You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The change to support carry flag optimization on ARM64 (go.dev/cl/432275) caused a large regression in the crypto/ed25519 benchmarks. I missed this regression when reviewing the patch.
The optimization created by using the CarryChainTail priority fails when schedulable memory ops get intertwined with schedulable CarryChainTail ops like ADDZ on PPC64. This leads to many avoidable CA bit clobbers.
pmur
changed the title
cmd/compile: performance regression compiling crypto/internal/edwards25519/field.feMulGeneric
cmd/compile: performance regression compiling crypto/internal/edwards25519/field.feMulGeneric on PPC64
Oct 31, 2022
The change to support carry flag optimization on ARM64 (go.dev/cl/432275) caused a large regression in the crypto/ed25519 benchmarks. I missed this regression when reviewing the patch.
The optimization created by using the CarryChainTail priority fails when schedulable memory ops get intertwined with schedulable CarryChainTail ops like ADDZ on PPC64. This leads to many avoidable CA bit clobbers.
What did you expect to see?
No difference in benchmark tests.
What did you see instead?
The text was updated successfully, but these errors were encountered: