mirror of
https://github.com/OPCFoundation/UA-.NETStandard.git
synced 2026-10-16 21:02:51 +08:00
Marc Schier
76d33cb139
# Description The two OPC UA source generators were built on `SourceGenerator.Foundations` (SGF), whose only job was getting the generator's dependency closure loaded inside the compiler. It does so with a **process-wide, first-generator-wins `AppDomain.AssemblyResolve` hook backed by a non-concurrent `Dictionary` mutated from `AssemblyLoad` callbacks**. Concurrent compilations in a shared `VBCSCompiler` race it, which is the root cause of the intermittent clean-build failure: ``` CSC : error CS8784: Generator 'StackSourceGeneratorHoist' failed to initialize. ... Exception was of type 'FileNotFoundException' with message 'Could not load file or assembly 'SourceGenerator.Foundations.Contracts, Version=2.0.14.0, ...' ``` The repository already solves that problem deterministically — #4041 made the NuGet packages ship the full closure under `analyzers/dotnet/.../cs`, and `Directory.Build.targets` registers the same closure as `Analyzer` items for in-repo `ProjectReference` consumers. SGF was therefore pure liability and is removed. ## What changed **1. SourceGenerator.Foundations removed.** `ModelSourceGenerator` and `StackSourceGenerator` are plain `[Generator(LanguageNames.CSharp)] IIncrementalGenerator` implementations again. SGF's implicit crash handling is replaced by an explicit `SourceGenerator.Guard` that reports `MODELGEN003` / `STACKGEN003` and rethrows cancellation. Its logger sink was already dead code and is dropped. `Opc.Ua.SourceGeneration.Stack.dll` shrank from **18 MB** (embedded closure) to **29 KB**. **2. Analyzers now ship under Roslyn-versioned folders.** `analyzers/dotnet/roslyn4.14/cs` and `analyzers/dotnet/roslyn5.0/cs` instead of `analyzers/dotnet/cs`, driven by `roslyn.props` at the repository root, with new `*.Pack` projects owning the package identity. The versioned folder matters: a host older than the band **skips** the analyzer instead of loading it and failing inside generator initialization. **3. Dependencies flow automatically.** Both delivery paths discover the payload by globbing the generator's output directory, so adding a package to a generator's dependency tree requires **no** `.csproj`, `.targets`, `.nuspec` or validation-script edit. Verified by temporarily adding a `PackageReference` and watching it appear in the payload with no other file touched. ## Two Roslyn bands: VS2022 and VS2026 Analyzers ship twice, once per band. The .NET SDK loads the highest band its compiler supports and **silently skips** anything above it, so a single-band package leaves every consumer below that band with no diagnostics and no generated code. | Band | Built against | Loaded by | | --- | --- | --- | | `roslyn4.14` | `Microsoft.CodeAnalysis.CSharp 4.14.0` | Visual Studio 2022 17.14+ / .NET 9 SDK | | `roslyn5.0` | `Microsoft.CodeAnalysis.CSharp 5.0.0` | Visual Studio 2026 18.0+ / .NET 10 SDK | Getting this wrong is invisible, which is most of why it took three attempts. The rule: > An analyzer must reference the **lowest** `System.Collections.Immutable` / `System.Reflection.Metadata` across all hosts it claims to support, and must never ship them. .NET binds a reference *upward* but never downward. Every failure mode is a compiler **warning**: | Diagnostic | Cause | | --- | --- | | `CS9057` | Built against a newer Roslyn than the host, so it is skipped entirely | | `CS8784` | Loaded, then `MissingMethodException` - a shipped S.C.I bound a second `ImmutableArray<T>` | | `CS8032` | Analyzer instance could not be created | This PR first shipped a `roslyn4.8` band that **never loaded** - confirmed against a real Roslyn 4.8 compiler: ``` warning CS8784: Generator 'ModelSourceGenerator' failed to initialize ... MissingMethodException: Method not found: 'IncrementalValueProvider`1<ImmutableArray`1<!!0>> IncrementalValueProviderExtensions.Collect(IncrementalValuesProvider`1<!!0>)' ``` That band was removed, and **this is not a regression**: master had the same defect, pinned to Roslyn 4.14 while referencing `System.Collections.Immutable 10.0.0.0`, so no 4.x host ever worked there either. The band that landed