cmd/compile, runtime: emit only GC data for stack objects

Currently, for stack objects, the compiler emits metadata that
includes the offset and type descriptor for each object. The type
descriptor symbol has many fields, and it references many other
symbols, e.g. field/element types, equality functions, names.

Observe that what we actually need at runtime is only the GC
metadata that are needed to scan the object, and the GC metadata
are "leaf" symbols (which doesn't reference other symbols). Emit
only the GC data instead. This avoids bringing live the type
descriptor as well as things referenced by it (if it is not
otherwise live).

This reduces binary sizes:

                     old          new
hello (println)    1187776      1133856 (-4.5%)
hello (fmt)        1902448      1844416 (-3.1%)
cmd/compile       22670432     22438576 (-1.0%)
cmd/link           6346272      6225408 (-1.9%)

No significant change in compiler speed.

name        old time/op       new time/op       delta
Template          184ms ± 2%        186ms ± 5%    ~     (p=0.905 n=9+10)
Unicode          78.4ms ± 5%       76.3ms ± 3%  -2.60%  (p=0.009 n=10+10)
GoTypes           1.09s ± 1%        1.08s ± 1%  -0.73%  (p=0.027 n=10+8)
Compiler         85.6ms ± 3%       84.6ms ± 4%    ~     (p=0.143 n=10+10)
SSA               7.23s ± 1%        7.25s ± 1%    ~     (p=0.780 n=10+9)
Flate             116ms ± 5%        115ms ± 6%    ~     (p=0.912 n=10+10)
GoParser          201ms ± 4%        195ms ± 1%    ~     (p=0.089 n=10+10)
Reflect           455ms ± 1%        458ms ± 2%    ~     (p=0.050 n=9+9)
Tar               155ms ± 2%        155ms ± 3%    ~     (p=0.436 n=10+10)
XML               202ms ± 2%        200ms ± 2%    ~     (p=0.053 n=10+9)

Change-Id: I33a7f383d79afba1a482cac6da0cf5b7de9c0ec4
Reviewed-on: https://go-review.googlesource.com/c/go/+/313514
Trust: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>

5 files changed, 128 insertions, 48 deletions

diff --git a/src/cmd/compile/internal/liveness/plive.go b/src/cmd/compile/internal/liveness/plive.go
index 424d248590..f5c2ef7709 100644
--- a/src/cmd/compile/internal/liveness/plive.go
+++ b/src/cmd/compile/internal/liveness/plive.go
@@ -1431,8 +1431,26 @@ func (lv *liveness) emitStackObjects() *obj.LSym {
 		// Note: arguments and return values have non-negative Xoffset,
 		// in which case the offset is relative to argp.
 		// Locals have a negative Xoffset, in which case the offset is relative to varp.
-		off = objw.Uintptr(x, off, uint64(v.FrameOffset()))
-		off = objw.SymPtr(x, off, reflectdata.TypeLinksym(v.Type()), 0)
+		// We already limit the frame size, so the offset and the object size
+		// should not be too big.
+		frameOffset := v.FrameOffset()
+		if frameOffset != int64(int32(frameOffset)) {
+			base.Fatalf("frame offset too big: %v %d", v, frameOffset)
+		}
+		off = objw.Uint32(x, off, uint32(frameOffset))
+
+		t := v.Type()
+		sz := t.Width
+		if sz != int64(int32(sz)) {
+			base.Fatalf("stack object too big: %v of type %v, size %d", v, t, sz)
+		}
+		lsym, useGCProg, ptrdata := reflectdata.GCSym(t)
+		if useGCProg {
+			ptrdata = -ptrdata
+		}
+		off = objw.Uint32(x, off, uint32(sz))
+		off = objw.Uint32(x, off, uint32(ptrdata))
+		off = objw.SymPtr(x, off, lsym, 0)
 	}
 
 	if base.Flag.Live != 0 {
diff --git a/src/cmd/compile/internal/reflectdata/reflect.go b/src/cmd/compile/internal/reflectdata/reflect.go
index 06c4986cf4..01eaf26a0a 100644
--- a/src/cmd/compile/internal/reflectdata/reflect.go
+++ b/src/cmd/compile/internal/reflectdata/reflect.go
@@ -52,6 +52,9 @@ var (
 	signatset   = make(map[*types.Type]struct{})
 	signatslice []*types.Type
 
+	gcsymmu  sync.Mutex // protects gcsymset and gcsymslice
+	gcsymset = make(map[*types.Type]struct{})
+
 	itabs []itabEntry
 	ptabs []*ir.Name
 )
@@ -694,7 +697,8 @@ func dcommontype(lsym *obj.LSym, t *types.Type) int {
 		sptr = writeType(tptr)
 	}
 
-	gcsym, useGCProg, ptrdata := dgcsym(t)
+	gcsym, useGCProg, ptrdata := dgcsym(t, true)
+	delete(gcsymset, t)
 
