runtime: separate spans of noscan objects

Currently, we mix objects with pointers and objects without pointers
("noscan" objects) together in memory. As a result, for every object
we grey, we have to check that object's heap bits to find out if it's
noscan, which adds to the per-object cost of GC. This also hurts the
TLB footprint of the garbage collector because it decreases the
density of scannable objects at the page level.

This commit improves the situation by using separate spans for noscan
objects. This will allow a much simpler noscan check (in a follow up
CL), eliminate the need to clear the bitmap of noscan objects (in a
follow up CL), and improves TLB footprint by increasing the density of
scannable objects.

This is also a step toward eliminating dead bits, since the current
noscan check depends on checking the dead bit of the first word.

This has no effect on the heap size of the garbage benchmark.

We'll measure the performance change of this after the follow-up
optimizations.

This is a cherry-pick from dev.garbage commit d491e550. The only
non-trivial merge conflict was in updatememstats in mstats.go, where
we now have to separate the per-spanclass stats from the per-sizeclass
stats.

Change-Id: I13bdc4869538ece5649a8d2a41c6605371618e40
Reviewed-on: https://go-review.googlesource.com/41251
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>

src/runtime/export_test.go

@@ -300,13 +300,13 @@ func ReadMemStatsSlow() (base, slow MemStats) {
 			if s.state != mSpanInUse {
 				continue
 			}
-			if s.sizeclass == 0 {
+			if sizeclass := s.spanclass.sizeclass(); sizeclass == 0 {
 				slow.Mallocs++
 				slow.Alloc += uint64(s.elemsize)
 			} else {
 				slow.Mallocs += uint64(s.allocCount)
 				slow.Alloc += uint64(s.allocCount) * uint64(s.elemsize)
-				bySize[s.sizeclass].Mallocs += uint64(s.allocCount)
+				bySize[sizeclass].Mallocs += uint64(s.allocCount)
 			}
 		}
 

src/runtime/malloc.go

@@ -518,8 +518,8 @@ func nextFreeFast(s *mspan) gclinkptr {
 // weight allocation. If it is a heavy weight allocation the caller must
 // determine whether a new GC cycle needs to be started or if the GC is active
 // whether this goroutine needs to assist the GC.
-func (c *mcache) nextFree(sizeclass uint8) (v gclinkptr, s *mspan, shouldhelpgc bool) {
-	s = c.alloc[sizeclass]
+func (c *mcache) nextFree(spc spanClass) (v gclinkptr, s *mspan, shouldhelpgc bool) {
+	s = c.alloc[spc]
 	shouldhelpgc = false
 	freeIndex := s.nextFreeIndex()
 	if freeIndex == s.nelems {
@@ -529,10 +529,10 @@ func (c *mcache) nextFree(sizeclass uint8) (v gclinkptr, s *mspan, shouldhelpgc
 			throw("s.allocCount != s.nelems && freeIndex == s.nelems")
 		}
 		systemstack(func() {
-			c.refill(int32(sizeclass))
+			c.refill(spc)
 		})
 		shouldhelpgc = true
-		s = c.alloc[sizeclass]
+		s = c.alloc[spc]
 
 		freeIndex = s.nextFreeIndex()
 	}
@@ -656,10 +656,10 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 				return x
 			}
 			// Allocate a new maxTinySize block.
-			span := c.alloc[tinySizeClass]
+			span := c.alloc[tinySpanClass]
 			v := nextFreeFast(span)
 			if v == 0 {
-				v, _, shouldhelpgc = c.nextFree(tinySizeClass)
+				v, _, shouldhelpgc = c.nextFree(tinySpanClass)
 			}
 			x = unsafe.Pointer(v)
 			(*[2]uint64)(x)[0] = 0
@@ -679,10 +679,11 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 				sizeclass = size_to_class128[(size-smallSizeMax+largeSizeDiv-1)/largeSizeDiv]
 			}
 			size = uintptr(class_to_size[sizeclass])
-			span := c.alloc[sizeclass]
+			spc := makeSpanClass(sizeclass, noscan)
+			span := c.alloc[spc]
 			v := nextFreeFast(span)
 			if v == 0 {
-				v, span, shouldhelpgc = c.nextFree(sizeclass)
+				v, span, shouldhelpgc = c.nextFree(spc)
 			}
 			x = unsafe.Pointer(v)
 			if needzero && span.needzero != 0 {
@@ -693,7 +694,7 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 		var s *mspan
 		shouldhelpgc = true
 		systemstack(func() {
-			s = largeAlloc(size, needzero)
+			s = largeAlloc(size, needzero, noscan)
 		})
 		s.freeindex = 1
 		s.allocCount = 1
@@ -784,7 +785,7 @@ func mallocgc(size uintptr, typ *_type, needzero bool) unsafe.Pointer {
 	return x
 }
 
