CoreCLR源码
前一篇我们看到了CoreCLR中对Object的定义，这一篇我们将会看CoreCLR中对new的定义和处理
new对于.Net程序员们来说同样是耳熟能详的关键词，我们每天都会用到new，然而new究竟是什么？
因为篇幅限制和避免难度跳的太高，这一篇将不会详细讲解以下的内容，请耐心等待后续的文章
- GC如何分配内存
- JIT如何解析IL
- JIT如何生成机器码
使用到的名词和缩写
以下的内容将会使用到一些名词和缩写，如果碰到看不懂的可以到这里来对照
BasicBlock: 在同一个分支(Branch)的一群指令，使用双向链表连接
GenTree: 语句树，节点类型以GT开头
Importation: 从BasicBlock生成GenTree的过程
Lowering: 具体化语句树，让语句树的各个节点可以明确的转换到机器码
SSA: Static Single Assignment
R2R: Ready To Run
Phases: JIT编译IL到机器码经过的各个阶段
JIT: Just In Time
CEE: CLR Execute Engine
ee: Execute Engine
EH: Exception Handling
Cor: CoreCLR
comp: Compiler
fg: FlowGraph
imp: Import
LDLOCA: Load Local Variable
gt: Generate
hlp: Help
Ftn: Function
MP: Multi Process
CER: Constrained Execution Regions
TLS: Thread Local Storage.Net中的三种new
请看图中的代码和生成的IL，我们可以看到尽管同样是new，却生成了三种不同的IL代码
- 对class的new，IL指令是newobj
- 对array的new，IL指令是newarr
- 对struct的new，因为myStruct已经在本地变量里面了，new的时候仅仅是调用ldloca加载然后调用构造函数
我们先来看newobj和newarr这两个指令在coreclr中是怎么定义的
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/opcode.def#L153
OPDEF(CEE_NEWOBJ, "newobj", VarPop, PushRef, InlineMethod, IObjModel, 1, 0xFF, 0x73, CALL)
OPDEF(CEE_NEWARR, "newarr", PopI, PushRef, InlineType, IObjModel, 1, 0xFF, 0x8D, NEXT)我们可以看到这两个指令的定义，名称分别是CEE_NEWOBJ和CEE_NEWARR，请记住这两个名称
第一种new(对class的new)生成了什么机器码
接下来我们将看看coreclr是如何把CEE_NEWOBJ指令变为机器码的
在讲解之前请先大概了解JIT的工作流程，JIT编译按函数为单位，当调用函数时会自动触发JIT编译
- 把函数的IL转换为BasicBlock(基本代码块)
- 从BasicBlock(基本代码块)生成GenTree(语句树)
- 对GenTree(语句树)进行Morph(变形)
- 对GenTree(语句树)进行Lowering(具体化)
- 根据GenTree(语句树)生成机器码
下面的代码虽然进过努力的提取，但仍然比较长，请耐心阅读
我们从JIT的入口函数开始看，这个函数会被EE(运行引擎)调用
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/corjit.h#L350
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/ee_il_dll.cpp#L279
注: 按微软文档中说CILJit是32位上的实现，PreJit是64位上的实现，但实际我找不到PreJit在哪里
CorJitResult CILJit::compileMethod(
ICorJitInfo* compHnd, CORINFO_METHOD_INFO* methodInfo, unsigned flags, BYTE** entryAddress, ULONG* nativeSizeOfCode)
{
// 省略部分代码......
assert(methodInfo->ILCode);
result = jitNativeCode(methodHandle, methodInfo->scope, compHnd, methodInfo, &methodCodePtr, nativeSizeOfCode,
&jitFlags, nullptr);
// 省略部分代码......
return CorJitResult(result);
}jitNativeCode是一个负责使用JIT编译单个函数的静态函数，会在内部为编译的函数创建单独的Compiler实例
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.cpp#L6075
int jitNativeCode(CORINFO_METHOD_HANDLE methodHnd,
CORINFO_MODULE_HANDLE classPtr,
COMP_HANDLE compHnd,
CORINFO_METHOD_INFO* methodInfo,
void** methodCodePtr,
ULONG* methodCodeSize,
JitFlags* compileFlags,
void* inlineInfoPtr)
{
// 省略部分代码......
pParam->pComp->compInit(pParam->pAlloc, pParam->inlineInfo);
pParam->pComp->jitFallbackCompile = pParam->jitFallbackCompile;
// Now generate the code
pParam->result =
pParam->pComp->compCompile(pParam->methodHnd, pParam->classPtr, pParam->compHnd, pParam->methodInfo,
pParam->methodCodePtr, pParam->methodCodeSize, pParam->compileFlags);
// 省略部分代码......
return result;
}Compiler::compCompile是Compiler类提供的入口函数，作用同样是编译函数
注意这个函数有7个参数，等一会还会有一个同名但只有3个参数的函数
这个函数主要调用了Compiler::compCompileHelper函数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.cpp#L4693
int Compiler::compCompile(CORINFO_METHOD_HANDLE methodHnd,
CORINFO_MODULE_HANDLE classPtr,
COMP_HANDLE compHnd,
CORINFO_METHOD_INFO* methodInfo,
void** methodCodePtr,
ULONG* methodCodeSize,
JitFlags* compileFlags)
{
// 省略部分代码......
pParam->result = pParam->pThis->compCompileHelper(pParam->classPtr, pParam->compHnd, pParam->methodInfo,
pParam->methodCodePtr, pParam->methodCodeSize,
pParam->compileFlags, pParam->instVerInfo);
// 省略部分代码......
return param.result;
}让我们继续看Compiler::compCompileHelper
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.cpp#L5294
int Compiler::compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
COMP_HANDLE compHnd,
CORINFO_METHOD_INFO* methodInfo,
void** methodCodePtr,
ULONG* methodCodeSize,
JitFlags* compileFlags,
CorInfoInstantiationVerification instVerInfo)
{
// 省略部分代码......
