cmd/compile, cmd/link, runtime: make defers low-cost through inline code and extra funcdata

Generate inline code at defer time to save the args of defer calls to unique
(autotmp) stack slots, and generate inline code at exit time to check which defer
calls were made and make the associated function/method/interface calls. We
remember that a particular defer statement was reached by storing in the deferBits
variable (always stored on the stack). At exit time, we check the bits of the
deferBits variable to determine which defer function calls to make (in reverse
order). These low-cost defers are only used for functions where no defers
appear in loops. In addition, we don't do these low-cost defers if there are too
many defer statements or too many exits in a function (to limit code increase).

When a function uses open-coded defers, we produce extra
FUNCDATA_OpenCodedDeferInfo information that specifies the number of defers, and
for each defer, the stack slots where the closure and associated args have been
stored. The funcdata also includes the location of the deferBits variable.
Therefore, for panics, we can use this funcdata to determine exactly which defers
are active, and call the appropriate functions/methods/closures with the correct
arguments for each active defer.

In order to unwind the stack correctly after a recover(), we need to add an extra
code segment to functions with open-coded defers that simply calls deferreturn()
and returns. This segment is not reachable by the normal function, but is returned
to by the runtime during recovery. We set the liveness information of this
deferreturn() to be the same as the liveness at the first function call during the
last defer exit code (so all return values and all stack slots needed by the defer
calls will be live).

I needed to increase the stackguard constant from 880 to 896, because of a small
amount of new code in deferreturn().

The -N flag disables open-coded defers. '-d defer' prints out the kind of defer
being used at each defer statement (heap-allocated, stack-allocated, or
open-coded).

Cost of defer statement  [ go test -run NONE -bench BenchmarkDefer$ runtime ]
  With normal (stack-allocated) defers only:         35.4  ns/op
  With open-coded defers:                             5.6  ns/op
  Cost of function call alone (remove defer keyword): 4.4  ns/op

Text size increase (including funcdata) for go cmd without/with open-coded defers:  0.09%

The average size increase (including funcdata) for only the functions that use
open-coded defers is 1.1%.

The cost of a panic followed by a recover got noticeably slower, since panic
processing now requires a scan of the stack for open-coded defer frames. This scan
is required, even if no frames are using open-coded defers:

Cost of panic and recover [ go test -run NONE -bench BenchmarkPanicRecover runtime ]
  Without open-coded defers:        62.0 ns/op
  With open-coded defers:           255  ns/op

A CGO Go-to-C-to-Go benchmark got noticeably faster because of open-coded defers:

CGO Go-to-C-to-Go benchmark [cd misc/cgo/test; go test -run NONE -bench BenchmarkCGoCallback ]
  Without open-coded defers:        443 ns/op
  With open-coded defers:           347 ns/op

Updates #14939 (defer performance)
Updates #34481 (design doc)

Change-Id: I51a389860b9676cfa1b84722f5fb84d3c4ee9e28
Reviewed-on: https://go-review.googlesource.com/c/go/+/190098
Reviewed-by: Austin Clements <austin@google.com>

16 files changed, 659 insertions, 86 deletions

diff --git a/src/cmd/compile/internal/gc/escape.go b/src/cmd/compile/internal/gc/escape.go
index b855f4a174..3dc6b98123 100644
--- a/src/cmd/compile/internal/gc/escape.go
+++ b/src/cmd/compile/internal/gc/escape.go
@@ -371,6 +371,7 @@ func (e *Escape) stmt(n *Node) {
 		e.stmts(n.Right.Ninit)
 		e.call(e.addrs(n.List), n.Right, nil)
 	case ORETURN:
+		e.curfn.Func.numReturns++
 		results := e.curfn.Type.Results().FieldSlice()
 		for i, v := range n.List.Slice() {
 			e.assign(asNode(results[i].Nname), v, "return", n)
@@ -378,6 +379,16 @@ func (e *Escape) stmt(n *Node) {
 	case OCALLFUNC, OCALLMETH, OCALLINTER, OCLOSE, OCOPY, ODELETE, OPANIC, OPRINT, OPRINTN, ORECOVER:
 		e.call(nil, n, nil)
 	case OGO, ODEFER:
+		if n.Op == ODEFER {
+			e.curfn.Func.SetHasDefer(true)
+			e.curfn.Func.numDefers++
+			if e.curfn.Func.numDefers > maxOpenDefers {
+				// Don't allow open defers if there are more than
+				// 8 defers in the function, since we use a single
+				// byte to record active defers.
+				e.curfn.Func.SetOpenCodedDeferDisallowed(true)
+			}
+		}
 		e.stmts(n.Left.Ninit)
 		e.call(nil, n.Left, n)
 
@@ -872,8 +883,13 @@ func (e *Escape) augmentParamHole(k EscHole, where *Node) EscHole {
 	// non-transient location to avoid arguments from being
 	// transiently allocated.
 	if where.Op == ODEFER && e.loopDepth == 1 {
-		where.Esc = EscNever // force stack allocation of defer record (see ssa.go)
+		// force stack allocation of defer record, unless open-coded
+		// defers are used (see ssa.go)
+		where.Esc = EscNever
 		return e.later(k)
+	} else if where.Op == ODEFER {
+		// If any defer occurs in a loop, open-coded defers cannot be used
+		e.curfn.Func.SetOpenCodedDeferDisallowed(true)
 	}
 
 	return e.heapHole()
diff --git a/src/cmd/compile/internal/gc/main.go b/src/cmd/compile/internal/gc/main.go
index 05aac9ecb2..8806386707 100644
--- a/src/cmd/compile/internal/gc/main.go
+++ b/src/cmd/compile/internal/gc/main.go
@@ -52,6 +52,7 @@ var (
 	Debug_typecheckinl int
 	Debug_gendwarfinl  int
 	Debug_softfloat    int
+	Debug_defer        int
 )
 
 // Debug arguments.
@@ -81,6 +82,7 @@ var debugtab = []struct {
 	{"typecheckinl", "eager typechecking of inline function bodies", &Debug_typecheckinl},
 	{"dwarfinl", "print information about DWARF inlined function creation", &Debug_gendwarfinl},
 	{"softfloat", "force compiler to emit soft-float code", &Debug_softfloat},
+	{"defer", "print information about defer compilation", &Debug_defer},
 }
 
 const debugHelpHeader = `usage: -d arg[,arg]* and arg is <key>[=<value>]
diff --git a/src/cmd/compile/internal/gc/obj.go b/src/cmd/compile/internal/gc/obj.go
index be13b27892..83371fabf5 100644
--- a/src/cmd/compile/internal/gc/obj.go
+++ b/src/cmd/compile/internal/gc/obj.go
@@ -294,6 +294,9 @@ func addGCLocals() {
 			}
 			ggloblsym(x, int32(len(x.P)), attr)
 		}
+		if x := s.Func.OpenCodedDeferInfo; x != nil {
+			ggloblsym(x, int32(len(x.P)), obj.RODATA|obj.DUPOK)
+		}
 	}
 }
 
