Performance optimizations for EscapingUtilities by DustinCampbell · Pull Request #13426 · dotnet/msbuild

DustinCampbell · 2026-03-20T23:27:57Z

Tip

I recommend reviewing this pull request commit-by-commit. I made sure that each commit is distinct, cohesive and has a detailed description.

This PR rewrites and optimizes EscapingUtilities.Escape and EscapingUtilities.UnescapeAll — two methods called heavily throughout MSBuild evaluation — to reduce allocations and improve throughput on both .NET and .NET Framework 4.7.2.

The largest wins are in Escape (especially the no-special-chars fast path) and strings with invalid escape sequences, with speed improvements of up to 3.3× on .NET 10.0 and 3.0× on .NET Framework 4.8.1, and the UnescapeAll allocation for strings with invalid escape sequences eliminated entirely on both runtimes.

Highlights

🚀 Speed Improvements (.NET 10.0)

Benchmark	Before	After	Speedup
`Escape_NoSpecialChars`	17.96 ns	5.37 ns	3.3×
`Escape_FewSpecialChars`	191.8 ns	84.9 ns	2.3×
`Escape_ManySpecialChars`	892.0 ns	391.9 ns	2.3×
`UnescapeAll_InvalidEscapeSequences`	38.5 ns	19.6 ns	2.0×
`EscapeWithCaching_FewSpecialChars`	48.9 ns	27.9 ns	1.8×
`EscapeWithCaching_ManySpecialChars`	50.0 ns	27.7 ns	1.8×
`RoundTrip_EscapeThenUnescape`	245.3 ns	148.4 ns	1.7×

🧹 Allocation Reductions (.NET 10.0)

Benchmark	Before	After	Reduction
`UnescapeAll_InvalidEscapeSequences`	48 B	0 B	100%

📊 .NET Framework 4.8.1

Benchmark	Before	After	Speedup
`UnescapeAll_InvalidEscapeSequences`	89.7 ns	30.3 ns	3.0×
`Escape_NoSpecialChars`	79.9 ns	36.6 ns	2.2×
`Escape_ManySpecialChars`	1,073.6 ns	463.7 ns	2.3×
`Escape_FewSpecialChars`	349.5 ns	198.6 ns	1.8×
`RoundTrip_EscapeThenUnescape`	541.7 ns	371.0 ns	1.5×

The UnescapeAll_InvalidEscapeSequences allocation is also eliminated on .NET Framework (40 B → 0 B, 100% reduction).

Detailed Summary of Changes

Infrastructure

Add BufferScope<T>: a new ref struct that manages a stack-allocated buffer with ArrayPool<T> fallback for heap overflow, enabling low-allocation temporary storage.
Add RefArrayBuilder<T>: a new ref struct for cheaply building arrays using BufferScope<T>, used by the Escape two-pass algorithm.
Add benchmarks for EscapingUtilities: BenchmarkDotNet benchmarks covering Escape and UnescapeAll across typical MSBuild string patterns.

`Escape` optimizations

Use SearchValues<char> on .NET: replaces the IndexOfAny(char[]) scan with a SearchValues<char>-based scan, enabling vectorized SIMD search.
Replace IndexOfAny with a bitmask scan on .NET Framework: all 9 escapable characters fall in the ASCII range ['$'...'@'], so a 29-bit bitmask replaces the O(k) per-character array scan with a single range check + bit test.
Reuse first IndexOfAnyEscapeChar result: avoids a redundant scan pass by feeding the initial hit directly into the collection loop.
Refactor to a two-pass direct string allocation: a first pass collects all special character indices into a RefArrayBuilder<int>, then a second pass allocates a single exact-sized string (via string.Create on .NET; unsafe fixed+Buffer.MemoryCopy on .NET Framework), eliminating the StringBuilder entirely.

`UnescapeAll` optimizations

Acquire StringBuilder lazily: the StringBuilder is only rented from StringBuilderCache when the first decodable %XX sequence is actually found, so strings with % signs that aren't valid escape sequences return the original string with zero allocations.
Scope IndexOf('%') to the active trim window: when trim: true, the percent-sign search is bounded to [startIndex, endIndex) computed from the trimmed range, avoiding scanning whitespace that will be discarded.
Replace TryDecodeHexDigit arithmetic with HexConverter lookup table: delegates to the internal HexConverter.FromChar lookup table, replacing the multi-branch switch with a single table lookup.

Test improvements

Clean up EscapingUtilities_Tests: migrated to [Theory]/[InlineData] with Shouldly assertions and removed redundant test helpers.
Add more test coverage: expanded cases for UnescapeAll, UnescapeAll(trim: true), Escape, round-trip Escape↔Unescape, and ContainsEscapedWildcards.
Clean up EscapingUtilities in preparation for performance work: modernized XML docs, renamed parameters for clarity, and tightened nullability annotations ahead of the algorithmic changes.

Benchmarks

Initial Results (baselines)

.NET 10.0

Method	Mean	Error	StdDev	Gen0	Allocated
UnescapeAll_NoSpecialChars	2.612 ns	0.0785 ns	0.0696 ns	-	-
UnescapeAll_FewEscapeSequences	88.216 ns	0.7230 ns	0.6409 ns	0.0153	96 B
UnescapeAll_ManyEscapeSequences	176.489 ns	0.6002 ns	0.5012 ns	0.0100	64 B
UnescapeAll_InvalidEscapeSequences	38.494 ns	0.4281 ns	0.3795 ns	0.0076	48 B
UnescapeAll_WithTrim	56.222 ns	1.1704 ns	2.4171 ns	0.0076	48 B
Escape_NoSpecialChars	17.957 ns	0.4183 ns	0.4108 ns	-	-
Escape_FewSpecialChars	191.810 ns	2.7928 ns	2.4758 ns	0.0176	112 B
Escape_ManySpecialChars	892.012 ns	16.6366 ns	15.5619 ns	0.0324	208 B
EscapeWithCaching_FewSpecialChars	48.888 ns	0.9908 ns	0.9731 ns	-	-
EscapeWithCaching_ManySpecialChars	49.986 ns	0.9944 ns	1.3611 ns	-	-
ContainsEscapedWildcards_NoPercent	2.009 ns	0.0231 ns	0.0216 ns	-	-
ContainsEscapedWildcards_HasWildcards	2.895 ns	0.0433 ns	0.0405 ns	-	-
ContainsEscapedWildcards_LongNoWildcards	20.120 ns	0.1630 ns	0.1525 ns	-	-
RoundTrip_EscapeThenUnescape	245.305 ns	2.0098 ns	1.8799 ns	0.0329	208 B

.NET Framework 4.8.1

Method	Mean	Error	StdDev	Gen0	Allocated
UnescapeAll_NoSpecialChars	19.242 ns	0.2443 ns	0.2285 ns	-	-
UnescapeAll_FewEscapeSequences	182.132 ns	1.5881 ns	1.4855 ns	0.0160	84 B
UnescapeAll_ManyEscapeSequences	314.347 ns	2.0764 ns	1.9422 ns	0.0105	56 B
UnescapeAll_InvalidEscapeSequences	89.696 ns	0.6176 ns	0.5777 ns	0.0076	40 B
UnescapeAll_WithTrim	119.322 ns	0.8753 ns	0.8188 ns	0.0067	36 B
Escape_NoSpecialChars	79.902 ns	0.5406 ns	0.4793 ns	-	-
Escape_FewSpecialChars	349.474 ns	2.9539 ns	2.7631 ns	0.0191	100 B
Escape_ManySpecialChars	1,073.623 ns	5.7130 ns	5.3440 ns	0.0362	196 B
EscapeWithCaching_FewSpecialChars	75.406 ns	0.5825 ns	0.5449 ns	-	-
EscapeWithCaching_ManySpecialChars	63.058 ns	0.6023 ns	0.5634 ns	-	-
ContainsEscapedWildcards_NoPercent	18.369 ns	0.1220 ns	0.1081 ns	-	-
ContainsEscapedWildcards_HasWildcards	9.071 ns	0.0870 ns	0.0814 ns	-	-
ContainsEscapedWildcards_LongNoWildcards	56.651 ns	1.1451 ns	1.1247 ns	-	-
RoundTrip_EscapeThenUnescape	541.701 ns	10.4909 ns	11.6606 ns	0.0343	184 B

