Repository navigation
Run Unity 2D projects as Codename One applications - #5976
shai-almog wants to merge 19 commits into
Conversation
A Unity project's Assets and ProjectSettings in common/src/main/unity build unmodified: the .NET SDK compiles the C# scripts, maven/cil-translator turns the assembly into Java class files and compiles every scene, prefab, animator controller, clip, tile map and particle system to reflection-free Java, and maven/unity-compat is the runtime they run against (a .NET base library, the UnityEngine API over com.codename1.gaming, and our own TextMesh Pro and Cinemachine implementations). Maven (compile-unity, import-unity-project) and Gradle (compileUnity) share UnityProjectBuilder. Both modules sit behind the unity-compat profile and are published by the release. Float results have to match C# on every target, so native builds get -ffp-contract=off and the JavaScript target rounds float results with Math.fround. Bugs this uncovered and fixes outside the new modules, each with a test: - ParparVM JavaScript: wide stack shuffles, MULTIANEWARRAY dimension pops, a call result compared with FCMP/DCMP evaluated more than once, SWAP losing the lower value. - ParparVM native: arraycopy/clone of an array of primitive arrays took the primitive path. - GameView: a key repeat released a held key; a focused hosted child kept the multi-touch arrays from the view. - Box2D: chain against chain threw in World.popContact; the broad phase allocated a merge buffer on every step above 31 pairs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Developer Guide build artifacts are available for download from this workflow run:
Developer Guide quality checks: |
|
Compared 193 screenshots: 193 matched. Benchmark ResultsDetailed Performance Metrics
|
|
Compared 178 screenshots: 178 matched. Native Android coverage
✅ Native Android screenshot tests passed. Native Android coverage
Benchmark ResultsDetailed Performance Metrics
|
Cloudflare Preview
|
✅ Continuous Quality ReportTest & Coverage
Static Analysis
Generated automatically by the PR CI workflow. |
|
Compared 193 screenshots: 193 matched. Benchmark ResultsDetailed Performance Metrics
ParparVM vs HotSpot (JDK 25): Windows x64Runner CPU: Intel64 Family 6 Model 106 Stepping 6, GenuineIntel (baseline 1 performance regression:
Ratios are ParparVM / JDK 25: below 1.00x ParparVM is faster (time) or smaller (RAM). Median of 5 interleaved, paired rounds; every run's output was verified. A ratio more than 15% (time) / 15% (RAM) away from its baseline in
A change that moves performance ON PURPOSE records the new baseline in Result: performance regression |
|
Compared 193 screenshots: 193 matched. ParparVM vs HotSpot (JDK 25): Linux x64Runner CPU: Intel(R) Xeon(R) Platinum 8370C CPU @ 2.80GHz (baseline Ratios are ParparVM / JDK 25: below 1.00x ParparVM is faster (time) or smaller (RAM). Median of 5 interleaved, paired rounds; every run's output was verified. A ratio more than 15% (time) / 15% (RAM) away from its baseline in
Result: no regression |
|
Compared 193 screenshots: 193 matched. ParparVM vs HotSpot (JDK 25): Linux arm64Runner CPU: Neoverse-N2 (baseline Ratios are ParparVM / JDK 25: below 1.00x ParparVM is faster (time) or smaller (RAM). Median of 5 interleaved, paired rounds; every run's output was verified. A ratio more than 15% (time) / 15% (RAM) away from its baseline in
Result: no regression |
|
Compared 193 screenshots: 193 matched. Benchmark ResultsDetailed Performance Metrics
ParparVM vs HotSpot (JDK 25): Windows arm64Runner CPU: ARMv8 (64-bit) Family 8 Model D49 Revision 0, MICROSOFT CORPORATION (baseline Ratios are ParparVM / JDK 25: below 1.00x ParparVM is faster (time) or smaller (RAM). Median of 5 interleaved, paired rounds; every run's output was verified. A ratio more than 15% (time) / 15% (RAM) away from its baseline in
Result: no regression |
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
✅ ByteCodeTranslator Quality ReportTest & Coverage
Benchmark Results
Static Analysis
Generated automatically by the PR CI workflow. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5639acb341
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| UnityRuntime.begin(); | ||
|
|
||
| view = new UnityGameView(); |
There was a problem hiding this comment.
Calling UnityRuntime.begin() before the UnityGameView is created and initialized runs every script's Awake/OnEnable while PlayerPrefs still uses its transient in-memory fallback and AudioSource has no output. Consequently, saved preferences read in Awake appear missing, writes made there disappear once the store is installed, and every scene AudioSource with Play On Awake marks itself playing without ever sending audio to the device. Initialize these view-backed services before waking the first scene.
Useful? React with 👍 / 👎.
| if (negative) { | ||
| value = -value; |
There was a problem hiding this comment.
For the valid input "-2147483648", negating the scanned magnitude produces Long.MIN_VALUE, which is also used as the invalid-input sentinel. Parse therefore throws FormatException, and TryParse returns false, instead of producing Integer.MIN_VALUE. Use a distinct parse-status indicator so the minimum Int32 remains representable.
Useful? React with 👍 / 👎.
| c.resumeAt = 0f; | ||
| c.phase = 0; |
There was a problem hiding this comment.
When StartCoroutine is called from Start or Update, it immediately advances the iterator and adds it with phase == 0 and resumeAt == 0; the same frame's later runCoroutines(0) then advances it again. Thus a first yield return null resumes in the current frame rather than the next frame as promised, so initialization or frame-based loops can execute twice in one frame. Record the frame at which a null/default yield may next resume.
