Add net9.0 TFM with span-based Task.WhenAll + BenchmarkDotNet harness by buvinghausen · Pull Request #42 · buvinghausen/TaskTupleAwaiter

buvinghausen · 2026-05-13T22:39:31Z

Summary

Adds net10.0 to the library's TargetFrameworks, giving .NET 10+ consumers a build where Task.WhenAll(...) calls inside the generated awaiters bind to Task.WhenAll(ReadOnlySpan<Task>) and stack-allocate the buffer, eliminating the per-await Task[] heap
allocation that the params overload requires.
Generator change: Items helper now wraps its output as a collection expression so call sites emit Task.WhenAll([tasks.Item1, ..., tasks.ItemN]). On netstandard2.0/net462/net8.0 this still lowers to new Task[]{...} and binds to Task.WhenAll(params Task[]) — identical IL to before. On net10.0 it binds to the span overload.
Adds test/TaskTupleAwaiter.Benchmarks BenchmarkDotNet harness covering arities 2/4/8/16 across typed/non-generic tuples and pre-completed/async completion modes, with ConfigureAwait(bool) and ConfigureAwait(ConfigureAwaitOptions) variants. [MemoryDiagnoser]
on every class so allocation deltas are visible.
AOT smoke test adds net10.0 to its TFM list so the new code path is exercised under NativeAOT publish in CI.
test/Directory.Build.props scoped to TaskTupleAwaiter.Tests only — the multi-TFM (net11.0;...;net472) and xUnit/Shouldly inheritance previously leaked into BenchmarkDotNet's auto-generated wrapper csproj (which lives under bin/) and broke restore with
NU1201. Other test-tree projects now provide their own TFMs explicitly.
Library jumps from 3 lib folders to 4 in the NuGet package: lib/netstandard2.0, lib/net462, lib/net8.0, lib/net10.0. .NET 9 consumers continue to receive the lib/net8.0 asset.

Empirical finding (full numbers in `docs/superpowers/specs/2026-05-13-net10-span-whenall-benchmark-results.md`)

The C# 13+ compiler's overload-preference rule already binds positional Task.WhenAll(t1, t2, ..., tN) to Task.WhenAll(ReadOnlySpan<Task>) on net10.0 whenever both overloads are visible. That means adding the net10.0 TFM alone delivered the allocation
reduction — the generator's collection-expression syntax leaves IL on net10.0 byte-for-byte unchanged. The change is preserved as belt-and-suspenders insurance against future overload-resolution behavior.

Per-op Allocated results (TypedTuple Arity2_PreCompleted through Arity16_PreCompleted):

TFM	A2	A4	A8	A16
net8.0 (baseline & after)	120 B	136 B	168 B	808 B
net10.0 (baseline & after)	72 B	72 B	72 B	648 B

Test plan

CI green on all build matrix entries.
AOT smoke test publishes successfully on net8.0, net10.0, and net11.0.
All 2431 tests pass on net472, net8.0, net9.0, net10.0, net11.0.
Sanity check the produced .nupkg lists four lib/ folders.

Notes

Spec: docs/superpowers/specs/2026-05-13-net10-span-whenall-design.md
Plan: docs/superpowers/plans/2026-05-13-net10-span-whenall.md
Benchmark numbers (before + after): docs/superpowers/specs/2026-05-13-net10-span-whenall-benchmark-results.md

