Inference results lost after multiple codegen invocations

[maleadt]

#21677 seems to have caused a regression in irgen idempotence (verified by bisect):

julia> exec_dummy() = return nothing
exec_dummy (generic function with 1 method)

julia> @code_llvm exec_dummy()

; Function Attrs: sspstrong
define void @julia_exec_dummy_66504() #0 !dbg !5 {
top:
  ret void
}

julia> @code_llvm exec_dummy()

; Function Attrs: sspstrong
define void @julia_exec_dummy_66528() #0 !dbg !5 {
top:
  ret void
}

julia> # third time's a charm
julia> @code_llvm exec_dummy()

; Function Attrs: sspstrong
define i8** @japi1_exec_dummy_66528(i8**, i8***, i32) #0 !dbg !5 {
top:
  %3 = alloca i8***, align 8
  store volatile i8*** %1, i8**** %3, align 8
  ret i8** inttoptr (i64 140093546754056 to i8**)
}

The japi1 emission is caused because upon third emission specsig is false due to src->inferred being false too. This breaks CUDAnative, where we can only handle specialized julia_* signatures. (Note we use _dump_function with CodegenParams(cached=false) over there.)

Sorry for not doing much debugging here myself, but I'm short on time.
cc @vtjnash @jrevels

[vtjnash]

diff --git a/src/gf.c b/src/gf.c
index 3be37fc..c732014 100644
--- a/src/gf.c
+++ b/src/gf.c
@@ -262,7 +262,7 @@ jl_code_info_t *jl_type_infer(jl_method_instance_t **pli, size_t world, int forc
     li->inInference = 1;
     jl_svec_t *linfo_src_rettype = (jl_svec_t*)jl_apply_with_saved_exception_state(fargs, 3, 0);
     ptls->world_age = last_age;
-    assert((jl_is_method(li->def.method) || li->inInference == 0) && "inference failed on a toplevel expr");
+    li->inInference = 0;
 
     jl_code_info_t *src = NULL;
     if (jl_is_svec(linfo_src_rettype) && jl_svec_len(linfo_src_rettype) == 3 &&

[maleadt]

Great, thanks! Is this good-to-go like that? 9ebc8f5

[vtjnash]

Yes