JIT: "default(T) == null" does not emit a constant when T is Nullable<>

[michkot]

After toying with efficient detection of reference types / value types / nullable types I triggered a discovery of generic context issue #34641.

I now also find out that default(T) == null for T being Nullable<XXX> in the following snippet does not produce a constant in ASM but rather builds what seems to be a composed "struct + HasValue bool" on stack and then cmps to 0x0.
sharplab.io link with longer example

public static class C<TVal>  {
  public static bool Test(TVal val)
  {
     if (default(TVal) == null)
     {
        return true;
     }
     return false;
  }
}

The produced assembly is (I guess) valid, although quite inefficient. (And I did not find other good way to determine whether type is reference of a non-nullable valuetype.)

[EgorBo]

I now also find out that default(T) == null for T being Nullable in the following snippet does not produce a constant in ASM

why should it produce a constant? a variable of a Nullable type can be null (or not)

[khellang]

@michkot Have you seen #1157 and #32098?

[benaadams]

why should it produce a constant? a variable of a Nullable type can be null (or not)

default of a nullable is always null so its constant?

e.g. default(int?) is always null

[EgorBo]

why should it produce a constant? a variable of a Nullable type can be null (or not)

default of a nullable is always null so its constant?

e.g. default(int?) is always null

Ah, oops.

[EgorBo]

@benaadams @michkot anyway I can't repro it on my master, probably is already fixed (e.g. here? #32269)

It's not the first time people use sharplab with .NET Core 3.0 to file issues against RyuJIT-master 😐

[michkot]

@benaadams @michkot anyway I can't repro it on my master, probably is already fixed (e.g. here? #32269)

It's not the first time people use sharplab with .NET Core 3.0 to file issues against RyuJIT-master 😐

Thanks for looking into it. I will close the issue now then 👍 . Can you point me at some resources/docs so I can try it myself at current master @EgorBo ? (I had luck the last time with sharplab, so I tried it again).

[michkot]

@michkot Have you seen #1157 and #32098?

I did not explicitly, but I noticted that work being done in other issue. That feature/opt. is also nice when you need reference/value type distinction. I trully care about null being a default here though :)

[EgorBo]

@michkot
my workflow is this:

git clone git@github.com:dotnet/runtime.git
cd runtime/src/coreclr
./build.sh -checked

once it finishes:

cd myConsoleApp
dotnet publish -c Release -r osx-x64
cp ~/runtime/artifacts/bin/coreclr/OSX.x64.Checked/* ~/myConsoleApp/bin/netcoreapp3.1/osx-x64/publish
cd ~/myConsoleApp/bin/netcoreapp3.1/osx-x64/publish
COMPlus_JitDisasm=MyNamespace.MyProgram:MyMethod ./myConsoleApp

Also, let me promote my add-in for that: https://github.com/EgorBo/Disasmo (it still needs a pre-built runtime in Checked config)