From 1adf0d08b8a5fd8fc04491d77efab72eee343b7c Mon Sep 17 00:00:00 2001
From: "duanyi.aster" <duanyi.aster@bytedance.com>
Date: Thu, 1 Feb 2024 01:23:39 +0800
Subject: [PATCH] reserve x28

---
 internal/native/neon/f32toa_arm64.s           | 10 +--
 internal/native/neon/f32toa_subr_arm64.go     |  2 +-
 internal/native/neon/f64toa_arm64.s           | 10 +--
 internal/native/neon/f64toa_subr_arm64.go     |  2 +-
 ...tfloat_arm64_test.go => fastfloat_test.go} |  0
 ...{fastint_arm64_test.go => fastint_test.go} |  0
 internal/native/neon/get_by_path_arm64.s      |  8 +-
 .../native/neon/get_by_path_subr_arm64.go     |  2 +-
 internal/native/neon/html_escape_arm64.s      |  8 +-
 .../native/neon/html_escape_subr_arm64.go     |  2 +-
 internal/native/neon/i64toa_arm64.s           |  8 +-
 internal/native/neon/i64toa_subr_arm64.go     |  2 +-
 internal/native/neon/lspace_arm64.s           |  8 +-
 internal/native/neon/lspace_subr_arm64.go     |  2 +-
 internal/native/neon/native_arm64_test.go     | 36 ++++----
 internal/native/neon/quote_arm64.s            |  8 +-
 internal/native/neon/quote_subr_arm64.go      |  2 +-
 internal/native/neon/recover_arm64_test.go    | 10 +--
 internal/native/neon/skip_array_arm64.s       |  8 +-
 internal/native/neon/skip_array_subr_arm64.go |  2 +-
 internal/native/neon/skip_number_arm64.s      |  8 +-
 .../native/neon/skip_number_subr_arm64.go     |  2 +-
 internal/native/neon/skip_object_arm64.s      |  8 +-
 .../native/neon/skip_object_subr_arm64.go     |  2 +-
 internal/native/neon/skip_one_arm64.s         |  8 +-
 internal/native/neon/skip_one_fast_arm64.s    |  8 +-
 .../native/neon/skip_one_fast_subr_arm64.go   |  2 +-
 internal/native/neon/skip_one_subr_arm64.go   |  2 +-
 internal/native/neon/u64toa_arm64.s           |  8 +-
 internal/native/neon/u64toa_subr_arm64.go     |  2 +-
 internal/native/neon/unquote_arm64.s          |  8 +-
 internal/native/neon/unquote_subr_arm64.go    |  2 +-
 internal/native/neon/validate_one_arm64.s     |  8 +-
 .../native/neon/validate_one_subr_arm64.go    |  2 +-
 internal/native/neon/validate_utf8_arm64.s    |  8 +-
 .../native/neon/validate_utf8_fast_arm64.s    |  8 +-
 .../neon/validate_utf8_fast_subr_arm64.go     |  2 +-
 .../native/neon/validate_utf8_subr_arm64.go   |  2 +-
 internal/native/neon/value_arm64.s            |  8 +-
 internal/native/neon/value_subr_arm64.go      |  2 +-
 internal/native/neon/vnumber_arm64.s          |  8 +-
 internal/native/neon/vnumber_subr_arm64.go    |  2 +-
 internal/native/neon/vsigned_arm64.s          |  8 +-
 internal/native/neon/vsigned_subr_arm64.go    |  2 +-
 internal/native/neon/vstring_arm64.s          |  8 +-
 internal/native/neon/vstring_subr_arm64.go    |  2 +-
 internal/native/neon/vunsigned_arm64.s        |  8 +-
 internal/native/neon/vunsigned_subr_arm64.go  |  2 +-
 tools/asm2arm/arm.py                          | 85 +++++++++----------
 49 files changed, 177 insertions(+), 178 deletions(-)
 rename internal/native/neon/{fastfloat_arm64_test.go => fastfloat_test.go} (100%)
 rename internal/native/neon/{fastint_arm64_test.go => fastint_test.go} (100%)

diff --git a/internal/native/neon/f32toa_arm64.s b/internal/native/neon/f32toa_arm64.s
index fff239bb5..035b0f8ea 100644
--- a/internal/native/neon/f32toa_arm64.s
+++ b/internal/native/neon/f32toa_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__f32toa_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _f32toa:
 	WORD $0x1e260009  // fmov	w9, s0
@@ -902,7 +900,7 @@ _Digits:
 	WORD $0x37393639  // .ascii 4, '96979899'
 	WORD $0x39393839  // .ascii 4, '9899'
 	WORD $0x00000000  // .p2align 3, 0x00
-_LB_412ab4da: // _pow10_ceil_sig_f32.g
+_LB_9ffba3ae: // _pow10_ceil_sig_f32.g
 	WORD $0x4b43fcf5; WORD $0x81ceb32c  // .quad -9093133594791772939
 	WORD $0x5e14fc32; WORD $0xa2425ff7  // .quad -6754730975062328270
 	WORD $0x359a3b3f; WORD $0xcad2f7f5  // .quad -3831727700400522433
@@ -981,7 +979,7 @@ _LB_412ab4da: // _pow10_ceil_sig_f32.g
 	WORD $0x7bea5c70; WORD $0x8f7e32ce  // .quad -8106986416796705680
 	WORD $0x1ae4f38c; WORD $0xb35dbf82  // .quad -5522047002568494196
 
