If a path to the module-info.java file is passed as an argument, we
should treat all other source files passed as arguments (either as
individual source files or inside a source directory) as members of that
module. Previously we treated other source files as members of the
unnamed module, and this resulted in incorrect errors when using a
member exported with a qualification from another named module, for
example
#KT-18598 In Fixed
Note that javac reports a nice error in this case ("package foo is
declared in module lib, which is not in the module graph"), but we only
report "unresolved reference" because the corresponding modules are not
added to classpath roots. We should improve this in the future
#KT-18598 In Progress
Files like ant-javafx.jar, deploy.jar, java.jnlp.jar, javafx-swt.jar etc
should not be added to the classpath if JDK home points to a JDK 9
distribution
In this commit, only IDE tests are added, because we look for module
declarations in the IDE across the whole project, whereas in the
compiler we should do this on the module path only and that requires
separate work (KT-18599) which is done in the following commits.
(The change in Cache.kt is needed so that
JvmModuleAccessibilityChecker.ClassifierUsage, which is an inner class,
would be injected properly.)
#KT-18598 In Progress
#KT-18599 In Progress
Previously we traversed all notLoadedRoots on each request for package
parts with the given package FQ name. Since notLoadedRoots might contain
a lot of roots (which never transition into "loadedModules" because e.g.
they are not Kotlin libraries, but just Java libraries or SDK roots with
the META-INF directory), this was potentially hurting performance. It
seems it's more optimal to compute everything eagerly once
JvmPackagePartProvider is constructed.
Another problem with the previous version of JvmPackagePartProvider was
that it did not support "updateable classpath" which is used by REPL and
kapt2, it only used the initial roots provided in the
CompilerConfiguration. In REPL specifically, we would thus fail to
resolve top-level callables from libraries which were dynamically added
to the execution classpath (via some kind of a @DependsOn annotation).
In the new code, JvmPackagePartProvider no longer depends on
CompilerConfiguration to avoid this sort of confusion, but rather relies
on the object that constructed it (KotlinCoreEnvironment in this case)
to provide the correct roots. This is also beneficial because the
computation of actual VirtualFile-based roots from the ones in the
CompilerConfiguration might get trickier with modular Java 9 roots
No tests added because the only way to test this seems to be to launch a
new process and send "Ctrl+D" into it, which seems cumbersome for such
small change
If the compiler is invoked through JPS, an instance of
XmlMessageRenderer is used and the output should be a valid XML.
Previously, we reported the argument parse error (and the note to try
"-help") before calling messageRenderer.renderPreamble, which resulted
in invalid XML. The same was happening for the usage printed on "-help".
Both of these problems are fixed
#KT-14848 Fixed
Invalid compiler arguments should lead to a compilation error, not
internal error. This has been fixed by splitting the
"parseArguments(...)" call (which does parsing + validation) in
CompileServiceImpl into two calls of parseCommandLineArguments +
validateArguments, and returning with a compilation error if the latter
did not succeed
#KT-16057 Fixed
#KT-14848 In Progress
According to the spec, "java.se" and every other non-"java.*" module
that exports at least one package without qualification, is a root.
Currently we only support compilation of a single unnamed module, and
apparently unnamed module should read all root modules.
#KT-18180 Fixed
Write the modular JDK (9+) path to the module.xml file passed to the
compiler from the JPS plugin. This path is then recorded in the compiler
configuration in KotlinToJVMBytecodeCompiler.configureSourceRoots. This
is needed because in JPS plugin, we pass "-no-jdk" and thus no JDK home
path was recorded in the compiler configuration in
K2JVMCompiler.setupJdkClasspathRoots. Presence of JDK home path in the
configuration is crucial for JDK 9 support (see
KotlinCoreEnvironment.Companion.createApplicationEnvironment), because
classes there can only be loaded with the special "jrt" file system, not
as .class files in .jar files
#KT-17801 Fixed
E.g. "kotlinc foo.kt test/Bar.java" will compile foo.kt, and
declarations from Bar.java will be accessible to Kotlin code in foo.kt.
The change in AbstractTopLevelMembersInvocationTest is needed because an
incorrect configuration was created in that test where a library jar was
also a Java source root (the compiler is never configured this way in
production), which led to an exception in
JavaCoreProjectEnvironment#addSourcesToClasspath
#KT-17697 Fixed
In this mode, javac AST and Symbol files are used during
Kotlin compilation instead of PSI / binary stuff.
Later, they are reused for Java file compilation.
javac in this mode is integrated into kotlinc.
Otherwise it might lead to the problems when reading class files
from the source scope (e.g. a groovy file or an already compiled java file)
Note that it may be reproduced only with Maven because in JPS
and Gradle we reading source java files even when they're shouldn't be
recompiled, while in Maven we read class files in the case
#KT-17897 Fixed
Backend: If kotlin class extends kotlin.collection.List
write it as it's super interface (light class mode only)
IDE: Provide wrapper classes to java resolve
that try to emulate backend behaviour
For example if kotlin class implements kotlin.collections.Map,
we provide a superinterface that has abstract 'getEntries' method
and 'entrySet' method that is considered default.
In reality all those methods are generated in the class itself.
In IDE supporting this case without hacks is not feasible performance-wise
since kotlin.collection.* may not be an immediate supertype and we need
to compute all supertypes just to calculate own methods of the class
- use nullable type "T?" instead of HandleResult consisting of the found
value and the "should continue" flag
- inline all local functions because they don't add any value now
Friend modules should be provided using the -Xfriend-modules flag
in the same format as -libraries. No manual configuration required for
JPS, Gradle and Maven plugins.
Friend modules could be switched off using the -Xfriend-modules-disabled
flag. Doing that will
* prevent internal declarations from being exported,
* values provided by -Xfriend-modules ignored,
* raise a compilation error on attemps to use internal declarations from other modules
Fixes #KT-15135 and #KT-16568.
Alexander Udalov