runtime: process ptr bitmaps one word at a time

Read the bitmaps one uintptr at a time instead of one byte at a time.

Performance so far:
 Allocation heavy, no retention: ~30% faster in heapBitsSetType
 Scan heavy, ~no allocation: ~even in scanobject

Change-Id: I40d492b50d7f89d1b4261c2de58f6d255fa5e93e
Reviewed-on: https://go-review.googlesource.com/c/go/+/407036
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
This commit is contained in:
Keith Randall 2022-05-06 18:40:17 -07:00 committed by Keith Randall
parent b589208c8c
commit c3833a5543
3 changed files with 91 additions and 17 deletions

View file

@ -1549,7 +1549,11 @@ func dgcsym(t *types.Type, write bool) (lsym *obj.LSym, useGCProg bool, ptrdata
// dgcptrmask emits and returns the symbol containing a pointer mask for type t.
func dgcptrmask(t *types.Type, write bool) *obj.LSym {
ptrmask := make([]byte, (types.PtrDataSize(t)/int64(types.PtrSize)+7)/8)
// Bytes we need for the ptrmask.
n := (types.PtrDataSize(t)/int64(types.PtrSize) + 7) / 8
// Runtime wants ptrmasks padded to a multiple of uintptr in size.
n = (n + int64(types.PtrSize) - 1) &^ (int64(types.PtrSize) - 1)
ptrmask := make([]byte, n)
fillptrmask(t, ptrmask)
p := fmt.Sprintf("runtime.gcbits.%x", ptrmask)

View file

@ -2265,7 +2265,10 @@ func bucketOf(ktyp, etyp *rtype) *rtype {
if ktyp.ptrdata != 0 || etyp.ptrdata != 0 {
nptr := (bucketSize*(1+ktyp.size+etyp.size) + goarch.PtrSize) / goarch.PtrSize
mask := make([]byte, (nptr+7)/8)
n := (nptr + 7) / 8
// Runtime needs pointer masks to be a multiple of uintptr in size.
n = (n + goarch.PtrSize - 1) &^ (goarch.PtrSize - 1)
mask := make([]byte, n)
base := bucketSize / goarch.PtrSize
if ktyp.ptrdata != 0 {
@ -2971,7 +2974,10 @@ func ArrayOf(length int, elem Type) Type {
// Element is small with pointer mask; array is still small.
// Create direct pointer mask by turning each 1 bit in elem
// into length 1 bits in larger mask.
mask := make([]byte, (array.ptrdata/goarch.PtrSize+7)/8)
n := (array.ptrdata/goarch.PtrSize + 7) / 8
// Runtime needs pointer masks to be a multiple of uintptr in size.
n = (n + goarch.PtrSize - 1) &^ (goarch.PtrSize - 1)
mask := make([]byte, n)
emitGCMask(mask, 0, typ, array.len)
array.gcdata = &mask[0]
@ -3140,8 +3146,13 @@ type bitVector struct {
// append a bit to the bitmap.
func (bv *bitVector) append(bit uint8) {
if bv.n%8 == 0 {
bv.data = append(bv.data, 0)
if bv.n%(8*goarch.PtrSize) == 0 {
// Runtime needs pointer masks to be a multiple of uintptr in size.
// Since reflect passes bv.data directly to the runtime as a pointer mask,
// we append a full uintptr of zeros at a time.
for i := 0; i < goarch.PtrSize; i++ {
bv.data = append(bv.data, 0)
}
}
bv.data[bv.n/8] |= bit << (bv.n % 8)
bv.n++

View file

@ -893,6 +893,19 @@ func (h writeHeapBits) flush(addr, size uintptr) {
}
}
// Read the bytes starting at the aligned pointer p into a uintptr.
// Read is little-endian.
func readUintptr(p *byte) uintptr {
x := *(*uintptr)(unsafe.Pointer(p))
if goarch.BigEndian {
if goarch.PtrSize == 8 {
return uintptr(sys.Bswap64(uint64(x)))
}
return uintptr(sys.Bswap32(uint32(x)))
}
return x
}
// heapBitsSetType records that the new allocation [x, x+size)
// holds in [x, x+dataSize) one or more values of type typ.
// (The number of values is given by dataSize / typ.size.)
@ -917,7 +930,7 @@ func (h writeHeapBits) flush(addr, size uintptr) {
// machines, callers must execute a store/store (publication) barrier
// between calling this function and making the object reachable.
func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
const doubleCheck = true // slow but helpful; enable to test modifications to this code
const doubleCheck = false // slow but helpful; enable to test modifications to this code
if doubleCheck && dataSize%typ.size != 0 {
throw("heapBitsSetType: dataSize not a multiple of typ.size")
@ -995,19 +1008,65 @@ func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
// objects with scalar tails, all but the last tail does have to
// be initialized, because there is no way to say "skip forward".
for i := uintptr(0); true; i += typ.size {
p := typ.gcdata
var j uintptr
for j = 0; j+8*goarch.PtrSize < typ.ptrdata; j += 8 * goarch.PtrSize {
h = h.write(uintptr(*p), 8)
p = add1(p)
ptrs := typ.ptrdata / goarch.PtrSize
if typ.size == dataSize { // Single element
if ptrs <= ptrBits { // Single small element
m := readUintptr(typ.gcdata)
h = h.write(m, ptrs)
} else { // Single large element
p := typ.gcdata
for {
h = h.write(readUintptr(p), ptrBits)
p = addb(p, ptrBits/8)
ptrs -= ptrBits
if ptrs <= ptrBits {
break
}
}
m := readUintptr(p)
h = h.write(m, ptrs)
}
h = h.write(uintptr(*p), (typ.ptrdata-j)/goarch.PtrSize)
if i+typ.size == dataSize {
break // don't need the trailing nonptr bits on the last element.
} else { // Repeated element
words := typ.size / goarch.PtrSize // total words, including scalar tail
if words <= ptrBits { // Repeated small element
n := dataSize / typ.size
m := readUintptr(typ.gcdata)
// Make larger unit to repeat
for words <= ptrBits/2 {
if n&1 != 0 {
h = h.write(m, words)
}
n /= 2
m |= m << words
ptrs += words
words *= 2
if n == 1 {
break
}
}
for n > 1 {
h = h.write(m, words)
n--
}
h = h.write(m, ptrs)
} else { // Repeated large element
for i := uintptr(0); true; i += typ.size {
p := typ.gcdata
j := ptrs
for j > ptrBits {
h = h.write(readUintptr(p), ptrBits)
p = addb(p, ptrBits/8)
j -= ptrBits
}
m := readUintptr(p)
h = h.write(m, j)
if i+typ.size == dataSize {
break // don't need the trailing nonptr bits on the last element.
}
// Pad with zeros to the start of the next element.
h = h.pad(typ.size - typ.ptrdata)
}
}
// Pad with zeros to the start of the next element.
h = h.pad(typ.size - typ.ptrdata)
}
h.flush(x, size)