-TEXT ·__f32toa(SB), $0-24
+TEXT ·__f32toa(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -993,7 +991,9 @@ _entry:
 _f32toa:
 	MOVD out+0(FP), R0
 	FMOVD val+8(FP), F0
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__f32toa_entry__+16(SB)  // _f32toa
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/f32toa_subr_arm64.go b/internal/native/neon/f32toa_subr_arm64.go
index a30a7ce02..df71102ac 100644
--- a/internal/native/neon/f32toa_subr_arm64.go
+++ b/internal/native/neon/f32toa_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __f32toa_entry__() uintptr
 
 var (
-    _subr__f32toa uintptr = 16
+    _subr__f32toa uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/f64toa_arm64.s b/internal/native/neon/f64toa_arm64.s
index 2a554fa47..5112840c0 100644
--- a/internal/native/neon/f64toa_arm64.s
+++ b/internal/native/neon/f64toa_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__f64toa_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _f64toa:
 	WORD $0x9e660009  // fmov	x9, d0
@@ -1220,7 +1218,7 @@ _Digits:
 	WORD $0x37393639  // .ascii 4, '96979899'
 	WORD $0x39393839  // .ascii 4, '9899'
 	  // .p2align 3, 0x00
-_LB_9b3e13db: // _pow10_ceil_sig.g
+_LB_03fda5ca: // _pow10_ceil_sig.g
 	WORD $0xbebcdc4f; WORD $0xff77b1fc  // .quad -38366372719436721
 	WORD $0x13bb0f7b; WORD $0x25e8e89c  // .quad 2731688931043774331
 	WORD $0xf73609b1; WORD $0x9faacf3d  // .quad -6941508010590729807
@@ -2456,7 +2454,7 @@ _LB_9b3e13db: // _pow10_ceil_sig.g
 	WORD $0xb4e31ba9; WORD $0x9e19db92  // .quad -7054365918152680535
 	WORD $0x6a8346d2; WORD $0x6c07a2c2  // .quad 7784369436827535058
 
-TEXT ·__f64toa(SB), $0-24
+TEXT ·__f64toa(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -2468,7 +2466,9 @@ _entry:
 _f64toa:
 	MOVD out+0(FP), R0
 	FMOVD val+8(FP), F0
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__f64toa_entry__+16(SB)  // _f64toa
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/f64toa_subr_arm64.go b/internal/native/neon/f64toa_subr_arm64.go
index 841d0424e..f0c980647 100644
--- a/internal/native/neon/f64toa_subr_arm64.go
+++ b/internal/native/neon/f64toa_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __f64toa_entry__() uintptr
 
 var (
-    _subr__f64toa uintptr = 16
+    _subr__f64toa uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/fastfloat_arm64_test.go b/internal/native/neon/fastfloat_test.go
similarity index 100%
rename from internal/native/neon/fastfloat_arm64_test.go
rename to internal/native/neon/fastfloat_test.go
diff --git a/internal/native/neon/fastint_arm64_test.go b/internal/native/neon/fastint_test.go
similarity index 100%
rename from internal/native/neon/fastint_arm64_test.go
rename to internal/native/neon/fastint_test.go
diff --git a/internal/native/neon/get_by_path_arm64.s b/internal/native/neon/get_by_path_arm64.s
index e3609fc47..9f5c49c30 100644
--- a/internal/native/neon/get_by_path_arm64.s
+++ b/internal/native/neon/get_by_path_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__get_by_path_entry__(SB), NOSPLIT, $192
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -4553,7 +4551,7 @@ __UnquoteTab:
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00'
 
-TEXT ·__get_by_path(SB), $0-40
+TEXT ·__get_by_path(SB), NOSPLIT, $0-40
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -4567,7 +4565,9 @@ _get_by_path:
 	MOVD p+8(FP), R1
 	MOVD path+16(FP), R2
 	MOVD m+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__get_by_path_entry__+64(SB)  // _get_by_path
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+32(FP)
 	RET
 
diff --git a/internal/native/neon/get_by_path_subr_arm64.go b/internal/native/neon/get_by_path_subr_arm64.go
index bf9945ab2..c2aca952e 100644
--- a/internal/native/neon/get_by_path_subr_arm64.go
+++ b/internal/native/neon/get_by_path_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __get_by_path_entry__() uintptr
 
 var (
-    _subr__get_by_path uintptr = 64
+    _subr__get_by_path uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/html_escape_arm64.s b/internal/native/neon/html_escape_arm64.s
index e609a7333..2227891fc 100644
--- a/internal/native/neon/html_escape_arm64.s
+++ b/internal/native/neon/html_escape_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__html_escape_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -1376,7 +1374,7 @@ __HtmlQuoteTab:
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00'
 
-TEXT ·__html_escape(SB), $0-40
+TEXT ·__html_escape(SB), NOSPLIT, $0-40
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -1390,7 +1388,9 @@ _html_escape:
 	MOVD nb+8(FP), R1
 	MOVD dp+16(FP), R2
 	MOVD dn+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__html_escape_entry__+48(SB)  // _html_escape
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+32(FP)
 	RET
 
diff --git a/internal/native/neon/html_escape_subr_arm64.go b/internal/native/neon/html_escape_subr_arm64.go
index d98fc4ff8..9882fb6ad 100644
--- a/internal/native/neon/html_escape_subr_arm64.go
+++ b/internal/native/neon/html_escape_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __html_escape_entry__() uintptr
 
 var (
-    _subr__html_escape uintptr = 48
+    _subr__html_escape uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/i64toa_arm64.s b/internal/native/neon/i64toa_arm64.s
index e72f676a9..9260bca9d 100644
--- a/internal/native/neon/i64toa_arm64.s
+++ b/internal/native/neon/i64toa_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__i64toa_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 3, 0x00
 lCPI0_0:
 	WORD $0x147b20c5
@@ -952,7 +950,7 @@ _VecShiftShuffles:
 	WORD $0xffffffff  // .ascii 4, '\xff\xff\xff\xff\xff\xff\xff\xff'
 	WORD $0xffffffff  // .ascii 4, '\xff\xff\xff\xff'
 
