[
https://issues.apache.org/jira/browse/GROOVY-12555?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18145966#comment-18145966
]
ASF GitHub Bot commented on GROOVY-12555:
-----------------------------------------
jdaugherty commented on PR #3084:
URL: https://github.com/apache/groovy/pull/3084#issuecomment-6931817145
Thanks, Paul. I took the proposal as written. For item 2 I kept one default
everywhere and documented GSE instead. The branch is rebased on master, and the
new commits are `cafadeda54` (option, lookup and tests) and `3b35ce66af`
(notes). The 6.0.x port branch has the same change.
**Loader history.** Agreed: it applies with or without sources, so I've
dropped it as an argument. The notes now say that a reused `GroovyShell` or
GSE, groovysh and the console can resolve a name to a class that the loader
compiled earlier. With `defaultPackageSources=false` this goes away as well,
since the class-file lookup doesn't use `loadClass`.
`testClassCompiledEarlierByTheClassLoader` covers it.
**1. `defaultPackageSources`.**
`CompilerConfiguration.DEFAULT_PACKAGE_SOURCES` (`defaultPackageSources`,
system property `groovy.resolve.default.package.sources`) is on unless set to
`false`, and ignored when `defaultPackageFirst` is `false`. When it is `false`,
`resolveFromDefaultPackage` calls a new package-private
`ClassNodeResolver.resolveClassFile`. That method reuses the existing
`readClassFile`: `loader.getResource(name + ".class")`, then `new
DecompiledClassNode(AsmDecompiler.parseClass(url), new
AsmReferenceResolver(this, compilationUnit))`, with the same name-mismatch
check.
- A hit goes through `cacheClass`.
- A miss is not cached as `NO_CLASS`. It goes into a separate set in the
resolver, so a name that occurs many times in one unit doesn't call
`getResource` each time, and the later `resolveToOuter` still finds a sibling
source after the default imports
(`testDefaultPackageSourceFoundWithoutSources`).
- It doesn't go through `findClassNode`, so GSE's override and its roots are
not involved.
I counted `ResourceConnector` calls for a GSE script that uses 12
default-import types: 76 with the sources on, 52 with them off, and 52 with
`defaultPackageFirst=false`. That matches your two extra lookups per type, and
none with the sources off.
**2. GSE default.** I kept one default everywhere. Having GSE resolve the
same script differently from `groovyc` seemed worse than documenting the
behaviour. A GSE-only default would also have to copy the caller's
`CompilerConfiguration` in `ScriptClassLoader` for every compile, to avoid
changing a configuration it doesn't own. `COMPATIBILITY.md` describes the mixed
results, the stale results until the engine is recreated, and the two extra
root lookups per name, with `defaultPackageSources=false` as the remedy.
`testScriptEngineRootSource` checks both: by default the root's `Date`, with
the sources off `java.util.Date`.
**3. Gradle.** Agreed. The Groovy 7 entry now says an incremental Gradle
build gives the same result either way, and gives the route
`groovyOptions.optimizationOptions['defaultPackageSources']`.
**4. Tests.** With the sources off, `Groovy12555` now covers:
- a compiled class, which still comes before the default import;
- a sibling source, which now comes after it;
- a sibling source whose name isn't a default import, which is still found;
- a nested class.
It also adds the GSE case and the `GroovyShell` loader case, and
`CompilerConfigurationTest` covers the new system property. The tests for the
sibling source, GSE and the loader fail without the change. Core `:test` passes
locally on both branches.
**5. Notes.** The Groovy 7 and 6.0.1 entries in `COMPATIBILITY.md` and the
note in `core-program-structure.adoc` now cover the option and the cases above.
In the 6.0.x port both options are off unless set to `true`, so
`defaultPackageFirst=true` alone gives the class-file-only lookup there. That
is enough for the Gradle case.
**6. Housekeeping.** Rebased on current master. In both conflicts,
`COMPATIBILITY.md` and the system-property table, I kept both sides. The 6.0.x
port is rebased on `GROOVY_6_0_X` as well.
@daniellansun, could you take another look? Since your approval there's the
nested-name commit (`289dfdeca2` after the rebase, `b96fa00c30` before it) and
the two commits above.
