Vector64 and Vector128.Create doesn't generate movi for long immediates

[tannergooding]

Vector64.Create and Vector128.Create don't generate movi as the other types do for immediate inputs. This appears to be because the logic is using IsCnsIntOrI whioch will miss GT_CNS_LNG: https://github.com/dotnet/runtime/blob/master/src/coreclr/src/jit/lowerarmarch.cpp#L933-L957

CC. @kunalspathak, @CarolEidt, @echesakovMSFT

[tannergooding]

I was going to suggest using IsIntegralConst, but It looks like IsIntegralConst it (https://github.com/dotnet/runtime/blob/master/src/coreclr/src/jit/gentree.h#L7315-L7322) looks to only check GT_CNS_INT on 64-bit, which means we'll miss the case where a GT_CNS_LNG was created on 64-bit (even though the underlying field is size_t and could have been a GT_CNS_INT)

[kunalspathak]

Thanks for reporting it @tannergooding . Perhaps .Create() is the only API that gets affected by using IsCnsIntOrI.

[tannergooding]

Possibly, but it seems like something that is very easy to hit if you have an API that can take either an int or long.

I'd think that you really don't care if it is an int or long most of the time, you just care what the value is and if it can "fit" in the encoding. For example, a GT_CNS_LNG that is -1 is just as encodable as a GT_CNS_INT of the same value when the immediate is sign-extended.

[echesakov]

@tannergooding Can you give an example when Vector128.Create does not lower to movi?

Regarding your comment

This appears to be because the logic is using IsCnsIntOrI whioch will miss GT_CNS_LNG

I don't think this is a case. GT_CNS_LNG should never appear in the IR on 64-bit target
https://github.com/dotnet/runtime/blob/master/src/coreclr/src/jit/gentree.h#L2920-L2926
since GT_CNS_INT is stored in size_t which is capable to hold both int and long.

[tannergooding]

GT_CNS_LNG should never appear in the IR on 64-bit target

Ah I see, gtNewLconNode just creates a GT_CNS_INT so you should only get one if you manually create a GT_CNS_LNG...

Looks like its missing because we get:

               [000007] -----+------              *  HWINTRINSIC simd16 long DuplicateToVector128
               [000006] -----+------              \--*  CAST      long <- int
               [000005] -----+------                 \--*  CNS_INT   int    31

[AndyAyersMS]

@kunalspathak milestone? Seems like 5.0, so putting it there for now.

[kunalspathak]

@tannergooding - Can you confirm if you will be optimizing long as part of #36267?

[tannergooding]

I will not be. Optimizing the casts is a more general issue and one we hit on x86/x64 as well.

[tannergooding]

I can, however, see if it is relatively easy to handle and address it in a follow up PR.