runtime (gc): add garbage collector that uses an external allocator

compileopts/config.go

@@ -96,7 +96,7 @@ func (c *Config) CgoEnabled() bool {
 }
 
 // GC returns the garbage collection strategy in use on this platform. Valid
-// values are "none", "leaking", and "conservative".
+// values are "none", "leaking", "extalloc", and "conservative".
 func (c *Config) GC() string {
 	if c.Options.GC != "" {
 		return c.Options.GC
@@ -104,22 +104,29 @@ func (c *Config) GC() string {
 	if c.Target.GC != "" {
 		return c.Target.GC
 	}
-	return "conservative"
+	for _, tag := range c.Target.BuildTags {
+		if tag == "baremetal" || tag == "wasm" {
+			return "conservative"
+		}
+	}
+	return "extalloc"
 }
 
 // NeedsStackObjects returns true if the compiler should insert stack objects
 // that can be traced by the garbage collector.
 func (c *Config) NeedsStackObjects() bool {
-	if c.GC() != "conservative" {
+	switch c.GC() {
+	case "conservative", "extalloc":
+		for _, tag := range c.BuildTags() {
+			if tag == "baremetal" {
+				return false
+			}
+		}
+
+		return true
+	default:
 		return false
 	}
-	for _, tag := range c.BuildTags() {
-		if tag == "baremetal" {
-			return false
-		}
-	}
-
-	return true
 }
 
 // Scheduler returns the scheduler implementation. Valid values are "coroutines"

main.go

@@ -701,7 +701,7 @@ func handleCompilerError(err error) {
 func main() {
 	outpath := flag.String("o", "", "output filename")
 	opt := flag.String("opt", "z", "optimization level: 0, 1, 2, s, z")
-	gc := flag.String("gc", "", "garbage collector to use (none, leaking, conservative)")
+	gc := flag.String("gc", "", "garbage collector to use (none, leaking, extalloc, conservative)")
 	panicStrategy := flag.String("panic", "print", "panic strategy (print, trap)")
 	scheduler := flag.String("scheduler", "", "which scheduler to use (coroutines, tasks)")
 	printIR := flag.Bool("printir", false, "print LLVM IR")

