All of these involve conditionals that check some static, final debug
flag or debug level. The code will indeed be dead if WALA is built
with those debug facilities turned off. But we still want the code
present in case someone needs to turn some aspect of debugging on for
a while.
This fixes 33 out of 37 Eclipse "Potential resource leak: '...' may
not be closed" warnings. It also fixes 3 out of 37 Eclipse "Resource
'...' should be managed by try-with-resource" warnings, although that
was not the main focus of this effort.
The remaining 4 warnings about potential resource leaks all involve a
leaked JarFile instance that is passed to a JarFileModule constructor
call. JarFileModile never attempts to close its underlying JarFile;
this code is written as though JarFile cleanup were the caller's
responsibility. However, the JarFile often cannot be closed by the
code that creates the JarFileModule either, since the JarFile needs to
remain open while the JarFileModule is in use, and some of these
JarFileModules stay around beyond the lifetime of the code that
created them. Truly fixing this would essentially require making
JarFileModule implement Closeable, which in turn would probably
require that Module implement Closeable, which in turn would require
changes to lots of code that deals with Module instances to arrange
for them to be properly closed. That's more invasive than I'm
prepared to take on right now.
Instead, rely on Java's ability to infer type parameters in many
contexts. This removes 665 Eclipse warnings.
Note: a few of these changes are to files under "test" subdirectories.
Presumably those are files that serve as test inputs rather than being
part of WALA code proper. As far as I can tell, these changes do not
break any WALA tests. But if any of those tests were specifically
intended to exercise WALA on code with non-inferred generic type
parameters, then I really should be leaving those alone.
Dalvik bytecode represents 'null' as 0 which may lead to problems
in phi instructions. This variant of TypeInference fixes these
problems by
- tracking whether an SSA value is constant zero and
- ignoring constant zeros in the transfer function of phi instructions
when meeting with non-primitive types
Some of these might have proper DTDs or XML Schema definitions
floating around somewhere that we could use. Presumably many do not.
Rather than hand-craft such definitions myself, I'm just giving each a
minimal stub DOCTYPE declaration. That's enough to satisfy Eclipse's
XML validator, which otherwise complains that these files lack grammar
constraints.
If a client violates these restrictions, I prefer that their code fail
at compile time instead of run time. Changing a few key types from
`Class` to `Class<? extends AbstractAndroidModel>` gives us precisely
the static enforcement we need and lets us remove an
`AbstractAndroidModel.class.isAssignableFrom` run-time check.
However, this does change the public API of `AndroidEntryPointManager`
in two ways. The `getModelBehavior` and `setModelBehavior` methods now
respectively accept and return `Class<? extends AbstractAndroidModel>`
instead of `Class`. Is tightening up a public API in this manner
considered OK?
Ben Liblit
Julian Dolby
Martin Mohr
Martin Hecker
Julian Dolby
Manu Sridharan