[Arm64] Don't generate hardware intrinsic method bodies

[echesakov]

Fixes #39618

[tannergooding]

Why don't we need something similar for Vector128<T> and Vector256<T> on x86?

[tannergooding]

(I might just be missing where we cover it, if it exists somewhere)

[echesakov]

My understanding (that could be incorrect) is that on x86 in R2R code we want to treat all calls outside of System.Runtime.Intrinsics.X86 as regular calls since there is no guarantee what ISAs will be supported on target platform while on Arm64 - ArmBase and AdvSimd are baseline and always supported.

@davidwrighton Is it correct interpretation of discussion in #38060?

[tannergooding]

Ah right. I remember this reasoning now. It always confuses me at first 😄

[kunalspathak]

you are missing |=. It should be fTreatAsRegularMethodCall |= (strcmp(className, "Vector641") != 0) && (strcmp(className, "Vector1281") != 0);

[echesakov]

No, I am not.

Consider this:

1. If fTreatAsRegularMethodCall=true before line 2160, then it means fIsPlatformHWIntrinsic=true and the if-condition is false.
2. If fTreatAsRegularMethodCall=false before line 2160, then fTreatAsRegularMethodCall = fTreatAsRegularMethodCall || (if-cond-expr) is equvalent to fTreatAsRegularMethodCall = (if-cond-expr).

I could've written this as

fTreatAsRegularMethodCall = (fIsGetIsSupportedMethod && fIsPlatformHWIntrinsic) || (!fIsPlatformHWIntrinsic && fIsHWIntrinsic 
#if defined(TARGET_ARM64)
    && ((strcmp(className, "Vector64`1") == 0) || (strcmp(className, "Vector128`1") == 0))
#endif
);

[kunalspathak]

I could've written this as

Yes either that or something like this:

        fTreatAsRegularMethodCall = (fIsGetIsSupportedMethod && fIsPlatformHWIntrinsic);

        if (!fIsPlatformHWIntrinsic && fIsHWIntrinsic)
        {
#if defined(TARGET_ARM64)
            // For ARM64, treat as method calls if HWIntrinsic is not the one supported.
            fTreatAsRegularMethodCall |= (strcmp(className, "Vector64`1") != 0) && (strcmp(className, "Vector128`1") != 0);
#else
            // For XArch, always treat HWIntrinsic as method calls.
  	    fTreatAsRegularMethodCall |= true;
#endif 
        }

Currently, it just feels like we override the fTreatAsRegularMethodCall without carefully inspecting the code the way you pointed.

[echesakov]

I pushed the simpler change. I think it's better to repeat (!fIsPlatformHWIntrinsic && fIsHWIntrinsic) twice.

[echesakov]

@AntonLapounov Can you please show me the place when I can disable pre-generation of the intrinsics bodies on Arm64 in CrossGen2?

[AntonLapounov]

I think CorInfoImpl.ShouldSkipCompilation controls that.

[echesakov]

It looks it already does this right

runtime/src/coreclr/src/tools/aot/ILCompiler.ReadyToRun/JitInterface/CorInfoImpl.ReadyToRun.cs

Lines 190 to 193 in 557c7af

	if (HardwareIntrinsicHelpers.IsHardwareIntrinsic(methodNeedingCode))
	{
	return true;
	}

Thanks!

[kunalspathak]

LGTM