diff --git a/src/cmd/compile/internal/gc/reflect.go b/src/cmd/compile/internal/gc/reflect.go
index 9e3dca25c8..00a24f2dff 100644
--- a/src/cmd/compile/internal/gc/reflect.go
+++ b/src/cmd/compile/internal/gc/reflect.go
@@ -338,6 +338,7 @@ func deferstruct(stksize int64) *types.Type {
 		makefield("siz", types.Types[TUINT32]),
 		makefield("started", types.Types[TBOOL]),
 		makefield("heap", types.Types[TBOOL]),
+		makefield("openDefer", types.Types[TBOOL]),
 		makefield("sp", types.Types[TUINTPTR]),
 		makefield("pc", types.Types[TUINTPTR]),
 		// Note: the types here don't really matter. Defer structures
@@ -346,6 +347,9 @@ func deferstruct(stksize int64) *types.Type {
 		makefield("fn", types.Types[TUINTPTR]),
 		makefield("_panic", types.Types[TUINTPTR]),
 		makefield("link", types.Types[TUINTPTR]),
+		makefield("framepc", types.Types[TUINTPTR]),
+		makefield("varp", types.Types[TUINTPTR]),
+		makefield("fd", types.Types[TUINTPTR]),
 		makefield("args", argtype),
 	}
 
diff --git a/src/cmd/compile/internal/gc/sizeof_test.go b/src/cmd/compile/internal/gc/sizeof_test.go
index f4725c0eb2..ce4a216c2e 100644
--- a/src/cmd/compile/internal/gc/sizeof_test.go
+++ b/src/cmd/compile/internal/gc/sizeof_test.go
@@ -20,7 +20,7 @@ func TestSizeof(t *testing.T) {
 		_32bit uintptr     // size on 32bit platforms
 		_64bit uintptr     // size on 64bit platforms
 	}{
-		{Func{}, 116, 208},
+		{Func{}, 124, 224},
 		{Name{}, 32, 56},
 		{Param{}, 24, 48},
 		{Node{}, 76, 128},
diff --git a/src/cmd/compile/internal/gc/ssa.go b/src/cmd/compile/internal/gc/ssa.go
index dd8dacd149..200cca1063 100644
--- a/src/cmd/compile/internal/gc/ssa.go
+++ b/src/cmd/compile/internal/gc/ssa.go
@@ -29,6 +29,10 @@ var ssaDumpStdout bool // whether to dump to stdout
 var ssaDumpCFG string  // generate CFGs for these phases
 const ssaDumpFile = "ssa.html"
 
+// The max number of defers in a function using open-coded defers. We enforce this
+// limit because the deferBits bitmask is currently a single byte (to minimize code size)
+const maxOpenDefers = 8
+
 // ssaDumpInlined holds all inlined functions when ssaDump contains a function name.
 var ssaDumpInlined []*Node
 
@@ -165,6 +169,107 @@ func initssaconfig() {
 	SigPanic = sysfunc("sigpanic")
 }
 
+// getParam returns the Field of ith param of node n (which is a
+// function/method/interface call), where the receiver of a method call is
+// considered as the 0th parameter. This does not include the receiver of an
+// interface call.
+func getParam(n *Node, i int) *types.Field {
+	t := n.Left.Type
+	if n.Op == OCALLMETH {
+		if i == 0 {
+			return t.Recv()
+		}
+		return t.Params().Field(i - 1)
+	}
+	return t.Params().Field(i)
+}
+
+// dvarint writes a varint v to the funcdata in symbol x and returns the new offset
+func dvarint(x *obj.LSym, off int, v int64) int {
+	if v < 0 || v > 1e9 {
+		panic(fmt.Sprintf("dvarint: bad offset for funcdata - %v", v))
+	}
+	if v < 1<<7 {
+		return duint8(x, off, uint8(v))
+	}
+	off = duint8(x, off, uint8((v&127)|128))
+	if v < 1<<14 {
+		return duint8(x, off, uint8(v>>7))
+	}
+	off = duint8(x, off, uint8(((v>>7)&127)|128))
+	if v < 1<<21 {
+		return duint8(x, off, uint8(v>>14))
+	}
+	off = duint8(x, off, uint8(((v>>14)&127)|128))
+	if v < 1<<28 {
+		return duint8(x, off, uint8(v>>21))
+	}
+	off = duint8(x, off, uint8(((v>>21)&127)|128))
+	return duint8(x, off, uint8(v>>28))
+}
+
+// emitOpenDeferInfo emits FUNCDATA information about the defers in a function
+// that is using open-coded defers.  This funcdata is used to determine the active
+// defers in a function and execute those defers during panic processing.
+//
+// The funcdata is all encoded in varints (since values will almost always be less
+// than 128, but stack offsets could potentially be up to 2Gbyte). All "locations"
+// for stack variables are specified as the number of bytes below varp for their
+// starting address. The format is:
+//
+//  - Max total argument size among all the defers
+//  - Location of the deferBits variable
+//  - Number of defers in the function
+//  - Information about each defer call, in reverse order of appearance in the function:
+//    - Total argument size of the call
+//    - Location of the closure value to call
+//    - 1 or 0 to indicate if there is a receiver for the call
+//      - If yes, then the location of the receiver value
+//    - Number of arguments
+//    - Information about each argument
+//      - Location of the stored defer argument in this function's frame
+//      - Size of the argument
+//      - Offset of where argument should be placed in the args frame when making call
+func emitOpenDeferInfo(s state) {
+	x := Ctxt.Lookup(s.curfn.Func.lsym.Name + ".opendefer")
+	s.curfn.Func.lsym.Func.OpenCodedDeferInfo = x
+	off := 0
+
+	// Compute maxargsize (max size of arguments for all defers)
+	// first, so we can output it first to the funcdata
+	var maxargsize int64
+	for i := len(s.opendefers) - 1; i >= 0; i-- {
+		r := s.opendefers[i]
+		argsize := r.n.Left.Type.ArgWidth()
+		if argsize > maxargsize {
+			maxargsize = argsize
+		}
+	}
+	off = dvarint(x, off, maxargsize)
+	off = dvarint(x, off, -s.deferBitsTemp.Xoffset)
+	off = dvarint(x, off, int64(len(s.opendefers)))
+
+	// Write in reverse-order, for ease of running in that order at runtime
+	for i := len(s.opendefers) - 1; i >= 0; i-- {
+		r := s.opendefers[i]
+		off = dvarint(x, off, r.n.Left.Type.ArgWidth())
+		off = dvarint(x, off, -r.closureNode.Xoffset)
+		if r.rcvrNode != nil {
+			off = dvarint(x, off, 1)
+			off = dvarint(x, off, -r.rcvrNode.Xoffset)
+		} else {
+			off = dvarint(x, off, 0)
+		}
+		off = dvarint(x, off, int64(len(r.argNodes)))
+		for j, arg := range r.argNodes {
+			f := getParam(r.n, j)
+			off = dvarint(x, off, -arg.Xoffset)
+			off = dvarint(x, off, f.Type.Size())
+			off = dvarint(x, off, f.Offset)
+		}
+	}
+}
+
 // buildssa builds an SSA function for fn.
 // worker indicates which of the backend workers is doing the processing.
 func buildssa(fn *Node, worker int) *ssa.Func {
@@ -227,11 +332,48 @@ func buildssa(fn *Node, worker int) *ssa.Func {
 	s.labeledNodes = map[*Node]*ssaLabel{}
 	s.fwdVars = map[*Node]*ssa.Value{}
 	s.startmem = s.entryNewValue0(ssa.OpInitMem, types.TypeMem)
+
+	s.hasOpenDefers = Debug['N'] == 0 && s.hasdefer && !s.curfn.Func.OpenCodedDeferDisallowed()
+	if s.hasOpenDefers && s.curfn.Func.Exit.Len() > 0 {
+		// Skip doing open defers if there is any extra exit code (likely
+		// copying heap-allocated return values or race detection), since
+		// we will not generate that code in the case of the extra
+		// deferreturn/ret segment.
+		s.hasOpenDefers = false
+	}
+	if s.hasOpenDefers &&
+		s.curfn.Func.numReturns*s.curfn.Func.numDefers > 15 {
+		// Since we are generating defer calls at every exit for
+		// open-coded defers, skip doing open-coded defers if there are
+		// too many returns (especially if there are multiple defers).
+		// Open-coded defers are most important for improving performance
+		// for smaller functions (which don't have many returns).
+		s.hasOpenDefers = false
+	}
+
 	s.sp = s.entryNewValue0(ssa.OpSP, types.Types[TUINTPTR]) // TODO: use generic pointer type (unsafe.Pointer?) instead
 	s.sb = s.entryNewValue0(ssa.OpSB, types.Types[TUINTPTR])
 
