perf(mcache): no alloc (#227)

goos: darwin
goarch: arm64
pkg: github.com/bytedance/gopkg/lang/mcache
                      │   ./old.out   │              ./new.out              │
                      │    sec/op     │    sec/op     vs base               │
MCache4096-12           26.270n ±  2%   9.533n ±  1%  -63.71% (p=0.002 n=6)
MCache10M-12            25.885n ±  2%   9.486n ±  0%  -63.36% (p=0.002 n=6)
MCache4096Parallel-12    105.3µ ± 17%   113.4µ ±  7%        ~ (p=0.065 n=6)
MCache10MParallel-12     123.7µ ±  5%   120.8µ ± 17%        ~ (p=0.310 n=6)
geomean                  1.725µ         1.055µ        -38.83%

                      │    ./old.out    │               ./new.out                │
                      │      B/op       │    B/op     vs base                    │
MCache4096-12               24.00 ±  0%    0.00 ± 0%  -100.00% (p=0.002 n=6)
MCache10M-12                27.50 ±  5%    0.00 ± 0%  -100.00% (p=0.002 n=6)
MCache4096Parallel-12     234.7Ki ± 23%   0.0Ki ± 0%  -100.00% (p=0.002 n=6)
MCache10MParallel-12    804424.50 ±  3%   74.50 ±  ?   -99.99% (p=0.002 n=6)
geomean                   3.282Ki                     ?                      ¹ ²
¹ summaries must be >0 to compute geomean
² ratios must be >0 to compute geomean

                      │  ./old.out   │               ./new.out                │
                      │  allocs/op   │ allocs/op   vs base                    │
MCache4096-12            1.000 ±  0%   0.000 ± 0%  -100.00% (p=0.002 n=6)
MCache10M-12             1.000 ±  0%   0.000 ± 0%  -100.00% (p=0.002 n=6)
MCache4096Parallel-12   10.00k ± 23%   0.00k ± 0%  -100.00% (p=0.002 n=6)
MCache10MParallel-12    32.48k ±  4%   0.00k ± 0%  -100.00% (p=0.002 n=6)
geomean                  134.3                     ?                      ¹ ²
¹ summaries must be >0 to compute geomean
² ratios must be >0 to compute geomean

lang/mcache/mcache.go

@@ -16,6 +16,7 @@ package mcache
 
 import (
  "sync"
+ "unsafe"
 
  "github.com/bytedance/gopkg/lang/dirtmake"
 )
@@ -25,12 +26,19 @@ const maxSize = 46
 // index contains []byte which cap is 1<<index
 var caches [maxSize]sync.Pool
 
+type bytesHeader struct {
+ Data *byte
+ Len int
+ Cap int
+}
+
 func init() {
  for i := 0; i < maxSize; i++ {
  size := 1 << i
  caches[i].New = func() interface{} {
  buf := dirtmake.Bytes(0, size)
- return buf
+ h := (*bytesHeader)(unsafe.Pointer(&buf))
+ return h.Data
  }
  }
 }
@@ -57,8 +65,14 @@ func Malloc(size int, capacity ...int) []byte {
  if len(capacity) > 0 && capacity[0] > size {
  c = capacity[0]
  }
- var ret = caches[calcIndex(c)].Get().([]byte)
- ret = ret[:size]
+
+ i := calcIndex(c)
+
+ ret := []byte{}
+ h := (*bytesHeader)(unsafe.Pointer(&ret))
+ h.Len = size
+ h.Cap = 1 << i
+ h.Data = caches[i].Get().(*byte)
  return ret
 }
 
@@ -68,6 +82,6 @@ func Free(buf []byte) {
  if !isPowerOfTwo(size) {
  return
  }
- buf = buf[:0]
- caches[bsr(size)].Put(buf)
+ h := (*bytesHeader)(unsafe.Pointer(&buf))
+ caches[bsr(size)].Put(h.Data)
 }

lang/mcache/utils.go

@@ -21,5 +21,5 @@ func bsr(x int) int {
 }
 
 func isPowerOfTwo(x int) bool {
- return (x & (-x)) == x
+ return (x != 0) && ((x & (-x)) == x)
 }