// 初始化本地变量表
lvaInitTypeRef();
// 省略部分代码......
// 查找所有BasicBlock
fgFindBasicBlocks();
// 省略部分代码......
// 调用3个参数的compCompile函数，注意不是7个函数的compCompile函数
compCompile(methodCodePtr, methodCodeSize, compileFlags);
// 省略部分代码......
return CORJIT_OK;
}现在到了3个参数的compCompile，这个函数被微软认为是JIT最被感兴趣的入口函数
你可以额外阅读一下微软的JIT介绍文档
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.cpp#L4078
//*********************************************************************************************
// #Phases
//
// This is the most interesting 'toplevel' function in the JIT. It goes through the operations of
// importing, morphing, optimizations and code generation. This is called from the EE through the
// code:CILJit::compileMethod function.
//
// For an overview of the structure of the JIT, see:
// https://github.com/dotnet/coreclr/blob/master/Documentation/botr/ryujit-overview.md
//
void Compiler::compCompile(void** methodCodePtr, ULONG* methodCodeSize, JitFlags* compileFlags)
{
// 省略部分代码......
// 转换BasicBlock(基本代码块)到GenTree(语句树)
fgImport();
// 省略部分代码......
// 这里会进行各个处理步骤(Phases)，如Inline和优化等
// 省略部分代码......
// 转换GT_ALLOCOBJ节点到GT_CALL节点(分配内存=调用帮助函数)
ObjectAllocator objectAllocator(this);
objectAllocator.Run();
// 省略部分代码......
// 创建本地变量表和计算各个变量的引用计数
lvaMarkLocalVars();
// 省略部分代码......
// 具体化语句树
Lowering lower(this, m_pLinearScan); // PHASE_LOWERING
lower.Run();
// 省略部分代码......
// 生成机器码
codeGen->genGenerateCode(methodCodePtr, methodCodeSize);
}到这里你应该大概知道JIT在总体上做了什么事情
接下来我们来看Compiler::fgImport函数，这个函数负责把BasicBlock(基本代码块)转换到GenTree(语句树)
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/flowgraph.cpp#L6663
void Compiler::fgImport()
{
// 省略部分代码......
impImport(fgFirstBB);
// 省略部分代码......
}再看Compiler::impImport
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/importer.cpp#L9207
void Compiler::impImport(BasicBlock* method)
{
// 省略部分代码......
/* Import blocks in the worker-list until there are no more */
while (impPendingList)
{
PendingDsc* dsc = impPendingList;
impPendingList = impPendingList->pdNext;
// 省略部分代码......
/* Now import the block */
impImportBlock(dsc->pdBB);
}
}再看Compiler::impImportBlock
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/importer.cpp#L15321
//***************************************************************
// Import the instructions for the given basic block. Perform
// verification, throwing an exception on failure. Push any successor blocks that are enabled for the first
// time, or whose verification pre-state is changed.
void Compiler::impImportBlock(BasicBlock* block)
{
// 省略部分代码......
pParam->pThis->impImportBlockCode(pParam->block);
}在接下来的Compiler::impImportBlockCode函数里面我们终于可以看到对CEE_NEWOBJ指令的处理了
这个函数有5000多行，推荐直接搜索case CEE_NEWOBJ来看以下的部分
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/importer.cpp#L9207
/*****************************************************************************
* Import the instr for the given basic block
*/
void Compiler::impImportBlockCode(BasicBlock* block)
{
// 省略部分代码......
// 处理CEE_NEWOBJ指令
case CEE_NEWOBJ:
// 在这里微软给出了有三种情况
// 一种是对象是array，一种是对象有活动的长度（例如string），一种是普通的class
// 在这里我们只分析第三种情况
// There are three different cases for new
// Object size is variable (depends on arguments)
// 1) Object is an array (arrays treated specially by the EE)
// 2) Object is some other variable sized object (e.g. String)
// 3) Class Size can be determined beforehand (normal case)
// In the first case, we need to call a NEWOBJ helper (multinewarray)
// in the second case we call the constructor with a '0' this pointer
// In the third case we alloc the memory, then call the constuctor
// 省略部分代码......
// 创建一个GT_ALLOCOBJ类型的GenTree(语句树)节点，用于分配内存
op1 = gtNewAllocObjNode(info.compCompHnd->getNewHelper(&resolvedToken, info.compMethodHnd),
resolvedToken.hClass, TYP_REF, op1);
// 省略部分代码......
// 因为GT_ALLOCOBJ仅负责分配内存，我们还需要调用构造函数
// 这里复用了CEE_CALL指令的处理
goto CALL;
// 省略部分代码......
CALL: // memberRef should be set.
// 省略部分代码......
// 创建一个GT_CALL类型的GenTree(语句树)节点，用于调用构造函数
callTyp = impImportCall(opcode, &resolvedToken, constraintCall ? &constrainedResolvedToken : nullptr,
newObjThisPtr, prefixFlags, &callInfo, opcodeOffs);请记住上面代码中新建的两个GenTree(语句树)节点
- 节点GT_ALLOCOBJ用于分配内存
- 节点GT_CALL用于调用构造函数
在上面的代码我们可以看到在生成GT_ALLOCOBJ类型的节点时还传入了一个newHelper参数，这个newHelper正是分配内存函数的一个标识(索引值)
在CoreCLR中有很多HelperFunc(帮助函数)供JIT生成的代码调用
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp#L5894
CorInfoHelpFunc CEEInfo::getNewHelper(CORINFO_RESOLVED_TOKEN * pResolvedToken, CORINFO_METHOD_HANDLE callerHandle)
{
// 省略部分代码......