Final Results

.NET 10.0

Method	Mean	Error	StdDev	Gen0	Allocated
UnescapeAll_NoSpecialChars	2.705 ns	0.0772 ns	0.0722 ns	-	-
UnescapeAll_FewEscapeSequences	83.407 ns	1.7204 ns	1.6897 ns	0.0153	96 B
UnescapeAll_ManyEscapeSequences	176.586 ns	1.5432 ns	1.4435 ns	0.0100	64 B
UnescapeAll_InvalidEscapeSequences	19.590 ns	0.1503 ns	0.1333 ns	-	-
UnescapeAll_WithTrim	56.048 ns	0.8981 ns	0.8400 ns	0.0076	48 B
Escape_NoSpecialChars	5.368 ns	0.0523 ns	0.0489 ns	-	-
Escape_FewSpecialChars	84.876 ns	0.9482 ns	0.8870 ns	0.0178	112 B
Escape_ManySpecialChars	391.947 ns	2.9983 ns	2.8047 ns	0.0329	208 B
EscapeWithCaching_FewSpecialChars	27.894 ns	0.3138 ns	0.2935 ns	-	-
EscapeWithCaching_ManySpecialChars	27.667 ns	0.1934 ns	0.1809 ns	-	-
ContainsEscapedWildcards_NoPercent	1.990 ns	0.0231 ns	0.0216 ns	-	-
ContainsEscapedWildcards_HasWildcards	3.982 ns	0.0205 ns	0.0182 ns	-	-
ContainsEscapedWildcards_LongNoWildcards	19.671 ns	0.1140 ns	0.1066 ns	-	-
RoundTrip_EscapeThenUnescape	148.380 ns	1.9688 ns	1.8416 ns	0.0331	208 B

.NET Framework 4.8.1

Method	Mean	Error	StdDev	Gen0	Allocated
UnescapeAll_NoSpecialChars	19.585 ns	0.1821 ns	0.1520 ns	-	-
UnescapeAll_FewEscapeSequences	175.258 ns	1.3275 ns	1.1768 ns	0.0160	84 B
UnescapeAll_ManyEscapeSequences	299.969 ns	1.9503 ns	1.7289 ns	0.0105	56 B
UnescapeAll_InvalidEscapeSequences	30.317 ns	0.4839 ns	0.4527 ns	-	-
UnescapeAll_WithTrim	115.244 ns	1.0802 ns	1.0104 ns	0.0067	36 B
Escape_NoSpecialChars	36.594 ns	0.3431 ns	0.3209 ns	-	-
Escape_FewSpecialChars	198.554 ns	1.2490 ns	0.9751 ns	0.0191	100 B
Escape_ManySpecialChars	463.734 ns	4.5158 ns	4.0031 ns	0.0372	196 B
EscapeWithCaching_FewSpecialChars	61.413 ns	1.0921 ns	1.0215 ns	-	-
EscapeWithCaching_ManySpecialChars	51.776 ns	0.8385 ns	0.7843 ns	-	-
ContainsEscapedWildcards_NoPercent	18.567 ns	0.0863 ns	0.0720 ns	-	-
ContainsEscapedWildcards_HasWildcards	9.044 ns	0.0935 ns	0.0829 ns	-	-
ContainsEscapedWildcards_LongNoWildcards	54.810 ns	0.3122 ns	0.2767 ns	-	-
RoundTrip_EscapeThenUnescape	371.026 ns	5.3404 ns	4.9954 ns	0.0348	184 B

- Convert Fact-based tests to Theory with InlineData - Replace xUnit assertions with Shouldly - Remove empty XML doc comments and #nullable disable - Use expression-bodied test methods

Cover UnescapeAll, Escape, EscapeWithCaching, ContainsEscapedWildcards, and round-trip scenarios with MemoryDiagnoser to establish a performance baseline before optimization.

@JeremyKuhne

Many thanks to @JeremyKuhne for providing this BufferScope<T> implementation.

- Make public-facing methods `public` (was `internal`) - Enable nullable reference types; add `[return: NotNullIfNotNull]` to `UnescapeAll` and `Escape` - Use file-scoped namespace - Merge `Escape` + `EscapeWithCaching` into `Escape(string? value, bool cache = false)` - Inline `AppendEscapedChar` and `AppendEscapedString` helpers into `Escape` - Rename field `s_unescapedToEscapedStrings` → `s_escapedStringCache` - Rename parameters to `value` consistently across all public methods - Rename local variables for clarity (`percentIndex`, `startIndex`, `endIndex`, `hi`/`lo`, `specialCharIndex`, etc.) - Use collection expression for `s_charsToEscape`; target-typed `new` for cache dictionary - Use expression bodies for simple members (`HexDigitChar`, `Escape`, etc.) - Use relational pattern matching in `TryDecodeHexDigit`; list patterns in `ContainsEscapedWildcards` - Use `do...while` in `UnescapeAll` since `percentIndex` is already known non-negative on entry - Replace `ch / 0x10` with `ch >> 4` in nibble extraction - Clean up XML doc comments throughout - Remove stale and redundant inline comments

- Address unnecessary allocation when a string contains a '%' but no valid escape sequence. - Add test for edge case when a string contains a '%' but no valid escape sequence and 'trim' is set to true. In that case, the trimmed string should still be returned.

