These are all problems that Eclipse can detect, but that it detects no
instances of right now. Treating these as warnings instead of errors
should help prevent us from slipping backward in the future.
Casting to `Foo<Bar>` results in an unchecked-cast warning due to Java
generics type erasure. However, sometimes we don't really need a
`Foo<Bar>`, but could simply use any `Foo<?>`. Casting to the latter
creates no warning.
We already have plenty of examples of Serializable classes with this
field, and the vast majority of those fields have generated IDs rather
than "1L". From this I infer that using proper serialVersionUID
fields is considered appropriate WALA coding style.
One such annotation was unnecessary because the thing it was
suppressing no longer happens. Any future unnecessary warning
suppressions of this kind will now be treated as errors.
The other annotations were unnecessary because the corresponding
warnings have been disabled entirely in the Eclipse projects'
configurations. There seems to be no way to tell Eclipse to treat
these as anything other than "info" diagnostics in the future, so
that's how they will remain.
In general, my approach was to try to eliminate each unused parameter
using Eclipse's "Change Method Signature" refactoring. That did not
always succeed: a parameter may be unused in some base class method,
but then be used in subclass's override of that method. In cases
where refactoring to eliminate a parameter failed, I instead annotated
the parameter with '@SuppressWarnings("unused")' to silence the
warning.
Note: this group of changes creates a significant risk of
incompatibility for third-party WALA code. Some removed parameters
change externally-visible APIs. Furthermore, these changes do not
necessarily lead to Java compilation errors. For example, suppose
third-party code subclasses a WALA class or interface, overrides a
method, but does not annotate that method as @Override. Removing a
parameter means that the third-party method no longer overrides. This
can quietly change code behavior without compile-time errors or
warnings. This is exactly why one should use @Override wherever
possible, but we cannot guarantee that third-party WALA users have
done that.
Unnecessary "throws" declarations tend to cascade. If foo() calls
bar() and bar() falsely declares that it might throw IOException, that
often leads a programmer to declare that foo() might throw IOException
as well. Fixing the bar() throws declaration then reveals that we can
fix the foo() throws declaration too. By the time we reach a fixed
point with cleaning these up, we have removed roughly 320 unnecessary
throws declarations.
In a few cases, this cleanup even lets us remove entire "try
... catch" statements where the only thing being caught was an
exception that we now statically know cannot be thrown. Nice!
In Eclipse project configurations, upgrade any future such shenanigans
from warnings to errors. Now that we've fixed this, we don't want it
coming back again.
There is a potential drawback to this change. Conceivably some public
WALA API entry point might have declared that it could throw some
exception merely to reserve the *option* of throwing that exception in
third-party code that subclasses and overrides the API entry point in
question. I have no idea whether this is a significant concern in
practice, though.
The fix is to add "static" where appropriate, of course. I've also
simplified calls to such methods to reflect the fact that they no
longer need a specific object to call the method on.
In projects that contain test inputs, I've left the non-static
declarations unchanged, and instead downgraded the warning to be
ignored. In all other projects, this warning has been upgraded to an
error.
In general, these diagnostics are now errors in projects for which all
such warnings have been fixed. There are three unfixed warnings in
two projects, so this diagnostic remains a warning (not an error) in
those projects.
There are also many places where rwa-types-usage warnings have been
locally suppressed using @SuppressWarnings annotations. I haven't
systematically revisited those to see if any can be fixed properly.
So for those projects this diagnostic must also remain a warning (not
an error), since @SuppressWarnings does not work on things Eclipse is
configured to treat as errors.
Along the way, I also converted many "for (;;)" loops into modern
"for (:)" loops. I didn't systematically look for all opportunities
to do this, though. I merely made this change where I was already
converting raw Iterator uses into modern Iterator<...> uses.
Better use of generics also allowed many casts to become statically
redundant. I have removed all such redundant casts.
Only three raw-types warnings remain after this batch of fixes. All
three involve raw uses of CallGraphBuilder. I've tried to fix these
too, but it quickly snowballs into a cascade of changes that may or
may not eventually reach a statically-type-save fixed point. I may
give these last few problem areas another go in the future. For now,
though, the hundreds of other fixes seem worth keeping even if there
are a few stragglers.
This commit may change some public APIs, but only by making weaker
type signatures stronger by replacing raw types with generic types.
For example, we may change something like "Set" into "Set<String>",
but we're not adding new arguments, changing any
underlying (post-generics-erasure) types, etc.
There are two such diagnostics: one for collection methods and one for
equals(). See
<https://www.eclipse.org/eclipse/news/4.7/jdt.php#unlikely-argument-types>
for more information about these two new diagnostics.
For each of these diagnostics, I've set the severity level to
"warning" in projects that have some instances of the suspicious code,
or to "error" in projects that have no instances of the suspicious
code.
In regular application code, these warnings should be taken seriously
and fixed. But for test code, it's better to keep things simple and
not add any methods that aren't strictly needed by the test.
These should mostly be things that we've already decided earlier that
we explicitly don't want to "fix" because they simply disagree with
the WALA project's coding style.
The additional diagnostics are ones that were previously being
ignored, but which we seem to have been ignoring by default rather
than as a conscious choice.
For diagnostics of which we currently have *zero* instances, treat
these as errors rather than merely warnings. The intent is to
permanently lock out future regressions of things we've completely
fixed. In the future, whenever we fix the last instance of a given
warning in a given Eclipse project, we should also promote that
diagnostic to an error to keep things clean into the future.
If a method is private, there's no risk that a subclass elsewhere
might be overriding it and depending on dynamic dispatch to choose the
right implementation. So all of these private methods can safely be
declared static without risk of regression in either WALA code or
unseen third-party code.
The "potentially" qualifier is here because these methods are visible
outside the WALA source tree. These methods may seem OK to be static
based on the code we have here, but we have no way of knowing whether
third-party code expected to be able to subclass and override. I'm
going to play it safe and assume that we want to allow that.
Note that we are still allowing Eclipse warnings about methods that
can *definitely* be declared static; a different configuration option
controls these. For private methods, final methods, and methods in
final classes, if the code seems static-safe based on what we have
here, then that's good enough: we don't need to worry about
third-party overrides.
In certain cases, one may want to ignore inter-procedural control
dependence. Consider the following example:
flag = getFlagVal();
if (flag) {
doStuff();
}
If we are ignoring interprocedural control dependence, a forward slice
from the first statement will *not* include statements inside doStuff()
and its transitive callees.
This option is useful in scenarios where the effects of statements
inside control-dependent callees can be accounted for via some cheaper
effect analysis. E.g., if you only care about heap effects of control-
dependent callees, you can compute that using mod-ref analysis,
rather than sucking all the control-dependent callee statements into the
slice.
Also added some more detailed comments, a new unit test, and removed
some trailing whitespace.
Manu Sridharan
Julian Dolby
Ben Liblit
Raffi Khatchadourian
Manu Sridharan