diff --git a/examples/sha1/asm.go b/examples/sha1/asm.go
index c370cdd1..85f7f1e3 100644
--- a/examples/sha1/asm.go
+++ b/examples/sha1/asm.go
@@ -18,13 +18,13 @@ func main() {
 	w := AllocLocal(64)
 	W := func(r int) Mem { return w.Offset((r % 16) * 4) }
 
-	// Load initial hash.
+	Comment("Load initial hash.")
 	hash := [5]Register{GP32(), GP32(), GP32(), GP32(), GP32()}
 	for i, r := range hash {
 		MOVL(h.Offset(4*i), r)
 	}
 
-	// Initialize registers.
+	Comment("Initialize registers.")
 	a, b, c, d, e := GP32(), GP32(), GP32(), GP32(), GP32()
 	for i, r := range []Register{a, b, c, d, e} {
 		MOVL(hash[i], r)
@@ -42,6 +42,7 @@ func main() {
 	}
 
 	for r := 0; r < 80; r++ {
+		Commentf("Round %d.", r)
 		q := quarter[r/20]
 
 		// Load message value.
@@ -72,12 +73,12 @@ func main() {
 		a, b, c, d, e = t, a, b, c, d
 	}
 
-	// Final add.
+	Comment("Final add.")
 	for i, r := range []Register{a, b, c, d, e} {
 		ADDL(r, hash[i])
 	}
 
-	// Store results back.
+	Comment("Store results back.")
 	for i, r := range hash {
 		MOVL(r, h.Offset(4*i))
 	}
diff --git a/examples/sha1/sha1.s b/examples/sha1/sha1.s
index 4f16ebac..518372db 100644
--- a/examples/sha1/sha1.s
+++ b/examples/sha1/sha1.s
@@ -2,18 +2,24 @@
 
 // func block(h *[5]uint32, m []byte)
 TEXT ·block(SB), $64-32
-	MOVQ   h(FP), AX
-	MOVQ   m_base+8(FP), CX
-	MOVL   (AX), DX
-	MOVL   4(AX), BX
-	MOVL   8(AX), BP
-	MOVL   12(AX), SI
-	MOVL   16(AX), DI
-	MOVL   DX, R8
-	MOVL   BX, R9
-	MOVL   BP, R10
-	MOVL   SI, R11
-	MOVL   DI, R12
+	MOVQ h(FP), AX
+	MOVQ m_base+8(FP), CX
+
+	// Load initial hash.
+	MOVL (AX), DX
+	MOVL 4(AX), BX
+	MOVL 8(AX), BP
+	MOVL 12(AX), SI
+	MOVL 16(AX), DI
+
+	// Initialize registers.
+	MOVL DX, R8
+	MOVL BX, R9
+	MOVL BP, R10
+	MOVL SI, R11
+	MOVL DI, R12
+
+	// Round 0.
 	MOVL   (CX), R13
 	BSWAPL R13
 	MOVL   R13, (SP)
@@ -28,6 +34,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R14
 	ADDL   R13, R14
 	ROLL   $0x1e, R9
+
+	// Round 1.
 	MOVL   4(CX), R12
 	BSWAPL R12
 	MOVL   R12, 4(SP)
@@ -42,6 +50,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R13
 	ADDL   R12, R13
 	ROLL   $0x1e, R8
+
+	// Round 2.
 	MOVL   8(CX), R11
 	BSWAPL R11
 	MOVL   R11, 8(SP)
@@ -56,6 +66,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R12
 	ADDL   R11, R12
 	ROLL   $0x1e, R14
+
+	// Round 3.
 	MOVL   12(CX), R10
 	BSWAPL R10
 	MOVL   R10, 12(SP)
@@ -70,6 +82,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R11
 	ADDL   R10, R11
 	ROLL   $0x1e, R13
+
+	// Round 4.
 	MOVL   16(CX), R9
 	BSWAPL R9
 	MOVL   R9, 16(SP)
@@ -84,6 +98,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R10
 	ADDL   R9, R10
 	ROLL   $0x1e, R12
+
+	// Round 5.
 	MOVL   20(CX), R8
 	BSWAPL R8
 	MOVL   R8, 20(SP)
@@ -98,6 +114,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R9
 	ADDL   R8, R9
 	ROLL   $0x1e, R11
+
+	// Round 6.
 	MOVL   24(CX), R8
 	BSWAPL R8
 	MOVL   R8, 24(SP)
@@ -112,6 +130,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R14
 	ADDL   R8, R14
 	ROLL   $0x1e, R10
+
+	// Round 7.
 	MOVL   28(CX), R8
 	BSWAPL R8
 	MOVL   R8, 28(SP)
@@ -126,6 +146,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R13
 	ADDL   R8, R13
 	ROLL   $0x1e, R9
+
+	// Round 8.
 	MOVL   32(CX), R8
 	BSWAPL R8
 	MOVL   R8, 32(SP)
@@ -140,6 +162,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R12
 	ADDL   R8, R12
 	ROLL   $0x1e, R14
+
+	// Round 9.
 	MOVL   36(CX), R8
 	BSWAPL R8
 	MOVL   R8, 36(SP)
@@ -154,6 +178,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R11
 	ADDL   R8, R11
 	ROLL   $0x1e, R13
+
+	// Round 10.
 	MOVL   40(CX), R8
 	BSWAPL R8
 	MOVL   R8, 40(SP)
@@ -168,6 +194,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R10
 	ADDL   R8, R10
 	ROLL   $0x1e, R12
+
+	// Round 11.
