diff --git a/sys/crypto/openssl/aarch64/armv8-mont.S b/sys/crypto/openssl/aarch64/armv8-mont.S
index 99f3a3d07ff94..4bdba95f80f8c 100644
--- a/sys/crypto/openssl/aarch64/armv8-mont.S
+++ b/sys/crypto/openssl/aarch64/armv8-mont.S
@@ -1,14 +1,32 @@
 /* Do not modify. This file is auto-generated from armv8-mont.pl. */
+#ifndef	__KERNEL__
+# include "arm_arch.h"
+
+.hidden	OPENSSL_armv8_rsa_neonized
+#endif
 .text
 
 .globl	bn_mul_mont
 .type	bn_mul_mont,%function
 .align	5
 bn_mul_mont:
+.Lbn_mul_mont:
+	tst	x5,#3
+	b.ne	.Lmul_mont
+	cmp	x5,#32
+	b.le	.Lscalar_impl
+#ifndef	__KERNEL__
+	adrp	x17,OPENSSL_armv8_rsa_neonized
+	ldr	w17,[x17,#:lo12:OPENSSL_armv8_rsa_neonized]
+	cbnz	w17, bn_mul8x_mont_neon
+#endif
+
+.Lscalar_impl:
 	tst	x5,#7
 	b.eq	__bn_sqr8x_mont
 	tst	x5,#3
 	b.eq	__bn_mul4x_mont
+
 .Lmul_mont:
 	stp	x29,x30,[sp,#-64]!
 	add	x29,sp,#0
@@ -132,7 +150,7 @@ bn_mul_mont:
 	mul	x16,x14,x15		// np[j]*m1
 	adds	x12,x12,x6
 	umulh	x17,x14,x15
-	str	x12,[x22,#-16]		// tp[j-1]
+	stur	x12,[x22,#-16]		// tp[j-1]
 	cbnz	x21,.Linner
 
 .Linner_skip:
@@ -188,13 +206,13 @@ bn_mul_mont:
 	csel	x14,x23,x8,lo		// did it borrow?
 	ldr	x23,[x22],#8
 	ldr	x8,[x0],#8
-	str	xzr,[x22,#-16]		// wipe tp
-	str	x14,[x0,#-16]
+	stur	xzr,[x22,#-16]		// wipe tp
+	stur	x14,[x0,#-16]
 	cbnz	x5,.Lcond_copy
 
 	csel	x14,x23,x8,lo
-	str	xzr,[x22,#-8]		// wipe tp
-	str	x14,[x0,#-8]
+	stur	xzr,[x22,#-8]		// wipe tp
+	stur	x14,[x0,#-8]
 
 	ldp	x19,x20,[x29,#16]
 	mov	sp,x29
@@ -204,6 +222,704 @@ bn_mul_mont:
 	ldr	x29,[sp],#64
 	ret
 .size	bn_mul_mont,.-bn_mul_mont
+.type	bn_mul8x_mont_neon,%function
+.align	5
+bn_mul8x_mont_neon:
+	stp	x29,x30,[sp,#-80]!
+	mov	x16,sp
+	stp	d8,d9,[sp,#16]
+	stp	d10,d11,[sp,#32]
+	stp	d12,d13,[sp,#48]
+	stp	d14,d15,[sp,#64]
+	lsl	x5,x5,#1
+	eor	v14.16b,v14.16b,v14.16b
+
+.align	4
+.LNEON_8n:
+	eor	v6.16b,v6.16b,v6.16b
+	sub	x7,sp,#128
+	eor	v7.16b,v7.16b,v7.16b
+	sub	x7,x7,x5,lsl#4
+	eor	v8.16b,v8.16b,v8.16b
+	and	x7,x7,#-64
+	eor	v9.16b,v9.16b,v9.16b
+	mov	sp,x7		// alloca
+	eor	v10.16b,v10.16b,v10.16b
+	add	x7,x7,#256
+	eor	v11.16b,v11.16b,v11.16b
+	sub	x8,x5,#8
+	eor	v12.16b,v12.16b,v12.16b
+	eor	v13.16b,v13.16b,v13.16b
+
+.LNEON_8n_init:
+	st1	{v6.2d,v7.2d},[x7],#32
+	subs	x8,x8,#8
+	st1	{v8.2d,v9.2d},[x7],#32
+	st1	{v10.2d,v11.2d},[x7],#32
+	st1	{v12.2d,v13.2d},[x7],#32
+	bne	.LNEON_8n_init
+
+	add	x6,sp,#256
+	ld1	{v0.4s,v1.4s},[x1],#32
+	add	x10,sp,#8
+	ldr	s30,[x4],#4
+	mov	x9,x5
+	b	.LNEON_8n_outer
+
+.align	4
+.LNEON_8n_outer:
+	ldr	s28,[x2],#4   // *b++
+	uxtl	v28.4s,v28.4h
+	add	x7,sp,#128
+	ld1	{v2.4s,v3.4s},[x3],#32
+
+	umlal	v6.2d,v28.2s,v0.s[0]
+	umlal	v7.2d,v28.2s,v0.s[1]
+	umlal	v8.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v6.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v9.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v6.2d
+	umlal	v10.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v11.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[sp]		// put aside smashed b[8*i+0]
+	umlal	v12.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v13.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v6.2d,v29.2s,v2.s[0]
+	umlal	v7.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v8.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v6.2d,#16
+	umlal	v9.2d,v29.2s,v2.s[3]
+	umlal	v10.2d,v29.2s,v3.s[0]
+	ext	v6.16b,v6.16b,v6.16b,#8
+	add	v6.2d,v6.2d,v15.2d
+	umlal	v11.2d,v29.2s,v3.s[1]
+	ushr	v6.2d,v6.2d,#16
+	umlal	v12.2d,v29.2s,v3.s[2]
+	umlal	v13.2d,v29.2s,v3.s[3]
+	add	v16.2d,v7.2d,v6.2d
+	ins	v7.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+0]
+	umlal	v7.2d,v28.2s,v0.s[0]
+	ld1	{v6.2d},[x6],#16
+	umlal	v8.2d,v28.2s,v0.s[1]
+	umlal	v9.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v7.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v10.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v7.2d
+	umlal	v11.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v12.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+1]
+	umlal	v13.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v6.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v7.2d,v29.2s,v2.s[0]
+	umlal	v8.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v9.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v7.2d,#16
+	umlal	v10.2d,v29.2s,v2.s[3]
+	umlal	v11.2d,v29.2s,v3.s[0]
+	ext	v7.16b,v7.16b,v7.16b,#8
+	add	v7.2d,v7.2d,v15.2d
+	umlal	v12.2d,v29.2s,v3.s[1]
+	ushr	v7.2d,v7.2d,#16
+	umlal	v13.2d,v29.2s,v3.s[2]
+	umlal	v6.2d,v29.2s,v3.s[3]
+	add	v16.2d,v8.2d,v7.2d
+	ins	v8.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+1]
+	umlal	v8.2d,v28.2s,v0.s[0]
+	ld1	{v7.2d},[x6],#16
+	umlal	v9.2d,v28.2s,v0.s[1]
+	umlal	v10.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v8.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v11.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v8.2d
+	umlal	v12.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v13.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+2]
+	umlal	v6.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v7.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v8.2d,v29.2s,v2.s[0]
+	umlal	v9.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v10.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v8.2d,#16
+	umlal	v11.2d,v29.2s,v2.s[3]
+	umlal	v12.2d,v29.2s,v3.s[0]
+	ext	v8.16b,v8.16b,v8.16b,#8
+	add	v8.2d,v8.2d,v15.2d
+	umlal	v13.2d,v29.2s,v3.s[1]
+	ushr	v8.2d,v8.2d,#16
+	umlal	v6.2d,v29.2s,v3.s[2]
+	umlal	v7.2d,v29.2s,v3.s[3]
+	add	v16.2d,v9.2d,v8.2d
+	ins	v9.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+2]
+	umlal	v9.2d,v28.2s,v0.s[0]
+	ld1	{v8.2d},[x6],#16
+	umlal	v10.2d,v28.2s,v0.s[1]
+	umlal	v11.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v9.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v12.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v9.2d
+	umlal	v13.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v6.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+3]
+	umlal	v7.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v8.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v9.2d,v29.2s,v2.s[0]
+	umlal	v10.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v11.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v9.2d,#16
+	umlal	v12.2d,v29.2s,v2.s[3]
+	umlal	v13.2d,v29.2s,v3.s[0]
+	ext	v9.16b,v9.16b,v9.16b,#8
+	add	v9.2d,v9.2d,v15.2d
+	umlal	v6.2d,v29.2s,v3.s[1]
+	ushr	v9.2d,v9.2d,#16
+	umlal	v7.2d,v29.2s,v3.s[2]
+	umlal	v8.2d,v29.2s,v3.s[3]
+	add	v16.2d,v10.2d,v9.2d
+	ins	v10.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+3]
+	umlal	v10.2d,v28.2s,v0.s[0]
+	ld1	{v9.2d},[x6],#16
+	umlal	v11.2d,v28.2s,v0.s[1]
+	umlal	v12.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v10.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v13.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v10.2d
+	umlal	v6.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v7.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+4]
+	umlal	v8.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v9.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v10.2d,v29.2s,v2.s[0]
+	umlal	v11.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v12.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v10.2d,#16
+	umlal	v13.2d,v29.2s,v2.s[3]
+	umlal	v6.2d,v29.2s,v3.s[0]
+	ext	v10.16b,v10.16b,v10.16b,#8
+	add	v10.2d,v10.2d,v15.2d
+	umlal	v7.2d,v29.2s,v3.s[1]
+	ushr	v10.2d,v10.2d,#16
+	umlal	v8.2d,v29.2s,v3.s[2]
+	umlal	v9.2d,v29.2s,v3.s[3]
+	add	v16.2d,v11.2d,v10.2d
+	ins	v11.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+4]
+	umlal	v11.2d,v28.2s,v0.s[0]
+	ld1	{v10.2d},[x6],#16
+	umlal	v12.2d,v28.2s,v0.s[1]
+	umlal	v13.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v11.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v6.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v11.2d
+	umlal	v7.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v8.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+5]
+	umlal	v9.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v10.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v11.2d,v29.2s,v2.s[0]
+	umlal	v12.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v13.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v11.2d,#16
+	umlal	v6.2d,v29.2s,v2.s[3]
+	umlal	v7.2d,v29.2s,v3.s[0]
+	ext	v11.16b,v11.16b,v11.16b,#8
+	add	v11.2d,v11.2d,v15.2d
+	umlal	v8.2d,v29.2s,v3.s[1]
+	ushr	v11.2d,v11.2d,#16
+	umlal	v9.2d,v29.2s,v3.s[2]
+	umlal	v10.2d,v29.2s,v3.s[3]
+	add	v16.2d,v12.2d,v11.2d
+	ins	v12.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+5]
+	umlal	v12.2d,v28.2s,v0.s[0]
+	ld1	{v11.2d},[x6],#16
+	umlal	v13.2d,v28.2s,v0.s[1]
+	umlal	v6.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v12.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v7.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v12.2d
+	umlal	v8.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v9.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+6]
+	umlal	v10.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v11.2d,v28.2s,v1.s[3]
+	ldr	s28,[x2],#4   // *b++
+	umlal	v12.2d,v29.2s,v2.s[0]
+	umlal	v13.2d,v29.2s,v2.s[1]
+	uxtl	v28.4s,v28.4h
+	umlal	v6.2d,v29.2s,v2.s[2]
+	ushr	v15.2d,v12.2d,#16
+	umlal	v7.2d,v29.2s,v2.s[3]
+	umlal	v8.2d,v29.2s,v3.s[0]
+	ext	v12.16b,v12.16b,v12.16b,#8
+	add	v12.2d,v12.2d,v15.2d
+	umlal	v9.2d,v29.2s,v3.s[1]
+	ushr	v12.2d,v12.2d,#16
+	umlal	v10.2d,v29.2s,v3.s[2]
+	umlal	v11.2d,v29.2s,v3.s[3]
+	add	v16.2d,v13.2d,v12.2d
+	ins	v13.d[0],v16.d[0]
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+6]
+	umlal	v13.2d,v28.2s,v0.s[0]
+	ld1	{v12.2d},[x6],#16
+	umlal	v6.2d,v28.2s,v0.s[1]
+	umlal	v7.2d,v28.2s,v0.s[2]
+	shl	v29.2d,v13.2d,#16
+	ext	v29.16b,v29.16b,v29.16b,#8
+	umlal	v8.2d,v28.2s,v0.s[3]
+	add	v29.2d,v29.2d,v13.2d
+	umlal	v9.2d,v28.2s,v1.s[0]
+	mul	v29.2s,v29.2s,v30.2s
+	umlal	v10.2d,v28.2s,v1.s[1]
+	st1	{v28.2s},[x10],#8	// put aside smashed b[8*i+7]
+	umlal	v11.2d,v28.2s,v1.s[2]
+	uxtl	v29.4s,v29.4h
+	umlal	v12.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[sp]		// pull smashed b[8*i+0]
+	umlal	v13.2d,v29.2s,v2.s[0]
+	ld1	{v0.4s,v1.4s},[x1],#32
+	umlal	v6.2d,v29.2s,v2.s[1]
+	umlal	v7.2d,v29.2s,v2.s[2]
+	mov	v5.16b,v13.16b
+	ushr	v5.2d,v5.2d,#16
+	ext	v13.16b,v13.16b,v13.16b,#8
+	umlal	v8.2d,v29.2s,v2.s[3]
+	umlal	v9.2d,v29.2s,v3.s[0]
+	add	v13.2d,v13.2d,v5.2d
+	umlal	v10.2d,v29.2s,v3.s[1]
+	ushr	v13.2d,v13.2d,#16
+	eor	v15.16b,v15.16b,v15.16b
+	ins	v13.d[1],v15.d[0]
+	umlal	v11.2d,v29.2s,v3.s[2]
+	umlal	v12.2d,v29.2s,v3.s[3]
+	add	v6.2d,v6.2d,v13.2d
+	st1	{v29.2s},[x10],#8	// put aside smashed m[8*i+7]
+	add	x10,sp,#8		// rewind
+	sub	x8,x5,#8
+	b	.LNEON_8n_inner
+
+.align	4
+.LNEON_8n_inner:
+	subs	x8,x8,#8
+	umlal	v6.2d,v28.2s,v0.s[0]
+	ld1	{v13.2d},[x6]
+	umlal	v7.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+0]
+	umlal	v8.2d,v28.2s,v0.s[2]
+	ld1	{v2.4s,v3.4s},[x3],#32
+	umlal	v9.2d,v28.2s,v0.s[3]
+	b.eq	.LInner_jump
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump:
+	umlal	v10.2d,v28.2s,v1.s[0]
+	umlal	v11.2d,v28.2s,v1.s[1]
+	umlal	v12.2d,v28.2s,v1.s[2]
+	umlal	v13.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+1]
+	umlal	v6.2d,v29.2s,v2.s[0]
+	umlal	v7.2d,v29.2s,v2.s[1]
+	umlal	v8.2d,v29.2s,v2.s[2]
+	umlal	v9.2d,v29.2s,v2.s[3]
+	umlal	v10.2d,v29.2s,v3.s[0]
+	umlal	v11.2d,v29.2s,v3.s[1]
+	umlal	v12.2d,v29.2s,v3.s[2]
+	umlal	v13.2d,v29.2s,v3.s[3]
+	st1	{v6.2d},[x7],#16
+	umlal	v7.2d,v28.2s,v0.s[0]
+	ld1	{v6.2d},[x6]
+	umlal	v8.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+1]
+	umlal	v9.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump1
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump1:
+	umlal	v10.2d,v28.2s,v0.s[3]
+	umlal	v11.2d,v28.2s,v1.s[0]
+	umlal	v12.2d,v28.2s,v1.s[1]
+	umlal	v13.2d,v28.2s,v1.s[2]
+	umlal	v6.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+2]
+	umlal	v7.2d,v29.2s,v2.s[0]
+	umlal	v8.2d,v29.2s,v2.s[1]
+	umlal	v9.2d,v29.2s,v2.s[2]
+	umlal	v10.2d,v29.2s,v2.s[3]
+	umlal	v11.2d,v29.2s,v3.s[0]
+	umlal	v12.2d,v29.2s,v3.s[1]
+	umlal	v13.2d,v29.2s,v3.s[2]
+	umlal	v6.2d,v29.2s,v3.s[3]
+	st1	{v7.2d},[x7],#16
+	umlal	v8.2d,v28.2s,v0.s[0]
+	ld1	{v7.2d},[x6]
+	umlal	v9.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+2]
+	umlal	v10.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump2
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump2:
+	umlal	v11.2d,v28.2s,v0.s[3]
+	umlal	v12.2d,v28.2s,v1.s[0]
+	umlal	v13.2d,v28.2s,v1.s[1]
+	umlal	v6.2d,v28.2s,v1.s[2]
+	umlal	v7.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+3]
+	umlal	v8.2d,v29.2s,v2.s[0]
+	umlal	v9.2d,v29.2s,v2.s[1]
+	umlal	v10.2d,v29.2s,v2.s[2]
+	umlal	v11.2d,v29.2s,v2.s[3]
+	umlal	v12.2d,v29.2s,v3.s[0]
+	umlal	v13.2d,v29.2s,v3.s[1]
+	umlal	v6.2d,v29.2s,v3.s[2]
+	umlal	v7.2d,v29.2s,v3.s[3]
+	st1	{v8.2d},[x7],#16
+	umlal	v9.2d,v28.2s,v0.s[0]
+	ld1	{v8.2d},[x6]
+	umlal	v10.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+3]
+	umlal	v11.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump3
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump3:
+	umlal	v12.2d,v28.2s,v0.s[3]
+	umlal	v13.2d,v28.2s,v1.s[0]
+	umlal	v6.2d,v28.2s,v1.s[1]
+	umlal	v7.2d,v28.2s,v1.s[2]
+	umlal	v8.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+4]
+	umlal	v9.2d,v29.2s,v2.s[0]
+	umlal	v10.2d,v29.2s,v2.s[1]
+	umlal	v11.2d,v29.2s,v2.s[2]
+	umlal	v12.2d,v29.2s,v2.s[3]
+	umlal	v13.2d,v29.2s,v3.s[0]
+	umlal	v6.2d,v29.2s,v3.s[1]
+	umlal	v7.2d,v29.2s,v3.s[2]
+	umlal	v8.2d,v29.2s,v3.s[3]
+	st1	{v9.2d},[x7],#16
+	umlal	v10.2d,v28.2s,v0.s[0]
+	ld1	{v9.2d},[x6]
+	umlal	v11.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+4]
+	umlal	v12.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump4
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump4:
+	umlal	v13.2d,v28.2s,v0.s[3]
+	umlal	v6.2d,v28.2s,v1.s[0]
+	umlal	v7.2d,v28.2s,v1.s[1]
+	umlal	v8.2d,v28.2s,v1.s[2]
+	umlal	v9.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+5]
+	umlal	v10.2d,v29.2s,v2.s[0]
+	umlal	v11.2d,v29.2s,v2.s[1]
+	umlal	v12.2d,v29.2s,v2.s[2]
+	umlal	v13.2d,v29.2s,v2.s[3]
+	umlal	v6.2d,v29.2s,v3.s[0]
+	umlal	v7.2d,v29.2s,v3.s[1]
+	umlal	v8.2d,v29.2s,v3.s[2]
+	umlal	v9.2d,v29.2s,v3.s[3]
+	st1	{v10.2d},[x7],#16
+	umlal	v11.2d,v28.2s,v0.s[0]
+	ld1	{v10.2d},[x6]
+	umlal	v12.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+5]
+	umlal	v13.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump5
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump5:
+	umlal	v6.2d,v28.2s,v0.s[3]
+	umlal	v7.2d,v28.2s,v1.s[0]
+	umlal	v8.2d,v28.2s,v1.s[1]
+	umlal	v9.2d,v28.2s,v1.s[2]
+	umlal	v10.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+6]
+	umlal	v11.2d,v29.2s,v2.s[0]
+	umlal	v12.2d,v29.2s,v2.s[1]
+	umlal	v13.2d,v29.2s,v2.s[2]
+	umlal	v6.2d,v29.2s,v2.s[3]
+	umlal	v7.2d,v29.2s,v3.s[0]
+	umlal	v8.2d,v29.2s,v3.s[1]
+	umlal	v9.2d,v29.2s,v3.s[2]
+	umlal	v10.2d,v29.2s,v3.s[3]
+	st1	{v11.2d},[x7],#16
+	umlal	v12.2d,v28.2s,v0.s[0]
+	ld1	{v11.2d},[x6]
+	umlal	v13.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+6]
+	umlal	v6.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump6
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump6:
+	umlal	v7.2d,v28.2s,v0.s[3]
+	umlal	v8.2d,v28.2s,v1.s[0]
+	umlal	v9.2d,v28.2s,v1.s[1]
+	umlal	v10.2d,v28.2s,v1.s[2]
+	umlal	v11.2d,v28.2s,v1.s[3]
+	ld1	{v28.2s},[x10],#8	// pull smashed b[8*i+7]
+	umlal	v12.2d,v29.2s,v2.s[0]
+	umlal	v13.2d,v29.2s,v2.s[1]
+	umlal	v6.2d,v29.2s,v2.s[2]
+	umlal	v7.2d,v29.2s,v2.s[3]
+	umlal	v8.2d,v29.2s,v3.s[0]
+	umlal	v9.2d,v29.2s,v3.s[1]
+	umlal	v10.2d,v29.2s,v3.s[2]
+	umlal	v11.2d,v29.2s,v3.s[3]
+	st1	{v12.2d},[x7],#16
+	umlal	v13.2d,v28.2s,v0.s[0]
+	ld1	{v12.2d},[x6]
+	umlal	v6.2d,v28.2s,v0.s[1]
+	ld1	{v29.2s},[x10],#8	// pull smashed m[8*i+7]
+	umlal	v7.2d,v28.2s,v0.s[2]
+	b.eq	.LInner_jump7
+	add	x6,x6,#16	// don't advance in last iteration
+.LInner_jump7:
+	umlal	v8.2d,v28.2s,v0.s[3]
+	umlal	v9.2d,v28.2s,v1.s[0]
+	umlal	v10.2d,v28.2s,v1.s[1]
+	umlal	v11.2d,v28.2s,v1.s[2]
+	umlal	v12.2d,v28.2s,v1.s[3]
+	b.ne	.LInner_after_rewind8
+	sub	x1,x1,x5,lsl#2	// rewind
+.LInner_after_rewind8:
+	umlal	v13.2d,v29.2s,v2.s[0]
+	ld1	{v28.2s},[sp]		// pull smashed b[8*i+0]
+	umlal	v6.2d,v29.2s,v2.s[1]
+	ld1	{v0.4s,v1.4s},[x1],#32
+	umlal	v7.2d,v29.2s,v2.s[2]
+	add	x10,sp,#8		// rewind
+	umlal	v8.2d,v29.2s,v2.s[3]
+	umlal	v9.2d,v29.2s,v3.s[0]
+	umlal	v10.2d,v29.2s,v3.s[1]
+	umlal	v11.2d,v29.2s,v3.s[2]
+	st1	{v13.2d},[x7],#16
+	umlal	v12.2d,v29.2s,v3.s[3]
+
+	bne	.LNEON_8n_inner
+	add	x6,sp,#128
+	st1	{v6.2d,v7.2d},[x7],#32
+	eor	v2.16b,v2.16b,v2.16b	// v2
+	st1	{v8.2d,v9.2d},[x7],#32
+	eor	v3.16b,v3.16b,v3.16b	// v3
+	st1	{v10.2d,v11.2d},[x7],#32
+	st1	{v12.2d},[x7]
+
+	subs	x9,x9,#8
+	ld1	{v6.2d,v7.2d},[x6],#32
+	ld1	{v8.2d,v9.2d},[x6],#32
+	ld1	{v10.2d,v11.2d},[x6],#32
+	ld1	{v12.2d,v13.2d},[x6],#32
+
+	b.eq	.LInner_8n_jump_2steps
+	sub	x3,x3,x5,lsl#2	// rewind
+	b	.LNEON_8n_outer
+
+.LInner_8n_jump_2steps:
+	add	x7,sp,#128
+	st1	{v2.2d,v3.2d}, [sp],#32	// start wiping stack frame
+	mov	v5.16b,v6.16b
+	ushr	v15.2d,v6.2d,#16
+	ext	v6.16b,v6.16b,v6.16b,#8
+	st1	{v2.2d,v3.2d}, [sp],#32
+	add	v6.2d,v6.2d,v15.2d
+	st1	{v2.2d,v3.2d}, [sp],#32
+	ushr	v15.2d,v6.2d,#16
+	st1	{v2.2d,v3.2d}, [sp],#32
+	zip1	v6.4h,v5.4h,v6.4h
+	ins	v15.d[1],v14.d[0]
+
+	mov	x8,x5
+	b	.LNEON_tail_entry
+
+.align	4
+.LNEON_tail:
+	add	v6.2d,v6.2d,v15.2d
+	mov	v5.16b,v6.16b
+	ushr	v15.2d,v6.2d,#16
+	ext	v6.16b,v6.16b,v6.16b,#8
+	ld1	{v8.2d,v9.2d}, [x6],#32
+	add	v6.2d,v6.2d,v15.2d
+	ld1	{v10.2d,v11.2d}, [x6],#32
+	ushr	v15.2d,v6.2d,#16
+	ld1	{v12.2d,v13.2d}, [x6],#32
+	zip1	v6.4h,v5.4h,v6.4h
+	ins	v15.d[1],v14.d[0]
+
+.LNEON_tail_entry:
+	add	v7.2d,v7.2d,v15.2d
+	st1	{v6.s}[0], [x7],#4
+	ushr	v15.2d,v7.2d,#16
+	mov	v5.16b,v7.16b
+	ext	v7.16b,v7.16b,v7.16b,#8
+	add	v7.2d,v7.2d,v15.2d
+	ushr	v15.2d,v7.2d,#16
+	zip1	v7.4h,v5.4h,v7.4h
+	ins	v15.d[1],v14.d[0]
+	add	v8.2d,v8.2d,v15.2d
+	st1	{v7.s}[0], [x7],#4
+	ushr	v15.2d,v8.2d,#16
+	mov	v5.16b,v8.16b
+	ext	v8.16b,v8.16b,v8.16b,#8
+	add	v8.2d,v8.2d,v15.2d
+	ushr	v15.2d,v8.2d,#16
+	zip1	v8.4h,v5.4h,v8.4h
+	ins	v15.d[1],v14.d[0]
+	add	v9.2d,v9.2d,v15.2d
+	st1	{v8.s}[0], [x7],#4
+	ushr	v15.2d,v9.2d,#16
+	mov	v5.16b,v9.16b
+	ext	v9.16b,v9.16b,v9.16b,#8
+	add	v9.2d,v9.2d,v15.2d
+	ushr	v15.2d,v9.2d,#16
+	zip1	v9.4h,v5.4h,v9.4h
+	ins	v15.d[1],v14.d[0]
+	add	v10.2d,v10.2d,v15.2d
+	st1	{v9.s}[0], [x7],#4
+	ushr	v15.2d,v10.2d,#16
+	mov	v5.16b,v10.16b
+	ext	v10.16b,v10.16b,v10.16b,#8
+	add	v10.2d,v10.2d,v15.2d
+	ushr	v15.2d,v10.2d,#16
+	zip1	v10.4h,v5.4h,v10.4h
+	ins	v15.d[1],v14.d[0]
+	add	v11.2d,v11.2d,v15.2d
+	st1	{v10.s}[0], [x7],#4
+	ushr	v15.2d,v11.2d,#16
+	mov	v5.16b,v11.16b
+	ext	v11.16b,v11.16b,v11.16b,#8
+	add	v11.2d,v11.2d,v15.2d
+	ushr	v15.2d,v11.2d,#16
+	zip1	v11.4h,v5.4h,v11.4h
+	ins	v15.d[1],v14.d[0]
+	add	v12.2d,v12.2d,v15.2d
+	st1	{v11.s}[0], [x7],#4
+	ushr	v15.2d,v12.2d,#16
+	mov	v5.16b,v12.16b
+	ext	v12.16b,v12.16b,v12.16b,#8
+	add	v12.2d,v12.2d,v15.2d
+	ushr	v15.2d,v12.2d,#16
+	zip1	v12.4h,v5.4h,v12.4h
+	ins	v15.d[1],v14.d[0]
+	add	v13.2d,v13.2d,v15.2d
+	st1	{v12.s}[0], [x7],#4
+	ushr	v15.2d,v13.2d,#16
+	mov	v5.16b,v13.16b
+	ext	v13.16b,v13.16b,v13.16b,#8
+	add	v13.2d,v13.2d,v15.2d
+	ushr	v15.2d,v13.2d,#16
+	zip1	v13.4h,v5.4h,v13.4h
+	ins	v15.d[1],v14.d[0]
+	ld1	{v6.2d,v7.2d}, [x6],#32
+	subs	x8,x8,#8
+	st1	{v13.s}[0], [x7],#4
+	bne	.LNEON_tail
+
+	st1	{v15.s}[0], [x7],#4	// top-most bit
+	sub	x3,x3,x5,lsl#2		// rewind x3
+	subs	x1,sp,#0			// clear carry flag
+	add	x2,sp,x5,lsl#2
+
+.LNEON_sub:
+	ldp	w4,w5,[x1],#8
+	ldp	w6,w7,[x1],#8
+	ldp	w8,w9,[x3],#8
+	ldp	w10,w11,[x3],#8
+	sbcs	w8,w4,w8
+	sbcs	w9,w5,w9
+	sbcs	w10,w6,w10
+	sbcs	w11,w7,w11
+	sub	x17,x2,x1
+	stp	w8,w9,[x0],#8
+	stp	w10,w11,[x0],#8
+	cbnz	x17,.LNEON_sub
+
+	ldr	w10, [x1]		// load top-most bit
+	mov	x11,sp
+	eor	v0.16b,v0.16b,v0.16b
+	sub	x11,x2,x11		// this is num*4
+	eor	v1.16b,v1.16b,v1.16b
+	mov	x1,sp
+	sub	x0,x0,x11		// rewind x0
+	mov	x3,x2		// second 3/4th of frame
+	sbcs	w10,w10,wzr		// result is carry flag
+
+.LNEON_copy_n_zap:
+	ldp	w4,w5,[x1],#8
+	ldp	w6,w7,[x1],#8
+	ldp	w8,w9,[x0],#8
+	ldp	w10,w11,[x0]
+	sub	x0,x0,#8
+	b.cs	.LCopy_1
+	mov	w8,w4
+	mov	w9,w5
+	mov	w10,w6
+	mov	w11,w7
+.LCopy_1:
+	st1	{v0.2d,v1.2d}, [x3],#32		// wipe
+	st1	{v0.2d,v1.2d}, [x3],#32		// wipe
+	ldp	w4,w5,[x1],#8
+	ldp	w6,w7,[x1],#8
+	stp	w8,w9,[x0],#8
+	stp	w10,w11,[x0],#8
+	sub	x1,x1,#32
+	ldp	w8,w9,[x0],#8
+	ldp	w10,w11,[x0]
+	sub	x0,x0,#8
+	b.cs	.LCopy_2
+	mov	w8, w4
+	mov	w9, w5
+	mov	w10, w6
+	mov	w11, w7
+.LCopy_2:
+	st1	{v0.2d,v1.2d}, [x1],#32		// wipe
+	st1	{v0.2d,v1.2d}, [x3],#32		// wipe
+	sub	x17,x2,x1		// preserves carry
+	stp	w8,w9,[x0],#8
+	stp	w10,w11,[x0],#8
+	cbnz	x17,.LNEON_copy_n_zap
+
+	mov	sp,x16
+	ldp	d14,d15,[sp,#64]
+	ldp	d12,d13,[sp,#48]
+	ldp	d10,d11,[sp,#32]
+	ldp	d8,d9,[sp,#16]
+	ldr	x29,[sp],#80
+	ret	// bx lr
+
+.size	bn_mul8x_mont_neon,.-bn_mul8x_mont_neon
 .type	__bn_sqr8x_mont,%function
 .align	5
 __bn_sqr8x_mont:
@@ -519,7 +1235,7 @@ __bn_sqr8x_mont:
 	ldp	x10,x11,[x2,#8*4]
 	ldp	x12,x13,[x2,#8*6]
 	adds	x19,x19,x6
-	ldr	x4,[x0,#-8*8]
+	ldur	x4,[x0,#-8*8]
 	adcs	x20,x20,x7
 	ldp	x6,x7,[x1,#8*0]
 	adcs	x21,x21,x8
@@ -714,7 +1430,7 @@ __bn_sqr8x_mont:
 	//adc	x28,xzr,xzr		// moved below
 	cbz	x27,.Lsqr8x8_post_condition
 
-	ldr	x4,[x2,#-8*8]
+	ldur	x4,[x2,#-8*8]
 	ldp	x6,x7,[x1,#8*0]
 	ldp	x8,x9,[x1,#8*2]
 	ldp	x10,x11,[x1,#8*4]
@@ -772,7 +1488,7 @@ __bn_sqr8x_mont:
 	ldp	x12,x13,[x2,#8*6]
 	cbz	x27,.Lsqr8x_tail_break
 
-	ldr	x4,[x0,#-8*8]
+	ldur	x4,[x0,#-8*8]
 	adds	x19,x19,x6
 	adcs	x20,x20,x7
 	ldp	x6,x7,[x1,#8*0]
diff --git a/sys/crypto/openssl/aarch64/chacha-armv8.S b/sys/crypto/openssl/aarch64/chacha-armv8.S
index 24db5ad307319..609e34f422a26 100644
--- a/sys/crypto/openssl/aarch64/chacha-armv8.S
+++ b/sys/crypto/openssl/aarch64/chacha-armv8.S
@@ -1,23 +1,20 @@
 /* Do not modify. This file is auto-generated from chacha-armv8.pl. */
-#include "arm_arch.h"
-
-.text
-
+#ifndef	__KERNEL__
+# include "arm_arch.h"
 
 .hidden	OPENSSL_armcap_P
+#endif
+
+.text
 
 .align	5
 .Lsigma:
 .quad	0x3320646e61707865,0x6b20657479622d32		// endian-neutral
 .Lone:
-.long	1,0,0,0
-.LOPENSSL_armcap_P:
-#ifdef	__ILP32__
-.long	OPENSSL_armcap_P-.
-#else
-.quad	OPENSSL_armcap_P-.
-#endif
-.byte	67,104,97,67,104,97,50,48,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.long	1,2,3,4
+.Lrot24:
+.long	0x02010003,0x06050407,0x0a09080b,0x0e0d0c0f
+.byte	67,104,97,67,104,97,50,48,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,64,100,111,116,45,97,115,109,0
 .align	2
 
 .globl	ChaCha20_ctr32
@@ -25,17 +22,15 @@
 .align	5
 ChaCha20_ctr32:
 	cbz	x2,.Labort
-	adr	x5,.LOPENSSL_armcap_P
 	cmp	x2,#192
 	b.lo	.Lshort
-#ifdef	__ILP32__
-	ldrsw	x6,[x5]
-#else
-	ldr	x6,[x5]
-#endif
-	ldr	w17,[x6,x5]
+
+#ifndef	__KERNEL__
+	adrp	x17,OPENSSL_armcap_P
+	ldr	w17,[x17,#:lo12:OPENSSL_armcap_P]
 	tst	w17,#ARMV7_NEON
-	b.ne	ChaCha20_neon
+	b.ne	.LChaCha20_neon
+#endif
 
 .Lshort:
 .inst	0xd503233f			// paciasp
@@ -54,7 +49,7 @@ ChaCha20_ctr32:
 	ldp	x24,x25,[x3]		// load key
 	ldp	x26,x27,[x3,#16]
 	ldp	x28,x30,[x4]		// load counter
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x24,x24,#32
 	ror	x25,x25,#32
 	ror	x26,x26,#32
@@ -215,7 +210,7 @@ ChaCha20_ctr32:
 	add	x20,x20,x21,lsl#32
 	ldp	x19,x21,[x1,#48]
 	add	x1,x1,#64
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -272,7 +267,7 @@ ChaCha20_ctr32:
 	add	x15,x15,x16,lsl#32
 	add	x17,x17,x19,lsl#32
 	add	x20,x20,x21,lsl#32
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -311,9 +306,13 @@ ChaCha20_ctr32:
 	ret
 .size	ChaCha20_ctr32,.-ChaCha20_ctr32
 
+#ifdef	__KERNEL__
+.globl	ChaCha20_neon
+#endif
 .type	ChaCha20_neon,%function
 .align	5
 ChaCha20_neon:
+.LChaCha20_neon:
 .inst	0xd503233f			// paciasp
 	stp	x29,x30,[sp,#-96]!
 	add	x29,sp,#0
@@ -330,15 +329,16 @@ ChaCha20_neon:
 	sub	sp,sp,#64
 
 	ldp	x22,x23,[x5]		// load sigma
-	ld1	{v24.4s},[x5],#16
+	ld1	{v0.4s},[x5],#16
 	ldp	x24,x25,[x3]		// load key
 	ldp	x26,x27,[x3,#16]
-	ld1	{v25.4s,v26.4s},[x3]
+	ld1	{v1.4s,v2.4s},[x3]
 	ldp	x28,x30,[x4]		// load counter
-	ld1	{v27.4s},[x4]
-	ld1	{v31.4s},[x5]
-#ifdef	__ARMEB__
-	rev64	v24.4s,v24.4s
+	ld1	{v3.4s},[x4]
+	stp	d8,d9,[sp]			// meet ABI requirements
+	ld1	{v8.4s,v9.4s},[x5]
+#ifdef	__AARCH64EB__
+	rev64	v0.4s,v0.4s
 	ror	x24,x24,#32
 	ror	x25,x25,#32
 	ror	x26,x26,#32
@@ -346,296 +346,330 @@ ChaCha20_neon:
 	ror	x28,x28,#32
 	ror	x30,x30,#32
 #endif
-	add	v27.4s,v27.4s,v31.4s		// += 1
-	add	v28.4s,v27.4s,v31.4s
-	add	v29.4s,v28.4s,v31.4s
-	shl	v31.4s,v31.4s,#2			// 1 -> 4
 
 .Loop_outer_neon:
-	mov	w5,w22			// unpack key block
+	dup	v16.4s,v0.s[0]			// unpack key block
+	mov	w5,w22
+	dup	v20.4s,v0.s[1]
 	lsr	x6,x22,#32
-	mov	v0.16b,v24.16b
+	dup	v24.4s,v0.s[2]
 	mov	w7,w23
+	dup	v28.4s,v0.s[3]
 	lsr	x8,x23,#32
-	mov	v4.16b,v24.16b
+	dup	v17.4s,v1.s[0]
 	mov	w9,w24
+	dup	v21.4s,v1.s[1]
 	lsr	x10,x24,#32
-	mov	v16.16b,v24.16b
+	dup	v25.4s,v1.s[2]
 	mov	w11,w25
-	mov	v1.16b,v25.16b
+	dup	v29.4s,v1.s[3]
 	lsr	x12,x25,#32
-	mov	v5.16b,v25.16b
+	dup	v19.4s,v3.s[0]
 	mov	w13,w26
-	mov	v17.16b,v25.16b
+	dup	v23.4s,v3.s[1]
 	lsr	x14,x26,#32
-	mov	v3.16b,v27.16b
+	dup	v27.4s,v3.s[2]
 	mov	w15,w27
-	mov	v7.16b,v28.16b
+	dup	v31.4s,v3.s[3]
 	lsr	x16,x27,#32
-	mov	v19.16b,v29.16b
+	add	v19.4s,v19.4s,v8.4s
 	mov	w17,w28
-	mov	v2.16b,v26.16b
+	dup	v18.4s,v2.s[0]
 	lsr	x19,x28,#32
-	mov	v6.16b,v26.16b
+	dup	v22.4s,v2.s[1]
 	mov	w20,w30
-	mov	v18.16b,v26.16b
+	dup	v26.4s,v2.s[2]
 	lsr	x21,x30,#32
+	dup	v30.4s,v2.s[3]
 
 	mov	x4,#10
-	subs	x2,x2,#256
+	subs	x2,x2,#320
 .Loop_neon:
 	sub	x4,x4,#1
-	add	v0.4s,v0.4s,v1.4s
+	add	v16.4s,v16.4s,v17.4s
 	add	w5,w5,w9
-	add	v4.4s,v4.4s,v5.4s
+	add	v20.4s,v20.4s,v21.4s
 	add	w6,w6,w10
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	add	w7,w7,w11
-	eor	v3.16b,v3.16b,v0.16b
+	add	v28.4s,v28.4s,v29.4s
 	add	w8,w8,w12
-	eor	v7.16b,v7.16b,v4.16b
-	eor	w17,w17,w5
 	eor	v19.16b,v19.16b,v16.16b
+	eor	w17,w17,w5
+	eor	v23.16b,v23.16b,v20.16b
 	eor	w19,w19,w6
-	rev32	v3.8h,v3.8h
+	eor	v27.16b,v27.16b,v24.16b
 	eor	w20,w20,w7
-	rev32	v7.8h,v7.8h
+	eor	v31.16b,v31.16b,v28.16b
 	eor	w21,w21,w8
 	rev32	v19.8h,v19.8h
 	ror	w17,w17,#16
-	add	v2.4s,v2.4s,v3.4s
+	rev32	v23.8h,v23.8h
 	ror	w19,w19,#16
-	add	v6.4s,v6.4s,v7.4s
+	rev32	v27.8h,v27.8h
 	ror	w20,w20,#16
-	add	v18.4s,v18.4s,v19.4s
+	rev32	v31.8h,v31.8h
 	ror	w21,w21,#16
-	eor	v20.16b,v1.16b,v2.16b
+	add	v18.4s,v18.4s,v19.4s
 	add	w13,w13,w17
-	eor	v21.16b,v5.16b,v6.16b
+	add	v22.4s,v22.4s,v23.4s
 	add	w14,w14,w19
-	eor	v22.16b,v17.16b,v18.16b
+	add	v26.4s,v26.4s,v27.4s
 	add	w15,w15,w20
-	ushr	v1.4s,v20.4s,#20
+	add	v30.4s,v30.4s,v31.4s
 	add	w16,w16,w21
-	ushr	v5.4s,v21.4s,#20
+	eor	v4.16b,v17.16b,v18.16b
 	eor	w9,w9,w13
-	ushr	v17.4s,v22.4s,#20
+	eor	v5.16b,v21.16b,v22.16b
 	eor	w10,w10,w14
-	sli	v1.4s,v20.4s,#12
+	eor	v6.16b,v25.16b,v26.16b
 	eor	w11,w11,w15
-	sli	v5.4s,v21.4s,#12
+	eor	v7.16b,v29.16b,v30.16b
 	eor	w12,w12,w16
-	sli	v17.4s,v22.4s,#12
+	ushr	v17.4s,v4.4s,#20
 	ror	w9,w9,#20
-	add	v0.4s,v0.4s,v1.4s
+	ushr	v21.4s,v5.4s,#20
 	ror	w10,w10,#20
-	add	v4.4s,v4.4s,v5.4s
+	ushr	v25.4s,v6.4s,#20
 	ror	w11,w11,#20
-	add	v16.4s,v16.4s,v17.4s
+	ushr	v29.4s,v7.4s,#20
 	ror	w12,w12,#20
-	eor	v20.16b,v3.16b,v0.16b
+	sli	v17.4s,v4.4s,#12
 	add	w5,w5,w9
-	eor	v21.16b,v7.16b,v4.16b
+	sli	v21.4s,v5.4s,#12
 	add	w6,w6,w10
-	eor	v22.16b,v19.16b,v16.16b
+	sli	v25.4s,v6.4s,#12
 	add	w7,w7,w11
-	ushr	v3.4s,v20.4s,#24
+	sli	v29.4s,v7.4s,#12
 	add	w8,w8,w12
-	ushr	v7.4s,v21.4s,#24
+	add	v16.4s,v16.4s,v17.4s
 	eor	w17,w17,w5
-	ushr	v19.4s,v22.4s,#24
+	add	v20.4s,v20.4s,v21.4s
 	eor	w19,w19,w6
-	sli	v3.4s,v20.4s,#8
+	add	v24.4s,v24.4s,v25.4s
 	eor	w20,w20,w7
-	sli	v7.4s,v21.4s,#8
+	add	v28.4s,v28.4s,v29.4s
 	eor	w21,w21,w8
-	sli	v19.4s,v22.4s,#8
+	eor	v4.16b,v19.16b,v16.16b
 	ror	w17,w17,#24
-	add	v2.4s,v2.4s,v3.4s
+	eor	v5.16b,v23.16b,v20.16b
 	ror	w19,w19,#24
-	add	v6.4s,v6.4s,v7.4s
+	eor	v6.16b,v27.16b,v24.16b
 	ror	w20,w20,#24
-	add	v18.4s,v18.4s,v19.4s
+	eor	v7.16b,v31.16b,v28.16b
 	ror	w21,w21,#24
-	eor	v20.16b,v1.16b,v2.16b
+	tbl	v19.16b,{v4.16b},v9.16b
 	add	w13,w13,w17
-	eor	v21.16b,v5.16b,v6.16b
+	tbl	v23.16b,{v5.16b},v9.16b
 	add	w14,w14,w19
-	eor	v22.16b,v17.16b,v18.16b
+	tbl	v27.16b,{v6.16b},v9.16b
 	add	w15,w15,w20
-	ushr	v1.4s,v20.4s,#25
+	tbl	v31.16b,{v7.16b},v9.16b
 	add	w16,w16,w21
-	ushr	v5.4s,v21.4s,#25
+	add	v18.4s,v18.4s,v19.4s
 	eor	w9,w9,w13
-	ushr	v17.4s,v22.4s,#25
+	add	v22.4s,v22.4s,v23.4s
 	eor	w10,w10,w14
-	sli	v1.4s,v20.4s,#7
+	add	v26.4s,v26.4s,v27.4s
 	eor	w11,w11,w15
-	sli	v5.4s,v21.4s,#7
+	add	v30.4s,v30.4s,v31.4s
 	eor	w12,w12,w16
-	sli	v17.4s,v22.4s,#7
+	eor	v4.16b,v17.16b,v18.16b
 	ror	w9,w9,#25
-	ext	v2.16b,v2.16b,v2.16b,#8
+	eor	v5.16b,v21.16b,v22.16b
 	ror	w10,w10,#25
-	ext	v6.16b,v6.16b,v6.16b,#8
+	eor	v6.16b,v25.16b,v26.16b
 	ror	w11,w11,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	eor	v7.16b,v29.16b,v30.16b
 	ror	w12,w12,#25
-	ext	v3.16b,v3.16b,v3.16b,#12
-	ext	v7.16b,v7.16b,v7.16b,#12
-	ext	v19.16b,v19.16b,v19.16b,#12
-	ext	v1.16b,v1.16b,v1.16b,#4
-	ext	v5.16b,v5.16b,v5.16b,#4
-	ext	v17.16b,v17.16b,v17.16b,#4
-	add	v0.4s,v0.4s,v1.4s
+	ushr	v17.4s,v4.4s,#25
+	ushr	v21.4s,v5.4s,#25
+	ushr	v25.4s,v6.4s,#25
+	ushr	v29.4s,v7.4s,#25
+	sli	v17.4s,v4.4s,#7
+	sli	v21.4s,v5.4s,#7
+	sli	v25.4s,v6.4s,#7
+	sli	v29.4s,v7.4s,#7
+	add	v16.4s,v16.4s,v21.4s
 	add	w5,w5,w10
-	add	v4.4s,v4.4s,v5.4s
+	add	v20.4s,v20.4s,v25.4s
 	add	w6,w6,w11
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v29.4s
 	add	w7,w7,w12
-	eor	v3.16b,v3.16b,v0.16b
+	add	v28.4s,v28.4s,v17.4s
 	add	w8,w8,w9
-	eor	v7.16b,v7.16b,v4.16b
+	eor	v31.16b,v31.16b,v16.16b
 	eor	w21,w21,w5
-	eor	v19.16b,v19.16b,v16.16b
+	eor	v19.16b,v19.16b,v20.16b
 	eor	w17,w17,w6
-	rev32	v3.8h,v3.8h
+	eor	v23.16b,v23.16b,v24.16b
 	eor	w19,w19,w7
-	rev32	v7.8h,v7.8h
+	eor	v27.16b,v27.16b,v28.16b
 	eor	w20,w20,w8
-	rev32	v19.8h,v19.8h
+	rev32	v31.8h,v31.8h
 	ror	w21,w21,#16
-	add	v2.4s,v2.4s,v3.4s
+	rev32	v19.8h,v19.8h
 	ror	w17,w17,#16
-	add	v6.4s,v6.4s,v7.4s
+	rev32	v23.8h,v23.8h
 	ror	w19,w19,#16
-	add	v18.4s,v18.4s,v19.4s
+	rev32	v27.8h,v27.8h
 	ror	w20,w20,#16
-	eor	v20.16b,v1.16b,v2.16b
+	add	v26.4s,v26.4s,v31.4s
 	add	w15,w15,w21
-	eor	v21.16b,v5.16b,v6.16b
+	add	v30.4s,v30.4s,v19.4s
 	add	w16,w16,w17
-	eor	v22.16b,v17.16b,v18.16b
+	add	v18.4s,v18.4s,v23.4s
 	add	w13,w13,w19
-	ushr	v1.4s,v20.4s,#20
+	add	v22.4s,v22.4s,v27.4s
 	add	w14,w14,w20
-	ushr	v5.4s,v21.4s,#20
+	eor	v4.16b,v21.16b,v26.16b
 	eor	w10,w10,w15
-	ushr	v17.4s,v22.4s,#20
+	eor	v5.16b,v25.16b,v30.16b
 	eor	w11,w11,w16
-	sli	v1.4s,v20.4s,#12
+	eor	v6.16b,v29.16b,v18.16b
 	eor	w12,w12,w13
-	sli	v5.4s,v21.4s,#12
+	eor	v7.16b,v17.16b,v22.16b
 	eor	w9,w9,w14
-	sli	v17.4s,v22.4s,#12
+	ushr	v21.4s,v4.4s,#20
 	ror	w10,w10,#20
-	add	v0.4s,v0.4s,v1.4s
+	ushr	v25.4s,v5.4s,#20
 	ror	w11,w11,#20
-	add	v4.4s,v4.4s,v5.4s
+	ushr	v29.4s,v6.4s,#20
 	ror	w12,w12,#20
-	add	v16.4s,v16.4s,v17.4s
+	ushr	v17.4s,v7.4s,#20
 	ror	w9,w9,#20
-	eor	v20.16b,v3.16b,v0.16b
+	sli	v21.4s,v4.4s,#12
 	add	w5,w5,w10
-	eor	v21.16b,v7.16b,v4.16b
+	sli	v25.4s,v5.4s,#12
 	add	w6,w6,w11
-	eor	v22.16b,v19.16b,v16.16b
+	sli	v29.4s,v6.4s,#12
 	add	w7,w7,w12
-	ushr	v3.4s,v20.4s,#24
+	sli	v17.4s,v7.4s,#12
 	add	w8,w8,w9
-	ushr	v7.4s,v21.4s,#24
+	add	v16.4s,v16.4s,v21.4s
 	eor	w21,w21,w5
-	ushr	v19.4s,v22.4s,#24
+	add	v20.4s,v20.4s,v25.4s
 	eor	w17,w17,w6
-	sli	v3.4s,v20.4s,#8
+	add	v24.4s,v24.4s,v29.4s
 	eor	w19,w19,w7
-	sli	v7.4s,v21.4s,#8
+	add	v28.4s,v28.4s,v17.4s
 	eor	w20,w20,w8
-	sli	v19.4s,v22.4s,#8
+	eor	v4.16b,v31.16b,v16.16b
 	ror	w21,w21,#24
-	add	v2.4s,v2.4s,v3.4s
+	eor	v5.16b,v19.16b,v20.16b
 	ror	w17,w17,#24
-	add	v6.4s,v6.4s,v7.4s
+	eor	v6.16b,v23.16b,v24.16b
 	ror	w19,w19,#24
-	add	v18.4s,v18.4s,v19.4s
+	eor	v7.16b,v27.16b,v28.16b
 	ror	w20,w20,#24
-	eor	v20.16b,v1.16b,v2.16b
+	tbl	v31.16b,{v4.16b},v9.16b
 	add	w15,w15,w21
-	eor	v21.16b,v5.16b,v6.16b
+	tbl	v19.16b,{v5.16b},v9.16b
 	add	w16,w16,w17
-	eor	v22.16b,v17.16b,v18.16b
+	tbl	v23.16b,{v6.16b},v9.16b
 	add	w13,w13,w19
-	ushr	v1.4s,v20.4s,#25
+	tbl	v27.16b,{v7.16b},v9.16b
 	add	w14,w14,w20
-	ushr	v5.4s,v21.4s,#25
+	add	v26.4s,v26.4s,v31.4s
 	eor	w10,w10,w15
-	ushr	v17.4s,v22.4s,#25
+	add	v30.4s,v30.4s,v19.4s
 	eor	w11,w11,w16
-	sli	v1.4s,v20.4s,#7
+	add	v18.4s,v18.4s,v23.4s
 	eor	w12,w12,w13
-	sli	v5.4s,v21.4s,#7
+	add	v22.4s,v22.4s,v27.4s
 	eor	w9,w9,w14
-	sli	v17.4s,v22.4s,#7
+	eor	v4.16b,v21.16b,v26.16b
 	ror	w10,w10,#25
-	ext	v2.16b,v2.16b,v2.16b,#8
+	eor	v5.16b,v25.16b,v30.16b
 	ror	w11,w11,#25
-	ext	v6.16b,v6.16b,v6.16b,#8
+	eor	v6.16b,v29.16b,v18.16b
 	ror	w12,w12,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	eor	v7.16b,v17.16b,v22.16b
 	ror	w9,w9,#25
-	ext	v3.16b,v3.16b,v3.16b,#4
-	ext	v7.16b,v7.16b,v7.16b,#4
-	ext	v19.16b,v19.16b,v19.16b,#4
-	ext	v1.16b,v1.16b,v1.16b,#12
-	ext	v5.16b,v5.16b,v5.16b,#12
-	ext	v17.16b,v17.16b,v17.16b,#12
+	ushr	v21.4s,v4.4s,#25
+	ushr	v25.4s,v5.4s,#25
+	ushr	v29.4s,v6.4s,#25
+	ushr	v17.4s,v7.4s,#25
+	sli	v21.4s,v4.4s,#7
+	sli	v25.4s,v5.4s,#7
+	sli	v29.4s,v6.4s,#7
+	sli	v17.4s,v7.4s,#7
 	cbnz	x4,.Loop_neon
 
+	add	v19.4s,v19.4s,v8.4s
+
+	zip1	v4.4s,v16.4s,v20.4s			// transpose data
+	zip1	v5.4s,v24.4s,v28.4s
+	zip2	v6.4s,v16.4s,v20.4s
+	zip2	v7.4s,v24.4s,v28.4s
+	zip1	v16.2d,v4.2d,v5.2d
+	zip2	v20.2d,v4.2d,v5.2d
+	zip1	v24.2d,v6.2d,v7.2d
+	zip2	v28.2d,v6.2d,v7.2d
+
+	zip1	v4.4s,v17.4s,v21.4s
+	zip1	v5.4s,v25.4s,v29.4s
+	zip2	v6.4s,v17.4s,v21.4s
+	zip2	v7.4s,v25.4s,v29.4s
+	zip1	v17.2d,v4.2d,v5.2d
+	zip2	v21.2d,v4.2d,v5.2d
+	zip1	v25.2d,v6.2d,v7.2d
+	zip2	v29.2d,v6.2d,v7.2d
+
+	zip1	v4.4s,v18.4s,v22.4s
 	add	w5,w5,w22		// accumulate key block
-	add	v0.4s,v0.4s,v24.4s
+	zip1	v5.4s,v26.4s,v30.4s
 	add	x6,x6,x22,lsr#32
-	add	v4.4s,v4.4s,v24.4s
+	zip2	v6.4s,v18.4s,v22.4s
 	add	w7,w7,w23
-	add	v16.4s,v16.4s,v24.4s
+	zip2	v7.4s,v26.4s,v30.4s
 	add	x8,x8,x23,lsr#32
-	add	v2.4s,v2.4s,v26.4s
+	zip1	v18.2d,v4.2d,v5.2d
 	add	w9,w9,w24
-	add	v6.4s,v6.4s,v26.4s
+	zip2	v22.2d,v4.2d,v5.2d
 	add	x10,x10,x24,lsr#32
-	add	v18.4s,v18.4s,v26.4s
+	zip1	v26.2d,v6.2d,v7.2d
 	add	w11,w11,w25
-	add	v3.4s,v3.4s,v27.4s
+	zip2	v30.2d,v6.2d,v7.2d
 	add	x12,x12,x25,lsr#32
+
+	zip1	v4.4s,v19.4s,v23.4s
 	add	w13,w13,w26
-	add	v7.4s,v7.4s,v28.4s
+	zip1	v5.4s,v27.4s,v31.4s
 	add	x14,x14,x26,lsr#32
+	zip2	v6.4s,v19.4s,v23.4s
 	add	w15,w15,w27
-	add	v19.4s,v19.4s,v29.4s
+	zip2	v7.4s,v27.4s,v31.4s
 	add	x16,x16,x27,lsr#32
+	zip1	v19.2d,v4.2d,v5.2d
 	add	w17,w17,w28
-	add	v1.4s,v1.4s,v25.4s
+	zip2	v23.2d,v4.2d,v5.2d
 	add	x19,x19,x28,lsr#32
+	zip1	v27.2d,v6.2d,v7.2d
 	add	w20,w20,w30
-	add	v5.4s,v5.4s,v25.4s
+	zip2	v31.2d,v6.2d,v7.2d
 	add	x21,x21,x30,lsr#32
-	add	v17.4s,v17.4s,v25.4s
 
 	b.lo	.Ltail_neon
 
 	add	x5,x5,x6,lsl#32	// pack
 	add	x7,x7,x8,lsl#32
 	ldp	x6,x8,[x1,#0]		// load input
+	add	v16.4s,v16.4s,v0.4s			// accumulate key block
 	add	x9,x9,x10,lsl#32
 	add	x11,x11,x12,lsl#32
 	ldp	x10,x12,[x1,#16]
+	add	v17.4s,v17.4s,v1.4s
 	add	x13,x13,x14,lsl#32
 	add	x15,x15,x16,lsl#32
 	ldp	x14,x16,[x1,#32]
+	add	v18.4s,v18.4s,v2.4s
 	add	x17,x17,x19,lsl#32
 	add	x20,x20,x21,lsl#32
 	ldp	x19,x21,[x1,#48]
+	add	v19.4s,v19.4s,v3.4s
 	add	x1,x1,#64
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -645,48 +679,68 @@ ChaCha20_neon:
 	rev	x17,x17
 	rev	x20,x20
 #endif
-	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
+	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
 	eor	x5,x5,x6
+	add	v20.4s,v20.4s,v0.4s
 	eor	x7,x7,x8
+	add	v21.4s,v21.4s,v1.4s
 	eor	x9,x9,x10
+	add	v22.4s,v22.4s,v2.4s
 	eor	x11,x11,x12
+	add	v23.4s,v23.4s,v3.4s
 	eor	x13,x13,x14
-	eor	v0.16b,v0.16b,v20.16b
+	eor	v16.16b,v16.16b,v4.16b
+	movi	v4.4s,#5
 	eor	x15,x15,x16
-	eor	v1.16b,v1.16b,v21.16b
+	eor	v17.16b,v17.16b,v5.16b
 	eor	x17,x17,x19
-	eor	v2.16b,v2.16b,v22.16b
+	eor	v18.16b,v18.16b,v6.16b
 	eor	x20,x20,x21
-	eor	v3.16b,v3.16b,v23.16b
-	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
+	eor	v19.16b,v19.16b,v7.16b
+	add	v8.4s,v8.4s,v4.4s			// += 5
+	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
 
 	stp	x5,x7,[x0,#0]		// store output
-	add	x28,x28,#4			// increment counter
+	add	x28,x28,#5			// increment counter
 	stp	x9,x11,[x0,#16]
-	add	v27.4s,v27.4s,v31.4s		// += 4
 	stp	x13,x15,[x0,#32]
-	add	v28.4s,v28.4s,v31.4s
 	stp	x17,x20,[x0,#48]
-	add	v29.4s,v29.4s,v31.4s
 	add	x0,x0,#64
 
-	st1	{v0.16b,v1.16b,v2.16b,v3.16b},[x0],#64
-	ld1	{v0.16b,v1.16b,v2.16b,v3.16b},[x1],#64
+	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[x0],#64
+	add	v24.4s,v24.4s,v0.4s
+	add	v25.4s,v25.4s,v1.4s
+	add	v26.4s,v26.4s,v2.4s
+	add	v27.4s,v27.4s,v3.4s
+	ld1	{v16.16b,v17.16b,v18.16b,v19.16b},[x1],#64
+
+	eor	v20.16b,v20.16b,v4.16b
+	eor	v21.16b,v21.16b,v5.16b
+	eor	v22.16b,v22.16b,v6.16b
+	eor	v23.16b,v23.16b,v7.16b
+	st1	{v20.16b,v21.16b,v22.16b,v23.16b},[x0],#64
+	add	v28.4s,v28.4s,v0.4s
+	add	v29.4s,v29.4s,v1.4s
+	add	v30.4s,v30.4s,v2.4s
+	add	v31.4s,v31.4s,v3.4s
+	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
 
-	eor	v4.16b,v4.16b,v20.16b
-	eor	v5.16b,v5.16b,v21.16b
-	eor	v6.16b,v6.16b,v22.16b
-	eor	v7.16b,v7.16b,v23.16b
-	st1	{v4.16b,v5.16b,v6.16b,v7.16b},[x0],#64
+	eor	v24.16b,v24.16b,v16.16b
+	eor	v25.16b,v25.16b,v17.16b
+	eor	v26.16b,v26.16b,v18.16b
+	eor	v27.16b,v27.16b,v19.16b
+	st1	{v24.16b,v25.16b,v26.16b,v27.16b},[x0],#64
 
-	eor	v16.16b,v16.16b,v0.16b
-	eor	v17.16b,v17.16b,v1.16b
-	eor	v18.16b,v18.16b,v2.16b
-	eor	v19.16b,v19.16b,v3.16b
-	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[x0],#64
+	eor	v28.16b,v28.16b,v20.16b
+	eor	v29.16b,v29.16b,v21.16b
+	eor	v30.16b,v30.16b,v22.16b
+	eor	v31.16b,v31.16b,v23.16b
+	st1	{v28.16b,v29.16b,v30.16b,v31.16b},[x0],#64
 
 	b.hi	.Loop_outer_neon
 
+	ldp	d8,d9,[sp]			// meet ABI requirements
+
 	ldp	x19,x20,[x29,#16]
 	add	sp,sp,#64
 	ldp	x21,x22,[x29,#32]
@@ -697,8 +751,10 @@ ChaCha20_neon:
 .inst	0xd50323bf			// autiasp
 	ret
 
+.align	4
 .Ltail_neon:
-	add	x2,x2,#256
+	add	x2,x2,#320
+	ldp	d8,d9,[sp]			// meet ABI requirements
 	cmp	x2,#64
 	b.lo	.Less_than_64
 
@@ -715,7 +771,7 @@ ChaCha20_neon:
 	add	x20,x20,x21,lsl#32
 	ldp	x19,x21,[x1,#48]
 	add	x1,x1,#64
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -735,48 +791,68 @@ ChaCha20_neon:
 	eor	x20,x20,x21
 
 	stp	x5,x7,[x0,#0]		// store output
-	add	x28,x28,#4			// increment counter
+	add	v16.4s,v16.4s,v0.4s			// accumulate key block
 	stp	x9,x11,[x0,#16]
+	add	v17.4s,v17.4s,v1.4s
 	stp	x13,x15,[x0,#32]
+	add	v18.4s,v18.4s,v2.4s
 	stp	x17,x20,[x0,#48]
+	add	v19.4s,v19.4s,v3.4s
 	add	x0,x0,#64
 	b.eq	.Ldone_neon
 	sub	x2,x2,#64
 	cmp	x2,#64
-	b.lo	.Less_than_128
+	b.lo	.Last_neon
 
-	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
-	eor	v0.16b,v0.16b,v20.16b
-	eor	v1.16b,v1.16b,v21.16b
-	eor	v2.16b,v2.16b,v22.16b
-	eor	v3.16b,v3.16b,v23.16b
-	st1	{v0.16b,v1.16b,v2.16b,v3.16b},[x0],#64
+	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
+	eor	v16.16b,v16.16b,v4.16b
+	eor	v17.16b,v17.16b,v5.16b
+	eor	v18.16b,v18.16b,v6.16b
+	eor	v19.16b,v19.16b,v7.16b
+	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[x0],#64
 	b.eq	.Ldone_neon
+
+	add	v16.4s,v20.4s,v0.4s
+	add	v17.4s,v21.4s,v1.4s
 	sub	x2,x2,#64
+	add	v18.4s,v22.4s,v2.4s
 	cmp	x2,#64
-	b.lo	.Less_than_192
+	add	v19.4s,v23.4s,v3.4s
+	b.lo	.Last_neon
 
-	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
-	eor	v4.16b,v4.16b,v20.16b
-	eor	v5.16b,v5.16b,v21.16b
-	eor	v6.16b,v6.16b,v22.16b
-	eor	v7.16b,v7.16b,v23.16b
-	st1	{v4.16b,v5.16b,v6.16b,v7.16b},[x0],#64
+	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
+	eor	v20.16b,v16.16b,v4.16b
+	eor	v21.16b,v17.16b,v5.16b
+	eor	v22.16b,v18.16b,v6.16b
+	eor	v23.16b,v19.16b,v7.16b
+	st1	{v20.16b,v21.16b,v22.16b,v23.16b},[x0],#64
 	b.eq	.Ldone_neon
+
+	add	v16.4s,v24.4s,v0.4s
+	add	v17.4s,v25.4s,v1.4s
 	sub	x2,x2,#64
+	add	v18.4s,v26.4s,v2.4s
+	cmp	x2,#64
+	add	v19.4s,v27.4s,v3.4s
+	b.lo	.Last_neon
 
-	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[sp]
-	b	.Last_neon
+	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
+	eor	v24.16b,v16.16b,v4.16b
+	eor	v25.16b,v17.16b,v5.16b
+	eor	v26.16b,v18.16b,v6.16b
+	eor	v27.16b,v19.16b,v7.16b
+	st1	{v24.16b,v25.16b,v26.16b,v27.16b},[x0],#64
+	b.eq	.Ldone_neon
 
-.Less_than_128:
-	st1	{v0.16b,v1.16b,v2.16b,v3.16b},[sp]
-	b	.Last_neon
-.Less_than_192:
-	st1	{v4.16b,v5.16b,v6.16b,v7.16b},[sp]
-	b	.Last_neon
+	add	v16.4s,v28.4s,v0.4s
+	add	v17.4s,v29.4s,v1.4s
+	add	v18.4s,v30.4s,v2.4s
+	add	v19.4s,v31.4s,v3.4s
+	sub	x2,x2,#64
 
-.align	4
 .Last_neon:
+	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[sp]
+
 	sub	x0,x0,#1
 	add	x1,x1,x2
 	add	x0,x0,x2
@@ -824,16 +900,18 @@ ChaCha20_512_neon:
 .L512_or_more_neon:
 	sub	sp,sp,#128+64
 
+	eor	v7.16b,v7.16b,v7.16b
 	ldp	x22,x23,[x5]		// load sigma
-	ld1	{v24.4s},[x5],#16
+	ld1	{v0.4s},[x5],#16
 	ldp	x24,x25,[x3]		// load key
 	ldp	x26,x27,[x3,#16]
-	ld1	{v25.4s,v26.4s},[x3]
+	ld1	{v1.4s,v2.4s},[x3]
 	ldp	x28,x30,[x4]		// load counter
-	ld1	{v27.4s},[x4]
-	ld1	{v31.4s},[x5]
-#ifdef	__ARMEB__
-	rev64	v24.4s,v24.4s
+	ld1	{v3.4s},[x4]
+	ld1	{v7.s}[0],[x5]
+	add	x3,x5,#16			// .Lrot24
+#ifdef	__AARCH64EB__
+	rev64	v0.4s,v0.4s
 	ror	x24,x24,#32
 	ror	x25,x25,#32
 	ror	x26,x26,#32
@@ -841,14 +919,14 @@ ChaCha20_512_neon:
 	ror	x28,x28,#32
 	ror	x30,x30,#32
 #endif
-	add	v27.4s,v27.4s,v31.4s		// += 1
-	stp	q24,q25,[sp,#0]		// off-load key block, invariant part
-	add	v27.4s,v27.4s,v31.4s		// not typo
-	str	q26,[sp,#32]
-	add	v28.4s,v27.4s,v31.4s
-	add	v29.4s,v28.4s,v31.4s
-	add	v30.4s,v29.4s,v31.4s
-	shl	v31.4s,v31.4s,#2			// 1 -> 4
+	add	v3.4s,v3.4s,v7.4s		// += 1
+	stp	q0,q1,[sp,#0]		// off-load key block, invariant part
+	add	v3.4s,v3.4s,v7.4s		// not typo
+	str	q2,[sp,#32]
+	add	v4.4s,v3.4s,v7.4s
+	add	v5.4s,v4.4s,v7.4s
+	add	v6.4s,v5.4s,v7.4s
+	shl	v7.4s,v7.4s,#2			// 1 -> 4
 
 	stp	d8,d9,[sp,#128+0]		// meet ABI requirements
 	stp	d10,d11,[sp,#128+16]
@@ -858,461 +936,450 @@ ChaCha20_512_neon:
 	sub	x2,x2,#512			// not typo
 
