diff --git a/Gopkg.lock b/Gopkg.lock
index 73e2350b..bc2364e2 100644
--- a/Gopkg.lock
+++ b/Gopkg.lock
@@ -23,7 +23,7 @@
   branch = "master"
   name = "github.com/aead/chacha20"
   packages = ["chacha"]
-  revision = "0a02eb8dd59b95937590a847165c4d0b97032422"
+  revision = "8457f65383c5be6183d33e992fbf1786d6ab3e76"
 
 [[projects]]
   branch = "master"
diff --git a/vendor/github.com/aead/chacha20/chacha/chachaAVX2_amd64.s b/vendor/github.com/aead/chacha20/chacha/chachaAVX2_amd64.s
index 291da7ff..c2b5f52f 100644
--- a/vendor/github.com/aead/chacha20/chacha/chachaAVX2_amd64.s
+++ b/vendor/github.com/aead/chacha20/chacha/chachaAVX2_amd64.s
@@ -21,10 +21,10 @@
 TEXT ·xorKeyStreamAVX2(SB), 4, $320-80
 	MOVQ dst_base+0(FP), DI
 	MOVQ src_base+24(FP), SI
-	MOVQ src_len+32(FP), CX
 	MOVQ block+48(FP), BX
 	MOVQ state+56(FP), AX
 	MOVQ rounds+64(FP), DX
+	MOVQ src_len+32(FP), CX
 
 	MOVQ SP, R8
 	ADDQ $32, SP
diff --git a/vendor/github.com/aead/chacha20/chacha/chacha_386.s b/vendor/github.com/aead/chacha20/chacha/chacha_386.s
index 934e69b2..a2658c5e 100644
--- a/vendor/github.com/aead/chacha20/chacha/chacha_386.s
+++ b/vendor/github.com/aead/chacha20/chacha/chacha_386.s
@@ -47,7 +47,7 @@ TEXT ·supportsSSSE3(SB), NOSPLIT, $0-1
 #define Dst DI
 #define Nonce AX
 #define Key BX
-#define Rounds CX
+#define Rounds DX
 
 // func hChaCha20SSE2(out *[32]byte, nonce *[16]byte, key *[32]byte)
 TEXT ·hChaCha20SSE2(SB), 4, $0-12
@@ -117,9 +117,9 @@ chacha_loop:
 TEXT ·xorKeyStreamSSE2(SB), 4, $0-40
 	MOVL dst_base+0(FP), Dst
 	MOVL src_base+12(FP), Src
-	MOVL src_len+16(FP), Len
 	MOVL state+28(FP), State
 	MOVL rounds+32(FP), Rounds
+	MOVL src_len+16(FP), Len
 
 	MOVOU 0*16(State), X0
 	MOVOU 1*16(State), X1
@@ -199,9 +199,9 @@ DONE:
 TEXT ·xorKeyStreamSSSE3(SB), 4, $80-40
 	MOVL dst_base+0(FP), Dst
 	MOVL src_base+12(FP), Src
-	MOVL src_len+16(FP), Len
 	MOVL state+28(FP), Tmp0
 	MOVL rounds+32(FP), Rounds
+	MOVL src_len+16(FP), Len
 
 	MOVL Stack, Tmp2 // save stack pointer
 	ADDL $16, Stack  // ensure 16 byte stack alignment
@@ -211,6 +211,7 @@ TEXT ·xorKeyStreamSSSE3(SB), 4, $80-40
 	MOVOU 1*16(Tmp0), X1
 	MOVOU 2*16(Tmp0), X2
 	MOVOU 3*16(Tmp0), X3
+
 	TESTL Len, Len
 	JZ    DONE
 
diff --git a/vendor/github.com/aead/chacha20/chacha/chacha_amd64.s b/vendor/github.com/aead/chacha20/chacha/chacha_amd64.s
index 96b0680a..262f9ade 100644
--- a/vendor/github.com/aead/chacha20/chacha/chacha_amd64.s
+++ b/vendor/github.com/aead/chacha20/chacha/chacha_amd64.s
@@ -49,7 +49,7 @@ TEXT ·supportsAVX2(SB), 4, $0-1
 #define Dst DI
 #define Nonce AX
 #define Key BX
-#define Rounds CX
+#define Rounds DX
 
 // func initialize(state *[64]byte, key []byte, nonce *[16]byte)
 TEXT ·initialize(SB), 4, $0-40
@@ -166,10 +166,10 @@ chacha_loop:
 TEXT ·xorKeyStreamSSE2(SB), 4, $112-80
 	MOVQ dst_base+0(FP), Dst
 	MOVQ src_base+24(FP), Src
-	MOVQ src_len+32(FP), Len
 	MOVQ block+48(FP), Buffer
 	MOVQ state+56(FP), State
 	MOVQ rounds+64(FP), Rounds
+	MOVQ src_len+32(FP), Len
 
 	MOVOU 0*16(State), X0
 	MOVOU 1*16(State), X1
@@ -462,10 +462,10 @@ DONE:
 TEXT ·xorKeyStreamSSSE3(SB), 4, $144-80
 	MOVQ dst_base+0(FP), Dst
 	MOVQ src_base+24(FP), Src
-	MOVQ src_len+32(FP), Len
 	MOVQ block+48(FP), Buffer
 	MOVQ state+56(FP), State
 	MOVQ rounds+64(FP), Rounds
+	MOVQ src_len+32(FP), Len
 
 	MOVOU 0*16(State), X0
 	MOVOU 1*16(State), X1
@@ -773,10 +773,10 @@ DONE:
 TEXT ·xorKeyStreamAVX(SB), 4, $144-80
 	MOVQ dst_base+0(FP), Dst
 	MOVQ src_base+24(FP), Src
-	MOVQ src_len+32(FP), Len
 	MOVQ block+48(FP), Buffer
 	MOVQ state+56(FP), State
 	MOVQ rounds+64(FP), Rounds
+	MOVQ src_len+32(FP), Len
 
 	VMOVDQU 0*16(State), X0
 	VMOVDQU 1*16(State), X1