Unnecessary GC frame when the result is returned directly

[yuyichao]

On 0.4.3,

julia> f2() = ccall(:jl_get_cpu_name, Ref{ByteString}, ())
f2 (generic function with 1 method)

julia> @code_llvm f2()

define %jl_value_t* @julia_f2_21304() {
top:
  %0 = call %jl_value_t* inttoptr (i64 140514339702960 to %jl_value_t* ()*)()
  ret %jl_value_t* %0
}

On current master

julia> f2() = ccall(:jl_get_cpu_name, Ref{ByteString}, ())
f2 (generic function with 1 method)

julia> @code_llvm f2()

define %jl_value_t* @julia_f2_22854() #0 {
top:
  %0 = alloca [3 x %jl_value_t*], align 8
  %.sub = getelementptr inbounds [3 x %jl_value_t*], [3 x %jl_value_t*]* %0, i64 0, i64 0
  %1 = getelementptr [3 x %jl_value_t*], [3 x %jl_value_t*]* %0, i64 0, i64 2
  store %jl_value_t* null, %jl_value_t** %1, align 8
  %2 = bitcast [3 x %jl_value_t*]* %0 to i64*
  store i64 2, i64* %2, align 8
  %3 = getelementptr [3 x %jl_value_t*], [3 x %jl_value_t*]* %0, i64 0, i64 1
  %4 = load i64, i64* bitcast (%jl_value_t*** @jl_tls_states to i64*), align 8
  %5 = bitcast %jl_value_t** %3 to i64*
  store i64 %4, i64* %5, align 8
  store %jl_value_t** %.sub, %jl_value_t*** @jl_tls_states, align 8
  %6 = call %jl_value_t* inttoptr (i64 139747097742064 to %jl_value_t* ()*)()
  store %jl_value_t* %6, %jl_value_t** %1, align 8
  %7 = load i64, i64* %5, align 8
  store i64 %7, i64* bitcast (%jl_value_t*** @jl_tls_states to i64*), align 8
  ret %jl_value_t* %6
}

[yuyichao]

Move to #15369