Useful? React with 👍 / 👎.
| public static float get_timeSinceLevelLoad() { | ||
| return time; |
There was a problem hiding this comment.
get_timeSinceLevelLoad() returns the application-wide Time.time, but changeScene() never resets that clock. After loading a second scene, scripts therefore see time since application startup rather than time since the current scene was loaded, which breaks scene-local timers and spawn schedules. Track a separate scene-load timestamp or reset a dedicated level clock on every load.
Useful? React with 👍 / 👎.
| public static boolean get_isMobilePlatform() { | ||
| return false; |
There was a problem hiding this comment.
Application.isMobilePlatform is hard-coded to false, including when the generated application runs on Android or iOS. Unity scripts commonly use this property to enable touch controls, mobile layouts, or device-specific behavior, so those paths remain disabled on the very device builds this compatibility layer targets. Have the host install or derive the actual platform value instead of returning a constant.
Useful? React with 👍 / 👎.
| xvfb-run -a mvn deploy -Psign-artifacts -Dgpg.passphrase=$MAVEN_GPG_PASSPHRASE \ | ||
| -Plocal-dev-javase -Darchetype.test.skip=true -DskipTests \ | ||
| -Plocal-dev-javase -Dunity-compat -Darchetype.test.skip=true -DskipTests \ |
There was a problem hiding this comment.
The release step first runs update-version.sh, whose mvn versions:set invocation does not activate unity-compat, and only activates this profile in the later deploy command. These two modules are therefore absent from the version-update reactor and retain their explicit 8.0-SNAPSHOT parent and artifact versions; once the root POM has the tag version, activating them for deploy produces a parent-version mismatch or publishes under snapshot coordinates instead of the tag. Run the version update with this profile active or explicitly update these two POMs before deployment.
Useful? React with 👍 / 👎.
| mv.visitJumpInsn(IFNE, ok); | ||
| invoke(INVOKESTATIC, Names.INTEROP, "divideByZero", "()V", false); |
There was a problem hiding this comment.
The generated check covers only a zero divisor. For int.MinValue / -1 and long.MinValue / -1, CIL signed division must throw OverflowException, while JVM IDIV/LDIV silently returns the minimum value. Translated C# therefore continues with a corrupt arithmetic result and cannot catch the expected exception; add the minimum-value/negative-one check before emitting the division opcode.
Useful? React with 👍 / 👎.
| public static int Abs(int v) { | ||
| return v < 0 ? -v : v; |
There was a problem hiding this comment.
System.Math.Abs(int.MinValue) must throw OverflowException, because the positive magnitude is not representable, but negating it here wraps and returns the same negative value; the long overload has the same defect. Scripts using the documented System.Math surface can therefore propagate a negative value from an operation guaranteed to be nonnegative. Detect the minimum values and throw the runtime's OverflowException before negating.
Useful? React with 👍 / 👎.
… perf baseline On the Ubuntu runner dotnet is linked into /usr/bin, so dropping every PATH directory that holds it also dropped env, bash and mvn. The directory is now shadowed by links to everything in it except dotnet. objectAllocation RAM on Neoverse-N2 is bimodal; the row is recalibrated from this run and three master runs with a wider RAM tolerance. No improvement is claimed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review findings (rationale for the ones left alone is in comments at the code concerned): - Install the view's services (PlayerPrefs store, audio, platform) before the first scene wakes. - A coroutine started from Start or Update resumes its first null yield on the next frame. - Time.timeSinceLevelLoad restarts on a scene load. - Application.platform and isMobilePlatform come from the host platform. - Signed division and remainder of MinValue by -1 throw OverflowException, as does System.Math.Abs of MinValue. JavaScript screenshots: XYChart rotated the canvas and rotated it back, which with float-rounded angles leaves a transform 1e-14 off identity, and the JavaScript port draws DOM text only for an exact identity. XYChart and ChartComponent now restore the transform they were handed, and the port's DOM text decision tolerates that residue. A differential fixture against the JVM found four older JavaScript target bugs: int to float/double read an un-normalized int, float/double to int wrapped instead of saturating, a static read before its first write was undefined, and Float.toString printed the double. The offscreen check now runs on every scene sample in the integration test. It found: a flick release delivered at the press point because CodenameOneImplementation aliased the shared pointer arrays, GameView missing single finger presses after the first drag, and a sub-image cache keyed by corner only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
…e numeric fixture The JavaScript target now truncates a float product the way the JVM does, so the six theme screens that size from one move by a pixel; the goldens are this run's CI captures. JsNumericDifferentialApp compares narrowing compound assignments against the JVM on purpose and joins the other compiler fixtures CodeQL does not read. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Compared 214 screenshots: 214 matched. |
…lLoad Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Apple TV (tvOS / Metal)Compared 171 screenshots: 163 matched, 8 missing references.
|
|
Compared 167 screenshots: 167 matched. Benchmark Results
Detailed Performance Metrics
|
|
Compared 236 screenshots: 236 matched. |
|
@codex review |
|
Compared 168 screenshots: 168 matched. Benchmark Results
Build and Run Timing
Detailed Performance Metrics
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8a51fcb4b1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| public static boolean Any(IEnumerable_1 source, Class t) { | ||
| return walk(source).MoveNext(); |
There was a problem hiding this comment.
