmp1->in(2))->isa_int();
if (t12 && t12->is_con() && t112 && t112->is_con() &&
t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) {
Node *ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0)));
return new BoolNode(ncmp, _test.negate());
}
}
// Same for long type: change "bool eq/ne (cmp (and X 16) 16)" into "bool ne/eq (cmp (and X 16) 0)".
if (cop == Op_CmpL &&
(_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
cmp1->Opcode() == Op_AndL && cmp2->Opcode() == Op_ConL &&
cmp1->in(2)->Opcode() == Op_ConL) {
const TypeLong *t12 = phase->type(cmp2)->isa_long();
const TypeLong *t112 = phase->type(cmp1->in(2))->isa_long();
if (t12 && t12->is_con() && t112 && t112->is_con() &&
t12->get_con() == t112->get_con() && is_power_of_2(t12->get_con())) {
Node *ncmp = phase->transform(new CmpLNode(cmp1, phase->longcon(0)));
return new BoolNode(ncmp, _test.negate());
}
}
// Change "bool eq/ne (cmp (xor X 1) 0)" into "bool ne/eq (cmp X 0)".
// The XOR-1 is an idiom used to flip the sense of a bool. We flip the
// test instead.
int cmp1_op = cmp1->Opcode();
const TypeInt* cmp2_type = phase->type(cmp2)->isa_int();
if (cmp2_type == NULL) return NULL;
Node* j_xor = cmp1;
if( cmp2_type == TypeInt::ZERO &&
cmp1_op == Op_XorI &&
j_xor->in(1) != j_xor && // An xor of itself is dead
phase->type( j_xor->in(1) ) == TypeInt::BOOL &&
phase->type( j_xor->in(2) ) == TypeInt::ONE &&
(_test._test == BoolTest::eq ||
_test._test == BoolTest::ne) ) {
Node *ncmp = phase->transform(new CmpINode(j_xor->in(1),cmp2));
return new BoolNode( ncmp, _test.negate() );
}
// Change ((x & m) u<= m) or ((m & x) u<= m) to always true
// Same with ((x & m) u< m+1) and ((m & x) u< m+1)
if (cop == Op_CmpU &&
cmp1_op == Op_AndI) {
Node* bound = NULL;
if (_test._test == BoolTest::le) {
bound = cmp2;
} else if (_test._test == BoolTest::lt &&
cmp2->Opcode() == Op_AddI &&
cmp2->in(2)->find_int_con(0) == 1) {
bound = cmp2->in(1);
}
if (cmp1->in(2) == bound || cmp1->in(1) == bound) {
return ConINode::make(1);
}
}
// Change ((x & (m - 1)) u< m) into (m > 0)
// This is the off-by-one variant of the above
if (cop == Op_CmpU &&
_test._test == BoolTest::lt &&
cmp1_op == Op_AndI) {
Node* l = cmp1->in(1);
Node* r = cmp1->in(2);
for (int repeat = 0; repeat < 2; repeat++) {
bool match = r->Opcode() == Op_AddI && r->in(2)->find_int_con(0) == -1 &&
r->in(1) == cmp2;
if (match) {
// arraylength known to be non-negative, so a (arraylength != 0) is sufficient,
// but to be compatible with the array range check pattern, use (arraylength u> 0)
Node* ncmp = cmp2->Opcode() == Op_LoadRange
? phase->transform(new CmpUNode(cmp2, phase->intcon(0)))
: phase->transform(new CmpINode(cmp2, phase->intcon(0)));
return new BoolNode(ncmp, BoolTest::gt);
} else {
// commute and try again
l = cmp1->in(2);
r = cmp1->in(1);
}
}
}
// Change x u< 1 or x u<= 0 to x == 0
if (cop == Op_CmpU &&
cmp1_op != Op_LoadRange &&
((_test._test == BoolTest::lt &&
cmp2->find_int_con(-1) == 1) ||
(_test._test == BoolTest::le &&
cmp2->find_int_con(-1) == 0))) {
Node* ncmp = phase->transform(new CmpINode(cmp1, phase->intcon(0)));
return new BoolNode(ncmp, BoolTest::eq);
}
// Change (arraylength <= 0) or (arraylength == 0)
// into (arraylength u<= 0)
// Also change (arraylength != 0) into (arraylength u> 0)
// The latter version matches the code pattern generated for
// array range checks, which will more likely be optimized later.
if (cop == Op_CmpI &&
cmp1_op == Op_LoadRange &&
cmp2->find_int_con(-1) == 0) {
if (_test._test == BoolTest::le || _test._test == BoolTest::eq) {
Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
return new BoolNode(ncmp, BoolTest::le);
} else if (_test._test == BoolTest::ne) {
Node* ncmp = phase->transform(new CmpUNode(cmp1, cmp2));
return new BoolNode(ncmp, BoolTest::gt);
}
}
// Change "bool eq/ne (cmp (Conv2B X) 0)" into "bool eq/ne (cmp X 0)".
