Fix is and pattern-matching behavior in presence of implicit UD conversion
#24547
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…version and also an explicit reference conversion. User-defined conversions should never be considered for `is` and pattern-matching. Fixes dotnet#24522
Member
Author
|
@agocke @dotnet/roslyn-compiler may I please have a couple of reviews of this pattern-matching bug fix? |
AlekseyTs
reviewed
Feb 1, 2018
| HashSet<DiagnosticInfo> discarded = null; | ||
| Conversion c = Compilation.Conversions.ClassifyConversionFromExpression(operand, type, ref discarded); | ||
| const bool preferBuiltinConversion = true; | ||
| Conversion c = Compilation.Conversions.ClassifyConversionFromExpression(operand, type, ref discarded, preferBuiltinConversion); |
Contributor
There was a problem hiding this comment.
preferBuiltinConversion [](start = 114, length = 23)
It doesn't look like passing this value guarantees that user defined conversions are ignored completely. #Closed
Member
Author
There was a problem hiding this comment.
That isn't the intended behavior. It just needs to prioritize builtin conversions over user-defined conversions.
Contributor
There was a problem hiding this comment.
It just needs to prioritize builtin conversions over user-defined conversions.
What would be a valid scenario when a user defined conversion should be accepted here? #Closed
Member
Author
There was a problem hiding this comment.
I don't know what you mean by "accepted". The conversion is required in the node but isn't used subsequent to lowering. The is operator in the language does not actually require that a conversion exist. The conversion is stored to assist with semantics (specifically warnings) and lowering when the is operator appears in source. Those don't apply to the node produced here.
If it would make it more clear, we can use NoConversion here.
Member
Author
There was a problem hiding this comment.
Or, if you prefer, we can compute a built-in conversion from type.
Contributor
|
Done with review pass (iteration 1). #Closed |
Member
Author
|
Done with responses to review pass (iteration 1). Do you have any other comments? |
Member
Author
|
@agocke Can you please review this pattern-matching bug fix? |
1 similar comment
Member
Author
|
@agocke Can you please review this pattern-matching bug fix? |
Member
|
@gafter Looking now. |
agocke
approved these changes
Feb 2, 2018
heejaechang
pushed a commit
that referenced
this pull request
Feb 7, 2018
* Remove duplicate lock DocumentState.s_syntaxTreeToIdMapLock This lock is only being used to protect access to an instance which contains internal synchronization. * Better handle surrounding directives when inlining a local variable. * Add tests. * Share code between VB and C#. * Reduce allocations in UnboundLambda Fixes #23463 * Restore ReturnInferenceCacheKey as the key for _returnInferenceCache * Update code to more closely follow patterns of the original code * Cleanup from code review * Verify MSBuild version in Developer CMD prompt Roslyn is designed to have the simplest possible contribution story: clone then build. Every pre-req needed is either located on the machine or bootstrapped via NuGet. All the way down to using an xcopy MSBuild if needed. The one case which causes a problem is the VS command prompt. In this case MSBuild is pre-installed on the machine and may or may not be suitable for building Roslyn. Previously when building from a VS command prompt we just used whatever MSBuild was provided. The assumption being a developer command prompt was an explicit statement of whath MSBuild you wanted to use. Based on all of our customer reports though this does not seem to be the assumption that consumers of our repo have. The build gave them no explicit errors about the provided toolset and hence when the build failed they assigned flakiness to our repo. Going forward we are applying the same version validation to MSBuild when provided via a developer command prompt. If it doesn't match we will refuse to build asking the user to upgrade VS or build from a normal command prompt. * Remove unneeded debugging line * Comment about pre-release * Added minimum version * Add Omit If Default style option * Add space to be like test without the omit * Add/Remove without needing a property * Reformat * PR feedback * Fix VB diagnostic based on feedback * Handle case of NotApplicable modifier and field declaration list * Fix tests * PR feedback * PR feedback * Support negative null-check when we are suggesting to inline type checks Fixes #21097 Fixes #24286 * fix a case where persistent storage registration fails and some clean… (#24458) * fix a case where persistent storage registration fails and some clean up code around it. * added readonly * address PR feedback * removed comments no longer relevant * renamed lock name * moved waiter from diagnostics.dll to features.dll where all interfaces are defined. (#24512) * put listener change back in (#24120) * leave old types in legacy folder until partner teams move to new interface * added legacy waiter to support partner teams * Remove methods indirecting access to _metadataFileNameToConvertedProjectReference This field is documented as being written and read from any thread, but in practice all uses are guarded by an AssertIsForeground(). Thus we can get rid of the helper methods that are trying to "help" by locking before accessing the fields, making it really hard to track all the real uses of it. * Make method static that doesn't need state * Fix up tests of P2P to metadata reference conversion It turns out we had some tests, but the tests were disabled. This was because the tests weren't working properly anyways: they were calling into UpdateProjectBinPath which only updated some (but not all) of the project state. That was an internal helper method that shouldn't be used by tests. Updating the tests to use SetBinOutputPathAndRelatedData works better. * Delete debug-only reference validation This was some legacy code that tried to verify that the references we have from the project system match up to what DTE and other sources say. This was debug-only, and the actual asserts were commented out. This is deadweight at this point, so delete it. * added and cleaned up logs around build and live diagnostics. (#24551) also added RoslynActivityLogger that can be enabled through project-system-tool * Avoid closure allocations on the BindSyntaxTreeToId fast path * CS1628 error text mentions in parameters; fixes #24584 * Update optimization data to 2.7.0-beta3-62526-01... * Small cleanup of completion logic. * Locate implementations for reference assemblies using the process binding path * Use GlobalAssemblyCache helper to locate assemblies directly in the GAC * Update InteractiveEditorFeatures to account for a second definition of GlobalAssemblyCache * Move to xunit.console for CoreClr tests Previously we were using xunit.console for desktop tests and dotnet-xunit for our CoreClr tests. This change unifies us on top of xunit.console (now that it has a netcoreapp2.0 version available). * Move unix builds to xunit.runner.console as well * Fixes 559223 Fix and re-enable test that would catch this error * Update LanguageServices training data again... * Get actual directory name, not file * Fix dir name issue * Cleanup based on code review feedback * Check fully-qualified names for SuppressIldasmAttribute and ReferenceAssemblyAttribute * Use correct reference location, or fail decompilation if it's not available * Fix typo... * Don't use inferred member name if that creates duplicates (#24632) * Fixes #23983 * Added test for unique IDEDiagnosticIDs * Fixed capitalization on local variable * Fix `is` and pattern-matching behavior in presence of implicit UD conversion (#24547) * Fix `is` and pattern-matching behavior in presence of implicit UD conversion and also an explicit reference conversion. User-defined conversions should never be considered for `is` and pattern-matching. Fixes #24522
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fix
isand pattern-matching behavior in presence of implicit UD conversionand also an explicit reference conversion between the same types.
User-defined conversions should never be considered for
isand pattern-matching.Customer scenario
See #24522
Bugs this fixes
Fixes #24522
Workarounds, if any
Casting the left-hand-side to
objectis often a work-around.Risk
Low, as the fix is quite localized.
Performance impact
Should be slight performance improvement, as we now only check built-in conversions for certain type-check operations.
Is this a regression from a previous update?
Yes. C#6 did not have this bug.
Root cause analysis
We did not realize that it is possible for an implicit user-defined operator to be present when an explicit reference conversion applies.
We now have tests for the ways that symptoms can arise.
How was the bug found?
Customer reported.
Test documentation updated?
N/A