When the source is a custom enumerator or a compiler-generated iterator with a finally block, terminal operations that stop early—such as this Any, First, Contains, ElementAt, and Take—abandon the enumerator without calling Dispose(), even though IEnumerator_1 implements IDisposable. The iterator's cleanup therefore never runs after an early result or predicate exception; wrap enumeration in try/finally and dispose the enumerator on every exit path.
Useful? React with 👍 / 👎.
| public static Object GetValues(Type enumType) { | ||
| Type t = enumType(enumType); | ||
| int[] out = new int[t.values.length]; | ||
| for (int i = 0; i < out.length; i++) { | ||
| out[i] = (int) t.values[i]; |
There was a problem hiding this comment.
For any enum whose underlying type is not int—for example, enum E : long { Huge = 1L << 40 }—Enum.GetValues always returns an int[] and truncates every member. Translated callers cast the result to the enum's normalized JVM array type (for this example, long[]), so the usual (E[])Enum.GetValues(typeof(E)) either fails with a cast error or receives corrupted values. Construct the primitive array corresponding to the enum's recorded underlying type instead.
Useful? React with 👍 / 👎.
| for (int i = 1; i < out.length; i++) { | ||
| int v = out[i]; | ||
| int j = i - 1; | ||
| while (j >= 0 && out[j] > v) { | ||
| out[j + 1] = out[j]; |
There was a problem hiding this comment.
When an enum contains negative members, such as the common flags value All = -1, this signed comparison places that member before zero. Enum.GetValues orders members by the unsigned magnitude of their binary representation, so { All = -1, None = 0, One = 1 } must enumerate as None, One, All, whereas this code returns All, None, One. Use an unsigned comparison appropriate to the enum's underlying width.
Useful? React with 👍 / 👎.
| } else if (c == '}') { | ||
| if (i + 1 < n && format.charAt(i + 1) == '}') { | ||
| i++; | ||
| } | ||
| sb.append('}'); |
There was a problem hiding this comment.
For a malformed composite format such as "value }", .NET throws FormatException, but this branch appends the unmatched brace and returns normally. This changes control flow for scripts that validate or recover from user-provided format strings; only a doubled }} should emit a literal closing brace, while a standalone } must be rejected.
Useful? React with 👍 / 👎.
| public static Type $of(Class type, String fullName) { | ||
| Type t = (Type) byClass.get(type); | ||
| if (t == null) { | ||
| t = new Type(fullName, type, null, null, false); | ||
| byClass.put(type, t); |
There was a problem hiding this comment.
The cache key is only the JVM Class, but multiple distinct .NET primitives share one box class: int and uint both use Integer.class, long and ulong use Long.class, and byte and sbyte use Byte.class. Consequently, whichever type is requested first is returned for both, making typeof(int) == typeof(uint) true and potentially reporting the wrong FullName. Include the .NET type identity in the cache key rather than keying solely by the erased Java class.
Useful? React with 👍 / 👎.
| static long number(Object value) { | ||
| if (value instanceof EnumBox) { | ||
| return ((EnumBox) value).value; |
There was a problem hiding this comment.
This conversion discards the EnumBox.type, so APIs such as Enum.IsDefined, Enum.GetName, HasFlag, and CompareTo accept a boxed member of an unrelated enum and operate only on its numeric value. For example, firstEnum.HasFlag(secondEnum) can return true when their bits overlap, whereas .NET throws ArgumentException for differing enum types. Validate the box's enum type against the requested or receiver type before extracting its number.
Useful? React with 👍 / 👎.
| public void set_time(float value) { | ||
| startedAt = Time.unscaledTime - value; | ||
| } |
There was a problem hiding this comment.
When a playing source executes audio.time = value, this setter only changes the runtime's bookkeeping and never seeks the device's Media, so the audible clip continues from its old position while scripts observe a different position. The bookkeeping is also immediately wrong when pitch != 1, because get_time() multiplies elapsed time by pitch but this setter does not divide the requested clip position by pitch. Add a seek operation to AudioOutput and derive startedAt consistently with the playback rate.
Useful? React with 👍 / 👎.
| public static String Trim(String s) { | ||
| return s.trim(); | ||
| } |
There was a problem hiding this comment.
String.trim() only removes characters at or below U+0020, whereas System.String.Trim() removes Unicode whitespace. Inputs surrounded by characters such as non-breaking space (U+00A0) or em space (U+2003) therefore remain untrimmed, even though this runtime already recognizes them in Char_.IsWhiteSpace. Scan the ends using the .NET whitespace predicate instead of delegating to Java's trim().
Useful? React with 👍 / 👎.
| String wanted = value.trim(); | ||
| for (int i = 0; i < t.names.length; i++) { | ||
| if (ignoreCase ? t.names[i].equalsIgnoreCase(wanted) : t.names[i].equals(wanted)) { | ||
| return new EnumBox(t, t.values[i]); | ||
| } | ||
| } | ||
| try { | ||
| return new EnumBox(t, Long.parseLong(wanted)); |
There was a problem hiding this comment.
Enum.Parse accepts a comma-separated list of member names, which is the standard way to parse flags such as "Read, Write"; this implementation only checks the entire string against one member and then attempts numeric parsing. Valid combined values therefore throw ArgumentException. Split on commas, trim each name, resolve each with the requested case sensitivity, and combine their underlying values.
Useful? React with 👍 / 👎.
| if (t instanceof NegativeArraySizeException) { | ||
| return new OverflowException(); | ||
| } | ||
| return t; |
There was a problem hiding this comment.