-func largeAlloc(size uintptr, needzero bool) *mspan {
+func largeAlloc(size uintptr, needzero bool, noscan bool) *mspan {
 	// print("largeAlloc size=", size, "\n")
 
 	if size+_PageSize < size {
@@ -800,7 +801,7 @@ func largeAlloc(size uintptr, needzero bool) *mspan {
 	// pays the debt down to npage pages.
 	deductSweepCredit(npages*_PageSize, npages)
 
-	s := mheap_.alloc(npages, 0, true, needzero)
+	s := mheap_.alloc(npages, makeSpanClass(0, noscan), true, needzero)
 	if s == nil {
 		throw("out of memory")
 	}

src/runtime/mbitmap.go

@@ -499,6 +499,7 @@ func (h heapBits) isPointer() bool {
 // It must be told how large the object at h is for efficiency.
 // h must describe the initial word of the object.
 func (h heapBits) hasPointers(size uintptr) bool {
+	// TODO: Use span.noScan instead of the heap bitmap.
 	if size == sys.PtrSize { // 1-word objects are always pointers
 		return true
 	}

src/runtime/mcache.go

@@ -33,7 +33,8 @@ type mcache struct {
 	local_tinyallocs uintptr // number of tiny allocs not counted in other stats
 
 	// The rest is not accessed on every malloc.
-	alloc [_NumSizeClasses]*mspan // spans to allocate from
+
+	alloc [numSpanClasses]*mspan // spans to allocate from, indexed by spanClass
 
 	stackcache [_NumStackOrders]stackfreelist
 
@@ -77,7 +78,7 @@ func allocmcache() *mcache {
 	lock(&mheap_.lock)
 	c := (*mcache)(mheap_.cachealloc.alloc())
 	unlock(&mheap_.lock)
-	for i := 0; i < _NumSizeClasses; i++ {
+	for i := range c.alloc {
 		c.alloc[i] = &emptymspan
 	}
 	c.next_sample = nextSample()
@@ -103,12 +104,12 @@ func freemcache(c *mcache) {
 
 // Gets a span that has a free object in it and assigns it
 // to be the cached span for the given sizeclass. Returns this span.
-func (c *mcache) refill(sizeclass int32) *mspan {
+func (c *mcache) refill(spc spanClass) *mspan {
 	_g_ := getg()
 
 	_g_.m.locks++
 	// Return the current cached span to the central lists.
-	s := c.alloc[sizeclass]
+	s := c.alloc[spc]
 
 	if uintptr(s.allocCount) != s.nelems {
 		throw("refill of span with free space remaining")
@@ -119,7 +120,7 @@ func (c *mcache) refill(sizeclass int32) *mspan {
 	}
 
 	// Get a new cached span from the central lists.
-	s = mheap_.central[sizeclass].mcentral.cacheSpan()
+	s = mheap_.central[spc].mcentral.cacheSpan()
 	if s == nil {
 		throw("out of memory")
 	}
@@ -128,13 +129,13 @@ func (c *mcache) refill(sizeclass int32) *mspan {
 		throw("span has no free space")
 	}
 
