aboutsummaryrefslogtreecommitdiff
path: root/openssl/crypto/ec/asm/ecp_nistz256-avx2.pl
diff options
context:
space:
mode:
Diffstat (limited to 'openssl/crypto/ec/asm/ecp_nistz256-avx2.pl')
-rwxr-xr-xopenssl/crypto/ec/asm/ecp_nistz256-avx2.pl2093
1 files changed, 2093 insertions, 0 deletions
diff --git a/openssl/crypto/ec/asm/ecp_nistz256-avx2.pl b/openssl/crypto/ec/asm/ecp_nistz256-avx2.pl
new file mode 100755
index 000000000..4c220aa64
--- /dev/null
+++ b/openssl/crypto/ec/asm/ecp_nistz256-avx2.pl
@@ -0,0 +1,2093 @@
+#!/usr/bin/env perl
+
+##############################################################################
+# #
+# Copyright 2014 Intel Corporation #
+# #
+# Licensed under the Apache License, Version 2.0 (the "License"); #
+# you may not use this file except in compliance with the License. #
+# You may obtain a copy of the License at #
+# #
+# http://www.apache.org/licenses/LICENSE-2.0 #
+# #
+# Unless required by applicable law or agreed to in writing, software #
+# distributed under the License is distributed on an "AS IS" BASIS, #
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #
+# See the License for the specific language governing permissions and #
+# limitations under the License. #
+# #
+##############################################################################
+# #
+# Developers and authors: #
+# Shay Gueron (1, 2), and Vlad Krasnov (1) #
+# (1) Intel Corporation, Israel Development Center #
+# (2) University of Haifa #
+# Reference: #
+# S.Gueron and V.Krasnov, "Fast Prime Field Elliptic Curve Cryptography with#
+# 256 Bit Primes" #
+# #
+##############################################################################
+
+$flavour = shift;
+$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 $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
+
+if (`$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);
+ $addx = ($1>=2.23);
+}
+
+if (!$addx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
+ `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
+ $avx = ($1>=2.09) + ($1>=2.10);
+ $addx = ($1>=2.10);
+}
+
+if (!$addx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
+ `ml64 2>&1` =~ /Version ([0-9]+)\./) {
+ $avx = ($1>=10) + ($1>=11);
+ $addx = ($1>=12);
+}
+
+if (!$addx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9])\.([0-9]+)/) {
+ my $ver = $2 + $3/100.0; # 3.1->3.01, 3.10->3.10
+ $avx = ($ver>=3.0) + ($ver>=3.01);
+ $addx = ($ver>=3.03);
+}
+
+if ($avx>=2) {{
+$digit_size = "\$29";
+$n_digits = "\$9";
+
+$code.=<<___;
+.text
+
+.align 64
+.LAVX2_AND_MASK:
+.LAVX2_POLY:
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x000001ff, 0x000001ff, 0x000001ff, 0x000001ff
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x00040000, 0x00040000, 0x00040000, 0x00040000
+.quad 0x1fe00000, 0x1fe00000, 0x1fe00000, 0x1fe00000
+.quad 0x00ffffff, 0x00ffffff, 0x00ffffff, 0x00ffffff
+
+.LAVX2_POLY_x2:
+.quad 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC
+.quad 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC
+.quad 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC
+.quad 0x400007FC, 0x400007FC, 0x400007FC, 0x400007FC
+.quad 0x3FFFFFFE, 0x3FFFFFFE, 0x3FFFFFFE, 0x3FFFFFFE
+.quad 0x3FFFFFFE, 0x3FFFFFFE, 0x3FFFFFFE, 0x3FFFFFFE
+.quad 0x400FFFFE, 0x400FFFFE, 0x400FFFFE, 0x400FFFFE
+.quad 0x7F7FFFFE, 0x7F7FFFFE, 0x7F7FFFFE, 0x7F7FFFFE
+.quad 0x03FFFFFC, 0x03FFFFFC, 0x03FFFFFC, 0x03FFFFFC
+
+.LAVX2_POLY_x8:
+.quad 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8
+.quad 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8
+.quad 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8, 0xFFFFFFF8
+.quad 0x80000FF8, 0x80000FF8, 0x80000FF8, 0x80000FF8
+.quad 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC
+.quad 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC, 0x7FFFFFFC
+.quad 0x801FFFFC, 0x801FFFFC, 0x801FFFFC, 0x801FFFFC
+.quad 0xFEFFFFFC, 0xFEFFFFFC, 0xFEFFFFFC, 0xFEFFFFFC
+.quad 0x07FFFFF8, 0x07FFFFF8, 0x07FFFFF8, 0x07FFFFF8
+
+.LONE:
+.quad 0x00000020, 0x00000020, 0x00000020, 0x00000020
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x1fffc000, 0x1fffc000, 0x1fffc000, 0x1fffc000
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1f7fffff, 0x1f7fffff, 0x1f7fffff, 0x1f7fffff
+.quad 0x03ffffff, 0x03ffffff, 0x03ffffff, 0x03ffffff
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+
+# RR = 2^266 mod p in AVX2 format, to transform from the native OpenSSL
+# Montgomery form (*2^256) to our format (*2^261)
+
+.LTO_MONT_AVX2:
+.quad 0x00000400, 0x00000400, 0x00000400, 0x00000400
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x1ff80000, 0x1ff80000, 0x1ff80000, 0x1ff80000
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x0fffffff, 0x0fffffff, 0x0fffffff, 0x0fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x00000003, 0x00000003, 0x00000003, 0x00000003
+
+.LFROM_MONT_AVX2:
+.quad 0x00000001, 0x00000001, 0x00000001, 0x00000001
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+.quad 0x1ffffe00, 0x1ffffe00, 0x1ffffe00, 0x1ffffe00
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1fffffff, 0x1fffffff, 0x1fffffff, 0x1fffffff
+.quad 0x1ffbffff, 0x1ffbffff, 0x1ffbffff, 0x1ffbffff
+.quad 0x001fffff, 0x001fffff, 0x001fffff, 0x001fffff
+.quad 0x00000000, 0x00000000, 0x00000000, 0x00000000
+
+.LIntOne:
+.long 1,1,1,1,1,1,1,1
+___
+
+{
+# This function recieves a pointer to an array of four affine points
+# (X, Y, <1>) and rearanges the data for AVX2 execution, while
+# converting it to 2^29 radix redundant form
+
+my ($X0,$X1,$X2,$X3, $Y0,$Y1,$Y2,$Y3,
+ $T0,$T1,$T2,$T3, $T4,$T5,$T6,$T7)=map("%ymm$_",(0..15));
+
+$code.=<<___;
+.globl ecp_nistz256_avx2_transpose_convert
+.type ecp_nistz256_avx2_transpose_convert,\@function,2
+.align 64
+ecp_nistz256_avx2_transpose_convert:
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -8-16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ # Load the data
+ vmovdqa 32*0(%rsi), $X0
+ lea 112(%rsi), %rax # size optimization
+ vmovdqa 32*1(%rsi), $Y0
+ lea .LAVX2_AND_MASK(%rip), %rdx
+ vmovdqa 32*2(%rsi), $X1
+ vmovdqa 32*3(%rsi), $Y1
+ vmovdqa 32*4-112(%rax), $X2
+ vmovdqa 32*5-112(%rax), $Y2
+ vmovdqa 32*6-112(%rax), $X3
+ vmovdqa 32*7-112(%rax), $Y3
+
+ # Transpose X and Y independently
+ vpunpcklqdq $X1, $X0, $T0 # T0 = [B2 A2 B0 A0]
+ vpunpcklqdq $X3, $X2, $T1 # T1 = [D2 C2 D0 C0]
+ vpunpckhqdq $X1, $X0, $T2 # T2 = [B3 A3 B1 A1]
+ vpunpckhqdq $X3, $X2, $T3 # T3 = [D3 C3 D1 C1]
+