Several public runtime APIs throw java.lang.IllegalArgumentException for invalid C# calls, including PlayerPrefs, SceneManager, Input, physics queries, and Instantiate, but this dispatcher leaves that throwable unchanged. Because generated catch clauses test against the custom System.Exception hierarchy, even catch (Exception) cannot catch these failures and the raw JVM exception escapes the script. Throw the corresponding compatibility exception at each API or map Java argument exceptions here.
Useful? React with 👍 / 👎.
|
Compared 168 screenshots: 168 matched. Benchmark Results
Build and Run Timing
Detailed Performance Metrics
|
…seline Base library: LINQ operators dispose their enumerators on every exit, Enum values keep their underlying type and unsigned order, Enum.Parse takes comma lists, HasFlag/CompareTo/IsDefined refuse another enum's member, String.Format rejects a lone closing brace, Trim follows Char.IsWhiteSpace, typeof(int) and typeof(uint) are distinct, a float key's two zeros are one key, and a Java exception thrown by the runtime is caught by a script's catch (Exception). AudioSource.time seeks the device's media and accounts for pitch. The macOS draw-gradient golden is this branch's CI capture: with -ffp-contract=off the mutable-image ellipse edge now equals the on-screen one. Linux x64 perf baseline rows recalibrated per runner CPU from PR and master runs; the shift is present on master at the base commit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f0abf10218
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| Image image = (Image) paintedTexts.get(key); | ||
| if (image == null) { | ||
| image = paint(d, width, height); | ||
| paintedTexts.put(key, image); |
There was a problem hiding this comment.
When a Text component displays a changing value such as a timer, score, or frame counter, every distinct string creates an Image and stores it permanently in paintedTexts; the map is never cleared or evicted. A frequently updated label therefore retains thousands of raster images and can exhaust device memory during an otherwise normal game session. Keep only current/recent entries or use a bounded cache.
Useful? React with 👍 / 👎.
| if (yielded instanceof WaitForSeconds) { | ||
| c.resumeAt = Time.time + ((WaitForSeconds) yielded).seconds; | ||
| } else if (yielded instanceof Coroutine) { | ||
| c.waitingFor = (Coroutine) yielded; | ||
| } else if (yielded instanceof WaitForFixedUpdate) { |
There was a problem hiding this comment.
When a coroutine uses the common yield return ChildRoutine() form, the yielded value is an IEnumerator, but this chain recognizes only a started Coroutine; it consequently treats the child iterator like null, never executes it, and resumes the parent on the next frame. Add nested-enumerator handling so the parent waits while the yielded iterator is driven to completion.
Useful? React with 👍 / 👎.
| public static float get_unscaledDeltaTime() { | ||
| return unscaledDeltaTime; |
There was a problem hiding this comment.
When Time.unscaledDeltaTime is read from FixedUpdate, Unity reports the unscaled duration of that fixed step, but this getter always returns the enclosing render frame's dt. At nonmatching render/physics rates—or when several fixed steps run in one frame—fixed-update timers and movement therefore advance by the render delta once per physics step, producing speed and scheduling errors.
Useful? React with 👍 / 👎.
| if (f.isDirectory()) { | ||
| collectStamped(f, prefix + f.getName() + "/", out); | ||
| } else { | ||
| out.add(prefix + f.getName() + ":" + f.length() + ":" + f.lastModified()); |
There was a problem hiding this comment.
In the new Unity incremental build cache, a source or asset rewritten with the same byte length and preserved modification time—such as after restoring or synchronizing files with timestamps intact—produces the same stamp even if its contents changed. run() then accepts state.txt and skips compilation, leaving stale translated classes or resources in the application; hash the file bytes rather than only name, size, and timestamp.
Useful? React with 👍 / 👎.
| public List_1() { | ||
| items = new Object[4]; | ||
| } | ||
|
|
||
| public List_1(int capacity) { | ||
| items = new Object[capacity < 4 ? 4 : capacity]; |
There was a problem hiding this comment.
When valid C# uses new List<T>(source)—including the new List<int>(numbers) array case explicitly anticipated by ArrayView—the translator emits a call to the erased List_1(IEnumerable_1) constructor, but this runtime class exposes only the parameterless and capacity constructors. The Unity project therefore fails translation with a missing runtime method instead of constructing the list; add the enumerable constructor and dispose its enumerator after copying.
Useful? React with 👍 / 👎.
| String simple = meta.getName().substring(0, meta.getName().length() - ".cs.meta".length()); | ||
| for (CilAssembly assembly : scripts) { | ||
| for (TypeDef t : assembly.types()) { | ||
| String full = t.fullName(); | ||
| if (full.equals(simple) || full.endsWith("." + simple)) { | ||
| scriptsByGuid.put(guid, t); |
There was a problem hiding this comment.
When a valid Unity project contains two scripts with the same filename in different folders and namespaces, each .cs.meta GUID is mapped by simple class name and this loop always selects the first matching TypeDef. Both GUIDs can therefore resolve to the same script class, causing scene objects to receive the wrong behaviour and serialized fields from the other script to be dropped. Use source/PDB information or another unambiguous file-to-type mapping.
Useful? React with 👍 / 👎.
| GameObject go = (GameObject) objects.get(i); | ||
| if (go.destroyed || !go.activeInHierarchy()) { | ||
| continue; | ||
| } | ||
| if ((name != null && !go.name.equals(name)) || (tag != null && !go.tag.equals(tag))) { | ||
| continue; |
There was a problem hiding this comment.