 	s.startBlock(s.f.Entry)
 	s.vars[&memVar] = s.startmem
+	if s.hasOpenDefers {
+		// Create the deferBits variable and stack slot.  deferBits is a
+		// bitmask showing which of the open-coded defers in this function
+		// have been activated.
+		deferBitsTemp := tempAt(src.NoXPos, s.curfn, types.Types[TUINT8])
+		s.deferBitsTemp = deferBitsTemp
+		// For this value, AuxInt is initialized to zero by default
+		startDeferBits := s.entryNewValue0(ssa.OpConst8, types.Types[TUINT8])
+		s.vars[&deferBitsVar] = startDeferBits
+		s.deferBitsAddr = s.addr(deferBitsTemp, false)
+		s.store(types.Types[TUINT8], s.deferBitsAddr, startDeferBits)
+		// Make sure that the deferBits stack slot is kept alive (for use
+		// by panics) and stores to deferBits are not eliminated, even if
+		// all checking code on deferBits in the function exit can be
+		// eliminated, because the defer statements were all
+		// unconditional.
+		s.vars[&memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
+	}
 
 	// Generate addresses of local declarations
 	s.decladdrs = map[*Node]*ssa.Value{}
@@ -287,6 +429,11 @@ func buildssa(fn *Node, worker int) *ssa.Func {
 
 	// Main call to ssa package to compile function
 	ssa.Compile(s.f)
+
+	if s.hasOpenDefers {
+		emitOpenDeferInfo(s)
+	}
+
 	return s.f
 }
 
@@ -375,6 +522,29 @@ func (s *state) updateUnsetPredPos(b *ssa.Block) {
 	}
 }
 
+// Information about each open-coded defer.
+type openDeferInfo struct {
+	// The ODEFER node representing the function call of the defer
+	n *Node
+	// If defer call is closure call, the address of the argtmp where the
+	// closure is stored.
+	closure *ssa.Value
+	// The node representing the argtmp where the closure is stored - used for
+	// function, method, or interface call, to store a closure that panic
+	// processing can use for this defer.
+	closureNode *Node
+	// If defer call is interface call, the address of the argtmp where the
+	// receiver is stored
+	rcvr *ssa.Value
+	// The node representing the argtmp where the receiver is stored
+	rcvrNode *Node
+	// The addresses of the argtmps where the evaluated arguments of the defer
+	// function call are stored.
+	argVals []*ssa.Value
+	// The nodes representing the argtmps where the args of the defer are stored
+	argNodes []*Node
+}
+
 type state struct {
 	// configuration (arch) information
 	config *ssa.Config
@@ -416,6 +586,9 @@ type state struct {
 	startmem *ssa.Value
 	sp       *ssa.Value
 	sb       *ssa.Value
+	// value representing address of where deferBits autotmp is stored
+	deferBitsAddr *ssa.Value
+	deferBitsTemp *Node
 
 	// line number stack. The current line number is top of stack
 	line []src.XPos
@@ -432,6 +605,19 @@ type state struct {
 	cgoUnsafeArgs bool
 	hasdefer      bool // whether the function contains a defer statement
 	softFloat     bool
+	hasOpenDefers bool // whether we are doing open-coded defers
+
+	// If doing open-coded defers, list of info about the defer calls in
+	// scanning order. Hence, at exit we should run these defers in reverse
+	// order of this list
+	opendefers []*openDeferInfo
+	// For open-coded defers, this is the beginning and end blocks of the last
+	// defer exit code that we have generated so far. We use these to share
+	// code between exits if the shareDeferExits option (disabled by default)
+	// is on.
+	lastDeferExit       *ssa.Block // Entry block of last defer exit code we generated
+	lastDeferFinalBlock *ssa.Block // Final block of last defer exit code we generated
+	lastDeferCount      int        // Number of defers encountered at that point
 }
 
 type funcLine struct {
@@ -469,12 +655,13 @@ var (
 	memVar = Node{Op: ONAME, Sym: &types.Sym{Name: "mem"}}
 
 	// dummy nodes for temporary variables
-	ptrVar    = Node{Op: ONAME, Sym: &types.Sym{Name: "ptr"}}
-	lenVar    = Node{Op: ONAME, Sym: &types.Sym{Name: "len"}}
-	newlenVar = Node{Op: ONAME, Sym: &types.Sym{Name: "newlen"}}
-	capVar    = Node{Op: ONAME, Sym: &types.Sym{Name: "cap"}}
-	typVar    = Node{Op: ONAME, Sym: &types.Sym{Name: "typ"}}
-	okVar     = Node{Op: ONAME, Sym: &types.Sym{Name: "ok"}}
+	ptrVar       = Node{Op: ONAME, Sym: &types.Sym{Name: "ptr"}}
+	lenVar       = Node{Op: ONAME, Sym: &types.Sym{Name: "len"}}
+	newlenVar    = Node{Op: ONAME, Sym: &types.Sym{Name: "newlen"}}
+	capVar       = Node{Op: ONAME, Sym: &types.Sym{Name: "cap"}}
+	typVar       = Node{Op: ONAME, Sym: &types.Sym{Name: "typ"}}
+	okVar        = Node{Op: ONAME, Sym: &types.Sym{Name: "ok"}}
+	deferBitsVar = Node{Op: ONAME, Sym: &types.Sym{Name: "deferBits"}}
 )
 
 // startBlock sets the current block we're generating code in to b.
@@ -865,11 +1052,27 @@ func (s *state) stmt(n *Node) {
 			}
 		}
 	case ODEFER:
-		d := callDefer
-		if n.Esc == EscNever {
-			d = callDeferStack
+		if Debug_defer > 0 {
+			var defertype string
+			if s.hasOpenDefers {
+				defertype = "open-coded"
+			} else if n.Esc == EscNever {
+				defertype = "stack-allocated"
+			} else {
+				defertype = "heap-allocated"
+			}
+			Warnl(n.Pos, "defer: %s defer in function %s",
+				defertype, s.curfn.funcname())
+		}
+		if s.hasOpenDefers {
+			s.openDeferRecord(n.Left)
+		} else {
+			d := callDefer
+			if n.Esc == EscNever {
+				d = callDeferStack
+			}
+			s.call(n.Left, d)
 		}
-		s.call(n.Left, d)
 	case OGO:
 		s.call(n.Left, callGo)
 
