exists_equal tuning.

Make sure we don't forget the bound in `env`.
And this makes `subtype` slightly faster by reducing heavy re-subtyping.
But of course we will consume more Runion stack with this change.

src/subtype.c

@@ -1461,6 +1461,7 @@ static int forall_exists_equal(jl_value_t *x, jl_value_t *y, jl_stenv_t *e)
         (is_definite_length_tuple_type(x) && is_indefinite_length_tuple_type(y)))
         return 0;
 
+    jl_saved_unionstate_t oldLunions; push_unionstate(&oldLunions, &e->Lunions);
 
     if ((jl_is_uniontype(x) && jl_is_uniontype(y))) {
         // For 2 unions, first try a more efficient greedy algorithm that compares the unions
@@ -1472,7 +1473,6 @@ static int forall_exists_equal(jl_value_t *x, jl_value_t *y, jl_stenv_t *e)
         // }
 
         // Optimization
-        jl_saved_unionstate_t oldLunions; push_unionstate(&oldLunions, &e->Lunions);
         int16_t old2Rmore  = e->Runions.more;
         if (pick_union_decision(e, 1) == 0) {
             int16_t oldRmore = e->Runions.more;
@@ -1506,7 +1506,6 @@ static int forall_exists_equal(jl_value_t *x, jl_value_t *y, jl_stenv_t *e)
     if ((!jl_has_free_typevars(x) || !jl_has_free_typevars(y)))
         sub = local_forall_exists_subtype(x, y, e, 2);
     else {
-        jl_saved_unionstate_t oldLunions; push_unionstate(&oldLunions, &e->Lunions);
         e->Lunions.used = 0;
         while (1) {
             e->Lunions.more = 0;
@@ -1518,7 +1517,19 @@ static int forall_exists_equal(jl_value_t *x, jl_value_t *y, jl_stenv_t *e)
         pop_unionstate(&e->Lunions, &oldLunions);
     }
 
-    return sub && subtype(y, x, e, 0);
+    if (sub) {
+        e->Lunions.used = 0;
+        while (1) {
+            e->Lunions.more = 0;
+            e->Lunions.depth = 0;
+            sub = subtype(y, x, e, 0);
+            if (!sub || !next_union_state(e, 0))
+                break;
+        }
+        pop_unionstate(&e->Lunions, &oldLunions);
+    }
+
+    return sub;
 }
 
 static int exists_subtype(jl_value_t *x, jl_value_t *y, jl_stenv_t *e, jl_value_t *saved, jl_savedenv_t *se, int param)

test/subtype.jl

@@ -1480,10 +1480,10 @@ f24521(::Type{T}, ::Type{T}) where {T} = T
 @test f24521(Tuple{Vararg{Int64}}, Tuple{Vararg{Int64,N}} where N) == Tuple{Vararg{Int64,N}} where N
 
 # issue #26654
-@test_broken !(Ref{Union{Int64, Ref{Number}}} <: Ref{Union{Ref{T}, T}} where T)
-@test_broken !(Ref{Union{Int64, Val{Number}}} <: Ref{Union{Val{T}, T}} where T)
-@test_broken !(Ref{Union{Ref{Number}, Int64}} <: Ref{Union{Ref{T}, T}} where T)
-@test_broken !(Ref{Union{Val{Number}, Int64}} <: Ref{Union{Val{T}, T}} where T)
+@test !(Ref{Union{Int64, Ref{Number}}} <: Ref{Union{Ref{T}, T}} where T)
+@test !(Ref{Union{Int64, Val{Number}}} <: Ref{Union{Val{T}, T}} where T)
+@test !(Ref{Union{Ref{Number}, Int64}} <: Ref{Union{Ref{T}, T}} where T)
+@test !(Ref{Union{Val{Number}, Int64}} <: Ref{Union{Val{T}, T}} where T)
 
 # issue #26180
 @test !(Ref{Union{Ref{Int64}, Ref{Number}}} <: Ref{Ref{T}} where T)