 	MOVL   44(CX), R8
 	BSWAPL R8
 	MOVL   R8, 44(SP)
@@ -182,6 +210,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R9
 	ADDL   R8, R9
 	ROLL   $0x1e, R11
+
+	// Round 12.
 	MOVL   48(CX), R8
 	BSWAPL R8
 	MOVL   R8, 48(SP)
@@ -196,6 +226,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R14
 	ADDL   R8, R14
 	ROLL   $0x1e, R10
+
+	// Round 13.
 	MOVL   52(CX), R8
 	BSWAPL R8
 	MOVL   R8, 52(SP)
@@ -210,6 +242,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R13
 	ADDL   R8, R13
 	ROLL   $0x1e, R9
+
+	// Round 14.
 	MOVL   56(CX), R8
 	BSWAPL R8
 	MOVL   R8, 56(SP)
@@ -224,6 +258,8 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R12
 	ADDL   R8, R12
 	ROLL   $0x1e, R14
+
+	// Round 15.
 	MOVL   60(CX), CX
 	BSWAPL CX
 	MOVL   CX, 60(SP)
@@ -238,1102 +274,1234 @@ TEXT ·block(SB), $64-32
 	ADDL   $0x5a827999, R8
 	ADDL   CX, R8
 	ROLL   $0x1e, R13
-	MOVL   52(SP), CX
-	XORL   32(SP), CX
-	XORL   8(SP), CX
-	XORL   (SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, (SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R14, R11
-	XORL   R13, R11
-	ANDL   R12, R11
-	XORL   R14, R11
-	ADDL   R11, R10
-	ADDL   R9, R10
-	ADDL   $0x5a827999, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   56(SP), CX
-	XORL   36(SP), CX
-	XORL   12(SP), CX
-	XORL   4(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 4(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R13, R11
-	XORL   R12, R11
-	ANDL   R8, R11
-	XORL   R13, R11
-	ADDL   R11, R9
-	ADDL   R14, R9
-	ADDL   $0x5a827999, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   60(SP), CX
-	XORL   40(SP), CX
-	XORL   16(SP), CX
-	XORL   8(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 8(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R12, R14
-	XORL   R8, R14
-	ANDL   R10, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0x5a827999, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   (SP), CX
-	XORL   44(SP), CX
-	XORL   20(SP), CX
-	XORL   12(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 12(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R8, R14
-	XORL   R10, R14
-	ANDL   R9, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0x5a827999, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   4(SP), CX
-	XORL   48(SP), CX
-	XORL   24(SP), CX
-	XORL   16(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 16(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0x6ed9eba1, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   8(SP), CX
-	XORL   52(SP), CX
-	XORL   28(SP), CX
-	XORL   20(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 20(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0x6ed9eba1, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   12(SP), CX
-	XORL   56(SP), CX
-	XORL   32(SP), CX
-	XORL   24(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 24(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0x6ed9eba1, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   16(SP), CX
-	XORL   60(SP), CX
-	XORL   36(SP), CX
-	XORL   28(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 28(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0x6ed9eba1, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   20(SP), CX
-	XORL   (SP), CX
-	XORL   40(SP), CX
-	XORL   32(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 32(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0x6ed9eba1, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   24(SP), CX
-	XORL   4(SP), CX
-	XORL   44(SP), CX
-	XORL   36(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 36(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0x6ed9eba1, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   28(SP), CX
-	XORL   8(SP), CX
-	XORL   48(SP), CX
-	XORL   40(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 40(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0x6ed9eba1, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   32(SP), CX
-	XORL   12(SP), CX
-	XORL   52(SP), CX
-	XORL   44(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 44(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0x6ed9eba1, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   36(SP), CX
-	XORL   16(SP), CX
-	XORL   56(SP), CX
-	XORL   48(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 48(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0x6ed9eba1, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   40(SP), CX
-	XORL   20(SP), CX
-	XORL   60(SP), CX
-	XORL   52(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 52(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0x6ed9eba1, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   44(SP), CX
-	XORL   24(SP), CX
-	XORL   (SP), CX
-	XORL   56(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 56(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0x6ed9eba1, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   48(SP), CX
-	XORL   28(SP), CX
-	XORL   4(SP), CX
-	XORL   60(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 60(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0x6ed9eba1, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   52(SP), CX
-	XORL   32(SP), CX
-	XORL   8(SP), CX
-	XORL   (SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, (SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0x6ed9eba1, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   56(SP), CX
-	XORL   36(SP), CX
-	XORL   12(SP), CX
-	XORL   4(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 4(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0x6ed9eba1, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   60(SP), CX
-	XORL   40(SP), CX
-	XORL   16(SP), CX
-	XORL   8(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 8(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0x6ed9eba1, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   (SP), CX