@@ -1286,12 +1489,28 @@ func (s *state) stmt(n *Node) {
 	}
 }
 
+// If true, share as many open-coded defer exits as possible (with the downside of
+// worse line-number information)
+const shareDeferExits = false
+
 // exit processes any code that needs to be generated just before returning.
 // It returns a BlockRet block that ends the control flow. Its control value
 // will be set to the final memory state.
 func (s *state) exit() *ssa.Block {
 	if s.hasdefer {
-		s.rtcall(Deferreturn, true, nil)
+		if s.hasOpenDefers {
+			if shareDeferExits && s.lastDeferExit != nil && len(s.opendefers) == s.lastDeferCount {
+				if s.curBlock.Kind != ssa.BlockPlain {
+					panic("Block for an exit should be BlockPlain")
+				}
+				s.curBlock.AddEdgeTo(s.lastDeferExit)
+				s.endBlock()
+				return s.lastDeferFinalBlock
+			}
+			s.openDeferExit()
+		} else {
+			s.rtcall(Deferreturn, true, nil)
+		}
 	}
 
 	// Run exit code. Typically, this code copies heap-allocated PPARAMOUT
@@ -1314,6 +1533,9 @@ func (s *state) exit() *ssa.Block {
 	b := s.endBlock()
 	b.Kind = ssa.BlockRet
 	b.SetControl(m)
+	if s.hasdefer && s.hasOpenDefers {
+		s.lastDeferFinalBlock = b
+	}
 	return b
 }
 
@@ -3764,6 +3986,230 @@ func (s *state) intrinsicArgs(n *Node) []*ssa.Value {
 	return args
 }
 