 .Loop_outer_512_neon:
-	mov	v0.16b,v24.16b
-	mov	v4.16b,v24.16b
-	mov	v8.16b,v24.16b
-	mov	v12.16b,v24.16b
-	mov	v16.16b,v24.16b
-	mov	v20.16b,v24.16b
-	mov	v1.16b,v25.16b
+	mov	v8.16b,v0.16b
+	mov	v12.16b,v0.16b
+	mov	v16.16b,v0.16b
+	mov	v20.16b,v0.16b
+	mov	v24.16b,v0.16b
+	mov	v28.16b,v0.16b
+	mov	v9.16b,v1.16b
 	mov	w5,w22			// unpack key block
-	mov	v5.16b,v25.16b
+	mov	v13.16b,v1.16b
 	lsr	x6,x22,#32
-	mov	v9.16b,v25.16b
+	mov	v17.16b,v1.16b
 	mov	w7,w23
-	mov	v13.16b,v25.16b
+	mov	v21.16b,v1.16b
 	lsr	x8,x23,#32
-	mov	v17.16b,v25.16b
+	mov	v25.16b,v1.16b
 	mov	w9,w24
-	mov	v21.16b,v25.16b
+	mov	v29.16b,v1.16b
 	lsr	x10,x24,#32
-	mov	v3.16b,v27.16b
+	mov	v11.16b,v3.16b
 	mov	w11,w25
-	mov	v7.16b,v28.16b
+	mov	v15.16b,v4.16b
 	lsr	x12,x25,#32
-	mov	v11.16b,v29.16b
+	mov	v19.16b,v5.16b
 	mov	w13,w26
-	mov	v15.16b,v30.16b
+	mov	v23.16b,v6.16b
 	lsr	x14,x26,#32
-	mov	v2.16b,v26.16b
+	mov	v10.16b,v2.16b
 	mov	w15,w27
-	mov	v6.16b,v26.16b
+	mov	v14.16b,v2.16b
 	lsr	x16,x27,#32
-	add	v19.4s,v3.4s,v31.4s			// +4
+	add	v27.4s,v11.4s,v7.4s			// +4
 	mov	w17,w28
-	add	v23.4s,v7.4s,v31.4s			// +4
+	add	v31.4s,v15.4s,v7.4s			// +4
 	lsr	x19,x28,#32
-	mov	v10.16b,v26.16b
+	mov	v18.16b,v2.16b
 	mov	w20,w30
-	mov	v14.16b,v26.16b
+	mov	v22.16b,v2.16b
 	lsr	x21,x30,#32
-	mov	v18.16b,v26.16b
-	stp	q27,q28,[sp,#48]		// off-load key block, variable part
-	mov	v22.16b,v26.16b
-	str	q29,[sp,#80]
+	mov	v26.16b,v2.16b
+	stp	q3,q4,[sp,#48]		// off-load key block, variable part
+	mov	v30.16b,v2.16b
+	stp	q5,q6,[sp,#80]
 
 	mov	x4,#5
+	ld1	{v6.4s},[x3]
 	subs	x2,x2,#512
 .Loop_upper_neon:
 	sub	x4,x4,#1
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	add	w5,w5,w9
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	add	w6,w6,w10
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	add	w7,w7,w11
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	add	w8,w8,w12
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	eor	w17,w17,w5
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	eor	w19,w19,w6
-	eor	v3.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	eor	w20,w20,w7
-	eor	v7.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	eor	w21,w21,w8
-	eor	v11.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	ror	w17,w17,#16
-	eor	v15.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	ror	w19,w19,#16
-	eor	v19.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	ror	w20,w20,#16
-	eor	v23.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	ror	w21,w21,#16
-	rev32	v3.8h,v3.8h
+	rev32	v11.8h,v11.8h
 	add	w13,w13,w17
-	rev32	v7.8h,v7.8h
+	rev32	v15.8h,v15.8h
 	add	w14,w14,w19
-	rev32	v11.8h,v11.8h
+	rev32	v19.8h,v19.8h
 	add	w15,w15,w20
-	rev32	v15.8h,v15.8h
+	rev32	v23.8h,v23.8h
 	add	w16,w16,w21
-	rev32	v19.8h,v19.8h
+	rev32	v27.8h,v27.8h
 	eor	w9,w9,w13
-	rev32	v23.8h,v23.8h
+	rev32	v31.8h,v31.8h
 	eor	w10,w10,w14
-	add	v2.4s,v2.4s,v3.4s
+	add	v10.4s,v10.4s,v11.4s
 	eor	w11,w11,w15
-	add	v6.4s,v6.4s,v7.4s
+	add	v14.4s,v14.4s,v15.4s
 	eor	w12,w12,w16
-	add	v10.4s,v10.4s,v11.4s
+	add	v18.4s,v18.4s,v19.4s
 	ror	w9,w9,#20
-	add	v14.4s,v14.4s,v15.4s
+	add	v22.4s,v22.4s,v23.4s
 	ror	w10,w10,#20
-	add	v18.4s,v18.4s,v19.4s
+	add	v26.4s,v26.4s,v27.4s
 	ror	w11,w11,#20
-	add	v22.4s,v22.4s,v23.4s
+	add	v30.4s,v30.4s,v31.4s
 	ror	w12,w12,#20
-	eor	v24.16b,v1.16b,v2.16b
+	eor	v0.16b,v9.16b,v10.16b
 	add	w5,w5,w9
-	eor	v25.16b,v5.16b,v6.16b
+	eor	v1.16b,v13.16b,v14.16b
 	add	w6,w6,w10
-	eor	v26.16b,v9.16b,v10.16b
+	eor	v2.16b,v17.16b,v18.16b
 	add	w7,w7,w11
-	eor	v27.16b,v13.16b,v14.16b
+	eor	v3.16b,v21.16b,v22.16b
 	add	w8,w8,w12
-	eor	v28.16b,v17.16b,v18.16b
+	eor	v4.16b,v25.16b,v26.16b
 	eor	w17,w17,w5
-	eor	v29.16b,v21.16b,v22.16b
+	eor	v5.16b,v29.16b,v30.16b
 	eor	w19,w19,w6
-	ushr	v1.4s,v24.4s,#20
+	ushr	v9.4s,v0.4s,#20
 	eor	w20,w20,w7
-	ushr	v5.4s,v25.4s,#20
+	ushr	v13.4s,v1.4s,#20
 	eor	w21,w21,w8
-	ushr	v9.4s,v26.4s,#20
+	ushr	v17.4s,v2.4s,#20
 	ror	w17,w17,#24
-	ushr	v13.4s,v27.4s,#20
+	ushr	v21.4s,v3.4s,#20
 	ror	w19,w19,#24
-	ushr	v17.4s,v28.4s,#20
+	ushr	v25.4s,v4.4s,#20
 	ror	w20,w20,#24
-	ushr	v21.4s,v29.4s,#20
+	ushr	v29.4s,v5.4s,#20
 	ror	w21,w21,#24
-	sli	v1.4s,v24.4s,#12
+	sli	v9.4s,v0.4s,#12
 	add	w13,w13,w17
-	sli	v5.4s,v25.4s,#12
+	sli	v13.4s,v1.4s,#12
 	add	w14,w14,w19
-	sli	v9.4s,v26.4s,#12
+	sli	v17.4s,v2.4s,#12
 	add	w15,w15,w20
-	sli	v13.4s,v27.4s,#12
+	sli	v21.4s,v3.4s,#12
 	add	w16,w16,w21
-	sli	v17.4s,v28.4s,#12
+	sli	v25.4s,v4.4s,#12
 	eor	w9,w9,w13
-	sli	v21.4s,v29.4s,#12
+	sli	v29.4s,v5.4s,#12
 	eor	w10,w10,w14
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	eor	w11,w11,w15
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	eor	w12,w12,w16
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	ror	w9,w9,#25
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	ror	w10,w10,#25
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	ror	w11,w11,#25
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	ror	w12,w12,#25
-	eor	v24.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	add	w5,w5,w10
-	eor	v25.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	add	w6,w6,w11
-	eor	v26.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	add	w7,w7,w12
-	eor	v27.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	add	w8,w8,w9
-	eor	v28.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	eor	w21,w21,w5
-	eor	v29.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	eor	w17,w17,w6
-	ushr	v3.4s,v24.4s,#24
+	tbl	v11.16b,{v11.16b},v6.16b
 	eor	w19,w19,w7
-	ushr	v7.4s,v25.4s,#24
+	tbl	v15.16b,{v15.16b},v6.16b
 	eor	w20,w20,w8
-	ushr	v11.4s,v26.4s,#24
+	tbl	v19.16b,{v19.16b},v6.16b
 	ror	w21,w21,#16
-	ushr	v15.4s,v27.4s,#24
+	tbl	v23.16b,{v23.16b},v6.16b
 	ror	w17,w17,#16
-	ushr	v19.4s,v28.4s,#24
+	tbl	v27.16b,{v27.16b},v6.16b
 	ror	w19,w19,#16
-	ushr	v23.4s,v29.4s,#24
+	tbl	v31.16b,{v31.16b},v6.16b
 	ror	w20,w20,#16
-	sli	v3.4s,v24.4s,#8
+	add	v10.4s,v10.4s,v11.4s
 	add	w15,w15,w21
-	sli	v7.4s,v25.4s,#8
+	add	v14.4s,v14.4s,v15.4s
 	add	w16,w16,w17
-	sli	v11.4s,v26.4s,#8
+	add	v18.4s,v18.4s,v19.4s
 	add	w13,w13,w19
-	sli	v15.4s,v27.4s,#8
+	add	v22.4s,v22.4s,v23.4s
 	add	w14,w14,w20
-	sli	v19.4s,v28.4s,#8
+	add	v26.4s,v26.4s,v27.4s
 	eor	w10,w10,w15
-	sli	v23.4s,v29.4s,#8
+	add	v30.4s,v30.4s,v31.4s
 	eor	w11,w11,w16
-	add	v2.4s,v2.4s,v3.4s
+	eor	v0.16b,v9.16b,v10.16b
 	eor	w12,w12,w13
-	add	v6.4s,v6.4s,v7.4s
+	eor	v1.16b,v13.16b,v14.16b
 	eor	w9,w9,w14
-	add	v10.4s,v10.4s,v11.4s
+	eor	v2.16b,v17.16b,v18.16b
 	ror	w10,w10,#20
-	add	v14.4s,v14.4s,v15.4s
+	eor	v3.16b,v21.16b,v22.16b
 	ror	w11,w11,#20
-	add	v18.4s,v18.4s,v19.4s
+	eor	v4.16b,v25.16b,v26.16b
 	ror	w12,w12,#20
-	add	v22.4s,v22.4s,v23.4s
+	eor	v5.16b,v29.16b,v30.16b
 	ror	w9,w9,#20
-	eor	v24.16b,v1.16b,v2.16b
+	ushr	v9.4s,v0.4s,#25
 	add	w5,w5,w10
-	eor	v25.16b,v5.16b,v6.16b
+	ushr	v13.4s,v1.4s,#25
 	add	w6,w6,w11
-	eor	v26.16b,v9.16b,v10.16b
+	ushr	v17.4s,v2.4s,#25
 	add	w7,w7,w12
-	eor	v27.16b,v13.16b,v14.16b
+	ushr	v21.4s,v3.4s,#25
 	add	w8,w8,w9
-	eor	v28.16b,v17.16b,v18.16b
+	ushr	v25.4s,v4.4s,#25
 	eor	w21,w21,w5
-	eor	v29.16b,v21.16b,v22.16b
+	ushr	v29.4s,v5.4s,#25
 	eor	w17,w17,w6
-	ushr	v1.4s,v24.4s,#25
+	sli	v9.4s,v0.4s,#7
 	eor	w19,w19,w7
-	ushr	v5.4s,v25.4s,#25
+	sli	v13.4s,v1.4s,#7
 	eor	w20,w20,w8
-	ushr	v9.4s,v26.4s,#25
+	sli	v17.4s,v2.4s,#7
 	ror	w21,w21,#24
-	ushr	v13.4s,v27.4s,#25
+	sli	v21.4s,v3.4s,#7
 	ror	w17,w17,#24
-	ushr	v17.4s,v28.4s,#25
+	sli	v25.4s,v4.4s,#7
 	ror	w19,w19,#24
-	ushr	v21.4s,v29.4s,#25
+	sli	v29.4s,v5.4s,#7
 	ror	w20,w20,#24
-	sli	v1.4s,v24.4s,#7
+	ext	v10.16b,v10.16b,v10.16b,#8
 	add	w15,w15,w21
-	sli	v5.4s,v25.4s,#7
+	ext	v14.16b,v14.16b,v14.16b,#8
 	add	w16,w16,w17
-	sli	v9.4s,v26.4s,#7
+	ext	v18.16b,v18.16b,v18.16b,#8
 	add	w13,w13,w19
-	sli	v13.4s,v27.4s,#7
+	ext	v22.16b,v22.16b,v22.16b,#8
 	add	w14,w14,w20
-	sli	v17.4s,v28.4s,#7
+	ext	v26.16b,v26.16b,v26.16b,#8
 	eor	w10,w10,w15
-	sli	v21.4s,v29.4s,#7
+	ext	v30.16b,v30.16b,v30.16b,#8
 	eor	w11,w11,w16
-	ext	v2.16b,v2.16b,v2.16b,#8
+	ext	v11.16b,v11.16b,v11.16b,#12
 	eor	w12,w12,w13
-	ext	v6.16b,v6.16b,v6.16b,#8
+	ext	v15.16b,v15.16b,v15.16b,#12
 	eor	w9,w9,w14
-	ext	v10.16b,v10.16b,v10.16b,#8
+	ext	v19.16b,v19.16b,v19.16b,#12
 	ror	w10,w10,#25
-	ext	v14.16b,v14.16b,v14.16b,#8
+	ext	v23.16b,v23.16b,v23.16b,#12
 	ror	w11,w11,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	ext	v27.16b,v27.16b,v27.16b,#12
 	ror	w12,w12,#25
-	ext	v22.16b,v22.16b,v22.16b,#8
+	ext	v31.16b,v31.16b,v31.16b,#12
 	ror	w9,w9,#25
-	ext	v3.16b,v3.16b,v3.16b,#12
-	ext	v7.16b,v7.16b,v7.16b,#12
-	ext	v11.16b,v11.16b,v11.16b,#12
-	ext	v15.16b,v15.16b,v15.16b,#12
-	ext	v19.16b,v19.16b,v19.16b,#12
-	ext	v23.16b,v23.16b,v23.16b,#12
-	ext	v1.16b,v1.16b,v1.16b,#4
-	ext	v5.16b,v5.16b,v5.16b,#4
 	ext	v9.16b,v9.16b,v9.16b,#4
 	ext	v13.16b,v13.16b,v13.16b,#4
 	ext	v17.16b,v17.16b,v17.16b,#4
 	ext	v21.16b,v21.16b,v21.16b,#4
-	add	v0.4s,v0.4s,v1.4s
+	ext	v25.16b,v25.16b,v25.16b,#4
+	ext	v29.16b,v29.16b,v29.16b,#4
+	add	v8.4s,v8.4s,v9.4s
 	add	w5,w5,w9
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	add	w6,w6,w10
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	add	w7,w7,w11
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	add	w8,w8,w12
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	eor	w17,w17,w5
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	eor	w19,w19,w6
-	eor	v3.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	eor	w20,w20,w7
-	eor	v7.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	eor	w21,w21,w8
-	eor	v11.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	ror	w17,w17,#16
-	eor	v15.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	ror	w19,w19,#16
-	eor	v19.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	ror	w20,w20,#16
-	eor	v23.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	ror	w21,w21,#16
-	rev32	v3.8h,v3.8h
+	rev32	v11.8h,v11.8h
 	add	w13,w13,w17
-	rev32	v7.8h,v7.8h
+	rev32	v15.8h,v15.8h
 	add	w14,w14,w19
-	rev32	v11.8h,v11.8h
+	rev32	v19.8h,v19.8h
 	add	w15,w15,w20
-	rev32	v15.8h,v15.8h
+	rev32	v23.8h,v23.8h
 	add	w16,w16,w21
-	rev32	v19.8h,v19.8h
+	rev32	v27.8h,v27.8h
 	eor	w9,w9,w13
-	rev32	v23.8h,v23.8h
+	rev32	v31.8h,v31.8h
 	eor	w10,w10,w14
-	add	v2.4s,v2.4s,v3.4s
+	add	v10.4s,v10.4s,v11.4s
 	eor	w11,w11,w15
-	add	v6.4s,v6.4s,v7.4s
+	add	v14.4s,v14.4s,v15.4s
 	eor	w12,w12,w16
-	add	v10.4s,v10.4s,v11.4s
+	add	v18.4s,v18.4s,v19.4s
 	ror	w9,w9,#20
-	add	v14.4s,v14.4s,v15.4s
+	add	v22.4s,v22.4s,v23.4s
 	ror	w10,w10,#20
-	add	v18.4s,v18.4s,v19.4s
+	add	v26.4s,v26.4s,v27.4s
 	ror	w11,w11,#20
-	add	v22.4s,v22.4s,v23.4s
+	add	v30.4s,v30.4s,v31.4s
 	ror	w12,w12,#20
-	eor	v24.16b,v1.16b,v2.16b
+	eor	v0.16b,v9.16b,v10.16b
 	add	w5,w5,w9
-	eor	v25.16b,v5.16b,v6.16b
+	eor	v1.16b,v13.16b,v14.16b
 	add	w6,w6,w10
-	eor	v26.16b,v9.16b,v10.16b
+	eor	v2.16b,v17.16b,v18.16b
 	add	w7,w7,w11
-	eor	v27.16b,v13.16b,v14.16b
+	eor	v3.16b,v21.16b,v22.16b
 	add	w8,w8,w12
-	eor	v28.16b,v17.16b,v18.16b
+	eor	v4.16b,v25.16b,v26.16b
 	eor	w17,w17,w5
-	eor	v29.16b,v21.16b,v22.16b
+	eor	v5.16b,v29.16b,v30.16b
 	eor	w19,w19,w6
-	ushr	v1.4s,v24.4s,#20
+	ushr	v9.4s,v0.4s,#20
 	eor	w20,w20,w7
-	ushr	v5.4s,v25.4s,#20
+	ushr	v13.4s,v1.4s,#20
 	eor	w21,w21,w8
-	ushr	v9.4s,v26.4s,#20
+	ushr	v17.4s,v2.4s,#20
 	ror	w17,w17,#24
-	ushr	v13.4s,v27.4s,#20
+	ushr	v21.4s,v3.4s,#20
 	ror	w19,w19,#24
-	ushr	v17.4s,v28.4s,#20
+	ushr	v25.4s,v4.4s,#20
 	ror	w20,w20,#24
-	ushr	v21.4s,v29.4s,#20
+	ushr	v29.4s,v5.4s,#20
 	ror	w21,w21,#24
-	sli	v1.4s,v24.4s,#12
+	sli	v9.4s,v0.4s,#12
 	add	w13,w13,w17
-	sli	v5.4s,v25.4s,#12
+	sli	v13.4s,v1.4s,#12
 	add	w14,w14,w19
-	sli	v9.4s,v26.4s,#12
+	sli	v17.4s,v2.4s,#12
 	add	w15,w15,w20
-	sli	v13.4s,v27.4s,#12
+	sli	v21.4s,v3.4s,#12
 	add	w16,w16,w21
-	sli	v17.4s,v28.4s,#12
+	sli	v25.4s,v4.4s,#12
 	eor	w9,w9,w13
-	sli	v21.4s,v29.4s,#12
+	sli	v29.4s,v5.4s,#12
 	eor	w10,w10,w14
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	eor	w11,w11,w15
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	eor	w12,w12,w16
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	ror	w9,w9,#25
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	ror	w10,w10,#25
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	ror	w11,w11,#25
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	ror	w12,w12,#25
-	eor	v24.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	add	w5,w5,w10
-	eor	v25.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	add	w6,w6,w11
-	eor	v26.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	add	w7,w7,w12
-	eor	v27.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	add	w8,w8,w9
-	eor	v28.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	eor	w21,w21,w5
-	eor	v29.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	eor	w17,w17,w6
-	ushr	v3.4s,v24.4s,#24
+	tbl	v11.16b,{v11.16b},v6.16b
 	eor	w19,w19,w7
-	ushr	v7.4s,v25.4s,#24
+	tbl	v15.16b,{v15.16b},v6.16b
 	eor	w20,w20,w8
-	ushr	v11.4s,v26.4s,#24
+	tbl	v19.16b,{v19.16b},v6.16b
 	ror	w21,w21,#16
-	ushr	v15.4s,v27.4s,#24
+	tbl	v23.16b,{v23.16b},v6.16b
 	ror	w17,w17,#16
-	ushr	v19.4s,v28.4s,#24
+	tbl	v27.16b,{v27.16b},v6.16b
 	ror	w19,w19,#16
-	ushr	v23.4s,v29.4s,#24
+	tbl	v31.16b,{v31.16b},v6.16b
 	ror	w20,w20,#16
-	sli	v3.4s,v24.4s,#8
+	add	v10.4s,v10.4s,v11.4s
 	add	w15,w15,w21
-	sli	v7.4s,v25.4s,#8
+	add	v14.4s,v14.4s,v15.4s
 	add	w16,w16,w17
-	sli	v11.4s,v26.4s,#8
+	add	v18.4s,v18.4s,v19.4s
 	add	w13,w13,w19
-	sli	v15.4s,v27.4s,#8
+	add	v22.4s,v22.4s,v23.4s
 	add	w14,w14,w20
-	sli	v19.4s,v28.4s,#8
+	add	v26.4s,v26.4s,v27.4s
 	eor	w10,w10,w15
-	sli	v23.4s,v29.4s,#8
+	add	v30.4s,v30.4s,v31.4s
 	eor	w11,w11,w16
-	add	v2.4s,v2.4s,v3.4s
+	eor	v0.16b,v9.16b,v10.16b
 	eor	w12,w12,w13
-	add	v6.4s,v6.4s,v7.4s
+	eor	v1.16b,v13.16b,v14.16b
 	eor	w9,w9,w14
-	add	v10.4s,v10.4s,v11.4s
+	eor	v2.16b,v17.16b,v18.16b
 	ror	w10,w10,#20
-	add	v14.4s,v14.4s,v15.4s
+	eor	v3.16b,v21.16b,v22.16b
 	ror	w11,w11,#20
-	add	v18.4s,v18.4s,v19.4s
+	eor	v4.16b,v25.16b,v26.16b
 	ror	w12,w12,#20
-	add	v22.4s,v22.4s,v23.4s
+	eor	v5.16b,v29.16b,v30.16b
 	ror	w9,w9,#20
-	eor	v24.16b,v1.16b,v2.16b
+	ushr	v9.4s,v0.4s,#25
 	add	w5,w5,w10
-	eor	v25.16b,v5.16b,v6.16b
+	ushr	v13.4s,v1.4s,#25
 	add	w6,w6,w11
-	eor	v26.16b,v9.16b,v10.16b
+	ushr	v17.4s,v2.4s,#25
 	add	w7,w7,w12
-	eor	v27.16b,v13.16b,v14.16b
+	ushr	v21.4s,v3.4s,#25
 	add	w8,w8,w9
-	eor	v28.16b,v17.16b,v18.16b
+	ushr	v25.4s,v4.4s,#25
 	eor	w21,w21,w5
-	eor	v29.16b,v21.16b,v22.16b
+	ushr	v29.4s,v5.4s,#25
 	eor	w17,w17,w6
-	ushr	v1.4s,v24.4s,#25
+	sli	v9.4s,v0.4s,#7
 	eor	w19,w19,w7
-	ushr	v5.4s,v25.4s,#25
+	sli	v13.4s,v1.4s,#7
 	eor	w20,w20,w8
-	ushr	v9.4s,v26.4s,#25
+	sli	v17.4s,v2.4s,#7
 	ror	w21,w21,#24
-	ushr	v13.4s,v27.4s,#25
+	sli	v21.4s,v3.4s,#7
 	ror	w17,w17,#24
-	ushr	v17.4s,v28.4s,#25
+	sli	v25.4s,v4.4s,#7
 	ror	w19,w19,#24
-	ushr	v21.4s,v29.4s,#25
+	sli	v29.4s,v5.4s,#7
 	ror	w20,w20,#24
-	sli	v1.4s,v24.4s,#7
+	ext	v10.16b,v10.16b,v10.16b,#8
 	add	w15,w15,w21
-	sli	v5.4s,v25.4s,#7
+	ext	v14.16b,v14.16b,v14.16b,#8
 	add	w16,w16,w17
-	sli	v9.4s,v26.4s,#7
+	ext	v18.16b,v18.16b,v18.16b,#8
 	add	w13,w13,w19
-	sli	v13.4s,v27.4s,#7
+	ext	v22.16b,v22.16b,v22.16b,#8
 	add	w14,w14,w20
-	sli	v17.4s,v28.4s,#7
+	ext	v26.16b,v26.16b,v26.16b,#8
 	eor	w10,w10,w15
-	sli	v21.4s,v29.4s,#7
+	ext	v30.16b,v30.16b,v30.16b,#8
 	eor	w11,w11,w16
-	ext	v2.16b,v2.16b,v2.16b,#8
+	ext	v11.16b,v11.16b,v11.16b,#4
 	eor	w12,w12,w13
-	ext	v6.16b,v6.16b,v6.16b,#8
+	ext	v15.16b,v15.16b,v15.16b,#4
 	eor	w9,w9,w14
-	ext	v10.16b,v10.16b,v10.16b,#8
+	ext	v19.16b,v19.16b,v19.16b,#4
 	ror	w10,w10,#25
-	ext	v14.16b,v14.16b,v14.16b,#8
+	ext	v23.16b,v23.16b,v23.16b,#4
 	ror	w11,w11,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	ext	v27.16b,v27.16b,v27.16b,#4
 	ror	w12,w12,#25
-	ext	v22.16b,v22.16b,v22.16b,#8
+	ext	v31.16b,v31.16b,v31.16b,#4
 	ror	w9,w9,#25
-	ext	v3.16b,v3.16b,v3.16b,#4
-	ext	v7.16b,v7.16b,v7.16b,#4
-	ext	v11.16b,v11.16b,v11.16b,#4
-	ext	v15.16b,v15.16b,v15.16b,#4
-	ext	v19.16b,v19.16b,v19.16b,#4
-	ext	v23.16b,v23.16b,v23.16b,#4
-	ext	v1.16b,v1.16b,v1.16b,#12
-	ext	v5.16b,v5.16b,v5.16b,#12
 	ext	v9.16b,v9.16b,v9.16b,#12
 	ext	v13.16b,v13.16b,v13.16b,#12
 	ext	v17.16b,v17.16b,v17.16b,#12
 	ext	v21.16b,v21.16b,v21.16b,#12
+	ext	v25.16b,v25.16b,v25.16b,#12
+	ext	v29.16b,v29.16b,v29.16b,#12
 	cbnz	x4,.Loop_upper_neon
 
 	add	w5,w5,w22		// accumulate key block
@@ -1345,7 +1412,7 @@ ChaCha20_512_neon:
 	add	x20,x20,x21,lsl#32
 	ldp	x19,x21,[x1,#48]
 	add	x1,x1,#64
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -1390,476 +1457,465 @@ ChaCha20_512_neon:
 	mov	x4,#5
 .Loop_lower_neon:
 	sub	x4,x4,#1
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	add	w5,w5,w9
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	add	w6,w6,w10
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	add	w7,w7,w11
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	add	w8,w8,w12
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	eor	w17,w17,w5
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	eor	w19,w19,w6
-	eor	v3.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	eor	w20,w20,w7
-	eor	v7.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	eor	w21,w21,w8
-	eor	v11.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	ror	w17,w17,#16
-	eor	v15.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	ror	w19,w19,#16
-	eor	v19.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	ror	w20,w20,#16
-	eor	v23.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	ror	w21,w21,#16
-	rev32	v3.8h,v3.8h
+	rev32	v11.8h,v11.8h
 	add	w13,w13,w17
-	rev32	v7.8h,v7.8h
+	rev32	v15.8h,v15.8h
 	add	w14,w14,w19
-	rev32	v11.8h,v11.8h
+	rev32	v19.8h,v19.8h
 	add	w15,w15,w20
-	rev32	v15.8h,v15.8h
+	rev32	v23.8h,v23.8h
 	add	w16,w16,w21
-	rev32	v19.8h,v19.8h
+	rev32	v27.8h,v27.8h
 	eor	w9,w9,w13
-	rev32	v23.8h,v23.8h
+	rev32	v31.8h,v31.8h
 	eor	w10,w10,w14
-	add	v2.4s,v2.4s,v3.4s
+	add	v10.4s,v10.4s,v11.4s
 	eor	w11,w11,w15
-	add	v6.4s,v6.4s,v7.4s
+	add	v14.4s,v14.4s,v15.4s
 	eor	w12,w12,w16
-	add	v10.4s,v10.4s,v11.4s
+	add	v18.4s,v18.4s,v19.4s
 	ror	w9,w9,#20
-	add	v14.4s,v14.4s,v15.4s
+	add	v22.4s,v22.4s,v23.4s
 	ror	w10,w10,#20
-	add	v18.4s,v18.4s,v19.4s
+	add	v26.4s,v26.4s,v27.4s
 	ror	w11,w11,#20
-	add	v22.4s,v22.4s,v23.4s
+	add	v30.4s,v30.4s,v31.4s
 	ror	w12,w12,#20
-	eor	v24.16b,v1.16b,v2.16b
+	eor	v0.16b,v9.16b,v10.16b
 	add	w5,w5,w9
-	eor	v25.16b,v5.16b,v6.16b
+	eor	v1.16b,v13.16b,v14.16b
 	add	w6,w6,w10
-	eor	v26.16b,v9.16b,v10.16b
+	eor	v2.16b,v17.16b,v18.16b
 	add	w7,w7,w11
-	eor	v27.16b,v13.16b,v14.16b
+	eor	v3.16b,v21.16b,v22.16b
 	add	w8,w8,w12
-	eor	v28.16b,v17.16b,v18.16b
+	eor	v4.16b,v25.16b,v26.16b
 	eor	w17,w17,w5
-	eor	v29.16b,v21.16b,v22.16b
+	eor	v5.16b,v29.16b,v30.16b
 	eor	w19,w19,w6
-	ushr	v1.4s,v24.4s,#20
+	ushr	v9.4s,v0.4s,#20
 	eor	w20,w20,w7
-	ushr	v5.4s,v25.4s,#20
+	ushr	v13.4s,v1.4s,#20
 	eor	w21,w21,w8
-	ushr	v9.4s,v26.4s,#20
+	ushr	v17.4s,v2.4s,#20
 	ror	w17,w17,#24
-	ushr	v13.4s,v27.4s,#20
+	ushr	v21.4s,v3.4s,#20
 	ror	w19,w19,#24
-	ushr	v17.4s,v28.4s,#20
+	ushr	v25.4s,v4.4s,#20
 	ror	w20,w20,#24
-	ushr	v21.4s,v29.4s,#20
+	ushr	v29.4s,v5.4s,#20
 	ror	w21,w21,#24
-	sli	v1.4s,v24.4s,#12
+	sli	v9.4s,v0.4s,#12
 	add	w13,w13,w17
-	sli	v5.4s,v25.4s,#12
+	sli	v13.4s,v1.4s,#12
 	add	w14,w14,w19
-	sli	v9.4s,v26.4s,#12
+	sli	v17.4s,v2.4s,#12
 	add	w15,w15,w20
-	sli	v13.4s,v27.4s,#12
+	sli	v21.4s,v3.4s,#12
 	add	w16,w16,w21
-	sli	v17.4s,v28.4s,#12
+	sli	v25.4s,v4.4s,#12
 	eor	w9,w9,w13
-	sli	v21.4s,v29.4s,#12
+	sli	v29.4s,v5.4s,#12
 	eor	w10,w10,w14
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	eor	w11,w11,w15
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	eor	w12,w12,w16
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	ror	w9,w9,#25
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	ror	w10,w10,#25
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	ror	w11,w11,#25
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	ror	w12,w12,#25
-	eor	v24.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	add	w5,w5,w10
-	eor	v25.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	add	w6,w6,w11
-	eor	v26.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	add	w7,w7,w12
-	eor	v27.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	add	w8,w8,w9
-	eor	v28.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	eor	w21,w21,w5
-	eor	v29.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	eor	w17,w17,w6
-	ushr	v3.4s,v24.4s,#24
+	tbl	v11.16b,{v11.16b},v6.16b
 	eor	w19,w19,w7
-	ushr	v7.4s,v25.4s,#24
+	tbl	v15.16b,{v15.16b},v6.16b
 	eor	w20,w20,w8
-	ushr	v11.4s,v26.4s,#24
+	tbl	v19.16b,{v19.16b},v6.16b
 	ror	w21,w21,#16
-	ushr	v15.4s,v27.4s,#24
+	tbl	v23.16b,{v23.16b},v6.16b
 	ror	w17,w17,#16
-	ushr	v19.4s,v28.4s,#24
+	tbl	v27.16b,{v27.16b},v6.16b
 	ror	w19,w19,#16
-	ushr	v23.4s,v29.4s,#24
+	tbl	v31.16b,{v31.16b},v6.16b
 	ror	w20,w20,#16
-	sli	v3.4s,v24.4s,#8
+	add	v10.4s,v10.4s,v11.4s
 	add	w15,w15,w21
-	sli	v7.4s,v25.4s,#8
+	add	v14.4s,v14.4s,v15.4s
 	add	w16,w16,w17
-	sli	v11.4s,v26.4s,#8
+	add	v18.4s,v18.4s,v19.4s
 	add	w13,w13,w19
-	sli	v15.4s,v27.4s,#8
+	add	v22.4s,v22.4s,v23.4s
 	add	w14,w14,w20
-	sli	v19.4s,v28.4s,#8
+	add	v26.4s,v26.4s,v27.4s
 	eor	w10,w10,w15
-	sli	v23.4s,v29.4s,#8
+	add	v30.4s,v30.4s,v31.4s
 	eor	w11,w11,w16
-	add	v2.4s,v2.4s,v3.4s
+	eor	v0.16b,v9.16b,v10.16b
 	eor	w12,w12,w13
-	add	v6.4s,v6.4s,v7.4s
+	eor	v1.16b,v13.16b,v14.16b
 	eor	w9,w9,w14
-	add	v10.4s,v10.4s,v11.4s
+	eor	v2.16b,v17.16b,v18.16b
 	ror	w10,w10,#20
-	add	v14.4s,v14.4s,v15.4s
+	eor	v3.16b,v21.16b,v22.16b
 	ror	w11,w11,#20
-	add	v18.4s,v18.4s,v19.4s
+	eor	v4.16b,v25.16b,v26.16b
 	ror	w12,w12,#20
-	add	v22.4s,v22.4s,v23.4s
+	eor	v5.16b,v29.16b,v30.16b
 	ror	w9,w9,#20
-	eor	v24.16b,v1.16b,v2.16b
+	ushr	v9.4s,v0.4s,#25
 	add	w5,w5,w10
-	eor	v25.16b,v5.16b,v6.16b
+	ushr	v13.4s,v1.4s,#25
 	add	w6,w6,w11
-	eor	v26.16b,v9.16b,v10.16b
+	ushr	v17.4s,v2.4s,#25
 	add	w7,w7,w12
-	eor	v27.16b,v13.16b,v14.16b
+	ushr	v21.4s,v3.4s,#25
 	add	w8,w8,w9
-	eor	v28.16b,v17.16b,v18.16b
+	ushr	v25.4s,v4.4s,#25
 	eor	w21,w21,w5
-	eor	v29.16b,v21.16b,v22.16b
+	ushr	v29.4s,v5.4s,#25
 	eor	w17,w17,w6
-	ushr	v1.4s,v24.4s,#25
+	sli	v9.4s,v0.4s,#7
 	eor	w19,w19,w7
-	ushr	v5.4s,v25.4s,#25
+	sli	v13.4s,v1.4s,#7
 	eor	w20,w20,w8
-	ushr	v9.4s,v26.4s,#25
+	sli	v17.4s,v2.4s,#7
 	ror	w21,w21,#24
-	ushr	v13.4s,v27.4s,#25
+	sli	v21.4s,v3.4s,#7
 	ror	w17,w17,#24
-	ushr	v17.4s,v28.4s,#25
+	sli	v25.4s,v4.4s,#7
 	ror	w19,w19,#24
-	ushr	v21.4s,v29.4s,#25
+	sli	v29.4s,v5.4s,#7
 	ror	w20,w20,#24
-	sli	v1.4s,v24.4s,#7
+	ext	v10.16b,v10.16b,v10.16b,#8
 	add	w15,w15,w21
-	sli	v5.4s,v25.4s,#7
+	ext	v14.16b,v14.16b,v14.16b,#8
 	add	w16,w16,w17
-	sli	v9.4s,v26.4s,#7
+	ext	v18.16b,v18.16b,v18.16b,#8
 	add	w13,w13,w19
-	sli	v13.4s,v27.4s,#7
+	ext	v22.16b,v22.16b,v22.16b,#8
 	add	w14,w14,w20
-	sli	v17.4s,v28.4s,#7
+	ext	v26.16b,v26.16b,v26.16b,#8
 	eor	w10,w10,w15
-	sli	v21.4s,v29.4s,#7
+	ext	v30.16b,v30.16b,v30.16b,#8
 	eor	w11,w11,w16
-	ext	v2.16b,v2.16b,v2.16b,#8
+	ext	v11.16b,v11.16b,v11.16b,#12
 	eor	w12,w12,w13
-	ext	v6.16b,v6.16b,v6.16b,#8
+	ext	v15.16b,v15.16b,v15.16b,#12
 	eor	w9,w9,w14
-	ext	v10.16b,v10.16b,v10.16b,#8
+	ext	v19.16b,v19.16b,v19.16b,#12
 	ror	w10,w10,#25
-	ext	v14.16b,v14.16b,v14.16b,#8
+	ext	v23.16b,v23.16b,v23.16b,#12
 	ror	w11,w11,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	ext	v27.16b,v27.16b,v27.16b,#12
 	ror	w12,w12,#25
-	ext	v22.16b,v22.16b,v22.16b,#8
+	ext	v31.16b,v31.16b,v31.16b,#12
 	ror	w9,w9,#25
-	ext	v3.16b,v3.16b,v3.16b,#12
-	ext	v7.16b,v7.16b,v7.16b,#12
-	ext	v11.16b,v11.16b,v11.16b,#12
-	ext	v15.16b,v15.16b,v15.16b,#12
-	ext	v19.16b,v19.16b,v19.16b,#12
-	ext	v23.16b,v23.16b,v23.16b,#12
-	ext	v1.16b,v1.16b,v1.16b,#4
-	ext	v5.16b,v5.16b,v5.16b,#4
 	ext	v9.16b,v9.16b,v9.16b,#4
 	ext	v13.16b,v13.16b,v13.16b,#4
 	ext	v17.16b,v17.16b,v17.16b,#4
 	ext	v21.16b,v21.16b,v21.16b,#4
-	add	v0.4s,v0.4s,v1.4s
+	ext	v25.16b,v25.16b,v25.16b,#4
+	ext	v29.16b,v29.16b,v29.16b,#4
+	add	v8.4s,v8.4s,v9.4s
 	add	w5,w5,w9
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	add	w6,w6,w10
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	add	w7,w7,w11
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	add	w8,w8,w12
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	eor	w17,w17,w5
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	eor	w19,w19,w6
-	eor	v3.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	eor	w20,w20,w7
-	eor	v7.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	eor	w21,w21,w8
-	eor	v11.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	ror	w17,w17,#16
-	eor	v15.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	ror	w19,w19,#16
-	eor	v19.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	ror	w20,w20,#16
-	eor	v23.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	ror	w21,w21,#16
-	rev32	v3.8h,v3.8h
+	rev32	v11.8h,v11.8h
 	add	w13,w13,w17
-	rev32	v7.8h,v7.8h
+	rev32	v15.8h,v15.8h
 	add	w14,w14,w19
-	rev32	v11.8h,v11.8h
+	rev32	v19.8h,v19.8h
 	add	w15,w15,w20
-	rev32	v15.8h,v15.8h
+	rev32	v23.8h,v23.8h
 	add	w16,w16,w21
-	rev32	v19.8h,v19.8h
+	rev32	v27.8h,v27.8h
 	eor	w9,w9,w13
-	rev32	v23.8h,v23.8h
+	rev32	v31.8h,v31.8h
 	eor	w10,w10,w14
-	add	v2.4s,v2.4s,v3.4s
+	add	v10.4s,v10.4s,v11.4s
 	eor	w11,w11,w15
-	add	v6.4s,v6.4s,v7.4s
+	add	v14.4s,v14.4s,v15.4s
 	eor	w12,w12,w16
-	add	v10.4s,v10.4s,v11.4s
+	add	v18.4s,v18.4s,v19.4s
 	ror	w9,w9,#20
-	add	v14.4s,v14.4s,v15.4s
+	add	v22.4s,v22.4s,v23.4s
 	ror	w10,w10,#20
-	add	v18.4s,v18.4s,v19.4s
+	add	v26.4s,v26.4s,v27.4s
 	ror	w11,w11,#20
-	add	v22.4s,v22.4s,v23.4s
+	add	v30.4s,v30.4s,v31.4s
 	ror	w12,w12,#20
-	eor	v24.16b,v1.16b,v2.16b
+	eor	v0.16b,v9.16b,v10.16b
 	add	w5,w5,w9
-	eor	v25.16b,v5.16b,v6.16b
+	eor	v1.16b,v13.16b,v14.16b
 	add	w6,w6,w10
-	eor	v26.16b,v9.16b,v10.16b
+	eor	v2.16b,v17.16b,v18.16b
 	add	w7,w7,w11
-	eor	v27.16b,v13.16b,v14.16b
+	eor	v3.16b,v21.16b,v22.16b
 	add	w8,w8,w12
-	eor	v28.16b,v17.16b,v18.16b
+	eor	v4.16b,v25.16b,v26.16b
 	eor	w17,w17,w5
-	eor	v29.16b,v21.16b,v22.16b
+	eor	v5.16b,v29.16b,v30.16b
 	eor	w19,w19,w6
-	ushr	v1.4s,v24.4s,#20
+	ushr	v9.4s,v0.4s,#20
 	eor	w20,w20,w7
-	ushr	v5.4s,v25.4s,#20
+	ushr	v13.4s,v1.4s,#20
 	eor	w21,w21,w8
-	ushr	v9.4s,v26.4s,#20
+	ushr	v17.4s,v2.4s,#20
 	ror	w17,w17,#24
-	ushr	v13.4s,v27.4s,#20
+	ushr	v21.4s,v3.4s,#20
 	ror	w19,w19,#24
-	ushr	v17.4s,v28.4s,#20
+	ushr	v25.4s,v4.4s,#20
 	ror	w20,w20,#24
-	ushr	v21.4s,v29.4s,#20
+	ushr	v29.4s,v5.4s,#20
 	ror	w21,w21,#24
-	sli	v1.4s,v24.4s,#12
+	sli	v9.4s,v0.4s,#12
 	add	w13,w13,w17
-	sli	v5.4s,v25.4s,#12
+	sli	v13.4s,v1.4s,#12
 	add	w14,w14,w19
-	sli	v9.4s,v26.4s,#12
+	sli	v17.4s,v2.4s,#12
 	add	w15,w15,w20
-	sli	v13.4s,v27.4s,#12
+	sli	v21.4s,v3.4s,#12
 	add	w16,w16,w21
-	sli	v17.4s,v28.4s,#12
+	sli	v25.4s,v4.4s,#12
 	eor	w9,w9,w13
-	sli	v21.4s,v29.4s,#12
+	sli	v29.4s,v5.4s,#12
 	eor	w10,w10,w14
-	add	v0.4s,v0.4s,v1.4s
+	add	v8.4s,v8.4s,v9.4s
 	eor	w11,w11,w15
-	add	v4.4s,v4.4s,v5.4s
+	add	v12.4s,v12.4s,v13.4s
 	eor	w12,w12,w16
-	add	v8.4s,v8.4s,v9.4s
+	add	v16.4s,v16.4s,v17.4s
 	ror	w9,w9,#25
-	add	v12.4s,v12.4s,v13.4s
+	add	v20.4s,v20.4s,v21.4s
 	ror	w10,w10,#25
-	add	v16.4s,v16.4s,v17.4s
+	add	v24.4s,v24.4s,v25.4s
 	ror	w11,w11,#25
-	add	v20.4s,v20.4s,v21.4s
+	add	v28.4s,v28.4s,v29.4s
 	ror	w12,w12,#25
-	eor	v24.16b,v3.16b,v0.16b
+	eor	v11.16b,v11.16b,v8.16b
 	add	w5,w5,w10
-	eor	v25.16b,v7.16b,v4.16b
+	eor	v15.16b,v15.16b,v12.16b
 	add	w6,w6,w11
-	eor	v26.16b,v11.16b,v8.16b
+	eor	v19.16b,v19.16b,v16.16b
 	add	w7,w7,w12
-	eor	v27.16b,v15.16b,v12.16b
+	eor	v23.16b,v23.16b,v20.16b
 	add	w8,w8,w9
-	eor	v28.16b,v19.16b,v16.16b
+	eor	v27.16b,v27.16b,v24.16b
 	eor	w21,w21,w5
-	eor	v29.16b,v23.16b,v20.16b
+	eor	v31.16b,v31.16b,v28.16b
 	eor	w17,w17,w6
-	ushr	v3.4s,v24.4s,#24
+	tbl	v11.16b,{v11.16b},v6.16b
 	eor	w19,w19,w7
-	ushr	v7.4s,v25.4s,#24
+	tbl	v15.16b,{v15.16b},v6.16b
 	eor	w20,w20,w8
-	ushr	v11.4s,v26.4s,#24
+	tbl	v19.16b,{v19.16b},v6.16b
 	ror	w21,w21,#16
-	ushr	v15.4s,v27.4s,#24
+	tbl	v23.16b,{v23.16b},v6.16b
 	ror	w17,w17,#16
-	ushr	v19.4s,v28.4s,#24
+	tbl	v27.16b,{v27.16b},v6.16b
 	ror	w19,w19,#16
-	ushr	v23.4s,v29.4s,#24
+	tbl	v31.16b,{v31.16b},v6.16b
 	ror	w20,w20,#16
-	sli	v3.4s,v24.4s,#8
+	add	v10.4s,v10.4s,v11.4s
 	add	w15,w15,w21
-	sli	v7.4s,v25.4s,#8
+	add	v14.4s,v14.4s,v15.4s
 	add	w16,w16,w17
-	sli	v11.4s,v26.4s,#8
+	add	v18.4s,v18.4s,v19.4s
 	add	w13,w13,w19
-	sli	v15.4s,v27.4s,#8
+	add	v22.4s,v22.4s,v23.4s
 	add	w14,w14,w20
-	sli	v19.4s,v28.4s,#8
+	add	v26.4s,v26.4s,v27.4s
 	eor	w10,w10,w15
-	sli	v23.4s,v29.4s,#8
+	add	v30.4s,v30.4s,v31.4s
 	eor	w11,w11,w16
-	add	v2.4s,v2.4s,v3.4s
+	eor	v0.16b,v9.16b,v10.16b
 	eor	w12,w12,w13
-	add	v6.4s,v6.4s,v7.4s
+	eor	v1.16b,v13.16b,v14.16b
 	eor	w9,w9,w14
-	add	v10.4s,v10.4s,v11.4s
+	eor	v2.16b,v17.16b,v18.16b
 	ror	w10,w10,#20
-	add	v14.4s,v14.4s,v15.4s
+	eor	v3.16b,v21.16b,v22.16b
 	ror	w11,w11,#20
-	add	v18.4s,v18.4s,v19.4s
+	eor	v4.16b,v25.16b,v26.16b
 	ror	w12,w12,#20
-	add	v22.4s,v22.4s,v23.4s
+	eor	v5.16b,v29.16b,v30.16b
 	ror	w9,w9,#20
-	eor	v24.16b,v1.16b,v2.16b
+	ushr	v9.4s,v0.4s,#25
 	add	w5,w5,w10
-	eor	v25.16b,v5.16b,v6.16b
+	ushr	v13.4s,v1.4s,#25
 	add	w6,w6,w11
-	eor	v26.16b,v9.16b,v10.16b
+	ushr	v17.4s,v2.4s,#25
 	add	w7,w7,w12
-	eor	v27.16b,v13.16b,v14.16b
+	ushr	v21.4s,v3.4s,#25
 	add	w8,w8,w9
-	eor	v28.16b,v17.16b,v18.16b
+	ushr	v25.4s,v4.4s,#25
 	eor	w21,w21,w5
-	eor	v29.16b,v21.16b,v22.16b
+	ushr	v29.4s,v5.4s,#25
 	eor	w17,w17,w6
-	ushr	v1.4s,v24.4s,#25
+	sli	v9.4s,v0.4s,#7
 	eor	w19,w19,w7
-	ushr	v5.4s,v25.4s,#25
+	sli	v13.4s,v1.4s,#7
 	eor	w20,w20,w8
-	ushr	v9.4s,v26.4s,#25
+	sli	v17.4s,v2.4s,#7
 	ror	w21,w21,#24
-	ushr	v13.4s,v27.4s,#25
+	sli	v21.4s,v3.4s,#7
 	ror	w17,w17,#24
-	ushr	v17.4s,v28.4s,#25
+	sli	v25.4s,v4.4s,#7
 	ror	w19,w19,#24
-	ushr	v21.4s,v29.4s,#25
+	sli	v29.4s,v5.4s,#7
 	ror	w20,w20,#24
-	sli	v1.4s,v24.4s,#7
+	ext	v10.16b,v10.16b,v10.16b,#8
 	add	w15,w15,w21
-	sli	v5.4s,v25.4s,#7
+	ext	v14.16b,v14.16b,v14.16b,#8
 	add	w16,w16,w17
-	sli	v9.4s,v26.4s,#7
+	ext	v18.16b,v18.16b,v18.16b,#8
 	add	w13,w13,w19
-	sli	v13.4s,v27.4s,#7
+	ext	v22.16b,v22.16b,v22.16b,#8
 	add	w14,w14,w20
-	sli	v17.4s,v28.4s,#7
+	ext	v26.16b,v26.16b,v26.16b,#8
 	eor	w10,w10,w15
-	sli	v21.4s,v29.4s,#7
+	ext	v30.16b,v30.16b,v30.16b,#8
 	eor	w11,w11,w16
-	ext	v2.16b,v2.16b,v2.16b,#8
+	ext	v11.16b,v11.16b,v11.16b,#4
 	eor	w12,w12,w13
-	ext	v6.16b,v6.16b,v6.16b,#8
+	ext	v15.16b,v15.16b,v15.16b,#4
 	eor	w9,w9,w14
-	ext	v10.16b,v10.16b,v10.16b,#8
+	ext	v19.16b,v19.16b,v19.16b,#4
 	ror	w10,w10,#25
-	ext	v14.16b,v14.16b,v14.16b,#8
+	ext	v23.16b,v23.16b,v23.16b,#4
 	ror	w11,w11,#25
-	ext	v18.16b,v18.16b,v18.16b,#8
+	ext	v27.16b,v27.16b,v27.16b,#4
 	ror	w12,w12,#25
-	ext	v22.16b,v22.16b,v22.16b,#8
+	ext	v31.16b,v31.16b,v31.16b,#4
 	ror	w9,w9,#25
-	ext	v3.16b,v3.16b,v3.16b,#4
-	ext	v7.16b,v7.16b,v7.16b,#4
-	ext	v11.16b,v11.16b,v11.16b,#4
-	ext	v15.16b,v15.16b,v15.16b,#4
-	ext	v19.16b,v19.16b,v19.16b,#4
-	ext	v23.16b,v23.16b,v23.16b,#4
-	ext	v1.16b,v1.16b,v1.16b,#12
-	ext	v5.16b,v5.16b,v5.16b,#12
 	ext	v9.16b,v9.16b,v9.16b,#12
 	ext	v13.16b,v13.16b,v13.16b,#12
 	ext	v17.16b,v17.16b,v17.16b,#12
 	ext	v21.16b,v21.16b,v21.16b,#12
+	ext	v25.16b,v25.16b,v25.16b,#12
+	ext	v29.16b,v29.16b,v29.16b,#12
 	cbnz	x4,.Loop_lower_neon
 
 	add	w5,w5,w22		// accumulate key block
-	ldp	q24,q25,[sp,#0]
+	ldp	q0,q1,[sp,#0]
 	add	x6,x6,x22,lsr#32
-	ldp	q26,q27,[sp,#32]
+	ldp	q2,q3,[sp,#32]
 	add	w7,w7,w23
-	ldp	q28,q29,[sp,#64]
+	ldp	q4,q5,[sp,#64]
 	add	x8,x8,x23,lsr#32
-	add	v0.4s,v0.4s,v24.4s
+	ldr	q6,[sp,#96]
+	add	v8.4s,v8.4s,v0.4s
 	add	w9,w9,w24
-	add	v4.4s,v4.4s,v24.4s
+	add	v12.4s,v12.4s,v0.4s
 	add	x10,x10,x24,lsr#32
-	add	v8.4s,v8.4s,v24.4s
+	add	v16.4s,v16.4s,v0.4s
 	add	w11,w11,w25
-	add	v12.4s,v12.4s,v24.4s
+	add	v20.4s,v20.4s,v0.4s
 	add	x12,x12,x25,lsr#32
-	add	v16.4s,v16.4s,v24.4s
+	add	v24.4s,v24.4s,v0.4s
 	add	w13,w13,w26
-	add	v20.4s,v20.4s,v24.4s
+	add	v28.4s,v28.4s,v0.4s
 	add	x14,x14,x26,lsr#32
-	add	v2.4s,v2.4s,v26.4s
+	add	v10.4s,v10.4s,v2.4s
 	add	w15,w15,w27
-	add	v6.4s,v6.4s,v26.4s
+	add	v14.4s,v14.4s,v2.4s
 	add	x16,x16,x27,lsr#32
-	add	v10.4s,v10.4s,v26.4s
+	add	v18.4s,v18.4s,v2.4s
 	add	w17,w17,w28
-	add	v14.4s,v14.4s,v26.4s
+	add	v22.4s,v22.4s,v2.4s
 	add	x19,x19,x28,lsr#32
-	add	v18.4s,v18.4s,v26.4s
+	add	v26.4s,v26.4s,v2.4s
 	add	w20,w20,w30
-	add	v22.4s,v22.4s,v26.4s
+	add	v30.4s,v30.4s,v2.4s
 	add	x21,x21,x30,lsr#32
-	add	v19.4s,v19.4s,v31.4s			// +4
+	add	v27.4s,v27.4s,v7.4s			// +4
 	add	x5,x5,x6,lsl#32	// pack
-	add	v23.4s,v23.4s,v31.4s			// +4
+	add	v31.4s,v31.4s,v7.4s			// +4
 	add	x7,x7,x8,lsl#32
-	add	v3.4s,v3.4s,v27.4s
+	add	v11.4s,v11.4s,v3.4s
 	ldp	x6,x8,[x1,#0]		// load input
-	add	v7.4s,v7.4s,v28.4s
+	add	v15.4s,v15.4s,v4.4s
 	add	x9,x9,x10,lsl#32
-	add	v11.4s,v11.4s,v29.4s
+	add	v19.4s,v19.4s,v5.4s
 	add	x11,x11,x12,lsl#32
-	add	v15.4s,v15.4s,v30.4s
+	add	v23.4s,v23.4s,v6.4s
 	ldp	x10,x12,[x1,#16]
-	add	v19.4s,v19.4s,v27.4s
+	add	v27.4s,v27.4s,v3.4s
 	add	x13,x13,x14,lsl#32
-	add	v23.4s,v23.4s,v28.4s
+	add	v31.4s,v31.4s,v4.4s
 	add	x15,x15,x16,lsl#32
-	add	v1.4s,v1.4s,v25.4s
+	add	v9.4s,v9.4s,v1.4s
 	ldp	x14,x16,[x1,#32]
-	add	v5.4s,v5.4s,v25.4s
+	add	v13.4s,v13.4s,v1.4s
 	add	x17,x17,x19,lsl#32
-	add	v9.4s,v9.4s,v25.4s
+	add	v17.4s,v17.4s,v1.4s
 	add	x20,x20,x21,lsl#32
-	add	v13.4s,v13.4s,v25.4s
+	add	v21.4s,v21.4s,v1.4s
 	ldp	x19,x21,[x1,#48]
-	add	v17.4s,v17.4s,v25.4s
+	add	v25.4s,v25.4s,v1.4s
 	add	x1,x1,#64
-	add	v21.4s,v21.4s,v25.4s
+	add	v29.4s,v29.4s,v1.4s
 
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	rev	x5,x5
 	rev	x7,x7
 	rev	x9,x9
@@ -1869,20 +1925,20 @@ ChaCha20_512_neon:
 	rev	x17,x17
 	rev	x20,x20
 #endif
-	ld1	{v24.16b,v25.16b,v26.16b,v27.16b},[x1],#64
+	ld1	{v0.16b,v1.16b,v2.16b,v3.16b},[x1],#64
 	eor	x5,x5,x6
 	eor	x7,x7,x8
 	eor	x9,x9,x10
 	eor	x11,x11,x12
 	eor	x13,x13,x14
-	eor	v0.16b,v0.16b,v24.16b
+	eor	v8.16b,v8.16b,v0.16b
 	eor	x15,x15,x16
-	eor	v1.16b,v1.16b,v25.16b
+	eor	v9.16b,v9.16b,v1.16b
 	eor	x17,x17,x19
-	eor	v2.16b,v2.16b,v26.16b
+	eor	v10.16b,v10.16b,v2.16b
 	eor	x20,x20,x21
-	eor	v3.16b,v3.16b,v27.16b
-	ld1	{v24.16b,v25.16b,v26.16b,v27.16b},[x1],#64
+	eor	v11.16b,v11.16b,v3.16b
+	ld1	{v0.16b,v1.16b,v2.16b,v3.16b},[x1],#64
 
 	stp	x5,x7,[x0,#0]		// store output
 	add	x28,x28,#7			// increment counter
@@ -1890,82 +1946,83 @@ ChaCha20_512_neon:
 	stp	x13,x15,[x0,#32]
 	stp	x17,x20,[x0,#48]
 	add	x0,x0,#64
-	st1	{v0.16b,v1.16b,v2.16b,v3.16b},[x0],#64
-
-	ld1	{v0.16b,v1.16b,v2.16b,v3.16b},[x1],#64
-	eor	v4.16b,v4.16b,v24.16b
-	eor	v5.16b,v5.16b,v25.16b
-	eor	v6.16b,v6.16b,v26.16b
-	eor	v7.16b,v7.16b,v27.16b
-	st1	{v4.16b,v5.16b,v6.16b,v7.16b},[x0],#64
-
-	ld1	{v4.16b,v5.16b,v6.16b,v7.16b},[x1],#64
-	eor	v8.16b,v8.16b,v0.16b
-	ldp	q24,q25,[sp,#0]
-	eor	v9.16b,v9.16b,v1.16b
-	ldp	q26,q27,[sp,#32]
-	eor	v10.16b,v10.16b,v2.16b
-	eor	v11.16b,v11.16b,v3.16b
 	st1	{v8.16b,v9.16b,v10.16b,v11.16b},[x0],#64
 
 	ld1	{v8.16b,v9.16b,v10.16b,v11.16b},[x1],#64
-	eor	v12.16b,v12.16b,v4.16b
-	eor	v13.16b,v13.16b,v5.16b
-	eor	v14.16b,v14.16b,v6.16b
-	eor	v15.16b,v15.16b,v7.16b
+	eor	v12.16b,v12.16b,v0.16b
+	eor	v13.16b,v13.16b,v1.16b
+	eor	v14.16b,v14.16b,v2.16b
+	eor	v15.16b,v15.16b,v3.16b
 	st1	{v12.16b,v13.16b,v14.16b,v15.16b},[x0],#64
 
 	ld1	{v12.16b,v13.16b,v14.16b,v15.16b},[x1],#64
 	eor	v16.16b,v16.16b,v8.16b
+	ldp	q0,q1,[sp,#0]
 	eor	v17.16b,v17.16b,v9.16b
+	ldp	q2,q3,[sp,#32]
 	eor	v18.16b,v18.16b,v10.16b
 	eor	v19.16b,v19.16b,v11.16b
 	st1	{v16.16b,v17.16b,v18.16b,v19.16b},[x0],#64
 
-	shl	v0.4s,v31.4s,#1			// 4 -> 8
+	ld1	{v16.16b,v17.16b,v18.16b,v19.16b},[x1],#64
 	eor	v20.16b,v20.16b,v12.16b
 	eor	v21.16b,v21.16b,v13.16b
 	eor	v22.16b,v22.16b,v14.16b
 	eor	v23.16b,v23.16b,v15.16b
 	st1	{v20.16b,v21.16b,v22.16b,v23.16b},[x0],#64
 
-	add	v27.4s,v27.4s,v0.4s			// += 8
-	add	v28.4s,v28.4s,v0.4s
-	add	v29.4s,v29.4s,v0.4s
-	add	v30.4s,v30.4s,v0.4s
+	ld1	{v20.16b,v21.16b,v22.16b,v23.16b},[x1],#64
+	eor	v24.16b,v24.16b,v16.16b
+	eor	v25.16b,v25.16b,v17.16b
+	eor	v26.16b,v26.16b,v18.16b
+	eor	v27.16b,v27.16b,v19.16b
+	st1	{v24.16b,v25.16b,v26.16b,v27.16b},[x0],#64
+
+	shl	v8.4s,v7.4s,#1			// 4 -> 8
+	eor	v28.16b,v28.16b,v20.16b
+	eor	v29.16b,v29.16b,v21.16b
+	eor	v30.16b,v30.16b,v22.16b
+	eor	v31.16b,v31.16b,v23.16b
+	st1	{v28.16b,v29.16b,v30.16b,v31.16b},[x0],#64
+
+	add	v3.4s,v3.4s,v8.4s			// += 8
+	add	v4.4s,v4.4s,v8.4s
+	add	v5.4s,v5.4s,v8.4s
+	add	v6.4s,v6.4s,v8.4s
 
 	b.hs	.Loop_outer_512_neon
 
 	adds	x2,x2,#512
-	ushr	v0.4s,v31.4s,#2			// 4 -> 1
+	ushr	v7.4s,v7.4s,#1			// 4 -> 2
 
-	ldp	d8,d9,[sp,#128+0]		// meet ABI requirements
-	ldp	d10,d11,[sp,#128+16]
+	ldp	d10,d11,[sp,#128+16]		// meet ABI requirements
 	ldp	d12,d13,[sp,#128+32]
 	ldp	d14,d15,[sp,#128+48]
 
-	stp	q24,q31,[sp,#0]		// wipe off-load area
-	stp	q24,q31,[sp,#32]
-	stp	q24,q31,[sp,#64]
+	stp	q0,q0,[sp,#0]		// wipe off-load area
+	stp	q0,q0,[sp,#32]
+	stp	q0,q0,[sp,#64]
 
 	b.eq	.Ldone_512_neon
 
+	sub	x3,x3,#16			// .Lone
 	cmp	x2,#192
-	sub	v27.4s,v27.4s,v0.4s			// -= 1
-	sub	v28.4s,v28.4s,v0.4s
-	sub	v29.4s,v29.4s,v0.4s
 	add	sp,sp,#128
+	sub	v3.4s,v3.4s,v7.4s		// -= 2
+	ld1	{v8.4s,v9.4s},[x3]
 	b.hs	.Loop_outer_neon
 
-	eor	v25.16b,v25.16b,v25.16b
-	eor	v26.16b,v26.16b,v26.16b
-	eor	v27.16b,v27.16b,v27.16b
-	eor	v28.16b,v28.16b,v28.16b
-	eor	v29.16b,v29.16b,v29.16b
-	eor	v30.16b,v30.16b,v30.16b
+	ldp	d8,d9,[sp,#0]			// meet ABI requirements
+	eor	v1.16b,v1.16b,v1.16b
+	eor	v2.16b,v2.16b,v2.16b
+	eor	v3.16b,v3.16b,v3.16b
+	eor	v4.16b,v4.16b,v4.16b
+	eor	v5.16b,v5.16b,v5.16b
+	eor	v6.16b,v6.16b,v6.16b
 	b	.Loop_outer
 
 .Ldone_512_neon:
+	ldp	d8,d9,[sp,#128+0]		// meet ABI requirements
 	ldp	x19,x20,[x29,#16]
 	add	sp,sp,#128+64
 	ldp	x21,x22,[x29,#32]
diff --git a/sys/crypto/openssl/aarch64/ecp_nistz256-armv8.S b/sys/crypto/openssl/aarch64/ecp_nistz256-armv8.S
index fc74b7095cb84..d297ac15605cf 100644
--- a/sys/crypto/openssl/aarch64/ecp_nistz256-armv8.S
+++ b/sys/crypto/openssl/aarch64/ecp_nistz256-armv8.S
@@ -3823,7 +3823,7 @@ ecp_nistz256_ord_mul_mont:
 
 ////////////////////////////////////////////////////////////////////////
 // void ecp_nistz256_ord_sqr_mont(uint64_t res[4], uint64_t a[4],
-//                                int rep);
+//                                uint64_t rep);
 .globl	ecp_nistz256_ord_sqr_mont
 .type	ecp_nistz256_ord_sqr_mont,%function
 .align	4
@@ -4023,7 +4023,7 @@ ecp_nistz256_scatter_w5:
 
 	ldp	x4,x5,[x1]		// X
 	ldp	x6,x7,[x1,#16]
-	str	w4,[x0,#64*0-4]
+	stur	w4,[x0,#64*0-4]
 	lsr	x4,x4,#32
 	str	w5,[x0,#64*1-4]
 	lsr	x5,x5,#32
@@ -4039,7 +4039,7 @@ ecp_nistz256_scatter_w5:
 
 	ldp	x4,x5,[x1,#32]	// Y
 	ldp	x6,x7,[x1,#48]
-	str	w4,[x0,#64*0-4]
+	stur	w4,[x0,#64*0-4]
 	lsr	x4,x4,#32
 	str	w5,[x0,#64*1-4]
 	lsr	x5,x5,#32
@@ -4055,7 +4055,7 @@ ecp_nistz256_scatter_w5:
 
 	ldp	x4,x5,[x1,#64]	// Z
 	ldp	x6,x7,[x1,#80]
-	str	w4,[x0,#64*0-4]
+	stur	w4,[x0,#64*0-4]
 	lsr	x4,x4,#32
 	str	w5,[x0,#64*1-4]
 	lsr	x5,x5,#32
diff --git a/sys/crypto/openssl/aarch64/ghashv8-armx.S b/sys/crypto/openssl/aarch64/ghashv8-armx.S
index 387d3c1f27b86..fe5cd49a25aa3 100644
--- a/sys/crypto/openssl/aarch64/ghashv8-armx.S
+++ b/sys/crypto/openssl/aarch64/ghashv8-armx.S
@@ -2,6 +2,7 @@
 #include "arm_arch.h"
 
 #if __ARM_MAX_ARCH__>=7
+.arch	armv8-a+crypto
 .text
 .globl	gcm_init_v8
 .type	gcm_init_v8,%function
diff --git a/sys/crypto/openssl/aarch64/keccak1600-armv8.S b/sys/crypto/openssl/aarch64/keccak1600-armv8.S
index d52a88d8d0fd9..cf63318a8d087 100644
--- a/sys/crypto/openssl/aarch64/keccak1600-armv8.S
+++ b/sys/crypto/openssl/aarch64/keccak1600-armv8.S
@@ -574,22 +574,22 @@ SHA3_squeeze:
 .type	KeccakF1600_ce,%function
 .align	5
 KeccakF1600_ce:
-	mov	x9,#12
+	mov	x9,#24
 	adr	x10,iotas
 	b	.Loop_ce
 .align	4
 .Loop_ce:
 	////////////////////////////////////////////////// Theta
-.inst	0xce052819	//eor3 v25.16b,v0.16b,v5.16b,v10.16b
-.inst	0xce062c3a	//eor3 v26.16b,v1.16b,v6.16b,v11.16b
-.inst	0xce07305b	//eor3 v27.16b,v2.16b,v7.16b,v12.16b
-.inst	0xce08347c	//eor3 v28.16b,v3.16b,v8.16b,v13.16b
-.inst	0xce09389d	//eor3 v29.16b,v4.16b,v9.16b,v14.16b
-.inst	0xce0f5339	//eor3 v25.16b,v25.16b,   v15.16b,v20.16b
-.inst	0xce10575a	//eor3 v26.16b,v26.16b,   v16.16b,v21.16b
-.inst	0xce115b7b	//eor3 v27.16b,v27.16b,   v17.16b,v22.16b
-.inst	0xce125f9c	//eor3 v28.16b,v28.16b,   v18.16b,v23.16b
-.inst	0xce1363bd	//eor3 v29.16b,v29.16b,   v19.16b,v24.16b
+.inst	0xce0f2a99	//eor3 v25.16b,v20.16b,v15.16b,v10.16b
+.inst	0xce102eba	//eor3 v26.16b,v21.16b,v16.16b,v11.16b
+.inst	0xce1132db	//eor3 v27.16b,v22.16b,v17.16b,v12.16b
+.inst	0xce1236fc	//eor3 v28.16b,v23.16b,v18.16b,v13.16b
+.inst	0xce133b1d	//eor3 v29.16b,v24.16b,v19.16b,v14.16b
+.inst	0xce050339	//eor3 v25.16b,v25.16b,   v5.16b,v0.16b
+.inst	0xce06075a	//eor3 v26.16b,v26.16b,   v6.16b,v1.16b
+.inst	0xce070b7b	//eor3 v27.16b,v27.16b,   v7.16b,v2.16b
+.inst	0xce080f9c	//eor3 v28.16b,v28.16b,   v8.16b,v3.16b
+.inst	0xce0913bd	//eor3 v29.16b,v29.16b,   v9.16b,v4.16b
 
 .inst	0xce7b8f3e	//rax1 v30.16b,v25.16b,v27.16b			// D[1]
 .inst	0xce7c8f5f	//rax1 v31.16b,v26.16b,v28.16b			// D[2]
@@ -598,13 +598,15 @@ KeccakF1600_ce:
 .inst	0xce7a8fbd	//rax1 v29.16b,v29.16b,v26.16b			// D[0]
 
 	////////////////////////////////////////////////// Theta+Rho+Pi
-.inst	0xce9e50d9	//xar v25.16b,   v6.16b,v30.16b,#64-44	// C[0]=A[0][1]
+.inst	0xce9efc39	//xar v25.16b,   v1.16b,v30.16b,#64-1 // C[0]=A[2][0]
+
+.inst	0xce9e50c1	//xar v1.16b,v6.16b,v30.16b,#64-44
 .inst	0xce9cb126	//xar v6.16b,v9.16b,v28.16b,#64-20
 .inst	0xce9f0ec9	//xar v9.16b,v22.16b,v31.16b,#64-61
 .inst	0xce9c65d6	//xar v22.16b,v14.16b,v28.16b,#64-39
 .inst	0xce9dba8e	//xar v14.16b,v20.16b,v29.16b,#64-18
 
-.inst	0xce9f0854	//xar v20.16b,v2.16b,v31.16b,#64-62
+.inst	0xce9f085a	//xar v26.16b,   v2.16b,v31.16b,#64-62 // C[1]=A[4][0]
 
 .inst	0xce9f5582	//xar v2.16b,v12.16b,v31.16b,#64-43
 .inst	0xce9b9dac	//xar v12.16b,v13.16b,v27.16b,#64-25
@@ -614,145 +616,57 @@ KeccakF1600_ce:
 