-	c.alloc[sizeclass] = s
+	c.alloc[spc] = s
 	_g_.m.locks--
 	return s
 }
 
 func (c *mcache) releaseAll() {
-	for i := 0; i < _NumSizeClasses; i++ {
+	for i := range c.alloc {
 		s := c.alloc[i]
 		if s != &emptymspan {
 			mheap_.central[i].mcentral.uncacheSpan(s)

src/runtime/mcentral.go

@@ -19,7 +19,7 @@ import "runtime/internal/atomic"
 //go:notinheap
 type mcentral struct {
 	lock      mutex
-	sizeclass int32
+	spanclass spanClass
 	nonempty  mSpanList // list of spans with a free object, ie a nonempty free list
 	empty     mSpanList // list of spans with no free objects (or cached in an mcache)
 
@@ -30,8 +30,8 @@ type mcentral struct {
 }
 
 // Initialize a single central free list.
-func (c *mcentral) init(sizeclass int32) {
-	c.sizeclass = sizeclass
+func (c *mcentral) init(spc spanClass) {
+	c.spanclass = spc
 	c.nonempty.init()
 	c.empty.init()
 }
@@ -39,7 +39,7 @@ func (c *mcentral) init(sizeclass int32) {
 // Allocate a span to use in an MCache.
 func (c *mcentral) cacheSpan() *mspan {
 	// Deduct credit for this span allocation and sweep if necessary.
-	spanBytes := uintptr(class_to_allocnpages[c.sizeclass]) * _PageSize
+	spanBytes := uintptr(class_to_allocnpages[c.spanclass.sizeclass()]) * _PageSize
 	deductSweepCredit(spanBytes, 0)
 
 	lock(&c.lock)
@@ -225,11 +225,11 @@ func (c *mcentral) freeSpan(s *mspan, preserve bool, wasempty bool) bool {
 
 // grow allocates a new empty span from the heap and initializes it for c's size class.
 func (c *mcentral) grow() *mspan {
-	npages := uintptr(class_to_allocnpages[c.sizeclass])
-	size := uintptr(class_to_size[c.sizeclass])
+	npages := uintptr(class_to_allocnpages[c.spanclass.sizeclass()])
+	size := uintptr(class_to_size[c.spanclass.sizeclass()])
 	n := (npages << _PageShift) / size
 
-	s := mheap_.alloc(npages, c.sizeclass, false, true)
+	s := mheap_.alloc(npages, c.spanclass, false, true)
 	if s == nil {
 		return nil
 	}

src/runtime/mfinal.go

@@ -455,7 +455,7 @@ func findObject(v unsafe.Pointer) (s *mspan, x unsafe.Pointer, n uintptr) {
 	}
 
 	n = s.elemsize
-	if s.sizeclass != 0 {
+	if s.spanclass.sizeclass() != 0 {
 		x = add(x, (uintptr(v)-uintptr(x))/n*n)
 	}
 	return

src/runtime/mgcmark.go

@@ -1290,7 +1290,7 @@ func gcDumpObject(label string, obj, off uintptr) {
 		print(" s=nil\n")
 		return
 	}
-	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.sizeclass=", s.sizeclass, " s.elemsize=", s.elemsize, " s.state=")
+	print(" s.base()=", hex(s.base()), " s.limit=", hex(s.limit), " s.spanclass=", s.spanclass, " s.elemsize=", s.elemsize, " s.state=")
 	if 0 <= s.state && int(s.state) < len(mSpanStateNames) {
 		print(mSpanStateNames[s.state], "\n")
 	} else {

src/runtime/mgcsweep.go

@@ -195,7 +195,7 @@ func (s *mspan) sweep(preserve bool) bool {
 
 	atomic.Xadd64(&mheap_.pagesSwept, int64(s.npages))
 
-	cl := s.sizeclass
+	spc := s.spanclass
 	size := s.elemsize
 	res := false
 
@@ -288,7 +288,7 @@ func (s *mspan) sweep(preserve bool) bool {
 
 	// Count the number of free objects in this span.
 	nalloc := uint16(s.countAlloc())
-	if cl == 0 && nalloc == 0 {
+	if spc.sizeclass() == 0 && nalloc == 0 {
 		s.needzero = 1
 		freeToHeap = true
 	}
@@ -331,9 +331,9 @@ func (s *mspan) sweep(preserve bool) bool {
 		atomic.Store(&s.sweepgen, sweepgen)
 	}
 
