//=============================================================================
uint TypeNode::size_of() const { return sizeof(*this); }
#ifndef PRODUCT
void TypeNode::dump_spec(outputStream *st) const {
if( !Verbose && !WizardMode ) {
// standard dump does this in Verbose and WizardMode
st->print(" #"); _type->dump_on(st);
}
}
#endif
uint TypeNode::hash() const {
return Node::hash() + _type->hash();
}
uint TypeNode::cmp( const Node &n ) const
{ return !Type::cmp( _type, ((TypeNode&)n)._type ); }
const Type *TypeNode::bottom_type() const { return _type; }
const Type *TypeNode::Value( PhaseTransform * ) const { return _type; }
//------------------------------ideal_reg--------------------------------------
uint TypeNode::ideal_reg() const {
return _type->ideal_reg();
}
C:\hotspot-69087d08d473\src\share\vm/opto/node.hpp
/*
* Copyright (c) 1997, 2016, 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.
*
*/
#ifndef SHARE_VM_OPTO_NODE_HPP
#define SHARE_VM_OPTO_NODE_HPP
#include "libadt/port.hpp"
#include "libadt/vectset.hpp"
#include "opto/compile.hpp"
#include "opto/type.hpp"
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
class AbstractLockNode;
class AddNode;
class AddPNode;
class AliasInfo;
class AllocateArrayNode;
class AllocateNode;
class Block;
class BoolNode;
class BoxLockNode;
class CMoveNode;
class CallDynamicJavaNode;
class CallJavaNode;
class CallLeafNode;
class CallNode;
class CallRuntimeNode;
class CallStaticJavaNode;
class CatchNode;
class CatchProjNode;
class CheckCastPPNode;
class CastIINode;
class ClearArrayNode;
class CmpNode;
class CodeBuffer;
class ConstraintCastNode;
class ConNode;
class CountedLoopNode;
class CountedLoopEndNode;
class DecodeNarrowPtrNode;
class DecodeNNode;
class DecodeNKlassNode;
class EncodeNarrowPtrNode;
class EncodePNode;
class EncodePKlassNode;
class FastLockNode;
class FastUnlockNode;
class IfNode;
class IfFalseNode;
class IfTrueNode;
class InitializeNode;
class JVMState;
class JumpNode;
class JumpProjNode;
class LoadNode;
class LoadStoreNode;
class LockNode;
class LoopNode;
class MachBranchNode;
class MachCallDynamicJavaNode;
class MachCallJavaNode;
class MachCallLeafNode;
class MachCallNode;
class MachCallRuntimeNode;
class MachCallStaticJavaNode;
class MachConstantBaseNode;
class MachConstantNode;
class MachGotoNode;
class MachIfNode;
class MachNode;
class MachNullCheckNode;
class MachProjNode;
class MachReturnNode;
class MachSafePointNode;
class MachSpillCopyNode;
class MachTempNode;
class MachMergeNode;
class Matcher;
class MemBarNode;
class MemBarStoreStoreNode;
class MemNode;
class MergeMemNode;
class MulNode;
class MultiNode;
class MultiBranchNode;
class NeverBranchNode;
class Node;
class Node_Array;
class Node_List;
class Node_Stack;
class NullCheckNode;
class OopMap;
class ParmNode;
class PCTableNode;
class PhaseCCP;
class PhaseGVN;
class PhaseIterGVN;
class PhaseRegAlloc;
class PhaseTransform;
class PhaseValues;
class PhiNode;
class Pipeline;
class ProjNode;
class RegMask;
class RegionNode;
class RootNode;
class SafePointNode;
class SafePointScalarObjectNode;
class StartNode;
class State;
class StoreNode;
class SubNode;
class Type;
class TypeNode;
class UnlockNode;
class VectorNode;
class LoadVectorNode;
class StoreVectorNode;
class VectorSet;
typedef void (*NFunc)(Node&,void*);
extern "C" {
typedef int (*C_sort_func_t)(const void *, const void *);
}
// The type of all node counts and indexes.
// It must hold at least 16 bits, but must also be fast to load and store.
// This type, if less than 32 bits, could limit the number of possible nodes.
// (To make this type platform-specific, move to globalDefinitions_xxx.hpp.)
typedef unsigned int node_idx_t;
#ifndef OPTO_DU_ITERATOR_ASSERT
#ifdef ASSERT
#define OPTO_DU_ITERATOR_ASSERT 1
#else
#define OPTO_DU_ITERATOR_ASSERT 0
#endif
#endif //OPTO_DU_ITERATOR_ASSERT
#if OPTO_DU_ITERATOR_ASSERT
class DUIterator;
class DUIterator_Fast;
class DUIterator_Last;
#else
typedef uint DUIterator;
typedef Node** DUIterator_Fast;
typedef Node** DUIterator_Last;
#endif
// Node Sentinel
#define NodeSentinel (Node*)-1
// Unknown count frequency
#define COUNT_UNKNOWN (-1.0f)
//------------------------------Node-------------------------------------------
// Nodes define actions in the program. They create values, which have types.
// They are both vertices in a directed graph and program primitives. Nodes
// are labeled; the label is the "opcode", the primitive function in the lambda
// calculus sense that gives meaning to the Node. Node inputs are ordered (so
// that "a-b" is different from "b-a"). The inputs to a Node are the inputs to
// the Node's function. These inputs also define a Type equation for the Node.
// Solving these Type equations amounts to doing dataflow analysis.
// Control and data are uniformly represented in the graph. Finally, Nodes
// have a unique dense integer index which is used to index into side arrays
// whenever I have phase-specific information.
class Node {
friend class VMStructs;
// Lots of restrictions on cloning Nodes
Node(const Node&); // not defined; linker error to use these
Node &operator=(const Node &rhs);
public:
friend class Compile;
#if OPTO_DU_ITERATOR_ASSERT
friend class DUIterator_Common;
friend class DUIterator;
friend class DUIterator_Fast;
friend class DUIterator_Last;
#endif
// Because Nodes come and go, I define an Arena of Node structures to pull
// from. This should allow fast access to node creation & deletion. This
// field is a local cache of a value defined in some "program fragment" for
// which these Nodes are just a part of.
// New Operator that takes a Compile pointer, this will eventually
// be the "new" New operator.
inline void* operator new( size_t x, Compile* C) throw() {
Node* n = (Node*)C->node_arena()->Amalloc_D(x);
#ifdef ASSERT
n->_in = (Node**)n; // magic cookie for assertion check
#endif
n->_out = (Node**)C;
return (void*)n;
}
// Delete is a NOP
void operator delete( void *ptr ) {}
// Fancy destructor; eagerly attempt to reclaim Node numberings and storage
void destruct();
// Create a new Node. Required is the number is of inputs required for
// semantic correctness.
Node( uint required );
// Create a new Node with given input edges.
// This version requires use of the "edge-count" new.
// E.g. new (C,3) FooNode( C, NULL, left, right );
Node( Node *n0 );
Node( Node *n0, Node *n1 );
Node( Node *n0, Node *n1, Node *n2 );
Node( Node *n0, Node *n1, Node *n2, Node *n3 );
Node( Node *n0, Node *n1, Node *n2, Node *n3, Node *n4 );
Node( Node *n0, Node *n1, Node *n2, Node *n3, Node *n4, Node *n5 );
Node( Node *n0, Node *n1, Node *n2, Node *n3,
Node *n4, Node *n5, Node *n6 );
// Clone an inherited Node given only the base Node type.
Node* clone() const;
// Clone a Node, immediately supplying one or two new edges.
// The first and second arguments, if non-null, replace in(1) and in(2),
// respectively.
Node* clone_with_data_edge(Node* in1, Node* in2 = NULL) const {
Node* nn = clone();
if (in1 != NULL) nn->set_req(1, in1);
if (in2 != NULL) nn->set_req(2, in2);
return nn;
}
private:
// Shared setup for the above constructors.
// Handles all interactions with Compile::current.
// Puts initial values in all Node fields except _idx.
// Returns the initial value for _idx, which cannot
// be initialized by assignment.
inline int Init(int req, Compile* C);
//----------------- input edge handling
protected:
friend class PhaseCFG; // Access to address of _in array elements
Node **_in; // Array of use-def references to Nodes
Node **_out; // Array of def-use references to Nodes
// Input edges are split into two categories. Required edges are required
// for semantic correctness; order is important and NULLs are allowed.
// Precedence edges are used to help determine execution order and are
// added, e.g., for scheduling purposes. They are unordered and not
// duplicated; they have no embedded NULLs. Edges from 0 to _cnt-1
// are required, from _cnt to _max-1 are precedence edges.
node_idx_t _cnt; // Total number of required Node inputs.
node_idx_t _max; // Actual length of input array.
// Output edges are an unordered list of def-use edges which exactly
// correspond to required input edges which point from other nodes
// to this one. Thus the count of the output edges is the number of
// users of this node.
node_idx_t _outcnt; // Total number of Node outputs.
node_idx_t _outmax; // Actual length of output array.
// Grow the actual input array to the next larger power-of-2 bigger than len.
void grow( uint len );
// Grow the output array to the next larger power-of-2 bigger than len.
void out_grow( uint len );
public:
// Each Node is assigned a unique small/dense number. This number is used
// to index into auxiliary arrays of data and bit vectors.
// The field _idx is declared constant to defend against inadvertent assignments,
// since it is used by clients as a naked field. However, the field's value can be
// changed using the set_idx() method.
//
// The PhaseRenumberLive phase renumbers nodes based on liveness information.
// Therefore, it updates the value of the _idx field. The parse-time _idx is
// preserved in _parse_idx.
const node_idx_t _idx;
DEBUG_ONLY(const node_idx_t _parse_idx;)
// Get the (read-only) number of input edges
uint req() const { return _cnt; }
uint len() const { return _max; }
// Get the (read-only) number of output edges
uint outcnt() const { return _outcnt; }
#if OPTO_DU_ITERATOR_ASSERT
// Iterate over the out-edges of this node. Deletions are illegal.
inline DUIterator outs() const;
// Use this when the out array might have changed to suppress asserts.
inline DUIterator& refresh_out_pos(DUIterator& i) const;
// Does the node have an out at this position? (Used for iteration.)
inline bool has_out(DUIterator& i) const;
inline Node* out(DUIterator& i) const;
// Iterate over the out-edges of this node. All changes are illegal.
inline DUIterator_Fast fast_outs(DUIterator_Fast& max) const;
inline Node* fast_out(DUIterator_Fast& i) const;
// Iterate over the out-edges of this node, deleting one at a time.
inline DUIterator_Last last_outs(DUIterator_Last& min) const;
inline Node* last_out(DUIterator_Last& i) const;
// The inline bodies of all these methods are after the iterator definitions.
#else
// Iterate over the out-edges of this node. Deletions are illegal.
// This iteration uses integral indexes, to decouple from array reallocations.
DUIterator outs() const { return 0; }
// Use this when the out array might have changed to suppress asserts.
DUIterator refresh_out_pos(DUIterator i) const { return i; }
// Reference to the i'th output Node. Error if out of bounds.
Node* out(DUIterator i) const { assert(i < _outcnt, "oob"); return _out[i]; }
// Does the node have an out at this position? (Used for iteration.)
bool has_out(DUIterator i) const { return i < _outcnt; }
// Iterate over the out-edges of this node. All changes are illegal.
// This iteration uses a pointer internal to the out array.
DUIterator_Fast fast_outs(DUIterator_Fast& max) const {
Node** out = _out;
// Assign a limit pointer to the reference argument:
max = out + (ptrdiff_t)_outcnt;
// Return the base pointer:
return out;
}
Node* fast_out(DUIterator_Fast i) const { return *i; }
// Iterate over the out-edges of this node, deleting one at a time.
// This iteration uses a pointer internal to the out array.
DUIterator_Last last_outs(DUIterator_Last& min) const {
Node** out = _out;
// Assign a limit pointer to the reference argument:
min = out;
// Return the pointer to the start of the iteration:
return out + (ptrdiff_t)_outcnt - 1;
}
Node* last_out(DUIterator_Last i) const { return *i; }
#endif
// Reference to the i'th input Node. Error if out of bounds.
Node* in(uint i) const { assert(i < _max, err_msg_res("oob: i=%d, _max=%d", i, _max)); return _in[i]; }
// Reference to the i'th input Node. NULL if out of bounds.
Node* lookup(uint i) const { return ((i < _max) ? _in[i] : NULL); }
// Reference to the i'th output Node. Error if out of bounds.
// Use this accessor sparingly. We are going trying to use iterators instead.
Node* raw_out(uint i) const { assert(i < _outcnt,"oob"); return _out[i]; }
// Return the unique out edge.
Node* unique_out() const { assert(_outcnt==1,"not unique"); return _out[0]; }
// Delete out edge at position 'i' by moving last out edge to position 'i'
void raw_del_out(uint i) {
assert(i < _outcnt,"oob");
assert(_outcnt > 0,"oob");
#if OPTO_DU_ITERATOR_ASSERT
// Record that a change happened here.
debug_only(_last_del = _out[i]; ++_del_tick);
#endif
_out[i] = _out[--_outcnt];
// Smash the old edge so it can't be used accidentally.
debug_only(_out[_outcnt] = (Node *)(uintptr_t)0xdeadbeef);
}
#ifdef ASSERT
bool is_dead() const;
#define is_not_dead(n) ((n) == NULL || !VerifyIterativeGVN || !((n)->is_dead()))
#endif
// Check whether node has become unreachable
bool is_unreachable(PhaseIterGVN &igvn) const;
// Set a required input edge, also updates corresponding output edge
void add_req( Node *n ); // Append a NEW required input
void add_req( Node *n0, Node *n1 ) {
add_req(n0); add_req(n1); }
void add_req( Node *n0, Node *n1, Node *n2 ) {
add_req(n0); add_req(n1); add_req(n2); }
void add_req_batch( Node* n, uint m ); // Append m NEW required inputs (all n).
void del_req( uint idx ); // Delete required edge & compact
void del_req_ordered( uint idx ); // Delete required edge & compact with preserved order
void ins_req( uint i, Node *n ); // Insert a NEW required input
void set_req( uint i, Node *n ) {
assert( is_not_dead(n), "can not use dead node");
assert( i < _cnt, err_msg_res("oob: i=%d, _cnt=%d", i, _cnt));
assert( !VerifyHashTableKeys || _hash_lock == 0,
"remove node from hash table before modifying it");
Node** p = &_in[i]; // cache this._in, across the del_out call
if (*p != NULL) (*p)->del_out((Node *)this);
(*p) = n;
if (n != NULL) n->add_out((Node *)this);
}
// Light version of set_req() to init inputs after node creation.
void init_req( uint i, Node *n ) {
assert( i == 0 && this == n ||
is_not_dead(n), "can not use dead node");
assert( i < _cnt, "oob");
assert( !VerifyHashTableKeys || _hash_lock == 0,
"remove node from hash table before modifying it");
assert( _in[i] == NULL, "sanity");
_in[i] = n;
if (n != NULL) n->add_out((Node *)this);
}
// Find first occurrence of n among my edges:
int find_edge(Node* n);
int find_prec_edge(Node* n) {
for (uint i = req(); i < len(); i++) {
if (_in[i] == n) return i;
if (_in[i] == NULL) {
DEBUG_ONLY( while ((++i) < len()) assert(_in[i] == NULL, "Gap in prec edges!"); )
break;
}
}
return -1;
}
int replace_edge(Node* old, Node* neww);
int replace_edges_in_range(Node* old, Node* neww, int start, int end);
// NULL out all inputs to eliminate incoming Def-Use edges.
// Return the number of edges between 'n' and 'this'
int disconnect_inputs(Node *n, Compile *c);
// Quickly, return true if and only if I am Compile::current()->top().
bool is_top() const {
assert((this == (Node*) Compile::current()->top()) == (_out == NULL), "");
return (_out == NULL);
}
// Reaffirm invariants for is_top. (Only from Compile::set_cached_top_node.)
void setup_is_top();
// Strip away casting. (It is depth-limited.)
Node* uncast() const;
// Return whether two Nodes are equivalent, after stripping casting.
bool eqv_uncast(const Node* n) const {
return (this->uncast() == n->uncast());
}
// Find out of current node that matches opcode.
Node* find_out_with(int opcode);
private:
static Node* uncast_helper(const Node* n);
// Add an output edge to the end of the list
void add_out( Node *n ) {
if (is_top()) return;
if( _outcnt == _outmax ) out_grow(_outcnt);
_out[_outcnt++] = n;
}
// Delete an output edge
void del_out( Node *n ) {
if (is_top()) return;
Node** outp = &_out[_outcnt];
// Find and remove n
do {
assert(outp > _out, "Missing Def-Use edge");
} while (*--outp != n);
*outp = _out[--_outcnt];
// Smash the old edge so it can't be used accidentally.
debug_only(_out[_outcnt] = (Node *)(uintptr_t)0xdeadbeef);
// Record that a change happened here.
#if OPTO_DU_ITERATOR_ASSERT
debug_only(_last_del = n; ++_del_tick);
#endif
}
// Close gap after removing edge.
void close_prec_gap_at(uint gap) {
assert(_cnt <= gap && gap < _max, "no valid prec edge");
uint i = gap;
Node *last = NULL;
for (; i < _max-1; ++i) {
Node *next = _in[i+1];
if (next == NULL) break;
last = next;
}
_in[gap] = last; // Move last slot to empty one.
_in[i] = NULL; // NULL out last slot.
}
public:
// Globally replace this node by a given new node, updating all uses.
void replace_by(Node* new_node);
// Globally replace this node by a given new node, updating all uses
// and cutting input edges of old node.
void subsume_by(Node* new_node, Compile* c) {
replace_by(new_node);
disconnect_inputs(NULL, c);
}
void set_req_X( uint i, Node *n, PhaseIterGVN *igvn );
// Find the one non-null required input. RegionNode only
Node *nonnull_req() const;
// Add or remove precedence edges
void add_prec( Node *n );
void rm_prec( uint i );
// Note: prec(i) will not necessarily point to n if edge already exists.
void set_prec( uint i, Node *n ) {
assert(i < _max, err_msg("oob: i=%d, _max=%d", i, _max));
assert(is_not_dead(n), "can not use dead node");
assert(i >= _cnt, "not a precedence edge");
// Avoid spec violation: duplicated prec edge.
if (_in[i] == n) return;
if (n == NULL || find_prec_edge(n) != -1) {
rm_prec(i);
return;
}
if (_in[i] != NULL) _in[i]->del_out((Node *)this);
_in[i] = n;
if (n != NULL) n->add_out((Node *)this);
}
// Set this node's index, used by cisc_version to replace current node
void set_idx(uint new_idx) {
const node_idx_t* ref = &_idx;
*(node_idx_t*)ref = new_idx;
}
// Swap input edge order. (Edge indexes i1 and i2 are usually 1 and 2.)
void swap_edges(uint i1, uint i2) {
debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH);
// Def-Use info is unchanged
Node* n1 = in(i1);
Node* n2 = in(i2);
_in[i1] = n2;
_in[i2] = n1;
// If this node is in the hash table, make sure it doesn't need a rehash.
assert(check_hash == NO_HASH || check_hash == hash(), "edge swap must preserve hash code");
}
// Iterators over input Nodes for a Node X are written as:
// for( i = 0; i < X.req(); i++ ) ... X[i] ...
// NOTE: Required edges can contain embedded NULL pointers.
//----------------- Other Node Properties
// Generate class id for some ideal nodes to avoid virtual query
// methods is_<Node>().
// Class id is the set of bits corresponded to the node class and all its
// super classes so that queries for super classes are also valid.
// Subclasses of the same super class have different assigned bit
// (the third parameter in the macro DEFINE_CLASS_ID).
// Classes with deeper hierarchy are declared first.
// Classes with the same hierarchy depth are sorted by usage frequency.
//
// The query method masks the bits to cut off bits of subclasses
// and then compare the result with the class id
// (see the macro DEFINE_CLASS_QUERY below).
//
// Class_MachCall=30, ClassMask_MachCall=31
// 12 8 4 0
// 0 0 0 0 0 0 0 0 1 1 1 1 0
// | | | |
// | | | Bit_Mach=2
// | | Bit_MachReturn=4
// | Bit_MachSafePoint=8
// Bit_MachCall=16
//
// Class_CountedLoop=56, ClassMask_CountedLoop=63
// 12 8 4 0
// 0 0 0 0 0 0 0 1 1 1 0 0 0
// | | |
// | | Bit_Region=8
// | Bit_Loop=16
// Bit_CountedLoop=32
#define DEFINE_CLASS_ID(cl, supcl, subn) \
Bit_##cl = (Class_##supcl == 0) ? 1 << subn : (Bit_##supcl) << (1 + subn) , \
Class_##cl = Class_##supcl + Bit_##cl , \
ClassMask_##cl = ((Bit_##cl << 1) - 1) ,
// This enum is used only for C2 ideal and mach nodes with is_<node>() methods
// so that it's values fits into 16 bits.
enum NodeClasses {
Bit_Node = 0x0000,
Class_Node = 0x0000,
ClassMask_Node = 0xFFFF,
DEFINE_CLASS_ID(Multi, Node, 0)
DEFINE_CLASS_ID(SafePoint, Multi, 0)
DEFINE_CLASS_ID(Call, SafePoint, 0)
DEFINE_CLASS_ID(CallJava, Call, 0)
DEFINE_CLASS_ID(CallStaticJava, CallJava, 0)
DEFINE_CLASS_ID(CallDynamicJava, CallJava, 1)
DEFINE_CLASS_ID(CallRuntime, Call, 1)
DEFINE_CLASS_ID(CallLeaf, CallRuntime, 0)
DEFINE_CLASS_ID(Allocate, Call, 2)
DEFINE_CLASS_ID(AllocateArray, Allocate, 0)
DEFINE_CLASS_ID(AbstractLock, Call, 3)
DEFINE_CLASS_ID(Lock, AbstractLock, 0)
DEFINE_CLASS_ID(Unlock, AbstractLock, 1)
DEFINE_CLASS_ID(MultiBranch, Multi, 1)
DEFINE_CLASS_ID(PCTable, MultiBranch, 0)
DEFINE_CLASS_ID(Catch, PCTable, 0)
DEFINE_CLASS_ID(Jump, PCTable, 1)
DEFINE_CLASS_ID(If, MultiBranch, 1)
DEFINE_CLASS_ID(CountedLoopEnd, If, 0)
DEFINE_CLASS_ID(NeverBranch, MultiBranch, 2)
DEFINE_CLASS_ID(Start, Multi, 2)
DEFINE_CLASS_ID(MemBar, Multi, 3)
DEFINE_CLASS_ID(Initialize, MemBar, 0)
DEFINE_CLASS_ID(MemBarStoreStore, MemBar, 1)
DEFINE_CLASS_ID(Mach, Node, 1)
DEFINE_CLASS_ID(MachReturn, Mach, 0)
DEFINE_CLASS_ID(MachSafePoint, MachReturn, 0)
DEFINE_CLASS_ID(MachCall, MachSafePoint, 0)
DEFINE_CLASS_ID(MachCallJava, MachCall, 0)
DEFINE_CLASS_ID(MachCallStaticJava, MachCallJava, 0)
DEFINE_CLASS_ID(MachCallDynamicJava, MachCallJava, 1)
DEFINE_CLASS_ID(MachCallRuntime, MachCall, 1)
DEFINE_CLASS_ID(MachCallLeaf, MachCallRuntime, 0)
DEFINE_CLASS_ID(MachBranch, Mach, 1)
DEFINE_CLASS_ID(MachIf, MachBranch, 0)
DEFINE_CLASS_ID(MachGoto, MachBranch, 1)
DEFINE_CLASS_ID(MachNullCheck, MachBranch, 2)
DEFINE_CLASS_ID(MachSpillCopy, Mach, 2)
DEFINE_CLASS_ID(MachTemp, Mach, 3)
DEFINE_CLASS_ID(MachConstantBase, Mach, 4)
DEFINE_CLASS_ID(MachConstant, Mach, 5)
DEFINE_CLASS_ID(MachMerge, Mach, 6)
DEFINE_CLASS_ID(Type, Node, 2)
DEFINE_CLASS_ID(Phi, Type, 0)
DEFINE_CLASS_ID(ConstraintCast, Type, 1)
DEFINE_CLASS_ID(CastII, ConstraintCast, 0)
DEFINE_CLASS_ID(CheckCastPP, Type, 2)
DEFINE_CLASS_ID(CMove, Type, 3)
DEFINE_CLASS_ID(SafePointScalarObject, Type, 4)
DEFINE_CLASS_ID(DecodeNarrowPtr, Type, 5)
DEFINE_CLASS_ID(DecodeN, DecodeNarrowPtr, 0)
DEFINE_CLASS_ID(DecodeNKlass, DecodeNarrowPtr, 1)
DEFINE_CLASS_ID(EncodeNarrowPtr, Type, 6)
DEFINE_CLASS_ID(EncodeP, EncodeNarrowPtr, 0)
DEFINE_CLASS_ID(EncodePKlass, EncodeNarrowPtr, 1)
DEFINE_CLASS_ID(Proj, Node, 3)
DEFINE_CLASS_ID(CatchProj, Proj, 0)
DEFINE_CLASS_ID(JumpProj, Proj, 1)
DEFINE_CLASS_ID(IfTrue, Proj, 2)
DEFINE_CLASS_ID(IfFalse, Proj, 3)
DEFINE_CLASS_ID(Parm, Proj, 4)
DEFINE_CLASS_ID(MachProj, Proj, 5)
DEFINE_CLASS_ID(Mem, Node, 4)
DEFINE_CLASS_ID(Load, Mem, 0)
DEFINE_CLASS_ID(LoadVector, Load, 0)
DEFINE_CLASS_ID(Store, Mem, 1)
DEFINE_CLASS_ID(StoreVector, Store, 0)
DEFINE_CLASS_ID(LoadStore, Mem, 2)
DEFINE_CLASS_ID(Region, Node, 5)
DEFINE_CLASS_ID(Loop, Region, 0)
DEFINE_CLASS_ID(Root, Loop, 0)
DEFINE_CLASS_ID(CountedLoop, Loop, 1)
DEFINE_CLASS_ID(Sub, Node, 6)
DEFINE_CLASS_ID(Cmp, Sub, 0)
DEFINE_CLASS_ID(FastLock, Cmp, 0)
DEFINE_CLASS_ID(FastUnlock, Cmp, 1)
DEFINE_CLASS_ID(MergeMem, Node, 7)
DEFINE_CLASS_ID(Bool, Node, 8)
DEFINE_CLASS_ID(AddP, Node, 9)
DEFINE_CLASS_ID(BoxLock, Node, 10)
DEFINE_CLASS_ID(Add, Node, 11)
DEFINE_CLASS_ID(Mul, Node, 12)
DEFINE_CLASS_ID(Vector, Node, 13)
DEFINE_CLASS_ID(ClearArray, Node, 14)
_max_classes = ClassMask_ClearArray
};
#undef DEFINE_CLASS_ID
// Flags are sorted by usage frequency.
