[LlvmIrGenerator] emit c"..." literals for string blobs

[jonathanpeppers]

Stacked on #12279 — review that first; this PR's diff is one file, LlvmIrGenerator.cs.

WriteStringBlobArray() emitted one LLVM IR element per byte of the type map string blobs:

@type_map_managed_type_names = ... constant [3035518 x i8] [
	; 'Android.OS.AsyncTask' @ 0
	i8 u0x41, i8 u0x6e, ...
]

For a hello world MAUI app that's ~3.6 million elements and a 52.6MB typemaps.arm64-v8a.ll, which is slow to write and much slower for llc to consume.

Emit a single constant byte string instead — the same form the generator already uses for constant string literals:

@type_map_managed_type_names = ... constant [3035518 x i8] c"Android.OS.AsyncTask\00..."

Bytes stream through a pooled char [] rather than millions of single-character TextWriter writes.

Results

dotnet new maui -sc, Debug, clean build, -nodeReuse:false. The machine is shared and noisy, so raw per-run values are shown rather than a single median:

Metric	Before	After
typemaps.arm64-v8a.ll	52.6MB	15.2MB
CompileNativeAssembly (llc)	2239, 2287, 2177ms	224, 168, 203, 201, 221, 203ms
GenerateTypeMappings	1489, 1649, 1414ms	645, 1187, 1206, 543, 576, 643ms

CompileNativeAssembly is the headline: ~10x faster, with no overlap between the before and after ranges. That directly addresses the CompileNativeAssembly regression seen in the CoreCLR-vs-Mono binlog comparison that motivated this work.

GenerateTypeMappings is genuinely faster — every "after" run beats every "before" run — but it's noisy enough (543-1206ms across identical binaries) that I won't quote a percentage.

LinkApplicationSharedLibraries and total build time are unchanged. The linker consumes a byte-identical .o, so there is nothing there to win.

Correctness

typemaps.arm64-v8a.o is byte-for-byte identical before and after (SHA256 73E4A982E8C8098FB7909EA550CDEE392E19AAB6938E145A4C362BE0F0417DCF, 4161824 bytes).

One caveat for anyone reproducing this: the -ci.<branch>.<n> SDK version string is embedded in the assembly names blob and changes on every SDK rebuild, so .o hashes only compare within a matched before/after pair.

Additional checks:

• All three globals decode from the .ll to exactly their declared [N x i8] length — type_map_assembly_names 2707, type_map_managed_type_names 3035518, type_map_java_type_names 592060 — and the decoded bytes appear verbatim in the .o.
• All 28,225 i32 offsets in the file (16,567 java + 11,658 managed) resolve correctly against the decoded blobs, zero unresolved. This matters because those offsets index into the blobs.
• Microsoft.Android.Build.BaseTasks-Tests 129/129, LlvmIrGeneratorTests 9/9, GenerateNativeApplicationConfigSourcesTests 4/4. The last of those runs llc on a generated .ll and parses the resulting .s, so it round-trips a string blob end to end.

Nothing parses blob contents out of the .ll, so no test changes were needed. TrimmableTypeMapBuildTests only matches the array type ([N x), which is unchanged.

Tradeoffs

• Per-string ; 'Foo' @ 123 comments are gone. A constant string literal has to fit on a single line and ; comments extend to the end of the line, so they can't be interleaved. They were also a large fraction of the file size.
• The managed type names line is ~3MB wide. llc handles it fine; text editors and diff tools will not enjoy it.
• 15.2MB, not ~5MB. The remaining bulk is the i32 offset/index arrays, which this change doesn't touch — the obvious next target.
• This removes the last caller of HexUtilities.WriteHex (TextWriter, ...) that [LlvmIrGenerator] ~4x faster type map generation #12279 added for the old hot path. The overload is public and unit-tested, so it's left in place.

Squash on merge — the 4 commits are kept separate for readability.

[Copilot]

Pull request overview

This PR updates LlvmIrGenerator to emit type-map string blobs as LLVM IR constant byte strings (c"...") instead of per-byte i8 u0xNN array elements, dramatically reducing .ll size and improving llc compile time for large type maps.

Changes:

• Reworked WriteStringBlobArray() to stream a single c"..." literal for the blob, escaping bytes as needed.
• Switched the hot output path to use a pooled char[] chunk buffer to avoid millions of tiny TextWriter writes.
• Removed per-string inline comments in blob emission (no longer representable within a single-line constant string literal).

[jonathanpeppers]

/review

[github-actions]

✅ Android PR Reviewer completed successfully!

[github-actions]

✅ LGTM

Solid, well-scoped performance change to WriteStringBlobArray. Emitting a single c"..." byte-string constant instead of millions of i8 u0xNN array elements is the right approach, and the reasoning is clearly documented in the method comment.

Correctness review (independent):

• LLVM IR escaping is correct: ", \, and everything outside printable ASCII [32,127) are hex-escaped; 127 (DEL) and 0 (NUL) fall through to the escape path as expected.
• Buffer bounds are safe: the literal path flushes at chunkUsed == chunkCapacity (needs 1 char) and the escape path at chunkUsed + 3 > chunkCapacity (needs \ + 2 hex). After Flush() the 4096-char rented buffer always has room, so no overflow of chunk.AsSpan(chunkUsed, 2).
• Per-string terminating NUL is preserved (WriteByte(0)), so total byte count still matches the declared [N x i8] length.
• ArrayPool<char> rent/return is correctly wrapped in try/finally, and using chunk.Length as the capacity handles Rent over-allocation.
• No test impact: EnvironmentHelper.ReadStringBlob parses the llc-produced .asciz output, which is downstream of the .ll and byte-identical per the PR's .o hash verification.

Verification: The PR body's byte-identical .o (SHA256 match) plus offset-resolution and unit-test checks are exactly the right evidence for a change like this — thank you for the thoroughness.

Two non-blocking 💡 suggestions inline (magic-number naming; uppercase-hex cosmetic note). Neither blocks merge.

CI note: at review time the pipeline shows pending with no completed checks — confirm the dotnet-android build/tests go green before merging.

Generated by Android PR Reviewer for #12280 · 91.4 AIC · ⌖ 18.7 AIC · ⊞ 6.9K
_{Comment /review to run again}