-TEXT ·__i64toa(SB), $0-24
+TEXT ·__i64toa(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -964,7 +962,9 @@ _entry:
 _i64toa:
 	MOVD out+0(FP), R0
 	MOVD val+8(FP), R1
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__i64toa_entry__+64(SB)  // _i64toa
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/i64toa_subr_arm64.go b/internal/native/neon/i64toa_subr_arm64.go
index e6970363f..442305e36 100644
--- a/internal/native/neon/i64toa_subr_arm64.go
+++ b/internal/native/neon/i64toa_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __i64toa_entry__() uintptr
 
 var (
-    _subr__i64toa uintptr = 64
+    _subr__i64toa uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/lspace_arm64.s b/internal/native/neon/lspace_arm64.s
index d48ba901b..a8d688e98 100644
--- a/internal/native/neon/lspace_arm64.s
+++ b/internal/native/neon/lspace_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__lspace_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _lspace:
 	WORD $0xa9be7bfd  // stp	fp, lr, [sp, #-32]!
@@ -41,7 +39,7 @@ LBB0_6:
 	WORD $0xa8c27bfd  // ldp	fp, lr, [sp], #32
 	WORD $0xd65f03c0  // ret
 
-TEXT ·__lspace(SB), $0-32
+TEXT ·__lspace(SB), NOSPLIT, $0-32
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -54,7 +52,9 @@ _lspace:
 	MOVD sp+0(FP), R0
 	MOVD nb+8(FP), R1
 	MOVD off+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__lspace_entry__+16(SB)  // _lspace
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+24(FP)
 	RET
 
diff --git a/internal/native/neon/lspace_subr_arm64.go b/internal/native/neon/lspace_subr_arm64.go
index 8ce268321..89fbf2d7c 100644
--- a/internal/native/neon/lspace_subr_arm64.go
+++ b/internal/native/neon/lspace_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __lspace_entry__() uintptr
 
 var (
-    _subr__lspace uintptr = 16
+    _subr__lspace uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/native_arm64_test.go b/internal/native/neon/native_arm64_test.go
index df1246214..bd377f93f 100644
--- a/internal/native/neon/native_arm64_test.go
+++ b/internal/native/neon/native_arm64_test.go
@@ -35,10 +35,16 @@ import (
     `github.com/stretchr/testify/require`
 )
 
+//go:noinline
+func BREAK() bool {
+    return true
+}
+
 func TestNative_Value(t *testing.T) {
     var v types.JsonState
     s := `   -12345`
     p := (*rt.GoString)(unsafe.Pointer(&s))
+    BREAK()
     x := value(p.Ptr, p.Len, 0, &v, 0)
     assert.Equal(t, 9, x)
     assert.Equal(t, types.V_INTEGER, v.Vt)
@@ -236,20 +242,6 @@ func TestNative_HTMLEscapeNoMem(t *testing.T) {
     assert.Equal(t, `hello`, string(d))
 }
 
-func TestNative_Vstring(t *testing.T) {
-    var v types.JsonState
-    i := 0
-    s := `test"test\n2"`
-    vstring(&s, &i, &v, 0)
-    assert.Equal(t, 5, i)
-    assert.Equal(t, -1, v.Ep)
-    assert.Equal(t, int64(0), v.Iv)
-    vstring(&s, &i, &v, 0)
-    assert.Equal(t, 13, i)
-    assert.Equal(t, 9, v.Ep)
-    assert.Equal(t, int64(5), v.Iv)
-}
-
 func TestNative_Vstring_ValidUnescapedChars(t *testing.T) {
     var v types.JsonState
     valid := uint64(types.F_VALIDATE_STRING)
@@ -632,4 +624,18 @@ func BenchmarkNative_GetByPath(b *testing.B) {
         _ = get_by_path(&s, &p, &path, sm)
         types.FreeStateMachine(sm)
     }
-}
\ No newline at end of file
+}
+
+func TestNative_Vstring(t *testing.T) {
+    var v types.JsonState
+    i := 0
+    s := `test"test\n2"`
+    vstring(&s, &i, &v, 0)
+    assert.Equal(t, 5, i)
+    assert.Equal(t, -1, v.Ep)
+    assert.Equal(t, int64(0), v.Iv)
+    vstring(&s, &i, &v, 0)
+    assert.Equal(t, 13, i)
+    assert.Equal(t, 9, v.Ep)
+    assert.Equal(t, int64(5), v.Iv)
+}
diff --git a/internal/native/neon/quote_arm64.s b/internal/native/neon/quote_arm64.s
index a96bd24f1..8f1f5d545 100644
--- a/internal/native/neon/quote_arm64.s
+++ b/internal/native/neon/quote_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__quote_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -2531,7 +2529,7 @@ __EscTab:
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00'
 