 .inst	0xce9c948f	//xar v15.16b,v4.16b,v28.16b,#64-27
 
-	eor	v0.16b,v0.16b,v29.16b
-	ldr	x11,[x10],#8
-
-.inst	0xce9bae5a	//xar v26.16b,   v18.16b,v27.16b,#64-21	// C[1]=A[0][3]
-.inst	0xce9fc632	//xar v18.16b,v17.16b,v31.16b,#64-15
-.inst	0xce9ed971	//xar v17.16b,v11.16b,v30.16b,#64-10
-.inst	0xce9fe8eb	//xar v11.16b,v7.16b,v31.16b,#64-6
-.inst	0xce9df547	//xar v7.16b,v10.16b,v29.16b,#64-3
-
-.inst	0xce9efc2a	//xar v10.16b,v1.16b,v30.16b,#64-1	// *
-
-.inst	0xce9ccb04	//xar v4.16b,v24.16b,v28.16b,#64-14
+.inst	0xce9ccb1c	//xar v28.16b,   v24.16b,v28.16b,#64-14 // D[4]=A[0][4]
 .inst	0xce9efab8	//xar v24.16b,v21.16b,v30.16b,#64-2
-.inst	0xce9b2515	//xar v21.16b,v8.16b,v27.16b,#64-55
-.inst	0xce9e4e08	//xar v8.16b,v16.16b,v30.16b,#64-45
+.inst	0xce9b2508	//xar v8.16b,v8.16b,v27.16b,#64-55 // A[1][3]=A[4][1]
+.inst	0xce9e4e04	//xar v4.16b,v16.16b,v30.16b,#64-45 // A[0][4]=A[1][3]
 .inst	0xce9d70b0	//xar v16.16b,v5.16b,v29.16b,#64-36
 
-.inst	0xce9b907b	//xar v27.16b,   v3.16b,v27.16b,#64-28	// C[2]=A[1][0]
+.inst	0xce9b9065	//xar v5.16b,v3.16b,v27.16b,#64-28
 
-	////////////////////////////////////////////////// Chi+Iota
-	dup	v31.2d,x11				// borrow C[6]
-.inst	0xce22641c	//bcax v28.16b,   v0.16b,v2.16b,v25.16b	// *
-.inst	0xce3a0b21	//bcax v1.16b,v25.16b,   v26.16b,   v2.16b	// *
-.inst	0xce246842	//bcax v2.16b,v2.16b,v4.16b,v26.16b
-.inst	0xce201343	//bcax v3.16b,v26.16b,   v0.16b,v4.16b
-.inst	0xce390084	//bcax v4.16b,v4.16b,v25.16b,   v0.16b
-
-.inst	0xce271b65	//bcax v5.16b,v27.16b,   v7.16b,v6.16b	// *
-.inst	0xce281cd9	//bcax v25.16b,   v6.16b,v8.16b,v7.16b	// *
-.inst	0xce2920e7	//bcax v7.16b,v7.16b,v9.16b,v8.16b
-.inst	0xce3b2508	//bcax v8.16b,v8.16b,v27.16b,   v9.16b
-.inst	0xce266d29	//bcax v9.16b,v9.16b,v6.16b,v27.16b
-
-	eor	v0.16b,v28.16b,v31.16b			// Iota
-
-.inst	0xce2c2d5a	//bcax v26.16b,   v10.16b,v12.16b,v11.16b	// *
-.inst	0xce2d317b	//bcax v27.16b,   v11.16b,v13.16b,v12.16b	// *
-.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
-.inst	0xce2a39ad	//bcax v13.16b,v13.16b,v10.16b,v14.16b
-.inst	0xce2b29ce	//bcax v14.16b,v14.16b,v11.16b,v10.16b
+	eor	v0.16b,v0.16b,v29.16b
 
-.inst	0xce3141fc	//bcax v28.16b,   v15.16b,v17.16b,v16.16b	// *
-.inst	0xce32461d	//bcax v29.16b,   v16.16b,v18.16b,v17.16b	// *
-.inst	0xce334a31	//bcax v17.16b,v17.16b,v19.16b,v18.16b
-.inst	0xce2f4e52	//bcax v18.16b,v18.16b,v15.16b,v19.16b
-.inst	0xce303e73	//bcax v19.16b,v19.16b,v16.16b,v15.16b
+.inst	0xce9bae5b	//xar v27.16b,   v18.16b,v27.16b,#64-21 // D[3]=A[0][3]
+.inst	0xce9fc623	//xar v3.16b,v17.16b,v31.16b,#64-15 // A[0][3]=A[3][3]
+.inst	0xce9ed97e	//xar v30.16b,   v11.16b,v30.16b,#64-10 // D[1]=A[3][2]
+.inst	0xce9fe8ff	//xar v31.16b,   v7.16b,v31.16b,#64-6 // D[2]=A[2][1]
+.inst	0xce9df55d	//xar v29.16b,   v10.16b,v29.16b,#64-3 // D[0]=A[1][2]
 
-.inst	0xce36569e	//bcax v30.16b,   v20.16b,v22.16b,v21.16b	// *
-.inst	0xce375abf	//bcax v31.16b,   v21.16b,v23.16b,v22.16b	// *
+	////////////////////////////////////////////////// Chi+Iota
+.inst	0xce362354	//bcax v20.16b,v26.16b,   v22.16b,v8.16b	// A[1][3]=A[4][1]
+.inst	0xce375915	//bcax v21.16b,v8.16b,v23.16b,v22.16b	// A[1][3]=A[4][1]
 .inst	0xce385ed6	//bcax v22.16b,v22.16b,v24.16b,v23.16b
-.inst	0xce3462f7	//bcax v23.16b,v23.16b,v20.16b,v24.16b
-.inst	0xce355318	//bcax v24.16b,v24.16b,v21.16b,v20.16b
-	////////////////////////////////////////////////// Theta
-.inst	0xce056806	//eor3 v6.16b,v0.16b,v5.16b,v26.16b
-.inst	0xce196c2a	//eor3 v10.16b,v1.16b,v25.16b,v27.16b
-.inst	0xce07304b	//eor3 v11.16b,v2.16b,v7.16b,v12.16b
-.inst	0xce08346f	//eor3 v15.16b,v3.16b,v8.16b,v13.16b
-.inst	0xce093890	//eor3 v16.16b,v4.16b,v9.16b,v14.16b
-.inst	0xce1c78c6	//eor3 v6.16b,v6.16b,   v28.16b,v30.16b
-.inst	0xce1d7d4a	//eor3 v10.16b,v10.16b,   v29.16b,v31.16b
-.inst	0xce11596b	//eor3 v11.16b,v11.16b,   v17.16b,v22.16b
-.inst	0xce125def	//eor3 v15.16b,v15.16b,   v18.16b,v23.16b
-.inst	0xce136210	//eor3 v16.16b,v16.16b,   v19.16b,v24.16b
-
-.inst	0xce6b8cd4	//rax1 v20.16b,v6.16b,v11.16b			// D[1]
-.inst	0xce6f8d55	//rax1 v21.16b,v10.16b,v15.16b			// D[2]
-.inst	0xce708d6b	//rax1 v11.16b,v11.16b,v16.16b			// D[3]
-.inst	0xce668def	//rax1 v15.16b,v15.16b,v6.16b			// D[4]
-.inst	0xce6a8e10	//rax1 v16.16b,v16.16b,v10.16b			// D[0]
+.inst	0xce3a62f7	//bcax v23.16b,v23.16b,v26.16b,   v24.16b
+.inst	0xce286b18	//bcax v24.16b,v24.16b,v8.16b,v26.16b	// A[1][3]=A[4][1]
 
-	////////////////////////////////////////////////// Theta+Rho+Pi
-.inst	0xce945326	//xar v6.16b,   v25.16b,v20.16b,#64-44	// C[0]=A[0][1]
-.inst	0xce8fb139	//xar v25.16b,v9.16b,v15.16b,#64-20
-.inst	0xce950ec9	//xar v9.16b,v22.16b,v21.16b,#64-61
-.inst	0xce8f65d6	//xar v22.16b,v14.16b,v15.16b,#64-39
-.inst	0xce90bbce	//xar v14.16b,v30.16b,v16.16b,#64-18
-
-.inst	0xce95085e	//xar v30.16b,v2.16b,v21.16b,#64-62
-
-.inst	0xce955582	//xar v2.16b,v12.16b,v21.16b,#64-43
-.inst	0xce8b9dac	//xar v12.16b,v13.16b,v11.16b,#64-25
-.inst	0xce8fe26d	//xar v13.16b,v19.16b,v15.16b,#64-8
-.inst	0xce8b22f3	//xar v19.16b,v23.16b,v11.16b,#64-56
-.inst	0xce905f97	//xar v23.16b,v28.16b,v16.16b,#64-41
-
-.inst	0xce8f949c	//xar v28.16b,v4.16b,v15.16b,#64-27
-
-	eor	v0.16b,v0.16b,v16.16b
-	ldr	x11,[x10],#8
-
-.inst	0xce8bae4a	//xar v10.16b,   v18.16b,v11.16b,#64-21	// C[1]=A[0][3]
-.inst	0xce95c632	//xar v18.16b,v17.16b,v21.16b,#64-15
-.inst	0xce94db71	//xar v17.16b,v27.16b,v20.16b,#64-10
-.inst	0xce95e8fb	//xar v27.16b,v7.16b,v21.16b,#64-6
-.inst	0xce90f747	//xar v7.16b,v26.16b,v16.16b,#64-3
+	ld1r	{v26.2d},[x10],#8
 
-.inst	0xce94fc3a	//xar v26.16b,v1.16b,v20.16b,#64-1	// *
+.inst	0xce330fd1	//bcax v17.16b,v30.16b,   v19.16b,v3.16b	// A[0][3]=A[3][3]
+.inst	0xce2f4c72	//bcax v18.16b,v3.16b,v15.16b,v19.16b	// A[0][3]=A[3][3]
+.inst	0xce303e73	//bcax v19.16b,v19.16b,v16.16b,v15.16b
+.inst	0xce3e41ef	//bcax v15.16b,v15.16b,v30.16b,   v16.16b
+.inst	0xce237a10	//bcax v16.16b,v16.16b,v3.16b,v30.16b	// A[0][3]=A[3][3]
 
-.inst	0xce8fcb04	//xar v4.16b,v24.16b,v15.16b,#64-14
-.inst	0xce94fbf8	//xar v24.16b,v31.16b,v20.16b,#64-2
-.inst	0xce8b251f	//xar v31.16b,v8.16b,v11.16b,#64-55
-.inst	0xce944fa8	//xar v8.16b,v29.16b,v20.16b,#64-45
-.inst	0xce9070bd	//xar v29.16b,v5.16b,v16.16b,#64-36
+.inst	0xce2c7f2a	//bcax v10.16b,v25.16b,   v12.16b,v31.16b
+.inst	0xce2d33eb	//bcax v11.16b,v31.16b,   v13.16b,v12.16b
+.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
+.inst	0xce3939ad	//bcax v13.16b,v13.16b,v25.16b,   v14.16b
+.inst	0xce3f65ce	//bcax v14.16b,v14.16b,v31.16b,   v25.16b
 
-.inst	0xce8b906b	//xar v11.16b,   v3.16b,v11.16b,#64-28	// C[2]=A[1][0]
+.inst	0xce2913a7	//bcax v7.16b,v29.16b,   v9.16b,v4.16b	// A[0][4]=A[1][3]
+.inst	0xce252488	//bcax v8.16b,v4.16b,v5.16b,v9.16b	// A[0][4]=A[1][3]
+.inst	0xce261529	//bcax v9.16b,v9.16b,v6.16b,v5.16b
+.inst	0xce3d18a5	//bcax v5.16b,v5.16b,v29.16b,   v6.16b
+.inst	0xce2474c6	//bcax v6.16b,v6.16b,v4.16b,v29.16b	// A[0][4]=A[1][3]
 
-	////////////////////////////////////////////////// Chi+Iota
-	dup	v21.2d,x11				// borrow C[6]
-.inst	0xce22180f	//bcax v15.16b,   v0.16b,v2.16b,v6.16b	// *
-.inst	0xce2a08c1	//bcax v1.16b,v6.16b,   v10.16b,   v2.16b	// *
-.inst	0xce242842	//bcax v2.16b,v2.16b,v4.16b,v10.16b
-.inst	0xce201143	//bcax v3.16b,v10.16b,   v0.16b,v4.16b
-.inst	0xce260084	//bcax v4.16b,v4.16b,v6.16b,   v0.16b
-
-.inst	0xce276565	//bcax v5.16b,v11.16b,   v7.16b,v25.16b	// *
-.inst	0xce281f26	//bcax v6.16b,   v25.16b,v8.16b,v7.16b	// *
-.inst	0xce2920e7	//bcax v7.16b,v7.16b,v9.16b,v8.16b
-.inst	0xce2b2508	//bcax v8.16b,v8.16b,v11.16b,   v9.16b
-.inst	0xce392d29	//bcax v9.16b,v9.16b,v25.16b,v11.16b
-
-	eor	v0.16b,v15.16b,v21.16b			// Iota
-
-.inst	0xce2c6f4a	//bcax v10.16b,   v26.16b,v12.16b,v27.16b	// *
-.inst	0xce2d336b	//bcax v11.16b,   v27.16b,v13.16b,v12.16b	// *
-.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
-.inst	0xce3a39ad	//bcax v13.16b,v13.16b,v26.16b,v14.16b
-.inst	0xce3b69ce	//bcax v14.16b,v14.16b,v27.16b,v26.16b
+.inst	0xce207363	//bcax v3.16b,v27.16b,   v0.16b,v28.16b
+.inst	0xce210384	//bcax v4.16b,v28.16b,   v1.16b,v0.16b
+.inst	0xce220400	//bcax v0.16b,v0.16b,v2.16b,v1.16b
+.inst	0xce3b0821	//bcax v1.16b,v1.16b,v27.16b,   v2.16b
+.inst	0xce3c6c42	//bcax v2.16b,v2.16b,v28.16b,   v27.16b
 
-.inst	0xce31778f	//bcax v15.16b,   v28.16b,v17.16b,v29.16b	// *
-.inst	0xce3247b0	//bcax v16.16b,   v29.16b,v18.16b,v17.16b	// *
-.inst	0xce334a31	//bcax v17.16b,v17.16b,v19.16b,v18.16b
-.inst	0xce3c4e52	//bcax v18.16b,v18.16b,v28.16b,v19.16b
-.inst	0xce3d7273	//bcax v19.16b,v19.16b,v29.16b,v28.16b
+	eor	v0.16b,v0.16b,v26.16b
 
-.inst	0xce367fd4	//bcax v20.16b,   v30.16b,v22.16b,v31.16b	// *
-.inst	0xce375bf5	//bcax v21.16b,   v31.16b,v23.16b,v22.16b	// *
-.inst	0xce385ed6	//bcax v22.16b,v22.16b,v24.16b,v23.16b
-.inst	0xce3e62f7	//bcax v23.16b,v23.16b,v30.16b,v24.16b
-.inst	0xce3f7b18	//bcax v24.16b,v24.16b,v31.16b,v30.16b
 	subs	x9,x9,#1
 	bne	.Loop_ce
 
diff --git a/sys/crypto/openssl/aarch64/poly1305-armv8.S b/sys/crypto/openssl/aarch64/poly1305-armv8.S
index 8ea86cb222fa6..101e4fb40552d 100644
--- a/sys/crypto/openssl/aarch64/poly1305-armv8.S
+++ b/sys/crypto/openssl/aarch64/poly1305-armv8.S
@@ -23,17 +23,12 @@ poly1305_init:
 	csel	x0,xzr,x0,eq
 	b.eq	.Lno_key
 
-#ifdef	__ILP32__
-	ldrsw	x11,.LOPENSSL_armcap_P
-#else
-	ldr	x11,.LOPENSSL_armcap_P
-#endif
-	adr	x10,.LOPENSSL_armcap_P
+	adrp	x17,OPENSSL_armcap_P
+	ldr	w17,[x17,#:lo12:OPENSSL_armcap_P]
 
 	ldp	x7,x8,[x1]		// load key
 	mov	x9,#0xfffffffc0fffffff
 	movk	x9,#0x0fff,lsl#48
-	ldr	w17,[x10,x11]
 #ifdef	__ARMEB__
 	rev	x7,x7			// flip bytes
 	rev	x8,x8
@@ -45,10 +40,10 @@ poly1305_init:
 
 	tst	w17,#ARMV7_NEON
 
-	adr	x12,poly1305_blocks
-	adr	x7,poly1305_blocks_neon
-	adr	x13,poly1305_emit
-	adr	x8,poly1305_emit_neon
+	adr	x12,.Lpoly1305_blocks
+	adr	x7,.Lpoly1305_blocks_neon
+	adr	x13,.Lpoly1305_emit
+	adr	x8,.Lpoly1305_emit_neon
 
 	csel	x12,x12,x7,eq
 	csel	x13,x13,x8,eq
@@ -67,6 +62,7 @@ poly1305_init:
 .type	poly1305_blocks,%function
 .align	5
 poly1305_blocks:
+.Lpoly1305_blocks:
 	ands	x2,x2,#-16
 	b.eq	.Lno_data
 
@@ -131,6 +127,7 @@ poly1305_blocks:
 .type	poly1305_emit,%function
 .align	5
 poly1305_emit:
+.Lpoly1305_emit:
 	ldp	x4,x5,[x0]		// load hash base 2^64
 	ldr	x6,[x0,#16]
 	ldp	x10,x11,[x2]	// load nonce
@@ -225,10 +222,11 @@ poly1305_splat:
 .type	poly1305_blocks_neon,%function
 .align	5
 poly1305_blocks_neon:
+.Lpoly1305_blocks_neon:
 	ldr	x17,[x0,#24]
 	cmp	x2,#128
 	b.hs	.Lblocks_neon
-	cbz	x17,poly1305_blocks
+	cbz	x17,.Lpoly1305_blocks
 
 .Lblocks_neon:
 .inst	0xd503233f		// paciasp
@@ -371,7 +369,7 @@ poly1305_blocks_neon:
 	csel	x16,x17,x16,lo
 
 	mov	x4,#1
-	str	x4,[x0,#-24]		// set is_base2_26
+	stur	x4,[x0,#-24]		// set is_base2_26
 	sub	x0,x0,#48		// restore original x0
 	b	.Ldo_neon
 
@@ -808,6 +806,7 @@ poly1305_blocks_neon:
 .type	poly1305_emit_neon,%function
 .align	5
 poly1305_emit_neon:
+.Lpoly1305_emit_neon:
 	ldr	x17,[x0,#24]
 	cbz	x17,poly1305_emit
 
@@ -860,12 +859,6 @@ poly1305_emit_neon:
 .align	5
 .Lzeros:
 .long	0,0,0,0,0,0,0,0
-.LOPENSSL_armcap_P:
-#ifdef	__ILP32__
-.long	OPENSSL_armcap_P-.
-#else
-.quad	OPENSSL_armcap_P-.
-#endif
 .byte	80,111,108,121,49,51,48,53,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
 .align	2
diff --git a/sys/crypto/openssl/aarch64/sha1-armv8.S b/sys/crypto/openssl/aarch64/sha1-armv8.S
index 847de3818ea66..a9e1d81d58177 100644
--- a/sys/crypto/openssl/aarch64/sha1-armv8.S
+++ b/sys/crypto/openssl/aarch64/sha1-armv8.S
@@ -1,22 +1,18 @@
 /* Do not modify. This file is auto-generated from sha1-armv8.pl. */
-#include "arm_arch.h"
+#ifndef	__KERNEL__
+# include "arm_arch.h"
 
-.text
+.hidden	OPENSSL_armcap_P
+#endif
 
+.text
 
-.hidden	OPENSSL_armcap_P
 .globl	sha1_block_data_order
 .type	sha1_block_data_order,%function
 .align	6
 sha1_block_data_order:
-#ifdef	__ILP32__
-	ldrsw	x16,.LOPENSSL_armcap_P
-#else
-	ldr	x16,.LOPENSSL_armcap_P
-#endif
-	adr	x17,.LOPENSSL_armcap_P
-	add	x16,x16,x17
-	ldr	w16,[x16]
+	adrp	x16,OPENSSL_armcap_P
+	ldr	w16,[x16,#:lo12:OPENSSL_armcap_P]
 	tst	w16,#ARMV8_SHA1
 	b.ne	.Lv8_entry
 
@@ -37,7 +33,7 @@ sha1_block_data_order:
 	movz	w28,#0x7999
 	sub	x2,x2,#1
 	movk	w28,#0x5a82,lsl#16
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x3,x3,#32
 #else
 	rev32	x3,x3
@@ -45,7 +41,7 @@ sha1_block_data_order:
 	add	w24,w24,w28		// warm it up
 	add	w24,w24,w3
 	lsr	x4,x3,#32
-	ldr	x5,[x1,#-56]
+	ldur	x5,[x1,#-56]
 	bic	w25,w23,w21
 	and	w26,w22,w21
 	ror	w27,w20,#27
@@ -55,7 +51,7 @@ sha1_block_data_order:
 	ror	w21,w21,#2
 	add	w23,w23,w4	// future e+=X[i]
 	add	w24,w24,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x5,x5,#32
 #else
 	rev32	x5,x5
@@ -70,7 +66,7 @@ sha1_block_data_order:
 	add	w22,w22,w5	// future e+=X[i]
 	add	w23,w23,w25		// e+=F(b,c,d)
 	lsr	x6,x5,#32
-	ldr	x7,[x1,#-48]
+	ldur	x7,[x1,#-48]
 	bic	w25,w21,w24
 	and	w26,w20,w24
 	ror	w27,w23,#27
@@ -80,7 +76,7 @@ sha1_block_data_order:
 	ror	w24,w24,#2
 	add	w21,w21,w6	// future e+=X[i]
 	add	w22,w22,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x7,x7,#32
 #else
 	rev32	x7,x7
@@ -95,7 +91,7 @@ sha1_block_data_order:
 	add	w20,w20,w7	// future e+=X[i]
 	add	w21,w21,w25		// e+=F(b,c,d)
 	lsr	x8,x7,#32
-	ldr	x9,[x1,#-40]
+	ldur	x9,[x1,#-40]
 	bic	w25,w24,w22
 	and	w26,w23,w22
 	ror	w27,w21,#27
@@ -105,7 +101,7 @@ sha1_block_data_order:
 	ror	w22,w22,#2
 	add	w24,w24,w8	// future e+=X[i]
 	add	w20,w20,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x9,x9,#32
 #else
 	rev32	x9,x9
@@ -120,7 +116,7 @@ sha1_block_data_order:
 	add	w23,w23,w9	// future e+=X[i]
 	add	w24,w24,w25		// e+=F(b,c,d)
 	lsr	x10,x9,#32
-	ldr	x11,[x1,#-32]
+	ldur	x11,[x1,#-32]
 	bic	w25,w22,w20
 	and	w26,w21,w20
 	ror	w27,w24,#27
@@ -130,7 +126,7 @@ sha1_block_data_order:
 	ror	w20,w20,#2
 	add	w22,w22,w10	// future e+=X[i]
 	add	w23,w23,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x11,x11,#32
 #else
 	rev32	x11,x11
@@ -145,7 +141,7 @@ sha1_block_data_order:
 	add	w21,w21,w11	// future e+=X[i]
 	add	w22,w22,w25		// e+=F(b,c,d)
 	lsr	x12,x11,#32
-	ldr	x13,[x1,#-24]
+	ldur	x13,[x1,#-24]
 	bic	w25,w20,w23
 	and	w26,w24,w23
 	ror	w27,w22,#27
@@ -155,7 +151,7 @@ sha1_block_data_order:
 	ror	w23,w23,#2
 	add	w20,w20,w12	// future e+=X[i]
 	add	w21,w21,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x13,x13,#32
 #else
 	rev32	x13,x13
@@ -170,7 +166,7 @@ sha1_block_data_order:
 	add	w24,w24,w13	// future e+=X[i]
 	add	w20,w20,w25		// e+=F(b,c,d)
 	lsr	x14,x13,#32
-	ldr	x15,[x1,#-16]
+	ldur	x15,[x1,#-16]
 	bic	w25,w23,w21
 	and	w26,w22,w21
 	ror	w27,w20,#27
@@ -180,7 +176,7 @@ sha1_block_data_order:
 	ror	w21,w21,#2
 	add	w23,w23,w14	// future e+=X[i]
 	add	w24,w24,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x15,x15,#32
 #else
 	rev32	x15,x15
@@ -195,7 +191,7 @@ sha1_block_data_order:
 	add	w22,w22,w15	// future e+=X[i]
 	add	w23,w23,w25		// e+=F(b,c,d)
 	lsr	x16,x15,#32
-	ldr	x17,[x1,#-8]
+	ldur	x17,[x1,#-8]
 	bic	w25,w21,w24
 	and	w26,w20,w24
 	ror	w27,w23,#27
@@ -205,7 +201,7 @@ sha1_block_data_order:
 	ror	w24,w24,#2
 	add	w21,w21,w16	// future e+=X[i]
 	add	w22,w22,w25		// e+=F(b,c,d)
-#ifdef	__ARMEB__
+#ifdef	__AARCH64EB__
 	ror	x17,x17,#32
 #else
 	rev32	x17,x17
@@ -1211,12 +1207,6 @@ sha1_block_armv8:
 .long	0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1	//K_20_39
 .long	0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc	//K_40_59
 .long	0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6	//K_60_79
-.LOPENSSL_armcap_P:
-#ifdef	__ILP32__
-.long	OPENSSL_armcap_P-.
-#else
-.quad	OPENSSL_armcap_P-.
-#endif
 .byte	83,72,65,49,32,98,108,111,99,107,32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
 .align	2
diff --git a/sys/crypto/openssl/aarch64/sha256-armv8.S b/sys/crypto/openssl/aarch64/sha256-armv8.S
index e3be48716e07d..f0c4a14610b8c 100644
--- a/sys/crypto/openssl/aarch64/sha256-armv8.S
+++ b/sys/crypto/openssl/aarch64/sha256-armv8.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from sha512-armv8.pl. */
 // Copyright 2014-2020 The OpenSSL Project Authors. All Rights Reserved.
 //
-// Licensed under the OpenSSL license (the "License").  You may not use
+// Licensed under the Apache License 2.0 (the "License").  You may not use
 // this file except in compliance with the License.  You can obtain a copy
 // in the file LICENSE in the source distribution or at
 // https://www.openssl.org/source/license.html
@@ -28,6 +28,7 @@
 // X-Gene			20.0 (+100%)	12.8 (+300%(***))
 // Mongoose	2.36		13.0 (+50%)	8.36 (+33%)
 // Kryo		1.92		17.4 (+30%)	11.2 (+8%)
+// ThunderX2	2.54		13.2 (+40%)	8.40 (+18%)
 //
 // (*)	Software SHA256 results are of lesser relevance, presented
 //	mostly for informational purposes.
@@ -53,27 +54,23 @@
 // deliver much less improvement, likely *negative* on Cortex-A5x.
 // Which is why NEON support is limited to SHA256.]
 
+// $output is the last argument if it looks like a file (it has an extension)
+// $flavour is the first argument if it doesn't look like a file
 #ifndef	__KERNEL__
 # include "arm_arch.h"
+
+.hidden	OPENSSL_armcap_P
 #endif
 
 .text
 
-
-.hidden	OPENSSL_armcap_P
 .globl	sha256_block_data_order
 .type	sha256_block_data_order,%function
 .align	6
 sha256_block_data_order:
 #ifndef	__KERNEL__
-# ifdef	__ILP32__
-	ldrsw	x16,.LOPENSSL_armcap_P
-# else
-	ldr	x16,.LOPENSSL_armcap_P
-# endif
-	adr	x17,.LOPENSSL_armcap_P
-	add	x16,x16,x17
-	ldr	w16,[x16]
+	adrp	x16,OPENSSL_armcap_P
+	ldr	w16,[x16,#:lo12:OPENSSL_armcap_P]
 	tst	w16,#ARMV8_SHA256
 	b.ne	.Lv8_entry
 	tst	w16,#ARMV7_NEON
@@ -1064,15 +1061,6 @@ sha256_block_data_order:
 .long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
 .long	0	//terminator
 .size	.LK256,.-.LK256
-#ifndef	__KERNEL__
-.align	3
-.LOPENSSL_armcap_P:
-# ifdef	__ILP32__
-.long	OPENSSL_armcap_P-.
-# else
-.quad	OPENSSL_armcap_P-.
-# endif
-#endif
 .byte	83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
 .align	2
diff --git a/sys/crypto/openssl/aarch64/sha512-armv8.S b/sys/crypto/openssl/aarch64/sha512-armv8.S
index 2103672a35ff8..ded34d2a96b0d 100644
--- a/sys/crypto/openssl/aarch64/sha512-armv8.S
+++ b/sys/crypto/openssl/aarch64/sha512-armv8.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from sha512-armv8.pl. */
 // Copyright 2014-2020 The OpenSSL Project Authors. All Rights Reserved.
 //
-// Licensed under the OpenSSL license (the "License").  You may not use
+// Licensed under the Apache License 2.0 (the "License").  You may not use
 // this file except in compliance with the License.  You can obtain a copy
 // in the file LICENSE in the source distribution or at
 // https://www.openssl.org/source/license.html
@@ -28,6 +28,7 @@
 // X-Gene			20.0 (+100%)	12.8 (+300%(***))
 // Mongoose	2.36		13.0 (+50%)	8.36 (+33%)
 // Kryo		1.92		17.4 (+30%)	11.2 (+8%)
+// ThunderX2	2.54		13.2 (+40%)	8.40 (+18%)
 //
 // (*)	Software SHA256 results are of lesser relevance, presented
 //	mostly for informational purposes.
@@ -53,27 +54,23 @@
 // deliver much less improvement, likely *negative* on Cortex-A5x.
 // Which is why NEON support is limited to SHA256.]
 
+// $output is the last argument if it looks like a file (it has an extension)
+// $flavour is the first argument if it doesn't look like a file
 #ifndef	__KERNEL__
 # include "arm_arch.h"
+
+.hidden	OPENSSL_armcap_P
 #endif
 
 .text
 
-
-.hidden	OPENSSL_armcap_P
 .globl	sha512_block_data_order
 .type	sha512_block_data_order,%function
 .align	6
 sha512_block_data_order:
 #ifndef	__KERNEL__
-# ifdef	__ILP32__
-	ldrsw	x16,.LOPENSSL_armcap_P
-# else
-	ldr	x16,.LOPENSSL_armcap_P
-# endif
-	adr	x17,.LOPENSSL_armcap_P
-	add	x16,x16,x17
-	ldr	w16,[x16]
+	adrp	x16,OPENSSL_armcap_P
+	ldr	w16,[x16,#:lo12:OPENSSL_armcap_P]
 	tst	w16,#ARMV8_SHA512
 	b.ne	.Lv8_entry
 #endif
@@ -1086,15 +1083,6 @@ sha512_block_data_order:
 .quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
 .quad	0	// terminator
 .size	.LK512,.-.LK512
-#ifndef	__KERNEL__
-.align	3
-.LOPENSSL_armcap_P:
-# ifdef	__ILP32__
-.long	OPENSSL_armcap_P-.
-# else
-.quad	OPENSSL_armcap_P-.
-# endif
-#endif
 .byte	83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
 .align	2
diff --git a/sys/crypto/openssl/aarch64/vpaes-armv8.S b/sys/crypto/openssl/aarch64/vpaes-armv8.S
index 7ec5b5a83538e..07c08bfe7daac 100644
--- a/sys/crypto/openssl/aarch64/vpaes-armv8.S
+++ b/sys/crypto/openssl/aarch64/vpaes-armv8.S
@@ -91,12 +91,12 @@ _vpaes_consts:
 .align	2
 .size	_vpaes_consts,.-_vpaes_consts
 .align	6
-##
-##  _aes_preheat
-##
-##  Fills register %r10 -> .aes_consts (so you can -fPIC)
-##  and %xmm9-%xmm15 as specified below.
-##
+//
+//  _aes_preheat
+//
+//  Fills register %r10 -> .aes_consts (so you can -fPIC)
+//  and %xmm9-%xmm15 as specified below.
+//
 .type	_vpaes_encrypt_preheat,%function
 .align	4
 _vpaes_encrypt_preheat:
@@ -108,21 +108,21 @@ _vpaes_encrypt_preheat:
 	ret
 .size	_vpaes_encrypt_preheat,.-_vpaes_encrypt_preheat
 
-##
-##  _aes_encrypt_core
-##
-##  AES-encrypt %xmm0.
-##
-##  Inputs:
-##     %xmm0 = input
-##     %xmm9-%xmm15 as in _vpaes_preheat
-##    (%rdx) = scheduled keys
-##
-##  Output in %xmm0
-##  Clobbers  %xmm1-%xmm5, %r9, %r10, %r11, %rax
-##  Preserves %xmm6 - %xmm8 so you get some local vectors
-##
-##
+//
+//  _aes_encrypt_core
+//
+//  AES-encrypt %xmm0.
+//
+//  Inputs:
+//     %xmm0 = input
+//     %xmm9-%xmm15 as in _vpaes_preheat
+//    (%rdx) = scheduled keys
+//
+//  Output in %xmm0
+//  Clobbers  %xmm1-%xmm5, %r9, %r10, %r11, %rax
+//  Preserves %xmm6 - %xmm8 so you get some local vectors
+//
+//
 .type	_vpaes_encrypt_core,%function
 .align	4
 _vpaes_encrypt_core:
@@ -328,11 +328,11 @@ _vpaes_decrypt_preheat:
 	ret
 .size	_vpaes_decrypt_preheat,.-_vpaes_decrypt_preheat
 
-##
-##  Decryption core
-##
-##  Same API as encryption core.
-##
+//
+//  Decryption core
+//
+//  Same API as encryption core.
+//
 .type	_vpaes_decrypt_core,%function
 .align	4
 _vpaes_decrypt_core:
@@ -577,11 +577,11 @@ _vpaes_decrypt_2x:
 	tbl	v1.16b,  {v8.16b},v2.16b
 	ret
 .size	_vpaes_decrypt_2x,.-_vpaes_decrypt_2x
-########################################################
-##                                                    ##
-##                  AES key schedule                  ##
-##                                                    ##
-########################################################
+////////////////////////////////////////////////////////
+//                                                    //
+//                  AES key schedule                  //
+//                                                    //
+////////////////////////////////////////////////////////
 .type	_vpaes_key_preheat,%function
 .align	4
 _vpaes_key_preheat:
@@ -637,14 +637,14 @@ _vpaes_schedule_core:
 	b.eq	.Lschedule_192
 	// 128: fall though
 
-##
-##  .schedule_128
-##
-##  128-bit specific part of key schedule.
-##
-##  This schedule is really simple, because all its parts
-##  are accomplished by the subroutines.
-##
+//
+//  .schedule_128
+//
+//  128-bit specific part of key schedule.
+//
+//  This schedule is really simple, because all its parts
+//  are accomplished by the subroutines.
+//
 .Lschedule_128:
 	mov	x0, #10			// mov	$10, %esi
 
@@ -655,21 +655,21 @@ _vpaes_schedule_core:
 	bl	_vpaes_schedule_mangle		// write output
 	b	.Loop_schedule_128
 
-##
-##  .aes_schedule_192
-##
-##  192-bit specific part of key schedule.
-##
-##  The main body of this schedule is the same as the 128-bit
-##  schedule, but with more smearing.  The long, high side is
-##  stored in %xmm7 as before, and the short, low side is in
-##  the high bits of %xmm6.
-##
-##  This schedule is somewhat nastier, however, because each
-##  round produces 192 bits of key material, or 1.5 round keys.
-##  Therefore, on each cycle we do 2 rounds and produce 3 round
-##  keys.
-##
+//
+//  .aes_schedule_192
+//
+//  192-bit specific part of key schedule.
+//
+//  The main body of this schedule is the same as the 128-bit
+//  schedule, but with more smearing.  The long, high side is
+//  stored in %xmm7 as before, and the short, low side is in
+//  the high bits of %xmm6.
+//
+//  This schedule is somewhat nastier, however, because each
+//  round produces 192 bits of key material, or 1.5 round keys.
+//  Therefore, on each cycle we do 2 rounds and produce 3 round
+//  keys.
+//
 .align	4
 .Lschedule_192:
 	sub	x0, x0, #8
@@ -693,16 +693,16 @@ _vpaes_schedule_core:
 	bl	_vpaes_schedule_192_smear
 	b	.Loop_schedule_192
 
-##
-##  .aes_schedule_256
-##
-##  256-bit specific part of key schedule.
-##
-##  The structure here is very similar to the 128-bit
-##  schedule, but with an additional "low side" in
-##  %xmm6.  The low side's rounds are the same as the
-##  high side's, except no rcon and no rotation.
-##
+//
+//  .aes_schedule_256
+//
+//  256-bit specific part of key schedule.
+//
+//  The structure here is very similar to the 128-bit
+//  schedule, but with an additional "low side" in
+//  %xmm6.  The low side's rounds are the same as the
+//  high side's, except no rcon and no rotation.
+//
 .align	4
 .Lschedule_256:
 	ld1	{v0.16b}, [x0]		// vmovdqu	16(%rdi),%xmm0		# load key part 2 (unaligned)
@@ -729,16 +729,16 @@ _vpaes_schedule_core:
 
 	b	.Loop_schedule_256
 
-##
-##  .aes_schedule_mangle_last
-##
-##  Mangler for last round of key schedule
-##  Mangles %xmm0
-##    when encrypting, outputs out(%xmm0) ^ 63
-##    when decrypting, outputs unskew(%xmm0)
-##
-##  Always called right before return... jumps to cleanup and exits
-##
+//
+//  .aes_schedule_mangle_last
+//
+//  Mangler for last round of key schedule
+//  Mangles %xmm0
+//    when encrypting, outputs out(%xmm0) ^ 63
+//    when decrypting, outputs unskew(%xmm0)
+//
+//  Always called right before return... jumps to cleanup and exits
+//
 .align	4
 .Lschedule_mangle_last:
 	// schedule last round key from xmm0
@@ -772,20 +772,20 @@ _vpaes_schedule_core:
 	ret
 .size	_vpaes_schedule_core,.-_vpaes_schedule_core
 
-##
-##  .aes_schedule_192_smear
-##
-##  Smear the short, low side in the 192-bit key schedule.
-##
-##  Inputs:
-##    %xmm7: high side, b  a  x  y
-##    %xmm6:  low side, d  c  0  0
-##    %xmm13: 0
-##
-##  Outputs:
-##    %xmm6: b+c+d  b+c  0  0
-##    %xmm0: b+c+d  b+c  b  a
-##
+//
+//  .aes_schedule_192_smear
+//
+//  Smear the short, low side in the 192-bit key schedule.
+//
+//  Inputs:
+//    %xmm7: high side, b  a  x  y
+//    %xmm6:  low side, d  c  0  0
+//    %xmm13: 0
+//
+//  Outputs:
+//    %xmm6: b+c+d  b+c  0  0
+//    %xmm0: b+c+d  b+c  b  a
+//
 .type	_vpaes_schedule_192_smear,%function
 .align	4
 _vpaes_schedule_192_smear:
@@ -801,24 +801,24 @@ _vpaes_schedule_192_smear:
 	ret
 .size	_vpaes_schedule_192_smear,.-_vpaes_schedule_192_smear
 
-##
-##  .aes_schedule_round
-##
-##  Runs one main round of the key schedule on %xmm0, %xmm7
-##
-##  Specifically, runs subbytes on the high dword of %xmm0
-##  then rotates it by one byte and xors into the low dword of
-##  %xmm7.
-##
-##  Adds rcon from low byte of %xmm8, then rotates %xmm8 for
-##  next rcon.
-##
-##  Smears the dwords of %xmm7 by xoring the low into the
-##  second low, result into third, result into highest.
-##
-##  Returns results in %xmm7 = %xmm0.
-##  Clobbers %xmm1-%xmm4, %r11.
-##
+//
+//  .aes_schedule_round
+//
+//  Runs one main round of the key schedule on %xmm0, %xmm7
+//
+//  Specifically, runs subbytes on the high dword of %xmm0
+//  then rotates it by one byte and xors into the low dword of
+//  %xmm7.
+//
+//  Adds rcon from low byte of %xmm8, then rotates %xmm8 for
+//  next rcon.
+//
+//  Smears the dwords of %xmm7 by xoring the low into the
+//  second low, result into third, result into highest.
+//
+//  Returns results in %xmm7 = %xmm0.
+//  Clobbers %xmm1-%xmm4, %r11.
+//
 .type	_vpaes_schedule_round,%function
 .align	4
 _vpaes_schedule_round:
@@ -866,15 +866,15 @@ _vpaes_schedule_low_round:
 	ret
 .size	_vpaes_schedule_round,.-_vpaes_schedule_round
 
-##
-##  .aes_schedule_transform
-##
-##  Linear-transform %xmm0 according to tables at (%r11)
-##
-##  Requires that %xmm9 = 0x0F0F... as in preheat
-##  Output in %xmm0
-##  Clobbers %xmm1, %xmm2
-##
+//
+//  .aes_schedule_transform
+//
+//  Linear-transform %xmm0 according to tables at (%r11)
+//
+//  Requires that %xmm9 = 0x0F0F... as in preheat
+//  Output in %xmm0
+//  Clobbers %xmm1, %xmm2
+//
 .type	_vpaes_schedule_transform,%function
 .align	4
 _vpaes_schedule_transform:
@@ -888,29 +888,29 @@ _vpaes_schedule_transform:
 	ret
 .size	_vpaes_schedule_transform,.-_vpaes_schedule_transform
 
-##
-##  .aes_schedule_mangle
-##
-##  Mangle xmm0 from (basis-transformed) standard version
-##  to our version.
-##
-##  On encrypt,
-##    xor with 0x63
-##    multiply by circulant 0,1,1,1
-##    apply shiftrows transform
-##
-##  On decrypt,
-##    xor with 0x63
-##    multiply by "inverse mixcolumns" circulant E,B,D,9
-##    deskew
-##    apply shiftrows transform
-##
-##
-##  Writes out to (%rdx), and increments or decrements it
-##  Keeps track of round number mod 4 in %r8
-##  Preserves xmm0
-##  Clobbers xmm1-xmm5
-##
+//
+//  .aes_schedule_mangle
+//
+//  Mangle xmm0 from (basis-transformed) standard version
+//  to our version.
+//
+//  On encrypt,
+//    xor with 0x63
+//    multiply by circulant 0,1,1,1
+//    apply shiftrows transform
+//
+//  On decrypt,
+//    xor with 0x63
+//    multiply by "inverse mixcolumns" circulant E,B,D,9
+//    deskew
+//    apply shiftrows transform
+//
+//
+//  Writes out to (%rdx), and increments or decrements it
+//  Keeps track of round number mod 4 in %r8
+//  Preserves xmm0
+//  Clobbers xmm1-xmm5
+//
 .type	_vpaes_schedule_mangle,%function
 .align	4
 _vpaes_schedule_mangle:
diff --git a/sys/crypto/openssl/arm/aes-armv4.S b/sys/crypto/openssl/arm/aes-armv4.S
index c0c01485aaf92..723b54f4374aa 100644
--- a/sys/crypto/openssl/arm/aes-armv4.S
+++ b/sys/crypto/openssl/arm/aes-armv4.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from aes-armv4.pl. */
 @ Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
 @
-@ Licensed under the OpenSSL license (the "License").  You may not use
+@ Licensed under the Apache License 2.0 (the "License").  You may not use
 @ this file except in compliance with the License.  You can obtain a copy
 @ in the file LICENSE in the source distribution or at
 @ https://www.openssl.org/source/license.html
@@ -39,13 +39,14 @@
 @ Profiler-assisted and platform-specific optimization resulted in 16%
 @ improvement on Cortex A8 core and ~21.5 cycles per byte.
 
+@ $output is the last argument if it looks like a file (it has an extension)
+@ $flavour is the first argument if it doesn't look like a file
 #ifndef __KERNEL__
 # include "arm_arch.h"
 #else
 # define __ARM_ARCH__ __LINUX_ARM_ARCH__
 #endif
 
-.text
 #if defined(__thumb2__) && !defined(__APPLE__)
 .syntax	unified
 .thumb
@@ -54,6 +55,8 @@
 #undef __thumb2__
 #endif
 
+.text
+
 .type	AES_Te,%object
 .align	5
 AES_Te:
diff --git a/sys/crypto/openssl/arm/aesv8-armx.S b/sys/crypto/openssl/arm/aesv8-armx.S
index 569f1c6ba4001..b59badae62afb 100644
--- a/sys/crypto/openssl/arm/aesv8-armx.S
+++ b/sys/crypto/openssl/arm/aesv8-armx.S
@@ -2,11 +2,18 @@
 #include "arm_arch.h"
 
 #if __ARM_MAX_ARCH__>=7
-.text
 .arch	armv7-a	@ don't confuse not-so-latest binutils with argv8 :-)
 .fpu	neon
+#ifdef	__thumb2__
+.syntax	unified
+.thumb
+# define INST(a,b,c,d)	.byte	c,d|0xc,a,b
+#else
 .code	32
-#undef	__thumb2__
+# define INST(a,b,c,d)	.byte	a,b,c,d
+#endif
+
+.text
 .align	5
 .Lrcon:
 .long	0x01,0x01,0x01,0x01
@@ -49,7 +56,7 @@ aes_v8_set_encrypt_key:
 	vtbl.8	d21,{q3},d5
 	vext.8	q9,q0,q3,#12
 	vst1.32	{q3},[r2]!
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 	subs	r1,r1,#1
 
 	veor	q3,q3,q9
@@ -68,7 +75,7 @@ aes_v8_set_encrypt_key:
 	vtbl.8	d21,{q3},d5
 	vext.8	q9,q0,q3,#12
 	vst1.32	{q3},[r2]!
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 
 	veor	q3,q3,q9
 	vext.8	q9,q0,q9,#12
@@ -83,7 +90,7 @@ aes_v8_set_encrypt_key:
 	vtbl.8	d21,{q3},d5
 	vext.8	q9,q0,q3,#12
 	vst1.32	{q3},[r2]!
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 
 	veor	q3,q3,q9
 	vext.8	q9,q0,q9,#12
@@ -115,7 +122,7 @@ aes_v8_set_encrypt_key:
 #else
 	vst1.32	{d16},[r2]!
 #endif
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 	subs	r1,r1,#1
 
 	veor	q3,q3,q9
@@ -151,7 +158,7 @@ aes_v8_set_encrypt_key:
 	vtbl.8	d21,{q8},d5
 	vext.8	q9,q0,q3,#12
 	vst1.32	{q8},[r2]!
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 	subs	r1,r1,#1
 
 	veor	q3,q3,q9
@@ -167,7 +174,7 @@ aes_v8_set_encrypt_key:
 
 	vdup.32	q10,d7[1]
 	vext.8	q9,q0,q8,#12
-.byte	0x00,0x43,0xf0,0xf3	@ aese q10,q0
+	INST(0x00,0x43,0xf0,0xf3)	@ aese q10,q0
 
 	veor	q8,q8,q9
 	vext.8	q9,q0,q9,#12
@@ -210,15 +217,15 @@ aes_v8_set_decrypt_key:
 .Loop_imc:
 	vld1.32	{q0},[r2]
 	vld1.32	{q1},[r0]
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
 	vst1.32	{q0},[r0],r4
 	vst1.32	{q1},[r2]!
 	cmp	r0,r2
 	bhi	.Loop_imc
 
 	vld1.32	{q0},[r2]
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
 	vst1.32	{q0},[r0]
 
 	eor	r0,r0,r0		@ return value
@@ -236,19 +243,19 @@ aes_v8_encrypt:
 	vld1.32	{q1},[r2]!
 
 .Loop_enc:
-.byte	0x00,0x43,0xb0,0xf3	@ aese q2,q0
-.byte	0x84,0x43,0xb0,0xf3	@ aesmc q2,q2
+	INST(0x00,0x43,0xb0,0xf3)	@ aese q2,q0
+	INST(0x84,0x43,0xb0,0xf3)	@ aesmc q2,q2
 	vld1.32	{q0},[r2]!
 	subs	r3,r3,#2
-.byte	0x02,0x43,0xb0,0xf3	@ aese q2,q1
-.byte	0x84,0x43,0xb0,0xf3	@ aesmc q2,q2
+	INST(0x02,0x43,0xb0,0xf3)	@ aese q2,q1
+	INST(0x84,0x43,0xb0,0xf3)	@ aesmc q2,q2
 	vld1.32	{q1},[r2]!
 	bgt	.Loop_enc
 
-.byte	0x00,0x43,0xb0,0xf3	@ aese q2,q0
-.byte	0x84,0x43,0xb0,0xf3	@ aesmc q2,q2
+	INST(0x00,0x43,0xb0,0xf3)	@ aese q2,q0
+	INST(0x84,0x43,0xb0,0xf3)	@ aesmc q2,q2
 	vld1.32	{q0},[r2]
-.byte	0x02,0x43,0xb0,0xf3	@ aese q2,q1
+	INST(0x02,0x43,0xb0,0xf3)	@ aese q2,q1
 	veor	q2,q2,q0
 
 	vst1.8	{q2},[r1]
@@ -265,24 +272,336 @@ aes_v8_decrypt:
 	vld1.32	{q1},[r2]!
 
 .Loop_dec:
-.byte	0x40,0x43,0xb0,0xf3	@ aesd q2,q0
-.byte	0xc4,0x43,0xb0,0xf3	@ aesimc q2,q2
+	INST(0x40,0x43,0xb0,0xf3)	@ aesd q2,q0
+	INST(0xc4,0x43,0xb0,0xf3)	@ aesimc q2,q2
 	vld1.32	{q0},[r2]!
 	subs	r3,r3,#2
-.byte	0x42,0x43,0xb0,0xf3	@ aesd q2,q1
-.byte	0xc4,0x43,0xb0,0xf3	@ aesimc q2,q2
+	INST(0x42,0x43,0xb0,0xf3)	@ aesd q2,q1
+	INST(0xc4,0x43,0xb0,0xf3)	@ aesimc q2,q2
 	vld1.32	{q1},[r2]!
 	bgt	.Loop_dec
 
-.byte	0x40,0x43,0xb0,0xf3	@ aesd q2,q0
-.byte	0xc4,0x43,0xb0,0xf3	@ aesimc q2,q2
+	INST(0x40,0x43,0xb0,0xf3)	@ aesd q2,q0
+	INST(0xc4,0x43,0xb0,0xf3)	@ aesimc q2,q2
 	vld1.32	{q0},[r2]
-.byte	0x42,0x43,0xb0,0xf3	@ aesd q2,q1
+	INST(0x42,0x43,0xb0,0xf3)	@ aesd q2,q1
 	veor	q2,q2,q0
 
 	vst1.8	{q2},[r1]
 	bx	lr
 .size	aes_v8_decrypt,.-aes_v8_decrypt
+.globl	aes_v8_ecb_encrypt
+.type	aes_v8_ecb_encrypt,%function
+.align	5
+aes_v8_ecb_encrypt:
+	mov	ip,sp
+	stmdb	sp!,{r4,r5,r6,r7,r8,lr}
+	vstmdb	sp!,{d8,d9,d10,d11,d12,d13,d14,d15}			@ ABI specification says so
+	ldmia	ip,{r4,r5}			@ load remaining args
+	subs	r2,r2,#16
+	mov	r8,#16
+	blo	.Lecb_done
+	it	eq
+	moveq	r8,#0
+
+	cmp	r4,#0					@ en- or decrypting?
+	ldr	r5,[r3,#240]
+	and	r2,r2,#-16
+	vld1.8	{q0},[r0],r8
+
+	vld1.32	{q8,q9},[r3]				@ load key schedule...
+	sub	r5,r5,#6
+	add	r7,r3,r5,lsl#4				@ pointer to last 7 round keys
+	sub	r5,r5,#2
+	vld1.32	{q10,q11},[r7]!
+	vld1.32	{q12,q13},[r7]!
+	vld1.32	{q14,q15},[r7]!
+	vld1.32	{q7},[r7]
+
+	add	r7,r3,#32
+	mov	r6,r5
+	beq	.Lecb_dec
+
+	vld1.8	{q1},[r0]!
+	subs	r2,r2,#32				@ bias
+	add	r6,r5,#2
+	vorr	q3,q1,q1
+	vorr	q10,q1,q1
+	vorr	q1,q0,q0
+	blo	.Lecb_enc_tail
+
+	vorr	q1,q3,q3
+	vld1.8	{q10},[r0]!
+.Loop3x_ecb_enc:
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.32	{q8},[r7]!
+	subs	r6,r6,#2
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.32	{q9},[r7]!
+	bgt	.Loop3x_ecb_enc
+
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	subs	r2,r2,#0x30
+	it	lo
+	movlo	r6,r2				@ r6, r6, is zero at this point
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	add	r0,r0,r6			@ r0 is adjusted in such way that
+						@ at exit from the loop q1-q10
+						@ are loaded with last "words"
+	mov	r7,r3
+	INST(0x28,0x03,0xb0,0xf3)	@ aese q0,q12
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x28,0x23,0xb0,0xf3)	@ aese q1,q12
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x28,0x43,0xf0,0xf3)	@ aese q10,q12
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.8	{q2},[r0]!
+	INST(0x2a,0x03,0xb0,0xf3)	@ aese q0,q13
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2a,0x23,0xb0,0xf3)	@ aese q1,q13
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x2a,0x43,0xf0,0xf3)	@ aese q10,q13
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.8	{q3},[r0]!
+	INST(0x2c,0x03,0xb0,0xf3)	@ aese q0,q14
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2c,0x23,0xb0,0xf3)	@ aese q1,q14
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x2c,0x43,0xf0,0xf3)	@ aese q10,q14
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.8	{q11},[r0]!
+	INST(0x2e,0x03,0xb0,0xf3)	@ aese q0,q15
+	INST(0x2e,0x23,0xb0,0xf3)	@ aese q1,q15
+	INST(0x2e,0x43,0xf0,0xf3)	@ aese q10,q15
+	vld1.32	{q8},[r7]!		@ re-pre-load rndkey[0]
+	add	r6,r5,#2
+	veor	q4,q7,q0
+	veor	q5,q7,q1
+	veor	q10,q10,q7
+	vld1.32	{q9},[r7]!		@ re-pre-load rndkey[1]
+	vst1.8	{q4},[r1]!
+	vorr	q0,q2,q2
+	vst1.8	{q5},[r1]!
+	vorr	q1,q3,q3
+	vst1.8	{q10},[r1]!
+	vorr	q10,q11,q11
+	bhs	.Loop3x_ecb_enc
+
+	cmn	r2,#0x30
+	beq	.Lecb_done
+	nop
+
+.Lecb_enc_tail:
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.32	{q8},[r7]!
+	subs	r6,r6,#2
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	vld1.32	{q9},[r7]!
+	bgt	.Lecb_enc_tail
+
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	INST(0x28,0x23,0xb0,0xf3)	@ aese q1,q12
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x28,0x43,0xf0,0xf3)	@ aese q10,q12
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	cmn	r2,#0x20
+	INST(0x2a,0x23,0xb0,0xf3)	@ aese q1,q13
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x2a,0x43,0xf0,0xf3)	@ aese q10,q13
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	INST(0x2c,0x23,0xb0,0xf3)	@ aese q1,q14
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x2c,0x43,0xf0,0xf3)	@ aese q10,q14
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
+	INST(0x2e,0x23,0xb0,0xf3)	@ aese q1,q15
+	INST(0x2e,0x43,0xf0,0xf3)	@ aese q10,q15
+	beq	.Lecb_enc_one
+	veor	q5,q7,q1
+	veor	q9,q7,q10
+	vst1.8	{q5},[r1]!
+	vst1.8	{q9},[r1]!
+	b	.Lecb_done
+
+.Lecb_enc_one:
+	veor	q5,q7,q10
+	vst1.8	{q5},[r1]!
+	b	.Lecb_done
+.align	5
+.Lecb_dec:
+	vld1.8	{q1},[r0]!
+	subs	r2,r2,#32			@ bias
+	add	r6,r5,#2
+	vorr	q3,q1,q1
+	vorr	q10,q1,q1
+	vorr	q1,q0,q0
+	blo	.Lecb_dec_tail
+
+	vorr	q1,q3,q3
+	vld1.8	{q10},[r0]!
+.Loop3x_ecb_dec:
+	INST(0x60,0x03,0xb0,0xf3)	@ aesd q0,q8
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.32	{q8},[r7]!
+	subs	r6,r6,#2
+	INST(0x62,0x03,0xb0,0xf3)	@ aesd q0,q9
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.32	{q9},[r7]!
+	bgt	.Loop3x_ecb_dec
+
+	INST(0x60,0x03,0xb0,0xf3)	@ aesd q0,q8
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	subs	r2,r2,#0x30
+	it	lo
+	movlo	r6,r2				@ r6, r6, is zero at this point
+	INST(0x62,0x03,0xb0,0xf3)	@ aesd q0,q9
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	add	r0,r0,r6 			@ r0 is adjusted in such way that
+						@ at exit from the loop q1-q10
+						@ are loaded with last "words"
+	mov	r7,r3
+	INST(0x68,0x03,0xb0,0xf3)	@ aesd q0,q12
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x68,0x23,0xb0,0xf3)	@ aesd q1,q12
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x68,0x43,0xf0,0xf3)	@ aesd q10,q12
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.8	{q2},[r0]!
+	INST(0x6a,0x03,0xb0,0xf3)	@ aesd q0,q13
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x6a,0x23,0xb0,0xf3)	@ aesd q1,q13
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6a,0x43,0xf0,0xf3)	@ aesd q10,q13
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.8	{q3},[r0]!
+	INST(0x6c,0x03,0xb0,0xf3)	@ aesd q0,q14
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x6c,0x23,0xb0,0xf3)	@ aesd q1,q14
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6c,0x43,0xf0,0xf3)	@ aesd q10,q14
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.8	{q11},[r0]!
+	INST(0x6e,0x03,0xb0,0xf3)	@ aesd q0,q15
+	INST(0x6e,0x23,0xb0,0xf3)	@ aesd q1,q15
+	INST(0x6e,0x43,0xf0,0xf3)	@ aesd q10,q15
+	vld1.32	{q8},[r7]!			@ re-pre-load rndkey[0]
+	add	r6,r5,#2
+	veor	q4,q7,q0
+	veor	q5,q7,q1
+	veor	q10,q10,q7
+	vld1.32	{q9},[r7]!			@ re-pre-load rndkey[1]
+	vst1.8	{q4},[r1]!
+	vorr	q0,q2,q2
+	vst1.8	{q5},[r1]!
+	vorr	q1,q3,q3
+	vst1.8	{q10},[r1]!
+	vorr	q10,q11,q11
+	bhs	.Loop3x_ecb_dec
+
+	cmn	r2,#0x30
+	beq	.Lecb_done
+	nop
+
+.Lecb_dec_tail:
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.32	{q8},[r7]!
+	subs	r6,r6,#2
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	vld1.32	{q9},[r7]!
+	bgt	.Lecb_dec_tail
+
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x68,0x23,0xb0,0xf3)	@ aesd q1,q12
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x68,0x43,0xf0,0xf3)	@ aesd q10,q12
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	cmn	r2,#0x20
+	INST(0x6a,0x23,0xb0,0xf3)	@ aesd q1,q13
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6a,0x43,0xf0,0xf3)	@ aesd q10,q13
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x6c,0x23,0xb0,0xf3)	@ aesd q1,q14
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6c,0x43,0xf0,0xf3)	@ aesd q10,q14
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x6e,0x23,0xb0,0xf3)	@ aesd q1,q15
+	INST(0x6e,0x43,0xf0,0xf3)	@ aesd q10,q15
+	beq	.Lecb_dec_one
+	veor	q5,q7,q1
+	veor	q9,q7,q10
+	vst1.8	{q5},[r1]!
+	vst1.8	{q9},[r1]!
+	b	.Lecb_done
+
+.Lecb_dec_one:
+	veor	q5,q7,q10
+	vst1.8	{q5},[r1]!
+
+.Lecb_done:
+	vldmia	sp!,{d8,d9,d10,d11,d12,d13,d14,d15}
+	ldmia	sp!,{r4,r5,r6,r7,r8,pc}
+.size	aes_v8_ecb_encrypt,.-aes_v8_ecb_encrypt
 .globl	aes_v8_cbc_encrypt
 .type	aes_v8_cbc_encrypt,%function
 .align	5
@@ -294,6 +613,7 @@ aes_v8_cbc_encrypt:
 	subs	r2,r2,#16
 	mov	r8,#16
 	blo	.Lcbc_abort
+	it	eq
 	moveq	r8,#0
 
 	cmp	r5,#0			@ en- or decrypting?
@@ -324,58 +644,59 @@ aes_v8_cbc_encrypt:
 	add	r7,r3,#16
 	add	r6,r3,#16*4
 	add	r12,r3,#16*5
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	add	r14,r3,#16*6
 	add	r3,r3,#16*7
 	b	.Lenter_cbc_enc
 
 .align	4
 .Loop_cbc_enc:
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vst1.8	{q6},[r1]!
 .Lenter_cbc_enc:
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x04,0x03,0xb0,0xf3	@ aese q0,q2
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x04,0x03,0xb0,0xf3)	@ aese q0,q2
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.32	{q8},[r6]
 	cmp	r5,#4
-.byte	0x06,0x03,0xb0,0xf3	@ aese q0,q3
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x06,0x03,0xb0,0xf3)	@ aese q0,q3
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.32	{q9},[r12]
 	beq	.Lcbc_enc192
 
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.32	{q8},[r14]
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.32	{q9},[r3]
 	nop
 
 .Lcbc_enc192:
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	subs	r2,r2,#16
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	it	eq
 	moveq	r8,#0
-.byte	0x24,0x03,0xb0,0xf3	@ aese q0,q10
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x26,0x03,0xb0,0xf3	@ aese q0,q11
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x24,0x03,0xb0,0xf3)	@ aese q0,q10
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x26,0x03,0xb0,0xf3)	@ aese q0,q11
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.8	{q8},[r0],r8
-.byte	0x28,0x03,0xb0,0xf3	@ aese q0,q12
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x28,0x03,0xb0,0xf3)	@ aese q0,q12
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	veor	q8,q8,q5
-.byte	0x2a,0x03,0xb0,0xf3	@ aese q0,q13
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x2a,0x03,0xb0,0xf3)	@ aese q0,q13
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.32	{q9},[r7]		@ re-pre-load rndkey[1]
-.byte	0x2c,0x03,0xb0,0xf3	@ aese q0,q14
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x2e,0x03,0xb0,0xf3	@ aese q0,q15
+	INST(0x2c,0x03,0xb0,0xf3)	@ aese q0,q14
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2e,0x03,0xb0,0xf3)	@ aese q0,q15
 	veor	q6,q0,q7
 	bhs	.Loop_cbc_enc
 
@@ -385,35 +706,36 @@ aes_v8_cbc_encrypt:
 .align	5
 .Lcbc_enc128:
 	vld1.32	{q2,q3},[r7]
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	b	.Lenter_cbc_enc128
 .Loop_cbc_enc128:
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vst1.8	{q6},[r1]!
 .Lenter_cbc_enc128:
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	subs	r2,r2,#16
-.byte	0x04,0x03,0xb0,0xf3	@ aese q0,q2
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x04,0x03,0xb0,0xf3)	@ aese q0,q2
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	it	eq
 	moveq	r8,#0
-.byte	0x06,0x03,0xb0,0xf3	@ aese q0,q3
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x24,0x03,0xb0,0xf3	@ aese q0,q10
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x26,0x03,0xb0,0xf3	@ aese q0,q11
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x06,0x03,0xb0,0xf3)	@ aese q0,q3
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x24,0x03,0xb0,0xf3)	@ aese q0,q10
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x26,0x03,0xb0,0xf3)	@ aese q0,q11
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	vld1.8	{q8},[r0],r8
-.byte	0x28,0x03,0xb0,0xf3	@ aese q0,q12
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x2a,0x03,0xb0,0xf3	@ aese q0,q13
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x2c,0x03,0xb0,0xf3	@ aese q0,q14
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
+	INST(0x28,0x03,0xb0,0xf3)	@ aese q0,q12
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2a,0x03,0xb0,0xf3)	@ aese q0,q13
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2c,0x03,0xb0,0xf3)	@ aese q0,q14
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
 	veor	q8,q8,q5
-.byte	0x2e,0x03,0xb0,0xf3	@ aese q0,q15
+	INST(0x2e,0x03,0xb0,0xf3)	@ aese q0,q15
 	veor	q6,q0,q7
 	bhs	.Loop_cbc_enc128
 
@@ -434,71 +756,71 @@ aes_v8_cbc_encrypt:
 	vorr	q2,q0,q0
 	vorr	q3,q1,q1
 	vorr	q11,q10,q10
-
 .Loop3x_cbc_dec:
-.byte	0x60,0x03,0xb0,0xf3	@ aesd q0,q8
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x60,0x23,0xb0,0xf3	@ aesd q1,q8
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x60,0x43,0xf0,0xf3	@ aesd q10,q8
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x60,0x03,0xb0,0xf3)	@ aesd q0,q8
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.32	{q8},[r7]!
 	subs	r6,r6,#2
-.byte	0x62,0x03,0xb0,0xf3	@ aesd q0,q9
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x62,0x23,0xb0,0xf3	@ aesd q1,q9
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x62,0x43,0xf0,0xf3	@ aesd q10,q9
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x62,0x03,0xb0,0xf3)	@ aesd q0,q9
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.32	{q9},[r7]!
 	bgt	.Loop3x_cbc_dec
 
-.byte	0x60,0x03,0xb0,0xf3	@ aesd q0,q8
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x60,0x23,0xb0,0xf3	@ aesd q1,q8
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x60,0x43,0xf0,0xf3	@ aesd q10,q8
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x60,0x03,0xb0,0xf3)	@ aesd q0,q8
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	veor	q4,q6,q7
 	subs	r2,r2,#0x30
 	veor	q5,q2,q7
+	it	lo
 	movlo	r6,r2			@ r6, r6, is zero at this point
-.byte	0x62,0x03,0xb0,0xf3	@ aesd q0,q9
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x62,0x23,0xb0,0xf3	@ aesd q1,q9
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x62,0x43,0xf0,0xf3	@ aesd q10,q9
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x62,0x03,0xb0,0xf3)	@ aesd q0,q9
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	veor	q9,q3,q7
 	add	r0,r0,r6		@ r0 is adjusted in such way that
 					@ at exit from the loop q1-q10
 					@ are loaded with last "words"
 	vorr	q6,q11,q11
 	mov	r7,r3
-.byte	0x68,0x03,0xb0,0xf3	@ aesd q0,q12
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x68,0x23,0xb0,0xf3	@ aesd q1,q12
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x68,0x43,0xf0,0xf3	@ aesd q10,q12
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x68,0x03,0xb0,0xf3)	@ aesd q0,q12
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x68,0x23,0xb0,0xf3)	@ aesd q1,q12
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x68,0x43,0xf0,0xf3)	@ aesd q10,q12
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.8	{q2},[r0]!
-.byte	0x6a,0x03,0xb0,0xf3	@ aesd q0,q13
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x6a,0x23,0xb0,0xf3	@ aesd q1,q13
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x6a,0x43,0xf0,0xf3	@ aesd q10,q13
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x6a,0x03,0xb0,0xf3)	@ aesd q0,q13
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x6a,0x23,0xb0,0xf3)	@ aesd q1,q13
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6a,0x43,0xf0,0xf3)	@ aesd q10,q13
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.8	{q3},[r0]!
-.byte	0x6c,0x03,0xb0,0xf3	@ aesd q0,q14
-.byte	0xc0,0x03,0xb0,0xf3	@ aesimc q0,q0
-.byte	0x6c,0x23,0xb0,0xf3	@ aesd q1,q14
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x6c,0x43,0xf0,0xf3	@ aesd q10,q14
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x6c,0x03,0xb0,0xf3)	@ aesd q0,q14
+	INST(0xc0,0x03,0xb0,0xf3)	@ aesimc q0,q0
+	INST(0x6c,0x23,0xb0,0xf3)	@ aesd q1,q14
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6c,0x43,0xf0,0xf3)	@ aesd q10,q14
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.8	{q11},[r0]!
-.byte	0x6e,0x03,0xb0,0xf3	@ aesd q0,q15
-.byte	0x6e,0x23,0xb0,0xf3	@ aesd q1,q15
-.byte	0x6e,0x43,0xf0,0xf3	@ aesd q10,q15
+	INST(0x6e,0x03,0xb0,0xf3)	@ aesd q0,q15
+	INST(0x6e,0x23,0xb0,0xf3)	@ aesd q1,q15
+	INST(0x6e,0x43,0xf0,0xf3)	@ aesd q10,q15
 	vld1.32	{q8},[r7]!	@ re-pre-load rndkey[0]
 	add	r6,r5,#2
 	veor	q4,q4,q0
@@ -518,44 +840,44 @@ aes_v8_cbc_encrypt:
 	nop
 
 .Lcbc_dec_tail:
-.byte	0x60,0x23,0xb0,0xf3	@ aesd q1,q8
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x60,0x43,0xf0,0xf3	@ aesd q10,q8
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.32	{q8},[r7]!
 	subs	r6,r6,#2
-.byte	0x62,0x23,0xb0,0xf3	@ aesd q1,q9
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x62,0x43,0xf0,0xf3	@ aesd q10,q9
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	vld1.32	{q9},[r7]!
 	bgt	.Lcbc_dec_tail
 
-.byte	0x60,0x23,0xb0,0xf3	@ aesd q1,q8
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x60,0x43,0xf0,0xf3	@ aesd q10,q8
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
-.byte	0x62,0x23,0xb0,0xf3	@ aesd q1,q9
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x62,0x43,0xf0,0xf3	@ aesd q10,q9
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
-.byte	0x68,0x23,0xb0,0xf3	@ aesd q1,q12
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x68,0x43,0xf0,0xf3	@ aesd q10,q12
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x60,0x23,0xb0,0xf3)	@ aesd q1,q8
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x60,0x43,0xf0,0xf3)	@ aesd q10,q8
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x62,0x23,0xb0,0xf3)	@ aesd q1,q9
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x62,0x43,0xf0,0xf3)	@ aesd q10,q9
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
+	INST(0x68,0x23,0xb0,0xf3)	@ aesd q1,q12
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x68,0x43,0xf0,0xf3)	@ aesd q10,q12
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	cmn	r2,#0x20
-.byte	0x6a,0x23,0xb0,0xf3	@ aesd q1,q13
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x6a,0x43,0xf0,0xf3	@ aesd q10,q13
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x6a,0x23,0xb0,0xf3)	@ aesd q1,q13
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6a,0x43,0xf0,0xf3)	@ aesd q10,q13
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	veor	q5,q6,q7
-.byte	0x6c,0x23,0xb0,0xf3	@ aesd q1,q14
-.byte	0xc2,0x23,0xb0,0xf3	@ aesimc q1,q1
-.byte	0x6c,0x43,0xf0,0xf3	@ aesd q10,q14
-.byte	0xe4,0x43,0xf0,0xf3	@ aesimc q10,q10
+	INST(0x6c,0x23,0xb0,0xf3)	@ aesd q1,q14
+	INST(0xc2,0x23,0xb0,0xf3)	@ aesimc q1,q1
+	INST(0x6c,0x43,0xf0,0xf3)	@ aesd q10,q14
+	INST(0xe4,0x43,0xf0,0xf3)	@ aesimc q10,q10
 	veor	q9,q3,q7
-.byte	0x6e,0x23,0xb0,0xf3	@ aesd q1,q15
-.byte	0x6e,0x43,0xf0,0xf3	@ aesd q10,q15
+	INST(0x6e,0x23,0xb0,0xf3)	@ aesd q1,q15
+	INST(0x6e,0x43,0xf0,0xf3)	@ aesd q10,q15
 	beq	.Lcbc_dec_one
 	veor	q5,q5,q1
 	veor	q9,q9,q10
@@ -602,6 +924,7 @@ aes_v8_ctr32_encrypt_blocks:
 	vld1.32	{q7},[r7]
 	add	r7,r3,#32
 	mov	r6,r5
+	it	lo
 	movlo	r12,#0
 #ifndef __ARMEB__
 	rev	r8, r8
@@ -621,76 +944,76 @@ aes_v8_ctr32_encrypt_blocks:
 
