Extensions: proposal to use type inference during lookup (#9126)

Author: jcouv

meetings/working-groups/extensions/extensions-lookup.md

@@ -0,0 +1,99 @@
+There are two parts to this proposal:
+1. adjust how we find compatible substituted extension containers
+2. align with current implementation of extension methods
+
+# Finding a compatible substituted extension container
+
+The proposal here is to look at `extension<extensionTypeParameters>(receiverParameter)` like a method signature, 
+and apply current [type inference](https://github.com/dotnet/csharpstandard/blob/draft-v8/standard/expressions.md#1263-type-inference) 
+and [receiver applicability](https://github.com/dotnet/csharpstandard/blob/draft-v8/standard/expressions.md#128103-extension-method-invocations) rules to it, given the type of a receiver.
+
+The type inference step infers the extension type parameters (if possible).  
+The applicability step tells us whether the extension works with the given receiver, 
+using the applicability rules of `this` parameters.
+
+This can be applied both when the receiver is an instance or when it is a type.
+
+Re-using the existing type inference algorithm solves the variance problem we'd discussed in LDM.  
+It makes this scenario work as desired, because type inference is smarter than the implemented algorithm for extensions:
+```cs
+IEnumerable<string>.M();
+
+static class E
+{
+  extension(IEnumerable<object>)
+  {
+    public static void M() { }
+  }
+}
+```
+
+# Aligning with implementation of classic extension methods
+
+The above should bring the behavior of new extensions very close to classic extensions.  
+But there is still a small gap with the current implementation of classic extension methods, 
+when arguments beyond the receiver are required for type inference of the type parameters
+on the extension container.  
+The spec for classic extension methods specifies 2 phases (find candidates compatible with the receiver, then complete the overload resolution), 
+but the implementation only has 1 phase (find all candidates and do overload resolution with all the arguments including one for the receiver value).
+
+Example we had [discussed](https://github.com/dotnet/csharplang/blob/main/meetings/2024/LDM-2024-10-02.md#extensions):
+```cs
+public class C
+{
+    public void M(I<string> i, out object o)
+    {
+        i.M(out o); // infers E.M<object>
+        i.M2(out o); // error CS1503: Argument 1: cannot convert from 'out object' to 'out string'
+    }
+}
+public static class E
+{
+   public static void M<T>(this I<T> i, out T t) { t = default; }
+   extension<T>(I<T> i)
+   {
+      public void M2(out T t) { t = default; }
+   }
+}
+public interface I<out T> { }
+```
+
+My proposal is that the implementation continue to diverge from the spec: instead of doing 2-phase lookup 
+(as described in the section above, where we find compatible substituted extension containers, then find the candidate members in those)
+we could do a 1-phase lookup. We would only do this in invocation scenarios.
+
+For such invocation scenarios:
+1. we collect all the candidate methods (both classic extension methods and new ones, without excluding any extension containers)
+2. we combine all the type parameters and the parameters into a single signature
+3. we apply overload resolution to the resulting set
+
+The transformation at step2 would take a method like the following:
+```cs
+static class E
+{
+  extension<extensionTypeParameters>(receiverParameter)
+  {
+    void M<methodTypeParameters>(methodParameters);  
+  }
+}
+```
+and produce a signature like this:
+```
+static void M<extensionTypeParameters, methodTypeParameters>(this receiverParameter, methodParameters);
+```
+
+Note: for static scenarios, we would play the same trick as in the above section, where we take a type/static receiver and use it as an argument.
+
+# Recap
+
+If we accepted both parts of the proposal:
+- `instance.Method(...)` would behave exactly the same whether `Method` is a classic or new extension method 
+(from an implementation perspective)
+- `Type.Method(...)` would behave exactly like the instance scenario
+- other scenarios all use the new resolution method where we figure out the compatible substituted extension containers, then collect candidates
+- `instance.Property` and `Type.Property` 
+- `instance[...]`
+(we first figure out the compatible substituted extension container, then do overload resolution with the candidate indexers)
+- const, nested type, operators, ...
+
+