+// openDeferRecord adds code to evaluate and store the args for an open-code defer
+// call, and records info about the defer, so we can generate proper code on the
+// exit paths. n is the sub-node of the defer node that is the actual function
+// call. We will also record funcdata information on where the args are stored
+// (as well as the deferBits variable), and this will enable us to run the proper
+// defer calls during panics.
+func (s *state) openDeferRecord(n *Node) {
+	index := len(s.opendefers)
+
+	// Do any needed expression evaluation for the args (including the
+	// receiver, if any). This may be evaluating something like 'autotmp_3 =
+	// once.mutex'. Such a statement will create a mapping in s.vars[] from
+	// the autotmp name to the evaluated SSA arg value, but won't do any
+	// stores to the stack.
+	s.stmtList(n.List)
+
+	args := []*ssa.Value{}
+	argNodes := []*Node{}
+
+	opendefer := &openDeferInfo{
+		n: n,
+	}
+	fn := n.Left
+	if n.Op == OCALLFUNC {
+		// We must always store the function value in a stack slot for the
+		// runtime panic code to use. But in the defer exit code, we will
+		// call the function directly if it is a static function.
+		closureVal := s.expr(fn)
+		closure := s.openDeferSave(fn, fn.Type, closureVal)
+		opendefer.closureNode = closure.Aux.(*Node)
+		if !(fn.Op == ONAME && fn.Class() == PFUNC) {
+			opendefer.closure = closure
+		}
+	} else if n.Op == OCALLMETH {
+		if fn.Op != ODOTMETH {
+			Fatalf("OCALLMETH: n.Left not an ODOTMETH: %v", fn)
+		}
+		closureVal := s.getMethodClosure(fn)
+		// We must always store the function value in a stack slot for the
+		// runtime panic code to use. But in the defer exit code, we will
+		// call the method directly.
+		closure := s.openDeferSave(fn, fn.Type, closureVal)
+		opendefer.closureNode = closure.Aux.(*Node)
+	} else {
+		if fn.Op != ODOTINTER {
+			Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op)
+		}
+		closure, rcvr := s.getClosureAndRcvr(fn)
+		opendefer.closure = s.openDeferSave(fn, closure.Type, closure)
+		// Important to get the receiver type correct, so it is recognized
+		// as a pointer for GC purposes.
+		opendefer.rcvr = s.openDeferSave(nil, fn.Type.Recv().Type, rcvr)
+		opendefer.closureNode = opendefer.closure.Aux.(*Node)
+		opendefer.rcvrNode = opendefer.rcvr.Aux.(*Node)
+	}
+	for _, argn := range n.Rlist.Slice() {
+		v := s.openDeferSave(argn, argn.Type, s.expr(argn))
+		args = append(args, v)
+		argNodes = append(argNodes, v.Aux.(*Node))
+	}
+	opendefer.argVals = args
+	opendefer.argNodes = argNodes
+	s.opendefers = append(s.opendefers, opendefer)
+
+	// Update deferBits only after evaluation and storage to stack of
+	// args/receiver/interface is successful.
+	bitvalue := s.constInt8(types.Types[TUINT8], 1<<uint(index))
+	newDeferBits := s.newValue2(ssa.OpOr8, types.Types[TUINT8], s.variable(&deferBitsVar, types.Types[TUINT8]), bitvalue)
+	s.vars[&deferBitsVar] = newDeferBits
+	s.store(types.Types[TUINT8], s.deferBitsAddr, newDeferBits)
+}
+
+// openDeferSave generates SSA nodes to store a value val (with type t) for an
+// open-coded defer on the stack at an explicit autotmp location, so it can be
+// reloaded and used for the appropriate call on exit. n is the associated node,
+// which is only needed if the associated type is non-SSAable. It returns an SSA
+// value representing a pointer to the stack location.
+func (s *state) openDeferSave(n *Node, t *types.Type, val *ssa.Value) *ssa.Value {
+	argTemp := tempAt(val.Pos.WithNotStmt(), s.curfn, t)
+	var addrArgTemp *ssa.Value
+	// Use OpVarLive to make sure stack slots for the args, etc. are not
+	// removed by dead-store elimination
+	if s.curBlock.ID != s.f.Entry.ID {
+		// Force the argtmp storing this defer function/receiver/arg to be
+		// declared in the entry block, so that it will be live for the
+		// defer exit code (which will actually access it only if the
+		// associated defer call has been activated).
+		s.defvars[s.f.Entry.ID][&memVar] = s.entryNewValue1A(ssa.OpVarDef, types.TypeMem, argTemp, s.defvars[s.f.Entry.ID][&memVar])
+		s.defvars[s.f.Entry.ID][&memVar] = s.entryNewValue1A(ssa.OpVarLive, types.TypeMem, argTemp, s.defvars[s.f.Entry.ID][&memVar])
+		addrArgTemp = s.entryNewValue2A(ssa.OpLocalAddr, types.NewPtr(argTemp.Type), argTemp, s.sp, s.defvars[s.f.Entry.ID][&memVar])
+	} else {
+		// Special case if we're still in the entry block. We can't use
+		// the above code, since s.defvars[s.f.Entry.ID] isn't defined
+		// until we end the entry block with s.endBlock().
+		s.vars[&memVar] = s.newValue1Apos(ssa.OpVarDef, types.TypeMem, argTemp, s.mem(), false)
+		s.vars[&memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, argTemp, s.mem(), false)
+		addrArgTemp = s.newValue2Apos(ssa.OpLocalAddr, types.NewPtr(argTemp.Type), argTemp, s.sp, s.mem(), false)
+	}
+	if types.Haspointers(t) {
+		// Since we may use this argTemp during exit depending on the
+		// deferBits, we must define it unconditionally on entry.
+		// Therefore, we must make sure it is zeroed out in the entry
+		// block if it contains pointers, else GC may wrongly follow an
+		// uninitialized pointer value.
+		argTemp.Name.SetNeedzero(true)
+	}
+	if !canSSAType(t) {
+		if n.Op != ONAME {
+			panic(fmt.Sprintf("Non-SSAable value should be a named location: %v", n))
+		}
+		a := s.addr(n, false)
+		s.move(t, addrArgTemp, a)
+		return addrArgTemp
+	}
+	// We are storing to the stack, hence we can avoid the full checks in
+	// storeType() (no write barrier) and do a simple store().
+	s.store(t, addrArgTemp, val)
+	return addrArgTemp
+}
+
+// openDeferExit generates SSA for processing all the open coded defers at exit.
+// The code involves loading deferBits, and checking each of the bits to see if
+// the corresponding defer statement was executed. For each bit that is turned
+// on, the associated defer call is made.
+func (s *state) openDeferExit() {
+	deferExit := s.f.NewBlock(ssa.BlockPlain)
+	s.endBlock().AddEdgeTo(deferExit)
+	s.startBlock(deferExit)
+	s.lastDeferExit = deferExit
+	s.lastDeferCount = len(s.opendefers)
+	zeroval := s.constInt8(types.Types[TUINT8], 0)
+	// Test for and run defers in reverse order
+	for i := len(s.opendefers) - 1; i >= 0; i-- {
+		r := s.opendefers[i]
+		bCond := s.f.NewBlock(ssa.BlockPlain)
+		bEnd := s.f.NewBlock(ssa.BlockPlain)
+
+		deferBits := s.variable(&deferBitsVar, types.Types[TUINT8])
+		// Generate code to check if the bit associated with the current
+		// defer is set.
+		bitval := s.constInt8(types.Types[TUINT8], 1<<uint(i))
+		andval := s.newValue2(ssa.OpAnd8, types.Types[TUINT8], deferBits, bitval)
+		eqVal := s.newValue2(ssa.OpEq8, types.Types[TBOOL], andval, zeroval)
+		b := s.endBlock()
+		b.Kind = ssa.BlockIf
+		b.SetControl(eqVal)
+		b.AddEdgeTo(bEnd)
+		b.AddEdgeTo(bCond)
+		bCond.AddEdgeTo(bEnd)
+		s.startBlock(bCond)
+
+		// Clear this bit in deferBits and force store back to stack, so
+		// we will not try to re-run this defer call if this defer call panics.
+		nbitval := s.newValue1(ssa.OpCom8, types.Types[TUINT8], bitval)
+		maskedval := s.newValue2(ssa.OpAnd8, types.Types[TUINT8], deferBits, nbitval)
+		s.store(types.Types[TUINT8], s.deferBitsAddr, maskedval)
+		// Use this value for following tests, so we keep previous
+		// bits cleared.
+		s.vars[&deferBitsVar] = maskedval
+
+		// Generate code to call the function call of the defer, using the
+		// closure/receiver/args that were stored in argtmps at the point
+		// of the defer statement.
+		argStart := Ctxt.FixedFrameSize()
+		fn := r.n.Left
+		stksize := fn.Type.ArgWidth()
+		if r.rcvr != nil {
+			// rcvr in case of OCALLINTER
+			v := s.load(r.rcvr.Type.Elem(), r.rcvr)
+			addr := s.constOffPtrSP(s.f.Config.Types.UintptrPtr, argStart)
+			s.store(types.Types[TUINTPTR], addr, v)
+		}
+		for j, argAddrVal := range r.argVals {
+			f := getParam(r.n, j)
+			pt := types.NewPtr(f.Type)
+			addr := s.constOffPtrSP(pt, argStart+f.Offset)
+			if !canSSAType(f.Type) {
+				s.move(f.Type, addr, argAddrVal)
+			} else {
+				argVal := s.load(f.Type, argAddrVal)
+				s.storeType(f.Type, addr, argVal, 0, false)
+			}
+		}
+		var call *ssa.Value
+		if r.closure != nil {
+			v := s.load(r.closure.Type.Elem(), r.closure)
+			s.maybeNilCheckClosure(v, callDefer)
+			codeptr := s.rawLoad(types.Types[TUINTPTR], v)
+			call = s.newValue3(ssa.OpClosureCall, types.TypeMem, codeptr, v, s.mem())
+		} else {
+			// Do a static call if the original call was a static function or method
+			call = s.newValue1A(ssa.OpStaticCall, types.TypeMem, fn.Sym.Linksym(), s.mem())
+		}
+		call.AuxInt = stksize
+		s.vars[&memVar] = call
+		// Make sure that the stack slots with pointers are kept live
+		// through the call (which is a pre-emption point). Also, we will
+		// use the first call of the last defer exit to compute liveness
+		// for the deferreturn, so we want all stack slots to be live.
+		if r.closureNode != nil {
+			s.vars[&memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.closureNode, s.mem(), false)
+		}
+		if r.rcvrNode != nil {
+			if types.Haspointers(r.rcvrNode.Type) {
+				s.vars[&memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, r.rcvrNode, s.mem(), false)
+			}
+		}
+		for _, argNode := range r.argNodes {
+			if types.Haspointers(argNode.Type) {
+				s.vars[&memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, argNode, s.mem(), false)
+			}
+		}
+
+		if i == len(s.opendefers)-1 {
+			// Record the call of the first defer. This will be used
+			// to set liveness info for the deferreturn (which is also
+			// used for any location that causes a runtime panic)
+			s.f.LastDeferExit = call
+		}
+		s.endBlock()
+		s.startBlock(bEnd)
+	}
+}
+
 // Calls the function n using the specified call type.
 // Returns the address of the return value (or nil if none).
 func (s *state) call(n *Node, k callKind) *ssa.Value {
@@ -3779,11 +4225,10 @@ func (s *state) call(n *Node, k callKind) *ssa.Value {
 			break
 		}
 		closure = s.expr(fn)
-		if k != callDefer && k != callDeferStack && (thearch.LinkArch.Family == sys.Wasm || objabi.GOOS == "aix" && k != callGo) {
-			// Deferred nil function needs to panic when the function is invoked, not the point of defer statement.
-			// On AIX, the closure needs to be verified as fn can be nil, except if it's a call go. This needs to be handled by the runtime to have the "go of nil func value" error.
-			// TODO(neelance): On other architectures this should be eliminated by the optimization steps
-			s.nilCheck(closure)
+		if k != callDefer && k != callDeferStack {
+			// Deferred nil function needs to panic when the function is invoked,
+			// not the point of defer statement.
+			s.maybeNilCheckClosure(closure, k)
 		}
 	case OCALLMETH:
 		if fn.Op != ODOTMETH {
@@ -3793,35 +4238,20 @@ func (s *state) call(n *Node, k callKind) *ssa.Value {
 			sym = fn.Sym
 			break
 		}
-		// Make a name n2 for the function.
-		// fn.Sym might be sync.(*Mutex).Unlock.
-		// Make a PFUNC node out of that, then evaluate it.
-		// We get back an SSA value representing &sync.(*Mutex).Unlock·f.
-		// We can then pass that to defer or go.
-		n2 := newnamel(fn.Pos, fn.Sym)
-		n2.Name.Curfn = s.curfn
-		n2.SetClass(PFUNC)
-		// n2.Sym already existed, so it's already marked as a function.
-		n2.Pos = fn.Pos
-		n2.Type = types.Types[TUINT8] // dummy type for a static closure. Could use runtime.funcval if we had it.
-		closure = s.expr(n2)
+		closure = s.getMethodClosure(fn)
 		// Note: receiver is already present in n.Rlist, so we don't
 		// want to set it here.
 	case OCALLINTER:
 		if fn.Op != ODOTINTER {
 			s.Fatalf("OCALLINTER: n.Left not an ODOTINTER: %v", fn.Op)
 		}
-		i := s.expr(fn.Left)
-		itab := s.newValue1(ssa.OpITab, types.Types[TUINTPTR], i)
-		s.nilCheck(itab)
-		itabidx := fn.Xoffset + 2*int64(Widthptr) + 8 // offset of fun field in runtime.itab
-		itab = s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
+		var iclosure *ssa.Value
+		iclosure, rcvr = s.getClosureAndRcvr(fn)
 		if k == callNormal {
-			codeptr = s.load(types.Types[TUINTPTR], itab)
+			codeptr = s.load(types.Types[TUINTPTR], iclosure)
 		} else {
-			closure = itab
+			closure = iclosure
 		}
-		rcvr = s.newValue1(ssa.OpIData, types.Types[TUINTPTR], i)
 	}
 	dowidth(fn.Type)
 	stksize := fn.Type.ArgWidth() // includes receiver, args, and results
@@ -3847,18 +4277,22 @@ func (s *state) call(n *Node, k callKind) *ssa.Value {
 			s.constInt32(types.Types[TUINT32], int32(stksize)))
 		// 1: started, set in deferprocStack
 		// 2: heap, set in deferprocStack
-		// 3: sp, set in deferprocStack
-		// 4: pc, set in deferprocStack
-		// 5: fn
+		// 3: openDefer
+		// 4: sp, set in deferprocStack
+		// 5: pc, set in deferprocStack
+		// 6: fn
 		s.store(closure.Type,
-			s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(5), addr),
+			s.newValue1I(ssa.OpOffPtr, closure.Type.PtrTo(), t.FieldOff(6), addr),
 			closure)
-		// 6: panic, set in deferprocStack
-		// 7: link, set in deferprocStack
+		// 7: panic, set in deferprocStack
+		// 8: link, set in deferprocStack
+		// 9: framepc
+		// 10: varp
+		// 11: fd
 