 .align	4
 .Loop3x_ctr32:
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x20,0x23,0xb0,0xf3	@ aese q1,q8
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
-.byte	0x20,0x43,0xf0,0xf3	@ aese q10,q8
-.byte	0xa4,0x43,0xf0,0xf3	@ aesmc q10,q10
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
 	vld1.32	{q8},[r7]!
 	subs	r6,r6,#2
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x22,0x23,0xb0,0xf3	@ aese q1,q9
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
-.byte	0x22,0x43,0xf0,0xf3	@ aese q10,q9
-.byte	0xa4,0x43,0xf0,0xf3	@ aesmc q10,q10
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
 	vld1.32	{q9},[r7]!
 	bgt	.Loop3x_ctr32
 
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x83,0xb0,0xf3	@ aesmc q4,q0
-.byte	0x20,0x23,0xb0,0xf3	@ aese q1,q8
-.byte	0x82,0xa3,0xb0,0xf3	@ aesmc q5,q1
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x83,0xb0,0xf3)	@ aesmc q4,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0xa3,0xb0,0xf3)	@ aesmc q5,q1
 	vld1.8	{q2},[r0]!
 	add	r9,r8,#1
-.byte	0x20,0x43,0xf0,0xf3	@ aese q10,q8
-.byte	0xa4,0x43,0xf0,0xf3	@ aesmc q10,q10
+	INST(0x20,0x43,0xf0,0xf3)	@ aese q10,q8
+	INST(0xa4,0x43,0xf0,0xf3)	@ aesmc q10,q10
 	vld1.8	{q3},[r0]!
 	rev	r9,r9
-.byte	0x22,0x83,0xb0,0xf3	@ aese q4,q9
-.byte	0x88,0x83,0xb0,0xf3	@ aesmc q4,q4
-.byte	0x22,0xa3,0xb0,0xf3	@ aese q5,q9
-.byte	0x8a,0xa3,0xb0,0xf3	@ aesmc q5,q5
+	INST(0x22,0x83,0xb0,0xf3)	@ aese q4,q9
+	INST(0x88,0x83,0xb0,0xf3)	@ aesmc q4,q4
+	INST(0x22,0xa3,0xb0,0xf3)	@ aese q5,q9
+	INST(0x8a,0xa3,0xb0,0xf3)	@ aesmc q5,q5
 	vld1.8	{q11},[r0]!
 	mov	r7,r3
-.byte	0x22,0x43,0xf0,0xf3	@ aese q10,q9
-.byte	0xa4,0x23,0xf0,0xf3	@ aesmc q9,q10
-.byte	0x28,0x83,0xb0,0xf3	@ aese q4,q12
-.byte	0x88,0x83,0xb0,0xf3	@ aesmc q4,q4
-.byte	0x28,0xa3,0xb0,0xf3	@ aese q5,q12
-.byte	0x8a,0xa3,0xb0,0xf3	@ aesmc q5,q5
+	INST(0x22,0x43,0xf0,0xf3)	@ aese q10,q9
+	INST(0xa4,0x23,0xf0,0xf3)	@ aesmc q9,q10
+	INST(0x28,0x83,0xb0,0xf3)	@ aese q4,q12
+	INST(0x88,0x83,0xb0,0xf3)	@ aesmc q4,q4
+	INST(0x28,0xa3,0xb0,0xf3)	@ aese q5,q12
+	INST(0x8a,0xa3,0xb0,0xf3)	@ aesmc q5,q5
 	veor	q2,q2,q7
 	add	r10,r8,#2
-.byte	0x28,0x23,0xf0,0xf3	@ aese q9,q12
-.byte	0xa2,0x23,0xf0,0xf3	@ aesmc q9,q9
+	INST(0x28,0x23,0xf0,0xf3)	@ aese q9,q12
+	INST(0xa2,0x23,0xf0,0xf3)	@ aesmc q9,q9
 	veor	q3,q3,q7
 	add	r8,r8,#3
-.byte	0x2a,0x83,0xb0,0xf3	@ aese q4,q13
-.byte	0x88,0x83,0xb0,0xf3	@ aesmc q4,q4
-.byte	0x2a,0xa3,0xb0,0xf3	@ aese q5,q13
-.byte	0x8a,0xa3,0xb0,0xf3	@ aesmc q5,q5
+	INST(0x2a,0x83,0xb0,0xf3)	@ aese q4,q13
+	INST(0x88,0x83,0xb0,0xf3)	@ aesmc q4,q4
+	INST(0x2a,0xa3,0xb0,0xf3)	@ aese q5,q13
+	INST(0x8a,0xa3,0xb0,0xf3)	@ aesmc q5,q5
 	veor	q11,q11,q7
 	vmov.32	d13[1], r9
-.byte	0x2a,0x23,0xf0,0xf3	@ aese q9,q13
-.byte	0xa2,0x23,0xf0,0xf3	@ aesmc q9,q9
+	INST(0x2a,0x23,0xf0,0xf3)	@ aese q9,q13
+	INST(0xa2,0x23,0xf0,0xf3)	@ aesmc q9,q9
 	vorr	q0,q6,q6
 	rev	r10,r10
-.byte	0x2c,0x83,0xb0,0xf3	@ aese q4,q14
-.byte	0x88,0x83,0xb0,0xf3	@ aesmc q4,q4
+	INST(0x2c,0x83,0xb0,0xf3)	@ aese q4,q14
+	INST(0x88,0x83,0xb0,0xf3)	@ aesmc q4,q4
 	vmov.32	d13[1], r10
 	rev	r12,r8
-.byte	0x2c,0xa3,0xb0,0xf3	@ aese q5,q14
-.byte	0x8a,0xa3,0xb0,0xf3	@ aesmc q5,q5
+	INST(0x2c,0xa3,0xb0,0xf3)	@ aese q5,q14
+	INST(0x8a,0xa3,0xb0,0xf3)	@ aesmc q5,q5
 	vorr	q1,q6,q6
 	vmov.32	d13[1], r12
-.byte	0x2c,0x23,0xf0,0xf3	@ aese q9,q14
-.byte	0xa2,0x23,0xf0,0xf3	@ aesmc q9,q9
+	INST(0x2c,0x23,0xf0,0xf3)	@ aese q9,q14
+	INST(0xa2,0x23,0xf0,0xf3)	@ aesmc q9,q9
 	vorr	q10,q6,q6
 	subs	r2,r2,#3
-.byte	0x2e,0x83,0xb0,0xf3	@ aese q4,q15
-.byte	0x2e,0xa3,0xb0,0xf3	@ aese q5,q15
-.byte	0x2e,0x23,0xf0,0xf3	@ aese q9,q15
+	INST(0x2e,0x83,0xb0,0xf3)	@ aese q4,q15
+	INST(0x2e,0xa3,0xb0,0xf3)	@ aese q5,q15
+	INST(0x2e,0x23,0xf0,0xf3)	@ aese q9,q15
 
 	veor	q2,q2,q4
 	vld1.32	{q8},[r7]!	@ re-pre-load rndkey[0]
@@ -707,48 +1030,49 @@ aes_v8_ctr32_encrypt_blocks:
 	beq	.Lctr32_done
 	cmp	r2,#1
 	mov	r12,#16
+	it	eq
 	moveq	r12,#0
 
 .Lctr32_tail:
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x20,0x23,0xb0,0xf3	@ aese q1,q8
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	vld1.32	{q8},[r7]!
 	subs	r6,r6,#2
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x22,0x23,0xb0,0xf3	@ aese q1,q9
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	vld1.32	{q9},[r7]!
 	bgt	.Lctr32_tail
 
-.byte	0x20,0x03,0xb0,0xf3	@ aese q0,q8
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x20,0x23,0xb0,0xf3	@ aese q1,q8
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
-.byte	0x22,0x03,0xb0,0xf3	@ aese q0,q9
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x22,0x23,0xb0,0xf3	@ aese q1,q9
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x20,0x03,0xb0,0xf3)	@ aese q0,q8
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x20,0x23,0xb0,0xf3)	@ aese q1,q8
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
+	INST(0x22,0x03,0xb0,0xf3)	@ aese q0,q9
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x22,0x23,0xb0,0xf3)	@ aese q1,q9
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	vld1.8	{q2},[r0],r12
-.byte	0x28,0x03,0xb0,0xf3	@ aese q0,q12
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x28,0x23,0xb0,0xf3	@ aese q1,q12
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x28,0x03,0xb0,0xf3)	@ aese q0,q12
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x28,0x23,0xb0,0xf3)	@ aese q1,q12
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	vld1.8	{q3},[r0]
-.byte	0x2a,0x03,0xb0,0xf3	@ aese q0,q13
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x2a,0x23,0xb0,0xf3	@ aese q1,q13
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x2a,0x03,0xb0,0xf3)	@ aese q0,q13
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2a,0x23,0xb0,0xf3)	@ aese q1,q13
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	veor	q2,q2,q7
-.byte	0x2c,0x03,0xb0,0xf3	@ aese q0,q14
-.byte	0x80,0x03,0xb0,0xf3	@ aesmc q0,q0
-.byte	0x2c,0x23,0xb0,0xf3	@ aese q1,q14
-.byte	0x82,0x23,0xb0,0xf3	@ aesmc q1,q1
+	INST(0x2c,0x03,0xb0,0xf3)	@ aese q0,q14
+	INST(0x80,0x03,0xb0,0xf3)	@ aesmc q0,q0
+	INST(0x2c,0x23,0xb0,0xf3)	@ aese q1,q14
+	INST(0x82,0x23,0xb0,0xf3)	@ aesmc q1,q1
 	veor	q3,q3,q7
-.byte	0x2e,0x03,0xb0,0xf3	@ aese q0,q15
-.byte	0x2e,0x23,0xb0,0xf3	@ aese q1,q15
+	INST(0x2e,0x03,0xb0,0xf3)	@ aese q0,q15
+	INST(0x2e,0x23,0xb0,0xf3)	@ aese q1,q15
 
 	cmp	r2,#1
 	veor	q2,q2,q0
diff --git a/sys/crypto/openssl/arm/armv4-gf2m.S b/sys/crypto/openssl/arm/armv4-gf2m.S
index 5ee07d4bf9a35..ccf6212f6eb10 100644
--- a/sys/crypto/openssl/arm/armv4-gf2m.S
+++ b/sys/crypto/openssl/arm/armv4-gf2m.S
@@ -1,13 +1,14 @@
 /* Do not modify. This file is auto-generated from armv4-gf2m.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
 #else
 .code	32
 #endif
+
+.text
 .type	mul_1x1_ialu,%function
 .align	5
 mul_1x1_ialu:
@@ -100,11 +101,13 @@ bn_GF2m_mul_2x2:
 #if __ARM_MAX_ARCH__>=7
 	stmdb	sp!,{r10,lr}
 	ldr	r12,.LOPENSSL_armcap
+# if !defined(_WIN32)
 	adr	r10,.LOPENSSL_armcap
 	ldr	r12,[r12,r10]
-#ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r12,[r12]
-#endif
+# endif
 	tst	r12,#ARMV7_NEON
 	itt	ne
 	ldrne	r10,[sp],#8
@@ -218,7 +221,11 @@ bn_GF2m_mul_2x2:
 #if __ARM_MAX_ARCH__>=7
 .align	5
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.
+# endif
 #endif
 .byte	71,70,40,50,94,109,41,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,65,82,77,118,52,47,78,69,79,78,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
diff --git a/sys/crypto/openssl/arm/armv4-mont.S b/sys/crypto/openssl/arm/armv4-mont.S
index 69ca04f891192..460d5cb400cb8 100644
--- a/sys/crypto/openssl/arm/armv4-mont.S
+++ b/sys/crypto/openssl/arm/armv4-mont.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from armv4-mont.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -9,10 +8,16 @@
 .code	32
 #endif
 
+.text
+
 #if __ARM_MAX_ARCH__>=7
 .align	5
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.Lbn_mul_mont
+# endif
 #endif
 
 .globl	bn_mul_mont
@@ -26,12 +31,14 @@ bn_mul_mont:
 #if __ARM_MAX_ARCH__>=7
 	tst	ip,#7
 	bne	.Lialu
-	adr	r0,.Lbn_mul_mont
-	ldr	r2,.LOPENSSL_armcap
+	ldr	r0,.LOPENSSL_armcap
+#if !defined(_WIN32)
+	adr	r2,.Lbn_mul_mont
 	ldr	r0,[r0,r2]
-#ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r0,[r0]
-#endif
+# endif
 	tst	r0,#ARMV7_NEON		@ NEON available?
 	ldmia	sp, {r0,r2}
 	beq	.Lialu
diff --git a/sys/crypto/openssl/arm/armv4cpuid.S b/sys/crypto/openssl/arm/armv4cpuid.S
index f1f59a2de10ec..8436d1f079c4f 100644
--- a/sys/crypto/openssl/arm/armv4cpuid.S
+++ b/sys/crypto/openssl/arm/armv4cpuid.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from armv4cpuid.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__) && !defined(__APPLE__)
 .syntax	unified
 .thumb
@@ -10,6 +9,8 @@
 #undef	__thumb2__
 #endif
 
+.text
+
 .align	5
 .globl	OPENSSL_atomic_add
 .type	OPENSSL_atomic_add,%function
diff --git a/sys/crypto/openssl/arm/bsaes-armv7.S b/sys/crypto/openssl/arm/bsaes-armv7.S
index 2414da194a228..91261fc0b69d2 100644
--- a/sys/crypto/openssl/arm/bsaes-armv7.S
+++ b/sys/crypto/openssl/arm/bsaes-armv7.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from bsaes-armv7.pl. */
-@ Copyright 2012-2020 The OpenSSL Project Authors. All Rights Reserved.
+@ Copyright 2012-2023 The OpenSSL Project Authors. All Rights Reserved.
 @
-@ Licensed under the OpenSSL license (the "License").  You may not use
+@ Licensed under the Apache License 2.0 (the "License").  You may not use
 @ this file except in compliance with the License.  You can obtain a copy
 @ in the file LICENSE in the source distribution or at
 @ https://www.openssl.org/source/license.html
@@ -14,7 +14,7 @@
 @ details see http://www.openssl.org/~appro/cryptogams/.
 @
 @ Specific modes and adaptation for Linux kernel by Ard Biesheuvel
-@ of Linaro. Permission to use under GPL terms is granted.
+@ of Linaro.
 @ ====================================================================
 
 @ Bit-sliced AES for ARM NEON
@@ -50,6 +50,8 @@
 @ April-August 2013
 @ Add CBC, CTR and XTS subroutines and adapt for kernel use; courtesy of Ard.
 
+@ $output is the last argument if it looks like a file (it has an extension)
+@ $flavour is the first argument if it doesn't look like a file
 #ifndef __KERNEL__
 # include "arm_arch.h"
 
@@ -74,7 +76,6 @@
 .arch	armv7-a
 .fpu	neon
 
-.text
 .syntax	unified 	@ ARMv7-capable assembler is expected to handle this
 #if defined(__thumb2__) && !defined(__APPLE__)
 .thumb
@@ -83,6 +84,8 @@
 # undef __thumb2__
 #endif
 
+.text
+
 .type	_bsaes_decrypt8,%function
 .align	4
 _bsaes_decrypt8:
@@ -1071,18 +1074,18 @@ _bsaes_key_convert:
 
 
 
-.globl	bsaes_cbc_encrypt
-.type	bsaes_cbc_encrypt,%function
+.globl	ossl_bsaes_cbc_encrypt
+.type	ossl_bsaes_cbc_encrypt,%function
 .align	5
-bsaes_cbc_encrypt:
+ossl_bsaes_cbc_encrypt:
 #ifndef	__KERNEL__
 	cmp	r2, #128
 #ifndef	__thumb__
 	blo	AES_cbc_encrypt
 #else
-	bhs	1f
+	bhs	.Lcbc_do_bsaes
 	b	AES_cbc_encrypt
-1:
+.Lcbc_do_bsaes:
 #endif
 #endif
 
@@ -1336,12 +1339,12 @@ bsaes_cbc_encrypt:
 	vst1.8	{q15}, [r8]		@ return IV
 	VFP_ABI_POP
 	ldmia	sp!, {r4,r5,r6,r7,r8,r9,r10, pc}
-.size	bsaes_cbc_encrypt,.-bsaes_cbc_encrypt
+.size	ossl_bsaes_cbc_encrypt,.-ossl_bsaes_cbc_encrypt
 
-.globl	bsaes_ctr32_encrypt_blocks
-.type	bsaes_ctr32_encrypt_blocks,%function
+.globl	ossl_bsaes_ctr32_encrypt_blocks
+.type	ossl_bsaes_ctr32_encrypt_blocks,%function
 .align	5
-bsaes_ctr32_encrypt_blocks:
+ossl_bsaes_ctr32_encrypt_blocks:
 	cmp	r2, #8			@ use plain AES for
 	blo	.Lctr_enc_short			@ small sizes
 
@@ -1564,11 +1567,11 @@ bsaes_ctr32_encrypt_blocks:
 	vstmia	sp!, {q0,q1}
 
 	ldmia	sp!, {r4,r5,r6,r7,r8, pc}
-.size	bsaes_ctr32_encrypt_blocks,.-bsaes_ctr32_encrypt_blocks
-.globl	bsaes_xts_encrypt
-.type	bsaes_xts_encrypt,%function
+.size	ossl_bsaes_ctr32_encrypt_blocks,.-ossl_bsaes_ctr32_encrypt_blocks
+.globl	ossl_bsaes_xts_encrypt
+.type	ossl_bsaes_xts_encrypt,%function
 .align	4
-bsaes_xts_encrypt:
+ossl_bsaes_xts_encrypt:
 	mov	ip, sp
 	stmdb	sp!, {r4,r5,r6,r7,r8,r9,r10, lr}		@ 0x20
 	VFP_ABI_PUSH
@@ -2043,12 +2046,12 @@ bsaes_xts_encrypt:
 	VFP_ABI_POP
 	ldmia	sp!, {r4,r5,r6,r7,r8,r9,r10, pc}	@ return
 
-.size	bsaes_xts_encrypt,.-bsaes_xts_encrypt
+.size	ossl_bsaes_xts_encrypt,.-ossl_bsaes_xts_encrypt
 
-.globl	bsaes_xts_decrypt
-.type	bsaes_xts_decrypt,%function
+.globl	ossl_bsaes_xts_decrypt
+.type	ossl_bsaes_xts_decrypt,%function
 .align	4
-bsaes_xts_decrypt:
+ossl_bsaes_xts_decrypt:
 	mov	ip, sp
 	stmdb	sp!, {r4,r5,r6,r7,r8,r9,r10, lr}		@ 0x20
 	VFP_ABI_PUSH
@@ -2554,5 +2557,5 @@ bsaes_xts_decrypt:
 	VFP_ABI_POP
 	ldmia	sp!, {r4,r5,r6,r7,r8,r9,r10, pc}	@ return
 
-.size	bsaes_xts_decrypt,.-bsaes_xts_decrypt
+.size	ossl_bsaes_xts_decrypt,.-ossl_bsaes_xts_decrypt
 #endif
diff --git a/sys/crypto/openssl/arm/chacha-armv4.S b/sys/crypto/openssl/arm/chacha-armv4.S
index 077500d3c188f..9b673c4c981e4 100644
--- a/sys/crypto/openssl/arm/chacha-armv4.S
+++ b/sys/crypto/openssl/arm/chacha-armv4.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from chacha-armv4.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__) || defined(__clang__)
 .syntax	unified
 #endif
@@ -15,6 +14,8 @@
 #define ldrhsb	ldrbhs
 #endif
 
+.text
+
 .align	5
 .Lsigma:
 .long	0x61707865,0x3320646e,0x79622d32,0x6b206574	@ endian-neutral
@@ -22,7 +23,11 @@
 .long	1,0,0,0
 #if __ARM_MAX_ARCH__>=7
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.LChaCha20_ctr32
+# endif
 #else
 .word	-1
 #endif
@@ -49,8 +54,10 @@ ChaCha20_ctr32:
 	cmp	r2,#192			@ test len
 	bls	.Lshort
 	ldr	r4,[r14,#-32]
+# if !defined(_WIN32)
 	ldr	r4,[r14,r4]
-# ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r4,[r4]
 # endif
 	tst	r4,#ARMV7_NEON
diff --git a/sys/crypto/openssl/arm/ecp_nistz256-armv4.S b/sys/crypto/openssl/arm/ecp_nistz256-armv4.S
index 150b83583dea2..90338c24f733c 100644
--- a/sys/crypto/openssl/arm/ecp_nistz256-armv4.S
+++ b/sys/crypto/openssl/arm/ecp_nistz256-armv4.S
@@ -1,13 +1,13 @@
 /* Do not modify. This file is auto-generated from ecp_nistz256-armv4.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
 #else
 .code	32
 #endif
+.section	.rodata
 .globl	ecp_nistz256_precomputed
 .type	ecp_nistz256_precomputed,%object
 .align	12
@@ -2381,6 +2381,8 @@ ecp_nistz256_precomputed:
 .byte	0xec,0xf0,0x42,0x88,0xd0,0x81,0x51,0xf9,0x1b,0xbc,0x43,0xa4,0x37,0xf1,0xd7,0x90,0x21,0x7e,0xa0,0x3e,0x63,0xfb,0x21,0xfa,0x12,0xfb,0xde,0xc7,0xbf,0xb3,0x58,0xe7,0x76,0x42,0x20,0x01,0x3d,0x66,0x80,0xf1,0xb8,0xaf,0xfa,0x7d,0x96,0x89,0x36,0x48,0x95,0xd9,0x6e,0x6d,0xe6,0x4f,0xff,0x2a,0x47,0x61,0xf2,0x04,0xb7,0x83,0x14,0xce
 .byte	0x0a,0x3c,0x73,0x17,0x50,0x88,0x03,0x25,0x4a,0xe3,0x13,0x55,0x8b,0x7e,0x50,0x38,0xfc,0x14,0x0b,0x04,0x8e,0xa8,0x5b,0xd6,0x72,0x20,0x60,0xe9,0xaa,0x22,0x82,0x11,0xc6,0xc4,0xd7,0xb9,0xc8,0x0c,0x7e,0x05,0xfb,0x90,0xe4,0x9c,0x28,0x89,0x29,0x99,0x63,0x4d,0xec,0x7b,0x50,0xbd,0xd8,0xa3,0x5b,0x50,0x77,0x19,0x81,0x92,0xce,0x82
 .size	ecp_nistz256_precomputed,.-ecp_nistz256_precomputed
+
+.text
 .align	5
 .LRR:@ 2^512 mod P precomputed for NIST P256 polynomial
 .long	0x00000003, 0x00000000, 0xffffffff, 0xfffffffb
diff --git a/sys/crypto/openssl/arm/ghash-armv4.S b/sys/crypto/openssl/arm/ghash-armv4.S
index 413ab2ffed1d6..e665ba66797fa 100644
--- a/sys/crypto/openssl/arm/ghash-armv4.S
+++ b/sys/crypto/openssl/arm/ghash-armv4.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from ghash-armv4.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__) || defined(__clang__)
 .syntax	unified
 #define ldrplb  ldrbpl
@@ -13,6 +12,8 @@
 .code	32
 #endif
 
+.text
+
 .type	rem_4bit,%object
 .align	5
 rem_4bit:
diff --git a/sys/crypto/openssl/arm/ghashv8-armx.S b/sys/crypto/openssl/arm/ghashv8-armx.S
index f617d99814301..3234cc9ba497a 100644
--- a/sys/crypto/openssl/arm/ghashv8-armx.S
+++ b/sys/crypto/openssl/arm/ghashv8-armx.S
@@ -2,10 +2,17 @@
 #include "arm_arch.h"
 
 #if __ARM_MAX_ARCH__>=7
-.text
 .fpu	neon
+#ifdef __thumb2__
+.syntax	unified
+.thumb
+# define INST(a,b,c,d) .byte  c,0xef,a,b
+#else
 .code	32
-#undef	__thumb2__
+# define INST(a,b,c,d) .byte  a,b,c,0xf2
+#endif
+
+.text
 .globl	gcm_init_v8
 .type	gcm_init_v8,%function
 .align	4
@@ -29,23 +36,23 @@ gcm_init_v8:
 
 	@ calculate H^2
 	vext.8	q8,q12,q12,#8		@ Karatsuba pre-processing
-.byte	0xa8,0x0e,0xa8,0xf2	@ pmull q0,q12,q12
+	INST(0xa8,0x0e,0xa8,0xf2)	@ pmull q0,q12,q12
 	veor	q8,q8,q12
-.byte	0xa9,0x4e,0xa9,0xf2	@ pmull2 q2,q12,q12
-.byte	0xa0,0x2e,0xa0,0xf2	@ pmull q1,q8,q8
+	INST(0xa9,0x4e,0xa9,0xf2)	@ pmull2 q2,q12,q12
+	INST(0xa0,0x2e,0xa0,0xf2)	@ pmull q1,q8,q8
 
 	vext.8	q9,q0,q2,#8		@ Karatsuba post-processing
 	veor	q10,q0,q2
 	veor	q1,q1,q9
 	veor	q1,q1,q10
-.byte	0x26,0x4e,0xe0,0xf2	@ pmull q10,q0,q11		@ 1st phase
+	INST(0x26,0x4e,0xe0,0xf2)	@ pmull q10,q0,q11		@ 1st phase
 
 	vmov	d4,d3		@ Xh|Xm - 256-bit result
 	vmov	d3,d0		@ Xm is rotated Xl
 	veor	q0,q1,q10
 
 	vext.8	q10,q0,q0,#8		@ 2nd phase
-.byte	0x26,0x0e,0xa0,0xf2	@ pmull q0,q0,q11
+	INST(0x26,0x0e,0xa0,0xf2)	@ pmull q0,q0,q11
 	veor	q10,q10,q2
 	veor	q14,q0,q10
 
@@ -68,23 +75,23 @@ gcm_gmult_v8:
 #endif
 	vext.8	q3,q9,q9,#8
 
-.byte	0x86,0x0e,0xa8,0xf2	@ pmull q0,q12,q3		@ H.lo·Xi.lo
+	INST(0x86,0x0e,0xa8,0xf2)	@ pmull q0,q12,q3		@ H.lo·Xi.lo
 	veor	q9,q9,q3		@ Karatsuba pre-processing
-.byte	0x87,0x4e,0xa9,0xf2	@ pmull2 q2,q12,q3		@ H.hi·Xi.hi
-.byte	0xa2,0x2e,0xaa,0xf2	@ pmull q1,q13,q9		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
+	INST(0x87,0x4e,0xa9,0xf2)	@ pmull2 q2,q12,q3		@ H.hi·Xi.hi
+	INST(0xa2,0x2e,0xaa,0xf2)	@ pmull q1,q13,q9		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
 
 	vext.8	q9,q0,q2,#8		@ Karatsuba post-processing
 	veor	q10,q0,q2
 	veor	q1,q1,q9
 	veor	q1,q1,q10
-.byte	0x26,0x4e,0xe0,0xf2	@ pmull q10,q0,q11		@ 1st phase of reduction
+	INST(0x26,0x4e,0xe0,0xf2)	@ pmull q10,q0,q11		@ 1st phase of reduction
 
 	vmov	d4,d3		@ Xh|Xm - 256-bit result
 	vmov	d3,d0		@ Xm is rotated Xl
 	veor	q0,q1,q10
 
 	vext.8	q10,q0,q0,#8		@ 2nd phase of reduction
-.byte	0x26,0x0e,0xa0,0xf2	@ pmull q0,q0,q11
+	INST(0x26,0x0e,0xa0,0xf2)	@ pmull q0,q0,q11
 	veor	q10,q10,q2
 	veor	q0,q0,q10
 
@@ -120,6 +127,7 @@ gcm_ghash_v8:
 	vld1.64	{q12,q13},[r1]!	@ load twisted H, ..., H^2
 	vmov.i8	q11,#0xe1
 	vld1.64	{q14},[r1]
+	it	eq
 	moveq	r12,#0			@ is it time to zero r12?
 	vext.8	q0,q0,q0,#8		@ rotate Xi
 	vld1.64	{q8},[r2]!	@ load [rotated] I[0]
@@ -136,26 +144,28 @@ gcm_ghash_v8:
 #endif
 	vext.8	q7,q9,q9,#8
 	veor	q3,q3,q0		@ I[i]^=Xi
-.byte	0x8e,0x8e,0xa8,0xf2	@ pmull q4,q12,q7		@ H·Ii+1
+	INST(0x8e,0x8e,0xa8,0xf2)	@ pmull q4,q12,q7		@ H·Ii+1
 	veor	q9,q9,q7		@ Karatsuba pre-processing
-.byte	0x8f,0xce,0xa9,0xf2	@ pmull2 q6,q12,q7
+	INST(0x8f,0xce,0xa9,0xf2)	@ pmull2 q6,q12,q7
 	b	.Loop_mod2x_v8
 
 .align	4
 .Loop_mod2x_v8:
 	vext.8	q10,q3,q3,#8
 	subs	r3,r3,#32		@ is there more data?
-.byte	0x86,0x0e,0xac,0xf2	@ pmull q0,q14,q3		@ H^2.lo·Xi.lo
+	INST(0x86,0x0e,0xac,0xf2)	@ pmull q0,q14,q3		@ H^2.lo·Xi.lo
+	it	lo
 	movlo	r12,#0			@ is it time to zero r12?
 
-.byte	0xa2,0xae,0xaa,0xf2	@ pmull q5,q13,q9
+	INST(0xa2,0xae,0xaa,0xf2)	@ pmull q5,q13,q9
 	veor	q10,q10,q3		@ Karatsuba pre-processing
-.byte	0x87,0x4e,0xad,0xf2	@ pmull2 q2,q14,q3		@ H^2.hi·Xi.hi
+	INST(0x87,0x4e,0xad,0xf2)	@ pmull2 q2,q14,q3		@ H^2.hi·Xi.hi
 	veor	q0,q0,q4		@ accumulate
-.byte	0xa5,0x2e,0xab,0xf2	@ pmull2 q1,q13,q10		@ (H^2.lo+H^2.hi)·(Xi.lo+Xi.hi)
+	INST(0xa5,0x2e,0xab,0xf2)	@ pmull2 q1,q13,q10		@ (H^2.lo+H^2.hi)·(Xi.lo+Xi.hi)
 	vld1.64	{q8},[r2],r12	@ load [rotated] I[i+2]
 
 	veor	q2,q2,q6
+	it	eq
 	moveq	r12,#0			@ is it time to zero r12?
 	veor	q1,q1,q5
 
@@ -167,7 +177,7 @@ gcm_ghash_v8:
 	vrev64.8	q8,q8
 #endif
 	veor	q1,q1,q10
-.byte	0x26,0x4e,0xe0,0xf2	@ pmull q10,q0,q11		@ 1st phase of reduction
+	INST(0x26,0x4e,0xe0,0xf2)	@ pmull q10,q0,q11		@ 1st phase of reduction
 
 #ifndef __ARMEB__
 	vrev64.8	q9,q9
@@ -177,15 +187,15 @@ gcm_ghash_v8:
 	vext.8	q7,q9,q9,#8
 	vext.8	q3,q8,q8,#8
 	veor	q0,q1,q10
-.byte	0x8e,0x8e,0xa8,0xf2	@ pmull q4,q12,q7		@ H·Ii+1
+	INST(0x8e,0x8e,0xa8,0xf2)	@ pmull q4,q12,q7		@ H·Ii+1
 	veor	q3,q3,q2		@ accumulate q3 early
 
 	vext.8	q10,q0,q0,#8		@ 2nd phase of reduction
-.byte	0x26,0x0e,0xa0,0xf2	@ pmull q0,q0,q11
+	INST(0x26,0x0e,0xa0,0xf2)	@ pmull q0,q0,q11
 	veor	q3,q3,q10
 	veor	q9,q9,q7		@ Karatsuba pre-processing
 	veor	q3,q3,q0
-.byte	0x8f,0xce,0xa9,0xf2	@ pmull2 q6,q12,q7
+	INST(0x8f,0xce,0xa9,0xf2)	@ pmull2 q6,q12,q7
 	bhs	.Loop_mod2x_v8		@ there was at least 32 more bytes
 
 	veor	q2,q2,q10
@@ -198,23 +208,23 @@ gcm_ghash_v8:
 	veor	q3,q3,q0		@ inp^=Xi
 	veor	q9,q8,q10		@ q9 is rotated inp^Xi
 
-.byte	0x86,0x0e,0xa8,0xf2	@ pmull q0,q12,q3		@ H.lo·Xi.lo
+	INST(0x86,0x0e,0xa8,0xf2)	@ pmull q0,q12,q3		@ H.lo·Xi.lo
 	veor	q9,q9,q3		@ Karatsuba pre-processing
-.byte	0x87,0x4e,0xa9,0xf2	@ pmull2 q2,q12,q3		@ H.hi·Xi.hi
-.byte	0xa2,0x2e,0xaa,0xf2	@ pmull q1,q13,q9		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
+	INST(0x87,0x4e,0xa9,0xf2)	@ pmull2 q2,q12,q3		@ H.hi·Xi.hi
+	INST(0xa2,0x2e,0xaa,0xf2)	@ pmull q1,q13,q9		@ (H.lo+H.hi)·(Xi.lo+Xi.hi)
 
 	vext.8	q9,q0,q2,#8		@ Karatsuba post-processing
 	veor	q10,q0,q2
 	veor	q1,q1,q9
 	veor	q1,q1,q10
-.byte	0x26,0x4e,0xe0,0xf2	@ pmull q10,q0,q11		@ 1st phase of reduction
+	INST(0x26,0x4e,0xe0,0xf2)	@ pmull q10,q0,q11		@ 1st phase of reduction
 
 	vmov	d4,d3		@ Xh|Xm - 256-bit result
 	vmov	d3,d0		@ Xm is rotated Xl
 	veor	q0,q1,q10
 
 	vext.8	q10,q0,q0,#8		@ 2nd phase of reduction
-.byte	0x26,0x0e,0xa0,0xf2	@ pmull q0,q0,q11
+	INST(0x26,0x0e,0xa0,0xf2)	@ pmull q0,q0,q11
 	veor	q10,q10,q2
 	veor	q0,q0,q10
 
diff --git a/sys/crypto/openssl/arm/keccak1600-armv4.S b/sys/crypto/openssl/arm/keccak1600-armv4.S
index ae0cd9cd4b8b3..ddad05acad647 100644
--- a/sys/crypto/openssl/arm/keccak1600-armv4.S
+++ b/sys/crypto/openssl/arm/keccak1600-armv4.S
@@ -1,8 +1,6 @@
 /* Do not modify. This file is auto-generated from keccak1600-armv4.pl. */
 #include "arm_arch.h"
 
-.text
-
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -10,6 +8,8 @@
 .code	32
 #endif
 
+.text
+
 .type	iotas32, %object
 .align	5
 iotas32:
@@ -1826,7 +1826,14 @@ KeccakF1600_enter:
 #endif
 	blo	.Lround2x
 
+#if __ARM_ARCH__>=5
 	ldr	pc,[sp,#440]
+#else
+	ldr	lr,[sp,#440]
+	tst	lr,#1
+	moveq	pc,lr		@ be binary compatible with V4, yet
+.word	0xe12fff1e		@ interoperable with Thumb ISA:-)
+#endif
 .size	KeccakF1600_int,.-KeccakF1600_int
 
 .type	KeccakF1600, %function
@@ -1865,7 +1872,14 @@ KeccakF1600:
 	stmia	r11, {r0,r1,r2,r3,r4,r5,r6,r7,r8,r9}
 
 	add	sp,sp,#440+20
+#if __ARM_ARCH__>=5
 	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,pc}
+#else
+	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,lr}
+	tst	lr,#1
+	moveq	pc,lr		@ be binary compatible with V4, yet
+.word	0xe12fff1e		@ interoperable with Thumb ISA:-)
+#endif
 .size	KeccakF1600,.-KeccakF1600
 .globl	SHA3_absorb
 .type	SHA3_absorb,%function
@@ -2011,7 +2025,14 @@ SHA3_absorb:
 .Labsorb_abort:
 	add	sp,sp,#456+32
 	mov	r0,r12			@ return value
+#if __ARM_ARCH__>=5
 	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
+#else
+	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
+	tst	lr,#1
+	moveq	pc,lr		@ be binary compatible with V4, yet
+.word	0xe12fff1e		@ interoperable with Thumb ISA:-)
+#endif
 .size	SHA3_absorb,.-SHA3_absorb
 .globl	SHA3_squeeze
 .type	SHA3_squeeze,%function
@@ -2156,7 +2177,14 @@ SHA3_squeeze:
 .align	4
 .Lsqueeze_done:
 	add	sp,sp,#24
+#if __ARM_ARCH__>=5
 	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,pc}
+#else
+	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,lr}
+	tst	lr,#1
+	moveq	pc,lr		@ be binary compatible with V4, yet
+.word	0xe12fff1e		@ interoperable with Thumb ISA:-)
+#endif
 .size	SHA3_squeeze,.-SHA3_squeeze
 #if __ARM_MAX_ARCH__>=7
 .fpu	neon
@@ -2362,7 +2390,7 @@ KeccakF1600_neon:
 	subs	r3, r3, #1
 	bne	.Loop_neon
 
-.word	0xe12fff1e
+	bx	lr
 .size	KeccakF1600_neon,.-KeccakF1600_neon
 
 .globl	SHA3_absorb_neon
diff --git a/sys/crypto/openssl/arm/poly1305-armv4.S b/sys/crypto/openssl/arm/poly1305-armv4.S
index 64d0ef9b5a38a..41b7caf06cc66 100644
--- a/sys/crypto/openssl/arm/poly1305-armv4.S
+++ b/sys/crypto/openssl/arm/poly1305-armv4.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from poly1305-armv4.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -9,6 +8,8 @@
 .code	32
 #endif
 
+.text
+
 .globl	poly1305_emit
 .globl	poly1305_blocks
 .globl	poly1305_init
@@ -53,8 +54,10 @@ poly1305_init:
 	and	r4,r4,r10
 
 #if	__ARM_MAX_ARCH__>=7
+# if !defined(_WIN32)
 	ldr	r12,[r11,r12]		@ OPENSSL_armcap_P
-# ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r12,[r12]
 # endif
 #endif
@@ -69,32 +72,22 @@ poly1305_init:
 
 #if	__ARM_MAX_ARCH__>=7
 	tst	r12,#ARMV7_NEON		@ check for NEON
-# ifdef	__APPLE__
-	adr	r9,poly1305_blocks_neon
-	adr	r11,poly1305_blocks
-#  ifdef __thumb2__
-	it	ne
-#  endif
+# ifdef	__thumb2__
+	adr	r9,.Lpoly1305_blocks_neon
+	adr	r11,.Lpoly1305_blocks
+	adr	r12,.Lpoly1305_emit
+	adr	r10,.Lpoly1305_emit_neon
+	itt	ne
 	movne	r11,r9
-	adr	r12,poly1305_emit
-	adr	r10,poly1305_emit_neon
-#  ifdef __thumb2__
-	it	ne
-#  endif
 	movne	r12,r10
+	orr	r11,r11,#1	@ thumb-ify address
+	orr	r12,r12,#1
 # else
-#  ifdef __thumb2__
-	itete	eq
-#  endif
 	addeq	r12,r11,#(.Lpoly1305_emit-.Lpoly1305_init)
 	addne	r12,r11,#(.Lpoly1305_emit_neon-.Lpoly1305_init)
 	addeq	r11,r11,#(.Lpoly1305_blocks-.Lpoly1305_init)
 	addne	r11,r11,#(.Lpoly1305_blocks_neon-.Lpoly1305_init)
 # endif
-# ifdef	__thumb2__
-	orr	r12,r12,#1	@ thumb-ify address
-	orr	r11,r11,#1
-# endif
 #endif
 	ldrb	r9,[r1,#11]
 	orr	r6,r6,r7,lsl#8
@@ -1162,7 +1155,11 @@ poly1305_emit_neon:
 .Lzeros:
 .long	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.Lpoly1305_init
+# endif
 #endif
 .byte	80,111,108,121,49,51,48,53,32,102,111,114,32,65,82,77,118,52,47,78,69,79,78,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
diff --git a/sys/crypto/openssl/arm/sha1-armv4-large.S b/sys/crypto/openssl/arm/sha1-armv4-large.S
index 2492c049a2a0a..c23e1df6e38ed 100644
--- a/sys/crypto/openssl/arm/sha1-armv4-large.S
+++ b/sys/crypto/openssl/arm/sha1-armv4-large.S
@@ -1,7 +1,6 @@
 /* Do not modify. This file is auto-generated from sha1-armv4-large.pl. */
 #include "arm_arch.h"
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -9,6 +8,8 @@
 .code	32
 #endif
 
+.text
+
 .globl	sha1_block_data_order
 .type	sha1_block_data_order,%function
 
@@ -16,12 +17,14 @@
 sha1_block_data_order:
 #if __ARM_MAX_ARCH__>=7
 .Lsha1_block:
-	adr	r3,.Lsha1_block
 	ldr	r12,.LOPENSSL_armcap
+# if !defined(_WIN32)
+	adr	r3,.Lsha1_block
 	ldr	r12,[r3,r12]		@ OPENSSL_armcap_P
-#ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r12,[r12]
-#endif
+# endif
 	tst	r12,#ARMV8_SHA1
 	bne	.LARMv8
 	tst	r12,#ARMV7_NEON
@@ -486,7 +489,11 @@ sha1_block_data_order:
 .LK_60_79:.word	0xca62c1d6
 #if __ARM_MAX_ARCH__>=7
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.Lsha1_block
+# endif
 #endif
 .byte	83,72,65,49,32,98,108,111,99,107,32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,65,82,77,118,52,47,78,69,79,78,47,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .align	2
diff --git a/sys/crypto/openssl/arm/sha256-armv4.S b/sys/crypto/openssl/arm/sha256-armv4.S
index 5b541faf07f7d..c1b775f767c8a 100644
--- a/sys/crypto/openssl/arm/sha256-armv4.S
+++ b/sys/crypto/openssl/arm/sha256-armv4.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from sha256-armv4.pl. */
 @ Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
 @
-@ Licensed under the OpenSSL license (the "License").  You may not use
+@ Licensed under the Apache License 2.0 (the "License").  You may not use
 @ this file except in compliance with the License.  You can obtain a copy
 @ in the file LICENSE in the source distribution or at
 @ https://www.openssl.org/source/license.html
@@ -44,6 +44,8 @@
 @
 @ Add ARMv8 code path performing at 2.0 cpb on Apple A7.
 
+@ $output is the last argument if it looks like a file (it has an extension)
+@ $flavour is the first argument if it doesn't look like a file
 #ifndef __KERNEL__
 # include "arm_arch.h"
 #else
@@ -51,7 +53,6 @@
 # define __ARM_MAX_ARCH__ 7
 #endif
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -59,6 +60,8 @@
 .code	32
 #endif
 
+.text
+
 .type	K256,%object
 .align	5
 K256:
@@ -82,7 +85,11 @@ K256:
 .word	0				@ terminator
 #if __ARM_MAX_ARCH__>=7 && !defined(__KERNEL__)
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.Lsha256_block_data_order
+# endif
 #endif
 .align	5
 
@@ -97,10 +104,12 @@ sha256_block_data_order:
 #endif
 #if __ARM_MAX_ARCH__>=7 && !defined(__KERNEL__)
 	ldr	r12,.LOPENSSL_armcap
+# if !defined(_WIN32)
 	ldr	r12,[r3,r12]		@ OPENSSL_armcap_P
-#ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r12,[r12]
-#endif
+# endif
 	tst	r12,#ARMV8_SHA256
 	bne	.LARMv8
 	tst	r12,#ARMV7_NEON
diff --git a/sys/crypto/openssl/arm/sha512-armv4.S b/sys/crypto/openssl/arm/sha512-armv4.S
index f1522a8cbff73..1f84a8ff93fe9 100644
--- a/sys/crypto/openssl/arm/sha512-armv4.S
+++ b/sys/crypto/openssl/arm/sha512-armv4.S
@@ -1,7 +1,7 @@
 /* Do not modify. This file is auto-generated from sha512-armv4.pl. */
 @ Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
 @
-@ Licensed under the OpenSSL license (the "License").  You may not use
+@ Licensed under the Apache License 2.0 (the "License").  You may not use
 @ this file except in compliance with the License.  You can obtain a copy
 @ in the file LICENSE in the source distribution or at
 @ https://www.openssl.org/source/license.html
@@ -74,7 +74,6 @@
 # define WORD64(hi0,lo0,hi1,lo1)	.word	hi0,lo0, hi1,lo1
 #endif
 
-.text
 #if defined(__thumb2__)
 .syntax	unified
 .thumb
@@ -83,6 +82,8 @@
 .code	32
 #endif
 
