Skip to content

Avoid eliding pointer-to-ref conversions#79311

Merged
jjonescz merged 6 commits intodotnet:mainfrom
jjonescz:79051-retrack
Jul 17, 2025
Merged

Avoid eliding pointer-to-ref conversions#79311
jjonescz merged 6 commits intodotnet:mainfrom
jjonescz:79051-retrack

Conversation

@jjonescz
Copy link
Member

@jjonescz jjonescz commented Jul 9, 2025

Fixes #79051.

@jjonescz jjonescz changed the title Avoid eliding pointer to ref conversions Avoid eliding pointer-to-ref conversions Jul 9, 2025
@jjonescz jjonescz marked this pull request as ready for review July 9, 2025 15:58
@jjonescz jjonescz requested a review from a team as a code owner July 9, 2025 15:59
@jjonescz jjonescz requested a review from jaredpar July 9, 2025 15:59

// If this is a pointer-to-ref assignment, keep the local so GC knows to re-track it.
else if (localSymbol.RefKind != RefKind.None &&
localSymbol.SynthesizedKind == SynthesizedLocalKind.UserDefined &&
Copy link
Contributor

@AlekseyTs AlekseyTs Jul 9, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

localSymbol.SynthesizedKind == SynthesizedLocalKind.UserDefined

It is not obvious why this condition is important. #Closed

Copy link
Member Author

@jjonescz jjonescz Jul 9, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See test Retrack_PointerToRefLocal_Synthesized which would fail without it. The test has the reasoning. Basically I think code like (*ptr)++ which implicitly creates a synthesized ref local can continue to be optimized (the local can be scheduled to stack). In other words, we only want to avoid the optimization for user-defined locals.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do you think that?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I guess this doesn't have one correct answer, we could go in one of two ways about this.

  1. I chose to only preserve ref locals that user wrote and expects them to be tracked by GC. Main reason is that pointers are unsafe so GC doesn't need to know about any operations on them, but they can escape through ref locals to safe territory and hence we should let GC know about that transition.
  2. We could go further and make sure GC knows about any other dereferenced pointers, like (*ptr)++ or the example illustrated in test Retrack_PointerToRefLocal_Arg_EvaluationOrder. I think that can be implemented in addition to (1) if desired, but is probably going to be more complicated.

@jjonescz jjonescz requested a review from jkotas July 9, 2025 16:18
}
}
""";
CompileAndVerify(source, verify: Verification.Fails, options: TestOptions.UnsafeDebugDll).VerifyIL("C.M", """
Copy link
Contributor

@AlekseyTs AlekseyTs Jul 9, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

CompileAndVerify

I think we need to execute code and observe its behavior during execution. This comment applies to other similar places. #Closed

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I guess that is probably not necessary if release is now generating code similar to debug and debug code is not changed.

@AlekseyTs
Copy link
Contributor

AlekseyTs commented Jul 9, 2025

Done with review pass (commit 4) #Closed

Copy link
Contributor

@AlekseyTs AlekseyTs left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM (commit 5)

@jkotas
Copy link
Member

jkotas commented Jul 10, 2025

I assume that this change is only fixing the first example from #79051 (comment) . Is that correct?

LGTM

@jjonescz
Copy link
Member Author

I assume that this change is only fixing the first example from #79051 (comment) . Is that correct?

Yes, that's correct.

@jjonescz
Copy link
Member Author

@dotnet/roslyn-compiler for a second review, thanks

@jjonescz jjonescz merged commit 5db3521 into dotnet:main Jul 17, 2025
24 checks passed
@jjonescz jjonescz deleted the 79051-retrack branch July 17, 2025 10:34
@dotnet-policy-service dotnet-policy-service bot added this to the Next milestone Jul 17, 2025
333fred added a commit that referenced this pull request Jul 24, 2025
* Support interpolated string handlers in extension blocks (#78425)

* Support interpolated string handlers in extension blocks

Closes #78137.

* A few tests fail PEVerify

* Update handling of receivers in static scenarios and add tests for indexers/static methods

* Update after merge

* Ensure nullable index adjustment happens always, not just when the receiver parameter is used.

* Additional il verification progress

* PR feedback. The only remaining item is additional testing for ref safety to record the current state.

* Additional PR feedback

* More testing, suggested simplification

* Simplify and more testing.

