diff --git a/cl/_testdata/print/out.ll b/cl/_testdata/print/out.ll
index 9f89186a..8fbc290b 100644
--- a/cl/_testdata/print/out.ll
+++ b/cl/_testdata/print/out.ll
@@ -85,7 +85,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   store i64 0, ptr @main.minhexdigits, align 4
   br label %_llgo_2
 
@@ -1412,7 +1412,7 @@ declare i32 @printf(ptr, ...)
 
 declare void @"github.com/goplus/llgo/internal/runtime.init"()
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_float32, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testdata/vargs/out.ll b/cl/_testdata/vargs/out.ll
index a687457b..2dcb1327 100644
--- a/cl/_testdata/vargs/out.ll
+++ b/cl/_testdata/vargs/out.ll
@@ -19,7 +19,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -118,7 +118,7 @@ declare void @"github.com/goplus/llgo/internal/runtime.init"()
 
 declare ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64)
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_int, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testgo/defer/out.ll b/cl/_testgo/defer/out.ll
index b37f04be..6ef57dab 100644
--- a/cl/_testgo/defer/out.ll
+++ b/cl/_testgo/defer/out.ll
@@ -3,7 +3,7 @@ source_filename = "main"
 
 %"github.com/goplus/llgo/internal/runtime.String" = type { ptr, i64 }
 %"github.com/goplus/llgo/internal/runtime.eface" = type { ptr, ptr }
-%"github.com/goplus/llgo/internal/runtime.Defer" = type { { ptr, ptr }, i64, ptr }
+%"github.com/goplus/llgo/internal/runtime.Defer" = type { { ptr, ptr }, i64, ptr, i64 }
 
 @"main.init$guard" = global ptr null
 @0 = private unnamed_addr constant [6 x i8] c"error\00", align 1
@@ -67,75 +67,119 @@ _llgo_0:
   call void @main.init()
   %2 = load ptr, ptr @__llgo_defer, align 8
   %3 = call ptr @pthread_getspecific(ptr %2)
-  %4 = alloca i8, i64 32, align 1
+  %4 = alloca i8, i64 40, align 1
   %5 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 0
-  store ptr @"main$_llgo_defer", ptr %5, align 8
+  store ptr null, ptr %5, align 8
   %6 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
   store i64 0, ptr %6, align 4
   %7 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 2
   store ptr %3, ptr %7, align 8
   %8 = call i32 @pthread_setspecific(ptr %2, ptr %4)
   %9 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
-  %10 = load i64, ptr %9, align 4
-  %11 = or i64 %10, 1
-  store i64 %11, ptr %9, align 4
-  %12 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
-  %13 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %12, i32 0, i32 0
-  store ptr @1, ptr %13, align 8
-  %14 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %12, i32 0, i32 1
-  store i64 5, ptr %14, align 4
-  %15 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %12, align 8
-  %16 = call i1 @main.f(%"github.com/goplus/llgo/internal/runtime.String" %15)
-  br i1 %16, label %_llgo_2, label %_llgo_4
+  %10 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 3
+  %11 = load i64, ptr %9, align 4
+  %12 = or i64 %11, 1
+  store i64 %12, ptr %9, align 4
+  %13 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
+  %14 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %13, i32 0, i32 0
+  store ptr @1, ptr %14, align 8
+  %15 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %13, i32 0, i32 1
+  store i64 5, ptr %15, align 4
+  %16 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %13, align 8
+  %17 = call i1 @main.f(%"github.com/goplus/llgo/internal/runtime.String" %16)
+  br i1 %17, label %_llgo_2, label %_llgo_4
 
 _llgo_1:                                          ; No predecessors!
   ret i32 0
 
 _llgo_2:                                          ; preds = %_llgo_0
-  %17 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
-  %18 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %17, i32 0, i32 0
-  store ptr @2, ptr %18, align 8
-  %19 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %17, i32 0, i32 1
-  store i64 5, ptr %19, align 4
-  %20 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %17, align 8
-  %21 = load ptr, ptr @__llgo_defer, align 8
-  %22 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
-  %23 = load i64, ptr %22, align 4
-  %24 = or i64 %23, 2
-  store i64 %24, ptr %22, align 4
+  %18 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
+  %19 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %18, i32 0, i32 0
+  store ptr @2, ptr %19, align 8
+  %20 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %18, i32 0, i32 1
+  store i64 5, ptr %20, align 4
+  %21 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %18, align 8
+  %22 = load i64, ptr %9, align 4
+  %23 = or i64 %22, 2
+  store i64 %23, ptr %9, align 4
   br label %_llgo_3
 