-	XORL   44(SP), CX
-	XORL   20(SP), CX
-	XORL   12(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 12(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0x6ed9eba1, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   4(SP), CX
-	XORL   48(SP), CX
-	XORL   24(SP), CX
-	XORL   16(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 16(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0x6ed9eba1, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   8(SP), CX
-	XORL   52(SP), CX
-	XORL   28(SP), CX
-	XORL   20(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 20(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0x6ed9eba1, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   12(SP), CX
-	XORL   56(SP), CX
-	XORL   32(SP), CX
-	XORL   24(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 24(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0x6ed9eba1, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   16(SP), CX
-	XORL   60(SP), CX
-	XORL   36(SP), CX
-	XORL   28(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 28(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0x6ed9eba1, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   20(SP), CX
-	XORL   (SP), CX
-	XORL   40(SP), CX
-	XORL   32(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 32(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	ORL    R13, R14
-	ANDL   R11, R14
-	MOVL   R12, R15
-	ANDL   R13, R15
-	ORL    R14, R15
-	ADDL   R15, R10
-	ADDL   R9, R10
-	ADDL   $0x8f1bbcdc, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   24(SP), CX
-	XORL   4(SP), CX
-	XORL   44(SP), CX
-	XORL   36(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 36(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	ORL    R12, R14
-	ANDL   R13, R14
-	MOVL   R8, R15
-	ANDL   R12, R15
-	ORL    R14, R15
-	ADDL   R15, R9
-	ADDL   R11, R9
-	ADDL   $0x8f1bbcdc, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   28(SP), CX
-	XORL   8(SP), CX
-	XORL   48(SP), CX
-	XORL   40(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 40(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	ORL    R8, R14
-	ANDL   R12, R14
-	MOVL   R10, R15
-	ANDL   R8, R15
-	ORL    R14, R15
-	ADDL   R15, R11
-	ADDL   R13, R11
-	ADDL   $0x8f1bbcdc, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   32(SP), CX
-	XORL   12(SP), CX
-	XORL   52(SP), CX
-	XORL   44(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 44(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	ORL    R10, R14
-	ANDL   R8, R14
-	MOVL   R9, R15
-	ANDL   R10, R15
-	ORL    R14, R15
-	ADDL   R15, R13
-	ADDL   R12, R13
-	ADDL   $0x8f1bbcdc, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   36(SP), CX
-	XORL   16(SP), CX
-	XORL   56(SP), CX
-	XORL   48(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 48(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	ORL    R9, R14
-	ANDL   R10, R14
-	MOVL   R11, R15
-	ANDL   R9, R15
-	ORL    R14, R15
-	ADDL   R15, R12
-	ADDL   R8, R12
-	ADDL   $0x8f1bbcdc, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   40(SP), CX
-	XORL   20(SP), CX
-	XORL   60(SP), CX
-	XORL   52(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 52(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	ORL    R11, R14
-	ANDL   R9, R14
-	MOVL   R13, R15
-	ANDL   R11, R15
-	ORL    R14, R15
-	ADDL   R15, R8
-	ADDL   R10, R8
-	ADDL   $0x8f1bbcdc, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   44(SP), CX
-	XORL   24(SP), CX
-	XORL   (SP), CX
-	XORL   56(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 56(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	ORL    R13, R14
-	ANDL   R11, R14
-	MOVL   R12, R15
-	ANDL   R13, R15
-	ORL    R14, R15
-	ADDL   R15, R10
-	ADDL   R9, R10
-	ADDL   $0x8f1bbcdc, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   48(SP), CX
-	XORL   28(SP), CX
-	XORL   4(SP), CX
-	XORL   60(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 60(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	ORL    R12, R14
-	ANDL   R13, R14
-	MOVL   R8, R15
-	ANDL   R12, R15
-	ORL    R14, R15
-	ADDL   R15, R9
-	ADDL   R11, R9
-	ADDL   $0x8f1bbcdc, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   52(SP), CX
-	XORL   32(SP), CX
-	XORL   8(SP), CX
-	XORL   (SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, (SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	ORL    R8, R14
-	ANDL   R12, R14
-	MOVL   R10, R15
-	ANDL   R8, R15
-	ORL    R14, R15
-	ADDL   R15, R11
-	ADDL   R13, R11
-	ADDL   $0x8f1bbcdc, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   56(SP), CX
-	XORL   36(SP), CX
-	XORL   12(SP), CX
-	XORL   4(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 4(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	ORL    R10, R14
-	ANDL   R8, R14
-	MOVL   R9, R15
-	ANDL   R10, R15
-	ORL    R14, R15
-	ADDL   R15, R13
-	ADDL   R12, R13
-	ADDL   $0x8f1bbcdc, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   60(SP), CX
-	XORL   40(SP), CX
-	XORL   16(SP), CX
-	XORL   8(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 8(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	ORL    R9, R14
-	ANDL   R10, R14
-	MOVL   R11, R15
-	ANDL   R9, R15
-	ORL    R14, R15
-	ADDL   R15, R12
-	ADDL   R8, R12
-	ADDL   $0x8f1bbcdc, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   (SP), CX
-	XORL   44(SP), CX
-	XORL   20(SP), CX
-	XORL   12(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 12(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	ORL    R11, R14
-	ANDL   R9, R14
-	MOVL   R13, R15
-	ANDL   R11, R15
-	ORL    R14, R15
-	ADDL   R15, R8
-	ADDL   R10, R8
-	ADDL   $0x8f1bbcdc, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   4(SP), CX
-	XORL   48(SP), CX
-	XORL   24(SP), CX
