Improve MSBuild expansion performance by DustinCampbell · Pull Request #14697 · dotnet/msbuild

DustinCampbell · 2026-08-12T22:36:00Z

Important

This depends on #14660. Until that PR merges, this PR includes its benchmark commits; the expansion changes begin after bd61887.

Tip

Each commit is self-contained and has a good description. Consider reviewing commit-by-commit starting at 71d8b82.

Context

Expansion repeatedly scans strings for property, item, and metadata markers and performs parsing, dispatch, and context resolution for every match. These operations are pervasive during project evaluation, so even small per-expression or per-item costs accumulate across large solutions.

This change optimizes those hot paths while preserving existing expansion behavior.

Note

This is the first in a series of expansion performance PRs. It focuses on targeted, behavior-preserving reductions in CPU overhead; allocation-focused work and more fundamental architectural changes are intentionally left for follow-up PRs. There is still substantial low-hanging fruit in these paths.

Changes Made

Add specialized marker detection and indexing APIs for $(, @(, and %( expressions, replacing duplicated and broader searches throughout evaluation.
Replace the stateful item-expression enumerator with indexed capture parsing and avoid wrapping large item captures in Nullable<T>.
Specialize common quoted-transform shapes so literal, exact, embedded-single, and multiple-match expressions avoid general-purpose scanning and builders.
Separate property-specific parenthesis scanning from general function argument parsing and classify ordinary properties while scanning.
Encapsulate property expansion state in a readonly ref struct and cache invariant expansion policy.
Join explicitly separated item vectors directly with ValueStringBuilder on .NET and StringBuilderCache on .NET Framework, avoiding intermediate collections and LINQ.
Cache item-spec modifier classification while parsing quoted-transform metadata.
Dispatch implicit quoted transforms directly instead of synthesizing an intrinsic function call and argument array.
Resolve project-directory and defining-project context only for modifiers that consume those values.

Benchmark Results

The baseline is bd61887f13; the optimized implementation is 34c6229708. Negative timing percentages indicate faster execution. The four benchmark classes were unchanged between these commits.

🚀 Performance Highlights

Area	.NET 11	LegacyJIT x86	Allocation impact
🔍 Marker detection (`NoExpansion`)	30.9–51.1% faster	40.6–56.0% faster	➖ Unchanged
✨ Quoted transform, 100 items	48.3% faster	23.6% faster	📉 -32 B / -17 B
🔗 Explicit separator, 100 items	12.1% faster	28.9% faster	📉 -1,664 B / -844 B
🏷️ Qualified metadata	32.9% faster	19.1% faster	➖ Unchanged
⚙️ Multiple distinct properties	15.8% faster	14.7% faster	➖ Unchanged
🧩 Mixed property, metadata, and 100-item transform	59.3% faster	27.0% faster	📉 -32 B / -16 B

.NET 11.0 — RyuJIT x64

ItemExpansionBenchmark

Method	Items	Mean: baseline → latest	Time	Allocated: baseline → latest
NoExpansion	10	26.22 → 13.92 ns	-46.9%	0 → 0 B
Simple	10	573.49 → 557.89 ns	-2.7%	280 → 280 B
WithQuotedTransform	10	3,029.15 → 1,712.49 ns	-43.5%	1,024 → 992 B (-32 B)
WithSeparator	10	600.19 → 530.68 ns	-11.6%	768 → 544 B (-29.2%)
WithQuotedTransformAndUnescape	10	3,016.00 → 1,725.81 ns	-42.8%	1,024 → 992 B (-32 B)
NoExpansion	100	24.80 → 13.62 ns	-45.1%	0 → 0 B
Simple	100	3,923.71 → 3,847.24 ns	-1.9%	1,720 → 1,720 B
WithQuotedTransform	100	25,621.73 → 13,239.86 ns	-48.3%	7,505 → 7,473 B (-32 B)
WithSeparator	100	3,552.17 → 3,123.38 ns	-12.1%	5,624 → 3,960 B (-29.6%)
WithQuotedTransformAndUnescape	100	25,539.33 → 13,135.71 ns	-48.6%	7,505 → 7,473 B (-32 B)

MetadataExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated
NoExpansion	15.81 → 10.92 ns	-30.9%	0 → 0 B
Unqualified	167.10 → 116.72 ns	-30.1%	0 → 0 B
Qualified	227.22 → 152.43 ns	-32.9%	48 → 48 B
Multiple	290.44 → 224.46 ns	-22.7%	0 → 0 B
MultipleAndUnescape	307.20 → 224.48 ns	-26.9%	0 → 0 B

MixedExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated: baseline → latest
NoExpansion	38.72 → 18.94 ns	-51.1%	0 → 0 B
PropertyAndItem	17,827.80 → 7,470.08 ns	-58.1%	7,497 → 7,465 B (-32 B)
PropertyAndMetadata	525.56 → 405.57 ns	-22.8%	0 → 0 B
All	19,144.79 → 7,785.14 ns	-59.3%	7,665 → 7,633 B (-32 B)
AllAndUnescape	18,446.47 → 7,810.63 ns	-57.7%	7,665 → 7,633 B (-32 B)

PropertyExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated
NoExpansion	16.95 → 11.52 ns	-32.0%	0 → 0 B
Single	103.20 → 80.27 ns	-22.2%	0 → 0 B
Embedded	209.95 → 205.71 ns	-2.0%	40 → 40 B
MultipleDistinct	481.61 → 405.37 ns	-15.8%	0 → 0 B
MultipleRepeated	472.70 → 402.12 ns	-14.9%	0 → 0 B
Undefined	183.21 → 183.53 ns	+0.2%	40 → 40 B
LongValue	88.46 → 73.90 ns	-16.5%	0 → 0 B
MultipleAndUnescape	413.11 → 416.51 ns	+0.8%	0 → 0 B

.NET Framework 4.8.1 — LegacyJIT x86

ItemExpansionBenchmark

Method	Items	Mean: baseline → latest	Time	Allocated: baseline → latest
NoExpansion	10	191.4 → 112.6 ns	-41.2%	0 → 0 B
Simple	10	2,793.1 → 2,667.0 ns	-4.5%	152 → 152 B
WithQuotedTransform	10	13,727.2 → 10,449.7 ns	-23.9%	962 → 946 B (-16 B)
WithSeparator	10	2,246.3 → 1,674.0 ns	-25.5%	593 → 469 B (-20.9%)
WithQuotedTransformAndUnescape	10	13,778.2 → 10,467.6 ns	-24.0%	962 → 946 B (-16 B)
NoExpansion	100	190.9 → 112.4 ns	-41.1%	0 → 0 B
Simple	100	19,269.2 → 18,822.0 ns	-2.3%	874 → 874 B
WithQuotedTransform	100	120,620.9 → 92,212.8 ns	-23.6%	7,816 → 7,799 B (-17 B)
WithSeparator	100	10,977.8 → 7,810.4 ns	-28.9%	9,162 → 8,318 B (-9.2%)
WithQuotedTransformAndUnescape	100	121,049.3 → 91,529.3 ns	-24.4%	7,816 → 7,799 B (-17 B)

MetadataExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated
NoExpansion	172.0 → 102.1 ns	-40.6%	0 → 0 B
Unqualified	946.8 → 712.2 ns	-24.8%	0 → 0 B
Qualified	1,120.2 → 905.8 ns	-19.1%	40 → 40 B
Multiple	1,454.9 → 1,139.5 ns	-21.7%	0 → 0 B
MultipleAndUnescape	1,523.0 → 1,170.3 ns	-23.2%	0 → 0 B

MixedExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated: baseline → latest
NoExpansion	343.4 → 151.2 ns	-56.0%	0 → 0 B
PropertyAndItem	79,038.1 → 57,716.0 ns	-27.0%	7,806 → 7,790 B (-16 B)
PropertyAndMetadata	2,541.3 → 2,103.6 ns	-17.2%	0 → 0 B
All	80,833.6 → 58,976.5 ns	-27.0%	7,914 → 7,898 B (-16 B)
AllAndUnescape	81,588.6 → 59,358.2 ns	-27.2%	7,914 → 7,898 B (-16 B)

PropertyExpansionBenchmark

Method	Mean: baseline → latest	Time	Allocated
NoExpansion	166.0 → 88.49 ns	-46.7%	0 → 0 B
Single	459.8 → 333.3 ns	-27.5%	0 → 0 B
Embedded	1,037.3 → 879.82 ns	-15.2%	28 → 28 B
MultipleDistinct	2,269.6 → 1,935.24 ns	-14.7%	0 → 0 B
MultipleRepeated	2,194.5 → 1,921.38 ns	-12.4%	0 → 0 B
Undefined	959.3 → 827.33 ns	-13.8%	28 → 28 B
LongValue	441.7 → 300.15 ns	-32.0%	0 → 0 B
MultipleAndUnescape	2,301.5 → 1,908.8 ns	-17.1%	0 → 0 B

Analysis

Quoted item transforms improve by 42.8–48.6% on .NET 11 and 23.6–24.4% on LegacyJIT. Savings scale with item count, confirming that per-item modifier classification and context lookup work was removed.
Mixed expressions containing a 100-item quoted transform improve by 57.7–59.3% on .NET 11 and approximately 27% on LegacyJIT.
Metadata expansion improves by 19.1–32.9% across qualified, unqualified, multiple, and unescaping cases without allocation changes.
Property expansion improves by 12.4–32.0% on LegacyJIT. Common .NET 11 property shapes improve by 14.9–22.2%; the small changes to embedded, undefined, and unescaping cases are within benchmark noise.
Explicit separator expansion reduces allocations by approximately 29% on .NET 11 and saves 124–844 B on LegacyJIT while improving execution by 11.6–28.9%.
The fixed 32 B x64 and approximately 16 B x86 reduction in quoted and mixed item benchmarks matches removal of the synthetic transform argument array.
No benchmark case shows a meaningful timing regression or allocation increase.

Testing

Full test suite passes.
Captured BenchmarkDotNet baselines and optimized results for ItemExpansionBenchmark, MetadataExpansionBenchmark, MixedExpansionBenchmark, and PropertyExpansionBenchmark on .NET 11.0 x64 and .NET Framework 4.8.1 LegacyJIT x86.

Notes

This change intentionally preserves existing expansion semantics and public API behavior. The focused benchmark infrastructure is reviewed separately in #14660.

Replace the monolithic Expander benchmark with targeted property, item, metadata, mixed-expression, condition, shredder, and lazy-item suites. Add reusable setup helpers and make cache-state expectations explicit. Add a cross-TFM PowerShell runner with broad and granular benchmark sets, storing results under the repository artifacts directory.

Create fresh item instances for each iteration so cold and repeated modifier lookups measure the intended cache state. Clear the shared defining-project cache after setup to prevent state leaking between iterations. Batch item accesses and normalize results with OperationsPerInvoke. Move item discovery and materialization out of the measured code.

Separate cached escaping into explicit cache-hit and cache-miss cases. Prime hit inputs before measurement and generate unique miss inputs for each iteration so warmup cannot change the measured cache state. Batch operations to reduce harness overhead while preserving input length and escaping characteristics.

Expose launch count through the benchmark runner for higher-confidence comparisons across independent processes. Keep power-plan enforcement opt-in so dedicated machines retain their configured plan. Document smoke, exploratory, and confirmation runs, including execution cost and cross-framework comparison constraints.