- Add System.Buffers.SearchValues<char> field initialized from s_charsToEscape (#if NET) - Extract IndexOfAnyEscapeChar helper to abstract the platform difference, keeping Escape itself free of #if guards - Fall back to char[] / IndexOfAny on .NET Framework

- All chars in s_charsToEscape fall within the ASCII range ['$' (0x24) .. '@' (0x40)] - Encode membership as a 29-bit uint bitmask indexed by (c - '$') - Replace O(n×k) IndexOfAny array scan with an O(n) range check + bit test per char - Eliminates managed-to-native transition overhead on each IndexOfAnyEscapeChar call - No change on .NET (SearchValues path unchanged)

- Capture the result of the initial IndexOfAnyEscapeChar(value) fast-path check rather than discarding it and repeating the same scan at the start of the loop - Switch from while(true)/break to do...while since specialCharIndex >= 0 is established before entering the loop, eliminating a redundant branch per iteration

Collect all special char positions in a first pass using RefArrayBuilder<int>, compute the exact output length, then write directly into a freshly-allocated string with no intermediate buffer: - On .NET: string.Create writes directly into the output string without an intermediate buffer - On .NET Framework: new string('\0', length) + Buffer.MemoryCopy via unsafe fixed pointers for fast, native-speed chunk copies - Extract cache operations into TryGetFromCache and AddToCache helpers - Remove StringBuilderCache dependency from Escape entirely

- Compute trim bounds (startIndex/endIndex) before searching for '%' so the initial scan and inner-loop scan skip leading/trailing whitespace - Scope the inner-loop IndexOf to [percentIndex+1, endIndex) to match - Extract GetDefaultResult static local to deduplicate the no-escape- sequences return path (used in both the early-exit and sb-is-null cases)

- Add HexConverter with a 256-entry hex digit table (ReadOnlySpan<byte> on .NET; static byte[] on Framework — the Framework JIT can't elide bounds checks on ReadOnlySpan<byte> like the .NET Core JIT can) - TryDecodeHexDigit delegates to HexConverter.FromChar

Copilot

Pull request overview

This pull request refactors Microsoft.Build.Shared.EscapingUtilities hot paths (Escape, UnescapeAll) and adds supporting low-allocation infrastructure (pooled buffer + ref-struct array builder) plus benchmarks and expanded unit tests to validate the new behavior and performance characteristics.

Changes:

Rewrite EscapingUtilities.Escape and EscapingUtilities.UnescapeAll to reduce allocations and improve throughput (including new fast paths and cache-aware escape).
Add new low-allocation helper types (BufferScope<T>, RefArrayBuilder<T>, HexConverter, TypeInfo<T>) plus polyfills needed for newer language/runtime features across TFMs.
Add/expand benchmarks and unit tests for escaping/unescaping and the new infrastructure.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/Framework/EscapingUtilities.cs	Rewrites escape/unescape implementations for fewer allocations and faster scanning/encoding.
src/Framework/Utilities/BufferScope.cs	Adds pooled-buffer ref struct to support temporary storage with minimal allocations.
src/Framework/Collections/RefArrayBuilder.cs	Adds ref-struct builder over `BufferScope<T>` for efficient index collection / array building.
src/Framework/Utilities/HexConverter.cs	Adds lookup-table hex decoding helper used by unescape fast path.
src/Framework/Utilities/TypeInfo.cs	Adds type inspection helper used to decide whether to clear pooled arrays on return.
src/Framework/Polyfills/UnscopedRefAttribute.cs	Adds polyfill / forwarder for `UnscopedRefAttribute` across target frameworks.
src/Framework/Polyfills/StringExtensions.cs	Adds string null/empty/whitespace extension helpers used by updated code paths.
src/MSBuild.Benchmarks/EscapingUtilitiesBenchmark.cs	Adds BenchmarkDotNet coverage for `Escape`, `UnescapeAll`, and related helpers.
src/Framework.UnitTests/EscapingUtilities_Tests.cs	Expands and modernizes escaping/unescaping tests (Shouldly + theories).
src/Framework.UnitTests/BufferScopeTests.cs	Adds test coverage for pooled buffer behavior, growth, pinning, and disposal.
src/Framework.UnitTests/RefArrayBuilder_Tests.cs	Adds extensive coverage for builder operations (add/insert/remove/range/etc.).
src/Framework.UnitTests/TypeInfoTests.cs	Adds coverage for `TypeInfo<T>.IsReferenceOrContainsReferences()` behavior and caching.
src/Utilities/TaskItem.cs	Updates call site to use `Escape(..., cache: true)` instead of `EscapeWithCaching`.
src/Shared/TaskParameter.cs	Updates call site to use `Escape(..., cache: true)` instead of `EscapeWithCaching`.
src/Build/BackEnd/TaskExecutionHost/TaskExecutionHost.cs	Updates call site to use `Escape(..., cache: true)`; removes trailing whitespace.
src/Utilities.UnitTests/StringExtensions_Tests.cs	Disambiguates `StringExtensions.Replace` call site due to new `Microsoft.Build.StringExtensions`.

