LLVM lint: Memory reference address is misaligned

[sanxiyn]

This is a subissue of #7463.

pub enum H<A, B> {
    L(A),
    R(B),
}
pub fn test() -> H<i32, i64> {
    H::L(0)
}

Compile with rustc -C passes=lint and you get:

Undefined behavior: Memory reference address is misaligned
  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %2, i8* getelementptr inbounds ({ i8, i32, [8 x i8] }, { i8, i32, [8 x i8] }* @const3020, i32 0, i32 0), i64 16, i32 8, i1 false)

Manually setting alignment of const3020 to 8 in LLVM IR fixes the lint.

[dotdash]

Seems to be caused by using the sizing_type for alignment.

[dotdash]

Oh, nevermind, I completely misread the problem here and confused it with a different issue. The alignment of the sizing type is fine and the right thing would indeed be to add the alignment info to the constant.

[sanxiyn]

Eh, I am confused. Isn't the alignment of { i8, i32, [8 x i8] } 4, i32 being the most aligned? Then why are we passing 8 as alignment to memcpy intrinsic? Is this related to sizing type?

[dotdash]

{ i8, i32, [8 x i8] } is just the specific representation for the L variant. The type H<i32, i64> is represented as %"H<i32, i64>" = type { i8, [7 x i8], [1 x i64] }, so its alignment is 8. And I think all our code deals with types rather than concrete variants when checking for alignment.

We could probably extend that to allow for better packing of constants in some cases, but I'm not sure whether that's worth the trouble.