* Feedback

* Suggested optimization

* Add skipped test for #78433.

* Fix loss of nullability attributes when getting the 'constructed reduced from' method. (#79400)

* Add tracking issue for disabled assert (#79427)

Reference tracking issue #79426 for the disabled assert.

* Fix code gen for some compound assignment scenarios involving extension properties. (#79339)

* Extensions: extension marker type name (#79308)

* Move to xunit.runner.visualstudio 3.1.3

* Avoid eliding pointer-to-ref conversions (#79311)

* Avoid eliding pointer to ref conversions

* Consider only user-defined locals

* Update pre-existing tests

* Test evaluation order

* Explain user-defined local check

* Avoid reusing temps whose refs might be captured (#76009)

* Avoid reusing temps whose refs might be captured

* Revert some changes

* Simplify the heuristic

* Avoid reusing any local whose address has been taken

* Update tests

* Inline a function

* Revert unrelated change

* Remove non-reusable locals after checking for them

* Keep ref count for addressed locals list

* Extend an assert

* Add high-level comment to MightEscapeTemporaryRefs

* Filter non-reusable locals

* Simplify `int`s to `bool`s

* Replace coalesce with an assert

* Mark nested calls as always used

* Fix this parameter of nint methods

* Test chained call

* Simplify by using this parameter symbol

* Fix readonly targets

* Fixup a test

* Remove an unused parameter

* Fix nullability after merge

* [main] Update dependencies from dotnet/arcade (#79385)

[main] Update dependencies from dotnet/arcade


 - Revert .NET SDK bump

 - Set NUGET_PACKAGES for bootstrap builds in CI

 - Set NUGET_PACKAGES for Correctness_Build_Artifacts

 - Add tracking issue for workarounds

 - Call nuget restore with the -publish flag

 - We need to restore where we publish

 - Add comment for strange restore behavior

 - Update comment

* Fixes rename attribute bug (#79418)

* fix bug

* update test

* simplify

* comment

* clarity

* gracefully handle the document or languageservice being null

* Adds EqualityComparer<T>.Create polyfill

* Revert compiler changes

* Update dependencies from https://github.com/dotnet/arcade build 20250717.3

Microsoft.DotNet.Arcade.Sdk , Microsoft.DotNet.Helix.Sdk , Microsoft.DotNet.XliffTasks
 From Version 10.0.0-beta.25358.3 -> To Version 10.0.0-beta.25367.3

* Move to .NET 10p6 (#79438)

* Move to .NET 10p6

Let's try to get ahead of the next one.

* Update contributing docs as well.

* Allow razor to use SolutionChecksumUpdater (#79373)

* Allow razor to use SolutionChecksumUpdater

SolutionChecksumUpdater allows incremental updates of sourcetexts in the server process, but was previously limited to just documents in the solution. Instead, by loosening this to allow TextDocuments, razor editing can have incremental updates of the primary razor document when cohosting is turned on.

* use is/or instead of ==/||

* Remove dependency on ClassificationOptions and DefinitionItem from Semantic Search service (#79439)

* Update insertions for VS release changes

* Fix static extension method not showing up on enum type

* Add workitem

* cleanup

* Use ConfigureAwait(true) when in a blocking JTF run call

* GetTextSynchronously

* Update src/VisualStudio/Core/Def/LanguageService/AbstractLanguageService`2.VsLanguageDebugInfo.cs

Co-authored-by: Jason Malinowski <jason@jason-m.com>

* Make synchronous

* Docs

* Allow Razor to create a formatting options, and send it to OOP

* Update dependencies from https://github.com/dotnet/arcade build 20250717.4

Microsoft.DotNet.Arcade.Sdk , Microsoft.DotNet.Helix.Sdk , Microsoft.DotNet.XliffTasks
 From Version 10.0.0-beta.25358.3 -> To Version 10.0.0-beta.25367.4

* Remove assumption about a tree shape that compound assignment produces (#79432)

* Skip GetProjectXml_FileBasedProgram_SdkTooOld_01 test

* Bump Microsoft.VisualStudio.Extensibility.Testing.Xunit to 0.1.796-beta

This brings along some fixes needed for it to work under newer versions
of xunit.runner.visualstudio.

