提交 83cabf23 编写于 作者: K kvn

Merge

......@@ -160,6 +160,7 @@ class LibraryCallKit : public GraphKit {
bool inline_trans(vmIntrinsics::ID id);
bool inline_abs(vmIntrinsics::ID id);
bool inline_sqrt(vmIntrinsics::ID id);
void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
bool inline_pow(vmIntrinsics::ID id);
bool inline_exp(vmIntrinsics::ID id);
bool inline_min_max(vmIntrinsics::ID id);
......@@ -1535,40 +1536,79 @@ bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
return true;
}
void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
//-------------------
//result=(result.isNaN())? funcAddr():result;
// Check: If isNaN() by checking result!=result? then either trap
// or go to runtime
Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
// Build the boolean node
Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
{
BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
// End the current control-flow path
push_pair(x);
if (y != NULL) {
push_pair(y);
}
// The pow or exp intrinsic returned a NaN, which requires a call
// to the runtime. Recompile with the runtime call.
uncommon_trap(Deoptimization::Reason_intrinsic,
Deoptimization::Action_make_not_entrant);
}
push_pair(result);
} else {
// If this inlining ever returned NaN in the past, we compile a call
// to the runtime to properly handle corner cases
IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
Node* if_slow = _gvn.transform( new (C, 1) IfFalseNode(iff) );
Node* if_fast = _gvn.transform( new (C, 1) IfTrueNode(iff) );
if (!if_slow->is_top()) {
RegionNode* result_region = new(C, 3) RegionNode(3);
PhiNode* result_val = new (C, 3) PhiNode(result_region, Type::DOUBLE);
result_region->init_req(1, if_fast);
result_val->init_req(1, result);
set_control(if_slow);
const TypePtr* no_memory_effects = NULL;
Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
no_memory_effects,
x, top(), y, y ? top() : NULL);
Node* value = _gvn.transform(new (C, 1) ProjNode(rt, TypeFunc::Parms+0));
#ifdef ASSERT
Node* value_top = _gvn.transform(new (C, 1) ProjNode(rt, TypeFunc::Parms+1));
assert(value_top == top(), "second value must be top");
#endif
result_region->init_req(2, control());
result_val->init_req(2, value);
push_result(result_region, result_val);
} else {
push_pair(result);
}
}
}
//------------------------------inline_exp-------------------------------------
// Inline exp instructions, if possible. The Intel hardware only misses
// really odd corner cases (+/- Infinity). Just uncommon-trap them.
bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
assert(id == vmIntrinsics::_dexp, "Not exp");
// If this inlining ever returned NaN in the past, we do not intrinsify it
// every again. NaN results requires StrictMath.exp handling.
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
_sp += arg_size(); // restore stack pointer
Node *x = pop_math_arg();
Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
//-------------------
//result=(result.isNaN())? StrictMath::exp():result;
// Check: If isNaN() by checking result!=result? then go to Strict Math
Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
// Build the boolean node
Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
// End the current control-flow path
push_pair(x);
// Math.exp intrinsic returned a NaN, which requires StrictMath.exp
// to handle. Recompile without intrinsifying Math.exp
uncommon_trap(Deoptimization::Reason_intrinsic,
Deoptimization::Action_make_not_entrant);
}
finish_pow_exp(result, x, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
C->set_has_split_ifs(true); // Has chance for split-if optimization
push_pair(result);
return true;
}
......@@ -1577,17 +1617,12 @@ bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
assert(id == vmIntrinsics::_dpow, "Not pow");
// If this inlining ever returned NaN in the past, we do not intrinsify it
// every again. NaN results requires StrictMath.pow handling.
if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
// Do not intrinsify on older platforms which lack cmove.
if (ConditionalMoveLimit == 0) return false;
// Pseudocode for pow
// if (x <= 0.0) {
// if ((double)((int)y)==y) { // if y is int
// result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
// long longy = (long)y;
// if ((double)longy == y) { // if y is long
// if (y + 1 == y) longy = 0; // huge number: even
// result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
// } else {
// result = NaN;
// }
......@@ -1595,7 +1630,7 @@ bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
// result = DPow(x,y);
// }
// if (result != result)? {
// uncommon_trap();
// result = uncommon_trap() or runtime_call();
// }
// return result;
......@@ -1603,15 +1638,14 @@ bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
Node* y = pop_math_arg();
Node* x = pop_math_arg();
Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
Node* result = NULL;
// Short form: if not top-level (i.e., Math.pow but inlining Math.pow
// inside of something) then skip the fancy tests and just check for
// NaN result.
Node *result = NULL;
if( jvms()->depth() >= 1 ) {
result = fast_result;
if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
// Short form: skip the fancy tests and just check for NaN result.
result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
} else {
// If this inlining ever returned NaN in the past, include all
// checks + call to the runtime.