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Extends TaskTupleAwaiter library to target net10.0 in addition to existing TFMs. All 2431 tests pass on net472, net8.0, net9.0, net10.0, and net11.0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The xUnit/Shouldly imports and multi-TFM (net11.0;...;net472) list are only meaningful for the unit-test project. The AOT smoke test and the benchmarks project both override TFMs in their own csprojs and don't need the inherited xUnit packages — the Remove items they used to need become no-ops, but the bigger problem is BenchmarkDotNet's internal wrapper csproj. BDN generates that wrapper at runtime under bin/, and MSBuild's Directory.Build.props auto-import walks up to find this file; the wrapper inherits TargetFrameworks=...;net472 and then fails NU1201 trying to consume the benchmark project (which targets net8.0;net10.0). Scoping the multi-TFM PropertyGroup and the xUnit ItemGroup to MSBuildProjectName == TaskTupleAwaiter.Tests preserves existing test behavior and unblocks BDN runs without per-project workarounds. Also gitignore BenchmarkDotNet.Artifacts/ so reports don't pollute git status during benchmark runs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds TypedTupleAwaitBenchmarks with 8 methods covering arities 2/4/8/16 under pre-completed and async (Task.Yield) completion modes. The class is decorated with [MemoryDiagnoser] so BDN reports per-op allocations. Also suppress CA1707 and CA1822 in the benchmarks csproj — BenchmarkDotNet requires underscore-separated method names and instance benchmark methods, both of which the repo's analyzers (latest-Recommended + TreatWarningsAsErrors) would otherwise turn into build errors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Mirrors TypedTupleAwaitBenchmarks for awaitable tuples of plain Task (no result type). The first element of each tuple is cast to (Task) to force binding to the non-generic extension method overload group rather than any homogeneous typed inference path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Covers both the legacy bool overload and the .NET 8+ ConfigureAwaitOptions overload, in both the typed and non-generic tuple flavors, at arities 4 and 16. These measure the configured-await code path which builds a separate TupleConfiguredTaskAwaitable struct under the hood. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Empirically, the net10.0 baseline already shows substantially lower per-op allocations than net8.0 for low-arity pre-completed tuple awaits (120 B -> 72 B at arity 2 typed, 808 B -> 648 B at arity 16 typed). The C# 13+ compiler prefers Task.WhenAll(ReadOnlySpan<Task>) over the params Task[] overload for positional Task.WhenAll(t1, ..., tN) call sites whenever both are visible. Since the ReadOnlySpan overload only exists on net9+, the net8.0 library DLL still binds to params Task[] (heap allocation), while the net10.0 library DLL already binds to the span overload (stack allocation). Adding the net10.0 TFM in Task 1 was sufficient on its own to deliver the optimization — Task 9's collection-expression change is intent-clarifying rather than IL-changing. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Wraps the Items helper output with brackets so each generated call site becomes Task.WhenAll([tasks.Item1, ..., tasks.ItemN]) instead of Task.WhenAll(tasks.Item1, ..., tasks.ItemN). On netstandard2.0, net462, and net8.0 the compiler still binds to Task.WhenAll(params Task[]) and allocates the array — same IL as before. On net9+ the compiler binds to Task.WhenAll(ReadOnlySpan<Task>) and stack-allocates the buffer. Empirically this change is intent-clarifying rather than IL-changing on net10.0: the C# 13+ compiler's overload-preference rule already binds Task.WhenAll(t1, t2, ...) positional calls to the span overload when both are visible. The bracketed collection expression is the explicit form and remains stable across future C# overload-resolution changes. All 2431 tests still pass on all 5 test TFMs (net8.0, net9.0, net10.0, net11.0, net472). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Exercises the new generated collection-expression WhenAll code path under NativeAOT publish on net10.0, in addition to the existing net8.0 and net11.0 coverage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- CLAUDE.md: update library TFM list, repo layout, and the generator section to describe the collection-expression emission and per-TFM compiler binding behavior. - README.md: note the allocation-free WhenAll on .NET 10+ and add a net10 row to the Compatibility table. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Confirms the IL-level outcome predicted in the previous commit: - net10.0 after-change allocations match net10.0 baseline byte-for-byte across all 22 benchmarks (72 B / 648 B for pre-completed at the arities where the win applies). The collection-expression source has no IL impact on net10.0 — the compiler was already binding to the span overload via overload preference. - net8.0 after-change allocations match net8.0 baseline byte-for-byte (120 / 136 / 168 / 808 etc.). The collection-expression lowers to new Task[]{...} on net8.0 — identical IL to the original positional params form. Net win attributable to this branch: adding the net10.0 TFM (Task 1) delivers the per-await Task[] allocation reduction for .NET 10 consumers. The generator collection-expression syntax (Task 9) is preserved as belt-and-suspenders insurance against future C# overload resolution changes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- CLAUDE.md: remove phantom test/TaskTupleAwaiter.Benchmarks/Directory.Build.props entry from the repo layout. The isolation is implemented via the MSBuildProjectName condition in test/Directory.Build.props, not via a local props file. - Generator helper comment: replace "net9+" with "net10.0+" and note explicitly that the library does not ship a net9.0 asset, so .NET 9 consumers pick the net8.0 asset and don't see the optimization. - .gitignore: also ignore *.nupkg / *.snupkg so local `dotnet pack` runs don't show as untracked. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

buvinghausen · 2026-05-13T22:45:32Z

@jnm2 headed out to dinner haven't taken the time to review or do anything no rush on this will leave it open feel free to post musings.

