X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/d6b9dc043945e8b65ebcd84bfa2668e93041f598..3cb47d2759de69a9e4fdf0030f518ca513b59c7b:/symm/salsa20-arm-neon.S

diff --git a/symm/salsa20-arm-neon.S b/symm/salsa20-arm-neon.S
index f212f2fa..9cb40472 100644
--- a/symm/salsa20-arm-neon.S
+++ b/symm/salsa20-arm-neon.S
@@ -205,35 +205,36 @@ FUNC(salsa20_core_arm_neon)
 
 	b	0b
 
-	// Almost there.  Firstly the feedfoward addition, and then we have
-	// to write out the result.  Here we have to undo the permutation
-	// which was already applied to the input.
-9:	vadd.u32 q8, q8, q12
-	vadd.u32 q9, q9, q13
-	vadd.u32 q10, q10, q14
-	vadd.u32 q11, q11, q15
-
-	vst1.32	{d16[0]}, [r2 :32]!
-	vst1.32	{d22[1]}, [r2 :32]!
-	vst1.32	{d21[0]}, [r2 :32]!
-	vst1.32	{d19[1]}, [r2 :32]!
-
-	vst1.32	{d18[0]}, [r2 :32]!
-	vst1.32	{d16[1]}, [r2 :32]!
-	vst1.32	{d23[0]}, [r2 :32]!
-	vst1.32	{d21[1]}, [r2 :32]!
-
-	vst1.32	{d20[0]}, [r2 :32]!
-	vst1.32	{d18[1]}, [r2 :32]!
-	vst1.32	{d17[0]}, [r2 :32]!
-	vst1.32	{d23[1]}, [r2 :32]!
-
-	vst1.32	{d22[0]}, [r2 :32]!
-	vst1.32	{d20[1]}, [r2 :32]!
-	vst1.32	{d19[0]}, [r2 :32]!
-	vst1.32	{d17[1]}, [r2 :32]!
+	// Almost there.  Firstly the feedfoward addition.
+9:	vadd.u32 q0, q8, q12			//  0,  5, 10, 15
+	vadd.u32 q9, q9, q13			//  4,  9, 14,  3
+	vadd.u32 q10, q10, q14			//  8, 13,  2,  7
+	vadd.u32 q11, q11, q15			// 12,  1,  6, 11
+
+	// Next we must undo the permutation which was already applied to the
+	// input.  This can be done juggling values in registers, with the
+	// following fancy footwork: some row rotations, a transpose, and
+	// some more rotations.
+	vext.32	q9, q9, q9, #3			//  3,  4,  9, 14
+	vext.32	q10, q10, q10, #2		//  2,  7,  8, 13
+	vext.32 q11, q11, q11, #1		//  1,  6, 11, 12
+
+	vzip.32	q0, q10				//  0,  2,  5,  7
+						// 10,  8, 15, 13
+	vzip.32	q11, q9				//  1,  3,  6,  4
+						// 11,  9, 12, 14
+
+	vzip.32	q0, q11				//  0,  1,  2,  3
+						//  5,  6,  7,  4
+	vzip.32	q10, q9				// 10, 11,  8,  9
+						// 15, 12, 13, 14
+
+	vext.32	q1, q11, q11, #3		//  4,  5,  6,  7
+	vext.32	q2, q10, q10, #2		//  8,  9, 10, 11
+	vext.32	q3, q9, q9, #1			// 12, 13, 14, 15
 
 	// And with that, we're done.
+	vstmia	r2, {d0-d7}
 	bx	r14
 
 ENDFUNC