Skip to contentSkip to conversation

crypto: poly1305 improvements

PATCH Superseded Jason A. Donenfeld 2 participants 6 messages Jan 4, 2020 – Jan 5, 2020

Subsystems (6)
Arm (32-bit) x86 MIPS Crypto API +2 more subsystems

v7 of 8 Superseded Latest is v8 →

lore
Share & tools
lore

Jason A. Donenfeld

author maintainer
#0 lore raw

v7 cover crypto: poly1305 improvements

Version 7 incorporates suggestions from the mailing list on version 6. We now use a union type to handle the poly1305_core_key, per suggestion. And now the changes to the cryptogams code are nicely split out into its own commit, with detail on the (limited) scope of the changes. I believe this should address the last of issues brought up.

####

These are some improvements to the Poly1305 code that I think should be fairly uncontroversial. The first part, the new C implementations, adds cleaner code in two forms that can easily be compared and reviewed, and also results in performance speedups. The second part, the new x86_64 implementation, replaces an slow unvetted implementation with an extremely fast implementation that has received many eyeballs. Finally, we fix up some deadcode.

Jason A. Donenfeld (4):
crypto: poly1305 - add new 32 and 64-bit generic versions
crypto: x86_64/poly1305 - import unmodified cryptogams implementation
crypto: x86_64/poly1305 - wire up faster implementations for kernel
crypto: arm/arm64/mips/poly1305 - remove redundant non-reduction from
emit

 arch/arm/crypto/poly1305-glue.c        |   18 +-
 arch/arm64/crypto/poly1305-glue.c      |   18 +-
 arch/mips/crypto/poly1305-glue.c       |   18 +-
 arch/x86/crypto/.gitignore             |    1 +
 arch/x86/crypto/Makefile               |   11 +-
 arch/x86/crypto/poly1305-avx2-x86_64.S |  390 ---
 arch/x86/crypto/poly1305-sse2-x86_64.S |  590 ----
 arch/x86/crypto/poly1305-x86_64.pl     | 4265 ++++++++++++++++++++++++
 arch/x86/crypto/poly1305_glue.c        |  308 +-
 crypto/adiantum.c                      |    4 +-
 crypto/nhpoly1305.c                    |    2 +-
 crypto/poly1305_generic.c              |   25 +-
 include/crypto/internal/poly1305.h     |   45 +-
 include/crypto/nhpoly1305.h            |    4 +-
 include/crypto/poly1305.h              |   26 +-
 lib/crypto/Kconfig                     |    2 +-
 lib/crypto/Makefile                    |    4 +-
 lib/crypto/poly1305-donna32.c          |  204 ++
 lib/crypto/poly1305-donna64.c          |  185 +
 lib/crypto/poly1305.c                  |  169 +-
 20 files changed, 4924 insertions(+), 1365 deletions(-)
 create mode 100644 arch/x86/crypto/.gitignore
 delete mode 100644 arch/x86/crypto/poly1305-avx2-x86_64.S
 delete mode 100644 arch/x86/crypto/poly1305-sse2-x86_64.S
 create mode 100644 arch/x86/crypto/poly1305-x86_64.pl
 create mode 100644 lib/crypto/poly1305-donna32.c
 create mode 100644 lib/crypto/poly1305-donna64.c
Signature · 1 line
2.24.1
author maintainer
#1 lore raw

PATCH v7 1/4 crypto: poly1305 - add new 32 and 64-bit generic versions

These two C implementations from Zinc -- a 32x32 one and a 64x64 one, depending on the platform -- come from Andrew Moon's public domain poly1305-donna portable code, modified for usage in the kernel. The precomputation in the 32-bit version and the use of 64x64 multiplies in the 64-bit version make these perform better than the code it replaces. Moon's code is also very widespread and has received many eyeballs of scrutiny.

There's a bit of interference between the x86 implementation, which relies on internal details of the old scalar implementation. In the next commit, the x86 implementation will be replaced with a faster one that doesn't rely on this, so none of this matters much. But for now, to keep this passing the tests, we inline the bits of the old implementation that the x86 implementation relied on. Also, since we now support a slightly larger key space, via the union, some offsets had to be fixed up.

Nonce calculation was folded in with the emit function, to take advantage of 64x64 arithmetic. However, Adiantum appeared to rely on no nonce handling in emit, so this path was conditionalized. We also introduced a new struct, poly1305_core_key, to represent the precise amount of space that particular implementation uses.

Testing with kbench9000, depending on the CPU, the update function for the 32x32 version has been improved by 4%-7%, and for the 64x64 by 19%-30%. The 32x32 gains are small, but I think there's great value in having a parallel implementation to the 64x64 one so that the two can be compared side-by-side as nice stand-alone units.

---
 arch/x86/crypto/poly1305-avx2-x86_64.S |  20 +--
 arch/x86/crypto/poly1305_glue.c        | 215 +++++++++++++++++++++++--
 crypto/adiantum.c                      |   4 +-
 crypto/nhpoly1305.c                    |   2 +-
 crypto/poly1305_generic.c              |  25 ++-
 include/crypto/internal/poly1305.h     |  45 ++----
 include/crypto/nhpoly1305.h            |   4 +-
 include/crypto/poly1305.h              |  26 ++-
 lib/crypto/Makefile                    |   4 +-
 lib/crypto/poly1305-donna32.c          | 204 +++++++++++++++++++++++
 lib/crypto/poly1305-donna64.c          | 185 +++++++++++++++++++++
 lib/crypto/poly1305.c                  | 169 +------------------
 12 files changed, 675 insertions(+), 228 deletions(-)
 create mode 100644 lib/crypto/poly1305-donna32.c
 create mode 100644 lib/crypto/poly1305-donna64.c
arch/x86/crypto/poly1305-avx2-x86_64.S +10 -10
index d6063feda9da..8f56989ea599 100644
@@ -34,16 +34,16 @@ ORMASK:	.octa 0x00000000010000000000000001000000
 #define u2 0x08(%r8)
 #define u3 0x0c(%r8)
 #define u4 0x10(%r8)
-#define w0 0x14(%r8)
-#define w1 0x18(%r8)
-#define w2 0x1c(%r8)
-#define w3 0x20(%r8)
-#define w4 0x24(%r8)
-#define y0 0x28(%r8)
-#define y1 0x2c(%r8)
-#define y2 0x30(%r8)
-#define y3 0x34(%r8)
-#define y4 0x38(%r8)
+#define w0 0x18(%r8)
+#define w1 0x1c(%r8)
+#define w2 0x20(%r8)
+#define w3 0x24(%r8)
+#define w4 0x28(%r8)
+#define y0 0x30(%r8)
+#define y1 0x34(%r8)
+#define y2 0x38(%r8)
+#define y3 0x3c(%r8)
+#define y4 0x40(%r8)
 #define m %rsi
 #define hc0 %ymm0
 #define hc1 %ymm1
arch/x86/crypto/poly1305_glue.c +204 -11

252 lines not on this page · raw message

crypto/adiantum.c +2 -2
index aded26092268..c846a887abe1 100644
@@ -72,7 +72,7 @@ struct adiantum_tfm_ctx {
 	struct crypto_skcipher *streamcipher;
 	struct crypto_cipher *blockcipher;
 	struct crypto_shash *hash;
-	struct poly1305_key header_hash_key;
+	struct poly1305_core_key header_hash_key;
 };
 
 struct adiantum_request_ctx {
@@ -249,7 +249,7 @@ static void adiantum_hash_header(struct skcipher_request *req)
 	poly1305_core_blocks(&state, &tctx->header_hash_key, req->iv,
 			     TWEAK_SIZE / POLY1305_BLOCK_SIZE, 1);
 
-	poly1305_core_emit(&state, &rctx->header_hash);
+	poly1305_core_emit(&state, NULL, &rctx->header_hash);
 }
 
 /* Hash the left-hand part (the "bulk") of the message using NHPoly1305 */
crypto/nhpoly1305.c +1 -1
index f6b6a52092b4..8a3006c3b51b 100644
@@ -210,7 +210,7 @@ int crypto_nhpoly1305_final_helper(struct shash_desc *desc, u8 *dst, nh_t nh_fn)
 	if (state->nh_remaining)
 		process_nh_hash_value(state, key);
 
-	poly1305_core_emit(&state->poly_state, dst);
+	poly1305_core_emit(&state->poly_state, NULL, dst);
 	return 0;
 }
 EXPORT_SYMBOL(crypto_nhpoly1305_final_helper);
crypto/poly1305_generic.c +24 -1
index 21edbd8c99fb..94af47eb6fa6 100644
@@ -31,6 +31,29 @@ static int crypto_poly1305_init(struct shash_desc *desc)
 	return 0;
 }
 
+static unsigned int crypto_poly1305_setdesckey(struct poly1305_desc_ctx *dctx,
+					       const u8 *src, unsigned int srclen)
+{
+	if (!dctx->sset) {
+		if (!dctx->rset && srclen >= POLY1305_BLOCK_SIZE) {
+			poly1305_core_setkey(&dctx->core_r, src);
+			src += POLY1305_BLOCK_SIZE;
+			srclen -= POLY1305_BLOCK_SIZE;
+			dctx->rset = 2;
+		}
+		if (srclen >= POLY1305_BLOCK_SIZE) {
+			dctx->s[0] = get_unaligned_le32(src +  0);
+			dctx->s[1] = get_unaligned_le32(src +  4);
+			dctx->s[2] = get_unaligned_le32(src +  8);
+			dctx->s[3] = get_unaligned_le32(src + 12);
+			src += POLY1305_BLOCK_SIZE;
+			srclen -= POLY1305_BLOCK_SIZE;
+			dctx->sset = true;
+		}
+	}
+	return srclen;
+}
+
 static void poly1305_blocks(struct poly1305_desc_ctx *dctx, const u8 *src,
 			    unsigned int srclen)
 {
@@ -42,7 +65,7 @@ static void poly1305_blocks(struct poly1305_desc_ctx *dctx, const u8 *src,
 		srclen = datalen;
 	}
 
-	poly1305_core_blocks(&dctx->h, dctx->r, src,
+	poly1305_core_blocks(&dctx->h, &dctx->core_r, src,
 			     srclen / POLY1305_BLOCK_SIZE, 1);
 }
 
include/crypto/internal/poly1305.h +10 -35

58 lines not on this page · raw message

include/crypto/nhpoly1305.h +2 -2
index 53c04423c582..306925fea190 100644
@@ -7,7 +7,7 @@
 #define _NHPOLY1305_H
 
 #include <crypto/hash.h>
-#include <crypto/poly1305.h>
+#include <crypto/internal/poly1305.h>
 
 /* NH parameterization: */
 
@@ -33,7 +33,7 @@
 #define NHPOLY1305_KEY_SIZE	(POLY1305_BLOCK_SIZE + NH_KEY_BYTES)
 
 struct nhpoly1305_key {
-	struct poly1305_key poly_key;
+	struct poly1305_core_key poly_key;
 	u32 nh_key[NH_KEY_WORDS];
 };
 
include/crypto/poly1305.h +23 -3
index 74c6e1cd73ee..f1f67fc749cf 100644
@@ -13,12 +13,29 @@
 #define POLY1305_KEY_SIZE	32
 #define POLY1305_DIGEST_SIZE	16
 
+/* The poly1305_key and poly1305_state types are mostly opaque and
+ * implementation-defined. Limbs might be in base 2^64 or base 2^26, or
+ * different yet. The union type provided keeps these 64-bit aligned for the
+ * case in which this is implemented using 64x64 multiplies.
+ */
+
 struct poly1305_key {
-	u32 r[5];	/* key, base 2^26 */
+	union {
+		u32 r[5];
+		u64 r64[3];
+	};
+};
+
+struct poly1305_core_key {
+	struct poly1305_key key;
+	struct poly1305_key precomputed_s;
 };
 
 struct poly1305_state {
-	u32 h[5];	/* accumulator, base 2^26 */
+	union {
+		u32 h[5];
+		u64 h64[3];
+	};
 };
 
 struct poly1305_desc_ctx {
@@ -35,7 +52,10 @@ struct poly1305_desc_ctx {
 	/* accumulator */
 	struct poly1305_state h;
 	/* key */
-	struct poly1305_key r[CONFIG_CRYPTO_LIB_POLY1305_RSIZE];
+	union {
+		struct poly1305_key opaque_r[CONFIG_CRYPTO_LIB_POLY1305_RSIZE];
+		struct poly1305_core_key core_r;
+	};
 };
 
 void poly1305_init_arch(struct poly1305_desc_ctx *desc, const u8 *key);
lib/crypto/Makefile +3 -1
index 34a701ab8b92..d19e5195569e 100644
@@ -28,7 +28,9 @@ obj-$(CONFIG_CRYPTO_LIB_DES)			+= libdes.o
 libdes-y					:= des.o
 
 obj-$(CONFIG_CRYPTO_LIB_POLY1305_GENERIC)	+= libpoly1305.o
-libpoly1305-y					:= poly1305.o
+libpoly1305-y					:= poly1305-donna32.o
+libpoly1305-$(CONFIG_ARCH_SUPPORTS_INT128)	:= poly1305-donna64.o
+libpoly1305-y					+= poly1305.o
 
 obj-$(CONFIG_CRYPTO_LIB_SHA256)			+= libsha256.o
 libsha256-y					:= sha256.o
lib/crypto/poly1305-donna32.c new file +204 -0

204 lines not on this page · raw message

lib/crypto/poly1305-donna64.c new file +185 -0

185 lines not on this page · raw message

lib/crypto/poly1305.c +7 -162

206 lines not on this page · raw message

Signature · 1 line
2.24.1
author maintainer
#2 lore raw

PATCH v7 2/4 crypto: x86_64/poly1305 - import unmodified cryptogams implementation

In reply to Jason A. Donenfeld

These x86_64 vectorized implementations come from Andy Polyakov's CRYPTOGAMS implementation, and are included here in raw form without modification, so that subsequent commits that fix these up for the kernel can see how it has changed.