When a script calls GameObject.Find("Canvas/Panel/Button"), Unity interprets the slash-separated argument as a hierarchy path, but this implementation compares the whole argument only against each object's leaf name. Such lookups always return null unless an object literally contains slashes in its name, breaking a standard way scripts locate nested scene objects; resolve the root and traverse children as the existing Transform.Find implementation does.
Useful? React with 👍 / 👎.
| public java.lang.Object GetComponentInChildren(Class type) { | ||
| if (!active) { | ||
| return null; | ||
| } | ||
| java.lang.Object c = GetComponent(type); |
There was a problem hiding this comment.
When GetComponentInChildren is invoked on an inactive GameObject, Unity still searches that root object even though the one-argument overload excludes its inactive descendants, but this early return skips the root as well. Managers querying an inactive prefab or temporarily disabled object therefore receive null despite the requested component being attached directly to it; GetComponentsInChildren has the same issue through collect, so both traversals should always inspect their starting object.
Useful? React with 👍 / 👎.
| public static boolean IsDigit(char c) { | ||
| return c >= '0' && c <= '9'; | ||
| } |
There was a problem hiding this comment.
For decimal digits outside ASCII, such as Arabic-Indic ٣ or full-width ５, System.Char.IsDigit must return true because they have the Unicode decimal-digit category, but this implementation accepts only 0 through 9. Translated validation and parsing code therefore rejects digit strings that work in .NET; use a Unicode digit-category check rather than an ASCII range.
Useful? React with 👍 / 👎.
Text images are released when a label stops showing them: the core sprite renderer gains releaseTexture, and the Unity view's text cache is bounded. yield return of a bare IEnumerator runs the child before the parent resumes. Time's unscaled values follow the fixed step inside FixedUpdate. List and Dictionary copy constructors. GameObject.Find walks a path. GetComponentInChildren always searches the object asked. Char.IsDigit covers the Unicode decimal digits. Two scripts with one file name are told apart by their source path. The incremental build stamp hashes content and holds no times or absolute paths, so a built project can move between machines. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A gallery project of eight scenes, one per feature area, is compiled into the test app; each test advances a held clock a fixed number of frames and captures. scripts/unity-gallery-prebuilt.sh packs the translated project so a job without the .NET SDK can build the app. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y port unity-gallery-prebuilt.yml builds the pack behind a cache and uploads it; every workflow that builds the hellocodenameone app unpacks it before the build, so no macOS, Windows or Android job needs the .NET SDK. A consumer job in unity-compat.yml holds the no-SDK path. A build whose compiled output does not match its sources or jars now says so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the pack digest Linux, Windows and JavaScript builds of the test app listed their class directories by hand and never saw a dependency jar; the common module now stages its runtime class path where they read it. The compile-unity tool class path took the translator's test dependencies, so the pack's digest covered the consumer's own core jar; it is runtime scope now, and a mismatch names the part and the first entry that differs. The eight Unity tests are a port-status feature. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Unity gallery's Canvas scene has a panel that is black at 40% over the camera's background. Android and Windows rendered 0.6 x background; Linux rendered it LIGHTER than the background and JavaScript a third too light. Both used one glBlendFunc(SRC_ALPHA, ONE_MINUS_SRC_ALPHA) for all four channels, so the frame's alpha became sa*sa + (1-sa) = 0.76 under the panel. The colour was right; the frame was then composited with that alpha -- by cairo over the form on Linux, as a premultiplied canvas on JavaScript. The alpha channel now takes ONE / ONE_MINUS_SRC_ALPHA on Linux, JavaScript, Android, JavaSE (JOGL) and Metal, which is what Direct3D already did, and RenderState.BlendMode documents it. The JavaSE software device weights additive light by the source alpha, as every GPU backend does. GpuBlendContractTest runs the blend equation with each backend's factors; parparvm-tests.yml names the sources it reads. The offscreen scene check asserts the panel against the background. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Compared 12 screenshots: 12 matched. |
…s for
The Unity gallery's TextMeshPro scene draws a bold and an italic line, and
both came out in the regular face on native Linux and Windows. The runtime
asks the portable way -- createTrueTypeFont("native:MainBold", ...)
.derive(px, STYLE_PLAIN) -- so the style is in the name alone.
Linux mapped every native: name to regular Sans. Windows read the style at
load and then rebuilt the derived font from derive's weight only. Both now
keep it; graphics-draw-string on those ports draws its MainBold, MainBlack
and Italic* lines styled for the first time and will need reseeding.
NativeFontStyleContractTest compiles the two name parsers alone and holds
the derive paths by source; parparvm-tests.yml names the files it reads.
