Since SSAPhiInstructions are never visited by NullPointerTransferFunctionProvider.TransferFunctionSSAVisitor,
we now respect phi instructions present at a given block by providing additional NodeTransferFunctions, improving precision.
Formerly, meets would lead to incorrect results due to incorrect initialization of initial data flow facts.
These are now properly initialized, interpreting
"State.BOTH" to mean: both "null" and "non-null" are possible values for the given variable, and
"State.UNKNOWN" to be the absurd assertion.
The initial fact at the entry block assumes variables to be BOTH, other blocks are initialy assumed unreachable and hence their variables to be UNKNOWN.
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
The setting should comply with the comment. Plus,
turning it on seems to lead to some unsoundness because
exception points-to sets become empty but should not be
Access is provided via corresponding methods in FieldImpl, ShrikeCTMethod and ShrikeClass.
Since we do not currently have implementation of these methods for front-ends other than Shrike, these new methods are not yet made available in the corresponding interfaces.
Previously, the various Eclipse projects' Java configurations used
mixtures of 1.6, 1.7, and 1.8. Many were internally inconsistent,
such as requiring 1.7 in "MANIFEST.MF" but 1.6 in the Eclipse JDT
build preferences. The Travis-CI configuration tests against both 1.7
and 1.8, but does not test against 1.6.
Across all projects, the most common version was 1.7. So I'm going to
assume that 1.7 is the intended build target. This commit makes 1.7
the selected version nearly everywhere.
"com.ibm.wala.core.testdata" is the one exception. This specific
project uses a few features only found in 1.8, such as lambda
expressions. Previously, "com.ibm.wala.core.testdata" used 1.7 in
some aspects of its configuration but 1.8 in others. Now it
consistently targets 1.8. I wish this one project didn't need to be
inconsistent with the rest of WALA, but at least now it's consistent
with itself.
(Personally, I'd be happy to target 1.8 only. But my impression
across all of these configuration files is that the WALA developers
still want to be compatible with 1.7. If that is no longer a
requirement, let me know and I will adjust these changes accordingly
to target 1.8 only.)
This change eliminates 11 "There is no 'jre.compilation.profile' build
entry and the project has Java compliance preferences set" warnings
and 13 "The JRE container on the classpath is not a perfect match to
the 'JavaSE-1.7' execution environment" warnings. However, it also
adds 450 "Redundant specification of type arguments <...>" warnings
and 17 "Resource '...' should be managed by try-with-resource"
warnings. So this seems like a net step backward in my wish to reduce
WALA warnings. However, those new warnings concern Java 1.7 language
features that we were not previously using to good effect in projects
that targeted 1.6. If we all agree that we can now target 1.7
instead, then we can use these helpful features as the newly-added
warnings suggest. So I call that a step in the right direction.
Some source files here definitely use Hamcrest, so listing it as a
dependency seems reasonable. What I find confusing is the inconsistency
among my Eclipse installations. On some of my various machines, Eclipse
reports an error if this dependency is not listed. On others, Eclipse
finds the required jar and reports no error, even if this dependency is
not listed. I don't know why the latter works, or why the inconsistency
exists at all. Eclipse is a complex, subtle beast. What I can say is
that this change fixes the error for my Eclipses that were reporting an
error, and does not introduce any new errors for my Eclipses that were
already happy before this change.