Swtich from a sync.RWMutex to atomic.Value for cacheTypeFields.
On GOARCH=386, this recovers most of the remaining performance
difference from the 1.6 release. Compared with tip on linux/386:
name old time/op new time/op delta
CodeDecoder-40 92.8ms ± 1% 87.7ms ± 1% -5.50% (p=0.000 n=10+10)
name old speed new speed delta
CodeDecoder-40 20.9MB/s ± 1% 22.1MB/s ± 1% +5.83% (p=0.000 n=10+10)
With more time and care, I believe more of the JSON decoder's work
could be shifted so it is done before decoding, and independent of
the number of bytes processed. Maybe someone could explore that for
Go 1.8.
For #16117.
Change-Id: I049655b2e5b76384a0d5f4b90e3ec7cc8d8c4340
Reviewed-on: https://go-review.googlesource.com/24472
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
"strconv"
"strings"
"sync"
+ "sync/atomic"
"unicode"
"unicode/utf8"
)
}
var fieldCache struct {
- sync.RWMutex
- m map[reflect.Type][]field
+ value atomic.Value // map[reflect.Type][]field
+ mu sync.Mutex // used only by writers
}
// cachedTypeFields is like typeFields but uses a cache to avoid repeated work.
func cachedTypeFields(t reflect.Type) []field {
- fieldCache.RLock()
- f := fieldCache.m[t]
- fieldCache.RUnlock()
+ m, _ := fieldCache.value.Load().(map[reflect.Type][]field)
+ f := m[t]
if f != nil {
return f
}
f = []field{}
}
- fieldCache.Lock()
- if fieldCache.m == nil {
- fieldCache.m = map[reflect.Type][]field{}
+ fieldCache.mu.Lock()
+ m, _ = fieldCache.value.Load().(map[reflect.Type][]field)
+ newM := make(map[reflect.Type][]field, len(m)+1)
+ for k, v := range m {
+ newM[k] = v
}
- fieldCache.m[t] = f
- fieldCache.Unlock()
+ newM[t] = f
+ fieldCache.value.Store(newM)
+ fieldCache.mu.Unlock()
return f
}