+ vpunpcklqdq $Y1, $Y0, $T4
+ vpunpcklqdq $Y3, $Y2, $T5
+ vpunpckhqdq $Y1, $Y0, $T6
+ vpunpckhqdq $Y3, $Y2, $T7
+
+ vperm2i128 \$0x20, $T1, $T0, $X0 # X0 = [D0 C0 B0 A0]
+ vperm2i128 \$0x20, $T3, $T2, $X1 # X1 = [D1 C1 B1 A1]
+ vperm2i128 \$0x31, $T1, $T0, $X2 # X2 = [D2 C2 B2 A2]
+ vperm2i128 \$0x31, $T3, $T2, $X3 # X3 = [D3 C3 B3 A3]
+
+ vperm2i128 \$0x20, $T5, $T4, $Y0
+ vperm2i128 \$0x20, $T7, $T6, $Y1
+ vperm2i128 \$0x31, $T5, $T4, $Y2
+ vperm2i128 \$0x31, $T7, $T6, $Y3
+ vmovdqa (%rdx), $T7
+
+ vpand (%rdx), $X0, $T0 # out[0] = in[0] & mask;
+ vpsrlq \$29, $X0, $X0
+ vpand $T7, $X0, $T1 # out[1] = (in[0] >> shift) & mask;
+ vpsrlq \$29, $X0, $X0
+ vpsllq \$6, $X1, $T2
+ vpxor $X0, $T2, $T2
+ vpand $T7, $T2, $T2 # out[2] = ((in[0] >> (shift*2)) ^ (in[1] << (64-shift*2))) & mask;
+ vpsrlq \$23, $X1, $X1
+ vpand $T7, $X1, $T3 # out[3] = (in[1] >> ((shift*3)%64)) & mask;
+ vpsrlq \$29, $X1, $X1
+ vpsllq \$12, $X2, $T4
+ vpxor $X1, $T4, $T4
+ vpand $T7, $T4, $T4 # out[4] = ((in[1] >> ((shift*4)%64)) ^ (in[2] << (64*2-shift*4))) & mask;
+ vpsrlq \$17, $X2, $X2
+ vpand $T7, $X2, $T5 # out[5] = (in[2] >> ((shift*5)%64)) & mask;
+ vpsrlq \$29, $X2, $X2
+ vpsllq \$18, $X3, $T6
+ vpxor $X2, $T6, $T6
+ vpand $T7, $T6, $T6 # out[6] = ((in[2] >> ((shift*6)%64)) ^ (in[3] << (64*3-shift*6))) & mask;
+ vpsrlq \$11, $X3, $X3
+ vmovdqa $T0, 32*0(%rdi)
+ lea 112(%rdi), %rax # size optimization
+ vpand $T7, $X3, $T0 # out[7] = (in[3] >> ((shift*7)%64)) & mask;
+ vpsrlq \$29, $X3, $X3 # out[8] = (in[3] >> ((shift*8)%64)) & mask;
+
+ vmovdqa $T1, 32*1(%rdi)
+ vmovdqa $T2, 32*2(%rdi)
+ vmovdqa $T3, 32*3(%rdi)
+ vmovdqa $T4, 32*4-112(%rax)
+ vmovdqa $T5, 32*5-112(%rax)
+ vmovdqa $T6, 32*6-112(%rax)
+ vmovdqa $T0, 32*7-112(%rax)
+ vmovdqa $X3, 32*8-112(%rax)
+ lea 448(%rdi), %rax # size optimization
+
+ vpand $T7, $Y0, $T0 # out[0] = in[0] & mask;
+ vpsrlq \$29, $Y0, $Y0
+ vpand $T7, $Y0, $T1 # out[1] = (in[0] >> shift) & mask;
+ vpsrlq \$29, $Y0, $Y0
+ vpsllq \$6, $Y1, $T2
+ vpxor $Y0, $T2, $T2
+ vpand $T7, $T2, $T2 # out[2] = ((in[0] >> (shift*2)) ^ (in[1] << (64-shift*2))) & mask;
+ vpsrlq \$23, $Y1, $Y1
+ vpand $T7, $Y1, $T3 # out[3] = (in[1] >> ((shift*3)%64)) & mask;
+ vpsrlq \$29, $Y1, $Y1
+ vpsllq \$12, $Y2, $T4
+ vpxor $Y1, $T4, $T4
+ vpand $T7, $T4, $T4 # out[4] = ((in[1] >> ((shift*4)%64)) ^ (in[2] << (64*2-shift*4))) & mask;
+ vpsrlq \$17, $Y2, $Y2
+ vpand $T7, $Y2, $T5 # out[5] = (in[2] >> ((shift*5)%64)) & mask;
+ vpsrlq \$29, $Y2, $Y2
+ vpsllq \$18, $Y3, $T6
+ vpxor $Y2, $T6, $T6
+ vpand $T7, $T6, $T6 # out[6] = ((in[2] >> ((shift*6)%64)) ^ (in[3] << (64*3-shift*6))) & mask;
+ vpsrlq \$11, $Y3, $Y3
+ vmovdqa $T0, 32*9-448(%rax)
+ vpand $T7, $Y3, $T0 # out[7] = (in[3] >> ((shift*7)%64)) & mask;
+ vpsrlq \$29, $Y3, $Y3 # out[8] = (in[3] >> ((shift*8)%64)) & mask;
+
+ vmovdqa $T1, 32*10-448(%rax)
+ vmovdqa $T2, 32*11-448(%rax)
+ vmovdqa $T3, 32*12-448(%rax)
+ vmovdqa $T4, 32*13-448(%rax)
+ vmovdqa $T5, 32*14-448(%rax)
+ vmovdqa $T6, 32*15-448(%rax)
+ vmovdqa $T0, 32*16-448(%rax)
+ vmovdqa $Y3, 32*17-448(%rax)
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps 16*0(%rsp), %xmm6
+ movaps 16*1(%rsp), %xmm7
+ movaps 16*2(%rsp), %xmm8
+ movaps 16*3(%rsp), %xmm9
+ movaps 16*4(%rsp), %xmm10
+ movaps 16*5(%rsp), %xmm11
+ movaps 16*6(%rsp), %xmm12
+ movaps 16*7(%rsp), %xmm13
+ movaps 16*8(%rsp), %xmm14
+ movaps 16*9(%rsp), %xmm15
+ lea 8+16*10(%rsp), %rsp
+___
+$code.=<<___;
+ ret
+.size ecp_nistz256_avx2_transpose_convert,.-ecp_nistz256_avx2_transpose_convert
+___
+}
+{
+################################################################################
+# This function recieves a pointer to an array of four AVX2 formatted points
+# (X, Y, Z) convert the data to normal representation, and rearanges the data
+
+my ($D0,$D1,$D2,$D3, $D4,$D5,$D6,$D7, $D8)=map("%ymm$_",(0..8));
+my ($T0,$T1,$T2,$T3, $T4,$T5,$T6)=map("%ymm$_",(9..15));
+
+$code.=<<___;
+
+.globl ecp_nistz256_avx2_convert_transpose_back
+.type ecp_nistz256_avx2_convert_transpose_back,\@function,2
+.align 32
+ecp_nistz256_avx2_convert_transpose_back:
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -8-16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ mov \$3, %ecx
+
+.Lconv_loop:
+ vmovdqa 32*0(%rsi), $D0
+ lea 160(%rsi), %rax # size optimization
+ vmovdqa 32*1(%rsi), $D1
+ vmovdqa 32*2(%rsi), $D2
+ vmovdqa 32*3(%rsi), $D3
+ vmovdqa 32*4-160(%rax), $D4
+ vmovdqa 32*5-160(%rax), $D5
+ vmovdqa 32*6-160(%rax), $D6
+ vmovdqa 32*7-160(%rax), $D7
+ vmovdqa 32*8-160(%rax), $D8
+
+ vpsllq \$29, $D1, $D1
+ vpsllq \$58, $D2, $T0
+ vpaddq $D1, $D0, $D0
+ vpaddq $T0, $D0, $D0 # out[0] = (in[0]) ^ (in[1] << shift*1) ^ (in[2] << shift*2);
+
+ vpsrlq \$6, $D2, $D2
+ vpsllq \$23, $D3, $D3
+ vpsllq \$52, $D4, $T1
+ vpaddq $D2, $D3, $D3
+ vpaddq $D3, $T1, $D1 # out[1] = (in[2] >> (64*1-shift*2)) ^ (in[3] << shift*3%64) ^ (in[4] << shift*4%64);
+
+ vpsrlq \$12, $D4, $D4
+ vpsllq \$17, $D5, $D5
+ vpsllq \$46, $D6, $T2
+ vpaddq $D4, $D5, $D5
+ vpaddq $D5, $T2, $D2 # out[2] = (in[4] >> (64*2-shift*4)) ^ (in[5] << shift*5%64) ^ (in[6] << shift*6%64);
+
+ vpsrlq \$18, $D6, $D6
+ vpsllq \$11, $D7, $D7
+ vpsllq \$40, $D8, $T3
+ vpaddq $D6, $D7, $D7
+ vpaddq $D7, $T3, $D3 # out[3] = (in[6] >> (64*3-shift*6)) ^ (in[7] << shift*7%64) ^ (in[8] << shift*8%64);
+
+ vpunpcklqdq $D1, $D0, $T0 # T0 = [B2 A2 B0 A0]
+ vpunpcklqdq $D3, $D2, $T1 # T1 = [D2 C2 D0 C0]
+ vpunpckhqdq $D1, $D0, $T2 # T2 = [B3 A3 B1 A1]
+ vpunpckhqdq $D3, $D2, $T3 # T3 = [D3 C3 D1 C1]
+
+ vperm2i128 \$0x20, $T1, $T0, $D0 # X0 = [D0 C0 B0 A0]
+ vperm2i128 \$0x20, $T3, $T2, $D1 # X1 = [D1 C1 B1 A1]
+ vperm2i128 \$0x31, $T1, $T0, $D2 # X2 = [D2 C2 B2 A2]
+ vperm2i128 \$0x31, $T3, $T2, $D3 # X3 = [D3 C3 B3 A3]
+
+ vmovdqa $D0, 32*0(%rdi)
+ vmovdqa $D1, 32*3(%rdi)
+ vmovdqa $D2, 32*6(%rdi)
+ vmovdqa $D3, 32*9(%rdi)
+
+ lea 32*9(%rsi), %rsi
+ lea 32*1(%rdi), %rdi
+
+ dec %ecx
+ jnz .Lconv_loop
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps 16*0(%rsp), %xmm6
+ movaps 16*1(%rsp), %xmm7
+ movaps 16*2(%rsp), %xmm8
+ movaps 16*3(%rsp), %xmm9
+ movaps 16*4(%rsp), %xmm10
+ movaps 16*5(%rsp), %xmm11
+ movaps 16*6(%rsp), %xmm12
+ movaps 16*7(%rsp), %xmm13
+ movaps 16*8(%rsp), %xmm14
+ movaps 16*9(%rsp), %xmm15
+ lea 8+16*10(%rsp), %rsp
+___
+$code.=<<___;
+ ret
+.size ecp_nistz256_avx2_convert_transpose_back,.-ecp_nistz256_avx2_convert_transpose_back
+___
+}
+{
+my ($r_ptr,$a_ptr,$b_ptr,$itr)=("%rdi","%rsi","%rdx","%ecx");