+.text
+
 .type	K512,%object
 .align	5
 K512:
@@ -129,7 +130,11 @@ K512:
 .size	K512,.-K512
 #if __ARM_MAX_ARCH__>=7 && !defined(__KERNEL__)
 .LOPENSSL_armcap:
+# ifdef	_WIN32
+.word	OPENSSL_armcap_P
+# else
 .word	OPENSSL_armcap_P-.Lsha512_block_data_order
+# endif
 .skip	32-4
 #else
 .skip	32
@@ -146,10 +151,12 @@ sha512_block_data_order:
 #endif
 #if __ARM_MAX_ARCH__>=7 && !defined(__KERNEL__)
 	ldr	r12,.LOPENSSL_armcap
+# if !defined(_WIN32)
 	ldr	r12,[r3,r12]		@ OPENSSL_armcap_P
-#ifdef	__APPLE__
+# endif
+# if defined(__APPLE__) || defined(_WIN32)
 	ldr	r12,[r12]
-#endif
+# endif
 	tst	r12,#ARMV7_NEON
 	bne	.LNEON
 #endif
diff --git a/sys/crypto/openssl/i386/aes-586.S b/sys/crypto/openssl/i386/aes-586.S
new file mode 100644
index 0000000000000..861ee21e8400e
--- /dev/null
+++ b/sys/crypto/openssl/i386/aes-586.S
@@ -0,0 +1,6644 @@
+/* Do not modify. This file is auto-generated from aes-586.pl. */
+#ifdef PIC
+.text
+.type	_x86_AES_encrypt_compact,@function
+.align	16
+_x86_AES_encrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+.align	16
+.L000loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ch,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$8,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$24,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%ecx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	xorl	%esi,%ecx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%ecx
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%ecx
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%edx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	xorl	%esi,%edx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%edx
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%edx
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%eax,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	xorl	%esi,%eax
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%eax
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%eax
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%ebx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	xorl	%esi,%ebx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%ebx
+	xorl	%edi,%esi
+	xorl	%esi,%ebx
+	movl	20(%esp),%edi
+	movl	28(%esp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L000loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ch,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$8,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$24,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+
+	xorl	16(%edi),%eax
+	xorl	20(%edi),%ebx
+	xorl	24(%edi),%ecx
+	xorl	28(%edi),%edx
+	ret
+.size	_x86_AES_encrypt_compact,.-_x86_AES_encrypt_compact
+.type	_sse_AES_encrypt_compact,@function
+.align	16
+_sse_AES_encrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	$454761243,%eax
+	movl	%eax,8(%esp)
+	movl	%eax,12(%esp)
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+.align	16
+.L001loop:
+	pshufw	$8,%mm0,%mm1
+	pshufw	$13,%mm4,%mm5
+	movd	%mm1,%eax
+	movd	%mm5,%ebx
+	movl	%edi,20(%esp)
+	movzbl	%al,%esi
+	movzbl	%ah,%edx
+	pshufw	$13,%mm0,%mm2
+	movzbl	-128(%ebp,%esi,1),%ecx
+	movzbl	%bl,%edi
+	movzbl	-128(%ebp,%edx,1),%edx
+	shrl	$16,%eax
+	shll	$8,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$16,%esi
+	pshufw	$8,%mm4,%mm6
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$24,%esi
+	shrl	$16,%ebx
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$8,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	movd	%mm2,%eax
+	movd	%ecx,%mm0
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%ah,%edi
+	shll	$16,%ecx
+	movd	%mm6,%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	shll	$8,%esi
+	shrl	$16,%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shrl	$16,%eax
+	movd	%ecx,%mm1
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%ah,%edi
+	shll	$16,%ecx
+	andl	$255,%eax
+	orl	%esi,%ecx
+	punpckldq	%mm1,%mm0
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%eax,1),%eax
+	orl	%esi,%ecx
+	shll	$16,%eax
+	movzbl	-128(%ebp,%edi,1),%esi
+	orl	%eax,%edx
+	shll	$8,%esi
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	orl	%esi,%ecx
+	orl	%ebx,%edx
+	movl	20(%esp),%edi
+	movd	%ecx,%mm4
+	movd	%edx,%mm5
+	punpckldq	%mm5,%mm4
+	addl	$16,%edi
+	cmpl	24(%esp),%edi
+	ja	.L002out
+	movq	8(%esp),%mm2
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	movq	%mm0,%mm1
+	movq	%mm4,%mm5
+	pcmpgtb	%mm0,%mm3
+	pcmpgtb	%mm4,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	pshufw	$177,%mm0,%mm2
+	pshufw	$177,%mm4,%mm6
+	paddb	%mm0,%mm0
+	paddb	%mm4,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pshufw	$177,%mm2,%mm3
+	pshufw	$177,%mm6,%mm7
+	pxor	%mm0,%mm1
+	pxor	%mm4,%mm5
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	%mm3,%mm2
+	movq	%mm7,%mm6
+	pslld	$8,%mm3
+	pslld	$8,%mm7
+	psrld	$24,%mm2
+	psrld	$24,%mm6
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	movq	(%edi),%mm2
+	movq	8(%edi),%mm6
+	psrld	$8,%mm1
+	psrld	$8,%mm5
+	movl	-128(%ebp),%eax
+	pslld	$24,%mm3
+	pslld	$24,%mm7
+	movl	-64(%ebp),%ebx
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movl	(%ebp),%ecx
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movl	64(%ebp),%edx
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	jmp	.L001loop
+.align	16
+.L002out:
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	ret
+.size	_sse_AES_encrypt_compact,.-_sse_AES_encrypt_compact
+.type	_x86_AES_encrypt,@function
+.align	16
+_x86_AES_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+.align	16
+.L003loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%bh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%ch,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%dh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movzbl	%bh,%edi
+	xorl	1(%ebp,%edi,8),%esi
+
+	movl	20(%esp),%edi
+	movl	(%ebp,%edx,8),%edx
+	movzbl	%ah,%eax
+	xorl	3(%ebp,%eax,8),%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	xorl	2(%ebp,%ebx,8),%edx
+	movl	8(%esp),%ebx
+	xorl	1(%ebp,%ecx,8),%edx
+	movl	%esi,%ecx
+
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L003loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%bh,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%ch,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%dh,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movl	2(%ebp,%edx,8),%edx
+	andl	$255,%edx
+	movzbl	%ah,%eax
+	movl	(%ebp,%eax,8),%eax
+	andl	$65280,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movl	(%ebp,%ebx,8),%ebx
+	andl	$16711680,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movl	2(%ebp,%ecx,8),%ecx
+	andl	$4278190080,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	ret
+.align	64
+.LAES_Te:
+.long	2774754246,2774754246
+.long	2222750968,2222750968
+.long	2574743534,2574743534
+.long	2373680118,2373680118
+.long	234025727,234025727
+.long	3177933782,3177933782
+.long	2976870366,2976870366
+.long	1422247313,1422247313
+.long	1345335392,1345335392
+.long	50397442,50397442
+.long	2842126286,2842126286
+.long	2099981142,2099981142
+.long	436141799,436141799
+.long	1658312629,1658312629
+.long	3870010189,3870010189
+.long	2591454956,2591454956
+.long	1170918031,1170918031
+.long	2642575903,2642575903
+.long	1086966153,1086966153
+.long	2273148410,2273148410
+.long	368769775,368769775
+.long	3948501426,3948501426
+.long	3376891790,3376891790
+.long	200339707,200339707
+.long	3970805057,3970805057
+.long	1742001331,1742001331
+.long	4255294047,4255294047
+.long	3937382213,3937382213
+.long	3214711843,3214711843
+.long	4154762323,4154762323
+.long	2524082916,2524082916
+.long	1539358875,1539358875
+.long	3266819957,3266819957
+.long	486407649,486407649
+.long	2928907069,2928907069
+.long	1780885068,1780885068
+.long	1513502316,1513502316
+.long	1094664062,1094664062
+.long	49805301,49805301
+.long	1338821763,1338821763
+.long	1546925160,1546925160
+.long	4104496465,4104496465
+.long	887481809,887481809
+.long	150073849,150073849
+.long	2473685474,2473685474
+.long	1943591083,1943591083
+.long	1395732834,1395732834
+.long	1058346282,1058346282
+.long	201589768,201589768
+.long	1388824469,1388824469
+.long	1696801606,1696801606
+.long	1589887901,1589887901
+.long	672667696,672667696
+.long	2711000631,2711000631
+.long	251987210,251987210
+.long	3046808111,3046808111
+.long	151455502,151455502
+.long	907153956,907153956
+.long	2608889883,2608889883
+.long	1038279391,1038279391
+.long	652995533,652995533
+.long	1764173646,1764173646
+.long	3451040383,3451040383
+.long	2675275242,2675275242
+.long	453576978,453576978
+.long	2659418909,2659418909
+.long	1949051992,1949051992
+.long	773462580,773462580
+.long	756751158,756751158
+.long	2993581788,2993581788
+.long	3998898868,3998898868
+.long	4221608027,4221608027
+.long	4132590244,4132590244
+.long	1295727478,1295727478
+.long	1641469623,1641469623
+.long	3467883389,3467883389
+.long	2066295122,2066295122
+.long	1055122397,1055122397
+.long	1898917726,1898917726
+.long	2542044179,2542044179
+.long	4115878822,4115878822
+.long	1758581177,1758581177
+.long	0,0
+.long	753790401,753790401
+.long	1612718144,1612718144
+.long	536673507,536673507
+.long	3367088505,3367088505
+.long	3982187446,3982187446
+.long	3194645204,3194645204
+.long	1187761037,1187761037
+.long	3653156455,3653156455
+.long	1262041458,1262041458
+.long	3729410708,3729410708
+.long	3561770136,3561770136
+.long	3898103984,3898103984
+.long	1255133061,1255133061
+.long	1808847035,1808847035
+.long	720367557,720367557
+.long	3853167183,3853167183
+.long	385612781,385612781
+.long	3309519750,3309519750
+.long	3612167578,3612167578
+.long	1429418854,1429418854
+.long	2491778321,2491778321
+.long	3477423498,3477423498
+.long	284817897,284817897
+.long	100794884,100794884
+.long	2172616702,2172616702
+.long	4031795360,4031795360
+.long	1144798328,1144798328
+.long	3131023141,3131023141
+.long	3819481163,3819481163
+.long	4082192802,4082192802
+.long	4272137053,4272137053
+.long	3225436288,3225436288
+.long	2324664069,2324664069
+.long	2912064063,2912064063
+.long	3164445985,3164445985
+.long	1211644016,1211644016
+.long	83228145,83228145
+.long	3753688163,3753688163
+.long	3249976951,3249976951
+.long	1977277103,1977277103
+.long	1663115586,1663115586
+.long	806359072,806359072
+.long	452984805,452984805
+.long	250868733,250868733
+.long	1842533055,1842533055
+.long	1288555905,1288555905
+.long	336333848,336333848
+.long	890442534,890442534
+.long	804056259,804056259
+.long	3781124030,3781124030
+.long	2727843637,2727843637
+.long	3427026056,3427026056
+.long	957814574,957814574
+.long	1472513171,1472513171
+.long	4071073621,4071073621
+.long	2189328124,2189328124
+.long	1195195770,1195195770
+.long	2892260552,2892260552
+.long	3881655738,3881655738
+.long	723065138,723065138
+.long	2507371494,2507371494
+.long	2690670784,2690670784
+.long	2558624025,2558624025
+.long	3511635870,3511635870
+.long	2145180835,2145180835
+.long	1713513028,1713513028
+.long	2116692564,2116692564
+.long	2878378043,2878378043
+.long	2206763019,2206763019
+.long	3393603212,3393603212
+.long	703524551,703524551
+.long	3552098411,3552098411
+.long	1007948840,1007948840
+.long	2044649127,2044649127
+.long	3797835452,3797835452
+.long	487262998,487262998
+.long	1994120109,1994120109
+.long	1004593371,1004593371
+.long	1446130276,1446130276
+.long	1312438900,1312438900
+.long	503974420,503974420
+.long	3679013266,3679013266
+.long	168166924,168166924
+.long	1814307912,1814307912
+.long	3831258296,3831258296
+.long	1573044895,1573044895
+.long	1859376061,1859376061
+.long	4021070915,4021070915
+.long	2791465668,2791465668
+.long	2828112185,2828112185
+.long	2761266481,2761266481
+.long	937747667,937747667
+.long	2339994098,2339994098
+.long	854058965,854058965
+.long	1137232011,1137232011
+.long	1496790894,1496790894
+.long	3077402074,3077402074
+.long	2358086913,2358086913
+.long	1691735473,1691735473
+.long	3528347292,3528347292
+.long	3769215305,3769215305
+.long	3027004632,3027004632
+.long	4199962284,4199962284
+.long	133494003,133494003
+.long	636152527,636152527
+.long	2942657994,2942657994
+.long	2390391540,2390391540
+.long	3920539207,3920539207
+.long	403179536,403179536
+.long	3585784431,3585784431
+.long	2289596656,2289596656
+.long	1864705354,1864705354
+.long	1915629148,1915629148
+.long	605822008,605822008
+.long	4054230615,4054230615
+.long	3350508659,3350508659
+.long	1371981463,1371981463
+.long	602466507,602466507
+.long	2094914977,2094914977
+.long	2624877800,2624877800
+.long	555687742,555687742
+.long	3712699286,3712699286
+.long	3703422305,3703422305
+.long	2257292045,2257292045
+.long	2240449039,2240449039
+.long	2423288032,2423288032
+.long	1111375484,1111375484
+.long	3300242801,3300242801
+.long	2858837708,2858837708
+.long	3628615824,3628615824
+.long	84083462,84083462
+.long	32962295,32962295
+.long	302911004,302911004
+.long	2741068226,2741068226
+.long	1597322602,1597322602
+.long	4183250862,4183250862
+.long	3501832553,3501832553
+.long	2441512471,2441512471
+.long	1489093017,1489093017
+.long	656219450,656219450
+.long	3114180135,3114180135
+.long	954327513,954327513
+.long	335083755,335083755
+.long	3013122091,3013122091
+.long	856756514,856756514
+.long	3144247762,3144247762
+.long	1893325225,1893325225
+.long	2307821063,2307821063
+.long	2811532339,2811532339
+.long	3063651117,3063651117
+.long	572399164,572399164
+.long	2458355477,2458355477
+.long	552200649,552200649
+.long	1238290055,1238290055
+.long	4283782570,4283782570
+.long	2015897680,2015897680
+.long	2061492133,2061492133
+.long	2408352771,2408352771
+.long	4171342169,4171342169
+.long	2156497161,2156497161
+.long	386731290,386731290
+.long	3669999461,3669999461
+.long	837215959,837215959
+.long	3326231172,3326231172
+.long	3093850320,3093850320
+.long	3275833730,3275833730
+.long	2962856233,2962856233
+.long	1999449434,1999449434
+.long	286199582,286199582
+.long	3417354363,3417354363
+.long	4233385128,4233385128
+.long	3602627437,3602627437
+.long	974525996,974525996
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.long	1,2,4,8
+.long	16,32,64,128
+.long	27,54,0,0
+.long	0,0,0,0
+.size	_x86_AES_encrypt,.-_x86_AES_encrypt
+.globl	AES_encrypt
+.type	AES_encrypt,@function
+.align	16
+AES_encrypt:
+.L_AES_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	28(%esp),%edi
+	movl	%esp,%eax
+	subl	$36,%esp
+	andl	$-64,%esp
+	leal	-127(%edi),%ebx
+	subl	%esp,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esp
+	addl	$4,%esp
+	movl	%eax,28(%esp)
+	call	.L004pic_point
+.L004pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P-.L004pic_point(%ebp),%eax
+	leal	.LAES_Te-.L004pic_point(%ebp),%ebp
+	leal	764(%esp),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	btl	$25,(%eax)
+	jnc	.L005x86
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	call	_sse_AES_encrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	16
+.L005x86:
+	movl	%ebp,24(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	call	_x86_AES_encrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_encrypt,.-.L_AES_encrypt_begin
+.type	_x86_AES_decrypt_compact,@function
+.align	16
+_x86_AES_decrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+.align	16
+.L006loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	shrl	$24,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%eax
+	subl	%edi,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%esi,%eax
+	movl	$2155905152,%edi
+	andl	%eax,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%eax,%eax,1),%ebx
+	subl	%edi,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%ecx,%eax
+	xorl	%esi,%ebx
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%ecx,%ebx
+	roll	$8,%ecx
+	xorl	%esi,%ebp
+	xorl	%eax,%ecx
+	xorl	%ebp,%eax
+	xorl	%ebx,%ecx
+	xorl	%ebp,%ebx
+	roll	$24,%eax
+	xorl	%ebp,%ecx
+	roll	$16,%ebx
+	xorl	%eax,%ecx
+	roll	$8,%ebp
+	xorl	%ebx,%ecx
+	movl	4(%esp),%eax
+	xorl	%ebp,%ecx
+	movl	%ecx,12(%esp)
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebx
+	subl	%edi,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%esi,%ebx
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%edx,%ebx
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%edx,%ecx
+	roll	$8,%edx
+	xorl	%esi,%ebp
+	xorl	%ebx,%edx
+	xorl	%ebp,%ebx
+	xorl	%ecx,%edx
+	xorl	%ebp,%ecx
+	roll	$24,%ebx
+	xorl	%ebp,%edx
+	roll	$16,%ecx
+	xorl	%ebx,%edx
+	roll	$8,%ebp
+	xorl	%ecx,%edx
+	movl	8(%esp),%ebx
+	xorl	%ebp,%edx
+	movl	%edx,16(%esp)
+	movl	$2155905152,%edi
+	andl	%eax,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%eax,%eax,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%edx
+	subl	%edi,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%eax,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%eax,%edx
+	roll	$8,%eax
+	xorl	%esi,%ebp
+	xorl	%ecx,%eax
+	xorl	%ebp,%ecx
+	xorl	%edx,%eax
+	xorl	%ebp,%edx
+	roll	$24,%ecx
+	xorl	%ebp,%eax
+	roll	$16,%edx
+	xorl	%ecx,%eax
+	roll	$8,%ebp
+	xorl	%edx,%eax
+	xorl	%ebp,%eax
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%edx
+	subl	%edi,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%ebx,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%ebx,%edx
+	roll	$8,%ebx
+	xorl	%esi,%ebp
+	xorl	%ecx,%ebx
+	xorl	%ebp,%ecx
+	xorl	%edx,%ebx
+	xorl	%ebp,%edx
+	roll	$24,%ecx
+	xorl	%ebp,%ebx
+	roll	$16,%edx
+	xorl	%ecx,%ebx
+	roll	$8,%ebp
+	xorl	%edx,%ebx
+	movl	12(%esp),%ecx
+	xorl	%ebp,%ebx
+	movl	16(%esp),%edx
+	movl	20(%esp),%edi
+	movl	28(%esp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L006loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	xorl	16(%edi),%eax
+	xorl	20(%edi),%ebx
+	xorl	24(%edi),%ecx
+	xorl	28(%edi),%edx
+	ret
+.size	_x86_AES_decrypt_compact,.-_x86_AES_decrypt_compact
+.type	_sse_AES_decrypt_compact,@function
+.align	16
+_sse_AES_decrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	$454761243,%eax
+	movl	%eax,8(%esp)
+	movl	%eax,12(%esp)
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+.align	16
+.L007loop:
+	pshufw	$12,%mm0,%mm1
+	pshufw	$9,%mm4,%mm5
+	movd	%mm1,%eax
+	movd	%mm5,%ebx
+	movl	%edi,20(%esp)
+	movzbl	%al,%esi
+	movzbl	%ah,%edx
+	pshufw	$6,%mm0,%mm2
+	movzbl	-128(%ebp,%esi,1),%ecx
+	movzbl	%bl,%edi
+	movzbl	-128(%ebp,%edx,1),%edx
+	shrl	$16,%eax
+	shll	$8,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$16,%esi
+	pshufw	$3,%mm4,%mm6
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$24,%esi
+	shrl	$16,%ebx
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$8,%esi
+	movd	%mm2,%eax
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	shll	$16,%esi
+	movd	%mm6,%ebx
+	movd	%ecx,%mm0
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%al,%edi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$16,%esi
+	shrl	$16,%eax
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shrl	$16,%ebx
+	shll	$8,%esi
+	movd	%edx,%mm1
+	movzbl	-128(%ebp,%edi,1),%edx
+	movzbl	%bh,%edi
+	shll	$24,%edx
+	andl	$255,%ebx
+	orl	%esi,%edx
+	punpckldq	%mm1,%mm0
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$8,%esi
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	orl	%ebx,%edx
+	shll	$16,%esi
+	movzbl	-128(%ebp,%eax,1),%eax
+	orl	%esi,%edx
+	shll	$24,%eax
+	orl	%eax,%ecx
+	movl	20(%esp),%edi
+	movd	%edx,%mm4
+	movd	%ecx,%mm5
+	punpckldq	%mm5,%mm4
+	addl	$16,%edi
+	cmpl	24(%esp),%edi
+	ja	.L008out
+	movq	%mm0,%mm3
+	movq	%mm4,%mm7
+	pshufw	$228,%mm0,%mm2
+	pshufw	$228,%mm4,%mm6
+	movq	%mm0,%mm1
+	movq	%mm4,%mm5
+	pshufw	$177,%mm0,%mm0
+	pshufw	$177,%mm4,%mm4
+	pslld	$8,%mm2
+	pslld	$8,%mm6
+	psrld	$8,%mm3
+	psrld	$8,%mm7
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pslld	$16,%mm2
+	pslld	$16,%mm6
+	psrld	$16,%mm3
+	psrld	$16,%mm7
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movq	8(%esp),%mm3
+	pxor	%mm2,%mm2
+	pxor	%mm6,%mm6
+	pcmpgtb	%mm1,%mm2
+	pcmpgtb	%mm5,%mm6
+	pand	%mm3,%mm2
+	pand	%mm3,%mm6
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm2,%mm1
+	pxor	%mm6,%mm5
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	movq	%mm1,%mm2
+	movq	%mm5,%mm6
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pslld	$24,%mm3
+	pslld	$24,%mm7
+	psrld	$8,%mm2
+	psrld	$8,%mm6
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	8(%esp),%mm2
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	pcmpgtb	%mm1,%mm3
+	pcmpgtb	%mm5,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm3,%mm1
+	pxor	%mm7,%mm5
+	pshufw	$177,%mm1,%mm3
+	pshufw	$177,%mm5,%mm7
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	pcmpgtb	%mm1,%mm3
+	pcmpgtb	%mm5,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm3,%mm1
+	pxor	%mm7,%mm5
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	pshufw	$177,%mm1,%mm2
+	pshufw	$177,%mm5,%mm6
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pslld	$8,%mm1
+	pslld	$8,%mm5
+	psrld	$8,%mm3
+	psrld	$8,%mm7
+	movq	(%edi),%mm2
+	movq	8(%edi),%mm6
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movl	-128(%ebp),%eax
+	pslld	$16,%mm1
+	pslld	$16,%mm5
+	movl	-64(%ebp),%ebx
+	psrld	$16,%mm3
+	psrld	$16,%mm7
+	movl	(%ebp),%ecx
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movl	64(%ebp),%edx
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	jmp	.L007loop
+.align	16
+.L008out:
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	ret
+.size	_sse_AES_decrypt_compact,.-_sse_AES_decrypt_compact
+.type	_x86_AES_decrypt,@function
+.align	16
+_x86_AES_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+.align	16
+.L009loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%dh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%ah,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%bh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movl	(%ebp,%edx,8),%edx
+	movzbl	%ch,%ecx
+	xorl	3(%ebp,%ecx,8),%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	xorl	2(%ebp,%ebx,8),%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	xorl	1(%ebp,%eax,8),%edx
+	movl	4(%esp),%eax
+
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L009loop
+	leal	2176(%ebp),%ebp
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+	leal	-128(%ebp),%ebp
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	movzbl	(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	leal	-2048(%ebp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	ret
+.align	64
+.LAES_Td:
+.long	1353184337,1353184337
+.long	1399144830,1399144830
+.long	3282310938,3282310938
+.long	2522752826,2522752826
+.long	3412831035,3412831035
+.long	4047871263,4047871263
+.long	2874735276,2874735276
+.long	2466505547,2466505547
+.long	1442459680,1442459680
+.long	4134368941,4134368941
+.long	2440481928,2440481928
+.long	625738485,625738485
+.long	4242007375,4242007375
+.long	3620416197,3620416197
+.long	2151953702,2151953702
+.long	2409849525,2409849525
+.long	1230680542,1230680542
+.long	1729870373,1729870373
+.long	2551114309,2551114309
+.long	3787521629,3787521629
+.long	41234371,41234371
+.long	317738113,317738113
+.long	2744600205,2744600205
+.long	3338261355,3338261355
+.long	3881799427,3881799427
+.long	2510066197,2510066197
+.long	3950669247,3950669247
+.long	3663286933,3663286933
+.long	763608788,763608788
+.long	3542185048,3542185048
+.long	694804553,694804553
+.long	1154009486,1154009486
+.long	1787413109,1787413109
+.long	2021232372,2021232372
+.long	1799248025,1799248025
+.long	3715217703,3715217703
+.long	3058688446,3058688446
+.long	397248752,397248752
+.long	1722556617,1722556617
+.long	3023752829,3023752829
+.long	407560035,407560035
+.long	2184256229,2184256229
+.long	1613975959,1613975959
+.long	1165972322,1165972322
+.long	3765920945,3765920945
+.long	2226023355,2226023355
+.long	480281086,480281086
+.long	2485848313,2485848313
+.long	1483229296,1483229296
+.long	436028815,436028815
+.long	2272059028,2272059028
+.long	3086515026,3086515026
+.long	601060267,601060267
+.long	3791801202,3791801202
+.long	1468997603,1468997603
+.long	715871590,715871590
+.long	120122290,120122290
+.long	63092015,63092015
+.long	2591802758,2591802758
+.long	2768779219,2768779219
+.long	4068943920,4068943920
+.long	2997206819,2997206819
+.long	3127509762,3127509762
+.long	1552029421,1552029421
+.long	723308426,723308426
+.long	2461301159,2461301159
+.long	4042393587,4042393587
+.long	2715969870,2715969870
+.long	3455375973,3455375973
+.long	3586000134,3586000134
+.long	526529745,526529745
+.long	2331944644,2331944644
+.long	2639474228,2639474228
+.long	2689987490,2689987490
+.long	853641733,853641733
+.long	1978398372,1978398372
+.long	971801355,971801355
+.long	2867814464,2867814464
+.long	111112542,111112542
+.long	1360031421,1360031421
+.long	4186579262,4186579262
+.long	1023860118,1023860118
+.long	2919579357,2919579357
+.long	1186850381,1186850381
+.long	3045938321,3045938321
+.long	90031217,90031217
+.long	1876166148,1876166148
+.long	4279586912,4279586912
+.long	620468249,620468249
+.long	2548678102,2548678102
+.long	3426959497,3426959497
+.long	2006899047,2006899047
+.long	3175278768,3175278768
+.long	2290845959,2290845959
+.long	945494503,945494503
+.long	3689859193,3689859193
+.long	1191869601,1191869601
+.long	3910091388,3910091388
+.long	3374220536,3374220536
+.long	0,0
+.long	2206629897,2206629897
+.long	1223502642,1223502642
+.long	2893025566,2893025566
+.long	1316117100,1316117100
+.long	4227796733,4227796733
+.long	1446544655,1446544655
+.long	517320253,517320253
+.long	658058550,658058550
+.long	1691946762,1691946762
+.long	564550760,564550760
+.long	3511966619,3511966619
+.long	976107044,976107044
+.long	2976320012,2976320012
+.long	266819475,266819475
+.long	3533106868,3533106868
+.long	2660342555,2660342555
+.long	1338359936,1338359936
+.long	2720062561,2720062561
+.long	1766553434,1766553434
+.long	370807324,370807324
+.long	179999714,179999714
+.long	3844776128,3844776128
+.long	1138762300,1138762300
+.long	488053522,488053522
+.long	185403662,185403662
+.long	2915535858,2915535858
+.long	3114841645,3114841645
+.long	3366526484,3366526484
+.long	2233069911,2233069911
+.long	1275557295,1275557295
+.long	3151862254,3151862254
+.long	4250959779,4250959779
+.long	2670068215,2670068215
+.long	3170202204,3170202204
+.long	3309004356,3309004356
+.long	880737115,880737115
+.long	1982415755,1982415755
+.long	3703972811,3703972811
+.long	1761406390,1761406390
+.long	1676797112,1676797112
+.long	3403428311,3403428311
+.long	277177154,277177154
+.long	1076008723,1076008723
+.long	538035844,538035844
+.long	2099530373,2099530373
+.long	4164795346,4164795346
+.long	288553390,288553390
+.long	1839278535,1839278535
+.long	1261411869,1261411869
+.long	4080055004,4080055004
+.long	3964831245,3964831245
+.long	3504587127,3504587127
+.long	1813426987,1813426987
+.long	2579067049,2579067049
+.long	4199060497,4199060497
+.long	577038663,577038663
+.long	3297574056,3297574056
+.long	440397984,440397984
+.long	3626794326,3626794326
+.long	4019204898,4019204898
+.long	3343796615,3343796615
+.long	3251714265,3251714265
+.long	4272081548,4272081548
+.long	906744984,906744984
+.long	3481400742,3481400742
+.long	685669029,685669029
+.long	646887386,646887386
+.long	2764025151,2764025151
+.long	3835509292,3835509292
+.long	227702864,227702864
+.long	2613862250,2613862250
+.long	1648787028,1648787028
+.long	3256061430,3256061430
+.long	3904428176,3904428176
+.long	1593260334,1593260334
+.long	4121936770,4121936770
+.long	3196083615,3196083615
+.long	2090061929,2090061929
+.long	2838353263,2838353263
+.long	3004310991,3004310991
+.long	999926984,999926984
+.long	2809993232,2809993232
+.long	1852021992,1852021992
+.long	2075868123,2075868123
+.long	158869197,158869197
+.long	4095236462,4095236462
+.long	28809964,28809964
+.long	2828685187,2828685187
+.long	1701746150,1701746150
+.long	2129067946,2129067946
+.long	147831841,147831841
+.long	3873969647,3873969647
+.long	3650873274,3650873274
+.long	3459673930,3459673930
+.long	3557400554,3557400554
+.long	3598495785,3598495785
+.long	2947720241,2947720241
+.long	824393514,824393514
+.long	815048134,815048134
+.long	3227951669,3227951669
+.long	935087732,935087732
+.long	2798289660,2798289660
+.long	2966458592,2966458592
+.long	366520115,366520115
+.long	1251476721,1251476721
+.long	4158319681,4158319681
+.long	240176511,240176511
+.long	804688151,804688151
+.long	2379631990,2379631990
+.long	1303441219,1303441219
+.long	1414376140,1414376140
+.long	3741619940,3741619940
+.long	3820343710,3820343710
+.long	461924940,461924940
+.long	3089050817,3089050817
+.long	2136040774,2136040774
+.long	82468509,82468509
+.long	1563790337,1563790337
+.long	1937016826,1937016826
+.long	776014843,776014843
+.long	1511876531,1511876531
+.long	1389550482,1389550482
+.long	861278441,861278441
+.long	323475053,323475053
+.long	2355222426,2355222426
+.long	2047648055,2047648055
+.long	2383738969,2383738969
+.long	2302415851,2302415851
+.long	3995576782,3995576782
+.long	902390199,902390199
+.long	3991215329,3991215329
+.long	1018251130,1018251130
+.long	1507840668,1507840668
+.long	1064563285,1064563285
+.long	2043548696,2043548696
+.long	3208103795,3208103795
+.long	3939366739,3939366739
+.long	1537932639,1537932639
+.long	342834655,342834655
+.long	2262516856,2262516856
+.long	2180231114,2180231114
+.long	1053059257,1053059257
+.long	741614648,741614648
+.long	1598071746,1598071746
+.long	1925389590,1925389590
+.long	203809468,203809468
+.long	2336832552,2336832552
+.long	1100287487,1100287487
+.long	1895934009,1895934009
+.long	3736275976,3736275976
+.long	2632234200,2632234200
+.long	2428589668,2428589668
+.long	1636092795,1636092795
+.long	1890988757,1890988757
+.long	1952214088,1952214088
+.long	1113045200,1113045200
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.size	_x86_AES_decrypt,.-_x86_AES_decrypt
+.globl	AES_decrypt
+.type	AES_decrypt,@function
+.align	16
+AES_decrypt:
+.L_AES_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	28(%esp),%edi
+	movl	%esp,%eax
+	subl	$36,%esp
+	andl	$-64,%esp
+	leal	-127(%edi),%ebx
+	subl	%esp,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esp
+	addl	$4,%esp
+	movl	%eax,28(%esp)
+	call	.L010pic_point
+.L010pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P-.L010pic_point(%ebp),%eax
+	leal	.LAES_Td-.L010pic_point(%ebp),%ebp
+	leal	764(%esp),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	btl	$25,(%eax)
+	jnc	.L011x86
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	call	_sse_AES_decrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	16
+.L011x86:
+	movl	%ebp,24(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	call	_x86_AES_decrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_decrypt,.-.L_AES_decrypt_begin
+.globl	AES_cbc_encrypt
+.type	AES_cbc_encrypt,@function
+.align	16
+AES_cbc_encrypt:
+.L_AES_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ecx
+	cmpl	$0,%ecx
+	je	.L012drop_out
+	call	.L013pic_point
+.L013pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P-.L013pic_point(%ebp),%eax
+	cmpl	$0,40(%esp)
+	leal	.LAES_Te-.L013pic_point(%ebp),%ebp
+	jne	.L014picked_te
+	leal	.LAES_Td-.LAES_Te(%ebp),%ebp
+.L014picked_te:
+	pushfl
+	cld
+	cmpl	$512,%ecx
+	jb	.L015slow_way
+	testl	$15,%ecx
+	jnz	.L015slow_way
+	btl	$28,(%eax)
+	jc	.L015slow_way
+	leal	-324(%esp),%esi
+	andl	$-64,%esi
+	movl	%ebp,%eax
+	leal	2304(%ebp),%ebx
+	movl	%esi,%edx
+	andl	$4095,%eax
+	andl	$4095,%ebx
+	andl	$4095,%edx
+	cmpl	%ebx,%edx
+	jb	.L016tbl_break_out
+	subl	%ebx,%edx
+	subl	%edx,%esi
+	jmp	.L017tbl_ok
+.align	4
+.L016tbl_break_out:
+	subl	%eax,%edx
+	andl	$4095,%edx
+	addl	$384,%edx
+	subl	%edx,%esi
+.align	4
+.L017tbl_ok:
+	leal	24(%esp),%edx
+	xchgl	%esi,%esp
+	addl	$4,%esp
+	movl	%ebp,24(%esp)
+	movl	%esi,28(%esp)
+	movl	(%edx),%eax
+	movl	4(%edx),%ebx
+	movl	12(%edx),%edi
+	movl	16(%edx),%esi
+	movl	20(%edx),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,40(%esp)
+	movl	%edi,44(%esp)
+	movl	%esi,48(%esp)
+	movl	$0,316(%esp)
+	movl	%edi,%ebx
+	movl	$61,%ecx
+	subl	%ebp,%ebx
+	movl	%edi,%esi
+	andl	$4095,%ebx
+	leal	76(%esp),%edi
+	cmpl	$2304,%ebx
+	jb	.L018do_copy
+	cmpl	$3852,%ebx
+	jb	.L019skip_copy
+.align	4
+.L018do_copy:
+	movl	%edi,44(%esp)
+.long	2784229001
+.L019skip_copy:
+	movl	$16,%edi
+.align	4
+.L020prefetch_tbl:
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%esi
+	leal	128(%ebp),%ebp
+	subl	$1,%edi
+	jnz	.L020prefetch_tbl
+	subl	$2048,%ebp
+	movl	32(%esp),%esi
+	movl	48(%esp),%edi
+	cmpl	$0,%edx
+	je	.L021fast_decrypt
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+.align	16
+.L022fast_enc_loop:
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	xorl	(%esi),%eax
+	xorl	4(%esi),%ebx
+	xorl	8(%esi),%ecx
+	xorl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_encrypt
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	leal	16(%esi),%esi
+	movl	40(%esp),%ecx
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L022fast_enc_loop
+	movl	48(%esp),%esi
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	cmpl	$0,316(%esp)
+	movl	44(%esp),%edi
+	je	.L023skip_ezero
+	movl	$60,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2884892297
+.L023skip_ezero:
+	movl	28(%esp),%esp
+	popfl
+.L012drop_out:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L021fast_decrypt:
+	cmpl	36(%esp),%esi
+	je	.L024fast_dec_in_place
+	movl	%edi,52(%esp)
+.align	4
+.align	16
+.L025fast_dec_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt
+	movl	52(%esp),%edi
+	movl	40(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	36(%esp),%edi
+	movl	32(%esp),%esi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	movl	%esi,52(%esp)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edi
+	movl	%edi,36(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L025fast_dec_loop
+	movl	52(%esp),%edi
+	movl	48(%esp),%esi
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	jmp	.L026fast_dec_out
+.align	16
+.L024fast_dec_in_place:
+.L027fast_dec_in_place_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	leal	60(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt
+	movl	48(%esp),%edi
+	movl	36(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	leal	16(%esi),%esi
+	movl	%esi,36(%esp)
+	leal	60(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	32(%esp),%esi
+	movl	40(%esp),%ecx
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L027fast_dec_in_place_loop
+.align	4
+.L026fast_dec_out:
+	cmpl	$0,316(%esp)
+	movl	44(%esp),%edi
+	je	.L028skip_dzero
+	movl	$60,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2884892297
+.L028skip_dzero:
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L015slow_way:
+	movl	(%eax),%eax
+	movl	36(%esp),%edi
+	leal	-80(%esp),%esi
+	andl	$-64,%esi
+	leal	-143(%edi),%ebx
+	subl	%esi,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esi
+	leal	768(%esi),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	leal	24(%esp),%edx
+	xchgl	%esi,%esp
+	addl	$4,%esp
+	movl	%ebp,24(%esp)
+	movl	%esi,28(%esp)
+	movl	%eax,52(%esp)
+	movl	(%edx),%eax
+	movl	4(%edx),%ebx
+	movl	16(%edx),%esi
+	movl	20(%edx),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,40(%esp)
+	movl	%edi,44(%esp)
+	movl	%esi,48(%esp)
+	movl	%esi,%edi
+	movl	%eax,%esi
+	cmpl	$0,%edx
+	je	.L029slow_decrypt
+	cmpl	$16,%ecx
+	movl	%ebx,%edx
+	jb	.L030slow_enc_tail
+	btl	$25,52(%esp)
+	jnc	.L031slow_enc_x86
+	movq	(%edi),%mm0
+	movq	8(%edi),%mm4
+.align	16
+.L032slow_enc_loop_sse:
+	pxor	(%esi),%mm0
+	pxor	8(%esi),%mm4
+	movl	44(%esp),%edi
+	call	_sse_AES_encrypt_compact
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	40(%esp),%ecx
+	movq	%mm0,(%edi)
+	movq	%mm4,8(%edi)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	cmpl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jae	.L032slow_enc_loop_sse
+	testl	$15,%ecx
+	jnz	.L030slow_enc_tail
+	movl	48(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L031slow_enc_x86:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+.align	4
+.L033slow_enc_loop_x86:
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	xorl	(%esi),%eax
+	xorl	4(%esi),%ebx
+	xorl	8(%esi),%ecx
+	xorl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_encrypt_compact
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	cmpl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jae	.L033slow_enc_loop_x86
+	testl	$15,%ecx
+	jnz	.L030slow_enc_tail
+	movl	48(%esp),%esi
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L030slow_enc_tail:
+	emms
+	movl	%edx,%edi
+	movl	$16,%ebx
+	subl	%ecx,%ebx
+	cmpl	%esi,%edi
+	je	.L034enc_in_place
+.align	4
+.long	2767451785
+	jmp	.L035enc_skip_in_place
+.L034enc_in_place:
+	leal	(%edi,%ecx,1),%edi
+.L035enc_skip_in_place:
+	movl	%ebx,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2868115081
+	movl	48(%esp),%edi
+	movl	%edx,%esi
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	$16,40(%esp)
+	jmp	.L033slow_enc_loop_x86
+.align	16
+.L029slow_decrypt:
+	btl	$25,52(%esp)
+	jnc	.L036slow_dec_loop_x86
+.align	4
+.L037slow_dec_loop_sse:
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	movl	44(%esp),%edi
+	call	_sse_AES_decrypt_compact
+	movl	32(%esp),%esi
+	leal	60(%esp),%eax
+	movl	36(%esp),%ebx
+	movl	40(%esp),%ecx
+	movl	48(%esp),%edi
+	movq	(%esi),%mm1
+	movq	8(%esi),%mm5
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movq	%mm1,(%edi)
+	movq	%mm5,8(%edi)
+	subl	$16,%ecx
+	jc	.L038slow_dec_partial_sse
+	movq	%mm0,(%ebx)
+	movq	%mm4,8(%ebx)
+	leal	16(%ebx),%ebx
+	movl	%ebx,36(%esp)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	movl	%ecx,40(%esp)
+	jnz	.L037slow_dec_loop_sse
+	emms
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L038slow_dec_partial_sse:
+	movq	%mm0,(%eax)
+	movq	%mm4,8(%eax)
+	emms
+	addl	$16,%ecx
+	movl	%ebx,%edi
+	movl	%eax,%esi
+.align	4
+.long	2767451785
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L036slow_dec_loop_x86:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	leal	60(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt_compact
+	movl	48(%esp),%edi
+	movl	40(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	subl	$16,%esi
+	jc	.L039slow_dec_partial_x86
+	movl	%esi,40(%esp)
+	movl	36(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	leal	16(%esi),%esi
+	movl	%esi,36(%esp)
+	leal	60(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	32(%esp),%esi
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	jnz	.L036slow_dec_loop_x86
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L039slow_dec_partial_x86:
+	leal	60(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	32(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	movl	36(%esp),%edi
+	leal	60(%esp),%esi
+.align	4
+.long	2767451785
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_cbc_encrypt,.-.L_AES_cbc_encrypt_begin
+.type	_x86_AES_set_encrypt_key,@function
+.align	16
+_x86_AES_set_encrypt_key:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	24(%esp),%esi
+	movl	32(%esp),%edi
+	testl	$-1,%esi
+	jz	.L040badpointer
+	testl	$-1,%edi
+	jz	.L040badpointer
+	call	.L041pic_point
+.L041pic_point:
+	popl	%ebp
+	leal	.LAES_Te-.L041pic_point(%ebp),%ebp
+	leal	2176(%ebp),%ebp
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+	movl	28(%esp),%ecx
+	cmpl	$128,%ecx
+	je	.L04210rounds
+	cmpl	$192,%ecx
+	je	.L04312rounds
+	cmpl	$256,%ecx
+	je	.L04414rounds
+	movl	$-2,%eax
+	jmp	.L045exit
+.L04210rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L04610shortcut
+.align	4
+.L04710loop:
+	movl	(%edi),%eax
+	movl	12(%edi),%edx
+.L04610shortcut:
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,16(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,20(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,24(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,28(%edi)
+	incl	%ecx
+	addl	$16,%edi
+	cmpl	$10,%ecx
+	jl	.L04710loop
+	movl	$10,80(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L04312rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	16(%esi),%ecx
+	movl	20(%esi),%edx
+	movl	%ecx,16(%edi)
+	movl	%edx,20(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L04812shortcut
+.align	4
+.L04912loop:
+	movl	(%edi),%eax
+	movl	20(%edi),%edx
+.L04812shortcut:
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,24(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,28(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,32(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,36(%edi)
+	cmpl	$7,%ecx
+	je	.L05012break
+	incl	%ecx
+	xorl	16(%edi),%eax
+	movl	%eax,40(%edi)
+	xorl	20(%edi),%eax
+	movl	%eax,44(%edi)
+	addl	$24,%edi
+	jmp	.L04912loop
+.L05012break:
+	movl	$12,72(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L04414rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	movl	%eax,16(%edi)
+	movl	%ebx,20(%edi)
+	movl	%ecx,24(%edi)
+	movl	%edx,28(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L05114shortcut
+.align	4
+.L05214loop:
+	movl	28(%edi),%edx
+.L05114shortcut:
+	movl	(%edi),%eax
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,32(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,36(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,40(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,44(%edi)
+	cmpl	$6,%ecx
+	je	.L05314break
+	incl	%ecx
+	movl	%eax,%edx
+	movl	16(%edi),%eax
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	shll	$8,%ebx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movl	%eax,48(%edi)
+	xorl	20(%edi),%eax
+	movl	%eax,52(%edi)
+	xorl	24(%edi),%eax
+	movl	%eax,56(%edi)
+	xorl	28(%edi),%eax
+	movl	%eax,60(%edi)
+	addl	$32,%edi
+	jmp	.L05214loop
+.L05314break:
+	movl	$14,48(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L040badpointer:
+	movl	$-1,%eax
+.L045exit:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	_x86_AES_set_encrypt_key,.-_x86_AES_set_encrypt_key
+.globl	AES_set_encrypt_key
+.type	AES_set_encrypt_key,@function
+.align	16
+AES_set_encrypt_key:
+.L_AES_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	call	_x86_AES_set_encrypt_key
+	ret
+.size	AES_set_encrypt_key,.-.L_AES_set_encrypt_key_begin
+.globl	AES_set_decrypt_key
+.type	AES_set_decrypt_key,@function
+.align	16
+AES_set_decrypt_key:
+.L_AES_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	call	_x86_AES_set_encrypt_key
+	cmpl	$0,%eax
+	je	.L054proceed
+	ret
+.L054proceed:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%esi
+	movl	240(%esi),%ecx
+	leal	(,%ecx,4),%ecx
+	leal	(%esi,%ecx,4),%edi
+.align	4
+.L055invert:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	(%edi),%ecx
+	movl	4(%edi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,(%esi)
+	movl	%edx,4(%esi)
+	movl	8(%esi),%eax
+	movl	12(%esi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,8(%edi)
+	movl	%ebx,12(%edi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	addl	$16,%esi
+	subl	$16,%edi
+	cmpl	%edi,%esi
+	jne	.L055invert
+	movl	28(%esp),%edi
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,28(%esp)
+	movl	16(%edi),%eax
+.align	4
+.L056permute:
+	addl	$16,%edi
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%esi,%ebx
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%eax,%ebx
+	xorl	%esi,%ecx
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%eax,%ecx
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	roll	$8,%eax
+	xorl	%esi,%edx
+	movl	4(%edi),%ebp
+	xorl	%ebx,%eax
+	xorl	%edx,%ebx
+	xorl	%ecx,%eax
+	roll	$24,%ebx
+	xorl	%edx,%ecx
+	xorl	%edx,%eax
+	roll	$16,%ecx
+	xorl	%ebx,%eax
+	roll	$8,%edx
+	xorl	%ecx,%eax
+	movl	%ebp,%ebx
+	xorl	%edx,%eax
+	movl	%eax,(%edi)
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%ebx,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%ebx,%edx
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	roll	$8,%ebx
+	xorl	%esi,%eax
+	movl	8(%edi),%ebp
+	xorl	%ecx,%ebx
+	xorl	%eax,%ecx
+	xorl	%edx,%ebx
+	roll	$24,%ecx
+	xorl	%eax,%edx
+	xorl	%eax,%ebx
+	roll	$16,%edx
+	xorl	%ecx,%ebx
+	roll	$8,%eax
+	xorl	%edx,%ebx
+	movl	%ebp,%ecx
+	xorl	%eax,%ebx
+	movl	%ebx,4(%edi)
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%esi,%edx
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%ecx,%edx
+	xorl	%esi,%eax
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%ecx,%eax
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	roll	$8,%ecx
+	xorl	%esi,%ebx
+	movl	12(%edi),%ebp
+	xorl	%edx,%ecx
+	xorl	%ebx,%edx
+	xorl	%eax,%ecx
+	roll	$24,%edx
+	xorl	%ebx,%eax
+	xorl	%ebx,%ecx
+	roll	$16,%eax
+	xorl	%edx,%ecx
+	roll	$8,%ebx
+	xorl	%eax,%ecx
+	movl	%ebp,%edx
+	xorl	%ebx,%ecx
+	movl	%ecx,8(%edi)
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%esi,%eax
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%edx,%eax
+	xorl	%esi,%ebx
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%edx,%ebx
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	roll	$8,%edx
+	xorl	%esi,%ecx
+	movl	16(%edi),%ebp
+	xorl	%eax,%edx
+	xorl	%ecx,%eax
+	xorl	%ebx,%edx
+	roll	$24,%eax
+	xorl	%ecx,%ebx
+	xorl	%ecx,%edx
+	roll	$16,%ebx
+	xorl	%eax,%edx
+	roll	$8,%ecx
+	xorl	%ebx,%edx
+	movl	%ebp,%eax
+	xorl	%ecx,%edx
+	movl	%edx,12(%edi)
+	cmpl	28(%esp),%edi
+	jb	.L056permute
+	xorl	%eax,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_set_decrypt_key,.-.L_AES_set_decrypt_key_begin
+.byte	65,69,83,32,102,111,114,32,120,56,54,44,32,67,82,89
+.byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
+.byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
+#else
+.text
+.type	_x86_AES_encrypt_compact,@function
+.align	16
+_x86_AES_encrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+.align	16
+.L000loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ch,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$8,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$24,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%ecx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	xorl	%esi,%ecx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%ecx
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%ecx
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%edx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	xorl	%esi,%edx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%edx
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%edx
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%eax,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	xorl	%esi,%eax
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%eax
+	xorl	%edi,%esi
+	movl	$2155905152,%ebp
+	xorl	%esi,%eax
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%edi
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	andl	$4278124286,%edi
+	subl	%ebp,%esi
+	movl	%ebx,%ebp
+	andl	$454761243,%esi
+	rorl	$16,%ebp
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	xorl	%esi,%ebx
+	rorl	$24,%edi
+	xorl	%ebp,%esi
+	roll	$24,%ebx
+	xorl	%edi,%esi
+	xorl	%esi,%ebx
+	movl	20(%esp),%edi
+	movl	28(%esp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L000loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ch,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$8,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$24,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+
+	xorl	16(%edi),%eax
+	xorl	20(%edi),%ebx
+	xorl	24(%edi),%ecx
+	xorl	28(%edi),%edx
+	ret
+.size	_x86_AES_encrypt_compact,.-_x86_AES_encrypt_compact
+.type	_sse_AES_encrypt_compact,@function
+.align	16
+_sse_AES_encrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	$454761243,%eax
+	movl	%eax,8(%esp)
+	movl	%eax,12(%esp)
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+.align	16
+.L001loop:
+	pshufw	$8,%mm0,%mm1
+	pshufw	$13,%mm4,%mm5
+	movd	%mm1,%eax
+	movd	%mm5,%ebx
+	movl	%edi,20(%esp)
+	movzbl	%al,%esi
+	movzbl	%ah,%edx
+	pshufw	$13,%mm0,%mm2
+	movzbl	-128(%ebp,%esi,1),%ecx
+	movzbl	%bl,%edi
+	movzbl	-128(%ebp,%edx,1),%edx
+	shrl	$16,%eax
+	shll	$8,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$16,%esi
+	pshufw	$8,%mm4,%mm6
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$24,%esi
+	shrl	$16,%ebx
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$8,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	movd	%mm2,%eax
+	movd	%ecx,%mm0
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%ah,%edi
+	shll	$16,%ecx
+	movd	%mm6,%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	shll	$8,%esi
+	shrl	$16,%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shrl	$16,%eax
+	movd	%ecx,%mm1
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%ah,%edi
+	shll	$16,%ecx
+	andl	$255,%eax
+	orl	%esi,%ecx
+	punpckldq	%mm1,%mm0
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%eax,1),%eax
+	orl	%esi,%ecx
+	shll	$16,%eax
+	movzbl	-128(%ebp,%edi,1),%esi
+	orl	%eax,%edx
+	shll	$8,%esi
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	orl	%esi,%ecx
+	orl	%ebx,%edx
+	movl	20(%esp),%edi
+	movd	%ecx,%mm4
+	movd	%edx,%mm5
+	punpckldq	%mm5,%mm4
+	addl	$16,%edi
+	cmpl	24(%esp),%edi
+	ja	.L002out
+	movq	8(%esp),%mm2
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	movq	%mm0,%mm1
+	movq	%mm4,%mm5
+	pcmpgtb	%mm0,%mm3
+	pcmpgtb	%mm4,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	pshufw	$177,%mm0,%mm2
+	pshufw	$177,%mm4,%mm6
+	paddb	%mm0,%mm0
+	paddb	%mm4,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pshufw	$177,%mm2,%mm3
+	pshufw	$177,%mm6,%mm7
+	pxor	%mm0,%mm1
+	pxor	%mm4,%mm5
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	%mm3,%mm2
+	movq	%mm7,%mm6
+	pslld	$8,%mm3
+	pslld	$8,%mm7
+	psrld	$24,%mm2
+	psrld	$24,%mm6
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	movq	(%edi),%mm2
+	movq	8(%edi),%mm6
+	psrld	$8,%mm1
+	psrld	$8,%mm5
+	movl	-128(%ebp),%eax
+	pslld	$24,%mm3
+	pslld	$24,%mm7
+	movl	-64(%ebp),%ebx
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movl	(%ebp),%ecx
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movl	64(%ebp),%edx
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	jmp	.L001loop
+.align	16
+.L002out:
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	ret
+.size	_sse_AES_encrypt_compact,.-_sse_AES_encrypt_compact
+.type	_x86_AES_encrypt,@function
+.align	16
+_x86_AES_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+.align	16
+.L003loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%bh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%ch,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%dh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movzbl	%bh,%edi
+	xorl	1(%ebp,%edi,8),%esi
+
+	movl	20(%esp),%edi
+	movl	(%ebp,%edx,8),%edx
+	movzbl	%ah,%eax
+	xorl	3(%ebp,%eax,8),%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	xorl	2(%ebp,%ebx,8),%edx
+	movl	8(%esp),%ebx
+	xorl	1(%ebp,%ecx,8),%edx
+	movl	%esi,%ecx
+
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L003loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%bh,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	shrl	$16,%ebx
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%ch,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$24,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	shrl	$24,%ecx
+	movl	2(%ebp,%esi,8),%esi
+	andl	$255,%esi
+	movzbl	%dh,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$65280,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edx
+	andl	$255,%edi
+	movl	(%ebp,%edi,8),%edi
+	andl	$16711680,%edi
+	xorl	%edi,%esi
+	movzbl	%bh,%edi
+	movl	2(%ebp,%edi,8),%edi
+	andl	$4278190080,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movl	2(%ebp,%edx,8),%edx
+	andl	$255,%edx
+	movzbl	%ah,%eax
+	movl	(%ebp,%eax,8),%eax
+	andl	$65280,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	andl	$255,%ebx
+	movl	(%ebp,%ebx,8),%ebx
+	andl	$16711680,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	movl	2(%ebp,%ecx,8),%ecx
+	andl	$4278190080,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	ret
+.align	64
+.LAES_Te:
+.long	2774754246,2774754246
+.long	2222750968,2222750968
+.long	2574743534,2574743534
+.long	2373680118,2373680118
+.long	234025727,234025727
+.long	3177933782,3177933782
+.long	2976870366,2976870366
+.long	1422247313,1422247313
+.long	1345335392,1345335392
+.long	50397442,50397442
+.long	2842126286,2842126286
+.long	2099981142,2099981142
+.long	436141799,436141799
+.long	1658312629,1658312629
+.long	3870010189,3870010189
+.long	2591454956,2591454956
+.long	1170918031,1170918031
+.long	2642575903,2642575903
+.long	1086966153,1086966153
+.long	2273148410,2273148410
+.long	368769775,368769775
+.long	3948501426,3948501426
+.long	3376891790,3376891790
+.long	200339707,200339707
+.long	3970805057,3970805057
+.long	1742001331,1742001331
+.long	4255294047,4255294047
+.long	3937382213,3937382213
+.long	3214711843,3214711843
+.long	4154762323,4154762323
+.long	2524082916,2524082916
+.long	1539358875,1539358875
+.long	3266819957,3266819957
+.long	486407649,486407649
+.long	2928907069,2928907069
+.long	1780885068,1780885068
+.long	1513502316,1513502316
+.long	1094664062,1094664062
+.long	49805301,49805301
+.long	1338821763,1338821763
+.long	1546925160,1546925160
+.long	4104496465,4104496465
+.long	887481809,887481809
+.long	150073849,150073849
+.long	2473685474,2473685474
+.long	1943591083,1943591083
+.long	1395732834,1395732834
+.long	1058346282,1058346282
+.long	201589768,201589768
+.long	1388824469,1388824469
+.long	1696801606,1696801606
+.long	1589887901,1589887901
+.long	672667696,672667696
+.long	2711000631,2711000631
+.long	251987210,251987210
+.long	3046808111,3046808111
+.long	151455502,151455502
+.long	907153956,907153956
+.long	2608889883,2608889883
+.long	1038279391,1038279391
+.long	652995533,652995533
+.long	1764173646,1764173646
+.long	3451040383,3451040383
+.long	2675275242,2675275242
+.long	453576978,453576978
+.long	2659418909,2659418909
+.long	1949051992,1949051992
+.long	773462580,773462580
+.long	756751158,756751158
+.long	2993581788,2993581788
+.long	3998898868,3998898868
+.long	4221608027,4221608027
+.long	4132590244,4132590244
+.long	1295727478,1295727478
+.long	1641469623,1641469623
+.long	3467883389,3467883389
+.long	2066295122,2066295122
+.long	1055122397,1055122397
+.long	1898917726,1898917726
+.long	2542044179,2542044179
+.long	4115878822,4115878822
+.long	1758581177,1758581177
+.long	0,0
+.long	753790401,753790401
+.long	1612718144,1612718144
+.long	536673507,536673507
+.long	3367088505,3367088505
+.long	3982187446,3982187446
+.long	3194645204,3194645204
+.long	1187761037,1187761037
+.long	3653156455,3653156455
+.long	1262041458,1262041458
+.long	3729410708,3729410708
+.long	3561770136,3561770136
+.long	3898103984,3898103984
+.long	1255133061,1255133061
+.long	1808847035,1808847035
+.long	720367557,720367557
+.long	3853167183,3853167183
+.long	385612781,385612781
+.long	3309519750,3309519750
+.long	3612167578,3612167578
+.long	1429418854,1429418854
+.long	2491778321,2491778321
+.long	3477423498,3477423498
+.long	284817897,284817897
+.long	100794884,100794884
+.long	2172616702,2172616702
+.long	4031795360,4031795360
+.long	1144798328,1144798328
+.long	3131023141,3131023141
+.long	3819481163,3819481163
+.long	4082192802,4082192802
+.long	4272137053,4272137053
+.long	3225436288,3225436288
+.long	2324664069,2324664069
+.long	2912064063,2912064063
+.long	3164445985,3164445985
+.long	1211644016,1211644016
+.long	83228145,83228145
+.long	3753688163,3753688163
+.long	3249976951,3249976951
+.long	1977277103,1977277103
+.long	1663115586,1663115586
+.long	806359072,806359072
+.long	452984805,452984805
+.long	250868733,250868733
+.long	1842533055,1842533055
+.long	1288555905,1288555905
+.long	336333848,336333848
+.long	890442534,890442534
+.long	804056259,804056259
+.long	3781124030,3781124030
+.long	2727843637,2727843637
+.long	3427026056,3427026056
+.long	957814574,957814574
+.long	1472513171,1472513171
+.long	4071073621,4071073621
+.long	2189328124,2189328124
+.long	1195195770,1195195770
+.long	2892260552,2892260552
+.long	3881655738,3881655738
+.long	723065138,723065138
+.long	2507371494,2507371494
+.long	2690670784,2690670784
+.long	2558624025,2558624025
+.long	3511635870,3511635870
+.long	2145180835,2145180835
+.long	1713513028,1713513028
+.long	2116692564,2116692564
+.long	2878378043,2878378043
+.long	2206763019,2206763019
+.long	3393603212,3393603212
+.long	703524551,703524551
+.long	3552098411,3552098411
+.long	1007948840,1007948840
+.long	2044649127,2044649127
+.long	3797835452,3797835452
+.long	487262998,487262998
+.long	1994120109,1994120109
+.long	1004593371,1004593371
+.long	1446130276,1446130276
+.long	1312438900,1312438900
+.long	503974420,503974420
+.long	3679013266,3679013266
+.long	168166924,168166924
+.long	1814307912,1814307912
+.long	3831258296,3831258296
+.long	1573044895,1573044895
+.long	1859376061,1859376061
+.long	4021070915,4021070915
+.long	2791465668,2791465668
+.long	2828112185,2828112185
+.long	2761266481,2761266481
+.long	937747667,937747667
+.long	2339994098,2339994098
+.long	854058965,854058965
+.long	1137232011,1137232011
+.long	1496790894,1496790894
+.long	3077402074,3077402074
+.long	2358086913,2358086913
+.long	1691735473,1691735473
+.long	3528347292,3528347292
+.long	3769215305,3769215305
+.long	3027004632,3027004632
+.long	4199962284,4199962284
+.long	133494003,133494003
+.long	636152527,636152527
+.long	2942657994,2942657994
+.long	2390391540,2390391540
+.long	3920539207,3920539207
+.long	403179536,403179536
+.long	3585784431,3585784431
+.long	2289596656,2289596656
+.long	1864705354,1864705354
+.long	1915629148,1915629148
+.long	605822008,605822008
+.long	4054230615,4054230615
+.long	3350508659,3350508659
+.long	1371981463,1371981463
+.long	602466507,602466507
+.long	2094914977,2094914977
+.long	2624877800,2624877800
+.long	555687742,555687742
+.long	3712699286,3712699286
+.long	3703422305,3703422305
+.long	2257292045,2257292045
+.long	2240449039,2240449039
+.long	2423288032,2423288032
+.long	1111375484,1111375484
+.long	3300242801,3300242801
+.long	2858837708,2858837708
+.long	3628615824,3628615824
+.long	84083462,84083462
+.long	32962295,32962295
+.long	302911004,302911004
+.long	2741068226,2741068226
+.long	1597322602,1597322602
+.long	4183250862,4183250862
+.long	3501832553,3501832553
+.long	2441512471,2441512471
+.long	1489093017,1489093017
+.long	656219450,656219450
+.long	3114180135,3114180135
+.long	954327513,954327513
+.long	335083755,335083755
+.long	3013122091,3013122091
+.long	856756514,856756514
+.long	3144247762,3144247762
+.long	1893325225,1893325225
+.long	2307821063,2307821063
+.long	2811532339,2811532339
+.long	3063651117,3063651117
+.long	572399164,572399164
+.long	2458355477,2458355477
+.long	552200649,552200649
+.long	1238290055,1238290055
+.long	4283782570,4283782570
+.long	2015897680,2015897680
+.long	2061492133,2061492133
+.long	2408352771,2408352771
+.long	4171342169,4171342169
+.long	2156497161,2156497161
+.long	386731290,386731290
+.long	3669999461,3669999461
+.long	837215959,837215959
+.long	3326231172,3326231172
+.long	3093850320,3093850320
+.long	3275833730,3275833730
+.long	2962856233,2962856233
+.long	1999449434,1999449434
+.long	286199582,286199582
+.long	3417354363,3417354363
+.long	4233385128,4233385128
+.long	3602627437,3602627437
+.long	974525996,974525996
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.byte	99,124,119,123,242,107,111,197
+.byte	48,1,103,43,254,215,171,118
+.byte	202,130,201,125,250,89,71,240
+.byte	173,212,162,175,156,164,114,192
+.byte	183,253,147,38,54,63,247,204
+.byte	52,165,229,241,113,216,49,21
+.byte	4,199,35,195,24,150,5,154
+.byte	7,18,128,226,235,39,178,117
+.byte	9,131,44,26,27,110,90,160
+.byte	82,59,214,179,41,227,47,132
+.byte	83,209,0,237,32,252,177,91
+.byte	106,203,190,57,74,76,88,207
+.byte	208,239,170,251,67,77,51,133
+.byte	69,249,2,127,80,60,159,168
+.byte	81,163,64,143,146,157,56,245
+.byte	188,182,218,33,16,255,243,210
+.byte	205,12,19,236,95,151,68,23
+.byte	196,167,126,61,100,93,25,115
+.byte	96,129,79,220,34,42,144,136
+.byte	70,238,184,20,222,94,11,219
+.byte	224,50,58,10,73,6,36,92
+.byte	194,211,172,98,145,149,228,121
+.byte	231,200,55,109,141,213,78,169
+.byte	108,86,244,234,101,122,174,8
+.byte	186,120,37,46,28,166,180,198
+.byte	232,221,116,31,75,189,139,138
+.byte	112,62,181,102,72,3,246,14
+.byte	97,53,87,185,134,193,29,158
+.byte	225,248,152,17,105,217,142,148
+.byte	155,30,135,233,206,85,40,223
+.byte	140,161,137,13,191,230,66,104
+.byte	65,153,45,15,176,84,187,22
+.long	1,2,4,8
+.long	16,32,64,128
+.long	27,54,0,0
+.long	0,0,0,0
+.size	_x86_AES_encrypt,.-_x86_AES_encrypt
+.globl	AES_encrypt
+.type	AES_encrypt,@function
+.align	16
+AES_encrypt:
+.L_AES_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	28(%esp),%edi
+	movl	%esp,%eax
+	subl	$36,%esp
+	andl	$-64,%esp
+	leal	-127(%edi),%ebx
+	subl	%esp,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esp
+	addl	$4,%esp
+	movl	%eax,28(%esp)
+	call	.L004pic_point
+.L004pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P,%eax
+	leal	.LAES_Te-.L004pic_point(%ebp),%ebp
+	leal	764(%esp),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	btl	$25,(%eax)
+	jnc	.L005x86
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	call	_sse_AES_encrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	16
+.L005x86:
+	movl	%ebp,24(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	call	_x86_AES_encrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_encrypt,.-.L_AES_encrypt_begin
+.type	_x86_AES_decrypt_compact,@function
+.align	16
+_x86_AES_decrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+.align	16
+.L006loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	shrl	$24,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%eax
+	subl	%edi,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%esi,%eax
+	movl	$2155905152,%edi
+	andl	%eax,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%eax,%eax,1),%ebx
+	subl	%edi,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%ecx,%eax
+	xorl	%esi,%ebx
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%ecx,%ebx
+	roll	$8,%ecx
+	xorl	%esi,%ebp
+	xorl	%eax,%ecx
+	xorl	%ebp,%eax
+	xorl	%ebx,%ecx
+	xorl	%ebp,%ebx
+	roll	$24,%eax
+	xorl	%ebp,%ecx
+	roll	$16,%ebx
+	xorl	%eax,%ecx
+	roll	$8,%ebp
+	xorl	%ebx,%ecx
+	movl	4(%esp),%eax
+	xorl	%ebp,%ecx
+	movl	%ecx,12(%esp)
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebx
+	subl	%edi,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%esi,%ebx
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%edx,%ebx
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%edx,%ecx
+	roll	$8,%edx
+	xorl	%esi,%ebp
+	xorl	%ebx,%edx
+	xorl	%ebp,%ebx
+	xorl	%ecx,%edx
+	xorl	%ebp,%ecx
+	roll	$24,%ebx
+	xorl	%ebp,%edx
+	roll	$16,%ecx
+	xorl	%ebx,%edx
+	roll	$8,%ebp
+	xorl	%ecx,%edx
+	movl	8(%esp),%ebx
+	xorl	%ebp,%edx
+	movl	%edx,16(%esp)
+	movl	$2155905152,%edi
+	andl	%eax,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%eax,%eax,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%edx
+	subl	%edi,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%eax,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%eax,%edx
+	roll	$8,%eax
+	xorl	%esi,%ebp
+	xorl	%ecx,%eax
+	xorl	%ebp,%ecx
+	xorl	%edx,%eax
+	xorl	%ebp,%edx
+	roll	$24,%ecx
+	xorl	%ebp,%eax
+	roll	$16,%edx
+	xorl	%ecx,%eax
+	roll	$8,%ebp
+	xorl	%edx,%eax
+	xorl	%ebp,%eax
+	movl	$2155905152,%edi
+	andl	%ebx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ebx,%ebx,1),%ecx
+	subl	%edi,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%edi
+	andl	%ecx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%ecx,%ecx,1),%edx
+	subl	%edi,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%ebx,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%edi
+	andl	%edx,%edi
+	movl	%edi,%esi
+	shrl	$7,%edi
+	leal	(%edx,%edx,1),%ebp
+	subl	%edi,%esi
+	andl	$4278124286,%ebp
+	andl	$454761243,%esi
+	xorl	%ebx,%edx
+	roll	$8,%ebx
+	xorl	%esi,%ebp
+	xorl	%ecx,%ebx
+	xorl	%ebp,%ecx
+	xorl	%edx,%ebx
+	xorl	%ebp,%edx
+	roll	$24,%ecx
+	xorl	%ebp,%ebx
+	roll	$16,%edx
+	xorl	%ecx,%ebx
+	roll	$8,%ebp
+	xorl	%edx,%ebx
+	movl	12(%esp),%ecx
+	xorl	%ebp,%ebx
+	movl	16(%esp),%edx
+	movl	20(%esp),%edi
+	movl	28(%esp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L006loop
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	-128(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	-128(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	-128(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	-128(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	movzbl	-128(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	xorl	16(%edi),%eax
+	xorl	20(%edi),%ebx
+	xorl	24(%edi),%ecx
+	xorl	28(%edi),%edx
+	ret
+.size	_x86_AES_decrypt_compact,.-_x86_AES_decrypt_compact
+.type	_sse_AES_decrypt_compact,@function
+.align	16
+_sse_AES_decrypt_compact:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+	movl	$454761243,%eax
+	movl	%eax,8(%esp)
+	movl	%eax,12(%esp)
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+.align	16
+.L007loop:
+	pshufw	$12,%mm0,%mm1
+	pshufw	$9,%mm4,%mm5
+	movd	%mm1,%eax
+	movd	%mm5,%ebx
+	movl	%edi,20(%esp)
+	movzbl	%al,%esi
+	movzbl	%ah,%edx
+	pshufw	$6,%mm0,%mm2
+	movzbl	-128(%ebp,%esi,1),%ecx
+	movzbl	%bl,%edi
+	movzbl	-128(%ebp,%edx,1),%edx
+	shrl	$16,%eax
+	shll	$8,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$16,%esi
+	pshufw	$3,%mm4,%mm6
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$24,%esi
+	shrl	$16,%ebx
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shll	$24,%esi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$8,%esi
+	movd	%mm2,%eax
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	shll	$16,%esi
+	movd	%mm6,%ebx
+	movd	%ecx,%mm0
+	movzbl	-128(%ebp,%edi,1),%ecx
+	movzbl	%al,%edi
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bl,%edi
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%ah,%edi
+	shll	$16,%esi
+	shrl	$16,%eax
+	orl	%esi,%edx
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%bh,%edi
+	shrl	$16,%ebx
+	shll	$8,%esi
+	movd	%edx,%mm1
+	movzbl	-128(%ebp,%edi,1),%edx
+	movzbl	%bh,%edi
+	shll	$24,%edx
+	andl	$255,%ebx
+	orl	%esi,%edx
+	punpckldq	%mm1,%mm0
+	movzbl	-128(%ebp,%edi,1),%esi
+	movzbl	%al,%edi
+	shll	$8,%esi
+	movzbl	%ah,%eax
+	movzbl	-128(%ebp,%ebx,1),%ebx
+	orl	%esi,%ecx
+	movzbl	-128(%ebp,%edi,1),%esi
+	orl	%ebx,%edx
+	shll	$16,%esi
+	movzbl	-128(%ebp,%eax,1),%eax
+	orl	%esi,%edx
+	shll	$24,%eax
+	orl	%eax,%ecx
+	movl	20(%esp),%edi
+	movd	%edx,%mm4
+	movd	%ecx,%mm5
+	punpckldq	%mm5,%mm4
+	addl	$16,%edi
+	cmpl	24(%esp),%edi
+	ja	.L008out
+	movq	%mm0,%mm3
+	movq	%mm4,%mm7
+	pshufw	$228,%mm0,%mm2
+	pshufw	$228,%mm4,%mm6
+	movq	%mm0,%mm1
+	movq	%mm4,%mm5
+	pshufw	$177,%mm0,%mm0
+	pshufw	$177,%mm4,%mm4
+	pslld	$8,%mm2
+	pslld	$8,%mm6
+	psrld	$8,%mm3
+	psrld	$8,%mm7
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pslld	$16,%mm2
+	pslld	$16,%mm6
+	psrld	$16,%mm3
+	psrld	$16,%mm7
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movq	8(%esp),%mm3
+	pxor	%mm2,%mm2
+	pxor	%mm6,%mm6
+	pcmpgtb	%mm1,%mm2
+	pcmpgtb	%mm5,%mm6
+	pand	%mm3,%mm2
+	pand	%mm3,%mm6
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm2,%mm1
+	pxor	%mm6,%mm5
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	movq	%mm1,%mm2
+	movq	%mm5,%mm6
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pslld	$24,%mm3
+	pslld	$24,%mm7
+	psrld	$8,%mm2
+	psrld	$8,%mm6
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	movq	8(%esp),%mm2
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	pcmpgtb	%mm1,%mm3
+	pcmpgtb	%mm5,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm3,%mm1
+	pxor	%mm7,%mm5
+	pshufw	$177,%mm1,%mm3
+	pshufw	$177,%mm5,%mm7
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm3,%mm3
+	pxor	%mm7,%mm7
+	pcmpgtb	%mm1,%mm3
+	pcmpgtb	%mm5,%mm7
+	pand	%mm2,%mm3
+	pand	%mm2,%mm7
+	paddb	%mm1,%mm1
+	paddb	%mm5,%mm5
+	pxor	%mm3,%mm1
+	pxor	%mm7,%mm5
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movq	%mm1,%mm3
+	movq	%mm5,%mm7
+	pshufw	$177,%mm1,%mm2
+	pshufw	$177,%mm5,%mm6
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	pslld	$8,%mm1
+	pslld	$8,%mm5
+	psrld	$8,%mm3
+	psrld	$8,%mm7
+	movq	(%edi),%mm2
+	movq	8(%edi),%mm6
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	movl	-128(%ebp),%eax
+	pslld	$16,%mm1
+	pslld	$16,%mm5
+	movl	-64(%ebp),%ebx
+	psrld	$16,%mm3
+	psrld	$16,%mm7
+	movl	(%ebp),%ecx
+	pxor	%mm1,%mm0
+	pxor	%mm5,%mm4
+	movl	64(%ebp),%edx
+	pxor	%mm3,%mm0
+	pxor	%mm7,%mm4
+	pxor	%mm2,%mm0
+	pxor	%mm6,%mm4
+	jmp	.L007loop
+.align	16
+.L008out:
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	ret
+.size	_sse_AES_decrypt_compact,.-_sse_AES_decrypt_compact
+.type	_x86_AES_decrypt,@function
+.align	16
+_x86_AES_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	movl	%edi,20(%esp)
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,24(%esp)
+.align	16
+.L009loop:
+	movl	%eax,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%dh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,4(%esp)
+
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%ah,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+	movl	%esi,8(%esp)
+
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movl	(%ebp,%esi,8),%esi
+	movzbl	%bh,%edi
+	xorl	3(%ebp,%edi,8),%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	xorl	2(%ebp,%edi,8),%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	xorl	1(%ebp,%edi,8),%esi
+
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movl	(%ebp,%edx,8),%edx
+	movzbl	%ch,%ecx
+	xorl	3(%ebp,%ecx,8),%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	xorl	2(%ebp,%ebx,8),%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	xorl	1(%ebp,%eax,8),%edx
+	movl	4(%esp),%eax
+
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	cmpl	24(%esp),%edi
+	movl	%edi,20(%esp)
+	jb	.L009loop
+	leal	2176(%ebp),%ebp
+	movl	-128(%ebp),%edi
+	movl	-96(%ebp),%esi
+	movl	-64(%ebp),%edi
+	movl	-32(%ebp),%esi
+	movl	(%ebp),%edi
+	movl	32(%ebp),%esi
+	movl	64(%ebp),%edi
+	movl	96(%ebp),%esi
+	leal	-128(%ebp),%ebp
+	movl	%eax,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%dh,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ebx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,4(%esp)
+	movl	%ebx,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%ah,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%ecx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	%esi,8(%esp)
+	movl	%ecx,%esi
+	andl	$255,%esi
+	movzbl	(%ebp,%esi,1),%esi
+	movzbl	%bh,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$8,%edi
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	shrl	$16,%edi
+	andl	$255,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$16,%edi
+	xorl	%edi,%esi
+	movl	%edx,%edi
+	shrl	$24,%edi
+	movzbl	(%ebp,%edi,1),%edi
+	shll	$24,%edi
+	xorl	%edi,%esi
+	movl	20(%esp),%edi
+	andl	$255,%edx
+	movzbl	(%ebp,%edx,1),%edx
+	movzbl	%ch,%ecx
+	movzbl	(%ebp,%ecx,1),%ecx
+	shll	$8,%ecx
+	xorl	%ecx,%edx
+	movl	%esi,%ecx
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	movzbl	(%ebp,%ebx,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%edx
+	movl	8(%esp),%ebx
+	shrl	$24,%eax
+	movzbl	(%ebp,%eax,1),%eax
+	shll	$24,%eax
+	xorl	%eax,%edx
+	movl	4(%esp),%eax
+	leal	-2048(%ebp),%ebp
+	addl	$16,%edi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	ret
+.align	64
+.LAES_Td:
+.long	1353184337,1353184337
+.long	1399144830,1399144830
+.long	3282310938,3282310938
+.long	2522752826,2522752826
+.long	3412831035,3412831035
+.long	4047871263,4047871263
+.long	2874735276,2874735276
+.long	2466505547,2466505547
+.long	1442459680,1442459680
+.long	4134368941,4134368941
+.long	2440481928,2440481928
+.long	625738485,625738485
+.long	4242007375,4242007375
+.long	3620416197,3620416197
+.long	2151953702,2151953702
+.long	2409849525,2409849525
+.long	1230680542,1230680542
+.long	1729870373,1729870373
+.long	2551114309,2551114309
+.long	3787521629,3787521629
+.long	41234371,41234371
+.long	317738113,317738113
+.long	2744600205,2744600205
+.long	3338261355,3338261355
+.long	3881799427,3881799427
+.long	2510066197,2510066197
+.long	3950669247,3950669247
+.long	3663286933,3663286933
+.long	763608788,763608788
+.long	3542185048,3542185048
+.long	694804553,694804553
+.long	1154009486,1154009486
+.long	1787413109,1787413109
+.long	2021232372,2021232372
+.long	1799248025,1799248025
+.long	3715217703,3715217703
+.long	3058688446,3058688446
+.long	397248752,397248752
+.long	1722556617,1722556617
+.long	3023752829,3023752829
+.long	407560035,407560035
+.long	2184256229,2184256229
+.long	1613975959,1613975959
+.long	1165972322,1165972322
+.long	3765920945,3765920945
+.long	2226023355,2226023355
+.long	480281086,480281086
+.long	2485848313,2485848313
+.long	1483229296,1483229296
+.long	436028815,436028815
+.long	2272059028,2272059028
+.long	3086515026,3086515026
+.long	601060267,601060267
+.long	3791801202,3791801202
+.long	1468997603,1468997603
+.long	715871590,715871590
+.long	120122290,120122290
+.long	63092015,63092015
+.long	2591802758,2591802758
+.long	2768779219,2768779219
+.long	4068943920,4068943920
+.long	2997206819,2997206819
+.long	3127509762,3127509762
+.long	1552029421,1552029421
+.long	723308426,723308426
+.long	2461301159,2461301159
+.long	4042393587,4042393587
+.long	2715969870,2715969870
+.long	3455375973,3455375973
+.long	3586000134,3586000134
+.long	526529745,526529745
+.long	2331944644,2331944644
+.long	2639474228,2639474228
+.long	2689987490,2689987490
+.long	853641733,853641733
+.long	1978398372,1978398372
+.long	971801355,971801355
+.long	2867814464,2867814464
+.long	111112542,111112542
+.long	1360031421,1360031421
+.long	4186579262,4186579262
+.long	1023860118,1023860118
+.long	2919579357,2919579357
+.long	1186850381,1186850381
+.long	3045938321,3045938321
+.long	90031217,90031217
+.long	1876166148,1876166148
+.long	4279586912,4279586912
+.long	620468249,620468249
+.long	2548678102,2548678102
+.long	3426959497,3426959497
+.long	2006899047,2006899047
+.long	3175278768,3175278768
+.long	2290845959,2290845959
+.long	945494503,945494503
+.long	3689859193,3689859193
+.long	1191869601,1191869601
+.long	3910091388,3910091388
+.long	3374220536,3374220536
+.long	0,0
+.long	2206629897,2206629897
+.long	1223502642,1223502642
+.long	2893025566,2893025566
+.long	1316117100,1316117100
+.long	4227796733,4227796733
+.long	1446544655,1446544655
+.long	517320253,517320253
+.long	658058550,658058550
+.long	1691946762,1691946762
+.long	564550760,564550760
+.long	3511966619,3511966619
+.long	976107044,976107044
+.long	2976320012,2976320012
+.long	266819475,266819475
+.long	3533106868,3533106868
+.long	2660342555,2660342555
+.long	1338359936,1338359936
+.long	2720062561,2720062561
+.long	1766553434,1766553434
+.long	370807324,370807324
+.long	179999714,179999714
+.long	3844776128,3844776128
+.long	1138762300,1138762300
+.long	488053522,488053522
+.long	185403662,185403662
+.long	2915535858,2915535858
+.long	3114841645,3114841645
+.long	3366526484,3366526484
+.long	2233069911,2233069911
+.long	1275557295,1275557295
+.long	3151862254,3151862254
+.long	4250959779,4250959779
+.long	2670068215,2670068215
+.long	3170202204,3170202204
+.long	3309004356,3309004356
+.long	880737115,880737115
+.long	1982415755,1982415755
+.long	3703972811,3703972811
+.long	1761406390,1761406390
+.long	1676797112,1676797112
+.long	3403428311,3403428311
+.long	277177154,277177154
+.long	1076008723,1076008723
+.long	538035844,538035844
+.long	2099530373,2099530373
+.long	4164795346,4164795346
+.long	288553390,288553390
+.long	1839278535,1839278535
+.long	1261411869,1261411869
+.long	4080055004,4080055004
+.long	3964831245,3964831245
+.long	3504587127,3504587127
+.long	1813426987,1813426987
+.long	2579067049,2579067049
+.long	4199060497,4199060497
+.long	577038663,577038663
+.long	3297574056,3297574056
+.long	440397984,440397984
+.long	3626794326,3626794326
+.long	4019204898,4019204898
+.long	3343796615,3343796615
+.long	3251714265,3251714265
+.long	4272081548,4272081548
+.long	906744984,906744984
+.long	3481400742,3481400742
+.long	685669029,685669029
+.long	646887386,646887386
+.long	2764025151,2764025151
+.long	3835509292,3835509292
+.long	227702864,227702864
+.long	2613862250,2613862250
+.long	1648787028,1648787028
+.long	3256061430,3256061430
+.long	3904428176,3904428176
+.long	1593260334,1593260334
+.long	4121936770,4121936770
+.long	3196083615,3196083615
+.long	2090061929,2090061929
+.long	2838353263,2838353263
+.long	3004310991,3004310991
+.long	999926984,999926984
+.long	2809993232,2809993232
+.long	1852021992,1852021992
+.long	2075868123,2075868123
+.long	158869197,158869197
+.long	4095236462,4095236462
+.long	28809964,28809964
+.long	2828685187,2828685187
+.long	1701746150,1701746150
+.long	2129067946,2129067946
+.long	147831841,147831841
+.long	3873969647,3873969647
+.long	3650873274,3650873274
+.long	3459673930,3459673930
+.long	3557400554,3557400554
+.long	3598495785,3598495785
+.long	2947720241,2947720241
+.long	824393514,824393514
+.long	815048134,815048134
+.long	3227951669,3227951669
+.long	935087732,935087732
+.long	2798289660,2798289660
+.long	2966458592,2966458592
+.long	366520115,366520115
+.long	1251476721,1251476721
+.long	4158319681,4158319681
+.long	240176511,240176511
+.long	804688151,804688151
+.long	2379631990,2379631990
+.long	1303441219,1303441219
+.long	1414376140,1414376140
+.long	3741619940,3741619940
+.long	3820343710,3820343710
+.long	461924940,461924940
+.long	3089050817,3089050817
+.long	2136040774,2136040774
+.long	82468509,82468509
+.long	1563790337,1563790337
+.long	1937016826,1937016826
+.long	776014843,776014843
+.long	1511876531,1511876531
+.long	1389550482,1389550482
+.long	861278441,861278441
+.long	323475053,323475053
+.long	2355222426,2355222426
+.long	2047648055,2047648055
+.long	2383738969,2383738969
+.long	2302415851,2302415851
+.long	3995576782,3995576782
+.long	902390199,902390199
+.long	3991215329,3991215329
+.long	1018251130,1018251130
+.long	1507840668,1507840668
+.long	1064563285,1064563285
+.long	2043548696,2043548696
+.long	3208103795,3208103795
+.long	3939366739,3939366739
+.long	1537932639,1537932639
+.long	342834655,342834655
+.long	2262516856,2262516856
+.long	2180231114,2180231114
+.long	1053059257,1053059257
+.long	741614648,741614648
+.long	1598071746,1598071746
+.long	1925389590,1925389590
+.long	203809468,203809468
+.long	2336832552,2336832552
+.long	1100287487,1100287487
+.long	1895934009,1895934009
+.long	3736275976,3736275976
+.long	2632234200,2632234200
+.long	2428589668,2428589668
+.long	1636092795,1636092795
+.long	1890988757,1890988757
+.long	1952214088,1952214088
+.long	1113045200,1113045200
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.byte	82,9,106,213,48,54,165,56
+.byte	191,64,163,158,129,243,215,251
+.byte	124,227,57,130,155,47,255,135
+.byte	52,142,67,68,196,222,233,203
+.byte	84,123,148,50,166,194,35,61
+.byte	238,76,149,11,66,250,195,78
+.byte	8,46,161,102,40,217,36,178
+.byte	118,91,162,73,109,139,209,37
+.byte	114,248,246,100,134,104,152,22
+.byte	212,164,92,204,93,101,182,146
+.byte	108,112,72,80,253,237,185,218
+.byte	94,21,70,87,167,141,157,132
+.byte	144,216,171,0,140,188,211,10
+.byte	247,228,88,5,184,179,69,6
+.byte	208,44,30,143,202,63,15,2
+.byte	193,175,189,3,1,19,138,107
+.byte	58,145,17,65,79,103,220,234
+.byte	151,242,207,206,240,180,230,115
+.byte	150,172,116,34,231,173,53,133
+.byte	226,249,55,232,28,117,223,110
+.byte	71,241,26,113,29,41,197,137
+.byte	111,183,98,14,170,24,190,27
+.byte	252,86,62,75,198,210,121,32
+.byte	154,219,192,254,120,205,90,244
+.byte	31,221,168,51,136,7,199,49
+.byte	177,18,16,89,39,128,236,95
+.byte	96,81,127,169,25,181,74,13
+.byte	45,229,122,159,147,201,156,239
+.byte	160,224,59,77,174,42,245,176
+.byte	200,235,187,60,131,83,153,97
+.byte	23,43,4,126,186,119,214,38
+.byte	225,105,20,99,85,33,12,125
+.size	_x86_AES_decrypt,.-_x86_AES_decrypt
+.globl	AES_decrypt
+.type	AES_decrypt,@function
+.align	16
+AES_decrypt:
+.L_AES_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	28(%esp),%edi
+	movl	%esp,%eax
+	subl	$36,%esp
+	andl	$-64,%esp
+	leal	-127(%edi),%ebx
+	subl	%esp,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esp
+	addl	$4,%esp
+	movl	%eax,28(%esp)
+	call	.L010pic_point
+.L010pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P,%eax
+	leal	.LAES_Td-.L010pic_point(%ebp),%ebp
+	leal	764(%esp),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	btl	$25,(%eax)
+	jnc	.L011x86
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	call	_sse_AES_decrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	16
+.L011x86:
+	movl	%ebp,24(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	call	_x86_AES_decrypt_compact
+	movl	28(%esp),%esp
+	movl	24(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_decrypt,.-.L_AES_decrypt_begin
+.globl	AES_cbc_encrypt
+.type	AES_cbc_encrypt,@function
+.align	16
+AES_cbc_encrypt:
+.L_AES_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ecx
+	cmpl	$0,%ecx
+	je	.L012drop_out
+	call	.L013pic_point
+.L013pic_point:
+	popl	%ebp
+	leal	OPENSSL_ia32cap_P,%eax
+	cmpl	$0,40(%esp)
+	leal	.LAES_Te-.L013pic_point(%ebp),%ebp
+	jne	.L014picked_te
+	leal	.LAES_Td-.LAES_Te(%ebp),%ebp
+.L014picked_te:
+	pushfl
+	cld
+	cmpl	$512,%ecx
+	jb	.L015slow_way
+	testl	$15,%ecx
+	jnz	.L015slow_way
+	btl	$28,(%eax)
+	jc	.L015slow_way
+	leal	-324(%esp),%esi
+	andl	$-64,%esi
+	movl	%ebp,%eax
+	leal	2304(%ebp),%ebx
+	movl	%esi,%edx
+	andl	$4095,%eax
+	andl	$4095,%ebx
+	andl	$4095,%edx
+	cmpl	%ebx,%edx
+	jb	.L016tbl_break_out
+	subl	%ebx,%edx
+	subl	%edx,%esi
+	jmp	.L017tbl_ok
+.align	4
+.L016tbl_break_out:
+	subl	%eax,%edx
+	andl	$4095,%edx
+	addl	$384,%edx
+	subl	%edx,%esi
+.align	4
+.L017tbl_ok:
+	leal	24(%esp),%edx
+	xchgl	%esi,%esp
+	addl	$4,%esp
+	movl	%ebp,24(%esp)
+	movl	%esi,28(%esp)
+	movl	(%edx),%eax
+	movl	4(%edx),%ebx
+	movl	12(%edx),%edi
+	movl	16(%edx),%esi
+	movl	20(%edx),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,40(%esp)
+	movl	%edi,44(%esp)
+	movl	%esi,48(%esp)
+	movl	$0,316(%esp)
+	movl	%edi,%ebx
+	movl	$61,%ecx
+	subl	%ebp,%ebx
+	movl	%edi,%esi
+	andl	$4095,%ebx
+	leal	76(%esp),%edi
+	cmpl	$2304,%ebx
+	jb	.L018do_copy
+	cmpl	$3852,%ebx
+	jb	.L019skip_copy
+.align	4
+.L018do_copy:
+	movl	%edi,44(%esp)
+.long	2784229001
+.L019skip_copy:
+	movl	$16,%edi
+.align	4
+.L020prefetch_tbl:
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%esi
+	leal	128(%ebp),%ebp
+	subl	$1,%edi
+	jnz	.L020prefetch_tbl
+	subl	$2048,%ebp
+	movl	32(%esp),%esi
+	movl	48(%esp),%edi
+	cmpl	$0,%edx
+	je	.L021fast_decrypt
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+.align	16
+.L022fast_enc_loop:
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	xorl	(%esi),%eax
+	xorl	4(%esi),%ebx
+	xorl	8(%esi),%ecx
+	xorl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_encrypt
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	leal	16(%esi),%esi
+	movl	40(%esp),%ecx
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L022fast_enc_loop
+	movl	48(%esp),%esi
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	cmpl	$0,316(%esp)
+	movl	44(%esp),%edi
+	je	.L023skip_ezero
+	movl	$60,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2884892297
+.L023skip_ezero:
+	movl	28(%esp),%esp
+	popfl
+.L012drop_out:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L021fast_decrypt:
+	cmpl	36(%esp),%esi
+	je	.L024fast_dec_in_place
+	movl	%edi,52(%esp)
+.align	4
+.align	16
+.L025fast_dec_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt
+	movl	52(%esp),%edi
+	movl	40(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	36(%esp),%edi
+	movl	32(%esp),%esi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	movl	%esi,52(%esp)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edi
+	movl	%edi,36(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L025fast_dec_loop
+	movl	52(%esp),%edi
+	movl	48(%esp),%esi
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	jmp	.L026fast_dec_out
+.align	16
+.L024fast_dec_in_place:
+.L027fast_dec_in_place_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	leal	60(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt
+	movl	48(%esp),%edi
+	movl	36(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	leal	16(%esi),%esi
+	movl	%esi,36(%esp)
+	leal	60(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	32(%esp),%esi
+	movl	40(%esp),%ecx
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	subl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jnz	.L027fast_dec_in_place_loop
+.align	4
+.L026fast_dec_out:
+	cmpl	$0,316(%esp)
+	movl	44(%esp),%edi
+	je	.L028skip_dzero
+	movl	$60,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2884892297
+.L028skip_dzero:
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L015slow_way:
+	movl	(%eax),%eax
+	movl	36(%esp),%edi
+	leal	-80(%esp),%esi
+	andl	$-64,%esi
+	leal	-143(%edi),%ebx
+	subl	%esi,%ebx
+	negl	%ebx
+	andl	$960,%ebx
+	subl	%ebx,%esi
+	leal	768(%esi),%ebx
+	subl	%ebp,%ebx
+	andl	$768,%ebx
+	leal	2176(%ebp,%ebx,1),%ebp
+	leal	24(%esp),%edx
+	xchgl	%esi,%esp
+	addl	$4,%esp
+	movl	%ebp,24(%esp)
+	movl	%esi,28(%esp)
+	movl	%eax,52(%esp)
+	movl	(%edx),%eax
+	movl	4(%edx),%ebx
+	movl	16(%edx),%esi
+	movl	20(%edx),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,40(%esp)
+	movl	%edi,44(%esp)
+	movl	%esi,48(%esp)
+	movl	%esi,%edi
+	movl	%eax,%esi
+	cmpl	$0,%edx
+	je	.L029slow_decrypt
+	cmpl	$16,%ecx
+	movl	%ebx,%edx
+	jb	.L030slow_enc_tail
+	btl	$25,52(%esp)
+	jnc	.L031slow_enc_x86
+	movq	(%edi),%mm0
+	movq	8(%edi),%mm4
+.align	16
+.L032slow_enc_loop_sse:
+	pxor	(%esi),%mm0
+	pxor	8(%esi),%mm4
+	movl	44(%esp),%edi
+	call	_sse_AES_encrypt_compact
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	40(%esp),%ecx
+	movq	%mm0,(%edi)
+	movq	%mm4,8(%edi)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	cmpl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jae	.L032slow_enc_loop_sse
+	testl	$15,%ecx
+	jnz	.L030slow_enc_tail
+	movl	48(%esp),%esi
+	movq	%mm0,(%esi)
+	movq	%mm4,8(%esi)
+	emms
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L031slow_enc_x86:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+.align	4
+.L033slow_enc_loop_x86:
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	xorl	(%esi),%eax
+	xorl	4(%esi),%ebx
+	xorl	8(%esi),%ecx
+	xorl	12(%esi),%edx
+	movl	44(%esp),%edi
+	call	_x86_AES_encrypt_compact
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	leal	16(%edi),%edx
+	movl	%edx,36(%esp)
+	subl	$16,%ecx
+	cmpl	$16,%ecx
+	movl	%ecx,40(%esp)
+	jae	.L033slow_enc_loop_x86
+	testl	$15,%ecx
+	jnz	.L030slow_enc_tail
+	movl	48(%esp),%esi
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L030slow_enc_tail:
+	emms
+	movl	%edx,%edi
+	movl	$16,%ebx
+	subl	%ecx,%ebx
+	cmpl	%esi,%edi
+	je	.L034enc_in_place
+.align	4
+.long	2767451785
+	jmp	.L035enc_skip_in_place
+.L034enc_in_place:
+	leal	(%edi,%ecx,1),%edi
+.L035enc_skip_in_place:
+	movl	%ebx,%ecx
+	xorl	%eax,%eax
+.align	4
+.long	2868115081
+	movl	48(%esp),%edi
+	movl	%edx,%esi
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	$16,40(%esp)
+	jmp	.L033slow_enc_loop_x86
+.align	16
+.L029slow_decrypt:
+	btl	$25,52(%esp)
+	jnc	.L036slow_dec_loop_x86
+.align	4
+.L037slow_dec_loop_sse:
+	movq	(%esi),%mm0
+	movq	8(%esi),%mm4
+	movl	44(%esp),%edi
+	call	_sse_AES_decrypt_compact
+	movl	32(%esp),%esi
+	leal	60(%esp),%eax
+	movl	36(%esp),%ebx
+	movl	40(%esp),%ecx
+	movl	48(%esp),%edi
+	movq	(%esi),%mm1
+	movq	8(%esi),%mm5
+	pxor	(%edi),%mm0
+	pxor	8(%edi),%mm4
+	movq	%mm1,(%edi)
+	movq	%mm5,8(%edi)
+	subl	$16,%ecx
+	jc	.L038slow_dec_partial_sse
+	movq	%mm0,(%ebx)
+	movq	%mm4,8(%ebx)
+	leal	16(%ebx),%ebx
+	movl	%ebx,36(%esp)
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	movl	%ecx,40(%esp)
+	jnz	.L037slow_dec_loop_sse
+	emms
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L038slow_dec_partial_sse:
+	movq	%mm0,(%eax)
+	movq	%mm4,8(%eax)
+	emms
+	addl	$16,%ecx
+	movl	%ebx,%edi
+	movl	%eax,%esi
+.align	4
+.long	2767451785
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L036slow_dec_loop_x86:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	leal	60(%esp),%edi
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	44(%esp),%edi
+	call	_x86_AES_decrypt_compact
+	movl	48(%esp),%edi
+	movl	40(%esp),%esi
+	xorl	(%edi),%eax
+	xorl	4(%edi),%ebx
+	xorl	8(%edi),%ecx
+	xorl	12(%edi),%edx
+	subl	$16,%esi
+	jc	.L039slow_dec_partial_x86
+	movl	%esi,40(%esp)
+	movl	36(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	leal	16(%esi),%esi
+	movl	%esi,36(%esp)
+	leal	60(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	32(%esp),%esi
+	leal	16(%esi),%esi
+	movl	%esi,32(%esp)
+	jnz	.L036slow_dec_loop_x86
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+	pushfl
+.align	16
+.L039slow_dec_partial_x86:
+	leal	60(%esp),%esi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	32(%esp),%esi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	40(%esp),%ecx
+	movl	36(%esp),%edi
+	leal	60(%esp),%esi
+.align	4
+.long	2767451785
+	movl	28(%esp),%esp
+	popfl
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_cbc_encrypt,.-.L_AES_cbc_encrypt_begin
+.type	_x86_AES_set_encrypt_key,@function
+.align	16
+_x86_AES_set_encrypt_key:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	24(%esp),%esi
+	movl	32(%esp),%edi
+	testl	$-1,%esi
+	jz	.L040badpointer
+	testl	$-1,%edi
+	jz	.L040badpointer
+	call	.L041pic_point
+.L041pic_point:
+	popl	%ebp
+	leal	.LAES_Te-.L041pic_point(%ebp),%ebp
+	leal	2176(%ebp),%ebp
+	movl	-128(%ebp),%eax
+	movl	-96(%ebp),%ebx
+	movl	-64(%ebp),%ecx
+	movl	-32(%ebp),%edx
+	movl	(%ebp),%eax
+	movl	32(%ebp),%ebx
+	movl	64(%ebp),%ecx
+	movl	96(%ebp),%edx
+	movl	28(%esp),%ecx
+	cmpl	$128,%ecx
+	je	.L04210rounds
+	cmpl	$192,%ecx
+	je	.L04312rounds
+	cmpl	$256,%ecx
+	je	.L04414rounds
+	movl	$-2,%eax
+	jmp	.L045exit
+.L04210rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L04610shortcut
+.align	4
+.L04710loop:
+	movl	(%edi),%eax
+	movl	12(%edi),%edx
+.L04610shortcut:
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,16(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,20(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,24(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,28(%edi)
+	incl	%ecx
+	addl	$16,%edi
+	cmpl	$10,%ecx
+	jl	.L04710loop
+	movl	$10,80(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L04312rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	16(%esi),%ecx
+	movl	20(%esi),%edx
+	movl	%ecx,16(%edi)
+	movl	%edx,20(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L04812shortcut
+.align	4
+.L04912loop:
+	movl	(%edi),%eax
+	movl	20(%edi),%edx
+.L04812shortcut:
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,24(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,28(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,32(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,36(%edi)
+	cmpl	$7,%ecx
+	je	.L05012break
+	incl	%ecx
+	xorl	16(%edi),%eax
+	movl	%eax,40(%edi)
+	xorl	20(%edi),%eax
+	movl	%eax,44(%edi)
+	addl	$24,%edi
+	jmp	.L04912loop
+.L05012break:
+	movl	$12,72(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L04414rounds:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,8(%edi)
+	movl	%edx,12(%edi)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	movl	%eax,16(%edi)
+	movl	%ebx,20(%edi)
+	movl	%ecx,24(%edi)
+	movl	%edx,28(%edi)
+	xorl	%ecx,%ecx
+	jmp	.L05114shortcut
+.align	4
+.L05214loop:
+	movl	28(%edi),%edx
+.L05114shortcut:
+	movl	(%edi),%eax
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$8,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	xorl	896(%ebp,%ecx,4),%eax
+	movl	%eax,32(%edi)
+	xorl	4(%edi),%eax
+	movl	%eax,36(%edi)
+	xorl	8(%edi),%eax
+	movl	%eax,40(%edi)
+	xorl	12(%edi),%eax
+	movl	%eax,44(%edi)
+	cmpl	$6,%ecx
+	je	.L05314break
+	incl	%ecx
+	movl	%eax,%edx
+	movl	16(%edi),%eax
+	movzbl	%dl,%esi
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shrl	$16,%edx
+	shll	$8,%ebx
+	movzbl	%dl,%esi
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	movzbl	%dh,%esi
+	shll	$16,%ebx
+	xorl	%ebx,%eax
+	movzbl	-128(%ebp,%esi,1),%ebx
+	shll	$24,%ebx
+	xorl	%ebx,%eax
+	movl	%eax,48(%edi)
+	xorl	20(%edi),%eax
+	movl	%eax,52(%edi)
+	xorl	24(%edi),%eax
+	movl	%eax,56(%edi)
+	xorl	28(%edi),%eax
+	movl	%eax,60(%edi)
+	addl	$32,%edi
+	jmp	.L05214loop
+.L05314break:
+	movl	$14,48(%edi)
+	xorl	%eax,%eax
+	jmp	.L045exit
+.L040badpointer:
+	movl	$-1,%eax
+.L045exit:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	_x86_AES_set_encrypt_key,.-_x86_AES_set_encrypt_key
+.globl	AES_set_encrypt_key
+.type	AES_set_encrypt_key,@function
+.align	16
+AES_set_encrypt_key:
+.L_AES_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	call	_x86_AES_set_encrypt_key
+	ret
+.size	AES_set_encrypt_key,.-.L_AES_set_encrypt_key_begin
+.globl	AES_set_decrypt_key
+.type	AES_set_decrypt_key,@function
+.align	16
+AES_set_decrypt_key:
+.L_AES_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
+	call	_x86_AES_set_encrypt_key
+	cmpl	$0,%eax
+	je	.L054proceed
+	ret
+.L054proceed:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%esi
+	movl	240(%esi),%ecx
+	leal	(,%ecx,4),%ecx
+	leal	(%esi,%ecx,4),%edi
+.align	4
+.L055invert:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	(%edi),%ecx
+	movl	4(%edi),%edx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	movl	%ecx,(%esi)
+	movl	%edx,4(%esi)
+	movl	8(%esi),%eax
+	movl	12(%esi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	%eax,8(%edi)
+	movl	%ebx,12(%edi)
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	addl	$16,%esi
+	subl	$16,%edi
+	cmpl	%edi,%esi
+	jne	.L055invert
+	movl	28(%esp),%edi
+	movl	240(%edi),%esi
+	leal	-2(%esi,%esi,1),%esi
+	leal	(%edi,%esi,8),%esi
+	movl	%esi,28(%esp)
+	movl	16(%edi),%eax
+.align	4
+.L056permute:
+	addl	$16,%edi
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%esi,%ebx
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%eax,%ebx
+	xorl	%esi,%ecx
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%eax,%ecx
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	roll	$8,%eax
+	xorl	%esi,%edx
+	movl	4(%edi),%ebp
+	xorl	%ebx,%eax
+	xorl	%edx,%ebx
+	xorl	%ecx,%eax
+	roll	$24,%ebx
+	xorl	%edx,%ecx
+	xorl	%edx,%eax
+	roll	$16,%ecx
+	xorl	%ebx,%eax
+	roll	$8,%edx
+	xorl	%ecx,%eax
+	movl	%ebp,%ebx
+	xorl	%edx,%eax
+	movl	%eax,(%edi)
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	xorl	%esi,%ecx
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%ebx,%ecx
+	xorl	%esi,%edx
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%ebx,%edx
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	roll	$8,%ebx
+	xorl	%esi,%eax
+	movl	8(%edi),%ebp
+	xorl	%ecx,%ebx
+	xorl	%eax,%ecx
+	xorl	%edx,%ebx
+	roll	$24,%ecx
+	xorl	%eax,%edx
+	xorl	%eax,%ebx
+	roll	$16,%edx
+	xorl	%ecx,%ebx
+	roll	$8,%eax
+	xorl	%edx,%ebx
+	movl	%ebp,%ecx
+	xorl	%eax,%ebx
+	movl	%ebx,4(%edi)
+	movl	$2155905152,%ebp
+	andl	%ecx,%ebp
+	leal	(%ecx,%ecx,1),%edx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%edx
+	andl	$454761243,%esi
+	xorl	%esi,%edx
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%ecx,%edx
+	xorl	%esi,%eax
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%ecx,%eax
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	roll	$8,%ecx
+	xorl	%esi,%ebx
+	movl	12(%edi),%ebp
+	xorl	%edx,%ecx
+	xorl	%ebx,%edx
+	xorl	%eax,%ecx
+	roll	$24,%edx
+	xorl	%ebx,%eax
+	xorl	%ebx,%ecx
+	roll	$16,%eax
+	xorl	%edx,%ecx
+	roll	$8,%ebx
+	xorl	%eax,%ecx
+	movl	%ebp,%edx
+	xorl	%ebx,%ecx
+	movl	%ecx,8(%edi)
+	movl	$2155905152,%ebp
+	andl	%edx,%ebp
+	leal	(%edx,%edx,1),%eax
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%eax
+	andl	$454761243,%esi
+	xorl	%esi,%eax
+	movl	$2155905152,%ebp
+	andl	%eax,%ebp
+	leal	(%eax,%eax,1),%ebx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	subl	%ebp,%esi
+	andl	$4278124286,%ebx
+	andl	$454761243,%esi
+	xorl	%edx,%eax
+	xorl	%esi,%ebx
+	movl	$2155905152,%ebp
+	andl	%ebx,%ebp
+	leal	(%ebx,%ebx,1),%ecx
+	movl	%ebp,%esi
+	shrl	$7,%ebp
+	xorl	%edx,%ebx
+	subl	%ebp,%esi
+	andl	$4278124286,%ecx
+	andl	$454761243,%esi
+	roll	$8,%edx
+	xorl	%esi,%ecx
+	movl	16(%edi),%ebp
+	xorl	%eax,%edx
+	xorl	%ecx,%eax
+	xorl	%ebx,%edx
+	roll	$24,%eax
+	xorl	%ecx,%ebx
+	xorl	%ecx,%edx
+	roll	$16,%ebx
+	xorl	%eax,%edx
+	roll	$8,%ecx
+	xorl	%ebx,%edx
+	movl	%ebp,%eax
+	xorl	%ecx,%edx
+	movl	%edx,12(%edi)
+	cmpl	28(%esp),%edi
+	jb	.L056permute
+	xorl	%eax,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	AES_set_decrypt_key,.-.L_AES_set_decrypt_key_begin
+.byte	65,69,83,32,102,111,114,32,120,56,54,44,32,67,82,89
+.byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
+.byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
+#endif
diff --git a/sys/crypto/openssl/i386/aesni-x86.S b/sys/crypto/openssl/i386/aesni-x86.S
index ea2058f2bad8a..9dbcba730f33f 100644
--- a/sys/crypto/openssl/i386/aesni-x86.S
+++ b/sys/crypto/openssl/i386/aesni-x86.S
@@ -6,6 +6,11 @@
 .align	16
 aesni_encrypt:
 .L_aesni_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	12(%esp),%edx
 	movups	(%eax),%xmm2
@@ -33,6 +38,11 @@ aesni_encrypt:
 .align	16
 aesni_decrypt:
 .L_aesni_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	12(%esp),%edx
 	movups	(%eax),%xmm2
@@ -58,6 +68,11 @@ aesni_decrypt:
 .type	_aesni_encrypt2,@function
 .align	16
 _aesni_encrypt2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -85,6 +100,11 @@ _aesni_encrypt2:
 .type	_aesni_decrypt2,@function
 .align	16
 _aesni_decrypt2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -112,6 +132,11 @@ _aesni_decrypt2:
 .type	_aesni_encrypt3,@function
 .align	16
 _aesni_encrypt3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -144,6 +169,11 @@ _aesni_encrypt3:
 .type	_aesni_decrypt3,@function
 .align	16
 _aesni_decrypt3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -176,6 +206,11 @@ _aesni_decrypt3:
 .type	_aesni_encrypt4,@function
 .align	16
 _aesni_encrypt4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	movups	16(%edx),%xmm1
 	shll	$4,%ecx
@@ -214,6 +249,11 @@ _aesni_encrypt4:
 .type	_aesni_decrypt4,@function
 .align	16
 _aesni_decrypt4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	movups	16(%edx),%xmm1
 	shll	$4,%ecx
@@ -252,6 +292,11 @@ _aesni_decrypt4:
 .type	_aesni_encrypt6,@function
 .align	16
 _aesni_encrypt6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -306,6 +351,11 @@ _aesni_encrypt6:
 .type	_aesni_decrypt6,@function
 .align	16
 _aesni_decrypt6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -362,6 +412,11 @@ _aesni_decrypt6:
 .align	16
 aesni_ecb_encrypt:
 .L_aesni_ecb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -597,6 +652,11 @@ aesni_ecb_encrypt:
 .align	16
 aesni_ccm64_encrypt_blocks:
 .L_aesni_ccm64_encrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -685,6 +745,11 @@ aesni_ccm64_encrypt_blocks:
 .align	16
 aesni_ccm64_decrypt_blocks:
 .L_aesni_ccm64_decrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -808,6 +873,11 @@ aesni_ccm64_decrypt_blocks:
 .align	16
 aesni_ctr32_encrypt_blocks:
 .L_aesni_ctr32_encrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1046,6 +1116,11 @@ aesni_ctr32_encrypt_blocks:
 .align	16
 aesni_xts_encrypt:
 .L_aesni_xts_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1406,6 +1481,11 @@ aesni_xts_encrypt:
 .align	16
 aesni_xts_decrypt:
 .L_aesni_xts_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1796,6 +1876,11 @@ aesni_xts_decrypt:
 .align	16
 aesni_ocb_encrypt:
 .L_aesni_ocb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2191,6 +2276,11 @@ aesni_ocb_encrypt:
 .align	16
 aesni_ocb_decrypt:
 .L_aesni_ocb_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2586,6 +2676,11 @@ aesni_ocb_decrypt:
 .align	16
 aesni_cbc_encrypt:
 .L_aesni_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2845,6 +2940,11 @@ aesni_cbc_encrypt:
 .type	_aesni_set_encrypt_key,@function
 .align	16
 _aesni_set_encrypt_key:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	testl	%eax,%eax
@@ -3180,6 +3280,11 @@ _aesni_set_encrypt_key:
 .align	16
 aesni_set_encrypt_key:
 .L_aesni_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%ecx
 	movl	12(%esp),%edx
@@ -3191,6 +3296,11 @@ aesni_set_encrypt_key:
 .align	16
 aesni_set_decrypt_key:
 .L_aesni_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%ecx
 	movl	12(%esp),%edx
@@ -3237,6 +3347,23 @@ aesni_set_decrypt_key:
 .byte	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
 .byte	115,108,46,111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	aesni_encrypt
@@ -3244,6 +3371,11 @@ aesni_set_decrypt_key:
 .align	16
 aesni_encrypt:
 .L_aesni_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	12(%esp),%edx
 	movups	(%eax),%xmm2
@@ -3271,6 +3403,11 @@ aesni_encrypt:
 .align	16
 aesni_decrypt:
 .L_aesni_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	12(%esp),%edx
 	movups	(%eax),%xmm2
@@ -3296,6 +3433,11 @@ aesni_decrypt:
 .type	_aesni_encrypt2,@function
 .align	16
 _aesni_encrypt2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3323,6 +3465,11 @@ _aesni_encrypt2:
 .type	_aesni_decrypt2,@function
 .align	16
 _aesni_decrypt2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3350,6 +3497,11 @@ _aesni_decrypt2:
 .type	_aesni_encrypt3,@function
 .align	16
 _aesni_encrypt3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3382,6 +3534,11 @@ _aesni_encrypt3:
 .type	_aesni_decrypt3,@function
 .align	16
 _aesni_decrypt3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3414,6 +3571,11 @@ _aesni_decrypt3:
 .type	_aesni_encrypt4,@function
 .align	16
 _aesni_encrypt4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	movups	16(%edx),%xmm1
 	shll	$4,%ecx
@@ -3452,6 +3614,11 @@ _aesni_encrypt4:
 .type	_aesni_decrypt4,@function
 .align	16
 _aesni_decrypt4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	movups	16(%edx),%xmm1
 	shll	$4,%ecx
@@ -3490,6 +3657,11 @@ _aesni_decrypt4:
 .type	_aesni_encrypt6,@function
 .align	16
 _aesni_encrypt6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3544,6 +3716,11 @@ _aesni_encrypt6:
 .type	_aesni_decrypt6,@function
 .align	16
 _aesni_decrypt6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movups	(%edx),%xmm0
 	shll	$4,%ecx
 	movups	16(%edx),%xmm1
@@ -3600,6 +3777,11 @@ _aesni_decrypt6:
 .align	16
 aesni_ecb_encrypt:
 .L_aesni_ecb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3835,6 +4017,11 @@ aesni_ecb_encrypt:
 .align	16
 aesni_ccm64_encrypt_blocks:
 .L_aesni_ccm64_encrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3923,6 +4110,11 @@ aesni_ccm64_encrypt_blocks:
 .align	16
 aesni_ccm64_decrypt_blocks:
 .L_aesni_ccm64_decrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4046,6 +4238,11 @@ aesni_ccm64_decrypt_blocks:
 .align	16
 aesni_ctr32_encrypt_blocks:
 .L_aesni_ctr32_encrypt_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4284,6 +4481,11 @@ aesni_ctr32_encrypt_blocks:
 .align	16
 aesni_xts_encrypt:
 .L_aesni_xts_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4644,6 +4846,11 @@ aesni_xts_encrypt:
 .align	16
 aesni_xts_decrypt:
 .L_aesni_xts_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5034,6 +5241,11 @@ aesni_xts_decrypt:
 .align	16
 aesni_ocb_encrypt:
 .L_aesni_ocb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5429,6 +5641,11 @@ aesni_ocb_encrypt:
 .align	16
 aesni_ocb_decrypt:
 .L_aesni_ocb_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5824,6 +6041,11 @@ aesni_ocb_decrypt:
 .align	16
 aesni_cbc_encrypt:
 .L_aesni_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -6083,6 +6305,11 @@ aesni_cbc_encrypt:
 .type	_aesni_set_encrypt_key,@function
 .align	16
 _aesni_set_encrypt_key:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	testl	%eax,%eax
@@ -6418,6 +6645,11 @@ _aesni_set_encrypt_key:
 .align	16
 aesni_set_encrypt_key:
 .L_aesni_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%ecx
 	movl	12(%esp),%edx
@@ -6429,6 +6661,11 @@ aesni_set_encrypt_key:
 .align	16
 aesni_set_decrypt_key:
 .L_aesni_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%ecx
 	movl	12(%esp),%edx
@@ -6475,4 +6712,21 @@ aesni_set_decrypt_key:
 .byte	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
 .byte	115,108,46,111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/bf-586.S b/sys/crypto/openssl/i386/bf-586.S
index fd533bb993152..5f68ffad0f577 100644
--- a/sys/crypto/openssl/i386/bf-586.S
+++ b/sys/crypto/openssl/i386/bf-586.S
@@ -6,6 +6,11 @@
 .align	16
 BF_encrypt:
 .L_BF_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -356,6 +361,11 @@ BF_encrypt:
 .align	16
 BF_decrypt:
 .L_BF_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -706,6 +716,11 @@ BF_decrypt:
 .align	16
 BF_cbc_encrypt:
 .L_BF_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -769,21 +784,56 @@ BF_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L006ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L007ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L008ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L009ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L010ejend
 .L011ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L012ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L013ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L010ejend:
 	xorl	%ecx,%eax
@@ -895,6 +945,23 @@ BF_cbc_encrypt:
 .long	.L006ej7-.L004PIC_point
 .align	64
 .size	BF_cbc_encrypt,.-.L_BF_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	BF_encrypt
@@ -902,6 +969,11 @@ BF_cbc_encrypt:
 .align	16
 BF_encrypt:
 .L_BF_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1252,6 +1324,11 @@ BF_encrypt:
 .align	16
 BF_decrypt:
 .L_BF_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1602,6 +1679,11 @@ BF_decrypt:
 .align	16
 BF_cbc_encrypt:
 .L_BF_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1665,21 +1747,56 @@ BF_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L006ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L007ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L008ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L009ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L010ejend
 .L011ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L012ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L013ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L010ejend:
 	xorl	%ecx,%eax
@@ -1791,4 +1908,21 @@ BF_cbc_encrypt:
 .long	.L006ej7-.L004PIC_point
 .align	64
 .size	BF_cbc_encrypt,.-.L_BF_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/bn-586.S b/sys/crypto/openssl/i386/bn-586.S
index 512bcf3a82dbd..ebbea406881fb 100644
--- a/sys/crypto/openssl/i386/bn-586.S
+++ b/sys/crypto/openssl/i386/bn-586.S
@@ -6,6 +6,11 @@
 .align	16
 bn_mul_add_words:
 .L_bn_mul_add_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L000PIC_me_up
 .L000PIC_me_up:
 	popl	%eax
@@ -289,6 +294,11 @@ bn_mul_add_words:
 .align	16
 bn_mul_words:
 .L_bn_mul_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L010PIC_me_up
 .L010PIC_me_up:
 	popl	%eax
@@ -471,6 +481,11 @@ bn_mul_words:
 .align	16
 bn_sqr_words:
 .L_bn_sqr_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L017PIC_me_up
 .L017PIC_me_up:
 	popl	%eax
@@ -612,6 +627,11 @@ bn_sqr_words:
 .align	16
 bn_div_words:
 .L_bn_div_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	movl	12(%esp),%ecx
@@ -623,6 +643,11 @@ bn_div_words:
 .align	16
 bn_add_words:
 .L_bn_add_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -805,6 +830,11 @@ bn_add_words:
 .align	16
 bn_sub_words:
 .L_bn_sub_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -987,6 +1017,11 @@ bn_sub_words:
 .align	16
 bn_sub_part_words:
 .L_bn_sub_part_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1529,6 +1564,23 @@ bn_sub_part_words:
 	ret
 .size	bn_sub_part_words,.-.L_bn_sub_part_words_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	bn_mul_add_words
@@ -1536,6 +1588,11 @@ bn_sub_part_words:
 .align	16
 bn_mul_add_words:
 .L_bn_mul_add_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	OPENSSL_ia32cap_P,%eax
 	btl	$26,(%eax)
 	jnc	.L000maw_non_sse2
@@ -1816,6 +1873,11 @@ bn_mul_add_words:
 .align	16
 bn_mul_words:
 .L_bn_mul_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	OPENSSL_ia32cap_P,%eax
 	btl	$26,(%eax)
 	jnc	.L009mw_non_sse2
@@ -1995,6 +2057,11 @@ bn_mul_words:
 .align	16
 bn_sqr_words:
 .L_bn_sqr_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	OPENSSL_ia32cap_P,%eax
 	btl	$26,(%eax)
 	jnc	.L015sqr_non_sse2
@@ -2133,6 +2200,11 @@ bn_sqr_words:
 .align	16
 bn_div_words:
 .L_bn_div_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	movl	12(%esp),%ecx
@@ -2144,6 +2216,11 @@ bn_div_words:
 .align	16
 bn_add_words:
 .L_bn_add_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2326,6 +2403,11 @@ bn_add_words:
 .align	16
 bn_sub_words:
 .L_bn_sub_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2508,6 +2590,11 @@ bn_sub_words:
 .align	16
 bn_sub_part_words:
 .L_bn_sub_part_words_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3050,4 +3137,21 @@ bn_sub_part_words:
 	ret
 .size	bn_sub_part_words,.-.L_bn_sub_part_words_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/cast-586.S b/sys/crypto/openssl/i386/cast-586.S
index fea8b021a95f0..f016559e7fb42 100644
--- a/sys/crypto/openssl/i386/cast-586.S
+++ b/sys/crypto/openssl/i386/cast-586.S
@@ -6,6 +6,11 @@
 .align	16
 CAST_encrypt:
 .L_CAST_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -376,6 +381,11 @@ CAST_encrypt:
 .align	16
 CAST_decrypt:
 .L_CAST_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -743,6 +753,11 @@ CAST_decrypt:
 .align	16
 CAST_cbc_encrypt:
 .L_CAST_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -806,21 +821,56 @@ CAST_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L008ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L009ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L010ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L011ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L012ejend
 .L013ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L014ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L015ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L012ejend:
 	xorl	%ecx,%eax
@@ -932,6 +982,23 @@ CAST_cbc_encrypt:
 .long	.L008ej7-.L006PIC_point
 .align	64
 .size	CAST_cbc_encrypt,.-.L_CAST_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	CAST_encrypt
@@ -939,6 +1006,11 @@ CAST_cbc_encrypt:
 .align	16
 CAST_encrypt:
 .L_CAST_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1309,6 +1381,11 @@ CAST_encrypt:
 .align	16
 CAST_decrypt:
 .L_CAST_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1676,6 +1753,11 @@ CAST_decrypt:
 .align	16
 CAST_cbc_encrypt:
 .L_CAST_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1739,21 +1821,56 @@ CAST_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L008ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L009ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L010ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L011ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L012ejend
 .L013ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L014ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L015ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L012ejend:
 	xorl	%ecx,%eax
@@ -1865,4 +1982,21 @@ CAST_cbc_encrypt:
 .long	.L008ej7-.L006PIC_point
 .align	64
 .size	CAST_cbc_encrypt,.-.L_CAST_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/chacha-x86.S b/sys/crypto/openssl/i386/chacha-x86.S
index a45c3e3a37e1f..df2b4c3f4f95c 100644
--- a/sys/crypto/openssl/i386/chacha-x86.S
+++ b/sys/crypto/openssl/i386/chacha-x86.S
@@ -6,6 +6,11 @@
 .align	16
 ChaCha20_ctr32:
 .L_ChaCha20_ctr32_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -379,6 +384,11 @@ ChaCha20_ctr32:
 .align	16
 ChaCha20_ssse3:
 .L_ChaCha20_ssse3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -534,6 +544,11 @@ ChaCha20_ssse3:
 .align	16
 ChaCha20_xop:
 .L_ChaCha20_xop_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1008,6 +1023,23 @@ ChaCha20_xop:
 	ret
 .size	ChaCha20_xop,.-.L_ChaCha20_xop_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	ChaCha20_ctr32
@@ -1015,6 +1047,11 @@ ChaCha20_xop:
 .align	16
 ChaCha20_ctr32:
 .L_ChaCha20_ctr32_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1388,6 +1425,11 @@ ChaCha20_ctr32:
 .align	16
 ChaCha20_ssse3:
 .L_ChaCha20_ssse3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1543,6 +1585,11 @@ ChaCha20_ssse3:
 .align	16
 ChaCha20_xop:
 .L_ChaCha20_xop_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2017,4 +2064,21 @@ ChaCha20_xop:
 	ret
 .size	ChaCha20_xop,.-.L_ChaCha20_xop_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/cmll-x86.S b/sys/crypto/openssl/i386/cmll-x86.S
index 3a49ea43c2308..f47261016e40e 100644
--- a/sys/crypto/openssl/i386/cmll-x86.S
+++ b/sys/crypto/openssl/i386/cmll-x86.S
@@ -6,6 +6,11 @@
 .align	16
 Camellia_EncryptBlock_Rounds:
 .L_Camellia_EncryptBlock_Rounds_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -60,6 +65,11 @@ Camellia_EncryptBlock_Rounds:
 .align	16
 Camellia_EncryptBlock:
 .L_Camellia_EncryptBlock_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$128,%eax
 	subl	4(%esp),%eax
 	movl	$3,%eax
@@ -72,6 +82,11 @@ Camellia_EncryptBlock:
 .align	16
 Camellia_encrypt:
 .L_Camellia_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -124,6 +139,11 @@ Camellia_encrypt:
 .type	_x86_Camellia_encrypt,@function
 .align	16
 _x86_Camellia_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	(%edi),%eax
 	xorl	4(%edi),%ebx
 	xorl	8(%edi),%ecx
@@ -354,6 +374,11 @@ _x86_Camellia_encrypt:
 .align	16
 Camellia_DecryptBlock_Rounds:
 .L_Camellia_DecryptBlock_Rounds_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -408,6 +433,11 @@ Camellia_DecryptBlock_Rounds:
 .align	16
 Camellia_DecryptBlock:
 .L_Camellia_DecryptBlock_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$128,%eax
 	subl	4(%esp),%eax
 	movl	$3,%eax
@@ -420,6 +450,11 @@ Camellia_DecryptBlock:
 .align	16
 Camellia_decrypt:
 .L_Camellia_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -472,6 +507,11 @@ Camellia_decrypt:
 .type	_x86_Camellia_decrypt,@function
 .align	16
 _x86_Camellia_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	(%edi),%eax
 	xorl	4(%edi),%ebx
 	xorl	8(%edi),%ecx
@@ -702,6 +742,11 @@ _x86_Camellia_decrypt:
 .align	16
 Camellia_Ekeygen:
 .L_Camellia_Ekeygen_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1543,6 +1588,11 @@ Camellia_Ekeygen:
 .align	16
 Camellia_set_key:
 .L_Camellia_set_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ecx
 	movl	12(%esp),%ebx
@@ -2095,6 +2145,11 @@ Camellia_set_key:
 .align	16
 Camellia_cbc_encrypt:
 .L_Camellia_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2374,6 +2429,23 @@ Camellia_cbc_encrypt:
 .byte	67,97,109,101,108,108,105,97,32,102,111,114,32,120,56,54
 .byte	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
 .byte	115,108,46,111,114,103,62,0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	Camellia_EncryptBlock_Rounds
@@ -2381,6 +2453,11 @@ Camellia_cbc_encrypt:
 .align	16
 Camellia_EncryptBlock_Rounds:
 .L_Camellia_EncryptBlock_Rounds_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2435,6 +2512,11 @@ Camellia_EncryptBlock_Rounds:
 .align	16
 Camellia_EncryptBlock:
 .L_Camellia_EncryptBlock_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$128,%eax
 	subl	4(%esp),%eax
 	movl	$3,%eax
@@ -2447,6 +2529,11 @@ Camellia_EncryptBlock:
 .align	16
 Camellia_encrypt:
 .L_Camellia_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2499,6 +2586,11 @@ Camellia_encrypt:
 .type	_x86_Camellia_encrypt,@function
 .align	16
 _x86_Camellia_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	(%edi),%eax
 	xorl	4(%edi),%ebx
 	xorl	8(%edi),%ecx
@@ -2729,6 +2821,11 @@ _x86_Camellia_encrypt:
 .align	16
 Camellia_DecryptBlock_Rounds:
 .L_Camellia_DecryptBlock_Rounds_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2783,6 +2880,11 @@ Camellia_DecryptBlock_Rounds:
 .align	16
 Camellia_DecryptBlock:
 .L_Camellia_DecryptBlock_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$128,%eax
 	subl	4(%esp),%eax
 	movl	$3,%eax
@@ -2795,6 +2897,11 @@ Camellia_DecryptBlock:
 .align	16
 Camellia_decrypt:
 .L_Camellia_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2847,6 +2954,11 @@ Camellia_decrypt:
 .type	_x86_Camellia_decrypt,@function
 .align	16
 _x86_Camellia_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	(%edi),%eax
 	xorl	4(%edi),%ebx
 	xorl	8(%edi),%ecx
@@ -3077,6 +3189,11 @@ _x86_Camellia_decrypt:
 .align	16
 Camellia_Ekeygen:
 .L_Camellia_Ekeygen_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3918,6 +4035,11 @@ Camellia_Ekeygen:
 .align	16
 Camellia_set_key:
 .L_Camellia_set_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ecx
 	movl	12(%esp),%ebx
@@ -4470,6 +4592,11 @@ Camellia_set_key:
 .align	16
 Camellia_cbc_encrypt:
 .L_Camellia_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4749,4 +4876,21 @@ Camellia_cbc_encrypt:
 .byte	67,97,109,101,108,108,105,97,32,102,111,114,32,120,56,54
 .byte	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
 .byte	115,108,46,111,114,103,62,0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/co-586.S b/sys/crypto/openssl/i386/co-586.S
index e3f734245def6..ed8f2d14f2eea 100644
--- a/sys/crypto/openssl/i386/co-586.S
+++ b/sys/crypto/openssl/i386/co-586.S
@@ -6,6 +6,11 @@
 .align	16
 bn_mul_comba8:
 .L_bn_mul_comba8_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	movl	12(%esp),%esi
 	pushl	%edi
@@ -550,6 +555,11 @@ bn_mul_comba8:
 .align	16
 bn_mul_comba4:
 .L_bn_mul_comba4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	movl	12(%esp),%esi
 	pushl	%edi
@@ -718,6 +728,11 @@ bn_mul_comba4:
 .align	16
 bn_sqr_comba8:
 .L_bn_sqr_comba8_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	pushl	%ebp
@@ -1126,6 +1141,11 @@ bn_sqr_comba8:
 .align	16
 bn_sqr_comba4:
 .L_bn_sqr_comba4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	pushl	%ebp
@@ -1253,6 +1273,23 @@ bn_sqr_comba4:
 	popl	%esi
 	ret
 .size	bn_sqr_comba4,.-.L_bn_sqr_comba4_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	bn_mul_comba8
@@ -1260,6 +1297,11 @@ bn_sqr_comba4:
 .align	16
 bn_mul_comba8:
 .L_bn_mul_comba8_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	movl	12(%esp),%esi
 	pushl	%edi
@@ -1804,6 +1846,11 @@ bn_mul_comba8:
 .align	16
 bn_mul_comba4:
 .L_bn_mul_comba4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	movl	12(%esp),%esi
 	pushl	%edi
@@ -1972,6 +2019,11 @@ bn_mul_comba4:
 .align	16
 bn_sqr_comba8:
 .L_bn_sqr_comba8_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	pushl	%ebp
@@ -2380,6 +2432,11 @@ bn_sqr_comba8:
 .align	16
 bn_sqr_comba4:
 .L_bn_sqr_comba4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	pushl	%ebp
@@ -2507,4 +2564,21 @@ bn_sqr_comba4:
 	popl	%esi
 	ret
 .size	bn_sqr_comba4,.-.L_bn_sqr_comba4_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/crypt586.S b/sys/crypto/openssl/i386/crypt586.S
index fd5e72fd8abe7..b86cc00e680c9 100644
--- a/sys/crypto/openssl/i386/crypt586.S
+++ b/sys/crypto/openssl/i386/crypt586.S
@@ -6,6 +6,11 @@
 .align	16
 fcrypt_body:
 .L_fcrypt_body_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -878,6 +883,23 @@ fcrypt_body:
 	popl	%ebp
 	ret
 .size	fcrypt_body,.-.L_fcrypt_body_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	fcrypt_body
@@ -885,6 +907,11 @@ fcrypt_body:
 .align	16
 fcrypt_body:
 .L_fcrypt_body_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1753,4 +1780,21 @@ fcrypt_body:
 	popl	%ebp
 	ret
 .size	fcrypt_body,.-.L_fcrypt_body_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/des-586.S b/sys/crypto/openssl/i386/des-586.S
index fd9506d92c4c2..41c22a851e88a 100644
--- a/sys/crypto/openssl/i386/des-586.S
+++ b/sys/crypto/openssl/i386/des-586.S
@@ -5,6 +5,11 @@
 .type	_x86_DES_encrypt,@function
 .align	16
 _x86_DES_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ecx
 