Measure fixed property-expansion shapes with 10 and 100 unrelated properties in the backing bag. Include a no-expansion baseline to separate lookup effects from fixed benchmark overhead. Add a dedicated category and runner set while preserving the broader property-expansion scaling groups.

Resolve explicit relative artifact paths against PowerShell's current location instead of the process working directory, which may not track Set-Location. This ensures paths such as .\artifacts are created under the directory from which Run-Benchmarks.ps1 was invoked. Document the relative path behavior for the ArtifactsPath parameter.

Document the intentional cold-path and cache-isolation behavior of the focused evaluation benchmarks. Use the shared temporary-directory helper for modifier benchmarks to centralize setup and cleanup.

MSBuild repeatedly searches evaluation strings for the `$(`, `@(`, and `%(` syntax markers. Existing call sites used culture-aware searches, ordinal substring searches, broad single-character checks, and character-by-character loops. This duplicated work and made the common marker-absent path pay for slower two-character searches. Add named marker constants and specialized `Contains*Marker` and `IndexOf*Marker` APIs to `ExpressionShredder`. The index helpers search for a single-character prefix and check the following `(` directly. They support both unbounded and count-bounded ranges while leaving full expression validation to the existing parsers. Use the helpers throughout property, metadata, and item-vector expansion; SDK-reference property detection; item provenance; and intrinsic item metadata handling. This removes the unnecessary invariant-culture SDK-reference search, replaces broad `Contains('@')` checks with exact marker detection, and lets expansion loops jump directly between candidate references. Refactor `ReferencedItemExpressionsEnumerator` to jump between bounded `@(` markers while preserving resumable state. Refactor `GetReferencedItemNamesAndMetadata` to perform one bounded `IndexOfAny` scan for `@` and `%` instead of scanning the remaining input twice. Its explicit scan position preserves marker ordering, subrange boundaries, and malformed-expression recovery without manipulating a `for` loop counter. Add focused coverage for marker constants, all marker types, missing and trailing prefixes, rejected prefix characters, repeated markers, start indexes, bounded ranges, range boundaries, and containment checks.

Replace the stateful item expression enumerator with an indexed TryGetNextItemVectorExpression API. Update expansion callers, tests, and benchmarks to consume captures directly and avoid redundant marker scans.

Use the Try pattern with an out parameter when expanding a single item vector expression. This avoids wrapping the relatively large capture struct in Nullable<T> and copying it through Nullable<T>.Value.

Specialize literal, exact, embedded-single, and multiple-match transforms so common cases avoid general-purpose scanning and builders. Jump directly between metadata markers and hoist invariant boundaries while preserving existing expansion and allocation behavior.

Separate property-specific parenthesis scanning from the general function-argument scanner. Classify property and registry function candidates while scanning so ordinary properties go directly to lookup. Move uncommon compatibility and function handling into ExpandProperty, use string-based marker and quote scans, and replace magic comparison lengths with named constants. This reduces work in common property expansion paths and keeps specialized behavior isolated.

Encapsulate property expansion context in a readonly ref struct so helper methods no longer pass the same state through each call. Cache the truncation policy for reuse during expansion while preserving the existing static entry points.

Avoid materializing intermediate entries and LINQ iterators when joining item vectors with explicit separators. Use ValueStringBuilder on .NET and StringBuilderCache on .NET Framework. Preserve transform classification for empty item vectors while still evaluating Count and AnyHaveMetadataValue on empty lists.

Resolve item-spec modifier kinds when parsing quoted transform metadata instead of classifying the same name for every item. Add a non-caching enum overload of GetItemSpecModifier so quoted transforms can call the modifier implementation directly.

Dispatch unnamed quoted transforms directly instead of synthesizing an intrinsic function name and one-element argument array. This avoids modifier classification and intrinsic dictionary lookup. Keep explicitly named ExpandQuotedExpressionFunction invocations on the existing validated argument path.