ViktorHofer · 2026-03-30T17:31:42Z

Looks great. Trying to better understand the intent here. Did this path show up hot on a perf trace somewhere?

DustinCampbell · 2026-03-30T17:40:13Z

Looks great. Trying to better understand the intent here. Did this path show up hot on a perf trace somewhere?

I added some quick-and-dirty and found that these methods are called a lot. In my tests, I built a significant part of Roslyn and found 3,000,000+ invocations of these methods. I'm definitely intending to reduce the number of times these are called, but I decided to improve the performance of the methods themselves as a first step.

Ensure that Insert(...) only takes the fast path if "_count < _scope.Length" to potentially shifting past the end of the scope. Added tests to cover this issue, which is more likely to appear with a stack-allocated buffer is used, since a rented array might actually be larger than the requested minimum length.

ViktorHofer · 2026-03-30T18:16:28Z

Sweet. Are these changes already noticeable end-to-end in the roslyn build (the eval part of it)?

DustinCampbell · 2026-03-30T19:07:46Z

Sweet. Are these changes already noticeable end-to-end in the roslyn build (the eval part of it)?

What metric did you mean? Obviously, this change is reducing milliseconds and lowering allocations. Many such changes in aggregate would likely show wall clock improvements.

I missed adding string resources to SR.resx when cherry-picking RefArrayBuilder from another local branch.

JanKrivanek · 2026-04-15T14:29:47Z

/review

github-actions · 2026-04-15T14:30:28Z

✅ Expert Code Review (command) completed successfully!

Correctness review complete. All edge cases traced — no concrete failing scenario found. Dimension is LGTM.

github-actions · 2026-04-15T14:56:17Z

Expert Code Review — 24-Dimension Analysis

Excellent performance work! The algorithmic improvements are sound — bitmask scan, two-pass direct allocation, lazy StringBuilder, and SearchValues integration are all well-designed. The benchmarks are thorough and the speedups are impressive (up to 3.3× on .NET 10).

Dimension Summary

#	Dimension	Verdict
1	Backwards Compatibility	✅ LGTM — `EscapingUtilities` and all changed methods are `internal`; no external callers affected
2	ChangeWave Discipline	✅ LGTM — pure perf optimization, no behavioral change
3	Performance & Allocation	⚠️ NIT — `RefArrayBuilder(64)` always rents from ArrayPool; consider `stackalloc int[8]` for common case
4	Test Coverage	✅ LGTM — comprehensive Theory/InlineData coverage, good edge cases
5	Error Message Quality	✅ LGTM — new SR.resx strings are clear
6	Logging & Diagnostics	✅ LGTM — N/A for this change
7	String Comparison	✅ LGTM — all comparisons are ordinal/char-level
8	API Surface Discipline	❌ MAJOR — `TypeInfo<T>` is `public` but should be `internal`
9	MSBuild Target Authoring	✅ LGTM — N/A
10	Design Before Implementation	✅ LGTM — well-structured two-pass approach
11	Cross-Platform Correctness	✅ LGTM — bitmask and `SearchValues` produce identical results; unsafe code is sound
12	Code Simplification	✅ LGTM — cleaner than the old code
13	Concurrency & Thread Safety	✅ LGTM — lock-guarded cache, benign `TypeInfo` race, ref structs are stack-local
14	Naming Precision	✅ LGTM — clear parameter/method names
15	SDK Integration	✅ LGTM — N/A
16	Idiomatic C#	✅ LGTM — good use of pattern matching, ref structs, SearchValues
17	File I/O & Path Handling	✅ LGTM — N/A
18	Documentation	✅ LGTM — XML docs are thorough
19	Build Infrastructure	✅ LGTM — no build config changes
20	Scope & PR Discipline	⚠️ MODERATE — `RefArrayBuilder` ships ~750 lines of unused code + 13 XLF files
21	Evaluation Model Integrity	✅ LGTM — N/A
22	Correctness & Edge Cases	✅ LGTM — bitmask verified, UnescapeAll logic traced, ContainsEscapedWildcards bounds correct
23	Dependency Management	✅ LGTM — no new external dependencies
24	Security Awareness	✅ LGTM — unsafe code is bounded, no user-controlled indices