 	movl	(%ecx),%eax
@@ -476,6 +481,11 @@ _x86_DES_encrypt:
 .type	_x86_DES_decrypt,@function
 .align	16
 _x86_DES_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ecx
 
 	movl	120(%ecx),%eax
@@ -949,6 +959,11 @@ _x86_DES_decrypt:
 .align	16
 DES_encrypt1:
 .L_DES_encrypt1_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 
@@ -1062,6 +1077,11 @@ DES_encrypt1:
 .align	16
 DES_encrypt2:
 .L_DES_encrypt2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 
@@ -1105,6 +1125,11 @@ DES_encrypt2:
 .align	16
 DES_encrypt3:
 .L_DES_encrypt3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ebx
 	pushl	%ebp
@@ -1226,6 +1251,11 @@ DES_encrypt3:
 .align	16
 DES_decrypt3:
 .L_DES_decrypt3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ebx
 	pushl	%ebp
@@ -1347,6 +1377,11 @@ DES_decrypt3:
 .align	16
 DES_ncbc_encrypt:
 .L_DES_ncbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1408,21 +1443,56 @@ DES_ncbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L012ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L013ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L014ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L015ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L016ejend
 .L017ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L018ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L019ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L016ejend:
 	xorl	%ecx,%eax
@@ -1527,6 +1597,11 @@ DES_ncbc_encrypt:
 .align	16
 DES_ede3_cbc_encrypt:
 .L_DES_ede3_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1592,21 +1667,56 @@ DES_ede3_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L036ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L037ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L038ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L039ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L040ejend
 .L041ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L042ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L043ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L040ejend:
 	xorl	%ecx,%eax
@@ -1837,12 +1947,34 @@ DES_SPtrans:
 .long	8519680,131200,537002112,545259520
 .long	128,545390592,8519808,0
 .long	536870912,545259648,131072,8519808
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	DES_SPtrans
 .type	_x86_DES_encrypt,@function
 .align	16
 _x86_DES_encrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ecx
 
 	movl	(%ecx),%eax
@@ -2314,6 +2446,11 @@ _x86_DES_encrypt:
 .type	_x86_DES_decrypt,@function
 .align	16
 _x86_DES_decrypt:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ecx
 
