9bd0164237
Improve sync.SeqCount performance.
...
- Make sync.SeqCountEpoch not a struct. This allows sync.SeqCount.BeginRead()
to be inlined.
- Mark sync.SeqAtomicLoad<T> nosplit to mitigate the Go compiler's refusal to
inline it. (Best I could get was "cost 92 exceeds budget 80".)
- Use runtime-guided spinning in SeqCount.BeginRead().
Benchmarks:
name old time/op new time/op delta
pkg:pkg/sync/sync goos:linux goarch:amd64
SeqCountWriteUncontended-12 8.24ns ± 0% 11.40ns ± 0% +38.35% (p=0.000 n=10+10)
SeqCountReadUncontended-12 0.33ns ± 0% 0.14ns ± 3% -57.77% (p=0.000 n=7+8)
pkg:pkg/sync/seqatomictest/seqatomic goos:linux goarch:amd64
SeqAtomicLoadIntUncontended-12 0.64ns ± 1% 0.41ns ± 1% -36.40% (p=0.000 n=10+8)
SeqAtomicTryLoadIntUncontended-12 0.18ns ± 4% 0.18ns ± 1% ~ (p=0.206 n=10+8)
AtomicValueLoadIntUncontended-12 0.27ns ± 3% 0.27ns ± 0% -1.77% (p=0.000 n=10+8)
(atomic.Value.Load is, of course, inlined. We would expect an uncontended
inline SeqAtomicLoad<int> to perform identically to SeqAtomicTryLoad<int>.) The
"regression" in BenchmarkSeqCountWriteUncontended, despite this CL changing
nothing in that path, is attributed to microarchitectural subtlety; the
benchmark loop is unchanged except for its address:
Before this CL:
:0 0x4e62d1 48ffc2 INCQ DX
:0 0x4e62d4 48399110010000 CMPQ DX, 0x110(CX)
:0 0x4e62db 7e26 JLE 0x4e6303
:0 0x4e62dd 90 NOPL
:0 0x4e62de bb01000000 MOVL $0x1, BX
:0 0x4e62e3 f00fc118 LOCK XADDL BX, 0(AX)
:0 0x4e62e7 ffc3 INCL BX
:0 0x4e62e9 0fbae300 BTL $0x0, BX
:0 0x4e62ed 733a JAE 0x4e6329
:0 0x4e62ef 90 NOPL
:0 0x4e62f0 bb01000000 MOVL $0x1, BX
:0 0x4e62f5 f00fc118 LOCK XADDL BX, 0(AX)
:0 0x4e62f9 ffc3 INCL BX
:0 0x4e62fb 0fbae300 BTL $0x0, BX
:0 0x4e62ff 73d0 JAE 0x4e62d1
After this CL:
:0 0x4e6361 48ffc2 INCQ DX
:0 0x4e6364 48399110010000 CMPQ DX, 0x110(CX)
:0 0x4e636b 7e26 JLE 0x4e6393
:0 0x4e636d 90 NOPL
:0 0x4e636e bb01000000 MOVL $0x1, BX
:0 0x4e6373 f00fc118 LOCK XADDL BX, 0(AX)
:0 0x4e6377 ffc3 INCL BX
:0 0x4e6379 0fbae300 BTL $0x0, BX
:0 0x4e637d 733a JAE 0x4e63b9
:0 0x4e637f 90 NOPL
:0 0x4e6380 bb01000000 MOVL $0x1, BX
:0 0x4e6385 f00fc118 LOCK XADDL BX, 0(AX)
:0 0x4e6389 ffc3 INCL BX
:0 0x4e638b 0fbae300 BTL $0x0, BX
:0 0x4e638f 73d0 JAE 0x4e6361
PiperOrigin-RevId: 329754148
2020-09-02 11:37:31 -07:00
27500d529f
New sync package.
...
* Rename syncutil to sync.
* Add aliases to sync types.
* Replace existing usage of standard library sync package.
This will make it easier to swap out synchronization primitives. For example,
this will allow us to use primitives from github.com/sasha-s/go-deadlock to
check for lock ordering violations.
Updates #1472
PiperOrigin-RevId: 289033387
2020-01-09 22:02:24 -08:00
Jamie Liu
Ian Gudger
Michael Pratt
Googler