---
 arch/x86/crypto/poly1305-x86_64.pl | 4159 ++++++++++++++++++++++++++++
 1 file changed, 4159 insertions(+)
 create mode 100644 arch/x86/crypto/poly1305-x86_64.pl
arch/x86/crypto/poly1305-x86_64.pl new file +4159 -0
new file mode 100644
index 000000000000..342ad7f18aa7
@@ -0,0 +1,4159 @@
+#! /usr/bin/env perl
+# Copyright 2016-2018 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License").  You may not use
+# this file except in compliance with the License.  You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+#
+# ====================================================================
+# Written by Andy Polyakov <[email protected]> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# This module implements Poly1305 hash for x86_64.
+#
+# March 2015
+#
+# Initial release.
+#
+# December 2016
+#
+# Add AVX512F+VL+BW code path.
+#
+# November 2017
+#
+# Convert AVX512F+VL+BW code path to pure AVX512F, so that it can be
+# executed even on Knights Landing. Trigger for modification was
+# observation that AVX512 code paths can negatively affect overall
+# Skylake-X system performance. Since we are likely to suppress
+# AVX512F capability flag [at least on Skylake-X], conversion serves
+# as kind of "investment protection". Note that next *lake processor,
+# Cannolake, has AVX512IFMA code path to execute...
+#
+# Numbers are cycles per processed byte with poly1305_blocks alone,
+# measured with rdtsc at fixed clock frequency.
+#
+#		IALU/gcc-4.8(*)	AVX(**)		AVX2	AVX-512
+# P4		4.46/+120%	-
+# Core 2	2.41/+90%	-
+# Westmere	1.88/+120%	-
+# Sandy Bridge	1.39/+140%	1.10
+# Haswell	1.14/+175%	1.11		0.65
+# Skylake[-X]	1.13/+120%	0.96		0.51	[0.35]
+# Silvermont	2.83/+95%	-
+# Knights L	3.60/?		1.65		1.10	0.41(***)
+# Goldmont	1.70/+180%	-
+# VIA Nano	1.82/+150%	-
+# Sledgehammer	1.38/+160%	-
+# Bulldozer	2.30/+130%	0.97
+# Ryzen		1.15/+200%	1.08		1.18
+#
+# (*)	improvement coefficients relative to clang are more modest and
+#	are ~50% on most processors, in both cases we are comparing to
+#	__int128 code;
+# (**)	SSE2 implementation was attempted, but among non-AVX processors
+#	it was faster than integer-only code only on older Intel P4 and
+#	Core processors, 50-30%, less newer processor is, but slower on
+#	contemporary ones, for example almost 2x slower on Atom, and as
+#	former are naturally disappearing, SSE2 is deemed unnecessary;
+# (***)	strangely enough performance seems to vary from core to core,
+#	listed result is best case;
+
+$flavour = shift;
+$output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
+		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
+	$avx = ($1>=2.19) + ($1>=2.22) + ($1>=2.25) + ($1>=2.26);
+}
+
+if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
+	   `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)(?:\.([0-9]+))?/) {
+	$avx = ($1>=2.09) + ($1>=2.10) + 2 * ($1>=2.12);
+	$avx += 2 if ($1==2.11 && $2>=8);
+}
+
+if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
+	   `ml64 2>&1` =~ /Version ([0-9]+)\./) {
+	$avx = ($1>=10) + ($1>=12);
+}
+
+if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
+	$avx = ($2>=3.0) + ($2>3.0);
+}
+
+open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
+*STDOUT=*OUT;
+
+my ($ctx,$inp,$len,$padbit)=("%rdi","%rsi","%rdx","%rcx");
+my ($mac,$nonce)=($inp,$len);	# *_emit arguments
+my ($d1,$d2,$d3, $r0,$r1,$s1)=map("%r$_",(8..13));
+my ($h0,$h1,$h2)=("%r14","%rbx","%rbp");
+
+sub poly1305_iteration {
+# input:	copy of $r1 in %rax, $h0-$h2, $r0-$r1
+# output:	$h0-$h2 *= $r0-$r1
+$code.=<<___;
+	mulq	$h0			# h0*r1
+	mov	%rax,$d2
+	 mov	$r0,%rax
+	mov	%rdx,$d3
+
+	mulq	$h0			# h0*r0
+	mov	%rax,$h0		# future $h0
+	 mov	$r0,%rax
+	mov	%rdx,$d1
+
+	mulq	$h1			# h1*r0
+	add	%rax,$d2
+	 mov	$s1,%rax
+	adc	%rdx,$d3
+
+	mulq	$h1			# h1*s1
+	 mov	$h2,$h1			# borrow $h1
+	add	%rax,$h0
+	adc	%rdx,$d1
+
+	imulq	$s1,$h1			# h2*s1
+	add	$h1,$d2
+	 mov	$d1,$h1
+	adc	\$0,$d3
+
+	imulq	$r0,$h2			# h2*r0
+	add	$d2,$h1
+	mov	\$-4,%rax		# mask value
+	adc	$h2,$d3
+
+	and	$d3,%rax		# last reduction step
+	mov	$d3,$h2
+	shr	\$2,$d3
+	and	\$3,$h2
+	add	$d3,%rax
+	add	%rax,$h0
+	adc	\$0,$h1
+	adc	\$0,$h2
+___
+}
+
+########################################################################
+# Layout of opaque area is following.
+#
+#	unsigned __int64 h[3];		# current hash value base 2^64
+#	unsigned __int64 r[2];		# key value base 2^64
+
+$code.=<<___;
+.text
+
+.extern	OPENSSL_ia32cap_P
+
+.globl	poly1305_init
+.hidden	poly1305_init
+.globl	poly1305_blocks
+.hidden	poly1305_blocks
+.globl	poly1305_emit
+.hidden	poly1305_emit
+
+.type	poly1305_init,\@function,3
+.align	32
+poly1305_init:
+	xor	%rax,%rax
+	mov	%rax,0($ctx)		# initialize hash value
+	mov	%rax,8($ctx)
+	mov	%rax,16($ctx)
+
+	cmp	\$0,$inp
+	je	.Lno_key
+
+	lea	poly1305_blocks(%rip),%r10
+	lea	poly1305_emit(%rip),%r11
+___
+$code.=<<___	if ($avx);
+	mov	OPENSSL_ia32cap_P+4(%rip),%r9
+	lea	poly1305_blocks_avx(%rip),%rax
+	lea	poly1305_emit_avx(%rip),%rcx
+	bt	\$`60-32`,%r9		# AVX?
+	cmovc	%rax,%r10
+	cmovc	%rcx,%r11
+___
+$code.=<<___	if ($avx>1);
+	lea	poly1305_blocks_avx2(%rip),%rax
+	bt	\$`5+32`,%r9		# AVX2?
+	cmovc	%rax,%r10
+___
+$code.=<<___	if ($avx>3);
+	mov	\$`(1<<31|1<<21|1<<16)`,%rax
+	shr	\$32,%r9
+	and	%rax,%r9
+	cmp	%rax,%r9
+	je	.Linit_base2_44
+___
+$code.=<<___;
+	mov	\$0x0ffffffc0fffffff,%rax
+	mov	\$0x0ffffffc0ffffffc,%rcx
+	and	0($inp),%rax
+	and	8($inp),%rcx
+	mov	%rax,24($ctx)
+	mov	%rcx,32($ctx)
+___
+$code.=<<___	if ($flavour !~ /elf32/);
+	mov	%r10,0(%rdx)
+	mov	%r11,8(%rdx)
+___
+$code.=<<___	if ($flavour =~ /elf32/);
+	mov	%r10d,0(%rdx)
+	mov	%r11d,4(%rdx)
+___
+$code.=<<___;
+	mov	\$1,%eax
+.Lno_key:
+	ret
+.size	poly1305_init,.-poly1305_init
+
+.type	poly1305_blocks,\@function,4
+.align	32
+poly1305_blocks:
+.cfi_startproc
+.Lblocks:
+	shr	\$4,$len
+	jz	.Lno_data		# too short
+
+	push	%rbx
+.cfi_push	%rbx
+	push	%rbp
+.cfi_push	%rbp
+	push	%r12
+.cfi_push	%r12
+	push	%r13
+.cfi_push	%r13
+	push	%r14
+.cfi_push	%r14
+	push	%r15
+.cfi_push	%r15
+.Lblocks_body:
+
+	mov	$len,%r15		# reassign $len
+
+	mov	24($ctx),$r0		# load r
+	mov	32($ctx),$s1
+
+	mov	0($ctx),$h0		# load hash value
+	mov	8($ctx),$h1
+	mov	16($ctx),$h2
+
+	mov	$s1,$r1
+	shr	\$2,$s1
+	mov	$r1,%rax
+	add	$r1,$s1			# s1 = r1 + (r1 >> 2)
+	jmp	.Loop
+
+.align	32
+.Loop:
+	add	0($inp),$h0		# accumulate input
+	adc	8($inp),$h1
+	lea	16($inp),$inp
+	adc	$padbit,$h2
+___
+	&poly1305_iteration();
+$code.=<<___;
+	mov	$r1,%rax
+	dec	%r15			# len-=16
+	jnz	.Loop
+
+	mov	$h0,0($ctx)		# store hash value
+	mov	$h1,8($ctx)
+	mov	$h2,16($ctx)
+
+	mov	0(%rsp),%r15
+.cfi_restore	%r15
+	mov	8(%rsp),%r14
+.cfi_restore	%r14
+	mov	16(%rsp),%r13
+.cfi_restore	%r13
+	mov	24(%rsp),%r12
+.cfi_restore	%r12
+	mov	32(%rsp),%rbp
+.cfi_restore	%rbp
+	mov	40(%rsp),%rbx
+.cfi_restore	%rbx
+	lea	48(%rsp),%rsp
+.cfi_adjust_cfa_offset	-48
+.Lno_data:
+.Lblocks_epilogue:
+	ret
+.cfi_endproc
+.size	poly1305_blocks,.-poly1305_blocks
+
+.type	poly1305_emit,\@function,3
+.align	32
+poly1305_emit:
+.Lemit:
+	mov	0($ctx),%r8	# load hash value
+	mov	8($ctx),%r9
+	mov	16($ctx),%r10
+
+	mov	%r8,%rax
+	add	\$5,%r8		# compare to modulus
+	mov	%r9,%rcx
+	adc	\$0,%r9
+	adc	\$0,%r10
+	shr	\$2,%r10	# did 130-bit value overflow?
+	cmovnz	%r8,%rax
+	cmovnz	%r9,%rcx
+
+	add	0($nonce),%rax	# accumulate nonce
+	adc	8($nonce),%rcx
+	mov	%rax,0($mac)	# write result
+	mov	%rcx,8($mac)
+
+	ret
+.size	poly1305_emit,.-poly1305_emit
+___
+if ($avx) {
+
+########################################################################
+# Layout of opaque area is following.
+#
+#	unsigned __int32 h[5];		# current hash value base 2^26
+#	unsigned __int32 is_base2_26;
+#	unsigned __int64 r[2];		# key value base 2^64
+#	unsigned __int64 pad;
+#	struct { unsigned __int32 r^2, r^1, r^4, r^3; } r[9];
+#
+# where r^n are base 2^26 digits of degrees of multiplier key. There are
+# 5 digits, but last four are interleaved with multiples of 5, totalling
+# in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
+
+my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
+    map("%xmm$_",(0..15));
+
+$code.=<<___;
+.type	__poly1305_block,\@abi-omnipotent
+.align	32
+__poly1305_block:
+___
+	&poly1305_iteration();
+$code.=<<___;
+	ret
+.size	__poly1305_block,.-__poly1305_block
+
+.type	__poly1305_init_avx,\@abi-omnipotent
+.align	32
+__poly1305_init_avx:
+	mov	$r0,$h0
+	mov	$r1,$h1
+	xor	$h2,$h2
+
+	lea	48+64($ctx),$ctx	# size optimization
+
+	mov	$r1,%rax
+	call	__poly1305_block	# r^2
+
+	mov	\$0x3ffffff,%eax	# save interleaved r^2 and r base 2^26
+	mov	\$0x3ffffff,%edx
+	mov	$h0,$d1
+	and	$h0#d,%eax
+	mov	$r0,$d2
+	and	$r0#d,%edx
+	mov	%eax,`16*0+0-64`($ctx)
+	shr	\$26,$d1
+	mov	%edx,`16*0+4-64`($ctx)
+	shr	\$26,$d2
+
+	mov	\$0x3ffffff,%eax
+	mov	\$0x3ffffff,%edx
+	and	$d1#d,%eax
+	and	$d2#d,%edx
+	mov	%eax,`16*1+0-64`($ctx)
+	lea	(%rax,%rax,4),%eax	# *5
+	mov	%edx,`16*1+4-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	mov	%eax,`16*2+0-64`($ctx)
+	shr	\$26,$d1
+	mov	%edx,`16*2+4-64`($ctx)
+	shr	\$26,$d2
+
+	mov	$h1,%rax
+	mov	$r1,%rdx
+	shl	\$12,%rax
+	shl	\$12,%rdx
+	or	$d1,%rax
+	or	$d2,%rdx
+	and	\$0x3ffffff,%eax
+	and	\$0x3ffffff,%edx
+	mov	%eax,`16*3+0-64`($ctx)
+	lea	(%rax,%rax,4),%eax	# *5
+	mov	%edx,`16*3+4-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	mov	%eax,`16*4+0-64`($ctx)
+	mov	$h1,$d1
+	mov	%edx,`16*4+4-64`($ctx)
+	mov	$r1,$d2
+
+	mov	\$0x3ffffff,%eax
+	mov	\$0x3ffffff,%edx
+	shr	\$14,$d1
+	shr	\$14,$d2
+	and	$d1#d,%eax
+	and	$d2#d,%edx
+	mov	%eax,`16*5+0-64`($ctx)
+	lea	(%rax,%rax,4),%eax	# *5
+	mov	%edx,`16*5+4-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	mov	%eax,`16*6+0-64`($ctx)
+	shr	\$26,$d1
+	mov	%edx,`16*6+4-64`($ctx)
+	shr	\$26,$d2
+
+	mov	$h2,%rax
+	shl	\$24,%rax
+	or	%rax,$d1
+	mov	$d1#d,`16*7+0-64`($ctx)
+	lea	($d1,$d1,4),$d1		# *5
+	mov	$d2#d,`16*7+4-64`($ctx)
+	lea	($d2,$d2,4),$d2		# *5
+	mov	$d1#d,`16*8+0-64`($ctx)
+	mov	$d2#d,`16*8+4-64`($ctx)
+
+	mov	$r1,%rax
+	call	__poly1305_block	# r^3
+
+	mov	\$0x3ffffff,%eax	# save r^3 base 2^26
+	mov	$h0,$d1
+	and	$h0#d,%eax
+	shr	\$26,$d1
+	mov	%eax,`16*0+12-64`($ctx)
+
+	mov	\$0x3ffffff,%edx
+	and	$d1#d,%edx
+	mov	%edx,`16*1+12-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	shr	\$26,$d1
+	mov	%edx,`16*2+12-64`($ctx)
+
+	mov	$h1,%rax
+	shl	\$12,%rax
+	or	$d1,%rax
+	and	\$0x3ffffff,%eax
+	mov	%eax,`16*3+12-64`($ctx)
+	lea	(%rax,%rax,4),%eax	# *5
+	mov	$h1,$d1
+	mov	%eax,`16*4+12-64`($ctx)
+
+	mov	\$0x3ffffff,%edx
+	shr	\$14,$d1
+	and	$d1#d,%edx
+	mov	%edx,`16*5+12-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	shr	\$26,$d1
+	mov	%edx,`16*6+12-64`($ctx)
+
+	mov	$h2,%rax
+	shl	\$24,%rax
+	or	%rax,$d1
+	mov	$d1#d,`16*7+12-64`($ctx)
+	lea	($d1,$d1,4),$d1		# *5
+	mov	$d1#d,`16*8+12-64`($ctx)
+
+	mov	$r1,%rax
+	call	__poly1305_block	# r^4
+
+	mov	\$0x3ffffff,%eax	# save r^4 base 2^26
+	mov	$h0,$d1
+	and	$h0#d,%eax
+	shr	\$26,$d1
+	mov	%eax,`16*0+8-64`($ctx)
+
+	mov	\$0x3ffffff,%edx
+	and	$d1#d,%edx
+	mov	%edx,`16*1+8-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	shr	\$26,$d1
+	mov	%edx,`16*2+8-64`($ctx)
+
+	mov	$h1,%rax
+	shl	\$12,%rax
+	or	$d1,%rax
+	and	\$0x3ffffff,%eax
+	mov	%eax,`16*3+8-64`($ctx)
+	lea	(%rax,%rax,4),%eax	# *5
+	mov	$h1,$d1
+	mov	%eax,`16*4+8-64`($ctx)
+
+	mov	\$0x3ffffff,%edx
+	shr	\$14,$d1
+	and	$d1#d,%edx
+	mov	%edx,`16*5+8-64`($ctx)
+	lea	(%rdx,%rdx,4),%edx	# *5
+	shr	\$26,$d1
+	mov	%edx,`16*6+8-64`($ctx)
+
+	mov	$h2,%rax
+	shl	\$24,%rax
+	or	%rax,$d1
+	mov	$d1#d,`16*7+8-64`($ctx)
+	lea	($d1,$d1,4),$d1		# *5
+	mov	$d1#d,`16*8+8-64`($ctx)
+
+	lea	-48-64($ctx),$ctx	# size [de-]optimization
+	ret
+.size	__poly1305_init_avx,.-__poly1305_init_avx
+
+.type	poly1305_blocks_avx,\@function,4
+.align	32
+poly1305_blocks_avx:
+.cfi_startproc
+	mov	20($ctx),%r8d		# is_base2_26
+	cmp	\$128,$len
+	jae	.Lblocks_avx
+	test	%r8d,%r8d
+	jz	.Lblocks
+
+.Lblocks_avx:
+	and	\$-16,$len
+	jz	.Lno_data_avx
+
+	vzeroupper
+
+	test	%r8d,%r8d
+	jz	.Lbase2_64_avx
+
+	test	\$31,$len
+	jz	.Leven_avx
+
+	push	%rbx
+.cfi_push	%rbx
+	push	%rbp
+.cfi_push	%rbp
+	push	%r12
+.cfi_push	%r12
+	push	%r13
+.cfi_push	%r13
+	push	%r14
+.cfi_push	%r14
+	push	%r15
+.cfi_push	%r15
+.Lblocks_avx_body:
+
+	mov	$len,%r15		# reassign $len
+
+	mov	0($ctx),$d1		# load hash value
+	mov	8($ctx),$d2
+	mov	16($ctx),$h2#d
+
+	mov	24($ctx),$r0		# load r
+	mov	32($ctx),$s1
+
+	################################# base 2^26 -> base 2^64
+	mov	$d1#d,$h0#d
+	and	\$`-1*(1<<31)`,$d1
+	mov	$d2,$r1			# borrow $r1
+	mov	$d2#d,$h1#d
+	and	\$`-1*(1<<31)`,$d2
+
+	shr	\$6,$d1
+	shl	\$52,$r1
+	add	$d1,$h0