 	movl	120(%ecx),%eax
@@ -2787,6 +2924,11 @@ _x86_DES_decrypt:
 .align	16
 DES_encrypt1:
 .L_DES_encrypt1_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 
@@ -2900,6 +3042,11 @@ DES_encrypt1:
 .align	16
 DES_encrypt2:
 .L_DES_encrypt2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 
@@ -2943,6 +3090,11 @@ DES_encrypt2:
 .align	16
 DES_encrypt3:
 .L_DES_encrypt3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ebx
 	pushl	%ebp
@@ -3064,6 +3216,11 @@ DES_encrypt3:
 .align	16
 DES_decrypt3:
 .L_DES_decrypt3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	movl	8(%esp),%ebx
 	pushl	%ebp
@@ -3185,6 +3342,11 @@ DES_decrypt3:
 .align	16
 DES_ncbc_encrypt:
 .L_DES_ncbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -3246,21 +3408,56 @@ DES_ncbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L012ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L013ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L014ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L015ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L016ejend
 .L017ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L018ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L019ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L016ejend:
 	xorl	%ecx,%eax
@@ -3365,6 +3562,11 @@ DES_ncbc_encrypt:
 .align	16
 DES_ede3_cbc_encrypt:
 .L_DES_ede3_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -3430,21 +3632,56 @@ DES_ede3_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L036ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L037ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L038ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L039ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L040ejend
 .L041ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L042ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L043ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L040ejend:
 	xorl	%ecx,%eax
@@ -3675,4 +3912,21 @@ DES_SPtrans:
 .long	8519680,131200,537002112,545259520
 .long	128,545390592,8519808,0
 .long	536870912,545259648,131072,8519808
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/e_padlock-x86.S b/sys/crypto/openssl/i386/e_padlock-x86.S
index 66a152488e90f..164ebe3cdd80a 100644
--- a/sys/crypto/openssl/i386/e_padlock-x86.S
+++ b/sys/crypto/openssl/i386/e_padlock-x86.S
@@ -6,6 +6,11 @@
 .align	16
 padlock_capability:
 .L_padlock_capability_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	pushfl
 	popl	%eax
@@ -66,6 +71,11 @@ padlock_capability:
 .align	16
 padlock_key_bswap:
 .L_padlock_key_bswap_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	240(%edx),%ecx
 	incl	%ecx
@@ -84,6 +94,11 @@ padlock_key_bswap:
 .align	16
 padlock_verify_context:
 .L_padlock_verify_context_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	leal	.Lpadlock_saved_context-.L004verify_pic_point,%eax
 	pushfl
@@ -95,6 +110,11 @@ padlock_verify_context:
 .type	_padlock_verify_ctx,@function
 .align	16
 _padlock_verify_ctx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%eax
 	btl	$30,4(%esp)
 	jnc	.L005verified
@@ -111,6 +131,11 @@ _padlock_verify_ctx:
 .align	16
 padlock_reload_key:
 .L_padlock_reload_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushfl
 	popfl
 	ret
@@ -120,6 +145,11 @@ padlock_reload_key:
 .align	16
 padlock_aes_block:
 .L_padlock_aes_block_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	pushl	%ebx
@@ -140,6 +170,11 @@ padlock_aes_block:
 .align	16
 padlock_ecb_encrypt:
 .L_padlock_ecb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -319,6 +354,11 @@ padlock_ecb_encrypt:
 .align	16
 padlock_cbc_encrypt:
 .L_padlock_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -502,6 +542,11 @@ padlock_cbc_encrypt:
 .align	16
 padlock_cfb_encrypt:
 .L_padlock_cfb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -624,6 +669,11 @@ padlock_cfb_encrypt:
 .align	16
 padlock_ofb_encrypt:
 .L_padlock_ofb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -746,6 +796,11 @@ padlock_ofb_encrypt:
 .align	16
 padlock_ctr32_encrypt:
 .L_padlock_ctr32_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -853,6 +908,11 @@ padlock_ctr32_encrypt:
 .align	16
 padlock_xstore:
 .L_padlock_xstore_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	movl	8(%esp),%edi
 	movl	12(%esp),%edx
@@ -863,6 +923,11 @@ padlock_xstore:
 .type	_win32_segv_handler,@function
 .align	16
 _win32_segv_handler:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$1,%eax
 	movl	4(%esp),%edx
 	movl	12(%esp),%ecx
@@ -878,6 +943,11 @@ _win32_segv_handler:
 .align	16
 padlock_sha1_oneshot:
 .L_padlock_sha1_oneshot_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	xorl	%eax,%eax
@@ -909,6 +979,11 @@ padlock_sha1_oneshot:
 .align	16
 padlock_sha1_blocks:
 .L_padlock_sha1_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -939,6 +1014,11 @@ padlock_sha1_blocks:
 .align	16
 padlock_sha256_oneshot:
 .L_padlock_sha256_oneshot_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	xorl	%eax,%eax
@@ -970,6 +1050,11 @@ padlock_sha256_oneshot:
 .align	16
 padlock_sha256_blocks:
 .L_padlock_sha256_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -1000,6 +1085,11 @@ padlock_sha256_blocks:
 .align	16
 padlock_sha512_blocks:
 .L_padlock_sha512_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -1041,6 +1131,23 @@ padlock_sha512_blocks:
 .align	4
 .Lpadlock_saved_context:
 .long	0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	padlock_capability
@@ -1048,6 +1155,11 @@ padlock_sha512_blocks:
 .align	16
 padlock_capability:
 .L_padlock_capability_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebx
 	pushfl
 	popl	%eax
@@ -1108,6 +1220,11 @@ padlock_capability:
 .align	16
 padlock_key_bswap:
 .L_padlock_key_bswap_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	240(%edx),%ecx
 	incl	%ecx
@@ -1126,6 +1243,11 @@ padlock_key_bswap:
 .align	16
 padlock_verify_context:
 .L_padlock_verify_context_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	leal	.Lpadlock_saved_context-.L004verify_pic_point,%eax
 	pushfl
@@ -1137,6 +1259,11 @@ padlock_verify_context:
 .type	_padlock_verify_ctx,@function
 .align	16
 _padlock_verify_ctx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%eax
 	btl	$30,4(%esp)
 	jnc	.L005verified
@@ -1153,6 +1280,11 @@ _padlock_verify_ctx:
 .align	16
 padlock_reload_key:
 .L_padlock_reload_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushfl
 	popfl
 	ret
@@ -1162,6 +1294,11 @@ padlock_reload_key:
 .align	16
 padlock_aes_block:
 .L_padlock_aes_block_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	pushl	%ebx
@@ -1182,6 +1319,11 @@ padlock_aes_block:
 .align	16
 padlock_ecb_encrypt:
 .L_padlock_ecb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1361,6 +1503,11 @@ padlock_ecb_encrypt:
 .align	16
 padlock_cbc_encrypt:
 .L_padlock_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1544,6 +1691,11 @@ padlock_cbc_encrypt:
 .align	16
 padlock_cfb_encrypt:
 .L_padlock_cfb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1666,6 +1818,11 @@ padlock_cfb_encrypt:
 .align	16
 padlock_ofb_encrypt:
 .L_padlock_ofb_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1788,6 +1945,11 @@ padlock_ofb_encrypt:
 .align	16
 padlock_ctr32_encrypt:
 .L_padlock_ctr32_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1895,6 +2057,11 @@ padlock_ctr32_encrypt:
 .align	16
 padlock_xstore:
 .L_padlock_xstore_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	movl	8(%esp),%edi
 	movl	12(%esp),%edx
@@ -1905,6 +2072,11 @@ padlock_xstore:
 .type	_win32_segv_handler,@function
 .align	16
 _win32_segv_handler:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$1,%eax
 	movl	4(%esp),%edx
 	movl	12(%esp),%ecx
@@ -1920,6 +2092,11 @@ _win32_segv_handler:
 .align	16
 padlock_sha1_oneshot:
 .L_padlock_sha1_oneshot_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	xorl	%eax,%eax
@@ -1951,6 +2128,11 @@ padlock_sha1_oneshot:
 .align	16
 padlock_sha1_blocks:
 .L_padlock_sha1_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -1981,6 +2163,11 @@ padlock_sha1_blocks:
 .align	16
 padlock_sha256_oneshot:
 .L_padlock_sha256_oneshot_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	xorl	%eax,%eax
@@ -2012,6 +2199,11 @@ padlock_sha256_oneshot:
 .align	16
 padlock_sha256_blocks:
 .L_padlock_sha256_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -2042,6 +2234,11 @@ padlock_sha256_blocks:
 .align	16
 padlock_sha512_blocks:
 .L_padlock_sha512_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%esi
 	movl	12(%esp),%edi
@@ -2083,4 +2280,21 @@ padlock_sha512_blocks:
 .align	4
 .Lpadlock_saved_context:
 .long	0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/ecp_nistz256-x86.S b/sys/crypto/openssl/i386/ecp_nistz256-x86.S
index b72eda775c9db..d304619280c7a 100644
--- a/sys/crypto/openssl/i386/ecp_nistz256-x86.S
+++ b/sys/crypto/openssl/i386/ecp_nistz256-x86.S
@@ -2389,6 +2389,11 @@ ecp_nistz256_precomputed:
 .align	16
 ecp_nistz256_mul_by_2:
 .L_ecp_nistz256_mul_by_2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2408,6 +2413,11 @@ ecp_nistz256_mul_by_2:
 .align	16
 ecp_nistz256_mul_by_3:
 .L_ecp_nistz256_mul_by_3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2433,6 +2443,11 @@ ecp_nistz256_mul_by_3:
 .align	16
 ecp_nistz256_div_by_2:
 .L_ecp_nistz256_div_by_2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2449,6 +2464,11 @@ ecp_nistz256_div_by_2:
 .type	_ecp_nistz256_div_by_2,@function
 .align	16
 _ecp_nistz256_div_by_2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ebp
 	xorl	%edx,%edx
 	movl	4(%esi),%ebx
@@ -2532,6 +2552,11 @@ _ecp_nistz256_div_by_2:
 .align	16
 ecp_nistz256_add:
 .L_ecp_nistz256_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2549,6 +2574,11 @@ ecp_nistz256_add:
 .type	_ecp_nistz256_add,@function
 .align	16
 _ecp_nistz256_add:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
 	movl	8(%esi),%ecx
@@ -2626,6 +2656,11 @@ _ecp_nistz256_add:
 .align	16
 ecp_nistz256_sub:
 .L_ecp_nistz256_sub_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2643,6 +2678,11 @@ ecp_nistz256_sub:
 .type	_ecp_nistz256_sub,@function
 .align	16
 _ecp_nistz256_sub:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
 	movl	8(%esi),%ecx
@@ -2701,6 +2741,11 @@ _ecp_nistz256_sub:
 .align	16
 ecp_nistz256_neg:
 .L_ecp_nistz256_neg_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2729,6 +2774,11 @@ ecp_nistz256_neg:
 .type	_picup_eax,@function
 .align	16
 _picup_eax:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esp),%eax
 	ret
 .size	_picup_eax,.-_picup_eax
@@ -2737,6 +2787,11 @@ _picup_eax:
 .align	16
 ecp_nistz256_to_mont:
 .L_ecp_nistz256_to_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2760,6 +2815,11 @@ ecp_nistz256_to_mont:
 .align	16
 ecp_nistz256_from_mont:
 .L_ecp_nistz256_from_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2783,6 +2843,11 @@ ecp_nistz256_from_mont:
 .align	16
 ecp_nistz256_mul_mont:
 .L_ecp_nistz256_mul_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2806,6 +2871,11 @@ ecp_nistz256_mul_mont:
 .align	16
 ecp_nistz256_sqr_mont:
 .L_ecp_nistz256_sqr_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2827,6 +2897,11 @@ ecp_nistz256_sqr_mont:
 .type	_ecp_nistz256_mul_mont,@function
 .align	16
 _ecp_nistz256_mul_mont:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	andl	$83886080,%eax
 	cmpl	$83886080,%eax
 	jne	.L004mul_mont_ialu
@@ -3724,6 +3799,11 @@ _ecp_nistz256_mul_mont:
 .align	16
 ecp_nistz256_scatter_w5:
 .L_ecp_nistz256_scatter_w5_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3757,6 +3837,11 @@ ecp_nistz256_scatter_w5:
 .align	16
 ecp_nistz256_gather_w5:
 .L_ecp_nistz256_gather_w5_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3851,6 +3936,11 @@ ecp_nistz256_gather_w5:
 .align	16
 ecp_nistz256_scatter_w7:
 .L_ecp_nistz256_scatter_w7_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3882,6 +3972,11 @@ ecp_nistz256_scatter_w7:
 .align	16
 ecp_nistz256_gather_w7:
 .L_ecp_nistz256_gather_w7_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4096,6 +4191,11 @@ ecp_nistz256_gather_w7:
 .align	16
 ecp_nistz256_point_double:
 .L_ecp_nistz256_point_double_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4222,6 +4322,11 @@ ecp_nistz256_point_double:
 .align	16
 ecp_nistz256_point_add:
 .L_ecp_nistz256_point_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -4735,6 +4840,11 @@ ecp_nistz256_point_add:
 .align	16
 ecp_nistz256_point_add_affine:
 .L_ecp_nistz256_point_add_affine_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5163,6 +5273,23 @@ ecp_nistz256_point_add_affine:
 	ret
 .size	ecp_nistz256_point_add_affine,.-.L_ecp_nistz256_point_add_affine_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	ecp_nistz256_precomputed
@@ -7553,6 +7680,11 @@ ecp_nistz256_precomputed:
 .align	16
 ecp_nistz256_mul_by_2:
 .L_ecp_nistz256_mul_by_2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7572,6 +7704,11 @@ ecp_nistz256_mul_by_2:
 .align	16
 ecp_nistz256_mul_by_3:
 .L_ecp_nistz256_mul_by_3_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7597,6 +7734,11 @@ ecp_nistz256_mul_by_3:
 .align	16
 ecp_nistz256_div_by_2:
 .L_ecp_nistz256_div_by_2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7613,6 +7755,11 @@ ecp_nistz256_div_by_2:
 .type	_ecp_nistz256_div_by_2,@function
 .align	16
 _ecp_nistz256_div_by_2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ebp
 	xorl	%edx,%edx
 	movl	4(%esi),%ebx
@@ -7696,6 +7843,11 @@ _ecp_nistz256_div_by_2:
 .align	16
 ecp_nistz256_add:
 .L_ecp_nistz256_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7713,6 +7865,11 @@ ecp_nistz256_add:
 .type	_ecp_nistz256_add,@function
 .align	16
 _ecp_nistz256_add:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
 	movl	8(%esi),%ecx
@@ -7790,6 +7947,11 @@ _ecp_nistz256_add:
 .align	16
 ecp_nistz256_sub:
 .L_ecp_nistz256_sub_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7807,6 +7969,11 @@ ecp_nistz256_sub:
 .type	_ecp_nistz256_sub,@function
 .align	16
 _ecp_nistz256_sub:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
 	movl	8(%esi),%ecx
@@ -7865,6 +8032,11 @@ _ecp_nistz256_sub:
 .align	16
 ecp_nistz256_neg:
 .L_ecp_nistz256_neg_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7893,6 +8065,11 @@ ecp_nistz256_neg:
 .type	_picup_eax,@function
 .align	16
 _picup_eax:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esp),%eax
 	ret
 .size	_picup_eax,.-_picup_eax
@@ -7901,6 +8078,11 @@ _picup_eax:
 .align	16
 ecp_nistz256_to_mont:
 .L_ecp_nistz256_to_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7924,6 +8106,11 @@ ecp_nistz256_to_mont:
 .align	16
 ecp_nistz256_from_mont:
 .L_ecp_nistz256_from_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7947,6 +8134,11 @@ ecp_nistz256_from_mont:
 .align	16
 ecp_nistz256_mul_mont:
 .L_ecp_nistz256_mul_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7970,6 +8162,11 @@ ecp_nistz256_mul_mont:
 .align	16
 ecp_nistz256_sqr_mont:
 .L_ecp_nistz256_sqr_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7991,6 +8188,11 @@ ecp_nistz256_sqr_mont:
 .type	_ecp_nistz256_mul_mont,@function
 .align	16
 _ecp_nistz256_mul_mont:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	andl	$83886080,%eax
 	cmpl	$83886080,%eax
 	jne	.L004mul_mont_ialu
@@ -8888,6 +9090,11 @@ _ecp_nistz256_mul_mont:
 .align	16
 ecp_nistz256_scatter_w5:
 .L_ecp_nistz256_scatter_w5_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -8921,6 +9128,11 @@ ecp_nistz256_scatter_w5:
 .align	16
 ecp_nistz256_gather_w5:
 .L_ecp_nistz256_gather_w5_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -9015,6 +9227,11 @@ ecp_nistz256_gather_w5:
 .align	16
 ecp_nistz256_scatter_w7:
 .L_ecp_nistz256_scatter_w7_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -9046,6 +9263,11 @@ ecp_nistz256_scatter_w7:
 .align	16
 ecp_nistz256_gather_w7:
 .L_ecp_nistz256_gather_w7_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -9260,6 +9482,11 @@ ecp_nistz256_gather_w7:
 .align	16
 ecp_nistz256_point_double:
 .L_ecp_nistz256_point_double_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -9386,6 +9613,11 @@ ecp_nistz256_point_double:
 .align	16
 ecp_nistz256_point_add:
 .L_ecp_nistz256_point_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -9899,6 +10131,11 @@ ecp_nistz256_point_add:
 .align	16
 ecp_nistz256_point_add_affine:
 .L_ecp_nistz256_point_add_affine_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -10327,4 +10564,21 @@ ecp_nistz256_point_add_affine:
 	ret
 .size	ecp_nistz256_point_add_affine,.-.L_ecp_nistz256_point_add_affine_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/ghash-x86.S b/sys/crypto/openssl/i386/ghash-x86.S
index 2d031cdd47e58..e855db4c1b556 100644
--- a/sys/crypto/openssl/i386/ghash-x86.S
+++ b/sys/crypto/openssl/i386/ghash-x86.S
@@ -6,6 +6,11 @@
 .align	16
 gcm_gmult_4bit_x86:
 .L_gcm_gmult_4bit_x86_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -100,6 +105,11 @@ gcm_gmult_4bit_x86:
 .align	16
 gcm_ghash_4bit_x86:
 .L_gcm_ghash_4bit_x86_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -209,6 +219,11 @@ gcm_ghash_4bit_x86:
 .align	16
 gcm_gmult_4bit_mmx:
 .L_gcm_gmult_4bit_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -308,6 +323,11 @@ gcm_gmult_4bit_mmx:
 .align	16
 gcm_ghash_4bit_mmx:
 .L_gcm_ghash_4bit_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -912,6 +932,11 @@ gcm_ghash_4bit_mmx:
 .align	16
 gcm_init_clmul:
 .L_gcm_init_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	call	.L010pic
@@ -981,6 +1006,11 @@ gcm_init_clmul:
 .align	16
 gcm_gmult_clmul:
 .L_gcm_gmult_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%edx
 	call	.L011pic
@@ -1034,6 +1064,11 @@ gcm_gmult_clmul:
 .align	16
 gcm_ghash_clmul:
 .L_gcm_ghash_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1264,6 +1299,23 @@ gcm_ghash_clmul:
 .byte	82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112
 .byte	112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62
 .byte	0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	gcm_gmult_4bit_x86
@@ -1271,6 +1323,11 @@ gcm_ghash_clmul:
 .align	16
 gcm_gmult_4bit_x86:
 .L_gcm_gmult_4bit_x86_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1365,6 +1422,11 @@ gcm_gmult_4bit_x86:
 .align	16
 gcm_ghash_4bit_x86:
 .L_gcm_ghash_4bit_x86_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1474,6 +1536,11 @@ gcm_ghash_4bit_x86:
 .align	16
 gcm_gmult_4bit_mmx:
 .L_gcm_gmult_4bit_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1573,6 +1640,11 @@ gcm_gmult_4bit_mmx:
 .align	16
 gcm_ghash_4bit_mmx:
 .L_gcm_ghash_4bit_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2177,6 +2249,11 @@ gcm_ghash_4bit_mmx:
 .align	16
 gcm_init_clmul:
 .L_gcm_init_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	call	.L010pic
@@ -2246,6 +2323,11 @@ gcm_init_clmul:
 .align	16
 gcm_gmult_clmul:
 .L_gcm_gmult_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%eax
 	movl	8(%esp),%edx
 	call	.L011pic
@@ -2299,6 +2381,11 @@ gcm_gmult_clmul:
 .align	16
 gcm_ghash_clmul:
 .L_gcm_ghash_clmul_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2529,4 +2616,21 @@ gcm_ghash_clmul:
 .byte	82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112
 .byte	112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62
 .byte	0
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/md5-586.S b/sys/crypto/openssl/i386/md5-586.S
index 1b1a07ec96d19..2d249dd7230fc 100644
--- a/sys/crypto/openssl/i386/md5-586.S
+++ b/sys/crypto/openssl/i386/md5-586.S
@@ -1,11 +1,16 @@
 /* Do not modify. This file is auto-generated from md5-586.pl. */
 #ifdef PIC
 .text
-.globl	md5_block_asm_data_order
-.type	md5_block_asm_data_order,@function
+.globl	ossl_md5_block_asm_data_order
+.type	ossl_md5_block_asm_data_order,@function
 .align	16
-md5_block_asm_data_order:
-.L_md5_block_asm_data_order_begin:
+ossl_md5_block_asm_data_order:
+.L_ossl_md5_block_asm_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	movl	12(%esp),%edi
@@ -677,14 +682,36 @@ md5_block_asm_data_order:
 	popl	%edi
 	popl	%esi
 	ret
-.size	md5_block_asm_data_order,.-.L_md5_block_asm_data_order_begin
+.size	ossl_md5_block_asm_data_order,.-.L_ossl_md5_block_asm_data_order_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
-.globl	md5_block_asm_data_order
-.type	md5_block_asm_data_order,@function
+.globl	ossl_md5_block_asm_data_order
+.type	ossl_md5_block_asm_data_order,@function
 .align	16
-md5_block_asm_data_order:
-.L_md5_block_asm_data_order_begin:
+ossl_md5_block_asm_data_order:
+.L_ossl_md5_block_asm_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	movl	12(%esp),%edi
@@ -1356,5 +1383,22 @@ md5_block_asm_data_order:
 	popl	%edi
 	popl	%esi
 	ret
-.size	md5_block_asm_data_order,.-.L_md5_block_asm_data_order_begin
+.size	ossl_md5_block_asm_data_order,.-.L_ossl_md5_block_asm_data_order_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/poly1305-x86.S b/sys/crypto/openssl/i386/poly1305-x86.S
index f8a678b3dc310..38af67d4ffc98 100644
--- a/sys/crypto/openssl/i386/poly1305-x86.S
+++ b/sys/crypto/openssl/i386/poly1305-x86.S
@@ -7,6 +7,11 @@
 .align	16
 poly1305_init:
 .L_poly1305_init_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -68,6 +73,11 @@ poly1305_init:
 .align	16
 poly1305_blocks:
 .L_poly1305_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -236,6 +246,11 @@ poly1305_blocks:
 .align	16
 poly1305_emit:
 .L_poly1305_emit_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -295,6 +310,11 @@ poly1305_emit:
 .type	_poly1305_init_sse2,@function
 .align	16
 _poly1305_init_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqu	24(%edi),%xmm4
 	leal	48(%edi),%edi
 	movl	%esp,%ebp
@@ -497,6 +517,11 @@ _poly1305_init_sse2:
 .type	_poly1305_blocks_sse2,@function
 .align	16
 _poly1305_blocks_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1258,6 +1283,11 @@ _poly1305_blocks_sse2:
 .type	_poly1305_emit_sse2,@function
 .align	16
 _poly1305_emit_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1351,6 +1381,11 @@ _poly1305_emit_sse2:
 .type	_poly1305_init_avx2,@function
 .align	16
 _poly1305_init_avx2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	vmovdqu	24(%edi),%xmm4
 	leal	48(%edi),%edi
 	movl	%esp,%ebp
@@ -1522,6 +1557,11 @@ _poly1305_init_avx2:
 .type	_poly1305_blocks_avx2,@function
 .align	16
 _poly1305_blocks_avx2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1910,6 +1950,23 @@ _poly1305_blocks_avx2:
 .byte	114,103,62,0
 .align	4
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .align	64
@@ -1918,6 +1975,11 @@ _poly1305_blocks_avx2:
 .align	16
 poly1305_init:
 .L_poly1305_init_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1979,6 +2041,11 @@ poly1305_init:
 .align	16
 poly1305_blocks:
 .L_poly1305_blocks_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2147,6 +2214,11 @@ poly1305_blocks:
 .align	16
 poly1305_emit:
 .L_poly1305_emit_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2206,6 +2278,11 @@ poly1305_emit:
 .type	_poly1305_init_sse2,@function
 .align	16
 _poly1305_init_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqu	24(%edi),%xmm4
 	leal	48(%edi),%edi
 	movl	%esp,%ebp
@@ -2408,6 +2485,11 @@ _poly1305_init_sse2:
 .type	_poly1305_blocks_sse2,@function
 .align	16
 _poly1305_blocks_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3169,6 +3251,11 @@ _poly1305_blocks_sse2:
 .type	_poly1305_emit_sse2,@function
 .align	16
 _poly1305_emit_sse2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3262,6 +3349,11 @@ _poly1305_emit_sse2:
 .type	_poly1305_init_avx2,@function
 .align	16
 _poly1305_init_avx2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	vmovdqu	24(%edi),%xmm4
 	leal	48(%edi),%edi
 	movl	%esp,%ebp
@@ -3433,6 +3525,11 @@ _poly1305_init_avx2:
 .type	_poly1305_blocks_avx2,@function
 .align	16
 _poly1305_blocks_avx2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3821,4 +3918,21 @@ _poly1305_blocks_avx2:
 .byte	114,103,62,0
 .align	4
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/rc4-586.S b/sys/crypto/openssl/i386/rc4-586.S
index 1d076f3d6b101..98f5cacec9a69 100644
--- a/sys/crypto/openssl/i386/rc4-586.S
+++ b/sys/crypto/openssl/i386/rc4-586.S
@@ -6,6 +6,11 @@
 .align	16
 RC4:
 .L_RC4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -272,6 +277,11 @@ RC4:
 .align	16
 RC4_set_key:
 .L_RC4_set_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -350,6 +360,11 @@ RC4_set_key:
 .align	16
 RC4_options:
 .L_RC4_options_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L018pic_point
 .L018pic_point:
 	popl	%eax
@@ -380,6 +395,23 @@ RC4_options:
 .align	64
 .size	RC4_options,.-.L_RC4_options_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	RC4
@@ -387,6 +419,11 @@ RC4_options:
 .align	16
 RC4:
 .L_RC4_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -650,6 +687,11 @@ RC4:
 .align	16
 RC4_set_key:
 .L_RC4_set_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -725,6 +767,11 @@ RC4_set_key:
 .align	16
 RC4_options:
 .L_RC4_options_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L016pic_point
 .L016pic_point:
 	popl	%eax
@@ -752,4 +799,21 @@ RC4_options:
 .align	64
 .size	RC4_options,.-.L_RC4_options_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/rc5-586.S b/sys/crypto/openssl/i386/rc5-586.S
index 5ab6bea14b3a8..72db8f41c442f 100644
--- a/sys/crypto/openssl/i386/rc5-586.S
+++ b/sys/crypto/openssl/i386/rc5-586.S
@@ -6,6 +6,11 @@
 .align	16
 RC5_32_encrypt:
 .L_RC5_32_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%esi
@@ -197,6 +202,11 @@ RC5_32_encrypt:
 .align	16
 RC5_32_decrypt:
 .L_RC5_32_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%esi
@@ -390,6 +400,11 @@ RC5_32_decrypt:
 .align	16
 RC5_32_cbc_encrypt:
 .L_RC5_32_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -449,21 +464,56 @@ RC5_32_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L010ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L011ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L012ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L013ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L014ejend
 .L015ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L016ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L017ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L014ejend:
 	xorl	%ecx,%eax
@@ -563,6 +613,23 @@ RC5_32_cbc_encrypt:
 .long	.L010ej7-.L008PIC_point
 .align	64
 .size	RC5_32_cbc_encrypt,.-.L_RC5_32_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	RC5_32_encrypt
@@ -570,6 +637,11 @@ RC5_32_cbc_encrypt:
 .align	16
 RC5_32_encrypt:
 .L_RC5_32_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%esi
@@ -761,6 +833,11 @@ RC5_32_encrypt:
 .align	16
 RC5_32_decrypt:
 .L_RC5_32_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%esi
@@ -954,6 +1031,11 @@ RC5_32_decrypt:
 .align	16
 RC5_32_cbc_encrypt:
 .L_RC5_32_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 
 	pushl	%ebp
 	pushl	%ebx
@@ -1013,21 +1095,56 @@ RC5_32_cbc_encrypt:
 	xorl	%edx,%edx
 	jmp	*%ebp
 .L010ej7:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	6(%esi),%dh
 	shll	$8,%edx
 .L011ej6:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	5(%esi),%dh
 .L012ej5:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	4(%esi),%dl
 .L013ej4:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	(%esi),%ecx
 	jmp	.L014ejend
 .L015ej3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	2(%esi),%ch
 	shll	$8,%ecx
 .L016ej2:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	1(%esi),%ch
 .L017ej1:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movb	(%esi),%cl
 .L014ejend:
 	xorl	%ecx,%eax
@@ -1127,4 +1244,21 @@ RC5_32_cbc_encrypt:
 .long	.L010ej7-.L008PIC_point
 .align	64
 .size	RC5_32_cbc_encrypt,.-.L_RC5_32_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/rmd-586.S b/sys/crypto/openssl/i386/rmd-586.S
index 5b6be50568b5e..20f64bb75cc50 100644
--- a/sys/crypto/openssl/i386/rmd-586.S
+++ b/sys/crypto/openssl/i386/rmd-586.S
@@ -6,6 +6,11 @@
 .align	16
 ripemd160_block_asm_data_order:
 .L_ripemd160_block_asm_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	pushl	%esi
@@ -1964,6 +1969,23 @@ ripemd160_block_asm_data_order:
 	popl	%esi
 	ret
 .size	ripemd160_block_asm_data_order,.-.L_ripemd160_block_asm_data_order_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	ripemd160_block_asm_data_order
@@ -1971,6 +1993,11 @@ ripemd160_block_asm_data_order:
 .align	16
 ripemd160_block_asm_data_order:
 .L_ripemd160_block_asm_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%eax
 	pushl	%esi
@@ -3929,4 +3956,21 @@ ripemd160_block_asm_data_order:
 	popl	%esi
 	ret
 .size	ripemd160_block_asm_data_order,.-.L_ripemd160_block_asm_data_order_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/sha1-586.S b/sys/crypto/openssl/i386/sha1-586.S
index 1b6b1a420f52b..19e598226a1de 100644
--- a/sys/crypto/openssl/i386/sha1-586.S
+++ b/sys/crypto/openssl/i386/sha1-586.S
@@ -6,6 +6,11 @@
 .align	16
 sha1_block_data_order:
 .L_sha1_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1400,6 +1405,11 @@ sha1_block_data_order:
 .type	_sha1_block_data_order_shaext,@function
 .align	16
 _sha1_block_data_order_shaext:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1570,6 +1580,11 @@ _sha1_block_data_order_shaext:
 .type	_sha1_block_data_order_ssse3,@function
 .align	16
 _sha1_block_data_order_ssse3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2789,6 +2804,11 @@ _sha1_block_data_order_ssse3:
 .type	_sha1_block_data_order_avx,@function
 .align	16
 _sha1_block_data_order_avx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -3969,6 +3989,23 @@ _sha1_block_data_order_avx:
 .byte	89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112
 .byte	114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	sha1_block_data_order
@@ -3976,6 +4013,11 @@ _sha1_block_data_order_avx:
 .align	16
 sha1_block_data_order:
 .L_sha1_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5370,6 +5412,11 @@ sha1_block_data_order:
 .type	_sha1_block_data_order_shaext,@function
 .align	16
 _sha1_block_data_order_shaext:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5540,6 +5587,11 @@ _sha1_block_data_order_shaext:
 .type	_sha1_block_data_order_ssse3,@function
 .align	16
 _sha1_block_data_order_ssse3:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -6759,6 +6811,11 @@ _sha1_block_data_order_ssse3:
 .type	_sha1_block_data_order_avx,@function
 .align	16
 _sha1_block_data_order_avx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -7939,4 +7996,21 @@ _sha1_block_data_order_avx:
 .byte	89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112
 .byte	114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/sha256-586.S b/sys/crypto/openssl/i386/sha256-586.S
index 454030c18cc67..f0ed09ebf84a8 100644
--- a/sys/crypto/openssl/i386/sha256-586.S
+++ b/sys/crypto/openssl/i386/sha256-586.S
@@ -6,6 +6,11 @@
 .align	16
 sha256_block_data_order:
 .L_sha256_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -6782,6 +6787,23 @@ sha256_block_data_order:
 	ret
 .size	sha256_block_data_order,.-.L_sha256_block_data_order_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	sha256_block_data_order
@@ -6789,6 +6811,11 @@ sha256_block_data_order:
 .align	16
 sha256_block_data_order:
 .L_sha256_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -13565,4 +13592,21 @@ sha256_block_data_order:
 	ret
 .size	sha256_block_data_order,.-.L_sha256_block_data_order_begin
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/sha512-586.S b/sys/crypto/openssl/i386/sha512-586.S
index c5f6b718d56b7..a7385b55e0acf 100644
--- a/sys/crypto/openssl/i386/sha512-586.S
+++ b/sys/crypto/openssl/i386/sha512-586.S
@@ -6,6 +6,11 @@
 .align	16
 sha512_block_data_order:
 .L_sha512_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2828,6 +2833,23 @@ sha512_block_data_order:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	sha512_block_data_order
@@ -2835,6 +2857,11 @@ sha512_block_data_order:
 .align	16
 sha512_block_data_order:
 .L_sha512_block_data_order_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -5657,4 +5684,21 @@ sha512_block_data_order:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/vpaes-x86.S b/sys/crypto/openssl/i386/vpaes-x86.S
index 1c9cef1456a07..0d479de7c30de 100644
--- a/sys/crypto/openssl/i386/vpaes-x86.S
+++ b/sys/crypto/openssl/i386/vpaes-x86.S
@@ -61,6 +61,11 @@
 .type	_vpaes_preheat,@function
 .align	16
 _vpaes_preheat:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%ebp
 	movdqa	-48(%ebp),%xmm7
 	movdqa	-16(%ebp),%xmm6
@@ -69,6 +74,11 @@ _vpaes_preheat:
 .type	_vpaes_encrypt_core,@function
 .align	16
 _vpaes_encrypt_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$16,%ecx
 	movl	240(%edx),%eax
 	movdqa	%xmm6,%xmm1
@@ -146,6 +156,11 @@ _vpaes_encrypt_core:
 .type	_vpaes_decrypt_core,@function
 .align	16
 _vpaes_decrypt_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	608(%ebp),%ebx
 	movl	240(%edx),%eax
 	movdqa	%xmm6,%xmm1
@@ -234,6 +249,11 @@ _vpaes_decrypt_core:
 .type	_vpaes_schedule_core,@function
 .align	16
 _vpaes_schedule_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%ebp
 	movdqu	(%esi),%xmm0
 	movdqa	320(%ebp),%xmm2
@@ -328,6 +348,11 @@ _vpaes_schedule_core:
 .type	_vpaes_schedule_192_smear,@function
 .align	16
 _vpaes_schedule_192_smear:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pshufd	$128,%xmm6,%xmm1
 	pshufd	$254,%xmm7,%xmm0
 	pxor	%xmm1,%xmm6
@@ -340,6 +365,11 @@ _vpaes_schedule_192_smear:
 .type	_vpaes_schedule_round,@function
 .align	16
 _vpaes_schedule_round:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	8(%esp),%xmm2
 	pxor	%xmm1,%xmm1
 .byte	102,15,58,15,202,15
@@ -389,6 +419,11 @@ _vpaes_schedule_round:
 .type	_vpaes_schedule_transform,@function
 .align	16
 _vpaes_schedule_transform:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	-16(%ebp),%xmm2
 	movdqa	%xmm2,%xmm1
 	pandn	%xmm0,%xmm1
@@ -404,6 +439,11 @@ _vpaes_schedule_transform:
 .type	_vpaes_schedule_mangle,@function
 .align	16
 _vpaes_schedule_mangle:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	%xmm0,%xmm4
 	movdqa	128(%ebp),%xmm5
 	testl	%edi,%edi
@@ -465,6 +505,11 @@ _vpaes_schedule_mangle:
 .align	16
 vpaes_set_encrypt_key:
 .L_vpaes_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -498,6 +543,11 @@ vpaes_set_encrypt_key:
 .align	16
 vpaes_set_decrypt_key:
 .L_vpaes_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -536,6 +586,11 @@ vpaes_set_decrypt_key:
 .align	16
 vpaes_encrypt:
 .L_vpaes_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -565,6 +620,11 @@ vpaes_encrypt:
 .align	16
 vpaes_decrypt:
 .L_vpaes_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -594,6 +654,11 @@ vpaes_decrypt:
 .align	16
 vpaes_cbc_encrypt:
 .L_vpaes_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -660,6 +725,23 @@ vpaes_cbc_encrypt:
 	popl	%ebp
 	ret
 .size	vpaes_cbc_encrypt,.-.L_vpaes_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .align	64
@@ -722,6 +804,11 @@ vpaes_cbc_encrypt:
 .type	_vpaes_preheat,@function
 .align	16
 _vpaes_preheat:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%ebp
 	movdqa	-48(%ebp),%xmm7
 	movdqa	-16(%ebp),%xmm6
@@ -730,6 +817,11 @@ _vpaes_preheat:
 .type	_vpaes_encrypt_core,@function
 .align	16
 _vpaes_encrypt_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	$16,%ecx
 	movl	240(%edx),%eax
 	movdqa	%xmm6,%xmm1
@@ -807,6 +899,11 @@ _vpaes_encrypt_core:
 .type	_vpaes_decrypt_core,@function
 .align	16
 _vpaes_decrypt_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	608(%ebp),%ebx
 	movl	240(%edx),%eax
 	movdqa	%xmm6,%xmm1
@@ -895,6 +992,11 @@ _vpaes_decrypt_core:
 .type	_vpaes_schedule_core,@function
 .align	16
 _vpaes_schedule_core:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	addl	(%esp),%ebp
 	movdqu	(%esi),%xmm0
 	movdqa	320(%ebp),%xmm2
@@ -989,6 +1091,11 @@ _vpaes_schedule_core:
 .type	_vpaes_schedule_192_smear,@function
 .align	16
 _vpaes_schedule_192_smear:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pshufd	$128,%xmm6,%xmm1
 	pshufd	$254,%xmm7,%xmm0
 	pxor	%xmm1,%xmm6
@@ -1001,6 +1108,11 @@ _vpaes_schedule_192_smear:
 .type	_vpaes_schedule_round,@function
 .align	16
 _vpaes_schedule_round:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	8(%esp),%xmm2
 	pxor	%xmm1,%xmm1
 .byte	102,15,58,15,202,15
@@ -1050,6 +1162,11 @@ _vpaes_schedule_round:
 .type	_vpaes_schedule_transform,@function
 .align	16
 _vpaes_schedule_transform:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	-16(%ebp),%xmm2
 	movdqa	%xmm2,%xmm1
 	pandn	%xmm0,%xmm1
@@ -1065,6 +1182,11 @@ _vpaes_schedule_transform:
 .type	_vpaes_schedule_mangle,@function
 .align	16
 _vpaes_schedule_mangle:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movdqa	%xmm0,%xmm4
 	movdqa	128(%ebp),%xmm5
 	testl	%edi,%edi
@@ -1126,6 +1248,11 @@ _vpaes_schedule_mangle:
 .align	16
 vpaes_set_encrypt_key:
 .L_vpaes_set_encrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1159,6 +1286,11 @@ vpaes_set_encrypt_key:
 .align	16
 vpaes_set_decrypt_key:
 .L_vpaes_set_decrypt_key_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1197,6 +1329,11 @@ vpaes_set_decrypt_key:
 .align	16
 vpaes_encrypt:
 .L_vpaes_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1226,6 +1363,11 @@ vpaes_encrypt:
 .align	16
 vpaes_decrypt:
 .L_vpaes_decrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1255,6 +1397,11 @@ vpaes_decrypt:
 .align	16
 vpaes_cbc_encrypt:
 .L_vpaes_cbc_encrypt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1321,4 +1468,21 @@ vpaes_cbc_encrypt:
 	popl	%ebp
 	ret
 .size	vpaes_cbc_encrypt,.-.L_vpaes_cbc_encrypt_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/wp-mmx.S b/sys/crypto/openssl/i386/wp-mmx.S
index fc5df6274a9ca..407554c26515a 100644
--- a/sys/crypto/openssl/i386/wp-mmx.S
+++ b/sys/crypto/openssl/i386/wp-mmx.S
@@ -6,6 +6,11 @@
 .align	16
 whirlpool_block_mmx:
 .L_whirlpool_block_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -1106,6 +1111,23 @@ whirlpool_block_mmx:
 .byte	251,238,124,102,221,23,71,158
 .byte	202,45,191,7,173,90,131,51
 .size	whirlpool_block_mmx,.-.L_whirlpool_block_mmx_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	whirlpool_block_mmx
@@ -1113,6 +1135,11 @@ whirlpool_block_mmx:
 .align	16
 whirlpool_block_mmx:
 .L_whirlpool_block_mmx_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -2213,4 +2240,21 @@ whirlpool_block_mmx:
 .byte	251,238,124,102,221,23,71,158
 .byte	202,45,191,7,173,90,131,51
 .size	whirlpool_block_mmx,.-.L_whirlpool_block_mmx_begin
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/x86-gf2m.S b/sys/crypto/openssl/i386/x86-gf2m.S
index 9c0e251781049..093988d3d5e2d 100644
--- a/sys/crypto/openssl/i386/x86-gf2m.S
+++ b/sys/crypto/openssl/i386/x86-gf2m.S
@@ -4,6 +4,11 @@
 .type	_mul_1x1_mmx,@function
 .align	16
 _mul_1x1_mmx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	subl	$36,%esp
 	movl	%eax,%ecx
 	leal	(%eax,%eax,1),%edx
@@ -107,6 +112,11 @@ _mul_1x1_mmx:
 .type	_mul_1x1_ialu,@function
 .align	16
 _mul_1x1_ialu:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	subl	$36,%esp
 	movl	%eax,%ecx
 	leal	(%eax,%eax,1),%edx
@@ -241,6 +251,11 @@ _mul_1x1_ialu:
 .align	16
 bn_GF2m_mul_2x2:
 .L_bn_GF2m_mul_2x2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L000PIC_me_up
 .L000PIC_me_up:
 	popl	%edx
@@ -345,11 +360,33 @@ bn_GF2m_mul_2x2:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .type	_mul_1x1_mmx,@function
 .align	16
 _mul_1x1_mmx:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	subl	$36,%esp
 	movl	%eax,%ecx
 	leal	(%eax,%eax,1),%edx
@@ -453,6 +490,11 @@ _mul_1x1_mmx:
 .type	_mul_1x1_ialu,@function
 .align	16
 _mul_1x1_ialu:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	subl	$36,%esp
 	movl	%eax,%ecx
 	leal	(%eax,%eax,1),%edx
@@ -587,6 +629,11 @@ _mul_1x1_ialu:
 .align	16
 bn_GF2m_mul_2x2:
 .L_bn_GF2m_mul_2x2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	OPENSSL_ia32cap_P,%edx
 	movl	(%edx),%eax
 	movl	4(%edx),%edx
@@ -688,4 +735,21 @@ bn_GF2m_mul_2x2:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/x86-mont.S b/sys/crypto/openssl/i386/x86-mont.S
index 049c9e626e06f..4a5309b3e0db9 100644
--- a/sys/crypto/openssl/i386/x86-mont.S
+++ b/sys/crypto/openssl/i386/x86-mont.S
@@ -6,6 +6,11 @@
 .align	16
 bn_mul_mont:
 .L_bn_mul_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -475,6 +480,23 @@ bn_mul_mont:
 .byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
 .byte	111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	bn_mul_mont
@@ -482,6 +504,11 @@ bn_mul_mont:
 .align	16
 bn_mul_mont:
 .L_bn_mul_mont_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -948,4 +975,21 @@ bn_mul_mont:
 .byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
 .byte	111,114,103,62,0
 .comm	OPENSSL_ia32cap_P,16,4
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/i386/x86cpuid.S b/sys/crypto/openssl/i386/x86cpuid.S
index d232734b64baa..2a7f7188b65fc 100644
--- a/sys/crypto/openssl/i386/x86cpuid.S
+++ b/sys/crypto/openssl/i386/x86cpuid.S
@@ -6,6 +6,11 @@
 .align	16
 OPENSSL_ia32_cpuid:
 .L_OPENSSL_ia32_cpuid_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -150,6 +155,11 @@ OPENSSL_ia32_cpuid:
 .align	16
 OPENSSL_rdtsc:
 .L_OPENSSL_rdtsc_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	%eax,%eax
 	xorl	%edx,%edx
 	call	.L009PIC_me_up
@@ -167,6 +177,11 @@ OPENSSL_rdtsc:
 .align	16
 OPENSSL_instrument_halt:
 .L_OPENSSL_instrument_halt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	call	.L011PIC_me_up
 .L011PIC_me_up:
 	popl	%ecx
@@ -199,6 +214,11 @@ OPENSSL_instrument_halt:
 .align	16
 OPENSSL_far_spin:
 .L_OPENSSL_far_spin_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushfl
 	popl	%eax
 	btl	$9,%eax
@@ -226,6 +246,11 @@ OPENSSL_far_spin:
 .align	16
 OPENSSL_wipe_cpu:
 .L_OPENSSL_wipe_cpu_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	%eax,%eax
 	xorl	%edx,%edx
 	call	.L015PIC_me_up
@@ -257,6 +282,11 @@ OPENSSL_wipe_cpu:
 .align	16
 OPENSSL_atomic_add:
 .L_OPENSSL_atomic_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%ecx
 	pushl	%ebx
@@ -276,6 +306,11 @@ OPENSSL_atomic_add:
 .align	16
 OPENSSL_cleanse:
 .L_OPENSSL_cleanse_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%ecx
 	xorl	%eax,%eax
@@ -313,6 +348,11 @@ OPENSSL_cleanse:
 .align	16
 CRYPTO_memcmp:
 .L_CRYPTO_memcmp_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	movl	12(%esp),%esi
@@ -342,6 +382,11 @@ CRYPTO_memcmp:
 .align	16
 OPENSSL_instrument_bus:
 .L_OPENSSL_instrument_bus_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -390,6 +435,11 @@ OPENSSL_instrument_bus:
 .align	16
 OPENSSL_instrument_bus2:
 .L_OPENSSL_instrument_bus2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -451,6 +501,11 @@ OPENSSL_instrument_bus2:
 .align	16
 OPENSSL_ia32_rdrand_bytes:
 .L_OPENSSL_ia32_rdrand_bytes_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%ebx
 	xorl	%eax,%eax
@@ -494,6 +549,11 @@ OPENSSL_ia32_rdrand_bytes:
 .align	16
 OPENSSL_ia32_rdseed_bytes:
 .L_OPENSSL_ia32_rdseed_bytes_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%ebx
 	xorl	%eax,%eax
@@ -537,6 +597,23 @@ OPENSSL_ia32_rdseed_bytes:
 .comm	OPENSSL_ia32cap_P,16,4
 .section	.init
 	call	OPENSSL_cpuid_setup
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #else
 .text
 .globl	OPENSSL_ia32_cpuid
@@ -544,6 +621,11 @@ OPENSSL_ia32_rdseed_bytes:
 .align	16
 OPENSSL_ia32_cpuid:
 .L_OPENSSL_ia32_cpuid_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -688,6 +770,11 @@ OPENSSL_ia32_cpuid:
 .align	16
 OPENSSL_rdtsc:
 .L_OPENSSL_rdtsc_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	%eax,%eax
 	xorl	%edx,%edx
 	leal	OPENSSL_ia32cap_P,%ecx
@@ -702,6 +789,11 @@ OPENSSL_rdtsc:
 .align	16
 OPENSSL_instrument_halt:
 .L_OPENSSL_instrument_halt_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	leal	OPENSSL_ia32cap_P,%ecx
 	btl	$4,(%ecx)
 	jnc	.L010nohalt
@@ -731,6 +823,11 @@ OPENSSL_instrument_halt:
 .align	16
 OPENSSL_far_spin:
 .L_OPENSSL_far_spin_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushfl
 	popl	%eax
 	btl	$9,%eax
@@ -758,6 +855,11 @@ OPENSSL_far_spin:
 .align	16
 OPENSSL_wipe_cpu:
 .L_OPENSSL_wipe_cpu_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	xorl	%eax,%eax
 	xorl	%edx,%edx
 	leal	OPENSSL_ia32cap_P,%ecx
@@ -786,6 +888,11 @@ OPENSSL_wipe_cpu:
 .align	16
 OPENSSL_atomic_add:
 .L_OPENSSL_atomic_add_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%ecx
 	pushl	%ebx
@@ -805,6 +912,11 @@ OPENSSL_atomic_add:
 .align	16
 OPENSSL_cleanse:
 .L_OPENSSL_cleanse_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	movl	4(%esp),%edx
 	movl	8(%esp),%ecx
 	xorl	%eax,%eax
@@ -842,6 +954,11 @@ OPENSSL_cleanse:
 .align	16
 CRYPTO_memcmp:
 .L_CRYPTO_memcmp_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%esi
 	pushl	%edi
 	movl	12(%esp),%esi
@@ -871,6 +988,11 @@ CRYPTO_memcmp:
 .align	16
 OPENSSL_instrument_bus:
 .L_OPENSSL_instrument_bus_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -916,6 +1038,11 @@ OPENSSL_instrument_bus:
 .align	16
 OPENSSL_instrument_bus2:
 .L_OPENSSL_instrument_bus2_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%ebp
 	pushl	%ebx
 	pushl	%esi
@@ -974,6 +1101,11 @@ OPENSSL_instrument_bus2:
 .align	16
 OPENSSL_ia32_rdrand_bytes:
 .L_OPENSSL_ia32_rdrand_bytes_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%ebx
 	xorl	%eax,%eax
@@ -1017,6 +1149,11 @@ OPENSSL_ia32_rdrand_bytes:
 .align	16
 OPENSSL_ia32_rdseed_bytes:
 .L_OPENSSL_ia32_rdseed_bytes_begin:
+	#ifdef __CET__
+
+.byte	243,15,30,251
+	#endif
+
 	pushl	%edi
 	pushl	%ebx
 	xorl	%eax,%eax
@@ -1060,4 +1197,21 @@ OPENSSL_ia32_rdseed_bytes:
 .comm	OPENSSL_ia32cap_P,16,4
 .section	.init
 	call	OPENSSL_cpuid_setup
+
+	.section ".note.gnu.property", "a"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
 #endif
diff --git a/sys/crypto/openssl/powerpc/bn-ppc.S b/sys/crypto/openssl/powerpc/bn-ppc.S
new file mode 100644
index 0000000000000..166e2de1dddb2
--- /dev/null
+++ b/sys/crypto/openssl/powerpc/bn-ppc.S
@@ -0,0 +1,1855 @@
+/* Do not modify. This file is auto-generated from ppc.pl. */
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+.globl	bn_sqr_comba4
+.type	bn_sqr_comba4,@function
+.globl	bn_sqr_comba8
+.type	bn_sqr_comba8,@function
+.globl	bn_mul_comba4
+.type	bn_mul_comba4,@function
+.globl	bn_mul_comba8
+.type	bn_mul_comba8,@function
+.globl	bn_sub_words
+.type	bn_sub_words,@function
+.globl	bn_add_words
+.type	bn_add_words,@function
+.globl	bn_div_words
+.type	bn_div_words,@function
+.globl	bn_sqr_words
+.type	bn_sqr_words,@function
+.globl	bn_mul_words
+.type	bn_mul_words,@function
+.globl	bn_mul_add_words
+.type	bn_mul_add_words,@function
+
+
+
+.machine	"any"
+.text
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba4:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	lwz	5,0(4)
+	mullw	9,5,5
+	mulhwu	10,5,5
+
+
+
+
+	stw	9,0(3)
+
+	lwz	6,4(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,0
+
+
+	addc	10,7,10
+	addze	11,8
+	addze	9,9
+
+	stw	10,4(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	lwz	6,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	stw	11,8(3)
+
+	lwz	6,12(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,4(4)
+	lwz	6,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	stw	9,12(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	lwz	6,12(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	stw	10,16(3)
+
+	lwz	5,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	stw	11,20(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+
+	stw	9,24(3)
+	stw	10,28(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba4,.-bn_sqr_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba8:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	lwz	5,0(4)
+	mullw	9,5,5
+	mulhwu	10,5,5
+	stw	9,0(3)
+
+	lwz	6,4(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,0
+	addze	9,0
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	stw	10,4(3)
+
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	lwz	6,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	stw	11,8(3)
+
+	lwz	6,12(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,4(4)
+	lwz	6,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	stw	9,12(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	lwz	6,12(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	lwz	5,0(4)
+	lwz	6,16(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	stw	10,16(3)
+
+	lwz	6,20(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	lwz	5,4(4)
+	lwz	6,16(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	lwz	5,8(4)
+	lwz	6,12(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	stw	11,20(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	lwz	6,16(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,4(4)
+	lwz	6,20(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,0(4)
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	stw	9,24(3)
+
+	lwz	6,28(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	lwz	5,4(4)
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	lwz	5,8(4)
+	lwz	6,20(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	lwz	5,12(4)
+	lwz	6,16(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	stw	10,28(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	lwz	6,20(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	lwz	5,8(4)
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	lwz	5,4(4)
+	lwz	6,28(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	stw	11,32(3)
+
+	lwz	5,8(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,12(4)
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	lwz	5,16(4)
+	lwz	6,20(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	stw	9,36(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	lwz	5,12(4)
+	lwz	6,28(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	stw	10,40(3)
+
+	lwz	5,16(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	lwz	5,20(4)
+	lwz	6,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	stw	11,44(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	lwz	6,28(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	stw	9,48(3)
+
+
+	lwz	5,24(4)
+	mullw	7,5,6
+	mulhwu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	stw	10,52(3)
+
+	mullw	7,6,6
+	mulhwu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	stw	11,56(3)
+	stw	9, 60(3)
+
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba8,.-bn_sqr_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba4:
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+	lwz	6,0(4)
+	lwz	7,0(5)
+	mullw	10,6,7
+	mulhwu	11,6,7
+	stw	10,0(3)
+
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,8,11
+	adde	12,9,0
+	addze	10,0
+
+	lwz	6, 4(4)
+	lwz	7, 0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	stw	11,4(3)
+
+	lwz	6,8(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	lwz	6,4(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+
+	lwz	6,0(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	stw	12,8(3)
+
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,0
+
+	lwz	6,4(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	lwz	6,8(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	lwz	6,12(4)
+	lwz	7,0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+	stw	10,12(3)
+
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,0
+
+	lwz	6,8(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+
+	lwz	6,4(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	stw	11,16(3)
+
+	lwz	6,8(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	lwz	6,12(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	stw	12,20(3)
+
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+
+	stw	10,24(3)
+	stw	11,28(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba4,.-bn_mul_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba8:
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+	lwz	6,0(4)
+	lwz	7,0(5)
+	mullw	10,6,7
+	mulhwu	11,6,7
+	stw	10,0(3)
+
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	addze	12,9
+	addze	10,0
+
+	lwz	6,4(4)
+	lwz	7,0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	stw	11,4(3)
+
+	lwz	6,8(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	lwz	6,4(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,0(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	stw	12,8(3)
+
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	lwz	6,4(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+
+	lwz	6,8(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,12(4)
+	lwz	7,0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	stw	10,12(3)
+
+	lwz	6,16(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	lwz	6,12(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,8(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,4(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,0(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	stw	11,16(3)
+
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	lwz	6,4(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,8(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,12(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,16(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,20(4)
+	lwz	7,0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	stw	12,20(3)
+
+	lwz	6,24(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	lwz	6,20(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,16(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,12(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,8(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,4(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,0(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	stw	10,24(3)
+
+	lwz	7,28(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	lwz	6,4(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,8(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,12(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,16(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,20(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,24(4)
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,28(4)
+	lwz	7,0(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	stw	11,28(3)
+
+	lwz	7,4(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	lwz	6,24(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,20(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,16(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,12(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,8(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,4(4)
+	lwz	7,28(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	stw	12,32(3)
+
+	lwz	6,8(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	lwz	6,12(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,16(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,20(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,24(4)
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,28(4)
+	lwz	7,8(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	stw	10,36(3)
+
+	lwz	7,12(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	lwz	6,24(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,20(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,16(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	lwz	6,12(4)
+	lwz	7,28(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	stw	11,40(3)
+
+	lwz	6,16(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	lwz	6,20(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,24(4)
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	lwz	6,28(4)
+	lwz	7,16(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	stw	12,44(3)
+
+	lwz	7,20(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	lwz	6,24(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	lwz	6,20(4)
+	lwz	7,28(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	stw	10,48(3)
+
+	lwz	6,24(4)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	lwz	6,28(4)
+	lwz	7,24(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	stw	11,52(3)
+
+	lwz	7,28(5)
+	mullw	8,6,7
+	mulhwu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	stw	12,56(3)
+	stw	10,60(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba8,.-bn_mul_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_sub_words:
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	subfc.	7,0,6
+
+
+	beq	.Lppcasm_sub_adios
+	addi	4,4,-4
+	addi	3,3,-4
+	addi	5,5,-4
+	mtctr	6
+.Lppcasm_sub_mainloop:
+	lwzu	7,4(4)
+	lwzu	8,4(5)
+	subfe	6,8,7
+
+
+	stwu	6,4(3)
+	bdnz	.Lppcasm_sub_mainloop
+.Lppcasm_sub_adios:
+	subfze	3,0
+	andi.	3,3,1
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_sub_words,.-bn_sub_words
+
+
+
+
+
+
+
+
+.align	4
+bn_add_words:
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	addic.	6,6,0
+	beq	.Lppcasm_add_adios
+	addi	4,4,-4
+	addi	3,3,-4
+	addi	5,5,-4
+	mtctr	6
+.Lppcasm_add_mainloop:
+	lwzu	7,4(4)
+	lwzu	8,4(5)
+	adde	8,7,8
+	stwu	8,4(3)
+	bdnz	.Lppcasm_add_mainloop
+.Lppcasm_add_adios:
+	addze	3,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_add_words,.-bn_add_words
+
+
+
+
+
+
+
+
+.align	4
+bn_div_words:
+
+
+
+
+
+
+
+
+
+
+
+	cmplwi	0,5,0
+	bne	.Lppcasm_div1
+	li	3,-1
+	blr	
+.Lppcasm_div1:
+	xor	0,0,0
+	li	8,32
+	cntlzw.	7,5
+	beq	.Lppcasm_div2
+	subf	8,7,8
+	srw.	9,3,8
+	tw	16,9,0
+.Lppcasm_div2:
+	.long	0x7c032840
+	blt	.Lppcasm_div3
+	subf	3,5,3
+.Lppcasm_div3:
+	cmpi	0,0,7,0
+	beq	.Lppcasm_div4
+	slw	3,3,7
+	srw	8,4,8
+	slw	5,5,7
+	or	3,3,8
+	slw	4,4,7
+.Lppcasm_div4:
+	srwi	9,5,16
+
+
+	li	6,2
+	mtctr	6
+.Lppcasm_divouterloop:
+	srwi	8,3,16
+	srwi	11,4,16
+
+	.long	0x7c084840
+	bne	.Lppcasm_div5
+
+	li	8,-1
+	clrlwi	8,8,16
+	b	.Lppcasm_div6
+.Lppcasm_div5:
+	divwu	8,3,9
+.Lppcasm_div6:
+	mullw	12,9,8
+	clrlwi	10,5,16
+	mullw	6,8,10
+
+.Lppcasm_divinnerloop:
+	subf	10,12,3
+	srwi	7,10,16
+	addic.	7,7,0
+
+
+
+	slwi	7,10,16
+	or	7,7,11
+	.long	0x7c863840
+	bne	.Lppcasm_divinnerexit
+	ble	1,.Lppcasm_divinnerexit
+	addi	8,8,-1
+	subf	12,9,12
+	clrlwi	10,5,16
+	subf	6,10,6
+	b	.Lppcasm_divinnerloop
+.Lppcasm_divinnerexit:
+	srwi	10,6,16
+	slwi	11,6,16
+	.long	0x7c845840
+	add	12,12,10
+	bge	1,.Lppcasm_div7
+	addi	12,12,1
+.Lppcasm_div7:
+	subf	11,11,4
+	.long	0x7c836040
+	bge	1,.Lppcasm_div8
+	addi	8,8,-1
+	add	3,5,3
+.Lppcasm_div8:
+	subf	12,12,3
+	slwi	4,11,16
+
+
+
+	insrwi	11,12,16,16
+	rotlwi	3,11,16
+	bdz	.Lppcasm_div9
+	slwi	0,8,16
+	b	.Lppcasm_divouterloop
+.Lppcasm_div9:
+	or	3,8,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_div_words,.-bn_div_words
+
+
+
+
+
+
+
+.align	4
+bn_sqr_words:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	addic.	5,5,0
+	beq	.Lppcasm_sqr_adios
+	addi	4,4,-4
+	addi	3,3,-4
+	mtctr	5
+.Lppcasm_sqr_mainloop:
+
+	lwzu	6,4(4)
+	mullw	7,6,6
+	mulhwu	8,6,6
+	stwu	7,4(3)
+	stwu	8,4(3)
+	bdnz	.Lppcasm_sqr_mainloop
+.Lppcasm_sqr_adios:
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_sqr_words,.-bn_sqr_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_words:
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_mw_REM
+	mtctr	7
+.Lppcasm_mw_LOOP:
+
+	lwz	8,0(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	addc	9,9,12
+
+
+
+
+	stw	9,0(3)
+
+	lwz	8,4(4)
+	mullw	11,6,8
+	mulhwu	12,6,8
+	adde	11,11,10
+
+	stw	11,4(3)
+
+	lwz	8,8(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	adde	9,9,12
+
+	stw	9,8(3)
+
+	lwz	8,12(4)
+	mullw	11,6,8
+	mulhwu	12,6,8
+	adde	11,11,10
+	addze	12,12
+
+	stw	11,12(3)
+
+	addi	3,3,16
+	addi	4,4,16
+	bdnz	.Lppcasm_mw_LOOP
+
+.Lppcasm_mw_REM:
+	andi.	5,5,0x3
+	beq	.Lppcasm_mw_OVER
+
+	lwz	8,0(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	stw	9,0(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+
+	lwz	8,4(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	stw	9,4(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+	lwz	8,8(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	stw	9,8(3)
+	addi	12,10,0
+
+.Lppcasm_mw_OVER:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_words,.-bn_mul_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_add_words:
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_maw_leftover
+	mtctr	7
+.Lppcasm_maw_mainloop:
+
+	lwz	8,0(4)
+	lwz	11,0(3)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+
+
+
+
+
+	stw	9,0(3)
+
+
+	lwz	8,4(4)
+	lwz	9,4(3)
+	mullw	11,6,8
+	mulhwu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+
+	stw	11,4(3)
+
+
+	lwz	8,8(4)
+	mullw	9,6,8
+	lwz	11,8(3)
+	mulhwu	10,6,8
+	adde	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+	stw	9,8(3)
+
+
+	lwz	8,12(4)
+	mullw	11,6,8
+	lwz	9,12(3)
+	mulhwu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+	addze	12,12
+	stw	11,12(3)
+	addi	3,3,16
+	addi	4,4,16
+	bdnz	.Lppcasm_maw_mainloop
+
+.Lppcasm_maw_leftover:
+	andi.	5,5,0x3
+	beq	.Lppcasm_maw_adios
+	addi	3,3,-4
+	addi	4,4,-4
+
+	mtctr	5
+	lwzu	8,4(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	lwzu	11,4(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	stw	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	lwzu	8,4(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	lwzu	11,4(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	stw	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	lwzu	8,4(4)
+	mullw	9,6,8
+	mulhwu	10,6,8
+	lwzu	11,4(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	stw	9,0(3)
+
+.Lppcasm_maw_adios:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_add_words,.-bn_mul_add_words
+.align	4
diff --git a/sys/crypto/openssl/powerpc/poly1305-ppc.S b/sys/crypto/openssl/powerpc/poly1305-ppc.S
index 5b458a6ec4e52..d6fe346937246 100644
--- a/sys/crypto/openssl/powerpc/poly1305-ppc.S
+++ b/sys/crypto/openssl/powerpc/poly1305-ppc.S
@@ -11,6 +11,7 @@ poly1305_init_int:
 	stw	0,8(3)
 	stw	0,12(3)
 	stw	0,16(3)
+	stw	0,24(3)
 
 	.long	0x7c040040
 	beq-	.Lno_key
@@ -46,6 +47,7 @@ poly1305_init_int:
 .type	poly1305_blocks,@function
 .align	4
 poly1305_blocks:
+.Lpoly1305_blocks:
 	srwi.	5,5,4
 	beq-	.Labort
 
@@ -243,18 +245,110 @@ poly1305_blocks:
 .long	0
 .byte	0,12,4,1,0x80,18,4,0
 .size	poly1305_blocks,.-poly1305_blocks
-
 .globl	poly1305_emit
 .type	poly1305_emit,@function
-.align	4
+.align	5
 poly1305_emit:
-	stwu	1,-96(1)
-	mflr	0
-	stw	28,80(1)
-	stw	29,84(1)
-	stw	30,88(1)
-	stw	31,92(1)
-	stw	0,100(1)
+	lwz	0,24(3)
+	lwz	6,0(3)
+	lwz	7,4(3)
+	lwz	8,8(3)
+	lwz	9,12(3)
+	lwz	10,16(3)
+	cmplwi	0,0
+	beq	.Lemit_base2_32
+
+	slwi	11,7,26
+	srwi	7,7,6
+	slwi	12,8,20
+	srwi	8,8,12
+	addc	6,6,11
+	slwi	11,9,14
+	srwi	9,9,18
+	adde	7,7,12
+	slwi	12,10,8
+	srwi	10,10,24
+	adde	8,8,11
+	adde	9,9,12
+	addze	10,10
+
+.Lemit_base2_32:
+	addic	0,6,5
+	addze	0,7
+	addze	0,8
+	addze	0,9
+	addze	0,10
+
+	srwi	0,0,2
+	neg	0,0
+	andi.	0,0,5
+
+	addc	6,6,0
+	lwz	0,0(5)
+	addze	7,7
+	lwz	11,4(5)
+	addze	8,8
+	lwz	12,8(5)
+	addze	9,9
+	lwz	10,12(5)
+
+	addc	6,6,0
+	adde	7,7,11
+	adde	8,8,12
+	adde	9,9,10
+
+	addi	3,4,-1
+	addi	4,4,7
+
+	stbu	6,1(3)
+	srwi	6,6,8
+	stbu	8,1(4)
+	srwi	8,8,8
+
+	stbu	6,1(3)
+	srwi	6,6,8
+	stbu	8,1(4)
+	srwi	8,8,8
+
+	stbu	6,1(3)
+	srwi	6,6,8
+	stbu	8,1(4)
+	srwi	8,8,8
+
+	stbu	6,1(3)
+	stbu	8,1(4)
+
+	stbu	7,1(3)
+	srwi	7,7,8
+	stbu	9,1(4)
+	srwi	9,9,8
+
+	stbu	7,1(3)
+	srwi	7,7,8
+	stbu	9,1(4)
+	srwi	9,9,8
+
+	stbu	7,1(3)
+	srwi	7,7,8
+	stbu	9,1(4)
+	srwi	9,9,8
+
+	stbu	7,1(3)
+	stbu	9,1(4)
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.size	poly1305_emit,.-poly1305_emit
+.globl	poly1305_blocks_vsx
+.type	poly1305_blocks_vsx,@function
+.align	5
+poly1305_blocks_vsx:
+	lwz	7,24(3)
+	cmplwi	5,128
+	bge	__poly1305_blocks_vsx
+	cmplwi	7,0
+	beq	.Lpoly1305_blocks
 
 	lwz	7,0(3)
 	lwz	8,4(3)
@@ -262,51 +356,946 @@ poly1305_emit:
 	lwz	10,12(3)
 	lwz	11,16(3)
 
-	addic	28,7,5
-	addze	29,8
-	addze	30,9
-	addze	31,10
-	addze	0,11
+	slwi	0,8,26
+	srwi	8,8,6
+	slwi	12,9,20
+	srwi	9,9,12
+	addc	7,7,0
+	slwi	0,10,14
+	srwi	10,10,18
+	adde	8,8,12
+	slwi	12,11,8
+	srwi	11,11,24
+	adde	9,9,0
+	li	0,0
+	adde	10,10,12
+	addze	11,11
 
-	srwi	0,0,2
-	neg	0,0
+	stw	7,0(3)
+	stw	8,4(3)
+	stw	9,8(3)
+	stw	10,12(3)
+	stw	11,16(3)
+	stw	0,24(3)
 
-	andc	7,7,0
-	and	28,28,0
-	andc	8,8,0
-	and	29,29,0
-	or	7,7,28
-	lwz	28,0(5)
-	andc	9,9,0
-	and	30,30,0
-	or	8,8,29
-	lwz	29,4(5)
-	andc	10,10,0
-	and	31,31,0
-	or	9,9,30
-	lwz	30,8(5)
-	or	10,10,31
-	lwz	31,12(5)
+	b	.Lpoly1305_blocks
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.size	poly1305_blocks_vsx,.-poly1305_blocks_vsx
+
+.align	5
+__poly1305_mul:
+	.long	0x11E05088
+	.long	0x12015088
+	.long	0x12225088
+	.long	0x12435088
+	.long	0x12645088
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+	.long	0x12823888
+	.long	0x11EFA0C0
+	.long	0x12833888
+	.long	0x1210A0C0
+	.long	0x12843888
+	.long	0x1231A0C0
+	.long	0x12803088
+	.long	0x1252A0C0
+	.long	0x12813088
+	.long	0x1273A0C0
+
+	.long	0x12814888
+	.long	0x11EFA0C0
+	.long	0x12824888
+	.long	0x1210A0C0
+	.long	0x12834888
+	.long	0x1231A0C0
+	.long	0x12844888
+	.long	0x1252A0C0
+	.long	0x12804088
+	.long	0x1273A0C0
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
 
-	addc	7,7,28
-	adde	8,8,29
-	adde	9,9,30
-	adde	10,10,31
-	li	29,4
-	stwbrx	7,0,4
-	li	30,8
-	stwbrx	8,29,4
-	li	31,12
-	stwbrx	9,30,4
-	stwbrx	10,31,4
-	lwz	28,80(1)
-	lwz	29,84(1)
-	lwz	30,88(1)
-	lwz	31,92(1)
-	addi	1,1,96
 	blr	
 .long	0
-.byte	0,12,4,1,0x80,4,3,0
-.size	poly1305_emit,.-poly1305_emit
-.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.byte	0,12,0x14,0,0,0,0,0
+.size	__poly1305_mul,.-__poly1305_mul
+
+.align	5
+__poly1305_blocks_vsx:
+	stwu	1,-384(1)
+	mflr	0
+	li	10,167
+	li	11,183
+	mfspr	12,256
+	stvx	20,10,1
+	addi	10,10,32
+	stvx	21,11,1
+	addi	11,11,32
+	stvx	22,10,1
+	addi	10,10,32
+	stvx	23,10,1
+	addi	10,10,32
+	stvx	24,11,1
+	addi	11,11,32
+	stvx	25,10,1
+	addi	10,10,32
+	stvx	26,10,1
+	addi	10,10,32
+	stvx	27,11,1
+	addi	11,11,32
+	stvx	28,10,1
+	addi	10,10,32
+	stvx	29,11,1
+	addi	11,11,32
+	stvx	30,10,1
+	stvx	31,11,1
+	stw	12,360(1)
+	li	12,-1
+	mtspr	256,12
+	stw	27,364(1)
+	stw	28,368(1)
+	stw	29,372(1)
+	stw	30,376(1)
+	stw	31,380(1)
+	stw	0,388(1)
+
+	bl	.LPICmeup
+
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	.long	0x7FA06699
+	.long	0x7F3B6699
+	.long	0x7F7C6699
+	.long	0x7FFD6699
+	.long	0x7FDE6699
+
+	cmplwi	7,0
+	bne	.Lskip_init_vsx
+
+	lwz	8,32(3)
+	lwz	9,36(3)
+	lwz	10,40(3)
+	lwz	11,44(3)
+
+	extrwi	7,8,26,6
+	extrwi	8,8,6,0
+	insrwi	8,9,20,6
+	extrwi	9,9,12,0
+	insrwi	9,10,14,6
+	extrwi	10,10,18,0
+	insrwi	10,11,8,6
+	extrwi	11,11,24,0
+
+	.long	0x7D4701E7
+	slwi	7,8,2
+	.long	0x7D6801E7
+	add	8,8,7
+	.long	0x7D8801E7
+	slwi	8,9,2
+	.long	0x7DA901E7
+	add	9,9,8
+	.long	0x7DC901E7
+	slwi	9,10,2
+	.long	0x7CCA01E7
+	add	10,10,9
+	.long	0x7CEA01E7
+	slwi	10,11,2
+	.long	0x7D0B01E7
+	add	11,11,10
+	.long	0x7D2B01E7
+
+	vor	0,10,10
+	vor	1,11,11
+	vor	2,13,13
+	vor	3,6,6
+	vor	4,8,8
+
+	bl	__poly1305_mul
+
+	.long	0xF1405057
+	.long	0xF1615857
+	.long	0xF1A26857
+	.long	0xF0C33057
+	.long	0xF1044057
+	.long	0xF0000057
+	.long	0xF0210857
+	.long	0xF0421057
+	.long	0xF0631857
+	.long	0xF0842057
+	.long	0x118BA5C4
+	.long	0x11CDA5C4
+	.long	0x10E6A5C4
+	.long	0x1128A5C4
+	.long	0x118C58C0
+	.long	0x11CE68C0
+	.long	0x10E730C0
+	.long	0x112940C0
+
+	bl	__poly1305_mul
+
+	addi	7,3,0x60
+	lwz	8,0(3)
+	lwz	9,4(3)
+	lwz	10,8(3)
+	lwz	11,12(3)
+	lwz	0,16(3)
+
+	.long	0x114A068C
+	.long	0x116B0E8C
+	.long	0x11AD168C
+	.long	0x10C61E8C
+	.long	0x1108268C
+	vslw	12,11,20
+	vslw	14,13,20
+	vslw	7,6,20
+	vslw	9,8,20
+	vadduwm	12,12,11
+	vadduwm	14,14,13
+	vadduwm	7,7,6
+	vadduwm	9,9,8
+
+	.long	0x7D5D1F99
+	.long	0x7D7E1F99
+	.long	0x7D9F1F99
+	.long	0x7DA03F99
+	.long	0x7DDB3F99
+	.long	0x7CDC3F99
+	.long	0x7CFD3F99
+	.long	0x7D1E3F99
+	.long	0x7D3F3F99
+
+	extrwi	7,8,26,6
+	extrwi	8,8,6,0
+	.long	0x7C0701E7
+	insrwi	8,9,20,6
+	extrwi	9,9,12,0
+	.long	0x7C2801E7
+	insrwi	9,10,14,6
+	extrwi	10,10,18,0
+	.long	0x7C4901E7
+	insrwi	10,11,8,6
+	extrwi	11,11,24,0
+	.long	0x7C6A01E7
+	insrwi	11,0,3,5
+	.long	0x7C8B01E7
+	li	0,1
+	stw	0,24(3)
+	b	.Loaded_vsx
+
+.align	4
+.Lskip_init_vsx:
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001819
+	.long	0x7C3B1819
+	.long	0x7C5C1819
+	.long	0x7C7D1819
+	.long	0x7C9E1819
+
+.Loaded_vsx:
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	li	7,0x60
+	li	8,0x70
+	addi	10,3,64
+	addi	11,1,39
+
+	vxor	20,20,20
+	.long	0xF000A057
+	.long	0xF021A057
+	.long	0xF042A057
+	.long	0xF063A057
+	.long	0xF084A057
+
+	.long	0x7F5F6699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	vspltisb	28,14
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x1295CEC4
+	.long	0x12D6D6C4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+	vand	21,21,29
+	vand	20,20,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10D4368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x7D5D1A99
+	.long	0x7D605299
+	.long	0x7D9B5299
+	.long	0x7DBC5299
+	.long	0x7DDD5299
+	.long	0x7EBE5299
+	.long	0x7EDF5299
+	.long	0x7EE75299
+	.long	0x7F085299
+	stvx	11,0,11
+	stvx	12,27,11
+	stvx	13,28,11
+	stvx	14,29,11
+	stvx	21,30,11
+	stvx	22,31,11
+	stvx	23,7,11
+	stvx	24,8,11
+
+	addi	4,4,0x40
+	addi	12,12,0x50
+	addi	0,5,-64
+	srdi	0,0,6
+	mtctr	0
+	b	.Loop_vsx
+
+.align	4
+.Loop_vsx:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	.long	0x11E55288
+	.long	0x12055A88
+	.long	0x12256A88
+	.long	0x12466A88
+
+	.long	0x12865288
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12676A88
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	lvx	12,31,11
+	.long	0x12885A88
+	.long	0x1273A0C0
+	lvx	11,30,11
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12875288
+	.long	0x1231A0C0
+	.long	0x12885288
+	.long	0x1252A0C0
+	lvx	14,8,11
+	.long	0x12895288
+	.long	0x1273A0C0
+	lvx	13,7,11
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+	.long	0x7F406699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	lvx	12,27,11
+	.long	0x12815888
+	.long	0x1273A0C0
+	lvx	11,0,11
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	lvx	14,29,11
+	.long	0x12806888
+	.long	0x1273A0C0
+	lvx	13,28,11
+
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12D6D6C4
+	.long	0x1355CEC4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+	vand	21,21,29
+	vand	26,26,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10DA368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	addi	4,4,0x40
+	bdnz	.Loop_vsx
+
+	neg	5,5
+	andi.	5,5,0x30
+	sub	4,4,5
+
+	.long	0x7D5D1E99
+	.long	0x7D605699
+	.long	0x7D9B5699
+	.long	0x7DBC5699
+	.long	0x7DDD5699
+
+.Last_vsx:
+	.long	0x11E55288
+	.long	0x12065288
+	.long	0x12275288
+	.long	0x12485288
+	.long	0x12695288
+
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12855A88
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	.long	0x7D9F5699
+	.long	0x12885A88
+	.long	0x1273A0C0
+	.long	0x7D7E5699
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12856A88
+	.long	0x1231A0C0
+	.long	0x12866A88
+	.long	0x1252A0C0
+	.long	0x7DC85699
+	.long	0x12876A88
+	.long	0x1273A0C0
+	.long	0x7DA75699
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	.long	0x7D9B5699
+	.long	0x12815888
+	.long	0x1273A0C0
+	.long	0x7D605699
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	.long	0x7DDD5699
+	.long	0x12806888
+	.long	0x1273A0C0
+	.long	0x7DBC5699
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+
+
+
+	.long	0xF00F7A57
+	.long	0xF0308257
+	.long	0xF0518A57
+	.long	0xF0729257
+	.long	0xF0939A57
+	.long	0x11EF00C0
+	.long	0x121008C0
+	.long	0x123110C0
+	.long	0x125218C0
+	.long	0x127320C0
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	beq	.Ldone_vsx
+
+	add	6,12,5
+
+	.long	0x7F406699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF297C057
+	vperm	21,23,24,31
+	.long	0xF2D7C357
+
+	.long	0x7DE03699
+	.long	0x7E1D3699
+
+	.long	0x12F4CEC4
+	.long	0x12B5D6C4
+	.long	0x1316DEC4
+	.long	0x12D6E6C4
+	vand	20,20,29
+	vand	23,23,29
+	vand	21,21,29
+	vand	22,22,29
+
+
+	.long	0x11384E8C
+	.long	0x10B42E8C
+	.long	0x10D7368C
+	.long	0x10F53E8C
+	.long	0x1116468C
+	vor	9,9,30
+
+	vperm	0,0,0,15
+	vand	5,5,    16
+	vperm	1,1,1,15
+	vand	6,6,    16
+	vperm	2,2,2,15
+	vand	7,7,    16
+	vperm	3,3,3,15
+	vand	8,8,    16
+	vperm	4,4,4,15
+	vand	9,9,    16
+
+	.long	0x10A500C0
+	vxor	0,0,0
+	.long	0x10C608C0
+	vxor	1,1,1
+	.long	0x10E710C0
+	vxor	2,2,2
+	.long	0x110818C0
+	vxor	3,3,3
+	.long	0x112920C0
+	vxor	4,4,4
+
+	xor.	5,5,5
+	b	.Last_vsx
+
+.align	4
+.Ldone_vsx:
+	lwz	0,388(1)
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001919
+	.long	0x7C3B1919
+	.long	0x7C5C1919
+	.long	0x7C7D1919
+	.long	0x7C9E1919
+
+	lwz	12,360(1)
+	mtlr	0
+	li	10,167
+	li	11,183
+	mtspr	256,12
+	lvx	20,10,1
+	addi	10,10,32
+	lvx	21,10,1
+	addi	10,10,32
+	lvx	22,11,1
+	addi	11,11,32
+	lvx	23,10,1
+	addi	10,10,32
+	lvx	24,11,1
+	addi	11,11,32
+	lvx	25,10,1
+	addi	10,10,32
+	lvx	26,11,1
+	addi	11,11,32
+	lvx	27,10,1
+	addi	10,10,32
+	lvx	28,11,1
+	addi	11,11,32
+	lvx	29,10,1
+	addi	10,10,32
+	lvx	30,11,1
+	lvx	31,10,1
+	lwz	27,364(1)
+	lwz	28,368(1)
+	lwz	29,372(1)
+	lwz	30,376(1)
+	lwz	31,380(1)
+	addi	1,1,384
+	blr	
+.long	0
+.byte	0,12,0x04,1,0x80,5,4,0
+.long	0
+.size	__poly1305_blocks_vsx,.-__poly1305_blocks_vsx
+
+.align	6
+.LPICmeup:
+	mflr	0
+	bcl	20,31,$+4
+	mflr	12
+	addi	12,12,56
+	mtlr	0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.space	28
+
+.long	0x00000000,0x03ffffff
+.long	0x00000000,0x03ffffff
+.long	0x00000000,0x0000001a
+.long	0x00000000,0x0000001a
+.long	0x00000000,0x00000028
+.long	0x00000000,0x00000028
+.long	0x00000000,0x0e0f0001
+.long	0x00000000,0x1e1f1011
+.long	0x01000000,0x01000000
+.long	0x01000000,0x01000000
+.long	0x07060504,0x03020100
+.long	0x0f0e0d0c,0x0b0a0908
+
+.long	0x00000000,0x00000000
+.long	0x00000000,0x04050607
+.long	0x04050607,0x00000000
+.long	0x00000000,0x00000000
+.long	0x00000000,0x00000000
+.long	0x04050607,0x00000000
+
+.long	0xffffffff,0x00000000
+.long	0xffffffff,0xffffffff
+.long	0xffffffff,0x00000000
+.long	0xffffffff,0x00000000
+.long	0x00000000,0x00000000
+.long	0xffffffff,0x00000000
+.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,64,100,111,116,45,97,115,109,0
 .align	2
diff --git a/sys/crypto/openssl/powerpc/vpaes-ppc.S b/sys/crypto/openssl/powerpc/vpaes-ppc.S
index fd11edc79389d..9a21c796aa814 100644
--- a/sys/crypto/openssl/powerpc/vpaes-ppc.S
+++ b/sys/crypto/openssl/powerpc/vpaes-ppc.S
@@ -667,7 +667,7 @@ vpaes_cbc_encrypt:
 	vor	24,0,0
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -719,7 +719,7 @@ vpaes_cbc_encrypt:
 	vor	24,25,25
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -1037,7 +1037,7 @@ _vpaes_schedule_core:
 
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1059,7 +1059,7 @@ _vpaes_schedule_core:
 	addi	9, 5, -15
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1150,7 +1150,7 @@ _vpaes_schedule_low_round:
 
 	vsldoi	1, 9, 7, 12
 	vxor	7, 7, 1
-	vspltisb	1, 0x0f
+	vspltisb	1,0x0f
 	vsldoi	4, 9, 7, 8
 
 
@@ -1246,7 +1246,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr	
@@ -1297,7 +1297,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr	
diff --git a/sys/crypto/openssl/powerpc64/bn-ppc.S b/sys/crypto/openssl/powerpc64/bn-ppc.S
new file mode 100644
index 0000000000000..50805d3d52bc7
--- /dev/null
+++ b/sys/crypto/openssl/powerpc64/bn-ppc.S
@@ -0,0 +1,1876 @@
+/* Do not modify. This file is auto-generated from ppc.pl. */
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+.globl	bn_sqr_comba4
+.type	bn_sqr_comba4,@function
+.globl	bn_sqr_comba8
+.type	bn_sqr_comba8,@function
+.globl	bn_mul_comba4
+.type	bn_mul_comba4,@function
+.globl	bn_mul_comba8
+.type	bn_mul_comba8,@function
+.globl	bn_sub_words
+.type	bn_sub_words,@function
+.globl	bn_add_words
+.type	bn_add_words,@function
+.globl	bn_div_words
+.type	bn_div_words,@function
+.globl	bn_sqr_words
+.type	bn_sqr_words,@function
+.globl	bn_mul_words
+.type	bn_mul_words,@function
+.globl	bn_mul_add_words
+.type	bn_mul_add_words,@function
+
+
+
+.machine	"any"
+.abiversion	2
+.text
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba4:
+.localentry	bn_sqr_comba4,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	ld	5,0(4)
+	mulld	9,5,5
+	mulhdu	10,5,5
+
+
+
+
+	std	9,0(3)
+
+	ld	6,8(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,0
+
+
+	addc	10,7,10
+	addze	11,8
+	addze	9,9
+
+	std	10,8(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,16(3)
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,24(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,32(3)
+
+	ld	5,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,40(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+
+	std	9,48(3)
+	std	10,56(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba4,.-bn_sqr_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba8:
+.localentry	bn_sqr_comba8,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	ld	5,0(4)
+	mulld	9,5,5
+	mulhdu	10,5,5
+	std	9,0(3)
+
+	ld	6,8(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,0
+	addze	9,0
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	std	10,8(3)
+
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	std	11,16(3)
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	std	9,24(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,0(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,32(3)
+
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,8(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,16(4)
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,40(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,0(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,48(3)
+
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,8(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,16(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,24(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,56(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,16(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,8(4)
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,64(3)
+
+	ld	5,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,24(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,32(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,72(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,24(4)
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,80(3)
+
+	ld	5,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,40(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,88(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,96(3)
+
+
+	ld	5,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,104(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	std	11,112(3)
+	std	9, 120(3)
+
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba8,.-bn_sqr_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba4:
+.localentry	bn_mul_comba4,0
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+	ld	6,0(4)
+	ld	7,0(5)
+	mulld	10,6,7
+	mulhdu	11,6,7
+	std	10,0(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,0
+	addze	10,0
+
+	ld	6, 8(4)
+	ld	7, 0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	std	11,8(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+
+	ld	6,0(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	std	12,16(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,0
+
+	ld	6,8(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	ld	6,16(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+	std	10,24(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,0
+
+	ld	6,16(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+
+	ld	6,8(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	std	11,32(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	ld	6,24(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	std	12,40(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+
+	std	10,48(3)
+	std	11,56(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba4,.-bn_mul_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba8:
+.localentry	bn_mul_comba8,0
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+	ld	6,0(4)
+	ld	7,0(5)
+	mulld	10,6,7
+	mulhdu	11,6,7
+	std	10,0(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	addze	12,9
+	addze	10,0
+
+	ld	6,8(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,8(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,0(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,16(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,8(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+
+	ld	6,16(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,24(3)
+
+	ld	6,32(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,24(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,16(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,8(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,0(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,32(3)
+
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,16(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,24(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,32(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,40(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,40(3)
+
+	ld	6,48(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,40(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,32(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,16(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,8(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,0(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,48(3)
+
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,8(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,16(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,24(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,32(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,40(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,48(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,56(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,56(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,48(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,40(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,32(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,24(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,16(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,8(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,64(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,24(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,32(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,40(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,48(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,56(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,72(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,48(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,40(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,32(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,24(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,80(3)
+
+	ld	6,32(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,40(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,48(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,56(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,88(3)
+
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,48(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,40(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,96(3)
+
+	ld	6,48(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,56(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,104(3)
+
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	std	12,112(3)
+	std	10,120(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba8,.-bn_mul_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_sub_words:
+.localentry	bn_sub_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	subfc.	7,0,6
+
+
+	beq	.Lppcasm_sub_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	addi	5,5,-8
+	mtctr	6
+.Lppcasm_sub_mainloop:
+	ldu	7,8(4)
+	ldu	8,8(5)
+	subfe	6,8,7
+
+
+	stdu	6,8(3)
+	bdnz	.Lppcasm_sub_mainloop
+.Lppcasm_sub_adios:
+	subfze	3,0
+	andi.	3,3,1
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_sub_words,.-bn_sub_words
+
+
+
+
+
+
+
+
+.align	4
+bn_add_words:
+.localentry	bn_add_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	addic.	6,6,0
+	beq	.Lppcasm_add_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	addi	5,5,-8
+	mtctr	6
+.Lppcasm_add_mainloop:
+	ldu	7,8(4)
+	ldu	8,8(5)
+	adde	8,7,8
+	stdu	8,8(3)
+	bdnz	.Lppcasm_add_mainloop
+.Lppcasm_add_adios:
+	addze	3,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_add_words,.-bn_add_words
+
+
+
+
+
+
+
+
+.align	4
+bn_div_words:
+.localentry	bn_div_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+	cmpldi	0,5,0
+	bne	.Lppcasm_div1
+	li	3,-1
+	blr	
+.Lppcasm_div1:
+	xor	0,0,0
+	li	8,64
+	cntlzd.	7,5
+	beq	.Lppcasm_div2
+	subf	8,7,8
+	srd.	9,3,8
+	td	16,9,0
+.Lppcasm_div2:
+	cmpld	0,3,5
+	blt	.Lppcasm_div3
+	subf	3,5,3
+.Lppcasm_div3:
+	cmpi	0,0,7,0
+	beq	.Lppcasm_div4
+	sld	3,3,7
+	srd	8,4,8
+	sld	5,5,7
+	or	3,3,8
+	sld	4,4,7
+.Lppcasm_div4:
+	srdi	9,5,32
+
+
+	li	6,2
+	mtctr	6
+.Lppcasm_divouterloop:
+	srdi	8,3,32
+	srdi	11,4,32
+
+	cmpld	0,8,9
+	bne	.Lppcasm_div5
+
+	li	8,-1
+	clrldi	8,8,32
+	b	.Lppcasm_div6
+.Lppcasm_div5:
+	divdu	8,3,9
+.Lppcasm_div6:
+	mulld	12,9,8
+	clrldi	10,5,32
+	mulld	6,8,10
+
+.Lppcasm_divinnerloop:
+	subf	10,12,3
+	srdi	7,10,32
+	addic.	7,7,0
+
+
+
+	sldi	7,10,32
+	or	7,7,11
+	cmpld	1,6,7
+	bne	.Lppcasm_divinnerexit
+	ble	1,.Lppcasm_divinnerexit
+	addi	8,8,-1
+	subf	12,9,12
+	clrldi	10,5,32
+	subf	6,10,6
+	b	.Lppcasm_divinnerloop
+.Lppcasm_divinnerexit:
+	srdi	10,6,32
+	sldi	11,6,32
+	cmpld	1,4,11
+	add	12,12,10
+	bge	1,.Lppcasm_div7
+	addi	12,12,1
+.Lppcasm_div7:
+	subf	11,11,4
+	cmpld	1,3,12
+	bge	1,.Lppcasm_div8
+	addi	8,8,-1
+	add	3,5,3
+.Lppcasm_div8:
+	subf	12,12,3
+	sldi	4,11,32
+
+
+
+	insrdi	11,12,32,32
+	rotldi	3,11,32
+	bdz	.Lppcasm_div9
+	sldi	0,8,32
+	b	.Lppcasm_divouterloop
+.Lppcasm_div9:
+	or	3,8,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_div_words,.-bn_div_words
+
+
+
+
+
+
+
+.align	4
+bn_sqr_words:
+.localentry	bn_sqr_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	addic.	5,5,0
+	beq	.Lppcasm_sqr_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	mtctr	5
+.Lppcasm_sqr_mainloop:
+
+	ldu	6,8(4)
+	mulld	7,6,6
+	mulhdu	8,6,6
+	stdu	7,8(3)
+	stdu	8,8(3)
+	bdnz	.Lppcasm_sqr_mainloop
+.Lppcasm_sqr_adios:
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_sqr_words,.-bn_sqr_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_words:
+.localentry	bn_mul_words,0
+
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_mw_REM
+	mtctr	7
+.Lppcasm_mw_LOOP:
+
+	ld	8,0(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+
+
+
+
+	std	9,0(3)
+
+	ld	8,8(4)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+
+	std	11,8(3)
+
+	ld	8,16(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	adde	9,9,12
+
+	std	9,16(3)
+
+	ld	8,24(4)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+
+	std	11,24(3)
+
+	addi	3,3,32
+	addi	4,4,32
+	bdnz	.Lppcasm_mw_LOOP
+
+.Lppcasm_mw_REM:
+	andi.	5,5,0x3
+	beq	.Lppcasm_mw_OVER
+
+	ld	8,0(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,0(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+
+	ld	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,8(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+	ld	8,16(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,16(3)
+	addi	12,10,0
+
+.Lppcasm_mw_OVER:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_words,.-bn_mul_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_add_words:
+.localentry	bn_mul_add_words,0
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_maw_leftover
+	mtctr	7
+.Lppcasm_maw_mainloop:
+
+	ld	8,0(4)
+	ld	11,0(3)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+
+
+
+
+
+	std	9,0(3)
+
+
+	ld	8,8(4)
+	ld	9,8(3)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+
+	std	11,8(3)
+
+
+	ld	8,16(4)
+	mulld	9,6,8
+	ld	11,16(3)
+	mulhdu	10,6,8
+	adde	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+	std	9,16(3)
+
+
+	ld	8,24(4)
+	mulld	11,6,8
+	ld	9,24(3)
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+	addze	12,12
+	std	11,24(3)
+	addi	3,3,32
+	addi	4,4,32
+	bdnz	.Lppcasm_maw_mainloop
+
+.Lppcasm_maw_leftover:
+	andi.	5,5,0x3
+	beq	.Lppcasm_maw_adios
+	addi	3,3,-8
+	addi	4,4,-8
+
+	mtctr	5
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+.Lppcasm_maw_adios:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_add_words,.-bn_mul_add_words
+.align	4
diff --git a/sys/crypto/openssl/powerpc64/ecp_nistp521-ppc64.S b/sys/crypto/openssl/powerpc64/ecp_nistp521-ppc64.S
new file mode 100644
index 0000000000000..5905180d168a6
--- /dev/null
+++ b/sys/crypto/openssl/powerpc64/ecp_nistp521-ppc64.S
@@ -0,0 +1,354 @@
+/* Do not modify. This file is auto-generated from ecp_nistp521-ppc64.pl. */
+.machine	"any"
+.abiversion	2
+.text
+
+.globl	p521_felem_mul
+.type	p521_felem_mul,@function
+.align	5
+p521_felem_mul:
+.localentry	p521_felem_mul,0
+
+
+	mr	12,1
+	stdu	1,-16*13(1)
+
+	stxv	52,-16*12(12)
+	stxv	53,-16*11(12)
+	stxv	54,-16*10(12)
+	stxv	55,-16*9(12)
+	stxv	56,-16*8(12)
+	stxv	57,-16*7(12)
+	stxv	58,-16*6(12)
+	stxv	59,-16*5(12)
+	stxv	60,-16*4(12)
+	stxv	61,-16*3(12)
+	stxv	62,-16*2(12)
+	stxv	63,-16*1(12)
+
+	vspltisw	0,0
+
+	lxsd	13,0(4)
+	lxsd	14,8(4)
+	lxsd	15,16(4)
+	lxsd	16,24(4)
+	lxsd	17,32(4)
+	lxsd	18,40(4)
+	lxsd	19,48(4)
+	lxsd	20,56(4)
+	lxsd	21,64(4)
+
+	lxsd	3,0(5)
+	lxsd	4,8(5)
+	lxsd	5,16(5)
+	lxsd	6,24(5)
+	lxsd	7,32(5)
+	lxsd	8,40(5)
+	lxsd	9,48(5)
+	lxsd	10,56(5)
+	lxsd	11,64(5)
+
+	.long	0x12ED1823
+
+	xxpermdi	33,45,46,0b00
+	xxpermdi	34,36,35,0b00
+	.long	0x13011023
+
+	xxpermdi	34,37,36,0b00
+	.long	0x13211023
+	.long	0x132F1E63
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13411023
+	xxpermdi	44,47,48,0b00
+	xxpermdi	54,36,35,0b00
+	.long	0x134CB6A3
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13611023
+	xxpermdi	54,37,36,0b00
+	.long	0x136CB6E3
+	.long	0x13711EE3
+
+	xxpermdi	34,40,39,0b00
+	.long	0x13811023
+	xxpermdi	54,38,37,0b00
+	.long	0x138CB723
+
+	xxpermdi	34,41,40,0b00
+	.long	0x13A11023
+	xxpermdi	54,39,38,0b00
+	.long	0x13ACB763
+
+	xxpermdi	34,42,41,0b00
+	.long	0x13C11023
+	xxpermdi	54,40,39,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,43,42,0b00
+	.long	0x13E11023
+	xxpermdi	54,41,40,0b00
+	.long	0x13ECB7E3
+
+	xxpermdi	33,49,50,0b00
+	xxpermdi	34,36,35,0b00
+	.long	0x13811723
+
+	xxpermdi	34,37,36,0b00
+	.long	0x13A11763
+	.long	0x13B31F63
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13C117A3
+	xxpermdi	44,51,52,0b00
+	xxpermdi	54,36,35,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13E117E3
+	xxpermdi	54,37,36,0b00
+	.long	0x13ECB7E3
+	.long	0x13F51FE3
+
+	li	8,0
+	li	9,1
+	mtvsrdd	33,9,8
+	.long	0x10630DC4
+	.long	0x10840DC4
+	.long	0x10A50DC4
+	.long	0x10C60DC4
+	.long	0x10E70DC4
+	.long	0x11080DC4
+	.long	0x11290DC4
+	.long	0x114A0DC4
+	.long	0x116B0DC4
+
+	.long	0x13D55FA3
+
+	xxpermdi	34,43,42,0b00
+	xxpermdi	33,52,53,0b00
+	.long	0x13A11763
+
+	xxpermdi	33,51,52,0b00
+	.long	0x13811723
+	.long	0x13954F23
+
+	xxpermdi	33,50,51,0b00
+	.long	0x136116E3
+	xxpermdi	54,41,40,0b00
+	xxpermdi	44,52,53,0b00
+	.long	0x136CB6E3
+
+	xxpermdi	33,49,50,0b00
+	.long	0x134116A3
+	xxpermdi	44,51,52,0b00
+	.long	0x134CB6A3
+	.long	0x13553EA3
+
+	xxpermdi	33,48,49,0b00
+	.long	0x13211663
+	xxpermdi	44,50,51,0b00
+	.long	0x132CB663
+
+	xxpermdi	33,47,48,0b00
+	.long	0x13011623
+	xxpermdi	44,49,50,0b00
+	.long	0x130CB623
+
+	xxpermdi	33,46,47,0b00
+	.long	0x12E115E3
+	xxpermdi	44,48,49,0b00
+	.long	0x12ECB5E3
+
+	xxpermdi	34,39,38,0b00
+	xxpermdi	33,52,53,0b00
+	.long	0x13211663
+
+	xxpermdi	33,51,52,0b00
+	.long	0x13011623
+	.long	0x13152E23
+
+	xxpermdi	33,50,51,0b00
+	.long	0x12E115E3
+	xxpermdi	54,37,36,0b00
+	xxpermdi	44,52,53,0b00
+	.long	0x12ECB5E3
+
+	stxv	55,0(3)
+	stxv	56,16(3)
+	stxv	57,32(3)
+	stxv	58,48(3)
+	stxv	59,64(3)
+	stxv	60,80(3)
+	stxv	61,96(3)
+	stxv	62,112(3)
+	stxv	63,128(3)
+
+	ld	12,0(1)
+	lxv	52,-16*12(12)
+	lxv	53,-16*11(12)
+	lxv	54,-16*10(12)
+	lxv	55,-16*9(12)
+	lxv	56,-16*8(12)
+	lxv	57,-16*7(12)
+	lxv	58,-16*6(12)
+	lxv	59,-16*5(12)
+	lxv	60,-16*4(12)
+	lxv	61,-16*3(12)
+	lxv	62,-16*2(12)
+	lxv	63,-16*1(12)
+	mr	1,12
+
+	blr	
+.size	p521_felem_mul,.-p521_felem_mul
+
+.globl	p521_felem_square
+.type	p521_felem_square,@function
+.align	5
+p521_felem_square:
+.localentry	p521_felem_square,0
+
+
+	mr	12,1
+	stdu	1,-16*13(1)
+
+	stxv	52,-16*12(12)
+	stxv	53,-16*11(12)
+	stxv	54,-16*10(12)
+	stxv	55,-16*9(12)
+	stxv	56,-16*8(12)
+	stxv	57,-16*7(12)
+	stxv	58,-16*6(12)
+	stxv	59,-16*5(12)
+	stxv	60,-16*4(12)
+	stxv	61,-16*3(12)
+	stxv	62,-16*2(12)
+	stxv	63,-16*1(12)
+
+	vspltisw	0,0
+
+	lxsd	13,0(4)
+	lxsd	14,8(4)
+	lxsd	15,16(4)
+	lxsd	16,24(4)
+	lxsd	17,32(4)
+	lxsd	18,40(4)
+	lxsd	19,48(4)
+	lxsd	20,56(4)
+	lxsd	21,64(4)
+
+	li	8,0
+	li	9,1
+	mtvsrdd	33,9,8
+	.long	0x106D0DC4
+	.long	0x108E0DC4
+	.long	0x10AF0DC4
+	.long	0x10D00DC4
+	.long	0x10F10DC4
+	.long	0x11120DC4
+	.long	0x11330DC4
+	.long	0x11540DC4
+	.long	0x11750DC4
+	.long	0x12ED6823
+
+	.long	0x130D2023
+
+	xxpermdi	33,45,46,0b00
+	xxpermdi	34,37,46,0b00
+	.long	0x13211023
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13411023
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13611023
+	.long	0x136F7EE3
+
+	xxpermdi	34,40,39,0b00
+	.long	0x13811023
+	.long	0x138F3723
+
+	xxpermdi	34,41,40,0b00
+	.long	0x13A11023
+	xxpermdi	44,47,48,0b00
+	xxpermdi	54,39,48,0b00
+	.long	0x13ACB763
+
+	xxpermdi	34,42,41,0b00
+	.long	0x13C11023
+	xxpermdi	54,40,39,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,43,42,0b00
+	.long	0x13E11023
+	xxpermdi	54,41,40,0b00
+	.long	0x13ECB7E3
+	.long	0x13F18FE3
+
+	.long	0x13124623
+
+	.long	0x13534EA3
+
+	.long	0x13945723
+
+	.long	0x13D55FA3
+
+	mtvsrdd	33,9,8
+	.long	0x11080DC4
+	.long	0x11290DC4
+	.long	0x114A0DC4
+	.long	0x116B0DC4
+
+	.long	0x13B45F63
+
+	.long	0x13935F23
+
+	xxpermdi	34,43,42,0b00
+	xxpermdi	33,50,51,0b00
+	.long	0x136116E3
+
+	xxpermdi	33,49,50,0b00
+	.long	0x134116A3
+
+	xxpermdi	33,48,49,0b00
+	.long	0x13211663
+	.long	0x13324E63
+
+	xxpermdi	33,47,48,0b00
+	.long	0x13011623
+	.long	0x13114E23
+
+	xxpermdi	33,46,47,0b00
+	.long	0x12E115E3
+	xxpermdi	34,41,40,0b00
+	xxpermdi	33,48,49,0b00
+	.long	0x12E115E3
+
+	stxv	55,0(3)
+	stxv	56,16(3)
+	stxv	57,32(3)
+	stxv	58,48(3)
+	stxv	59,64(3)
+	stxv	60,80(3)
+	stxv	61,96(3)
+	stxv	62,112(3)
+	stxv	63,128(3)
+
+	ld	12,0(1)
+	lxv	52,-16*12(12)
+	lxv	53,-16*11(12)
+	lxv	54,-16*10(12)
+	lxv	55,-16*9(12)
+	lxv	56,-16*8(12)
+	lxv	57,-16*7(12)
+	lxv	58,-16*6(12)
+	lxv	59,-16*5(12)
+	lxv	60,-16*4(12)
+	lxv	61,-16*3(12)
+	lxv	62,-16*2(12)
+	lxv	63,-16*1(12)
+	mr	1,12
+
+	blr	
+.size	p521_felem_square,.-p521_felem_square
+
diff --git a/sys/crypto/openssl/powerpc64/keccak1600-ppc64.S b/sys/crypto/openssl/powerpc64/keccak1600-ppc64.S
index bac074cdcf6f0..251f59855f5d4 100644
--- a/sys/crypto/openssl/powerpc64/keccak1600-ppc64.S
+++ b/sys/crypto/openssl/powerpc64/keccak1600-ppc64.S
@@ -304,19 +304,19 @@ KeccakF1600:
 dword_le_load:
 .localentry	dword_le_load,0
 