-	XORL   16(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 16(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	ORL    R13, R14
-	ANDL   R11, R14
-	MOVL   R12, R15
-	ANDL   R13, R15
-	ORL    R14, R15
-	ADDL   R15, R10
-	ADDL   R9, R10
-	ADDL   $0x8f1bbcdc, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   8(SP), CX
-	XORL   52(SP), CX
-	XORL   28(SP), CX
-	XORL   20(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 20(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	ORL    R12, R14
-	ANDL   R13, R14
-	MOVL   R8, R15
-	ANDL   R12, R15
-	ORL    R14, R15
-	ADDL   R15, R9
-	ADDL   R11, R9
-	ADDL   $0x8f1bbcdc, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   12(SP), CX
-	XORL   56(SP), CX
-	XORL   32(SP), CX
-	XORL   24(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 24(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	ORL    R8, R14
-	ANDL   R12, R14
-	MOVL   R10, R15
-	ANDL   R8, R15
-	ORL    R14, R15
-	ADDL   R15, R11
-	ADDL   R13, R11
-	ADDL   $0x8f1bbcdc, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   16(SP), CX
-	XORL   60(SP), CX
-	XORL   36(SP), CX
-	XORL   28(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 28(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	ORL    R10, R14
-	ANDL   R8, R14
-	MOVL   R9, R15
-	ANDL   R10, R15
-	ORL    R14, R15
-	ADDL   R15, R13
-	ADDL   R12, R13
-	ADDL   $0x8f1bbcdc, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   20(SP), CX
-	XORL   (SP), CX
-	XORL   40(SP), CX
-	XORL   32(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 32(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	ORL    R9, R14
-	ANDL   R10, R14
-	MOVL   R11, R15
-	ANDL   R9, R15
-	ORL    R14, R15
-	ADDL   R15, R12
-	ADDL   R8, R12
-	ADDL   $0x8f1bbcdc, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   24(SP), CX
-	XORL   4(SP), CX
-	XORL   44(SP), CX
-	XORL   36(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 36(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	ORL    R11, R14
-	ANDL   R9, R14
-	MOVL   R13, R15
-	ANDL   R11, R15
-	ORL    R14, R15
-	ADDL   R15, R8
-	ADDL   R10, R8
-	ADDL   $0x8f1bbcdc, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   28(SP), CX
-	XORL   8(SP), CX
-	XORL   48(SP), CX
-	XORL   40(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 40(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	ORL    R13, R14
-	ANDL   R11, R14
-	MOVL   R12, R15
-	ANDL   R13, R15
-	ORL    R14, R15
-	ADDL   R15, R10
-	ADDL   R9, R10
-	ADDL   $0x8f1bbcdc, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   32(SP), CX
-	XORL   12(SP), CX
-	XORL   52(SP), CX
-	XORL   44(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 44(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	ORL    R12, R14
-	ANDL   R13, R14
-	MOVL   R8, R15
-	ANDL   R12, R15
-	ORL    R14, R15
-	ADDL   R15, R9
-	ADDL   R11, R9
-	ADDL   $0x8f1bbcdc, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   36(SP), CX
-	XORL   16(SP), CX
-	XORL   56(SP), CX
-	XORL   48(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 48(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0xca62c1d6, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   40(SP), CX
-	XORL   20(SP), CX
-	XORL   60(SP), CX
-	XORL   52(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 52(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0xca62c1d6, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   44(SP), CX
-	XORL   24(SP), CX
-	XORL   (SP), CX
-	XORL   56(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 56(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0xca62c1d6, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   48(SP), CX
-	XORL   28(SP), CX
-	XORL   4(SP), CX
-	XORL   60(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 60(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0xca62c1d6, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   52(SP), CX
-	XORL   32(SP), CX
-	XORL   8(SP), CX
-	XORL   (SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, (SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0xca62c1d6, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   56(SP), CX
-	XORL   36(SP), CX
-	XORL   12(SP), CX
-	XORL   4(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 4(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0xca62c1d6, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   60(SP), CX
-	XORL   40(SP), CX
-	XORL   16(SP), CX
-	XORL   8(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 8(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0xca62c1d6, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   (SP), CX
-	XORL   44(SP), CX
-	XORL   20(SP), CX
-	XORL   12(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 12(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0xca62c1d6, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   4(SP), CX
-	XORL   48(SP), CX
-	XORL   24(SP), CX
-	XORL   16(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 16(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0xca62c1d6, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   8(SP), CX
-	XORL   52(SP), CX
-	XORL   28(SP), CX
-	XORL   20(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 20(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0xca62c1d6, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   12(SP), CX
-	XORL   56(SP), CX
-	XORL   32(SP), CX
-	XORL   24(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 24(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0xca62c1d6, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   16(SP), CX
-	XORL   60(SP), CX
-	XORL   36(SP), CX
-	XORL   28(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 28(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0xca62c1d6, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   20(SP), CX
-	XORL   (SP), CX