+	shr	\$12,$h1
+	shr	\$18,$d2
+	add	$r1,$h0
+	adc	$d2,$h1
+
+	mov	$h2,$d1
+	shl	\$40,$d1
+	shr	\$24,$h2
+	add	$d1,$h1
+	adc	\$0,$h2			# can be partially reduced...
+
+	mov	\$-4,$d2		# ... so reduce
+	mov	$h2,$d1
+	and	$h2,$d2
+	shr	\$2,$d1
+	and	\$3,$h2
+	add	$d2,$d1			# =*5
+	add	$d1,$h0
+	adc	\$0,$h1
+	adc	\$0,$h2
+
+	mov	$s1,$r1
+	mov	$s1,%rax
+	shr	\$2,$s1
+	add	$r1,$s1			# s1 = r1 + (r1 >> 2)
+
+	add	0($inp),$h0		# accumulate input
+	adc	8($inp),$h1
+	lea	16($inp),$inp
+	adc	$padbit,$h2
+
+	call	__poly1305_block
+
+	test	$padbit,$padbit		# if $padbit is zero,
+	jz	.Lstore_base2_64_avx	# store hash in base 2^64 format
+
+	################################# base 2^64 -> base 2^26
+	mov	$h0,%rax
+	mov	$h0,%rdx
+	shr	\$52,$h0
+	mov	$h1,$r0
+	mov	$h1,$r1
+	shr	\$26,%rdx
+	and	\$0x3ffffff,%rax	# h[0]
+	shl	\$12,$r0
+	and	\$0x3ffffff,%rdx	# h[1]
+	shr	\$14,$h1
+	or	$r0,$h0
+	shl	\$24,$h2
+	and	\$0x3ffffff,$h0		# h[2]
+	shr	\$40,$r1
+	and	\$0x3ffffff,$h1		# h[3]
+	or	$r1,$h2			# h[4]
+
+	sub	\$16,%r15
+	jz	.Lstore_base2_26_avx
+
+	vmovd	%rax#d,$H0
+	vmovd	%rdx#d,$H1
+	vmovd	$h0#d,$H2
+	vmovd	$h1#d,$H3
+	vmovd	$h2#d,$H4
+	jmp	.Lproceed_avx
+
+.align	32
+.Lstore_base2_64_avx:
+	mov	$h0,0($ctx)
+	mov	$h1,8($ctx)
+	mov	$h2,16($ctx)		# note that is_base2_26 is zeroed
+	jmp	.Ldone_avx
+
+.align	16
+.Lstore_base2_26_avx:
+	mov	%rax#d,0($ctx)		# store hash value base 2^26
+	mov	%rdx#d,4($ctx)
+	mov	$h0#d,8($ctx)
+	mov	$h1#d,12($ctx)
+	mov	$h2#d,16($ctx)
+.align	16
+.Ldone_avx:
+	mov	0(%rsp),%r15
+.cfi_restore	%r15
+	mov	8(%rsp),%r14
+.cfi_restore	%r14
+	mov	16(%rsp),%r13
+.cfi_restore	%r13
+	mov	24(%rsp),%r12
+.cfi_restore	%r12
+	mov	32(%rsp),%rbp
+.cfi_restore	%rbp
+	mov	40(%rsp),%rbx
+.cfi_restore	%rbx
+	lea	48(%rsp),%rsp
+.cfi_adjust_cfa_offset	-48
+.Lno_data_avx:
+.Lblocks_avx_epilogue:
+	ret
+.cfi_endproc
+
+.align	32
+.Lbase2_64_avx:
+.cfi_startproc
+	push	%rbx
+.cfi_push	%rbx
+	push	%rbp
+.cfi_push	%rbp
+	push	%r12
+.cfi_push	%r12
+	push	%r13
+.cfi_push	%r13
+	push	%r14
+.cfi_push	%r14
+	push	%r15
+.cfi_push	%r15
+.Lbase2_64_avx_body:
+
+	mov	$len,%r15		# reassign $len
+
+	mov	24($ctx),$r0		# load r
+	mov	32($ctx),$s1
+
+	mov	0($ctx),$h0		# load hash value
+	mov	8($ctx),$h1
+	mov	16($ctx),$h2#d
+
+	mov	$s1,$r1
+	mov	$s1,%rax
+	shr	\$2,$s1
+	add	$r1,$s1			# s1 = r1 + (r1 >> 2)
+
+	test	\$31,$len
+	jz	.Linit_avx
+
+	add	0($inp),$h0		# accumulate input
+	adc	8($inp),$h1
+	lea	16($inp),$inp
+	adc	$padbit,$h2
+	sub	\$16,%r15
+
+	call	__poly1305_block
+
+.Linit_avx:
+	################################# base 2^64 -> base 2^26
+	mov	$h0,%rax
+	mov	$h0,%rdx
+	shr	\$52,$h0
+	mov	$h1,$d1
+	mov	$h1,$d2
+	shr	\$26,%rdx
+	and	\$0x3ffffff,%rax	# h[0]
+	shl	\$12,$d1
+	and	\$0x3ffffff,%rdx	# h[1]
+	shr	\$14,$h1
+	or	$d1,$h0
+	shl	\$24,$h2
+	and	\$0x3ffffff,$h0		# h[2]
+	shr	\$40,$d2
+	and	\$0x3ffffff,$h1		# h[3]
+	or	$d2,$h2			# h[4]
+
+	vmovd	%rax#d,$H0
+	vmovd	%rdx#d,$H1
+	vmovd	$h0#d,$H2
+	vmovd	$h1#d,$H3
+	vmovd	$h2#d,$H4
+	movl	\$1,20($ctx)		# set is_base2_26
+
+	call	__poly1305_init_avx
+
+.Lproceed_avx:
+	mov	%r15,$len
+
+	mov	0(%rsp),%r15
+.cfi_restore	%r15
+	mov	8(%rsp),%r14
+.cfi_restore	%r14
+	mov	16(%rsp),%r13
+.cfi_restore	%r13
+	mov	24(%rsp),%r12
+.cfi_restore	%r12
+	mov	32(%rsp),%rbp
+.cfi_restore	%rbp
+	mov	40(%rsp),%rbx
+.cfi_restore	%rbx
+	lea	48(%rsp),%rax
+	lea	48(%rsp),%rsp
+.cfi_adjust_cfa_offset	-48
+.Lbase2_64_avx_epilogue:
+	jmp	.Ldo_avx
+.cfi_endproc
+
+.align	32
+.Leven_avx:
+.cfi_startproc
+	vmovd		4*0($ctx),$H0		# load hash value
+	vmovd		4*1($ctx),$H1
+	vmovd		4*2($ctx),$H2
+	vmovd		4*3($ctx),$H3
+	vmovd		4*4($ctx),$H4
+
+.Ldo_avx:
+___
+$code.=<<___	if (!$win64);
+	lea		-0x58(%rsp),%r11
+.cfi_def_cfa		%r11,0x60
+	sub		\$0x178,%rsp
+___
+$code.=<<___	if ($win64);
+	lea		-0xf8(%rsp),%r11
+	sub		\$0x218,%rsp
+	vmovdqa		%xmm6,0x50(%r11)
+	vmovdqa		%xmm7,0x60(%r11)
+	vmovdqa		%xmm8,0x70(%r11)
+	vmovdqa		%xmm9,0x80(%r11)
+	vmovdqa		%xmm10,0x90(%r11)
+	vmovdqa		%xmm11,0xa0(%r11)
+	vmovdqa		%xmm12,0xb0(%r11)
+	vmovdqa		%xmm13,0xc0(%r11)
+	vmovdqa		%xmm14,0xd0(%r11)
+	vmovdqa		%xmm15,0xe0(%r11)
+.Ldo_avx_body:
+___
+$code.=<<___;
+	sub		\$64,$len
+	lea		-32($inp),%rax
+	cmovc		%rax,$inp
+
+	vmovdqu		`16*3`($ctx),$D4	# preload r0^2
+	lea		`16*3+64`($ctx),$ctx	# size optimization
+	lea		.Lconst(%rip),%rcx
+
+	################################################################
+	# load input
+	vmovdqu		16*2($inp),$T0
+	vmovdqu		16*3($inp),$T1
+	vmovdqa		64(%rcx),$MASK		# .Lmask26
+
+	vpsrldq		\$6,$T0,$T2		# splat input
+	vpsrldq		\$6,$T1,$T3
+	vpunpckhqdq	$T1,$T0,$T4		# 4
+	vpunpcklqdq	$T1,$T0,$T0		# 0:1
+	vpunpcklqdq	$T3,$T2,$T3		# 2:3
+
+	vpsrlq		\$40,$T4,$T4		# 4
+	vpsrlq		\$26,$T0,$T1
+	vpand		$MASK,$T0,$T0		# 0
+	vpsrlq		\$4,$T3,$T2
+	vpand		$MASK,$T1,$T1		# 1
+	vpsrlq		\$30,$T3,$T3
+	vpand		$MASK,$T2,$T2		# 2
+	vpand		$MASK,$T3,$T3		# 3
+	vpor		32(%rcx),$T4,$T4	# padbit, yes, always
+
+	jbe		.Lskip_loop_avx
+
+	# expand and copy pre-calculated table to stack
+	vmovdqu		`16*1-64`($ctx),$D1
+	vmovdqu		`16*2-64`($ctx),$D2
+	vpshufd		\$0xEE,$D4,$D3		# 34xx -> 3434
+	vpshufd		\$0x44,$D4,$D0		# xx12 -> 1212
+	vmovdqa		$D3,-0x90(%r11)
+	vmovdqa		$D0,0x00(%rsp)
+	vpshufd		\$0xEE,$D1,$D4
+	vmovdqu		`16*3-64`($ctx),$D0
+	vpshufd		\$0x44,$D1,$D1
+	vmovdqa		$D4,-0x80(%r11)
+	vmovdqa		$D1,0x10(%rsp)
+	vpshufd		\$0xEE,$D2,$D3
+	vmovdqu		`16*4-64`($ctx),$D1
+	vpshufd		\$0x44,$D2,$D2
+	vmovdqa		$D3,-0x70(%r11)
+	vmovdqa		$D2,0x20(%rsp)
+	vpshufd		\$0xEE,$D0,$D4
+	vmovdqu		`16*5-64`($ctx),$D2
+	vpshufd		\$0x44,$D0,$D0
+	vmovdqa		$D4,-0x60(%r11)
+	vmovdqa		$D0,0x30(%rsp)
+	vpshufd		\$0xEE,$D1,$D3
+	vmovdqu		`16*6-64`($ctx),$D0
+	vpshufd		\$0x44,$D1,$D1
+	vmovdqa		$D3,-0x50(%r11)
+	vmovdqa		$D1,0x40(%rsp)
+	vpshufd		\$0xEE,$D2,$D4
+	vmovdqu		`16*7-64`($ctx),$D1
+	vpshufd		\$0x44,$D2,$D2
+	vmovdqa		$D4,-0x40(%r11)
+	vmovdqa		$D2,0x50(%rsp)
+	vpshufd		\$0xEE,$D0,$D3
+	vmovdqu		`16*8-64`($ctx),$D2
+	vpshufd		\$0x44,$D0,$D0
+	vmovdqa		$D3,-0x30(%r11)
+	vmovdqa		$D0,0x60(%rsp)
+	vpshufd		\$0xEE,$D1,$D4
+	vpshufd		\$0x44,$D1,$D1
+	vmovdqa		$D4,-0x20(%r11)
+	vmovdqa		$D1,0x70(%rsp)
+	vpshufd		\$0xEE,$D2,$D3
+	 vmovdqa	0x00(%rsp),$D4		# preload r0^2
+	vpshufd		\$0x44,$D2,$D2
+	vmovdqa		$D3,-0x10(%r11)
+	vmovdqa		$D2,0x80(%rsp)
+
+	jmp		.Loop_avx
+
+.align	32
+.Loop_avx:
+	################################################################
+	# ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
+	# ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
+	#   \___________________/
+	# ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
+	# ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
+	#   \___________________/ \____________________/
+	#
+	# Note that we start with inp[2:3]*r^2. This is because it
+	# doesn't depend on reduction in previous iteration.
+	################################################################
+	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
+	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
+	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
+	#
+	# though note that $Tx and $Hx are "reversed" in this section,
+	# and $D4 is preloaded with r0^2...
+
+	vpmuludq	$T0,$D4,$D0		# d0 = h0*r0
+	vpmuludq	$T1,$D4,$D1		# d1 = h1*r0
+	  vmovdqa	$H2,0x20(%r11)				# offload hash
+	vpmuludq	$T2,$D4,$D2		# d3 = h2*r0
+	 vmovdqa	0x10(%rsp),$H2		# r1^2
+	vpmuludq	$T3,$D4,$D3		# d3 = h3*r0
+	vpmuludq	$T4,$D4,$D4		# d4 = h4*r0
+
+	  vmovdqa	$H0,0x00(%r11)				#
+	vpmuludq	0x20(%rsp),$T4,$H0	# h4*s1
+	  vmovdqa	$H1,0x10(%r11)				#
+	vpmuludq	$T3,$H2,$H1		# h3*r1
+	vpaddq		$H0,$D0,$D0		# d0 += h4*s1
+	vpaddq		$H1,$D4,$D4		# d4 += h3*r1
+	  vmovdqa	$H3,0x30(%r11)				#
+	vpmuludq	$T2,$H2,$H0		# h2*r1
+	vpmuludq	$T1,$H2,$H1		# h1*r1
+	vpaddq		$H0,$D3,$D3		# d3 += h2*r1
+	 vmovdqa	0x30(%rsp),$H3		# r2^2
+	vpaddq		$H1,$D2,$D2		# d2 += h1*r1
+	  vmovdqa	$H4,0x40(%r11)				#
+	vpmuludq	$T0,$H2,$H2		# h0*r1
+	 vpmuludq	$T2,$H3,$H0		# h2*r2
+	vpaddq		$H2,$D1,$D1		# d1 += h0*r1
+
+	 vmovdqa	0x40(%rsp),$H4		# s2^2
+	vpaddq		$H0,$D4,$D4		# d4 += h2*r2
+	vpmuludq	$T1,$H3,$H1		# h1*r2
+	vpmuludq	$T0,$H3,$H3		# h0*r2
+	vpaddq		$H1,$D3,$D3		# d3 += h1*r2
+	 vmovdqa	0x50(%rsp),$H2		# r3^2
+	vpaddq		$H3,$D2,$D2		# d2 += h0*r2
+	vpmuludq	$T4,$H4,$H0		# h4*s2
+	vpmuludq	$T3,$H4,$H4		# h3*s2
+	vpaddq		$H0,$D1,$D1		# d1 += h4*s2
+	 vmovdqa	0x60(%rsp),$H3		# s3^2
+	vpaddq		$H4,$D0,$D0		# d0 += h3*s2
+
+	 vmovdqa	0x80(%rsp),$H4		# s4^2
+	vpmuludq	$T1,$H2,$H1		# h1*r3
+	vpmuludq	$T0,$H2,$H2		# h0*r3
+	vpaddq		$H1,$D4,$D4		# d4 += h1*r3
+	vpaddq		$H2,$D3,$D3		# d3 += h0*r3
+	vpmuludq	$T4,$H3,$H0		# h4*s3
+	vpmuludq	$T3,$H3,$H1		# h3*s3
+	vpaddq		$H0,$D2,$D2		# d2 += h4*s3
+	 vmovdqu	16*0($inp),$H0				# load input
+	vpaddq		$H1,$D1,$D1		# d1 += h3*s3
+	vpmuludq	$T2,$H3,$H3		# h2*s3
+	 vpmuludq	$T2,$H4,$T2		# h2*s4
+	vpaddq		$H3,$D0,$D0		# d0 += h2*s3
+
+	 vmovdqu	16*1($inp),$H1				#
+	vpaddq		$T2,$D1,$D1		# d1 += h2*s4
+	vpmuludq	$T3,$H4,$T3		# h3*s4
+	vpmuludq	$T4,$H4,$T4		# h4*s4
+	 vpsrldq	\$6,$H0,$H2				# splat input
+	vpaddq		$T3,$D2,$D2		# d2 += h3*s4
+	vpaddq		$T4,$D3,$D3		# d3 += h4*s4
+	 vpsrldq	\$6,$H1,$H3				#
+	vpmuludq	0x70(%rsp),$T0,$T4	# h0*r4
+	vpmuludq	$T1,$H4,$T0		# h1*s4
+	 vpunpckhqdq	$H1,$H0,$H4		# 4
+	vpaddq		$T4,$D4,$D4		# d4 += h0*r4
+	 vmovdqa	-0x90(%r11),$T4		# r0^4
+	vpaddq		$T0,$D0,$D0		# d0 += h1*s4
+
+	vpunpcklqdq	$H1,$H0,$H0		# 0:1
+	vpunpcklqdq	$H3,$H2,$H3		# 2:3
+
+	#vpsrlq		\$40,$H4,$H4		# 4
+	vpsrldq		\$`40/8`,$H4,$H4	# 4
+	vpsrlq		\$26,$H0,$H1
+	vpand		$MASK,$H0,$H0		# 0
+	vpsrlq		\$4,$H3,$H2
+	vpand		$MASK,$H1,$H1		# 1
+	vpand		0(%rcx),$H4,$H4		# .Lmask24
+	vpsrlq		\$30,$H3,$H3
+	vpand		$MASK,$H2,$H2		# 2
+	vpand		$MASK,$H3,$H3		# 3
+	vpor		32(%rcx),$H4,$H4	# padbit, yes, always
+
+	vpaddq		0x00(%r11),$H0,$H0	# add hash value
+	vpaddq		0x10(%r11),$H1,$H1
+	vpaddq		0x20(%r11),$H2,$H2
+	vpaddq		0x30(%r11),$H3,$H3
+	vpaddq		0x40(%r11),$H4,$H4
+
+	lea		16*2($inp),%rax
+	lea		16*4($inp),$inp
+	sub		\$64,$len
+	cmovc		%rax,$inp
+
+	################################################################
+	# Now we accumulate (inp[0:1]+hash)*r^4
+	################################################################
+	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
+	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
+	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
+
+	vpmuludq	$H0,$T4,$T0		# h0*r0
+	vpmuludq	$H1,$T4,$T1		# h1*r0
+	vpaddq		$T0,$D0,$D0
+	vpaddq		$T1,$D1,$D1
+	 vmovdqa	-0x80(%r11),$T2		# r1^4
+	vpmuludq	$H2,$T4,$T0		# h2*r0
+	vpmuludq	$H3,$T4,$T1		# h3*r0
+	vpaddq		$T0,$D2,$D2
+	vpaddq		$T1,$D3,$D3
+	vpmuludq	$H4,$T4,$T4		# h4*r0
+	 vpmuludq	-0x70(%r11),$H4,$T0	# h4*s1
+	vpaddq		$T4,$D4,$D4
+
+	vpaddq		$T0,$D0,$D0		# d0 += h4*s1
+	vpmuludq	$H2,$T2,$T1		# h2*r1
+	vpmuludq	$H3,$T2,$T0		# h3*r1
+	vpaddq		$T1,$D3,$D3		# d3 += h2*r1
+	 vmovdqa	-0x60(%r11),$T3		# r2^4
+	vpaddq		$T0,$D4,$D4		# d4 += h3*r1
+	vpmuludq	$H1,$T2,$T1		# h1*r1
+	vpmuludq	$H0,$T2,$T2		# h0*r1
+	vpaddq		$T1,$D2,$D2		# d2 += h1*r1
+	vpaddq		$T2,$D1,$D1		# d1 += h0*r1
+
+	 vmovdqa	-0x50(%r11),$T4		# s2^4
+	vpmuludq	$H2,$T3,$T0		# h2*r2
+	vpmuludq	$H1,$T3,$T1		# h1*r2
+	vpaddq		$T0,$D4,$D4		# d4 += h2*r2
+	vpaddq		$T1,$D3,$D3		# d3 += h1*r2
+	 vmovdqa	-0x40(%r11),$T2		# r3^4
+	vpmuludq	$H0,$T3,$T3		# h0*r2
+	vpmuludq	$H4,$T4,$T0		# h4*s2
+	vpaddq		$T3,$D2,$D2		# d2 += h0*r2
+	vpaddq		$T0,$D1,$D1		# d1 += h4*s2
+	 vmovdqa	-0x30(%r11),$T3		# s3^4
+	vpmuludq	$H3,$T4,$T4		# h3*s2
+	 vpmuludq	$H1,$T2,$T1		# h1*r3
+	vpaddq		$T4,$D0,$D0		# d0 += h3*s2
+
+	 vmovdqa	-0x10(%r11),$T4		# s4^4
+	vpaddq		$T1,$D4,$D4		# d4 += h1*r3
+	vpmuludq	$H0,$T2,$T2		# h0*r3
+	vpmuludq	$H4,$T3,$T0		# h4*s3
+	vpaddq		$T2,$D3,$D3		# d3 += h0*r3
+	vpaddq		$T0,$D2,$D2		# d2 += h4*s3
+	 vmovdqu	16*2($inp),$T0				# load input
+	vpmuludq	$H3,$T3,$T2		# h3*s3
+	vpmuludq	$H2,$T3,$T3		# h2*s3
+	vpaddq		$T2,$D1,$D1		# d1 += h3*s3
+	 vmovdqu	16*3($inp),$T1				#
+	vpaddq		$T3,$D0,$D0		# d0 += h2*s3
+
+	vpmuludq	$H2,$T4,$H2		# h2*s4
+	vpmuludq	$H3,$T4,$H3		# h3*s4
+	 vpsrldq	\$6,$T0,$T2				# splat input
+	vpaddq		$H2,$D1,$D1		# d1 += h2*s4
+	vpmuludq	$H4,$T4,$H4		# h4*s4
+	 vpsrldq	\$6,$T1,$T3				#
+	vpaddq		$H3,$D2,$H2		# h2 = d2 + h3*s4
+	vpaddq		$H4,$D3,$H3		# h3 = d3 + h4*s4
+	vpmuludq	-0x20(%r11),$H0,$H4	# h0*r4
+	vpmuludq	$H1,$T4,$H0
+	 vpunpckhqdq	$T1,$T0,$T4		# 4
+	vpaddq		$H4,$D4,$H4		# h4 = d4 + h0*r4
+	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
+
+	vpunpcklqdq	$T1,$T0,$T0		# 0:1
+	vpunpcklqdq	$T3,$T2,$T3		# 2:3
+
+	#vpsrlq		\$40,$T4,$T4		# 4
+	vpsrldq		\$`40/8`,$T4,$T4	# 4
+	vpsrlq		\$26,$T0,$T1
+	 vmovdqa	0x00(%rsp),$D4		# preload r0^2