+my ($ACC0,$ACC1,$ACC2,$ACC3,$ACC4,$ACC5,$ACC6,$ACC7,$ACC8)=map("%ymm$_",(0..8));
+my ($B,$Y,$T0,$AND_MASK,$OVERFLOW)=map("%ymm$_",(9..13));
+
+sub NORMALIZE {
+my $ret=<<___;
+ vpsrlq $digit_size, $ACC0, $T0
+ vpand $AND_MASK, $ACC0, $ACC0
+ vpaddq $T0, $ACC1, $ACC1
+
+ vpsrlq $digit_size, $ACC1, $T0
+ vpand $AND_MASK, $ACC1, $ACC1
+ vpaddq $T0, $ACC2, $ACC2
+
+ vpsrlq $digit_size, $ACC2, $T0
+ vpand $AND_MASK, $ACC2, $ACC2
+ vpaddq $T0, $ACC3, $ACC3
+
+ vpsrlq $digit_size, $ACC3, $T0
+ vpand $AND_MASK, $ACC3, $ACC3
+ vpaddq $T0, $ACC4, $ACC4
+
+ vpsrlq $digit_size, $ACC4, $T0
+ vpand $AND_MASK, $ACC4, $ACC4
+ vpaddq $T0, $ACC5, $ACC5
+
+ vpsrlq $digit_size, $ACC5, $T0
+ vpand $AND_MASK, $ACC5, $ACC5
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpsrlq $digit_size, $ACC6, $T0
+ vpand $AND_MASK, $ACC6, $ACC6
+ vpaddq $T0, $ACC7, $ACC7
+
+ vpsrlq $digit_size, $ACC7, $T0
+ vpand $AND_MASK, $ACC7, $ACC7
+ vpaddq $T0, $ACC8, $ACC8
+ #vpand $AND_MASK, $ACC8, $ACC8
+___
+ $ret;
+}
+
+sub STORE {
+my $ret=<<___;
+ vmovdqa $ACC0, 32*0(%rdi)
+ lea 160(%rdi), %rax # size optimization
+ vmovdqa $ACC1, 32*1(%rdi)
+ vmovdqa $ACC2, 32*2(%rdi)
+ vmovdqa $ACC3, 32*3(%rdi)
+ vmovdqa $ACC4, 32*4-160(%rax)
+ vmovdqa $ACC5, 32*5-160(%rax)
+ vmovdqa $ACC6, 32*6-160(%rax)
+ vmovdqa $ACC7, 32*7-160(%rax)
+ vmovdqa $ACC8, 32*8-160(%rax)
+___
+ $ret;
+}
+
+$code.=<<___;
+.type avx2_normalize,\@abi-omnipotent
+.align 32
+avx2_normalize:
+ vpsrlq $digit_size, $ACC0, $T0
+ vpand $AND_MASK, $ACC0, $ACC0
+ vpaddq $T0, $ACC1, $ACC1
+
+ vpsrlq $digit_size, $ACC1, $T0
+ vpand $AND_MASK, $ACC1, $ACC1
+ vpaddq $T0, $ACC2, $ACC2
+
+ vpsrlq $digit_size, $ACC2, $T0
+ vpand $AND_MASK, $ACC2, $ACC2
+ vpaddq $T0, $ACC3, $ACC3
+
+ vpsrlq $digit_size, $ACC3, $T0
+ vpand $AND_MASK, $ACC3, $ACC3
+ vpaddq $T0, $ACC4, $ACC4
+
+ vpsrlq $digit_size, $ACC4, $T0
+ vpand $AND_MASK, $ACC4, $ACC4
+ vpaddq $T0, $ACC5, $ACC5
+
+ vpsrlq $digit_size, $ACC5, $T0
+ vpand $AND_MASK, $ACC5, $ACC5
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpsrlq $digit_size, $ACC6, $T0
+ vpand $AND_MASK, $ACC6, $ACC6
+ vpaddq $T0, $ACC7, $ACC7
+
+ vpsrlq $digit_size, $ACC7, $T0
+ vpand $AND_MASK, $ACC7, $ACC7
+ vpaddq $T0, $ACC8, $ACC8
+ #vpand $AND_MASK, $ACC8, $ACC8
+
+ ret
+.size avx2_normalize,.-avx2_normalize
+
+.type avx2_normalize_n_store,\@abi-omnipotent
+.align 32
+avx2_normalize_n_store:
+ vpsrlq $digit_size, $ACC0, $T0
+ vpand $AND_MASK, $ACC0, $ACC0
+ vpaddq $T0, $ACC1, $ACC1
+
+ vpsrlq $digit_size, $ACC1, $T0
+ vpand $AND_MASK, $ACC1, $ACC1
+ vmovdqa $ACC0, 32*0(%rdi)
+ lea 160(%rdi), %rax # size optimization
+ vpaddq $T0, $ACC2, $ACC2
+
+ vpsrlq $digit_size, $ACC2, $T0
+ vpand $AND_MASK, $ACC2, $ACC2
+ vmovdqa $ACC1, 32*1(%rdi)
+ vpaddq $T0, $ACC3, $ACC3
+
+ vpsrlq $digit_size, $ACC3, $T0
+ vpand $AND_MASK, $ACC3, $ACC3
+ vmovdqa $ACC2, 32*2(%rdi)
+ vpaddq $T0, $ACC4, $ACC4
+
+ vpsrlq $digit_size, $ACC4, $T0
+ vpand $AND_MASK, $ACC4, $ACC4
+ vmovdqa $ACC3, 32*3(%rdi)
+ vpaddq $T0, $ACC5, $ACC5
+
+ vpsrlq $digit_size, $ACC5, $T0
+ vpand $AND_MASK, $ACC5, $ACC5
+ vmovdqa $ACC4, 32*4-160(%rax)
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpsrlq $digit_size, $ACC6, $T0
+ vpand $AND_MASK, $ACC6, $ACC6
+ vmovdqa $ACC5, 32*5-160(%rax)
+ vpaddq $T0, $ACC7, $ACC7
+
+ vpsrlq $digit_size, $ACC7, $T0
+ vpand $AND_MASK, $ACC7, $ACC7
+ vmovdqa $ACC6, 32*6-160(%rax)
+ vpaddq $T0, $ACC8, $ACC8
+ #vpand $AND_MASK, $ACC8, $ACC8
+ vmovdqa $ACC7, 32*7-160(%rax)
+ vmovdqa $ACC8, 32*8-160(%rax)
+
+ ret
+.size avx2_normalize_n_store,.-avx2_normalize_n_store
+
+################################################################################
+# void avx2_mul_x4(void* RESULTx4, void *Ax4, void *Bx4);
+.type avx2_mul_x4,\@abi-omnipotent
+.align 32
+avx2_mul_x4:
+ lea .LAVX2_POLY(%rip), %rax
+
+ vpxor $ACC0, $ACC0, $ACC0
+ vpxor $ACC1, $ACC1, $ACC1
+ vpxor $ACC2, $ACC2, $ACC2
+ vpxor $ACC3, $ACC3, $ACC3
+ vpxor $ACC4, $ACC4, $ACC4
+ vpxor $ACC5, $ACC5, $ACC5
+ vpxor $ACC6, $ACC6, $ACC6
+ vpxor $ACC7, $ACC7, $ACC7
+
+ vmovdqa 32*7(%rax), %ymm14
+ vmovdqa 32*8(%rax), %ymm15
+
+ mov $n_digits, $itr
+ lea -512($a_ptr), $a_ptr # strategic bias to control u-op density
+ jmp .Lavx2_mul_x4_loop
+
+.align 32
+.Lavx2_mul_x4_loop:
+ vmovdqa 32*0($b_ptr), $B
+ lea 32*1($b_ptr), $b_ptr
+
+ vpmuludq 32*0+512($a_ptr), $B, $T0
+ vpmuludq 32*1+512($a_ptr), $B, $OVERFLOW # borrow $OVERFLOW
+ vpaddq $T0, $ACC0, $ACC0
+ vpmuludq 32*2+512($a_ptr), $B, $T0
+ vpaddq $OVERFLOW, $ACC1, $ACC1
+ vpand $AND_MASK, $ACC0, $Y
+ vpmuludq 32*3+512($a_ptr), $B, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC2
+ vpmuludq 32*4+512($a_ptr), $B, $T0
+ vpaddq $OVERFLOW, $ACC3, $ACC3
+ vpmuludq 32*5+512($a_ptr), $B, $OVERFLOW
+ vpaddq $T0, $ACC4, $ACC4
+ vpmuludq 32*6+512($a_ptr), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*7+512($a_ptr), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ # Skip some multiplications, optimizing for the constant poly
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*8+512($a_ptr), $B, $ACC8
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ .byte 0x67
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $OVERFLOW
+ .byte 0x67
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $T0
+ vpaddq $OVERFLOW, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $OVERFLOW
+ vpaddq $T0, $ACC7, $ACC6
+ vpaddq $OVERFLOW, $ACC8, $ACC7
+
+ dec $itr
+ jnz .Lavx2_mul_x4_loop
+
+ vpxor $ACC8, $ACC8, $ACC8
+
+ ret
+.size avx2_mul_x4,.-avx2_mul_x4
+
+# Function optimized for the constant 1
+################################################################################
+# void avx2_mul_by1_x4(void* RESULTx4, void *Ax4);
+.type avx2_mul_by1_x4,\@abi-omnipotent
+.align 32
+avx2_mul_by1_x4:
+ lea .LAVX2_POLY(%rip), %rax
+
+ vpxor $ACC0, $ACC0, $ACC0
+ vpxor $ACC1, $ACC1, $ACC1
+ vpxor $ACC2, $ACC2, $ACC2
+ vpxor $ACC3, $ACC3, $ACC3
+ vpxor $ACC4, $ACC4, $ACC4
+ vpxor $ACC5, $ACC5, $ACC5
+ vpxor $ACC6, $ACC6, $ACC6
+ vpxor $ACC7, $ACC7, $ACC7
+ vpxor $ACC8, $ACC8, $ACC8
+
+ vmovdqa 32*3+.LONE(%rip), %ymm14
+ vmovdqa 32*7+.LONE(%rip), %ymm15
+
+ mov $n_digits, $itr
+ jmp .Lavx2_mul_by1_x4_loop
+
+.align 32
+.Lavx2_mul_by1_x4_loop:
+ vmovdqa 32*0($a_ptr), $B
+ .byte 0x48,0x8d,0xb6,0x20,0,0,0 # lea 32*1($a_ptr), $a_ptr
+
+ vpsllq \$5, $B, $OVERFLOW
+ vpmuludq %ymm14, $B, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC3
+ .byte 0x67
+ vpmuludq $AND_MASK, $B, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $T0, $ACC4, $ACC4
