For some reason I sort of blocked from my mind the fact that none of the actual
tests were running in GWT. Really I was just testing that things compiled.
So now the tests actually run in GWT. This means I had to split the tests that
rely on method/field lookup via reflection into a separate file because that
stuff will never work with GWT. I also changed a bunch of the other tests not
to surreptitiously use reflection because they were ostensibly testing
something else, so no point in those tests being GWT-incompatible just for
kicks.
This also means we can't use JUnit 4 annotations (sigh, I wish GWT would step
out of the stone age, but that's a whole other fiasco that I don't even want to
think about). So some of the tests that were using @Test(expected=FooException)
had to be rewritten to manually catch the exception and fail if it was not
thrown.
Anyhow, now most of the tests are in fact running in GWT (when you do "mvn
integration-test") and the test that I expected should fail in GWT is in fact
failing. Yay!
There's unfortunately no sane way to do the same reflection-like array access
in GWT, so we just have to enumerate all of the primitive types.
I could have just left this only in the Java backend, but I would prefer to
avoid differences between behavior on the JVM and behavior in GWT whenever
possible.
Fortunately this did not require using any JDK 1.8-specific APIs, just
searching (via reflection) through the set of interfaces implemented by a class
in addition to its super-classes.
One downside is that the transitive set of all implemented interfaces could be
much larger than the super-class chain, but this is the last thing we search,
so in theory we only even get here if we were about to fail to find the getter
entirely.
This also means that our tests have to be compiled and run with Java 8, because
I was not in the mood to figure out how to do all the complex Maven jockeying
to have a special JDK8 profile and disable JDK8 tests in the normal profile and
re-enable them in the JDK8 profile and blah blah blah, just shoot me now.
Section behavior used to be somewhat hackily determined from nullValue and
defaultValue, but it was messy. Now it has its own configuration and things are
more explicit.
Previously missing/null sections were treated as false, but if you set
nullValue to something, then missing sections became an error, but null
sections remained false. If you set defaultValue to something, missing sections
went back to being omitted. Kooky.
Now missing sections are omitted, unless you set strictSections to true, in
which case they become an error. nullValue and defaultValue have no impact on
section behavior.
There is no way to make null sections an error, but I don't see any demand for
that, so I'm going to leave that as a TODO.
Closes#60.
Mainly we also needed to trim around comment tags, but also there were a bunch
of tricky edge cases that were only partially addressed. This is some edgy
business!
The specs tests assumes an implementation will silently ignore invalid delims.
I think that's a bad idea, so I throw an exception on invalid delims. So I just
ignore those "tests".
This allows one to have a map with keys like "foo.bar.baz" even when not in
standards mode. If a whole key is missing, then we attempt to decompose it and
look it up piecewise.
Closes#55.
That's what collectors are for. This also avoids the needless expense of doing
Integer.parseInt on every single key (the vast majority of which will not
actually be numbers), every single time we do a context lookup.
Michael Bayne