enum NodeFlags {
Flag_is_Copy = 0x01, // should be first bit to avoid shift
Flag_rematerialize = Flag_is_Copy << 1,
Flag_needs_anti_dependence_check = Flag_rematerialize << 1,
Flag_is_macro = Flag_needs_anti_dependence_check << 1,
Flag_is_Con = Flag_is_macro << 1,
Flag_is_cisc_alternate = Flag_is_Con << 1,
Flag_is_dead_loop_safe = Flag_is_cisc_alternate << 1,
Flag_may_be_short_branch = Flag_is_dead_loop_safe << 1,
Flag_avoid_back_to_back_before = Flag_may_be_short_branch << 1,
Flag_avoid_back_to_back_after = Flag_avoid_back_to_back_before << 1,
Flag_has_call = Flag_avoid_back_to_back_after << 1,
Flag_is_expensive = Flag_has_call << 1,
_max_flags = (Flag_is_expensive << 1) - 1 // allow flags combination
};
private:
jushort _class_id;
jushort _flags;
protected:
// These methods should be called from constructors only.
void init_class_id(jushort c) {
assert(c <= _max_classes, "invalid node class");
_class_id = c; // cast out const
}
void init_flags(jushort fl) {
assert(fl <= _max_flags, "invalid node flag");
_flags |= fl;
}
void clear_flag(jushort fl) {
assert(fl <= _max_flags, "invalid node flag");
_flags &= ~fl;
}
public:
const jushort class_id() const { return _class_id; }
const jushort flags() const { return _flags; }
// Return a dense integer opcode number
virtual int Opcode() const;
// Virtual inherited Node size
virtual uint size_of() const;
// Other interesting Node properties
#define DEFINE_CLASS_QUERY(type) \
bool is_##type() const { \
return ((_class_id & ClassMask_##type) == Class_##type); \
} \
type##Node *as_##type() const { \
assert(is_##type(), "invalid node class"); \
return (type##Node*)this; \
} \
type##Node* isa_##type() const { \
return (is_##type()) ? as_##type() : NULL; \
}
DEFINE_CLASS_QUERY(AbstractLock)
DEFINE_CLASS_QUERY(Add)
DEFINE_CLASS_QUERY(AddP)
DEFINE_CLASS_QUERY(Allocate)
DEFINE_CLASS_QUERY(AllocateArray)
DEFINE_CLASS_QUERY(Bool)
DEFINE_CLASS_QUERY(BoxLock)
DEFINE_CLASS_QUERY(Call)
DEFINE_CLASS_QUERY(CallDynamicJava)
DEFINE_CLASS_QUERY(CallJava)
DEFINE_CLASS_QUERY(CallLeaf)
DEFINE_CLASS_QUERY(CallRuntime)
DEFINE_CLASS_QUERY(CallStaticJava)
DEFINE_CLASS_QUERY(Catch)
DEFINE_CLASS_QUERY(CatchProj)
DEFINE_CLASS_QUERY(CheckCastPP)
DEFINE_CLASS_QUERY(CastII)
DEFINE_CLASS_QUERY(ConstraintCast)
DEFINE_CLASS_QUERY(ClearArray)
DEFINE_CLASS_QUERY(CMove)
DEFINE_CLASS_QUERY(Cmp)
DEFINE_CLASS_QUERY(CountedLoop)
DEFINE_CLASS_QUERY(CountedLoopEnd)
DEFINE_CLASS_QUERY(DecodeNarrowPtr)
DEFINE_CLASS_QUERY(DecodeN)
DEFINE_CLASS_QUERY(DecodeNKlass)
DEFINE_CLASS_QUERY(EncodeNarrowPtr)
DEFINE_CLASS_QUERY(EncodeP)
DEFINE_CLASS_QUERY(EncodePKlass)
DEFINE_CLASS_QUERY(FastLock)
DEFINE_CLASS_QUERY(FastUnlock)
DEFINE_CLASS_QUERY(If)
DEFINE_CLASS_QUERY(IfFalse)
DEFINE_CLASS_QUERY(IfTrue)
DEFINE_CLASS_QUERY(Initialize)
DEFINE_CLASS_QUERY(Jump)
DEFINE_CLASS_QUERY(JumpProj)
DEFINE_CLASS_QUERY(Load)
DEFINE_CLASS_QUERY(LoadStore)
DEFINE_CLASS_QUERY(Lock)
DEFINE_CLASS_QUERY(Loop)
DEFINE_CLASS_QUERY(Mach)
DEFINE_CLASS_QUERY(MachBranch)
DEFINE_CLASS_QUERY(MachCall)
DEFINE_CLASS_QUERY(MachCallDynamicJava)
DEFINE_CLASS_QUERY(MachCallJava)
DEFINE_CLASS_QUERY(MachCallLeaf)
DEFINE_CLASS_QUERY(MachCallRuntime)
DEFINE_CLASS_QUERY(MachCallStaticJava)
DEFINE_CLASS_QUERY(MachConstantBase)
DEFINE_CLASS_QUERY(MachConstant)
DEFINE_CLASS_QUERY(MachGoto)
DEFINE_CLASS_QUERY(MachIf)
DEFINE_CLASS_QUERY(MachNullCheck)
DEFINE_CLASS_QUERY(MachProj)
DEFINE_CLASS_QUERY(MachReturn)
DEFINE_CLASS_QUERY(MachSafePoint)
DEFINE_CLASS_QUERY(MachSpillCopy)
DEFINE_CLASS_QUERY(MachTemp)
DEFINE_CLASS_QUERY(MachMerge)
DEFINE_CLASS_QUERY(Mem)
DEFINE_CLASS_QUERY(MemBar)
DEFINE_CLASS_QUERY(MemBarStoreStore)
DEFINE_CLASS_QUERY(MergeMem)
DEFINE_CLASS_QUERY(Mul)
DEFINE_CLASS_QUERY(Multi)
DEFINE_CLASS_QUERY(MultiBranch)
DEFINE_CLASS_QUERY(Parm)
DEFINE_CLASS_QUERY(PCTable)
DEFINE_CLASS_QUERY(Phi)
DEFINE_CLASS_QUERY(Proj)
DEFINE_CLASS_QUERY(Region)
DEFINE_CLASS_QUERY(Root)
DEFINE_CLASS_QUERY(SafePoint)
DEFINE_CLASS_QUERY(SafePointScalarObject)
DEFINE_CLASS_QUERY(Start)
DEFINE_CLASS_QUERY(Store)
DEFINE_CLASS_QUERY(Sub)
DEFINE_CLASS_QUERY(Type)
DEFINE_CLASS_QUERY(Vector)
DEFINE_CLASS_QUERY(LoadVector)
DEFINE_CLASS_QUERY(StoreVector)
DEFINE_CLASS_QUERY(Unlock)
#undef DEFINE_CLASS_QUERY
// duplicate of is_MachSpillCopy()
bool is_SpillCopy () const {
return ((_class_id & ClassMask_MachSpillCopy) == Class_MachSpillCopy);
}
bool is_Con () const { return (_flags & Flag_is_Con) != 0; }
// The data node which is safe to leave in dead loop during IGVN optimization.
bool is_dead_loop_safe() const {
return is_Phi() || (is_Proj() && in(0) == NULL) ||
((_flags & (Flag_is_dead_loop_safe | Flag_is_Con)) != 0 &&
(!is_Proj() || !in(0)->is_Allocate()));
}
// is_Copy() returns copied edge index (0 or 1)
uint is_Copy() const { return (_flags & Flag_is_Copy); }
virtual bool is_CFG() const { return false; }
// If this node is control-dependent on a test, can it be
// rerouted to a dominating equivalent test? This is usually
// true of non-CFG nodes, but can be false for operations which
// depend for their correct sequencing on more than one test.
// (In that case, hoisting to a dominating test may silently
// skip some other important test.)
virtual bool depends_only_on_test() const { assert(!is_CFG(), ""); return true; };
// When building basic blocks, I need to have a notion of block beginning
// Nodes, next block selector Nodes (block enders), and next block
// projections. These calls need to work on their machine equivalents. The
// Ideal beginning Nodes are RootNode, RegionNode and StartNode.
bool is_block_start() const {
if ( is_Region() )
return this == (const Node*)in(0);
else
return is_Start();
}
// The Ideal control projection Nodes are IfTrue/IfFalse, JumpProjNode, Root,
// Goto and Return. This call also returns the block ending Node.
virtual const Node *is_block_proj() const;
// The node is a "macro" node which needs to be expanded before matching
bool is_macro() const { return (_flags & Flag_is_macro) != 0; }
// The node is expensive: the best control is set during loop opts
bool is_expensive() const { return (_flags & Flag_is_expensive) != 0 && in(0) != NULL; }
//----------------- Optimization
// Get the worst-case Type output for this Node.
virtual const class Type *bottom_type() const;
// If we find a better type for a node, try to record it permanently.
// Return true if this node actually changed.
// Be sure to do the hash_delete game in the "rehash" variant.
void raise_bottom_type(const Type* new_type);
// Get the address type with which this node uses and/or defs memory,
// or NULL if none. The address type is conservatively wide.
// Returns non-null for calls, membars, loads, stores, etc.
// Returns TypePtr::BOTTOM if the node touches memory "broadly".
virtual const class TypePtr *adr_type() const { return NULL; }
// Return an existing node which computes the same function as this node.
// The optimistic combined algorithm requires this to return a Node which
// is a small number of steps away (e.g., one of my inputs).
virtual Node *Identity( PhaseTransform *phase );
// Return the set of values this Node can take on at runtime.
virtual const Type *Value( PhaseTransform *phase ) const;
// Return a node which is more "ideal" than the current node.
// The invariants on this call are subtle. If in doubt, read the
// treatise in node.cpp above the default implemention AND TEST WITH
// +VerifyIterativeGVN!
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
// Some nodes have specific Ideal subgraph transformations only if they are
// unique users of specific nodes. Such nodes should be put on IGVN worklist
// for the transformations to happen.
bool has_special_unique_user() const;
// Skip Proj and CatchProj nodes chains. Check for Null and Top.
Node* find_exact_control(Node* ctrl);
// Check if 'this' node dominates or equal to 'sub'.
bool dominates(Node* sub, Node_List &nlist);
protected:
bool remove_dead_region(PhaseGVN *phase, bool can_reshape);
public:
// Idealize graph, using DU info. Done after constant propagation
virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp );
// See if there is valid pipeline info
static const Pipeline *pipeline_class();
virtual const Pipeline *pipeline() const;
// Compute the latency from the def to this instruction of the ith input node
uint latency(uint i);
// Hash & compare functions, for pessimistic value numbering
// If the hash function returns the special sentinel value NO_HASH,
// the node is guaranteed never to compare equal to any other node.
// If we accidentally generate a hash with value NO_HASH the node
// won't go into the table and we'll lose a little optimization.
enum { NO_HASH = 0 };
virtual uint hash() const;
virtual uint cmp( const Node &n ) const;
// Operation appears to be iteratively computed (such as an induction variable)
// It is possible for this operation to return false for a loop-varying
// value, if it appears (by local graph inspection) to be computed by a simple conditional.
bool is_iteratively_computed();
// Determine if a node is Counted loop induction variable.
// The method is defined in loopnode.cpp.
const Node* is_loop_iv() const;
// Return a node with opcode "opc" and same inputs as "this" if one can
// be found; Otherwise return NULL;
Node* find_similar(int opc);
// Return the unique control out if only one. Null if none or more than one.
Node* unique_ctrl_out();
//----------------- Code Generation
// Ideal register class for Matching. Zero means unmatched instruction
// (these are cloned instead of converted to machine nodes).
virtual uint ideal_reg() const;
static const uint NotAMachineReg; // must be > max. machine register
// Do we Match on this edge index or not? Generally false for Control
// and true for everything else. Weird for calls & returns.
virtual uint match_edge(uint idx) const;
// Register class output is returned in
virtual const RegMask &out_RegMask() const;
// Register class input is expected in
virtual const RegMask &in_RegMask(uint) const;
// Should we clone rather than spill this instruction?
bool rematerialize() const;
// Return JVM State Object if this Node carries debug info, or NULL otherwise
virtual JVMState* jvms() const;
// Print as assembly
virtual void format( PhaseRegAlloc *, outputStream* st = tty ) const;
// Emit bytes starting at parameter 'ptr'
// Bump 'ptr' by the number of output bytes
virtual void emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const;
// Size of instruction in bytes
virtual uint size(PhaseRegAlloc *ra_) const;
// Convenience function to extract an integer constant from a node.
// If it is not an integer constant (either Con, CastII, or Mach),
// return value_if_unknown.
jint find_int_con(jint value_if_unknown) const {
const TypeInt* t = find_int_type();
return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown;
}
// Return the constant, knowing it is an integer constant already
jint get_int() const {
const TypeInt* t = find_int_type();
guarantee(t != NULL, "must be con");
return t->get_con();
}
// Here's where the work is done. Can produce non-constant int types too.
const TypeInt* find_int_type() const;
// Same thing for long (and intptr_t, via type.hpp):
jlong get_long() const {
const TypeLong* t = find_long_type();
guarantee(t != NULL, "must be con");
return t->get_con();
}
jlong find_long_con(jint value_if_unknown) const {
const TypeLong* t = find_long_type();
return (t != NULL && t->is_con()) ? t->get_con() : value_if_unknown;
}
const TypeLong* find_long_type() const;
const TypePtr* get_ptr_type() const;
// These guys are called by code generated by ADLC:
intptr_t get_ptr() const;
intptr_t get_narrowcon() const;
jdouble getd() const;
jfloat getf() const;
// Nodes which are pinned into basic blocks
virtual bool pinned() const { return false; }
// Nodes which use memory without consuming it, hence need antidependences
// More specifically, needs_anti_dependence_check returns true iff the node
// (a) does a load, and (b) does not perform a store (except perhaps to a
// stack slot or some other unaliased location).
bool needs_anti_dependence_check() const;
// Return which operand this instruction may cisc-spill. In other words,
// return operand position that can convert from reg to memory access
virtual int cisc_operand() const { return AdlcVMDeps::Not_cisc_spillable; }
bool is_cisc_alternate() const { return (_flags & Flag_is_cisc_alternate) != 0; }
//----------------- Graph walking
public:
// Walk and apply member functions recursively.
// Supplied (this) pointer is root.
void walk(NFunc pre, NFunc post, void *env);
static void nop(Node &, void*); // Dummy empty function
static void packregion( Node &n, void* );
private:
void walk_(NFunc pre, NFunc post, void *env, VectorSet &visited);
//----------------- Printing, etc
public:
#ifndef PRODUCT
Node* find(int idx) const; // Search the graph for the given idx.
Node* find_ctrl(int idx) const; // Search control ancestors for the given idx.
void dump() const { dump("\n"); } // Print this node.
void dump(const char* suffix, outputStream *st = tty) const;// Print this node.
void dump(int depth) const; // Print this node, recursively to depth d
void dump_ctrl(int depth) const; // Print control nodes, to depth d
virtual void dump_req(outputStream *st = tty) const; // Print required-edge info
virtual void dump_prec(outputStream *st = tty) const; // Print precedence-edge info
virtual void dump_out(outputStream *st = tty) const; // Print the output edge info
virtual void dump_spec(outputStream *st) const {}; // Print per-node info
void verify_edges(Unique_Node_List &visited); // Verify bi-directional edges
void verify() const; // Check Def-Use info for my subgraph
static void verify_recur(const Node *n, int verify_depth, VectorSet &old_space, VectorSet &new_space);
// This call defines a class-unique string used to identify class instances
virtual const char *Name() const;
void dump_format(PhaseRegAlloc *ra) const; // debug access to MachNode::format(...)
// RegMask Print Functions
void dump_in_regmask(int idx) { in_RegMask(idx).dump(); }
void dump_out_regmask() { out_RegMask().dump(); }
static bool in_dump() { return Compile::current()->_in_dump_cnt > 0; }
void fast_dump() const {
tty->print("%4d: %-17s", _idx, Name());
for (uint i = 0; i < len(); i++)
if (in(i))
tty->print(" %4d", in(i)->_idx);
else
tty->print(" NULL");
tty->print("\n");
}
#endif
#ifdef ASSERT
void verify_construction();
bool verify_jvms(const JVMState* jvms) const;
int _debug_idx; // Unique value assigned to every node.
int debug_idx() const { return _debug_idx; }
void set_debug_idx( int debug_idx ) { _debug_idx = debug_idx; }
Node* _debug_orig; // Original version of this, if any.
Node* debug_orig() const { return _debug_orig; }
void set_debug_orig(Node* orig); // _debug_orig = orig
int _hash_lock; // Barrier to modifications of nodes in the hash table
void enter_hash_lock() { ++_hash_lock; assert(_hash_lock < 99, "in too many hash tables?"); }
void exit_hash_lock() { --_hash_lock; assert(_hash_lock >= 0, "mispaired hash locks"); }
static void init_NodeProperty();
#if OPTO_DU_ITERATOR_ASSERT
const Node* _last_del; // The last deleted node.
uint _del_tick; // Bumped when a deletion happens..
#endif
#endif
};
//-----------------------------------------------------------------------------
// Iterators over DU info, and associated Node functions.
#if OPTO_DU_ITERATOR_ASSERT
// Common code for assertion checking on DU iterators.
class DUIterator_Common VALUE_OBJ_CLASS_SPEC {
#ifdef ASSERT
protected:
bool _vdui; // cached value of VerifyDUIterators
const Node* _node; // the node containing the _out array
uint _outcnt; // cached node->_outcnt
uint _del_tick; // cached node->_del_tick
Node* _last; // last value produced by the iterator
void sample(const Node* node); // used by c'tor to set up for verifies
void verify(const Node* node, bool at_end_ok = false);
void verify_resync();
void reset(const DUIterator_Common& that);
// The VDUI_ONLY macro protects code conditionalized on VerifyDUIterators
#define I_VDUI_ONLY(i,x) { if ((i)._vdui) { x; } }
#else
#define I_VDUI_ONLY(i,x) { }
#endif //ASSERT
};
#define VDUI_ONLY(x) I_VDUI_ONLY(*this, x)
// Default DU iterator. Allows appends onto the out array.
// Allows deletion from the out array only at the current point.
// Usage:
// for (DUIterator i = x->outs(); x->has_out(i); i++) {
// Node* y = x->out(i);
// ...
// }
// Compiles in product mode to a unsigned integer index, which indexes
// onto a repeatedly reloaded base pointer of x->_out. The loop predicate
// also reloads x->_outcnt. If you delete, you must perform "--i" just
// before continuing the loop. You must delete only the last-produced
// edge. You must delete only a single copy of the last-produced edge,
// or else you must delete all copies at once (the first time the edge
// is produced by the iterator).
class DUIterator : public DUIterator_Common {
friend class Node;
// This is the index which provides the product-mode behavior.
// Whatever the product-mode version of the system does to the
// DUI index is done to this index. All other fields in
// this class are used only for assertion checking.
uint _idx;
#ifdef ASSERT
uint _refresh_tick; // Records the refresh activity.
void sample(const Node* node); // Initialize _refresh_tick etc.
void verify(const Node* node, bool at_end_ok = false);
void verify_increment(); // Verify an increment operation.
void verify_resync(); // Verify that we can back up over a deletion.
void verify_finish(); // Verify that the loop terminated properly.
void refresh(); // Resample verification info.
void reset(const DUIterator& that); // Resample after assignment.
#endif
DUIterator(const Node* node, int dummy_to_avoid_conversion)
{ _idx = 0; debug_only(sample(node)); }
public:
// initialize to garbage; clear _vdui to disable asserts
DUIterator()
{ /*initialize to garbage*/ debug_only(_vdui = false); }
void operator++(int dummy_to_specify_postfix_op)
{ _idx++; VDUI_ONLY(verify_increment()); }
void operator--()
{ VDUI_ONLY(verify_resync()); --_idx; }
~DUIterator()
{ VDUI_ONLY(verify_finish()); }
void operator=(const DUIterator& that)
{ _idx = that._idx; debug_only(reset(that)); }
};
DUIterator Node::outs() const
{ return DUIterator(this, 0); }
DUIterator& Node::refresh_out_pos(DUIterator& i) const
{ I_VDUI_ONLY(i, i.refresh()); return i; }
bool Node::has_out(DUIterator& i) const
{ I_VDUI_ONLY(i, i.verify(this,true));return i._idx < _outcnt; }
Node* Node::out(DUIterator& i) const
{ I_VDUI_ONLY(i, i.verify(this)); return debug_only(i._last=) _out[i._idx]; }
// Faster DU iterator. Disallows insertions into the out array.
// Allows deletion from the out array only at the current point.
// Usage:
// for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
// Node* y = x->fast_out(i);
// ...
// }
// Compiles in product mode to raw Node** pointer arithmetic, with
// no reloading of pointers from the original node x. If you delete,
// you must perform "--i; --imax" just before continuing the loop.
// If you delete multiple copies of the same edge, you must decrement
// imax, but not i, multiple times: "--i, imax -= num_edges".
class DUIterator_Fast : public DUIterator_Common {
friend class Node;
friend class DUIterator_Last;
// This is the pointer which provides the product-mode behavior.
// Whatever the product-mode version of the system does to the
// DUI pointer is done to this pointer. All other fields in
// this class are used only for assertion checking.
Node** _outp;
#ifdef ASSERT
void verify(const Node* node, bool at_end_ok = false);
void verify_limit();
void verify_resync();
void verify_relimit(uint n);
void reset(const DUIterator_Fast& that);
#endif
// Note: offset must be signed, since -1 is sometimes passed
DUIterator_Fast(const Node* node, ptrdiff_t offset)
{ _outp = node->_out + offset; debug_only(sample(node)); }
public:
// initialize to garbage; clear _vdui to disable asserts
DUIterator_Fast()
{ /*initialize to garbage*/ debug_only(_vdui = false); }
void operator++(int dummy_to_specify_postfix_op)
{ _outp++; VDUI_ONLY(verify(_node, true)); }
void operator--()
{ VDUI_ONLY(verify_resync()); --_outp; }
void operator-=(uint n) // applied to the limit only
{ _outp -= n; VDUI_ONLY(verify_relimit(n)); }
bool operator<(DUIterator_Fast& limit) {
I_VDUI_ONLY(*this, this->verify(_node, true));
I_VDUI_ONLY(limit, limit.verify_limit());
return _outp < limit._outp;
}
void operator=(const DUIterator_Fast& that)
{ _outp = that._outp; debug_only(reset(that)); }
};
DUIterator_Fast Node::fast_outs(DUIterator_Fast& imax) const {
// Assign a limit pointer to the reference argument:
imax = DUIterator_Fast(this, (ptrdiff_t)_outcnt);
// Return the base pointer:
return DUIterator_Fast(this, 0);
}
Node* Node::fast_out(DUIterator_Fast& i) const {
I_VDUI_ONLY(i, i.verify(this));
return debug_only(i._last=) *i._outp;
}
// Faster DU iterator. Requires each successive edge to be removed.
// Does not allow insertion of any edges.
// Usage:
// for (DUIterator_Last imin, i = x->last_outs(imin); i >= imin; i -= num_edges) {
// Node* y = x->last_out(i);
// ...
// }
// Compiles in product mode to raw Node** pointer arithmetic, with
// no reloading of pointers from the original node x.
class DUIterator_Last : private DUIterator_Fast {
friend class Node;
#ifdef ASSERT
void verify(const Node* node, bool at_end_ok = false);
void verify_limit();
void verify_step(uint num_edges);
#endif
// Note: offset must be signed, since -1 is sometimes passed
DUIterator_Last(const Node* node, ptrdiff_t offset)
: DUIterator_Fast(node, offset) { }
void operator++(int dummy_to_specify_postfix_op) {} // do not use
void operator<(int) {} // do not use
public:
DUIterator_Last() { }
// initialize to garbage
void operator--()
{ _outp--; VDUI_ONLY(verify_step(1)); }
void operator-=(uint n)
{ _outp -= n; VDUI_ONLY(verify_step(n)); }
bool operator>=(DUIterator_Last& limit) {
I_VDUI_ONLY(*this, this->verify(_node, true));
I_VDUI_ONLY(limit, limit.verify_limit());
return _outp >= limit._outp;
}
void operator=(const DUIterator_Last& that)
{ DUIterator_Fast::operator=(that); }
};
DUIterator_Last Node::last_outs(DUIterator_Last& imin) const {
// Assign a limit pointer to the reference argument:
imin = DUIterator_Last(this, 0);
// Return the initial pointer:
return DUIterator_Last(this, (ptrdiff_t)_outcnt - 1);
}
Node* Node::last_out(DUIterator_Last& i) const {
I_VDUI_ONLY(i, i.verify(this));
return debug_only(i._last=) *i._outp;
}
#endif //OPTO_DU_ITERATOR_ASSERT
#undef I_VDUI_ONLY
#undef VDUI_ONLY
// An Iterator that truly follows the iterator pattern. Doesn't
// support deletion but could be made to.
//
// for (SimpleDUIterator i(n); i.has_next(); i.next()) {
// Node* m = i.get();
//
class SimpleDUIterator : public StackObj {
private:
Node* node;
DUIterator_Fast i;
DUIterator_Fast imax;
public:
SimpleDUIterator(Node* n): node(n), i(n->fast_outs(imax)) {}
bool has_next() { return i < imax; }
void next() { i++; }
Node* get() { return node->fast_out(i); }
};
//-----------------------------------------------------------------------------
// Map dense integer indices to Nodes. Uses classic doubling-array trick.
// Abstractly provides an infinite array of Node*'s, initialized to NULL.