 		// Then, store all the arguments of the defer call.
 		ft := fn.Type
-		off := t.FieldOff(8)
+		off := t.FieldOff(12)
 		args := n.Rlist.Slice()
 
 		// Set receiver (for interface calls). Always a pointer.
@@ -3973,6 +4407,44 @@ func (s *state) call(n *Node, k callKind) *ssa.Value {
 	return s.constOffPtrSP(types.NewPtr(fp.Type), fp.Offset+Ctxt.FixedFrameSize())
 }
 
+// maybeNilCheckClosure checks if a nil check of a closure is needed in some
+// architecture-dependent situations and, if so, emits the nil check.
+func (s *state) maybeNilCheckClosure(closure *ssa.Value, k callKind) {
+	if thearch.LinkArch.Family == sys.Wasm || objabi.GOOS == "aix" && k != callGo {
+		// On AIX, the closure needs to be verified as fn can be nil, except if it's a call go. This needs to be handled by the runtime to have the "go of nil func value" error.
+		// TODO(neelance): On other architectures this should be eliminated by the optimization steps
+		s.nilCheck(closure)
+	}
+}
+
+// getMethodClosure returns a value representing the closure for a method call
+func (s *state) getMethodClosure(fn *Node) *ssa.Value {
+	// Make a name n2 for the function.
+	// fn.Sym might be sync.(*Mutex).Unlock.
+	// Make a PFUNC node out of that, then evaluate it.
+	// We get back an SSA value representing &sync.(*Mutex).Unlock·f.
+	// We can then pass that to defer or go.
+	n2 := newnamel(fn.Pos, fn.Sym)
+	n2.Name.Curfn = s.curfn
+	n2.SetClass(PFUNC)
+	// n2.Sym already existed, so it's already marked as a function.
+	n2.Pos = fn.Pos
+	n2.Type = types.Types[TUINT8] // dummy type for a static closure. Could use runtime.funcval if we had it.
+	return s.expr(n2)
+}
+
+// getClosureAndRcvr returns values for the appropriate closure and receiver of an
+// interface call
+func (s *state) getClosureAndRcvr(fn *Node) (*ssa.Value, *ssa.Value) {
+	i := s.expr(fn.Left)
+	itab := s.newValue1(ssa.OpITab, types.Types[TUINTPTR], i)
+	s.nilCheck(itab)
+	itabidx := fn.Xoffset + 2*int64(Widthptr) + 8 // offset of fun field in runtime.itab
+	closure := s.newValue1I(ssa.OpOffPtr, s.f.Config.Types.UintptrPtr, itabidx, itab)
+	rcvr := s.newValue1(ssa.OpIData, types.Types[TUINTPTR], i)
+	return closure, rcvr
+}
+
 // etypesign returns the signed-ness of e, for integer/pointer etypes.
 // -1 means signed, +1 means unsigned, 0 means non-integer/non-pointer.
 func etypesign(e types.EType) int8 {
@@ -5146,6 +5618,16 @@ func (s *state) addNamedValue(n *Node, v *ssa.Value) {
 	s.f.NamedValues[loc] = append(values, v)
 }
 
+// Generate a disconnected call to a runtime routine and a return.
+func gencallret(pp *Progs, sym *obj.LSym) *obj.Prog {
+	p := pp.Prog(obj.ACALL)
+	p.To.Type = obj.TYPE_MEM
+	p.To.Name = obj.NAME_EXTERN
+	p.To.Sym = sym
+	p = pp.Prog(obj.ARET)
+	return p
+}
+
 // Branch is an unresolved branch.
 type Branch struct {
 	P *obj.Prog  // branch instruction
@@ -5181,6 +5663,11 @@ type SSAGenState struct {
 
 	// wasm: The number of values on the WebAssembly stack. This is only used as a safeguard.
 	OnWasmStackSkipped int
+
+	// Liveness index for the first function call in the final defer exit code
+	// path that we generated. All defer functions and args should be live at
+	// this point. This will be used to set the liveness for the deferreturn.
+	lastDeferLiveness LivenessIndex
 }
 
 // Prog appends a new Prog.
@@ -5308,6 +5795,17 @@ func genssa(f *ssa.Func, pp *Progs) {
 	s.livenessMap = liveness(e, f, pp)
 	emitStackObjects(e, pp)
 
+	openDeferInfo := e.curfn.Func.lsym.Func.OpenCodedDeferInfo
+	if openDeferInfo != nil {
+		// This function uses open-coded defers -- write out the funcdata
+		// info that we computed at the end of genssa.
+		p := pp.Prog(obj.AFUNCDATA)
+		Addrconst(&p.From, objabi.FUNCDATA_OpenCodedDeferInfo)
+		p.To.Type = obj.TYPE_MEM
+		p.To.Name = obj.NAME_EXTERN
+		p.To.Sym = openDeferInfo
+	}
+
 	// Remember where each block starts.
 	s.bstart = make([]*obj.Prog, f.NumBlocks())
 	s.pp = pp
@@ -5372,6 +5870,12 @@ func genssa(f *ssa.Func, pp *Progs) {
 			// Attach this safe point to the next
 			// instruction.
 			s.pp.nextLive = s.livenessMap.Get(v)
+
+			// Remember the liveness index of the first defer call of
+			// the last defer exit
+			if v.Block.Func.LastDeferExit != nil && v == v.Block.Func.LastDeferExit {
+				s.lastDeferLiveness = s.pp.nextLive
+			}
 			switch v.Op {
 			case ssa.OpInitMem:
 				// memory arg needs no code
@@ -5455,6 +5959,13 @@ func genssa(f *ssa.Func, pp *Progs) {
 		// nop (which will never execute) after the call.
 		thearch.Ginsnop(pp)
 	}
+	if openDeferInfo != nil {
+		// When doing open-coded defers, generate a disconnected call to
+		// deferreturn and a return. This will be used to during panic
+		// recovery to unwind the stack and return back to the runtime.
+		s.pp.nextLive = s.lastDeferLiveness
+		gencallret(pp, Deferreturn)
+	}
 
 	if inlMarks != nil {
 		// We have some inline marks. Try to find other instructions we're
diff --git a/src/cmd/compile/internal/gc/syntax.go b/src/cmd/compile/internal/gc/syntax.go
index 98903100fc..e0137acfb7 100644
--- a/src/cmd/compile/internal/gc/syntax.go
+++ b/src/cmd/compile/internal/gc/syntax.go
@@ -491,7 +491,9 @@ type Func struct {
 
