Put `Nothing` into an empty array rather than `null`. When running with `assert` on (unit tests), check the content of the array and `AssertError` quickly when a `null` is found.
Critical performance improvements after #4067
# Important Notes
- Replace if-then-else expressions in `Any.==` with case expressions.
- Fix caching in `EqualsNode`.
- This includes fixing specializations, along with fallback guard.
Creating two `findExceptionMessage` methods in `HostEnsoUtils` and in `VisualizationResult`. Why two? Because one of them is using `org.graalvm.polyglot` SDK as it runs in _"normal Java"_ mode. The other one is using Truffle API as it is running inside of partially evaluated instrument.
There is a `FindExceptionMessageTest` to guarantee consistency between the two methods. It simulates some exceptions in Enso code and checks that both methods extract the same _"message"_ from the exception. The tests verifies hosted and well as Enso exceptions - however testing other polyglot languages is only possible in other modules - as such I created `PolyglotFindExceptionMessageTest` - but that one doesn't have access to Truffle API - e.g. it doesn't really check the consistency - just that a reasonable message is extracted from a JavaScript exception.
# Important Notes
This is not full fix of #5260 - something needs to be done on the IDE side, as the IDE seems to ignore the delivered JSON message - even if it contains properly extracted exception message.
Implements the #5643 idea. As soon as `MainModule` creates `Context` for GraalVM execution, it schedules a background task to initialize JavaScript. The initialization finishes sooner than Enso compiler is ready to work, saving time when it is actually needed.
# Important Notes
Only modifies boot sequence of `MainModule` (used in the IDE) and `VerifyJavaScriptIsAvailableTest` (to verify the _"context passing logic"_ works OK between threads). Regular CLI execution remains unchanged for now assuming batch execution may not need JavaScript in all the cases and if it does the initialization speed isn't that critical.
Serialization of FQNs' metadata was broken - we attempted to serialiaze concrete modules and that's against the design.
Added a test illustrating the problem which would previously fail during serialization.
Closes#5037
# Important Notes
It is still hard to discover problems like this because SerializationManager creates system threads; when the exception occurs, it is typically not the real cause. Creating regular threads via `createThread` seems to be problematic for `native-image`. There is a simple workaround for the former but will leave it out for another PR to simplify the review.
Add `Comparator` type class emulation for all types. Migrate all the types in stdlib to this new `Comparator` API. The main documentation is in `Ordering.enso`.
Fixes these pivotals:
- https://www.pivotaltracker.com/story/show/183945328
- https://www.pivotaltracker.com/story/show/183958734
- https://www.pivotaltracker.com/story/show/184380208
# Important Notes
- The new Comparator API forces users to specify both `equals` and `hash` methods on their custom comparators.
- All the `compare_to` overrides were replaced by definition of a custom _ordered_ comparator.
- All the call sites of `x.compare_to y` method were replaced with `Ordering.compare x y`.
- `Ordering.compare` is essentially a shortcut for `Comparable.from x . compare x y`.
- The default comparator for `Any` is `Default_Unordered_Comparator`, which just forwards to the builtin `EqualsNode` and `HashCodeNode` nodes.
- For `x`, one can get its hash with `Comparable.from x . hash x`.
- This makes `hash` as _hidden_ as possible. There are no other public methods to get a hash code of an object.
- Comparing `x` and `y` can be done either by `Ordering.compare x y` or `Comparable.from x . compare x y` instead of `x.compare_to y`.
An exception encountered during serialization prevents engine from continuing because it enters an infinite loop(!).
# Important Notes
The aim of this PR is to make it possible for engine to recover from the serialization failures. Any failure would mean that we enter an infinite loop in deserialization which is in turn waiting for the serialization to finish (which will never happen).
In this particular case FQNs are [referencing concrete modules](https://github.com/enso-org/enso/issues/5037). A separate PR will address that.
Expressions returning polyglot values were not reporting the type of the result because we have to do additional magic that infers the correct Enso type. Since this is exactly what `TypeOfNode` does, I re-used the logic.
Straightforward solution failed in tests because of assertions:
```
[enso] WARNING: Execution of function main failed (Invalid library usage. Cached library must be adopted by a RootNode before it is executed.).
java.lang.AssertionError: Invalid library usage. Cached library must be adopted by a RootNode before it is executed.
```
That is why this PR replaces `ExecutionEventListener` with `ExecutionEventNodeFactory`.
# Important Notes
Usage of `TypeOfNode` for programs that **do not** import stdlib means that we report types that do not involve stdlib e.g.