+ vpaddq $T0, $ACC5, $ACC5
+ vpaddq $T0, $ACC6, $ACC6
+ vpsllq \$23, $B, $T0
+
+ .byte 0x67,0x67
+ vpmuludq %ymm15, $B, $OVERFLOW
+ vpsubq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ .byte 0x67,0x67
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $OVERFLOW
+ vmovdqa $ACC5, $ACC4
+ vpmuludq 32*7(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC6, $ACC5
+ vpaddq $T0, $ACC7, $ACC6
+ vpmuludq 32*8(%rax), $Y, $ACC7
+
+ dec $itr
+ jnz .Lavx2_mul_by1_x4_loop
+
+ ret
+.size avx2_mul_by1_x4,.-avx2_mul_by1_x4
+
+################################################################################
+# void avx2_sqr_x4(void* RESULTx4, void *Ax4, void *Bx4);
+.type avx2_sqr_x4,\@abi-omnipotent
+.align 32
+avx2_sqr_x4:
+ lea .LAVX2_POLY(%rip), %rax
+
+ vmovdqa 32*7(%rax), %ymm14
+ vmovdqa 32*8(%rax), %ymm15
+
+ vmovdqa 32*0($a_ptr), $B
+ vmovdqa 32*1($a_ptr), $ACC1
+ vmovdqa 32*2($a_ptr), $ACC2
+ vmovdqa 32*3($a_ptr), $ACC3
+ vmovdqa 32*4($a_ptr), $ACC4
+ vmovdqa 32*5($a_ptr), $ACC5
+ vmovdqa 32*6($a_ptr), $ACC6
+ vmovdqa 32*7($a_ptr), $ACC7
+ vpaddq $ACC1, $ACC1, $ACC1 # 2*$ACC0..7
+ vmovdqa 32*8($a_ptr), $ACC8
+ vpaddq $ACC2, $ACC2, $ACC2
+ vmovdqa $ACC1, 32*0(%rcx)
+ vpaddq $ACC3, $ACC3, $ACC3
+ vmovdqa $ACC2, 32*1(%rcx)
+ vpaddq $ACC4, $ACC4, $ACC4
+ vmovdqa $ACC3, 32*2(%rcx)
+ vpaddq $ACC5, $ACC5, $ACC5
+ vmovdqa $ACC4, 32*3(%rcx)
+ vpaddq $ACC6, $ACC6, $ACC6
+ vmovdqa $ACC5, 32*4(%rcx)
+ vpaddq $ACC7, $ACC7, $ACC7
+ vmovdqa $ACC6, 32*5(%rcx)
+ vpaddq $ACC8, $ACC8, $ACC8
+ vmovdqa $ACC7, 32*6(%rcx)
+ vmovdqa $ACC8, 32*7(%rcx)
+
+ #itr 1
+ vpmuludq $B, $B, $ACC0
+ vpmuludq $B, $ACC1, $ACC1
+ vpand $AND_MASK, $ACC0, $Y
+ vpmuludq $B, $ACC2, $ACC2
+ vpmuludq $B, $ACC3, $ACC3
+ vpmuludq $B, $ACC4, $ACC4
+ vpmuludq $B, $ACC5, $ACC5
+ vpmuludq $B, $ACC6, $ACC6
+ vpmuludq $AND_MASK, $Y, $T0
+ vpmuludq $B, $ACC7, $ACC7
+ vpmuludq $B, $ACC8, $ACC8
+ vmovdqa 32*1($a_ptr), $B
+
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 2
+ vpmuludq $B, $B, $OVERFLOW
+ vpand $AND_MASK, $ACC0, $Y
+ vpmuludq 32*1(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC1, $ACC1
+ vpmuludq 32*2(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC2
+ vpmuludq 32*3(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC3, $ACC3
+ vpmuludq 32*4(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC4, $ACC4
+ vpmuludq 32*5(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*2($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 3
+ vpmuludq $B, $B, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpmuludq 32*2(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC2
+ vpmuludq 32*3(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC3, $ACC3
+ vpmuludq 32*4(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC4, $ACC4
+ vpmuludq 32*5(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*3($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 4
+ vpmuludq $B, $B, $OVERFLOW
+ vpmuludq 32*3(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC3, $ACC3
+ vpmuludq 32*4(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC4, $ACC4
+ vpmuludq 32*5(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*4($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 5
+ vpmuludq $B, $B, $T0
+ vpmuludq 32*4(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC4, $ACC4
+ vpmuludq 32*5(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*5($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3+.LAVX2_POLY(%rip), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 6
+ vpmuludq $B, $B, $OVERFLOW
+ vpmuludq 32*5(%rcx), $B, $T0
+ vpaddq $OVERFLOW, $ACC5, $ACC5
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*6($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 7
+ vpmuludq $B, $B, $T0
+ vpmuludq 32*6(%rcx), $B, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC6
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*7($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 8
+ vpmuludq $B, $B, $OVERFLOW
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC7
+ vpmuludq 32*7(%rcx), $B, $ACC8
+ vmovdqa 32*8($a_ptr), $B
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpand $AND_MASK, $ACC0, $Y
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ #itr 9
+ vpmuludq $B, $B, $ACC8
+
+ vpmuludq $AND_MASK, $Y, $T0
+ vpaddq $T0, $ACC0, $OVERFLOW
+ vpsrlq $digit_size, $OVERFLOW, $OVERFLOW
+ vpaddq $T0, $ACC1, $ACC0
+ vpaddq $T0, $ACC2, $ACC1
+ vpmuludq 32*3(%rax), $Y, $T0
+ vpaddq $OVERFLOW, $ACC0, $ACC0
+ vpaddq $T0, $ACC3, $ACC2
+ vmovdqa $ACC4, $ACC3
+ vpsllq \$18, $Y, $T0
+ vmovdqa $ACC5, $ACC4
+ vpmuludq %ymm14, $Y, $OVERFLOW
+ vpaddq $T0, $ACC6, $ACC5
+ vpmuludq %ymm15, $Y, $T0
+ vpaddq $OVERFLOW, $ACC7, $ACC6
+ vpaddq $T0, $ACC8, $ACC7
+
+ vpxor $ACC8, $ACC8, $ACC8
+
+ ret
+.size avx2_sqr_x4,.-avx2_sqr_x4
+
+################################################################################
+# void avx2_sub_x4(void* RESULTx4, void *Ax4, void *Bx4);
+.type avx2_sub_x4,\@abi-omnipotent
+.align 32
+avx2_sub_x4:
+ vmovdqa 32*0($a_ptr), $ACC0
+ lea 160($a_ptr), $a_ptr
+ lea .LAVX2_POLY_x8+128(%rip), %rax
+ lea 128($b_ptr), $b_ptr
+ vmovdqa 32*1-160($a_ptr), $ACC1
+ vmovdqa 32*2-160($a_ptr), $ACC2
+ vmovdqa 32*3-160($a_ptr), $ACC3
+ vmovdqa 32*4-160($a_ptr), $ACC4
+ vmovdqa 32*5-160($a_ptr), $ACC5
+ vmovdqa 32*6-160($a_ptr), $ACC6
+ vmovdqa 32*7-160($a_ptr), $ACC7
+ vmovdqa 32*8-160($a_ptr), $ACC8
+
+ vpaddq 32*0-128(%rax), $ACC0, $ACC0
+ vpaddq 32*1-128(%rax), $ACC1, $ACC1
+ vpaddq 32*2-128(%rax), $ACC2, $ACC2
+ vpaddq 32*3-128(%rax), $ACC3, $ACC3
+ vpaddq 32*4-128(%rax), $ACC4, $ACC4
+ vpaddq 32*5-128(%rax), $ACC5, $ACC5
+ vpaddq 32*6-128(%rax), $ACC6, $ACC6
+ vpaddq 32*7-128(%rax), $ACC7, $ACC7
+ vpaddq 32*8-128(%rax), $ACC8, $ACC8
+
+ vpsubq 32*0-128($b_ptr), $ACC0, $ACC0
+ vpsubq 32*1-128($b_ptr), $ACC1, $ACC1
+ vpsubq 32*2-128($b_ptr), $ACC2, $ACC2
+ vpsubq 32*3-128($b_ptr), $ACC3, $ACC3
+ vpsubq 32*4-128($b_ptr), $ACC4, $ACC4
+ vpsubq 32*5-128($b_ptr), $ACC5, $ACC5
+ vpsubq 32*6-128($b_ptr), $ACC6, $ACC6
+ vpsubq 32*7-128($b_ptr), $ACC7, $ACC7
+ vpsubq 32*8-128($b_ptr), $ACC8, $ACC8
+
+ ret
+.size avx2_sub_x4,.-avx2_sub_x4
+
+.type avx2_select_n_store,\@abi-omnipotent
+.align 32
+avx2_select_n_store:
+ vmovdqa `8+32*9*8`(%rsp), $Y
+ vpor `8+32*9*8+32`(%rsp), $Y, $Y
+
+ vpandn $ACC0, $Y, $ACC0
+ vpandn $ACC1, $Y, $ACC1
+ vpandn $ACC2, $Y, $ACC2
+ vpandn $ACC3, $Y, $ACC3
+ vpandn $ACC4, $Y, $ACC4
+ vpandn $ACC5, $Y, $ACC5