-TEXT ·__quote(SB), $0-48
+TEXT ·__quote(SB), NOSPLIT, $0-48
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -2546,7 +2544,9 @@ _quote:
 	MOVD dp+16(FP), R2
 	MOVD dn+24(FP), R3
 	MOVD flags+32(FP), R4
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__quote_entry__+48(SB)  // _quote
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+40(FP)
 	RET
 
diff --git a/internal/native/neon/quote_subr_arm64.go b/internal/native/neon/quote_subr_arm64.go
index a13d7fcba..6f7a8162d 100644
--- a/internal/native/neon/quote_subr_arm64.go
+++ b/internal/native/neon/quote_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __quote_entry__() uintptr
 
 var (
-    _subr__quote uintptr = 48
+    _subr__quote uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/recover_arm64_test.go b/internal/native/neon/recover_arm64_test.go
index 5a8f9e1ec..970665597 100644
--- a/internal/native/neon/recover_arm64_test.go
+++ b/internal/native/neon/recover_arm64_test.go
@@ -20,10 +20,10 @@ package neon
 
 import (
 	`os`
+	`time`
 	`runtime`
 	`runtime/debug`
 	`testing`
-	`time`
 	`unsafe`
 
 	`github.com/bytedance/sonic/internal/native/types`
@@ -49,7 +49,7 @@ func TestMain(m *testing.M) {
 	m.Run()
 }
 
-func TestRecover_f64toa(t *testing.T) {
+func TestRecover_f32toa(t *testing.T) {
 	defer func() {
 		if r := recover(); r!= nil {
 			t.Log("recover: ", r)
@@ -57,10 +57,10 @@ func TestRecover_f64toa(t *testing.T) {
 			t.Fatal("no panic")
 		}
 	}()
-	_ = f64toa(nil, 123)
+	_ = f32toa(nil, 123)
 }
 
-func TestRecover_f32toa(t *testing.T) {
+func TestRecover_f64toa(t *testing.T) {
 	defer func() {
 		if r := recover(); r!= nil {
 			t.Log("recover: ", r)
@@ -68,7 +68,7 @@ func TestRecover_f32toa(t *testing.T) {
 			t.Fatal("no panic")
 		}
 	}()
-	_ = f32toa(nil, 123)
+	_ = f64toa(nil, 123)
 }
 
 func TestRecover_i64toa(t *testing.T) {
diff --git a/internal/native/neon/skip_array_arm64.s b/internal/native/neon/skip_array_arm64.s
index 69ae8772e..9986d49a5 100644
--- a/internal/native/neon/skip_array_arm64.s
+++ b/internal/native/neon/skip_array_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__skip_array_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x00000001; WORD $0x00000000  // .quad 1
@@ -2356,7 +2354,7 @@ LBB0_430:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__skip_array(SB), $0-40
+TEXT ·__skip_array(SB), NOSPLIT, $0-40
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -2370,7 +2368,9 @@ _skip_array:
 	MOVD p+8(FP), R1
 	MOVD m+16(FP), R2
 	MOVD flags+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__skip_array_entry__+64(SB)  // _skip_array
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+32(FP)
 	RET
 
diff --git a/internal/native/neon/skip_array_subr_arm64.go b/internal/native/neon/skip_array_subr_arm64.go
index 4622043d6..96ce7fdfc 100644
--- a/internal/native/neon/skip_array_subr_arm64.go
+++ b/internal/native/neon/skip_array_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __skip_array_entry__() uintptr
 
 var (
-    _subr__skip_array uintptr = 64
+    _subr__skip_array uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/skip_number_arm64.s b/internal/native/neon/skip_number_arm64.s
index 816a95f8e..232bb8072 100644
--- a/internal/native/neon/skip_number_arm64.s
+++ b/internal/native/neon/skip_number_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__skip_number_entry__(SB), NOSPLIT, $32
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -349,7 +347,7 @@ LBB0_60:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__skip_number(SB), $0-24
+TEXT ·__skip_number(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -361,7 +359,9 @@ _entry:
 _skip_number:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__skip_number_entry__+48(SB)  // _skip_number
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/skip_number_subr_arm64.go b/internal/native/neon/skip_number_subr_arm64.go
index 074ea8356..e0ff6f5b1 100644
--- a/internal/native/neon/skip_number_subr_arm64.go
+++ b/internal/native/neon/skip_number_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __skip_number_entry__() uintptr
 
 var (
-    _subr__skip_number uintptr = 48
+    _subr__skip_number uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/skip_object_arm64.s b/internal/native/neon/skip_object_arm64.s
index 8264d0660..7ff58bdd8 100644
--- a/internal/native/neon/skip_object_arm64.s
+++ b/internal/native/neon/skip_object_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__skip_object_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x00000001; WORD $0x00000000  // .quad 1
@@ -2356,7 +2354,7 @@ LBB0_430:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__skip_object(SB), $0-40
+TEXT ·__skip_object(SB), NOSPLIT, $0-40
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -2370,7 +2368,9 @@ _skip_object:
 	MOVD p+8(FP), R1
 	MOVD m+16(FP), R2
 	MOVD flags+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__skip_object_entry__+64(SB)  // _skip_object
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+32(FP)
 	RET
 
diff --git a/internal/native/neon/skip_object_subr_arm64.go b/internal/native/neon/skip_object_subr_arm64.go
index 0f20e72bc..e526aeb93 100644
--- a/internal/native/neon/skip_object_subr_arm64.go
+++ b/internal/native/neon/skip_object_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __skip_object_entry__() uintptr
 