At least the BDN stuff is there let me know which direction you want to take it.

jnm2

Looks great to me. We could also consider using array pooling to achieve similar benefits on pre-net-10 targets.

buvinghausen · 2026-05-14T01:24:26Z

@jnm2 ok shuffled the projects around so that Claude's MS build nightmare is over. Feel free to merge whenever you want.

buvinghausen · 2026-05-14T01:36:45Z

@jnm2 ok the build is back2good (DC Metro joke from their 2017/18 marketing campaign) the floor is yours

The synthesized record-struct members (Equals, GetHashCode, ToString, PrintMembers, == / !=) are never used by the awaiter protocol. Dropping the `record` keyword shrinks TaskTupleAwaiter.dll by ~30% across every TFM (net10.0: 92,160 -> 64,512 bytes; netstandard2.0: 80,896 -> 53,248). All 2,431 tests pass on every test TFM; benchmark allocation and time are unchanged within run-to-run noise. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Annotate the small forwarder methods in the generated awaiter and extension types with [MethodImpl(MethodImplOptions.AggressiveInlining)]: the GetAwaiter / ConfigureAwait extension methods (typed + non-generic arity 1 + typed arity 2..16), the IsCompleted / OnCompleted / UnsafeOnCompleted accessors on TupleTaskAwaiter and on TupleConfiguredTaskAwaitable.Awaiter, and the TupleConfiguredTaskAwaitable .GetAwaiter() helper. The constructor and GetResult are deliberately left un-annotated because their bodies grow with arity and forcing inline there is more likely to bloat than help. The IsCompleted properties are emitted with an explicit get accessor block because [MethodImpl] only attaches to methods/constructors, not property declarations. Measured on net10.0 (Intel Core Ultra 9 185H, BDN 0.15.8): - TypedTuple.Arity8_PreCompleted: 93.21 ns -> 80.00 ns (-14%) - TypedTuple.Arity16_Async: 6224.68 ns -> 5857.43 ns (-6%) - TypedTuple.Arity4_PreCompleted: 49.07 ns -> 46.77 ns (~5%) - Other arities: flat within run-to-run noise. DLL size delta: 0 B across all four TFMs (the new attribute metadata fits within existing PE 4 KB alignment padding). All 2431 tests still pass on all 5 test TFMs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Replaces the original baseline-vs-after framing (which mixed machines and intermediate code states) with a clean net10.0 vs net8.0 comparison captured on the same source tree (HEAD 7c98fd2) on a single machine. The new doc presents: - Full BDN tables for each of the three benchmark classes on both TFMs. - Per-method allocation deltas (net10 vs net8) so the user-facing win is obvious at a glance. - A headline summary covering both allocation and mean-time wins. - DLL size table for the shipped library. Headline: typed pre-completed tuple awaits drop 40-57% allocation at arity 2/4/8 (-20% at arity 16). Non-generic arity 16 drops 69%. Mean times improve 13-34% on the pre-completed path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

buvinghausen · 2026-05-14T06:47:38Z

Ok now officially in your court my dude

jnm2 · 2026-05-14T15:59:05Z

@copilot Please go through all .md files and code comments in this PR and remove all mentions in comments of collection expressions or [ as being related to whether the span overload is picked. Compiling against net10.0 was the necessary and sufficient condition for getting the span overload, and this is what should be stated everywhere rather than talking about collection expressions or syntactic explicitness. [ ] is not an explicit "pick span" syntax at all.

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/91413727-6185-498c-944c-8f40a2068b81 Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Co-authored-by: Joseph Musser <me@jnm2.com>

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/845db3d9-76cf-4ccb-a3fe-c876b9422e2d Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/f44ff7f8-bc2d-4d0a-be59-6ae4a07e1bf2 Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Co-authored-by: Joseph Musser <me@jnm2.com>

jnm2 · 2026-05-14T16:45:45Z

@buvinghausen I updated the rationale and also retargeted the change to net9, since that's where span-based Task.WhenAll first appears.

buvinghausen and others added 15 commits May 13, 2026 16:40

Add implementation plan for net10.0 span-WhenAll optimization

28fe6da

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Add net10.0 TFM to library

40e0dd8

Extends TaskTupleAwaiter library to target net10.0 in addition to existing TFMs. All 2431 tests pass on net472, net8.0, net9.0, net10.0, and net11.0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Scaffold TaskTupleAwaiter.Benchmarks project

6295ba4

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Add BenchmarkSwitcher entry point

387e4cf

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Document how to run the benchmarks