+	vpand		$MASK,$T0,$T0		# 0
+	vpsrlq		\$4,$T3,$T2
+	vpand		$MASK,$T1,$T1		# 1
+	vpand		0(%rcx),$T4,$T4		# .Lmask24
+	vpsrlq		\$30,$T3,$T3
+	vpand		$MASK,$T2,$T2		# 2
+	vpand		$MASK,$T3,$T3		# 3
+	vpor		32(%rcx),$T4,$T4	# padbit, yes, always
+
+	################################################################
+	# lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
+	# and P. Schwabe
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$D1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H4,$D0
+	vpand		$MASK,$H4,$H4
+
+	vpsrlq		\$26,$H1,$D1
+	vpand		$MASK,$H1,$H1
+	vpaddq		$D1,$H2,$H2		# h1 -> h2
+
+	vpaddq		$D0,$H0,$H0
+	vpsllq		\$2,$D0,$D0
+	vpaddq		$D0,$H0,$H0		# h4 -> h0
+
+	vpsrlq		\$26,$H2,$D2
+	vpand		$MASK,$H2,$H2
+	vpaddq		$D2,$H3,$H3		# h2 -> h3
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	ja		.Loop_avx
+
+.Lskip_loop_avx:
+	################################################################
+	# multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
+
+	vpshufd		\$0x10,$D4,$D4		# r0^n, xx12 -> x1x2
+	add		\$32,$len
+	jnz		.Long_tail_avx
+
+	vpaddq		$H2,$T2,$T2
+	vpaddq		$H0,$T0,$T0
+	vpaddq		$H1,$T1,$T1
+	vpaddq		$H3,$T3,$T3
+	vpaddq		$H4,$T4,$T4
+
+.Long_tail_avx:
+	vmovdqa		$H2,0x20(%r11)
+	vmovdqa		$H0,0x00(%r11)
+	vmovdqa		$H1,0x10(%r11)
+	vmovdqa		$H3,0x30(%r11)
+	vmovdqa		$H4,0x40(%r11)
+
+	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
+	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
+	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
+
+	vpmuludq	$T2,$D4,$D2		# d2 = h2*r0
+	vpmuludq	$T0,$D4,$D0		# d0 = h0*r0
+	 vpshufd	\$0x10,`16*1-64`($ctx),$H2		# r1^n
+	vpmuludq	$T1,$D4,$D1		# d1 = h1*r0
+	vpmuludq	$T3,$D4,$D3		# d3 = h3*r0
+	vpmuludq	$T4,$D4,$D4		# d4 = h4*r0
+
+	vpmuludq	$T3,$H2,$H0		# h3*r1
+	vpaddq		$H0,$D4,$D4		# d4 += h3*r1
+	 vpshufd	\$0x10,`16*2-64`($ctx),$H3		# s1^n
+	vpmuludq	$T2,$H2,$H1		# h2*r1
+	vpaddq		$H1,$D3,$D3		# d3 += h2*r1
+	 vpshufd	\$0x10,`16*3-64`($ctx),$H4		# r2^n
+	vpmuludq	$T1,$H2,$H0		# h1*r1
+	vpaddq		$H0,$D2,$D2		# d2 += h1*r1
+	vpmuludq	$T0,$H2,$H2		# h0*r1
+	vpaddq		$H2,$D1,$D1		# d1 += h0*r1
+	vpmuludq	$T4,$H3,$H3		# h4*s1
+	vpaddq		$H3,$D0,$D0		# d0 += h4*s1
+
+	 vpshufd	\$0x10,`16*4-64`($ctx),$H2		# s2^n
+	vpmuludq	$T2,$H4,$H1		# h2*r2
+	vpaddq		$H1,$D4,$D4		# d4 += h2*r2
+	vpmuludq	$T1,$H4,$H0		# h1*r2
+	vpaddq		$H0,$D3,$D3		# d3 += h1*r2
+	 vpshufd	\$0x10,`16*5-64`($ctx),$H3		# r3^n
+	vpmuludq	$T0,$H4,$H4		# h0*r2
+	vpaddq		$H4,$D2,$D2		# d2 += h0*r2
+	vpmuludq	$T4,$H2,$H1		# h4*s2
+	vpaddq		$H1,$D1,$D1		# d1 += h4*s2
+	 vpshufd	\$0x10,`16*6-64`($ctx),$H4		# s3^n
+	vpmuludq	$T3,$H2,$H2		# h3*s2
+	vpaddq		$H2,$D0,$D0		# d0 += h3*s2
+
+	vpmuludq	$T1,$H3,$H0		# h1*r3
+	vpaddq		$H0,$D4,$D4		# d4 += h1*r3
+	vpmuludq	$T0,$H3,$H3		# h0*r3
+	vpaddq		$H3,$D3,$D3		# d3 += h0*r3
+	 vpshufd	\$0x10,`16*7-64`($ctx),$H2		# r4^n
+	vpmuludq	$T4,$H4,$H1		# h4*s3
+	vpaddq		$H1,$D2,$D2		# d2 += h4*s3
+	 vpshufd	\$0x10,`16*8-64`($ctx),$H3		# s4^n
+	vpmuludq	$T3,$H4,$H0		# h3*s3
+	vpaddq		$H0,$D1,$D1		# d1 += h3*s3
+	vpmuludq	$T2,$H4,$H4		# h2*s3
+	vpaddq		$H4,$D0,$D0		# d0 += h2*s3
+
+	vpmuludq	$T0,$H2,$H2		# h0*r4
+	vpaddq		$H2,$D4,$D4		# h4 = d4 + h0*r4
+	vpmuludq	$T4,$H3,$H1		# h4*s4
+	vpaddq		$H1,$D3,$D3		# h3 = d3 + h4*s4
+	vpmuludq	$T3,$H3,$H0		# h3*s4
+	vpaddq		$H0,$D2,$D2		# h2 = d2 + h3*s4
+	vpmuludq	$T2,$H3,$H1		# h2*s4
+	vpaddq		$H1,$D1,$D1		# h1 = d1 + h2*s4
+	vpmuludq	$T1,$H3,$H3		# h1*s4
+	vpaddq		$H3,$D0,$D0		# h0 = d0 + h1*s4
+
+	jz		.Lshort_tail_avx
+
+	vmovdqu		16*0($inp),$H0		# load input
+	vmovdqu		16*1($inp),$H1
+
+	vpsrldq		\$6,$H0,$H2		# splat input
+	vpsrldq		\$6,$H1,$H3
+	vpunpckhqdq	$H1,$H0,$H4		# 4
+	vpunpcklqdq	$H1,$H0,$H0		# 0:1
+	vpunpcklqdq	$H3,$H2,$H3		# 2:3
+
+	vpsrlq		\$40,$H4,$H4		# 4
+	vpsrlq		\$26,$H0,$H1
+	vpand		$MASK,$H0,$H0		# 0
+	vpsrlq		\$4,$H3,$H2
+	vpand		$MASK,$H1,$H1		# 1
+	vpsrlq		\$30,$H3,$H3
+	vpand		$MASK,$H2,$H2		# 2
+	vpand		$MASK,$H3,$H3		# 3
+	vpor		32(%rcx),$H4,$H4	# padbit, yes, always
+
+	vpshufd		\$0x32,`16*0-64`($ctx),$T4	# r0^n, 34xx -> x3x4
+	vpaddq		0x00(%r11),$H0,$H0
+	vpaddq		0x10(%r11),$H1,$H1
+	vpaddq		0x20(%r11),$H2,$H2
+	vpaddq		0x30(%r11),$H3,$H3
+	vpaddq		0x40(%r11),$H4,$H4
+
+	################################################################
+	# multiply (inp[0:1]+hash) by r^4:r^3 and accumulate
+
+	vpmuludq	$H0,$T4,$T0		# h0*r0
+	vpaddq		$T0,$D0,$D0		# d0 += h0*r0
+	vpmuludq	$H1,$T4,$T1		# h1*r0
+	vpaddq		$T1,$D1,$D1		# d1 += h1*r0
+	vpmuludq	$H2,$T4,$T0		# h2*r0
+	vpaddq		$T0,$D2,$D2		# d2 += h2*r0
+	 vpshufd	\$0x32,`16*1-64`($ctx),$T2		# r1^n
+	vpmuludq	$H3,$T4,$T1		# h3*r0
+	vpaddq		$T1,$D3,$D3		# d3 += h3*r0
+	vpmuludq	$H4,$T4,$T4		# h4*r0
+	vpaddq		$T4,$D4,$D4		# d4 += h4*r0
+
+	vpmuludq	$H3,$T2,$T0		# h3*r1
+	vpaddq		$T0,$D4,$D4		# d4 += h3*r1
+	 vpshufd	\$0x32,`16*2-64`($ctx),$T3		# s1
+	vpmuludq	$H2,$T2,$T1		# h2*r1
+	vpaddq		$T1,$D3,$D3		# d3 += h2*r1
+	 vpshufd	\$0x32,`16*3-64`($ctx),$T4		# r2
+	vpmuludq	$H1,$T2,$T0		# h1*r1
+	vpaddq		$T0,$D2,$D2		# d2 += h1*r1
+	vpmuludq	$H0,$T2,$T2		# h0*r1
+	vpaddq		$T2,$D1,$D1		# d1 += h0*r1
+	vpmuludq	$H4,$T3,$T3		# h4*s1
+	vpaddq		$T3,$D0,$D0		# d0 += h4*s1
+
+	 vpshufd	\$0x32,`16*4-64`($ctx),$T2		# s2
+	vpmuludq	$H2,$T4,$T1		# h2*r2
+	vpaddq		$T1,$D4,$D4		# d4 += h2*r2
+	vpmuludq	$H1,$T4,$T0		# h1*r2
+	vpaddq		$T0,$D3,$D3		# d3 += h1*r2
+	 vpshufd	\$0x32,`16*5-64`($ctx),$T3		# r3
+	vpmuludq	$H0,$T4,$T4		# h0*r2
+	vpaddq		$T4,$D2,$D2		# d2 += h0*r2
+	vpmuludq	$H4,$T2,$T1		# h4*s2
+	vpaddq		$T1,$D1,$D1		# d1 += h4*s2
+	 vpshufd	\$0x32,`16*6-64`($ctx),$T4		# s3
+	vpmuludq	$H3,$T2,$T2		# h3*s2
+	vpaddq		$T2,$D0,$D0		# d0 += h3*s2
+
+	vpmuludq	$H1,$T3,$T0		# h1*r3
+	vpaddq		$T0,$D4,$D4		# d4 += h1*r3
+	vpmuludq	$H0,$T3,$T3		# h0*r3
+	vpaddq		$T3,$D3,$D3		# d3 += h0*r3
+	 vpshufd	\$0x32,`16*7-64`($ctx),$T2		# r4
+	vpmuludq	$H4,$T4,$T1		# h4*s3
+	vpaddq		$T1,$D2,$D2		# d2 += h4*s3
+	 vpshufd	\$0x32,`16*8-64`($ctx),$T3		# s4
+	vpmuludq	$H3,$T4,$T0		# h3*s3
+	vpaddq		$T0,$D1,$D1		# d1 += h3*s3
+	vpmuludq	$H2,$T4,$T4		# h2*s3
+	vpaddq		$T4,$D0,$D0		# d0 += h2*s3
+
+	vpmuludq	$H0,$T2,$T2		# h0*r4
+	vpaddq		$T2,$D4,$D4		# d4 += h0*r4
+	vpmuludq	$H4,$T3,$T1		# h4*s4
+	vpaddq		$T1,$D3,$D3		# d3 += h4*s4
+	vpmuludq	$H3,$T3,$T0		# h3*s4
+	vpaddq		$T0,$D2,$D2		# d2 += h3*s4
+	vpmuludq	$H2,$T3,$T1		# h2*s4
+	vpaddq		$T1,$D1,$D1		# d1 += h2*s4
+	vpmuludq	$H1,$T3,$T3		# h1*s4
+	vpaddq		$T3,$D0,$D0		# d0 += h1*s4
+
+.Lshort_tail_avx:
+	################################################################
+	# horizontal addition
+
+	vpsrldq		\$8,$D4,$T4
+	vpsrldq		\$8,$D3,$T3
+	vpsrldq		\$8,$D1,$T1
+	vpsrldq		\$8,$D0,$T0
+	vpsrldq		\$8,$D2,$T2
+	vpaddq		$T3,$D3,$D3
+	vpaddq		$T4,$D4,$D4
+	vpaddq		$T0,$D0,$D0
+	vpaddq		$T1,$D1,$D1
+	vpaddq		$T2,$D2,$D2
+
+	################################################################
+	# lazy reduction
+
+	vpsrlq		\$26,$D3,$H3
+	vpand		$MASK,$D3,$D3
+	vpaddq		$H3,$D4,$D4		# h3 -> h4
+
+	vpsrlq		\$26,$D0,$H0
+	vpand		$MASK,$D0,$D0
+	vpaddq		$H0,$D1,$D1		# h0 -> h1
+
+	vpsrlq		\$26,$D4,$H4
+	vpand		$MASK,$D4,$D4
+
+	vpsrlq		\$26,$D1,$H1
+	vpand		$MASK,$D1,$D1
+	vpaddq		$H1,$D2,$D2		# h1 -> h2
+
+	vpaddq		$H4,$D0,$D0
+	vpsllq		\$2,$H4,$H4
+	vpaddq		$H4,$D0,$D0		# h4 -> h0
+
+	vpsrlq		\$26,$D2,$H2
+	vpand		$MASK,$D2,$D2
+	vpaddq		$H2,$D3,$D3		# h2 -> h3
+
+	vpsrlq		\$26,$D0,$H0
+	vpand		$MASK,$D0,$D0
+	vpaddq		$H0,$D1,$D1		# h0 -> h1
+
+	vpsrlq		\$26,$D3,$H3
+	vpand		$MASK,$D3,$D3
+	vpaddq		$H3,$D4,$D4		# h3 -> h4
+
+	vmovd		$D0,`4*0-48-64`($ctx)	# save partially reduced
+	vmovd		$D1,`4*1-48-64`($ctx)
+	vmovd		$D2,`4*2-48-64`($ctx)
+	vmovd		$D3,`4*3-48-64`($ctx)
+	vmovd		$D4,`4*4-48-64`($ctx)
+___
+$code.=<<___	if ($win64);
+	vmovdqa		0x50(%r11),%xmm6
+	vmovdqa		0x60(%r11),%xmm7
+	vmovdqa		0x70(%r11),%xmm8
+	vmovdqa		0x80(%r11),%xmm9
+	vmovdqa		0x90(%r11),%xmm10
+	vmovdqa		0xa0(%r11),%xmm11
+	vmovdqa		0xb0(%r11),%xmm12
+	vmovdqa		0xc0(%r11),%xmm13
+	vmovdqa		0xd0(%r11),%xmm14
+	vmovdqa		0xe0(%r11),%xmm15
+	lea		0xf8(%r11),%rsp
+.Ldo_avx_epilogue:
+___
+$code.=<<___	if (!$win64);
+	lea		0x58(%r11),%rsp
+.cfi_def_cfa		%rsp,8
+___
+$code.=<<___;
+	vzeroupper
+	ret
+.cfi_endproc
+.size	poly1305_blocks_avx,.-poly1305_blocks_avx
+
+.type	poly1305_emit_avx,\@function,3
+.align	32
+poly1305_emit_avx:
+	cmpl	\$0,20($ctx)	# is_base2_26?
+	je	.Lemit
+
+	mov	0($ctx),%eax	# load hash value base 2^26
+	mov	4($ctx),%ecx
+	mov	8($ctx),%r8d
+	mov	12($ctx),%r11d
+	mov	16($ctx),%r10d
+
+	shl	\$26,%rcx	# base 2^26 -> base 2^64
+	mov	%r8,%r9
+	shl	\$52,%r8
+	add	%rcx,%rax
+	shr	\$12,%r9
+	add	%rax,%r8	# h0
+	adc	\$0,%r9
+
+	shl	\$14,%r11
+	mov	%r10,%rax
+	shr	\$24,%r10
+	add	%r11,%r9
+	shl	\$40,%rax
+	add	%rax,%r9	# h1
+	adc	\$0,%r10	# h2
+
+	mov	%r10,%rax	# could be partially reduced, so reduce
+	mov	%r10,%rcx
+	and	\$3,%r10
+	shr	\$2,%rax
+	and	\$-4,%rcx
+	add	%rcx,%rax
+	add	%rax,%r8
+	adc	\$0,%r9
+	adc	\$0,%r10
+
+	mov	%r8,%rax
+	add	\$5,%r8		# compare to modulus
+	mov	%r9,%rcx
+	adc	\$0,%r9
+	adc	\$0,%r10
+	shr	\$2,%r10	# did 130-bit value overflow?
+	cmovnz	%r8,%rax
+	cmovnz	%r9,%rcx
+
+	add	0($nonce),%rax	# accumulate nonce
+	adc	8($nonce),%rcx
+	mov	%rax,0($mac)	# write result
+	mov	%rcx,8($mac)
+
+	ret
+.size	poly1305_emit_avx,.-poly1305_emit_avx
+___
+
+if ($avx>1) {
+my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
+    map("%ymm$_",(0..15));
+my $S4=$MASK;
+
+$code.=<<___;
+.type	poly1305_blocks_avx2,\@function,4
+.align	32
+poly1305_blocks_avx2:
+.cfi_startproc
+	mov	20($ctx),%r8d		# is_base2_26
+	cmp	\$128,$len
+	jae	.Lblocks_avx2
+	test	%r8d,%r8d
+	jz	.Lblocks
+
+.Lblocks_avx2:
+	and	\$-16,$len
+	jz	.Lno_data_avx2
+
+	vzeroupper
+
+	test	%r8d,%r8d
+	jz	.Lbase2_64_avx2
+
+	test	\$63,$len
+	jz	.Leven_avx2
+
+	push	%rbx
+.cfi_push	%rbx
+	push	%rbp
+.cfi_push	%rbp
+	push	%r12
+.cfi_push	%r12
+	push	%r13
+.cfi_push	%r13
+	push	%r14
+.cfi_push	%r14
+	push	%r15
+.cfi_push	%r15
+.Lblocks_avx2_body:
+
+	mov	$len,%r15		# reassign $len
+
+	mov	0($ctx),$d1		# load hash value
+	mov	8($ctx),$d2
+	mov	16($ctx),$h2#d
+
+	mov	24($ctx),$r0		# load r
+	mov	32($ctx),$s1
+
+	################################# base 2^26 -> base 2^64
+	mov	$d1#d,$h0#d
+	and	\$`-1*(1<<31)`,$d1
+	mov	$d2,$r1			# borrow $r1
+	mov	$d2#d,$h1#d
+	and	\$`-1*(1<<31)`,$d2
+
+	shr	\$6,$d1
+	shl	\$52,$r1
+	add	$d1,$h0
+	shr	\$12,$h1
+	shr	\$18,$d2
+	add	$r1,$h0
+	adc	$d2,$h1
+
+	mov	$h2,$d1
+	shl	\$40,$d1
+	shr	\$24,$h2
+	add	$d1,$h1
+	adc	\$0,$h2			# can be partially reduced...
+
+	mov	\$-4,$d2		# ... so reduce
+	mov	$h2,$d1
+	and	$h2,$d2
+	shr	\$2,$d1
+	and	\$3,$h2
+	add	$d2,$d1			# =*5
+	add	$d1,$h0
+	adc	\$0,$h1
+	adc	\$0,$h2
+
+	mov	$s1,$r1
+	mov	$s1,%rax
+	shr	\$2,$s1
+	add	$r1,$s1			# s1 = r1 + (r1 >> 2)
+
+.Lbase2_26_pre_avx2:
+	add	0($inp),$h0		# accumulate input
+	adc	8($inp),$h1
+	lea	16($inp),$inp
+	adc	$padbit,$h2
+	sub	\$16,%r15
+
+	call	__poly1305_block
+	mov	$r1,%rax
+
+	test	\$63,%r15
+	jnz	.Lbase2_26_pre_avx2
+
+	test	$padbit,$padbit		# if $padbit is zero,
+	jz	.Lstore_base2_64_avx2	# store hash in base 2^64 format
+
+	################################# base 2^64 -> base 2^26
+	mov	$h0,%rax
+	mov	$h0,%rdx
+	shr	\$52,$h0
+	mov	$h1,$r0
+	mov	$h1,$r1
+	shr	\$26,%rdx
+	and	\$0x3ffffff,%rax	# h[0]
+	shl	\$12,$r0
+	and	\$0x3ffffff,%rdx	# h[1]
+	shr	\$14,$h1
+	or	$r0,$h0
+	shl	\$24,$h2
+	and	\$0x3ffffff,$h0		# h[2]
+	shr	\$40,$r1
+	and	\$0x3ffffff,$h1		# h[3]
+	or	$r1,$h2			# h[4]
+
+	test	%r15,%r15
+	jz	.Lstore_base2_26_avx2
+
+	vmovd	%rax#d,%x#$H0
+	vmovd	%rdx#d,%x#$H1
+	vmovd	$h0#d,%x#$H2
+	vmovd	$h1#d,%x#$H3
+	vmovd	$h2#d,%x#$H4
+	jmp	.Lproceed_avx2
+
+.align	32
+.Lstore_base2_64_avx2:
+	mov	$h0,0($ctx)
+	mov	$h1,8($ctx)
+	mov	$h2,16($ctx)		# note that is_base2_26 is zeroed
+	jmp	.Ldone_avx2
+
+.align	16
+.Lstore_base2_26_avx2:
+	mov	%rax#d,0($ctx)		# store hash value base 2^26
+	mov	%rdx#d,4($ctx)
+	mov	$h0#d,8($ctx)
+	mov	$h1#d,12($ctx)
+	mov	$h2#d,16($ctx)
+.align	16
+.Ldone_avx2:
+	mov	0(%rsp),%r15
+.cfi_restore	%r15
+	mov	8(%rsp),%r14
+.cfi_restore	%r14
+	mov	16(%rsp),%r13
+.cfi_restore	%r13
+	mov	24(%rsp),%r12
+.cfi_restore	%r12
+	mov	32(%rsp),%rbp
+.cfi_restore	%rbp
+	mov	40(%rsp),%rbx
+.cfi_restore	%rbx
+	lea	48(%rsp),%rsp
+.cfi_adjust_cfa_offset	-48
+.Lno_data_avx2:
+.Lblocks_avx2_epilogue:
+	ret
+.cfi_endproc
+
+.align	32
+.Lbase2_64_avx2:
+.cfi_startproc
+	push	%rbx
+.cfi_push	%rbx
+	push	%rbp
+.cfi_push	%rbp
+	push	%r12
+.cfi_push	%r12
+	push	%r13
+.cfi_push	%r13
+	push	%r14
+.cfi_push	%r14
+	push	%r15
+.cfi_push	%r15
+.Lbase2_64_avx2_body:
+
+	mov	$len,%r15		# reassign $len
+
+	mov	24($ctx),$r0		# load r
+	mov	32($ctx),$s1
+
