Mir: translation of indexing is incorrect

[nagisa]

const C: [u32; 5] = [0, 42, 112, 212, 433];

#[rustc_mir(graphviz="foo.gv")]
fn mir() {
    C[3];
}

fn main(){}

will fail to compile with an LLVM abort:

Stored value type does not match pointer operand type!
  store [5 x i32] %4, i32* %temp1, align 4
 i32Invalid bitcast
i8 addrspace(9)* bitcast (%str_slice { i8* getelementptr inbounds ([7 x i8], [7 x i8]* @str3466, i32 0, i32 0), i64 7 } to i8 addrspace(9)*)

Seems to be a mis-translation of indexing (trans/mir/lvalue at ProjectionElem::Index).

[nagisa]

This is not related to #30479.

According to @eddyb we might do something incorrectly with GEPs.

[eddyb]

That was my first guess, but it seems that C[3] is an rvalue with the type [u32; 5], suggesting C itself was loaded.

[eddyb]

Oh, my first guess was right: all GEPs are build::InBoundsGEP(bcx, tr_base.llval, &[llindex]) but that's not enough.
You see, GEP's first index is for the pointer itself. C[3] currently gets (mistakenly) translated as *(&C).offset(3) instead of the (correct) *(&C[0]).offset(3).
All GEP calls should use 0 as the first index and then the index of the array element/struct field.