 	// ../../../../reflect/type.go:/^type.rtype
 	// actual type structure
@@ -1321,6 +1325,16 @@ func WriteRuntimeTypes() {
 			}
 		}
 	}
+
+	// Emit GC data symbols.
+	gcsyms := make([]typeAndStr, 0, len(gcsymset))
+	for t := range gcsymset {
+		gcsyms = append(gcsyms, typeAndStr{t: t, short: types.TypeSymName(t), regular: t.String()})
+	}
+	sort.Sort(typesByString(gcsyms))
+	for _, ts := range gcsyms {
+		dgcsym(ts.t, true)
+	}
 }
 
 func WriteTabs() {
@@ -1490,29 +1504,46 @@ func (a typesByString) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
 //
 const maxPtrmaskBytes = 2048
 
-// dgcsym emits and returns a data symbol containing GC information for type t,
-// along with a boolean reporting whether the UseGCProg bit should be set in
-// the type kind, and the ptrdata field to record in the reflect type information.
-func dgcsym(t *types.Type) (lsym *obj.LSym, useGCProg bool, ptrdata int64) {
+// GCSym returns a data symbol containing GC information for type t, along
+// with a boolean reporting whether the UseGCProg bit should be set in the
+// type kind, and the ptrdata field to record in the reflect type information.
+// GCSym may be called in concurrent backend, so it does not emit the symbol
+// content.
+func GCSym(t *types.Type) (lsym *obj.LSym, useGCProg bool, ptrdata int64) {
+	// Record that we need to emit the GC symbol.
+	gcsymmu.Lock()
+	if _, ok := gcsymset[t]; !ok {
+		gcsymset[t] = struct{}{}
+	}
+	gcsymmu.Unlock()
+
+	return dgcsym(t, false)
+}
+
+// dgcsym returns a data symbol containing GC information for type t, along
+// with a boolean reporting whether the UseGCProg bit should be set in the
+// type kind, and the ptrdata field to record in the reflect type information.
+// When write is true, it writes the symbol data.
+func dgcsym(t *types.Type, write bool) (lsym *obj.LSym, useGCProg bool, ptrdata int64) {
 	ptrdata = types.PtrDataSize(t)
 	if ptrdata/int64(types.PtrSize) <= maxPtrmaskBytes*8 {
-		lsym = dgcptrmask(t)
+		lsym = dgcptrmask(t, write)
 		return
 	}
 
 	useGCProg = true
-	lsym, ptrdata = dgcprog(t)
+	lsym, ptrdata = dgcprog(t, write)
 	return
 }
 
 // dgcptrmask emits and returns the symbol containing a pointer mask for type t.
-func dgcptrmask(t *types.Type) *obj.LSym {
+func dgcptrmask(t *types.Type, write bool) *obj.LSym {
 	ptrmask := make([]byte, (types.PtrDataSize(t)/int64(types.PtrSize)+7)/8)
 	fillptrmask(t, ptrmask)
 	p := fmt.Sprintf("runtime.gcbits.%x", ptrmask)
 
 	lsym := base.Ctxt.Lookup(p)
-	if !lsym.OnList() {
+	if write && !lsym.OnList() {
 		for i, x := range ptrmask {
 			objw.Uint8(lsym, i, x)
 		}
@@ -1549,14 +1580,14 @@ func fillptrmask(t *types.Type, ptrmask []byte) {
 // [types.PtrDataSize(t), t.Width]).
 // In practice, the size is types.PtrDataSize(t) except for non-trivial arrays.
 // For non-trivial arrays, the program describes the full t.Width size.
-func dgcprog(t *types.Type) (*obj.LSym, int64) {
+func dgcprog(t *types.Type, write bool) (*obj.LSym, int64) {
 	types.CalcSize(t)
 	if t.Width == types.BADWIDTH {
 		base.Fatalf("dgcprog: %v badwidth", t)
 	}
 	lsym := TypeLinksymPrefix(".gcprog", t)
 	var p gcProg
-	p.init(lsym)
+	p.init(lsym, write)
 	p.emit(t, 0)
 	offset := p.w.BitIndex() * int64(types.PtrSize)
 	p.end()
@@ -1570,11 +1601,17 @@ type gcProg struct {
 	lsym   *obj.LSym
 	symoff int
 	w      gcprog.Writer
+	write  bool
 }
 
