RFC FS-1124 - Interfaces with static abstract methods

[vzarytovskii]

Support interfaces with static abstract methods suggestion, RFC

Things included in this PR that are not yet documented in RFC:

• Fix C<^int> to not need a space, including updating tests
• Allow ((^T or C) : (static member P: int) ()) SRTP call syntax (i.e. nominal types after or)

TODO:

• ILgen tests
• Make sure we don't permit anything extra, other than statics in interfaces.
• Consider glyphs for trait constraints
• Completion for 'T. IWSAMs + tests
• Go to definition for 'T.Name IWSAM (should work)
• Consider tooltips for 'T - may be a larger issue since these are missing for any generic type parameters at the moment
• Move tests/adhoc.fsx into individual unit tests including codegen
• Test creating delegates to IWSAM-constrained target methods
• Fix new SRTP call syntax for instance methods
• Fix trait with expression warning
• Test op_Implicit
• Test active patterns
• Test calling F#-defined function/method with IWSAM from C#
• Test (^T or C) : ... call syntax for SRTP
• Don - Fix type inference issue
• Don - Fix SRTP trait calls with byref parameters
• Petr/Vlad - Adhoc usage testing of autocomplete for trait calls + fix any bugs
• Petr/Vlad - Adhoc usage testing against actual latest .NET 7 generic math preview
• Petr - Test suppression of Systems.Numerics.I* interfaces on unitized types
• Don/Petr/Vlad - Get it green

Other cleanup items currently in this PR (could be factored out):

• Add ParamAttribs type in infos.fs for large tuple
• Pass extra g parameter to Freshen routines
• Remove wildcard matching when pattern matching on Item type, completing some cases in the obvious way
• Factor out AddUnsolvedMemberConstraint, TraitsAreRelated, EnforceConstraintConsistency, CheckConstraintImplication, EnforceConstraintSetConsistency, EliminateRedundantConstraints, cleanup SolveTypeChoice in ConstraintSolver
• Name fields of AssignedPropSetter
• Factor out CrackParamAttribsInfo in infos.fs
• Pipe into List.iter2 in ilxgen.fs
• Add DecideStaticOptimizations call to Optimizer.fs
• Some comments and cleanup in autocomplete implementation in FSharpCheckerResults.fs
• Replace some uses of direct matching with TType_app (tcref, _, _) with AbbrevOrAppTy tcref which looks through type inference equations.
• Some refactoring in SemanticClassification.fs to recude complexity
• Fix TTypeConvOp in Exprs.fs to use type equality
• Some cleanup for flags in TypedTree.fs
• Field names for TraitWitnessInfo
• Field names for ILMethSln
• Add documentation to decideStaticOptimizationConstraint
• Get HelpContextServiceTests.fs activated and working

Renamings:

• TcLongIdent --> TcLongIdent
• checkLanguageFeatureRuntimeErrorRecover --> checkLanguageFeatureRuntimeAndRecover`
• mkILNonGenericVirtualMethod --> mkILNonGenericVirtualInstanceMethod
• GetBaseClassCandidates --> IsInheritsCompletionCandidate
• GetInterfaceCandidates --> IsInterfaceCompletionCandidate
• IsInterfaceFile --> IsSignatureFile
• For traits:
  • MemberName --> MemberLogicalName
  • ArgumentTypes --> CompiledObjectAndArgumentTypes
  • ReturnType --> CompiledReturnType
• Microsoft.VisualStudio.FSharp.Editor.Tests.Roslyn --> VisualFSharp.UnitTests.Editor
• CreateDocument --> CreateSingleDocumentSolution

[dsyme]

@vzarytovskii@0101 I pushed a fix to allow byref arguments in the new syntax for SRTP trait calls, e.g. see tests/adhoc.fsx

These should be turned into proper tests please :)

There is a note about byref returns in member traits, which are not supported - they could be, but it is orthogonal to this PR

[dsyme]

These tests look good!

Probably add a test when the user defines an interface in System.Numerics too

[dsyme]

Also one where the user defines a new interface NOT in System.Numerics but derives from an interface in System.Numerics

[dsyme]

We should probably adjust the FSharp.Core.UnitTests and a few others to do this. I've noticed they use the Proto to build, perhaps it's better to just use the built compiler.

[vzarytovskii]

This looks good to me.
@KevinRansom@0101 give it another go pls and we can merge.

[0101]

Would be nice to include the switch to net7.0 to enable the tests for System.Numerics suppression but I suppose that can be done later.

[vzarytovskii]

Would be nice to include the switch to net7.0 to enable the tests for System.Numerics suppression but I suppose that can be done later.

Yeah, we probably shouldn't wait for it, can be done in parallel/later.

[cmeeren]

I can't find this info anywhere: When using ^a or ^b in the old syntax, what's the new syntax?

Here is the old syntax:

let inlineoldSyntax<^a,^bwhen(^a or ^b):(static member create:^a->^b)>(x:^a):^b =((^aor^b):(static membercreate:^a->^b) x)

But in F# 7, this doesn't compile:

let inlinenewSyntax<'a,'bwhen('a or 'b):(static member create:^a->'b)>(x:'a):'b =('a or 'b).create x

Neither does replacing the second line with ('a or b).create x.

What is the new syntax in this case?

[cmeeren]

It also seems like

let inlinef(x:'a):'b =(^a:(static membercreate: 'a -> 'b) x) fails compilation if replacing ^a with 'a.

fails if replacing the ^ in the function body with '. Is that intended? Can ' only be used when defining the SRTP constraints directly in the function definition, not the body?

[vzarytovskii]

It also seems like

let inlinef(x:'a):'b =(^a:(static membercreate: 'a -> 'b) x) fails compilation if replacing ^a with 'a.

fails if replacing the ^ in the function body with '. Is that intended? Can ' only be used when defining the SRTP constraints directly in the function definition, not the body?

Yeah, if the old syntax needs to be used, type ^ must be used, it's a known limitation.

[cmeeren]

Ok! But what about 'a or 'b? Does that still require the old syntax for invocation?

[Xyncgas]

how can I implement static abstract members in my derived classes