Common/fast lookupHWIntrinsic(...) API

[sdmaclea]

NamedIntrinsic Compiler::lookupNamedIntrinsic(CORINFO_METHOD_HANDLE method)

Currently uses this code to lookup intrinsics

#ifdef FEATURE_HW_INTRINSICS
#if defined(_TARGET_XARCH_)
    if ((namespaceName != nullptr) && strcmp(namespaceName, "System.Runtime.Intrinsics.X86") == 0)
    {
        InstructionSet isa = lookupHWIntrinsicISA(className);
        result             = lookupHWIntrinsic(methodName, isa);
    }
#elif defined(_TARGET_ARM64_)
    if ((namespaceName != nullptr) && strcmp(namespaceName, "System.Runtime.Intrinsics.Arm.Arm64") == 0)
    {
        result = lookupHWIntrinsic(className, methodName);
    }
#else // !defined(_TARGET_XARCH_) && !defined(_TARGET_ARM64_)
#error Unsupported platform
#endif // !defined(_TARGET_XARCH_) && !defined(_TARGET_ARM64_)
#endif // FEATURE_HW_INTRINSICS

It would be nice to simplify this to

#ifdef FEATURE_HW_INTRINSICS
    if (result == NI_Illegal)
    {
        result = lookupHWIntrinsic(namespaceName, className, methodName, method);
    }
#endif // FEATURE_HW_INTRINSICS

The actual parameters to lookupHWIntrinsic are arguable.

I proposed including method so that getMethodHash(CORINFO_METHOD_HANDLE) could be used for fast lookup from an unordered dictionary. Of course a hash of the strings could also be used, but I suspect method hash would be faster.

@4creators @tannergooding @CarolEidt @fiigii

category:implementation
theme:jit-coding-style
skill-level:intermediate
cost:small

[tannergooding]

This sounds reasonable to me, provided we can still keep it performant.

I believe current investigations lead us to believe we would need something like a binary search (or better), given the sheer number of intrinsics some ISAs are providing.

[CarolEidt]

This sounds reasonable to me, provided we can still keep it performant.

It's hard to imagine that it could be any worse than the current approach.

I believe current investigations lead us to believe we would need something like a binary search (or better), given the sheer number of intrinsics some ISAs are providing.

Binary search seems reasonable, since all the keys are known at compile-time. Whatever we do, we don't have to pay setup cost for the case where there are no intrinsics being used.

[fiigii]

I remember @AndyAyersMS suggested that the faster-lookup should handle all the new-style named intrinsic.

[4creators]

Binary search seems reasonable, since all the keys are known at compile-time

Than it should be possible to go for O(1) - a bit unconventional switch table.

[jkotas]

Duplicate of #13617