-	XORL   40(SP), CX
-	XORL   32(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 32(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0xca62c1d6, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   24(SP), CX
-	XORL   4(SP), CX
-	XORL   44(SP), CX
-	XORL   36(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 36(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0xca62c1d6, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	MOVL   28(SP), CX
-	XORL   8(SP), CX
-	XORL   48(SP), CX
-	XORL   40(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 40(SP)
-	MOVL   R13, R12
-	ROLL   $0x05, R12
-	MOVL   R11, R14
-	XORL   R9, R14
-	XORL   R10, R14
-	ADDL   R14, R12
-	ADDL   R8, R12
-	ADDL   $0xca62c1d6, R12
-	ADDL   CX, R12
-	ROLL   $0x1e, R11
-	MOVL   32(SP), CX
-	XORL   12(SP), CX
-	XORL   52(SP), CX
-	XORL   44(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 44(SP)
-	MOVL   R12, R8
-	ROLL   $0x05, R8
-	MOVL   R13, R14
-	XORL   R11, R14
-	XORL   R9, R14
-	ADDL   R14, R8
-	ADDL   R10, R8
-	ADDL   $0xca62c1d6, R8
-	ADDL   CX, R8
-	ROLL   $0x1e, R13
-	MOVL   36(SP), CX
-	XORL   16(SP), CX
-	XORL   56(SP), CX
-	XORL   48(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 48(SP)
-	MOVL   R8, R10
-	ROLL   $0x05, R10
-	MOVL   R12, R14
-	XORL   R13, R14
-	XORL   R11, R14
-	ADDL   R14, R10
-	ADDL   R9, R10
-	ADDL   $0xca62c1d6, R10
-	ADDL   CX, R10
-	ROLL   $0x1e, R12
-	MOVL   40(SP), CX
-	XORL   20(SP), CX
-	XORL   60(SP), CX
-	XORL   52(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 52(SP)
-	MOVL   R10, R9
-	ROLL   $0x05, R9
-	MOVL   R8, R14
-	XORL   R12, R14
-	XORL   R13, R14
-	ADDL   R14, R9
-	ADDL   R11, R9
-	ADDL   $0xca62c1d6, R9
-	ADDL   CX, R9
-	ROLL   $0x1e, R8
-	MOVL   44(SP), CX
-	XORL   24(SP), CX
-	XORL   (SP), CX
-	XORL   56(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 56(SP)
-	MOVL   R9, R11
-	ROLL   $0x05, R11
-	MOVL   R10, R14
-	XORL   R8, R14
-	XORL   R12, R14
-	ADDL   R14, R11
-	ADDL   R13, R11
-	ADDL   $0xca62c1d6, R11
-	ADDL   CX, R11
-	ROLL   $0x1e, R10
-	MOVL   48(SP), CX
-	XORL   28(SP), CX
-	XORL   4(SP), CX
-	XORL   60(SP), CX
-	ROLL   $0x01, CX
-	MOVL   CX, 60(SP)
-	MOVL   R11, R13
-	ROLL   $0x05, R13
-	MOVL   R9, R14
-	XORL   R10, R14
-	XORL   R8, R14
-	ADDL   R14, R13
-	ADDL   R12, R13
-	ADDL   $0xca62c1d6, R13
-	ADDL   CX, R13
-	ROLL   $0x1e, R9
-	ADDL   R13, DX
-	ADDL   R11, BX
-	ADDL   R9, BP
-	ADDL   R10, SI
-	ADDL   R8, DI
-	MOVL   DX, (AX)
-	MOVL   BX, 4(AX)
-	MOVL   BP, 8(AX)
-	MOVL   SI, 12(AX)
-	MOVL   DI, 16(AX)
+
+	// Round 16.
+	MOVL 52(SP), CX
+	XORL 32(SP), CX
+	XORL 8(SP), CX
+	XORL (SP), CX
+	ROLL $0x01, CX
+	MOVL CX, (SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R14, R11
+	XORL R13, R11
+	ANDL R12, R11
+	XORL R14, R11
+	ADDL R11, R10
+	ADDL R9, R10
+	ADDL $0x5a827999, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 17.
+	MOVL 56(SP), CX
+	XORL 36(SP), CX
+	XORL 12(SP), CX
+	XORL 4(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 4(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R13, R11
+	XORL R12, R11
+	ANDL R8, R11
+	XORL R13, R11
+	ADDL R11, R9
+	ADDL R14, R9
+	ADDL $0x5a827999, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 18.
+	MOVL 60(SP), CX
+	XORL 40(SP), CX
+	XORL 16(SP), CX
+	XORL 8(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 8(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R12, R14
+	XORL R8, R14
+	ANDL R10, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0x5a827999, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 19.
+	MOVL (SP), CX
+	XORL 44(SP), CX
+	XORL 20(SP), CX
+	XORL 12(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 12(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R8, R14
+	XORL R10, R14
+	ANDL R9, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0x5a827999, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 20.
+	MOVL 4(SP), CX
+	XORL 48(SP), CX
+	XORL 24(SP), CX
+	XORL 16(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 16(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0x6ed9eba1, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 21.
+	MOVL 8(SP), CX
+	XORL 52(SP), CX
+	XORL 28(SP), CX
+	XORL 20(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 20(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0x6ed9eba1, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 22.
+	MOVL 12(SP), CX
+	XORL 56(SP), CX
+	XORL 32(SP), CX
+	XORL 24(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 24(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0x6ed9eba1, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 23.
+	MOVL 16(SP), CX
+	XORL 60(SP), CX
+	XORL 36(SP), CX
+	XORL 28(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 28(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0x6ed9eba1, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 24.
+	MOVL 20(SP), CX
+	XORL (SP), CX
+	XORL 40(SP), CX
+	XORL 32(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 32(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0x6ed9eba1, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 25.
+	MOVL 24(SP), CX
+	XORL 4(SP), CX
+	XORL 44(SP), CX
+	XORL 36(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 36(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0x6ed9eba1, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 26.
+	MOVL 28(SP), CX
+	XORL 8(SP), CX
+	XORL 48(SP), CX
+	XORL 40(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 40(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0x6ed9eba1, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 27.
+	MOVL 32(SP), CX
+	XORL 12(SP), CX
+	XORL 52(SP), CX
+	XORL 44(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 44(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0x6ed9eba1, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 28.
