1 files changed, 56 insertions, 7 deletions

diff --git a/src/runtime/mbitmap.go b/src/runtime/mbitmap.go
index b342de600e..af89577703 100644
--- a/src/runtime/mbitmap.go
+++ b/src/runtime/mbitmap.go
@@ -88,6 +88,7 @@ const (
 
 // addb returns the byte pointer p+n.
 //go:nowritebarrier
+//go:nosplit
 func addb(p *byte, n uintptr) *byte {
 	// Note: wrote out full expression instead of calling add(p, n)
 	// to reduce the number of temporaries generated by the
@@ -99,6 +100,7 @@ func addb(p *byte, n uintptr) *byte {
 // subtractb is typically used when traversing the pointer tables referred to by hbits
 // which are arranged in reverse order.
 //go:nowritebarrier
+//go:nosplit
 func subtractb(p *byte, n uintptr) *byte {
 	// Note: wrote out full expression instead of calling add(p, -n)
 	// to reduce the number of temporaries generated by the
@@ -108,6 +110,7 @@ func subtractb(p *byte, n uintptr) *byte {
 
 // add1 returns the byte pointer p+1.
 //go:nowritebarrier
+//go:nosplit
 func add1(p *byte) *byte {
 	// Note: wrote out full expression instead of calling addb(p, 1)
 	// to reduce the number of temporaries generated by the
@@ -596,10 +599,10 @@ func (h heapBits) setCheckmarked(size uintptr) {
 
 // heapBitsBulkBarrier executes writebarrierptr_nostore
 // for every pointer slot in the memory range [p, p+size),
-// using the heap bitmap to locate those pointer slots.
+// using the heap, data, or BSS bitmap to locate those pointer slots.
 // This executes the write barriers necessary after a memmove.
 // Both p and size must be pointer-aligned.
-// The range [p, p+size) must lie within a single allocation.
+// The range [p, p+size) must lie within a single object.
 //
 // Callers should call heapBitsBulkBarrier immediately after
 // calling memmove(p, src, size). This function is marked nosplit
@@ -643,6 +646,22 @@ func heapBitsBulkBarrier(p, size uintptr) {
 			systemstack(func() {
 				gcUnwindBarriers(gp, p)
 			})
+			return
+		}
+
+		// If p is a global, use the data or BSS bitmaps to
+		// execute write barriers.
+		for datap := &firstmoduledata; datap != nil; datap = datap.next {
+			if datap.data <= p && p < datap.edata {
+				bulkBarrierBitmap(p, size, p-datap.data, datap.gcdatamask.bytedata)
+				return
+			}
+		}
+		for datap := &firstmoduledata; datap != nil; datap = datap.next {
+			if datap.bss <= p && p < datap.ebss {
+				bulkBarrierBitmap(p, size, p-datap.bss, datap.gcbssmask.bytedata)
+				return
+			}
 		}
 		return
 	}
@@ -657,6 +676,36 @@ func heapBitsBulkBarrier(p, size uintptr) {
 	}
 }
 
+// bulkBarrierBitmap executes write barriers for [p, p+size) using a
+// 1-bit pointer bitmap. p is assumed to start maskOffset bytes into
+// the data covered by the bitmap in bits.
+//
+// This is used by heapBitsBulkBarrier for writes to data and BSS.
+//
+//go:nosplit
+func bulkBarrierBitmap(p, size, maskOffset uintptr, bits *uint8) {
+	word := maskOffset / sys.PtrSize
+	bits = addb(bits, word/8)
+	mask := uint8(1) << (word % 8)
+
+	for i := uintptr(0); i < size; i += sys.PtrSize {
+		if mask == 0 {
+			bits = addb(bits, 1)
+			if *bits == 0 {
+				// Skip 8 words.
+				i += 7 * sys.PtrSize
+				continue
+			}
+			mask = 1
+		}
+		if *bits&mask != 0 {
+			x := (*uintptr)(unsafe.Pointer(p + i))
+			writebarrierptr_nostore(x, *x)
+		}
+		mask <<= 1
+	}
+}
+
 // typeBitsBulkBarrier executes writebarrierptr_nostore
 // for every pointer slot in the memory range [p, p+size),
 // using the type bitmap to locate those pointer slots.
@@ -676,11 +725,11 @@ func typeBitsBulkBarrier(typ *_type, p, size uintptr) {
 		throw("runtime: typeBitsBulkBarrier without type")
 	}
 	if typ.size != size {
-		println("runtime: typeBitsBulkBarrier with type ", typ._string, " of size ", typ.size, " but memory size", size)
+		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " of size ", typ.size, " but memory size", size)
 		throw("runtime: invalid typeBitsBulkBarrier")
 	}
 	if typ.kind&kindGCProg != 0 {
-		println("runtime: typeBitsBulkBarrier with type ", typ._string, " with GC prog")
+		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " with GC prog")
 		throw("runtime: invalid typeBitsBulkBarrier")
 	}
 	if !writeBarrier.needed {
@@ -1128,7 +1177,7 @@ func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
 	}
 	if nw == 0 {
 		// No pointers! Caller was supposed to check.
-		println("runtime: invalid type ", typ._string)
+		println("runtime: invalid type ", typ.string())
 		throw("heapBitsSetType: called with non-pointer type")
 		return
 	}
@@ -1314,7 +1363,7 @@ Phase4:
 	if doubleCheck {
 		end := heapBitsForAddr(x + size)
 		if typ.kind&kindGCProg == 0 && (hbitp != end.bitp || (w == nw+2) != (end.shift == 2)) {
-			println("ended at wrong bitmap byte for", typ._string, "x", dataSize/typ.size)
+			println("ended at wrong bitmap byte for", typ.string(), "x", dataSize/typ.size)
 			print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
 			print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
 			h0 := heapBitsForAddr(x)
@@ -1350,7 +1399,7 @@ Phase4:
 				}
 			}
 			if have != want {
-				println("mismatch writing bits for", typ._string, "x", dataSize/typ.size)
+				println("mismatch writing bits for", typ.string(), "x", dataSize/typ.size)
 				print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
 				print("kindGCProg=", typ.kind&kindGCProg != 0, "\n")
 				print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")