Manual Jni

Manual JNI surface

Regenerate docs/manual-jni.md — the inventory of every place in src/Microsoft.AndroidX.Compose/ that still uses hand-written JNI (JNIEnv.* calls, [Register("net/compose/…")] JCWs) instead of the source-generated [ComposeBridge] / [ComposeFacade] machinery. The report is produced by a single .NET 10 file-based program at scripts/manual-jni-report.cs. It scans every .cs file under src/Microsoft.AndroidX.Compose/, anchors each JNI call to its enclosing member, pulls a "Why not generated?" rationale from existing XML doc <remarks> and adjacent // ... comments, and writes a markdown report with summary metrics, per-file member tables, and a JCW catalogue.

Workflow

1. Run the script

From the repo root (Windows / PowerShell):

dotnet run scripts/manual-jni-report.cs

That's it — no .csproj, no #:package, no MSBuild target. Pure text/regex scan; runs in a couple of seconds.

Expected tail output:

Files scanned:           255
Files with manual JNI:   47
JNIEnv.* call sites:     196
JCW classes:             25 (4 with JNI, 21 pure-managed)
[ComposeBridge] partials: 202
[ComposeFacade] partials: 95
Report written: docs/manual-jni.md

2. Verify the diff

git diff --stat docs/manual-jni.md

The headline numbers live in the Summary section near the top. Quick sanity check:

Select-String -Path docs/manual-jni.md -Pattern '^- \*\*' | Select-Object -First 7

Re-running the script with no source change should produce a byte- identical file modulo the Generated by ... on YYYY-MM-DD HH:MM UTC timestamp. If it doesn't, see Troubleshooting.

3. Commit the regenerated report alongside the change that prompted it

The report is checked in deliberately so reviewers can see the manual surface shrink (or grow) in the PR diff. Don't commit the script run on its own — pair it with the bridge migration, new suspend bridge, or whatever change motivated the refresh.

git add docs/manual-jni.md
git commit --amend --no-edit  # or a fresh commit, depending on context

What the report contains

• Summary — files scanned / flagged, raw JNIEnv.* call count, total lines inside flagged members, JCW totals (split: with-JNI vs pure-managed delegate adapters), generator-emitted partials ([ComposeBridge] + [ComposeFacade]), manual : generated ratio.
• Per-file detail — one section per .cs file containing manual JNI or a JCW. Each section opens with the file's <summary> / leading comment as Purpose, notes how many [ComposeBridge] / [ComposeFacade] partials in the same file are excluded, then enumerates manual entry points in a table: member name, kind, line range, JNI call count, "Why not generated?", "Migration".
• JCW classes — every [Register("net/compose/…")] class in the project, with its Kotlin interface, source location, and whether the JCW body itself contains raw JNI.
• Caveats — heuristic limitations (regex-based detection, comment-driven rationale extraction).

Improving "Why not generated?" entries

The script extracts the rationale paragraph by walking upward from the member declaration (up to 30 lines), matching in this order:

1. An explicit Why raw JNI / Why manual marker — either form wins regardless of which comment style hosts it:
   • // Why raw JNI: … / // Why manual: … line.
   • /// XML doc <remarks> paragraph containing Why raw JNI.
2. Otherwise, the adjacent ///-doc or // comment block sitting directly above the declaration is used verbatim.
3. Falls back to TODO: document if there are no comments above the member.

If a member shows TODO: document in the table, that's a docs gap in the source itself. The fix is in the source file, not the script — add a // Why raw JNI: ... line above the method declaration, or flesh out the existing XML <remarks>. Re-run the script and the entry refreshes.

Don't add a banner just to please the report — only document members where the rationale is non-obvious from the surrounding context.

Troubleshooting

Member detection misses a multi-line declaration — the script expects member declarations to begin with the standard modifier prefix (public|internal|private|protected|static|...). If a method uses an unusual modifier order or layout ([MethodImpl(...)] first, then modifiers on the next line), the regex may skip it and the member's JNI calls land in a synthetic (file-level) row. Either restructure the declaration to put modifiers on the same line as the return type, or extend the modifier soup in rxMethod / rxCtor near the top of the script.

JCW count looks wrong — only Microsoft.AndroidX.Compose's own [Register("net/compose/…")] classes are counted. JCWs Mono.Android emits for upstream interfaces are deliberately excluded. If a new file registers a JCW under a different namespace prefix, update the rxRegister regex.

Generator partials counted as manual — the script excludes members decorated with [ComposeBridge] / [ComposeFacade] / [ComposeDefaults] by walking attribute lines immediately above the declaration. If a comment line splits an attribute from its target, the exclusion silently fails. Move the attribute adjacent to the declaration.

Report regen produced spurious diffs — every run except for the header timestamp should be byte-identical. Common causes if it isn't:

• A non-ASCII char drift (smart quotes, em-dashes) in a source comment got mangled by some other tool.
• Member detection picked up a new declaration that wasn't there before. To find which entry shifted, note the line range in the diffed row (e.g. Lines 221–246) and use git diff -U0 src/Microsoft.AndroidX.Compose/<file>.cs | rg -n '^@@' to locate the corresponding source change.

If diffs persist with no source change, the script has a bug; look at the order-sensitive sites (OrderBy(j.ClassName, StringComparer.Ordinal) on the JCW table, OrderBy(StartLine) inside per-file entries).

Don't do these things

• Don't hand-edit docs/manual-jni.md. It's regenerated every run and your edit will be lost.
• Don't run the script from inside a subdirectory; the output path (docs/manual-jni.md) and source-root path (src/Microsoft.AndroidX.Compose) are relative to the repo root.
• Don't add a // Why raw JNI: banner to a member that already has a clear <summary> / <remarks> block — the script's fallback reads adjacent doc comments fine.
• Don't include managed helpers with zero JNIEnv.* calls in the manual tally just because they live in ComposeBridges.cs. The report tracks raw JNI usage, not everything that isn't generator-decorated; a one-line managed helper isn't relevant surface.