> A default-package class resolves a precompiled default-package class to a
> default import of the same name
> ---------------------------------------------------------------------------------------------------------
>
> Key: GROOVY-12555
> URL: https://issues.apache.org/jira/browse/GROOVY-12555
> Project: Groovy
> Issue Type: Bug
> Components: Compiler
> Affects Versions: 4.0.33, 5.1.3
> Reporter: James Daugherty
> Priority: Major
>
> A class in the default package that refers to another class of the default
> package by its simple name finds that class while both are compiled together.
> When the other class is already compiled, the name resolves to a class of the
> default imports with the same simple name instead, here
> {{groovy.util.Factory}}, and the call fails at run time. In Java, a class of
> the same package, the unnamed one included, takes precedence over on-demand
> imports such as {{java.lang.*}}.
> {code:groovy}
> // Factory.groovy, default package
> class Factory {
> static String hello() { 'hello from the Factory of this project' }
> }
> // Caller.groovy, default package
> class Caller {
> static void main(String[] args) {
> println Factory.hello() // MissingMethodException for
> groovy.util.Factory when Factory is already compiled
> }
> }
> {code}
> Reproduced with 4.0.33, 5.1.3 and master. A reproducer with a groovyc script
> and an incremental Gradle build:
> https://github.com/jdaugherty/groovy-default-package-resolution
> *Cause*
> {{ResolveVisitor.resolve}} tries, in order, the classes of the module, of the
> compilation unit, of the default imports, and then the class path. For a
> module of a named package, the module step also looks up {{package.Name}} on
> the class path, so a compiled class of the same package comes first. For the
> default package there is no such step, so the default imports come before a
> compiled class of the default package.
> *Proposed fix*
> For a class of the default package, look up the simple name on the class path
> after the compilation unit and before the default imports, as is done for a
> named package. The lookup is behind the optimization option
> {{defaultPackageFirst}} ({{CompilerConfiguration.DEFAULT_PACKAGE_FIRST}};
> system property {{groovy.resolve.default.package.first}}; in Gradle,
> {{groovyOptions.optimizationOptions['defaultPackageFirst']}}, the only one of
> Gradle's settings that reaches the compiler's optimization options): on by
> default in Groovy 7, and an opt-in in 6.0.x.
> *Compatibility*
> For a compilation unit in the default package, a simple type name that is not
> declared or imported in the unit and not in the same {{CompilationUnit}} is
> looked up on the class path, and on the source path of the
> {{GroovyClassLoader}}, as a default-package class before the default imports
> ({{java.lang}}, {{java.util}}, {{java.io}}, {{java.net}}, {{java.time}},
> {{groovy.lang}}, {{groovy.util}}, {{BigInteger}}, {{BigDecimal}}). A
> default-package {{List}} or {{Node}} on the class path therefore shadows
> {{java.util.List}} or {{groovy.util.Node}} in default-package code, where the
> default import used to win. This matches JLS ยง6.4.1 (a type declared in the
> same package, the unnamed package included, shadows a type imported on
> demand), except that an explicit star import still takes precedence, as
> elsewhere in Groovy; it is also what Groovy already does when the class is in
> the same {{CompilationUnit}}, and what {{ResolveVisitor.resolveFromModule}}
> does for a named package. The change is not limited to incremental builds:
> the {{groovy}} command uses the current directory as the class path when none
> is given, so in a directory of scripts a sibling {{Node.groovy}},
> {{Timer.groovy}} or {{Math.groovy}} stands for that name in the other
> scripts, and a sibling source that does not compile becomes a compile error
> for a script that uses its name; the roots of a {{GroovyScriptEngine}} are
> source paths of the same kind. Each default-package compilation unit also
> pays one class-path lookup per default-import type it names, as a
> named-package unit always has, and a {{GroovyShell}}, {{Eval}} or
> {{GroovyScriptEngine}} pays it per script, since the lookup cache is per
> unit. The entries in COMPATIBILITY.md have the details and the remedies.
> Found while making incremental builds of a Grails application produce the
> same classes as a full build; see also GROOVY-12550 and GROOVY-12544.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)