MethodTable* pMT = VMClsHnd.AsMethodTable();
// 省略部分代码......
result = getNewHelperStatic(pMT);
// 省略部分代码......
return result;
}看CEEInfo::getNewHelperStatic
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp#L5941
CorInfoHelpFunc CEEInfo::getNewHelperStatic(MethodTable * pMT)
{
// 省略部分代码......
// 这里有很多判断，例如是否是Com对象或拥有析构函数，默认会返回CORINFO_HELP_NEWFAST
// Slow helper is the default
CorInfoHelpFunc helper = CORINFO_HELP_NEWFAST;
// 省略部分代码......
return helper;
}到这里，我们可以知道新建的两个节点带有以下的信息
- GT_ALLOCOBJ节点
- 分配内存的帮助函数标识，默认是CORINFO_HELP_NEWFAST
- GT_CALL节点
- 构造函数的句柄
在使用fgImport生成了GenTree(语句树)以后，还不能直接用这个树来生成机器代码，需要经过很多步的变换
其中的一步变换会把GT_ALLOCOBJ节点转换为GT_CALL节点，因为分配内存实际上是一个对JIT专用的帮助函数的调用
这个变换在ObjectAllocator中实现，ObjectAllocator是JIT编译过程中的一个阶段(Phase)
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/objectalloc.cpp#L27
void ObjectAllocator::DoPhase()
{
// 省略部分代码......
MorphAllocObjNodes();
}MorphAllocObjNodes用于查找所有节点，如果是GT_ALLOCOBJ则进行转换
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/objectalloc.cpp#L63
void ObjectAllocator::MorphAllocObjNodes()
{
// 省略部分代码......
for (GenTreeStmt* stmt = block->firstStmt(); stmt; stmt = stmt->gtNextStmt)
{
// 省略部分代码......
bool canonicalAllocObjFound = false;
// 省略部分代码......
if (op2->OperGet() == GT_ALLOCOBJ)
canonicalAllocObjFound = true;
// 省略部分代码......
if (canonicalAllocObjFound)
{
// 省略部分代码......
op2 = MorphAllocObjNodeIntoHelperCall(asAllocObj);
}
}
}MorphAllocObjNodeIntoHelperCall的定义
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/objectalloc.cpp#L152
// MorphAllocObjNodeIntoHelperCall: Morph a GT_ALLOCOBJ node into an
// allocation helper call.
GenTreePtr ObjectAllocator::MorphAllocObjNodeIntoHelperCall(GenTreeAllocObj* allocObj)
{
// 省略部分代码......
GenTreePtr helperCall = comp->fgMorphIntoHelperCall(allocObj, allocObj->gtNewHelper, comp->gtNewArgList(op1));
return helperCall;
}fgMorphIntoHelperCall的定义
这个函数转换GT_ALLOCOBJ节点到GT_CALL节点，并且获取指向分配内存的函数的指针
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/morph.cpp#L61
GenTreePtr Compiler::fgMorphIntoHelperCall(GenTreePtr tree, int helper, GenTreeArgList* args)
{
tree->ChangeOper(GT_CALL);
tree->gtFlags |= GTF_CALL;
// 省略部分代码......
// 如果GT_ALLOCOBJ中帮助函数的标识是CORINFO_HELP_NEWFAST，这里就是eeFindHelper(CORINFO_HELP_NEWFAST)
// eeFindHelper会把帮助函数的表示转换为帮助函数的句柄
tree->gtCall.gtCallType = CT_HELPER;
tree->gtCall.gtCallMethHnd = eeFindHelper(helper);
// 省略部分代码......
tree = fgMorphArgs(tree->AsCall());
return tree;
}到这里，我们可以知道新建的两个节点变成了这样
- GT_CALL节点 (调用帮助函数)
- 分配内存的帮助函数的句柄
- GT_CALL节点 (调用Managed函数)
- 构造函数的句柄
接下来JIT还会对GenTree(语句树)做出大量处理，这里省略说明，接下来我们来看机器码的生成
函数CodeGen::genCallInstruction负责把GT_CALL节点转换为汇编
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/codegenxarch.cpp#L5934
// Produce code for a GT_CALL node
void CodeGen::genCallInstruction(GenTreePtr node)
{
// 省略部分代码......
if (callType == CT_HELPER)
{
// 把句柄转换为帮助函数的句柄，默认是CORINFO_HELP_NEWFAST
helperNum = compiler->eeGetHelperNum(methHnd);
// 获取指向帮助函数的指针
// 这里等于调用compiler->compGetHelperFtn(CORINFO_HELP_NEWFAST, ...)
addr = compiler->compGetHelperFtn(helperNum, (void**)&pAddr);
}
else
{
// 调用普通函数
// Direct call to a non-virtual user function.
addr = call->gtDirectCallAddress;
}
}我们来看下compGetHelperFtn究竟把CORINFO_HELP_NEWFAST转换到了什么函数
compGetHelperFtn的定义
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.hpp#L1907
void* Compiler::compGetHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
void** ppIndirection) /* OUT */
{
// 省略部分代码......
addr = info.compCompHnd->getHelperFtn(ftnNum, ppIndirection);
return addr;
}getHelperFtn的定义
这里我们可以看到获取了hlpDynamicFuncTable这个函数表中的函数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp#L10369
void* CEEJitInfo::getHelperFtn(CorInfoHelpFunc ftnNum, /* IN */
void ** ppIndirection) /* OUT */
{
// 省略部分代码......
