Miscompile with C++23 constexpr-unknown constant evaluation

[efriedma-quic]

Testcase:

int &ff();
int &x = ff();
struct A { int& x; };
struct B { A x[20]; };
B f() { return {x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x}; }

Generates the following, which is wrong because the references point to x itself, not the referenced lvalue.

[...]
@constinit = private constant [20 x %struct.A] [%struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }, %struct.A { ptr @x }], align 8
[...]
define dso_local void @_Z1fv(ptr dead_on_unwind noalias writable sret(%struct.B) align 8 %agg.result) #2 {
entry:
  %x = getelementptr inbounds nuw %struct.B, ptr %agg.result, i32 0, i32 0
  call void @llvm.memcpy.p0.p0.i64(ptr align 8 %x, ptr align 8 @constinit, i64 160, i1 false)
  ret void
}

This is synthetic; I have no idea how likely it is for anyone to actually run into this. Filing as a followup to discussion on #128409.

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Eli Friedman (efriedma-quic)

Testcase:

int &ff();
int &x = ff();
struct A { int& x; };
struct B { A x[20]; };
B f() { return {x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x,x}; }

This is synthetic; I have no idea how likely it is for anyone to actually run into this. Filing as a followup to discussion on #128409.

[tbaederr]

This is also accepted by gcc and clang 19, why should it be rejected?

[efriedma-quic]

Updated description to be more clear.