+	mov	0($ctx),$h0		# load hash value
+	mov	8($ctx),$h1
+	mov	16($ctx),$h2#d
+
+	mov	$s1,$r1
+	mov	$s1,%rax
+	shr	\$2,$s1
+	add	$r1,$s1			# s1 = r1 + (r1 >> 2)
+
+	test	\$63,$len
+	jz	.Linit_avx2
+
+.Lbase2_64_pre_avx2:
+	add	0($inp),$h0		# accumulate input
+	adc	8($inp),$h1
+	lea	16($inp),$inp
+	adc	$padbit,$h2
+	sub	\$16,%r15
+
+	call	__poly1305_block
+	mov	$r1,%rax
+
+	test	\$63,%r15
+	jnz	.Lbase2_64_pre_avx2
+
+.Linit_avx2:
+	################################# base 2^64 -> base 2^26
+	mov	$h0,%rax
+	mov	$h0,%rdx
+	shr	\$52,$h0
+	mov	$h1,$d1
+	mov	$h1,$d2
+	shr	\$26,%rdx
+	and	\$0x3ffffff,%rax	# h[0]
+	shl	\$12,$d1
+	and	\$0x3ffffff,%rdx	# h[1]
+	shr	\$14,$h1
+	or	$d1,$h0
+	shl	\$24,$h2
+	and	\$0x3ffffff,$h0		# h[2]
+	shr	\$40,$d2
+	and	\$0x3ffffff,$h1		# h[3]
+	or	$d2,$h2			# h[4]
+
+	vmovd	%rax#d,%x#$H0
+	vmovd	%rdx#d,%x#$H1
+	vmovd	$h0#d,%x#$H2
+	vmovd	$h1#d,%x#$H3
+	vmovd	$h2#d,%x#$H4
+	movl	\$1,20($ctx)		# set is_base2_26
+
+	call	__poly1305_init_avx
+
+.Lproceed_avx2:
+	mov	%r15,$len			# restore $len
+	mov	OPENSSL_ia32cap_P+8(%rip),%r10d
+	mov	\$`(1<<31|1<<30|1<<16)`,%r11d
+
+	mov	0(%rsp),%r15
+.cfi_restore	%r15
+	mov	8(%rsp),%r14
+.cfi_restore	%r14
+	mov	16(%rsp),%r13
+.cfi_restore	%r13
+	mov	24(%rsp),%r12
+.cfi_restore	%r12
+	mov	32(%rsp),%rbp
+.cfi_restore	%rbp
+	mov	40(%rsp),%rbx
+.cfi_restore	%rbx
+	lea	48(%rsp),%rax
+	lea	48(%rsp),%rsp
+.cfi_adjust_cfa_offset	-48
+.Lbase2_64_avx2_epilogue:
+	jmp	.Ldo_avx2
+.cfi_endproc
+
+.align	32
+.Leven_avx2:
+.cfi_startproc
+	mov		OPENSSL_ia32cap_P+8(%rip),%r10d
+	vmovd		4*0($ctx),%x#$H0	# load hash value base 2^26
+	vmovd		4*1($ctx),%x#$H1
+	vmovd		4*2($ctx),%x#$H2
+	vmovd		4*3($ctx),%x#$H3
+	vmovd		4*4($ctx),%x#$H4
+
+.Ldo_avx2:
+___
+$code.=<<___		if ($avx>2);
+	cmp		\$512,$len
+	jb		.Lskip_avx512
+	and		%r11d,%r10d
+	test		\$`1<<16`,%r10d		# check for AVX512F
+	jnz		.Lblocks_avx512
+.Lskip_avx512:
+___
+$code.=<<___	if (!$win64);
+	lea		-8(%rsp),%r11
+.cfi_def_cfa		%r11,16
+	sub		\$0x128,%rsp
+___
+$code.=<<___	if ($win64);
+	lea		-0xf8(%rsp),%r11
+	sub		\$0x1c8,%rsp
+	vmovdqa		%xmm6,0x50(%r11)
+	vmovdqa		%xmm7,0x60(%r11)
+	vmovdqa		%xmm8,0x70(%r11)
+	vmovdqa		%xmm9,0x80(%r11)
+	vmovdqa		%xmm10,0x90(%r11)
+	vmovdqa		%xmm11,0xa0(%r11)
+	vmovdqa		%xmm12,0xb0(%r11)
+	vmovdqa		%xmm13,0xc0(%r11)
+	vmovdqa		%xmm14,0xd0(%r11)
+	vmovdqa		%xmm15,0xe0(%r11)
+.Ldo_avx2_body:
+___
+$code.=<<___;
+	lea		.Lconst(%rip),%rcx
+	lea		48+64($ctx),$ctx	# size optimization
+	vmovdqa		96(%rcx),$T0		# .Lpermd_avx2
+
+	# expand and copy pre-calculated table to stack
+	vmovdqu		`16*0-64`($ctx),%x#$T2
+	and		\$-512,%rsp
+	vmovdqu		`16*1-64`($ctx),%x#$T3
+	vmovdqu		`16*2-64`($ctx),%x#$T4
+	vmovdqu		`16*3-64`($ctx),%x#$D0
+	vmovdqu		`16*4-64`($ctx),%x#$D1
+	vmovdqu		`16*5-64`($ctx),%x#$D2
+	lea		0x90(%rsp),%rax		# size optimization
+	vmovdqu		`16*6-64`($ctx),%x#$D3
+	vpermd		$T2,$T0,$T2		# 00003412 -> 14243444
+	vmovdqu		`16*7-64`($ctx),%x#$D4
+	vpermd		$T3,$T0,$T3
+	vmovdqu		`16*8-64`($ctx),%x#$MASK
+	vpermd		$T4,$T0,$T4
+	vmovdqa		$T2,0x00(%rsp)
+	vpermd		$D0,$T0,$D0
+	vmovdqa		$T3,0x20-0x90(%rax)
+	vpermd		$D1,$T0,$D1
+	vmovdqa		$T4,0x40-0x90(%rax)
+	vpermd		$D2,$T0,$D2
+	vmovdqa		$D0,0x60-0x90(%rax)
+	vpermd		$D3,$T0,$D3
+	vmovdqa		$D1,0x80-0x90(%rax)
+	vpermd		$D4,$T0,$D4
+	vmovdqa		$D2,0xa0-0x90(%rax)
+	vpermd		$MASK,$T0,$MASK
+	vmovdqa		$D3,0xc0-0x90(%rax)
+	vmovdqa		$D4,0xe0-0x90(%rax)
+	vmovdqa		$MASK,0x100-0x90(%rax)
+	vmovdqa		64(%rcx),$MASK		# .Lmask26
+
+	################################################################
+	# load input
+	vmovdqu		16*0($inp),%x#$T0
+	vmovdqu		16*1($inp),%x#$T1
+	vinserti128	\$1,16*2($inp),$T0,$T0
+	vinserti128	\$1,16*3($inp),$T1,$T1
+	lea		16*4($inp),$inp
+
+	vpsrldq		\$6,$T0,$T2		# splat input
+	vpsrldq		\$6,$T1,$T3
+	vpunpckhqdq	$T1,$T0,$T4		# 4
+	vpunpcklqdq	$T3,$T2,$T2		# 2:3
+	vpunpcklqdq	$T1,$T0,$T0		# 0:1
+
+	vpsrlq		\$30,$T2,$T3
+	vpsrlq		\$4,$T2,$T2
+	vpsrlq		\$26,$T0,$T1
+	vpsrlq		\$40,$T4,$T4		# 4
+	vpand		$MASK,$T2,$T2		# 2
+	vpand		$MASK,$T0,$T0		# 0
+	vpand		$MASK,$T1,$T1		# 1
+	vpand		$MASK,$T3,$T3		# 3
+	vpor		32(%rcx),$T4,$T4	# padbit, yes, always
+
+	vpaddq		$H2,$T2,$H2		# accumulate input
+	sub		\$64,$len
+	jz		.Ltail_avx2
+	jmp		.Loop_avx2
+
+.align	32
+.Loop_avx2:
+	################################################################
+	# ((inp[0]*r^4+inp[4])*r^4+inp[ 8])*r^4
+	# ((inp[1]*r^4+inp[5])*r^4+inp[ 9])*r^3
+	# ((inp[2]*r^4+inp[6])*r^4+inp[10])*r^2
+	# ((inp[3]*r^4+inp[7])*r^4+inp[11])*r^1
+	#   \________/\__________/
+	################################################################
+	#vpaddq		$H2,$T2,$H2		# accumulate input
+	vpaddq		$H0,$T0,$H0
+	vmovdqa		`32*0`(%rsp),$T0	# r0^4
+	vpaddq		$H1,$T1,$H1
+	vmovdqa		`32*1`(%rsp),$T1	# r1^4
+	vpaddq		$H3,$T3,$H3
+	vmovdqa		`32*3`(%rsp),$T2	# r2^4
+	vpaddq		$H4,$T4,$H4
+	vmovdqa		`32*6-0x90`(%rax),$T3	# s3^4
+	vmovdqa		`32*8-0x90`(%rax),$S4	# s4^4
+
+	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
+	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
+	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
+	#
+	# however, as h2 is "chronologically" first one available pull
+	# corresponding operations up, so it's
+	#
+	# d4 = h2*r2   + h4*r0 + h3*r1             + h1*r3   + h0*r4
+	# d3 = h2*r1   + h3*r0           + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0           + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h2*5*r4 + h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3
+	# d0 = h2*5*r3 + h0*r0 + h4*5*r1 + h3*5*r2           + h1*5*r4
+
+	vpmuludq	$H2,$T0,$D2		# d2 = h2*r0
+	vpmuludq	$H2,$T1,$D3		# d3 = h2*r1
+	vpmuludq	$H2,$T2,$D4		# d4 = h2*r2
+	vpmuludq	$H2,$T3,$D0		# d0 = h2*s3
+	vpmuludq	$H2,$S4,$D1		# d1 = h2*s4
+
+	vpmuludq	$H0,$T1,$T4		# h0*r1
+	vpmuludq	$H1,$T1,$H2		# h1*r1, borrow $H2 as temp
+	vpaddq		$T4,$D1,$D1		# d1 += h0*r1
+	vpaddq		$H2,$D2,$D2		# d2 += h1*r1
+	vpmuludq	$H3,$T1,$T4		# h3*r1
+	vpmuludq	`32*2`(%rsp),$H4,$H2	# h4*s1
+	vpaddq		$T4,$D4,$D4		# d4 += h3*r1
+	vpaddq		$H2,$D0,$D0		# d0 += h4*s1
+	 vmovdqa	`32*4-0x90`(%rax),$T1	# s2
+
+	vpmuludq	$H0,$T0,$T4		# h0*r0
+	vpmuludq	$H1,$T0,$H2		# h1*r0
+	vpaddq		$T4,$D0,$D0		# d0 += h0*r0
+	vpaddq		$H2,$D1,$D1		# d1 += h1*r0
+	vpmuludq	$H3,$T0,$T4		# h3*r0
+	vpmuludq	$H4,$T0,$H2		# h4*r0
+	 vmovdqu	16*0($inp),%x#$T0	# load input
+	vpaddq		$T4,$D3,$D3		# d3 += h3*r0
+	vpaddq		$H2,$D4,$D4		# d4 += h4*r0
+	 vinserti128	\$1,16*2($inp),$T0,$T0
+
+	vpmuludq	$H3,$T1,$T4		# h3*s2
+	vpmuludq	$H4,$T1,$H2		# h4*s2
+	 vmovdqu	16*1($inp),%x#$T1
+	vpaddq		$T4,$D0,$D0		# d0 += h3*s2
+	vpaddq		$H2,$D1,$D1		# d1 += h4*s2
+	 vmovdqa	`32*5-0x90`(%rax),$H2	# r3
+	vpmuludq	$H1,$T2,$T4		# h1*r2
+	vpmuludq	$H0,$T2,$T2		# h0*r2
+	vpaddq		$T4,$D3,$D3		# d3 += h1*r2
+	vpaddq		$T2,$D2,$D2		# d2 += h0*r2
+	 vinserti128	\$1,16*3($inp),$T1,$T1
+	 lea		16*4($inp),$inp
+
+	vpmuludq	$H1,$H2,$T4		# h1*r3
+	vpmuludq	$H0,$H2,$H2		# h0*r3
+	 vpsrldq	\$6,$T0,$T2		# splat input
+	vpaddq		$T4,$D4,$D4		# d4 += h1*r3
+	vpaddq		$H2,$D3,$D3		# d3 += h0*r3
+	vpmuludq	$H3,$T3,$T4		# h3*s3
+	vpmuludq	$H4,$T3,$H2		# h4*s3
+	 vpsrldq	\$6,$T1,$T3
+	vpaddq		$T4,$D1,$D1		# d1 += h3*s3
+	vpaddq		$H2,$D2,$D2		# d2 += h4*s3
+	 vpunpckhqdq	$T1,$T0,$T4		# 4
+
+	vpmuludq	$H3,$S4,$H3		# h3*s4
+	vpmuludq	$H4,$S4,$H4		# h4*s4
+	 vpunpcklqdq	$T1,$T0,$T0		# 0:1
+	vpaddq		$H3,$D2,$H2		# h2 = d2 + h3*r4
+	vpaddq		$H4,$D3,$H3		# h3 = d3 + h4*r4
+	 vpunpcklqdq	$T3,$T2,$T3		# 2:3
+	vpmuludq	`32*7-0x90`(%rax),$H0,$H4	# h0*r4
+	vpmuludq	$H1,$S4,$H0		# h1*s4
+	vmovdqa		64(%rcx),$MASK		# .Lmask26
+	vpaddq		$H4,$D4,$H4		# h4 = d4 + h0*r4
+	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
+
+	################################################################
+	# lazy reduction (interleaved with tail of input splat)
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$D1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H4,$D4
+	vpand		$MASK,$H4,$H4
+
+	 vpsrlq		\$4,$T3,$T2
+
+	vpsrlq		\$26,$H1,$D1
+	vpand		$MASK,$H1,$H1
+	vpaddq		$D1,$H2,$H2		# h1 -> h2
+
+	vpaddq		$D4,$H0,$H0
+	vpsllq		\$2,$D4,$D4
+	vpaddq		$D4,$H0,$H0		# h4 -> h0
+
+	 vpand		$MASK,$T2,$T2		# 2
+	 vpsrlq		\$26,$T0,$T1
+
+	vpsrlq		\$26,$H2,$D2
+	vpand		$MASK,$H2,$H2
+	vpaddq		$D2,$H3,$H3		# h2 -> h3
+
+	 vpaddq		$T2,$H2,$H2		# modulo-scheduled
+	 vpsrlq		\$30,$T3,$T3
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	 vpsrlq		\$40,$T4,$T4		# 4
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	 vpand		$MASK,$T0,$T0		# 0
+	 vpand		$MASK,$T1,$T1		# 1
+	 vpand		$MASK,$T3,$T3		# 3
+	 vpor		32(%rcx),$T4,$T4	# padbit, yes, always
+
+	sub		\$64,$len
+	jnz		.Loop_avx2
+
+	.byte		0x66,0x90
+.Ltail_avx2:
+	################################################################
+	# while above multiplications were by r^4 in all lanes, in last
+	# iteration we multiply least significant lane by r^4 and most
+	# significant one by r, so copy of above except that references
+	# to the precomputed table are displaced by 4...
+
+	#vpaddq		$H2,$T2,$H2		# accumulate input
+	vpaddq		$H0,$T0,$H0
+	vmovdqu		`32*0+4`(%rsp),$T0	# r0^4
+	vpaddq		$H1,$T1,$H1
+	vmovdqu		`32*1+4`(%rsp),$T1	# r1^4
+	vpaddq		$H3,$T3,$H3
+	vmovdqu		`32*3+4`(%rsp),$T2	# r2^4
+	vpaddq		$H4,$T4,$H4
+	vmovdqu		`32*6+4-0x90`(%rax),$T3	# s3^4
+	vmovdqu		`32*8+4-0x90`(%rax),$S4	# s4^4
+
+	vpmuludq	$H2,$T0,$D2		# d2 = h2*r0
+	vpmuludq	$H2,$T1,$D3		# d3 = h2*r1
+	vpmuludq	$H2,$T2,$D4		# d4 = h2*r2
+	vpmuludq	$H2,$T3,$D0		# d0 = h2*s3
+	vpmuludq	$H2,$S4,$D1		# d1 = h2*s4
+
+	vpmuludq	$H0,$T1,$T4		# h0*r1
+	vpmuludq	$H1,$T1,$H2		# h1*r1
+	vpaddq		$T4,$D1,$D1		# d1 += h0*r1
+	vpaddq		$H2,$D2,$D2		# d2 += h1*r1
+	vpmuludq	$H3,$T1,$T4		# h3*r1
+	vpmuludq	`32*2+4`(%rsp),$H4,$H2	# h4*s1
+	vpaddq		$T4,$D4,$D4		# d4 += h3*r1
+	vpaddq		$H2,$D0,$D0		# d0 += h4*s1
+
+	vpmuludq	$H0,$T0,$T4		# h0*r0
+	vpmuludq	$H1,$T0,$H2		# h1*r0
+	vpaddq		$T4,$D0,$D0		# d0 += h0*r0
+	 vmovdqu	`32*4+4-0x90`(%rax),$T1	# s2
+	vpaddq		$H2,$D1,$D1		# d1 += h1*r0
+	vpmuludq	$H3,$T0,$T4		# h3*r0
+	vpmuludq	$H4,$T0,$H2		# h4*r0
+	vpaddq		$T4,$D3,$D3		# d3 += h3*r0
+	vpaddq		$H2,$D4,$D4		# d4 += h4*r0
+
+	vpmuludq	$H3,$T1,$T4		# h3*s2
+	vpmuludq	$H4,$T1,$H2		# h4*s2
+	vpaddq		$T4,$D0,$D0		# d0 += h3*s2
+	vpaddq		$H2,$D1,$D1		# d1 += h4*s2
+	 vmovdqu	`32*5+4-0x90`(%rax),$H2	# r3
+	vpmuludq	$H1,$T2,$T4		# h1*r2
+	vpmuludq	$H0,$T2,$T2		# h0*r2
+	vpaddq		$T4,$D3,$D3		# d3 += h1*r2
+	vpaddq		$T2,$D2,$D2		# d2 += h0*r2
+
+	vpmuludq	$H1,$H2,$T4		# h1*r3
+	vpmuludq	$H0,$H2,$H2		# h0*r3
+	vpaddq		$T4,$D4,$D4		# d4 += h1*r3
+	vpaddq		$H2,$D3,$D3		# d3 += h0*r3
+	vpmuludq	$H3,$T3,$T4		# h3*s3
+	vpmuludq	$H4,$T3,$H2		# h4*s3
+	vpaddq		$T4,$D1,$D1		# d1 += h3*s3
+	vpaddq		$H2,$D2,$D2		# d2 += h4*s3
+
+	vpmuludq	$H3,$S4,$H3		# h3*s4
+	vpmuludq	$H4,$S4,$H4		# h4*s4
+	vpaddq		$H3,$D2,$H2		# h2 = d2 + h3*r4
+	vpaddq		$H4,$D3,$H3		# h3 = d3 + h4*r4
+	vpmuludq	`32*7+4-0x90`(%rax),$H0,$H4		# h0*r4
+	vpmuludq	$H1,$S4,$H0		# h1*s4
+	vmovdqa		64(%rcx),$MASK		# .Lmask26
+	vpaddq		$H4,$D4,$H4		# h4 = d4 + h0*r4
+	vpaddq		$H0,$D0,$H0		# h0 = d0 + h1*s4
+
+	################################################################
+	# horizontal addition
+
+	vpsrldq		\$8,$D1,$T1
+	vpsrldq		\$8,$H2,$T2
+	vpsrldq		\$8,$H3,$T3
+	vpsrldq		\$8,$H4,$T4
+	vpsrldq		\$8,$H0,$T0
+	vpaddq		$T1,$D1,$D1
+	vpaddq		$T2,$H2,$H2
+	vpaddq		$T3,$H3,$H3
+	vpaddq		$T4,$H4,$H4
+	vpaddq		$T0,$H0,$H0
+
+	vpermq		\$0x2,$H3,$T3
+	vpermq		\$0x2,$H4,$T4
+	vpermq		\$0x2,$H0,$T0
+	vpermq		\$0x2,$D1,$T1
+	vpermq		\$0x2,$H2,$T2
+	vpaddq		$T3,$H3,$H3
+	vpaddq		$T4,$H4,$H4
+	vpaddq		$T0,$H0,$H0
+	vpaddq		$T1,$D1,$D1
+	vpaddq		$T2,$H2,$H2
+
+	################################################################
+	# lazy reduction
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$D1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H4,$D4
+	vpand		$MASK,$H4,$H4
+
+	vpsrlq		\$26,$H1,$D1
+	vpand		$MASK,$H1,$H1
+	vpaddq		$D1,$H2,$H2		# h1 -> h2
+
+	vpaddq		$D4,$H0,$H0
+	vpsllq		\$2,$D4,$D4
+	vpaddq		$D4,$H0,$H0		# h4 -> h0
+
+	vpsrlq		\$26,$H2,$D2
+	vpand		$MASK,$H2,$H2
+	vpaddq		$D2,$H3,$H3		# h2 -> h3
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	vmovd		%x#$H0,`4*0-48-64`($ctx)# save partially reduced
+	vmovd		%x#$H1,`4*1-48-64`($ctx)
+	vmovd		%x#$H2,`4*2-48-64`($ctx)
+	vmovd		%x#$H3,`4*3-48-64`($ctx)
+	vmovd		%x#$H4,`4*4-48-64`($ctx)
+___
+$code.=<<___	if ($win64);
+	vmovdqa		0x50(%r11),%xmm6
+	vmovdqa		0x60(%r11),%xmm7
+	vmovdqa		0x70(%r11),%xmm8
+	vmovdqa		0x80(%r11),%xmm9
+	vmovdqa		0x90(%r11),%xmm10
+	vmovdqa		0xa0(%r11),%xmm11
+	vmovdqa		0xb0(%r11),%xmm12
+	vmovdqa		0xc0(%r11),%xmm13
+	vmovdqa		0xd0(%r11),%xmm14