pfnHelper = hlpDynamicFuncTable[dynamicFtnNum].pfnHelper;
// 省略部分代码......
result = (LPVOID)GetEEFuncEntryPoint(pfnHelper);
return result;
}hlpDynamicFuncTable函数表使用了jithelpers.h中的定义，其中CORINFO_HELP_NEWFAST对应的函数如下
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/jithelpers.h#L78
JITHELPER(CORINFO_HELP_NEWFAST, JIT_New, CORINFO_HELP_SIG_REG_ONLY)可以看到对应了JIT_New，这个就是JIT生成的代码调用分配内存的函数了，JIT_New的定义如下
需要注意的是函数表中的JIT_New在满足一定条件时会被替换为更快的实现，但作用和JIT_New是一样的，这一块将在后面提及
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jithelpers.cpp#L2908
HCIMPL1(Object*, JIT_New, CORINFO_CLASS_HANDLE typeHnd_)
{
// 省略部分代码......
MethodTable *pMT = typeHnd.AsMethodTable();
// 省略部分代码......
// AllocateObject是分配内存的函数，这个函数供CoreCLR的内部代码或非托管代码调用
// JIT_New是对这个函数的一个包装，仅供JIT生成的代码调用
newobj = AllocateObject(pMT);
// 省略部分代码......
return(OBJECTREFToObject(newobj));
}
HCIMPLEND总结:
JIT从CEE_NEWOBJ生成了两段代码，一段是调用JIT_New函数分配内存的代码，一段是调用构造函数的代码
第二种new(对array的new)生成了什么机器码
我们来看一下CEE_NEWARR指令是怎样处理的，因为前面已经花了很大篇幅介绍对CEE_NEWOBJ的处理，这里仅列出不同的部分
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/importer.cpp#L13334
/*****************************************************************************
* Import the instr for the given basic block
*/
void Compiler::impImportBlockCode(BasicBlock* block)
{
// 省略部分代码......
// 处理CEE_NEWARR指令
case CEE_NEWARR:
// 省略部分代码......
args = gtNewArgList(op1, op2);
// 生成GT_CALL类型的节点调用帮助函数
/* Create a call to 'new' */
// Note that this only works for shared generic code because the same helper is used for all
// reference array types
op1 = gtNewHelperCallNode(info.compCompHnd->getNewArrHelper(resolvedToken.hClass), TYP_REF, 0, args);
}我们可以看到CEE_NEWARR直接生成了GT_CALL节点，不像CEE_NEWOBJ需要进一步的转换
getNewArrHelper返回了调用的帮助函数，我们来看一下getNewArrHelper
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp#L6035
/***********************************************************************/
// <REVIEW> this only works for shared generic code because all the
// helpers are actually the same. If they were different then things might
// break because the same helper would end up getting used for different but
// representation-compatible arrays (e.g. one with a default constructor
// and one without) </REVIEW>
CorInfoHelpFunc CEEInfo::getNewArrHelper (CORINFO_CLASS_HANDLE arrayClsHnd)
{
// 省略部分代码......
TypeHandle arrayType(arrayClsHnd);
result = getNewArrHelperStatic(arrayType);
// 省略部分代码......
return result;
}再看getNewArrHelperStatic，我们可以看到一般情况下会返回CORINFO_HELP_NEWARR_1_OBJ
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp#L6060
CorInfoHelpFunc CEEInfo::getNewArrHelperStatic(TypeHandle clsHnd)
{
// 省略部分代码......
if (CorTypeInfo::IsGenericVariable(elemType))
{
result = CORINFO_HELP_NEWARR_1_OBJ;
}
else if (CorTypeInfo::IsObjRef(elemType))
{
// It is an array of object refs
result = CORINFO_HELP_NEWARR_1_OBJ;
}
else
{
// These cases always must use the slow helper
// 省略部分代码......
}
return result;
{CORINFO_HELP_NEWARR_1_OBJ对应的函数如下
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/jithelpers.h#L86
DYNAMICJITHELPER(CORINFO_HELP_NEWARR_1_OBJ, JIT_NewArr1,CORINFO_HELP_SIG_REG_ONLY)可以看到对应了JIT_NewArr1这个包装给JIT调用的帮助函数
和JIT_New一样，在满足一定条件时会被替换为更快的实现
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jithelpers.cpp#L3303
HCIMPL2(Object*, JIT_NewArr1, CORINFO_CLASS_HANDLE arrayTypeHnd_, INT_PTR size)
{
// 省略部分代码......
CorElementType elemType = pArrayClassRef->GetArrayElementTypeHandle().GetSignatureCorElementType();
if (CorTypeInfo::IsPrimitiveType(elemType)
{
// 省略部分代码......
// 如果类型是基元类型(int, double等)则使用更快的FastAllocatePrimitiveArray函数
newArray = FastAllocatePrimitiveArray(pArrayClassRef->GetMethodTable(), static_cast<DWORD>(size), bAllocateInLargeHeap);
}
else
{
// 省略部分代码......