+	MOVL 36(SP), CX
+	XORL 16(SP), CX
+	XORL 56(SP), CX
+	XORL 48(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 48(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0x6ed9eba1, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 29.
+	MOVL 40(SP), CX
+	XORL 20(SP), CX
+	XORL 60(SP), CX
+	XORL 52(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 52(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0x6ed9eba1, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 30.
+	MOVL 44(SP), CX
+	XORL 24(SP), CX
+	XORL (SP), CX
+	XORL 56(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 56(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0x6ed9eba1, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 31.
+	MOVL 48(SP), CX
+	XORL 28(SP), CX
+	XORL 4(SP), CX
+	XORL 60(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 60(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0x6ed9eba1, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 32.
+	MOVL 52(SP), CX
+	XORL 32(SP), CX
+	XORL 8(SP), CX
+	XORL (SP), CX
+	ROLL $0x01, CX
+	MOVL CX, (SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0x6ed9eba1, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 33.
+	MOVL 56(SP), CX
+	XORL 36(SP), CX
+	XORL 12(SP), CX
+	XORL 4(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 4(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0x6ed9eba1, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 34.
+	MOVL 60(SP), CX
+	XORL 40(SP), CX
+	XORL 16(SP), CX
+	XORL 8(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 8(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0x6ed9eba1, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 35.
+	MOVL (SP), CX
+	XORL 44(SP), CX
+	XORL 20(SP), CX
+	XORL 12(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 12(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0x6ed9eba1, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 36.
+	MOVL 4(SP), CX
+	XORL 48(SP), CX
+	XORL 24(SP), CX
+	XORL 16(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 16(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0x6ed9eba1, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 37.
+	MOVL 8(SP), CX
+	XORL 52(SP), CX
+	XORL 28(SP), CX
+	XORL 20(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 20(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0x6ed9eba1, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 38.
+	MOVL 12(SP), CX
+	XORL 56(SP), CX
+	XORL 32(SP), CX
+	XORL 24(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 24(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0x6ed9eba1, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 39.
+	MOVL 16(SP), CX
+	XORL 60(SP), CX
+	XORL 36(SP), CX
+	XORL 28(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 28(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0x6ed9eba1, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 40.
+	MOVL 20(SP), CX
+	XORL (SP), CX
+	XORL 40(SP), CX
+	XORL 32(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 32(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	ORL  R13, R14
+	ANDL R11, R14
+	MOVL R12, R15
+	ANDL R13, R15
+	ORL  R14, R15
+	ADDL R15, R10
+	ADDL R9, R10
+	ADDL $0x8f1bbcdc, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 41.
+	MOVL 24(SP), CX
+	XORL 4(SP), CX
+	XORL 44(SP), CX
+	XORL 36(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 36(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	ORL  R12, R14
+	ANDL R13, R14
+	MOVL R8, R15
+	ANDL R12, R15
+	ORL  R14, R15
+	ADDL R15, R9
+	ADDL R11, R9
+	ADDL $0x8f1bbcdc, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 42.
+	MOVL 28(SP), CX
+	XORL 8(SP), CX
+	XORL 48(SP), CX
+	XORL 40(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 40(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	ORL  R8, R14
+	ANDL R12, R14
+	MOVL R10, R15
+	ANDL R8, R15
+	ORL  R14, R15
+	ADDL R15, R11
+	ADDL R13, R11
+	ADDL $0x8f1bbcdc, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 43.
+	MOVL 32(SP), CX
+	XORL 12(SP), CX
+	XORL 52(SP), CX
+	XORL 44(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 44(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	ORL  R10, R14
+	ANDL R8, R14
+	MOVL R9, R15
+	ANDL R10, R15
+	ORL  R14, R15
+	ADDL R15, R13
+	ADDL R12, R13
+	ADDL $0x8f1bbcdc, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 44.
+	MOVL 36(SP), CX
+	XORL 16(SP), CX
+	XORL 56(SP), CX
+	XORL 48(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 48(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	ORL  R9, R14
+	ANDL R10, R14
+	MOVL R11, R15
+	ANDL R9, R15
+	ORL  R14, R15
+	ADDL R15, R12
+	ADDL R8, R12
+	ADDL $0x8f1bbcdc, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 45.
+	MOVL 40(SP), CX
+	XORL 20(SP), CX
+	XORL 60(SP), CX
+	XORL 52(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 52(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	ORL  R11, R14
+	ANDL R9, R14
+	MOVL R13, R15
+	ANDL R11, R15
+	ORL  R14, R15
+	ADDL R15, R8
+	ADDL R10, R8
+	ADDL $0x8f1bbcdc, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 46.
+	MOVL 44(SP), CX
+	XORL 24(SP), CX
+	XORL (SP), CX
+	XORL 56(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 56(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	ORL  R13, R14
+	ANDL R11, R14
+	MOVL R12, R15
+	ANDL R13, R15
+	ORL  R14, R15
+	ADDL R15, R10
+	ADDL R9, R10
+	ADDL $0x8f1bbcdc, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 47.
+	MOVL 48(SP), CX
+	XORL 28(SP), CX
+	XORL 4(SP), CX
+	XORL 60(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 60(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	ORL  R12, R14
+	ANDL R13, R14
+	MOVL R8, R15
+	ANDL R12, R15
+	ORL  R14, R15
+	ADDL R15, R9
+	ADDL R11, R9
+	ADDL $0x8f1bbcdc, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 48.