 _llgo_3:                                          ; preds = %_llgo_4, %_llgo_2
-  %25 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
-  %26 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %25, i32 0, i32 0
-  store ptr @3, ptr %26, align 8
-  %27 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %25, i32 0, i32 1
-  store i64 3, ptr %27, align 4
-  %28 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %25, align 8
-  %29 = load ptr, ptr @__llgo_defer, align 8
-  %30 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
-  %31 = load i64, ptr %30, align 4
-  %32 = or i64 %31, 4
-  store i64 %32, ptr %30, align 4
-  %33 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
-  %34 = load i64, ptr %33, align 4
-  call void @"main$_llgo_defer"(i64 %34)
-  ret i32 0
+  %24 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
+  %25 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %24, i32 0, i32 0
+  store ptr @3, ptr %25, align 8
+  %26 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %24, i32 0, i32 1
+  store i64 3, ptr %26, align 4
+  %27 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %24, align 8
+  %28 = load i64, ptr %9, align 4
+  %29 = or i64 %28, 4
+  store i64 %29, ptr %9, align 4
+  store i64 0, ptr %10, align 4
+  br label %_llgo_5
 
 _llgo_4:                                          ; preds = %_llgo_0
-  %35 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
-  %36 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %35, i32 0, i32 0
-  store ptr @4, ptr %36, align 8
-  %37 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %35, i32 0, i32 1
-  store i64 5, ptr %37, align 4
-  %38 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %35, align 8
-  %39 = load ptr, ptr @__llgo_defer, align 8
-  %40 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, i32 0, i32 1
-  %41 = load i64, ptr %40, align 4
-  %42 = or i64 %41, 8
-  store i64 %42, ptr %40, align 4
+  %30 = alloca %"github.com/goplus/llgo/internal/runtime.String", align 8
+  %31 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %30, i32 0, i32 0
+  store ptr @4, ptr %31, align 8
+  %32 = getelementptr inbounds %"github.com/goplus/llgo/internal/runtime.String", ptr %30, i32 0, i32 1
+  store i64 5, ptr %32, align 4
+  %33 = load %"github.com/goplus/llgo/internal/runtime.String", ptr %30, align 8
+  %34 = load i64, ptr %9, align 4
+  %35 = or i64 %34, 8
+  store i64 %35, ptr %9, align 4
   call void @main.fail()
   br label %_llgo_3
+
+_llgo_5:                                          ; preds = %_llgo_3
+  %36 = load i64, ptr %9, align 4
+  %37 = and i64 %36, 8
+  %38 = icmp ne i64 %37, 0
+  br i1 %38, label %_llgo_7, label %_llgo_8
+
+_llgo_6:                                          ; preds = %_llgo_14
+  ret i32 0
+
+_llgo_7:                                          ; preds = %_llgo_5
+  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %33)
+  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
+  br label %_llgo_8
+
+_llgo_8:                                          ; preds = %_llgo_7, %_llgo_5
+  %39 = and i64 %36, 4
+  %40 = icmp ne i64 %39, 0
+  br i1 %40, label %_llgo_9, label %_llgo_10
+
+_llgo_9:                                          ; preds = %_llgo_8
+  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %27)
+  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
+  br label %_llgo_10
+
+_llgo_10:                                         ; preds = %_llgo_9, %_llgo_8
+  %41 = and i64 %36, 2
+  %42 = icmp ne i64 %41, 0
+  br i1 %42, label %_llgo_11, label %_llgo_12
+
+_llgo_11:                                         ; preds = %_llgo_10
+  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %21)
+  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
+  br label %_llgo_12
+
+_llgo_12:                                         ; preds = %_llgo_11, %_llgo_10
+  %43 = and i64 %36, 1
+  %44 = icmp ne i64 %43, 0
+  br i1 %44, label %_llgo_13, label %_llgo_14
+
+_llgo_13:                                         ; preds = %_llgo_12
+  call void @"main.main$1"()
+  br label %_llgo_14
+
+_llgo_14:                                         ; preds = %_llgo_13, %_llgo_12
+  %45 = load %"github.com/goplus/llgo/internal/runtime.Defer", ptr %4, align 8
+  %46 = extractvalue %"github.com/goplus/llgo/internal/runtime.Defer" %45, 2
+  %47 = call i32 @pthread_setspecific(ptr %2, ptr %46)
+  %48 = load i64, ptr %10, align 4
+  switch i64 %48, label %_llgo_6 [
+  ]
 }
 