+	vmovdqa		0xe0(%r11),%xmm15
+	lea		0xf8(%r11),%rsp
+.Ldo_avx2_epilogue:
+___
+$code.=<<___	if (!$win64);
+	lea		8(%r11),%rsp
+.cfi_def_cfa		%rsp,8
+___
+$code.=<<___;
+	vzeroupper
+	ret
+.cfi_endproc
+.size	poly1305_blocks_avx2,.-poly1305_blocks_avx2
+___
+#######################################################################
+if ($avx>2) {
+# On entry we have input length divisible by 64. But since inner loop
+# processes 128 bytes per iteration, cases when length is not divisible
+# by 128 are handled by passing tail 64 bytes to .Ltail_avx2. For this
+# reason stack layout is kept identical to poly1305_blocks_avx2. If not
+# for this tail, we wouldn't have to even allocate stack frame...
+
+my ($R0,$R1,$R2,$R3,$R4, $S1,$S2,$S3,$S4) = map("%zmm$_",(16..24));
+my ($M0,$M1,$M2,$M3,$M4) = map("%zmm$_",(25..29));
+my $PADBIT="%zmm30";
+
+map(s/%y/%z/,($T4,$T0,$T1,$T2,$T3));		# switch to %zmm domain
+map(s/%y/%z/,($D0,$D1,$D2,$D3,$D4));
+map(s/%y/%z/,($H0,$H1,$H2,$H3,$H4));
+map(s/%y/%z/,($MASK));
+
+$code.=<<___;
+.type	poly1305_blocks_avx512,\@function,4
+.align	32
+poly1305_blocks_avx512:
+.cfi_startproc
+.Lblocks_avx512:
+	mov		\$15,%eax
+	kmovw		%eax,%k2
+___
+$code.=<<___	if (!$win64);
+	lea		-8(%rsp),%r11
+.cfi_def_cfa		%r11,16
+	sub		\$0x128,%rsp
+___
+$code.=<<___	if ($win64);
+	lea		-0xf8(%rsp),%r11
+	sub		\$0x1c8,%rsp
+	vmovdqa		%xmm6,0x50(%r11)
+	vmovdqa		%xmm7,0x60(%r11)
+	vmovdqa		%xmm8,0x70(%r11)
+	vmovdqa		%xmm9,0x80(%r11)
+	vmovdqa		%xmm10,0x90(%r11)
+	vmovdqa		%xmm11,0xa0(%r11)
+	vmovdqa		%xmm12,0xb0(%r11)
+	vmovdqa		%xmm13,0xc0(%r11)
+	vmovdqa		%xmm14,0xd0(%r11)
+	vmovdqa		%xmm15,0xe0(%r11)
+.Ldo_avx512_body:
+___
+$code.=<<___;
+	lea		.Lconst(%rip),%rcx
+	lea		48+64($ctx),$ctx	# size optimization
+	vmovdqa		96(%rcx),%y#$T2		# .Lpermd_avx2
+
+	# expand pre-calculated table
+	vmovdqu		`16*0-64`($ctx),%x#$D0	# will become expanded ${R0}
+	and		\$-512,%rsp
+	vmovdqu		`16*1-64`($ctx),%x#$D1	# will become ... ${R1}
+	mov		\$0x20,%rax
+	vmovdqu		`16*2-64`($ctx),%x#$T0	# ... ${S1}
+	vmovdqu		`16*3-64`($ctx),%x#$D2	# ... ${R2}
+	vmovdqu		`16*4-64`($ctx),%x#$T1	# ... ${S2}
+	vmovdqu		`16*5-64`($ctx),%x#$D3	# ... ${R3}
+	vmovdqu		`16*6-64`($ctx),%x#$T3	# ... ${S3}
+	vmovdqu		`16*7-64`($ctx),%x#$D4	# ... ${R4}
+	vmovdqu		`16*8-64`($ctx),%x#$T4	# ... ${S4}
+	vpermd		$D0,$T2,$R0		# 00003412 -> 14243444
+	vpbroadcastq	64(%rcx),$MASK		# .Lmask26
+	vpermd		$D1,$T2,$R1
+	vpermd		$T0,$T2,$S1
+	vpermd		$D2,$T2,$R2
+	vmovdqa64	$R0,0x00(%rsp){%k2}	# save in case $len%128 != 0
+	 vpsrlq		\$32,$R0,$T0		# 14243444 -> 01020304
+	vpermd		$T1,$T2,$S2
+	vmovdqu64	$R1,0x00(%rsp,%rax){%k2}
+	 vpsrlq		\$32,$R1,$T1
+	vpermd		$D3,$T2,$R3
+	vmovdqa64	$S1,0x40(%rsp){%k2}
+	vpermd		$T3,$T2,$S3
+	vpermd		$D4,$T2,$R4
+	vmovdqu64	$R2,0x40(%rsp,%rax){%k2}
+	vpermd		$T4,$T2,$S4
+	vmovdqa64	$S2,0x80(%rsp){%k2}
+	vmovdqu64	$R3,0x80(%rsp,%rax){%k2}
+	vmovdqa64	$S3,0xc0(%rsp){%k2}
+	vmovdqu64	$R4,0xc0(%rsp,%rax){%k2}
+	vmovdqa64	$S4,0x100(%rsp){%k2}
+
+	################################################################
+	# calculate 5th through 8th powers of the key
+	#
+	# d0 = r0'*r0 + r1'*5*r4 + r2'*5*r3 + r3'*5*r2 + r4'*5*r1
+	# d1 = r0'*r1 + r1'*r0   + r2'*5*r4 + r3'*5*r3 + r4'*5*r2
+	# d2 = r0'*r2 + r1'*r1   + r2'*r0   + r3'*5*r4 + r4'*5*r3
+	# d3 = r0'*r3 + r1'*r2   + r2'*r1   + r3'*r0   + r4'*5*r4
+	# d4 = r0'*r4 + r1'*r3   + r2'*r2   + r3'*r1   + r4'*r0
+
+	vpmuludq	$T0,$R0,$D0		# d0 = r0'*r0
+	vpmuludq	$T0,$R1,$D1		# d1 = r0'*r1
+	vpmuludq	$T0,$R2,$D2		# d2 = r0'*r2
+	vpmuludq	$T0,$R3,$D3		# d3 = r0'*r3
+	vpmuludq	$T0,$R4,$D4		# d4 = r0'*r4
+	 vpsrlq		\$32,$R2,$T2
+
+	vpmuludq	$T1,$S4,$M0
+	vpmuludq	$T1,$R0,$M1
+	vpmuludq	$T1,$R1,$M2
+	vpmuludq	$T1,$R2,$M3
+	vpmuludq	$T1,$R3,$M4
+	 vpsrlq		\$32,$R3,$T3
+	vpaddq		$M0,$D0,$D0		# d0 += r1'*5*r4
+	vpaddq		$M1,$D1,$D1		# d1 += r1'*r0
+	vpaddq		$M2,$D2,$D2		# d2 += r1'*r1
+	vpaddq		$M3,$D3,$D3		# d3 += r1'*r2
+	vpaddq		$M4,$D4,$D4		# d4 += r1'*r3
+
+	vpmuludq	$T2,$S3,$M0
+	vpmuludq	$T2,$S4,$M1
+	vpmuludq	$T2,$R1,$M3
+	vpmuludq	$T2,$R2,$M4
+	vpmuludq	$T2,$R0,$M2
+	 vpsrlq		\$32,$R4,$T4
+	vpaddq		$M0,$D0,$D0		# d0 += r2'*5*r3
+	vpaddq		$M1,$D1,$D1		# d1 += r2'*5*r4
+	vpaddq		$M3,$D3,$D3		# d3 += r2'*r1
+	vpaddq		$M4,$D4,$D4		# d4 += r2'*r2
+	vpaddq		$M2,$D2,$D2		# d2 += r2'*r0
+
+	vpmuludq	$T3,$S2,$M0
+	vpmuludq	$T3,$R0,$M3
+	vpmuludq	$T3,$R1,$M4
+	vpmuludq	$T3,$S3,$M1
+	vpmuludq	$T3,$S4,$M2
+	vpaddq		$M0,$D0,$D0		# d0 += r3'*5*r2
+	vpaddq		$M3,$D3,$D3		# d3 += r3'*r0
+	vpaddq		$M4,$D4,$D4		# d4 += r3'*r1
+	vpaddq		$M1,$D1,$D1		# d1 += r3'*5*r3
+	vpaddq		$M2,$D2,$D2		# d2 += r3'*5*r4
+
+	vpmuludq	$T4,$S4,$M3
+	vpmuludq	$T4,$R0,$M4
+	vpmuludq	$T4,$S1,$M0
+	vpmuludq	$T4,$S2,$M1
+	vpmuludq	$T4,$S3,$M2
+	vpaddq		$M3,$D3,$D3		# d3 += r2'*5*r4
+	vpaddq		$M4,$D4,$D4		# d4 += r2'*r0
+	vpaddq		$M0,$D0,$D0		# d0 += r2'*5*r1
+	vpaddq		$M1,$D1,$D1		# d1 += r2'*5*r2
+	vpaddq		$M2,$D2,$D2		# d2 += r2'*5*r3
+
+	################################################################
+	# load input
+	vmovdqu64	16*0($inp),%z#$T3
+	vmovdqu64	16*4($inp),%z#$T4
+	lea		16*8($inp),$inp
+
+	################################################################
+	# lazy reduction
+
+	vpsrlq		\$26,$D3,$M3
+	vpandq		$MASK,$D3,$D3
+	vpaddq		$M3,$D4,$D4		# d3 -> d4
+
+	vpsrlq		\$26,$D0,$M0
+	vpandq		$MASK,$D0,$D0
+	vpaddq		$M0,$D1,$D1		# d0 -> d1
+
+	vpsrlq		\$26,$D4,$M4
+	vpandq		$MASK,$D4,$D4
+
+	vpsrlq		\$26,$D1,$M1
+	vpandq		$MASK,$D1,$D1
+	vpaddq		$M1,$D2,$D2		# d1 -> d2
+
+	vpaddq		$M4,$D0,$D0
+	vpsllq		\$2,$M4,$M4
+	vpaddq		$M4,$D0,$D0		# d4 -> d0
+
+	vpsrlq		\$26,$D2,$M2
+	vpandq		$MASK,$D2,$D2
+	vpaddq		$M2,$D3,$D3		# d2 -> d3
+
+	vpsrlq		\$26,$D0,$M0
+	vpandq		$MASK,$D0,$D0
+	vpaddq		$M0,$D1,$D1		# d0 -> d1
+
+	vpsrlq		\$26,$D3,$M3
+	vpandq		$MASK,$D3,$D3
+	vpaddq		$M3,$D4,$D4		# d3 -> d4
+
+	################################################################
+	# at this point we have 14243444 in $R0-$S4 and 05060708 in
+	# $D0-$D4, ...
+
+	vpunpcklqdq	$T4,$T3,$T0	# transpose input
+	vpunpckhqdq	$T4,$T3,$T4
+
+	# ... since input 64-bit lanes are ordered as 73625140, we could
+	# "vperm" it to 76543210 (here and in each loop iteration), *or*
+	# we could just flow along, hence the goal for $R0-$S4 is
+	# 1858286838784888 ...
+
+	vmovdqa32	128(%rcx),$M0		# .Lpermd_avx512:
+	mov		\$0x7777,%eax
+	kmovw		%eax,%k1
+
+	vpermd		$R0,$M0,$R0		# 14243444 -> 1---2---3---4---
+	vpermd		$R1,$M0,$R1
+	vpermd		$R2,$M0,$R2
+	vpermd		$R3,$M0,$R3
+	vpermd		$R4,$M0,$R4
+
+	vpermd		$D0,$M0,${R0}{%k1}	# 05060708 -> 1858286838784888
+	vpermd		$D1,$M0,${R1}{%k1}
+	vpermd		$D2,$M0,${R2}{%k1}
+	vpermd		$D3,$M0,${R3}{%k1}
+	vpermd		$D4,$M0,${R4}{%k1}
+
+	vpslld		\$2,$R1,$S1		# *5
+	vpslld		\$2,$R2,$S2
+	vpslld		\$2,$R3,$S3
+	vpslld		\$2,$R4,$S4
+	vpaddd		$R1,$S1,$S1
+	vpaddd		$R2,$S2,$S2
+	vpaddd		$R3,$S3,$S3
+	vpaddd		$R4,$S4,$S4
+
+	vpbroadcastq	32(%rcx),$PADBIT	# .L129
+
+	vpsrlq		\$52,$T0,$T2		# splat input
+	vpsllq		\$12,$T4,$T3
+	vporq		$T3,$T2,$T2
+	vpsrlq		\$26,$T0,$T1
+	vpsrlq		\$14,$T4,$T3
+	vpsrlq		\$40,$T4,$T4		# 4
+	vpandq		$MASK,$T2,$T2		# 2
+	vpandq		$MASK,$T0,$T0		# 0
+	#vpandq		$MASK,$T1,$T1		# 1
+	#vpandq		$MASK,$T3,$T3		# 3
+	#vporq		$PADBIT,$T4,$T4		# padbit, yes, always
+
+	vpaddq		$H2,$T2,$H2		# accumulate input
+	sub		\$192,$len
+	jbe		.Ltail_avx512
+	jmp		.Loop_avx512
+
+.align	32
+.Loop_avx512:
+	################################################################
+	# ((inp[0]*r^8+inp[ 8])*r^8+inp[16])*r^8
+	# ((inp[1]*r^8+inp[ 9])*r^8+inp[17])*r^7
+	# ((inp[2]*r^8+inp[10])*r^8+inp[18])*r^6
+	# ((inp[3]*r^8+inp[11])*r^8+inp[19])*r^5
+	# ((inp[4]*r^8+inp[12])*r^8+inp[20])*r^4
+	# ((inp[5]*r^8+inp[13])*r^8+inp[21])*r^3
+	# ((inp[6]*r^8+inp[14])*r^8+inp[22])*r^2
+	# ((inp[7]*r^8+inp[15])*r^8+inp[23])*r^1
+	#   \________/\___________/
+	################################################################
+	#vpaddq		$H2,$T2,$H2		# accumulate input
+
+	# d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
+	# d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
+	# d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
+	# d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
+	# d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
+	#
+	# however, as h2 is "chronologically" first one available pull
+	# corresponding operations up, so it's
+	#
+	# d3 = h2*r1   + h0*r3 + h1*r2   + h3*r0 + h4*5*r4
+	# d4 = h2*r2   + h0*r4 + h1*r3   + h3*r1 + h4*r0
+	# d0 = h2*5*r3 + h0*r0 + h1*5*r4         + h3*5*r2 + h4*5*r1
+	# d1 = h2*5*r4 + h0*r1           + h1*r0 + h3*5*r3 + h4*5*r2
+	# d2 = h2*r0           + h0*r2   + h1*r1 + h3*5*r4 + h4*5*r3
+
+	vpmuludq	$H2,$R1,$D3		# d3 = h2*r1
+	 vpaddq		$H0,$T0,$H0
+	vpmuludq	$H2,$R2,$D4		# d4 = h2*r2
+	 vpandq		$MASK,$T1,$T1		# 1
+	vpmuludq	$H2,$S3,$D0		# d0 = h2*s3
+	 vpandq		$MASK,$T3,$T3		# 3
+	vpmuludq	$H2,$S4,$D1		# d1 = h2*s4
+	 vporq		$PADBIT,$T4,$T4		# padbit, yes, always
+	vpmuludq	$H2,$R0,$D2		# d2 = h2*r0
+	 vpaddq		$H1,$T1,$H1		# accumulate input
+	 vpaddq		$H3,$T3,$H3
+	 vpaddq		$H4,$T4,$H4
+
+	  vmovdqu64	16*0($inp),$T3		# load input
+	  vmovdqu64	16*4($inp),$T4
+	  lea		16*8($inp),$inp
+	vpmuludq	$H0,$R3,$M3
+	vpmuludq	$H0,$R4,$M4
+	vpmuludq	$H0,$R0,$M0
+	vpmuludq	$H0,$R1,$M1
+	vpaddq		$M3,$D3,$D3		# d3 += h0*r3
+	vpaddq		$M4,$D4,$D4		# d4 += h0*r4
+	vpaddq		$M0,$D0,$D0		# d0 += h0*r0
+	vpaddq		$M1,$D1,$D1		# d1 += h0*r1
+
+	vpmuludq	$H1,$R2,$M3
+	vpmuludq	$H1,$R3,$M4
+	vpmuludq	$H1,$S4,$M0
+	vpmuludq	$H0,$R2,$M2
+	vpaddq		$M3,$D3,$D3		# d3 += h1*r2
+	vpaddq		$M4,$D4,$D4		# d4 += h1*r3
+	vpaddq		$M0,$D0,$D0		# d0 += h1*s4
+	vpaddq		$M2,$D2,$D2		# d2 += h0*r2
+
+	  vpunpcklqdq	$T4,$T3,$T0		# transpose input
+	  vpunpckhqdq	$T4,$T3,$T4
+
+	vpmuludq	$H3,$R0,$M3
+	vpmuludq	$H3,$R1,$M4
+	vpmuludq	$H1,$R0,$M1
+	vpmuludq	$H1,$R1,$M2
+	vpaddq		$M3,$D3,$D3		# d3 += h3*r0
+	vpaddq		$M4,$D4,$D4		# d4 += h3*r1
+	vpaddq		$M1,$D1,$D1		# d1 += h1*r0
+	vpaddq		$M2,$D2,$D2		# d2 += h1*r1
+
+	vpmuludq	$H4,$S4,$M3
+	vpmuludq	$H4,$R0,$M4
+	vpmuludq	$H3,$S2,$M0
+	vpmuludq	$H3,$S3,$M1
+	vpaddq		$M3,$D3,$D3		# d3 += h4*s4
+	vpmuludq	$H3,$S4,$M2
+	vpaddq		$M4,$D4,$D4		# d4 += h4*r0
+	vpaddq		$M0,$D0,$D0		# d0 += h3*s2
+	vpaddq		$M1,$D1,$D1		# d1 += h3*s3
+	vpaddq		$M2,$D2,$D2		# d2 += h3*s4
+
+	vpmuludq	$H4,$S1,$M0
+	vpmuludq	$H4,$S2,$M1
+	vpmuludq	$H4,$S3,$M2
+	vpaddq		$M0,$D0,$H0		# h0 = d0 + h4*s1
+	vpaddq		$M1,$D1,$H1		# h1 = d2 + h4*s2
+	vpaddq		$M2,$D2,$H2		# h2 = d3 + h4*s3
+
+	################################################################
+	# lazy reduction (interleaved with input splat)
+
+	 vpsrlq		\$52,$T0,$T2		# splat input
+	 vpsllq		\$12,$T4,$T3
+
+	vpsrlq		\$26,$D3,$H3
+	vpandq		$MASK,$D3,$D3
+	vpaddq		$H3,$D4,$H4		# h3 -> h4
+
+	 vporq		$T3,$T2,$T2
+
+	vpsrlq		\$26,$H0,$D0
+	vpandq		$MASK,$H0,$H0
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	 vpandq		$MASK,$T2,$T2		# 2
+
+	vpsrlq		\$26,$H4,$D4
+	vpandq		$MASK,$H4,$H4
+
+	vpsrlq		\$26,$H1,$D1
+	vpandq		$MASK,$H1,$H1
+	vpaddq		$D1,$H2,$H2		# h1 -> h2
+
+	vpaddq		$D4,$H0,$H0
+	vpsllq		\$2,$D4,$D4
+	vpaddq		$D4,$H0,$H0		# h4 -> h0
+
+	 vpaddq		$T2,$H2,$H2		# modulo-scheduled
+	 vpsrlq		\$26,$T0,$T1
+
+	vpsrlq		\$26,$H2,$D2
+	vpandq		$MASK,$H2,$H2
+	vpaddq		$D2,$D3,$H3		# h2 -> h3
+
+	 vpsrlq		\$14,$T4,$T3
+
+	vpsrlq		\$26,$H0,$D0
+	vpandq		$MASK,$H0,$H0
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	 vpsrlq		\$40,$T4,$T4		# 4
+
+	vpsrlq		\$26,$H3,$D3
+	vpandq		$MASK,$H3,$H3
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	 vpandq		$MASK,$T0,$T0		# 0
+	 #vpandq	$MASK,$T1,$T1		# 1
+	 #vpandq	$MASK,$T3,$T3		# 3
+	 #vporq		$PADBIT,$T4,$T4		# padbit, yes, always
+
+	sub		\$128,$len
+	ja		.Loop_avx512
+
+.Ltail_avx512:
+	################################################################
+	# while above multiplications were by r^8 in all lanes, in last
+	# iteration we multiply least significant lane by r^8 and most
+	# significant one by r, that's why table gets shifted...
+
+	vpsrlq		\$32,$R0,$R0		# 0105020603070408
+	vpsrlq		\$32,$R1,$R1
+	vpsrlq		\$32,$R2,$R2
+	vpsrlq		\$32,$S3,$S3
+	vpsrlq		\$32,$S4,$S4
+	vpsrlq		\$32,$R3,$R3
+	vpsrlq		\$32,$R4,$R4
+	vpsrlq		\$32,$S1,$S1
+	vpsrlq		\$32,$S2,$S2
+
+	################################################################
+	# load either next or last 64 byte of input
+	lea		($inp,$len),$inp
+
+	#vpaddq		$H2,$T2,$H2		# accumulate input
+	vpaddq		$H0,$T0,$H0
+
+	vpmuludq	$H2,$R1,$D3		# d3 = h2*r1
+	vpmuludq	$H2,$R2,$D4		# d4 = h2*r2
+	vpmuludq	$H2,$S3,$D0		# d0 = h2*s3
+	 vpandq		$MASK,$T1,$T1		# 1
+	vpmuludq	$H2,$S4,$D1		# d1 = h2*s4
+	 vpandq		$MASK,$T3,$T3		# 3
+	vpmuludq	$H2,$R0,$D2		# d2 = h2*r0