// This is a standard idiom for branching on a boolean value.
Node *c2b = cmp1;
if( cmp2_type == TypeInt::ZERO &&
cmp1_op == Op_Conv2B &&
(_test._test == BoolTest::eq ||
_test._test == BoolTest::ne) ) {
Node *ncmp = phase->transform(phase->type(c2b->in(1))->isa_int()
? (Node*)new CmpINode(c2b->in(1),cmp2)
: (Node*)new CmpPNode(c2b->in(1),phase->makecon(TypePtr::NULL_PTR))
);
return new BoolNode( ncmp, _test._test );
}
// Comparing a SubI against a zero is equal to comparing the SubI
// arguments directly. This only works for eq and ne comparisons
// due to possible integer overflow.
if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
(cop == Op_CmpI) &&
(cmp1_op == Op_SubI) &&
( cmp2_type == TypeInt::ZERO ) ) {
Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),cmp1->in(2)));
return new BoolNode( ncmp, _test._test );
}
// Same as above but with and AddI of a constant
if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
cop == Op_CmpI &&
cmp1_op == Op_AddI &&
cmp1->in(2) != NULL &&
phase->type(cmp1->in(2))->isa_int() &&
phase->type(cmp1->in(2))->is_int()->is_con() &&
cmp2_type == TypeInt::ZERO &&
!is_counted_loop_cmp(cmp) // modifying the exit test of a counted loop messes the counted loop shape
) {
const TypeInt* cmp1_in2 = phase->type(cmp1->in(2))->is_int();
Node *ncmp = phase->transform( new CmpINode(cmp1->in(1),phase->intcon(-cmp1_in2->_hi)));
return new BoolNode( ncmp, _test._test );
}
// Change "bool eq/ne (cmp (phi (X -X) 0))" into "bool eq/ne (cmp X 0)"
// since zero check of conditional negation of an integer is equal to
// zero check of the integer directly.
if ((_test._test == BoolTest::eq || _test._test == BoolTest::ne) &&
(cop == Op_CmpI) &&
(cmp2_type == TypeInt::ZERO) &&
(cmp1_op == Op_Phi)) {
// There should be a diamond phi with true path at index 1 or 2
PhiNode *phi = cmp1->as_Phi();
int idx_true = phi->is_diamond_phi();
if (idx_true != 0) {
// True input is in(idx_true) while false input is in(3 - idx_true)
Node *tin = phi->in(idx_true);
Node *fin = phi->in(3 - idx_true);
if ((tin->Opcode() == Op_SubI) &&
(phase->type(tin->in(1)) == TypeInt::ZERO) &&
(tin->in(2) == fin)) {
// Found conditional negation at true path, create a new CmpINode without that
Node *ncmp = phase->transform(new CmpINode(fin, cmp2));
return new BoolNode(ncmp, _test._test);
}
if ((fin->Opcode() == Op_SubI) &&
(phase->type(fin->in(1)) == TypeInt::ZERO) &&
(fin->in(2) == tin)) {
// Found conditional negation at false path, create a new CmpINode without that
Node *ncmp = phase->transform(new CmpINode(tin, cmp2));
return new BoolNode(ncmp, _test._test);
}
}
}
// Change (-A vs 0) into (A vs 0) by commuting the test. Disallow in the
// most general case because negating 0x80000000 does nothing. Needed for
// the CmpF3/SubI/CmpI idiom.
if( cop == Op_CmpI &&
cmp1_op == Op_SubI &&
cmp2_type == TypeInt::ZERO &&
phase->type( cmp1->in(1) ) == TypeInt::ZERO &&
phase->type( cmp1->in(2) )->higher_equal(TypeInt::SYMINT) ) {
Node *ncmp = phase->transform( new CmpINode(cmp1->in(2),cmp2));
return new BoolNode( ncmp, _test.commute() );
}
// Try to optimize signed integer comparison
return fold_cmpI(phase, cmp->as_Sub(), cmp1, cop, cmp1_op, cmp2_type);
// The transformation below is not valid for either signed or unsigned
// comparisons due to wraparound concerns at MAX_VALUE and MIN_VALUE.