 define void @"main.init$after"() {
@@ -183,50 +227,6 @@ _llgo_0:
   ret void
 }
 
-define void @"main$_llgo_defer"(i64 %0) {
-_llgo_0:
-  %1 = and i64 %0, 1
-  %2 = icmp ne i64 %1, 0
-  br i1 %2, label %_llgo_1, label %_llgo_2
-
-_llgo_1:                                          ; preds = %_llgo_0
-  call void @"main.main$1"()
-  br label %_llgo_2
-
-_llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
-  %3 = and i64 %0, 2
-  %4 = icmp ne i64 %3, 0
-  br i1 %4, label %_llgo_3, label %_llgo_4
-
-_llgo_3:                                          ; preds = %_llgo_2
-  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %20)
-  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
-  br label %_llgo_4
-
-_llgo_4:                                          ; preds = %_llgo_3, %_llgo_2
-  %5 = and i64 %0, 4
-  %6 = icmp ne i64 %5, 0
-  br i1 %6, label %_llgo_5, label %_llgo_6
-
-_llgo_5:                                          ; preds = %_llgo_4
-  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %28)
-  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
-  br label %_llgo_6
-
-_llgo_6:                                          ; preds = %_llgo_5, %_llgo_4
-  %7 = and i64 %0, 8
-  %8 = icmp ne i64 %7, 0
-  br i1 %8, label %_llgo_7, label %_llgo_8
-
-_llgo_7:                                          ; preds = %_llgo_6
-  call void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com/goplus/llgo/internal/runtime.String" %38)
-  call void @"github.com/goplus/llgo/internal/runtime.PrintByte"(i8 10)
-  br label %_llgo_8
-
-_llgo_8:                                          ; preds = %_llgo_7, %_llgo_6
-  ret void
-}
-
 declare ptr @pthread_getspecific(i32)
 
 declare i32 @pthread_setspecific(i32, ptr)
diff --git a/cl/_testgo/errors/out.ll b/cl/_testgo/errors/out.ll
index 599cc745..56719002 100644
--- a/cl/_testgo/errors/out.ll
+++ b/cl/_testgo/errors/out.ll
@@ -60,7 +60,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -102,7 +102,7 @@ _llgo_0:
 
 declare ptr @"github.com/goplus/llgo/internal/runtime.AllocZ"(i64)
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = call ptr @"github.com/goplus/llgo/internal/runtime.NewNamed"(i64 25, i64 0, i64 1)
   store ptr %0, ptr @_llgo_main.errorString, align 8
diff --git a/cl/_testgo/strucintf/out.ll b/cl/_testgo/strucintf/out.ll
index f23f046d..57632d34 100644
--- a/cl/_testgo/strucintf/out.ll
+++ b/cl/_testgo/strucintf/out.ll
@@ -49,7 +49,7 @@ _llgo_0:
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
   call void @"github.com/goplus/llgo/cl/internal/foo.init"()
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -241,7 +241,7 @@ _llgo_18:                                         ; preds = %_llgo_17, %_llgo_16
 
