如何在不优化的情况下生成 LLVM IR

Question

I am writing an LLVM PASS to analyze info in registers. It seems that IRBuilder optimized my code automatically, making an expression to be an operand. For example, I write down below code to generate LLVM IR.

// %reg = getelementptr inbounds ([128 x i256], [128 x i256]* @mstk, i256 0, i256 0
std::vector<llvm::Value*> indices(2, llvm::ConstantInt::get(Type::Int256Ty, 0));
llvm::ArrayRef<llvm::Value *> indicesRef(indices);
llvm::Value* m_sp = m_builder.CreateGEP(conArray, indicesRef, "spPtr"); 

// store 0, *%m_sp
m_builder.CreateStore(llvm::ConstantInt::get(Type::Int256Ty, 0), m_sp); 

The expected IR should consist of two registers. (see below)

%reg = getelementptr inbounds ([128 x i256], [128 x i256]* @mstk, i256 0, i256 0)
store i256 0, i256* reg

Unfortunately, IRBuilder optimizes the IR by combining the registers.

store i256 0, i256* getelementptr inbounds ([128 x i256], [128 x i256]* @mstk, i256 0, i256 0)

Is it possible to disable the IR optimization? I have made sure that I turned off all PASS. Thanks.

Answer 1

I solved this problem by disabling the constant folder of IR builder. See Disable constant folding for LLVM 10 C++ API