+	 vporq		$PADBIT,$T4,$T4		# padbit, yes, always
+	 vpaddq		$H1,$T1,$H1		# accumulate input
+	 vpaddq		$H3,$T3,$H3
+	 vpaddq		$H4,$T4,$H4
+
+	  vmovdqu	16*0($inp),%x#$T0
+	vpmuludq	$H0,$R3,$M3
+	vpmuludq	$H0,$R4,$M4
+	vpmuludq	$H0,$R0,$M0
+	vpmuludq	$H0,$R1,$M1
+	vpaddq		$M3,$D3,$D3		# d3 += h0*r3
+	vpaddq		$M4,$D4,$D4		# d4 += h0*r4
+	vpaddq		$M0,$D0,$D0		# d0 += h0*r0
+	vpaddq		$M1,$D1,$D1		# d1 += h0*r1
+
+	  vmovdqu	16*1($inp),%x#$T1
+	vpmuludq	$H1,$R2,$M3
+	vpmuludq	$H1,$R3,$M4
+	vpmuludq	$H1,$S4,$M0
+	vpmuludq	$H0,$R2,$M2
+	vpaddq		$M3,$D3,$D3		# d3 += h1*r2
+	vpaddq		$M4,$D4,$D4		# d4 += h1*r3
+	vpaddq		$M0,$D0,$D0		# d0 += h1*s4
+	vpaddq		$M2,$D2,$D2		# d2 += h0*r2
+
+	  vinserti128	\$1,16*2($inp),%y#$T0,%y#$T0
+	vpmuludq	$H3,$R0,$M3
+	vpmuludq	$H3,$R1,$M4
+	vpmuludq	$H1,$R0,$M1
+	vpmuludq	$H1,$R1,$M2
+	vpaddq		$M3,$D3,$D3		# d3 += h3*r0
+	vpaddq		$M4,$D4,$D4		# d4 += h3*r1
+	vpaddq		$M1,$D1,$D1		# d1 += h1*r0
+	vpaddq		$M2,$D2,$D2		# d2 += h1*r1
+
+	  vinserti128	\$1,16*3($inp),%y#$T1,%y#$T1
+	vpmuludq	$H4,$S4,$M3
+	vpmuludq	$H4,$R0,$M4
+	vpmuludq	$H3,$S2,$M0
+	vpmuludq	$H3,$S3,$M1
+	vpmuludq	$H3,$S4,$M2
+	vpaddq		$M3,$D3,$H3		# h3 = d3 + h4*s4
+	vpaddq		$M4,$D4,$D4		# d4 += h4*r0
+	vpaddq		$M0,$D0,$D0		# d0 += h3*s2
+	vpaddq		$M1,$D1,$D1		# d1 += h3*s3
+	vpaddq		$M2,$D2,$D2		# d2 += h3*s4
+
+	vpmuludq	$H4,$S1,$M0
+	vpmuludq	$H4,$S2,$M1
+	vpmuludq	$H4,$S3,$M2
+	vpaddq		$M0,$D0,$H0		# h0 = d0 + h4*s1
+	vpaddq		$M1,$D1,$H1		# h1 = d2 + h4*s2
+	vpaddq		$M2,$D2,$H2		# h2 = d3 + h4*s3
+
+	################################################################
+	# horizontal addition
+
+	mov		\$1,%eax
+	vpermq		\$0xb1,$H3,$D3
+	vpermq		\$0xb1,$D4,$H4
+	vpermq		\$0xb1,$H0,$D0
+	vpermq		\$0xb1,$H1,$D1
+	vpermq		\$0xb1,$H2,$D2
+	vpaddq		$D3,$H3,$H3
+	vpaddq		$D4,$H4,$H4
+	vpaddq		$D0,$H0,$H0
+	vpaddq		$D1,$H1,$H1
+	vpaddq		$D2,$H2,$H2
+
+	kmovw		%eax,%k3
+	vpermq		\$0x2,$H3,$D3
+	vpermq		\$0x2,$H4,$D4
+	vpermq		\$0x2,$H0,$D0
+	vpermq		\$0x2,$H1,$D1
+	vpermq		\$0x2,$H2,$D2
+	vpaddq		$D3,$H3,$H3
+	vpaddq		$D4,$H4,$H4
+	vpaddq		$D0,$H0,$H0
+	vpaddq		$D1,$H1,$H1
+	vpaddq		$D2,$H2,$H2
+
+	vextracti64x4	\$0x1,$H3,%y#$D3
+	vextracti64x4	\$0x1,$H4,%y#$D4
+	vextracti64x4	\$0x1,$H0,%y#$D0
+	vextracti64x4	\$0x1,$H1,%y#$D1
+	vextracti64x4	\$0x1,$H2,%y#$D2
+	vpaddq		$D3,$H3,${H3}{%k3}{z}	# keep single qword in case
+	vpaddq		$D4,$H4,${H4}{%k3}{z}	# it's passed to .Ltail_avx2
+	vpaddq		$D0,$H0,${H0}{%k3}{z}
+	vpaddq		$D1,$H1,${H1}{%k3}{z}
+	vpaddq		$D2,$H2,${H2}{%k3}{z}
+___
+map(s/%z/%y/,($T0,$T1,$T2,$T3,$T4, $PADBIT));
+map(s/%z/%y/,($H0,$H1,$H2,$H3,$H4, $D0,$D1,$D2,$D3,$D4, $MASK));
+$code.=<<___;
+	################################################################
+	# lazy reduction (interleaved with input splat)
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	 vpsrldq	\$6,$T0,$T2		# splat input
+	 vpsrldq	\$6,$T1,$T3
+	 vpunpckhqdq	$T1,$T0,$T4		# 4
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	 vpunpcklqdq	$T3,$T2,$T2		# 2:3
+	 vpunpcklqdq	$T1,$T0,$T0		# 0:1
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H4,$D4
+	vpand		$MASK,$H4,$H4
+
+	vpsrlq		\$26,$H1,$D1
+	vpand		$MASK,$H1,$H1
+	 vpsrlq		\$30,$T2,$T3
+	 vpsrlq		\$4,$T2,$T2
+	vpaddq		$D1,$H2,$H2		# h1 -> h2
+
+	vpaddq		$D4,$H0,$H0
+	vpsllq		\$2,$D4,$D4
+	 vpsrlq		\$26,$T0,$T1
+	 vpsrlq		\$40,$T4,$T4		# 4
+	vpaddq		$D4,$H0,$H0		# h4 -> h0
+
+	vpsrlq		\$26,$H2,$D2
+	vpand		$MASK,$H2,$H2
+	 vpand		$MASK,$T2,$T2		# 2
+	 vpand		$MASK,$T0,$T0		# 0
+	vpaddq		$D2,$H3,$H3		# h2 -> h3
+
+	vpsrlq		\$26,$H0,$D0
+	vpand		$MASK,$H0,$H0
+	 vpaddq		$H2,$T2,$H2		# accumulate input for .Ltail_avx2
+	 vpand		$MASK,$T1,$T1		# 1
+	vpaddq		$D0,$H1,$H1		# h0 -> h1
+
+	vpsrlq		\$26,$H3,$D3
+	vpand		$MASK,$H3,$H3
+	 vpand		$MASK,$T3,$T3		# 3
+	 vpor		32(%rcx),$T4,$T4	# padbit, yes, always
+	vpaddq		$D3,$H4,$H4		# h3 -> h4
+
+	lea		0x90(%rsp),%rax		# size optimization for .Ltail_avx2
+	add		\$64,$len
+	jnz		.Ltail_avx2
+
+	vpsubq		$T2,$H2,$H2		# undo input accumulation
+	vmovd		%x#$H0,`4*0-48-64`($ctx)# save partially reduced
+	vmovd		%x#$H1,`4*1-48-64`($ctx)
+	vmovd		%x#$H2,`4*2-48-64`($ctx)
+	vmovd		%x#$H3,`4*3-48-64`($ctx)
+	vmovd		%x#$H4,`4*4-48-64`($ctx)
+	vzeroall
+___
+$code.=<<___	if ($win64);
+	movdqa		0x50(%r11),%xmm6
+	movdqa		0x60(%r11),%xmm7
+	movdqa		0x70(%r11),%xmm8
+	movdqa		0x80(%r11),%xmm9
+	movdqa		0x90(%r11),%xmm10
+	movdqa		0xa0(%r11),%xmm11
+	movdqa		0xb0(%r11),%xmm12
+	movdqa		0xc0(%r11),%xmm13
+	movdqa		0xd0(%r11),%xmm14
+	movdqa		0xe0(%r11),%xmm15
+	lea		0xf8(%r11),%rsp
+.Ldo_avx512_epilogue:
+___
+$code.=<<___	if (!$win64);
+	lea		8(%r11),%rsp
+.cfi_def_cfa		%rsp,8
+___
+$code.=<<___;
+	ret
+.cfi_endproc
+.size	poly1305_blocks_avx512,.-poly1305_blocks_avx512
+___
+if ($avx>3) {
+########################################################################
+# VPMADD52 version using 2^44 radix.
+#
+# One can argue that base 2^52 would be more natural. Well, even though
+# some operations would be more natural, one has to recognize couple of
+# things. Base 2^52 doesn't provide advantage over base 2^44 if you look
+# at amount of multiply-n-accumulate operations. Secondly, it makes it
+# impossible to pre-compute multiples of 5 [referred to as s[]/sN in
+# reference implementations], which means that more such operations
+# would have to be performed in inner loop, which in turn makes critical
+# path longer. In other words, even though base 2^44 reduction might
+# look less elegant, overall critical path is actually shorter...
+
+########################################################################
+# Layout of opaque area is following.
+#
+#	unsigned __int64 h[3];		# current hash value base 2^44
+#	unsigned __int64 s[2];		# key value*20 base 2^44
+#	unsigned __int64 r[3];		# key value base 2^44
+#	struct { unsigned __int64 r^1, r^3, r^2, r^4; } R[4];
+#					# r^n positions reflect
+#					# placement in register, not
+#					# memory, R[3] is R[1]*20
+
+$code.=<<___;
+.type	poly1305_init_base2_44,\@function,3
+.align	32
+poly1305_init_base2_44:
+	xor	%rax,%rax
+	mov	%rax,0($ctx)		# initialize hash value
+	mov	%rax,8($ctx)
+	mov	%rax,16($ctx)
+
+.Linit_base2_44:
+	lea	poly1305_blocks_vpmadd52(%rip),%r10
+	lea	poly1305_emit_base2_44(%rip),%r11
+
+	mov	\$0x0ffffffc0fffffff,%rax
+	mov	\$0x0ffffffc0ffffffc,%rcx
+	and	0($inp),%rax
+	mov	\$0x00000fffffffffff,%r8
+	and	8($inp),%rcx
+	mov	\$0x00000fffffffffff,%r9
+	and	%rax,%r8
+	shrd	\$44,%rcx,%rax
+	mov	%r8,40($ctx)		# r0
+	and	%r9,%rax
+	shr	\$24,%rcx
+	mov	%rax,48($ctx)		# r1
+	lea	(%rax,%rax,4),%rax	# *5
+	mov	%rcx,56($ctx)		# r2
+	shl	\$2,%rax		# magic <<2
+	lea	(%rcx,%rcx,4),%rcx	# *5
+	shl	\$2,%rcx		# magic <<2
+	mov	%rax,24($ctx)		# s1
+	mov	%rcx,32($ctx)		# s2
+	movq	\$-1,64($ctx)		# write impossible value
+___
+$code.=<<___	if ($flavour !~ /elf32/);
+	mov	%r10,0(%rdx)
+	mov	%r11,8(%rdx)
+___
+$code.=<<___	if ($flavour =~ /elf32/);
+	mov	%r10d,0(%rdx)
+	mov	%r11d,4(%rdx)
+___
+$code.=<<___;
+	mov	\$1,%eax
+	ret
+.size	poly1305_init_base2_44,.-poly1305_init_base2_44
+___
+{
+my ($H0,$H1,$H2,$r2r1r0,$r1r0s2,$r0s2s1,$Dlo,$Dhi) = map("%ymm$_",(0..5,16,17));
+my ($T0,$inp_permd,$inp_shift,$PAD) = map("%ymm$_",(18..21));
+my ($reduc_mask,$reduc_rght,$reduc_left) = map("%ymm$_",(22..25));
+
+$code.=<<___;
+.type	poly1305_blocks_vpmadd52,\@function,4
+.align	32
+poly1305_blocks_vpmadd52:
+	shr	\$4,$len
+	jz	.Lno_data_vpmadd52		# too short
+
+	shl	\$40,$padbit
+	mov	64($ctx),%r8			# peek on power of the key
+
+	# if powers of the key are not calculated yet, process up to 3
+	# blocks with this single-block subroutine, otherwise ensure that
+	# length is divisible by 2 blocks and pass the rest down to next
+	# subroutine...
+
+	mov	\$3,%rax
+	mov	\$1,%r10
+	cmp	\$4,$len			# is input long
+	cmovae	%r10,%rax
+	test	%r8,%r8				# is power value impossible?
+	cmovns	%r10,%rax
+
+	and	$len,%rax			# is input of favourable length?
+	jz	.Lblocks_vpmadd52_4x
+
+	sub		%rax,$len
+	mov		\$7,%r10d
+	mov		\$1,%r11d
+	kmovw		%r10d,%k7
+	lea		.L2_44_inp_permd(%rip),%r10
+	kmovw		%r11d,%k1
+
+	vmovq		$padbit,%x#$PAD
+	vmovdqa64	0(%r10),$inp_permd	# .L2_44_inp_permd
+	vmovdqa64	32(%r10),$inp_shift	# .L2_44_inp_shift
+	vpermq		\$0xcf,$PAD,$PAD
+	vmovdqa64	64(%r10),$reduc_mask	# .L2_44_mask
+
+	vmovdqu64	0($ctx),${Dlo}{%k7}{z}		# load hash value
+	vmovdqu64	40($ctx),${r2r1r0}{%k7}{z}	# load keys
+	vmovdqu64	32($ctx),${r1r0s2}{%k7}{z}
+	vmovdqu64	24($ctx),${r0s2s1}{%k7}{z}
+
+	vmovdqa64	96(%r10),$reduc_rght	# .L2_44_shift_rgt
+	vmovdqa64	128(%r10),$reduc_left	# .L2_44_shift_lft
+
+	jmp		.Loop_vpmadd52
+
+.align	32
+.Loop_vpmadd52:
+	vmovdqu32	0($inp),%x#$T0		# load input as ----3210
+	lea		16($inp),$inp
+
+	vpermd		$T0,$inp_permd,$T0	# ----3210 -> --322110
+	vpsrlvq		$inp_shift,$T0,$T0
+	vpandq		$reduc_mask,$T0,$T0
+	vporq		$PAD,$T0,$T0
+
+	vpaddq		$T0,$Dlo,$Dlo		# accumulate input
+
+	vpermq		\$0,$Dlo,${H0}{%k7}{z}	# smash hash value
+	vpermq		\$0b01010101,$Dlo,${H1}{%k7}{z}
+	vpermq		\$0b10101010,$Dlo,${H2}{%k7}{z}
+
+	vpxord		$Dlo,$Dlo,$Dlo
+	vpxord		$Dhi,$Dhi,$Dhi
+
+	vpmadd52luq	$r2r1r0,$H0,$Dlo
+	vpmadd52huq	$r2r1r0,$H0,$Dhi
+
+	vpmadd52luq	$r1r0s2,$H1,$Dlo
+	vpmadd52huq	$r1r0s2,$H1,$Dhi
+
+	vpmadd52luq	$r0s2s1,$H2,$Dlo
+	vpmadd52huq	$r0s2s1,$H2,$Dhi
+
+	vpsrlvq		$reduc_rght,$Dlo,$T0	# 0 in topmost qword
+	vpsllvq		$reduc_left,$Dhi,$Dhi	# 0 in topmost qword
+	vpandq		$reduc_mask,$Dlo,$Dlo
+
+	vpaddq		$T0,$Dhi,$Dhi
+
+	vpermq		\$0b10010011,$Dhi,$Dhi	# 0 in lowest qword
+
+	vpaddq		$Dhi,$Dlo,$Dlo		# note topmost qword :-)
+
+	vpsrlvq		$reduc_rght,$Dlo,$T0	# 0 in topmost word
+	vpandq		$reduc_mask,$Dlo,$Dlo
+
+	vpermq		\$0b10010011,$T0,$T0
+
+	vpaddq		$T0,$Dlo,$Dlo
+
+	vpermq		\$0b10010011,$Dlo,${T0}{%k1}{z}
+
+	vpaddq		$T0,$Dlo,$Dlo
+	vpsllq		\$2,$T0,$T0
+
+	vpaddq		$T0,$Dlo,$Dlo
+
+	dec		%rax			# len-=16
+	jnz		.Loop_vpmadd52
+
+	vmovdqu64	$Dlo,0($ctx){%k7}	# store hash value
+
+	test		$len,$len
+	jnz		.Lblocks_vpmadd52_4x
+
+.Lno_data_vpmadd52:
+	ret
+.size	poly1305_blocks_vpmadd52,.-poly1305_blocks_vpmadd52
+___
+}
+{
+########################################################################
+# As implied by its name 4x subroutine processes 4 blocks in parallel
+# (but handles even 4*n+2 blocks lengths). It takes up to 4th key power
+# and is handled in 256-bit %ymm registers.
+
+my ($H0,$H1,$H2,$R0,$R1,$R2,$S1,$S2) = map("%ymm$_",(0..5,16,17));
+my ($D0lo,$D0hi,$D1lo,$D1hi,$D2lo,$D2hi) = map("%ymm$_",(18..23));
+my ($T0,$T1,$T2,$T3,$mask44,$mask42,$tmp,$PAD) = map("%ymm$_",(24..31));
+
+$code.=<<___;
+.type	poly1305_blocks_vpmadd52_4x,\@function,4
+.align	32
+poly1305_blocks_vpmadd52_4x:
+	shr	\$4,$len
+	jz	.Lno_data_vpmadd52_4x		# too short
+
+	shl	\$40,$padbit
+	mov	64($ctx),%r8			# peek on power of the key
+
+.Lblocks_vpmadd52_4x:
+	vpbroadcastq	$padbit,$PAD
+
+	vmovdqa64	.Lx_mask44(%rip),$mask44
+	mov		\$5,%eax
+	vmovdqa64	.Lx_mask42(%rip),$mask42
+	kmovw		%eax,%k1		# used in 2x path
+
+	test		%r8,%r8			# is power value impossible?
+	js		.Linit_vpmadd52		# if it is, then init R[4]
+
+	vmovq		0($ctx),%x#$H0		# load current hash value
+	vmovq		8($ctx),%x#$H1
+	vmovq		16($ctx),%x#$H2
+
+	test		\$3,$len		# is length 4*n+2?
+	jnz		.Lblocks_vpmadd52_2x_do
+
+.Lblocks_vpmadd52_4x_do:
+	vpbroadcastq	64($ctx),$R0		# load 4th power of the key
+	vpbroadcastq	96($ctx),$R1
+	vpbroadcastq	128($ctx),$R2
+	vpbroadcastq	160($ctx),$S1
+
+.Lblocks_vpmadd52_4x_key_loaded:
+	vpsllq		\$2,$R2,$S2		# S2 = R2*5*4
+	vpaddq		$R2,$S2,$S2
+	vpsllq		\$2,$S2,$S2
+
+	test		\$7,$len		# is len 8*n?
+	jz		.Lblocks_vpmadd52_8x
+
+	vmovdqu64	16*0($inp),$T2		# load data
+	vmovdqu64	16*2($inp),$T3
+	lea		16*4($inp),$inp
+
+	vpunpcklqdq	$T3,$T2,$T1		# transpose data
+	vpunpckhqdq	$T3,$T2,$T3
+
+	# at this point 64-bit lanes are ordered as 3-1-2-0
+
+	vpsrlq		\$24,$T3,$T2		# splat the data
+	vporq		$PAD,$T2,$T2
+	 vpaddq		$T2,$H2,$H2		# accumulate input
+	vpandq		$mask44,$T1,$T0
+	vpsrlq		\$44,$T1,$T1
+	vpsllq		\$20,$T3,$T3
+	vporq		$T3,$T1,$T1
+	vpandq		$mask44,$T1,$T1
+
+	sub		\$4,$len
+	jz		.Ltail_vpmadd52_4x
+	jmp		.Loop_vpmadd52_4x
+	ud2
+
+.align	32
+.Linit_vpmadd52:
+	vmovq		24($ctx),%x#$S1		# load key
+	vmovq		56($ctx),%x#$H2
+	vmovq		32($ctx),%x#$S2
+	vmovq		40($ctx),%x#$R0
+	vmovq		48($ctx),%x#$R1
+
+	vmovdqa		$R0,$H0
+	vmovdqa		$R1,$H1
+	vmovdqa		$H2,$R2
+
+	mov		\$2,%eax
+
+.Lmul_init_vpmadd52:
+	vpxorq		$D0lo,$D0lo,$D0lo
+	vpmadd52luq	$H2,$S1,$D0lo
+	vpxorq		$D0hi,$D0hi,$D0hi
+	vpmadd52huq	$H2,$S1,$D0hi
+	vpxorq		$D1lo,$D1lo,$D1lo
+	vpmadd52luq	$H2,$S2,$D1lo
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Signature · 1 line
2.24.1

