diff --git a/misc/cgo/test/cgo_linux_test.go b/misc/cgo/test/cgo_linux_test.go
index a9746b552ee3f..7c4628c49330c 100644
--- a/misc/cgo/test/cgo_linux_test.go
+++ b/misc/cgo/test/cgo_linux_test.go
@@ -15,6 +15,14 @@ func TestSetgid(t *testing.T) {
 	}
 	testSetgid(t)
 }
+
+func TestSetgidStress(t *testing.T) {
+	if runtime.GOOS == "android" {
+		t.Skip("unsupported on Android")
+	}
+	testSetgidStress(t)
+}
+
 func Test1435(t *testing.T)    { test1435(t) }
 func Test6997(t *testing.T)    { test6997(t) }
 func TestBuildID(t *testing.T) { testBuildID(t) }
diff --git a/misc/cgo/test/setgid2_linux.go b/misc/cgo/test/setgid2_linux.go
new file mode 100644
index 0000000000000..d239893f43ed2
--- /dev/null
+++ b/misc/cgo/test/setgid2_linux.go
@@ -0,0 +1,35 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Stress test setgid and thread creation. A thread
+// can get a SIGSETXID signal early on at thread
+// initialization, causing crash. See issue 53374.
+
+package cgotest
+
+/*
+#include <sys/types.h>
+#include <unistd.h>
+*/
+import "C"
+
+import (
+	"runtime"
+	"testing"
+)
+
+func testSetgidStress(t *testing.T) {
+	const N = 1000
+	ch := make(chan int, N)
+	for i := 0; i < N; i++ {
+		go func() {
+			C.setgid(0)
+			ch <- 1
+			runtime.LockOSThread() // so every goroutine uses a new thread
+		}()
+	}
+	for i := 0; i < N; i++ {
+		<-ch
+	}
+}
diff --git a/src/cmd/internal/obj/arm64/obj7.go b/src/cmd/internal/obj/arm64/obj7.go
index 1f2625d54f85d..83ae64a19b8e6 100644
--- a/src/cmd/internal/obj/arm64/obj7.go
+++ b/src/cmd/internal/obj/arm64/obj7.go
@@ -609,17 +609,17 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 			var prologueEnd *obj.Prog
 
 			aoffset := c.autosize
-			if aoffset > 0x1f0 {
-				// LDP offset variant range is -512 to 504, SP should be 16-byte aligned,
-				// so the maximum aoffset value is 496.
-				aoffset = 0x1f0
+			if aoffset > 0xf0 {
+				// MOVD.W offset variant range is -0x100 to 0xf8, SP should be 16-byte aligned.
+				// so the maximum aoffset value is 0xf0.
+				aoffset = 0xf0
 			}
 
 			// Frame is non-empty. Make sure to save link register, even if
 			// it is a leaf function, so that traceback works.
 			q = p
 			if c.autosize > aoffset {
-				// Frame size is too large for a STP instruction. Store the frame pointer
+				// Frame size is too large for a MOVD.W instruction. Store the frame pointer
 				// register and link register before decrementing SP, so if a signal comes
 				// during the execution of the function prologue, the traceback code will
 				// not see a half-updated stack frame.
@@ -679,50 +679,37 @@ func preprocess(ctxt *obj.Link, cursym *obj.LSym, newprog obj.ProgAlloc) {
 					q1.To.Offset = -8
 				}
 			} else {
-				// small frame, save FP and LR with one STP instruction, then update SP.
-				// Store first, so if a signal comes during the execution of the function
-				// prologue, the traceback code will not see a half-updated stack frame.
-				// STP (R29, R30), -aoffset-8(RSP)
+				// small frame, update SP and save LR in a single MOVD.W instruction.
+				// So if a signal comes during the execution of the function prologue,
+				// the traceback code will not see a half-updated stack frame.
+				// Also, on Linux, in a cgo binary we may get a SIGSETXID signal
+				// early on before the signal stack is set, as glibc doesn't allow
+				// us to block SIGSETXID. So it is important that we don't write below
+				// the SP until the signal stack is set.
+				// Luckily, all the functions from thread entry to setting the signal
+				// stack have small frames.
 				q1 = obj.Appendp(q, c.newprog)
-				q1.As = ASTP
+				q1.As = AMOVD
 				q1.Pos = p.Pos
-				q1.From.Type = obj.TYPE_REGREG
-				q1.From.Reg = REGFP
-				q1.From.Offset = REGLINK
+				q1.From.Type = obj.TYPE_REG
+				q1.From.Reg = REGLINK
 				q1.To.Type = obj.TYPE_MEM
-				q1.To.Offset = int64(-aoffset - 8)
+				q1.Scond = C_XPRE
+				q1.To.Offset = int64(-aoffset)
 				q1.To.Reg = REGSP
+				q1.Spadj = aoffset
 
 				prologueEnd = q1
 
-				q1 = c.ctxt.StartUnsafePoint(q1, c.newprog)
-				// This instruction is not async preemptible, see the above comment.
-				// SUB $aoffset, RSP, RSP
+				// Frame pointer.
 				q1 = obj.Appendp(q1, c.newprog)
 				q1.Pos = p.Pos
-				q1.As = ASUB
-				q1.From.Type = obj.TYPE_CONST
-				q1.From.Offset = int64(aoffset)
-				q1.Reg = REGSP
-				q1.To.Type = obj.TYPE_REG
+				q1.As = AMOVD
+				q1.From.Type = obj.TYPE_REG
+				q1.From.Reg = REGFP
+				q1.To.Type = obj.TYPE_MEM
 				q1.To.Reg = REGSP
-				q1.Spadj = aoffset
-
-				q1 = c.ctxt.EndUnsafePoint(q1, c.newprog, -1)
-
-				if buildcfg.GOOS == "ios" {
-					// See the above comment.
-					// STP (R29, R30), -8(RSP)
-					q1 = obj.Appendp(q1, c.newprog)
-					q1.As = ASTP
-					q1.Pos = p.Pos
-					q1.From.Type = obj.TYPE_REGREG
-					q1.From.Reg = REGFP
-					q1.From.Offset = REGLINK
-					q1.To.Type = obj.TYPE_MEM
-					q1.To.Offset = int64(-8)
-					q1.To.Reg = REGSP
-				}
+				q1.To.Offset = -8
 			}
 
 			prologueEnd.Pos = prologueEnd.Pos.WithXlogue(src.PosPrologueEnd)