Resolve project directory and defining-project metadata only for item-spec modifiers that consume those values. Common quoted transforms such as Filename now skip both per-item lookups.

github-actions

No blocking issues found in the snippet reviewed. GetItemSpecModifier already tolerates a null currentDirectory for Filename, Extension, RelativeDir, and Identity, and RecursiveDir is not derivable (TryGetDerivableModifierKind excludes it), so the new switch does not appear to introduce a functional regression. The change is likely performance-positive on this hot path because it avoids eagerly computing project context and DefiningProjectFullPath for modifiers that do not need them.

One concern: I could not find targeted regression coverage for this new dispatch in GetMetadataValueFromMatch, especially for quoted transform expansion of %(Filename), %(RelativeDir), and the %(DefiningProject*) modifiers. Existing tests cover ItemSpecModifiers.GetItemSpecModifier directly and some transform behavior, but not this exact expander-layer routing. A focused table-driven test here would better lock in the intended mapping and guard future additions to derivable modifiers.

Generated by Expert Code Review (on open) for #14697 · sonnet46 · 63.4 AIC · ⌖ 7.62 AIC · ⊞ 5.3K

Copilot

Pull request overview

This pull request optimizes MSBuild’s evaluation-time expansion hot paths (properties, item vectors, metadata, and quoted transforms) to reduce per-expression CPU overhead, and updates/expands the BenchmarkDotNet benchmark suite + runner script to measure these scenarios more directly.

Changes:

Refactors expansion marker detection and item-vector parsing (new indexing APIs in ExpressionShredder, new item-vector capture flow in Expander).
Optimizes item transform execution paths (quoted transforms dispatch, separator joins, metadata-match classification, and reduced context resolution for item-spec modifiers).
Reworks benchmark infrastructure and adds focused benchmark suites plus a cross-TFM runner script and updated documentation.