// Set the merge point for If node with condition of (x <= 0.0)
// There are four possible paths to region node and phi node
......@@ -1627,55 +1661,95 @@ bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
// Branch either way
IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
Node *opt_test = _gvn.transform(if1);
//assert( opt_test->is_If(), "Expect an IfNode");
IfNode *opt_if1 = (IfNode*)opt_test;
// Fast path taken; set region slot 3
Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(if1) );
r->init_req(3,fast_taken); // Capture fast-control
// Fast path not-taken, i.e. slow path
Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(if1) );
// Set fast path result
Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
phi->init_req(3, fast_result);
// Complex path
// Build the second if node (if y is int)
// Node for (int)y
Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
// Node for (double)((int) y)
Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
// Check (double)((int) y) : y
Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
// Check if (y isn't int) then go to slow path
Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
// Build the second if node (if y is long)
// Node for (long)y
Node *longy = _gvn.transform( new (C, 2) ConvD2LNode(y));
// Node for (double)((long) y)
Node *doublelongy= _gvn.transform( new (C, 2) ConvL2DNode(longy));
// Check (double)((long) y) : y
Node *cmplongy= _gvn.transform(new (C, 3) CmpDNode(doublelongy, y));
// Check if (y isn't long) then go to slow path
Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmplongy, BoolTest::ne ) );
// Branch either way
IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
Node *slow_path = opt_iff(r,if2); // Set region path 2
Node* ylong_path = _gvn.transform( new (C, 1) IfFalseNode(if2));
Node *slow_path = _gvn.transform( new (C, 1) IfTrueNode(if2) );
// Calculate DPow(abs(x), y)*(1 & (int)y)
// Calculate DPow(abs(x), y)*(1 & (long)y)
// Node for constant 1
Node *conone = intcon(1);
// 1& (int)y
Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
Node *conone = longcon(1);
// 1& (long)y
Node *signnode= _gvn.transform( new (C, 3) AndLNode(conone, longy) );
// A huge number is always even. Detect a huge number by checking
// if y + 1 == y and set integer to be tested for parity to 0.
// Required for corner case:
// (long)9.223372036854776E18 = max_jlong
// (double)(long)9.223372036854776E18 = 9.223372036854776E18
// max_jlong is odd but 9.223372036854776E18 is even
Node* yplus1 = _gvn.transform( new (C, 3) AddDNode(y, makecon(TypeD::make(1))));
Node *cmpyplus1= _gvn.transform(new (C, 3) CmpDNode(yplus1, y));
Node *bolyplus1 = _gvn.transform( new (C, 2) BoolNode( cmpyplus1, BoolTest::eq ) );
Node* correctedsign = NULL;
if (ConditionalMoveLimit != 0) {
correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
} else {
IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
RegionNode *r = new (C, 3) RegionNode(3);
Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
r->init_req(1, _gvn.transform( new (C, 1) IfFalseNode(ifyplus1)));
r->init_req(2, _gvn.transform( new (C, 1) IfTrueNode(ifyplus1)));
phi->init_req(1, signnode);
phi->init_req(2, longcon(0));
correctedsign = _gvn.transform(phi);
ylong_path = _gvn.transform(r);
record_for_igvn(r);
}
// zero node
Node *conzero = intcon(0);
// Check (1&(int)y)==0?
Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
// Check if (1&(int)y)!=0?, if so the result is negative
Node *conzero = longcon(0);
// Check (1&(long)y)==0?
Node *cmpeq1 = _gvn.transform(new (C, 3) CmpLNode(correctedsign, conzero));
// Check if (1&(long)y)!=0?, if so the result is negative
Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
// abs(x)
Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
// abs(x)^y
Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, absx, y) );
// -abs(x)^y
Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
// (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
// (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
Node *signresult = NULL;
if (ConditionalMoveLimit != 0) {
signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
} else {
IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
RegionNode *r = new (C, 3) RegionNode(3);
Node *phi = new (C, 3) PhiNode(r, Type::DOUBLE);
r->init_req(1, _gvn.transform( new (C, 1) IfFalseNode(ifyeven)));
r->init_req(2, _gvn.transform( new (C, 1) IfTrueNode(ifyeven)));
phi->init_req(1, absxpowy);
phi->init_req(2, negabsxpowy);
signresult = _gvn.transform(phi);
ylong_path = _gvn.transform(r);
record_for_igvn(r);
}
// Set complex path fast result
r->init_req(2, ylong_path);
phi->init_req(2, signresult);
static const jlong nan_bits = CONST64(0x7ff8000000000000);
......@@ -1689,27 +1763,10 @@ bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
result=_gvn.transform(phi);
}
//-------------------
//result=(result.isNaN())? uncommon_trap():result;
// Check: If isNaN() by checking result!=result? then go to Strict Math
Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
// Build the boolean node
Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
// End the current control-flow path
push_pair(x);
push_pair(y);
// Math.pow intrinsic returned a NaN, which requires StrictMath.pow
// to handle. Recompile without intrinsifying Math.pow.
uncommon_trap(Deoptimization::Reason_intrinsic,
Deoptimization::Action_make_not_entrant);
}
finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
C->set_has_split_ifs(true); // Has chance for split-if optimization
push_pair(result);
return true;
}
......