 var (
-    _subr__skip_object uintptr = 64
+    _subr__skip_object uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/skip_one_arm64.s b/internal/native/neon/skip_one_arm64.s
index 1de8f4ebd..e45105c10 100644
--- a/internal/native/neon/skip_one_arm64.s
+++ b/internal/native/neon/skip_one_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__skip_one_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x00000001; WORD $0x00000000  // .quad 1
@@ -2356,7 +2354,7 @@ LBB0_430:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__skip_one(SB), $0-40
+TEXT ·__skip_one(SB), NOSPLIT, $0-40
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -2370,7 +2368,9 @@ _skip_one:
 	MOVD p+8(FP), R1
 	MOVD m+16(FP), R2
 	MOVD flags+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__skip_one_entry__+64(SB)  // _skip_one
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+32(FP)
 	RET
 
diff --git a/internal/native/neon/skip_one_fast_arm64.s b/internal/native/neon/skip_one_fast_arm64.s
index 652b71b51..98dc26c4f 100644
--- a/internal/native/neon/skip_one_fast_arm64.s
+++ b/internal/native/neon/skip_one_fast_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__skip_one_fast_entry__(SB), NOSPLIT, $176
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -974,7 +972,7 @@ LBB0_124:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__skip_one_fast(SB), $0-24
+TEXT ·__skip_one_fast(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -986,7 +984,9 @@ _entry:
 _skip_one_fast:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__skip_one_fast_entry__+48(SB)  // _skip_one_fast
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/skip_one_fast_subr_arm64.go b/internal/native/neon/skip_one_fast_subr_arm64.go
index 78403de09..b4c2f5d07 100644
--- a/internal/native/neon/skip_one_fast_subr_arm64.go
+++ b/internal/native/neon/skip_one_fast_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __skip_one_fast_entry__() uintptr
 
 var (
-    _subr__skip_one_fast uintptr = 48
+    _subr__skip_one_fast uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/skip_one_subr_arm64.go b/internal/native/neon/skip_one_subr_arm64.go
index 9a91cbf3c..87c40c784 100644
--- a/internal/native/neon/skip_one_subr_arm64.go
+++ b/internal/native/neon/skip_one_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __skip_one_entry__() uintptr
 
 var (
-    _subr__skip_one uintptr = 64
+    _subr__skip_one uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/u64toa_arm64.s b/internal/native/neon/u64toa_arm64.s
index 6df9837ce..8b025665e 100644
--- a/internal/native/neon/u64toa_arm64.s
+++ b/internal/native/neon/u64toa_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__u64toa_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 3, 0x00
 lCPI0_0:
 	WORD $0x147b20c5
@@ -566,7 +564,7 @@ _VecShiftShuffles:
 	WORD $0xffffffff  // .ascii 4, '\xff\xff\xff\xff\xff\xff\xff\xff'
 	WORD $0xffffffff  // .ascii 4, '\xff\xff\xff\xff'
 
-TEXT ·__u64toa(SB), $0-24
+TEXT ·__u64toa(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -578,7 +576,9 @@ _entry:
 _u64toa:
 	MOVD out+0(FP), R0
 	MOVD val+8(FP), R1
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__u64toa_entry__+64(SB)  // _u64toa
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+16(FP)
 	RET
 
diff --git a/internal/native/neon/u64toa_subr_arm64.go b/internal/native/neon/u64toa_subr_arm64.go
index 84c5fff98..3cd0f3d7d 100644
--- a/internal/native/neon/u64toa_subr_arm64.go
+++ b/internal/native/neon/u64toa_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __u64toa_entry__() uintptr
 
 var (
-    _subr__u64toa uintptr = 64
+    _subr__u64toa uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/unquote_arm64.s b/internal/native/neon/unquote_arm64.s
index bdc7f503f..a799bb673 100644
--- a/internal/native/neon/unquote_arm64.s
+++ b/internal/native/neon/unquote_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__unquote_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -586,7 +584,7 @@ __UnquoteTab:
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .space 4, '\x00\x00\x00\x00'
 
-TEXT ·__unquote(SB), $0-48
+TEXT ·__unquote(SB), NOSPLIT, $0-48
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -601,7 +599,9 @@ _unquote:
 	MOVD dp+16(FP), R2
 	MOVD ep+24(FP), R3
 	MOVD flags+32(FP), R4
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__unquote_entry__+48(SB)  // _unquote
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+40(FP)
 	RET
 
diff --git a/internal/native/neon/unquote_subr_arm64.go b/internal/native/neon/unquote_subr_arm64.go
index 7adb5ca46..7f5c3e37b 100644
--- a/internal/native/neon/unquote_subr_arm64.go
+++ b/internal/native/neon/unquote_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __unquote_entry__() uintptr
 
 var (
-    _subr__unquote uintptr = 48
+    _subr__unquote uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/validate_one_arm64.s b/internal/native/neon/validate_one_arm64.s
index 4ec6d254f..58c57ea4e 100644
--- a/internal/native/neon/validate_one_arm64.s
+++ b/internal/native/neon/validate_one_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__validate_one_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x00000001; WORD $0x00000000  // .quad 1
@@ -1833,7 +1831,7 @@ LBB0_337:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__validate_one(SB), $0-32
+TEXT ·__validate_one(SB), NOSPLIT, $0-32
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -1846,7 +1844,9 @@ _validate_one:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
 	MOVD m+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__validate_one_entry__+64(SB)  // _validate_one
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+24(FP)
 	RET
 
diff --git a/internal/native/neon/validate_one_subr_arm64.go b/internal/native/neon/validate_one_subr_arm64.go
index 0b195c1d7..2fa1a5215 100644
--- a/internal/native/neon/validate_one_subr_arm64.go
+++ b/internal/native/neon/validate_one_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __validate_one_entry__() uintptr
 