+ vpandn $ACC6, $Y, $ACC6
+ vmovdqa `8+32*9*8+32`(%rsp), $B
+ vpandn $ACC7, $Y, $ACC7
+ vpandn `8+32*9*8`(%rsp), $B, $B
+ vpandn $ACC8, $Y, $ACC8
+
+ vpand 32*0(%rsi), $B, $T0
+ lea 160(%rsi), %rax
+ vpand 32*1(%rsi), $B, $Y
+ vpxor $T0, $ACC0, $ACC0
+ vpand 32*2(%rsi), $B, $T0
+ vpxor $Y, $ACC1, $ACC1
+ vpand 32*3(%rsi), $B, $Y
+ vpxor $T0, $ACC2, $ACC2
+ vpand 32*4-160(%rax), $B, $T0
+ vpxor $Y, $ACC3, $ACC3
+ vpand 32*5-160(%rax), $B, $Y
+ vpxor $T0, $ACC4, $ACC4
+ vpand 32*6-160(%rax), $B, $T0
+ vpxor $Y, $ACC5, $ACC5
+ vpand 32*7-160(%rax), $B, $Y
+ vpxor $T0, $ACC6, $ACC6
+ vpand 32*8-160(%rax), $B, $T0
+ vmovdqa `8+32*9*8+32`(%rsp), $B
+ vpxor $Y, $ACC7, $ACC7
+
+ vpand 32*0(%rdx), $B, $Y
+ lea 160(%rdx), %rax
+ vpxor $T0, $ACC8, $ACC8
+ vpand 32*1(%rdx), $B, $T0
+ vpxor $Y, $ACC0, $ACC0
+ vpand 32*2(%rdx), $B, $Y
+ vpxor $T0, $ACC1, $ACC1
+ vpand 32*3(%rdx), $B, $T0
+ vpxor $Y, $ACC2, $ACC2
+ vpand 32*4-160(%rax), $B, $Y
+ vpxor $T0, $ACC3, $ACC3
+ vpand 32*5-160(%rax), $B, $T0
+ vpxor $Y, $ACC4, $ACC4
+ vpand 32*6-160(%rax), $B, $Y
+ vpxor $T0, $ACC5, $ACC5
+ vpand 32*7-160(%rax), $B, $T0
+ vpxor $Y, $ACC6, $ACC6
+ vpand 32*8-160(%rax), $B, $Y
+ vpxor $T0, $ACC7, $ACC7
+ vpxor $Y, $ACC8, $ACC8
+ `&STORE`
+
+ ret
+.size avx2_select_n_store,.-avx2_select_n_store
+___
+$code.=<<___ if (0); # inlined
+################################################################################
+# void avx2_mul_by2_x4(void* RESULTx4, void *Ax4);
+.type avx2_mul_by2_x4,\@abi-omnipotent
+.align 32
+avx2_mul_by2_x4:
+ vmovdqa 32*0($a_ptr), $ACC0
+ lea 160($a_ptr), %rax
+ vmovdqa 32*1($a_ptr), $ACC1
+ vmovdqa 32*2($a_ptr), $ACC2
+ vmovdqa 32*3($a_ptr), $ACC3
+ vmovdqa 32*4-160(%rax), $ACC4
+ vmovdqa 32*5-160(%rax), $ACC5
+ vmovdqa 32*6-160(%rax), $ACC6
+ vmovdqa 32*7-160(%rax), $ACC7
+ vmovdqa 32*8-160(%rax), $ACC8
+
+ vpaddq $ACC0, $ACC0, $ACC0
+ vpaddq $ACC1, $ACC1, $ACC1
+ vpaddq $ACC2, $ACC2, $ACC2
+ vpaddq $ACC3, $ACC3, $ACC3
+ vpaddq $ACC4, $ACC4, $ACC4
+ vpaddq $ACC5, $ACC5, $ACC5
+ vpaddq $ACC6, $ACC6, $ACC6
+ vpaddq $ACC7, $ACC7, $ACC7
+ vpaddq $ACC8, $ACC8, $ACC8
+
+ ret
+.size avx2_mul_by2_x4,.-avx2_mul_by2_x4
+___
+my ($r_ptr_in,$a_ptr_in,$b_ptr_in)=("%rdi","%rsi","%rdx");
+my ($r_ptr,$a_ptr,$b_ptr)=("%r8","%r9","%r10");
+
+$code.=<<___;
+################################################################################
+# void ecp_nistz256_avx2_point_add_affine_x4(void* RESULTx4, void *Ax4, void *Bx4);
+.globl ecp_nistz256_avx2_point_add_affine_x4
+.type ecp_nistz256_avx2_point_add_affine_x4,\@function,3
+.align 32
+ecp_nistz256_avx2_point_add_affine_x4:
+ mov %rsp, %rax
+ push %rbp
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ lea -8(%rax), %rbp
+
+# Result + 32*0 = Result.X
+# Result + 32*9 = Result.Y
+# Result + 32*18 = Result.Z
+
+# A + 32*0 = A.X
+# A + 32*9 = A.Y
+# A + 32*18 = A.Z
+
+# B + 32*0 = B.X
+# B + 32*9 = B.Y
+
+ sub \$`32*9*8+32*2+32*8`, %rsp
+ and \$-64, %rsp
+
+ mov $r_ptr_in, $r_ptr
+ mov $a_ptr_in, $a_ptr
+ mov $b_ptr_in, $b_ptr
+
+ vmovdqa 32*0($a_ptr_in), %ymm0
+ vmovdqa .LAVX2_AND_MASK(%rip), $AND_MASK
+ vpxor %ymm1, %ymm1, %ymm1
+ lea 256($a_ptr_in), %rax # size optimization
+ vpor 32*1($a_ptr_in), %ymm0, %ymm0
+ vpor 32*2($a_ptr_in), %ymm0, %ymm0
+ vpor 32*3($a_ptr_in), %ymm0, %ymm0
+ vpor 32*4-256(%rax), %ymm0, %ymm0
+ lea 256(%rax), %rcx # size optimization
+ vpor 32*5-256(%rax), %ymm0, %ymm0
+ vpor 32*6-256(%rax), %ymm0, %ymm0
+ vpor 32*7-256(%rax), %ymm0, %ymm0
+ vpor 32*8-256(%rax), %ymm0, %ymm0
+ vpor 32*9-256(%rax), %ymm0, %ymm0
+ vpor 32*10-256(%rax), %ymm0, %ymm0
+ vpor 32*11-256(%rax), %ymm0, %ymm0
+ vpor 32*12-512(%rcx), %ymm0, %ymm0
+ vpor 32*13-512(%rcx), %ymm0, %ymm0
+ vpor 32*14-512(%rcx), %ymm0, %ymm0
+ vpor 32*15-512(%rcx), %ymm0, %ymm0
+ vpor 32*16-512(%rcx), %ymm0, %ymm0
+ vpor 32*17-512(%rcx), %ymm0, %ymm0
+ vpcmpeqq %ymm1, %ymm0, %ymm0
+ vmovdqa %ymm0, `32*9*8`(%rsp)
+
+ vpxor %ymm1, %ymm1, %ymm1
+ vmovdqa 32*0($b_ptr), %ymm0
+ lea 256($b_ptr), %rax # size optimization
+ vpor 32*1($b_ptr), %ymm0, %ymm0
+ vpor 32*2($b_ptr), %ymm0, %ymm0
+ vpor 32*3($b_ptr), %ymm0, %ymm0
+ vpor 32*4-256(%rax), %ymm0, %ymm0
+ lea 256(%rax), %rcx # size optimization
+ vpor 32*5-256(%rax), %ymm0, %ymm0
+ vpor 32*6-256(%rax), %ymm0, %ymm0
+ vpor 32*7-256(%rax), %ymm0, %ymm0
+ vpor 32*8-256(%rax), %ymm0, %ymm0
+ vpor 32*9-256(%rax), %ymm0, %ymm0
+ vpor 32*10-256(%rax), %ymm0, %ymm0
+ vpor 32*11-256(%rax), %ymm0, %ymm0
+ vpor 32*12-512(%rcx), %ymm0, %ymm0
+ vpor 32*13-512(%rcx), %ymm0, %ymm0
+ vpor 32*14-512(%rcx), %ymm0, %ymm0
+ vpor 32*15-512(%rcx), %ymm0, %ymm0
+ vpor 32*16-512(%rcx), %ymm0, %ymm0
+ vpor 32*17-512(%rcx), %ymm0, %ymm0
+ vpcmpeqq %ymm1, %ymm0, %ymm0
+ vmovdqa %ymm0, `32*9*8+32`(%rsp)
+
+ # Z1^2 = Z1*Z1
+ lea `32*9*2`($a_ptr), %rsi
+ lea `32*9*2`(%rsp), %rdi
+ lea `32*9*8+32*2`(%rsp), %rcx # temporary vector
+ call avx2_sqr_x4
+ call avx2_normalize_n_store
+
+ # U2 = X2*Z1^2
+ lea `32*9*0`($b_ptr), %rsi
+ lea `32*9*2`(%rsp), %rdx
+ lea `32*9*0`(%rsp), %rdi
+ call avx2_mul_x4
+ #call avx2_normalize
+ `&STORE`
+
+ # S2 = Z1*Z1^2 = Z1^3
+ lea `32*9*2`($a_ptr), %rsi
+ lea `32*9*2`(%rsp), %rdx
+ lea `32*9*1`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ # S2 = S2*Y2 = Y2*Z1^3
+ lea `32*9*1`($b_ptr), %rsi
+ lea `32*9*1`(%rsp), %rdx
+ lea `32*9*1`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ # H = U2 - U1 = U2 - X1
+ lea `32*9*0`(%rsp), %rsi
+ lea `32*9*0`($a_ptr), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ # R = S2 - S1 = S2 - Y1
+ lea `32*9*1`(%rsp), %rsi
+ lea `32*9*1`($a_ptr), %rdx
+ lea `32*9*4`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ # Z3 = H*Z1*Z2
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*2`($a_ptr), %rdx
+ lea `32*9*2`($r_ptr), %rdi
+ call avx2_mul_x4
+ call avx2_normalize
+
+ lea .LONE(%rip), %rsi
+ lea `32*9*2`($a_ptr), %rdx
+ call avx2_select_n_store
+
+ # R^2 = R^2
+ lea `32*9*4`(%rsp), %rsi
+ lea `32*9*6`(%rsp), %rdi
+ lea `32*9*8+32*2`(%rsp), %rcx # temporary vector
+ call avx2_sqr_x4
+ call avx2_normalize_n_store
+
+ # H^2 = H^2
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*5`(%rsp), %rdi
+ call avx2_sqr_x4
+ call avx2_normalize_n_store
+
+ # H^3 = H^2*H
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*5`(%rsp), %rdx
+ lea `32*9*7`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ # U2 = U1*H^2
+ lea `32*9*0`($a_ptr), %rsi
+ lea `32*9*5`(%rsp), %rdx
+ lea `32*9*0`(%rsp), %rdi
+ call avx2_mul_x4
