hashing of general objects was far too simple (fixes #20744) (#20767)

* hashing of general objects was far too simple (fixes #20744)

* hash_uint(3h - object_id(x)) is a somewhat better hash

If we'd used `3h - object_id(x)` previously, the original issue
would never have come up since `hash(a, hash(b, hash(c, h)))`
wouldn't have produced `3A - 3B + 3C - h` but would instead have given
the value `27h -9A - 3B - C`, which is asymmetrical in `a` and `c`.
Passing the result to `hash_uint` fixes that issue since that function
is non-linear, but defense in depth, right? We may as well do the
same operations in such a way that the passed-through value gets the
most operations done on it.

Author: StefanKarpinski

base/hashing.jl

@@ -7,7 +7,7 @@ hash(w::WeakRef, h::UInt) = hash(w.value, h)
 
 ## hashing general objects ##
 
-hash(x::ANY, h::UInt) = 3*object_id(x) - h
+hash(x::ANY, h::UInt) = hash_uint(3h - object_id(x))
 
 ## core data hashing functions ##
 
@@ -55,7 +55,6 @@ end
 
 ## symbol & expression hashing ##
 
-hash(x::Symbol, h::UInt) = 3*object_id(x) - h
 if UInt === UInt64
     hash(x::Expr, h::UInt) = hash(x.args, hash(x.head, h + 0x83c7900696d26dc6))
 else

test/hashing.jl

@@ -117,3 +117,6 @@ for prec in [3, 11, 15, 16, 31, 32, 33, 63, 64, 65, 254, 255, 256, 257, 258, 102
         @test isequal(x, y)
     end
 end
+
+# issue #20744
+@test hash(:c, hash(:b, hash(:a))) != hash(:a, hash(:b, hash(:c)))