Sometimes we see a thrown IllegalStateException when CPU profiling,
where the underlying cause is a ConcurrentModificationException in
the underlying GraalVM internals.
This attempts to minimize the chances of this exception being thrown
by starting proiling after the evaluator and its context have been
created.
Motivation:
- VmObjectLike.iterateMemberValues() is an internal iterator whose loop body is a Java lambda.
It suffers from the following limitations:
- `continue` and `break` cannot be used.
- `return` becomes an (awkward) replacement for `continue` and `break`.
It cannot be used to return from the enclosing method.
- Mutating variables requires hoisting them (MutableBoolean, etc.).
- Node.execute() calls are very slow because they are behind a Truffle boundary.
- Callers of iterateMemberValues() do not specify what and how they want to iterate.
This obscures the iteration code and makes it difficult (often impossible) to optimize.
Changes:
- Add interface VmObjectCursor, an external iterator that iterates over VmObject members.
Its design is inspired by org.graalvm.collections.MapCursor.
- Add VmObjectCursor implementations specialized for different members, VmObject subclasses, and cursor options.
- Replace VmObjectLike.iterateMemberValues() with the following methods that return VmObjectCursor:
- properties()
- properties(CursorOption)
- elements()
- elements(CursorOption)
- entries()
- entries(CursorOption)
- members
- members(CursorOption)
- Replace all usages of iterateMemberValues() with one of the above methods.
Before: object.iterateMembersValues((key, member, value) -> { if (member.isProp()) {...} });
After : for (var cursor = object.properties(); cursor.advance(); ) {...}
Result:
- Iteration code is much simpler and clearer.
- Iteration code that calls Node.execute() is much faster because it is no longer behind a Truffle boundary.
- Other iteration code is potentially faster and can be better optimized in the future.
Co-authored-by: translatenix <119817707+translatenix@users.noreply.github.com>
Co-authored-by: Jen Basch <jbasch94@gmail.com>
This introduces a way to run CPU profiling.
The profiler produces output in pprof format.
CLI users can enable CPU profiling via command line flags.
Profiling is otherwise not offered as an API.
Co-authored-by: Jen Basch <jbasch94@gmail.com>
Co-authored-by: Islon Scherer <islonscherer@gmail.com>
* Add tasks to generate source jars for java executables
* native source bundles
* source bundles in separate plugin
* Wire up sourceBundle task
* Produce bundles on a per-variant basis (Java fat jar, each native variant)
* Introduce `pklNativeSourceBundle` and `pklJvmSourceBundle` plugins for handling fat jar vs. GraalVM bits
* Download compiled classes list, and exclude Java sources for executable stuff specifically
* Include native sources when appropriate (GraalVM, JDK, zlib, musl, libpkl)
---------
Co-authored-by: Jen Basch <jbasch@apple.com>
This adds an option which allows users to control exactly how
`@ConfigurationProperties` annotations are attached to classes.
This helps when:
* Spring Boot users don't have a single root class, but multiple
classes that are injected in disparate places
* Pkl's module structure doesn't match the resulting `.properties`
layout
This changes it to render the qualified name instead of the simple name.
* Module classes render the module name
* Classes within module render the module name and the class name joined by `#`
* Rendering of members of `pkl:base` does not change
For example:
* `new k8s.core.v1.Service {}` instead of `new ModuleClass {}`.
* `new k8s.core.v1.Service#ServiceSpec {}` instead of `new ServiceSpec {}`
* `new pkl.Project {}` instead of `new ModuleClass {}`
This affects the following:
* `toString()`
* `trace()`
* Error messages
* REPL
Currently the module cache directory in CLI commands (and thus in Gradle tasks, since they use CLI commands under the hood) is only used for the "primary" task. For example, for an eval command, the module cache directory will be used for the evaluation of the target module.
However, most commands also work with projects, and for projects the `PklProject` file must be evaluated. This is done via `Project.loadFromPath`, which internally builds an evaluator with the default module cache path. Therefore, all evaluations of project manifests always use default cache path instead of whatever is configured through the CLI option/Gradle task property. This, in turn, may result in unexpected behavior: in general, the user would expect that if they specify the cache directory, it will be used for all evaluations (e.g. when they pre-populate the cache and expect evaluations to work offline).
The same is true for settings evaluation, although for settings it is significantly less important since settings scripts can't have remote dependencies.
This PR propagates cache path into the project evaluator (as well as the settings evaluator, for consistency). This will result in the user-specified cache directory being used for `PklProject` evaluations as well.
Patterns that originate from the language runtime will add
`Pattern.UNICODE_CHARACTER_CLASS | Pattern.UNICODE_CASE` as implicit
regex flags, so we need to do the same on the binary decoding side.
* Compute a derived regex expression with `\z` anchor
* Ensure derived regex can't be messed up with comments (in extended mode) or quote start
Follow up to #1880
This adds logic to UnresolvedTypeNode to create a `UnionOfStringLiterals`
type node if a member is a typealias who is a string or a string literal union.
Also, this fixes an issue where marking a duplicate string literal with the
default marker throws an index out of bounds exception.
This makes various improvements to help truffle do partial
evaluation of code related to typechecking.
* Split `UnionTypeNode` into `UnionTypeNodeExploded` and
`UnionTypeNodeLooped` to avoid gigantic graphs. This especially helps
in the case of typealiases, because they are inlined into their usage
sites.
* Move constraint-failure exception construction in
`TypeConstraintNode` behind `@TruffleBoundary` helper methods
instead of `transferToInterpreterAndInvalidate`, since building the
exception is a rare, off-hot-path operation.
* Move `ClassNode`'s one-time class-initialization logic behind a
`@TruffleBoundary` method (`initClass`), store `VmLanguage` as a
field instead of recomputing it via `VmLanguage.get(this)`, and use
`transferToInterpreterAndInvalidate` since class init never needs to
re-run on the compiled path.
* Pass `VmLanguage` explicitly into `TypeTestNode` and `ClassNode`
(via `AstBuilder`) rather than resolving it with
`VmLanguage.get(this)` on a hot path.
* Fix various issues where the partial evaluator determines that a frame
should have been materialized but was not.
* Pre-initialize call targets in `ObjectMember.setMemberNode()` and
`FunctionLiteralNode` so the first invocation doesn't trigger a
compilation invalidation.
* Fix truffle node tree issues (don't mark `@RootNode` as child,
deepCopy instead of passing an already adopted node).
* Remove unnecessary interpreter transfers when throwing type mismatch exceptions
Islon Scherer
Jen Basch
Daniel Chao
Aashish Mahato
Goutam Adwant
Sam Gammon
Aagam Shah
Carpe-WangandJen Basch
Maycon Campos
Alexandre Kohler
Vladimir Matveev
dependabot[bot]andJen Basch
Rushikesh
apple-ospo[bot]