Make Type.IsGenericType and GetGenericTypeDefinition() JIT intrinsics (and JIT time constants)

[Sergio0694]

Overview

This is one more codegen improvement we could leverage for CsWinRT, related to #95929. One scenario we have (here) is to check whether a given type T is some KeyValuePair<,> instantiation. This needs to be marshalled in a specific way to match the expectations of WinRT. Right now however, the only way we have to check for this case is to use IsGenericType and GetGenericTypeDefinition(), which are not JIT intrinsics. This makes the linker not see this special case, and therefore not correctly trim the code away (or just specialize it on the other hand). This is both marginally slower, but most importantly leaves some codegen size reduction opportunities on the table. We're seeing a bunch of KeyValuePair<__Canon, __Canon> instantiations in sizoscope even in sample WinRT components that never use KeyValuePair<,> at all anywhere. It would be nice if this could just be properly inlined as a constant instead.

Codegen

using SharpLab.Runtime;

[JitGeneric(typeof(System.Collections.Generic.KeyValuePair<int, int>))]
[JitGeneric(typeof(int))]
static bool IsKVP<T>()
{
    return typeof(T).IsGenericType && typeof(T).GetGenericTypeDefinition() == typeof(System.Collections.Generic.KeyValuePair<,>);
}

Current codegen on .NET 8 x64 (sharplab):

Program.<<Main>$>g__IsKVP|0_0[[System.Collections.Generic.KeyValuePair`2[[System.Int32, System.Private.CoreLib],[System.Int32, System.Private.CoreLib]], System.Private.CoreLib]]()
    L0000: sub rsp, 0x28
    L0004: mov rcx, 0x23e784e3260
    L000e: call qword ptr [0x7ffb1060a518]
    L0014: test eax, eax
    L0016: je short L0040
    L0018: mov rcx, 0x23e784e3260
    L0022: call qword ptr [0x7ffb1060a568]
    L0028: mov rcx, 0x23e783e2728
    L0032: cmp rax, rcx
    L0035: sete al
    L0038: movzx eax, al
    L003b: add rsp, 0x28
    L003f: ret
    L0040: xor eax, eax
    L0042: add rsp, 0x28
    L0046: ret

Program.<<Main>$>g__IsKVP|0_0[[System.Int32, System.Private.CoreLib]]()
    ; pretty much the same as above

Expected codegen:

Program.<<Main>$>g__IsKVP|0_0[[System.Collections.Generic.KeyValuePair`2[[System.Int32, System.Private.CoreLib],[System.Int32, System.Private.CoreLib]], System.Private.CoreLib]]()
    L0000: mov eax, 1
    L0005: ret

Program.<<Main>$>g__IsKVP|0_0[[System.Int32, System.Private.CoreLib]]()
    L0000: xor eax, eax
    L0002: ret

cc. @jkoritzinsky @MichalStrehovsky @EgorBo

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Overview

This is one more codegen improvement we could leverage for CsWinRT, related to #95929. One scenario we have (here) is to check whether a given type T is some KeyValuePair<,> instantiation. This needs to be marshalled in a specific way to match the expectations of WinRT. Right now however, the only way we have to check for this case is to use IsGenericType and GetGenericTypeDefinition(), which are not JIT intrinsics. This makes the linker not see this special case, and therefore not correctly trim the code away (or just specialize it on the other hand). This is both marginally slower, but most importantly leaves some codegen size reduction opportunities on the table. We're seeing a bunch of KeyValuePair<__Canon, __Canon> instantiations in sizoscope even in sample WinRT components that never use KeyValuePair<,> at all anywhere. It would be nice if this could just be properly inlined as a constant instead.

Codegen

using SharpLab.Runtime;

[JitGeneric(typeof(System.Collections.Generic.KeyValuePair<int, int>))]
[JitGeneric(typeof(int))]
static bool IsKVP<T>()
{
    return typeof(T).IsGenericType && typeof(T).GetGenericTypeDefinition() == typeof(System.Collections.Generic.KeyValuePair<,>);
}

Current codegen on .NET 8 x64 (sharplab):

Program.<<Main>$>g__IsKVP|0_0[[System.Collections.Generic.KeyValuePair`2[[System.Int32, System.Private.CoreLib],[System.Int32, System.Private.CoreLib]], System.Private.CoreLib]]()
    L0000: sub rsp, 0x28
    L0004: mov rcx, 0x23e784e3260
    L000e: call qword ptr [0x7ffb1060a518]
    L0014: test eax, eax
    L0016: je short L0040
    L0018: mov rcx, 0x23e784e3260
    L0022: call qword ptr [0x7ffb1060a568]
    L0028: mov rcx, 0x23e783e2728
    L0032: cmp rax, rcx
    L0035: sete al
    L0038: movzx eax, al
    L003b: add rsp, 0x28
    L003f: ret
    L0040: xor eax, eax
    L0042: add rsp, 0x28
    L0046: ret

Program.<<Main>$>g__IsKVP|0_0[[System.Int32, System.Private.CoreLib]]()
    ; pretty much the same as above

Expected codegen:

Program.<<Main>$>g__IsKVP|0_0[[System.Collections.Generic.KeyValuePair`2[[System.Int32, System.Private.CoreLib],[System.Int32, System.Private.CoreLib]], System.Private.CoreLib]]()
    L0000: mov eax, 1
    L0005: ret

Program.<<Main>$>g__IsKVP|0_0[[System.Int32, System.Private.CoreLib]]()
    L0000: xor eax, eax
    L0002: ret

cc. @jkoritzinsky @MichalStrehovsky @EgorBo

Author:	Sergio0694
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[Sergio0694]

Sharing some notes from Discord:

• @MichalPetryka mentioned that IsConstructedGenericType would be easier to make an intrinsic compared to IsGenericType, and it would still be sufficient for this scenario. Either of them would be completely fine, really.
• @SingleAccretion mentioned that making GetGenericTypeDefinition() could be difficult to make an intrinsic, but it would be possible to just make direct comparisons over that return value be an intrinsic. That would also work in this scenario.

[EgorBo]

grep.app:

• https://grep.app/search?q=typeof%5C%28%2A.%5C%29.IsGenericType&regexp=true&filter[lang][0]=C%23
• https://grep.app/search?q=typeof%5C%28%2A.%5C%29.GetGenericTypeDefinition&regexp=true&filter[lang][0]=C%23

[Sergio0694]

I agree there's not many hits, but this would be beneficial for CsWinRT. We're really trying to minimize the binary size in minimal applications and WinRT components, to make the adoption of C# easier to justify in place of C++ (and binary size is one of the main problems you immediately hit when doing so). We have lots of heavily generic code especially in our marshallers, and they all end up being instantiated over KeyValuePair<,> today because we can't tell the linker that branch is really not used. Having a linker-friendly way of doing that check would be nice to have, is all 🙂

[GerardSmit]

There are more hits with GitHub Search:

• https://github.com/search?q=%2Ftypeof%5C%28%5BA-Za-z0-9._%3C%3E+%5D*%5C%29%5C.IsGenericType%2F+language%3AC%23&type=code
• https://github.com/search?q=%2Ftypeof%5C%28%5BA-Za-z0-9._%3C%3E+%5D*%5C%29%5C.GetGenericTypeDefinition%5C%28+*%5C%29%2F+language%3AC%23&type=code

[teo-tsirpanis]

It would also help with F#'s typedefof.

[hez2010]

Can we also intrinsify typeof(T).IsConstructedGenericType as well?

[Sergio0694]

Thank you for volunteering! 😄