-	if nfreed > 0 && cl != 0 {
-		c.local_nsmallfree[cl] += uintptr(nfreed)
-		res = mheap_.central[cl].mcentral.freeSpan(s, preserve, wasempty)
+	if nfreed > 0 && spc.sizeclass() != 0 {
+		c.local_nsmallfree[spc.sizeclass()] += uintptr(nfreed)
+		res = mheap_.central[spc].mcentral.freeSpan(s, preserve, wasempty)
 		// MCentral_FreeSpan updates sweepgen
 	} else if freeToHeap {
 		// Free large span to heap

src/runtime/mheap.go

@@ -119,7 +119,8 @@ type mheap struct {
 	// the padding makes sure that the MCentrals are
 	// spaced CacheLineSize bytes apart, so that each MCentral.lock
 	// gets its own cache line.
-	central [_NumSizeClasses]struct {
+	// central is indexed by spanClass.
+	central [numSpanClasses]struct {
 		mcentral mcentral
 		pad      [sys.CacheLineSize - unsafe.Sizeof(mcentral{})%sys.CacheLineSize]byte
 	}
@@ -260,7 +261,7 @@ type mspan struct {
 	divMul      uint16     // for divide by elemsize - divMagic.mul
 	baseMask    uint16     // if non-0, elemsize is a power of 2, & this will get object allocation base
 	allocCount  uint16     // number of allocated objects
-	sizeclass   uint8      // size class
+	spanclass   spanClass  // size class and noscan (uint8)
 	incache     bool       // being used by an mcache
 	state       mSpanState // mspaninuse etc
 	needzero    uint8      // needs to be zeroed before allocation
@@ -315,6 +316,31 @@ func recordspan(vh unsafe.Pointer, p unsafe.Pointer) {
 	h.allspans = append(h.allspans, s)
 }
 
+// A spanClass represents the size class and noscan-ness of a span.
+//
+// Each size class has a noscan spanClass and a scan spanClass. The
+// noscan spanClass contains only noscan objects, which do not contain
+// pointers and thus do not need to be scanned by the garbage
+// collector.
+type spanClass uint8
+
+const (
+	numSpanClasses = _NumSizeClasses << 1
+	tinySpanClass  = tinySizeClass<<1 | 1
+)
+
+func makeSpanClass(sizeclass uint8, noscan bool) spanClass {
+	return spanClass(sizeclass<<1) | spanClass(bool2int(noscan))
+}
+
+func (sc spanClass) sizeclass() int8 {
+	return int8(sc >> 1)
+}
+
+func (sc spanClass) noscan() bool {
+	return sc&1 != 0
+}
+
 // inheap reports whether b is a pointer into a (potentially dead) heap object.
 // It returns false for pointers into _MSpanManual spans.
 // Non-preemptible because it is used by write barriers.
@@ -399,7 +425,7 @@ func mlookup(v uintptr, base *uintptr, size *uintptr, sp **mspan) int32 {
 	}
 
 	p := s.base()
-	if s.sizeclass == 0 {
+	if s.spanclass.sizeclass() == 0 {
 		// Large object.
 		if base != nil {
 			*base = p
@@ -447,7 +473,7 @@ func (h *mheap) init(spansStart, spansBytes uintptr) {
 
 	h.busylarge.init()
 	for i := range h.central {
-		h.central[i].mcentral.init(int32(i))
+		h.central[i].mcentral.init(spanClass(i))
 	}
 
 	sp := (*slice)(unsafe.Pointer(&h.spans))
@@ -577,7 +603,7 @@ func (h *mheap) reclaim(npage uintptr) {
 
