Repository navigation
[Mono] Intrinsify Vector128:ExtractMostSignificantBits on ARM64 and AMD64 #76025
Description
Activity
Did a little checking, on x86 we can use
_mm_movemask_epi8/ps/pd, these are SSE and SSE2, so this is perfectly fine with x86-64. There is no such operation for int16 - if we want it, we can_mm_shuffle_epi8so that odd bytes are packed in the lower half of the reg, then extract the byte mask and keep only the lower 8 bits of that. That requires SSSE3 and one load for the shuffle indices. It seems CoreCLR does it this way also.For arm64, there is no instruction to extract mask. CoreCLR does this by
anding away all but the MSB from each element, shift the MSBs andaddvthem to form the mask. This seems like the best option right now.Reacted by Ivan Povazan- added a commit that references this issue
on Apr 14, 2023 cc @matouskozak for further work.
Reacted by Ivan PovazanClosing this issue as Vector128::ExtractMostSignificantBits intrinsics support is complete both for miniJIT and AOT (LLVM) on ARM64 and AMD64. Remaining work for Vector64::ExtractMostSignificantBits is tracked in #90402.
- ghost locked as resolved and limited conversation to collaborators
on Sep 10, 2023
WASM has added intrinsics support for
Vector128:ExtractMostSignificantBits, where they saw big performance gain on a few Span and Json API's.Implement intrinsics support for
ExtractMostSignificantBitson:NOTE: Additionally, investigate if
Vector64variant could be easily handled by the change.