[aarch64][win] Applying alignment to a stack object causes a mis-compilation when using exception on Arm64 Windows

[dpaoliello]

Given the following code:

// Large object: alignment seems to be important?
struct alignas(128) BigObj {
    int value;
    // Destructor so it's kept alive.
    ~BigObj() { }
};

// Exception type need to be large enough to not fit in a register.
struct Error {
    int value;
    int padding[3];
};

int main() {
    BigObj bo{};

    try {
        throw Error { 42, {0, 0, 0} };
    } catch (const Error& e) {
        return e.value;
    }
    return 0;
}

This program crashes when built and run on Arm64 Windows:

> & 'C:\Program Files\LLVM\bin\clang.exe' .\test.cpp
> .\a.exe
> $LASTEXITCODE
-1073741819

The alignas seems to be important: if I remove that and replace it with a very large array within BigObj then the issue no longer reproduces.

When debugging, the slot for the exception in the catch is still null and nothing writes to it. In the VC Runtime during unwind, it writes the pointer to the exception to a completely different location.

I'm guessing that asjusting for the alignment may happen after calculating the CatchObjOffset in the exception data?

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Daniel Paoliello (dpaoliello)

Given the following code:

// Large object: alignment seems to be important?
struct alignas(128) BigObj {
    int value;
    // Destructor so it's kept alive.
    ~BigObj() { }
};

// Exception type need to be large enough to not fit in a register.
struct Error {
    int value;
    int padding[3];
};

int main() {
    BigObj bo{};

    try {
        throw Error { 42, {0, 0, 0} };
    } catch (const Error& e) {
        return e.value;
    }
    return 0;
}

This program crashes when built and run on Arm64 Windows:

> & 'C:\Program Files\LLVM\bin\clang.exe' .\test.cpp
> .\a.exe
> $LASTEXITCODE
-1073741819

The alignas seems to be important: if I remove that and replace it with a very large array within BigObj then the issue no longer reproduces.

[efriedma-quic]

Using alignas() in this context triggers stack realignment, which then triggers the use of a base pointer (x19) to refer to the stack because there's a variable gap between local variables and the frame pointer. That could mess up the offset math, so the catch is different from what exception handling expects.

It might be wrong for AArch64TargetLowering::needsFixedCatchObjects() to return false, if catch objects need to be allocated before the realignment gap.