 	Pragma syntax.Pragma // go:xxx function annotations
 
-	flags bitset16
+	flags      bitset16
+	numDefers  int // number of defer calls in the function
+	numReturns int // number of explicit returns in the function
 
 	// nwbrCalls records the LSyms of functions called by this
 	// function for go:nowritebarrierrec analysis. Only filled in
@@ -527,34 +529,37 @@ const (
 	funcNeedctxt                  // function uses context register (has closure variables)
 	funcReflectMethod             // function calls reflect.Type.Method or MethodByName
 	funcIsHiddenClosure
-	funcHasDefer            // contains a defer statement
-	funcNilCheckDisabled    // disable nil checks when compiling this function
-	funcInlinabilityChecked // inliner has already determined whether the function is inlinable
-	funcExportInline        // include inline body in export data
-	funcInstrumentBody      // add race/msan instrumentation during SSA construction
+	funcHasDefer                 // contains a defer statement
+	funcNilCheckDisabled         // disable nil checks when compiling this function
+	funcInlinabilityChecked      // inliner has already determined whether the function is inlinable
+	funcExportInline             // include inline body in export data
+	funcInstrumentBody           // add race/msan instrumentation during SSA construction
+	funcOpenCodedDeferDisallowed // can't do open-coded defers
 )
 
-func (f *Func) Dupok() bool               { return f.flags&funcDupok != 0 }
-func (f *Func) Wrapper() bool             { return f.flags&funcWrapper != 0 }
-func (f *Func) Needctxt() bool            { return f.flags&funcNeedctxt != 0 }
-func (f *Func) ReflectMethod() bool       { return f.flags&funcReflectMethod != 0 }
-func (f *Func) IsHiddenClosure() bool     { return f.flags&funcIsHiddenClosure != 0 }
-func (f *Func) HasDefer() bool            { return f.flags&funcHasDefer != 0 }
-func (f *Func) NilCheckDisabled() bool    { return f.flags&funcNilCheckDisabled != 0 }
-func (f *Func) InlinabilityChecked() bool { return f.flags&funcInlinabilityChecked != 0 }
-func (f *Func) ExportInline() bool        { return f.flags&funcExportInline != 0 }
-func (f *Func) InstrumentBody() bool      { return f.flags&funcInstrumentBody != 0 }
+func (f *Func) Dupok() bool                    { return f.flags&funcDupok != 0 }
+func (f *Func) Wrapper() bool                  { return f.flags&funcWrapper != 0 }
+func (f *Func) Needctxt() bool                 { return f.flags&funcNeedctxt != 0 }
+func (f *Func) ReflectMethod() bool            { return f.flags&funcReflectMethod != 0 }
+func (f *Func) IsHiddenClosure() bool          { return f.flags&funcIsHiddenClosure != 0 }
+func (f *Func) HasDefer() bool                 { return f.flags&funcHasDefer != 0 }
+func (f *Func) NilCheckDisabled() bool         { return f.flags&funcNilCheckDisabled != 0 }
+func (f *Func) InlinabilityChecked() bool      { return f.flags&funcInlinabilityChecked != 0 }
+func (f *Func) ExportInline() bool             { return f.flags&funcExportInline != 0 }
+func (f *Func) InstrumentBody() bool           { return f.flags&funcInstrumentBody != 0 }
+func (f *Func) OpenCodedDeferDisallowed() bool { return f.flags&funcOpenCodedDeferDisallowed != 0 }
 
-func (f *Func) SetDupok(b bool)               { f.flags.set(funcDupok, b) }
-func (f *Func) SetWrapper(b bool)             { f.flags.set(funcWrapper, b) }
-func (f *Func) SetNeedctxt(b bool)            { f.flags.set(funcNeedctxt, b) }
-func (f *Func) SetReflectMethod(b bool)       { f.flags.set(funcReflectMethod, b) }
-func (f *Func) SetIsHiddenClosure(b bool)     { f.flags.set(funcIsHiddenClosure, b) }
-func (f *Func) SetHasDefer(b bool)            { f.flags.set(funcHasDefer, b) }
-func (f *Func) SetNilCheckDisabled(b bool)    { f.flags.set(funcNilCheckDisabled, b) }
-func (f *Func) SetInlinabilityChecked(b bool) { f.flags.set(funcInlinabilityChecked, b) }
-func (f *Func) SetExportInline(b bool)        { f.flags.set(funcExportInline, b) }
-func (f *Func) SetInstrumentBody(b bool)      { f.flags.set(funcInstrumentBody, b) }
+func (f *Func) SetDupok(b bool)                    { f.flags.set(funcDupok, b) }
+func (f *Func) SetWrapper(b bool)                  { f.flags.set(funcWrapper, b) }
+func (f *Func) SetNeedctxt(b bool)                 { f.flags.set(funcNeedctxt, b) }
+func (f *Func) SetReflectMethod(b bool)            { f.flags.set(funcReflectMethod, b) }
+func (f *Func) SetIsHiddenClosure(b bool)          { f.flags.set(funcIsHiddenClosure, b) }
+func (f *Func) SetHasDefer(b bool)                 { f.flags.set(funcHasDefer, b) }
+func (f *Func) SetNilCheckDisabled(b bool)         { f.flags.set(funcNilCheckDisabled, b) }
+func (f *Func) SetInlinabilityChecked(b bool)      { f.flags.set(funcInlinabilityChecked, b) }
+func (f *Func) SetExportInline(b bool)             { f.flags.set(funcExportInline, b) }
+func (f *Func) SetInstrumentBody(b bool)           { f.flags.set(funcInstrumentBody, b) }
+func (f *Func) SetOpenCodedDeferDisallowed(b bool) { f.flags.set(funcOpenCodedDeferDisallowed, b) }
 
 func (f *Func) setWBPos(pos src.XPos) {
 	if Debug_wb != 0 {
diff --git a/src/cmd/compile/internal/gc/walk.go b/src/cmd/compile/internal/gc/walk.go
index 39d1ab689d..b064e7d71c 100644
--- a/src/cmd/compile/internal/gc/walk.go
+++ b/src/cmd/compile/internal/gc/walk.go
@@ -213,7 +213,6 @@ func walkstmt(n *Node) *Node {
 		yyerror("case statement out of place")
 