-func (p *gcProg) init(lsym *obj.LSym) {
+func (p *gcProg) init(lsym *obj.LSym, write bool) {
 	p.lsym = lsym
+	p.write = write && !lsym.OnList()
 	p.symoff = 4 // first 4 bytes hold program length
+	if !write {
+		p.w.Init(func(byte) {})
+		return
+	}
 	p.w.Init(p.writeByte)
 	if base.Debug.GCProg > 0 {
 		fmt.Fprintf(os.Stderr, "compile: start GCProg for %v\n", lsym)
@@ -1588,6 +1625,9 @@ func (p *gcProg) writeByte(x byte) {
 
 func (p *gcProg) end() {
 	p.w.End()
+	if !p.write {
+		return
+	}
 	objw.Uint32(p.lsym, 0, uint32(p.symoff-4))
 	objw.Global(p.lsym, int32(p.symoff), obj.DUPOK|obj.RODATA|obj.LOCAL)
 	p.lsym.Set(obj.AttrContentAddressable, true)
diff --git a/src/runtime/mgcmark.go b/src/runtime/mgcmark.go
index 719b21055b..1fd0732d62 100644
--- a/src/runtime/mgcmark.go
+++ b/src/runtime/mgcmark.go
@@ -792,24 +792,24 @@ func scanstack(gp *g, gcw *gcWork) {
 		if obj == nil {
 			continue
 		}
-		t := obj.typ
-		if t == nil {
+		r := obj.r
+		if r == nil {
 			// We've already scanned this object.
 			continue
 		}
-		obj.setType(nil) // Don't scan it again.
+		obj.setRecord(nil) // Don't scan it again.
 		if stackTraceDebug {
 			printlock()
-			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of type", t.string())
+			print("  live stkobj at", hex(state.stack.lo+uintptr(obj.off)), "of size", obj.size)
 			if conservative {
 				print(" (conservative)")
 			}
 			println()
 			printunlock()
 		}
-		gcdata := t.gcdata
+		gcdata := r.gcdata
 		var s *mspan
-		if t.kind&kindGCProg != 0 {
+		if r.useGCProg() {
 			// This path is pretty unlikely, an object large enough
 			// to have a GC program allocated on the stack.
 			// We need some space to unpack the program into a straight
@@ -819,15 +819,15 @@ func scanstack(gp *g, gcw *gcWork) {
 			// to change from a Lempel-Ziv style program to something else.
 			// Or we can forbid putting objects on stacks if they require
 			// a gc program (see issue 27447).
-			s = materializeGCProg(t.ptrdata, gcdata)
+			s = materializeGCProg(r.ptrdata(), gcdata)
 			gcdata = (*byte)(unsafe.Pointer(s.startAddr))
 		}
 
 		b := state.stack.lo + uintptr(obj.off)
 		if conservative {
-			scanConservative(b, t.ptrdata, gcdata, gcw, &state)
+			scanConservative(b, r.ptrdata(), gcdata, gcw, &state)
 		} else {
-			scanblock(b, t.ptrdata, gcdata, gcw, &state)
+			scanblock(b, r.ptrdata(), gcdata, gcw, &state)
 		}
 
 		if s != nil {
@@ -843,10 +843,10 @@ func scanstack(gp *g, gcw *gcWork) {
 		if stackTraceDebug {
 			for i := 0; i < x.nobj; i++ {
 				obj := &x.obj[i]
-				if obj.typ == nil { // reachable
+				if obj.r == nil { // reachable
 					continue
 				}
-				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of type", obj.typ.string())
+				println("  dead stkobj at", hex(gp.stack.lo+uintptr(obj.off)), "of size", obj.r.size)
 				// Note: not necessarily really dead - only reachable-from-ptr dead.
 			}
 		}
@@ -927,7 +927,7 @@ func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
 		// varp is 0 for defers, where there are no locals.
 		// In that case, there can't be a pointer to its args, either.
 		// (And all args would be scanned above anyway.)
-		for _, obj := range objs {
+		for i, obj := range objs {
 			off := obj.off
 			base := frame.varp // locals base pointer
 			if off >= 0 {
@@ -939,9 +939,9 @@ func scanframeworker(frame *stkframe, state *stackScanState, gcw *gcWork) {
 				continue
 			}
 			if stackTraceDebug {
-				println("stkobj at", hex(ptr), "of type", obj.typ.string())
+				println("stkobj at", hex(ptr), "of size", obj.size)
 			}
-			state.addObject(ptr, obj.typ)
+			state.addObject(ptr, &objs[i])
 		}
 	}
 }
diff --git a/src/runtime/mgcstack.go b/src/runtime/mgcstack.go
index 8eb941a328..92d58485f6 100644
--- a/src/runtime/mgcstack.go
+++ b/src/runtime/mgcstack.go
@@ -150,19 +150,19 @@ func init() {
 //
 //go:notinheap
 type stackObject struct {
-	off   uint32       // offset above stack.lo
-	size  uint32       // size of object
-	typ   *_type       // type info (for ptr/nonptr bits). nil if object has been scanned.
-	left  *stackObject // objects with lower addresses
-	right *stackObject // objects with higher addresses
+	off   uint32             // offset above stack.lo
+	size  uint32             // size of object
+	r     *stackObjectRecord // info of the object (for ptr/nonptr bits). nil if object has been scanned.
+	left  *stackObject       // objects with lower addresses
+	right *stackObject       // objects with higher addresses
 }
 