// Note that the constructor just zeros things, and since I use Arena
// allocation I do not need a destructor to reclaim storage.
class Node_Array : public ResourceObj {
friend class VMStructs;
protected:
Arena *_a; // Arena to allocate in
uint _max;
Node **_nodes;
void grow( uint i ); // Grow array node to fit
public:
Node_Array(Arena *a) : _a(a), _max(OptoNodeListSize) {
_nodes = NEW_ARENA_ARRAY( a, Node *, OptoNodeListSize );
for( int i = 0; i < OptoNodeListSize; i++ ) {
_nodes[i] = NULL;
}
}
Node_Array(Node_Array *na) : _a(na->_a), _max(na->_max), _nodes(na->_nodes) {}
Node *operator[] ( uint i ) const // Lookup, or NULL for not mapped
{ return (i<_max) ? _nodes[i] : (Node*)NULL; }
Node *at( uint i ) const { assert(i<_max,"oob"); return _nodes[i]; }
Node **adr() { return _nodes; }
// Extend the mapping: index i maps to Node *n.
void map( uint i, Node *n ) { if( i>=_max ) grow(i); _nodes[i] = n; }
void insert( uint i, Node *n );
void remove( uint i ); // Remove, preserving order
void sort( C_sort_func_t func);
void reset( Arena *new_a ); // Zap mapping to empty; reclaim storage
void clear(); // Set all entries to NULL, keep storage
uint Size() const { return _max; }
void dump() const;
};
class Node_List : public Node_Array {
friend class VMStructs;
uint _cnt;
public:
Node_List() : Node_Array(Thread::current()->resource_area()), _cnt(0) {}
Node_List(Arena *a) : Node_Array(a), _cnt(0) {}
bool contains(const Node* n) const {
for (uint e = 0; e < size(); e++) {
if (at(e) == n) return true;
}
return false;
}
void insert( uint i, Node *n ) { Node_Array::insert(i,n); _cnt++; }
void remove( uint i ) { Node_Array::remove(i); _cnt--; }
void push( Node *b ) { map(_cnt++,b); }
void yank( Node *n ); // Find and remove
Node *pop() { return _nodes[--_cnt]; }
Node *rpop() { Node *b = _nodes[0]; _nodes[0]=_nodes[--_cnt]; return b;}
void clear() { _cnt = 0; Node_Array::clear(); } // retain storage
uint size() const { return _cnt; }
void dump() const;
void dump_simple() const;
};
//------------------------------Unique_Node_List-------------------------------
class Unique_Node_List : public Node_List {
friend class VMStructs;
VectorSet _in_worklist;
uint _clock_index; // Index in list where to pop from next
public:
Unique_Node_List() : Node_List(), _in_worklist(Thread::current()->resource_area()), _clock_index(0) {}
Unique_Node_List(Arena *a) : Node_List(a), _in_worklist(a), _clock_index(0) {}
void remove( Node *n );
bool member( Node *n ) { return _in_worklist.test(n->_idx) != 0; }
VectorSet &member_set(){ return _in_worklist; }
void push( Node *b ) {
if( !_in_worklist.test_set(b->_idx) )
Node_List::push(b);
}
Node *pop() {
if( _clock_index >= size() ) _clock_index = 0;
Node *b = at(_clock_index);
map( _clock_index, Node_List::pop());
if (size() != 0) _clock_index++; // Always start from 0
_in_worklist >>= b->_idx;
return b;
}
Node *remove( uint i ) {
Node *b = Node_List::at(i);
_in_worklist >>= b->_idx;
map(i,Node_List::pop());
return b;
}
void yank( Node *n ) { _in_worklist >>= n->_idx; Node_List::yank(n); }
void clear() {
_in_worklist.Clear(); // Discards storage but grows automatically
Node_List::clear();
_clock_index = 0;
}
// Used after parsing to remove useless nodes before Iterative GVN
void remove_useless_nodes(VectorSet &useful);
#ifndef PRODUCT
void print_set() const { _in_worklist.print(); }
#endif
};
// Inline definition of Compile::record_for_igvn must be deferred to this point.
inline void Compile::record_for_igvn(Node* n) {
_for_igvn->push(n);
}
//------------------------------Node_Stack-------------------------------------
class Node_Stack {
friend class VMStructs;
protected:
struct INode {
Node *node; // Processed node
uint indx; // Index of next node's child
};
INode *_inode_top; // tos, stack grows up
INode *_inode_max; // End of _inodes == _inodes + _max
INode *_inodes; // Array storage for the stack
Arena *_a; // Arena to allocate in
void grow();
public:
Node_Stack(int size) {
size_t max = (size > OptoNodeListSize) ? size : OptoNodeListSize;
_a = Thread::current()->resource_area();
_inodes = NEW_ARENA_ARRAY( _a, INode, max );
_inode_max = _inodes + max;
_inode_top = _inodes - 1; // stack is empty
}
Node_Stack(Arena *a, int size) : _a(a) {
size_t max = (size > OptoNodeListSize) ? size : OptoNodeListSize;
_inodes = NEW_ARENA_ARRAY( _a, INode, max );
_inode_max = _inodes + max;
_inode_top = _inodes - 1; // stack is empty
}
void pop() {
assert(_inode_top >= _inodes, "node stack underflow");
--_inode_top;
}
void push(Node *n, uint i) {
++_inode_top;
if (_inode_top >= _inode_max) grow();
INode *top = _inode_top; // optimization
top->node = n;
top->indx = i;
}
Node *node() const {
return _inode_top->node;
}
Node* node_at(uint i) const {
assert(_inodes + i <= _inode_top, "in range");
return _inodes[i].node;
}
uint index() const {
return _inode_top->indx;
}
uint index_at(uint i) const {
assert(_inodes + i <= _inode_top, "in range");
return _inodes[i].indx;
}
void set_node(Node *n) {
_inode_top->node = n;
}
void set_index(uint i) {
_inode_top->indx = i;
}
uint size_max() const { return (uint)pointer_delta(_inode_max, _inodes, sizeof(INode)); } // Max size
uint size() const { return (uint)pointer_delta((_inode_top+1), _inodes, sizeof(INode)); } // Current size
bool is_nonempty() const { return (_inode_top >= _inodes); }
bool is_empty() const { return (_inode_top < _inodes); }
void clear() { _inode_top = _inodes - 1; } // retain storage
// Node_Stack is used to map nodes.
Node* find(uint idx) const;
};
//-----------------------------Node_Notes--------------------------------------
// Debugging or profiling annotations loosely and sparsely associated
// with some nodes. See Compile::node_notes_at for the accessor.
class Node_Notes VALUE_OBJ_CLASS_SPEC {
friend class VMStructs;
JVMState* _jvms;
public:
Node_Notes(JVMState* jvms = NULL) {
_jvms = jvms;
}
JVMState* jvms() { return _jvms; }
void set_jvms(JVMState* x) { _jvms = x; }
// True if there is nothing here.
bool is_clear() {
return (_jvms == NULL);
}
// Make there be nothing here.
void clear() {
_jvms = NULL;
}
// Make a new, clean node notes.
static Node_Notes* make(Compile* C) {
Node_Notes* nn = NEW_ARENA_ARRAY(C->comp_arena(), Node_Notes, 1);
nn->clear();
return nn;
}
Node_Notes* clone(Compile* C) {
Node_Notes* nn = NEW_ARENA_ARRAY(C->comp_arena(), Node_Notes, 1);
(*nn) = (*this);
return nn;
}
// Absorb any information from source.
bool update_from(Node_Notes* source) {
bool changed = false;
if (source != NULL) {
if (source->jvms() != NULL) {
set_jvms(source->jvms());
changed = true;
}
}
return changed;
}
};
// Inlined accessors for Compile::node_nodes that require the preceding class:
inline Node_Notes*
Compile::locate_node_notes(GrowableArray<Node_Notes*>* arr,
int idx, bool can_grow) {
assert(idx >= 0, "oob");
int block_idx = (idx >> _log2_node_notes_block_size);
int grow_by = (block_idx - (arr == NULL? 0: arr->length()));
if (grow_by >= 0) {
if (!can_grow) return NULL;
grow_node_notes(arr, grow_by + 1);
}
// (Every element of arr is a sub-array of length _node_notes_block_size.)
return arr->at(block_idx) + (idx & (_node_notes_block_size-1));
}
inline bool
Compile::set_node_notes_at(int idx, Node_Notes* value) {
if (value == NULL || value->is_clear())
return false; // nothing to write => write nothing
Node_Notes* loc = locate_node_notes(_node_note_array, idx, true);
assert(loc != NULL, "");
return loc->update_from(value);
}
//------------------------------TypeNode---------------------------------------
// Node with a Type constant.
class TypeNode : public Node {
protected:
virtual uint hash() const; // Check the type
virtual uint cmp( const Node &n ) const;
virtual uint size_of() const; // Size is bigger
const Type* const _type;
public:
void set_type(const Type* t) {
assert(t != NULL, "sanity");
debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH);
*(const Type**)&_type = t; // cast away const-ness
// If this node is in the hash table, make sure it doesn't need a rehash.
assert(check_hash == NO_HASH || check_hash == hash(), "type change must preserve hash code");
}
const Type* type() const { assert(_type != NULL, "sanity"); return _type; };
TypeNode( const Type *t, uint required ) : Node(required), _type(t) {
init_class_id(Class_Type);
}
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const;
virtual uint ideal_reg() const;
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const;
#endif
};
#endif // SHARE_VM_OPTO_NODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/opcodes.cpp
/*
* Copyright (c) 1998, 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.
*
*/
// no precompiled headers
// ----------------------------------------------------------------------------
// Build a table of class names as strings. Used both for debugging printouts
// and in the ADL machine descriptions.
#define macro(x) #x,
const char *NodeClassNames[] = {
"Node",
"Set",
"RegN",
"RegI",
"RegP",
"RegF",
"RegD",
"RegL",
"RegFlags",
"VecS",
"VecD",
"VecX",
"VecY",
"_last_machine_leaf",
#include "classes.hpp"
"_last_class_name",
};
#undef macro
C:\hotspot-69087d08d473\src\share\vm/opto/opcodes.hpp
/*
* Copyright (c) 1997, 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.
*
*/
#ifndef SHARE_VM_OPTO_OPCODES_HPP
#define SHARE_VM_OPTO_OPCODES_HPP
// Build a big enum of class names to give them dense integer indices
#define macro(x) Op_##x,
enum Opcodes {
Op_Node = 0,
macro(Set) // Instruction selection match rule
macro(RegN) // Machine narrow oop register
macro(RegI) // Machine integer register
macro(RegP) // Machine pointer register
macro(RegF) // Machine float register
macro(RegD) // Machine double register
macro(RegL) // Machine long register
macro(VecS) // Machine vectors register
macro(VecD) // Machine vectord register
macro(VecX) // Machine vectorx register
macro(VecY) // Machine vectory register
macro(RegFlags) // Machine flags register
_last_machine_leaf, // Split between regular opcodes and machine
#include "classes.hpp"
_last_opcode
};
#undef macro
// Table of names, indexed by Opcode
extern const char *NodeClassNames[];
#endif // SHARE_VM_OPTO_OPCODES_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/optoreg.hpp
/*
* Copyright (c) 2006, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_OPTOREG_HPP
#define SHARE_VM_OPTO_OPTOREG_HPP
//------------------------------OptoReg----------------------------------------
// We eventually need Registers for the Real World. Registers are essentially
// non-SSA names. A Register is represented as a number. Non-regular values
// (e.g., Control, Memory, I/O) use the Special register. The actual machine
// registers (as described in the ADL file for a machine) start at zero.
// Stack-slots (spill locations) start at the nest Chunk past the last machine
// register.
//
// Note that stack spill-slots are treated as a very large register set.
// They have all the correct properties for a Register: not aliased (unique
// named). There is some simple mapping from a stack-slot register number
// to the actual location on the stack; this mapping depends on the calling
// conventions and is described in the ADL.
//
// Note that Name is not enum. C++ standard defines that the range of enum
// is the range of smallest bit-field that can represent all enumerators
// declared in the enum. The result of assigning a value to enum is undefined
// if the value is outside the enumeration's valid range. OptoReg::Name is
// typedef'ed as int, because it needs to be able to represent spill-slots.
//
class OptoReg VALUE_OBJ_CLASS_SPEC {
friend class C2Compiler;
public:
typedef int Name;
enum {
// Chunk 0
Physical = AdlcVMDeps::Physical, // Start of physical regs
// A few oddballs at the edge of the world
Special = -2, // All special (not allocated) values
Bad = -1 // Not a register
};
private:
static const VMReg opto2vm[REG_COUNT];
static Name vm2opto[ConcreteRegisterImpl::number_of_registers];
public:
// Stack pointer register
static OptoReg::Name c_frame_pointer;
// Increment a register number. As in:
// "for ( OptoReg::Name i; i=Control; i = add(i,1) ) ..."
static Name add( Name x, int y ) { return Name(x+y); }
// (We would like to have an operator+ for RegName, but it is not
// a class, so this would be illegal in C++.)
static void dump(int, outputStream *st = tty);
// Get the stack slot number of an OptoReg::Name
static unsigned int reg2stack( OptoReg::Name r) {
assert( r >= stack0(), " must be");
return r - stack0();
}
// convert a stack slot number into an OptoReg::Name
static OptoReg::Name stack2reg( int idx) {
return Name(stack0() + idx);
}
static bool is_stack(Name n) {
return n >= stack0();
}
static bool is_valid(Name n) {
return (n != Bad);
}
static bool is_reg(Name n) {
return is_valid(n) && !is_stack(n);
}
static VMReg as_VMReg(OptoReg::Name n) {
if (is_reg(n)) {
// Must use table, it'd be nice if Bad was indexable...
return opto2vm[n];
} else {
assert(!is_stack(n), "must un warp");
return VMRegImpl::Bad();
}
}
// Can un-warp a stack slot or convert a register or Bad
static VMReg as_VMReg(OptoReg::Name n, int frame_size, int arg_count) {
if (is_reg(n)) {
// Must use table, it'd be nice if Bad was indexable...
return opto2vm[n];
} else if (is_stack(n)) {
int stack_slot = reg2stack(n);
if (stack_slot < arg_count) {
return VMRegImpl::stack2reg(stack_slot + frame_size);
}
return VMRegImpl::stack2reg(stack_slot - arg_count);
// return return VMRegImpl::stack2reg(reg2stack(OptoReg::add(n, -arg_count)));
} else {
return VMRegImpl::Bad();
}
}
static OptoReg::Name as_OptoReg(VMReg r) {
if (r->is_stack()) {
assert(false, "must warp");
return stack2reg(r->reg2stack());
} else if (r->is_valid()) {
// Must use table, it'd be nice if Bad was indexable...
return vm2opto[r->value()];
} else {
return Bad;
}
}
static OptoReg::Name stack0() {
return VMRegImpl::stack0->value();
}
static const char* regname(OptoReg::Name n) {
return as_VMReg(n)->name();
}
};
//---------------------------OptoRegPair-------------------------------------------
// Pairs of 32-bit registers for the allocator.
// This is a very similar class to VMRegPair. C2 only interfaces with VMRegPair
// via the calling convention code which is shared between the compilers.
// Since C2 uses OptoRegs for register allocation it is more efficient to use
// VMRegPair internally for nodes that can contain a pair of OptoRegs rather
// than use VMRegPair and continually be converting back and forth. So normally
// C2 will take in a VMRegPair from the calling convention code and immediately
// convert them to an OptoRegPair and stay in the OptoReg world. The only over
// conversion between OptoRegs and VMRegs is for debug info and oopMaps. This
// is not a high bandwidth spot and so it is not an issue.
// Note that onde other consequence of staying in the OptoReg world with OptoRegPairs
// is that there are "physical" OptoRegs that are not representable in the VMReg
// world, notably flags. [ But by design there is "space" in the VMReg world
// for such registers they just may not be concrete ]. So if we were to use VMRegPair
// then the VMReg world would have to have a representation for these registers
// so that a OptoReg->VMReg->OptoReg would reproduce ther original OptoReg. As it
// stands if you convert a flag (condition code) to a VMReg you will get VMRegImpl::Bad
// and converting that will return OptoReg::Bad losing the identity of the OptoReg.
class OptoRegPair {
friend class VMStructs;
private:
short _second;
short _first;
public:
void set_bad ( ) { _second = OptoReg::Bad; _first = OptoReg::Bad; }
void set1 ( OptoReg::Name n ) { _second = OptoReg::Bad; _first = n; }
void set2 ( OptoReg::Name n ) { _second = n + 1; _first = n; }
void set_pair( OptoReg::Name second, OptoReg::Name first ) { _second= second; _first= first; }
void set_ptr ( OptoReg::Name ptr ) {
#ifdef _LP64
_second = ptr+1;
#else
_second = OptoReg::Bad;
#endif
_first = ptr;
}
OptoReg::Name second() const { return _second; }
OptoReg::Name first() const { return _first; }
OptoRegPair(OptoReg::Name second, OptoReg::Name first) { _second = second; _first = first; }
OptoRegPair(OptoReg::Name f) { _second = OptoReg::Bad; _first = f; }
OptoRegPair() { _second = OptoReg::Bad; _first = OptoReg::Bad; }
};
#endif // SHARE_VM_OPTO_OPTOREG_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/output.cpp
/*
* Copyright (c) 1998, 2018, 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 "asm/assembler.inline.hpp"
#include "code/compiledIC.hpp"
#include "code/debugInfo.hpp"
#include "code/debugInfoRec.hpp"
#include "compiler/compileBroker.hpp"
#include "compiler/disassembler.hpp"
#include "compiler/oopMap.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/locknode.hpp"
#include "opto/machnode.hpp"
#include "opto/output.hpp"
#include "opto/regalloc.hpp"
#include "opto/runtime.hpp"
#include "opto/subnode.hpp"
#include "opto/type.hpp"
#include "runtime/handles.inline.hpp"
#include "utilities/xmlstream.hpp"
#ifndef PRODUCT
#define DEBUG_ARG(x) , x
#else
#define DEBUG_ARG(x)
#endif
// Convert Nodes to instruction bits and pass off to the VM
void Compile::Output() {
// RootNode goes
assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
// The number of new nodes (mostly MachNop) is proportional to
// the number of java calls and inner loops which are aligned.
if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
C->inner_loops()*(OptoLoopAlignment-1)),
"out of nodes before code generation" ) ) {
return;
}
// Make sure I can find the Start Node
Block *entry = _cfg->get_block(1);
Block *broot = _cfg->get_root_block();
const StartNode *start = entry->head()->as_Start();
// Replace StartNode with prolog
MachPrologNode *prolog = new (this) MachPrologNode();
entry->map_node(prolog, 0);
_cfg->map_node_to_block(prolog, entry);
_cfg->unmap_node_from_block(start); // start is no longer in any block
// Virtual methods need an unverified entry point
if( is_osr_compilation() ) {
if( PoisonOSREntry ) {
// TODO: Should use a ShouldNotReachHereNode...
_cfg->insert( broot, 0, new (this) MachBreakpointNode() );
}
} else {
if( _method && !_method->flags().is_static() ) {
// Insert unvalidated entry point
_cfg->insert( broot, 0, new (this) MachUEPNode() );
}
}
// Break before main entry point
if( (_method && _method->break_at_execute())
#ifndef PRODUCT
||(OptoBreakpoint && is_method_compilation())
||(OptoBreakpointOSR && is_osr_compilation())
||(OptoBreakpointC2R && !_method)
#endif
) {
// checking for _method means that OptoBreakpoint does not apply to
// runtime stubs or frame converters
_cfg->insert( entry, 1, new (this) MachBreakpointNode() );
}
// Insert epilogs before every return
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* block = _cfg->get_block(i);
if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
Node* m = block->end();
if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
block->add_inst(epilog);
_cfg->map_node_to_block(epilog, block);
}
}
}
# ifdef ENABLE_ZAP_DEAD_LOCALS
if (ZapDeadCompiledLocals) {
Insert_zap_nodes();
}
# endif
uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
blk_starts[0] = 0;
// Initialize code buffer and process short branches.
CodeBuffer* cb = init_buffer(blk_starts);
if (cb == NULL || failing()) {
return;
}
ScheduleAndBundle();
#ifndef PRODUCT
if (trace_opto_output()) {
tty->print("\n---- After ScheduleAndBundle ----\n");
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
tty->print("\nBB#%03d:\n", i);
Block* block = _cfg->get_block(i);
for (uint j = 0; j < block->number_of_nodes(); j++) {
Node* n = block->get_node(j);
OptoReg::Name reg = _regalloc->get_reg_first(n);
tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
n->dump();
}
}
}
#endif
if (failing()) {
return;
}
BuildOopMaps();
if (failing()) {
return;
}
fill_buffer(cb, blk_starts);
}
bool Compile::need_stack_bang(int frame_size_in_bytes) const {
// Determine if we need to generate a stack overflow check.
// Do it if the method is not a stub function and
// has java calls or has frame size > vm_page_size/8.
// The debug VM checks that deoptimization doesn't trigger an
// unexpected stack overflow (compiled method stack banging should
// guarantee it doesn't happen) so we always need the stack bang in
// a debug VM.
return (UseStackBanging && stub_function() == NULL &&
(has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
DEBUG_ONLY(|| true)));
}
bool Compile::need_register_stack_bang() const {
// Determine if we need to generate a register stack overflow check.
// This is only used on architectures which have split register
// and memory stacks (ie. IA64).
// Bang if the method is not a stub function and has java calls
return (stub_function() == NULL && has_java_calls());
}
# ifdef ENABLE_ZAP_DEAD_LOCALS
// In order to catch compiler oop-map bugs, we have implemented
// a debugging mode called ZapDeadCompilerLocals.
// This mode causes the compiler to insert a call to a runtime routine,
// "zap_dead_locals", right before each place in compiled code
// that could potentially be a gc-point (i.e., a safepoint or oop map point).
// The runtime routine checks that locations mapped as oops are really
// oops, that locations mapped as values do not look like oops,
// and that locations mapped as dead are not used later
// (by zapping them to an invalid address).
int Compile::_CompiledZap_count = 0;
void Compile::Insert_zap_nodes() {
bool skip = false;
// Dink with static counts because code code without the extra
// runtime calls is MUCH faster for debugging purposes
if ( CompileZapFirst == 0 ) ; // nothing special
else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
else if ( CompileZapFirst == CompiledZap_count() )
warning("starting zap compilation after skipping");
if ( CompileZapLast == -1 ) ; // nothing special
else if ( CompileZapLast < CompiledZap_count() ) skip = true;
else if ( CompileZapLast == CompiledZap_count() )
warning("about to compile last zap");
++_CompiledZap_count; // counts skipped zaps, too
if ( skip ) return;
if ( _method == NULL )
return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
// Insert call to zap runtime stub before every node with an oop map
for( uint i=0; i<_cfg->number_of_blocks(); i++ ) {
Block *b = _cfg->get_block(i);
for ( uint j = 0; j < b->number_of_nodes(); ++j ) {
Node *n = b->get_node(j);
// Determining if we should insert a zap-a-lot node in output.
// We do that for all nodes that has oopmap info, except for calls
// to allocation. Calls to allocation passes in the old top-of-eden pointer
// and expect the C code to reset it. Hence, there can be no safepoints between
// the inlined-allocation and the call to new_Java, etc.
// We also cannot zap monitor calls, as they must hold the microlock
// during the call to Zap, which also wants to grab the microlock.
bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
if ( insert ) { // it is MachSafePoint
if ( !n->is_MachCall() ) {
insert = false;
} else if ( n->is_MachCall() ) {
MachCallNode* call = n->as_MachCall();
if (call->entry_point() == OptoRuntime::new_instance_Java() ||
call->entry_point() == OptoRuntime::new_array_Java() ||
call->entry_point() == OptoRuntime::multianewarray2_Java() ||
call->entry_point() == OptoRuntime::multianewarray3_Java() ||
call->entry_point() == OptoRuntime::multianewarray4_Java() ||
call->entry_point() == OptoRuntime::multianewarray5_Java() ||
call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
) {
insert = false;
}
}
if (insert) {
Node *zap = call_zap_node(n->as_MachSafePoint(), i);
b->insert_node(zap, j);
_cfg->map_node_to_block(zap, b);
++j;
}
}
}
}
}
Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
CallStaticJavaNode* ideal_node =
new (this) CallStaticJavaNode( tf,
OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
"call zap dead locals stub", 0, TypePtr::BOTTOM);
// We need to copy the OopMap from the site we're zapping at.
// We have to make a copy, because the zap site might not be
// a call site, and zap_dead is a call site.
OopMap* clone = node_to_check->oop_map()->deep_copy();
// Add the cloned OopMap to the zap node
ideal_node->set_oop_map(clone);
return _matcher->match_sfpt(ideal_node);
}
bool Compile::is_node_getting_a_safepoint( Node* n) {
// This code duplicates the logic prior to the call of add_safepoint
// below in this file.
if( n->is_MachSafePoint() ) return true;
return false;
}
# endif // ENABLE_ZAP_DEAD_LOCALS
// Compute the size of first NumberOfLoopInstrToAlign instructions at the top
// of a loop. When aligning a loop we need to provide enough instructions
// in cpu's fetch buffer to feed decoders. The loop alignment could be
// avoided if we have enough instructions in fetch buffer at the head of a loop.
// By default, the size is set to 999999 by Block's constructor so that
// a loop will be aligned if the size is not reset here.
//
// Note: Mach instructions could contain several HW instructions
// so the size is estimated only.
//
void Compile::compute_loop_first_inst_sizes() {
// The next condition is used to gate the loop alignment optimization.
// Don't aligned a loop if there are enough instructions at the head of a loop
// or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
// is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
// equal to 11 bytes which is the largest address NOP instruction.
if (MaxLoopPad < OptoLoopAlignment - 1) {
uint last_block = _cfg->number_of_blocks() - 1;
for (uint i = 1; i <= last_block; i++) {
Block* block = _cfg->get_block(i);
// Check the first loop's block which requires an alignment.
if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
uint sum_size = 0;
uint inst_cnt = NumberOfLoopInstrToAlign;
inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
// Check subsequent fallthrough blocks if the loop's first
// block(s) does not have enough instructions.
Block *nb = block;
while(inst_cnt > 0 &&
i < last_block &&
!_cfg->get_block(i + 1)->has_loop_alignment() &&
!nb->has_successor(block)) {
i++;
nb = _cfg->get_block(i);
inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
} // while( inst_cnt > 0 && i < last_block )
block->set_first_inst_size(sum_size);
} // f( b->head()->is_Loop() )
} // for( i <= last_block )
} // if( MaxLoopPad < OptoLoopAlignment-1 )
}
// The architecture description provides short branch variants for some long
// branch instructions. Replace eligible long branches with short branches.
void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
// Compute size of each block, method size, and relocation information size
uint nblocks = _cfg->number_of_blocks();
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
// Collect worst case block paddings
int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
memset(block_worst_case_pad, 0, nblocks * sizeof(int));
DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
bool has_short_branch_candidate = false;
// Initialize the sizes to 0
code_size = 0; // Size in bytes of generated code
stub_size = 0; // Size in bytes of all stub entries
// Size in bytes of all relocation entries, including those in local stubs.
// Start with 2-bytes of reloc info for the unvalidated entry point
reloc_size = 1; // Number of relocation entries
// Make three passes. The first computes pessimistic blk_starts,
// relative jmp_offset and reloc_size information. The second performs
// short branch substitution using the pessimistic sizing. The
// third inserts nops where needed.
// Step one, perform a pessimistic sizing pass.
uint last_call_adr = max_uint;
uint last_avoid_back_to_back_adr = max_uint;
uint nop_size = (new (this) MachNopNode())->size(_regalloc);
for (uint i = 0; i < nblocks; i++) { // For all blocks
Block* block = _cfg->get_block(i);
// During short branch replacement, we store the relative (to blk_starts)
// offset of jump in jmp_offset, rather than the absolute offset of jump.
// This is so that we do not need to recompute sizes of all nodes when
// we compute correct blk_starts in our next sizing pass.
jmp_offset[i] = 0;
jmp_size[i] = 0;
jmp_nidx[i] = -1;
DEBUG_ONLY( jmp_target[i] = 0; )
DEBUG_ONLY( jmp_rule[i] = 0; )
// Sum all instruction sizes to compute block size
uint last_inst = block->number_of_nodes();
uint blk_size = 0;
for (uint j = 0; j < last_inst; j++) {
Node* nj = block->get_node(j);
// Handle machine instruction nodes
if (nj->is_Mach()) {
MachNode *mach = nj->as_Mach();
blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
reloc_size += mach->reloc();
if (mach->is_MachCall()) {
// add size information for trampoline stub
// class CallStubImpl is platform-specific and defined in the *.ad files.
stub_size += CallStubImpl::size_call_trampoline();
reloc_size += CallStubImpl::reloc_call_trampoline();
MachCallNode *mcall = mach->as_MachCall();
// This destination address is NOT PC-relative
mcall->method_set((intptr_t)mcall->entry_point());
if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
stub_size += CompiledStaticCall::to_interp_stub_size();
reloc_size += CompiledStaticCall::reloc_to_interp_stub();
}
} else if (mach->is_MachSafePoint()) {
// If call/safepoint are adjacent, account for possible
// nop to disambiguate the two safepoints.