* Reduce path length of Roslyn ServiceHub Services folder (#79460)

I've been hitting lots of gold bars in VS indicating various Roslyn features are unavailable when running when F5'ing the roslyn project. I finally spent some time digging into why this happens on my machine, and it turns out that out service hub service paths are just a bit too long. This PR just reduces that path length by 14 characters, which is enough for all the services to be under the 260 character long path limit.

* downgrade SDK

* Fix Typo

* Simplify stale project handling (#79386)

* Update System.Memory to prevent bootstrap failure

* Include dependencies of System.Memory in VSIX deployment project

* remove unnecessary additional reference

* Update dependencies from https://github.com/dotnet/arcade build 20250717.5

Microsoft.DotNet.Arcade.Sdk , Microsoft.DotNet.Helix.Sdk , Microsoft.DotNet.XliffTasks
 From Version 10.0.0-beta.25358.3 -> To Version 10.0.0-beta.25367.5

* Use new compiler API

* Correctly change return type of partial definition part

* Feedback

* Move RoslynParallel and ProducerConsumer to threading source package (#79474)

* Downgrade SDK

* Do not remove usings/imports in code-cleanup when file has syntax errors

* Add testws

* Find add-using results in SG docs

* Docs

* Do not place SuppressMessage attributes on an extension block

* remove asynchrony from service registration

* Ignore all targets of MS.CA.Threading.Package in rebuild (#79495)

* Don't mark special compiler synthesized members as obsolete

* Add test

* Use globs in .projitems for source packages (#79480)

* Fix name declaration completion in primary constructor parameters

* nrt

* nrt

* Generate complogs on build failure (#79446)

* Generate complogs on build failure

This should make it easier to reproduce CI compiler failures locally.

* Test bad change for complog creation

* Use --yes

* Include full exception

* Revert "Test bad change for complog creation"

This reverts commit a465981.

* remove from pkgdefs

* Extensions: fix issue with GetParameterOrExtensionParameter (#79402)

* Fix exception when requesting additional file diagnostics

* Do not do full cleanup when we are formatting a new document

The CodeAction.CleanupDocumentAsync was added in
#59091 to ensure the document
had elastic trivia formatted before we applied the result to the
file. It appears the intent was to do just the trivia formatting, but
the CleanupDocumentAsync does a lot of expensive operations around
simplification and adding imports, which aren't needed in this case.

Fixes https://devdiv.visualstudio.com/DevDiv/_workitems/edit/1540532

* Extensions: add test for function type scenario (#79323)

* Reenable analyzer that was disabled when we moved to .Net 10 Preview 5

* Allow captured primary constructor parameters to be still used in base-list

* Simplify ancestor checks

* Fix 'this' keyword not being recomended where allowed

* in progress

* Fix 'switch' keyword recomendations after an expression

* Fix escaping of enum members in completion

* Suggest proper 'Async'-suffixed name for Task-returning symbols

* Fix location of 'partial' modifier when generating new partial members

* Update VMR orchestrator property (#79517)

* Make attachdebugger flag work on non-Windows (#79499)

These compiler flags used Debugger.Launch, which only works on windows. I borrowed the equivalent code from the LSP launcher, which already has this functionality (https://github.com/dotnet/roslyn/blob/cf34acbc5d0f64ff87013cdd02ae61cff562daaf/src/LanguageServer/Microsoft.CodeAnalysis.LanguageServer/Program.cs#L75-L89).

* Downgrade System.Memory to make VS happy

* Fix auth issue in PR val builds

* Fix

* Update Basic.Reference.Assemblies (#79525)

* Update Basic.Reference.Assemblies

Update to version 1.8.3 of Basic.Reference.Assemblies and make references to .NET 10p6 available in our tests.

* Feedback

* Feedback

* Allow add using to work on Razor generated documents

* Improve compiler server/client logging (#79335)

* Remove unused constant

* Include timestamp in logs

* Log messages to msbuild as well

* Log server's process ID

* Log process creation on Unix

* Exclude some messages from the binlog

* Fix test loggers

* Revert "Exclude some messages from the binlog"

This reverts commit 12c1050.