`Standard.Builtins.Main.Integer` instead of `Standard.Base.Data.Numbers.Integer`. While surprising, this is correct and I would say desirable. While reviewing the code, notice the difference in expectations in our runtime tests.
- Fixes the display of Date, Time_Of_Day and Date_Time so doesn't wrap.
- Adjust serialization of large integer values for JS and display within table.
- Workaround for issue with using `.lines` in the Table (new bug filed).
- Disabled warning on no specified `separator` on `Concatenate`.
Does not include fix for aggregation on integer values outside of `long` range.
Closes#5038
- Use the proper widget structure.
- Provide new method `get_widget_json` with whole structure, but keep `get_full_annotations_json` in old form.
- Start to get to some reusable functions.
- Added widget to JS_Object field selections.
- New `set` function design - takes a `Column` and works with that more easily and supports control of `Set_Mode`.
- New simple `parse` API on `Column`.
- Separated expression support for `filter` to new `filter_by_expression` on `Table`.
- New `compute` function allowing creation of a column from an expression.
- Added case sensitivity argument to `Column` based on `starts_with`, `ends_with` and `contains`.
- Added case sensitivity argument to `Filter_Condition` for `Starts_With`, `Ends_With`, `Contains` and `Not_Contains`.
- Fixed the issue in JS Table visualisation where JavaScript date was incorrectly set.
- Some dynamic dropdown expressions - experimenting with ways to use them.
- Fixed issue with `.pretty` that wasn't escaping `\`.
- Changed default Postgres DB to `postgres`.
- Fixed SQLite support for starts_with, ends_with and contains to be consistent (using GLOB not LIKE).
https://github.com/enso-org/enso/pull/3764 introduced static wrappers for instance methods. Except it had a limitation to only be allowed for types with at least a single constructor.
That excluded builtin types as well which, by default, don't have them. This limitation is problematic for Array/Vector consolidation and makes builtin types somehow second-citizens.
This change lifts the limitation for builtin types only. Note that we do want to share the implementation of the generated builtin methods. At the same time due to the additional argument we have to adjust the starting index of the arguments.
This change avoids messing with the existing dispatch logic, to avoid unnecessary complexity.
As a result it is now possible to call builtin types' instance methods, statically:
```
arr = Array.new_1 42
Array.length arr
```
That would previously lead to missing method exception in runtime.
# Important Notes
The only exception is `Nothing`. Primarily because it requires `Nothing` to have a proper eigentype (`Nothing.type`) which would messed up a lot of existing logic for no obvious benefit (no more calling of `foo=Nothing` in parameters being one example).
For graying out the nodes, the engine need to send IDE a set of values that will be computed before executing the program (and the IDE colors them gray). In general, it is tricky to do because we cannot know for sure which exactly nodes will be computed without running the program. But we can estimate based on the invalidated values, which nodes are expected to be executed during the next run, and send them to the IDE. This logic is simpler than the previous approach, and turned out working pretty well in practice.
[Peek-gray-out.webm](https://user-images.githubusercontent.com/357683/215092755-0010e41d-a2cf-447a-900e-4619408effa0.webm)
Enso unit tests were running without `-ea` check enabled and as such various invariant checks in Truffle code were not executed. Let's turn the `-ea` flag on and fix all the code misbehaves.
LS needs to notify runtime that modules' sources need to be reloaded from FS, once its own buffer has been reloaded as well.
# Important Notes
Discovered during integration of https://github.com/enso-org/enso/pull/4050.
The test illustrates the problem if we don't reload module's sources - the sources essentially become stale even though they have changed.
- add: `GeneralAnnotation` IR node for `@name expression` annotations
- update: compilation pipeline to process the annotation expressions
- update: rewrite `OverloadsResolution` compiler pass so that it keeps the order of module definitions
- add: `Meta.get_annotation` builtin function that returns the result of annotation expression
- misc: improvements (private methods, lazy arguments, build.sbt cleanup)
Introduces unboxed (and arity-specialized) storage schemes for Atoms. It results in improvements both in memory consumption and runtime.
Memory wise: instead of using an array, we now use object fields. We also enable unboxing. This cuts a good few pointers in an unboxed object. E.g. a quadruple of integers is now 64 bytes (4x8 bytes for long fields + 16 bytes for layout and constructor pointers + 16 bytes for a class header). It used to be 168 bytes (4x24 bytes for boxed Longs + 16 bytes for array header + 32 bytes for array contents + 8 bytes for constructor ptr + 16 bytes for class header), so we're saving 104 bytes a piece. In the least impressive scenarios (all-boxed fields) we're saving 8 bytes per object (saving 16 bytes for array header, using 8 bytes for the new layout field). In the most-benchmarked case (list of longs), we save 32 bytes per cons-cell.