Key Findings

MAJOR — TypeInfo<T> is public (src/Framework/Utilities/TypeInfo.cs:16): Only used by internal types. Should be internal to avoid permanently exposing implementation details on the Framework NuGet package surface. See inline comment.
MODERATE — RefArrayBuilder<T> scope (src/Framework/Collections/RefArrayBuilder.cs): 851 lines added but only Add()/AsSpan()/Dispose() used. The unused LINQ-like API also pulls in 2 SR.resx strings → 13 XLF file changes. Consider trimming to what's needed now.
NIT — Unused variable (src/Framework/Utilities/TypeInfo.cs:52): Type type = typeof(T) is assigned but never read. May cause CS0219 warning/error.
NIT — ArrayPool allocation (src/Framework/EscapingUtilities.cs:195): new RefArrayBuilder<int>(initialCapacity: 64) could use stackalloc int[8] to avoid pool overhead for the >99% common case.

Generated by Expert Code Review (command) for issue #13426 · ● 26.6M · ◷

github-actions

Performance & Allocation Awareness

All other patterns are clean:

Two-pass Escape (string.Create / unsafe fixed) — exact-size allocation, no waste. ✅
Lazy StringBuilder in UnescapeAll — acquire/release pattern is correct; sb is only acquired on the first valid %XX hit, so purely-invalid or escape-free strings allocate nothing. ✅
IndexOfAnyEscapeChar — SearchValues<char> on .NET (SIMD), bitmask scan on .NET Framework. Both are tighter than the old IndexOfAny(char[]). ✅
HexConverter — ReadOnlySpan<byte> from PE data section on .NET (no allocation); static readonly byte[] on .NET Framework (single allocation, cached). ✅
BufferScope.Dispose — clears only when TypeInfo<T>.IsReferenceOrContainsReferences(), avoiding unnecessary zeroing for value types like int. ✅

One issue raised inline: RefArrayBuilder(initialCapacity: 64) always rents from ArrayPool<int>.Shared even for the common case of 1–2 special chars. A stackalloc int[8] scratch buffer via the existing Span<T> constructor eliminates that overhead for the common case.

Generated by Expert Code Review (command) for issue #13426 · ● 26.6M

- Change TypeInfo<T> from public to internal - Remove unused "type" variable - Prefer null-coalescing assignment operator

DustinCampbell · 2026-04-15T16:35:13Z

MAJOR — TypeInfo<T> is public (src/Framework/Utilities/TypeInfo.cs:16): Only used by internal types. Should be internal to avoid permanently exposing implementation details on the Framework NuGet package surface. See inline comment.

Thanks for catching this! It's now fixed.

MODERATE — RefArrayBuilder<T> scope (src/Framework/Collections/RefArrayBuilder.cs): 851 lines added but only Add()/AsSpan()/Dispose() used. The unused LINQ-like API also pulls in 2 SR.resx strings → 13 XLF file changes. Consider trimming to what's needed now.

This is intentional. Because it is tricky to implement, RefArrayBuilder<T> is intended to have a fairly complete API. It will be used more throughout the MSBuild code base in the near future.

NIT — Unused variable (src/Framework/Utilities/TypeInfo.cs:52): Type type = typeof(T) is assigned but never read. May cause CS0219 warning/error.

Fixed.

NIT — ArrayPool allocation (src/Framework/EscapingUtilities.cs:195): new RefArrayBuilder<int>(initialCapacity: 64) could use stackalloc int[8] to avoid pool overhead for the >99% common case.

The initial capacity has been reduced from 64 to 16. However, passing stackalloc int[8] results in slower code on .NET Framework because it increases the stack frame for every call to Escape by 32-bytes. Using stackalloc here noticeably degrades benchmarks for Escape.

DustinCampbell · 2026-04-15T18:25:53Z

@JanKrivanek: The expert code review feedback has been addressed.

