使用LLVM JIT代码

时间:2015-07-17 06:52:51

标签: c++ llvm llvm-ir llvm-c++-api

我希望我的大部分程序都是一个通常编译的C ++程序。所述程序使用一块连续的内存用于堆栈。堆栈的顶部由普通指针维护。

我想与通过LLVM JIT生成的代码共享该指针。例如,给定:

llvm::InitializeNativeTarget();

llvm::LLVMContext ctx;
std::unique_ptr<llvm::Module> uptr_module = llvm::make_unique<llvm::Module>( "lt", ctx );
llvm::Module *const module = uptr_module.get();

int *const stack = new int[100];
int *top = stack;                 // I want this pointer to be shared with JIT'd code

llvm::Function *const func = llvm::cast<llvm::Function>(
    module->getOrInsertFunction( "func", llvm::Type::getVoidTy( ctx ), (llvm::Type*)0 )
);
llvm::BasicBlock *const block = llvm::BasicBlock::Create( ctx, "entry", func );

pointerInc( &top, block );        // Increment the pointer in JIT'd code

llvm::ReturnInst::Create( ctx, block );
llvm::verifyFunction( *func, &llvm::outs() );
llvm::verifyModule( *module, &llvm::outs() );
module->dump();

llvm::EngineBuilder eb( std::move( uptr_module ) );
llvm::ExecutionEngine *const exec = eb.create();
assert( exec );

void *const func_ptr = exec->getPointerToFunction( func );
assert( func_ptr );
typedef void (*PFv_v)();
(*(PFv_v)func_ptr)();             // Call JIT'd function

其中pointerInc()会将JIT代码插入当前BasicBlock以增加toppointerInc()代码为:

// Convert a raw C++ pointer into an LLVM Constant*.
template<typename T>
inline llvm::Value* ptrToValue( T **pptr, llvm::LLVMContext &ctx ) {
    return return llvm::ConstantInt::get( llvm::Type::getInt64Ty( ctx ), (uint64_t)pptr );
}

void pointerInc( llvm::Constant *pptrAsInt64, llvm::ConstantInt *sizeof_T,
                 llvm::BasicBlock *block ) {
    llvm::LLVMContext &ctx = block->getContext();

    llvm::Constant *const intToPtr8 = llvm::ConstantExpr::getIntToPtr(
        pptrAsInt64, llvm::PointerType::getUnqual( llvm::Type::getInt8Ty( ctx ) )
    );

    llvm::GetElementPtrInst *const inc =
        llvm::GetElementPtrInst::Create( intToPtr8, sizeof_T, "inc", block );

    llvm::CastInst *const cast = llvm::CastInst::CreatePointerCast(
        inc, llvm::Type::getInt64Ty( ctx ), "cast", block
    );

    llvm::Constant *const intToPtr64 = llvm::ConstantExpr::getIntToPtr(
        pptrAsInt64, llvm::PointerType::getUnqual( llvm::Type::getInt64Ty( ctx ) )
    );

    llvm::StoreInst *const store = new llvm::StoreInst( cast, intToPtr64, false, block );
    store->setAlignment( 8 );
}

template<typename T>
inline void pointerInc( T **pptr, llvm::BasicBlock *block ) {
    llvm::LLVMContext &ctx = block->getContext();
    llvm::ConstantInt *const sizeof_T =
        llvm::ConstantInt::get( llvm::Type::getInt64Ty( ctx ), sizeof( T ) );
    pointerInc( ptrToValue( pptr, ctx ), sizeof_T, block );
}

不幸的是,这不起作用。这是(较大的)pointerInc()的主体是错的。代码实际上是从llc在普通C ++程序上生成的LLVM C ++ API代码派生而来的,该程序会增加指针。

运行时,程序会打印:

&p = 140734551679784
--------------------
; ModuleID = 'lt'

define void @func() {
entry:
  %inc = getelementptr i8* inttoptr (i64 140734551679784 to i8*), i64 4
  %cast = ptrtoint i8* %inc to i64
  store i64 %cast, i64* inttoptr (i64 140734551679784 to i64*), align 8
  ret void
}
Segmentation fault: 11 (core dumped)

有两个问题:

  1. 这是对的吗?我可以做我想做的事情,即用JIT代码共享原始C ++指针吗?
  2. 为什么要倾销核心?
  3. 即使我使JIT'd函数为空,代码仍然核心转储在调用该函数的行。 LLVM JIT设置代码看起来就像我看到的所有示例,所以我也没有看到它有什么问题。

    一点帮助?

    更新

    如果我更改弃用的行:

    void *const func_ptr = exec->getPointerToFunction( func );
    

    到新行:

    uint64_t const func_ptr = exec->getFunctionAddress( "func" );
    

    然后func_ptr为空。

1 个答案:

答案 0 :(得分:1)

在使用lcc(以及使用更好的C ++代码加入其中)之后玩了很多次,我有它的工作:

llvm::Value* pointerToPointer( void *ptr, llvm::BasicBlock *block ) {
    using namespace llvm;
    LLVMContext &ctx = block->getContext();
    ConstantInt *const ptrAsInt =
        ConstantInt::get( IntegerType::get( ctx, 64 ), (uint64_t)ptr );
    PointerType *const Int8Ptr_type = Type::getInt8PtrTy( ctx );
    PointerType *const Int8PtrPtr_type = PointerType::getUnqual( Int8Ptr_type );
    return new IntToPtrInst( ptrAsInt, Int8PtrPtr_type, "pptr", block );
}

void pointerInc( llvm::Value *pptr, llvm::ConstantInt *sizeof_T,
                llvm::BasicBlock *block ) {
    using namespace llvm;
    LLVMContext &ctx = block->getContext();

    LoadInst *const ptr = new LoadInst( pptr, "ptr", block );
    ptr->setAlignment( sizeof(void*) );

    GetElementPtrInst *const inc =
        GetElementPtrInst::Create( ptr, sizeof_T, "inc", block );

    StoreInst *const store = new StoreInst( inc, pptr, block );
    store->setAlignment(sizeof(void*));
}

template<typename T>
inline void pointerInc( T **pptr, llvm::BasicBlock *block ) {
    using namespace llvm;
    LLVMContext &ctx = block->getContext();
    ConstantInt *const sizeof_T = ConstantInt::get(
        IntegerType::get( ctx, 64 ), (uint64_t)sizeof( T )
    );
    pointerInc( pointerToPointer( pptr, block ), sizeof_T, block );
}

但是,只有通过以下方式调用JIT&lt; d / p>函数时,程序才能成功运行。

vector<GenericValue> noargs;
exec->runFunction( func, noargs );

使用getFunctionAddress()getPointerToFunction()转储核心。我仍然没有答案。