cmd/compile/internal/types2: underlying type of a type parameter is its constraint interface
Until now, the type checker operated with the definition that the
underlying type of a type parameter is itself. This leads to some
inconcistencies and caused us to disallow type declarations where
the RHS is a stand-alone type parameter.
This change implements an alernative definition: the underlying
type of a type parameter is the underlying type of its constraint;
i.e., the underlying type of a type parameter is always an interface
(because constraints must be interfaces). This matches the theory
closely and also resolves some inconsistencies. For example, we
don't need to prohibit stand-alone type parameters on the RHS of
a type declaration (though, for the sake of keeping the tests the
same, we still do in this CL). We also get a clear understanding of
what it would mean to use a type assertion or type switch on a type
parameter (still disabled with this CL). Finally, the declaration
of a type parameter now very closely matches the definition of an
ordinary type.
The main consequence is that the rules for assignment need to be
slightly modified: even though a type parameter is an interface,
we cannot simply assign to it per the rules for interfaces: the
type parameter's type is fixed for the instantiation and we need
to reflect that accordingly when checking for assignability.
This CL does not enable the new mode, it implements it in parallel
to the existing mode; the internal flag tparamIsIface is used to
switch between the modes.
The changes to the code are numerous, but straight-forward: when-
ever we deal with an underlying type that might be a type parameter
(or newly, an interface), we need to act slightly differently. For
the time being this leads to some code duplication because the code
supports both modes.
While some of the code for the new mode seems more complicated
(e.g., when we have an interface, the code checks that it is not
the underlying type of a type parameter), in reality many of the
extra checks are redundant and only present because of an abundance
of caution: interfaces with specific type sets are not permitted as
types for ordinary variables, and so even if we were to hit those
cases w/o excluding type parameters the behavior would be the same.
Runs all tests with tparamIsIface enabled and disabled.
Current setting: disabled.
Change-Id: I7bb6453f4fe2569d92face222058fb4e17b12f25
Reviewed-on: https://go-review.googlesource.com/c/go/+/359016
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Griesemer <gri@golang.org>
TryBot-Result: Go Bot <gobot@golang.org> Reviewed-by: Robert Findley <rfindley@google.com>