DustinCampbell · 2026-04-15T18:26:21Z

@JeremyKuhne: Here's the PR that includes BufferScope<T> that you had asked about.

Replace hand-written [DllImport] declarations and COM [ComImport] interfaces with CsWin32 source-generated interop. This enables Native AOT scenarios where built-in COM marshalling and System.Runtime.InteropServices marshalling infrastructure do not exist. The approach follows patterns established in WinForms, WPF, and the dotnet/sdk repository. CsWin32 is configured in Microsoft.Build.Framework with allowMarshaling: false and useSafeHandles: false, producing raw pointer signatures that are AOT-compatible. Other projects consume generated types via InternalsVisibleTo. Key changes: Infrastructure: - Add FEATURE_WINDOWSINTEROP compile-time gate in Directory.BeforeCommon.targets, disabled for source-only builds - Add CsWin32 (Microsoft.Windows.CsWin32) and PolySharp package references to Framework.csproj - Add NativeMethods.txt and NativeMethods.json for CsWin32 configuration - Add .editorconfig suppression for CS3019 in Windows/ folders - Add cswin32-interop skill document for agent guidance New utility types: - BufferScope<T>: ref struct for stack-allocated buffers with ArrayPool fallback, replacing manual ArrayPool rent/return and stackalloc patterns throughout interop code - TypeInfo<T>: cached RuntimeHelpers.IsReferenceOrContainsReferences polyfill for net472 - ComScope<T>: COM pointer lifetime management (using-disposable) - ComClassFactory: AOT-compatible COM activation without Activator.CreateInstance - IID: generic IID lookup via IComIID interface - VARIANT, BSTR, HRESULT, FILETIME extensions: CsWin32 partial type augmentations for safe usage patterns NativeMethods.cs cleanup: - Delete ~130 hand-written constants, enums, structs, and [DllImport] declarations replaced by CsWin32 typed equivalents - Replace MemoryStatus class with MEMORYSTATUSEX struct via TryGetMemoryStatus - Replace GetCurrentDirectory/GetFullPath/GetShortPathName/ GetLongPathName with BufferScope-based implementations - Replace StreamHandleType enum with bool useStandardError parameter - Delete DirectoryExists/FileExists/FileOrDirectoryExists wrappers; callers use PInvoke.GetFileAttributes directly - Delete HResultSucceeded/HResultFailed; callers use HRESULT.Failed - Add [UnsupportedOSPlatformGuard("windows")] to IsUnixLike - Version all [SupportedOSPlatform] attributes to "windows6.1" COM interop modernization: - Replace [ComImport] WMI interfaces (IWbemLocator, IWbemServices, IEnumWbemClassObject, IWbemClassObject) with struct-based COM using delegate* unmanaged[Stdcall] vtables - Add IDebugClient4-based command line retrieval as alternative to WMI (CommandLineSource.DebugEngine) - Move IFixedTypeInfo from NativeMethods.cs to its own file in Tasks Platform guard pattern: - Apply dual-guard pattern: #if FEATURE_WINDOWSINTEROP wraps runtime IsWindows checks, eliminating dead code in source builds - Use IsUnixLike (positive guard) instead of !IsWindows for CA1416 platform compatibility analysis - Files excluded at project level via <Compile Remove> when FeatureWindowsInterop is off (WindowsFileSystem, Windows/ folder, WMI structs) Behavioral changes: - Unzip.cs: !IsWindows → IsUnixLike for correct CA1416 analysis - ProcessExtensions: ArrayPool<T> replaced with BufferScope<T> - AssemblyInformation/GlobalAssemblyCache: remove dead non-Windows code paths in net472-only (#if !FEATURE_ASSEMBLYLOADCONTEXT) blocks, which only run on Windows Some related changes are also in dotnet#13426.

Base automatically changed from dev/dustinca/itemspecmodifiers-perf to main March 24, 2026 13:05

SimaTian requested a review from a team as a code owner March 24, 2026 13:05

AR-May assigned AR-May and SimaTian and unassigned AR-May Mar 24, 2026

DustinCampbell added 13 commits March 30, 2026 10:08

Clean up EscapingUtilities_Tests

de94291

- Convert Fact-based tests to Theory with InlineData - Replace xUnit assertions with Shouldly - Remove empty XML doc comments and #nullable disable - Use expression-bodied test methods

Add benchmarks for EscapingUtilities

fac5a67

Cover UnescapeAll, Escape, EscapeWithCaching, ContainsEscapedWildcards, and round-trip scenarios with MemoryDiagnoser to establish a performance baseline before optimization.

