C# 7.0 tuples - Invalid code generation for expression trees

[bartdesmet]

Repro steps:

C:\Projects\roslyn\Binaries\Debug>type e.cs
using System;
using System.Linq.Expressions;

class E
{
    static void Main()
    {
        Expression<Func<(int, int)>> f = () => new (int, int)(1, 2);
    }
}
C:\Projects\roslyn\Binaries\Debug>csc e.cs ..\..\src\Compilers\Test\Resources\Core\NetFX\ValueTuple\*.cs
Microsoft (R) Visual C# Compiler version 42.42.42.42
Copyright (C) Microsoft Corporation. All rights reserved.


C:\Projects\roslyn\Binaries\Debug>e

Unhandled Exception: System.ArgumentException: Cannot resolve method Void .ctor(Int32, Int32) because the declaring type of the method handle System.ValueTuple`2[T1,T2] is generic. Explicitly provide the declaring type to GetMethodFromHandle.
   at System.Reflection.MethodBase.GetMethodFromHandle(RuntimeMethodHandle handle)
   at E.Main()

C:\Projects\roslyn\Binaries\Debug>ildasm e.exe

Quick inspection shows that the generated IL code has:

  IL_0001:  ldtoken    method instance void valuetype System.ValueTuple`2<int32,int32>::.ctor(!0,
                                                                                              !1)
  IL_0006:  call       class [mscorlib]System.Reflection.MethodBase [mscorlib]System.Reflection.MethodBase::GetMethodFromHandle(valuetype [mscorlib]System.RuntimeMethodHandle)

This should call the GetMethodFromHandle overload with a RuntimeTypeHandle parameter that should be set to ldtoken valuetype System.ValueTuple2<int32,int32>`. In contrast, the following works fine:

        Expression<Func<(int, int)>> f = () => (1, 2);

It seems the difference is noticed here:

NamedTypeSymbol.AllTypeArgumentCount()
SyntheticBoundNodeFactory.GetMethodFromHandleMethod(...)
SyntheticBoundNodeFactory.ConstructorInfo(...)
ExpressionLambdaRewriter.VisitObjectCreationExpressionInternal(...)

That method returns 2 for the use of the tuple literal (1, 2) but 0 for the use of new (int, int)(1, 2). Looking higher up in the LocalRewriter it seems that VisitObjectCreationExpression receives the constructor of the tuple type but it never gets retargeted to the TupleUnderlyingType unlike what happens in the VisitTupleLiteral case.

[bartdesmet]

At first sight, it seems a substitution a la node.Constructor.TupleUnderlyingMethod ?? node.Constructor inside the if (_inExpressionLambda) in VisitObjectCreationExpression can address the issue. It'd require a change from node.UpdateArgumentsAndInitializer to the full-blown node.Update to pass in the rewritten constructor.

[bartdesmet]

I have a spot fix in my fork for expression tree futures @ bartdesmet@994f964. I haven't studied the code base extensively for the places where tuple types are retargeted to their underlying types, so the right fix may be in a different location. I'm planning to play around more with the intersection of expression trees and tuple types in the next few weeks.

[gafter]

The syntax new ( is to be forbidden.

[bartdesmet]

Great, that will fix it :-). Do we have an up-to-date document on the C# 7.0 features and their language surface, reflecting latest decisions? I'm basing my investigations on the docs/features MD files right now.

[VSadov]

#10891 should make this problem unreachable since the code would be cause a compile error