diff options
Diffstat (limited to 'vendor/google.golang.org/protobuf/encoding')
| -rw-r--r-- | vendor/google.golang.org/protobuf/encoding/protowire/wire.go | 26 |
1 files changed, 25 insertions, 1 deletions
diff --git a/vendor/google.golang.org/protobuf/encoding/protowire/wire.go b/vendor/google.golang.org/protobuf/encoding/protowire/wire.go index e942bc983..743bfb81d 100644 --- a/vendor/google.golang.org/protobuf/encoding/protowire/wire.go +++ b/vendor/google.golang.org/protobuf/encoding/protowire/wire.go @@ -371,7 +371,31 @@ func ConsumeVarint(b []byte) (v uint64, n int) { func SizeVarint(v uint64) int { // This computes 1 + (bits.Len64(v)-1)/7. // 9/64 is a good enough approximation of 1/7 - return int(9*uint32(bits.Len64(v))+64) / 64 + // + // The Go compiler can translate the bits.LeadingZeros64 call into the LZCNT + // instruction, which is very fast on CPUs from the last few years. The + // specific way of expressing the calculation matches C++ Protobuf, see + // https://godbolt.org/z/4P3h53oM4 for the C++ code and how gcc/clang + // optimize that function for GOAMD64=v1 and GOAMD64=v3 (-march=haswell). + + // By OR'ing v with 1, we guarantee that v is never 0, without changing the + // result of SizeVarint. LZCNT is not defined for 0, meaning the compiler + // needs to add extra instructions to handle that case. + // + // The Go compiler currently (go1.24.4) does not make use of this knowledge. + // This opportunity (removing the XOR instruction, which handles the 0 case) + // results in a small (1%) performance win across CPU architectures. + // + // Independently of avoiding the 0 case, we need the v |= 1 line because + // it allows the Go compiler to eliminate an extra XCHGL barrier. + v |= 1 + + // It would be clearer to write log2value := 63 - uint32(...), but + // writing uint32(...) ^ 63 is much more efficient (-14% ARM, -20% Intel). + // Proof of identity for our value range [0..63]: + // https://go.dev/play/p/Pdn9hEWYakX + log2value := uint32(bits.LeadingZeros64(v)) ^ 63 + return int((log2value*9 + (64 + 9)) / 64) } // AppendFixed32 appends v to b as a little-endian uint32. |