 	case ODEFER:
-		Curfn.Func.SetHasDefer(true)
 		fallthrough
 	case OGO:
 		switch n.Left.Op {
diff --git a/src/cmd/compile/internal/ssa/deadstore.go b/src/cmd/compile/internal/ssa/deadstore.go
index 6b9bcedadb..88af7a6f4a 100644
--- a/src/cmd/compile/internal/ssa/deadstore.go
+++ b/src/cmd/compile/internal/ssa/deadstore.go
@@ -170,6 +170,11 @@ func elimDeadAutosGeneric(f *Func) {
 			return
 		case OpVarLive:
 			// Don't delete the auto if it needs to be kept alive.
+
+			// We depend on this check to keep the autotmp stack slots
+			// for open-coded defers from being removed (since they
+			// may not be used by the inline code, but will be used by
+			// panic processing).
 			n, ok := v.Aux.(GCNode)
 			if !ok || n.StorageClass() != ClassAuto {
 				return
diff --git a/src/cmd/compile/internal/ssa/func.go b/src/cmd/compile/internal/ssa/func.go
index cdd5161913..332e201899 100644
--- a/src/cmd/compile/internal/ssa/func.go
+++ b/src/cmd/compile/internal/ssa/func.go
@@ -32,8 +32,16 @@ type Func struct {
 	Type   *types.Type // type signature of the function.
 	Blocks []*Block    // unordered set of all basic blocks (note: not indexable by ID)
 	Entry  *Block      // the entry basic block
-	bid    idAlloc     // block ID allocator
-	vid    idAlloc     // value ID allocator
+
+	// If we are using open-coded defers, this is the first call to a deferred
+	// function in the final defer exit sequence that we generated. This call
+	// should be after all defer statements, and will have all args, etc. of
+	// all defer calls as live. The liveness info of this call will be used
+	// for the deferreturn/ret segment generated for functions with open-coded
+	// defers.
+	LastDeferExit *Value
+	bid           idAlloc // block ID allocator
+	vid           idAlloc // value ID allocator
 
 	// Given an environment variable used for debug hash match,
 	// what file (if any) receives the yes/no logging?
diff --git a/src/cmd/internal/obj/link.go b/src/cmd/internal/obj/link.go
index 1c101bfc27..468e9402ee 100644
--- a/src/cmd/internal/obj/link.go
+++ b/src/cmd/internal/obj/link.go
@@ -405,10 +405,11 @@ type FuncInfo struct {
 	dwarfAbsFnSym      *LSym
 	dwarfDebugLinesSym *LSym
 
-	GCArgs       *LSym
-	GCLocals     *LSym
-	GCRegs       *LSym
-	StackObjects *LSym
+	GCArgs             *LSym
+	GCLocals           *LSym
+	GCRegs             *LSym
+	StackObjects       *LSym
+	OpenCodedDeferInfo *LSym
 }
 
 type InlMark struct {
diff --git a/src/cmd/internal/objabi/funcdata.go b/src/cmd/internal/objabi/funcdata.go
index addbd2ac88..08b75eb9fe 100644
--- a/src/cmd/internal/objabi/funcdata.go
+++ b/src/cmd/internal/objabi/funcdata.go
@@ -15,11 +15,12 @@ const (
 	PCDATA_StackMapIndex = 1
 	PCDATA_InlTreeIndex  = 2
 
-	FUNCDATA_ArgsPointerMaps   = 0
-	FUNCDATA_LocalsPointerMaps = 1
-	FUNCDATA_RegPointerMaps    = 2
-	FUNCDATA_StackObjects      = 3
-	FUNCDATA_InlTree           = 4
+	FUNCDATA_ArgsPointerMaps    = 0
+	FUNCDATA_LocalsPointerMaps  = 1
+	FUNCDATA_RegPointerMaps     = 2
+	FUNCDATA_StackObjects       = 3
+	FUNCDATA_InlTree            = 4
+	FUNCDATA_OpenCodedDeferInfo = 5
 
 	// ArgsSizeUnknown is set in Func.argsize to mark all functions
 	// whose argument size is unknown (C vararg functions, and
diff --git a/src/cmd/internal/objabi/funcid.go b/src/cmd/internal/objabi/funcid.go
index c13c3cb458..487f009830 100644
--- a/src/cmd/internal/objabi/funcid.go
+++ b/src/cmd/internal/objabi/funcid.go
@@ -85,6 +85,12 @@ func GetFuncID(name, file string) FuncID {
 		return FuncID_panicwrap
 	case "runtime.handleAsyncEvent":
 		return FuncID_handleAsyncEvent
+	case "runtime.deferreturn":
+		// Don't show in the call stack (used when invoking defer functions)
+		return FuncID_wrapper
+	case "runtime.runOpenDeferFrame":
+		// Don't show in the call stack (used when invoking defer functions)
+		return FuncID_wrapper
 	}
 	if file == "<autogenerated>" {
 		return FuncID_wrapper
diff --git a/src/cmd/internal/objabi/stack.go b/src/cmd/internal/objabi/stack.go
index 62ab0398a6..7320dbf365 100644
--- a/src/cmd/internal/objabi/stack.go
+++ b/src/cmd/internal/objabi/stack.go
@@ -18,7 +18,7 @@ const (
 )
 
 // Initialize StackGuard and StackLimit according to target system.
-var StackGuard = 880*stackGuardMultiplier() + StackSystem
+var StackGuard = 896*stackGuardMultiplier() + StackSystem
 var StackLimit = StackGuard - StackSystem - StackSmall
 
 // stackGuardMultiplier returns a multiplier to apply to the default
diff --git a/src/cmd/link/internal/ld/pcln.go b/src/cmd/link/internal/ld/pcln.go
index d9904f9093..3d7b42b187 100644
--- a/src/cmd/link/internal/ld/pcln.go
+++ b/src/cmd/link/internal/ld/pcln.go
@@ -11,6 +11,7 @@ import (
 	"cmd/internal/sys"
 	"cmd/link/internal/sym"
 	"encoding/binary"
+	"fmt"
 	"log"
 	"os"
 	"path/filepath"
@@ -255,13 +256,23 @@ func (ctxt *Link) pclntab() {
 			}
 			if r.Type.IsDirectJump() && r.Sym != nil && r.Sym.Name == "runtime.deferreturn" {
 				if ctxt.Arch.Family == sys.Wasm {
-					deferreturn = lastWasmAddr
+					deferreturn = lastWasmAddr - 1
 				} else {
 					// Note: the relocation target is in the call instruction, but
 					// is not necessarily the whole instruction (for instance, on
 					// x86 the relocation applies to bytes [1:5] of the 5 byte call
 					// instruction).
 					deferreturn = uint32(r.Off)
+					switch ctxt.Arch.Family {
+					case sys.AMD64, sys.I386, sys.MIPS, sys.MIPS64, sys.RISCV64:
+						deferreturn--
+					case sys.PPC64, sys.ARM, sys.ARM64:
+						// no change
+					case sys.S390X:
+						deferreturn -= 2
+					default:
+						panic(fmt.Sprint("Unhandled architecture:", ctxt.Arch.Family))
+					}
 				}
 				break // only need one
 			}
diff --git a/src/cmd/link/internal/ld/symtab.go b/src/cmd/link/internal/ld/symtab.go
index d686a8a476..b4236a5239 100644
--- a/src/cmd/link/internal/ld/symtab.go
+++ b/src/cmd/link/internal/ld/symtab.go
@@ -498,7 +498,8 @@ func (ctxt *Link) symtab() {
 		case strings.HasPrefix(s.Name, "gcargs."),
 			strings.HasPrefix(s.Name, "gclocals."),
 			strings.HasPrefix(s.Name, "gclocals·"),
-			strings.HasPrefix(s.Name, "inltree."):
+			strings.HasPrefix(s.Name, "inltree."),
+			strings.HasSuffix(s.Name, ".opendefer"):
 			s.Type = sym.SGOFUNC
 			s.Attr |= sym.AttrNotInSymbolTable
 			s.Outer = symgofunc