 var (
-    _subr__validate_one uintptr = 64
+    _subr__validate_one uintptr = 48
 )
 
 const (
diff --git a/internal/native/neon/validate_utf8_arm64.s b/internal/native/neon/validate_utf8_arm64.s
index feabfc9eb..ca87d00b3 100644
--- a/internal/native/neon/validate_utf8_arm64.s
+++ b/internal/native/neon/validate_utf8_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__validate_utf8_entry__(SB), NOSPLIT, $48
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _validate_utf8:
 	WORD $0xd10103ff  // sub	sp, sp, #64
@@ -197,7 +195,7 @@ LBB0_33:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__validate_utf8(SB), $0-32
+TEXT ·__validate_utf8(SB), NOSPLIT, $0-32
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -210,7 +208,9 @@ _validate_utf8:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
 	MOVD m+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__validate_utf8_entry__+16(SB)  // _validate_utf8
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+24(FP)
 	RET
 
diff --git a/internal/native/neon/validate_utf8_fast_arm64.s b/internal/native/neon/validate_utf8_fast_arm64.s
index aafde624f..4296b6a4f 100644
--- a/internal/native/neon/validate_utf8_fast_arm64.s
+++ b/internal/native/neon/validate_utf8_fast_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__validate_utf8_fast_entry__(SB), NOSPLIT, $32
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _validate_utf8_fast:
 	WORD $0xd100c3ff  // sub	sp, sp, #48
@@ -159,7 +157,7 @@ LBB0_30:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__validate_utf8_fast(SB), $0-16
+TEXT ·__validate_utf8_fast(SB), NOSPLIT, $0-16
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -170,7 +168,9 @@ _entry:
 
 _validate_utf8_fast:
 	MOVD s+0(FP), R0
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__validate_utf8_fast_entry__+16(SB)  // _validate_utf8_fast
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+8(FP)
 	RET
 
diff --git a/internal/native/neon/validate_utf8_fast_subr_arm64.go b/internal/native/neon/validate_utf8_fast_subr_arm64.go
index e72eceb46..a127e874a 100644
--- a/internal/native/neon/validate_utf8_fast_subr_arm64.go
+++ b/internal/native/neon/validate_utf8_fast_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __validate_utf8_fast_entry__() uintptr
 
 var (
-    _subr__validate_utf8_fast uintptr = 16
+    _subr__validate_utf8_fast uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/validate_utf8_subr_arm64.go b/internal/native/neon/validate_utf8_subr_arm64.go
index 8878908a5..ac8a68285 100644
--- a/internal/native/neon/validate_utf8_subr_arm64.go
+++ b/internal/native/neon/validate_utf8_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __validate_utf8_entry__() uintptr
 
 var (
-    _subr__validate_utf8 uintptr = 16
+    _subr__validate_utf8 uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/value_arm64.s b/internal/native/neon/value_arm64.s
index bc6fe7388..a3c79debd 100644
--- a/internal/native/neon/value_arm64.s
+++ b/internal/native/neon/value_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__value_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -5939,7 +5937,7 @@ _LSHIFT_TAB:
 	WORD $0x00000000  // .asciz 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .asciz 4, '\x00\x00\x00\x00'
 
-TEXT ·__value(SB), $0-48
+TEXT ·__value(SB), NOSPLIT, $0-48
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -5954,7 +5952,9 @@ _value:
 	MOVD p+16(FP), R2
 	MOVD v+24(FP), R3
 	MOVD flags+32(FP), R4
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__value_entry__+48(SB)  // _value
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	MOVD R0, ret+40(FP)
 	RET
 
diff --git a/internal/native/neon/value_subr_arm64.go b/internal/native/neon/value_subr_arm64.go
index c8e82e2e0..238bcf742 100644
--- a/internal/native/neon/value_subr_arm64.go
+++ b/internal/native/neon/value_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __value_entry__() uintptr
 
 var (
-    _subr__value uintptr = 48
+    _subr__value uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/vnumber_arm64.s b/internal/native/neon/vnumber_arm64.s
index b79e177fa..ecaefae50 100644
--- a/internal/native/neon/vnumber_arm64.s
+++ b/internal/native/neon/vnumber_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__vnumber_entry__(SB), NOSPLIT, $96
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _vnumber:
 	WORD $0xf9400031  // ldr	x17, [x1]
@@ -4669,7 +4667,7 @@ _LSHIFT_TAB:
 	WORD $0x00000000  // .asciz 4, '\x00\x00\x00\x00\x00\x00\x00\x00'
 	WORD $0x00000000  // .asciz 4, '\x00\x00\x00\x00'
 
-TEXT ·__vnumber(SB), $0-24
+TEXT ·__vnumber(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -4682,7 +4680,9 @@ _vnumber:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
 	MOVD v+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__vnumber_entry__+16(SB)  // _vnumber
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	RET
 
 _stack_grow:
diff --git a/internal/native/neon/vnumber_subr_arm64.go b/internal/native/neon/vnumber_subr_arm64.go
index 242a7a544..2f8086dc1 100644
--- a/internal/native/neon/vnumber_subr_arm64.go
+++ b/internal/native/neon/vnumber_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __vnumber_entry__() uintptr
 
 var (
-    _subr__vnumber uintptr = 16
+    _subr__vnumber uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/vsigned_arm64.s b/internal/native/neon/vsigned_arm64.s
index f17d147ca..8963eb869 100644
--- a/internal/native/neon/vsigned_arm64.s
+++ b/internal/native/neon/vsigned_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__vsigned_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _vsigned:
 	WORD $0xf9400028  // ldr	x8, [x1]
@@ -130,7 +128,7 @@ LBB0_23:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__vsigned(SB), $0-24
+TEXT ·__vsigned(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -143,7 +141,9 @@ _vsigned:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
 	MOVD v+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__vsigned_entry__+16(SB)  // _vsigned
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	RET
 
