The optimization, and therefore the change in the behavior
of Relation, is now needed by the class Logic, and cannot
be reverted.
This patch (written by Josh) therefore changes the
implementation of setAll() so that _1s is no longer used.
The previous implementation of TextAnalysisFormat contained the list
of all the existing relations that sbt knew of, and used this
information to write to and read from the disk the persisted analyses.
In this knew implementation, TextAnalysisFormat gets from the
Relations object what are the existing relations, and then persists
them to disk.
The previous situation was not optimal since it meant that, in order
to add a new dependency kind, one had to modify both the Relations
and TextAnalysisFormat.
Using this new implementation, no change to TextAnalysisFormat is
required whenever a new dependency kind is added.
The implementation of Relation should in theory make no difference
whether an element is unmapped, or whether it is mapped to an empty
set. One of the changes in 322f6de655
introduced an optimization to the '+' operation on Relations that,
in theory, should have made no difference to the semantic.
The result of that optimization is that some mappings of the form
"elem -> Set()" are no longer inserted in the forwardMap of the
Relation.
Unfortunately, the change resulted in the breakage of #1430,
causing "set every" to behave incorrectly. There must be, somewhere
in the code, a test on the presence of a key rather than an access
via <relation>.get(), or some other access that bypasses the
supposed semantic equivalence described above. I spent several
hours trying to track down exactly the offending test, without
success.
By undoing the relevant change in 322f6de655, "set every"
works again. That however offers no guarantee that everything else
will keep working correctly; the underlying quirk in the code that
depends on this supposedly inessential detail is also still
lurking in the code, which is less than ideal.
In some cases, expanded macros report that their original tree and
its expansion are the same, thus creating a cyclic chain. This chain
may then produce a SOE during dependencies or used names extraction.
This kind of problem was already reported in sbt/sbt#1237 and
sbt/sbt#1408. Unfortunately, the fix that was applied to the
dependencies extraction part was not sufficient.
Mark test 'source-dependencies/macro' as passing
Fixes#1544
It turns out there was a very subtle, and evil, issue sitting the Ivy/maven configuration, and it
related to dependency mapping. A mapping of `foo->bar(*)` means that the local configuration
`foo` depends on the remote configuration `bar`, if it exists, or *ALL CONFIGURATIONS* if `bar`
does not exist. Since the default Ivy configuration mapping was using the random `master`
configuration, which AFAICT is NEVER specified, just an assumed default, this would cause leaks
between maven + ivy projects.
i.e. if a maven POM depends on a module denoted by an ivy.xml file, then you'd wind up accidentally
bleeding ALL the ivy module's configurations into the maven module's configurations.
This fix works around the issue, by assuming that if there is no `master` configuration, than the
maven default of `compile` is intended. As sbt forces generated `ivy.xml` files to abide by
maven conventions, this works in all of our test cases. The only scenario where it wouldn't work
is those who have custom ivy.xml files *and* have pom.xml files which rely on those custom ivy.xml files,
a very unlikely situation where the workaround is: "define a master configuration".
Includes a test demonstrating the issue.
The issue is that when you manually set a ScalaInstance, i.e. not one from Ivy, the
classpath which is returned for any given configuration ONLY uses Ivy. This means that
the legitimate Scala JAR files that need to be on the classpath are missing from the list.
For some reason, the way we instantiate tests uses an unfiltered classloader against the
ScalaInstance, *BUT* the thread-context-classloader DOES use a filtered instance by
classpath. This add the hook into the TestFramework runner creation so that
the classpath accurately reflects the jars needed.
cc @rkuhn
Josh Suereth
eugene yokota
Jean-Rémi Desjardins
Antonio Cunei
Grzegorz Kossakowski
Martin Duhem