instead is `4.14`, and it is nearly free: `Microsoft.CodeAnalysis 4.14` depends on the **same S.C.I 9.0.0 as 5.0.0** and its `csc` runs on `net9.0`. Both bands therefore share `$(RoslynRuntimeVersion)`, so `Opc.Ua.Types` and `Opc.Ua.SourceGeneration.Core` are compiled **once** and serve both; only the five Roslyn-referencing assemblies get a second build, each from a 12-line project that glob-links the primary sources and overrides `RoslynApiVersion`. Going lower still - 4.8 wants S.C.I 7.x - would mean rebuilding that whole closure, `Opc.Ua.Types` included, and remains out of scope. (`System.Collections.Immutable 7.0.0` also does not resolve in this repository: `NU1603`, 8.0.0 substituted.) **Nothing else in the repository exercises the down-level payload**, and since every failure is a warning, a broken band would ship silently - exactly as the 4.8 one did. So `validate-source-generator-packages.ps1` now restores a real `Microsoft.Net.Compilers.Toolset 4.14.0`, runs its `csc.dll` over the packed `roslyn4.14` payload, and asserts **both** that none of `CS9057`/`CS8784`/`CS8032`/`CS8034` is emitted **and** that `/reportanalyzer` lists the generator as having executed - silence alone would also pass if the generator were never handed to the compiler. The migration analyzer, whose payload comes from a hand-written nuspec rather than `SourceGeneratorPack.targets`, gets the same check plus a band/assembly manifest assertion. To stop the trap recurring, `Microsoft.CodeAnalysis*`, `System.Collections.Immutable` and `System.Reflection.Metadata` are excluded from both the package payload and the in-repo analyzer closure, and asserted against in the validation script. `roslyn.props` and `docs/DeveloperGuide.md` document what adding a band requires. ## The shipped package set is unchanged Evaluating `PackageId`/`IsPackable` across every project in `preview-pack.slnx` gives the **same 47 package ids** on `master` and on this branch, and a full Release build plus `dotnet pack --no-build` produces exactly those 47. Only the owning project changed: | Package | master | this PR | | --- | --- | --- | | `…Opc.Ua.SourceGeneration` | `Opc.Ua.SourceGeneration.csproj` | `Opc.Ua.SourceGeneration.Pack.csproj` | | `…Opc.Ua.SourceGeneration.Stack` | `Opc.Ua.SourceGeneration.Stack.csproj` | `Opc.Ua.SourceGeneration.Stack.Pack.csproj` | `.azurepipelines/expected-packages.txt` pins the expected set and the validation script fails when the packed output drifts. ## Two new package checks **`build/<PackageId>.props` naming.** NuGet only auto-imports that exact name, and the file declares every `CompilerVisibleProperty` / `CompilerVisibleItemMetadata` the model generator relies on. It was previously packed as `build/OPCFoundation.Opc.Ua.SourceGeneration.props`, which does **not** match the package id — so **no NuGet consumer ever imported it**, and every `ModelSourceGenerator*` property and `AdditionalFiles` metadata was silently invisible. In-repo projects were unaffected because they import it by relative path. **An end-to-end consumer that actually drives the generator.** The pre-existing clean-consumer check only proved the analyzer *loads*. `Test-SourceGeneratingConsumer` builds a standalone project that feeds a real NodeSet2 in as an `AdditionalFile`, pins a `ModelSourceGeneratorPrefix` that cannot be derived from the model, and references the emitted types from hand-written code. It deliberately does **not** import the props by path, so the package's auto-import is what is under test. Both checks were verified to fail against a package rebuilt with the misnamed props. ## CI: parallelism re-enabled `buildandtest.yml` pinned three `dotnet build` invocations to `/p:UseSharedCompilation=false` and `-maxcpucount:1`. `dotnet build` defaults to all logical processors, so `-maxcpucount:1` actively disabled parallelism. Both flags were introduced silently in unrelated PRs (`f358078b1`, `208c601b0`) and suppress exactly the conditions the SGF race needed — now structurally gone. Compilation is ~72 % of what that workflow executes (233 of 323 job-minutes in run `29916768976`), so the queue-independent `Build` step durations are the baseline: | Measurement | Baseline | | --- | --- | | `Build` step, 26 `test-*` jobs | mean 5.53 min, sum 143.9 min | | `build-linux-all-tfm` | 56.3 min | | `build-windows-all-tfm` | 32.5 min | `codeql-analysis.yml` keeps the flag — the CodeQL tracer requires it. ## Incidental fixes - `Microsoft.CodeAnalysis.CSharp` declares an exact `[version]` range on `Microsoft.CodeAnalysis.Common`, so with `CentralPackageTransitivePinningEnabled` the central pins must move in lockstep or restore fails with `NU1109`. They now derive from `$(RoslynApiVersionVS2026)`. - `MigrationGeneratorTests` created its generator driver without parse options — silently fine under Roslyn 4.14, `Inconsistent language versions` under 5.0. - The migration analyzer props injected `<Analyzer>` items by hand from `analyzers/dotnet/cs`; NuGet registers analyzers itself and selects the matching folder, so the injection was obsolete and would now force the wrong assembly to load. - Exception diagnostics reported `ex.StackTrace` (null for an exception that never left its throw site, and carrying neither type nor inner chain) — now `ex.ToString()` at all three call sites. - The pack guard tested the flattened assembly list, so it only fired when *no* variant produced output; it now checks each declared variant contributed assemblies. - `GeneratePackageOnBuild` is disabled for the generators; packing happens in CI or on an explicit `dotnet pack`. ## Note for reviewers **Building this repository requires a Roslyn 5.x host** (the .NET 10 SDK or Visual Studio 2026). That is already the documented prerequisite and every CI leg installs `10.0.x` last, but it is now a hard requirement. ## Related Issues _No tracking issue: this is a follow-up to the CS8784 clean-build failure reported against #4041, which fixed the packaged delivery path but left the in-repo project-reference path relying on the SGF resolver._ - Follow-up to #4041 ## Checklist _Put an `x` in the boxes that apply. You can complete these step by step after opening the PR._ - [x] I have signed the [CLA](https://opcfoundation.org/license/cla/ContributorLicenseAgreementv1.0.pdf) and read the [CONTRIBUTING](https://github.com/OPCFoundation/UA-.NETStandard/blob/master/CONTRIBUTING.md) doc. - [x] I have added tests that prove my fix is effective or that my feature works and increased code coverage. - [x] I have added all necessary documentation. - [x] I have verified that my changes do not introduce (new) build or analyzer warnings. - [x] I ran **all** tests locally using the **UA.slnx** solution against at least .net **framework** and .net **10**, and all passed. - [ ] I fixed **all** failing and flaky tests in the CI pipelines and **all** CodeQL warnings. - [ ] I have addressed **all** PR feedback received. ### Verification detail | Check | Result | | --- | --- | | `dotnet build UA.slnx` (clean, and after merging master) | 0 errors; 18 warnings, all pre-existing `CA1873` in `Opc.Ua.Client` | | `Opc.Ua.SourceGeneration.Tests` | 81/81 on net10.0 and net48 | | `Opc.Ua.SourceGeneration.Stack.Tests` | 92/92 on net10.0, 89/89 on net48 | | `Opc.Ua.SourceGeneration.Core.Tests` | 3722 passed, 8 skipped | | `Opc.Ua.MigrationAnalyzer.Tests` | 121/121 on net10.0 and net48 | | Package id set, master vs branch | identical, 47 packages | | Release build + `dotnet pack --no-build` | 47 packages, matching the manifest; both `roslyn4.14` and `roslyn5.0` present; no host-provided assemblies shipped | | `validate-source-generator-packages.ps1` | passes, including manifest, props-naming, end-to-end generator consumer and down-level host checks | | Down-level band under real csc 4.14.0 | all three analyzer packages load **and run** | | Negative tests | misnamed props → both checks fail as designed; hidden variant output → pack guard fires; 5.0 build placed in `roslyn4.14` → real `CS9057`; non-running generator name → execution assertion fires; expecting a `roslyn4.8` band → manifest check fires | --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 9ed53262-5cfa-429a-820e-bf6146a00d8b
58 lines
3.5 KiB
XML
58 lines
3.5 KiB
XML
<Project>
|
|
<!--
|
|
The Roslyn API version the OPC UA analyzers and source generators are built
|
|
against, and the runtime assemblies that version implies.