 _stack_grow:
diff --git a/internal/native/neon/vsigned_subr_arm64.go b/internal/native/neon/vsigned_subr_arm64.go
index 611be3f3a..b436af3da 100644
--- a/internal/native/neon/vsigned_subr_arm64.go
+++ b/internal/native/neon/vsigned_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __vsigned_entry__() uintptr
 
 var (
-    _subr__vsigned uintptr = 16
+    _subr__vsigned uintptr = 0
 )
 
 const (
diff --git a/internal/native/neon/vstring_arm64.s b/internal/native/neon/vstring_arm64.s
index 77bd91682..e47fa4b32 100644
--- a/internal/native/neon/vstring_arm64.s
+++ b/internal/native/neon/vstring_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__vstring_entry__(SB), NOSPLIT, $32
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 4, 0x00
 lCPI0_0:
 	WORD $0x08040201
@@ -618,7 +616,7 @@ LBB0_77:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__vstring(SB), $0-32
+TEXT ·__vstring(SB), NOSPLIT, $0-32
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -632,7 +630,9 @@ _vstring:
 	MOVD p+8(FP), R1
 	MOVD v+16(FP), R2
 	MOVD flags+24(FP), R3
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__vstring_entry__+48(SB)  // _vstring
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	RET
 
 _stack_grow:
diff --git a/internal/native/neon/vstring_subr_arm64.go b/internal/native/neon/vstring_subr_arm64.go
index d0c984d87..aa3e88709 100644
--- a/internal/native/neon/vstring_subr_arm64.go
+++ b/internal/native/neon/vstring_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __vstring_entry__() uintptr
 
 var (
-    _subr__vstring uintptr = 48
+    _subr__vstring uintptr = 32
 )
 
 const (
diff --git a/internal/native/neon/vunsigned_arm64.s b/internal/native/neon/vunsigned_arm64.s
index e0cb44193..ec3149e07 100644
--- a/internal/native/neon/vunsigned_arm64.s
+++ b/internal/native/neon/vunsigned_arm64.s
@@ -7,9 +7,7 @@
 
 TEXT ·__vunsigned_entry__(SB), NOSPLIT, $16
 	NO_LOCAL_POINTERS
-	WORD $0x10000000  // adr x0, . $0(%rip)
-	WORD $0xd65f03c0  // ret
-	WORD $0x00000000; WORD $0x00000000  // .p2align 4, 0x00
+	PCALIGN $16
 	  // .p2align 2, 0x00
 _vunsigned:
 	WORD $0xf9400028  // ldr	x8, [x1]
@@ -118,7 +116,7 @@ LBB0_21:
 _MASK_USE_NUMBER:
 	WORD $0x00000002  // .long 2
 
-TEXT ·__vunsigned(SB), $0-24
+TEXT ·__vunsigned(SB), NOSPLIT, $0-24
 	NO_LOCAL_POINTERS
 
 _entry:
@@ -131,7 +129,9 @@ _vunsigned:
 	MOVD s+0(FP), R0
 	MOVD p+8(FP), R1
 	MOVD v+16(FP), R2
+	WORD $0xf90007fc // str x28, [sp, #8]
 	CALL ·__vunsigned_entry__+16(SB)  // _vunsigned
+	WORD $0xf94007fc // ldr x28, [sp, #8]
 	RET
 
 _stack_grow:
diff --git a/internal/native/neon/vunsigned_subr_arm64.go b/internal/native/neon/vunsigned_subr_arm64.go
index 6c16b762f..f32d6af5a 100644
--- a/internal/native/neon/vunsigned_subr_arm64.go
+++ b/internal/native/neon/vunsigned_subr_arm64.go
@@ -9,7 +9,7 @@ package neon
 func __vunsigned_entry__() uintptr
 
 var (
-    _subr__vunsigned uintptr = 16
+    _subr__vunsigned uintptr = 0
 )
 
 const (
diff --git a/tools/asm2arm/arm.py b/tools/asm2arm/arm.py
index 9118a207a..50a5a8faf 100644
--- a/tools/asm2arm/arm.py
+++ b/tools/asm2arm/arm.py
@@ -2018,6 +2018,7 @@ def _declare_body(self, name: str):
         size = 0 if size < 0 else size
         self.out.append('TEXT ·_%s_entry__(SB), NOSPLIT, $%d' % (name, size))
         self.out.append('\tNO_LOCAL_POINTERS')
+        self.out.append('\tPCALIGN $16')
         self._LE_4bytes_IntIntr_2_RawIntr()
         self._reloc()
 
@@ -2047,6 +2048,12 @@ def _declare_function(self, name: str, proto: Prototype):
         addr = self.code.get(subr)
         self.subr[subr] = addr
         size = self.code.stacksize(subr)
+        
+        # Notice: golang will insert 3 instructions when stacksize > 0.
+        # And since we use PCALIGN 16, thus we add 16 bytes PC
+        if size > 0:
+            addr += 16
+
         m_size = size + 64
         # rsp_sub_size = size + 16
 