src/runtime/gc_extalloc.go

@@ -0,0 +1,604 @@
+// +build gc.extalloc
+
+package runtime
+
+import "unsafe"
+
+// This garbage collector implementation allows TinyGo to use an external memory allocator.
+// It appends a header to the end of every allocation which the garbage collector uses for tracking purposes.
+// This is also a conservative collector.
+
+const (
+	gcDebug   = false
+	gcAsserts = false
+)
+
+func initHeap() {}
+
+// memTreap is a treap which is used to track allocations for the garbage collector.
+type memTreap struct {
+	root *memTreapNode
+}
+
+// printNode recursively prints a subtree at a given indentation depth.
+func (t *memTreap) printNode(n *memTreapNode, depth int) {
+	for i := 0; i < depth; i++ {
+		print(" ")
+	}
+	println(n, n.priority())
+	if n == nil {
+		return
+	}
+	if gcAsserts && n.parent == nil && t.root != n {
+		runtimePanic("parent missing")
+	}
+	t.printNode(n.left, depth+1)
+	t.printNode(n.right, depth+1)
+}
+
+// print the treap.
+func (t *memTreap) print() {
+	println("treap:")
+	t.printNode(t.root, 1)
+}
+
+// empty returns whether the treap contains any nodes.
+func (t *memTreap) empty() bool {
+	return t.root == nil
+}
+
+// minAddr returns the lowest address contained in an allocation in the treap.
+func (t *memTreap) minAddr() uintptr {
+	// Find the rightmost node.
+	n := t.root
+	for n.right != nil {
+		n = n.right
+	}
+
+	// The lowest address is the base of the rightmost node.
+	return uintptr(unsafe.Pointer(&n.base))
+}
+
+// maxAddr returns the highest address contained in an allocation in the treap.
+func (t *memTreap) maxAddr() uintptr {
+	// Find the leftmost node.
+	n := t.root
+	for n.left != nil {
+		n = n.left
+	}
+
+	// The highest address is the end of the leftmost node.
+	return uintptr(unsafe.Pointer(&n.base)) + n.size
+}
+
+// rotateRight does a right rotation of p and q.
+// https://en.wikipedia.org/wiki/Tree_rotation#/media/File:Tree_rotation.png
+func (t *memTreap) rotateRight(p, q *memTreapNode) {
+	if t.root == q {
+		t.root = p
+	} else {
+		*q.parentSlot() = p
+	}
+
+	//a := p.left
+	b := p.right
+	//c := q.right
+
+	p.parent = q.parent
+	p.right = q
+
+	q.parent = p
+	q.left = b
+
+	if b != nil {
+		b.parent = q
+	}
+}
+
+// rotateLeft does a left rotation of p and q.
+// https://en.wikipedia.org/wiki/Tree_rotation#/media/File:Tree_rotation.png
+func (t *memTreap) rotateLeft(p, q *memTreapNode) {
+	if t.root == p {
+		t.root = q
+	} else {
+		*p.parentSlot() = q
+	}
+
+	//a := p.left
+	b := q.left
+	//c := q.right
+
+	q.parent = p.parent
+	q.left = p
+
+	p.parent = q
+	p.right = b
+
+	if b != nil {
+		b.parent = p
+	}
+}
+
+// rotate rotates a lower node up to its parent.
+// The node n must be a child of m, and will be the parent of m after the rotation.
+func (t *memTreap) rotate(n, m *memTreapNode) {
+	// https://en.wikipedia.org/wiki/Tree_rotation#/media/File:Tree_rotation.png
+	if uintptr(unsafe.Pointer(n)) > uintptr(unsafe.Pointer(m)) {
+		t.rotateRight(n, m)
+	} else {
+		t.rotateLeft(m, n)
+	}
+}
+
+// insert a node into the treap.
+func (t *memTreap) insert(n *memTreapNode) {
+	if gcAsserts && (n.parent != nil || n.left != nil || n.right != nil) {
+		runtimePanic("tried to insert unzeroed treap node")
+	}
+
+	if t.root == nil {
+		// This is the first node, and can be inserted directly into the root.
+		t.root = n
+		return
+	}
+
+	// Insert like a regular binary search tree.
+	for n.parent = t.root; *n.parentSlot() != nil; n.parent = *n.parentSlot() {
+	}
+	*n.parentSlot() = n
+
+	// Rotate the tree to restore the heap invariant.
+	priority := n.priority()
+	for n.parent != nil && priority > n.parent.priority() {
+		t.rotate(n, n.parent)
+	}
+}
+
+// lookupAddr finds the treap node with the allocation containing the specified address.
+// If the address is not contained in any allocations in this treap, nil is returned.
+// NOTE: fields of memTreapNodes are not considered part of the allocations.
+func (t *memTreap) lookupAddr(addr uintptr) *memTreapNode {
+	n := t.root
+	for n != nil && !n.contains(addr) {
+		if addr > uintptr(unsafe.Pointer(n)) {
+			n = n.left
+		} else {
+			n = n.right
+		}
+	}
+
+	return n
+}
+
+// replace a node with another node on the treap.
+func (t *memTreap) replace(old, new *memTreapNode) {
+	if gcAsserts && (old == nil || new == nil) {
+		if gcDebug {
+			println("tried to replace:", old, "->", new)
+		}
+		runtimePanic("invalid replacement")
+	}
+	if gcAsserts && old.parent == nil && old != t.root {
+		if gcDebug {
+			println("tried to replace:", old, "->", new)
+			t.print()
+		}
+		runtimePanic("corrupted tree")
+	}
+	new.parent = old.parent
+	if old == t.root {
+		t.root = new
+	} else {
+		*new.parentSlot() = new
+	}
+}
+
+// remove a node from the treap.
+// This does not free the allocation.
+func (t *memTreap) remove(n *memTreapNode) {
+scan:
+	for {
+		switch {
+		case n.left == nil && n.right == nil && n.parent == nil:
+			// This is the only node - uproot it.
+			t.root = nil
+			break scan
+		case n.left == nil && n.right == nil:
+			// There are no nodes beneath here, so just remove this node from the parent.
+			*n.parentSlot() = nil
+			break scan