|
|
|
|
Lives at the repository root, not under tools/, because Directory.Packages.props
|
|
imports it: a partial-copy context that only takes the root files (the sample
|
|
Dockerfiles do exactly that) must still be able to restore.
|
|
|
|
Each analyzer ships under `analyzers/dotnet/<folder>/cs`. The .NET SDK loads the
|
|
highest folder its compiler supports and silently ignores folders above it. Using a
|
|
versioned folder rather than a bare `analyzers/dotnet/cs` is deliberate: a host older
|
|
than the band below skips the analyzer instead of loading it and then failing at
|
|
generator-initialization time.
|
|
|
|
See https://learn.microsoft.com/visualstudio/extensibility/roslyn-version-support
|
|
for the Roslyn package version to Visual Studio mapping.
|
|
|
|
RoslynRuntimeVersion is the critical part. The compiler supplies
|
|
System.Collections.Immutable and System.Reflection.Metadata itself, and the version it
|
|
supplies is fixed by its Roslyn band (4.8 -> 7.x, 4.14 -> 9.x, 5.0 -> 9.x). An analyzer
|
|
must therefore reference a version at or BELOW its band's - .NET satisfies a reference
|
|
from a higher assembly version but never from a lower one - and must never ship a copy
|
|
of its own, because a second copy binds a separate ImmutableArray<T> identity and the
|
|
generator dies with MissingMethodException on its first Roslyn call. Either failure
|
|
surfaces only as warning CS8784, so the consumer silently gets no generated code.
|
|
|
|
Directory.Packages.props pins both packages to $(RoslynRuntimeVersion) and central
|
|
transitive pinning propagates that through the whole analyzer closure - the generator,
|
|
Opc.Ua.SourceGeneration.Core and Opc.Ua.Types. Keeping the pin here rather than as a
|
|
loose version literal is what stops the closure drifting above the band again.
|
|
|
|
Adding a band below the current floor is therefore not just another entry in this
|
|
file. It only stays cheap while the new band shares $(RoslynRuntimeVersion): Roslyn
|
|
4.14 and 5.0 both depend on System.Collections.Immutable 9.0.0, so one build of the
|
|
non-Roslyn part of the closure serves both. Going lower - Roslyn 4.8 wants 7.x -
|
|
would require building that whole closure, Opc.Ua.Types included, a second time
|
|
against the older runtime.
|
|
-->
|
|
<PropertyGroup>
|
|
<!-- Visual Studio 2026 18.0+ / .NET 10 SDK. -->
|
|
<RoslynApiVersionVS2026>5.0.0</RoslynApiVersionVS2026>
|
|
<!-- Visual Studio 2022 17.14+ / .NET 9 SDK. -->
|
|
<RoslynApiVersionVS2022>4.14.0</RoslynApiVersionVS2022>
|
|
<!-- System.Collections.Immutable / System.Reflection.Metadata that both
|
|
Microsoft.CodeAnalysis $(RoslynApiVersionVS2022) and $(RoslynApiVersionVS2026)
|
|
depend on. The two bands share it, so a single build of the non-Roslyn part of
|
|
the closure (Opc.Ua.SourceGeneration.Core, Opc.Ua.Types) serves both. Roslyn
|
|
4.14's csc runs on net9.0 and takes these from the shared framework, which is
|
|
9.0.0.0 - matching this pin exactly. -->
|
|
<RoslynRuntimeVersion>9.0.0</RoslynRuntimeVersion>
|
|
</PropertyGroup>
|
|
<PropertyGroup>
|
|
<RoslynAnalyzerFolderVS2026>roslyn$(RoslynApiVersionVS2026.Substring(0, $(RoslynApiVersionVS2026.LastIndexOf('.'))))</RoslynAnalyzerFolderVS2026>
|
|
<RoslynAnalyzerFolderVS2022>roslyn$(RoslynApiVersionVS2022.Substring(0, $(RoslynApiVersionVS2022.LastIndexOf('.'))))</RoslynAnalyzerFolderVS2022>
|
|
</PropertyGroup>
|
|
</Project>
|