Also: the complete license header on four files the blend fix touched, and
the Unity gallery goldens for Android and JavaScript, plus the six scenes
without text for Linux (x64, arm64) and Windows, from CI run captures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Compared 179 screenshots: 179 matched. Benchmark Results
Detailed Performance Metrics
ParparVM vs HotSpot (JDK 25): macOS arm64Runner CPU: Apple M1 (Virtual) (baseline Ratios are ParparVM / JDK 25: below 1.00x ParparVM is faster (time) or smaller (RAM). Median of 5 interleaved, paired rounds; every run's output was verified. A ratio more than 15% (time) / 15% (RAM) away from its baseline in
Result: no regression |
…rame runs A text command was painted on the frame's thread. On iOS that is the main thread, while the port keeps one "image being drawn into" for the whole process: a component painted to the screen meanwhile ends it, and a string on its way to the text's image is dropped silently. A text is painted once and cached, so a line lost that way never came back -- the right-aligned line of the TextMeshPro gallery scene, in the iOS capture. A frame off the event dispatch thread now asks it for the text by callSerially and never waits; the pixels come back as an RGBImage. Until they arrive a label shows its previous text. A frame on the event dispatch thread paints at once, as before. SceneShots.checkOffEdt holds the event dispatch thread and draws frames off it for every gallery scene; the offscreen checks that assert a whole frame now draw it on the event dispatch thread. Also: Unity goldens for iOS (Metal, newest runtime) and the Canvas and TextMeshPro scenes on Linux and Windows; reseeds on Linux and Windows for the screens whose native: bold and italic fonts are now honoured; and a recalibrated Windows x64 translator row in the PR's perf overlay. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
https://github.com/codenameone/CodenameOne/blob/1a18fdf9e9e3e23549793fcd0dacc20684b5ab59/maven/unity-compat/src/main/csharp/UnityEngine/Values.cs#L711
The fixed 1e-5f floor makes Mathf.Approximately treat almost every pair near zero as equal—for example, Approximately(0f, 1e-6f) returns true. Unity's comparison uses the relative 1e-6 term with only a few floating-point epsilon units as its absolute floor, so gameplay checks around zero can now trigger roughly ten units too early. Replace this floor with the epsilon-based tolerance.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (names.norm(tp).isPrimitive()) { | ||
| Jvm.unbox(mv, names, tp); | ||
| } else { | ||
| mv.visitTypeInsn(CHECKCAST, names.classRef(tp)); |
There was a problem hiding this comment.
When handwritten generic runtime code invokes a delegate with a struct value, this adapter only casts the erased object and passes the same instance to the target. For example, LINQ obtains a Vector2 directly from List_1 and calls a Func<Vector2, bool>; if the lambda mutates its by-value parameter, it consequently mutates the element stored in the source list, unlike C# value semantics. Copy concrete struct arguments here before invoking the translated target.
Useful? React with 👍 / 👎.
| public void set_loop(boolean value) { | ||
| loop = value; | ||
| } |
There was a problem hiding this comment.
When AudioSource.loop is changed during playback, this setter updates only runtime bookkeeping and never changes UnityGameAudio's looping map. Turning looping on leaves the device voice non-looping while isPlaying remains true indefinitely; turning it off leaves the device restarting the clip even after runtime bookkeeping declares it finished. Update or restart the active output voice when this property changes.
Useful? React with 👍 / 👎.
| public static void set_pause(boolean value) { | ||
| paused = value; | ||
| } |
There was a problem hiding this comment.
When a game pauses audio through AudioListener.pause, this setter merely records the flag: it neither pauses device voices nor freezes each AudioSource's playback clock. Audio therefore continues audibly, and non-looping sources can expire while the listener is supposedly paused, breaking the common pause-menu flow. Propagate this state to active sources and compensate their timing when playback resumes.
Useful? React with 👍 / 👎.
|
|
||
| public static bool Approximately(float a, float b) | ||
| { | ||
| return Abs(b - a) < Max(1e-6f * Max(Abs(a), Abs(b)), 1e-5f); |
There was a problem hiding this comment.
The fixed 1e-5f floor makes Mathf.Approximately treat almost every pair near zero as equal—for example, Approximately(0f, 1e-6f) returns true. Unity's comparison uses the relative 1e-6 term with only a few floating-point epsilon units as its absolute floor, so gameplay checks around zero can now trigger roughly ten units too early. Replace this floor with the epsilon-based tolerance.
Useful? React with 👍 / 👎.
| } | ||
|
|
||
| public void Play(String stateName) { | ||
| Play(stateName, 0, 0f); |
There was a problem hiding this comment.
When a script repeatedly calls Play for the state that is already active, this overload supplies normalized time zero and restarts the animation on every call. Unity's omitted normalizedTime value is a negative-infinity sentinel that preserves the current state's time, so calling Play("Run") from Update, for example, should not freeze the animation on its first frame. Pass and honor the preserve-time sentinel instead of forcing zero.
Useful? React with 👍 / 👎.
| Item1 = other.Item1; | ||
| Item2 = other.Item2; |
There was a problem hiding this comment.
When either tuple element is itself a translated struct, $copyValue() reaches this assignment and preserves the same mutable element object in both tuple copies. Mutating that element through the copied tuple therefore also changes the original tuple, whereas copying a ValueTuple<Vector2, int> in C# copies the nested Vector2 value. Copy elements implementing Struct instead of assigning their erased references directly.
Useful? React with 👍 / 👎.
| if (hasParent && parent != null) { | ||
| t.SetParent(parent, position == null && worldSpace); | ||
| } |
There was a problem hiding this comment.
When original is a child object and the caller uses Instantiate(original) or requests world-space parenting, copyTree has copied the child's local transform onto an unparented root, and this block either leaves that local pose as its world pose or preserves it while assigning the new parent. The clone consequently jumps relative to the source's parent instead of retaining the source's world position and rotation. Capture and apply the source world pose for the no-parent and worldSpace cases.
Useful? React with 👍 / 👎.
| // Unity copies a serializable class field by field; | ||
| // here the two scripts share the one object. | ||
| helper = def != null && (def.flags & TYPE_SERIALIZABLE) != 0 ? "" : null; |
There was a problem hiding this comment.