+ #call avx2_normalize
+ `&STORE`
+
+ # Hsqr = U2*2
+ #lea 32*9*0(%rsp), %rsi
+ #lea 32*9*5(%rsp), %rdi
+ #call avx2_mul_by2_x4
+
+ vpaddq $ACC0, $ACC0, $ACC0 # inlined avx2_mul_by2_x4
+ lea `32*9*5`(%rsp), %rdi
+ vpaddq $ACC1, $ACC1, $ACC1
+ vpaddq $ACC2, $ACC2, $ACC2
+ vpaddq $ACC3, $ACC3, $ACC3
+ vpaddq $ACC4, $ACC4, $ACC4
+ vpaddq $ACC5, $ACC5, $ACC5
+ vpaddq $ACC6, $ACC6, $ACC6
+ vpaddq $ACC7, $ACC7, $ACC7
+ vpaddq $ACC8, $ACC8, $ACC8
+ call avx2_normalize_n_store
+
+ # X3 = R^2 - H^3
+ #lea 32*9*6(%rsp), %rsi
+ #lea 32*9*7(%rsp), %rdx
+ #lea 32*9*5(%rsp), %rcx
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_sub_x4
+ #NORMALIZE
+ #STORE
+
+ # X3 = X3 - U2*2
+ #lea 32*9*0($r_ptr), %rsi
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_sub_x4
+ #NORMALIZE
+ #STORE
+
+ lea `32*9*6+128`(%rsp), %rsi
+ lea .LAVX2_POLY_x2+128(%rip), %rax
+ lea `32*9*7+128`(%rsp), %rdx
+ lea `32*9*5+128`(%rsp), %rcx
+ lea `32*9*0`($r_ptr), %rdi
+
+ vmovdqa 32*0-128(%rsi), $ACC0
+ vmovdqa 32*1-128(%rsi), $ACC1
+ vmovdqa 32*2-128(%rsi), $ACC2
+ vmovdqa 32*3-128(%rsi), $ACC3
+ vmovdqa 32*4-128(%rsi), $ACC4
+ vmovdqa 32*5-128(%rsi), $ACC5
+ vmovdqa 32*6-128(%rsi), $ACC6
+ vmovdqa 32*7-128(%rsi), $ACC7
+ vmovdqa 32*8-128(%rsi), $ACC8
+
+ vpaddq 32*0-128(%rax), $ACC0, $ACC0
+ vpaddq 32*1-128(%rax), $ACC1, $ACC1
+ vpaddq 32*2-128(%rax), $ACC2, $ACC2
+ vpaddq 32*3-128(%rax), $ACC3, $ACC3
+ vpaddq 32*4-128(%rax), $ACC4, $ACC4
+ vpaddq 32*5-128(%rax), $ACC5, $ACC5
+ vpaddq 32*6-128(%rax), $ACC6, $ACC6
+ vpaddq 32*7-128(%rax), $ACC7, $ACC7
+ vpaddq 32*8-128(%rax), $ACC8, $ACC8
+
+ vpsubq 32*0-128(%rdx), $ACC0, $ACC0
+ vpsubq 32*1-128(%rdx), $ACC1, $ACC1
+ vpsubq 32*2-128(%rdx), $ACC2, $ACC2
+ vpsubq 32*3-128(%rdx), $ACC3, $ACC3
+ vpsubq 32*4-128(%rdx), $ACC4, $ACC4
+ vpsubq 32*5-128(%rdx), $ACC5, $ACC5
+ vpsubq 32*6-128(%rdx), $ACC6, $ACC6
+ vpsubq 32*7-128(%rdx), $ACC7, $ACC7
+ vpsubq 32*8-128(%rdx), $ACC8, $ACC8
+
+ vpsubq 32*0-128(%rcx), $ACC0, $ACC0
+ vpsubq 32*1-128(%rcx), $ACC1, $ACC1
+ vpsubq 32*2-128(%rcx), $ACC2, $ACC2
+ vpsubq 32*3-128(%rcx), $ACC3, $ACC3
+ vpsubq 32*4-128(%rcx), $ACC4, $ACC4
+ vpsubq 32*5-128(%rcx), $ACC5, $ACC5
+ vpsubq 32*6-128(%rcx), $ACC6, $ACC6
+ vpsubq 32*7-128(%rcx), $ACC7, $ACC7
+ vpsubq 32*8-128(%rcx), $ACC8, $ACC8
+ call avx2_normalize
+
+ lea 32*0($b_ptr), %rsi
+ lea 32*0($a_ptr), %rdx
+ call avx2_select_n_store
+
+ # H = U2 - X3
+ lea `32*9*0`(%rsp), %rsi
+ lea `32*9*0`($r_ptr), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ #
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*4`(%rsp), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ #
+ lea `32*9*7`(%rsp), %rsi
+ lea `32*9*1`($a_ptr), %rdx
+ lea `32*9*1`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ #
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*1`(%rsp), %rdx
+ lea `32*9*1`($r_ptr), %rdi
+ call avx2_sub_x4
+ call avx2_normalize
+
+ lea 32*9($b_ptr), %rsi
+ lea 32*9($a_ptr), %rdx
+ call avx2_select_n_store
+
+ #lea 32*9*0($r_ptr), %rsi
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_mul_by1_x4
+ #NORMALIZE
+ #STORE
+
+ lea `32*9*1`($r_ptr), %rsi
+ lea `32*9*1`($r_ptr), %rdi
+ call avx2_mul_by1_x4
+ call avx2_normalize_n_store
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps %xmm6, -16*10(%rbp)
+ movaps %xmm7, -16*9(%rbp)
+ movaps %xmm8, -16*8(%rbp)
+ movaps %xmm9, -16*7(%rbp)
+ movaps %xmm10, -16*6(%rbp)
+ movaps %xmm11, -16*5(%rbp)
+ movaps %xmm12, -16*4(%rbp)
+ movaps %xmm13, -16*3(%rbp)
+ movaps %xmm14, -16*2(%rbp)
+ movaps %xmm15, -16*1(%rbp)
+___
+$code.=<<___;
+ mov %rbp, %rsp
+ pop %rbp
+ ret
+.size ecp_nistz256_avx2_point_add_affine_x4,.-ecp_nistz256_avx2_point_add_affine_x4
+
+################################################################################
+# void ecp_nistz256_avx2_point_add_affines_x4(void* RESULTx4, void *Ax4, void *Bx4);
+.globl ecp_nistz256_avx2_point_add_affines_x4
+.type ecp_nistz256_avx2_point_add_affines_x4,\@function,3
+.align 32
+ecp_nistz256_avx2_point_add_affines_x4:
+ mov %rsp, %rax
+ push %rbp
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ lea -8(%rax), %rbp
+
+# Result + 32*0 = Result.X
+# Result + 32*9 = Result.Y
+# Result + 32*18 = Result.Z
+
+# A + 32*0 = A.X
+# A + 32*9 = A.Y
+
+# B + 32*0 = B.X
+# B + 32*9 = B.Y
+
+ sub \$`32*9*8+32*2+32*8`, %rsp
+ and \$-64, %rsp
+
+ mov $r_ptr_in, $r_ptr
+ mov $a_ptr_in, $a_ptr
+ mov $b_ptr_in, $b_ptr
+
+ vmovdqa 32*0($a_ptr_in), %ymm0
+ vmovdqa .LAVX2_AND_MASK(%rip), $AND_MASK
+ vpxor %ymm1, %ymm1, %ymm1
+ lea 256($a_ptr_in), %rax # size optimization
+ vpor 32*1($a_ptr_in), %ymm0, %ymm0
+ vpor 32*2($a_ptr_in), %ymm0, %ymm0
+ vpor 32*3($a_ptr_in), %ymm0, %ymm0
+ vpor 32*4-256(%rax), %ymm0, %ymm0
+ lea 256(%rax), %rcx # size optimization
+ vpor 32*5-256(%rax), %ymm0, %ymm0
+ vpor 32*6-256(%rax), %ymm0, %ymm0
+ vpor 32*7-256(%rax), %ymm0, %ymm0
+ vpor 32*8-256(%rax), %ymm0, %ymm0
+ vpor 32*9-256(%rax), %ymm0, %ymm0
+ vpor 32*10-256(%rax), %ymm0, %ymm0
+ vpor 32*11-256(%rax), %ymm0, %ymm0
+ vpor 32*12-512(%rcx), %ymm0, %ymm0
+ vpor 32*13-512(%rcx), %ymm0, %ymm0
+ vpor 32*14-512(%rcx), %ymm0, %ymm0
+ vpor 32*15-512(%rcx), %ymm0, %ymm0
+ vpor 32*16-512(%rcx), %ymm0, %ymm0
+ vpor 32*17-512(%rcx), %ymm0, %ymm0
+ vpcmpeqq %ymm1, %ymm0, %ymm0
+ vmovdqa %ymm0, `32*9*8`(%rsp)
+
+ vpxor %ymm1, %ymm1, %ymm1
+ vmovdqa 32*0($b_ptr), %ymm0
+ lea 256($b_ptr), %rax # size optimization
+ vpor 32*1($b_ptr), %ymm0, %ymm0
+ vpor 32*2($b_ptr), %ymm0, %ymm0
+ vpor 32*3($b_ptr), %ymm0, %ymm0
+ vpor 32*4-256(%rax), %ymm0, %ymm0
+ lea 256(%rax), %rcx # size optimization
+ vpor 32*5-256(%rax), %ymm0, %ymm0
+ vpor 32*6-256(%rax), %ymm0, %ymm0
+ vpor 32*7-256(%rax), %ymm0, %ymm0
+ vpor 32*8-256(%rax), %ymm0, %ymm0
+ vpor 32*9-256(%rax), %ymm0, %ymm0
+ vpor 32*10-256(%rax), %ymm0, %ymm0
+ vpor 32*11-256(%rax), %ymm0, %ymm0
+ vpor 32*12-512(%rcx), %ymm0, %ymm0
+ vpor 32*13-512(%rcx), %ymm0, %ymm0
+ vpor 32*14-512(%rcx), %ymm0, %ymm0
+ vpor 32*15-512(%rcx), %ymm0, %ymm0
+ vpor 32*16-512(%rcx), %ymm0, %ymm0
+ vpor 32*17-512(%rcx), %ymm0, %ymm0
+ vpcmpeqq %ymm1, %ymm0, %ymm0
+ vmovdqa %ymm0, `32*9*8+32`(%rsp)
+
+ # H = U2 - U1 = X2 - X1
+ lea `32*9*0`($b_ptr), %rsi
+ lea `32*9*0`($a_ptr), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ # R = S2 - S1 = Y2 - Y1
+ lea `32*9*1`($b_ptr), %rsi
+ lea `32*9*1`($a_ptr), %rdx
+ lea `32*9*4`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ # Z3 = H*Z1*Z2 = H
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*2`($r_ptr), %rdi