Add BufferScope<T> to manage stack and ArrayPool<T> buffers

8a07ffd

Many thanks to @JeremyKuhne for providing this BufferScope<T> implementation.

Add RefArrayBuilder<T> for building arrays cheaply

58dff16

Add more test coverage to EscapingUtilities_Tests

73f2150

DustinCampbell force-pushed the dev/dustinca/escapingutilities-perf branch from e06a889 to 2ff888d Compare March 30, 2026 17:24

Copilot AI review requested due to automatic review settings March 30, 2026 17:24

Copilot started reviewing on behalf of DustinCampbell March 30, 2026 17:24 View session

Copilot AI reviewed Mar 30, 2026

View reviewed changes

Comment thread src/Framework/Collections/RefArrayBuilder.cs Outdated

DustinCampbell added 3 commits March 31, 2026 08:39

Merge branch 'main' into escapingutilities-perf

39929a1

Merge branch 'main' into escapingutilities-perf

e7222d5

Add string resources for RefArrayBuilder

3f580df

I missed adding string resources to SR.resx when cherry-picking RefArrayBuilder from another local branch.

DustinCampbell added 5 commits April 7, 2026 08:38

Merge branch 'main' into escapingutilities-perf

9ba9649

Merge branch 'main' into escapingutilities-perf

dc7edc4

Merge branch 'main' into escapingutilities-perf

b336ca0

Merge branch 'main' into escapingutilities-perf

606494c

Merge branch 'main' into escapingutilities-perf

4a463ab

github-actions bot requested changes Apr 15, 2026

View reviewed changes

DustinCampbell added 3 commits April 15, 2026 09:11

Merge branch 'main' into escapingutilities-perf

4255f02

Fix TypeInfo accessibility and clean up

414899a

- Change TypeInfo<T> from public to internal - Remove unused "type" variable - Prefer null-coalescing assignment operator

EscapingUtilities: Reduce initial capacity of RefArrayBuilder<T> to 16

4fb6252

Merge branch 'main' into escapingutilities-perf

4c60b18

Merge branch 'main' into escapingutilities-perf

d7c2ac0

JeremyKuhne mentioned this pull request Apr 20, 2026

Migrate Windows interop to CsWin32 source-generated P/Invoke #13576

Open

Merge branch 'main' into escapingutilities-perf

3d6782b

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Performance optimizations for EscapingUtilities#13426

Performance optimizations for EscapingUtilities#13426
DustinCampbell wants to merge 28 commits intomainfrom
dev/dustinca/escapingutilities-perf

DustinCampbell commented Mar 20, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

ViktorHofer commented Mar 30, 2026

Uh oh!

DustinCampbell commented Mar 30, 2026

Uh oh!

ViktorHofer commented Mar 30, 2026

Uh oh!

DustinCampbell commented Mar 30, 2026 •

edited

Loading

Uh oh!

JanKrivanek commented Apr 15, 2026

Uh oh!

github-actions bot commented Apr 15, 2026 •

edited

Loading

Uh oh!

github-actions bot commented Apr 15, 2026

Uh oh!

github-actions bot left a comment

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Conversation

DustinCampbell commented Mar 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Highlights

🚀 Speed Improvements (.NET 10.0)

🧹 Allocation Reductions (.NET 10.0)

📊 .NET Framework 4.8.1

Detailed Summary of Changes

Infrastructure

Escape optimizations

UnescapeAll optimizations

Test improvements

Benchmarks

Initial Results (baselines)

.NET 10.0

.NET Framework 4.8.1

Final Results

.NET 10.0

.NET Framework 4.8.1

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

ViktorHofer commented Mar 30, 2026

Uh oh!

DustinCampbell commented Mar 30, 2026

Uh oh!

ViktorHofer commented Mar 30, 2026

Uh oh!

DustinCampbell commented Mar 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

JanKrivanek commented Apr 15, 2026

Uh oh!

github-actions bot commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Apr 15, 2026

Expert Code Review — 24-Dimension Analysis

Dimension Summary

Key Findings

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Performance & Allocation Awareness

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

DustinCampbell commented Apr 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

DustinCampbell commented Mar 20, 2026 •

edited

Loading

`Escape` optimizations

`UnescapeAll` optimizations

DustinCampbell commented Mar 30, 2026 •

edited

Loading

github-actions bot commented Apr 15, 2026 •

edited

Loading