-// obj.typ = typ, but with no write barrier.
+// obj.r = r, but with no write barrier.
 //go:nowritebarrier
-func (obj *stackObject) setType(typ *_type) {
+func (obj *stackObject) setRecord(r *stackObjectRecord) {
 	// Types of stack objects are always in read-only memory, not the heap.
 	// So not using a write barrier is ok.
-	*(*uintptr)(unsafe.Pointer(&obj.typ)) = uintptr(unsafe.Pointer(typ))
+	*(*uintptr)(unsafe.Pointer(&obj.r)) = uintptr(unsafe.Pointer(r))
 }
 
 // A stackScanState keeps track of the state used during the GC walk
@@ -271,7 +271,7 @@ func (s *stackScanState) getPtr() (p uintptr, conservative bool) {
 }
 
 // addObject adds a stack object at addr of type typ to the set of stack objects.
-func (s *stackScanState) addObject(addr uintptr, typ *_type) {
+func (s *stackScanState) addObject(addr uintptr, r *stackObjectRecord) {
 	x := s.tail
 	if x == nil {
 		// initial setup
@@ -294,8 +294,8 @@ func (s *stackScanState) addObject(addr uintptr, typ *_type) {
 	obj := &x.obj[x.nobj]
 	x.nobj++
 	obj.off = uint32(addr - s.stack.lo)
-	obj.size = uint32(typ.size)
-	obj.setType(typ)
+	obj.size = uint32(r.size)
+	obj.setRecord(r)
 	// obj.left and obj.right will be initialized by buildIndex before use.
 	s.nobjs++
 }
diff --git a/src/runtime/stack.go b/src/runtime/stack.go
index babfdfccf0..b21c9c9518 100644
--- a/src/runtime/stack.go
+++ b/src/runtime/stack.go
@@ -702,15 +702,15 @@ func adjustframe(frame *stkframe, arg unsafe.Pointer) bool {
 				// we call into morestack.)
 				continue
 			}
-			t := obj.typ
-			gcdata := t.gcdata
+			ptrdata := obj.ptrdata()
+			gcdata := obj.gcdata
 			var s *mspan
-			if t.kind&kindGCProg != 0 {
+			if obj.useGCProg() {
 				// See comments in mgcmark.go:scanstack
-				s = materializeGCProg(t.ptrdata, gcdata)
+				s = materializeGCProg(ptrdata, gcdata)
 				gcdata = (*byte)(unsafe.Pointer(s.startAddr))
 			}
-			for i := uintptr(0); i < t.ptrdata; i += sys.PtrSize {
+			for i := uintptr(0); i < ptrdata; i += sys.PtrSize {
 				if *addb(gcdata, i/(8*sys.PtrSize))>>(i/sys.PtrSize&7)&1 != 0 {
 					adjustpointer(adjinfo, unsafe.Pointer(p+i))
 				}
@@ -1346,20 +1346,42 @@ var (
 	abiRegArgsType           *_type      = efaceOf(&abiRegArgsEface)._type
 	methodValueCallFrameObjs             = []stackObjectRecord{
 		{
-			off: -int(alignUp(abiRegArgsType.size, 8)), // It's always the highest address local.
-			typ: abiRegArgsType,
+			off:      -int32(alignUp(abiRegArgsType.size, 8)), // It's always the highest address local.
+			size:     int32(abiRegArgsType.size),
+			_ptrdata: int32(abiRegArgsType.ptrdata),
+			gcdata:   abiRegArgsType.gcdata,
 		},
 	}
 )
 
+func init() {
+	if abiRegArgsType.kind&kindGCProg != 0 {
+		throw("abiRegArgsType needs GC Prog, update methodValueCallFrameObjs")
+	}
+}
+
 // A stackObjectRecord is generated by the compiler for each stack object in a stack frame.
 // This record must match the generator code in cmd/compile/internal/liveness/plive.go:emitStackObjects.
 type stackObjectRecord struct {
 	// offset in frame
 	// if negative, offset from varp
 	// if non-negative, offset from argp
-	off int
-	typ *_type
+	off      int32
+	size     int32
+	_ptrdata int32 // ptrdata, or -ptrdata is GC prog is used
+	gcdata   *byte // pointer map or GC prog of the type
+}
+
+func (r *stackObjectRecord) useGCProg() bool {
+	return r._ptrdata < 0
+}
+
+func (r *stackObjectRecord) ptrdata() uintptr {
+	x := r._ptrdata
+	if x < 0 {
+		return uintptr(-x)
+	}
+	return uintptr(x)
 }
 
 // This is exported as ABI0 via linkname so obj can call it.