Running code_typed changes result of code_llvm

[topolarity]

julia> using NetworkOptions

julia> code_llvm(NetworkOptions.__init__, (); raw=true)
; Function Signature: __init__()
;  @ /build/julia/usr/share/julia/stdlib/v1.11/NetworkOptions/src/ssh_options.jl:163 within `__init__`
define swiftcc void @julia___init___1092(ptr nonnull swiftself %pgcstack) #0 !dbg !5 {
top:
  %jlcallframe1 = alloca [2 x ptr], align 8
  %ptls_field = getelementptr inbounds ptr, ptr %pgcstack, i64 2
  %ptls_load = load ptr, ptr %ptls_field, align 8, !tbaa !14
  %0 = getelementptr inbounds ptr, ptr %ptls_load, i64 2
  %safepoint = load ptr, ptr %0, align 8, !tbaa !18
  fence syncscope("singlethread") seq_cst
  %1 = load volatile i64, ptr %safepoint, align 8, !dbg !20
  fence syncscope("singlethread") seq_cst
;  @ /build/julia/usr/share/julia/stdlib/v1.11/NetworkOptions/src/ssh_options.jl:165 within `__init__`
  store ptr @"jl_global#1098.jit", ptr %jlcallframe1, align 8, !dbg !21
  %2 = getelementptr inbounds ptr, ptr %jlcallframe1, i64 1, !dbg !21
  store ptr @"jl_global#1099.jit", ptr %2, align 8, !dbg !21
  %3 = call nonnull ptr @ijl_invoke(ptr nonnull @"jl_global#1097.jit", ptr nonnull %jlcallframe1, i32 2, ptr nonnull @"-Main.Base.lock#1096.jit"), !dbg !21
  ret void, !dbg !21
}

versus if you call code_typed first:

julia> using NetworkOptions

julia> code_typed(NetworkOptions.__init__, ())
1-element Vector{Any}:
 CodeInfo(
1 ─     invoke NetworkOptions.lock(NetworkOptions.var"#7#8"()::NetworkOptions.var"#7#8", NetworkOptions.BUNDLED_KNOWN_HOSTS_LOCK::ReentrantLock)::Nothing
└──     return nothing
) => Nothing

julia> code_llvm(NetworkOptions.__init__, (); raw=true)
; Function Signature: __init__()
;  @ /build/julia/usr/share/julia/stdlib/v1.11/NetworkOptions/src/ssh_options.jl:163 within `__init__`
define swiftcc void @julia___init___2493(ptr nonnull swiftself %pgcstack) #0 !dbg !5 {
top:
  %ptls_field = getelementptr inbounds ptr, ptr %pgcstack, i64 2
  %ptls_load = load ptr, ptr %ptls_field, align 8, !tbaa !14
  %0 = getelementptr inbounds ptr, ptr %ptls_load, i64 2
  %safepoint = load ptr, ptr %0, align 8, !tbaa !18
  fence syncscope("singlethread") seq_cst
  %1 = load volatile i64, ptr %safepoint, align 8, !dbg !20
  fence syncscope("singlethread") seq_cst
;  @ /build/julia/usr/share/julia/stdlib/v1.11/NetworkOptions/src/ssh_options.jl:165 within `__init__`
  call swiftcc void @j_lock_2497(ptr nonnull swiftself %pgcstack, ptr nonnull @"jl_global#2498.jit"), !dbg !21
  ret void, !dbg !21
}

Notice the jl_invoke that became a proper static call.

[vtjnash]

I think this is the same inference bug as I ran into here: #52069

[topolarity]

I believe @gbaraldi pointed out another unresolved mystery here, which is why we end up with this "constant invoke" in the first place

[gbaraldi]

This is mostly an artifact of the fact that the first time we compile that code the global doesn't exist, but if we recompile it we know it exists and assume it to always exist (potentially incorrectly)

[vtjnash]

Looking closer, this is not a bug at all, just a suboptimal inference result (which is never a bug)