$ the-wire · showcase
SIMD API churn lands as grpc gets a security bump
By RepoJournal · Filed · About Go · Composed from the cited sources · methodology
The simd and compiler work in golang/go is mostly cleanup and deprecation-shuffling, while the one change with real consequences is pkgsite-metrics moving grpc from v1.56.2 to v1.85.0 to clear a scanner-reported vulnerability.
all: update grpc to v1.85.0 golang/pkgsite-metrics
grpc moves from v1.56.2 to v1.85.0 to fix a vulnerability flagged by the security scanner. The upgrade drags a newer OpenTelemetry with it, which dropped the metric API used by golang.org/x/exp/event/otel and the GCP metric exporter, so internal/observe had to be rewritten against the new exporter API and the opentelemetry-operations-go exporters and propagator were bumped too.
simd: drop runtime feature check from variant FromArch golang/go
Generated FromArch functions on variant architectures no longer carry a runtime feature check: sequential comma-ok assertions against the base and variant bridge types already fold at compile time to the specialization midway generates. On arm64 FromArch now compiles to zero instructions and tofrom_arm64.go no longer references HasHardwareCarrylessMultiply.
simd: rename VectorBitSize to VectorBits golang/go
The simd API renames VectorBitSize to VectorBits, with the old name kept as a deprecated wrapper marked //go:fix inline so gopls' go fix rewrites callers in place. The compiler's midway rewrite now calls VectorBits directly.
go/analysis/passes/modernize: newexpr: skip arguments containing a func type golang/tools
Since the newexpr check began re-typechecking every argument with types.CheckExpr, arguments containing a func type, such as func literals, made the type checker write to the scope tree that analyzers like rangeint also read, a data race. newexpr now simply offers no fix for those calls.
hash/adler32: use current archsimd APIs golang/go
The long tail of the same SIMD cleanup: hash/adler32 drops the deprecated AsUint32x8, ExtendLo8ToUint16, and ExtendLo4ToUint32 APIs in favor of ReshapeToUint32s, ToBits, ReduceSum, ExtendToUint16, and ExtendToUint32, and the compiler learns to skip an amd64 Enlarge move when the input register is already zero-extended by a prior VEX or EVEX instruction.