// 默认使用AllocateArrayEx函数
INT32 size32 = (INT32)size;
newArray = AllocateArrayEx(typeHnd, &size32, 1);
}
// 省略部分代码......
return(OBJECTREFToObject(newArray));
}
HCIMPLEND总结:
JIT从CEE_NEWARR只生成了一段代码，就是调用JIT_NewArr1函数的代码
第三种new(对struct的new)生成了什么机器码
这种new会在栈(stack)分配内存，所以不需要调用任何分配内存的函数
在一开始的例子中，myStruct在编译时就已经定义为一个本地变量，对本地变量的需要的内存会在函数刚进入的时候一并分配
这里我们先来看本地变量所需要的内存是怎么计算的
先看Compiler::lvaAssignVirtualFrameOffsetsToLocals
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/lclvars.cpp#L4863
/*****************************************************************************
* lvaAssignVirtualFrameOffsetsToLocals() : Assign virtual stack offsets to
* locals, temps, and anything else. These will all be negative offsets
* (stack grows down) relative to the virtual '0'/return address
*/
void Compiler::lvaAssignVirtualFrameOffsetsToLocals()
{
// 省略部分代码......
for (cur = 0; alloc_order[cur]; cur++)
{
// 省略部分代码......
for (lclNum = 0, varDsc = lvaTable; lclNum < lvaCount; lclNum++, varDsc++)
{
// 省略部分代码......
// Reserve the stack space for this variable
stkOffs = lvaAllocLocalAndSetVirtualOffset(lclNum, lvaLclSize(lclNum), stkOffs);
}
}
}再看Compiler::lvaAllocLocalAndSetVirtualOffset
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/lclvars.cpp#L5537
int Compiler::lvaAllocLocalAndSetVirtualOffset(unsigned lclNum, unsigned size, int stkOffs)
{
// 省略部分代码......
/* Reserve space on the stack by bumping the frame size */
lvaIncrementFrameSize(size);
stkOffs -= size;
lvaTable[lclNum].lvStkOffs = stkOffs;
// 省略部分代码......
return stkOffs;
}再看Compiler::lvaIncrementFrameSize
我们可以看到最终会加到compLclFrameSize这个变量中，这个变量就是当前函数总共需要在栈(Stack)分配的内存大小
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/lclvars.cpp#L3528
inline void Compiler::lvaIncrementFrameSize(unsigned size)
{
if (size > MAX_FrameSize || compLclFrameSize + size > MAX_FrameSize)
{
BADCODE("Frame size overflow");
}
compLclFrameSize += size;
}现在来看生成机器码的代码，在栈分配内存的代码会在CodeGen::genFnProlog生成
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/codegencommon.cpp#L8140
void CodeGen::genFnProlog()
{
// 省略部分代码......
// ARM64和其他平台的调用时机不一样，但是参数一样
genAllocLclFrame(compiler->compLclFrameSize, initReg, &initRegZeroed, intRegState.rsCalleeRegArgMaskLiveIn);
}再看CodeGen::genAllocLclFrame，这里就是分配栈内存的代码了，简单的rsp(esp)减去了frameSize
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/codegencommon.cpp#L5846
/*-----------------------------------------------------------------------------
*
* Probe the stack and allocate the local stack frame: subtract from SP.
* On ARM64, this only does the probing; allocating the frame is done when callee-saved registers are saved.
*/
void CodeGen::genAllocLclFrame(unsigned frameSize, regNumber initReg, bool* pInitRegZeroed, regMaskTP maskArgRegsLiveIn)
{
// 省略部分代码......
// sub esp, frameSize 6
inst_RV_IV(INS_sub, REG_SPBASE, frameSize, EA_PTRSIZE);
}总结:
JIT对struct的new会生成统一在栈分配内存的代码，所以你在IL中看不到new struct的指令
调用构造函数的代码会从后面的call指令生成
第一种new(对class的new)做了什么
从上面的分析我们可以知道第一种new先调用JIT_New分配内存，然后调用构造函数
在上面JIT_New的源代码中可以看到，JIT_New内部调用了AllocateObject
先看AllocateObject函数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/gchelpers.cpp#L931
// AllocateObject will throw OutOfMemoryException so don't need to check
// for NULL return value from it.
OBJECTREF AllocateObject(MethodTable *pMT
#ifdef FEATURE_COMINTEROP
, bool fHandleCom
#endif
)
{
// 省略部分代码......
Object *orObject = NULL;
// 如果类型有重要的析构函数，预编译所有相关的函数(详细可以搜索CER)
// 同一个类型只会处理一次
if (pMT->HasCriticalFinalizer())
PrepareCriticalFinalizerObject(pMT);
// 省略部分代码......
DWORD baseSize = pMT->GetBaseSize();
// 调用gc的帮助函数分配内存，如果需要向8对齐则调用AllocAlign8，否则调用Alloc
if (pMT->RequiresAlign8())
{
// 省略部分代码......
orObject = (Object *) AllocAlign8(baseSize,
pMT->HasFinalizer(),
pMT->ContainsPointers(),
pMT->IsValueType());
}
else
{
orObject = (Object *) Alloc(baseSize,
pMT->HasFinalizer(),
pMT->ContainsPointers());
}
// 检查同步块索引(SyncBlock)是否为0
// verify zero'd memory (at least for sync block)
_ASSERTE( orObject->HasEmptySyncBlockInfo() );
// 设置类型信息(MethodTable)
if ((baseSize >= LARGE_OBJECT_SIZE))
{
orObject->SetMethodTableForLargeObject(pMT);
GCHeap::GetGCHeap()->PublishObject((BYTE*)orObject);
}
else
{
orObject->SetMethodTable(pMT);
}
// 省略部分代码......
return UNCHECKED_OBJECTREF_TO_OBJECTREF(oref);
}再看Alloc函数
源代码:
// There are only three ways to get into allocate an object.
// * Call optimized helpers that were generated on the fly. This is how JIT compiled code does most
// allocations, however they fall back code:Alloc, when for all but the most common code paths. These
// helpers are NOT used if profiler has asked to track GC allocation (see code:TrackAllocations)
// * Call code:Alloc - When the jit helpers fall back, or we do allocations within the runtime code
// itself, we ultimately call here.