Part of the diff is not shown here. View the raw message.

author maintainer
#3 lore raw

PATCH v7 3/4 crypto: x86_64/poly1305 - wire up faster implementations for kernel

In reply to Jason A. Donenfeld

These x86_64 vectorized implementations support AVX, AVX-2, and AVX512F. The AVX-512F implementation is disabled on Skylake, due to throttling, but it is quite fast on >= Cannonlake.

On the left is cycle counts on a Core i7 6700HQ using the AVX-2 codepath, comparing this implementation ("new") to the implementation in the current crypto api ("old"). On the right are benchmarks on a Xeon Gold 5120 using the AVX-512 codepath. The new implementation is faster on all benchmarks.

        AVX-2                  AVX-512
      ---------              -----------

    size    old     new      size   old     new
    ----    ----    ----     ----   ----    ----
    0       70      68       0      74      70
    16      92      90       16     96      92
    32      134     104      32     136     106
    48      172     120      48     184     124
    64      218     136      64     218     138
    80      254     158      80     260     160
    96      298     174      96     300     176
    112     342     192      112    342     194
    128     388     212      128    384     212
    144     428     228      144    420     226
    160     466     246      160    464     248
    176     510     264      176    504     264
    192     550     282      192    544     282
    208     594     302      208    582     300
    224     628     316      224    624     318
    240     676     334      240    662     338
    256     716     354      256    708     358
    272     764     374      272    748     372
    288     802     352      288    788     358
    304     420     366      304    422     370
    320     428     360      320    432     364
    336     484     378      336    486     380
    352     426     384      352    434     390
    368     478     400      368    480     408
    384     488     394      384    490     398
    400     542     408      400    542     412
    416     486     416      416    492     426
    432     534     430      432    538     436
    448     544     422      448    546     432
    464     600     438      464    600     448
    480     540     448      480    548     456
    496     594     464      496    594     476
    512     602     456      512    606     470
    528     656     476      528    656     480
    544     600     480      544    606     498
    560     650     494      560    652     512
    576     664     490      576    662     508
    592     714     508      592    716     522
    608     656     514      608    664     538
    624     708     532      624    710     552
    640     716     524      640    720     516
    656     770     536      656    772     526
    672     716     548      672    722     544
    688     770     562      688    768     556
    704     774     552      704    778     556
    720     826     568      720    832     568
    736     768     574      736    780     584
    752     822     592      752    826     600
    768     830     584      768    836     560
    784     884     602      784    888     572
    800     828     610      800    838     588
    816     884     628      816    884     604
    832     888     618      832    894     598
    848     942     632      848    946     612
    864     884     644      864    896     628
    880     936     660      880    942     644
    896     948     652      896    952     608
    912     1000    664      912    1004    616
    928     942     676      928    954     634
    944     994     690      944    1000    646
    960     1002    680      960    1008    646
    976     1054    694      976    1062    658
    992     1002    706      992    1012    674
    1008    1052    720      1008   1058    690

This commit wires in the prior implementation from Andy, and makes the following changes to be suitable for kernel land.

- Some cosmetic and structural changes, like renaming labels to .Lname, constants, and other Linux conventions, as well as making the code easy for us to maintain moving forward.

- CPU feature checking is done in C by the glue code.

- We avoid jumping into the middle of functions, to appease objtool, and instead parameterize shared code.

- We maintain frame pointers so that stack traces make sense.

- We remove the dependency on the perl xlate code, which transforms the output into things that assemblers we don't care about use.

Importantly, none of our changes affect the arithmetic or core code, but just involve the differing environment of kernel space.

---
 arch/x86/crypto/.gitignore             |   1 +
 arch/x86/crypto/Makefile               |  11 +-
 arch/x86/crypto/poly1305-avx2-x86_64.S | 390 --------------
 arch/x86/crypto/poly1305-sse2-x86_64.S | 590 ---------------------
 arch/x86/crypto/poly1305-x86_64.pl     | 682 ++++++++++++++-----------
 arch/x86/crypto/poly1305_glue.c        | 473 ++++++-----------
 lib/crypto/Kconfig                     |   2 +-
 7 files changed, 572 insertions(+), 1577 deletions(-)
 create mode 100644 arch/x86/crypto/.gitignore
 delete mode 100644 arch/x86/crypto/poly1305-avx2-x86_64.S
 delete mode 100644 arch/x86/crypto/poly1305-sse2-x86_64.S
arch/x86/crypto/.gitignore new file +1 -0
new file mode 100644
index 000000000000..c406ea6571fa
@@ -0,0 +1 @@
+poly1305-x86_64.S
arch/x86/crypto/Makefile +9 -2
index 958440eae27e..6982a2f8863f 100644
@@ -73,6 +73,10 @@ aegis128-aesni-y := aegis128-aesni-asm.o aegis128-aesni-glue.o
 
 nhpoly1305-sse2-y := nh-sse2-x86_64.o nhpoly1305-sse2-glue.o
 blake2s-x86_64-y := blake2s-core.o blake2s-glue.o
+poly1305-x86_64-y := poly1305-x86_64.o poly1305_glue.o
+ifneq ($(CONFIG_CRYPTO_POLY1305_X86_64),)
+targets += poly1305-x86_64.S
+endif
 
 ifeq ($(avx_supported),yes)
 	camellia-aesni-avx-x86_64-y := camellia-aesni-avx-asm_64.o \
@@ -101,10 +105,8 @@ aesni-intel-y := aesni-intel_asm.o aesni-intel_glue.o
 aesni-intel-$(CONFIG_64BIT) += aesni-intel_avx-x86_64.o aes_ctrby8_avx-x86_64.o
 ghash-clmulni-intel-y := ghash-clmulni-intel_asm.o ghash-clmulni-intel_glue.o
 sha1-ssse3-y := sha1_ssse3_asm.o sha1_ssse3_glue.o
-poly1305-x86_64-y := poly1305-sse2-x86_64.o poly1305_glue.o
 ifeq ($(avx2_supported),yes)
 sha1-ssse3-y += sha1_avx2_x86_64_asm.o
-poly1305-x86_64-y += poly1305-avx2-x86_64.o
 endif
 ifeq ($(sha1_ni_supported),yes)
 sha1-ssse3-y += sha1_ni_asm.o
@@ -118,3 +120,8 @@ sha256-ssse3-y += sha256_ni_asm.o
 endif
 sha512-ssse3-y := sha512-ssse3-asm.o sha512-avx-asm.o sha512-avx2-asm.o sha512_ssse3_glue.o
 crct10dif-pclmul-y := crct10dif-pcl-asm_64.o crct10dif-pclmul_glue.o
+
+quiet_cmd_perlasm = PERLASM $@
+      cmd_perlasm = $(PERL) $< > $@
+$(obj)/%.S: $(src)/%.pl FORCE
+	$(call if_changed,perlasm)
arch/x86/crypto/poly1305-avx2-x86_64.S deleted +0 -390

390 lines not on this page · raw message

arch/x86/crypto/poly1305-sse2-x86_64.S deleted +0 -590

590 lines not on this page · raw message

arch/x86/crypto/poly1305-x86_64.pl +394 -288

1,127 lines not on this page · raw message

arch/x86/crypto/poly1305_glue.c +167 -306

578 lines not on this page · raw message

lib/crypto/Kconfig +1 -1
index 0b2c4fce26d9..14c032de276e 100644
@@ -90,7 +90,7 @@ config CRYPTO_LIB_DES
 config CRYPTO_LIB_POLY1305_RSIZE
 	int
 	default 2 if MIPS
-	default 4 if X86_64
+	default 11 if X86_64
 	default 9 if ARM || ARM64
 	default 1
 
Signature · 1 line
2.24.1
#4 lore raw

Re: [PATCH v7 3/4] crypto: x86_64/poly1305 - wire up faster implementations for kernel

I love your patch! Yet something to improve:

Hi "Jason,

I love your patch! Yet something to improve:

[auto build test ERROR on cryptodev/master]
[also build test ERROR on v5.5-rc4 next-20191220]
[cannot apply to crypto/master]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system. BTW, we also suggest to use '--base' option to specify the base tree in git format-patch, please see https://stackoverflow.com/a/37406982]

url:    https://github.com/0day-ci/linux/commits/Jason-A-Donenfeld/crypto-poly1305-improvements/20200105-084721
base:   https://git.kernel.org/pub/scm/linux/kernel/git/herbert/cryptodev-2.6.git master
config: x86_64-fedora-25 (attached as .config)
compiler: gcc-7 (Debian 7.5.0-3) 7.5.0
reproduce:
        # save the attached .config to linux build tree
        make ARCH=x86_64 

If you fix the issue, kindly add following tag

All errors (new ones prefixed by >>):

>> ERROR: "poly1305_init_x86_64" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_blocks_avx2" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_blocks_avx512" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_emit_avx" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_blocks_avx" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_emit_x86_64" [arch/x86/crypto/poly1305-x86_64.ko] undefined!
>> ERROR: "poly1305_blocks_x86_64" [arch/x86/crypto/poly1305-x86_64.ko] undefined!

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org Intel Corporation

author maintainer
#5 lore raw

PATCH v7 4/4 crypto: arm/arm64/mips/poly1305 - remove redundant non-reduction from emit

In reply to Jason A. Donenfeld

This appears to be some kind of copy and paste error, and is actually dead code.

Pre: f = 0 ⇒ (f >> 32) = 0

    f = (f >> 32) + le32_to_cpu(digest[0]);
Post: 0 ≤ f < 2³²
    put_unaligned_le32(f, dst);

Pre: 0 ≤ f < 2³² ⇒ (f >> 32) = 0

    f = (f >> 32) + le32_to_cpu(digest[1]);
Post: 0 ≤ f < 2³²
    put_unaligned_le32(f, dst + 4);

Pre: 0 ≤ f < 2³² ⇒ (f >> 32) = 0

    f = (f >> 32) + le32_to_cpu(digest[2]);
Post: 0 ≤ f < 2³²
    put_unaligned_le32(f, dst + 8);

Pre: 0 ≤ f < 2³² ⇒ (f >> 32) = 0

    f = (f >> 32) + le32_to_cpu(digest[3]);
Post: 0 ≤ f < 2³²
    put_unaligned_le32(f, dst + 12);

Therefore this sequence is redundant. And Andy's code appears to handle misalignment acceptably.

---
 arch/arm/crypto/poly1305-glue.c   | 18 ++----------------
 arch/arm64/crypto/poly1305-glue.c | 18 ++----------------
 arch/mips/crypto/poly1305-glue.c  | 18 ++----------------
 3 files changed, 6 insertions(+), 48 deletions(-)
arch/arm/crypto/poly1305-glue.c +2 -16
index abe3f2d587dc..ceec04ec2f40 100644
@@ -20,7 +20,7 @@
 
 void poly1305_init_arm(void *state, const u8 *key);
 void poly1305_blocks_arm(void *state, const u8 *src, u32 len, u32 hibit);
-void poly1305_emit_arm(void *state, __le32 *digest, const u32 *nonce);
+void poly1305_emit_arm(void *state, u8 *digest, const u32 *nonce);
 
 void __weak poly1305_blocks_neon(void *state, const u8 *src, u32 len, u32 hibit)
 {
@@ -179,9 +179,6 @@ EXPORT_SYMBOL(poly1305_update_arch);
 
 void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 {
-	__le32 digest[4];
-	u64 f = 0;
-
 	if (unlikely(dctx->buflen)) {
 		dctx->buf[dctx->buflen++] = 1;
 		memset(dctx->buf + dctx->buflen, 0,
@@ -189,18 +186,7 @@ void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 		poly1305_blocks_arm(&dctx->h, dctx->buf, POLY1305_BLOCK_SIZE, 0);
 	}
 
-	poly1305_emit_arm(&dctx->h, digest, dctx->s);
-
-	/* mac = (h + s) % (2^128) */
-	f = (f >> 32) + le32_to_cpu(digest[0]);
-	put_unaligned_le32(f, dst);
-	f = (f >> 32) + le32_to_cpu(digest[1]);
-	put_unaligned_le32(f, dst + 4);
-	f = (f >> 32) + le32_to_cpu(digest[2]);
-	put_unaligned_le32(f, dst + 8);
-	f = (f >> 32) + le32_to_cpu(digest[3]);
-	put_unaligned_le32(f, dst + 12);
-
+	poly1305_emit_arm(&dctx->h, dst, dctx->s);
 	*dctx = (struct poly1305_desc_ctx){};
 }
 EXPORT_SYMBOL(poly1305_final_arch);
arch/arm64/crypto/poly1305-glue.c +2 -16
index 83a2338a8826..e97b092f56b8 100644
@@ -21,7 +21,7 @@
 asmlinkage void poly1305_init_arm64(void *state, const u8 *key);
 asmlinkage void poly1305_blocks(void *state, const u8 *src, u32 len, u32 hibit);
 asmlinkage void poly1305_blocks_neon(void *state, const u8 *src, u32 len, u32 hibit);
-asmlinkage void poly1305_emit(void *state, __le32 *digest, const u32 *nonce);
+asmlinkage void poly1305_emit(void *state, u8 *digest, const u32 *nonce);
 
 static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_neon);
 
@@ -162,9 +162,6 @@ EXPORT_SYMBOL(poly1305_update_arch);
 
 void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 {
-	__le32 digest[4];
-	u64 f = 0;
-
 	if (unlikely(dctx->buflen)) {
 		dctx->buf[dctx->buflen++] = 1;
 		memset(dctx->buf + dctx->buflen, 0,
@@ -172,18 +169,7 @@ void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 		poly1305_blocks(&dctx->h, dctx->buf, POLY1305_BLOCK_SIZE, 0);
 	}
 
-	poly1305_emit(&dctx->h, digest, dctx->s);
-
-	/* mac = (h + s) % (2^128) */
-	f = (f >> 32) + le32_to_cpu(digest[0]);
-	put_unaligned_le32(f, dst);
-	f = (f >> 32) + le32_to_cpu(digest[1]);
-	put_unaligned_le32(f, dst + 4);
-	f = (f >> 32) + le32_to_cpu(digest[2]);
-	put_unaligned_le32(f, dst + 8);
-	f = (f >> 32) + le32_to_cpu(digest[3]);
-	put_unaligned_le32(f, dst + 12);
-
+	poly1305_emit(&dctx->h, dst, dctx->s);
 	*dctx = (struct poly1305_desc_ctx){};
 }
 EXPORT_SYMBOL(poly1305_final_arch);
arch/mips/crypto/poly1305-glue.c +2 -16
index b37d29cf5d0a..fc881b46d911 100644
@@ -15,7 +15,7 @@
 
 asmlinkage void poly1305_init_mips(void *state, const u8 *key);
 asmlinkage void poly1305_blocks_mips(void *state, const u8 *src, u32 len, u32 hibit);
-asmlinkage void poly1305_emit_mips(void *state, __le32 *digest, const u32 *nonce);
+asmlinkage void poly1305_emit_mips(void *state, u8 *digest, const u32 *nonce);
 
 void poly1305_init_arch(struct poly1305_desc_ctx *dctx, const u8 *key)
 {
@@ -134,9 +134,6 @@ EXPORT_SYMBOL(poly1305_update_arch);
 
 void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 {
-	__le32 digest[4];
-	u64 f = 0;
-
 	if (unlikely(dctx->buflen)) {
 		dctx->buf[dctx->buflen++] = 1;
 		memset(dctx->buf + dctx->buflen, 0,
@@ -144,18 +141,7 @@ void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst)
 		poly1305_blocks_mips(&dctx->h, dctx->buf, POLY1305_BLOCK_SIZE, 0);
 	}
 
-	poly1305_emit_mips(&dctx->h, digest, dctx->s);
-
-	/* mac = (h + s) % (2^128) */
-	f = (f >> 32) + le32_to_cpu(digest[0]);
-	put_unaligned_le32(f, dst);
-	f = (f >> 32) + le32_to_cpu(digest[1]);
-	put_unaligned_le32(f, dst + 4);
-	f = (f >> 32) + le32_to_cpu(digest[2]);
-	put_unaligned_le32(f, dst + 8);
-	f = (f >> 32) + le32_to_cpu(digest[3]);
-	put_unaligned_le32(f, dst + 12);
-
+	poly1305_emit_mips(&dctx->h, dst, dctx->s);
 	*dctx = (struct poly1305_desc_ctx){};
 }
 EXPORT_SYMBOL(poly1305_final_arch);
Signature · 1 line
2.24.1