// ScheduleAndBundle() can rearrange nodes in a block,
// check for all offsets inside this block.
if (last_call_adr >= blk_starts[i]) {
blk_size += nop_size;
}
}
if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
// Nop is inserted between "avoid back to back" instructions.
// ScheduleAndBundle() can rearrange nodes in a block,
// check for all offsets inside this block.
if (last_avoid_back_to_back_adr >= blk_starts[i]) {
blk_size += nop_size;
}
}
if (mach->may_be_short_branch()) {
if (!nj->is_MachBranch()) {
#ifndef PRODUCT
nj->dump(3);
#endif
Unimplemented();
}
assert(jmp_nidx[i] == -1, "block should have only one branch");
jmp_offset[i] = blk_size;
jmp_size[i] = nj->size(_regalloc);
jmp_nidx[i] = j;
has_short_branch_candidate = true;
}
}
blk_size += nj->size(_regalloc);
// Remember end of call offset
if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
last_call_adr = blk_starts[i]+blk_size;
}
// Remember end of avoid_back_to_back offset
if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
}
}
// When the next block starts a loop, we may insert pad NOP
// instructions. Since we cannot know our future alignment,
// assume the worst.
if (i < nblocks - 1) {
Block* nb = _cfg->get_block(i + 1);
int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
if (max_loop_pad > 0) {
assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
// Adjust last_call_adr and/or last_avoid_back_to_back_adr.
// If either is the last instruction in this block, bump by
// max_loop_pad in lock-step with blk_size, so sizing
// calculations in subsequent blocks still can conservatively
// detect that it may the last instruction in this block.
if (last_call_adr == blk_starts[i]+blk_size) {
last_call_adr += max_loop_pad;
}
if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
last_avoid_back_to_back_adr += max_loop_pad;
}
blk_size += max_loop_pad;
block_worst_case_pad[i + 1] = max_loop_pad;
}
}
// Save block size; update total method size
blk_starts[i+1] = blk_starts[i]+blk_size;
}
// Step two, replace eligible long jumps.
bool progress = true;
uint last_may_be_short_branch_adr = max_uint;
while (has_short_branch_candidate && progress) {
progress = false;
has_short_branch_candidate = false;
int adjust_block_start = 0;
for (uint i = 0; i < nblocks; i++) {
Block* block = _cfg->get_block(i);
int idx = jmp_nidx[i];
MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
if (mach != NULL && mach->may_be_short_branch()) {
#ifdef ASSERT
assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
int j;
// Find the branch; ignore trailing NOPs.
for (j = block->number_of_nodes()-1; j>=0; j--) {
Node* n = block->get_node(j);
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
break;
}
assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
#endif
int br_size = jmp_size[i];
int br_offs = blk_starts[i] + jmp_offset[i];
// This requires the TRUE branch target be in succs[0]
uint bnum = block->non_connector_successor(0)->_pre_order;
int offset = blk_starts[bnum] - br_offs;
if (bnum > i) { // adjust following block's offset
offset -= adjust_block_start;
}
// This block can be a loop header, account for the padding
// in the previous block.
int block_padding = block_worst_case_pad[i];
assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
// In the following code a nop could be inserted before
// the branch which will increase the backward distance.
bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
if (needs_padding && offset <= 0)
offset -= nop_size;
if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
// We've got a winner. Replace this branch.
MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
// Update the jmp_size.
int new_size = replacement->size(_regalloc);
int diff = br_size - new_size;
assert(diff >= (int)nop_size, "short_branch size should be smaller");
// Conservatively take into account padding between
// avoid_back_to_back branches. Previous branch could be
// converted into avoid_back_to_back branch during next
// rounds.
if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
jmp_offset[i] += nop_size;
diff -= nop_size;
}
adjust_block_start += diff;
block->map_node(replacement, idx);
mach->subsume_by(replacement, C);
mach = replacement;
progress = true;
jmp_size[i] = new_size;
DEBUG_ONLY( jmp_target[i] = bnum; );
DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
} else {
// The jump distance is not short, try again during next iteration.
has_short_branch_candidate = true;
}
} // (mach->may_be_short_branch())
if (mach != NULL && (mach->may_be_short_branch() ||
mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
}
blk_starts[i+1] -= adjust_block_start;
}
}
#ifdef ASSERT
for (uint i = 0; i < nblocks; i++) { // For all blocks
if (jmp_target[i] != 0) {
int br_size = jmp_size[i];
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
}
assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
}
}
#endif
// Step 3, compute the offsets of all blocks, will be done in fill_buffer()
// after ScheduleAndBundle().
// ------------------
// Compute size for code buffer
code_size = blk_starts[nblocks];
// Relocation records
reloc_size += 1; // Relo entry for exception handler
// Adjust reloc_size to number of record of relocation info
// Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
// a relocation index.
// The CodeBuffer will expand the locs array if this estimate is too low.
reloc_size *= 10 / sizeof(relocInfo);
}
//------------------------------FillLocArray-----------------------------------
// Create a bit of debug info and append it to the array. The mapping is from
// Java local or expression stack to constant, register or stack-slot. For
// doubles, insert 2 mappings and return 1 (to tell the caller that the next
// entry has been taken care of and caller should skip it).
static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
// This should never have accepted Bad before
assert(OptoReg::is_valid(regnum), "location must be valid");
return (OptoReg::is_reg(regnum))
? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
: new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
}
ObjectValue*
Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
for (int i = 0; i < objs->length(); i++) {
assert(objs->at(i)->is_object(), "corrupt object cache");
ObjectValue* sv = (ObjectValue*) objs->at(i);
if (sv->id() == id) {
return sv;
}
}
// Otherwise..
return NULL;
}
void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
ObjectValue* sv ) {
assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
objs->append(sv);
}
void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
GrowableArray<ScopeValue*> *array,
GrowableArray<ScopeValue*> *objs ) {
assert( local, "use _top instead of null" );
if (array->length() != idx) {
assert(array->length() == idx + 1, "Unexpected array count");
// Old functionality:
// return
// New functionality:
// Assert if the local is not top. In product mode let the new node
// override the old entry.
assert(local == top(), "LocArray collision");
if (local == top()) {
return;
}
array->pop();
}
const Type *t = local->bottom_type();
// Is it a safepoint scalar object node?
if (local->is_SafePointScalarObject()) {
SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
if (sv == NULL) {
ciKlass* cik = t->is_oopptr()->klass();
assert(cik->is_instance_klass() ||
cik->is_array_klass(), "Not supported allocation.");
sv = new ObjectValue(spobj->_idx,
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
Compile::set_sv_for_object_node(objs, sv);
uint first_ind = spobj->first_index(sfpt->jvms());
for (uint i = 0; i < spobj->n_fields(); i++) {
Node* fld_node = sfpt->in(first_ind+i);
(void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
}
}
array->append(sv);
return;
}
// Grab the register number for the local
OptoReg::Name regnum = _regalloc->get_reg_first(local);
if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
// Record the double as two float registers.
// The register mask for such a value always specifies two adjacent
// float registers, with the lower register number even.
// Normally, the allocation of high and low words to these registers
// is irrelevant, because nearly all operations on register pairs
// (e.g., StoreD) treat them as a single unit.
// Here, we assume in addition that the words in these two registers
// stored "naturally" (by operations like StoreD and double stores
// within the interpreter) such that the lower-numbered register
// is written to the lower memory address. This may seem like
// a machine dependency, but it is not--it is a requirement on
// the author of the <arch>.ad file to ensure that, for every
// even/odd double-register pair to which a double may be allocated,
// the word in the even single-register is stored to the first
// memory word. (Note that register numbers are completely
// arbitrary, and are not tied to any machine-level encodings.)
#ifdef _LP64
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
array->append(new ConstantIntValue(0));
array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
} else if ( t->base() == Type::Long ) {
array->append(new ConstantIntValue(0));
array->append(new_loc_value( _regalloc, regnum, Location::lng ));
} else if ( t->base() == Type::RawPtr ) {
// jsr/ret return address which must be restored into a the full
// width 64-bit stack slot.
array->append(new_loc_value( _regalloc, regnum, Location::lng ));
}
#else //_LP64
#ifdef SPARC
if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
// For SPARC we have to swap high and low words for
// long values stored in a single-register (g0-g7).
array->append(new_loc_value( _regalloc, regnum , Location::normal ));
array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
} else
#endif //SPARC
if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
// Repack the double/long as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
array->append(new_loc_value( _regalloc, regnum , Location::normal ));
}
#endif //_LP64
else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
OptoReg::is_reg(regnum) ) {
array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
? Location::float_in_dbl : Location::normal ));
} else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
? Location::int_in_long : Location::normal ));
} else if( t->base() == Type::NarrowOop ) {
array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
} else {
array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
}
return;
}
// No register. It must be constant data.
switch (t->base()) {
case Type::Half: // Second half of a double
ShouldNotReachHere(); // Caller should skip 2nd halves
break;
case Type::AnyPtr:
array->append(new ConstantOopWriteValue(NULL));
break;
case Type::AryPtr:
case Type::InstPtr: // fall through
array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
break;
case Type::NarrowOop:
if (t == TypeNarrowOop::NULL_PTR) {
array->append(new ConstantOopWriteValue(NULL));
} else {
array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
}
break;
case Type::Int:
array->append(new ConstantIntValue(t->is_int()->get_con()));
break;
case Type::RawPtr:
// A return address (T_ADDRESS).
assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
#ifdef _LP64
// Must be restored to the full-width 64-bit stack slot.
array->append(new ConstantLongValue(t->is_ptr()->get_con()));
#else
array->append(new ConstantIntValue(t->is_ptr()->get_con()));
#endif
break;
case Type::FloatCon: {
float f = t->is_float_constant()->getf();
array->append(new ConstantIntValue(jint_cast(f)));
break;
}
case Type::DoubleCon: {
jdouble d = t->is_double_constant()->getd();
#ifdef _LP64
array->append(new ConstantIntValue(0));
array->append(new ConstantDoubleValue(d));
#else
// Repack the double as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
jlong_accessor acc;
acc.long_value = jlong_cast(d);
array->append(new ConstantIntValue(acc.words[1]));
array->append(new ConstantIntValue(acc.words[0]));
#endif
break;
}
case Type::Long: {
jlong d = t->is_long()->get_con();
#ifdef _LP64
array->append(new ConstantIntValue(0));
array->append(new ConstantLongValue(d));
#else
// Repack the long as two jints.
// The convention the interpreter uses is that the second local
// holds the first raw word of the native double representation.
// This is actually reasonable, since locals and stack arrays
// grow downwards in all implementations.
// (If, on some machine, the interpreter's Java locals or stack
// were to grow upwards, the embedded doubles would be word-swapped.)
jlong_accessor acc;
acc.long_value = d;
array->append(new ConstantIntValue(acc.words[1]));
array->append(new ConstantIntValue(acc.words[0]));
#endif
break;
}
case Type::Top: // Add an illegal value here
array->append(new LocationValue(Location()));
break;
default:
ShouldNotReachHere();
break;
}
}
// Determine if this node starts a bundle
bool Compile::starts_bundle(const Node *n) const {
return (_node_bundling_limit > n->_idx &&
_node_bundling_base[n->_idx].starts_bundle());
}
//--------------------------Process_OopMap_Node--------------------------------
void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
// Handle special safepoint nodes for synchronization
MachSafePointNode *sfn = mach->as_MachSafePoint();
MachCallNode *mcall;
#ifdef ENABLE_ZAP_DEAD_LOCALS
assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
#endif
int safepoint_pc_offset = current_offset;
bool is_method_handle_invoke = false;
bool return_oop = false;
// Add the safepoint in the DebugInfoRecorder
if( !mach->is_MachCall() ) {
mcall = NULL;
debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
} else {
mcall = mach->as_MachCall();
// Is the call a MethodHandle call?
if (mcall->is_MachCallJava()) {
if (mcall->as_MachCallJava()->_method_handle_invoke) {
assert(has_method_handle_invokes(), "must have been set during call generation");
is_method_handle_invoke = true;
}
}
// Check if a call returns an object.
if (mcall->returns_pointer()) {
return_oop = true;
}
safepoint_pc_offset += mcall->ret_addr_offset();
debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
}
// Loop over the JVMState list to add scope information
// Do not skip safepoints with a NULL method, they need monitor info
JVMState* youngest_jvms = sfn->jvms();
int max_depth = youngest_jvms->depth();
// Allocate the object pool for scalar-replaced objects -- the map from
// small-integer keys (which can be recorded in the local and ostack
// arrays) to descriptions of the object state.
GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
// Visit scopes from oldest to youngest.
for (int depth = 1; depth <= max_depth; depth++) {
JVMState* jvms = youngest_jvms->of_depth(depth);
int idx;
ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
// Safepoints that do not have method() set only provide oop-map and monitor info
// to support GC; these do not support deoptimization.
int num_locs = (method == NULL) ? 0 : jvms->loc_size();
int num_exps = (method == NULL) ? 0 : jvms->stk_size();
int num_mon = jvms->nof_monitors();
assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
"JVMS local count must match that of the method");
// Add Local and Expression Stack Information
// Insert locals into the locarray
GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
for( idx = 0; idx < num_locs; idx++ ) {
FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
}
// Insert expression stack entries into the exparray
GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
for( idx = 0; idx < num_exps; idx++ ) {
FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
}
// Add in mappings of the monitors
assert( !method ||
!method->is_synchronized() ||
method->is_native() ||
num_mon > 0 ||
!GenerateSynchronizationCode,
"monitors must always exist for synchronized methods");
// Build the growable array of ScopeValues for exp stack
GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
// Loop over monitors and insert into array
for (idx = 0; idx < num_mon; idx++) {
// Grab the node that defines this monitor
Node* box_node = sfn->monitor_box(jvms, idx);
Node* obj_node = sfn->monitor_obj(jvms, idx);
// Create ScopeValue for object
ScopeValue *scval = NULL;
if (obj_node->is_SafePointScalarObject()) {
SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
scval = Compile::sv_for_node_id(objs, spobj->_idx);
if (scval == NULL) {
const Type *t = spobj->bottom_type();
ciKlass* cik = t->is_oopptr()->klass();
assert(cik->is_instance_klass() ||
cik->is_array_klass(), "Not supported allocation.");
ObjectValue* sv = new ObjectValue(spobj->_idx,
new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
Compile::set_sv_for_object_node(objs, sv);
uint first_ind = spobj->first_index(youngest_jvms);
for (uint i = 0; i < spobj->n_fields(); i++) {
Node* fld_node = sfn->in(first_ind+i);
(void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
}
scval = sv;
}
} else if (!obj_node->is_Con()) {
OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
} else {
scval = new_loc_value( _regalloc, obj_reg, Location::oop );
}
} else {
const TypePtr *tp = obj_node->get_ptr_type();
scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
}
OptoReg::Name box_reg = BoxLockNode::reg(box_node);
Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
monarray->append(new MonitorValue(scval, basic_lock, eliminated));
}
// We dump the object pool first, since deoptimization reads it in first.
debug_info()->dump_object_pool(objs);
// Build first class objects to pass to scope
DebugToken *locvals = debug_info()->create_scope_values(locarray);
DebugToken *expvals = debug_info()->create_scope_values(exparray);
DebugToken *monvals = debug_info()->create_monitor_values(monarray);
// Make method available for all Safepoints
ciMethod* scope_method = method ? method : _method;
// Describe the scope here
assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
// Now we can describe the scope.
debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
} // End jvms loop
// Mark the end of the scope set.
debug_info()->end_safepoint(safepoint_pc_offset);
}
// A simplified version of Process_OopMap_Node, to handle non-safepoints.
class NonSafepointEmitter {
Compile* C;
JVMState* _pending_jvms;
int _pending_offset;
void emit_non_safepoint();
public:
NonSafepointEmitter(Compile* compile) {
this->C = compile;
_pending_jvms = NULL;
_pending_offset = 0;
}
void observe_instruction(Node* n, int pc_offset) {
if (!C->debug_info()->recording_non_safepoints()) return;
Node_Notes* nn = C->node_notes_at(n->_idx);
if (nn == NULL || nn->jvms() == NULL) return;
if (_pending_jvms != NULL &&
_pending_jvms->same_calls_as(nn->jvms())) {
// Repeated JVMS? Stretch it up here.
_pending_offset = pc_offset;
} else {
if (_pending_jvms != NULL &&
_pending_offset < pc_offset) {
emit_non_safepoint();
}
_pending_jvms = NULL;
if (pc_offset > C->debug_info()->last_pc_offset()) {
// This is the only way _pending_jvms can become non-NULL:
_pending_jvms = nn->jvms();
_pending_offset = pc_offset;
}
}
}
// Stay out of the way of real safepoints:
void observe_safepoint(JVMState* jvms, int pc_offset) {
if (_pending_jvms != NULL &&
!_pending_jvms->same_calls_as(jvms) &&
_pending_offset < pc_offset) {
emit_non_safepoint();
}
_pending_jvms = NULL;
}
void flush_at_end() {
if (_pending_jvms != NULL) {
emit_non_safepoint();
}
_pending_jvms = NULL;
}
};
void NonSafepointEmitter::emit_non_safepoint() {
JVMState* youngest_jvms = _pending_jvms;
int pc_offset = _pending_offset;
// Clear it now:
_pending_jvms = NULL;
DebugInformationRecorder* debug_info = C->debug_info();
assert(debug_info->recording_non_safepoints(), "sanity");
debug_info->add_non_safepoint(pc_offset);
int max_depth = youngest_jvms->depth();
// Visit scopes from oldest to youngest.
for (int depth = 1; depth <= max_depth; depth++) {
JVMState* jvms = youngest_jvms->of_depth(depth);
ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
}
// Mark the end of the scope set.
debug_info->end_non_safepoint(pc_offset);
}
//------------------------------init_buffer------------------------------------
CodeBuffer* Compile::init_buffer(uint* blk_starts) {
// Set the initially allocated size
int code_req = initial_code_capacity;
int locs_req = initial_locs_capacity;
int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
int const_req = initial_const_capacity;
int pad_req = NativeCall::instruction_size;
// The extra spacing after the code is necessary on some platforms.
// Sometimes we need to patch in a jump after the last instruction,
// if the nmethod has been deoptimized. (See 4932387, 4894843.)
// Compute the byte offset where we can store the deopt pc.
if (fixed_slots() != 0) {
_orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
}
// Compute prolog code size
_method_size = 0;
_frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
#if defined(IA64) && !defined(AIX)
if (save_argument_registers()) {
// 4815101: this is a stub with implicit and unknown precision fp args.
// The usual spill mechanism can only generate stfd's in this case, which
// doesn't work if the fp reg to spill contains a single-precision denorm.
// Instead, we hack around the normal spill mechanism using stfspill's and
// ldffill's in the MachProlog and MachEpilog emit methods. We allocate
// space here for the fp arg regs (f8-f15) we're going to thusly spill.
//
// If we ever implement 16-byte 'registers' == stack slots, we can
// get rid of this hack and have SpillCopy generate stfspill/ldffill
// instead of stfd/stfs/ldfd/ldfs.
_frame_slots += 8*(16/BytesPerInt);
}
#endif
assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
if (has_mach_constant_base_node()) {
uint add_size = 0;
// Fill the constant table.
// Note: This must happen before shorten_branches.
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* b = _cfg->get_block(i);
for (uint j = 0; j < b->number_of_nodes(); j++) {
Node* n = b->get_node(j);
// If the node is a MachConstantNode evaluate the constant
// value section.
if (n->is_MachConstant()) {
MachConstantNode* machcon = n->as_MachConstant();
machcon->eval_constant(C);
} else if (n->is_Mach()) {
// On Power there are more nodes that issue constants.
add_size += (n->as_Mach()->ins_num_consts() * 8);
}
}
}
// Calculate the offsets of the constants and the size of the
// constant table (including the padding to the next section).
constant_table().calculate_offsets_and_size();
const_req = constant_table().size() + add_size;
}
// Initialize the space for the BufferBlob used to find and verify
// instruction size in MachNode::emit_size()
init_scratch_buffer_blob(const_req);
if (failing()) return NULL; // Out of memory
// Pre-compute the length of blocks and replace
// long branches with short if machine supports it.
shorten_branches(blk_starts, code_req, locs_req, stub_req);
// nmethod and CodeBuffer count stubs & constants as part of method's code.
// class HandlerImpl is platform-specific and defined in the *.ad files.
int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
stub_req += MAX_stubs_size; // ensure per-stub margin
code_req += MAX_inst_size; // ensure per-instruction margin
if (StressCodeBuffers)
code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
int total_req =
const_req +
code_req +
pad_req +
stub_req +
exception_handler_req +
deopt_handler_req; // deopt handler
if (has_method_handle_invokes())
total_req += deopt_handler_req; // deopt MH handler
CodeBuffer* cb = code_buffer();
cb->initialize(total_req, locs_req);
// Have we run out of code space?
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return NULL;
}
// Configure the code buffer.
cb->initialize_consts_size(const_req);
cb->initialize_stubs_size(stub_req);
cb->initialize_oop_recorder(env()->oop_recorder());
// fill in the nop array for bundling computations
MachNode *_nop_list[Bundle::_nop_count];
Bundle::initialize_nops(_nop_list, this);
return cb;
}
//------------------------------fill_buffer------------------------------------
void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
// blk_starts[] contains offsets calculated during short branches processing,
// offsets should not be increased during following steps.
// Compute the size of first NumberOfLoopInstrToAlign instructions at head
// of a loop. It is used to determine the padding for loop alignment.
compute_loop_first_inst_sizes();
// Create oopmap set.
_oop_map_set = new OopMapSet();
// !!!!! This preserves old handling of oopmaps for now
debug_info()->set_oopmaps(_oop_map_set);
uint nblocks = _cfg->number_of_blocks();
// Count and start of implicit null check instructions
uint inct_cnt = 0;
uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
// Count and start of calls
uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
uint return_offset = 0;
int nop_size = (new (this) MachNopNode())->size(_regalloc);
int previous_offset = 0;
int current_offset = 0;
int last_call_offset = -1;
int last_avoid_back_to_back_offset = -1;
#ifdef ASSERT
uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
#endif
// Create an array of unused labels, one for each basic block, if printing is enabled
#ifndef PRODUCT
int *node_offsets = NULL;
uint node_offset_limit = unique();
if (print_assembly())
node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
#endif
NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
// Emit the constant table.
if (has_mach_constant_base_node()) {
constant_table().emit(*cb);
}
// Create an array of labels, one for each basic block
Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
for (uint i=0; i <= nblocks; i++) {
blk_labels[i].init();
}
// ------------------
// Now fill in the code buffer
Node *delay_slot = NULL;
for (uint i = 0; i < nblocks; i++) {
Block* block = _cfg->get_block(i);
Node* head = block->head();
// If this block needs to start aligned (i.e, can be reached other
// than by falling-thru from the previous block), then force the
// start of a new bundle.
if (Pipeline::requires_bundling() && starts_bundle(head)) {
cb->flush_bundle(true);
}
#ifdef ASSERT
if (!block->is_connector()) {
stringStream st;
block->dump_head(_cfg, &st);
MacroAssembler(cb).block_comment(st.as_string());
}
jmp_target[i] = 0;
jmp_offset[i] = 0;
jmp_size[i] = 0;
jmp_rule[i] = 0;
#endif
int blk_offset = current_offset;
// Define the label at the beginning of the basic block
MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
uint last_inst = block->number_of_nodes();
// Emit block normally, except for last instruction.
// Emit means "dump code bits into code buffer".
for (uint j = 0; j<last_inst; j++) {
// Get the node
Node* n = block->get_node(j);
// See if delay slots are supported
if (valid_bundle_info(n) &&
node_bundling(n)->used_in_unconditional_delay()) {
assert(delay_slot == NULL, "no use of delay slot node");
assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
delay_slot = n;
continue;
}
// If this starts a new instruction group, then flush the current one
// (but allow split bundles)
if (Pipeline::requires_bundling() && starts_bundle(n))
cb->flush_bundle(false);
// The following logic is duplicated in the code ifdeffed for
// ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
// should be factored out. Or maybe dispersed to the nodes?
// Special handling for SafePoint/Call Nodes
bool is_mcall = false;
if (n->is_Mach()) {
MachNode *mach = n->as_Mach();
is_mcall = n->is_MachCall();
bool is_sfn = n->is_MachSafePoint();
// If this requires all previous instructions be flushed, then do so
if (is_sfn || is_mcall || mach->alignment_required() != 1) {
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
// A padding may be needed again since a previous instruction
// could be moved to delay slot.
// align the instruction if necessary
int padding = mach->compute_padding(current_offset);
// Make sure safepoint node for polling is distinct from a call's
// return by adding a nop if needed.
if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
padding = nop_size;
}
if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
current_offset == last_avoid_back_to_back_offset) {
// Avoid back to back some instructions.
padding = nop_size;
}
if(padding > 0) {
assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
int nops_cnt = padding / nop_size;
MachNode *nop = new (this) MachNopNode(nops_cnt);
block->insert_node(nop, j++);
last_inst++;
_cfg->map_node_to_block(nop, block);
nop->emit(*cb, _regalloc);
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
// Remember the start of the last call in a basic block
if (is_mcall) {
MachCallNode *mcall = mach->as_MachCall();
// This destination address is NOT PC-relative
mcall->method_set((intptr_t)mcall->entry_point());
// Save the return address
call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
if (mcall->is_MachCallLeaf()) {
is_mcall = false;
is_sfn = false;
}
}
// sfn will be valid whenever mcall is valid now because of inheritance
if (is_sfn || is_mcall) {
// Handle special safepoint nodes for synchronization
if (!is_mcall) {
MachSafePointNode *sfn = mach->as_MachSafePoint();
// !!!!! Stubs only need an oopmap right now, so bail out
if (sfn->jvms()->method() == NULL) {
// Write the oopmap directly to the code blob??!!
# ifdef ENABLE_ZAP_DEAD_LOCALS
assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
# endif
continue;
}
} // End synchronization
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
current_offset);
Process_OopMap_Node(mach, current_offset);
} // End if safepoint
// If this is a null check, then add the start of the previous instruction to the list
else if( mach->is_MachNullCheck() ) {
inct_starts[inct_cnt++] = previous_offset;
}
// If this is a branch, then fill in the label with the target BB's label
else if (mach->is_MachBranch()) {
// This requires the TRUE branch target be in succs[0]
uint block_num = block->non_connector_successor(0)->_pre_order;
// Try to replace long branch if delay slot is not used,
// it is mostly for back branches since forward branch's
// distance is not updated yet.
bool delay_slot_is_used = valid_bundle_info(n) &&
node_bundling(n)->use_unconditional_delay();
if (!delay_slot_is_used && mach->may_be_short_branch()) {
assert(delay_slot == NULL, "not expecting delay slot node");
int br_size = n->size(_regalloc);
int offset = blk_starts[block_num] - current_offset;
if (block_num >= i) {
// Current and following block's offset are not
// finalized yet, adjust distance by the difference
// between calculated and final offsets of current block.
offset -= (blk_starts[i] - blk_offset);
}
// In the following code a nop could be inserted before
// the branch which will increase the backward distance.
bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
if (needs_padding && offset <= 0)
offset -= nop_size;
if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
// We've got a winner. Replace this branch.