Reviewed changes

Copilot reviewed 43 out of 43 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/MSBuild.Benchmarks/Run-Benchmarks.ps1	Adds a cross-TFM benchmark runner with filter/set selection and common options.
src/MSBuild.Benchmarks/readme.md	Documents the new runner, named sets, and run modes.
src/MSBuild.Benchmarks/Program.cs	Adds `--enforce-power-plan` plumbing into the BenchmarkDotNet config overrides.
src/MSBuild.Benchmarks/ItemSpecModifiersCachingBenchmark.cs	Updates modifier-caching benchmark setup + per-iteration item creation.
src/MSBuild.Benchmarks/DefiningProjectModifiersBenchmark.cs	Updates defining-project modifier benchmarks to use per-iteration setup and new temp-dir infra.
src/MSBuild.Benchmarks/EscapingUtilitiesBenchmark.cs	Removes the caching cases (now covered by the dedicated caching benchmark).
src/MSBuild.Benchmarks/EscapingUtilitiesCachingBenchmark.cs	Adds cache-hit vs cache-miss escaping benchmarks with per-iteration input generation.
src/MSBuild.Benchmarks/Infrastructure/TemporaryDirectory.cs	Adds reusable temp-directory lifetime management for benchmarks.
src/MSBuild.Benchmarks/Infrastructure/BenchmarkProject.cs	Adds helper for creating project instances/items for expander benchmarks.
src/MSBuild.Benchmarks/Infrastructure/ExpanderBuilder.cs	Adds a builder for constructing expanders with explicit properties/items/metadata.
src/MSBuild.Benchmarks/Infrastructure/ExpanderBenchmarkFixture.cs	Adds a fixture to keep expander + project state alive for benchmark lifetime.
src/MSBuild.Benchmarks/ExpanderBenchmark.cs	Removes the previous monolithic expander benchmark (replaced by focused suites).
src/MSBuild.Benchmarks/Evaluation/Expansion/PropertyExpansionBenchmark.cs	Adds focused property-shape expansion benchmarks.
src/MSBuild.Benchmarks/Evaluation/Expansion/PropertyFunctionExpansionBenchmark.cs	Adds focused property-function benchmarks.
src/MSBuild.Benchmarks/Evaluation/Expansion/PropertyExpansionScalingBenchmark.cs	Adds scaling benchmarks for property reference counts.
src/MSBuild.Benchmarks/Evaluation/Expansion/PropertyBagCardinalityBenchmark.cs	Adds scaling benchmarks for unrelated property-bag cardinality.
src/MSBuild.Benchmarks/Evaluation/Expansion/ItemExpansionBenchmark.cs	Adds focused item-vector benchmarks (incl. quoted transforms and separators).
src/MSBuild.Benchmarks/Evaluation/Expansion/ItemFunctionExpansionBenchmark.cs	Adds focused item-function/transform benchmarks.
src/MSBuild.Benchmarks/Evaluation/Expansion/MetadataExpansionBenchmark.cs	Adds focused metadata expansion benchmarks.
src/MSBuild.Benchmarks/Evaluation/Expansion/MetadataExpansionScalingBenchmark.cs	Adds scaling benchmarks for metadata reference counts.
src/MSBuild.Benchmarks/Evaluation/Expansion/MixedExpansionBenchmark.cs	Adds focused mixed property/item/metadata expansion benchmarks.
src/MSBuild.Benchmarks/Evaluation/ExpressionShredder/ExpressionShredderBenchmark.cs	Adds focused ExpressionShredder throughput benchmarks.
src/MSBuild.Benchmarks/Evaluation/ExpressionShredder/ExpressionShredderAllocationBenchmark.cs	Adds opt-in cold-cache allocation diagnostics for ExpressionShredder.
src/MSBuild.Benchmarks/Evaluation/Conditions/ConditionStrings.cs	Adds shared condition strings for parsing/evaluation benchmarks.
src/MSBuild.Benchmarks/Evaluation/Conditions/ConditionParsingBenchmark.cs	Adds focused condition parsing benchmarks.
src/MSBuild.Benchmarks/Evaluation/Conditions/ConditionEvaluationBenchmark.cs	Adds focused end-to-end condition evaluation benchmarks with warmed production cache.
src/MSBuild.Benchmarks/Evaluation/Items/LazyItemEvaluationBenchmark.cs	Adds end-to-end evaluation benchmarks for lazy item operations.
src/Framework/ItemSpecModifiers.cs	Adds an overload using `ItemSpecModifierKind` to avoid repeated name classification and enable more direct dispatch.
src/Build/Evaluation/ExpressionShredder.cs	Adds specialized marker indexing APIs and a new indexed item-vector parser.
src/Build/Evaluation/Expander.cs	Updates item-vector detection API usage and simplifies parenthesis scanning signature usage in function parsing.
src/Build/Evaluation/Expander.Function.cs	Updates closing-paren scanning call site after signature change.
src/Build/Evaluation/Expander.PropertyExpander.cs	Refactors property expansion into a `readonly ref struct` and uses new marker indexing + specialized paren scan.
src/Build/Evaluation/Expander.MetadataExpander.cs	Uses new marker detection + indexed item-vector scanning to expand metadata in gaps.
src/Build/Evaluation/Expander.ItemExpander.cs	Introduces `TryExpandSingleItemVectorExpression`, refactors item-vector expansion/joining, and optimizes quoted-transform dispatch.
src/Build/Evaluation/Expander.ItemExpander.Transforms.MetadataMatch.cs	Caches derivable modifier classification/kind per metadata match.
src/Build/Evaluation/Expander.ItemExpander.Transforms.cs	Special-cases quoted transform shapes, reduces allocation, and avoids unnecessary context resolution for modifiers.
src/Build/Evaluation/LazyItemEvaluator.cs	Switches to the new `TryExpandSingleItemVectorExpression` API for item reference detection.
src/Build/Evaluation/ItemSpec.cs	Switches item-spec parsing to the new `TryExpandSingleItemVectorExpression` API.
src/Build/Evaluation/Evaluator.cs	Uses new property-marker detection API for SDK reference property expansion gating.
src/Build/Definition/Project.cs	Uses new property-marker detection API for provenance checks.
src/Build/BackEnd/Components/RequestBuilder/IntrinsicTasks/ItemGroupIntrinsicTask.cs	Uses new item-vector marker detection API for metadata self-reference logging behavior.
src/Build.UnitTests/Evaluation/Expander_Tests.cs	Adds coverage to ensure empty item vectors still report whether a transform was present.