 declare ptr @"github.com/goplus/llgo/internal/runtime.Zeroinit"(ptr, i64)
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_int, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testgo/struczero/out.ll b/cl/_testgo/struczero/out.ll
index 2bd65e65..48894dd4 100644
--- a/cl/_testgo/struczero/out.ll
+++ b/cl/_testgo/struczero/out.ll
@@ -117,7 +117,7 @@ _llgo_0:
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
   call void @"github.com/goplus/llgo/cl/internal/foo.init"()
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -185,7 +185,7 @@ _llgo_0:
   ret i32 0
 }
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @"_llgo_github.com/goplus/llgo/cl/internal/foo.Foo", align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testrt/any/out.ll b/cl/_testrt/any/out.ll
index d2f59f3f..83e277c1 100644
--- a/cl/_testrt/any/out.ll
+++ b/cl/_testrt/any/out.ll
@@ -68,7 +68,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -101,7 +101,7 @@ _llgo_0:
   ret i32 0
 }
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_int8, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testrt/builtin/out.ll b/cl/_testrt/builtin/out.ll
index 8045b41c..ffd5bc49 100644
--- a/cl/_testrt/builtin/out.ll
+++ b/cl/_testrt/builtin/out.ll
@@ -86,7 +86,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   store i64 9223372036854775807, ptr @main.a, align 4
   store i64 -9223372036854775808, ptr @main.b, align 4
   store i64 -1, ptr @main.n, align 4
@@ -643,7 +643,7 @@ declare void @"github.com/goplus/llgo/internal/runtime.PrintString"(%"github.com
 
 declare %"github.com/goplus/llgo/internal/runtime.Slice" @"github.com/goplus/llgo/internal/runtime.SliceAppend"(%"github.com/goplus/llgo/internal/runtime.Slice", ptr, i64, i64)
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_int, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testrt/cast/out.ll b/cl/_testrt/cast/out.ll
index 1f6b0f30..5225b11c 100644
--- a/cl/_testrt/cast/out.ll
+++ b/cl/_testrt/cast/out.ll
@@ -401,7 +401,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -472,7 +472,7 @@ _llgo_0:
   ret i32 0
 }
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_string, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testrt/eface/out.ll b/cl/_testrt/eface/out.ll
index af51095f..c809135b 100644
--- a/cl/_testrt/eface/out.ll
+++ b/cl/_testrt/eface/out.ll
@@ -199,7 +199,7 @@ _llgo_0:
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
   call void @"github.com/goplus/llgo/internal/abi.init"()
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -436,7 +436,7 @@ declare void @"github.com/goplus/llgo/internal/abi.init"()
 
 declare void @"github.com/goplus/llgo/internal/runtime.init"()
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_bool, align 8
   %1 = icmp eq ptr %0, null
diff --git a/cl/_testrt/panic/out.ll b/cl/_testrt/panic/out.ll
index 4bac3016..ce254403 100644
--- a/cl/_testrt/panic/out.ll
+++ b/cl/_testrt/panic/out.ll
@@ -17,7 +17,7 @@ _llgo_0:
 
 _llgo_1:                                          ; preds = %_llgo_0
   store i1 true, ptr @"main.init$guard", align 1
-  call void @"main.init$abi"()
+  call void @"main.init$after"()
   br label %_llgo_2
 
 _llgo_2:                                          ; preds = %_llgo_1, %_llgo_0
@@ -51,7 +51,7 @@ _llgo_0:
 
 declare void @"github.com/goplus/llgo/internal/runtime.init"()
 
-define void @"main.init$abi"() {
+define void @"main.init$after"() {
 _llgo_0:
   %0 = load ptr, ptr @_llgo_string, align 8
   %1 = icmp eq ptr %0, null
diff --git a/internal/abi/llgo_autogen.lla b/internal/abi/llgo_autogen.lla
index 747da4e5..eb0833a5 100644
Binary files a/internal/abi/llgo_autogen.lla and b/internal/abi/llgo_autogen.lla differ
diff --git a/internal/runtime/llgo_autogen.lla b/internal/runtime/llgo_autogen.lla
index 6574534d..cfd6dabb 100644
Binary files a/internal/runtime/llgo_autogen.lla and b/internal/runtime/llgo_autogen.lla differ
diff --git a/internal/runtime/z_defer.go b/internal/runtime/z_defer.go
index 7f59abb4..007d4888 100644
--- a/internal/runtime/z_defer.go
+++ b/internal/runtime/z_defer.go
@@ -21,6 +21,7 @@ type Defer struct {
 	proc func(uintptr)
 	bits uintptr
 	link *Defer
+	rund int // index of RunDefers
 }
 
