erroneous reinterpret only fails with --code-coverage=user

[maleadt]

$ julia                      -e 'reinterpret(Nothing, nothing)'                               
                                               
$ julia --code-coverage=user -e 'reinterpret(Nothing, nothing)'                                                         
ERROR: bitcast: target type not a leaf primitive type
Stacktrace:
 [1] reinterpret(::Type{Nothing}, ::Nothing) at ./essentials.jl:413
 [2] top-level scope at none:1

Started behaving differently with 1.1, on 1.0 both fail. Arguably this use of reinterpret should be allowed, but the change in behavior based on code coverage had me stumped for a while.

[vtjnash]

reinterpret(T, nothing) should fail consistently for all T, but we're in need of merging #34539

[vtjnash]

Duplicate of #34482

[maleadt]

reinterpret(T, nothing) should fail consistently for all T, but we're in need of merging #34539

What's the recommended way to create an instance of a singleton type then? This for implementing the GPU version of pointerset; calling T() seems fragile (user might have defined a constructor, and I just want an instance to make codegen happy and generate a ret void).

[vtjnash]

unsafe_load could generate one (or its wrappers, such as Array{T,0}(undef)[]), though in general the system doesn't like to materialize undefined values from just their types

[maleadt]

I was trying to mimick Base.pointerref which does handle that case:

julia> foo(ptr) = Base.pointerref(ptr,1,1)
foo (generic function with 1 method)

julia> code_llvm(foo, Tuple{Ptr{Nothing}})

;  @ REPL[4]:1 within `foo'
define void @julia_foo_17046(i64) {
top:
  ret void
}