When a live object with a field such as [Serializable] Stats stats is instantiated, this branch deliberately selects a plain field assignment for the serialized class. The source and clone therefore share the same Stats instance, so changing runtime state on one clone changes the source and every other clone, whereas Unity serializes these classes by value when cloning an object. Recreate the serialized object graph and copy its serialized fields instead of retaining the reference.
Useful? React with 👍 / 👎.
| } else if (yielded instanceof WaitForFixedUpdate) { | ||
| c.phase = 1; | ||
| } else if (yielded instanceof WaitForEndOfFrame) { |
There was a problem hiding this comment.
When StartCoroutine is called from FixedUpdate and its first yield is WaitForFixedUpdate, advance marks it for phase 1 and the same physics step's later runCoroutines(1) immediately resumes it. The yield is supposed to wait until the following fixed update, so physics logic after the yield runs twice in one step. Track the fixed-step generation just as normal next-frame yields track frameCount.
Useful? React with 👍 / 👎.
| return y >= 0 ? Atan(y / x) + PI : Atan(y / x) - PI; | ||
| } | ||
| return y > 0 ? PI / 2 : y < 0 ? -PI / 2 : 0; |
There was a problem hiding this comment.
For signed-zero inputs, these ordinary comparisons discard the quadrant encoded in the sign bits: for example, Math.Atan2(0.0, -0.0) returns zero here instead of positive π, and Math.Atan2(-0.0, -1.0) returns positive π instead of negative π. This also affects Mathf.Atan2 and angle calculations at axis boundaries; handle the IEEE signed-zero cases explicitly.
Useful? React with 👍 / 👎.
… pause and loop Translator and base library, each held by a console line compared with dotnet or a sample trace line: - Instantiate gives a clone its own copy of a [Serializable] class field, and of the elements of an array or list of them. - Generated struct equality treats NaN fields as equal, as ValueType.Equals does. - A view over a struct array hands out copies: LINQ over one saw later writes to the array, and ToList shared its values. - A null out of erased code reads as default(T), for a struct and for a primitive: FirstOrDefault on an empty list of structs was null. - Copying a value tuple copies a struct it holds. - Atan2 and Atan keep the sign of zero and answer for infinities. - Mathf.Approximately uses the relative term with a floor of eight epsilons; the floor of 1e-5 made everything near zero equal. Runtime: - WaitForFixedUpdate asked for inside a fixed step waits for the next. - Instantiate of a child, with no parent or in world space, keeps the original's world position and rotation. - Animator.Play without a time leaves a state that is playing alone. - AudioListener.pause stops every source's clock and holds the output's voices; AudioSource.loop set while playing reaches the voice. A delegate adapter does not copy a struct argument: a translated target already copies one it writes, and the comment there says so. Also: the prebuilt gallery workflow listed its archive into grep -q, which failed the pipeline whenever grep left first; and hello RAM rows for two Windows runners in the PR's perf overlay, where the suite now runs the Unity scenes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The eight gallery scenes on iOS (Metal), tvOS and Mac Catalyst, and the TextMeshPro scene on the newest iOS runtime, which was held back until its right-aligned line was drawn. All from CI captures. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The row moved hello RAM on Xeon Platinum 8573C from 0.732x to 0.905x because master at the base commit measured the same. Master's base now carries 0.876x for it at the same 20% tolerance, which this pull request's 0.905x is inside, so the overlay no longer needs to say anything about that row -- and saying it from the old value is what the gate refuses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
port-status-nightly.yml: master's run lines without xvfb, with this branch's unpack of the Unity gallery prebuilt between them. The perf overlay drops its translator RAM row for EPYC 7763: master's pr/6006.json rebaselines the same row from the same value, and the gate refuses two overlays that do. The row is to be measured again on top of that one if this pull request's runs fall outside it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…XCTest The iOS screenshot legs sit at their 95 minute limit, on master as on this branch. Measured, the screenshot suite is 7.5 minutes of it; the rest is compiling, and two thirds of what is compiled is not the iOS app. - The embedded watch app was built for arm64 AND x86_64 whenever the destination was an iPhone: 9,880 of 14,834 compile units, 21 to 34 minutes. The CI builds now hold it to the host's slice through the xcconfig they already pass. The generated project is not changed. - The screenshot test server was built by every Apple job, 3 to 10 minutes each. The port build makes it once and the jobs take it from the bundle, as Linux and Windows already do; a job without it builds it as before. - The simulator booted ahead of the Xcode 27 leg's compile is shut down once the boot is proven; the test script boots it when it needs it. - The packaging job no longer runs the local notification XCTest, which the native job runs from the same sources: 23 to 38 minutes of a second compile for the same 0.17 seconds of tests. No timeout is raised. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Native fidelity (Android, Material 3)54 pairs compared -- median 95.6%, worst 91.3% ( Distribution --
Geometry vs native (bbox offset / size ratio / center offset / corner radius) -- gated separately from the visual score
Side-by-side comparisons (worst first)
|
Native fidelity (iOS Modern, Metal)68 pairs compared -- median 95.0%, worst 83.5% ( Distribution --
Geometry vs native (bbox offset / size ratio / center offset / corner radius) -- gated separately from the visual score
Side-by-side comparisons (worst first)
|
Native fidelity (ios-27-metal)68 pairs compared -- median 94.6%, worst 74.5% ( Distribution --
Geometry vs native (bbox offset / size ratio / center offset / corner radius) -- gated separately from the visual score
Side-by-side comparisons (worst first)
|
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 55dbb30cf4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| default: | ||
| return "java/lang/Integer"; |
There was a problem hiding this comment.