......@@ -1278,9 +1278,9 @@ void Parse::sharpen_type_after_if(BoolTest::mask btest,
// or the narrowOop equivalent.
const Type* obj_type = _gvn.type(obj);
const TypeOopPtr* tboth = obj_type->join(con_type)->isa_oopptr();
if (tboth != NULL && tboth != obj_type && tboth->higher_equal(obj_type)) {
if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
tboth->higher_equal(obj_type)) {
// obj has to be of the exact type Foo if the CmpP succeeds.
assert(tboth->klass_is_exact(), "klass should be exact");
int obj_in_map = map()->find_edge(obj);
JVMState* jvms = this->jvms();
if (obj_in_map >= 0 &&
......
......@@ -554,9 +554,7 @@ const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const {
return TypeInt::CC_GE;
} else if (hi0 <= lo1) {
// Check for special case in Hashtable::get. (See below.)
if ((jint)lo0 >= 0 && (jint)lo1 >= 0 &&
in(1)->Opcode() == Op_ModI &&
in(1)->in(2) == in(2) )
if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
return TypeInt::CC_LT;
return TypeInt::CC_LE;
}
......@@ -567,13 +565,17 @@ const Type *CmpUNode::sub( const Type *t1, const Type *t2 ) const {
// 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 &&
in(1)->Opcode() == Op_ModI &&
in(1)->in(2)->uncast() == in(2)->uncast())
if ((jint)lo0 >= 0 && (jint)lo1 >= 0 && is_index_range_check())
return TypeInt::CC_LT;
return TypeInt::CC; // else use worst case results
}
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)) {
......
......@@ -158,6 +158,7 @@ public:
CmpUNode( Node *in1, Node *in2 ) : CmpNode(in1,in2) {}
virtual int Opcode() const;
virtual const Type *sub( const Type *, const Type * ) const;
bool is_index_range_check() const;
};
//------------------------------CmpPNode---------------------------------------
......
/*
* Copyright (c) 2012, 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.
*
*/
/*
* Micro-benchmark for Math.pow() and Math.exp()
*/
import java.util.*;
public class Test7177917 {
static double d;
static Random r = new Random(0);
static long m_pow(double[][] values) {
double res = 0;
long start = System.nanoTime();
for (int i = 0; i < values.length; i++) {
res += Math.pow(values[i][0], values[i][1]);
}
long stop = System.nanoTime();
d = res;
return (stop - start) / 1000;
}
static long m_exp(double[] values) {
double res = 0;
long start = System.nanoTime();
for (int i = 0; i < values.length; i++) {
res += Math.exp(values[i]);
}
long stop = System.nanoTime();
d = res;
return (stop - start) / 1000;
}
static double[][] pow_values(int nb) {
double[][] res = new double[nb][2];
for (int i = 0; i < nb; i++) {
double ylogx = (1 + (r.nextDouble() * 2045)) - 1023; // 2045 rather than 2046 as a safety margin
double x = Math.abs(Double.longBitsToDouble(r.nextLong()));
while (x != x) {
x = Math.abs(Double.longBitsToDouble(r.nextLong()));
}
double logx = Math.log(x) / Math.log(2);
double y = ylogx / logx;
res[i][0] = x;
res[i][1] = y;
}
return res;
}
static double[] exp_values(int nb) {
double[] res = new double[nb];
for (int i = 0; i < nb; i++) {
double ylogx = (1 + (r.nextDouble() * 2045)) - 1023; // 2045 rather than 2046 as a safety margin
double x = Math.E;
double logx = Math.log(x) / Math.log(2);
double y = ylogx / logx;
res[i] = y;
}
return res;
}
static public void main(String[] args) {
{
// warmup
double[][] warmup_values = pow_values(10);
m_pow(warmup_values);
for (int i = 0; i < 20000; i++) {
m_pow(warmup_values);
}
// test pow perf
double[][] values = pow_values(1000000);
System.out.println("==> POW " + m_pow(values));
// force uncommon trap
double[][] nan_values = new double[1][2];
nan_values[0][0] = Double.NaN;
nan_values[0][1] = Double.NaN;
m_pow(nan_values);
// force recompilation
for (int i = 0; i < 20000; i++) {
m_pow(warmup_values);
}
// test pow perf again
System.out.println("==> POW " + m_pow(values));
}
{
// warmup
double[] warmup_values = exp_values(10);
m_exp(warmup_values);
for (int i = 0; i < 20000; i++) {
m_exp(warmup_values);
}
// test pow perf
double[] values = exp_values(1000000);
System.out.println("==> EXP " + m_exp(values));
// force uncommon trap
double[] nan_values = new double[1];
nan_values[0] = Double.NaN;
m_exp(nan_values);
// force recompilation
for (int i = 0; i < 20000; i++) {
m_exp(warmup_values);
}
// test pow perf again
System.out.println("==> EXP " + m_exp(values));
}
}
}
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册