MachNode* replacement = mach->as_MachBranch()->short_branch_version(this);
// Update the jmp_size.
int new_size = replacement->size(_regalloc);
assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
// Insert padding between avoid_back_to_back branches.
if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
MachNode *nop = new (this) MachNopNode();
block->insert_node(nop, j++);
_cfg->map_node_to_block(nop, block);
last_inst++;
nop->emit(*cb, _regalloc);
cb->flush_bundle(true);
current_offset = cb->insts_size();
}
#ifdef ASSERT
jmp_target[i] = block_num;
jmp_offset[i] = current_offset - blk_offset;
jmp_size[i] = new_size;
jmp_rule[i] = mach->rule();
#endif
block->map_node(replacement, j);
mach->subsume_by(replacement, C);
n = replacement;
mach = replacement;
}
}
mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
} else if (mach->ideal_Opcode() == Op_Jump) {
for (uint h = 0; h < block->_num_succs; h++) {
Block* succs_block = block->_succs[h];
for (uint j = 1; j < succs_block->num_preds(); j++) {
Node* jpn = succs_block->pred(j);
if (jpn->is_JumpProj() && jpn->in(0) == mach) {
uint block_num = succs_block->non_connector()->_pre_order;
Label *blkLabel = &blk_labels[block_num];
mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
}
}
}
}
#ifdef ASSERT
// Check that oop-store precedes the card-mark
else if (mach->ideal_Opcode() == Op_StoreCM) {
uint storeCM_idx = j;
int count = 0;
for (uint prec = mach->req(); prec < mach->len(); prec++) {
Node *oop_store = mach->in(prec); // Precedence edge
if (oop_store == NULL) continue;
count++;
uint i4;
for (i4 = 0; i4 < last_inst; ++i4) {
if (block->get_node(i4) == oop_store) {
break;
}
}
// Note: This test can provide a false failure if other precedence
// edges have been added to the storeCMNode.
assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
}
assert(count > 0, "storeCM expects at least one precedence edge");
}
#endif
else if (!n->is_Proj()) {
// Remember the beginning of the previous instruction, in case
// it's followed by a flag-kill and a null-check. Happens on
// Intel all the time, with add-to-memory kind of opcodes.
previous_offset = current_offset;
}
// Not an else-if!
// If this is a trap based cmp then add its offset to the list.
if (mach->is_TrapBasedCheckNode()) {
inct_starts[inct_cnt++] = current_offset;
}
}
// Verify that there is sufficient space remaining
cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return;
}
// Save the offset for the listing
#ifndef PRODUCT
if (node_offsets && n->_idx < node_offset_limit)
node_offsets[n->_idx] = cb->insts_size();
#endif
// "Normal" instruction case
DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
n->emit(*cb, _regalloc);
current_offset = cb->insts_size();
// Above we only verified that there is enough space in the instruction section.
// However, the instruction may emit stubs that cause code buffer expansion.
// Bail out here if expansion failed due to a lack of code cache space.
if (failing()) {
return;
}
#ifdef ASSERT
uint n_size = n->size(_regalloc);
if (n_size < (current_offset-instr_offset)) {
MachNode* mach = n->as_Mach();
n->dump();
mach->dump_format(_regalloc, tty);
tty->print_cr(" n_size (%d), current_offset (%d), instr_offset (%d)", n_size, current_offset, instr_offset);
Disassembler::decode(cb->insts_begin() + instr_offset, cb->insts_begin() + current_offset + 1, tty);
tty->print_cr(" ------------------- ");
BufferBlob* blob = this->scratch_buffer_blob();
address blob_begin = blob->content_begin();
Disassembler::decode(blob_begin, blob_begin + n_size + 1, tty);
assert(false, "wrong size of mach node");
}
#endif
non_safepoints.observe_instruction(n, current_offset);
// mcall is last "call" that can be a safepoint
// record it so we can see if a poll will directly follow it
// in which case we'll need a pad to make the PcDesc sites unique
// see 5010568. This can be slightly inaccurate but conservative
// in the case that return address is not actually at current_offset.
// This is a small price to pay.
if (is_mcall) {
last_call_offset = current_offset;
}
if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
// Avoid back to back some instructions.
last_avoid_back_to_back_offset = current_offset;
}
// See if this instruction has a delay slot
if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
assert(delay_slot != NULL, "expecting delay slot node");
// Back up 1 instruction
cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
// Save the offset for the listing
#ifndef PRODUCT
if (node_offsets && delay_slot->_idx < node_offset_limit)
node_offsets[delay_slot->_idx] = cb->insts_size();
#endif
// Support a SafePoint in the delay slot
if (delay_slot->is_MachSafePoint()) {
MachNode *mach = delay_slot->as_Mach();
// !!!!! Stubs only need an oopmap right now, so bail out
if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
// Write the oopmap directly to the code blob??!!
# ifdef ENABLE_ZAP_DEAD_LOCALS
assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
# endif
delay_slot = NULL;
continue;
}
int adjusted_offset = current_offset - Pipeline::instr_unit_size();
non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
adjusted_offset);
// Generate an OopMap entry
Process_OopMap_Node(mach, adjusted_offset);
}
// Insert the delay slot instruction
delay_slot->emit(*cb, _regalloc);
// Don't reuse it
delay_slot = NULL;
}
} // End for all instructions in block
// If the next block is the top of a loop, pad this block out to align
// the loop top a little. Helps prevent pipe stalls at loop back branches.
if (i < nblocks-1) {
Block *nb = _cfg->get_block(i + 1);
int padding = nb->alignment_padding(current_offset);
if( padding > 0 ) {
MachNode *nop = new (this) MachNopNode(padding / nop_size);
block->insert_node(nop, block->number_of_nodes());
_cfg->map_node_to_block(nop, block);
nop->emit(*cb, _regalloc);
current_offset = cb->insts_size();
}
}
// Verify that the distance for generated before forward
// short branches is still valid.
guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
// Save new block start offset
blk_starts[i] = blk_offset;
} // End of for all blocks
blk_starts[nblocks] = current_offset;
non_safepoints.flush_at_end();
// Offset too large?
if (failing()) return;
// Define a pseudo-label at the end of the code
MacroAssembler(cb).bind( blk_labels[nblocks] );
// Compute the size of the first block
_first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
assert(cb->insts_size() < 500000, "method is unreasonably large");
#ifdef ASSERT
for (uint i = 0; i < nblocks; i++) { // For all blocks
if (jmp_target[i] != 0) {
int br_size = jmp_size[i];
int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
assert(false, "Displacement too large for short jmp");
}
}
}
#endif
#ifndef PRODUCT
// Information on the size of the method, without the extraneous code
Scheduling::increment_method_size(cb->insts_size());
#endif
// ------------------
// Fill in exception table entries.
FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
// Only java methods have exception handlers and deopt handlers
// class HandlerImpl is platform-specific and defined in the *.ad files.
if (_method) {
// Emit the exception handler code.
_code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
if (failing()) {
return; // CodeBuffer::expand failed
}
// Emit the deopt handler code.
_code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
// Emit the MethodHandle deopt handler code (if required).
if (has_method_handle_invokes() && !failing()) {
// We can use the same code as for the normal deopt handler, we
// just need a different entry point address.
_code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
}
}
// One last check for failed CodeBuffer::expand:
if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
C->record_failure("CodeCache is full");
return;
}
#ifndef PRODUCT
// Dump the assembly code, including basic-block numbers
if (print_assembly()) {
ttyLocker ttyl; // keep the following output all in one block
if (!VMThread::should_terminate()) { // test this under the tty lock
// This output goes directly to the tty, not the compiler log.
// To enable tools to match it up with the compilation activity,
// be sure to tag this tty output with the compile ID.
if (xtty != NULL) {
xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
is_osr_compilation() ? " compile_kind='osr'" :
"");
}
if (method() != NULL) {
method()->print_metadata();
} else if (stub_name() != NULL) {
tty->print_cr("Generating RuntimeStub - %s", stub_name());
}
dump_asm(node_offsets, node_offset_limit);
if (xtty != NULL) {
xtty->tail("opto_assembly");
}
}
}
#endif
}
void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
_inc_table.set_size(cnt);
uint inct_cnt = 0;
for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
Block* block = _cfg->get_block(i);
Node *n = NULL;
int j;
// Find the branch; ignore trailing NOPs.
for (j = block->number_of_nodes() - 1; j >= 0; j--) {
n = block->get_node(j);
if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
break;
}
}
// If we didn't find anything, continue
if (j < 0) {
continue;
}
// Compute ExceptionHandlerTable subtable entry and add it
// (skip empty blocks)
if (n->is_Catch()) {
// Get the offset of the return from the call
uint call_return = call_returns[block->_pre_order];
#ifdef ASSERT
assert( call_return > 0, "no call seen for this basic block" );
while (block->get_node(--j)->is_MachProj()) ;
assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
#endif
// last instruction is a CatchNode, find it's CatchProjNodes
int nof_succs = block->_num_succs;
// allocate space
GrowableArray<intptr_t> handler_bcis(nof_succs);
GrowableArray<intptr_t> handler_pcos(nof_succs);
// iterate through all successors
for (int j = 0; j < nof_succs; j++) {
Block* s = block->_succs[j];
bool found_p = false;
for (uint k = 1; k < s->num_preds(); k++) {
Node* pk = s->pred(k);
if (pk->is_CatchProj() && pk->in(0) == n) {
const CatchProjNode* p = pk->as_CatchProj();
found_p = true;
// add the corresponding handler bci & pco information
if (p->_con != CatchProjNode::fall_through_index) {
// p leads to an exception handler (and is not fall through)
assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
// no duplicates, please
if (!handler_bcis.contains(p->handler_bci())) {
uint block_num = s->non_connector()->_pre_order;
handler_bcis.append(p->handler_bci());
handler_pcos.append(blk_labels[block_num].loc_pos());
}
}
}
}
assert(found_p, "no matching predecessor found");
// Note: Due to empty block removal, one block may have
// several CatchProj inputs, from the same Catch.
}
// Set the offset of the return from the call
_handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
continue;
}
// Handle implicit null exception table updates
if (n->is_MachNullCheck()) {
uint block_num = block->non_connector_successor(0)->_pre_order;
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
continue;
}
// Handle implicit exception table updates: trap instructions.
if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
uint block_num = block->non_connector_successor(0)->_pre_order;
_inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
continue;
}
} // End of for all blocks fill in exception table entries
}
// Static Variables
#ifndef PRODUCT
uint Scheduling::_total_nop_size = 0;
uint Scheduling::_total_method_size = 0;
uint Scheduling::_total_branches = 0;
uint Scheduling::_total_unconditional_delays = 0;
uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
#endif
// Initializer for class Scheduling
Scheduling::Scheduling(Arena *arena, Compile &compile)
: _arena(arena),
_cfg(compile.cfg()),
_regalloc(compile.regalloc()),
_reg_node(arena),
_bundle_instr_count(0),
_bundle_cycle_number(0),
_scheduled(arena),
_available(arena),
_next_node(NULL),
_bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
_pinch_free_list(arena)
#ifndef PRODUCT
, _branches(0)
, _unconditional_delays(0)
#endif
{
// Create a MachNopNode
_nop = new (&compile) MachNopNode();
// Now that the nops are in the array, save the count
// (but allow entries for the nops)
_node_bundling_limit = compile.unique();
uint node_max = _regalloc->node_regs_max_index();
compile.set_node_bundling_limit(_node_bundling_limit);
// This one is persistent within the Compile class
_node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
// Allocate space for fixed-size arrays
_node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
_uses = NEW_ARENA_ARRAY(arena, short, node_max);
_current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
// Clear the arrays
memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
memset(_node_latency, 0, node_max * sizeof(unsigned short));
memset(_uses, 0, node_max * sizeof(short));
memset(_current_latency, 0, node_max * sizeof(unsigned short));
// Clear the bundling information
memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
// Get the last node
Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
_next_node = block->get_node(block->number_of_nodes() - 1);
}
#ifndef PRODUCT
// Scheduling destructor
Scheduling::~Scheduling() {
_total_branches += _branches;
_total_unconditional_delays += _unconditional_delays;
}
#endif
// Step ahead "i" cycles
void Scheduling::step(uint i) {
Bundle *bundle = node_bundling(_next_node);
bundle->set_starts_bundle();
// Update the bundle record, but leave the flags information alone
if (_bundle_instr_count > 0) {
bundle->set_instr_count(_bundle_instr_count);
bundle->set_resources_used(_bundle_use.resourcesUsed());
}
// Update the state information
_bundle_instr_count = 0;
_bundle_cycle_number += i;
_bundle_use.step(i);
}
void Scheduling::step_and_clear() {
Bundle *bundle = node_bundling(_next_node);
bundle->set_starts_bundle();
// Update the bundle record
if (_bundle_instr_count > 0) {
bundle->set_instr_count(_bundle_instr_count);
bundle->set_resources_used(_bundle_use.resourcesUsed());
_bundle_cycle_number += 1;
}
// Clear the bundling information
_bundle_instr_count = 0;
_bundle_use.reset();
memcpy(_bundle_use_elements,
Pipeline_Use::elaborated_elements,
sizeof(Pipeline_Use::elaborated_elements));
}
// Perform instruction scheduling and bundling over the sequence of
// instructions in backwards order.
void Compile::ScheduleAndBundle() {
// Don't optimize this if it isn't a method
if (!_method)
return;
// Don't optimize this if scheduling is disabled
if (!do_scheduling())
return;
// Scheduling code works only with pairs (8 bytes) maximum.
if (max_vector_size() > 8)
return;
NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
// Create a data structure for all the scheduling information
Scheduling scheduling(Thread::current()->resource_area(), *this);
// Walk backwards over each basic block, computing the needed alignment
// Walk over all the basic blocks
scheduling.DoScheduling();
}
// Compute the latency of all the instructions. This is fairly simple,
// because we already have a legal ordering. Walk over the instructions
// from first to last, and compute the latency of the instruction based
// on the latency of the preceding instruction(s).
void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> ComputeLocalLatenciesForward\n");
#endif
// Walk over all the schedulable instructions
for( uint j=_bb_start; j < _bb_end; j++ ) {
// This is a kludge, forcing all latency calculations to start at 1.
// Used to allow latency 0 to force an instruction to the beginning
// of the bb
uint latency = 1;
Node *use = bb->get_node(j);
uint nlen = use->len();
// Walk over all the inputs
for ( uint k=0; k < nlen; k++ ) {
Node *def = use->in(k);
if (!def)
continue;
uint l = _node_latency[def->_idx] + use->latency(k);
if (latency < l)
latency = l;
}
_node_latency[use->_idx] = latency;
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# latency %4d: ", latency);
use->dump();
}
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- ComputeLocalLatenciesForward\n");
#endif
} // end ComputeLocalLatenciesForward
// See if this node fits into the present instruction bundle
bool Scheduling::NodeFitsInBundle(Node *n) {
uint n_idx = n->_idx;
// If this is the unconditional delay instruction, then it fits
if (n == _unconditional_delay_slot) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
#endif
return (true);
}
// If the node cannot be scheduled this cycle, skip it
if (_current_latency[n_idx] > _bundle_cycle_number) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
n->_idx, _current_latency[n_idx], _bundle_cycle_number);
#endif
return (false);
}
const Pipeline *node_pipeline = n->pipeline();
uint instruction_count = node_pipeline->instructionCount();
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
instruction_count = 0;
else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
instruction_count++;
if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
#endif
return (false);
}
// Don't allow non-machine nodes to be handled this way
if (!n->is_Mach() && instruction_count == 0)
return (false);
// See if there is any overlap
uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
if (delay > 0) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
#endif
return false;
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
#endif
return true;
}
Node * Scheduling::ChooseNodeToBundle() {
uint siz = _available.size();
if (siz == 0) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# ChooseNodeToBundle: NULL\n");
#endif
return (NULL);
}
// Fast path, if only 1 instruction in the bundle
if (siz == 1) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle (only 1): ");
_available[0]->dump();
}
#endif
return (_available[0]);
}
// Don't bother, if the bundle is already full
if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
for ( uint i = 0; i < siz; i++ ) {
Node *n = _available[i];
// Skip projections, we'll handle them another way
if (n->is_Proj())
continue;
// This presupposed that instructions are inserted into the
// available list in a legality order; i.e. instructions that
// must be inserted first are at the head of the list
if (NodeFitsInBundle(n)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle: ");
n->dump();
}
#endif
return (n);
}
}
}
// Nothing fits in this bundle, choose the highest priority
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# ChooseNodeToBundle: ");
_available[0]->dump();
}
#endif
return _available[0];
}
void Scheduling::AddNodeToAvailableList(Node *n) {
assert( !n->is_Proj(), "projections never directly made available" );
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# AddNodeToAvailableList: ");
n->dump();
}
#endif
int latency = _current_latency[n->_idx];
// Insert in latency order (insertion sort)
uint i;
for ( i=0; i < _available.size(); i++ )
if (_current_latency[_available[i]->_idx] > latency)
break;
// Special Check for compares following branches
if( n->is_Mach() && _scheduled.size() > 0 ) {
int op = n->as_Mach()->ideal_Opcode();
Node *last = _scheduled[0];
if( last->is_MachIf() && last->in(1) == n &&
( op == Op_CmpI ||
op == Op_CmpU ||
op == Op_CmpUL ||
op == Op_CmpP ||
op == Op_CmpF ||
op == Op_CmpD ||
op == Op_CmpL ) ) {
// Recalculate position, moving to front of same latency
for ( i=0 ; i < _available.size(); i++ )
if (_current_latency[_available[i]->_idx] >= latency)
break;
}
}
// Insert the node in the available list
_available.insert(i, n);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
dump_available();
#endif
}
void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
for ( uint i=0; i < n->len(); i++ ) {
Node *def = n->in(i);
if (!def) continue;
if( def->is_Proj() ) // If this is a machine projection, then
def = def->in(0); // propagate usage thru to the base instruction
if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
continue;
}
// Compute the latency
uint l = _bundle_cycle_number + n->latency(i);
if (_current_latency[def->_idx] < l)
_current_latency[def->_idx] = l;
// If this does not have uses then schedule it
if ((--_uses[def->_idx]) == 0)
AddNodeToAvailableList(def);
}
}
void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# AddNodeToBundle: ");
n->dump();
}
#endif
// Remove this from the available list
uint i;
for (i = 0; i < _available.size(); i++)
if (_available[i] == n)
break;
assert(i < _available.size(), "entry in _available list not found");
_available.remove(i);
// See if this fits in the current bundle
const Pipeline *node_pipeline = n->pipeline();
const Pipeline_Use& node_usage = node_pipeline->resourceUse();
// Check for instructions to be placed in the delay slot. We
// do this before we actually schedule the current instruction,
// because the delay slot follows the current instruction.
if (Pipeline::_branch_has_delay_slot &&
node_pipeline->hasBranchDelay() &&
!_unconditional_delay_slot) {
uint siz = _available.size();
// Conditional branches can support an instruction that
// is unconditionally executed and not dependent by the
// branch, OR a conditionally executed instruction if
// the branch is taken. In practice, this means that
// the first instruction at the branch target is
// copied to the delay slot, and the branch goes to
// the instruction after that at the branch target
if ( n->is_MachBranch() ) {
assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
assert( !n->is_Catch(), "should not look for delay slot for Catch" );
#ifndef PRODUCT
_branches++;
#endif
// At least 1 instruction is on the available list
// that is not dependent on the branch
for (uint i = 0; i < siz; i++) {
Node *d = _available[i];
const Pipeline *avail_pipeline = d->pipeline();
// Don't allow safepoints in the branch shadow, that will
// cause a number of difficulties
if ( avail_pipeline->instructionCount() == 1 &&
!avail_pipeline->hasMultipleBundles() &&
!avail_pipeline->hasBranchDelay() &&
Pipeline::instr_has_unit_size() &&
d->size(_regalloc) == Pipeline::instr_unit_size() &&
NodeFitsInBundle(d) &&
!node_bundling(d)->used_in_delay()) {
if (d->is_Mach() && !d->is_MachSafePoint()) {
// A node that fits in the delay slot was found, so we need to
// set the appropriate bits in the bundle pipeline information so
// that it correctly indicates resource usage. Later, when we
// attempt to add this instruction to the bundle, we will skip
// setting the resource usage.
_unconditional_delay_slot = d;
node_bundling(n)->set_use_unconditional_delay();
node_bundling(d)->set_used_in_unconditional_delay();
_bundle_use.add_usage(avail_pipeline->resourceUse());
_current_latency[d->_idx] = _bundle_cycle_number;
_next_node = d;
++_bundle_instr_count;
#ifndef PRODUCT
_unconditional_delays++;
#endif
break;
}
}
}
}
// No delay slot, add a nop to the usage
if (!_unconditional_delay_slot) {
// See if adding an instruction in the delay slot will overflow
// the bundle.
if (!NodeFitsInBundle(_nop)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(1 instruction for delay slot) ***\n");
#endif
step(1);
}
_bundle_use.add_usage(_nop->pipeline()->resourceUse());
_next_node = _nop;
++_bundle_instr_count;
}
// See if the instruction in the delay slot requires a
// step of the bundles
if (!NodeFitsInBundle(n)) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(branch won't fit) ***\n");
#endif
// Update the state information
_bundle_instr_count = 0;
_bundle_cycle_number += 1;
_bundle_use.step(1);
}
}
// Get the number of instructions
uint instruction_count = node_pipeline->instructionCount();
if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
instruction_count = 0;
// Compute the latency information
uint delay = 0;
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
if (relative_latency < 0)
relative_latency = 0;
delay = _bundle_use.full_latency(relative_latency, node_usage);
// Does not fit in this bundle, start a new one
if (delay > 0) {
step(delay);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(%d) ***\n", delay);
#endif
}
}
// If this was placed in the delay slot, ignore it
if (n != _unconditional_delay_slot) {
if (delay == 0) {
if (node_pipeline->hasMultipleBundles()) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(multiple instructions) ***\n");
#endif
step(1);
}
else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# *** STEP(%d >= %d instructions) ***\n",
instruction_count + _bundle_instr_count,
Pipeline::_max_instrs_per_cycle);
#endif
step(1);
}
}
if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
_bundle_instr_count++;
// Set the node's latency
_current_latency[n->_idx] = _bundle_cycle_number;
// Now merge the functional unit information
if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
_bundle_use.add_usage(node_usage);
// Increment the number of instructions in this bundle
_bundle_instr_count += instruction_count;
// Remember this node for later
if (n->is_Mach())
_next_node = n;
}
// It's possible to have a BoxLock in the graph and in the _bbs mapping but
// not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
// 'Schedule' them (basically ignore in the schedule) but do not insert them
// into the block. All other scheduled nodes get put in the schedule here.
int op = n->Opcode();
if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
(op != Op_Node && // Not an unused antidepedence node and
// not an unallocated boxlock
(OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
// Push any trailing projections
if( bb->get_node(bb->number_of_nodes()-1) != n ) {
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node *foi = n->fast_out(i);
if( foi->is_Proj() )
_scheduled.push(foi);
}
}
// Put the instruction in the schedule list
_scheduled.push(n);
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
dump_available();
#endif
// Walk all the definitions, decrementing use counts, and
// if a definition has a 0 use count, place it in the available list.
DecrementUseCounts(n,bb);
}
// This method sets the use count within a basic block. We will ignore all
// uses outside the current basic block. As we are doing a backwards walk,
// any node we reach that has a use count of 0 may be scheduled. This also
// avoids the problem of cyclic references from phi nodes, as long as phi
// nodes are at the front of the basic block. This method also initializes
// the available list to the set of instructions that have no uses within this
// basic block.
void Scheduling::ComputeUseCount(const Block *bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> ComputeUseCount\n");
#endif
// Clear the list of available and scheduled instructions, just in case
_available.clear();
_scheduled.clear();
// No delay slot specified
_unconditional_delay_slot = NULL;
#ifdef ASSERT
for( uint i=0; i < bb->number_of_nodes(); i++ )
assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
#endif
// Force the _uses count to never go to zero for unscheduable pieces
// of the block
for( uint k = 0; k < _bb_start; k++ )
_uses[bb->get_node(k)->_idx] = 1;
for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
_uses[bb->get_node(l)->_idx] = 1;
// Iterate backwards over the instructions in the block. Don't count the
// branch projections at end or the block header instructions.
for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
Node *n = bb->get_node(j);
if( n->is_Proj() ) continue; // Projections handled another way
// Account for all uses
for ( uint k = 0; k < n->len(); k++ ) {
Node *inp = n->in(k);
if (!inp) continue;
assert(inp != n, "no cycles allowed" );
if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
if (inp->is_Proj()) { // Skip through Proj's
inp = inp->in(0);
}
++_uses[inp->_idx]; // Count 1 block-local use
}
}
// If this instruction has a 0 use count, then it is available
if (!_uses[n->_idx]) {
_current_latency[n->_idx] = _bundle_cycle_number;
AddNodeToAvailableList(n);
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# uses: %3d: ", _uses[n->_idx]);
n->dump();
}
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- ComputeUseCount\n");
#endif
}
// This routine performs scheduling on each basic block in reverse order,
// using instruction latencies and taking into account function unit
// availability.
void Scheduling::DoScheduling() {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# -> DoScheduling\n");
#endif
Block *succ_bb = NULL;
Block *bb;
// Walk over all the basic blocks in reverse order
for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
bb = _cfg->get_block(i);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# Schedule BB#%03d (initial)\n", i);
for (uint j = 0; j < bb->number_of_nodes(); j++) {
bb->get_node(j)->dump();
}
}
#endif
// On the head node, skip processing
if (bb == _cfg->get_root_block()) {
continue;
}
// Skip empty, connector blocks
if (bb->is_connector())
continue;
// If the following block is not the sole successor of
// this one, then reset the pipeline information
if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
_next_node->_idx, _bundle_instr_count);
}
#endif
step_and_clear();
}
// Leave untouched the starting instruction, any Phis, a CreateEx node
// or Top. bb->get_node(_bb_start) is the first schedulable instruction.
_bb_end = bb->number_of_nodes()-1;
for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
Node *n = bb->get_node(_bb_start);
// Things not matched, like Phinodes and ProjNodes don't get scheduled.
// Also, MachIdealNodes do not get scheduled
if( !n->is_Mach() ) continue; // Skip non-machine nodes
MachNode *mach = n->as_Mach();
int iop = mach->ideal_Opcode();
if( iop == Op_CreateEx ) continue; // CreateEx is pinned
if( iop == Op_Con ) continue; // Do not schedule Top
if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
mach->pipeline() == MachNode::pipeline_class() &&
!n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
continue;
break; // Funny loop structure to be sure...