// * Call code:AllocLHeap - Used very rarely to force allocation to be on the large object heap.
//
// While this is a choke point into allocating an object, it is primitive (it does not want to know about
// MethodTable and thus does not initialize that poitner. It also does not know if the object is finalizable
// or contains pointers. Thus we quickly wrap this function in more user-friendly ones that know about
// MethodTables etc. (see code:FastAllocatePrimitiveArray code:AllocateArrayEx code:AllocateObject)
//
// You can get an exhaustive list of code sites that allocate GC objects by finding all calls to
// code:ProfilerObjectAllocatedCallback (since the profiler has to hook them all).
inline Object* Alloc(size_t size, BOOL bFinalize, BOOL bContainsPointers )
{
// 省略部分代码......
// We don't want to throw an SO during the GC, so make sure we have plenty
// of stack before calling in.
INTERIOR_STACK_PROBE_FOR(GetThread(), static_cast<unsigned>(DEFAULT_ENTRY_PROBE_AMOUNT * 1.5));
if (GCHeapUtilities::UseAllocationContexts())
retVal = GCHeapUtilities::GetGCHeap()->Alloc(GetThreadAllocContext(), size, flags);
else
retVal = GCHeapUtilities::GetGCHeap()->Alloc(size, flags);
if (!retVal)
{
ThrowOutOfMemory();
}
END_INTERIOR_STACK_PROBE;
return retVal;
}总结:
第一种new做的事情主要有
- 调用JIT_New
- 从GCHeap中申请一块内存
- 设置类型信息(MethodTable)
- 同步块索引默认为0，不需要设置
- 调用构造函数
第二种new(对array的new)做了什么
第二种new只调用了JIT_NewArr1，从上面JIT_NewArr1的源代码可以看到
如果元素的类型是基元类型(int, double等)则会调用FastAllocatePrimitiveArray，否则会调用AllocateArrayEx
先看FastAllocatePrimitiveArray函数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/gchelpers.cpp#L563
/*
* Allocates a single dimensional array of primitive types.
*/
OBJECTREF FastAllocatePrimitiveArray(MethodTable* pMT, DWORD cElements, BOOL bAllocateInLargeHeap)
{
// 省略部分代码......
// 检查元素数量不能大于一个硬性限制
SIZE_T componentSize = pMT->GetComponentSize();
if (cElements > MaxArrayLength(componentSize))
ThrowOutOfMemory();
// 检查总大小不能溢出
S_SIZE_T safeTotalSize = S_SIZE_T(cElements) * S_SIZE_T(componentSize) + S_SIZE_T(pMT->GetBaseSize());
if (safeTotalSize.IsOverflow())
ThrowOutOfMemory();
size_t totalSize = safeTotalSize.Value();
// 省略部分代码......
// 调用gc的帮助函数分配内存
ArrayBase* orObject;
if (bAllocateInLargeHeap)
{
orObject = (ArrayBase*) AllocLHeap(totalSize, FALSE, FALSE);
}
else
{
ArrayTypeDesc *pArrayR8TypeDesc = g_pPredefinedArrayTypes[ELEMENT_TYPE_R8];
if (DATA_ALIGNMENT < sizeof(double) && pArrayR8TypeDesc != NULL && pMT == pArrayR8TypeDesc->GetMethodTable() && totalSize < LARGE_OBJECT_SIZE - MIN_OBJECT_SIZE)
{
// Creation of an array of doubles, not in the large object heap.
// We want to align the doubles to 8 byte boundaries, but the GC gives us pointers aligned
// to 4 bytes only (on 32 bit platforms). To align, we ask for 12 bytes more to fill with a
// dummy object.
// If the GC gives us a 8 byte aligned address, we use it for the array and place the dummy
// object after the array, otherwise we put the dummy object first, shifting the base of
// the array to an 8 byte aligned address.
// Note: on 64 bit platforms, the GC always returns 8 byte aligned addresses, and we don't
// execute this code because DATA_ALIGNMENT < sizeof(double) is false.
_ASSERTE(DATA_ALIGNMENT == sizeof(double)/2);
_ASSERTE((MIN_OBJECT_SIZE % sizeof(double)) == DATA_ALIGNMENT); // used to change alignment
_ASSERTE(pMT->GetComponentSize() == sizeof(double));
_ASSERTE(g_pObjectClass->GetBaseSize() == MIN_OBJECT_SIZE);
_ASSERTE(totalSize < totalSize + MIN_OBJECT_SIZE);
orObject = (ArrayBase*) Alloc(totalSize + MIN_OBJECT_SIZE, FALSE, FALSE);
Object *orDummyObject;
if((size_t)orObject % sizeof(double))
{
orDummyObject = orObject;
orObject = (ArrayBase*) ((size_t)orObject + MIN_OBJECT_SIZE);
}
else
{
orDummyObject = (Object*) ((size_t)orObject + totalSize);
}
_ASSERTE(((size_t)orObject % sizeof(double)) == 0);
orDummyObject->SetMethodTable(g_pObjectClass);
}
else
{
orObject = (ArrayBase*) Alloc(totalSize, FALSE, FALSE);
bPublish = (totalSize >= LARGE_OBJECT_SIZE);
}
}
// 设置类型信息(MethodTable)
// Initialize Object
orObject->SetMethodTable( pMT );
_ASSERTE(orObject->GetMethodTable() != NULL);
// 设置数组长度
orObject->m_NumComponents = cElements;
// 省略部分代码......
return( ObjectToOBJECTREF((Object*)orObject) );
}再看AllocateArrayEx函数，这个函数比起上面的函数多出了对多维数组的处理
JIT_NewArr1调用AllocateArrayEx时传了3个参数，剩下2个参数是可选参数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/gchelpers.cpp#L282
// Handles arrays of arbitrary dimensions
//
// If dwNumArgs is set to greater than 1 for a SZARRAY this function will recursively
// allocate sub-arrays and fill them in.