* Cleanup

* Simplify

* Simplify

* Simplify

* Add tests

* Add tests

* Simplify

* Simplify

* initial change

* Renames

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Extensions: address or split remaining open issues directly associated with test plan (#79452)

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* Update tests

* In progress

* in progress

* in progress

* in progress

* in progress

* in progress

* in progress

* in progress

* Cleanup split between lexer and sliding-text-window (#79205)

Co-authored-by: Jan Jones <jan.jones.cz@gmail.com>

* Add docs

* Fix name qualification when moving a using alias outside of a file scoped namespace

* Fix strucutre out of bounds

* Tests

* Add Version.Details.props (#79539)

* Fix issue with 'use explicit type' and nullable tuples

* Add back

* Convert ValueTaskFactory methods to static extensions (#79541)

* Update logic

* Also downgrade System.Numerics.Vectors for VS

* Add Enum static extensions (#79546)

* Fix main build break

* Fix main build break AGAIN

* Projects to deploy (#79430)

* [main] Source code updates from dotnet/dotnet (#79483)

[main] Source code updates from dotnet/dotnet


 - Workaround cref resolve issue due to new RuntimeHelpers in SCI

* [main] Source code updates from dotnet/dotnet (#79563)

[main] Source code updates from dotnet/dotnet

* Revert "Fix main build break AGAIN (#79559)"

This reverts commit e67e2a4, reversing
changes made to 1a74e66.

* Tweak build break fix

* Tweak build break fix

* inline test strings

* Cleanup test methods

* Fix issue where we weren't properly adding elastic trivia to newly generated members

* Fix behavior

* Add test

* Mark SyntaxTokenParser as non-experimental (#79521)

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>

---------

Co-authored-by: Cyrus Najmabadi <cyrus.najmabadi@gmail.com>
Co-authored-by: AlekseyTs <AlekseyTs@users.noreply.github.com>
Co-authored-by: Julien Couvreur <julien.couvreur@gmail.com>
Co-authored-by: Joey Robichaud <joseph.robichaud@microsoft.com>
Co-authored-by: Jan Jones <janjones@microsoft.com>
Co-authored-by: dotnet-maestro[bot] <42748379+dotnet-maestro[bot]@users.noreply.github.com>
Co-authored-by: Joey Robichaud <jorobich@microsoft.com>
Co-authored-by: Ankita Khera <40616383+akhera99@users.noreply.github.com>
Co-authored-by: tmat <tomas.matousek@microsoft.com>
Co-authored-by: Tomáš Matoušek <tmat@users.noreply.github.com>
Co-authored-by: dotnet-maestro[bot] <dotnet-maestro[bot]@users.noreply.github.com>
Co-authored-by: Todd Grunke <toddgrun@microsoft.com>
Co-authored-by: David Barbet <dabarbet@microsoft.com>
Co-authored-by: Cyrus Najmabadi <cyrusn@microsoft.com>
Co-authored-by: Jason Malinowski <jason@jason-m.com>
Co-authored-by: David Wengier <david.wengier@microsoft.com>
Co-authored-by: Jason Malinowski <jason.malinowski@microsoft.com>
Co-authored-by: DoctorKrolic <mapmyp03@gmail.com>
Co-authored-by: Matt Mitchell <mmitche@microsoft.com>
Co-authored-by: Jan Jones <jan.jones.cz@gmail.com>
Co-authored-by: Djuradj Kurepa <91743470+dkurepa@users.noreply.github.com>
Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com>
@RikkiGibson RikkiGibson modified the milestones: Next, 18.0 P1 Aug 19, 2025
jjonescz added a commit that referenced this pull request Mar 1, 2026
Fixes #82137.
Follow up on #79311 to handle more
cases.

### Some background

If there is a pointer-to-ref conversion, we need to preserve the by-ref
local in IL so the CLR can "re-track" it, i.e., not garbage-collect it
while it's in a managed ref-local. Previously we would sometimes
optimize the local away so in IL the code would be just pointer
manipulation which GC doesn't track.

For example, in Release config, this code

```cs
ref byte b = ref p->F;
b.ToString();
```

would be lowered to

```cs
p->F.ToString()
```

but that's wrong if GC happens in between `p->F` and `.ToString()` (if
CLR sees the former code it has to track that local `b` and cannot
collect it but not in the latter case). More real-world examples are in
the linked issue. We previously
[fixed](#79311) this for some cases
but not for some more nested scenarios.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug] Invalid IL sequence emitted for re-track from pointer to byref

5 participants