}
// Compute last "interesting" instruction in block - last instruction we
// might schedule. _bb_end points just after last schedulable inst. We
// normally schedule conditional branches (despite them being forced last
// in the block), because they have delay slots we can fill. Calls all
// have their delay slots filled in the template expansions, so we don't
// bother scheduling them.
Node *last = bb->get_node(_bb_end);
// Ignore trailing NOPs.
while (_bb_end > 0 && last->is_Mach() &&
last->as_Mach()->ideal_Opcode() == Op_Con) {
last = bb->get_node(--_bb_end);
}
assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
if( last->is_Catch() ||
// Exclude unreachable path case when Halt node is in a separate block.
(_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
// There must be a prior call. Skip it.
while( !bb->get_node(--_bb_end)->is_MachCall() ) {
assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
}
} else if( last->is_MachNullCheck() ) {
// Backup so the last null-checked memory instruction is
// outside the schedulable range. Skip over the nullcheck,
// projection, and the memory nodes.
Node *mem = last->in(1);
do {
_bb_end--;
} while (mem != bb->get_node(_bb_end));
} else {
// Set _bb_end to point after last schedulable inst.
_bb_end++;
}
assert( _bb_start <= _bb_end, "inverted block ends" );
// Compute the register antidependencies for the basic block
ComputeRegisterAntidependencies(bb);
if (_cfg->C->failing()) return; // too many D-U pinch points
// Compute intra-bb latencies for the nodes
ComputeLocalLatenciesForward(bb);
// Compute the usage within the block, and set the list of all nodes
// in the block that have no uses within the block.
ComputeUseCount(bb);
// Schedule the remaining instructions in the block
while ( _available.size() > 0 ) {
Node *n = ChooseNodeToBundle();
guarantee(n != NULL, "no nodes available");
AddNodeToBundle(n,bb);
}
assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
#ifdef ASSERT
for( uint l = _bb_start; l < _bb_end; l++ ) {
Node *n = bb->get_node(l);
uint m;
for( m = 0; m < _bb_end-_bb_start; m++ )
if( _scheduled[m] == n )
break;
assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
}
#endif
// Now copy the instructions (in reverse order) back to the block
for ( uint k = _bb_start; k < _bb_end; k++ )
bb->map_node(_scheduled[_bb_end-k-1], k);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
tty->print("# Schedule BB#%03d (final)\n", i);
uint current = 0;
for (uint j = 0; j < bb->number_of_nodes(); j++) {
Node *n = bb->get_node(j);
if( valid_bundle_info(n) ) {
Bundle *bundle = node_bundling(n);
if (bundle->instr_count() > 0 || bundle->flags() > 0) {
tty->print("*** Bundle: ");
bundle->dump();
}
n->dump();
}
}
}
#endif
#ifdef ASSERT
verify_good_schedule(bb,"after block local scheduling");
#endif
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output())
tty->print("# <- DoScheduling\n");
#endif
// Record final node-bundling array location
_regalloc->C->set_node_bundling_base(_node_bundling_base);
} // end DoScheduling
// Verify that no live-range used in the block is killed in the block by a
// wrong DEF. This doesn't verify live-ranges that span blocks.
// Check for edge existence. Used to avoid adding redundant precedence edges.
static bool edge_from_to( Node *from, Node *to ) {
for( uint i=0; i<from->len(); i++ )
if( from->in(i) == to )
return true;
return false;
}
#ifdef ASSERT
void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
// Check for bad kills
if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
Node *prior_use = _reg_node[def];
if( prior_use && !edge_from_to(prior_use,n) ) {
tty->print("%s = ",OptoReg::as_VMReg(def)->name());
n->dump();
tty->print_cr("...");
prior_use->dump();
assert(edge_from_to(prior_use,n),msg);
}
_reg_node.map(def,NULL); // Kill live USEs
}
}
void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
// Zap to something reasonable for the verify code
_reg_node.clear();
// Walk over the block backwards. Check to make sure each DEF doesn't
// kill a live value (other than the one it's supposed to). Add each
// USE to the live set.
for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
Node *n = b->get_node(i);
int n_op = n->Opcode();
if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
// Fat-proj kills a slew of registers
RegMask rm = n->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
verify_do_def( n, kill, msg );
}
} else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
// Get DEF'd registers the normal way
verify_do_def( n, _regalloc->get_reg_first(n), msg );
verify_do_def( n, _regalloc->get_reg_second(n), msg );
}
// Now make all USEs live
for( uint i=1; i<n->req(); i++ ) {
Node *def = n->in(i);
assert(def != 0, "input edge required");
OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
if( OptoReg::is_valid(reg_lo) ) {
assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
_reg_node.map(reg_lo,n);
}
if( OptoReg::is_valid(reg_hi) ) {
assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
_reg_node.map(reg_hi,n);
}
}
}
// Zap to something reasonable for the Antidependence code
_reg_node.clear();
}
#endif
// Conditionally add precedence edges. Avoid putting edges on Projs.
static void add_prec_edge_from_to( Node *from, Node *to ) {
if( from->is_Proj() ) { // Put precedence edge on Proj's input
assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
from = from->in(0);
}
if( from != to && // No cycles (for things like LD L0,[L0+4] )
!edge_from_to( from, to ) ) // Avoid duplicate edge
from->add_prec(to);
}
void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
return;
Node *pinch = _reg_node[def_reg]; // Get pinch point
if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
is_def ) { // Check for a true def (not a kill)
_reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
return;
}
Node *kill = def; // Rename 'def' to more descriptive 'kill'
debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
// After some number of kills there _may_ be a later def
Node *later_def = NULL;
// Finding a kill requires a real pinch-point.
// Check for not already having a pinch-point.
// Pinch points are Op_Node's.
if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
later_def = pinch; // Must be def/kill as optimistic pinch-point
if ( _pinch_free_list.size() > 0) {
pinch = _pinch_free_list.pop();
} else {
pinch = new (_cfg->C) Node(1); // Pinch point to-be
}
if (pinch->_idx >= _regalloc->node_regs_max_index()) {
_cfg->C->record_method_not_compilable("too many D-U pinch points");
return;
}
_cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
_reg_node.map(def_reg,pinch); // Record pinch-point
//_regalloc->set_bad(pinch->_idx); // Already initialized this way.
if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
later_def = NULL; // and no later def
}
pinch->set_req(0,later_def); // Hook later def so we can find it
} else { // Else have valid pinch point
if( pinch->in(0) ) // If there is a later-def
later_def = pinch->in(0); // Get it
}
// Add output-dependence edge from later def to kill
if( later_def ) // If there is some original def
add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
// See if current kill is also a use, and so is forced to be the pinch-point.
if( pinch->Opcode() == Op_Node ) {
Node *uses = kill->is_Proj() ? kill->in(0) : kill;
for( uint i=1; i<uses->req(); i++ ) {
if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
_regalloc->get_reg_second(uses->in(i)) == def_reg ) {
// Yes, found a use/kill pinch-point
pinch->set_req(0,NULL); //
pinch->replace_by(kill); // Move anti-dep edges up
pinch = kill;
_reg_node.map(def_reg,pinch);
return;
}
}
}
// Add edge from kill to pinch-point
add_prec_edge_from_to(kill,pinch);
}
void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
return;
Node *pinch = _reg_node[use_reg]; // Get pinch point
// Check for no later def_reg/kill in block
if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
// Use has to be block-local as well
_cfg->get_block_for_node(use) == b) {
if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
pinch->req() == 1 ) { // pinch not yet in block?
pinch->del_req(0); // yank pointer to later-def, also set flag
// Insert the pinch-point in the block just after the last use
b->insert_node(pinch, b->find_node(use) + 1);
_bb_end++; // Increase size scheduled region in block
}
add_prec_edge_from_to(pinch,use);
}
}
// We insert antidependences between the reads and following write of
// allocated registers to prevent illegal code motion. Hopefully, the
// number of added references should be fairly small, especially as we
// are only adding references within the current basic block.
void Scheduling::ComputeRegisterAntidependencies(Block *b) {
#ifdef ASSERT
verify_good_schedule(b,"before block local scheduling");
#endif
// A valid schedule, for each register independently, is an endless cycle
// of: a def, then some uses (connected to the def by true dependencies),
// then some kills (defs with no uses), finally the cycle repeats with a new
// def. The uses are allowed to float relative to each other, as are the
// kills. No use is allowed to slide past a kill (or def). This requires
// antidependencies between all uses of a single def and all kills that
// follow, up to the next def. More edges are redundant, because later defs
// & kills are already serialized with true or antidependencies. To keep
// the edge count down, we add a 'pinch point' node if there's more than
// one use or more than one kill/def.
// We add dependencies in one bottom-up pass.
// For each instruction we handle it's DEFs/KILLs, then it's USEs.
// For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
// register. If not, we record the DEF/KILL in _reg_node, the
// register-to-def mapping. If there is a prior DEF/KILL, we insert a
// "pinch point", a new Node that's in the graph but not in the block.
// We put edges from the prior and current DEF/KILLs to the pinch point.
// We put the pinch point in _reg_node. If there's already a pinch point
// we merely add an edge from the current DEF/KILL to the pinch point.
// After doing the DEF/KILLs, we handle USEs. For each used register, we
// put an edge from the pinch point to the USE.
// To be expedient, the _reg_node array is pre-allocated for the whole
// compilation. _reg_node is lazily initialized; it either contains a NULL,
// or a valid def/kill/pinch-point, or a leftover node from some prior
// block. Leftover node from some prior block is treated like a NULL (no
// prior def, so no anti-dependence needed). Valid def is distinguished by
// it being in the current block.
bool fat_proj_seen = false;
uint last_safept = _bb_end-1;
Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
Node* last_safept_node = end_node;
for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
Node *n = b->get_node(i);
int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
// Fat-proj kills a slew of registers
// This can add edges to 'n' and obscure whether or not it was a def,
// hence the is_def flag.
fat_proj_seen = true;
RegMask rm = n->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
anti_do_def( b, n, kill, is_def );
}
} else {
// Get DEF'd registers the normal way
anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
}
// Kill projections on a branch should appear to occur on the
// branch, not afterwards, so grab the masks from the projections
// and process them.
if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) {
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* use = n->fast_out(i);
if (use->is_Proj()) {
RegMask rm = use->out_RegMask();// Make local copy
while( rm.is_NotEmpty() ) {
OptoReg::Name kill = rm.find_first_elem();
rm.Remove(kill);
anti_do_def( b, n, kill, false );
}
}
}
}
// Check each register used by this instruction for a following DEF/KILL
// that must occur afterward and requires an anti-dependence edge.
for( uint j=0; j<n->req(); j++ ) {
Node *def = n->in(j);
if( def ) {
assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
anti_do_use( b, n, _regalloc->get_reg_first(def) );
anti_do_use( b, n, _regalloc->get_reg_second(def) );
}
}
// Do not allow defs of new derived values to float above GC
// points unless the base is definitely available at the GC point.
Node *m = b->get_node(i);
// Add precedence edge from following safepoint to use of derived pointer
if( last_safept_node != end_node &&
m != last_safept_node) {
for (uint k = 1; k < m->req(); k++) {
const Type *t = m->in(k)->bottom_type();
if( t->isa_oop_ptr() &&
t->is_ptr()->offset() != 0 ) {
last_safept_node->add_prec( m );
break;
}
}
}
if( n->jvms() ) { // Precedence edge from derived to safept
// Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
if( b->get_node(last_safept) != last_safept_node ) {
last_safept = b->find_node(last_safept_node);
}
for( uint j=last_safept; j > i; j-- ) {
Node *mach = b->get_node(j);
if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
mach->add_prec( n );
}
last_safept = i;
last_safept_node = m;
}
}
if (fat_proj_seen) {
// Garbage collect pinch nodes that were not consumed.
// They are usually created by a fat kill MachProj for a call.
garbage_collect_pinch_nodes();
}
}
// Garbage collect pinch nodes for reuse by other blocks.
//
// The block scheduler's insertion of anti-dependence
// edges creates many pinch nodes when the block contains
// 2 or more Calls. A pinch node is used to prevent a
// combinatorial explosion of edges. If a set of kills for a
// register is anti-dependent on a set of uses (or defs), rather
// than adding an edge in the graph between each pair of kill
// and use (or def), a pinch is inserted between them:
//
// use1 use2 use3
// \ | /
// \ | /
// pinch
// / | \
// / | \
// kill1 kill2 kill3
//
// One pinch node is created per register killed when
// the second call is encountered during a backwards pass
// over the block. Most of these pinch nodes are never
// wired into the graph because the register is never
// used or def'ed in the block.
//
void Scheduling::garbage_collect_pinch_nodes() {
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
#endif
int trace_cnt = 0;
for (uint k = 0; k < _reg_node.Size(); k++) {
Node* pinch = _reg_node[k];
if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
// no predecence input edges
(pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
cleanup_pinch(pinch);
_pinch_free_list.push(pinch);
_reg_node.map(k, NULL);
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) {
trace_cnt++;
if (trace_cnt > 40) {
tty->print("\n");
trace_cnt = 0;
}
tty->print(" %d", pinch->_idx);
}
#endif
}
}
#ifndef PRODUCT
if (_cfg->C->trace_opto_output()) tty->print("\n");
#endif
}
// Clean up a pinch node for reuse.
void Scheduling::cleanup_pinch( Node *pinch ) {
assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
Node* use = pinch->last_out(i);
uint uses_found = 0;
for (uint j = use->req(); j < use->len(); j++) {
if (use->in(j) == pinch) {
use->rm_prec(j);
uses_found++;
}
}
assert(uses_found > 0, "must be a precedence edge");
i -= uses_found; // we deleted 1 or more copies of this edge
}
// May have a later_def entry
pinch->set_req(0, NULL);
}
#ifndef PRODUCT
void Scheduling::dump_available() const {
tty->print("#Availist ");
for (uint i = 0; i < _available.size(); i++)
tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
tty->cr();
}
// Print Scheduling Statistics
void Scheduling::print_statistics() {
// Print the size added by nops for bundling
tty->print("Nops added %d bytes to total of %d bytes",
_total_nop_size, _total_method_size);
if (_total_method_size > 0)
tty->print(", for %.2f%%",
((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
tty->print("\n");
// Print the number of branch shadows filled
if (Pipeline::_branch_has_delay_slot) {
tty->print("Of %d branches, %d had unconditional delay slots filled",
_total_branches, _total_unconditional_delays);
if (_total_branches > 0)
tty->print(", for %.2f%%",
((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
tty->print("\n");
}
uint total_instructions = 0, total_bundles = 0;
for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
uint bundle_count = _total_instructions_per_bundle[i];
total_instructions += bundle_count * i;
total_bundles += bundle_count;
}
if (total_bundles > 0)
tty->print("Average ILP (excluding nops) is %.2f\n",
((double)total_instructions) / ((double)total_bundles));
}
#endif
C:\hotspot-69087d08d473\src\share\vm/opto/output.hpp
/*
* Copyright (c) 2000, 2015, 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.
*
*/
#ifndef SHARE_VM_OPTO_OUTPUT_HPP
#define SHARE_VM_OPTO_OUTPUT_HPP
#include "opto/block.hpp"
#include "opto/node.hpp"
#if defined AD_MD_HPP
# include AD_MD_HPP
#elif defined TARGET_ARCH_MODEL_x86_32
# include "adfiles/ad_x86_32.hpp"
#elif defined TARGET_ARCH_MODEL_x86_64
# include "adfiles/ad_x86_64.hpp"
#elif defined TARGET_ARCH_MODEL_aarch64
# include "adfiles/ad_aarch64.hpp"
#elif defined TARGET_ARCH_MODEL_sparc
# include "adfiles/ad_sparc.hpp"
#elif defined TARGET_ARCH_MODEL_zero
# include "adfiles/ad_zero.hpp"
#elif defined TARGET_ARCH_MODEL_ppc_64
# include "adfiles/ad_ppc_64.hpp"
#endif
class Arena;
class Bundle;
class Block;
class Block_Array;
class Node;
class Node_Array;
class Node_List;
class PhaseCFG;
class PhaseChaitin;
class Pipeline_Use_Element;
class Pipeline_Use;
#ifndef PRODUCT
#define DEBUG_ARG(x) , x
#else
#define DEBUG_ARG(x)
#endif
// Define the initial sizes for allocation of the resizable code buffer
enum {
initial_code_capacity = 16 * 1024,
initial_stub_capacity = 4 * 1024,
initial_const_capacity = 4 * 1024,
initial_locs_capacity = 3 * 1024
};
//------------------------------Scheduling----------------------------------
// This class contains all the information necessary to implement instruction
// scheduling and bundling.
class Scheduling {
private:
// Arena to use
Arena *_arena;
// Control-Flow Graph info
PhaseCFG *_cfg;
// Register Allocation info
PhaseRegAlloc *_regalloc;
// Number of nodes in the method
uint _node_bundling_limit;
// List of scheduled nodes. Generated in reverse order
Node_List _scheduled;
// List of nodes currently available for choosing for scheduling
Node_List _available;
// For each instruction beginning a bundle, the number of following
// nodes to be bundled with it.
Bundle *_node_bundling_base;
// Mapping from register to Node
Node_List _reg_node;
// Free list for pinch nodes.
Node_List _pinch_free_list;
// Latency from the beginning of the containing basic block (base 1)
// for each node.
unsigned short *_node_latency;
// Number of uses of this node within the containing basic block.
short *_uses;
// Schedulable portion of current block. Skips Region/Phi/CreateEx up
// front, branch+proj at end. Also skips Catch/CProj (same as
// branch-at-end), plus just-prior exception-throwing call.
uint _bb_start, _bb_end;
// Latency from the end of the basic block as scheduled
unsigned short *_current_latency;
// Remember the next node
Node *_next_node;
// Use this for an unconditional branch delay slot
Node *_unconditional_delay_slot;
// Pointer to a Nop
MachNopNode *_nop;
// Length of the current bundle, in instructions
uint _bundle_instr_count;
// Current Cycle number, for computing latencies and bundling
uint _bundle_cycle_number;
// Bundle information
Pipeline_Use_Element _bundle_use_elements[resource_count];
Pipeline_Use _bundle_use;
// Dump the available list
void dump_available() const;
public:
Scheduling(Arena *arena, Compile &compile);
// Destructor
NOT_PRODUCT( ~Scheduling(); )
// Step ahead "i" cycles
void step(uint i);
// Step ahead 1 cycle, and clear the bundle state (for example,
// at a branch target)
void step_and_clear();
Bundle* node_bundling(const Node *n) {
assert(valid_bundle_info(n), "oob");
return (&_node_bundling_base[n->_idx]);
}
bool valid_bundle_info(const Node *n) const {
return (_node_bundling_limit > n->_idx);
}
bool starts_bundle(const Node *n) const {
return (_node_bundling_limit > n->_idx && _node_bundling_base[n->_idx].starts_bundle());
}
// Do the scheduling
void DoScheduling();
// Compute the local latencies walking forward over the list of
// nodes for a basic block
void ComputeLocalLatenciesForward(const Block *bb);
// Compute the register antidependencies within a basic block
void ComputeRegisterAntidependencies(Block *bb);
void verify_do_def( Node *n, OptoReg::Name def, const char *msg );
void verify_good_schedule( Block *b, const char *msg );
void anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def );
void anti_do_use( Block *b, Node *use, OptoReg::Name use_reg );
// Add a node to the current bundle
void AddNodeToBundle(Node *n, const Block *bb);
// Add a node to the list of available nodes
void AddNodeToAvailableList(Node *n);
// Compute the local use count for the nodes in a block, and compute
// the list of instructions with no uses in the block as available
void ComputeUseCount(const Block *bb);
// Choose an instruction from the available list to add to the bundle
Node * ChooseNodeToBundle();
// See if this Node fits into the currently accumulating bundle
bool NodeFitsInBundle(Node *n);
// Decrement the use count for a node
void DecrementUseCounts(Node *n, const Block *bb);
// Garbage collect pinch nodes for reuse by other blocks.
void garbage_collect_pinch_nodes();
// Clean up a pinch node for reuse (helper for above).
void cleanup_pinch( Node *pinch );
// Information for statistics gathering
#ifndef PRODUCT
private:
// Gather information on size of nops relative to total
uint _branches, _unconditional_delays;
static uint _total_nop_size, _total_method_size;
static uint _total_branches, _total_unconditional_delays;
static uint _total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
public:
static void print_statistics();
static void increment_instructions_per_bundle(uint i) {
_total_instructions_per_bundle[i]++;
}
static void increment_nop_size(uint s) {
_total_nop_size += s;
}
static void increment_method_size(uint s) {
_total_method_size += s;
}
#endif
};
#endif // SHARE_VM_OPTO_OUTPUT_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/parse.hpp
/*
* Copyright (c) 1997, 2013, 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.
*
*/
#ifndef SHARE_VM_OPTO_PARSE_HPP
#define SHARE_VM_OPTO_PARSE_HPP
#include "ci/ciMethodData.hpp"
#include "ci/ciTypeFlow.hpp"
#include "compiler/methodLiveness.hpp"
#include "libadt/vectset.hpp"
#include "oops/generateOopMap.hpp"
#include "opto/graphKit.hpp"
#include "opto/subnode.hpp"
class BytecodeParseHistogram;
class InlineTree;
class Parse;
class SwitchRange;
//------------------------------InlineTree-------------------------------------
class InlineTree : public ResourceObj {
friend class VMStructs;
Compile* C; // cache
JVMState* _caller_jvms; // state of caller
ciMethod* _method; // method being called by the caller_jvms
InlineTree* _caller_tree;
uint _count_inline_bcs; // Accumulated count of inlined bytecodes
// Call-site count / interpreter invocation count, scaled recursively.
// Always between 0.0 and 1.0. Represents the percentage of the method's
// total execution time used at this call site.
const float _site_invoke_ratio;
const int _max_inline_level; // the maximum inline level for this sub-tree (may be adjusted)
float compute_callee_frequency( int caller_bci ) const;
GrowableArray<InlineTree*> _subtrees;
void print_impl(outputStream* stj, int indent) const PRODUCT_RETURN;
const char* _msg;
protected:
InlineTree(Compile* C,
const InlineTree* caller_tree,
ciMethod* callee_method,
JVMState* caller_jvms,
int caller_bci,
float site_invoke_ratio,
int max_inline_level);
InlineTree *build_inline_tree_for_callee(ciMethod* callee_method,
JVMState* caller_jvms,
int caller_bci);
bool try_to_inline(ciMethod* callee_method,
ciMethod* caller_method,
int caller_bci,
JVMState* jvms,
ciCallProfile& profile,
WarmCallInfo* wci_result,
bool& should_delay);
bool should_inline(ciMethod* callee_method,
ciMethod* caller_method,
int caller_bci,
ciCallProfile& profile,
WarmCallInfo* wci_result);
bool should_not_inline(ciMethod* callee_method,
ciMethod* caller_method,
JVMState* jvms,
WarmCallInfo* wci_result);
void print_inlining(ciMethod* callee_method, int caller_bci,
ciMethod* caller_method, bool success) const;
InlineTree* caller_tree() const { return _caller_tree; }
InlineTree* callee_at(int bci, ciMethod* m) const;
int inline_level() const { return stack_depth(); }
int stack_depth() const { return _caller_jvms ? _caller_jvms->depth() : 0; }
const char* msg() const { return _msg; }
void set_msg(const char* msg) { _msg = msg; }
public:
static const char* check_can_parse(ciMethod* callee);
static InlineTree* build_inline_tree_root();
static InlineTree* find_subtree_from_root(InlineTree* root, JVMState* jvms, ciMethod* callee);
// For temporary (stack-allocated, stateless) ilts:
InlineTree(Compile* c, ciMethod* callee_method, JVMState* caller_jvms, float site_invoke_ratio, int max_inline_level);
// InlineTree enum
enum InlineStyle {
Inline_do_not_inline = 0, //
Inline_cha_is_monomorphic = 1, //
Inline_type_profile_monomorphic = 2 //
};
// See if it is OK to inline.
// The receiver is the inline tree for the caller.
//
// The result is a temperature indication. If it is hot or cold,
// inlining is immediate or undesirable. Otherwise, the info block
// returned is newly allocated and may be enqueued.
//
// If the method is inlinable, a new inline subtree is created on the fly,
// and may be accessed by find_subtree_from_root.
// The call_method is the dest_method for a special or static invocation.
// The call_method is an optimized virtual method candidate otherwise.
WarmCallInfo* ok_to_inline(ciMethod *call_method, JVMState* caller_jvms, ciCallProfile& profile, WarmCallInfo* wci, bool& should_delay);
// Information about inlined method
JVMState* caller_jvms() const { return _caller_jvms; }
ciMethod *method() const { return _method; }
int caller_bci() const { return _caller_jvms ? _caller_jvms->bci() : InvocationEntryBci; }
uint count_inline_bcs() const { return _count_inline_bcs; }
float site_invoke_ratio() const { return _site_invoke_ratio; };
#ifndef PRODUCT
private:
uint _count_inlines; // Count of inlined methods
public:
// Debug information collected during parse
uint count_inlines() const { return _count_inlines; };
#endif
GrowableArray<InlineTree*> subtrees() { return _subtrees; }
void print_value_on(outputStream* st) const PRODUCT_RETURN;
bool _forced_inline; // Inlining was forced by CompilerOracle, ciReplay or annotation
bool forced_inline() const { return _forced_inline; }
// Count number of nodes in this subtree
int count() const;
// Dump inlining replay data to the stream.
void dump_replay_data(outputStream* out);
};
//-----------------------------------------------------------------------------
//------------------------------Parse------------------------------------------
// Parse bytecodes, build a Graph
class Parse : public GraphKit {
public:
// Per-block information needed by the parser:
class Block {
private:
ciTypeFlow::Block* _flow;
int _pred_count; // how many predecessors in CFG?
int _preds_parsed; // how many of these have been parsed?
uint _count; // how many times executed? Currently only set by _goto's
bool _is_parsed; // has this block been parsed yet?
bool _is_handler; // is this block an exception handler?
bool _has_merged_backedge; // does this block have merged backedge?
SafePointNode* _start_map; // all values flowing into this block
MethodLivenessResult _live_locals; // lazily initialized liveness bitmap
int _num_successors; // Includes only normal control flow.
int _all_successors; // Include exception paths also.
Block** _successors;
// Use init_node/init_graph to initialize Blocks.