+	MOVL 52(SP), CX
+	XORL 32(SP), CX
+	XORL 8(SP), CX
+	XORL (SP), CX
+	ROLL $0x01, CX
+	MOVL CX, (SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	ORL  R8, R14
+	ANDL R12, R14
+	MOVL R10, R15
+	ANDL R8, R15
+	ORL  R14, R15
+	ADDL R15, R11
+	ADDL R13, R11
+	ADDL $0x8f1bbcdc, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 49.
+	MOVL 56(SP), CX
+	XORL 36(SP), CX
+	XORL 12(SP), CX
+	XORL 4(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 4(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	ORL  R10, R14
+	ANDL R8, R14
+	MOVL R9, R15
+	ANDL R10, R15
+	ORL  R14, R15
+	ADDL R15, R13
+	ADDL R12, R13
+	ADDL $0x8f1bbcdc, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 50.
+	MOVL 60(SP), CX
+	XORL 40(SP), CX
+	XORL 16(SP), CX
+	XORL 8(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 8(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	ORL  R9, R14
+	ANDL R10, R14
+	MOVL R11, R15
+	ANDL R9, R15
+	ORL  R14, R15
+	ADDL R15, R12
+	ADDL R8, R12
+	ADDL $0x8f1bbcdc, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 51.
+	MOVL (SP), CX
+	XORL 44(SP), CX
+	XORL 20(SP), CX
+	XORL 12(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 12(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	ORL  R11, R14
+	ANDL R9, R14
+	MOVL R13, R15
+	ANDL R11, R15
+	ORL  R14, R15
+	ADDL R15, R8
+	ADDL R10, R8
+	ADDL $0x8f1bbcdc, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 52.
+	MOVL 4(SP), CX
+	XORL 48(SP), CX
+	XORL 24(SP), CX
+	XORL 16(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 16(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	ORL  R13, R14
+	ANDL R11, R14
+	MOVL R12, R15
+	ANDL R13, R15
+	ORL  R14, R15
+	ADDL R15, R10
+	ADDL R9, R10
+	ADDL $0x8f1bbcdc, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 53.
+	MOVL 8(SP), CX
+	XORL 52(SP), CX
+	XORL 28(SP), CX
+	XORL 20(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 20(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	ORL  R12, R14
+	ANDL R13, R14
+	MOVL R8, R15
+	ANDL R12, R15
+	ORL  R14, R15
+	ADDL R15, R9
+	ADDL R11, R9
+	ADDL $0x8f1bbcdc, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 54.
+	MOVL 12(SP), CX
+	XORL 56(SP), CX
+	XORL 32(SP), CX
+	XORL 24(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 24(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	ORL  R8, R14
+	ANDL R12, R14
+	MOVL R10, R15
+	ANDL R8, R15
+	ORL  R14, R15
+	ADDL R15, R11
+	ADDL R13, R11
+	ADDL $0x8f1bbcdc, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 55.
+	MOVL 16(SP), CX
+	XORL 60(SP), CX
+	XORL 36(SP), CX
+	XORL 28(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 28(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	ORL  R10, R14
+	ANDL R8, R14
+	MOVL R9, R15
+	ANDL R10, R15
+	ORL  R14, R15
+	ADDL R15, R13
+	ADDL R12, R13
+	ADDL $0x8f1bbcdc, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 56.
+	MOVL 20(SP), CX
+	XORL (SP), CX
+	XORL 40(SP), CX
+	XORL 32(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 32(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	ORL  R9, R14
+	ANDL R10, R14
+	MOVL R11, R15
+	ANDL R9, R15
+	ORL  R14, R15
+	ADDL R15, R12
+	ADDL R8, R12
+	ADDL $0x8f1bbcdc, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 57.
+	MOVL 24(SP), CX
+	XORL 4(SP), CX
+	XORL 44(SP), CX
+	XORL 36(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 36(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	ORL  R11, R14
+	ANDL R9, R14
+	MOVL R13, R15
+	ANDL R11, R15
+	ORL  R14, R15
+	ADDL R15, R8
+	ADDL R10, R8
+	ADDL $0x8f1bbcdc, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 58.
+	MOVL 28(SP), CX
+	XORL 8(SP), CX
+	XORL 48(SP), CX
+	XORL 40(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 40(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	ORL  R13, R14
+	ANDL R11, R14
+	MOVL R12, R15
+	ANDL R13, R15
+	ORL  R14, R15
+	ADDL R15, R10
+	ADDL R9, R10
+	ADDL $0x8f1bbcdc, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 59.
+	MOVL 32(SP), CX
+	XORL 12(SP), CX
+	XORL 52(SP), CX
+	XORL 44(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 44(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	ORL  R12, R14
+	ANDL R13, R14
+	MOVL R8, R15
+	ANDL R12, R15
+	ORL  R14, R15
+	ADDL R15, R9
+	ADDL R11, R9
+	ADDL $0x8f1bbcdc, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 60.
+	MOVL 36(SP), CX
+	XORL 16(SP), CX
+	XORL 56(SP), CX
+	XORL 48(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 48(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0xca62c1d6, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 61.
+	MOVL 40(SP), CX
+	XORL 20(SP), CX
+	XORL 60(SP), CX
+	XORL 52(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 52(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0xca62c1d6, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 62.
+	MOVL 44(SP), CX
+	XORL 24(SP), CX
+	XORL (SP), CX
+	XORL 56(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 56(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0xca62c1d6, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 63.
