Arm64 SVE: re-enable use of predicate variants #116854
Conversation
Fixes dotnet#101970 in dotnet#115566 Adds a simple costing to fgMorphTryUseAllMaskVariant() and assumes nodes can always be converted to masks (using ConvertVectorToMask).
dotnet-policy-service
bot
added
the
community-contribution
github-actions
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
| @@ -1993,6 +1993,104 @@ SveMaskPattern EvaluateSimdMaskToPattern(var_types baseType, simdmask_t arg0) | |||
| } | |||
| } | |||
| } | |||
|
|
|||
| template <typename TSimd, typename TBase> | |||
There was a problem hiding this comment.
Same changes to this file also exist in #116852. I'll remove from one PR when the other is merged.
| // depending on if the corresponding input element | ||
| // has its least significant bit set | ||
|
|
||
| // For Arm64 we have count total bits to write, but they are sizeof(TBase) |
There was a problem hiding this comment.
These changes to EvaluateSimdCvtVectorToMask() fix the fuzzlyn errors in HEAD. In jitted code we use CMPNE against 0 to convert from vector to mask, therefore we need to match that here.
I've also made the same changes to the new function EvaluateSimdVectorToPattern()
|
/azp run Antigen,Fuzzlyn |
|
Azure Pipelines successfully started running 2 pipeline(s). |
|
Looks like Antigen and Fuzzlyn are still not happy |
|
Discussed offline: The fuzzlyn errors are mostly due to there being more opportunities for conversions between masks and vectors and there being issues in the way we handle that. We should pause this PR until we get that sorted out first. |
Fixes #101970
Predicate variants were implemented, and then turned off in #115566.
Adds a simple costing to fgMorphTryUseAllMaskVariant() and assumes nodes can always be converted to masks (using ConvertVectorToMask).