Thanks Peter, this is very interesting. Indeed I would not have anticipated (though it makes some sense) that calling append would outperform the manual loop, when the prefix is big, despite the overhead of a builtin func call.
On Tuesday, August 22, 2017 at 7:22:42 PM UTC+2, peterGo wrote: > > Val, > > That's a lot of speculation! > > The original benchmark applies to the original question: > > prefix := verylongstring[:3] > > If we change the parameters of the benchmark then we expect to get > different results. For example, read the Go gc compiler code. There's a > stack/heap optimization for 32 bytes or less. > > cmd/gc: allocate buffers for non-escaped strings on stack > commit e6fac08146df323eb95f46508bef937cdfb802fd > https://go.googlesource.com/go/+/e6fac08146df323eb95f46508bef937cdfb802fd > https://go-review.googlesource.com/c/go/+/3120 > > 3 byte prefix: > > BenchmarkNil-4 10000000 173 ns/op 16 B/op > 2 allocs/op > BenchmarkLiteral-4 10000000 171 ns/op 16 B/op > 2 allocs/op > BenchmarkConvert-4 20000000 76.9 ns/op 3 B/op > 1 allocs/op > BenchmarkCopy-4 20000000 67.1 ns/op 3 B/op > 1 allocs/op > BenchmarkLoop-4 20000000 65.7 ns/op 3 B/op > 1 allocs/op > > 32 byte prefix: > > BenchmarkNil-4 5000000 235 ns/op 64 B/op > 2 allocs/op > BenchmarkLiteral-4 10000000 232 ns/op 64 B/op > 2 allocs/op > BenchmarkConvert-4 10000000 129 ns/op 32 B/op > 1 allocs/op > BenchmarkCopy-4 10000000 118 ns/op 32 B/op > 1 allocs/op > BenchmarkLoop-4 10000000 183 ns/op 32 B/op > 1 allocs/op > > 33 byte prefix: > > BenchmarkNil-4 5000000 295 ns/op 96 B/op > 2 allocs/op > BenchmarkLiteral-4 5000000 285 ns/op 96 B/op > 2 allocs/op > BenchmarkConvert-4 5000000 251 ns/op 96 B/op > 2 allocs/op > BenchmarkCopy-4 10000000 143 ns/op 48 B/op > 1 allocs/op > BenchmarkLoop-4 10000000 188 ns/op 48 B/op > 1 allocs/op > > 256 byte prefix: > > BenchmarkNil-4 2000000 661 ns/op 512 B/op > 2 allocs/op > BenchmarkLiteral-4 2000000 665 ns/op 512 B/op > 2 allocs/op > BenchmarkConvert-4 2000000 659 ns/op 512 B/op > 2 allocs/op > BenchmarkCopy-4 5000000 369 ns/op 256 B/op > 1 allocs/op > BenchmarkLoop-4 2000000 883 ns/op 256 B/op > 1 allocs/op > > $ go version > go version devel +33484a6 Tue Aug 22 08:09:42 2017 +0000 linux/amd64 > > $ go test -run=! -bench=. -benchmem strslice_test.go > goos: linux > goarch: amd64 > > https://play.golang.org/p/tAYyecoQXY > > $ cat strslice_test.go > package main > > import ( > "strings" > "testing" > ) > > const pfxLen = 3 // 3, 32, 33, 256 > > var ( > s = strings.Repeat("a very, very long string", 4096) > prefix string > ) > > func BenchmarkNil(b *testing.B) { > for i := 0; i < b.N; i++ { > prefix = string(append([]byte(nil), s[:pfxLen]...)) > } > } > > func BenchmarkLiteral(b *testing.B) { > for i := 0; i < b.N; i++ { > prefix = string(append([]byte{}, s[:pfxLen]...)) > } > } > > func BenchmarkConvert(b *testing.B) { > for i := 0; i < b.N; i++ { > prefix = string([]byte(s[:pfxLen])) > } > } > > func BenchmarkCopy(b *testing.B) { > for i := 0; i < b.N; i++ { > buffer := make([]byte, pfxLen) > copy(buffer, s) > prefix = string(buffer) > } > } > > func BenchmarkLoop(b *testing.B) { > for i := 0; i < b.N; i++ { > buffer := make([]byte, pfxLen) > for i := 0; i < len(buffer); i++ { > buffer[i] = s[i] > } > prefix = string(buffer) > } > } > $ > > Peter > > > On Tuesday, August 22, 2017 at 8:07:01 AM UTC-4, Val wrote: >> >> FWIW, append is most often a small performance penalty when the number of >> elements is known ahead. >> And for some reason, using built-in func copy, or an explicit loop, is >> slightly faster on my workstation than BenchmarkConvert. >> Also, "small benchmarks are hard" so I suspect the small allocations in >> tight benchmark loops may not be representative of their real cost in a >> real program. >> >> >> func BenchmarkCopy(b *testing.B) { >> for i := 0; i < b.N; i++ { >> buffer := make([]byte, 3) >> copy(buffer, s[:3]) >> prefix = string(buffer) >> } >> } >> >> func BenchmarkLoop(b *testing.B) { >> for i := 0; i < b.N; i++ { >> buffer := make([]byte, 3) >> for i := 0; i < 3; i++ { >> buffer[i] = s[i] >> } >> prefix = string(buffer) >> } >> } >> >> >> BenchmarkNil-4 30000000 61.9 ns/op 16 >> B/op 2 allocs/op >> BenchmarkLiteral-4 30000000 59.6 ns/op 16 >> B/op 2 allocs/op >> BenchmarkConvert-4 50000000 37.4 ns/op 3 >> B/op 1 allocs/op >> BenchmarkCopy-4 50000000 30.8 ns/op 3 >> B/op 1 allocs/op >> BenchmarkLoop-4 50000000 29.3 ns/op 3 >> B/op 1 allocs/op >> >> Cheers >> Val >> >> On Tuesday, August 22, 2017 at 1:01:36 PM UTC+2, peterGo wrote: >>> >>> A benchmark; >>> >>> https://play.golang.org/p/oUyeldDG5Q >>> >>> $ cat strslice_test.go >>> package main >>> >>> import ( >>> "strings" >>> "testing" >>> ) >>> >>> var ( >>> s = strings.Repeat("a very, very long string", 4096) >>> prefix string >>> ) >>> >>> func BenchmarkNil(b *testing.B) { >>> for i := 0; i < b.N; i++ { >>> prefix = string(append([]byte(nil), s[:3]...)) >>> } >>> } >>> >>> func BenchmarkLiteral(b *testing.B) { >>> for i := 0; i < b.N; i++ { >>> prefix = string(append([]byte{}, s[:3]...)) >>> } >>> } >>> >>> func BenchmarkConvert(b *testing.B) { >>> for i := 0; i < b.N; i++ { >>> prefix = string([]byte(s[:3])) >>> } >>> } >>> >>> $ go test -run=! -bench=. -benchmem strslice_test.go >>> goos: linux >>> goarch: amd64 >>> BenchmarkNil-4 10000000 174 ns/op 16 >>> B/op 2 allocs/op >>> BenchmarkLiteral-4 10000000 174 ns/op 16 >>> B/op 2 allocs/op >>> BenchmarkConvert-4 20000000 77.0 ns/op 3 >>> B/op 1 allocs/op >>> PASS >>> ok command-line-arguments 5.512s >>> $ >>> >>> Peter >>> >>> On Tuesday, August 22, 2017 at 12:15:44 AM UTC-4, Tamás Gulácsi wrote: >>>> >>>> prefix := string([]byte(verylongstring[:3])) >>> >>> -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