-	lbzu	0,1(3)
-	lbzu	4,1(3)
-	lbzu	5,1(3)
+	lbz	0,1(3)
+	lbz	4,2(3)
+	lbz	5,3(3)
 	insrdi	0,4,8,48
-	lbzu	4,1(3)
+	lbz	4,4(3)
 	insrdi	0,5,8,40
-	lbzu	5,1(3)
+	lbz	5,5(3)
 	insrdi	0,4,8,32
-	lbzu	4,1(3)
+	lbz	4,6(3)
 	insrdi	0,5,8,24
-	lbzu	5,1(3)
+	lbz	5,7(3)
 	insrdi	0,4,8,16
-	lbzu	4,1(3)
+	lbzu	4,8(3)
 	insrdi	0,5,8,8
 	insrdi	0,4,8,0
 	blr	
@@ -579,21 +579,21 @@ SHA3_squeeze:
 	cmpldi	30,8
 	blt	.Lsqueeze_tail
 
-	stbu	0,1(29)
+	stb	0,1(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,2(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,3(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,4(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,5(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,6(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,7(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stbu	0,8(29)
 
 	subic.	30,30,8
 	beq	.Lsqueeze_done
diff --git a/sys/crypto/openssl/powerpc64/poly1305-ppc.S b/sys/crypto/openssl/powerpc64/poly1305-ppc.S
index d8b1066022933..523a590809cd6 100644
--- a/sys/crypto/openssl/powerpc64/poly1305-ppc.S
+++ b/sys/crypto/openssl/powerpc64/poly1305-ppc.S
@@ -12,6 +12,7 @@ poly1305_init_int:
 	std	0,0(3)
 	std	0,8(3)
 	std	0,16(3)
+	stw	0,24(3)
 
 	cmpld	4,0
 	beq-	.Lno_key
@@ -48,6 +49,7 @@ poly1305_init_int:
 poly1305_blocks:
 .localentry	poly1305_blocks,0
 
+.Lpoly1305_blocks:
 	srdi.	5,5,4
 	beq-	.Labort
 
@@ -138,48 +140,1003 @@ poly1305_blocks:
 .long	0
 .byte	0,12,4,1,0x80,5,4,0
 .size	poly1305_blocks,.-poly1305_blocks
-
 .globl	poly1305_emit
 .type	poly1305_emit,@function
-.align	4
+.align	5
 poly1305_emit:
 .localentry	poly1305_emit,0
 
-	ld	7,0(3)
-	ld	8,8(3)
-	ld	9,16(3)
-	ld	6,0(5)
-	ld	5,8(5)
+	lwz	7,0(3)
+	lwz	8,4(3)
+	lwz	9,8(3)
+	lwz	10,12(3)
+	lwz	11,16(3)
+	lwz	0,24(3)
+
+	sldi	8,8,26
+	sldi	12,9,52
+	srdi	9,9,12
+	sldi	10,10,14
+	add	7,7,8
+	addc	7,7,12
+	sldi	12,11,40
+	srdi	11,11,24
+	adde	8,9,10
+	addc	8,8,12
+	addze	9,11
+
+	ld	10,0(3)
+	ld	11,8(3)
+	ld	12,16(3)
+
+	neg	0,0
+	xor	7,7,10
+	xor	8,8,11
+	xor	9,9,12
+	and	7,7,0
+	and	8,8,0
+	and	9,9,0
+	xor	7,7,10
+	xor	8,8,11
+	xor	9,9,12
 
 	addic	10,7,5
 	addze	11,8
 	addze	12,9
 
-	srdi	0,12,2
-	neg	0,0
+	srdi	12,12,2
+	neg	12,12
 
-	andc	7,7,0
-	and	10,10,0
-	andc	8,8,0
-	and	11,11,0
+	andc	7,7,12
+	and	10,10,12
+	andc	8,8,12
+	and	11,11,12
 	or	7,7,10
 	or	8,8,11
-	rotldi	6,6,32
-	rotldi	5,5,32
-	addc	7,7,6
-	adde	8,8,5
-	rldicl	0,7,32,32
-	li	10,4
-	stwbrx	7,0,4
-	rldicl	7,8,32,32
-	li	11,8
-	stwbrx	0,10,4
-	li	12,12
-	stwbrx	8,11,4
-	stwbrx	7,12,4
+
+	lwz	12,4(5)
+	lwz	9,12(5)
+	lwz	10,0(5)
+	lwz	11,8(5)
+
+	insrdi	10,12,32,0
+	insrdi	11,9,32,0
+
+	addc	7,7,10
+	adde	8,8,11
+
+	addi	3,4,-1
+	addi	4,4,7
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	stbu	8,1(4)
+
 	blr	
 .long	0
 .byte	0,12,0x14,0,0,0,3,0
 .size	poly1305_emit,.-poly1305_emit
-.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.globl	poly1305_blocks_vsx
+.type	poly1305_blocks_vsx,@function
+.align	5
+poly1305_blocks_vsx:
+.localentry	poly1305_blocks_vsx,0
+
+	lwz	7,24(3)
+	cmpldi	5,128
+	bge	__poly1305_blocks_vsx
+
+	neg	0,7
+	lwz	7,0(3)
+	lwz	8,4(3)
+	lwz	9,8(3)
+	lwz	10,12(3)
+	lwz	11,16(3)
+
+	sldi	8,8,26
+	sldi	12,9,52
+	add	7,7,8
+	srdi	9,9,12
+	sldi	10,10,14
+	addc	7,7,12
+	sldi	8,11,40
+	adde	9,9,10
+	srdi	11,11,24
+	addc	9,9,8
+	addze	11,11
+
+	ld	8,0(3)
+	ld	10,8(3)
+	ld	12,16(3)
+
+	xor	7,7,8
+	xor	9,9,10
+	xor	11,11,12
+	and	7,7,0
+	and	9,9,0
+	and	11,11,0
+	xor	7,7,8
+	xor	9,9,10
+	xor	11,11,12
+
+	li	0,0
+	std	7,0(3)
+	std	9,8(3)
+	std	11,16(3)
+	stw	0,24(3)
+
+	b	.Lpoly1305_blocks
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.size	poly1305_blocks_vsx,.-poly1305_blocks_vsx
+
+.align	5
+__poly1305_mul:
+	mulld	9,6,27
+	mulhdu	10,6,27
+
+	mulld	30,7,29
+	mulhdu	31,7,29
+	addc	9,9,30
+	adde	10,10,31
+
+	mulld	30,6,28
+	mulhdu	11,6,28
+	addc	10,10,30
+	addze	11,11
+
+	mulld	30,7,27
+	mulhdu	31,7,27
+	addc	10,10,30
+	adde	11,11,31
+
+	mulld	30,8,29
+	mulld	31,8,27
+	addc	10,10,30
+	adde	11,11,31
+
+	andc	30,11,0
+	and	8,11,0
+	srdi	31,30,2
+	add	30,30,31
+	addc	6,9,30
+	addze	7,10
+	addze	8,8
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.size	__poly1305_mul,.-__poly1305_mul
+
+.align	5
+__poly1305_splat:
+	rldicl	9,6,0,38
+	rldicl	10,6,38,38
+	stw	9,0x00(31)
+
+	rldicl	11,6,12,52
+	slwi	9,10,2
+	stw	10,0x10(31)
+	add	9,9,10
+	stw	9,0x20(31)
+
+	insrdi	11,7,14,38
+	slwi	9,11,2
+	stw	11,0x30(31)
+	add	9,9,11
+	stw	9,0x40(31)
+
+	rldicl	10,7,50,38
+	rldicl	11,7,24,40
+	slwi	9,10,2
+	stw	10,0x50(31)
+	add	9,9,10
+	stw	9,0x60(31)
+
+	insrdi	11,8,3,37
+	slwi	9,11,2
+	stw	11,0x70(31)
+	add	9,9,11
+	stw	9,0x80(31)
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.size	__poly1305_splat,.-__poly1305_splat
+
+.align	5
+__poly1305_blocks_vsx:
+	stdu	1,-432(1)
+	mflr	0
+	li	10,191
+	li	11,207
+	li	12,-1
+	stvx	20,10,1
+	addi	10,10,32
+	stvx	21,11,1
+	addi	11,11,32
+	stvx	22,10,1
+	addi	10,10,32
+	stvx	23,10,1
+	addi	10,10,32
+	stvx	24,11,1
+	addi	11,11,32
+	stvx	25,10,1
+	addi	10,10,32
+	stvx	26,10,1
+	addi	10,10,32
+	stvx	27,11,1
+	addi	11,11,32
+	stvx	28,10,1
+	addi	10,10,32
+	stvx	29,11,1
+	addi	11,11,32
+	stvx	30,10,1
+	stvx	31,11,1
+	stw	12,388(1)
+	li	12,-1
+	or	12,12,12
+	std	27,392(1)
+	std	28,400(1)
+	std	29,408(1)
+	std	30,416(1)
+	std	31,424(1)
+	std	0,448(1)
+
+	bl	.LPICmeup
+
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	.long	0x7FA06699
+	.long	0x7F3B6699
+	.long	0x7F7C6699
+	.long	0x7FFD6699
+	.long	0x7FDE6699
+
+	cmplwi	7,0
+	bne	.Lskip_init_vsx
+
+	ld	27,32(3)
+	ld	28,40(3)
+	srdi	29,28,2
+	li	0,3
+	add	29,29,28
+
+	mr	6,27
+	mr	7,28
+	li	8,0
+	addi	31,3,56
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,48
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,60
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,52
+	bl	__poly1305_splat
+
+	ld	6,0(3)
+	ld	7,8(3)
+	ld	8,16(3)
+
+	rldicl	9,6,0,38
+	rldicl	10,6,38,38
+	rldicl	11,6,12,52
+	.long	0x7C0901E7
+	insrdi	11,7,14,38
+	.long	0x7C2A01E7
+	rldicl	10,7,50,38
+	.long	0x7C4B01E7
+	rldicl	11,7,24,40
+	.long	0x7C6A01E7
+	insrdi	11,8,3,37
+	.long	0x7C8B01E7
+	li	0,1
+	stw	0,24(3)
+	b	.Loaded_vsx
+
+.align	4
+.Lskip_init_vsx:
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001819
+	.long	0x7C3B1819
+	.long	0x7C5C1819
+	.long	0x7C7D1819
+	.long	0x7C9E1819
+
+.Loaded_vsx:
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	li	7,0x60
+	li	8,0x70
+	addi	10,3,64
+	addi	11,1,63
+
+	vxor	20,20,20
+	.long	0xF000A057
+	.long	0xF021A057
+	.long	0xF042A057
+	.long	0xF063A057
+	.long	0xF084A057
+
+	.long	0x7F5F6699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	vspltisb	28,14
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x1295CEC4
+	.long	0x12D6D6C4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+	vand	21,21,29
+	vand	20,20,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10D4368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x7D5D1A99
+	.long	0x7D605299
+	.long	0x7D9B5299
+	.long	0x7DBC5299
+	.long	0x7DDD5299
+	.long	0x7EBE5299
+	.long	0x7EDF5299
+	.long	0x7EE75299
+	.long	0x7F085299
+	stvx	11,0,11
+	stvx	12,27,11
+	stvx	13,28,11
+	stvx	14,29,11
+	stvx	21,30,11
+	stvx	22,31,11
+	stvx	23,7,11
+	stvx	24,8,11
+
+	addi	4,4,0x40
+	addi	12,12,0x50
+	addi	0,5,-64
+	srdi	0,0,6
+	mtctr	0
+	b	.Loop_vsx
+
+.align	4
+.Loop_vsx:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	.long	0x11E55288
+	.long	0x12055A88
+	.long	0x12256A88
+	.long	0x12466A88
+
+	.long	0x12865288
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12676A88
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	lvx	12,31,11
+	.long	0x12885A88
+	.long	0x1273A0C0
+	lvx	11,30,11
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12875288
+	.long	0x1231A0C0
+	.long	0x12885288
+	.long	0x1252A0C0
+	lvx	14,8,11
+	.long	0x12895288
+	.long	0x1273A0C0
+	lvx	13,7,11
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+	.long	0x7F406699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	lvx	12,27,11
+	.long	0x12815888
+	.long	0x1273A0C0
+	lvx	11,0,11
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	lvx	14,29,11
+	.long	0x12806888
+	.long	0x1273A0C0
+	lvx	13,28,11
+
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12D6D6C4
+	.long	0x1355CEC4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+	vand	21,21,29
+	vand	26,26,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10DA368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	addi	4,4,0x40
+	bdnz	.Loop_vsx
+
+	neg	5,5
+	andi.	5,5,0x30
+	sub	4,4,5
+
+	.long	0x7D5D1E99
+	.long	0x7D605699
+	.long	0x7D9B5699
+	.long	0x7DBC5699
+	.long	0x7DDD5699
+
+.Last_vsx:
+	.long	0x11E55288
+	.long	0x12065288
+	.long	0x12275288
+	.long	0x12485288
+	.long	0x12695288
+
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12855A88
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	.long	0x7D9F5699
+	.long	0x12885A88
+	.long	0x1273A0C0
+	.long	0x7D7E5699
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12856A88
+	.long	0x1231A0C0
+	.long	0x12866A88
+	.long	0x1252A0C0
+	.long	0x7DC85699
+	.long	0x12876A88
+	.long	0x1273A0C0
+	.long	0x7DA75699
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	.long	0x7D9B5699
+	.long	0x12815888
+	.long	0x1273A0C0
+	.long	0x7D605699
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	.long	0x7DDD5699
+	.long	0x12806888
+	.long	0x1273A0C0
+	.long	0x7DBC5699
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+
+
+
+	.long	0xF00F7A57
+	.long	0xF0308257
+	.long	0xF0518A57
+	.long	0xF0729257
+	.long	0xF0939A57
+	.long	0x11EF00C0
+	.long	0x121008C0
+	.long	0x123110C0
+	.long	0x125218C0
+	.long	0x127320C0
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	beq	.Ldone_vsx
+
+	add	6,12,5
+
+	.long	0x7F406699
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+	vperm	21,21,21,26
+	vperm	22,22,22,26
+	vperm	23,23,23,26
+	vperm	24,24,24,26
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF297C057
+	vperm	21,23,24,31
+	.long	0xF2D7C357
+
+	.long	0x7DE03699
+	.long	0x7E1D3699
+
+	.long	0x12F4CEC4
+	.long	0x12B5D6C4
+	.long	0x1316DEC4
+	.long	0x12D6E6C4
+	vand	20,20,29
+	vand	23,23,29
+	vand	21,21,29
+	vand	22,22,29
+
+
+	.long	0x11384E8C
+	.long	0x10B42E8C
+	.long	0x10D7368C
+	.long	0x10F53E8C
+	.long	0x1116468C
+	vor	9,9,30
+
+	vperm	0,0,0,15
+	vand	5,5,    16
+	vperm	1,1,1,15
+	vand	6,6,    16
+	vperm	2,2,2,15
+	vand	7,7,    16
+	vperm	3,3,3,15
+	vand	8,8,    16
+	vperm	4,4,4,15
+	vand	9,9,    16
+
+	.long	0x10A500C0
+	vxor	0,0,0
+	.long	0x10C608C0
+	vxor	1,1,1
+	.long	0x10E710C0
+	vxor	2,2,2
+	.long	0x110818C0
+	vxor	3,3,3
+	.long	0x112920C0
+	vxor	4,4,4
+
+	xor.	5,5,5
+	b	.Last_vsx
+
+.align	4
+.Ldone_vsx:
+	ld	0,448(1)
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001919
+	.long	0x7C3B1919
+	.long	0x7C5C1919
+	.long	0x7C7D1919
+	.long	0x7C9E1919
+
+	lwz	12,388(1)
+	mtlr	0
+	li	10,191
+	li	11,207
+	or	12,12,12
+	lvx	20,10,1
+	addi	10,10,32
+	lvx	21,10,1
+	addi	10,10,32
+	lvx	22,11,1
+	addi	11,11,32
+	lvx	23,10,1
+	addi	10,10,32
+	lvx	24,11,1
+	addi	11,11,32
+	lvx	25,10,1
+	addi	10,10,32
+	lvx	26,11,1
+	addi	11,11,32
+	lvx	27,10,1
+	addi	10,10,32
+	lvx	28,11,1
+	addi	11,11,32
+	lvx	29,10,1
+	addi	10,10,32
+	lvx	30,11,1
+	lvx	31,10,1
+	ld	27,392(1)
+	ld	28,400(1)
+	ld	29,408(1)
+	ld	30,416(1)
+	ld	31,424(1)
+	addi	1,1,432
+	blr	
+.long	0
+.byte	0,12,0x04,1,0x80,5,4,0
+.long	0
+.size	__poly1305_blocks_vsx,.-__poly1305_blocks_vsx
+
+.align	6
+.LPICmeup:
+	mflr	0
+	bcl	20,31,$+4
+	mflr	12
+	addi	12,12,56
+	mtlr	0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.space	28
+
+.long	0x00000000,0x03ffffff
+.long	0x00000000,0x03ffffff
+.long	0x00000000,0x0000001a
+.long	0x00000000,0x0000001a
+.long	0x00000000,0x00000028
+.long	0x00000000,0x00000028
+.long	0x00000000,0x0e0f0001
+.long	0x00000000,0x1e1f1011
+.long	0x01000000,0x01000000
+.long	0x01000000,0x01000000
+.long	0x07060504,0x03020100
+.long	0x0f0e0d0c,0x0b0a0908
+
+.long	0x00000000,0x00000000
+.long	0x00000000,0x04050607
+.long	0x04050607,0x00000000
+.long	0x00000000,0x00000000
+.long	0x00000000,0x00000000
+.long	0x04050607,0x00000000
+
+.long	0xffffffff,0x00000000
+.long	0xffffffff,0xffffffff
+.long	0xffffffff,0x00000000
+.long	0xffffffff,0x00000000
+.long	0x00000000,0x00000000
+.long	0xffffffff,0x00000000
+.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,64,100,111,116,45,97,115,109,0
 .align	2
diff --git a/sys/crypto/openssl/powerpc64/vpaes-ppc.S b/sys/crypto/openssl/powerpc64/vpaes-ppc.S
index bf0e1b3b89999..e92c9cd555262 100644
--- a/sys/crypto/openssl/powerpc64/vpaes-ppc.S
+++ b/sys/crypto/openssl/powerpc64/vpaes-ppc.S
@@ -674,7 +674,7 @@ vpaes_cbc_encrypt:
 	vor	24,0,0
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -726,7 +726,7 @@ vpaes_cbc_encrypt:
 	vor	24,25,25
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -1044,7 +1044,7 @@ _vpaes_schedule_core:
 
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1066,7 +1066,7 @@ _vpaes_schedule_core:
 	addi	9, 5, -15
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1157,7 +1157,7 @@ _vpaes_schedule_low_round:
 
 	vsldoi	1, 9, 7, 12
 	vxor	7, 7, 1
-	vspltisb	1, 0x0f
+	vspltisb	1,0x0f
 	vsldoi	4, 9, 7, 8
 
 
@@ -1253,7 +1253,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr	
@@ -1304,7 +1304,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr	
diff --git a/sys/crypto/openssl/powerpc64le/bn-ppc.S b/sys/crypto/openssl/powerpc64le/bn-ppc.S
new file mode 100644
index 0000000000000..50805d3d52bc7
--- /dev/null
+++ b/sys/crypto/openssl/powerpc64le/bn-ppc.S
@@ -0,0 +1,1876 @@
+/* Do not modify. This file is auto-generated from ppc.pl. */
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+.globl	bn_sqr_comba4
+.type	bn_sqr_comba4,@function
+.globl	bn_sqr_comba8
+.type	bn_sqr_comba8,@function
+.globl	bn_mul_comba4
+.type	bn_mul_comba4,@function
+.globl	bn_mul_comba8
+.type	bn_mul_comba8,@function
+.globl	bn_sub_words
+.type	bn_sub_words,@function
+.globl	bn_add_words
+.type	bn_add_words,@function
+.globl	bn_div_words
+.type	bn_div_words,@function
+.globl	bn_sqr_words
+.type	bn_sqr_words,@function
+.globl	bn_mul_words
+.type	bn_mul_words,@function
+.globl	bn_mul_add_words
+.type	bn_mul_add_words,@function
+
+
+
+.machine	"any"
+.abiversion	2
+.text
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba4:
+.localentry	bn_sqr_comba4,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	ld	5,0(4)
+	mulld	9,5,5
+	mulhdu	10,5,5
+
+
+
+
+	std	9,0(3)
+
+	ld	6,8(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,0
+
+
+	addc	10,7,10
+	addze	11,8
+	addze	9,9
+
+	std	10,8(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,16(3)
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	7,7,7
+	adde	8,8,8
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,24(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,32(3)
+
+	ld	5,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	7,7,7
+	adde	8,8,8
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,40(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+
+	std	9,48(3)
+	std	10,56(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba4,.-bn_sqr_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_sqr_comba8:
+.localentry	bn_sqr_comba8,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	ld	5,0(4)
+	mulld	9,5,5
+	mulhdu	10,5,5
+	std	9,0(3)
+
+	ld	6,8(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,0
+	addze	9,0
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	std	10,8(3)
+
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	std	11,16(3)
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	std	9,24(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,0(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,32(3)
+
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,8(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,16(4)
+	ld	6,24(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,40(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,8(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,0(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,48(3)
+
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,8(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,16(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,24(4)
+	ld	6,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,56(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,16(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,8(4)
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,64(3)
+
+	ld	5,16(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,24(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+
+	ld	5,32(4)
+	ld	6,40(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,72(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+
+	ld	5,24(4)
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,80(3)
+
+	ld	5,32(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,0
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+
+	ld	5,40(4)
+	ld	6,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	addc	11,7,11
+	adde	9,8,9
+	addze	10,10
+	std	11,88(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,0
+
+	ld	6,56(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	addc	9,7,9
+	adde	10,8,10
+	addze	11,11
+	std	9,96(3)
+
+
+	ld	5,48(4)
+	mulld	7,5,6
+	mulhdu	8,5,6
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,0
+	addc	10,7,10
+	adde	11,8,11
+	addze	9,9
+	std	10,104(3)
+
+	mulld	7,6,6
+	mulhdu	8,6,6
+	addc	11,7,11
+	adde	9,8,9
+	std	11,112(3)
+	std	9, 120(3)
+
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,2,0
+.long	0
+.size	bn_sqr_comba8,.-bn_sqr_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba4:
+.localentry	bn_mul_comba4,0
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+	ld	6,0(4)
+	ld	7,0(5)
+	mulld	10,6,7
+	mulhdu	11,6,7
+	std	10,0(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,0
+	addze	10,0
+
+	ld	6, 8(4)
+	ld	7, 0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	std	11,8(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+
+	ld	6,0(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	std	12,16(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,0
+
+	ld	6,8(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	ld	6,16(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+	addze	12,12
+	std	10,24(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,0
+
+	ld	6,16(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+
+	ld	6,8(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,8,11
+	adde	12,9,12
+	addze	10,10
+	std	11,32(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,0
+
+	ld	6,24(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,8,12
+	adde	10,9,10
+	addze	11,11
+	std	12,40(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,8,10
+	adde	11,9,11
+
+	std	10,48(3)
+	std	11,56(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba4,.-bn_mul_comba4
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_comba8:
+.localentry	bn_mul_comba8,0
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+	ld	6,0(4)
+	ld	7,0(5)
+	mulld	10,6,7
+	mulhdu	11,6,7
+	std	10,0(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	addze	12,9
+	addze	10,0
+
+	ld	6,8(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,8(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,0(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,16(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,8(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+
+	ld	6,16(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,24(3)
+
+	ld	6,32(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,24(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,16(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,8(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,0(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,32(3)
+
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,8(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,16(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,24(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,32(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,40(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,40(3)
+
+	ld	6,48(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,40(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,32(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,24(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,16(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,8(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,0(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,48(3)
+
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,8(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,16(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,24(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,32(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,40(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,48(4)
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,56(4)
+	ld	7,0(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,56(3)
+
+	ld	7,8(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,48(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,40(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,32(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,24(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,16(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,8(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,64(3)
+
+	ld	6,16(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,24(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,32(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,40(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,48(4)
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,56(4)
+	ld	7,16(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,72(3)
+
+	ld	7,24(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,48(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,40(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,32(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+
+	ld	6,24(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,80(3)
+
+	ld	6,32(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,0
+
+	ld	6,40(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,48(4)
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+
+	ld	6,56(4)
+	ld	7,32(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	addze	11,11
+	std	12,88(3)
+
+	ld	7,40(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,0
+
+	ld	6,48(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+
+	ld	6,40(4)
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	10,10,8
+	adde	11,11,9
+	addze	12,12
+	std	10,96(3)
+
+	ld	6,48(4)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,0
+
+	ld	6,56(4)
+	ld	7,48(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	11,11,8
+	adde	12,12,9
+	addze	10,10
+	std	11,104(3)
+
+	ld	7,56(5)
+	mulld	8,6,7
+	mulhdu	9,6,7
+	addc	12,12,8
+	adde	10,10,9
+	std	12,112(3)
+	std	10,120(3)
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_mul_comba8,.-bn_mul_comba8
+
+
+
+
+
+
+
+
+.align	4
+bn_sub_words:
+.localentry	bn_sub_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	subfc.	7,0,6
+
+
+	beq	.Lppcasm_sub_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	addi	5,5,-8
+	mtctr	6
+.Lppcasm_sub_mainloop:
+	ldu	7,8(4)
+	ldu	8,8(5)
+	subfe	6,8,7
+
+
+	stdu	6,8(3)
+	bdnz	.Lppcasm_sub_mainloop
+.Lppcasm_sub_adios:
+	subfze	3,0
+	andi.	3,3,1
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_sub_words,.-bn_sub_words
+
+
+
+
+
+
+
+
+.align	4
+bn_add_words:
+.localentry	bn_add_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+
+
+
+	addic.	6,6,0
+	beq	.Lppcasm_add_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	addi	5,5,-8
+	mtctr	6
+.Lppcasm_add_mainloop:
+	ldu	7,8(4)
+	ldu	8,8(5)
+	adde	8,7,8
+	stdu	8,8(3)
+	bdnz	.Lppcasm_add_mainloop
+.Lppcasm_add_adios:
+	addze	3,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_add_words,.-bn_add_words
+
+
+
+
+
+
+
+
+.align	4
+bn_div_words:
+.localentry	bn_div_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+	cmpldi	0,5,0
+	bne	.Lppcasm_div1
+	li	3,-1
+	blr	
+.Lppcasm_div1:
+	xor	0,0,0
+	li	8,64
+	cntlzd.	7,5
+	beq	.Lppcasm_div2
+	subf	8,7,8
+	srd.	9,3,8
+	td	16,9,0
+.Lppcasm_div2:
+	cmpld	0,3,5
+	blt	.Lppcasm_div3
+	subf	3,5,3
+.Lppcasm_div3:
+	cmpi	0,0,7,0
+	beq	.Lppcasm_div4
+	sld	3,3,7
+	srd	8,4,8
+	sld	5,5,7
+	or	3,3,8
+	sld	4,4,7
+.Lppcasm_div4:
+	srdi	9,5,32
+
+
+	li	6,2
+	mtctr	6
+.Lppcasm_divouterloop:
+	srdi	8,3,32
+	srdi	11,4,32
+
+	cmpld	0,8,9
+	bne	.Lppcasm_div5
+
+	li	8,-1
+	clrldi	8,8,32
+	b	.Lppcasm_div6
+.Lppcasm_div5:
+	divdu	8,3,9
+.Lppcasm_div6:
+	mulld	12,9,8
+	clrldi	10,5,32
+	mulld	6,8,10
+
+.Lppcasm_divinnerloop:
+	subf	10,12,3
+	srdi	7,10,32
+	addic.	7,7,0
+
+
+
+	sldi	7,10,32
+	or	7,7,11
+	cmpld	1,6,7
+	bne	.Lppcasm_divinnerexit
+	ble	1,.Lppcasm_divinnerexit
+	addi	8,8,-1
+	subf	12,9,12
+	clrldi	10,5,32
+	subf	6,10,6
+	b	.Lppcasm_divinnerloop
+.Lppcasm_divinnerexit:
+	srdi	10,6,32
+	sldi	11,6,32
+	cmpld	1,4,11
+	add	12,12,10
+	bge	1,.Lppcasm_div7
+	addi	12,12,1
+.Lppcasm_div7:
+	subf	11,11,4
+	cmpld	1,3,12
+	bge	1,.Lppcasm_div8
+	addi	8,8,-1
+	add	3,5,3
+.Lppcasm_div8:
+	subf	12,12,3
+	sldi	4,11,32
+
+
+
+	insrdi	11,12,32,32
+	rotldi	3,11,32
+	bdz	.Lppcasm_div9
+	sldi	0,8,32
+	b	.Lppcasm_divouterloop
+.Lppcasm_div9:
+	or	3,8,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_div_words,.-bn_div_words
+
+
+
+
+
+
+
+.align	4
+bn_sqr_words:
+.localentry	bn_sqr_words,0
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	addic.	5,5,0
+	beq	.Lppcasm_sqr_adios
+	addi	4,4,-8
+	addi	3,3,-8
+	mtctr	5
+.Lppcasm_sqr_mainloop:
+
+	ldu	6,8(4)
+	mulld	7,6,6
+	mulhdu	8,6,6
+	stdu	7,8(3)
+	stdu	8,8(3)
+	bdnz	.Lppcasm_sqr_mainloop
+.Lppcasm_sqr_adios:
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,3,0
+.long	0
+.size	bn_sqr_words,.-bn_sqr_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_words:
+.localentry	bn_mul_words,0
+
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_mw_REM
+	mtctr	7
+.Lppcasm_mw_LOOP:
+
+	ld	8,0(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+
+
+
+
+	std	9,0(3)
+
+	ld	8,8(4)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+
+	std	11,8(3)
+
+	ld	8,16(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	adde	9,9,12
+
+	std	9,16(3)
+
+	ld	8,24(4)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+
+	std	11,24(3)
+
+	addi	3,3,32
+	addi	4,4,32
+	bdnz	.Lppcasm_mw_LOOP
+
+.Lppcasm_mw_REM:
+	andi.	5,5,0x3
+	beq	.Lppcasm_mw_OVER
+
+	ld	8,0(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,0(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+
+	ld	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,8(3)
+	addi	12,10,0
+
+	addi	5,5,-1
+	cmpli	0,0,5,0
+	beq	.Lppcasm_mw_OVER
+
+
+	ld	8,16(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	std	9,16(3)
+	addi	12,10,0
+
+.Lppcasm_mw_OVER:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_words,.-bn_mul_words
+
+
+
+
+
+
+
+
+.align	4
+bn_mul_add_words:
+.localentry	bn_mul_add_words,0
+
+
+
+
+
+
+
+
+
+
+
+	xor	0,0,0
+	xor	12,12,12
+	rlwinm.	7,5,30,2,31
+	beq	.Lppcasm_maw_leftover
+	mtctr	7
+.Lppcasm_maw_mainloop:
+
+	ld	8,0(4)
+	ld	11,0(3)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	addc	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+
+
+
+
+
+	std	9,0(3)
+
+
+	ld	8,8(4)
+	ld	9,8(3)
+	mulld	11,6,8
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+
+	std	11,8(3)
+
+
+	ld	8,16(4)
+	mulld	9,6,8
+	ld	11,16(3)
+	mulhdu	10,6,8
+	adde	9,9,12
+	addze	10,10
+	addc	9,9,11
+
+	std	9,16(3)
+
+
+	ld	8,24(4)
+	mulld	11,6,8
+	ld	9,24(3)
+	mulhdu	12,6,8
+	adde	11,11,10
+	addze	12,12
+	addc	11,11,9
+	addze	12,12
+	std	11,24(3)
+	addi	3,3,32
+	addi	4,4,32
+	bdnz	.Lppcasm_maw_mainloop
+
+.Lppcasm_maw_leftover:
+	andi.	5,5,0x3
+	beq	.Lppcasm_maw_adios
+	addi	3,3,-8
+	addi	4,4,-8
+
+	mtctr	5
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+	bdz	.Lppcasm_maw_adios
+
+	ldu	8,8(4)
+	mulld	9,6,8
+	mulhdu	10,6,8
+	ldu	11,8(3)
+	addc	9,9,11
+	addze	10,10
+	addc	9,9,12
+	addze	12,10
+	std	9,0(3)
+
+.Lppcasm_maw_adios:
+	addi	3,12,0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.long	0
+.size	bn_mul_add_words,.-bn_mul_add_words
+.align	4
diff --git a/sys/crypto/openssl/powerpc64le/ecp_nistp521-ppc64.S b/sys/crypto/openssl/powerpc64le/ecp_nistp521-ppc64.S
new file mode 100644
index 0000000000000..5905180d168a6
--- /dev/null
+++ b/sys/crypto/openssl/powerpc64le/ecp_nistp521-ppc64.S
@@ -0,0 +1,354 @@
+/* Do not modify. This file is auto-generated from ecp_nistp521-ppc64.pl. */
+.machine	"any"
+.abiversion	2
+.text
+
+.globl	p521_felem_mul
+.type	p521_felem_mul,@function
+.align	5
+p521_felem_mul:
+.localentry	p521_felem_mul,0
+
+
+	mr	12,1
+	stdu	1,-16*13(1)
+
+	stxv	52,-16*12(12)
+	stxv	53,-16*11(12)
+	stxv	54,-16*10(12)
+	stxv	55,-16*9(12)
+	stxv	56,-16*8(12)
+	stxv	57,-16*7(12)
+	stxv	58,-16*6(12)
+	stxv	59,-16*5(12)
+	stxv	60,-16*4(12)
+	stxv	61,-16*3(12)
+	stxv	62,-16*2(12)
+	stxv	63,-16*1(12)
+
+	vspltisw	0,0
+
+	lxsd	13,0(4)
+	lxsd	14,8(4)
+	lxsd	15,16(4)
+	lxsd	16,24(4)
+	lxsd	17,32(4)
+	lxsd	18,40(4)
+	lxsd	19,48(4)
+	lxsd	20,56(4)
+	lxsd	21,64(4)
+
+	lxsd	3,0(5)
+	lxsd	4,8(5)
+	lxsd	5,16(5)
+	lxsd	6,24(5)
+	lxsd	7,32(5)
+	lxsd	8,40(5)
+	lxsd	9,48(5)
+	lxsd	10,56(5)
+	lxsd	11,64(5)
+
+	.long	0x12ED1823
+
+	xxpermdi	33,45,46,0b00
+	xxpermdi	34,36,35,0b00
+	.long	0x13011023
+
+	xxpermdi	34,37,36,0b00
+	.long	0x13211023
+	.long	0x132F1E63
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13411023
+	xxpermdi	44,47,48,0b00
+	xxpermdi	54,36,35,0b00
+	.long	0x134CB6A3
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13611023
+	xxpermdi	54,37,36,0b00
+	.long	0x136CB6E3
+	.long	0x13711EE3
+
+	xxpermdi	34,40,39,0b00
+	.long	0x13811023
+	xxpermdi	54,38,37,0b00
+	.long	0x138CB723
+
+	xxpermdi	34,41,40,0b00
+	.long	0x13A11023
+	xxpermdi	54,39,38,0b00
+	.long	0x13ACB763
+
+	xxpermdi	34,42,41,0b00
+	.long	0x13C11023
+	xxpermdi	54,40,39,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,43,42,0b00
+	.long	0x13E11023
+	xxpermdi	54,41,40,0b00
+	.long	0x13ECB7E3
+
+	xxpermdi	33,49,50,0b00
+	xxpermdi	34,36,35,0b00
+	.long	0x13811723
+
+	xxpermdi	34,37,36,0b00
+	.long	0x13A11763
+	.long	0x13B31F63
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13C117A3
+	xxpermdi	44,51,52,0b00
+	xxpermdi	54,36,35,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13E117E3
+	xxpermdi	54,37,36,0b00
+	.long	0x13ECB7E3
+	.long	0x13F51FE3
+
+	li	8,0
+	li	9,1
+	mtvsrdd	33,9,8
+	.long	0x10630DC4
+	.long	0x10840DC4
+	.long	0x10A50DC4
+	.long	0x10C60DC4
+	.long	0x10E70DC4
+	.long	0x11080DC4
+	.long	0x11290DC4
+	.long	0x114A0DC4
+	.long	0x116B0DC4
+
+	.long	0x13D55FA3
+
+	xxpermdi	34,43,42,0b00
+	xxpermdi	33,52,53,0b00
+	.long	0x13A11763
+
+	xxpermdi	33,51,52,0b00
+	.long	0x13811723
+	.long	0x13954F23
+
+	xxpermdi	33,50,51,0b00
+	.long	0x136116E3
+	xxpermdi	54,41,40,0b00
+	xxpermdi	44,52,53,0b00
+	.long	0x136CB6E3
+
+	xxpermdi	33,49,50,0b00
+	.long	0x134116A3
+	xxpermdi	44,51,52,0b00
+	.long	0x134CB6A3
+	.long	0x13553EA3
+
+	xxpermdi	33,48,49,0b00
+	.long	0x13211663
+	xxpermdi	44,50,51,0b00
+	.long	0x132CB663
+
+	xxpermdi	33,47,48,0b00
+	.long	0x13011623
+	xxpermdi	44,49,50,0b00
+	.long	0x130CB623
+
+	xxpermdi	33,46,47,0b00
+	.long	0x12E115E3
+	xxpermdi	44,48,49,0b00
+	.long	0x12ECB5E3
+
+	xxpermdi	34,39,38,0b00
+	xxpermdi	33,52,53,0b00
+	.long	0x13211663
+
+	xxpermdi	33,51,52,0b00
+	.long	0x13011623
+	.long	0x13152E23
+
+	xxpermdi	33,50,51,0b00
+	.long	0x12E115E3
+	xxpermdi	54,37,36,0b00
+	xxpermdi	44,52,53,0b00
+	.long	0x12ECB5E3
+
+	stxv	55,0(3)
+	stxv	56,16(3)
+	stxv	57,32(3)
+	stxv	58,48(3)
+	stxv	59,64(3)
+	stxv	60,80(3)
+	stxv	61,96(3)
+	stxv	62,112(3)
+	stxv	63,128(3)
+
+	ld	12,0(1)
+	lxv	52,-16*12(12)
+	lxv	53,-16*11(12)
+	lxv	54,-16*10(12)
+	lxv	55,-16*9(12)
+	lxv	56,-16*8(12)
+	lxv	57,-16*7(12)
+	lxv	58,-16*6(12)
+	lxv	59,-16*5(12)
+	lxv	60,-16*4(12)
+	lxv	61,-16*3(12)
+	lxv	62,-16*2(12)
+	lxv	63,-16*1(12)
+	mr	1,12
+
+	blr	
+.size	p521_felem_mul,.-p521_felem_mul
+
+.globl	p521_felem_square
+.type	p521_felem_square,@function
+.align	5
+p521_felem_square:
+.localentry	p521_felem_square,0
+
+
+	mr	12,1
+	stdu	1,-16*13(1)
+
+	stxv	52,-16*12(12)
+	stxv	53,-16*11(12)
+	stxv	54,-16*10(12)
+	stxv	55,-16*9(12)
+	stxv	56,-16*8(12)
+	stxv	57,-16*7(12)
+	stxv	58,-16*6(12)
+	stxv	59,-16*5(12)
+	stxv	60,-16*4(12)
+	stxv	61,-16*3(12)
+	stxv	62,-16*2(12)
+	stxv	63,-16*1(12)
+
+	vspltisw	0,0
+
+	lxsd	13,0(4)
+	lxsd	14,8(4)
+	lxsd	15,16(4)
+	lxsd	16,24(4)
+	lxsd	17,32(4)
+	lxsd	18,40(4)
+	lxsd	19,48(4)
+	lxsd	20,56(4)
+	lxsd	21,64(4)
+
+	li	8,0
+	li	9,1
+	mtvsrdd	33,9,8
+	.long	0x106D0DC4
+	.long	0x108E0DC4
+	.long	0x10AF0DC4
+	.long	0x10D00DC4
+	.long	0x10F10DC4
+	.long	0x11120DC4
+	.long	0x11330DC4
+	.long	0x11540DC4
+	.long	0x11750DC4
+	.long	0x12ED6823
+
+	.long	0x130D2023
+
+	xxpermdi	33,45,46,0b00
+	xxpermdi	34,37,46,0b00
+	.long	0x13211023
+
+	xxpermdi	34,38,37,0b00
+	.long	0x13411023
+
+	xxpermdi	34,39,38,0b00
+	.long	0x13611023
+	.long	0x136F7EE3
+
+	xxpermdi	34,40,39,0b00
+	.long	0x13811023
+	.long	0x138F3723
+
+	xxpermdi	34,41,40,0b00
+	.long	0x13A11023
+	xxpermdi	44,47,48,0b00
+	xxpermdi	54,39,48,0b00
+	.long	0x13ACB763
+
+	xxpermdi	34,42,41,0b00
+	.long	0x13C11023
+	xxpermdi	54,40,39,0b00
+	.long	0x13CCB7A3
+
+	xxpermdi	34,43,42,0b00
+	.long	0x13E11023
+	xxpermdi	54,41,40,0b00
+	.long	0x13ECB7E3
+	.long	0x13F18FE3
+
+	.long	0x13124623
+
+	.long	0x13534EA3
+
+	.long	0x13945723
+
+	.long	0x13D55FA3
+
+	mtvsrdd	33,9,8
+	.long	0x11080DC4
+	.long	0x11290DC4
+	.long	0x114A0DC4
+	.long	0x116B0DC4
+
+	.long	0x13B45F63
+
+	.long	0x13935F23
+
+	xxpermdi	34,43,42,0b00
+	xxpermdi	33,50,51,0b00
+	.long	0x136116E3
+
+	xxpermdi	33,49,50,0b00
+	.long	0x134116A3
+
+	xxpermdi	33,48,49,0b00
+	.long	0x13211663
+	.long	0x13324E63
+
+	xxpermdi	33,47,48,0b00
+	.long	0x13011623
+	.long	0x13114E23
+
+	xxpermdi	33,46,47,0b00
+	.long	0x12E115E3
+	xxpermdi	34,41,40,0b00
+	xxpermdi	33,48,49,0b00
+	.long	0x12E115E3
+
+	stxv	55,0(3)
+	stxv	56,16(3)
+	stxv	57,32(3)
+	stxv	58,48(3)
+	stxv	59,64(3)
+	stxv	60,80(3)
+	stxv	61,96(3)
+	stxv	62,112(3)
+	stxv	63,128(3)
+
+	ld	12,0(1)
+	lxv	52,-16*12(12)
+	lxv	53,-16*11(12)
+	lxv	54,-16*10(12)
+	lxv	55,-16*9(12)
+	lxv	56,-16*8(12)
+	lxv	57,-16*7(12)
+	lxv	58,-16*6(12)
+	lxv	59,-16*5(12)
+	lxv	60,-16*4(12)
+	lxv	61,-16*3(12)
+	lxv	62,-16*2(12)
+	lxv	63,-16*1(12)
+	mr	1,12
+
+	blr	
+.size	p521_felem_square,.-p521_felem_square
+
diff --git a/sys/crypto/openssl/powerpc64le/keccak1600-ppc64.S b/sys/crypto/openssl/powerpc64le/keccak1600-ppc64.S
index 1fb70dd39b256..3451068190b8e 100644
--- a/sys/crypto/openssl/powerpc64le/keccak1600-ppc64.S
+++ b/sys/crypto/openssl/powerpc64le/keccak1600-ppc64.S
@@ -304,19 +304,19 @@ KeccakF1600:
 dword_le_load:
 .localentry	dword_le_load,0
 
-	lbzu	0,1(3)
-	lbzu	4,1(3)
-	lbzu	5,1(3)
+	lbz	0,1(3)
+	lbz	4,2(3)
+	lbz	5,3(3)
 	insrdi	0,4,8,48
-	lbzu	4,1(3)
+	lbz	4,4(3)
 	insrdi	0,5,8,40
-	lbzu	5,1(3)
+	lbz	5,5(3)
 	insrdi	0,4,8,32
-	lbzu	4,1(3)
+	lbz	4,6(3)
 	insrdi	0,5,8,24
-	lbzu	5,1(3)
+	lbz	5,7(3)
 	insrdi	0,4,8,16
-	lbzu	4,1(3)
+	lbzu	4,8(3)
 	insrdi	0,5,8,8
 	insrdi	0,4,8,0
 	blr	
@@ -579,21 +579,21 @@ SHA3_squeeze:
 	cmpldi	30,8
 	blt	.Lsqueeze_tail
 
-	stbu	0,1(29)
+	stb	0,1(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,2(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,3(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,4(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,5(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,6(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stb	0,7(29)
 	srdi	0,0,8
-	stbu	0,1(29)
+	stbu	0,8(29)
 
 	subic.	30,30,8
 	beq	.Lsqueeze_done
diff --git a/sys/crypto/openssl/powerpc64le/poly1305-ppc.S b/sys/crypto/openssl/powerpc64le/poly1305-ppc.S
index 7ffdb4a3d39b1..69862b94b2c3b 100644
--- a/sys/crypto/openssl/powerpc64le/poly1305-ppc.S
+++ b/sys/crypto/openssl/powerpc64le/poly1305-ppc.S
@@ -12,6 +12,7 @@ poly1305_init_int:
 	std	0,0(3)
 	std	0,8(3)
 	std	0,16(3)
+	stw	0,24(3)
 
 	cmpld	4,0
 	beq-	.Lno_key
@@ -41,6 +42,7 @@ poly1305_init_int:
 poly1305_blocks:
 .localentry	poly1305_blocks,0
 
+.Lpoly1305_blocks:
 	srdi.	5,5,4
 	beq-	.Labort
 
@@ -124,39 +126,1003 @@ poly1305_blocks:
 .long	0
 .byte	0,12,4,1,0x80,5,4,0
 .size	poly1305_blocks,.-poly1305_blocks
-
 .globl	poly1305_emit
 .type	poly1305_emit,@function
-.align	4
+.align	5
 poly1305_emit:
 .localentry	poly1305_emit,0
 
-	ld	7,0(3)
-	ld	8,8(3)
-	ld	9,16(3)
-	ld	6,0(5)
-	ld	5,8(5)
+	lwz	7,0(3)
+	lwz	8,4(3)
+	lwz	9,8(3)
+	lwz	10,12(3)
+	lwz	11,16(3)
+	lwz	0,24(3)
+
+	sldi	8,8,26
+	sldi	12,9,52
+	srdi	9,9,12
+	sldi	10,10,14
+	add	7,7,8
+	addc	7,7,12
+	sldi	12,11,40
+	srdi	11,11,24
+	adde	8,9,10
+	addc	8,8,12
+	addze	9,11
+
+	ld	10,0(3)
+	ld	11,8(3)
+	ld	12,16(3)
+
+	neg	0,0
+	xor	7,7,10
+	xor	8,8,11
+	xor	9,9,12
+	and	7,7,0
+	and	8,8,0
+	and	9,9,0
+	xor	7,7,10
+	xor	8,8,11
+	xor	9,9,12
 
 	addic	10,7,5
 	addze	11,8
 	addze	12,9
 
-	srdi	0,12,2
-	neg	0,0
+	srdi	12,12,2
+	neg	12,12
 
-	andc	7,7,0
-	and	10,10,0
-	andc	8,8,0
-	and	11,11,0
+	andc	7,7,12
+	and	10,10,12
+	andc	8,8,12
+	and	11,11,12
 	or	7,7,10
 	or	8,8,11
-	addc	7,7,6
-	adde	8,8,5
-	std	7,0(4)
-	std	8,8(4)
+
+	lwz	12,4(5)
+	lwz	9,12(5)
+	lwz	10,0(5)
+	lwz	11,8(5)
+
+	insrdi	10,12,32,0
+	insrdi	11,9,32,0
+
+	addc	7,7,10
+	adde	8,8,11
+
+	addi	3,4,-1
+	addi	4,4,7
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	srdi	7,7,8
+	stbu	8,1(4)
+	srdi	8,8,8
+
+	stbu	7,1(3)
+	stbu	8,1(4)
+
 	blr	
 .long	0
 .byte	0,12,0x14,0,0,0,3,0
 .size	poly1305_emit,.-poly1305_emit
-.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.globl	poly1305_blocks_vsx
+.type	poly1305_blocks_vsx,@function
+.align	5
+poly1305_blocks_vsx:
+.localentry	poly1305_blocks_vsx,0
+
+	lwz	7,24(3)
+	cmpldi	5,128
+	bge	__poly1305_blocks_vsx
+
+	neg	0,7
+	lwz	7,0(3)
+	lwz	8,4(3)
+	lwz	9,8(3)
+	lwz	10,12(3)
+	lwz	11,16(3)
+
+	sldi	8,8,26
+	sldi	12,9,52
+	add	7,7,8
+	srdi	9,9,12
+	sldi	10,10,14
+	addc	7,7,12
+	sldi	8,11,40
+	adde	9,9,10
+	srdi	11,11,24
+	addc	9,9,8
+	addze	11,11
+
+	ld	8,0(3)
+	ld	10,8(3)
+	ld	12,16(3)
+
+	xor	7,7,8
+	xor	9,9,10
+	xor	11,11,12
+	and	7,7,0
+	and	9,9,0
+	and	11,11,0
+	xor	7,7,8
+	xor	9,9,10
+	xor	11,11,12
+
+	li	0,0
+	std	7,0(3)
+	std	9,8(3)
+	std	11,16(3)
+	stw	0,24(3)
+
+	b	.Lpoly1305_blocks
+.long	0
+.byte	0,12,0x14,0,0,0,4,0
+.size	poly1305_blocks_vsx,.-poly1305_blocks_vsx
+
+.align	5
+__poly1305_mul:
+	mulld	9,6,27
+	mulhdu	10,6,27
+
+	mulld	30,7,29
+	mulhdu	31,7,29
+	addc	9,9,30
+	adde	10,10,31
+
+	mulld	30,6,28
+	mulhdu	11,6,28
+	addc	10,10,30
+	addze	11,11
+
+	mulld	30,7,27
+	mulhdu	31,7,27
+	addc	10,10,30
+	adde	11,11,31
+
+	mulld	30,8,29
+	mulld	31,8,27
+	addc	10,10,30
+	adde	11,11,31
+
+	andc	30,11,0
+	and	8,11,0
+	srdi	31,30,2
+	add	30,30,31
+	addc	6,9,30
+	addze	7,10
+	addze	8,8
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.size	__poly1305_mul,.-__poly1305_mul
+
+.align	5
+__poly1305_splat:
+	rldicl	9,6,0,38
+	rldicl	10,6,38,38
+	stw	9,0x00(31)
+
+	rldicl	11,6,12,52
+	slwi	9,10,2
+	stw	10,0x10(31)
+	add	9,9,10
+	stw	9,0x20(31)
+
+	insrdi	11,7,14,38
+	slwi	9,11,2
+	stw	11,0x30(31)
+	add	9,9,11
+	stw	9,0x40(31)
+
+	rldicl	10,7,50,38
+	rldicl	11,7,24,40
+	slwi	9,10,2
+	stw	10,0x50(31)
+	add	9,9,10
+	stw	9,0x60(31)
+
+	insrdi	11,8,3,37
+	slwi	9,11,2
+	stw	11,0x70(31)
+	add	9,9,11
+	stw	9,0x80(31)
+
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.size	__poly1305_splat,.-__poly1305_splat
+
+.align	5
+__poly1305_blocks_vsx:
+	stdu	1,-432(1)
+	mflr	0
+	li	10,191
+	li	11,207
+	li	12,-1
+	stvx	20,10,1
+	addi	10,10,32
+	stvx	21,11,1
+	addi	11,11,32
+	stvx	22,10,1
+	addi	10,10,32
+	stvx	23,10,1
+	addi	10,10,32
+	stvx	24,11,1
+	addi	11,11,32
+	stvx	25,10,1
+	addi	10,10,32
+	stvx	26,10,1
+	addi	10,10,32
+	stvx	27,11,1
+	addi	11,11,32
+	stvx	28,10,1
+	addi	10,10,32
+	stvx	29,11,1
+	addi	11,11,32
+	stvx	30,10,1
+	stvx	31,11,1
+	stw	12,388(1)
+	li	12,-1
+	or	12,12,12
+	std	27,392(1)
+	std	28,400(1)
+	std	29,408(1)
+	std	30,416(1)
+	std	31,424(1)
+	std	0,448(1)
+
+	bl	.LPICmeup
+
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	.long	0x7FA06699
+	.long	0x7F3B6699
+	.long	0x7F7C6699
+	.long	0x7FFD6699
+	.long	0x7FDE6699
+
+	cmplwi	7,0
+	bne	.Lskip_init_vsx
+
+	ld	27,32(3)
+	ld	28,40(3)
+	srdi	29,28,2
+	li	0,3
+	add	29,29,28
+
+	mr	6,27
+	mr	7,28
+	li	8,0
+	addi	31,3,60
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,52
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,56
+	bl	__poly1305_splat
+
+	bl	__poly1305_mul
+	addi	31,3,48
+	bl	__poly1305_splat
+
+	ld	6,0(3)
+	ld	7,8(3)
+	ld	8,16(3)
+
+	rldicl	9,6,0,38
+	rldicl	10,6,38,38
+	rldicl	11,6,12,52
+	.long	0x7C0901E7
+	insrdi	11,7,14,38
+	.long	0x7C2A01E7
+	rldicl	10,7,50,38
+	.long	0x7C4B01E7
+	rldicl	11,7,24,40
+	.long	0x7C6A01E7
+	insrdi	11,8,3,37
+	.long	0x7C8B01E7
+	li	0,1
+	stw	0,24(3)
+	b	.Loaded_vsx
+
+.align	4
+.Lskip_init_vsx:
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001819
+	.long	0x7C3B1819
+	.long	0x7C5C1819
+	.long	0x7C7D1819
+	.long	0x7C9E1819
+
+.Loaded_vsx:
+	li	27,0x10
+	li	28,0x20
+	li	29,0x30
+	li	30,0x40
+	li	31,0x50
+	li	7,0x60
+	li	8,0x70
+	addi	10,3,64
+	addi	11,1,63
+
+	vxor	20,20,20
+	.long	0xF000A057
+	.long	0xF021A057
+	.long	0xF042A057
+	.long	0xF063A057
+	.long	0xF084A057
+
+
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+
+
+
+
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	vspltisb	28,14
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x1295CEC4
+	.long	0x12D6D6C4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+	vand	21,21,29
+	vand	20,20,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10D4368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x7D5D1A99
+	.long	0x7D605299
+	.long	0x7D9B5299
+	.long	0x7DBC5299
+	.long	0x7DDD5299
+	.long	0x7EBE5299
+	.long	0x7EDF5299
+	.long	0x7EE75299
+	.long	0x7F085299
+	stvx	11,0,11
+	stvx	12,27,11
+	stvx	13,28,11
+	stvx	14,29,11
+	stvx	21,30,11
+	stvx	22,31,11
+	stvx	23,7,11
+	stvx	24,8,11
+
+	addi	4,4,0x40
+	addi	12,12,0x50
+	addi	0,5,-64
+	srdi	0,0,6
+	mtctr	0
+	b	.Loop_vsx
+
+.align	4
+.Loop_vsx:
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	.long	0x11E55288
+	.long	0x12055A88
+	.long	0x12256A88
+	.long	0x12466A88
+
+	.long	0x12865288
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12676A88
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	lvx	12,31,11
+	.long	0x12885A88
+	.long	0x1273A0C0
+	lvx	11,30,11
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12875288
+	.long	0x1231A0C0
+	.long	0x12885288
+	.long	0x1252A0C0
+	lvx	14,8,11
+	.long	0x12895288
+	.long	0x1273A0C0
+	lvx	13,7,11
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+
+
+
+
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0xF2B7C057
+	vperm	22,23,24,31
+	.long	0xF2F7C357
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	lvx	12,27,11
+	.long	0x12815888
+	.long	0x1273A0C0
+	lvx	11,0,11
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	lvx	14,29,11
+	.long	0x12806888
+	.long	0x1273A0C0
+	lvx	13,28,11
+
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12D6D6C4
+	.long	0x1355CEC4
+	.long	0x1317DEC4
+	.long	0x12F7E6C4
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+	vand	21,21,29
+	vand	26,26,29
+	vand	22,22,29
+	vand	23,23,29
+
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x11384E8C
+	.long	0x10B52E8C
+	.long	0x10DA368C
+	.long	0x10F63E8C
+	.long	0x1117468C
+	vor	9,9,30
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	addi	4,4,0x40
+	bdnz	.Loop_vsx
+
+	neg	5,5
+	andi.	5,5,0x30
+	sub	4,4,5
+
+	.long	0x7D5D1E99
+	.long	0x7D605699
+	.long	0x7D9B5699
+	.long	0x7DBC5699
+	.long	0x7DDD5699
+
+.Last_vsx:
+	.long	0x11E55288
+	.long	0x12065288
+	.long	0x12275288
+	.long	0x12485288
+	.long	0x12695288
+
+	.long	0x12896288
+	.long	0x11EFA0C0
+	.long	0x12855A88
+	.long	0x1210A0C0
+	.long	0x12865A88
+	.long	0x1231A0C0
+	.long	0x12875A88
+	.long	0x1252A0C0
+	.long	0x7D9F5699
+	.long	0x12885A88
+	.long	0x1273A0C0
+	.long	0x7D7E5699
+
+	.long	0x104238C0
+	.long	0x100028C0
+	.long	0x106340C0
+	.long	0x102130C0
+	.long	0x108448C0
+
+	.long	0x12887288
+	.long	0x11EFA0C0
+	.long	0x12897288
+	.long	0x1210A0C0
+	.long	0x12856A88
+	.long	0x1231A0C0
+	.long	0x12866A88
+	.long	0x1252A0C0
+	.long	0x7DC85699
+	.long	0x12876A88
+	.long	0x1273A0C0
+	.long	0x7DA75699
+
+	.long	0x12876288
+	.long	0x11EFA0C0
+	.long	0x12886288
+	.long	0x1210A0C0
+	.long	0x12896288
+	.long	0x1231A0C0
+	.long	0x12855A88
+	.long	0x1252A0C0
+	.long	0x12865A88
+	.long	0x1273A0C0
+
+	.long	0x12867288
+	.long	0x11EFA0C0
+	.long	0x12877288
+	.long	0x1210A0C0
+	.long	0x12887288
+	.long	0x1231A0C0
+	.long	0x12897288
+	.long	0x1252A0C0
+	.long	0x12856A88
+	.long	0x1273A0C0
+
+
+	.long	0x12805088
+	.long	0x11EFA0C0
+	.long	0x12815088
+	.long	0x1210A0C0
+	.long	0x12825088
+	.long	0x1231A0C0
+	.long	0x12835088
+	.long	0x1252A0C0
+	.long	0x12845088
+	.long	0x1273A0C0
+
+	.long	0x12826088
+	.long	0x11EFA0C0
+	.long	0x12836088
+	.long	0x1210A0C0
+	.long	0x12846088
+	.long	0x1231A0C0
+	.long	0x12805888
+	.long	0x1252A0C0
+	.long	0x7D9B5699
+	.long	0x12815888
+	.long	0x1273A0C0
+	.long	0x7D605699
+
+	.long	0x12817088
+	.long	0x11EFA0C0
+	.long	0x12827088
+	.long	0x1210A0C0
+	.long	0x12837088
+	.long	0x1231A0C0
+	.long	0x12847088
+	.long	0x1252A0C0
+	.long	0x7DDD5699
+	.long	0x12806888
+	.long	0x1273A0C0
+	.long	0x7DBC5699
+
+	.long	0x12846088
+	.long	0x11EFA0C0
+	.long	0x12805888
+	.long	0x1210A0C0
+	.long	0x12815888
+	.long	0x1231A0C0
+	.long	0x12825888
+	.long	0x1252A0C0
+	.long	0x12835888
+	.long	0x1273A0C0
+
+	.long	0x12837088
+	.long	0x11EFA0C0
+	.long	0x12847088
+	.long	0x1210A0C0
+	.long	0x12806888
+	.long	0x1231A0C0
+	.long	0x12816888
+	.long	0x1252A0C0
+	.long	0x12826888
+	.long	0x1273A0C0
+
+
+
+
+	.long	0xF00F7A57
+	.long	0xF0308257
+	.long	0xF0518A57
+	.long	0xF0729257
+	.long	0xF0939A57
+	.long	0x11EF00C0
+	.long	0x121008C0
+	.long	0x123110C0
+	.long	0x125218C0
+	.long	0x127320C0
+
+
+
+
+	vspltisb	20,2
+	.long	0x1092CEC4
+	.long	0x102FCEC4
+	vand	3,18,29
+	vand	0,15,29
+	.long	0x108498C0
+	.long	0x102180C0
+
+	.long	0x1264CEC4
+	.long	0x1201CEC4
+	vand	4,4,29
+	vand	1,1,29
+	.long	0x100098C0
+	.long	0x105180C0
+
+	.long	0x1273A5C4
+	.long	0x1222CEC4
+	vand	2,2,29
+	.long	0x100098C0
+	.long	0x106388C0
+
+	.long	0x11E0CEC4
+	.long	0x1243CEC4
+	vand	0,0,29
+	vand	3,3,29
+	.long	0x102178C0
+	.long	0x108490C0
+
+	beq	.Ldone_vsx
+
+	add	6,12,5
+
+
+	.long	0x7EA02699
+	.long	0x7EDB2699
+	.long	0x7EFC2699
+	.long	0x7F1D2699
+
+
+
+
+
+	.long	0xF0B5B057
+	vspltisb	26,4
+	vperm	7,21,22,31
+	.long	0xF115B357
+
+	.long	0x10C5CEC4
+	.long	0x10E7D6C4
+	.long	0x1128DEC4
+	.long	0x1108E6C4
+	vand	5,5,29
+	vand	6,6,29
+	vand	7,7,29
+	vand	8,8,29
+
+	.long	0xF297C057
+	vperm	21,23,24,31
+	.long	0xF2D7C357
+
+	.long	0x7DE03699
+	.long	0x7E1D3699
+
+	.long	0x12F4CEC4
+	.long	0x12B5D6C4
+	.long	0x1316DEC4
+	.long	0x12D6E6C4
+	vand	20,20,29
+	vand	23,23,29
+	vand	21,21,29
+	vand	22,22,29
+
+
+	.long	0x11384E8C
+	.long	0x10B42E8C
+	.long	0x10D7368C
+	.long	0x10F53E8C
+	.long	0x1116468C
+	vor	9,9,30
+
+	vperm	0,0,0,15
+	vand	5,5,    16
+	vperm	1,1,1,15
+	vand	6,6,    16
+	vperm	2,2,2,15
+	vand	7,7,    16
+	vperm	3,3,3,15
+	vand	8,8,    16
+	vperm	4,4,4,15
+	vand	9,9,    16
+
+	.long	0x10A500C0
+	vxor	0,0,0
+	.long	0x10C608C0
+	vxor	1,1,1
+	.long	0x10E710C0
+	vxor	2,2,2
+	.long	0x110818C0
+	vxor	3,3,3
+	.long	0x112920C0
+	vxor	4,4,4
+
+	xor.	5,5,5
+	b	.Last_vsx
+
+.align	4
+.Ldone_vsx:
+	ld	0,448(1)
+	li	27,4
+	li	28,8
+	li	29,12
+	li	30,16
+	.long	0x7C001919
+	.long	0x7C3B1919
+	.long	0x7C5C1919
+	.long	0x7C7D1919
+	.long	0x7C9E1919
+
+	lwz	12,388(1)
+	mtlr	0
+	li	10,191
+	li	11,207
+	or	12,12,12
+	lvx	20,10,1
+	addi	10,10,32
+	lvx	21,10,1
+	addi	10,10,32
+	lvx	22,11,1
+	addi	11,11,32
+	lvx	23,10,1
+	addi	10,10,32
+	lvx	24,11,1
+	addi	11,11,32
+	lvx	25,10,1
+	addi	10,10,32
+	lvx	26,11,1
+	addi	11,11,32
+	lvx	27,10,1
+	addi	10,10,32
+	lvx	28,11,1
+	addi	11,11,32
+	lvx	29,10,1
+	addi	10,10,32
+	lvx	30,11,1
+	lvx	31,10,1
+	ld	27,392(1)
+	ld	28,400(1)
+	ld	29,408(1)
+	ld	30,416(1)
+	ld	31,424(1)
+	addi	1,1,432
+	blr	
+.long	0
+.byte	0,12,0x04,1,0x80,5,4,0
+.long	0
+.size	__poly1305_blocks_vsx,.-__poly1305_blocks_vsx
+
+.align	6
+.LPICmeup:
+	mflr	0
+	bcl	20,31,$+4
+	mflr	12
+	addi	12,12,56
+	mtlr	0
+	blr	
+.long	0
+.byte	0,12,0x14,0,0,0,0,0
+.space	28
+
+.long	0x03ffffff,0x00000000
+.long	0x03ffffff,0x00000000
+.long	0x0000001a,0x00000000
+.long	0x0000001a,0x00000000
+.long	0x00000028,0x00000000
+.long	0x00000028,0x00000000
+.long	0x0e0f0001,0x00000000
+.long	0x1e1f1011,0x00000000
+.long	0x01000000,0x01000000
+.long	0x01000000,0x01000000
+.long	0x03020100,0x07060504
+.long	0x0b0a0908,0x0f0e0d0c
+
+.long	0x00000000,0x00000000
+.long	0x04050607,0x00000000
+.long	0x00000000,0x04050607
+.long	0x00000000,0x00000000
+.long	0x00000000,0x00000000
+.long	0x00000000,0x04050607
+
+.long	0x00000000,0xffffffff
+.long	0xffffffff,0xffffffff
+.long	0x00000000,0xffffffff
+.long	0x00000000,0xffffffff
+.long	0x00000000,0x00000000
+.long	0x00000000,0xffffffff
+.byte	80,111,108,121,49,51,48,53,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,64,100,111,116,45,97,115,109,0
 .align	2
diff --git a/sys/crypto/openssl/powerpc64le/vpaes-ppc.S b/sys/crypto/openssl/powerpc64le/vpaes-ppc.S
index 718232627e6d4..751f67c33c565 100644
--- a/sys/crypto/openssl/powerpc64le/vpaes-ppc.S
+++ b/sys/crypto/openssl/powerpc64le/vpaes-ppc.S
@@ -674,7 +674,7 @@ vpaes_cbc_encrypt:
 	vor	24,0,0
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -726,7 +726,7 @@ vpaes_cbc_encrypt:
 	vor	24,25,25
 	sub.	30, 30, 0
 	vperm	0, 0, 0, 29
-	vsel	1, 28, 0, 30
+	vsel	1,28,0,30
 	vor	28,0,0
 	stvx	1, 0, 4
 	addi	4, 4, 16
@@ -1044,7 +1044,7 @@ _vpaes_schedule_core:
 
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1066,7 +1066,7 @@ _vpaes_schedule_core:
 	addi	9, 5, -15
 	vperm	0, 0, 0, 29
 	li	10, 4
-	vsel	2, 28, 0, 30
+	vsel	2,28,0,30
 	li	11, 8
 	stvx	2, 0, 5
 	li	12, 12
@@ -1157,7 +1157,7 @@ _vpaes_schedule_low_round:
 
 	vsldoi	1, 7, 9, 16-12
 	vxor	7, 7, 1
-	vspltisb	1, 0x0f
+	vspltisb	1,0x0f
 	vsldoi	4, 7, 9, 16-8
 
 
@@ -1253,7 +1253,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr	
@@ -1304,7 +1304,7 @@ _vpaes_schedule_mangle:
 
 
 	vperm	1, 3, 3, 29
-	vsel	2, 28, 1, 30
+	vsel	2,28,1,30
 	vor	28,1,1
 	stvx	2, 0, 5
 	blr