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

Cover every derivable item-spec modifier through quoted transform expansion. Verify each modifier reads only the project and defining project context it requires.

DustinCampbell · 2026-08-13T00:56:45Z

One concern: I could not find targeted regression coverage for this new dispatch in GetMetadataValueFromMatch, especially for quoted transform expansion of %(Filename), %(RelativeDir), and the %(DefiningProject*) modifiers. Existing tests cover ItemSpecModifiers.GetItemSpecModifier directly and some transform behavior, but not this exact expander-layer routing. A focused table-driven test here would better lock in the intended mapping and guard future additions to derivable modifiers.

Added coverage with d7eaa2f.

rainersigwald · 2026-08-13T20:12:56Z

Nice! lgtm

OvesN · 2026-08-18T13:10:05Z

https://dev.azure.com/devdiv/DevDiv/_git/VS/pullrequest/770759
experimental insertion seems fine

DustinCampbell · 2026-08-18T16:18:05Z

https://dev.azure.com/devdiv/DevDiv/_git/VS/pullrequest/770759 experimental insertion seems fine

It looks like there are wins related to this PR. Speedometer shows four highlights in CPlusPlus scenarios and it looks like all of them are attributable to vcxprojreader.

OvesN · 2026-08-19T11:16:54Z

+        /// <returns>
+        ///  The zero-based index of the marker, or <c>-1</c> if it is not found.
+        /// </returns>
+        public static int IndexOfPropertyMarker(string expression)


nit: Since we already have named constants for the full markers, should we also introduce PropertyMarkerPrefix , ItemVectorMarkerPrefix , and MetadataMarkerPrefix for the repeated $ , @ , and % ?

OvesN · 2026-08-19T11:21:47Z


            // PERF NOTE: pre-scanning the string for "%(" is cheaper than a full scan.
-            if (expression.IndexOf("%(", StringComparison.Ordinal) < 0)
+            if (!ExpressionShredder.ContainsMetadataMarker(expression))


Could we avoid scanning for the first metadata marker twice?ExpandMetadataLeaveEscaped first calls ContainsMetadataMarker, and ScanAndExpandMetadata later starts again with IndexOfMetadataMarker. Perhaps we could find the first index up front and pass it into the expansion path?

OvesN · 2026-08-19T11:37:13Z

    /// The CultureInfo from the invariant culture. Used to avoid allocations for
    /// performing IndexOf etc.
    /// </summary>
    private static readonly CompareInfo s_invariantCompareInfo = CultureInfo.InvariantCulture.CompareInfo;


I guess we do not need it now at all. Only remaining use of it is
if (s_invariantCompareInfo.IndexOf(_expression, "::", CompareOptions.OrdinalIgnoreCase) > -1)

DustinCampbell added 17 commits August 10, 2026 10:30

Clarify benchmark setup and cache behavior

bd61887

Document the intentional cold-path and cache-isolation behavior of the focused evaluation benchmarks. Use the shared temporary-directory helper for modifier benchmarks to centralize setup and cleanup.

Simplify item vector expression parsing

134f996

Replace the stateful item expression enumerator with an indexed TryGetNextItemVectorExpression API. Update expansion callers, tests, and benchmarks to consume captures directly and avoid redundant marker scans.