Time wise:
All list-summing benchmarks observe a ~2x speedup. List generation benchmarks get ~25x speedups, probably both due to less GC activity and better allocation characteristics (only allocating one object per Cons, rather than Cons + Object[] for fields). The "map-reverse" family gets a neat 10x speedup (part of the work is reading, which is 2x faster, the other is allocating, which is now 25x faster, we end up with 10x when combined).
`@` should not be legal to use as a binary operator. I accepted it in the parser because it occurred in the .enso sources, but it was actually used to create a syntax error to test error recovery.
See: https://www.pivotaltracker.com/story/show/184054024
* Hash codes prototype
* Remove Any.hash_code
* Improve caching of hashcode in atoms
* [WIP] Add Hash_Map type
* Implement Any.hash_code builtin for primitives and vectors
* Add some values to ValuesGenerator
* Fix example docs on Time_Zone.new
* [WIP] QuickFix for HashCodeTest before PR #3956 is merged
* Fix hash code contract in HashCodeTest
* Add times and dates values to HashCodeTest
* Fix docs
* Remove hashCodeForMetaInterop specialization
* Introduce snapshoting of HashMapBuilder
* Add unit tests for EnsoHashMap
* Remove duplicate test in Map_Spec.enso
* Hash_Map.to_vector caches result
* Hash_Map_Spec is a copy of Map_Spec
* Implement some methods in Hash_Map
* Add equalsHashMaps specialization to EqualsAnyNode
* get and insert operations are able to work with polyglot values
* Implement rest of Hash_Map API
* Add test that inserts elements with keys with same hash code
* EnsoHashMap.toDisplayString use builder storage directly
* Add separate specialization for host objects in EqualsAnyNode
* Fix specialization for host objects in EqualsAnyNode
* Add polyglot hash map tests
* EconomicMap keeps reference to EqualsNode and HashCodeNode.
Rather than passing these nodes to `get` and `insert` methods.
* HashMapTest run in polyglot context
* Fix containsKey index handling in snapshots
* Remove snapshots field from EnsoHashMapBuilder
* Prepare polyglot hash map handling.
- Hash_Map builtin methods are separate nodes
* Some bug fixes
* Remove ForeignMapWrapper.
We would have to wrap foreign maps in assignments for this to be efficient.
* Improve performance of Hash_Map.get_builtin
Also, if_nothing parameter is suspended
* Remove to_flat_vector.
Interop API requires nested vector (our previous to_vector implementation). Seems that I have misunderstood the docs the first time I read it.
- to_vector does not sort the vector by keys by default
* Fix polyglot hash maps method dispatch
* Add tests that effectively test hash code implementation.
Via hash map that behaves like a hash set.
* Remove Hashcode_Spec
* Add some polyglot tests
* Add Text.== tests for NFD normalization
* Fix NFD normalization bug in Text.java
* Improve performance of EqualsAnyNode.equalsTexts specialization
* Properly compute hash code for Atom and cache it
* Fix Text specialization in HashCodeAnyNode
* Add Hash_Map_Spec as part of all tests
* Remove HashMapTest.java
Providing all the infrastructure for all the needed Truffle nodes is no longer manageable.
* Remove rest of identityHashCode message implementations
* Replace old Map with Hash_Map
* Add some docs
* Add TruffleBoundaries
* Formatting
* Fix some tests to accept unsorted vector from Map.to_vector
* Delete Map.first and Map.last methods
* Add specialization for big integer hash
* Introduce proper HashCodeTest and EqualsTest.
- Use jUnit theories.
- Call nodes directly
* Fix some specializations for primitives in HashCodeAnyNode
* Fix host object specialization
* Remove Any.hash_code
* Fix import in Map.enso
* Update changelog
* Reformat
* Add truffle boundary to BigInteger.hashCode
* Fix performance of HashCodeTest - initialize DataPoints just once
* Fix MetaIsATest
* Fix ValuesGenerator.textual - Java's char is not Text
* Fix indent in Map_Spec.enso
* Add maps to datapoints in HashCodeTest
* Add specialization for maps in HashCodeAnyNode
* Add multiLevelAtoms to ValuesGenerator
* Provide a workaround for non-linear key inserts
* Fix specializations for double and BigInteger
* Cosmetics
* Add truffle boundaries
* Add allowInlining=true to some truffle boundaries.