// This transformation can be resurrected when we are able to
// make inferences about the range of values being subtracted from
// (or added to) relative to the wraparound point.
//
// // Remove +/-1's if possible.
// // "X <= Y-1" becomes "X < Y"
// // "X+1 <= Y" becomes "X < Y"
// // "X < Y+1" becomes "X <= Y"
// // "X-1 < Y" becomes "X <= Y"
// // Do not this to compares off of the counted-loop-end. These guys are
// // checking the trip counter and they want to use the post-incremented
// // counter. If they use the PRE-incremented counter, then the counter has
// // to be incremented in a private block on a loop backedge.
// if( du && du->cnt(this) && du->out(this)[0]->Opcode() == Op_CountedLoopEnd )
// return NULL;
// #ifndef PRODUCT
// // Do not do this in a wash GVN pass during verification.
// // Gets triggered by too many simple optimizations to be bothered with
// // re-trying it again and again.
// if( !phase->allow_progress() ) return NULL;
// #endif
// // Not valid for unsigned compare because of corner cases in involving zero.
// // For example, replacing "X-1 <u Y" with "X <=u Y" fails to throw an
// // exception in case X is 0 (because 0-1 turns into 4billion unsigned but
// // "0 <=u Y" is always true).
// if( cmp->Opcode() == Op_CmpU ) return NULL;
// int cmp2_op = cmp2->Opcode();
// if( _test._test == BoolTest::le ) {
// if( cmp1_op == Op_AddI &&
// phase->type( cmp1->in(2) ) == TypeInt::ONE )
// return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::lt );
// else if( cmp2_op == Op_AddI &&
// phase->type( cmp2->in(2) ) == TypeInt::MINUS_1 )
// return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::lt );
// } else if( _test._test == BoolTest::lt ) {
// if( cmp1_op == Op_AddI &&
// phase->type( cmp1->in(2) ) == TypeInt::MINUS_1 )
// return clone_cmp( cmp, cmp1->in(1), cmp2, phase, BoolTest::le );
// else if( cmp2_op == Op_AddI &&
// phase->type( cmp2->in(2) ) == TypeInt::ONE )
// return clone_cmp( cmp, cmp1, cmp2->in(1), phase, BoolTest::le );
// }
}
//------------------------------Value------------------------------------------
// Simplify a Bool (convert condition codes to boolean (1 or 0)) node,
// based on local information. If the input is constant, do it.
const Type* BoolNode::Value(PhaseGVN* phase) const {
return _test.cc2logical( phase->type( in(1) ) );
}
#ifndef PRODUCT
//------------------------------dump_spec--------------------------------------
// Dump special per-node info
void BoolNode::dump_spec(outputStream *st) const {
st->print("[");
_test.dump_on(st);
st->print("]");
}
//-------------------------------related---------------------------------------
// A BoolNode's related nodes are all of its data inputs, and all of its
// outputs until control nodes are hit, which are included. In compact
// representation, inputs till level 3 and immediate outputs are included.
void BoolNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
if (compact) {
this->collect_nodes(in_rel, 3, false, true);
this->collect_nodes(out_rel, -1, false, false);
} else {
this->collect_nodes_in_all_data(in_rel, false);
this->collect_nodes_out_all_ctrl_boundary(out_rel);
}
}
#endif
//----------------------is_counted_loop_exit_test------------------------------
// Returns true if node is used by a counted loop node.
bool BoolNode::is_counted_loop_exit_test() {
for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
Node* use = fast_out(i);
if (use->is_CountedLoopEnd()) {
return true;
}
}
return false;
}
//=============================================================================
//------------------------------Value------------------------------------------
// Compute sqrt
const Type* SqrtDNode::Value(PhaseGVN* phase) const {
const Type *t1 = phase->type( in(1) );
if( t1 == Type::TOP ) return Type::TOP;
if( t1->base() != Type::DoubleCon ) return Type::DOUBLE;
double d = t1->getd();
if( d < 0.0 ) return Type::DOUBLE;
return TypeD::make( sqrt( d ) );
}
const Type* SqrtFNode::Value(PhaseGVN* phase) const {
const Type *t1 = phase->type( in(1) );
if( t1 == Type::TOP ) return Type::TOP;
if( t1->base() != Type::FloatCon ) return Type::FLOAT;
float f = t1->getf();
if( f < 0.0f ) return Type::FLOAT;
return TypeF::make( (float)sqrt( (double)f ) );
}
Reference in New Issue View Git Blame Copy Permalink