Various tailcall test improvements

[jakobbotsch]

Add tests that test

• Tail calls to functions with small return values
• Tail calls to functions with retbuf
• Tail calls to functions with multi-reg returned structs
• Tail calls to functions with void returns
• Tail calls to abstract methods
• Tail calls to interface methods
• tail. calli sequences
• Tail calls to struct instance methods
• Tail calls involving refs
• Tail calls involving byrefs
• Tail calls involving generics

Remove some illegal tests that pass the address of variables on the local stack frame to a tail. call.

I have also made some changes to the tailcall return address hijacking test:

• On Linux x64, this test might not be forced to use helper because the struct ForceHelper is passed by value. I have added another arg to one of the tailcall helper functions that fixes this.
• I have changed the test to use fewer resources
• I am experimentally enabling it on ARM32

My new tests should fail on platforms without helper-based tailcalls and the hijacking one might also fail on ARM32/x64 Linux now. I will disable them once I see the CI results.

[jakobbotsch]

It looks like the new test is running into #26311 on x64 Windows. I will have to disable it on all platforms except x86 Windows.

[jakobbotsch]

It was strange to me that the issue in my previous comment was not more widespread since F# users should be running into it for many interface calls in tailcall position. I found out that this failure is a regression -- my new test works fine in .NET core 3.0 preview 8. After some bisecting I have tracked it down to #20643. I will leave it up to someone else whether we should try a targeted fix for it or wait for it to be fixed with #26418 considering that it is a regression.

cc @jashook

[jakobbotsch]

Here is a smaller F# example that access violates with master on x64 Windows:

open System
open System.Runtime.CompilerServices

type IFoo =
    abstract member TailCallee : int -> Guid -> int

[<MethodImpl(MethodImplOptions.NoInlining)>]
let TailCaller (foo : IFoo) =
    foo.TailCallee 0 (new Guid())

type Bar() =
    interface IFoo with
        member this.TailCallee _ _ = 0

[<EntryPoint>]
let main argv = TailCaller (new Bar())

[jashook]

@jakobbotsch this regression happens because we used to fast tail call, but we make the decision to not after #20643?

[jakobbotsch]

@jashook Both are slow tail calls. If I had to take a wild guess it would be related to the resetting of the arg info here:

coreclr/src/jit/morph.cpp

Lines 8309 to 8316 in f8f0c55

	if (!canFastTailCall)
	{
	fgMorphTailCall(call, pfnCopyArgs);
	// Force re-evaluating the argInfo. fgMorphTailCall will modify the
	// argument list, invalidating the argInfo.
	call->fgArgInfo = nullptr;
	}

But I have not at all tried to diagnose it, so it is a wild guess.

[jakobbotsch]

This should be good to go @jkotas.

[jakobbotsch]

1. Added the .cs file for the new test
2. Made TailcallVerifyWithPrefix pinvoke printf on Unix and Sleep on Windows. Clarified that it is now disabled because of no implicit/helper-based tailcalls where it is disabled. I will reenable it on the new supported platforms in [WIP] Implement portable tailcall helpers #26418.
3. Removed all the dead code from TailcallVerifyWithPrefixTest

[jashook]

lgtm minus two nits

[jashook]

Do you mind also adding a small readme into this folder explaining that this file is not used by testing, it is here to help regenerate the IL test.

[jakobbotsch]

Added a comment to the top of the file

[jashook]

Also, feel free to merge after the changes. This has passed CI. No reason to run them again.

[jakobbotsch]

Also, feel free to merge after the changes.

I don't have access to. 😄