+ call avx2_mul_by1_x4
+ call avx2_normalize
+
+ vmovdqa `32*9*8`(%rsp), $B
+ vpor `32*9*8+32`(%rsp), $B, $B
+
+ vpandn $ACC0, $B, $ACC0
+ lea .LONE+128(%rip), %rax
+ vpandn $ACC1, $B, $ACC1
+ vpandn $ACC2, $B, $ACC2
+ vpandn $ACC3, $B, $ACC3
+ vpandn $ACC4, $B, $ACC4
+ vpandn $ACC5, $B, $ACC5
+ vpandn $ACC6, $B, $ACC6
+ vpandn $ACC7, $B, $ACC7
+
+ vpand 32*0-128(%rax), $B, $T0
+ vpandn $ACC8, $B, $ACC8
+ vpand 32*1-128(%rax), $B, $Y
+ vpxor $T0, $ACC0, $ACC0
+ vpand 32*2-128(%rax), $B, $T0
+ vpxor $Y, $ACC1, $ACC1
+ vpand 32*3-128(%rax), $B, $Y
+ vpxor $T0, $ACC2, $ACC2
+ vpand 32*4-128(%rax), $B, $T0
+ vpxor $Y, $ACC3, $ACC3
+ vpand 32*5-128(%rax), $B, $Y
+ vpxor $T0, $ACC4, $ACC4
+ vpand 32*6-128(%rax), $B, $T0
+ vpxor $Y, $ACC5, $ACC5
+ vpand 32*7-128(%rax), $B, $Y
+ vpxor $T0, $ACC6, $ACC6
+ vpand 32*8-128(%rax), $B, $T0
+ vpxor $Y, $ACC7, $ACC7
+ vpxor $T0, $ACC8, $ACC8
+ `&STORE`
+
+ # R^2 = R^2
+ lea `32*9*4`(%rsp), %rsi
+ lea `32*9*6`(%rsp), %rdi
+ lea `32*9*8+32*2`(%rsp), %rcx # temporary vector
+ call avx2_sqr_x4
+ call avx2_normalize_n_store
+
+ # H^2 = H^2
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*5`(%rsp), %rdi
+ call avx2_sqr_x4
+ call avx2_normalize_n_store
+
+ # H^3 = H^2*H
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*5`(%rsp), %rdx
+ lea `32*9*7`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ # U2 = U1*H^2
+ lea `32*9*0`($a_ptr), %rsi
+ lea `32*9*5`(%rsp), %rdx
+ lea `32*9*0`(%rsp), %rdi
+ call avx2_mul_x4
+ #call avx2_normalize
+ `&STORE`
+
+ # Hsqr = U2*2
+ #lea 32*9*0(%rsp), %rsi
+ #lea 32*9*5(%rsp), %rdi
+ #call avx2_mul_by2_x4
+
+ vpaddq $ACC0, $ACC0, $ACC0 # inlined avx2_mul_by2_x4
+ lea `32*9*5`(%rsp), %rdi
+ vpaddq $ACC1, $ACC1, $ACC1
+ vpaddq $ACC2, $ACC2, $ACC2
+ vpaddq $ACC3, $ACC3, $ACC3
+ vpaddq $ACC4, $ACC4, $ACC4
+ vpaddq $ACC5, $ACC5, $ACC5
+ vpaddq $ACC6, $ACC6, $ACC6
+ vpaddq $ACC7, $ACC7, $ACC7
+ vpaddq $ACC8, $ACC8, $ACC8
+ call avx2_normalize_n_store
+
+ # X3 = R^2 - H^3
+ #lea 32*9*6(%rsp), %rsi
+ #lea 32*9*7(%rsp), %rdx
+ #lea 32*9*5(%rsp), %rcx
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_sub_x4
+ #NORMALIZE
+ #STORE
+
+ # X3 = X3 - U2*2
+ #lea 32*9*0($r_ptr), %rsi
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_sub_x4
+ #NORMALIZE
+ #STORE
+
+ lea `32*9*6+128`(%rsp), %rsi
+ lea .LAVX2_POLY_x2+128(%rip), %rax
+ lea `32*9*7+128`(%rsp), %rdx
+ lea `32*9*5+128`(%rsp), %rcx
+ lea `32*9*0`($r_ptr), %rdi
+
+ vmovdqa 32*0-128(%rsi), $ACC0
+ vmovdqa 32*1-128(%rsi), $ACC1
+ vmovdqa 32*2-128(%rsi), $ACC2
+ vmovdqa 32*3-128(%rsi), $ACC3
+ vmovdqa 32*4-128(%rsi), $ACC4
+ vmovdqa 32*5-128(%rsi), $ACC5
+ vmovdqa 32*6-128(%rsi), $ACC6
+ vmovdqa 32*7-128(%rsi), $ACC7
+ vmovdqa 32*8-128(%rsi), $ACC8
+
+ vpaddq 32*0-128(%rax), $ACC0, $ACC0
+ vpaddq 32*1-128(%rax), $ACC1, $ACC1
+ vpaddq 32*2-128(%rax), $ACC2, $ACC2
+ vpaddq 32*3-128(%rax), $ACC3, $ACC3
+ vpaddq 32*4-128(%rax), $ACC4, $ACC4
+ vpaddq 32*5-128(%rax), $ACC5, $ACC5
+ vpaddq 32*6-128(%rax), $ACC6, $ACC6
+ vpaddq 32*7-128(%rax), $ACC7, $ACC7
+ vpaddq 32*8-128(%rax), $ACC8, $ACC8
+
+ vpsubq 32*0-128(%rdx), $ACC0, $ACC0
+ vpsubq 32*1-128(%rdx), $ACC1, $ACC1
+ vpsubq 32*2-128(%rdx), $ACC2, $ACC2
+ vpsubq 32*3-128(%rdx), $ACC3, $ACC3
+ vpsubq 32*4-128(%rdx), $ACC4, $ACC4
+ vpsubq 32*5-128(%rdx), $ACC5, $ACC5
+ vpsubq 32*6-128(%rdx), $ACC6, $ACC6
+ vpsubq 32*7-128(%rdx), $ACC7, $ACC7
+ vpsubq 32*8-128(%rdx), $ACC8, $ACC8
+
+ vpsubq 32*0-128(%rcx), $ACC0, $ACC0
+ vpsubq 32*1-128(%rcx), $ACC1, $ACC1
+ vpsubq 32*2-128(%rcx), $ACC2, $ACC2
+ vpsubq 32*3-128(%rcx), $ACC3, $ACC3
+ vpsubq 32*4-128(%rcx), $ACC4, $ACC4
+ vpsubq 32*5-128(%rcx), $ACC5, $ACC5
+ vpsubq 32*6-128(%rcx), $ACC6, $ACC6
+ vpsubq 32*7-128(%rcx), $ACC7, $ACC7
+ vpsubq 32*8-128(%rcx), $ACC8, $ACC8
+ call avx2_normalize
+
+ lea 32*0($b_ptr), %rsi
+ lea 32*0($a_ptr), %rdx
+ call avx2_select_n_store
+
+ # H = U2 - X3
+ lea `32*9*0`(%rsp), %rsi
+ lea `32*9*0`($r_ptr), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_sub_x4
+ call avx2_normalize_n_store
+
+ # H = H*R
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*4`(%rsp), %rdx
+ lea `32*9*3`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ # S2 = S1 * H^3
+ lea `32*9*7`(%rsp), %rsi
+ lea `32*9*1`($a_ptr), %rdx
+ lea `32*9*1`(%rsp), %rdi
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ #
+ lea `32*9*3`(%rsp), %rsi
+ lea `32*9*1`(%rsp), %rdx
+ lea `32*9*1`($r_ptr), %rdi
+ call avx2_sub_x4
+ call avx2_normalize
+
+ lea 32*9($b_ptr), %rsi
+ lea 32*9($a_ptr), %rdx
+ call avx2_select_n_store
+
+ #lea 32*9*0($r_ptr), %rsi
+ #lea 32*9*0($r_ptr), %rdi
+ #call avx2_mul_by1_x4
+ #NORMALIZE
+ #STORE
+
+ lea `32*9*1`($r_ptr), %rsi
+ lea `32*9*1`($r_ptr), %rdi
+ call avx2_mul_by1_x4
+ call avx2_normalize_n_store
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps %xmm6, -16*10(%rbp)
+ movaps %xmm7, -16*9(%rbp)
+ movaps %xmm8, -16*8(%rbp)
+ movaps %xmm9, -16*7(%rbp)
+ movaps %xmm10, -16*6(%rbp)
+ movaps %xmm11, -16*5(%rbp)
+ movaps %xmm12, -16*4(%rbp)
+ movaps %xmm13, -16*3(%rbp)
+ movaps %xmm14, -16*2(%rbp)
+ movaps %xmm15, -16*1(%rbp)
+___
+$code.=<<___;
+ mov %rbp, %rsp
+ pop %rbp
+ ret
+.size ecp_nistz256_avx2_point_add_affines_x4,.-ecp_nistz256_avx2_point_add_affines_x4
+
+################################################################################
+# void ecp_nistz256_avx2_to_mont(void* RESULTx4, void *Ax4);
+.globl ecp_nistz256_avx2_to_mont
+.type ecp_nistz256_avx2_to_mont,\@function,2
+.align 32
+ecp_nistz256_avx2_to_mont:
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -8-16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ vmovdqa .LAVX2_AND_MASK(%rip), $AND_MASK
+ lea .LTO_MONT_AVX2(%rip), %rdx
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps 16*0(%rsp), %xmm6
+ movaps 16*1(%rsp), %xmm7
+ movaps 16*2(%rsp), %xmm8
+ movaps 16*3(%rsp), %xmm9
+ movaps 16*4(%rsp), %xmm10
+ movaps 16*5(%rsp), %xmm11
+ movaps 16*6(%rsp), %xmm12
+ movaps 16*7(%rsp), %xmm13
+ movaps 16*8(%rsp), %xmm14
+ movaps 16*9(%rsp), %xmm15
+ lea 8+16*10(%rsp), %rsp
+___
+$code.=<<___;
+ ret
+.size ecp_nistz256_avx2_to_mont,.-ecp_nistz256_avx2_to_mont
+
+################################################################################
+# void ecp_nistz256_avx2_from_mont(void* RESULTx4, void *Ax4);
+.globl ecp_nistz256_avx2_from_mont
+.type ecp_nistz256_avx2_from_mont,\@function,2
+.align 32
+ecp_nistz256_avx2_from_mont:
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -8-16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ vmovdqa .LAVX2_AND_MASK(%rip), $AND_MASK
+ lea .LFROM_MONT_AVX2(%rip), %rdx
+ call avx2_mul_x4
+ call avx2_normalize_n_store
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps 16*0(%rsp), %xmm6
+ movaps 16*1(%rsp), %xmm7
+ movaps 16*2(%rsp), %xmm8
+ movaps 16*3(%rsp), %xmm9
+ movaps 16*4(%rsp), %xmm10
+ movaps 16*5(%rsp), %xmm11
+ movaps 16*6(%rsp), %xmm12
+ movaps 16*7(%rsp), %xmm13
+ movaps 16*8(%rsp), %xmm14
+ movaps 16*9(%rsp), %xmm15
+ lea 8+16*10(%rsp), %rsp
+___
+$code.=<<___;
+ ret
+.size ecp_nistz256_avx2_from_mont,.-ecp_nistz256_avx2_from_mont
+
+################################################################################
+# void ecp_nistz256_avx2_set1(void* RESULTx4);
+.globl ecp_nistz256_avx2_set1
+.type ecp_nistz256_avx2_set1,\@function,1
+.align 32
+ecp_nistz256_avx2_set1:
+ lea .LONE+128(%rip), %rax
+ lea 128(%rdi), %rdi
+ vzeroupper
+ vmovdqa 32*0-128(%rax), %ymm0
+ vmovdqa 32*1-128(%rax), %ymm1
+ vmovdqa 32*2-128(%rax), %ymm2
+ vmovdqa 32*3-128(%rax), %ymm3
+ vmovdqa 32*4-128(%rax), %ymm4
+ vmovdqa 32*5-128(%rax), %ymm5
+ vmovdqa %ymm0, 32*0-128(%rdi)
+ vmovdqa 32*6-128(%rax), %ymm0
+ vmovdqa %ymm1, 32*1-128(%rdi)
+ vmovdqa 32*7-128(%rax), %ymm1
+ vmovdqa %ymm2, 32*2-128(%rdi)
+ vmovdqa 32*8-128(%rax), %ymm2
+ vmovdqa %ymm3, 32*3-128(%rdi)
+ vmovdqa %ymm4, 32*4-128(%rdi)
+ vmovdqa %ymm5, 32*5-128(%rdi)
+ vmovdqa %ymm0, 32*6-128(%rdi)
+ vmovdqa %ymm1, 32*7-128(%rdi)
+ vmovdqa %ymm2, 32*8-128(%rdi)
+
+ vzeroupper
+ ret
+.size ecp_nistz256_avx2_set1,.-ecp_nistz256_avx2_set1
+___
+}
+{
+################################################################################
+# void ecp_nistz256_avx2_multi_select_w7(void* RESULT, void *in,
+# int index0, int index1, int index2, int index3);
+################################################################################
+
+my ($val,$in_t,$index0,$index1,$index2,$index3)=("%rdi","%rsi","%edx","%ecx","%r8d","%r9d");
+my ($INDEX0,$INDEX1,$INDEX2,$INDEX3)=map("%ymm$_",(0..3));
+my ($R0a,$R0b,$R1a,$R1b,$R2a,$R2b,$R3a,$R3b)=map("%ymm$_",(4..11));
+my ($M0,$T0,$T1,$TMP0)=map("%ymm$_",(12..15));
+
+$code.=<<___;
+.globl ecp_nistz256_avx2_multi_select_w7
+.type ecp_nistz256_avx2_multi_select_w7,\@function,6
+.align 32
+ecp_nistz256_avx2_multi_select_w7:
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ lea -8-16*10(%rsp), %rsp
+ vmovaps %xmm6, -8-16*10(%rax)
+ vmovaps %xmm7, -8-16*9(%rax)
+ vmovaps %xmm8, -8-16*8(%rax)
+ vmovaps %xmm9, -8-16*7(%rax)
+ vmovaps %xmm10, -8-16*6(%rax)
+ vmovaps %xmm11, -8-16*5(%rax)
+ vmovaps %xmm12, -8-16*4(%rax)
+ vmovaps %xmm13, -8-16*3(%rax)
+ vmovaps %xmm14, -8-16*2(%rax)
+ vmovaps %xmm15, -8-16*1(%rax)
+___
+$code.=<<___;
+ lea .LIntOne(%rip), %rax
+
+ vmovd $index0, %xmm0
+ vmovd $index1, %xmm1
+ vmovd $index2, %xmm2
+ vmovd $index3, %xmm3
+
+ vpxor $R0a, $R0a, $R0a
+ vpxor $R0b, $R0b, $R0b
+ vpxor $R1a, $R1a, $R1a
+ vpxor $R1b, $R1b, $R1b
+ vpxor $R2a, $R2a, $R2a
+ vpxor $R2b, $R2b, $R2b
+ vpxor $R3a, $R3a, $R3a
+ vpxor $R3b, $R3b, $R3b
+ vmovdqa (%rax), $M0
+
+ vpermd $INDEX0, $R0a, $INDEX0
+ vpermd $INDEX1, $R0a, $INDEX1
+ vpermd $INDEX2, $R0a, $INDEX2
+ vpermd $INDEX3, $R0a, $INDEX3
+
+ mov \$64, %ecx
+ lea 112($val), $val # size optimization
+ jmp .Lmulti_select_loop_avx2
+
+# INDEX=0, corresponds to the point at infty (0,0)
+.align 32
+.Lmulti_select_loop_avx2:
+ vpcmpeqd $INDEX0, $M0, $TMP0
+
+ vmovdqa `32*0+32*64*2*0`($in_t), $T0
+ vmovdqa `32*1+32*64*2*0`($in_t), $T1
+ vpand $TMP0, $T0, $T0
+ vpand $TMP0, $T1, $T1
+ vpxor $T0, $R0a, $R0a
+ vpxor $T1, $R0b, $R0b
+
+ vpcmpeqd $INDEX1, $M0, $TMP0
+
+ vmovdqa `32*0+32*64*2*1`($in_t), $T0
+ vmovdqa `32*1+32*64*2*1`($in_t), $T1
+ vpand $TMP0, $T0, $T0
+ vpand $TMP0, $T1, $T1
+ vpxor $T0, $R1a, $R1a
+ vpxor $T1, $R1b, $R1b
+
+ vpcmpeqd $INDEX2, $M0, $TMP0
+
+ vmovdqa `32*0+32*64*2*2`($in_t), $T0
+ vmovdqa `32*1+32*64*2*2`($in_t), $T1
+ vpand $TMP0, $T0, $T0
+ vpand $TMP0, $T1, $T1
+ vpxor $T0, $R2a, $R2a
+ vpxor $T1, $R2b, $R2b
+
+ vpcmpeqd $INDEX3, $M0, $TMP0
+
+ vmovdqa `32*0+32*64*2*3`($in_t), $T0
+ vmovdqa `32*1+32*64*2*3`($in_t), $T1
+ vpand $TMP0, $T0, $T0
+ vpand $TMP0, $T1, $T1
+ vpxor $T0, $R3a, $R3a
+ vpxor $T1, $R3b, $R3b
+
+ vpaddd (%rax), $M0, $M0 # increment
+ lea 32*2($in_t), $in_t
+
+ dec %ecx
+ jnz .Lmulti_select_loop_avx2
+
+ vmovdqu $R0a, 32*0-112($val)
+ vmovdqu $R0b, 32*1-112($val)
+ vmovdqu $R1a, 32*2-112($val)
+ vmovdqu $R1b, 32*3-112($val)
+ vmovdqu $R2a, 32*4-112($val)
+ vmovdqu $R2b, 32*5-112($val)
+ vmovdqu $R3a, 32*6-112($val)
+ vmovdqu $R3b, 32*7-112($val)
+
+ vzeroupper
+___
+$code.=<<___ if ($win64);
+ movaps 16*0(%rsp), %xmm6
+ movaps 16*1(%rsp), %xmm7
+ movaps 16*2(%rsp), %xmm8
+ movaps 16*3(%rsp), %xmm9
+ movaps 16*4(%rsp), %xmm10
+ movaps 16*5(%rsp), %xmm11
+ movaps 16*6(%rsp), %xmm12
+ movaps 16*7(%rsp), %xmm13
+ movaps 16*8(%rsp), %xmm14
+ movaps 16*9(%rsp), %xmm15
+ lea 8+16*10(%rsp), %rsp
+___
+$code.=<<___;
+ ret
+.size ecp_nistz256_avx2_multi_select_w7,.-ecp_nistz256_avx2_multi_select_w7
+
+.extern OPENSSL_ia32cap_P
+.globl ecp_nistz_avx2_eligible
+.type ecp_nistz_avx2_eligible,\@abi-omnipotent
+.align 32
+ecp_nistz_avx2_eligible:
+ mov OPENSSL_ia32cap_P+8(%rip),%eax
+ shr \$5,%eax
+ and \$1,%eax
+ ret
+.size ecp_nistz_avx2_eligible,.-ecp_nistz_avx2_eligible
+___
+}
+}} else {{ # assembler is too old
+$code.=<<___;
+.text
+
+.globl ecp_nistz256_avx2_transpose_convert
+.globl ecp_nistz256_avx2_convert_transpose_back
+.globl ecp_nistz256_avx2_point_add_affine_x4
+.globl ecp_nistz256_avx2_point_add_affines_x4
+.globl ecp_nistz256_avx2_to_mont
+.globl ecp_nistz256_avx2_from_mont
+.globl ecp_nistz256_avx2_set1
+.globl ecp_nistz256_avx2_multi_select_w7
+.type ecp_nistz256_avx2_multi_select_w7,\@abi-omnipotent
+ecp_nistz256_avx2_transpose_convert:
+ecp_nistz256_avx2_convert_transpose_back:
+ecp_nistz256_avx2_point_add_affine_x4:
+ecp_nistz256_avx2_point_add_affines_x4:
+ecp_nistz256_avx2_to_mont:
+ecp_nistz256_avx2_from_mont:
+ecp_nistz256_avx2_set1:
+ecp_nistz256_avx2_multi_select_w7:
+ .byte 0x0f,0x0b # ud2
+ ret
+.size ecp_nistz256_avx2_multi_select_w7,.-ecp_nistz256_avx2_multi_select_w7
+
+.globl ecp_nistz_avx2_eligible
+.type ecp_nistz_avx2_eligible,\@abi-omnipotent
+ecp_nistz_avx2_eligible:
+ xor %eax,%eax
+ ret
+.size ecp_nistz_avx2_eligible,.-ecp_nistz_avx2_eligible
+___
+}}
+
+foreach (split("\n",$code)) {
+ s/\`([^\`]*)\`/eval($1)/geo;
+
+ print $_,"\n";
+}
+
+close STDOUT;