@@ -2057,7 +2064,7 @@ def _declare_function(self, name: str, proto: Prototype):
         self.out.append('')
         # frame size is 16 to store x29 and x30
         # self.out.append('TEXT ·%s(SB), NOSPLIT | NOFRAME, $0-%d' % (name, proto.argspace))
-        self.out.append('TEXT ·%s(SB), $%d-%d' % (name, 0, proto.argspace))
+        self.out.append('TEXT ·%s(SB), NOSPLIT, $%d-%d' % (name, 0, proto.argspace))
         self.out.append('\tNO_LOCAL_POINTERS')
 
         # add stack check if needed
@@ -2099,47 +2106,35 @@ def _declare_function(self, name: str, proto: Prototype):
 
         # Go ASM completely ignores the offset of the JMP instruction,
         # so we need to use indirect jumps instead for tail-call elimination
-        elif proto.retv is None:
-            # raise RuntimeError("UNIMPLEMENT FUNC: %s" % name)
-            print("%s return void." % name)
-            # self.out.append('\tLEAQ ·%s+%d(SB), AX  // %s' % (STUB_NAME, addr, subr))
-            # self.out.append('\tJMP AX')
-
-            # save LR(x30) and Frame Pointer(x29)
-
-            # self.out.append('\tADD $%d, RSP, RSP' % size)
-            # self.out.append('\tSTP.W (R29, R30), -16(RSP)')
-            # self.out.append('\tMOVD RSP, R29')
-            # self.out.append('\tSUB $16, RSP')
-            # self.out.append('\tADD $%d, RSP' % (size + 32))
-            self.out.append('\tCALL ·_%s_entry__+%d(SB)  // %s' % (subr, addr, subr))
-            # self.out.append('\tSUB $%d, RSP' % (size + 32))
-            # self.out.append('\tADD $16, RSP')
-            # self.out.append('\tLDP -8(RSP), (R29, R30)')
-            # self.out.append('\tADD $16, RSP')
-            # self.out.append('\tLDP.P 16(RSP), (R29, R30)')
-            # self.out.append('\tSUB $%d, RSP, RSP' % size)
-
-            # Restore LR and Frame Pointer
+        # raise RuntimeError("UNIMPLEMENT FUNC: %s" % name)
+        # self.out.append('\tLEAQ ·%s+%d(SB), AX  // %s' % (STUB_NAME, addr, subr))
+        # self.out.append('\tJMP AX')
+
+        # save LR(x30) and Frame Pointer(x29)
+
+        # self.out.append('\tADD $%d, RSP, RSP' % size)
+        # self.out.append('\tSTP.W (R29, R30), -16(RSP)')
+        # self.out.append('\tMOVD RSP, R29')
+        # self.out.append('\tSUB $16, RSP')
+        # self.out.append('\tADD $%d, RSP' % (size + 32))
+        self.out.append('\tWORD $0xf90007fc // str x28, [sp, #8]')
+        self.out.append('\tCALL ·_%s_entry__+%d(SB)  // %s' % (subr, addr, subr))
+        self.out.append('\tWORD $0xf94007fc // ldr x28, [sp, #8]')
+        # self.out.append('\tSUB $%d, RSP' % (size + 32))
+        # self.out.append('\tADD $16, RSP')
+        # self.out.append('\tLDP -8(RSP), (R29, R30)')
+        # self.out.append('\tADD $16, RSP')
+        # self.out.append('\tLDP.P 16(RSP), (R29, R30)')
+        # self.out.append('\tSUB $%d, RSP, RSP' % size)
+
+        # Restore LR and Frame Pointer
 
         # normal functions, call the real function, and return the result
-        else:
-            # save LR(x30) and Frame Pointer(x29)
-
-            # We need store return address manual in ARM
-            # self.out.append('\tSUB $16, RSP')
-            # self.out.append('\tADD $%d, RSP' % (size + 32))
-            self.out.append('\tCALL ·_%s_entry__+%d(SB)  // %s' % (subr, addr, subr))
-            # self.out.append('\tSUB $%d, RSP' % (size + 32))
-            # self.out.append('\tADD $16, RSP')
-            # self.out.append('\tSUB $%d, RSP, RSP' % size)
-            
+        if proto.retv is not None:
             self.out.append('\t%s, %s+%d(FP)' % (' '.join(REG_MAP[proto.retv.creg.reg]), proto.retv.name, offs))
-            
-            # Restore LR and Frame Pointer
-            # self.out.append('\tLDP -8(RSP), (R29, R30)')
-            # self.out.append('\tADD $16, RSP')
-
+        # Restore LR and Frame Pointer
+        # self.out.append('\tLDP -8(RSP), (R29, R30)')
+        # self.out.append('\tADD $16, RSP')
         
         self.out.append('\tRET')
 
@@ -2159,14 +2154,12 @@ def _declare_functions(self, protos: PrototypeMap):
                 raise SyntaxError('function prototype must have a "_" prefix: ' + repr(name))
 
     def parse(self, src: List[str], proto: PrototypeMap):
-        self.code.instr(Instruction('adr x0, .'))
+        # self.code.instr(Instruction('adr x0, .'))
         # self.code.instr(Instruction('add sp, sp, #%d'%self.code.stacksize(name)))
-        self.code.instr(Instruction('ret'))
-
-        # align 16 bytes
-        cmd = Command.parse(".p2align 4")
-        func = self._commands.get(cmd.cmd)
-        func(cmd.args)
+        # self.code.instr(Instruction('ret'))
+        # cmd = Command.parse(".p2align 4")
+        # func = self._commands.get(cmd.cmd)
+        # func(cmd.args)
 
         self._parse(src)
         self._declare(proto)