 // DeferProc calls deferred statements.
@@ -28,5 +29,6 @@ func DeferProc(d *Defer) {
 	for d != nil {
 		d.proc(d.bits)
 		d = d.link
+		_ = d.rund
 	}
 }
diff --git a/ssa/decl.go b/ssa/decl.go
index 4dc2badf..8abe0339 100644
--- a/ssa/decl.go
+++ b/ssa/decl.go
@@ -22,7 +22,6 @@ import (
 	"sort"
 	"strconv"
 	"strings"
-	"unsafe"
 
 	"github.com/goplus/llvm"
 )
@@ -181,16 +180,12 @@ type aFunction struct {
 
 	blks []BasicBlock
 
+	defer_ *aDefer
+
 	params   []Type
 	freeVars Expr
 	base     int // base = 1 if hasFreeVars; base = 0 otherwise
-
-	deferNextBit int // next defer bit
-	deferData    Expr
-	deferParam   Expr
-	deferb       unsafe.Pointer
-
-	hasVArg bool
+	hasVArg  bool
 }
 
 // Function represents a function or method.
diff --git a/ssa/package.go b/ssa/package.go
index 4cb779a3..98e471b6 100644
--- a/ssa/package.go
+++ b/ssa/package.go
@@ -179,7 +179,7 @@ type aProgram struct {
 	setSpecTy   *types.Signature
 	routineTy   *types.Signature
 	destructTy  *types.Signature
-	deferFnTy   *types.Signature
+	//deferFnTy *types.Signature
 
 	paramObjPtr_ *types.Var
 
diff --git a/ssa/stmt_builder.go b/ssa/stmt_builder.go
index 9f92efea..17449b86 100644
--- a/ssa/stmt_builder.go
+++ b/ssa/stmt_builder.go
@@ -23,7 +23,6 @@ import (
 	"go/types"
 	"log"
 	"strings"
-	"unsafe"
 
 	"github.com/goplus/llvm"
 )
@@ -65,7 +64,7 @@ type Builder = *aBuilder
 
 // EndBuild ends the build process of a function.
 func (b Builder) EndBuild() {
-	b.endDeferFunc()
+	b.Func.endDefer(b)
 }
 
 // Dispose disposes of the builder.
@@ -141,6 +140,18 @@ const (
 	deferKey = "__llgo_defer"
 )
 
+type aDefer struct {
+	nextBit  int        // next defer bit
+	key      Expr       // pthread TLS key
+	data     Expr       // pointer to runtime.Defer
+	bitsPtr  Expr       // pointer to defer bits
+	rundPtr  Expr       // pointer to RunDefers index
+	procBlk  BasicBlock // deferProc block
+	stmts    []func(bits Expr)
+	runsNext []BasicBlock // next blocks of RunDefers
+}
+
+/*
 // func(uintptr)
 func (p Program) tyDeferFunc() *types.Signature {
 	if p.deferFnTy == nil {
@@ -150,6 +161,7 @@ func (p Program) tyDeferFunc() *types.Signature {
 	}
 	return p.deferFnTy
 }
+*/
 
 func (p Package) deferInit(b Builder) {
 	keyVar := p.VarOf(deferKey)
@@ -175,23 +187,31 @@ func (b Builder) deferKey() Expr {
 	return b.Load(b.Pkg.newDeferKey().Expr)
 }
 
-// DeferFuncName returns the name of the defer procedure.
-func (p Function) DeferFuncName() string {
-	return p.Name() + "$_llgo_defer"
-}
-
-// DeferFunc returns the defer procedure of this function.
-func (p Function) DeferFunc() Function {
-	name := p.DeferFuncName()
-	return p.Pkg.NewFunc(name, p.Prog.tyDeferFunc(), InC)
-}
-
-func (b Builder) endDeferFunc() {
+func (b Builder) getDefer() *aDefer {
 	self := b.Func
-	if self.deferb != nil {
-		b = Builder(self.deferb)
-		b.Return()
+	if self.defer_ == nil {
+		// TODO(xsw): move to funtion start
+		// 0: proc func(uintptr)
+		// 1: bits uintptr
+		// 2: link *Defer
+		// 3: rund int
+		prog := b.Prog
+		key := b.deferKey()
+		deferfn := prog.Null(prog.VoidPtr())
+		zero := prog.Val(uintptr(0))
+		link := b.pthreadGetspecific(key)
+		ptr := b.aggregateAlloca(prog.Defer(), deferfn.impl, zero.impl, link.impl)
+		deferData := Expr{ptr, prog.DeferPtr()}
+		b.pthreadSetspecific(key, deferData)
+		self.defer_ = &aDefer{
+			key:     key,
+			data:    deferData,
+			bitsPtr: b.FieldAddr(deferData, 1),
+			rundPtr: b.FieldAddr(deferData, 3),
+			procBlk: self.MakeBlock(),
+		}
 	}
+	return self.defer_
 }
 
