From: Mark Wooding <mdw@distorted.org.uk>
Date: Mon, 7 Nov 2016 12:24:35 +0000 (+0000)
Subject: symm/salsa20-arm-neon.S: Improve output permutation still further.
X-Git-Tag: 2.3.0~15
X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~mdw/git/catacomb/commitdiff_plain/f48fb6a6b1e0f29a45b42beae638ef9886312579

symm/salsa20-arm-neon.S: Improve output permutation still further.
---

diff --git a/symm/salsa20-arm-neon.S b/symm/salsa20-arm-neon.S
index 9d553d3a..cea40192 100644
--- a/symm/salsa20-arm-neon.S
+++ b/symm/salsa20-arm-neon.S
@@ -205,33 +205,28 @@ FUNC(salsa20_core_arm_neon)
 
 	b	0b
 
-	// Almost there.  Firstly the feedfoward addition.
+	// Almost there.  Firstly the feedfoward addition.  Also, establish a
+	// constant which will be useful later.
 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
+	 vmov.i64 q12, #0xffffffff		// = (-1, 0, -1, 0)
+	vadd.u32 q1, q9, q13			//  4,  9, 14,  3
+	vadd.u32 q2, q10, q14			//  8, 13,  2,  7
+	vadd.u32 q3, 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
+	// input.  The core trick is from Dan Bernstein's `armneon3'
+	// implementation, but with a lot of liposuction.
+	vmov	q15, q0
+
+	// Sort out the columns by pairs.
+	vbif	q0, q3, q12			//  0,  1, 10, 11
+	vbif	q3, q2, q12			// 12, 13,  6,  7
+	vbif	q2, q1, q12			//  8,  9,  2,  3
+	vbif	q1, q15, q12			//  4,  5, 14, 15
+
+	// Now fix up the remaining discrepancies.
+	vswp	D1(q0), D1(q2)
+	vswp	D1(q1), D1(q3)
 
 	// And with that, we're done.
 	vstmia	r2, {QQ(q0, q3)}