When signed and unsigned primitives enter object or erased generic storage, this maps int and uint to the same Integer representation; the same collision occurs for long/ulong, byte/sbyte, and char/ushort. Consequently, ((object)(int)1).Equals((object)(uint)1) returns true instead of false, (uint)(object)(int)1 can unbox successfully instead of throwing, and object-keyed dictionaries conflate distinct .NET values. Use distinct box representations that retain the original primitive type.
Useful? React with 👍 / 👎.
| double whole = java.lang.Math.floor(b); | ||
| if (whole == b && b >= -1024 && b <= 1024) { |
There was a problem hiding this comment.
When the base is negative and the integral exponent has magnitude greater than 1024, this bound skips the integer-power branch and the later negative-base path returns NaN. For example, both Math.Pow(-1, 1025) and Mathf.Pow(-1f, 1025f) should return -1 but return NaN here; large powers also lose the correct signed infinity or zero result. Preserve integer parity for all representable integral exponents instead of routing them through the non-integral negative-base case.
Useful? React with 👍 / 👎.
| Delegate[] x = a.invocationList(); | ||
| Delegate[] y = b.invocationList(); |
There was a problem hiding this comment.
When both arguments are non-null but have different delegate types, .NET's Delegate.Combine throws ArgumentException; this code instead builds a multicast delegate using a's type and embeds b in its invocation list. A call such as Delegate.Combine(anAction, aFunc) therefore succeeds unexpectedly and only fails later during invocation with a cast error, changing both timing and exception handling. Validate that both operands have the same delegate type before constructing the combined list.
Useful? React with 👍 / 👎.
| t.getOutputDirectory().set(new File(layout.buildDir(), "generated/sources/cn1-unity")); | ||
| t.getClassesDirectory().set(new File(layout.buildDir(), "cn1-unity/classes")); | ||
| t.getStateDirectory().set(new File(layout.buildDir(), "cn1-unity/state")); | ||
| t.onlyIf(x -> com.codename1.maven.UnityProjectBuilder.isUnityProject(unityDir)); |
There was a problem hiding this comment.
In an incremental Gradle build where src/main/unity is deleted or ceases to contain both required directories after a successful Unity build, this onlyIf skips the task entirely and leaves its declared generated-source and classes directories untouched. Those directories are still registered with the main source set and application output, so stale generated scenes, translated classes, and assets can continue to compile or be packaged until the user happens to run clean; run a cleanup path when the Unity project is absent instead of skipping all task action.
Useful? React with 👍 / 👎.
2 participants
Summary
A Unity 2D project (C# scripts, scenes, prefabs, sprites, project settings) dropped into
common/src/main/unity, or imported withcn1:import-unity-project, builds unmodified as a Codename One application for every target. No Unity player and no .NET runtime ship with the app.maven/cil-translator: reads the assembly the C# compiler produced and writes one class file per C# type; compiles scenes, prefabs, animator controllers, clips, tile maps and particle systems to reflection-free Java (UnityAppImpl).maven/unity-compat: the runtime. A .NET base library in Java, theUnityEngineAPI overcom.codename1.gaming(2D rendering, Box2D physics with exact casts and overlap queries, Animator, tile maps, particles, uGUI, touch, audio,PlayerPrefs,Resources), plus our own implementations of the documented TextMesh Pro and Cinemachine API.UnityProjectBuilderinbuild-engine, thecompile-unityandimport-unity-projectMaven goals, the GradlecompileUnitytask, the archetype profile, and release publishing behind-Dunity-compat(the modules need the .NET SDK; the framework build does not).UnityApplication(createForm,onProjectInstalled,onStarted,callInFrame), so a game sits inside a Codename One form and uses the rest of the framework.Unity-Interop.asciidoc.Clean room: no Unity source, assemblies or package source were read or shipped; the stub
UnityEngine.dllis compiled from our own sources.Changes outside the new modules
Bugs this work uncovered, fixed at the root, each with a regression test:
POP2/DUP2/DUP_X2),MULTIANEWARRAYdimension pops, a call result compared viaFCMP/DCMPevaluated more than once,SWAPlosing the lower value.arraycopy/clone of an array of primitive arrays took the primitive path.GameView: a key repeat released a held key after 800 ms; a focused hosted child kept the multi-touch arrays from the view.World.popContact; the broad phase allocated a merge buffer every step above 31 pairs (now an in-place sort, checked againstArrays.sortover randomized pair sets).Behavior changes that affect every application:
-ffp-contract=off.floatresult withMath.fround.Both make float arithmetic match the JVM bit for bit. Their performance cost on ordinary applications has not been measured.
Verification
scripts/unity-compat-samples) produce byte-identical traces on the JVM, ParparVM native and JavaScript (maven/integration-tests/unity-compat-test.sh --parparvm).cil-translatorandunity-compat: SpotBugs 0, Checkstyle 0, PMD forbidden list clean, cast-semantics clean.MapViewMarkerLabelTesttimeout on a loaded machine; passes alone).Not verified
.github/workflows/unity-compat.ymland therelease.ymlchanges have never run on CI.GameInputstill reads edge flags that a press landing mid-frame can miss on ports that render off the EDT; left for a separate decision.🤖 Generated with Claude Code