+	MOVL 48(SP), CX
+	XORL 28(SP), CX
+	XORL 4(SP), CX
+	XORL 60(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 60(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0xca62c1d6, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 64.
+	MOVL 52(SP), CX
+	XORL 32(SP), CX
+	XORL 8(SP), CX
+	XORL (SP), CX
+	ROLL $0x01, CX
+	MOVL CX, (SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0xca62c1d6, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 65.
+	MOVL 56(SP), CX
+	XORL 36(SP), CX
+	XORL 12(SP), CX
+	XORL 4(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 4(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0xca62c1d6, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 66.
+	MOVL 60(SP), CX
+	XORL 40(SP), CX
+	XORL 16(SP), CX
+	XORL 8(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 8(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0xca62c1d6, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 67.
+	MOVL (SP), CX
+	XORL 44(SP), CX
+	XORL 20(SP), CX
+	XORL 12(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 12(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0xca62c1d6, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 68.
+	MOVL 4(SP), CX
+	XORL 48(SP), CX
+	XORL 24(SP), CX
+	XORL 16(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 16(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0xca62c1d6, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 69.
+	MOVL 8(SP), CX
+	XORL 52(SP), CX
+	XORL 28(SP), CX
+	XORL 20(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 20(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0xca62c1d6, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 70.
+	MOVL 12(SP), CX
+	XORL 56(SP), CX
+	XORL 32(SP), CX
+	XORL 24(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 24(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0xca62c1d6, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 71.
+	MOVL 16(SP), CX
+	XORL 60(SP), CX
+	XORL 36(SP), CX
+	XORL 28(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 28(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0xca62c1d6, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 72.
+	MOVL 20(SP), CX
+	XORL (SP), CX
+	XORL 40(SP), CX
+	XORL 32(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 32(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0xca62c1d6, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 73.
+	MOVL 24(SP), CX
+	XORL 4(SP), CX
+	XORL 44(SP), CX
+	XORL 36(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 36(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0xca62c1d6, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Round 74.
+	MOVL 28(SP), CX
+	XORL 8(SP), CX
+	XORL 48(SP), CX
+	XORL 40(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 40(SP)
+	MOVL R13, R12
+	ROLL $0x05, R12
+	MOVL R11, R14
+	XORL R9, R14
+	XORL R10, R14
+	ADDL R14, R12
+	ADDL R8, R12
+	ADDL $0xca62c1d6, R12
+	ADDL CX, R12
+	ROLL $0x1e, R11
+
+	// Round 75.
+	MOVL 32(SP), CX
+	XORL 12(SP), CX
+	XORL 52(SP), CX
+	XORL 44(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 44(SP)
+	MOVL R12, R8
+	ROLL $0x05, R8
+	MOVL R13, R14
+	XORL R11, R14
+	XORL R9, R14
+	ADDL R14, R8
+	ADDL R10, R8
+	ADDL $0xca62c1d6, R8
+	ADDL CX, R8
+	ROLL $0x1e, R13
+
+	// Round 76.
+	MOVL 36(SP), CX
+	XORL 16(SP), CX
+	XORL 56(SP), CX
+	XORL 48(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 48(SP)
+	MOVL R8, R10
+	ROLL $0x05, R10
+	MOVL R12, R14
+	XORL R13, R14
+	XORL R11, R14
+	ADDL R14, R10
+	ADDL R9, R10
+	ADDL $0xca62c1d6, R10
+	ADDL CX, R10
+	ROLL $0x1e, R12
+
+	// Round 77.
+	MOVL 40(SP), CX
+	XORL 20(SP), CX
+	XORL 60(SP), CX
+	XORL 52(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 52(SP)
+	MOVL R10, R9
+	ROLL $0x05, R9
+	MOVL R8, R14
+	XORL R12, R14
+	XORL R13, R14
+	ADDL R14, R9
+	ADDL R11, R9
+	ADDL $0xca62c1d6, R9
+	ADDL CX, R9
+	ROLL $0x1e, R8
+
+	// Round 78.
+	MOVL 44(SP), CX
+	XORL 24(SP), CX
+	XORL (SP), CX
+	XORL 56(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 56(SP)
+	MOVL R9, R11
+	ROLL $0x05, R11
+	MOVL R10, R14
+	XORL R8, R14
+	XORL R12, R14
+	ADDL R14, R11
+	ADDL R13, R11
+	ADDL $0xca62c1d6, R11
+	ADDL CX, R11
+	ROLL $0x1e, R10
+
+	// Round 79.
+	MOVL 48(SP), CX
+	XORL 28(SP), CX
+	XORL 4(SP), CX
+	XORL 60(SP), CX
+	ROLL $0x01, CX
+	MOVL CX, 60(SP)
+	MOVL R11, R13
+	ROLL $0x05, R13
+	MOVL R9, R14
+	XORL R10, R14
+	XORL R8, R14
+	ADDL R14, R13
+	ADDL R12, R13
+	ADDL $0xca62c1d6, R13
+	ADDL CX, R13
+	ROLL $0x1e, R9
+
+	// Final add.
+	ADDL R13, DX
+	ADDL R11, BX
+	ADDL R9, BP
+	ADDL R10, SI
+	ADDL R8, DI
+
+	// Store results back.
+	MOVL DX, (AX)
+	MOVL BX, 4(AX)
+	MOVL BP, 8(AX)
+	MOVL SI, 12(AX)
+	MOVL DI, 16(AX)
 	RET