+		case n.left != nil && n.right == nil:
+			t.replace(n, n.left)
+			break scan
+		case n.right != nil && n.left == nil:
+			t.replace(n, n.right)
+			break scan
+		default:
+			// Rotate this node downward.
+			if n.left.priority() > n.right.priority() {
+				t.rotate(n.left, n)
+			} else {
+				t.rotate(n.right, n)
+			}
+		}
+	}
+
+	n.left = nil
+	n.right = nil
+	n.parent = nil
+}
+
+// memTreapNode is a treap node used to track allocations for the garbage collector.
+// This struct is prepended to every allocation.
+type memTreapNode struct {
+	parent, left, right *memTreapNode
+	size                uintptr
+	base                struct{}
+}
+
+// priority computes a pseudo-random priority value for this treap node.
+// This value is a fibonacci hash (https://en.wikipedia.org/wiki/Hash_function#Fibonacci_hashing) of the node's memory address.
+func (n *memTreapNode) priority() uintptr {
+	// Select fibonacci multiplier for this bit-width.
+	var fibonacciMultiplier uint64
+	switch 8 * unsafe.Sizeof(uintptr(0)) {
+	case 16:
+		fibonacciMultiplier = 40503
+	case 32:
+		fibonacciMultiplier = 2654435769
+	case 64:
+		fibonacciMultiplier = 11400714819323198485
+	default:
+		runtimePanic("invalid size of uintptr")
+	}
+
+	// Hash the pointer.
+	return uintptr(fibonacciMultiplier) * uintptr(unsafe.Pointer(n))
+}
+
+// contains returns whether this allocation contains a given address.
+func (n *memTreapNode) contains(addr uintptr) bool {
+	return addr >= uintptr(unsafe.Pointer(&n.base)) && addr < uintptr(unsafe.Pointer(&n.base))+n.size
+}
+
+// parentSlot returns a pointer to the parent's reference to this node.
+func (n *memTreapNode) parentSlot() **memTreapNode {
+	if uintptr(unsafe.Pointer(n)) > uintptr(unsafe.Pointer(n.parent)) {
+		return &n.parent.left
+	} else {
+		return &n.parent.right
+	}
+}
+
+// memScanQueue is a queue of memTreapNodes.
+type memScanQueue struct {
+	head, tail *memTreapNode
+}
+
+// push adds an allocation onto the queue.
+func (q *memScanQueue) push(n *memTreapNode) {
+	if gcAsserts && (n.left != nil || n.right != nil || n.parent != nil) {
+		runtimePanic("tried to push a treap node that is in use")
+	}
+
+	if q.head == nil {
+		q.tail = n
+	} else {
+		q.head.left = n
+	}
+	n.right = q.head
+	q.head = n
+}
+
+// pop removes the next allocation from the queue.
+func (q *memScanQueue) pop() *memTreapNode {
+	n := q.tail
+	q.tail = n.left
+	if q.tail == nil {
+		q.head = nil
+	}
+	n.left = nil
+	n.right = nil
+	return n
+}
+
+// empty returns whether the queue contains any allocations.
+func (q *memScanQueue) empty() bool {
+	return q.tail == nil
+}
+
+// allocations is a treap containing all allocations.
+var allocations memTreap
+
+// usedMem is the total amount of allocated memory (including the space taken up by memory treap nodes).
+var usedMem uintptr
+
+// firstPtr and lastPtr are the bounds of memory used by the heap.
+// They are computed before the collector starts marking, and are used to quickly eliminate false positives.
+var firstPtr, lastPtr uintptr
+
+// scanQueue is a queue of marked allocations to scan.
+var scanQueue memScanQueue
+
+// mark searches for an allocation containing the given address and marks it if found.
+func mark(addr uintptr) bool {
+	if addr < firstPtr || addr > lastPtr {
+		// Pointer is outside of allocated bounds.
+		return false
+	}
+
+	node := allocations.lookupAddr(addr)
+	if node != nil {
+		if gcDebug {
+			println("mark:", addr)
+		}
+		allocations.remove(node)
+		scanQueue.push(node)
+	}
+
+	return node != nil
+}
+
+func markRoot(addr uintptr, root uintptr) {
+	marked := mark(root)
+	if gcDebug {
+		if marked {
+			println("marked root:", root, "at", addr)
+		} else if addr != 0 {
+			println("did not mark root:", root, "at", addr)
+		}
+	}
+}
+
+func markRoots(start uintptr, end uintptr) {
+	scan(start, end)
+}
+
+// scan loads all pointer-aligned words and marks any pointers that it finds.
+func scan(start uintptr, end uintptr) {
+	// Align start and end pointers.
+	start = (start + unsafe.Alignof(unsafe.Pointer(nil)) - 1) &^ (unsafe.Alignof(unsafe.Pointer(nil)) - 1)
+	end &^= unsafe.Alignof(unsafe.Pointer(nil)) - 1
+
+	// Mark all pointers.
+	for ptr := start; ptr < end; ptr += unsafe.Alignof(unsafe.Pointer(nil)) {
+		mark(*(*uintptr)(unsafe.Pointer(ptr)))
+	}
+}
+
+// scan marks all allocations referenced by this allocation.
+// This should only be invoked by the garbage collector.
+func (n *memTreapNode) scan() {
+	start := uintptr(unsafe.Pointer(&n.base))
+	end := start + n.size
+	scan(start, end)
+}
+
+// destroy removes and frees all allocations in the treap.
+func (t *memTreap) destroy() {
+	n := t.root
+	for n != nil {
+		switch {
+		case n.left != nil:
+			// Destroy the left subtree.
+			n = n.left
+		case n.right != nil:
+			// Destroy the right subtree.
+			n = n.right
+		default:
+			// This is a leaf node, so delete it and jump back to the parent.
+
+			// Save the parent to jump back to.
+			parent := n.parent
+
+			if parent != nil {
+				*n.parentSlot() = nil
+			} else {
+				t.root = nil
+			}
+
+			// Update used memory.
+			usedMem -= unsafe.Sizeof(memTreapNode{}) + n.size
+			if gcDebug {
+				println("collecting:", &n.base, "size:", n.size)
