--- /dev/null
+/// -*- mode: asm; asm-comment-char: ?/ -*-
+///
+/// Fancy SIMD implementation of ChaCha for ARM
+///
+/// (c) 2016 Straylight/Edgeware
+///
+
+///----- Licensing notice ---------------------------------------------------
+///
+/// This file is part of Catacomb.
+///
+/// Catacomb is free software; you can redistribute it and/or modify
+/// it under the terms of the GNU Library General Public License as
+/// published by the Free Software Foundation; either version 2 of the
+/// License, or (at your option) any later version.
+///
+/// Catacomb is distributed in the hope that it will be useful,
+/// but WITHOUT ANY WARRANTY; without even the implied warranty of
+/// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+/// GNU Library General Public License for more details.
+///
+/// You should have received a copy of the GNU Library General Public
+/// License along with Catacomb; if not, write to the Free
+/// Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
+/// MA 02111-1307, USA.
+
+///--------------------------------------------------------------------------
+/// External definitions.
+
+#include "config.h"
+#include "asm-common.h"
+
+///--------------------------------------------------------------------------
+/// Main.code.
+
+ .arch armv7-a
+ .fpu neon
+ .section .text
+
+FUNC(chacha_core_arm_neon)
+
+ // Arguments are in registers.
+ // r0 is the number of rounds to perform
+ // r1 points to the input matrix
+ // r2 points to the output matrix
+
+ // First job is to slurp the matrix into the SIMD registers. vldm
+ // and vstm work on word-aligned data, so this is fine.
+ //
+ // [ 0 1 2 3] (a, q8)
+ // [ 4 5 6 7] (b, q9)
+ // [ 8 9 10 11] (c, q10)
+ // [12 13 14 15] (d, q11)
+ //
+ // We need a copy for later. Rather than waste time copying them by
+ // hand, we'll use the three-address nature of the instruction set.
+ // But this means that the main loop is offset by a bit.
+ vldmia r1, {d24-d31}
+
+ // a += b; d ^= a; d <<<= 16
+ vadd.u32 q8, q12, q13
+ veor q11, q15, q8
+ vshl.u32 q0, q11, #16
+ vshr.u32 q11, q11, #16
+ vorr q11, q11, q0
+
+ // c += d; b ^= c; b <<<= 12
+ vadd.u32 q10, q14, q11
+ veor q9, q13, q10
+ vshl.u32 q0, q9, #12
+ vshr.u32 q9, q9, #20
+ vorr q9, q9, q0
+
+0:
+ // Apply (the rest of) a column quarterround to each of the columns
+ // simultaneously. Alas, there doesn't seem to be a packed word
+ // rotate, so we have to synthesize it.
+
+ // a += b; d ^= a; d <<<= 8
+ vadd.u32 q8, q8, q9
+ veor q11, q11, q8
+ vshl.u32 q0, q11, #8
+ vshr.u32 q11, q11, #24
+ vorr q11, q11, q0
+
+ // c += d; b ^= c; b <<<= 7
+ vadd.u32 q10, q10, q11
+ vext.32 q11, q11, q11, #3
+ veor q9, q9, q10
+ vext.32 q10, q10, q10, #2
+ vshl.u32 q0, q9, #7
+ vshr.u32 q9, q9, #25
+ vorr q9, q9, q0
+
+ // The not-quite-transpose conveniently only involves reordering
+ // elements of individual rows, which can be done quite easily. It
+ // doesn't involve any movement of elements between rows, or even
+ // renaming of the rows.
+ //
+ // [ 0 1 2 3] [ 0 1 2 3] (a, q8)
+ // [ 4 5 6 7] --> [ 5 6 7 4] (b, q9)
+ // [ 8 9 10 11] [10 11 8 9] (c, q10)
+ // [12 13 14 15] [15 12 13 14] (d, q11)
+ //
+ // The reorderings have for the most part been pushed upwards to
+ // reduce delays.
+ vext.32 q9, q9, q9, #1
+
+ // Apply the diagonal quarterround to each of the columns
+ // simultaneously.
+
+ // a += b; d ^= a; d <<<= 16
+ vadd.u32 q8, q8, q9
+ veor q11, q11, q8
+ vshl.u32 q0, q11, #16
+ vshr.u32 q11, q11, #16
+ vorr q11, q11, q0
+
+ // c += d; b ^= c; b <<<= 12
+ vadd.u32 q10, q10, q11
+ veor q9, q9, q10
+ vshl.u32 q0, q9, #12
+ vshr.u32 q9, q9, #20
+ vorr q9, q9, q0
+
+ // a += b; d ^= a; d <<<= 8
+ vadd.u32 q8, q8, q9
+ veor q11, q11, q8
+ vshl.u32 q0, q11, #8
+ vshr.u32 q11, q11, #24
+ vorr q11, q11, q0
+
+ // c += d; b ^= c; b <<<= 7
+ vadd.u32 q10, q10, q11
+ vext.32 q11, q11, q11, #1
+ veor q9, q9, q10
+ vext.32 q10, q10, q10, #2
+ vshl.u32 q0, q9, #7
+ vshr.u32 q9, q9, #25
+ vorr q9, q9, q0
+
+ // Finally finish off undoing the transpose, and we're done for this
+ // doubleround. Again, most of this was done above so we don't have
+ // to wait for the reorderings.
+ vext.32 q9, q9, q9, #3
+
+ // Decrement the loop counter and see if we should go round again.
+ subs r0, r0, #2
+ bls 9f
+
+ // Do the first part of the next round because this loop is offset.
+
+ // a += b; d ^= a; d <<<= 16
+ vadd.u32 q8, q8, q9
+ veor q11, q11, q8
+ vshl.u32 q0, q11, #16
+ vshr.u32 q11, q11, #16
+ vorr q11, q11, q0
+
+ // c += d; b ^= c; b <<<= 12
+ vadd.u32 q10, q10, q11
+ veor q9, q9, q10
+ vshl.u32 q0, q9, #12
+ vshr.u32 q9, q9, #20
+ vorr q9, q9, q0
+
+ b 0b
+
+ // Almost there. Firstly the feedfoward addition.
+9: vadd.u32 q8, q8, q12
+ vadd.u32 q9, q9, q13
+ vadd.u32 q10, q10, q14
+ vadd.u32 q11, q11, q15
+
+ // And now we write out the result.
+ vstmia r2, {d16-d23}
+
+ // And with that, we're done.
+ bx r14
+
+ENDFUNC
+
+///----- That's all, folks --------------------------------------------------