 // Allocate a new span of npage pages from the heap for GC'd memory
 // and record its size class in the HeapMap and HeapMapCache.
-func (h *mheap) alloc_m(npage uintptr, sizeclass int32, large bool) *mspan {
+func (h *mheap) alloc_m(npage uintptr, spanclass spanClass, large bool) *mspan {
 	_g_ := getg()
 	if _g_ != _g_.m.g0 {
 		throw("_mheap_alloc not on g0 stack")
@@ -617,8 +643,8 @@ func (h *mheap) alloc_m(npage uintptr, sizeclass int32, large bool) *mspan {
 		h.sweepSpans[h.sweepgen/2%2].push(s) // Add to swept in-use list.
 		s.state = _MSpanInUse
 		s.allocCount = 0
-		s.sizeclass = uint8(sizeclass)
-		if sizeclass == 0 {
+		s.spanclass = spanclass
+		if sizeclass := spanclass.sizeclass(); sizeclass == 0 {
 			s.elemsize = s.npages << _PageShift
 			s.divShift = 0
 			s.divMul = 0
@@ -670,13 +696,13 @@ func (h *mheap) alloc_m(npage uintptr, sizeclass int32, large bool) *mspan {
 	return s
 }
 
-func (h *mheap) alloc(npage uintptr, sizeclass int32, large bool, needzero bool) *mspan {
+func (h *mheap) alloc(npage uintptr, spanclass spanClass, large bool, needzero bool) *mspan {
 	// Don't do any operations that lock the heap on the G stack.
 	// It might trigger stack growth, and the stack growth code needs
 	// to be able to allocate heap.
 	var s *mspan
 	systemstack(func() {
-		s = h.alloc_m(npage, sizeclass, large)
+		s = h.alloc_m(npage, spanclass, large)
 	})
 
 	if s != nil {
@@ -710,7 +736,7 @@ func (h *mheap) allocManual(npage uintptr, stat *uint64) *mspan {
 		s.state = _MSpanManual
 		s.manualFreeList = 0
 		s.allocCount = 0
-		s.sizeclass = 0
+		s.spanclass = 0
 		s.nelems = 0
 		s.elemsize = 0
 		s.limit = s.base() + s.npages<<_PageShift
@@ -1144,7 +1170,7 @@ func (span *mspan) init(base uintptr, npages uintptr) {
 	span.startAddr = base
 	span.npages = npages
 	span.allocCount = 0
-	span.sizeclass = 0
+	span.spanclass = 0
 	span.incache = false
 	span.elemsize = 0
 	span.state = _MSpanDead

src/runtime/mstats.go

@@ -552,7 +552,18 @@ func updatememstats() {
 	// Collect allocation stats. This is safe and consistent
 	// because the world is stopped.
 	var smallFree, totalAlloc, totalFree uint64
-	for i := range mheap_.central {
+	// Collect per-spanclass stats.
+	for spc := range mheap_.central {
+		// The mcaches are now empty, so mcentral stats are
+		// up-to-date.
+		c := &mheap_.central[spc].mcentral
+		memstats.nmalloc += c.nmalloc
+		i := spanClass(spc).sizeclass()
+		memstats.by_size[i].nmalloc += c.nmalloc
+		totalAlloc += c.nmalloc * uint64(class_to_size[i])
+	}
+	// Collect per-sizeclass stats.
+	for i := 0; i < _NumSizeClasses; i++ {
 		if i == 0 {
 			memstats.nmalloc += mheap_.nlargealloc
 			totalAlloc += mheap_.largealloc
@@ -560,12 +571,6 @@ func updatememstats() {
 			memstats.nfree += mheap_.nlargefree
 			continue
 		}
-		// The mcaches are now empty, so mcentral stats are
-		// up-to-date.
-		c := &mheap_.central[i].mcentral
-		memstats.nmalloc += c.nmalloc
-		memstats.by_size[i].nmalloc += c.nmalloc
-		totalAlloc += c.nmalloc * uint64(class_to_size[i])
 
 		// The mcache stats have been flushed to mheap_.
 		memstats.nfree += mheap_.nsmallfree[i]

src/runtime/stubs.go

@@ -291,3 +291,10 @@ func checkASM() bool
 
 func memequal_varlen(a, b unsafe.Pointer) bool
 func eqstring(s1, s2 string) bool
+
+// bool2int returns 0 if x is false or 1 if x is true.
+func bool2int(x bool) int {
+	// Avoid branches. In the SSA compiler, this compiles to
+	// exactly what you would want it to.
+	return int(uint8(*(*uint8)(unsafe.Pointer(&x))))
+}