+				println("used memory:", usedMem)
+			}
+
+			// Free the node.
+			extfree(unsafe.Pointer(n))
+
+			// Jump back to the parent node.
+			n = parent
+		}
+	}
+}
+
+// gcrunning is used by gcAsserts to determine whether the garbage collector is running.
+// This is used to detect if the collector is invoking itself or trying to allocate memory.
+var gcrunning bool
+
+// activeMem is a treap used to store marked allocations which have already been scanned.
+// This is only used when the garbage collector is running.
+var activeMem memTreap
+
+func GC() {
+	if gcDebug {
+		println("running GC")
+	}
+	if allocations.empty() {
+		// Skip collection because the heap is empty.
+		if gcDebug {
+			println("nothing to collect")
+		}
+		return
+	}
+
+	if gcAsserts {
+		if gcrunning {
+			runtimePanic("GC called itself")
+		}
+		gcrunning = true
+	}
+
+	if gcDebug {
+		println("pre-GC allocations:")
+		allocations.print()
+	}
+
+	// Before scanning, find the lowest and highest allocated pointers.
+	// These can be quickly compared against to eliminate most false positives.
+	firstPtr, lastPtr = allocations.minAddr(), allocations.maxAddr()
+
+	// Start by scanning all of the global variables and the stack.
+	markGlobals()
+	markStack()
+
+	// Scan all referenced allocations, building a new treap with marked allocations.
+	// The marking process deletes the allocations from the old allocations treap, so they are only queued once.
+	for !scanQueue.empty() {
+		// Pop a marked node off of the scan queue.
+		n := scanQueue.pop()
+
+		// Scan and mark all nodes that this references.
+		n.scan()
+
+		// Insert this node into the new treap.
+		activeMem.insert(n)
+	}
+
+	// The allocations treap now only contains unreferenced nodes. Destroy them all.
+	allocations.destroy()
+	if gcAsserts && !allocations.empty() {
+		runtimePanic("failed to fully destroy allocations")
+	}
+
+	// Replace the allocations treap with the new treap.
+	allocations = activeMem
+	activeMem = memTreap{}
+
+	if gcDebug {
+		println("GC finished")
+	}
+
+	if gcAsserts {
+		gcrunning = false
+	}
+}
+
+// heapBound is used to control the growth of the heap.
+// When the heap exceeds this size, the garbage collector is run.
+// If the garbage collector cannot free up enough memory, the bound is doubled until the allocation fits.
+var heapBound uintptr = 4 * unsafe.Sizeof(memTreapNode{})
+
+// zeroSizedAlloc is just a sentinel that gets returned when allocating 0 bytes.
+var zeroSizedAlloc uint8
+
+// alloc tries to find some free space on the heap, possibly doing a garbage
+// collection cycle if needed. If no space is free, it panics.
+//go:noinline
+func alloc(size uintptr) unsafe.Pointer {
+	if size == 0 {
+		return unsafe.Pointer(&zeroSizedAlloc)
+	}
+
+	if gcAsserts && gcrunning {
+		runtimePanic("allocated inside the garbage collector")
+	}
+
+	// Calculate size of allocation including treap node.
+	allocSize := unsafe.Sizeof(memTreapNode{}) + size
+
+	var gcRan bool
+	for {
+		// Try to bound heap growth.
+		if usedMem+allocSize < usedMem {
+			if gcDebug {
+				println("current mem:", usedMem, "alloc size:", allocSize)
+			}
+			runtimePanic("target heap size exceeds address space size")
+		}
+		if usedMem+allocSize > heapBound {
+			if !gcRan {
+				// Run the garbage collector before growing the heap.
+				if gcDebug {
+					println("heap reached size limit")
+				}
+				GC()
+				gcRan = true
+				continue
+			} else {
+				// Grow the heap bound to fit the allocation.
+				for heapBound != 0 && usedMem+allocSize > heapBound {
+					heapBound <<= 1
+				}
+				if heapBound == 0 {
+					// This is only possible on hosted 32-bit systems.
+					// Allow the heap bound to encompass everything.
+					heapBound = ^uintptr(0)
+				}
+				if gcDebug {
+					println("raising heap size limit to", heapBound)
+				}
+			}
+		}
+
+		// Allocate the memory.
+		ptr := extalloc(allocSize)
+		if ptr == nil {
+			if gcDebug {
+				println("extalloc failed")
+			}
+			if gcRan {
+				// Garbage collector was not able to free up enough memory.
+				runtimePanic("out of memory")
+			} else {
+				// Run the garbage collector and try again.
+				GC()
+				gcRan = true
+				continue
+			}
+		}
+
+		// Initialize the memory treap node.
+		node := (*memTreapNode)(ptr)
+		*node = memTreapNode{
+			size: size,
+		}
+
+		// Insert allocation into the allocations treap.
+		allocations.insert(node)
+
+		// Extract the user's section of the allocation.
+		ptr = unsafe.Pointer(&node.base)
+		if gcAsserts && !node.contains(uintptr(ptr)) {
+			runtimePanic("node is not self-contained")
+		}
+		if gcAsserts {
+			check := allocations.lookupAddr(uintptr(ptr))
+			if check == nil {
+				if gcDebug {
+					println("failed to find:", ptr)
+					allocations.print()
+				}
+				runtimePanic("bad insert")
+			}
+		}
+
+		// Zero the allocation.
+		memzero(ptr, size)
+
+		// Update used memory.
+		usedMem += allocSize
+
+		if gcDebug {
+			println("allocated:", uintptr(ptr), "size:", size)
+			println("used memory:", usedMem)
+		}
+
+		return ptr
+	}
+}
+
+func free(ptr unsafe.Pointer) {
+	// Currently unimplemented due to bugs in coroutine lowering.
+}

