67 lines
1.4 KiB
Go
67 lines
1.4 KiB
Go
|
package xsync
|
||
|
|
|
||
|
|
import (
|
||
|
|
"math/bits"
|
||
|
|
"runtime"
|
||
|
|
_ "unsafe"
|
||
|
|
)
|
||
|
|
|
||
|
|
// test-only assert()-like flag
|
||
|
|
var assertionsEnabled = false
|
||
|
|
|
||
|
|
const (
|
||
|
|
// cacheLineSize is used in paddings to prevent false sharing;
|
||
|
|
// 64B are used instead of 128B as a compromise between
|
||
|
|
// memory footprint and performance; 128B usage may give ~30%
|
||
|
|
// improvement on NUMA machines.
|
||
|
|
cacheLineSize = 64
|
||
|
|
)
|
||
|
|
|
||
|
|
// nextPowOf2 computes the next highest power of 2 of 32-bit v.
|
||
|
|
// Source: https://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2
|
||
|
|
func nextPowOf2(v uint32) uint32 {
|
||
|
|
if v == 0 {
|
||
|
|
return 1
|
||
|
|
}
|
||
|
|
v--
|
||
|
|
v |= v >> 1
|
||
|
|
v |= v >> 2
|
||
|
|
v |= v >> 4
|
||
|
|
v |= v >> 8
|
||
|
|
v |= v >> 16
|
||
|
|
v++
|
||
|
|
return v
|
||
|
|
}
|
||
|
|
|
||
|
|
func parallelism() uint32 {
|
||
|
|
maxProcs := uint32(runtime.GOMAXPROCS(0))
|
||
|
|
numCores := uint32(runtime.NumCPU())
|
||
|
|
if maxProcs < numCores {
|
||
|
|
return maxProcs
|
||
|
|
}
|
||
|
|
return numCores
|
||
|
|
}
|
||
|
|
|
||
|
|
//go:noescape
|
||
|
|
//go:linkname runtime_fastrand runtime.fastrand
|
||
|
|
func runtime_fastrand() uint32
|
||
|
|
|
||
|
|
func broadcast(b uint8) uint64 {
|
||
|
|
return 0x101010101010101 * uint64(b)
|
||
|
|
}
|
||
|
|
|
||
|
|
func firstMarkedByteIndex(w uint64) int {
|
||
|
|
return bits.TrailingZeros64(w) >> 3
|
||
|
|
}
|
||
|
|
|
||
|
|
// SWAR byte search: may produce false positives, e.g. for 0x0100,
|
||
|
|
// so make sure to double-check bytes found by this function.
|
||
|
|
func markZeroBytes(w uint64) uint64 {
|
||
|
|
return ((w - 0x0101010101010101) & (^w) & 0x8080808080808080)
|
||
|
|
}
|
||
|
|
|
||
|
|
func setByte(w uint64, b uint8, idx int) uint64 {
|
||
|
|
shift := idx << 3
|
||
|
|
return (w &^ (0xff << shift)) | (uint64(b) << shift)
|
||
|
|
}
|