 // Defer emits a defer instruction.
@@ -200,43 +220,62 @@ func (b Builder) Defer(fn Expr, args ...Expr) {
 		logCall("Defer", fn, args)
 	}
 	prog := b.Prog
-	self := b.Func
-	next := self.deferNextBit
-	self.deferNextBit++
-	zero := prog.Val(uintptr(0))
-	key := b.deferKey()
-	if next == 0 {
-		deferfn := self.DeferFunc()
-		deferb := deferfn.MakeBody(1)
-		self.deferb = unsafe.Pointer(deferb)
-		self.deferParam = deferfn.Param(0)
-
-		// TODO(xsw): move to funtion start
-		// proc func(uintptr)
-		// bits uintptr
-		// link *Defer
-		link := b.pthreadGetspecific(key)
-		ptr := b.aggregateAlloca(prog.Defer(), deferfn.impl, zero.impl, link.impl)
-		self.deferData = Expr{ptr, prog.DeferPtr()}
-		b.pthreadSetspecific(key, self.deferData)
-	}
-	bitsPtr := b.FieldAddr(self.deferData, 1)
+	self := b.getDefer()
+	next := self.nextBit
+	self.nextBit++
+	bits := b.Load(self.bitsPtr)
 	nextbit := prog.Val(uintptr(1 << next))
-	b.Store(bitsPtr, b.BinOp(token.OR, b.Load(bitsPtr), nextbit))
+	b.Store(self.bitsPtr, b.BinOp(token.OR, bits, nextbit))
 
-	b = Builder(self.deferb)
-	has := b.BinOp(token.NEQ, b.BinOp(token.AND, self.deferParam, nextbit), zero)
-	b.IfThen(has, func() {
-		b.Call(fn, args...)
+	self.stmts = append(self.stmts, func(bits Expr) {
+		zero := prog.Val(uintptr(0))
+		has := b.BinOp(token.NEQ, b.BinOp(token.AND, bits, nextbit), zero)
+		b.IfThen(has, func() {
+			b.Call(fn, args...)
+		})
 	})
 }
 
 // RunDefers emits instructions to run deferred instructions.
 func (b Builder) RunDefers() {
-	self := b.Func
-	deferfn := self.DeferFunc()
-	bitsPtr := b.FieldAddr(self.deferData, 1)
-	b.Call(deferfn.Expr, b.Load(bitsPtr))
+	prog := b.Prog
+	self := b.getDefer()
+	b.Store(self.rundPtr, prog.Val(len(self.runsNext)))
+	b.Jump(self.procBlk)
+
+	blk := b.Func.MakeBlock()
+	self.runsNext = append(self.runsNext, blk)
+
+	b.SetBlockEx(blk, AtEnd, false)
+	b.blk.last = blk.last
+}
+
+func (p Function) endDefer(b Builder) {
+	self := p.defer_
+	if self == nil {
+		return
+	}
+	nexts := self.runsNext
+	n := len(nexts)
+	if n == 0 {
+		return
+	}
+	b.SetBlockEx(self.procBlk, AtEnd, true)
+	bits := b.Load(self.bitsPtr)
+	stmts := self.stmts
+	for i := len(stmts) - 1; i >= 0; i-- {
+		stmts[i](bits)
+	}
+
+	link := b.getField(b.Load(self.data), 2)
+	b.pthreadSetspecific(self.key, link)
+
+	prog := b.Prog
+	rund := b.Load(self.rundPtr)
+	sw := b.impl.CreateSwitch(rund.impl, nexts[0].first, n-1)
+	for i := 1; i < n; i++ {
+		sw.AddCase(prog.Val(i).impl, nexts[i].first)
+	}
 }
 
 // -----------------------------------------------------------------------------