// Block() : _live_locals((uintptr_t*)NULL,0) { ShouldNotReachHere(); }
Block() : _live_locals(NULL,0) { ShouldNotReachHere(); }
public:
// Set up the block data structure itself.
void init_node(Parse* outer, int po);
// Set up the block's relations to other blocks.
void init_graph(Parse* outer);
ciTypeFlow::Block* flow() const { return _flow; }
int pred_count() const { return _pred_count; }
int preds_parsed() const { return _preds_parsed; }
bool is_parsed() const { return _is_parsed; }
bool is_handler() const { return _is_handler; }
void set_count( uint x ) { _count = x; }
uint count() const { return _count; }
SafePointNode* start_map() const { assert(is_merged(),""); return _start_map; }
void set_start_map(SafePointNode* m) { assert(!is_merged(), ""); _start_map = m; }
// True after any predecessor flows control into this block
bool is_merged() const { return _start_map != NULL; }
#ifdef ASSERT
// True after backedge predecessor flows control into this block
bool has_merged_backedge() const { return _has_merged_backedge; }
void mark_merged_backedge(Block* pred) {
assert(is_SEL_head(), "should be loop head");
if (pred != NULL && is_SEL_backedge(pred)) {
assert(is_parsed(), "block should be parsed before merging backedges");
_has_merged_backedge = true;
}
}
#endif
// True when all non-exception predecessors have been parsed.
bool is_ready() const { return preds_parsed() == pred_count(); }
int num_successors() const { return _num_successors; }
int all_successors() const { return _all_successors; }
Block* successor_at(int i) const {
assert((uint)i < (uint)all_successors(), "");
return _successors[i];
}
Block* successor_for_bci(int bci);
int start() const { return flow()->start(); }
int limit() const { return flow()->limit(); }
int rpo() const { return flow()->rpo(); }
int start_sp() const { return flow()->stack_size(); }
bool is_loop_head() const { return flow()->is_loop_head(); }
bool is_SEL_head() const { return flow()->is_single_entry_loop_head(); }
bool is_SEL_backedge(Block* pred) const{ return is_SEL_head() && pred->rpo() >= rpo(); }
bool is_invariant_local(uint i) const {
const JVMState* jvms = start_map()->jvms();
if (!jvms->is_loc(i) || flow()->outer()->has_irreducible_entry()) return false;
return flow()->is_invariant_local(i - jvms->locoff());
}
bool can_elide_SEL_phi(uint i) const { assert(is_SEL_head(),""); return is_invariant_local(i); }
const Type* peek(int off=0) const { return stack_type_at(start_sp() - (off+1)); }
const Type* stack_type_at(int i) const;
const Type* local_type_at(int i) const;
static const Type* get_type(ciType* t) { return Type::get_typeflow_type(t); }
bool has_trap_at(int bci) const { return flow()->has_trap() && flow()->trap_bci() == bci; }
// Call this just before parsing a block.
void mark_parsed() {
assert(!_is_parsed, "must parse each block exactly once");
_is_parsed = true;
}
// Return the phi/region input index for the "current" pred,
// and bump the pred number. For historical reasons these index
// numbers are handed out in descending order. The last index is
// always PhiNode::Input (i.e., 1). The value returned is known
// as a "path number" because it distinguishes by which path we are
// entering the block.
int next_path_num() {
assert(preds_parsed() < pred_count(), "too many preds?");
return pred_count() - _preds_parsed++;
}
// Add a previously unaccounted predecessor to this block.
// This operates by increasing the size of the block's region
// and all its phi nodes (if any). The value returned is a
// path number ("pnum").
int add_new_path();
// Initialize me by recording the parser's map. My own map must be NULL.
void record_state(Parse* outer);
};
#ifndef PRODUCT
// BytecodeParseHistogram collects number of bytecodes parsed, nodes constructed, and transformations.
class BytecodeParseHistogram : public ResourceObj {
private:
enum BPHType {
BPH_transforms,
BPH_values
};
static bool _initialized;
static uint _bytecodes_parsed [Bytecodes::number_of_codes];
static uint _nodes_constructed[Bytecodes::number_of_codes];
static uint _nodes_transformed[Bytecodes::number_of_codes];
static uint _new_values [Bytecodes::number_of_codes];
Bytecodes::Code _initial_bytecode;
int _initial_node_count;
int _initial_transforms;
int _initial_values;
Parse *_parser;
Compile *_compiler;
// Initialization
static void reset();
// Return info being collected, select with global flag 'BytecodeParseInfo'
int current_count(BPHType info_selector);
public:
BytecodeParseHistogram(Parse *p, Compile *c);
static bool initialized();
// Record info when starting to parse one bytecode
void set_initial_state( Bytecodes::Code bc );
// Record results of parsing one bytecode
void record_change();
// Profile printing
static void print(float cutoff = 0.01F); // cutoff in percent
};
public:
// Record work done during parsing
BytecodeParseHistogram* _parse_histogram;
void set_parse_histogram(BytecodeParseHistogram *bph) { _parse_histogram = bph; }
BytecodeParseHistogram* parse_histogram() { return _parse_histogram; }
#endif
private:
friend class Block;
// Variables which characterize this compilation as a whole:
JVMState* _caller; // JVMS which carries incoming args & state.
float _expected_uses; // expected number of calls to this code
float _prof_factor; // discount applied to my profile counts
int _depth; // Inline tree depth, for debug printouts
const TypeFunc*_tf; // My kind of function type
int _entry_bci; // the osr bci or InvocationEntryBci
ciTypeFlow* _flow; // Results of previous flow pass.
Block* _blocks; // Array of basic-block structs.
int _block_count; // Number of elements in _blocks.
GraphKit _exits; // Record all normal returns and throws here.
bool _wrote_final; // Did we write a final field?
bool _wrote_volatile; // Did we write a volatile field?
bool _count_invocations; // update and test invocation counter
bool _method_data_update; // update method data oop
Node* _alloc_with_final; // An allocation node with final field
// Variables which track Java semantics during bytecode parsing:
Block* _block; // block currently getting parsed
ciBytecodeStream _iter; // stream of this method's bytecodes
int _blocks_merged; // Progress meter: state merges from BB preds
int _blocks_parsed; // Progress meter: BBs actually parsed
const FastLockNode* _synch_lock; // FastLockNode for synchronized method
#ifndef PRODUCT
int _max_switch_depth; // Debugging SwitchRanges.
int _est_switch_depth; // Debugging SwitchRanges.
#endif
bool _first_return; // true if return is the first to be parsed
bool _replaced_nodes_for_exceptions; // needs processing of replaced nodes in exception paths?
uint _new_idx; // any node with _idx above were new during this parsing. Used to trim the replaced nodes list.
public:
// Constructor
Parse(JVMState* caller, ciMethod* parse_method, float expected_uses);
virtual Parse* is_Parse() const { return (Parse*)this; }
// Accessors.
JVMState* caller() const { return _caller; }
float expected_uses() const { return _expected_uses; }
float prof_factor() const { return _prof_factor; }
int depth() const { return _depth; }
const TypeFunc* tf() const { return _tf; }
// entry_bci() -- see osr_bci, etc.
ciTypeFlow* flow() const { return _flow; }
// blocks() -- see rpo_at, start_block, etc.
int block_count() const { return _block_count; }
GraphKit& exits() { return _exits; }
bool wrote_final() const { return _wrote_final; }
void set_wrote_final(bool z) { _wrote_final = z; }
bool wrote_volatile() const { return _wrote_volatile; }
void set_wrote_volatile(bool z) { _wrote_volatile = z; }
bool count_invocations() const { return _count_invocations; }
bool method_data_update() const { return _method_data_update; }
Node* alloc_with_final() const { return _alloc_with_final; }
void set_alloc_with_final(Node* n) {
assert((_alloc_with_final == NULL) || (_alloc_with_final == n), "different init objects?");
_alloc_with_final = n;
}
Block* block() const { return _block; }
ciBytecodeStream& iter() { return _iter; }
Bytecodes::Code bc() const { return _iter.cur_bc(); }
void set_block(Block* b) { _block = b; }
// Derived accessors:
bool is_normal_parse() const { return _entry_bci == InvocationEntryBci; }
bool is_osr_parse() const { return _entry_bci != InvocationEntryBci; }
int osr_bci() const { assert(is_osr_parse(),""); return _entry_bci; }
void set_parse_bci(int bci);
// Must this parse be aborted?
bool failing() { return C->failing(); }
Block* rpo_at(int rpo) {
assert(0 <= rpo && rpo < _block_count, "oob");
return &_blocks[rpo];
}
Block* start_block() {
return rpo_at(flow()->start_block()->rpo());
}
// Can return NULL if the flow pass did not complete a block.
Block* successor_for_bci(int bci) {
return block()->successor_for_bci(bci);
}
private:
// Create a JVMS & map for the initial state of this method.
SafePointNode* create_entry_map();
// OSR helpers
Node *fetch_interpreter_state(int index, BasicType bt, Node *local_addrs, Node *local_addrs_base);
Node* check_interpreter_type(Node* l, const Type* type, SafePointNode* &bad_type_exit);
void load_interpreter_state(Node* osr_buf);
// Functions for managing basic blocks:
void init_blocks();
void load_state_from(Block* b);
void store_state_to(Block* b) { b->record_state(this); }
// Parse all the basic blocks.
void do_all_blocks();
// Parse the current basic block
void do_one_block();
// Raise an error if we get a bad ciTypeFlow CFG.
void handle_missing_successor(int bci);
// first actions (before BCI 0)
void do_method_entry();
// implementation of monitorenter/monitorexit
void do_monitor_enter();
void do_monitor_exit();
// Eagerly create phie throughout the state, to cope with back edges.
void ensure_phis_everywhere();
// Merge the current mapping into the basic block starting at bci
void merge( int target_bci);
// Same as plain merge, except that it allocates a new path number.
void merge_new_path( int target_bci);
// Merge the current mapping into an exception handler.
void merge_exception(int target_bci);
// Helper: Merge the current mapping into the given basic block
void merge_common(Block* target, int pnum);
// Helper functions for merging individual cells.
PhiNode *ensure_phi( int idx, bool nocreate = false);
PhiNode *ensure_memory_phi(int idx, bool nocreate = false);
// Helper to merge the current memory state into the given basic block
void merge_memory_edges(MergeMemNode* n, int pnum, bool nophi);
// Parse this bytecode, and alter the Parsers JVM->Node mapping
void do_one_bytecode();
// helper function to generate array store check
void array_store_check();
// Helper function to generate array load
void array_load(BasicType etype);
// Helper function to generate array store
void array_store(BasicType etype);
// Helper function to compute array addressing
Node* array_addressing(BasicType type, int vals, const Type* *result2=NULL);
void rtm_deopt();
// Pass current map to exits
void return_current(Node* value);
// Register finalizers on return from Object.<init>
void call_register_finalizer();
// Insert a compiler safepoint into the graph
void add_safepoint();
// Insert a compiler safepoint into the graph, if there is a back-branch.
void maybe_add_safepoint(int target_bci) {
if (UseLoopSafepoints && target_bci <= bci()) {
add_safepoint();
}
}
// Note: Intrinsic generation routines may be found in library_call.cpp.
// Helper function to setup Ideal Call nodes
void do_call();
// Helper function to uncommon-trap or bailout for non-compilable call-sites
bool can_not_compile_call_site(ciMethod *dest_method, ciInstanceKlass *klass);
// Helper function to setup for type-profile based inlining
bool prepare_type_profile_inline(ciInstanceKlass* prof_klass, ciMethod* prof_method);
// Helper functions for type checking bytecodes:
void do_checkcast();
void do_instanceof();
// Helper functions for shifting & arithmetic
void modf();
void modd();
void l2f();
void do_irem();
// implementation of _get* and _put* bytecodes
void do_getstatic() { do_field_access(true, false); }
void do_getfield () { do_field_access(true, true); }
void do_putstatic() { do_field_access(false, false); }
void do_putfield () { do_field_access(false, true); }
// common code for making initial checks and forming addresses
void do_field_access(bool is_get, bool is_field);
bool static_field_ok_in_clinit(ciField *field, ciMethod *method);
// common code for actually performing the load or store
void do_get_xxx(Node* obj, ciField* field, bool is_field);
void do_put_xxx(Node* obj, ciField* field, bool is_field);
// loading from a constant field or the constant pool
// returns false if push failed (non-perm field constants only, not ldcs)
bool push_constant(ciConstant con, bool require_constant = false, bool is_autobox_cache = false, const Type* basic_type = NULL);
// implementation of object creation bytecodes
void emit_guard_for_new(ciInstanceKlass* klass);
void do_new();
void do_newarray(BasicType elemtype);
void do_anewarray();
void do_multianewarray();
Node* expand_multianewarray(ciArrayKlass* array_klass, Node* *lengths, int ndimensions, int nargs);
// implementation of jsr/ret
void do_jsr();
void do_ret();
float dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test);
float branch_prediction(float &cnt, BoolTest::mask btest, int target_bci, Node* test);
bool seems_never_taken(float prob) const;
bool path_is_suitable_for_uncommon_trap(float prob) const;
bool seems_stable_comparison() const;
void do_ifnull(BoolTest::mask btest, Node* c);
void do_if(BoolTest::mask btest, Node* c);
int repush_if_args();
void adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
Block* path, Block* other_path);
void sharpen_type_after_if(BoolTest::mask btest,
Node* con, const Type* tcon,
Node* val, const Type* tval);
IfNode* jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask);
Node* jump_if_join(Node* iffalse, Node* iftrue);
void jump_if_true_fork(IfNode *ifNode, int dest_bci_if_true, int prof_table_index);
void jump_if_false_fork(IfNode *ifNode, int dest_bci_if_false, int prof_table_index);
void jump_if_always_fork(int dest_bci_if_true, int prof_table_index);
friend class SwitchRange;
void do_tableswitch();
void do_lookupswitch();
void jump_switch_ranges(Node* a, SwitchRange* lo, SwitchRange* hi, int depth = 0);
bool create_jump_tables(Node* a, SwitchRange* lo, SwitchRange* hi);
// helper functions for methodData style profiling
void test_counter_against_threshold(Node* cnt, int limit);
void increment_and_test_invocation_counter(int limit);
void test_for_osr_md_counter_at(ciMethodData* md, ciProfileData* data, ByteSize offset, int limit);
Node* method_data_addressing(ciMethodData* md, ciProfileData* data, ByteSize offset, Node* idx = NULL, uint stride = 0);
void increment_md_counter_at(ciMethodData* md, ciProfileData* data, ByteSize offset, Node* idx = NULL, uint stride = 0);
void set_md_flag_at(ciMethodData* md, ciProfileData* data, int flag_constant);
void profile_method_entry();
void profile_taken_branch(int target_bci, bool force_update = false);
void profile_not_taken_branch(bool force_update = false);
void profile_call(Node* receiver);
void profile_generic_call();
void profile_receiver_type(Node* receiver);
void profile_ret(int target_bci);
void profile_null_checkcast();
void profile_switch_case(int table_index);
// helper function for call statistics
void count_compiled_calls(bool at_method_entry, bool is_inline) PRODUCT_RETURN;
Node_Notes* make_node_notes(Node_Notes* caller_nn);
// Helper functions for handling normal and abnormal exits.
void build_exits();
// Fix up all exceptional control flow exiting a single bytecode.
void do_exceptions();
// Fix up all exiting control flow at the end of the parse.
void do_exits();
// Add Catch/CatchProjs
// The call is either a Java call or the VM's rethrow stub
void catch_call_exceptions(ciExceptionHandlerStream&);
// Handle all exceptions thrown by the inlined method.
// Also handles exceptions for individual bytecodes.
void catch_inline_exceptions(SafePointNode* ex_map);
// Merge the given map into correct exceptional exit state.
// Assumes that there is no applicable local handler.
void throw_to_exit(SafePointNode* ex_map);
// Use speculative type to optimize CmpP node
Node* optimize_cmp_with_klass(Node* c);
public:
#ifndef PRODUCT
// Handle PrintOpto, etc.
void show_parse_info();
void dump_map_adr_mem() const;
static void print_statistics(); // Print some performance counters
void dump();
void dump_bci(int bci);
#endif
};
#endif // SHARE_VM_OPTO_PARSE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/parse1.cpp
/*
* Copyright (c) 1997, 2016, 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 "interpreter/linkResolver.hpp"
#include "oops/method.hpp"
#include "opto/addnode.hpp"
#include "opto/c2compiler.hpp"
#include "opto/idealGraphPrinter.hpp"
#include "opto/locknode.hpp"
#include "opto/memnode.hpp"
#include "opto/parse.hpp"
#include "opto/rootnode.hpp"
#include "opto/runtime.hpp"
#include "runtime/arguments.hpp"
#include "runtime/handles.inline.hpp"
#include "runtime/sharedRuntime.hpp"
#include "utilities/copy.hpp"
// Static array so we can figure out which bytecodes stop us from compiling
// the most. Some of the non-static variables are needed in bytecodeInfo.cpp
// and eventually should be encapsulated in a proper class (gri 8/18/98).
int nodes_created = 0;
int methods_parsed = 0;
int methods_seen = 0;
int blocks_parsed = 0;
int blocks_seen = 0;
int explicit_null_checks_inserted = 0;
int explicit_null_checks_elided = 0;
int all_null_checks_found = 0, implicit_null_checks = 0;
int implicit_null_throws = 0;
int reclaim_idx = 0;
int reclaim_in = 0;
int reclaim_node = 0;
#ifndef PRODUCT
bool Parse::BytecodeParseHistogram::_initialized = false;
uint Parse::BytecodeParseHistogram::_bytecodes_parsed [Bytecodes::number_of_codes];
uint Parse::BytecodeParseHistogram::_nodes_constructed[Bytecodes::number_of_codes];
uint Parse::BytecodeParseHistogram::_nodes_transformed[Bytecodes::number_of_codes];
uint Parse::BytecodeParseHistogram::_new_values [Bytecodes::number_of_codes];
#endif
//------------------------------print_statistics-------------------------------
#ifndef PRODUCT
void Parse::print_statistics() {
tty->print_cr("--- Compiler Statistics ---");
tty->print("Methods seen: %d Methods parsed: %d", methods_seen, methods_parsed);
tty->print(" Nodes created: %d", nodes_created);
tty->cr();
if (methods_seen != methods_parsed)
tty->print_cr("Reasons for parse failures (NOT cumulative):");
tty->print_cr("Blocks parsed: %d Blocks seen: %d", blocks_parsed, blocks_seen);
if( explicit_null_checks_inserted )
tty->print_cr("%d original NULL checks - %d elided (%2d%%); optimizer leaves %d,", explicit_null_checks_inserted, explicit_null_checks_elided, (100*explicit_null_checks_elided)/explicit_null_checks_inserted, all_null_checks_found);
if( all_null_checks_found )
tty->print_cr("%d made implicit (%2d%%)", implicit_null_checks,
(100*implicit_null_checks)/all_null_checks_found);
if( implicit_null_throws )
tty->print_cr("%d implicit null exceptions at runtime",
implicit_null_throws);
if( PrintParseStatistics && BytecodeParseHistogram::initialized() ) {
BytecodeParseHistogram::print();
}
}
#endif
//------------------------------ON STACK REPLACEMENT---------------------------
// Construct a node which can be used to get incoming state for
// on stack replacement.
Node *Parse::fetch_interpreter_state(int index,
BasicType bt,
Node *local_addrs,
Node *local_addrs_base) {
Node *mem = memory(Compile::AliasIdxRaw);
Node *adr = basic_plus_adr( local_addrs_base, local_addrs, -index*wordSize );
Node *ctl = control();
// Very similar to LoadNode::make, except we handle un-aligned longs and
// doubles on Sparc. Intel can handle them just fine directly.
Node *l = NULL;
switch (bt) { // Signature is flattened
case T_INT: l = new (C) LoadINode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInt::INT, MemNode::unordered); break;
case T_FLOAT: l = new (C) LoadFNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::FLOAT, MemNode::unordered); break;
case T_ADDRESS: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); break;
case T_OBJECT: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInstPtr::BOTTOM, MemNode::unordered); break;
case T_LONG:
case T_DOUBLE: {
// Since arguments are in reverse order, the argument address 'adr'
// refers to the back half of the long/double. Recompute adr.
adr = basic_plus_adr(local_addrs_base, local_addrs, -(index+1)*wordSize);
if (Matcher::misaligned_doubles_ok) {
l = (bt == T_DOUBLE)
? (Node*)new (C) LoadDNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::DOUBLE, MemNode::unordered)
: (Node*)new (C) LoadLNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeLong::LONG, MemNode::unordered);
} else {
l = (bt == T_DOUBLE)
? (Node*)new (C) LoadD_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered)
: (Node*)new (C) LoadL_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered);
}
break;
}
default: ShouldNotReachHere();
}
return _gvn.transform(l);
}
// Helper routine to prevent the interpreter from handing
// unexpected typestate to an OSR method.
// The Node l is a value newly dug out of the interpreter frame.
// The type is the type predicted by ciTypeFlow. Note that it is
// not a general type, but can only come from Type::get_typeflow_type.
// The safepoint is a map which will feed an uncommon trap.
Node* Parse::check_interpreter_type(Node* l, const Type* type,
SafePointNode* &bad_type_exit) {
const TypeOopPtr* tp = type->isa_oopptr();
// TypeFlow may assert null-ness if a type appears unloaded.
if (type == TypePtr::NULL_PTR ||
(tp != NULL && !tp->klass()->is_loaded())) {
// Value must be null, not a real oop.
Node* chk = _gvn.transform( new (C) CmpPNode(l, null()) );
Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
IfNode* iff = create_and_map_if(control(), tst, PROB_MAX, COUNT_UNKNOWN);
set_control(_gvn.transform( new (C) IfTrueNode(iff) ));
Node* bad_type = _gvn.transform( new (C) IfFalseNode(iff) );
bad_type_exit->control()->add_req(bad_type);
l = null();
}
// Typeflow can also cut off paths from the CFG, based on
// types which appear unloaded, or call sites which appear unlinked.
// When paths are cut off, values at later merge points can rise
// toward more specific classes. Make sure these specific classes
// are still in effect.
if (tp != NULL && tp->klass() != C->env()->Object_klass()) {
// TypeFlow asserted a specific object type. Value must have that type.
Node* bad_type_ctrl = NULL;
l = gen_checkcast(l, makecon(TypeKlassPtr::make(tp->klass())), &bad_type_ctrl);
bad_type_exit->control()->add_req(bad_type_ctrl);
}
BasicType bt_l = _gvn.type(l)->basic_type();
BasicType bt_t = type->basic_type();
assert(_gvn.type(l)->higher_equal(type), "must constrain OSR typestate");
return l;
}
// Helper routine which sets up elements of the initial parser map when
// performing a parse for on stack replacement. Add values into map.
// The only parameter contains the address of a interpreter arguments.
void Parse::load_interpreter_state(Node* osr_buf) {
int index;
int max_locals = jvms()->loc_size();
int max_stack = jvms()->stk_size();
// Mismatch between method and jvms can occur since map briefly held
// an OSR entry state (which takes up one RawPtr word).
assert(max_locals == method()->max_locals(), "sanity");
assert(max_stack >= method()->max_stack(), "sanity");
assert((int)jvms()->endoff() == TypeFunc::Parms + max_locals + max_stack, "sanity");
assert((int)jvms()->endoff() == (int)map()->req(), "sanity");
// Find the start block.
Block* osr_block = start_block();
assert(osr_block->start() == osr_bci(), "sanity");
// Set initial BCI.
set_parse_bci(osr_block->start());
// Set initial stack depth.
set_sp(osr_block->start_sp());
// Check bailouts. We currently do not perform on stack replacement
// of loops in catch blocks or loops which branch with a non-empty stack.
if (sp() != 0) {
C->record_method_not_compilable("OSR starts with non-empty stack");
return;
}
// Do not OSR inside finally clauses:
if (osr_block->has_trap_at(osr_block->start())) {
C->record_method_not_compilable("OSR starts with an immediate trap");
return;
}
// Commute monitors from interpreter frame to compiler frame.
assert(jvms()->monitor_depth() == 0, "should be no active locks at beginning of osr");
int mcnt = osr_block->flow()->monitor_count();
Node *monitors_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals+mcnt*2-1)*wordSize);
for (index = 0; index < mcnt; index++) {
// Make a BoxLockNode for the monitor.
Node *box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
// Displaced headers and locked objects are interleaved in the
// temp OSR buffer. We only copy the locked objects out here.
// Fetch the locked object from the OSR temp buffer and copy to our fastlock node.
Node *lock_object = fetch_interpreter_state(index*2, T_OBJECT, monitors_addr, osr_buf);
// Try and copy the displaced header to the BoxNode
Node *displaced_hdr = fetch_interpreter_state((index*2) + 1, T_ADDRESS, monitors_addr, osr_buf);
store_to_memory(control(), box, displaced_hdr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
// Build a bogus FastLockNode (no code will be generated) and push the
// monitor into our debug info.
const FastLockNode *flock = _gvn.transform(new (C) FastLockNode( 0, lock_object, box ))->as_FastLock();
map()->push_monitor(flock);
// If the lock is our method synchronization lock, tuck it away in
// _sync_lock for return and rethrow exit paths.
if (index == 0 && method()->is_synchronized()) {
_synch_lock = flock;
}
}
// Use the raw liveness computation to make sure that unexpected
// values don't propagate into the OSR frame.
MethodLivenessResult live_locals = method()->liveness_at_bci(osr_bci());
if (!live_locals.is_valid()) {
// Degenerate or breakpointed method.
C->record_method_not_compilable("OSR in empty or breakpointed method");
return;
}
// Extract the needed locals from the interpreter frame.
Node *locals_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals-1)*wordSize);
// find all the locals that the interpreter thinks contain live oops
const BitMap live_oops = method()->live_local_oops_at_bci(osr_bci());
for (index = 0; index < max_locals; index++) {
if (!live_locals.at(index)) {
continue;
}
const Type *type = osr_block->local_type_at(index);
if (type->isa_oopptr() != NULL) {
// 6403625: Verify that the interpreter oopMap thinks that the oop is live
// else we might load a stale oop if the MethodLiveness disagrees with the
// result of the interpreter. If the interpreter says it is dead we agree
// by making the value go to top.
//
if (!live_oops.at(index)) {
if (C->log() != NULL) {
C->log()->elem("OSR_mismatch local_index='%d'",index);
}
set_local(index, null());
// and ignore it for the loads
continue;
}
}
// Filter out TOP, HALF, and BOTTOM. (Cf. ensure_phi.)
if (type == Type::TOP || type == Type::HALF) {
continue;
}
// If the type falls to bottom, then this must be a local that
// is mixing ints and oops or some such. Forcing it to top
// makes it go dead.
if (type == Type::BOTTOM) {
continue;
}
// Construct code to access the appropriate local.
BasicType bt = type->basic_type();
if (type == TypePtr::NULL_PTR) {
// Ptr types are mixed together with T_ADDRESS but NULL is
// really for T_OBJECT types so correct it.
bt = T_OBJECT;
}
Node *value = fetch_interpreter_state(index, bt, locals_addr, osr_buf);
set_local(index, value);
}
// Extract the needed stack entries from the interpreter frame.
for (index = 0; index < sp(); index++) {
const Type *type = osr_block->stack_type_at(index);
if (type != Type::TOP) {
// Currently the compiler bails out when attempting to on stack replace
// at a bci with a non-empty stack. We should not reach here.