Avoid nullable ItemExpressionCaptures

9bf312a

Use the Try pattern with an out parameter when expanding a single item vector expression. This avoids wrapping the relatively large capture struct in Nullable<T> and copying it through Nullable<T>.Value.

Refactor property expansion as a ref struct

5701db8

Encapsulate property expansion context in a readonly ref struct so helper methods no longer pass the same state through each call. Cache the truncation policy for reuse during expansion while preserving the existing static entry points.

Avoid unused item modifier context lookups

34c6229

Resolve project directory and defining-project metadata only for item-spec modifiers that consume those values. Common quoted transforms such as Filename now skip both per-item lookups.

Copilot AI lite review requested due to automatic review settings August 12, 2026 22:36

DustinCampbell temporarily deployed to copilot-pat-pool August 12, 2026 22:36 — with GitHub Actions Inactive

Copilot started reviewing on behalf of DustinCampbell August 12, 2026 22:36 View session

DustinCampbell temporarily deployed to copilot-pat-pool August 12, 2026 22:36 — with GitHub Actions Inactive

DustinCampbell temporarily deployed to copilot-pat-pool August 12, 2026 22:42 — with GitHub Actions Inactive

DustinCampbell temporarily deployed to copilot-pat-pool August 12, 2026 22:44 — with GitHub Actions Inactive

github-actions Bot reviewed Aug 12, 2026

View reviewed changes

DustinCampbell temporarily deployed to copilot-pat-pool August 12, 2026 22:45 — with GitHub Actions Inactive

Copilot AI reviewed Aug 12, 2026

View reviewed changes

Comment thread src/MSBuild.Benchmarks/EscapingUtilitiesCachingBenchmark.cs

Test quoted transform modifier context

d7eaa2f

Cover every derivable item-spec modifier through quoted transform expansion. Verify each modifier reads only the project and defining project context it requires.

OvesN self-assigned this Aug 14, 2026

OvesN added the wait-for-exp-insertion label Aug 14, 2026

dotnet deleted a comment from azure-pipelines Bot Aug 14, 2026

Merge branch 'main' into expansion-perf

01deeb5

OvesN removed the wait-for-exp-insertion label Aug 18, 2026

Merge branch 'main' into expansion-perf

18f09b2

OvesN reviewed Aug 19, 2026

View reviewed changes

Navigation Menu

Uh oh!

Improve MSBuild expansion performance - #14697

Improve MSBuild expansion performance#14697
DustinCampbell wants to merge 20 commits into
dotnet:mainfrom
DustinCampbell:expansion-perf

DustinCampbell commented Aug 12, 2026

Uh oh!

github-actions Bot left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

DustinCampbell commented Aug 13, 2026 •

edited

Loading

Uh oh!

rainersigwald commented Aug 13, 2026

Uh oh!

OvesN commented Aug 18, 2026

Uh oh!

DustinCampbell commented Aug 18, 2026

Uh oh!

OvesN Aug 19, 2026

Uh oh!

OvesN Aug 19, 2026

Uh oh!

OvesN Aug 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

DustinCampbell commented Aug 12, 2026

Context

Changes Made

Benchmark Results

🚀 Performance Highlights

ItemExpansionBenchmark

MetadataExpansionBenchmark

MixedExpansionBenchmark

PropertyExpansionBenchmark

ItemExpansionBenchmark

MetadataExpansionBenchmark

MixedExpansionBenchmark

PropertyExpansionBenchmark

Analysis

Testing

Notes

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

DustinCampbell commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rainersigwald commented Aug 13, 2026

Uh oh!

OvesN commented Aug 18, 2026

Uh oh!

DustinCampbell commented Aug 18, 2026

Uh oh!

OvesN Aug 19, 2026

Choose a reason for hiding this comment

Uh oh!

OvesN Aug 19, 2026

Choose a reason for hiding this comment

Uh oh!

OvesN Aug 19, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

DustinCampbell commented Aug 13, 2026 •

edited

Loading