Increases performance a lot.
* Increase the size of vectors, and warmup time for Vector.Distinct benchmark
* Various small performance fixes.
* Fix Geo_Spec tests to accept unsorted Map.to_vector
* Implement Map.remove
* FIx Visualization tests to accept unsorted Map.to_vector
* Treat java.util.Properties as Map
* Add truffle boundaries
* Invoke polyglot methods on java.util.Properties
* Ignore python tests if python lang is missing
Added a separate pass, `FullyQualifiedNames`, that partially resolves fully qualified names. The pass only resolves the library part of the name and replaces it with a reference to the `Main` module.
There are 2 scenarios that could be potentially:
1) the code uses a fully qualified name to a component that has been
parsed/compiled
2) the code uses a fully qualified name to a component that has **not** be
imported
For the former case, it is sufficient to just check `PackageRepository` for the presence of the library name.
In the latter we have to ensure that the library has been already parsed and all its imports are resolved. That would require the reference to `Compiler` in the `FullyQualifiedNames` pass, which could then trigger a full compilation for missing library. Since it has some undesired consequences (tracking of dependencies becomes rather complex) we decided to exclude that scenario until it is really needed.
# Important Notes
With this change, one can use a fully qualified name directly.
e.g.
```
import Standard.Base
main =
Standard.Base.IO.println "Hello world!"
```
Unsetting literal sources during the `text/fileClose` operation would also unlink sources from the actual file. That would be never set back during the reopening of the file, resulting in an NPE during application of changes at a later point.
# Important Notes
Revealed during https://github.com/enso-org/enso/pull/4050 which was the first(?) to do close/open of the file.
The exception showing up in GUI and preventing further actions is now gone.
The new `TreeToIr` conversion class inherited usage of `UnhandledEntity` to signal parser errors from the previous `AstToIr` Scala based convertor. Over the time we were improving error recovery - however only _step by step_. This pull request exterminates _all the panics_ for once and forever!
# Important Notes
Unlike Scala, Java has concept of [checked exceptions](http://wiki.apidesign.org/wiki/Checked_exception) - an exception that has to be caught. This PR introduces new checked `SyntaxException` and throws it instead of unchecked `UnhandledEntity`. Because the exception is checked, each method either needs to declare it in its signature or catch it and handle it. No exception can escape or disappear. The main conversion method `TreeToIr.translateModule` doesn't propagate the `SyntaxException`. That provably demonstrates - _all panic states_ are handled and reported as _syntax errors_ in the `IR`.
Component Groups may add additional modules to compilation. This change ensures that whatever modules result from import/export resolution, they are not duplicate. Not only could that lead to unnecessary compilation but also to multiple reports of the same error which would be confusing to users.
This fixes https://www.pivotaltracker.com/story/show/184189980
# Important Notes
No easy way to write unit tests for this, so skipping that on purpose for now.
Use `InteropLibrary.isString` and `asString` to convert any string value to `byte[]`
# Important Notes
Also contains a support for `Metadata.assertInCode` to help locating the right place in the code snippets.
**Vector**
- Adjusted `Vector.sort` to be `Vector.sort order on by`.
- Adjusted other sort to use `order` for direction argument.
- Added `insert`, `remove`, `index_of` and `last_index_of` to `Vector`.
- Added `start` and `if_missing` arguments to `find` on `Vector`, and adjusted default is `Not_Found` error.
- Added type checking to `+` on `Vector`.
- Altered `first`, `second` and `last` to error with `Index_Out_Of_Bounds` on `Vector`.
- Removed `sum`, `exists`, `head`, `init`, `tail`, `rest`, `append`, `prepend` from `Vector`.
**Pair**
- Added `last`, `any`, `all`, `contains`, `find`, `index_of`, `last_index_of`, `reverse`, `each`, `fold` and `reduce` to `Pair`.
- Added `get` to `Pair`.
**Range**
- Added `first`, `second`, `index_of`, `last_index_of`, `reverse` and `reduce` to `Range`.
- Added `at` and `get` to `Range`.
- Added `start` and `if_missing` arguments to `find` on `Range`.
- Simplified `last` and `length` of `Range`.
- Removed `exists` from `Range`.
**List**
- Added `second`, `find`, `index_of`, `last_index_of`, `reverse` and `reduce` to `Range`.
- Added `at` and `get` to `List`.
- Removed `exists` from `List`.
- Made `all` short-circuit if any fail on `List`.
- Altered `is_empty` to not compute the length of `List`.