ShouldNotReachHere();
}
}
// End the OSR migration
make_runtime_call(RC_LEAF, OptoRuntime::osr_end_Type(),
CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_end),
"OSR_migration_end", TypeRawPtr::BOTTOM,
osr_buf);
// Now that the interpreter state is loaded, make sure it will match
// at execution time what the compiler is expecting now:
SafePointNode* bad_type_exit = clone_map();
bad_type_exit->set_control(new (C) RegionNode(1));
assert(osr_block->flow()->jsrs()->size() == 0, "should be no jsrs live at osr point");
for (index = 0; index < max_locals; index++) {
if (stopped()) break;
Node* l = local(index);
if (l->is_top()) continue; // nothing here
const Type *type = osr_block->local_type_at(index);
if (type->isa_oopptr() != NULL) {
if (!live_oops.at(index)) {
// skip type check for dead oops
continue;
}
}
if (osr_block->flow()->local_type_at(index)->is_return_address()) {
// In our current system it's illegal for jsr addresses to be
// live into an OSR entry point because the compiler performs
// inlining of jsrs. ciTypeFlow has a bailout that detect this
// case and aborts the compile if addresses are live into an OSR
// entry point. Because of that we can assume that any address
// locals at the OSR entry point are dead. Method liveness
// isn't precise enought to figure out that they are dead in all
// cases so simply skip checking address locals all
// together. Any type check is guaranteed to fail since the
// interpreter type is the result of a load which might have any
// value and the expected type is a constant.
continue;
}
set_local(index, check_interpreter_type(l, type, bad_type_exit));
}
for (index = 0; index < sp(); index++) {
if (stopped()) break;
Node* l = stack(index);
if (l->is_top()) continue; // nothing here
const Type *type = osr_block->stack_type_at(index);
set_stack(index, check_interpreter_type(l, type, bad_type_exit));
}
if (bad_type_exit->control()->req() > 1) {
// Build an uncommon trap here, if any inputs can be unexpected.
bad_type_exit->set_control(_gvn.transform( bad_type_exit->control() ));
record_for_igvn(bad_type_exit->control());
SafePointNode* types_are_good = map();
set_map(bad_type_exit);
// The unexpected type happens because a new edge is active
// in the CFG, which typeflow had previously ignored.
// E.g., Object x = coldAtFirst() && notReached()? "str": new Integer(123).
// This x will be typed as Integer if notReached is not yet linked.
// It could also happen due to a problem in ciTypeFlow analysis.
uncommon_trap(Deoptimization::Reason_constraint,
Deoptimization::Action_reinterpret);
set_map(types_are_good);
}
}
//------------------------------Parse------------------------------------------
// Main parser constructor.
Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses)
: _exits(caller)
{
// Init some variables
_caller = caller;
_method = parse_method;
_expected_uses = expected_uses;
_depth = 1 + (caller->has_method() ? caller->depth() : 0);
_wrote_final = false;
_wrote_volatile = false;
_alloc_with_final = NULL;
_entry_bci = InvocationEntryBci;
_tf = NULL;
_block = NULL;
_first_return = true;
_replaced_nodes_for_exceptions = false;
_new_idx = C->unique();
debug_only(_block_count = -1);
debug_only(_blocks = (Block*)-1);
#ifndef PRODUCT
if (PrintCompilation || PrintOpto) {
// Make sure I have an inline tree, so I can print messages about it.
JVMState* ilt_caller = is_osr_parse() ? caller->caller() : caller;
InlineTree::find_subtree_from_root(C->ilt(), ilt_caller, parse_method);
}
_max_switch_depth = 0;
_est_switch_depth = 0;
#endif
_tf = TypeFunc::make(method());
_iter.reset_to_method(method());
_flow = method()->get_flow_analysis();
if (_flow->failing()) {
C->record_method_not_compilable_all_tiers(_flow->failure_reason());
}
#ifndef PRODUCT
if (_flow->has_irreducible_entry()) {
C->set_parsed_irreducible_loop(true);
}
#endif
if (_expected_uses <= 0) {
_prof_factor = 1;
} else {
float prof_total = parse_method->interpreter_invocation_count();
if (prof_total <= _expected_uses) {
_prof_factor = 1;
} else {
_prof_factor = _expected_uses / prof_total;
}
}
CompileLog* log = C->log();
if (log != NULL) {
log->begin_head("parse method='%d' uses='%g'",
log->identify(parse_method), expected_uses);
if (depth() == 1 && C->is_osr_compilation()) {
log->print(" osr_bci='%d'", C->entry_bci());
}
log->stamp();
log->end_head();
}
// Accumulate deoptimization counts.
// (The range_check and store_check counts are checked elsewhere.)
ciMethodData* md = method()->method_data();
for (uint reason = 0; reason < md->trap_reason_limit(); reason++) {
uint md_count = md->trap_count(reason);
if (md_count != 0) {
if (md_count == md->trap_count_limit())
md_count += md->overflow_trap_count();
uint total_count = C->trap_count(reason);
uint old_count = total_count;
total_count += md_count;
// Saturate the add if it overflows.
if (total_count < old_count || total_count < md_count)
total_count = (uint)-1;
C->set_trap_count(reason, total_count);
if (log != NULL)
log->elem("observe trap='%s' count='%d' total='%d'",
Deoptimization::trap_reason_name(reason),
md_count, total_count);
}
}
// Accumulate total sum of decompilations, also.
C->set_decompile_count(C->decompile_count() + md->decompile_count());
_count_invocations = C->do_count_invocations();
_method_data_update = C->do_method_data_update();
if (log != NULL && method()->has_exception_handlers()) {
log->elem("observe that='has_exception_handlers'");
}
assert(method()->can_be_compiled(), "Can not parse this method, cutout earlier");
assert(method()->has_balanced_monitors(), "Can not parse unbalanced monitors, cutout earlier");
// Always register dependence if JVMTI is enabled, because
// either breakpoint setting or hotswapping of methods may
// cause deoptimization.
if (C->env()->jvmti_can_hotswap_or_post_breakpoint()) {
C->dependencies()->assert_evol_method(method());
}
methods_seen++;
// Do some special top-level things.
if (depth() == 1 && C->is_osr_compilation()) {
_entry_bci = C->entry_bci();
_flow = method()->get_osr_flow_analysis(osr_bci());
if (_flow->failing()) {
C->record_method_not_compilable(_flow->failure_reason());
#ifndef PRODUCT
if (PrintOpto && (Verbose || WizardMode)) {
tty->print_cr("OSR @%d type flow bailout: %s", _entry_bci, _flow->failure_reason());
if (Verbose) {
method()->print();
method()->print_codes();
_flow->print();
}
}
#endif
}
_tf = C->tf(); // the OSR entry type is different
}
#ifdef ASSERT
if (depth() == 1) {
assert(C->is_osr_compilation() == this->is_osr_parse(), "OSR in sync");
if (C->tf() != tf()) {
MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
assert(C->env()->system_dictionary_modification_counter_changed(),
"Must invalidate if TypeFuncs differ");
}
} else {
assert(!this->is_osr_parse(), "no recursive OSR");
}
#endif
methods_parsed++;
#ifndef PRODUCT
// add method size here to guarantee that inlined methods are added too
if (TimeCompiler)
_total_bytes_compiled += method()->code_size();
show_parse_info();
#endif
if (failing()) {
if (log) log->done("parse");
return;
}
gvn().set_type(root(), root()->bottom_type());
gvn().transform(top());
// Import the results of the ciTypeFlow.
init_blocks();
// Merge point for all normal exits
build_exits();
// Setup the initial JVM state map.
SafePointNode* entry_map = create_entry_map();
// Check for bailouts during map initialization
if (failing() || entry_map == NULL) {
if (log) log->done("parse");
return;
}
Node_Notes* caller_nn = C->default_node_notes();
// Collect debug info for inlined calls unless -XX:-DebugInlinedCalls.
if (DebugInlinedCalls || depth() == 1) {
C->set_default_node_notes(make_node_notes(caller_nn));
}
if (is_osr_parse()) {
Node* osr_buf = entry_map->in(TypeFunc::Parms+0);
entry_map->set_req(TypeFunc::Parms+0, top());
set_map(entry_map);
load_interpreter_state(osr_buf);
} else {
set_map(entry_map);
do_method_entry();
}
if (depth() == 1 && !failing()) {
// Add check to deoptimize the nmethod if RTM state was changed
rtm_deopt();
}
// Check for bailouts during method entry or RTM state check setup.
if (failing()) {
if (log) log->done("parse");
C->set_default_node_notes(caller_nn);
return;
}
entry_map = map(); // capture any changes performed by method setup code
assert(jvms()->endoff() == map()->req(), "map matches JVMS layout");
// We begin parsing as if we have just encountered a jump to the
// method entry.
Block* entry_block = start_block();
assert(entry_block->start() == (is_osr_parse() ? osr_bci() : 0), "");
set_map_clone(entry_map);
merge_common(entry_block, entry_block->next_path_num());
#ifndef PRODUCT
BytecodeParseHistogram *parse_histogram_obj = new (C->env()->arena()) BytecodeParseHistogram(this, C);
set_parse_histogram( parse_histogram_obj );
#endif
// Parse all the basic blocks.
do_all_blocks();
C->set_default_node_notes(caller_nn);
// Check for bailouts during conversion to graph
if (failing()) {
if (log) log->done("parse");
return;
}
// Fix up all exiting control flow.
set_map(entry_map);
do_exits();
if (log) log->done("parse nodes='%d' live='%d' memory='" SIZE_FORMAT "'",
C->unique(), C->live_nodes(), C->node_arena()->used());
}
//---------------------------do_all_blocks-------------------------------------
void Parse::do_all_blocks() {
bool has_irreducible = flow()->has_irreducible_entry();
// Walk over all blocks in Reverse Post-Order.
while (true) {
bool progress = false;
for (int rpo = 0; rpo < block_count(); rpo++) {
Block* block = rpo_at(rpo);
if (block->is_parsed()) continue;
if (!block->is_merged()) {
// Dead block, no state reaches this block
continue;
}
// Prepare to parse this block.
load_state_from(block);
if (stopped()) {
// Block is dead.
continue;
}
blocks_parsed++;
progress = true;
if (block->is_loop_head() || block->is_handler() || has_irreducible && !block->is_ready()) {
// Not all preds have been parsed. We must build phis everywhere.
// (Note that dead locals do not get phis built, ever.)
ensure_phis_everywhere();
if (block->is_SEL_head() &&
(UseLoopPredicate || LoopLimitCheck)) {
// Add predicate to single entry (not irreducible) loop head.
assert(!block->has_merged_backedge(), "only entry paths should be merged for now");
// Need correct bci for predicate.
// It is fine to set it here since do_one_block() will set it anyway.
set_parse_bci(block->start());
add_predicate();
// Add new region for back branches.
int edges = block->pred_count() - block->preds_parsed() + 1; // +1 for original region
RegionNode *r = new (C) RegionNode(edges+1);
_gvn.set_type(r, Type::CONTROL);
record_for_igvn(r);
r->init_req(edges, control());
set_control(r);
// Add new phis.
ensure_phis_everywhere();
}
// Leave behind an undisturbed copy of the map, for future merges.
set_map(clone_map());
}
if (control()->is_Region() && !block->is_loop_head() && !has_irreducible && !block->is_handler()) {
// In the absence of irreducible loops, the Region and Phis
// associated with a merge that doesn't involve a backedge can
// be simplified now since the RPO parsing order guarantees
// that any path which was supposed to reach here has already
// been parsed or must be dead.
Node* c = control();
Node* result = _gvn.transform_no_reclaim(control());
if (c != result && TraceOptoParse) {
tty->print_cr("Block #%d replace %d with %d", block->rpo(), c->_idx, result->_idx);
}
if (result != top()) {
record_for_igvn(result);
}
}
// Parse the block.
do_one_block();
// Check for bailouts.
if (failing()) return;
}
// with irreducible loops multiple passes might be necessary to parse everything
if (!has_irreducible || !progress) {
break;
}
}
blocks_seen += block_count();
#ifndef PRODUCT
// Make sure there are no half-processed blocks remaining.
// Every remaining unprocessed block is dead and may be ignored now.
for (int rpo = 0; rpo < block_count(); rpo++) {
Block* block = rpo_at(rpo);
if (!block->is_parsed()) {
if (TraceOptoParse) {
tty->print_cr("Skipped dead block %d at bci:%d", rpo, block->start());
}
assert(!block->is_merged(), "no half-processed blocks");
}
}
#endif
}
static Node* mask_int_value(Node* v, BasicType bt, PhaseGVN* gvn) {
Compile* C = gvn->C;
switch (bt) {
case T_BYTE:
v = gvn->transform(new (C) LShiftINode(v, gvn->intcon(24)));
v = gvn->transform(new (C) RShiftINode(v, gvn->intcon(24)));
break;
case T_SHORT:
v = gvn->transform(new (C) LShiftINode(v, gvn->intcon(16)));
v = gvn->transform(new (C) RShiftINode(v, gvn->intcon(16)));
break;
case T_CHAR:
v = gvn->transform(new (C) AndINode(v, gvn->intcon(0xFFFF)));
break;
case T_BOOLEAN:
v = gvn->transform(new (C) AndINode(v, gvn->intcon(0x1)));
break;
}
return v;
}
//-------------------------------build_exits----------------------------------
// Build normal and exceptional exit merge points.
void Parse::build_exits() {
// make a clone of caller to prevent sharing of side-effects
_exits.set_map(_exits.clone_map());
_exits.clean_stack(_exits.sp());
_exits.sync_jvms();
RegionNode* region = new (C) RegionNode(1);
record_for_igvn(region);
gvn().set_type_bottom(region);
_exits.set_control(region);
// Note: iophi and memphi are not transformed until do_exits.
Node* iophi = new (C) PhiNode(region, Type::ABIO);
Node* memphi = new (C) PhiNode(region, Type::MEMORY, TypePtr::BOTTOM);
gvn().set_type_bottom(iophi);
gvn().set_type_bottom(memphi);
_exits.set_i_o(iophi);
_exits.set_all_memory(memphi);
// Add a return value to the exit state. (Do not push it yet.)
if (tf()->range()->cnt() > TypeFunc::Parms) {
const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms);
if (ret_type->isa_int()) {
BasicType ret_bt = method()->return_type()->basic_type();
if (ret_bt == T_BOOLEAN ||
ret_bt == T_CHAR ||
ret_bt == T_BYTE ||
ret_bt == T_SHORT) {
ret_type = TypeInt::INT;
}
}
// Don't "bind" an unloaded return klass to the ret_phi. If the klass
// becomes loaded during the subsequent parsing, the loaded and unloaded
// types will not join when we transform and push in do_exits().
const TypeOopPtr* ret_oop_type = ret_type->isa_oopptr();
if (ret_oop_type && !ret_oop_type->klass()->is_loaded()) {
ret_type = TypeOopPtr::BOTTOM;
}
int ret_size = type2size[ret_type->basic_type()];
Node* ret_phi = new (C) PhiNode(region, ret_type);
gvn().set_type_bottom(ret_phi);
_exits.ensure_stack(ret_size);
assert((int)(tf()->range()->cnt() - TypeFunc::Parms) == ret_size, "good tf range");
assert(method()->return_type()->size() == ret_size, "tf agrees w/ method");
_exits.set_argument(0, ret_phi); // here is where the parser finds it
// Note: ret_phi is not yet pushed, until do_exits.
}
}
//----------------------------build_start_state-------------------------------
// Construct a state which contains only the incoming arguments from an
// unknown caller. The method & bci will be NULL & InvocationEntryBci.
JVMState* Compile::build_start_state(StartNode* start, const TypeFunc* tf) {
int arg_size = tf->domain()->cnt();
int max_size = MAX2(arg_size, (int)tf->range()->cnt());
JVMState* jvms = new (this) JVMState(max_size - TypeFunc::Parms);
SafePointNode* map = new (this) SafePointNode(max_size, NULL);
record_for_igvn(map);
assert(arg_size == TypeFunc::Parms + (is_osr_compilation() ? 1 : method()->arg_size()), "correct arg_size");
Node_Notes* old_nn = default_node_notes();
if (old_nn != NULL && has_method()) {
Node_Notes* entry_nn = old_nn->clone(this);
JVMState* entry_jvms = new(this) JVMState(method(), old_nn->jvms());
entry_jvms->set_offsets(0);
entry_jvms->set_bci(entry_bci());
entry_nn->set_jvms(entry_jvms);
set_default_node_notes(entry_nn);
}
uint i;
for (i = 0; i < (uint)arg_size; i++) {
Node* parm = initial_gvn()->transform(new (this) ParmNode(start, i));
map->init_req(i, parm);
// Record all these guys for later GVN.
record_for_igvn(parm);
}
for (; i < map->req(); i++) {
map->init_req(i, top());
}
assert(jvms->argoff() == TypeFunc::Parms, "parser gets arguments here");
set_default_node_notes(old_nn);
map->set_jvms(jvms);
jvms->set_map(map);
return jvms;
}
//-----------------------------make_node_notes---------------------------------
Node_Notes* Parse::make_node_notes(Node_Notes* caller_nn) {
if (caller_nn == NULL) return NULL;
Node_Notes* nn = caller_nn->clone(C);
JVMState* caller_jvms = nn->jvms();
JVMState* jvms = new (C) JVMState(method(), caller_jvms);
jvms->set_offsets(0);
jvms->set_bci(_entry_bci);
nn->set_jvms(jvms);
return nn;
}
//--------------------------return_values--------------------------------------
void Compile::return_values(JVMState* jvms) {
GraphKit kit(jvms);
Node* ret = new (this) ReturnNode(TypeFunc::Parms,
kit.control(),
kit.i_o(),
kit.reset_memory(),
kit.frameptr(),
kit.returnadr());
// Add zero or 1 return values
int ret_size = tf()->range()->cnt() - TypeFunc::Parms;
if (ret_size > 0) {
kit.inc_sp(-ret_size); // pop the return value(s)
kit.sync_jvms();
ret->add_req(kit.argument(0));
// Note: The second dummy edge is not needed by a ReturnNode.
}
// bind it to root
root()->add_req(ret);
record_for_igvn(ret);
initial_gvn()->transform_no_reclaim(ret);
}
//------------------------rethrow_exceptions-----------------------------------
// Bind all exception states in the list into a single RethrowNode.
void Compile::rethrow_exceptions(JVMState* jvms) {
GraphKit kit(jvms);
if (!kit.has_exceptions()) return; // nothing to generate
// Load my combined exception state into the kit, with all phis transformed:
SafePointNode* ex_map = kit.combine_and_pop_all_exception_states();
Node* ex_oop = kit.use_exception_state(ex_map);
RethrowNode* exit = new (this) RethrowNode(kit.control(),
kit.i_o(), kit.reset_memory(),
kit.frameptr(), kit.returnadr(),
// like a return but with exception input
ex_oop);
// bind to root
root()->add_req(exit);
record_for_igvn(exit);
initial_gvn()->transform_no_reclaim(exit);
}
//---------------------------do_exceptions-------------------------------------
// Process exceptions arising from the current bytecode.
// Send caught exceptions to the proper handler within this method.
// Unhandled exceptions feed into _exit.
void Parse::do_exceptions() {
if (!has_exceptions()) return;
if (failing()) {
// Pop them all off and throw them away.
while (pop_exception_state() != NULL) ;
return;
}
PreserveJVMState pjvms(this, false);
SafePointNode* ex_map;
while ((ex_map = pop_exception_state()) != NULL) {
if (!method()->has_exception_handlers()) {
// Common case: Transfer control outward.
// Doing it this early allows the exceptions to common up
// even between adjacent method calls.
throw_to_exit(ex_map);
} else {
// Have to look at the exception first.
assert(stopped(), "catch_inline_exceptions trashes the map");
catch_inline_exceptions(ex_map);
stop_and_kill_map(); // we used up this exception state; kill it
}
}
// We now return to our regularly scheduled program:
}
//---------------------------throw_to_exit-------------------------------------
// Merge the given map into an exception exit from this method.
// The exception exit will handle any unlocking of receiver.
// The ex_oop must be saved within the ex_map, unlike merge_exception.
void Parse::throw_to_exit(SafePointNode* ex_map) {
// Pop the JVMS to (a copy of) the caller.
GraphKit caller;
caller.set_map_clone(_caller->map());
caller.set_bci(_caller->bci());
caller.set_sp(_caller->sp());
// Copy out the standard machine state:
for (uint i = 0; i < TypeFunc::Parms; i++) {
caller.map()->set_req(i, ex_map->in(i));
}
if (ex_map->has_replaced_nodes()) {
_replaced_nodes_for_exceptions = true;
}
caller.map()->transfer_replaced_nodes_from(ex_map, _new_idx);
// ...and the exception:
Node* ex_oop = saved_ex_oop(ex_map);
SafePointNode* caller_ex_map = caller.make_exception_state(ex_oop);
// Finally, collect the new exception state in my exits:
_exits.add_exception_state(caller_ex_map);
}
//------------------------------do_exits---------------------------------------
void Parse::do_exits() {
set_parse_bci(InvocationEntryBci);
// Now peephole on the return bits
Node* region = _exits.control();
_exits.set_control(gvn().transform(region));
Node* iophi = _exits.i_o();
_exits.set_i_o(gvn().transform(iophi));
// On PPC64, also add MemBarRelease for constructors which write
// volatile fields. As support_IRIW_for_not_multiple_copy_atomic_cpu
// is set on PPC64, no sync instruction is issued after volatile
// stores. We want to quarantee the same behaviour as on platforms
// with total store order, although this is not required by the Java
// memory model. So as with finals, we add a barrier here.
if (wrote_final() PPC64_ONLY(|| (wrote_volatile() && method()->is_initializer()))) {
// This method (which must be a constructor by the rules of Java)
// wrote a final. The effects of all initializations must be
// committed to memory before any code after the constructor
// publishes the reference to the newly constructor object.
// Rather than wait for the publication, we simply block the
// writes here. Rather than put a barrier on only those writes
// which are required to complete, we force all writes to complete.
//
// "All bets are off" unless the first publication occurs after a
// normal return from the constructor. We do not attempt to detect
// such unusual early publications. But no barrier is needed on
// exceptional returns, since they cannot publish normally.
//
_exits.insert_mem_bar(Op_MemBarRelease, alloc_with_final());
#ifndef PRODUCT
if (PrintOpto && (Verbose || WizardMode)) {
method()->print_name();
tty->print_cr(" writes finals and needs a memory barrier");
}
#endif
}
for (MergeMemStream mms(_exits.merged_memory()); mms.next_non_empty(); ) {
// transform each slice of the original memphi:
mms.set_memory(_gvn.transform(mms.memory()));
}
if (tf()->range()->cnt() > TypeFunc::Parms) {
const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms);
Node* ret_phi = _gvn.transform( _exits.argument(0) );
if (!_exits.control()->is_top() && _gvn.type(ret_phi)->empty()) {
// In case of concurrent class loading, the type we set for the
// ret_phi in build_exits() may have been too optimistic and the
// ret_phi may be top now.
// Otherwise, we've encountered an error and have to mark the method as
// not compilable. Just using an assertion instead would be dangerous
// as this could lead to an infinite compile loop in non-debug builds.
{
MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag);
if (C->env()->system_dictionary_modification_counter_changed()) {
C->record_failure(C2Compiler::retry_class_loading_during_parsing());
} else {
C->record_method_not_compilable("Can't determine return type.");
}
}
return;
}
if (ret_type->isa_int()) {
BasicType ret_bt = method()->return_type()->basic_type();
ret_phi = mask_int_value(ret_phi, ret_bt, &_gvn);
}
_exits.push_node(ret_type->basic_type(), ret_phi);
}
// Note: Logic for creating and optimizing the ReturnNode is in Compile.
// Unlock along the exceptional paths.
// This is done late so that we can common up equivalent exceptions
// (e.g., null checks) arising from multiple points within this method.
// See GraphKit::add_exception_state, which performs the commoning.
bool do_synch = method()->is_synchronized() && GenerateSynchronizationCode;
// record exit from a method if compiled while Dtrace is turned on.
if (do_synch || C->env()->dtrace_method_probes() || _replaced_nodes_for_exceptions) {
// First move the exception list out of _exits:
GraphKit kit(_exits.transfer_exceptions_into_jvms());
SafePointNode* normal_map = kit.map(); // keep this guy safe
// Now re-collect the exceptions into _exits:
SafePointNode* ex_map;
while ((ex_map = kit.pop_exception_state()) != NULL) {
Node* ex_oop = kit.use_exception_state(ex_map);
// Force the exiting JVM state to have this method at InvocationEntryBci.
// The exiting JVM state is otherwise a copy of the calling JVMS.
JVMState* caller = kit.jvms();
JVMState* ex_jvms = caller->clone_shallow(C);
ex_jvms->set_map(kit.clone_map());
ex_jvms->map()->set_jvms(ex_jvms);
ex_jvms->set_bci( InvocationEntryBci);
kit.set_jvms(ex_jvms);
if (do_synch) {
// Add on the synchronized-method box/object combo
kit.map()->push_monitor(_synch_lock);
// Unlock!
kit.shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node());
}
if (C->env()->dtrace_method_probes()) {
kit.make_dtrace_method_exit(method());
}
if (_replaced_nodes_for_exceptions) {
kit.map()->apply_replaced_nodes(_new_idx);
}
// Done with exception-path processing.
ex_map = kit.make_exception_state(ex_oop);
assert(ex_jvms->same_calls_as(ex_map->jvms()), "sanity");
// Pop the last vestige of this method:
ex_map->set_jvms(caller->clone_shallow(C));
ex_map->jvms()->set_map(ex_map);
_exits.push_exception_state(ex_map);
}
assert(_exits.map() == normal_map, "keep the same return state");
}
{
// Capture very early exceptions (receiver null checks) from caller JVMS
GraphKit caller(_caller);
SafePointNode* ex_map;
while ((ex_map = caller.pop_exception_state()) != NULL) {
_exits.add_exception_state(ex_map);
}
}
_exits.map()->apply_replaced_nodes(_new_idx);
}
//-----------------------------create_entry_map-------------------------------
// Initialize our parser map to contain the types at method entry.
// For OSR, the map contains a single RawPtr parameter.
// Initial monitor locking for sync. methods is performed by do_method_entry.
SafePointNode* Parse::create_entry_map() {
// Check for really stupid bail-out cases.
uint len = TypeFunc::Parms + method()->max_locals() + method()->max_stack();
if (len >= 32760) {
C->record_method_not_compilable_all_tiers("too many local variables");
return NULL;
}
// clear current replaced nodes that are of no use from here on (map was cloned in build_exits).
_caller->map()->delete_replaced_nodes();
// If this is an inlined method, we may have to do a receiver null check.
if (_caller->has_method() && is_normal_parse() && !method()->is_static()) {
GraphKit kit(_caller);
kit.null_check_receiver_before_call(method());
_caller = kit.transfer_exceptions_into_jvms();
if (kit.stopped()) {
_exits.add_exception_states_from(_caller);
_exits.set_jvms(_caller);
return NULL;
}
}
assert(method() != NULL, "parser must have a method");
// Create an initial safepoint to hold JVM state during parsing
JVMState* jvms = new (C) JVMState(method(), _caller->has_method() ? _caller : NULL);
set_map(new (C) SafePointNode(len, jvms));
jvms->set_map(map());
record_for_igvn(map());
assert(jvms->endoff() == len, "correct jvms sizing");
SafePointNode* inmap = _caller->map();
assert(inmap != NULL, "must have inmap");
// In case of null check on receiver above
map()->transfer_replaced_nodes_from(inmap, _new_idx);
uint i;
// Pass thru the predefined input parameters.
for (i = 0; i < TypeFunc::Parms; i++) {
map()->init_req(i, inmap->in(i));
}
08-26
10-28
1万+
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