src/runtime/gc_globals_conservative.go

@@ -1,4 +1,4 @@
-// +build gc.conservative
+// +build gc.conservative gc.extalloc
 // +build baremetal
 
 package runtime

src/runtime/gc_globals_precise.go

@@ -1,4 +1,4 @@
-// +build gc.conservative
+// +build gc.conservative gc.extalloc
 // +build !baremetal
 
 package runtime

src/runtime/gc_stack_portable.go

@@ -1,4 +1,4 @@
-// +build gc.conservative
+// +build gc.conservative gc.extalloc
 // +build !baremetal
 
 package runtime

src/runtime/gc_stack_raw.go

@@ -1,4 +1,4 @@
-// +build gc.conservative
+// +build gc.conservative gc.extalloc
 // +build baremetal
 
 package runtime

src/runtime/runtime_unix.go

@@ -24,10 +24,6 @@ func exit(code int)
 //go:export clock_gettime
 func clock_gettime(clk_id int32, ts *timespec)
 
-const heapSize = 1 * 1024 * 1024 // 1MB to start
-
-var heapStart, heapEnd uintptr
-
 type timeUnit int64
 
 const tickMicros = 1
@@ -47,8 +43,7 @@ func postinit() {}
 // Entry point for Go. Initialize all packages and call main.main().
 //go:export main
 func main() int {
-	heapStart = uintptr(malloc(heapSize))
-	heapEnd = heapStart + heapSize
+	preinit()
 
 	run()
 