//
// For arrays with lower bounds, pBounds is <lower bound 1>, <count 1>, <lower bound 2>, ...
OBJECTREF AllocateArrayEx(TypeHandle arrayType, INT32 *pArgs, DWORD dwNumArgs, BOOL bAllocateInLargeHeap
DEBUG_ARG(BOOL bDontSetAppDomain))
{
// 省略部分代码......
ArrayBase * orArray = NULL;
// 省略部分代码......
// 调用gc的帮助函数分配内存
if (bAllocateInLargeHeap)
{
orArray = (ArrayBase *) AllocLHeap(totalSize, FALSE, pArrayMT->ContainsPointers());
// 设置类型信息(MethodTable)
orArray->SetMethodTableForLargeObject(pArrayMT);
}
else
{
#ifdef FEATURE_64BIT_ALIGNMENT
MethodTable *pElementMT = arrayDesc->GetTypeParam().GetMethodTable();
if (pElementMT->RequiresAlign8() && pElementMT->IsValueType())
{
// This platform requires that certain fields are 8-byte aligned (and the runtime doesn't provide
// this guarantee implicitly, e.g. on 32-bit platforms). Since it's the array payload, not the
// header that requires alignment we need to be careful. However it just so happens that all the
// cases we care about (single and multi-dim arrays of value types) have an even number of DWORDs
// in their headers so the alignment requirements for the header and the payload are the same.
_ASSERTE(((pArrayMT->GetBaseSize() - SIZEOF_OBJHEADER) & 7) == 0);
orArray = (ArrayBase *) AllocAlign8(totalSize, FALSE, pArrayMT->ContainsPointers(), FALSE);
}
else
#endif
{
orArray = (ArrayBase *) Alloc(totalSize, FALSE, pArrayMT->ContainsPointers());
}
// 设置类型信息(MethodTable)
orArray->SetMethodTable(pArrayMT);
}
// 设置数组长度
// Initialize Object
orArray->m_NumComponents = cElements;
// 省略部分代码......
return ObjectToOBJECTREF((Object *) orArray);
}总结:
第二种new做的事情主要有
- 调用JIT_NewArr1
- 从GCHeap中申请一块内存
- 设置类型信息(MethodTable)
- 设置数组长度(m_NumComponents)
- 不会调用构造函数，所以所有内容都会为0（所有成员都会为默认值）
第三种new(对struct的new)做了什么
对struct的new不会从GCHeap申请内存，也不会设置类型信息(MethodTable)，所以可以直接进入总结
总结:
第三种new做的事情主要有
- 在进入函数时统一从栈(Stack)分配内存
- 分配的内存不会包含同步块索引(SyncBlock)和类型信息(MethodTable)
- 调用构造函数
验证第一种new(对class的new)
打开VS反汇编和内存窗口，让我们来看看第一种new实际做了什么事情
第一种new的反汇编结果如下，一共有两个call
00007FF919570B53 mov rcx,7FF9194161A0h // 设置第一个参数(指向MethodTable的指针)
00007FF919570B5D call 00007FF97905E350 // 调用分配内存的函数，默认是JIT_New
00007FF919570B62 mov qword ptr [rbp+38h],rax // 把地址设置到临时变量(rbp+38)
00007FF919570B66 mov r8,37BFC73068h
00007FF919570B70 mov r8,qword ptr [r8] // 设置第三个参数("hello")
00007FF919570B73 mov rcx,qword ptr [rbp+38h] // 设置第一个参数(this)
00007FF919570B77 mov edx,12345678h // 设置第二个参数(0x12345678)
00007FF919570B7C call 00007FF9195700B8 // 调用构造函数
00007FF919570B81 mov rcx,qword ptr [rbp+38h]
00007FF919570B85 mov qword ptr [rbp+50h],rcx // 把临时变量复制到myClass变量中第一个call是分配内存使用的帮助函数，默认调用JIT_New
但是这里实际调用的不是JIT_New而是JIT_TrialAllocSFastMP_InlineGetThread函数，这是一个优化版本允许从TLS(Thread Local Storage)中快速分配内存
我们来看一下JIT_TrialAllocSFastMP_InlineGetThread函数的定义
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/amd64/JitHelpers_InlineGetThread.asm#L59
; IN: rcx: MethodTable*
; OUT: rax: new object
LEAF_ENTRY JIT_TrialAllocSFastMP_InlineGetThread, _TEXT
mov edx, [rcx + OFFSET__MethodTable__m_BaseSize] // 从MethodTable获取需要分配的内存大小，放到edx
; m_BaseSize is guaranteed to be a multiple of 8.
