Method lookup with Tuple{DataType,Any} regression

[yuyichao]

Reduced from OpenCL.jl test failure.

function f2(::Type{T}, ::Int) where T
    1 + 1
end
@show code_typed(f2, Tuple{DataType,Any})

v = 1
T = UInt8
g() = f2(T::DataType, v)

@code_warntype g()
g()

Output:

code_typed(f2, Tuple{DataType, Any}) = 0-element Array{Any,1}
Variables:
  #self#::#g

Body:
  begin
      return (Main.f2)((Core.typeassert)(Main.T, Main.DataType)::DataType, Main.v)::Union{}
  end::Union{}
Unreachable reached at 0x7fb886a9814e

Doesn't happen on 0.6.

[yuyichao]

Seems the method lookup didn't regress (0.6 also return empty lookup result) but now type inference is exploiding that....

[vtjnash]

Looks like this case is wrong:

    frame #0: 0x00000001000ed473 libjulia.0.7.0.dylib`jl_type_intersection_env_s(a=<unavailable>, b=0x0000000109d180f0, penv=<unavailable>, issubty=<unavailable>) at subtype.c:2153 [opt]
   2150	                    if (jl_is_leaf_type(*ans)) {
   2151	                        for(i=0; i < sz; i++) {
   2152	                            if (jl_is_typevar(env[i])) {
-> 2153	                                *ans = jl_bottom_type; goto bot;
   2154	                            }
   2155	                        }
   2156	                    }

[vtjnash]

@JeffBezanson Why does this filter code exist? I don't see any tests fail if I remove it (although I only tested it on my branch where I also have #21026 fixed):

diff --git a/src/subtype.c b/src/subtype.c
index ac6bc07c01..95d91d8b4c 100644
--- a/src/subtype.c
+++ b/src/subtype.c
@@ -2142,20 +2142,9 @@ jl_value_t *jl_type_intersection_env_s(jl_value_t *a, jl_value_t *b, jl_svec_t *
         else {
             sz = szb;
             // TODO: compute better `env` directly during intersection.
-            // we assume that if the intersection is a leaf type, we have
-            // full information in `env`. however the intersection algorithm
-            // does not yet provide that in all cases so use subtype.
+            // for now, we attempt to compute env by using subtype on the intersection result
             if (szb > 0 && !jl_types_equal(b, (jl_value_t*)jl_type_type)) {
-                if (jl_subtype_env(*ans, b, env, szb)) {
-                    if (jl_is_leaf_type(*ans)) {
-                        for(i=0; i < sz; i++) {
-                            if (jl_is_typevar(env[i])) {
-                                *ans = jl_bottom_type; goto bot;
-                            }
-                        }
-                    }
-                }
-                else {
+                if (!jl_subtype_env(*ans, b, env, szb)) {
                     sz = 0;
                 }
             }

[JeffBezanson]

If anything, it would probably be an inference precision issue, where it's harder to get failures. It also might only be needed because of #21026.