- Altered `first`, `tail`, `head`, `init` and `last` to error with `Index_Out_Of_Bounds` on `List`.
**Others**
- Added `first`, `second`, `last`, `get` to `Text`.
- Added wrapper methods to the Random_Number_Generator so you can get random values more easily.
- Adjusted `Aggregate_Column` to operate on the first column by default.
- Added `contains_key` to `Map`.
- Added ALIAS to `row_count` and `order_by`.
Don't propagate errors from `toDisplayString` - construct an error message with `Atom.toString`.
# Important Notes
> currently a failure in to_text is swallowed by `toString` and we cannot detect that something went wrong during the serialization
Not sure how satisfying the solution is, but the error swallowing happens in Truffle and there is little to do with it. We can just catch the error ourselves and produce some meaningful string.
`Text.trim` `what` argument offered `Text` in `tagValues` - that's wrong. Using the `Text` type isn't allowed value for `Text`
# Important Notes
I had to update three other tests to match the new behavior.
Most of the problems with accessing `ArrayOverBuffer` have been resolved by using `CoerceArrayNode` (https://github.com/enso-org/enso/pull/3817). In `Array.sort` we still however specialized on Array which wasn't compatible with `ArrayOverBuffer`. Similarly sorting JS or Python arrays wouldn't work.
Added a specialization to `Array.sort` to deal with that case. A generic specialization (with `hasArrayElements`) not only handles `ArrayOverBuffer` but also polyglot arrays coming from JS or Python. We could have an additional specialization for `ArrayOverBuffer` only (removed in the last commit) that returns `ArrayOverBuffer` rather than `Array` although that adds additional complexity which so far is unnecessary.
Also fixed an example in `Array.enso` by providing a default argument.
Compiler performed name resolution of literals in type signatures but would silently fail to report any problems.
This meant that wrong names or forgotten imports would sneak in to stdlib.
This change introduces 2 main changes:
1) failed name resolutions are appended in `TypeNames` pass
2) `GatherDiagnostics` pass also collects and reports failures from type
signatures IR
Updated stdlib so that it passes given the correct gatekeepers in place.
Implements https://www.pivotaltracker.com/story/show/184032869
# Important Notes
- Currently we get failures in Full joins on Postgres which show a more serious problem - amending equality to ensure that `[NULL = NULL] == True` breaks hash/merge based indexing - so such joins will be extremely inefficient. All our joins currently rely on this notion of equality which will mean all of our DB joins will be extremely inefficient.
- We need to find a solution that will support nulls and still work OK with indices (but after exploring a few approaches: `COALESCE(a = b, a IS NULL AND b is NULL)`, `a IS NOT DISTINCT FROM b`, `(a = b) OR (a IS NULL AND b is NULL)`; all of which did not work (they all result in `ERROR: FULL JOIN is only supported with merge-joinable or hash-joinable join conditions`) I'm less certain that it is possible. Alternatively, we may need to change the NULL semantics to align it with SQL - this seems like likely the simpler solution, allowing us to generate simple, reliable SQL - the NULL=NULL solution will be cornering us into nasty workarounds very dependent on the particular backend.
`runtime-with-instruments` project sets `-Dgraalvm.locatorDisabled=true` that disables the discovery of available polyglot languages (installed with `gu`). On the other hand, enabling locator makes polyglot languages available, but also makes the program classes and the test classes loaded with different classloaders. This way we're unable to use `EnsoContext` in tests to observe internal context state (there is an exception when you try to cast to `EnsoContext`).
The solution is to move tests with enabled polyglot support, but disabled `EnsoContext` introspection to a separate project.
PR adds a flag to `Text` implementation tracking whether it is in a FCD normal form. Then this information can be used in the `Normalizer.compare` method.
| Benchmark name | Old (ms) | With flag (ms)
| --- | --- | ---
| Unicode very short | 40.29 | 40.04
| Unicode medium | 9.07 | 1.99
| Unicode big - random | 115.39 | 0.35
| Unicode big - early difference | 107.02 | 0.54
| Unicode big - late difference | 749.81 | 94.73
| ASCII very short | 28.13 | 31.13
| ASCII medium | 4.58 | 2.26
| ASCII big - random | 42.68 | 0.26
| ASCII big - early difference | 30.91 | 0.32
| ASCII big - late difference | 66.29 | 42.72
Full benchmark output.
[bench_old.txt](https://github.com/enso-org/enso/files/10325202/bench_old.txt)
[bench_new.txt](https://github.com/enso-org/enso/files/10325201/bench_new.txt)