PATCHABLE_INLINE_GETTHREAD r11, JIT_TrialAllocSFastMP_InlineGetThread__PatchTLSOffset
mov r10, [r11 + OFFSET__Thread__m_alloc_context__alloc_limit] // 获取从TLS分配内存的限制地址，放到r10
mov rax, [r11 + OFFSET__Thread__m_alloc_context__alloc_ptr] // 获取从TLS分配内存的当前地址，放到rax
add rdx, rax // 地址 + 需要分配的内存大小，放到rdx
cmp rdx, r10 // 判断是否可以从TLS分配内存
ja AllocFailed // if (rdx > r10)
mov [r11 + OFFSET__Thread__m_alloc_context__alloc_ptr], rdx // 设置新的当前地址
mov [rax], rcx // 给刚刚分配到的内存设置MethodTable
ifdef _DEBUG
call DEBUG_TrialAllocSetAppDomain_NoScratchArea
endif ; _DEBUG
ret // 分配成功，返回
AllocFailed:
jmp JIT_NEW // 分配失败，调用默认的JIT_New函数
LEAF_END JIT_TrialAllocSFastMP_InlineGetThread, _TEXT可以看到做的事情和JIT_New相同，但不是从堆而是从TLS中分配内存
第二个call调用构造函数，call的地址和下面的地址不一致可能是因为中间有一层包装，目前还未解明包装中的处理
最后一个call调用的是JIT_WriteBarrier
验证第二种new(对array的new)
反汇编可以看到第二种new只有一个call
00007FF919570B93 mov rcx,7FF9195B4CFAh // 设置第一个参数(指向MethodTable的指针)
00007FF919570B9D mov edx,378h // 设置第二个参数(数组的大小)
00007FF919570BA2 call 00007FF97905E440 // 调用分配内存的函数，默认是JIT_NewArr1
00007FF919570BA7 mov qword ptr [rbp+30h],rax // 设置到临时变量(rbp+30)
00007FF919570BAB mov rcx,qword ptr [rbp+30h]
00007FF919570BAF mov qword ptr [rbp+48h],rcx // 把临时变量复制到myArray变量中call实际调用的是JIT_NewArr1VC_MP_InlineGetThread这个函数
和JIT_TrialAllocSFastMP_InlineGetThread一样，同样是从TLS(Thread Local Storage)中快速分配内存的函数
源代码: https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/amd64/JitHelpers_InlineGetThread.asm#L207
具体代码这里就不再分析，有兴趣的可以去阅读上面的源代码
验证第三种new(对struct的new)
对struct的new会在函数进入的时候从栈分配内存，这里是减少rsp寄存器(栈顶)的值
00007FF919570B22 push rsi // 保存原rsi
00007FF919570B23 sub rsp,60h // 从栈分配内存
00007FF919570B27 mov rbp,rsp // 复制值到rbp
00007FF919570B2A mov rsi,rcx // 保存原rcx到rsi
00007FF919570B2D lea rdi,[rbp+28h] // rdi = rbp+28，有28个字节需要清零
00007FF919570B31 mov ecx,0Eh // rcx = 14 (计数)
00007FF919570B36 xor eax,eax // eax = 0
00007FF919570B38 rep stos dword ptr [rdi] // 把eax的值(short)设置到rdi直到rcx为0，总共清空14*2=28个字节
00007FF919570B3A mov rcx,rsi // 恢复原rcx因为分配的内存已经在栈里面，后面只需要直接调构造函数
00007FF919570BBD lea rcx,[rbp+40h] // 第一个参数 (this)
00007FF919570BC1 mov edx,55667788h // 第二个参数 (0x55667788)
00007FF919570BC6 call 00007FF9195700A0 // 调用构造函数构造函数的反编译
中间有一个call 00007FF97942E260调用的是JIT_DbgIsJustMyCode
在函数结束时会自动释放从栈分配的内存，在最后会让rsp = rbp + 0x60，这样rsp就恢复原值了
参考
http://stackoverflow.com/questions/1255803/does-the-net-clr-jit-compile-every-method-every-time
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/gchelpers.h
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/gchelpers.cpp#L986
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jithelpers.cpp#L2908
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterface.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/jitinterfacegen.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/vm/amd64/JitHelpers_InlineGetThread.asm
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/gc/gcinterface.h#L230
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/gc/gc.h
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/gc/gc.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/opcode.def#L153
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/readytorunhelpers.h#L46
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/readytorun.h#L236
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/corinfo.h##L1147
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/corjit.h#L350
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/ee_il_dll.cpp#L279
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/inc/jithelpers.h
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.hpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.h
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/compiler.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/flowgraph.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/importer.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/gentree.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/objectalloc.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/morph.cpp
https://github.com/dotnet/coreclr/blob/release/1.1.0/src/jit/codegenxarch.cpp#L8404
https://github.com/dotnet/coreclr/blob/release/1.1.0/Documentation/botr/ryujit-overview.md
https://github.com/dotnet/coreclr/blob/master/Documentation/building/viewing-jit-dumps.md
https://github.com/dotnet/coreclr/blob/master/Documentation/building/linux-instructions.md
https://en.wikipedia.org/wiki/Basic_block
https://en.wikipedia.org/wiki/Control_flow_graph
https://en.wikipedia.org/wiki/Static_single_assignment_form
https://msdn.microsoft.com/en-us/library/windows/hardware/ff561499(v=vs.85).aspx
https://msdn.microsoft.com/en-us/library/ms228973(v=vs.110).aspx
https://msdn.microsoft.com/en-us/library/system.runtime.constrainedexecution.criticalfinalizerobject(v=vs.110).aspx
https://msdn.microsoft.com/en-us/library/system.runtime.interopservices.safehandle(v=vs.110).aspx
https://msdn.microsoft.com/en-us/library/system.runtime.interopservices.criticalhandle(v=vs.110).aspx
https://dotnet.myget.org/feed/dotnet-core/package/nuget/runtime.win7-x64.Microsoft.NETCore.Runtime.CoreCLR
http://www.codemachine.com/article_x64deepdive.html
这一篇相对前一篇多了很多c++和汇编代码，也在表面上涉及到了JIT，你们可能会说看不懂
这是正常的，我也不是完全看懂这篇提到的所有处理
欢迎大神们勘误，也欢迎小白们提问
接下来我会重点分析GC分配内存的算法，敬请期待