708ac4e

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Add net10.0 to AOT smoke test TFMs

f7404c3

Exercises the new generated collection-expression WhenAll code path under NativeAOT publish on net10.0, in addition to the existing net8.0 and net11.0 coverage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

buvinghausen requested a review from jnm2 May 13, 2026 22:39

Merge branch 'master' into worktree-net10-span-whenall

3def907

jnm2 approved these changes May 14, 2026

View reviewed changes

Comment thread src/TaskTupleAwaiter.Generator/TaskTupleExtensionsGenerator.cs Outdated

Comment thread test/Directory.Build.props Outdated

Comment thread docs/superpowers/specs/2026-05-13-net10-span-whenall-benchmark-results.md Outdated

buvinghausen added 2 commits May 13, 2026 20:54

Fix claude puke in SLNX

b82de8a

End claude's msbuild nightmare

a9babb4

Make AOT smoke test more robust

8a50cdf

buvinghausen and others added 3 commits May 14, 2026 01:45

Copilot started work on behalf of jnm2 May 14, 2026 15:59 View session

Apply review wording suggestions in plan and benchmark docs

4961e1a

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/91413727-6185-498c-944c-8f40a2068b81 Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Copilot finished work on behalf of jnm2 May 14, 2026 16:13

Copilot AI requested a review from jnm2 May 14, 2026 16:13

Copilot started work on behalf of jnm2 May 14, 2026 16:15 View session

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread README.md Outdated

Apply suggestion from @jnm2

4d4ab54

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread README.md Outdated

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread README.md Outdated

Apply suggestions from code review

c8cfd60

Co-authored-by: Joseph Musser <me@jnm2.com>

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread src/TaskTupleAwaiter.Generator/TaskTupleExtensionsGenerator.cs

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread docs/superpowers/specs/2026-05-13-net10-span-whenall-design.md

jnm2 and others added 2 commits May 14, 2026 12:27

Apply suggestion from @jnm2

21a63e8

Retarget PR content from net10 to net9

5a6a164

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/845db3d9-76cf-4ccb-a3fe-c876b9422e2d Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Copilot finished work on behalf of jnm2 May 14, 2026 16:30

Copilot AI requested a review from jnm2 May 14, 2026 16:30

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread README.md Outdated

Copilot started work on behalf of jnm2 May 14, 2026 16:32 View session

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread docs/superpowers/specs/2026-05-13-net9-span-whenall-design.md Outdated

jnm2 and others added 2 commits May 14, 2026 12:33

Apply suggestion from @jnm2

6319c5f

Fix unresolved README merge markers

edeca49

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/f44ff7f8-bc2d-4d0a-be59-6ae4a07e1bf2 Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

jnm2 changed the title ~~Add net10.0 TFM with span-based Task.WhenAll + BenchmarkDotNet harness~~ Add net9.0 TFM with span-based Task.WhenAll + BenchmarkDotNet harness May 14, 2026

Clean up README conflict artifact and spec wording

b692d1b

Agent-Logs-Url: https://github.com/buvinghausen/TaskTupleAwaiter/sessions/f44ff7f8-bc2d-4d0a-be59-6ae4a07e1bf2 Co-authored-by: jnm2 <8040367+jnm2@users.noreply.github.com>

Copilot finished work on behalf of jnm2 May 14, 2026 16:36

Copilot AI requested a review from jnm2 May 14, 2026 16:36

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread README.md Outdated

jnm2 reviewed May 14, 2026

View reviewed changes

Comment thread tests/unit/Directory.Build.props Outdated

Apply suggestions from code review

826e42c

Co-authored-by: Joseph Musser <me@jnm2.com>

buvinghausen merged commit 712903a into master May 14, 2026
1 check passed

buvinghausen deleted the worktree-net10-span-whenall branch May 14, 2026 16:53

Conversation

buvinghausen commented May 13, 2026

Summary

Empirical finding (full numbers in docs/superpowers/specs/2026-05-13-net10-span-whenall-benchmark-results.md)

Test plan

Notes

Uh oh!

buvinghausen commented May 13, 2026

Uh oh!

jnm2 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

buvinghausen commented May 14, 2026

Uh oh!

buvinghausen commented May 14, 2026

Uh oh!

buvinghausen commented May 14, 2026

Uh oh!

jnm2 commented May 14, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jnm2 commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Empirical finding (full numbers in `docs/superpowers/specs/2026-05-13-net10-span-whenall-benchmark-results.md`)