@@ -83,3 +78,10 @@ func ticks() timeUnit {
 func syscall_Exit(code int) {
 	exit(code)
 }
+
+func extalloc(size uintptr) unsafe.Pointer {
+	return malloc(size)
+}
+
+//go:export free
+func extfree(ptr unsafe.Pointer)

src/runtime/runtime_unix_heap.go

@@ -0,0 +1,14 @@
+// +build darwin linux,!baremetal freebsd,!baremetal
+
+// +build gc.conservative gc.leaking
+
+package runtime
+
+const heapSize = 1 * 1024 * 1024 // 1MB to start
+
+var heapStart, heapEnd uintptr
+
+func preinit() {
+	heapStart = uintptr(malloc(heapSize))
+	heapEnd = heapStart + heapSize
+}

src/runtime/runtime_unix_noheap.go

@@ -0,0 +1,7 @@
+// +build darwin linux,!baremetal freebsd,!baremetal
+
+// +build gc.none gc.extalloc
+
+package runtime
+
+func preinit() {}

transform/coroutines.go

@@ -60,10 +60,11 @@ func LowerCoroutines(mod llvm.Module, needStackSlots bool) error {
 	defer target.Dispose()
 
 	pass := &coroutineLoweringPass{
-		mod:     mod,
-		ctx:     ctx,
-		builder: builder,
-		target:  target,
+		mod:            mod,
+		ctx:            ctx,
+		builder:        builder,
+		target:         target,
+		needStackSlots: needStackSlots,
 	}
 
 	err := pass.load()
@@ -149,6 +150,9 @@ type coroutineLoweringPass struct {
 
 	// llvm.coro intrinsics
 	coroId, coroSize, coroBegin, coroSuspend, coroEnd, coroFree, coroSave llvm.Value
+
+	trackPointer   llvm.Value
+	needStackSlots bool
 }
 
 // findAsyncFuncs finds all asynchronous functions.
@@ -265,6 +269,13 @@ func (c *coroutineLoweringPass) load() error {
 		return ErrMissingIntrinsic{"internal/task.createTask"}
 	}
 
+	if c.needStackSlots {
+		c.trackPointer = c.mod.NamedFunction("runtime.trackPointer")
+		if c.trackPointer.IsNil() {
+			return ErrMissingIntrinsic{"runtime.trackPointer"}
+		}
+	}
+
 	// Find async functions.
 	c.findAsyncFuncs()
 
@@ -297,6 +308,15 @@ func (c *coroutineLoweringPass) load() error {
 	return nil
 }
 
+func (c *coroutineLoweringPass) track(ptr llvm.Value) {
+	if c.needStackSlots {
+		if ptr.Type() != c.i8ptr {
+			ptr = c.builder.CreateBitCast(ptr, c.i8ptr, "track.bitcast")
+		}
+		c.builder.CreateCall(c.trackPointer, []llvm.Value{ptr, llvm.Undef(c.i8ptr), llvm.Undef(c.i8ptr)}, "")
+	}
+}
+
 // lowerStartSync lowers a goroutine start of a synchronous function to a synchronous call.
 func (c *coroutineLoweringPass) lowerStartSync(start llvm.Value) {
 	c.builder.SetInsertPointBefore(start)
@@ -662,6 +682,7 @@ func (c *coroutineLoweringPass) lowerFuncCoro(fn *asyncFunc) {
 	coroAlloc := c.builder.CreateCall(c.alloc, []llvm.Value{coroSize, llvm.Undef(c.i8ptr), llvm.Undef(c.i8ptr)}, "coro.alloc")
 	// %coro.state = call noalias i8* @llvm.coro.begin(token %coro.id, i8* %coro.alloc)
 	coroState := c.builder.CreateCall(c.coroBegin, []llvm.Value{coroId, coroAlloc}, "coro.state")
+	c.track(coroState)
 	// Store state into task.
 	task := c.builder.CreateCall(c.current, []llvm.Value{llvm.Undef(c.i8ptr), fn.rawTask}, "task")
 	parentState := c.builder.CreateCall(c.setState, []llvm.Value{task, coroState, llvm.Undef(c.i8ptr), llvm.Undef(c.i8ptr)}, "task.state.parent")
@@ -795,6 +816,9 @@ func (c *coroutineLoweringPass) lowerFuncCoro(fn *asyncFunc) {
 		if call.CalledValue() == c.pause {
 			call.EraseFromParentAsInstruction()
 		}
+
+		c.builder.SetInsertPointBefore(wakeup.FirstInstruction())
+		c.track(coroState)
 	}
 }