path: root/crypto/md5/asm/md5-aarch64.pl

#! /usr/bin/env perl
# Copyright 2021-2022 The OpenSSL Project Authors. All Rights Reserved.
#
# Licensed under the Apache License 2.0 (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

# MD5 optimized for aarch64.

use strict;

my $code;

#no warnings qw(uninitialized);
# $output is the last argument if it looks like a file (it has an extension)
# $flavour is the first argument if it doesn't look like a file
my $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
my $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;

$0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate;
( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
die "can't locate arm-xlate.pl";

open OUT,"| \"$^X\" $xlate $flavour \"$output\""
    or die "can't call $xlate: $1";
*STDOUT=*OUT;

$code .= <<EOF;

.text
.globl  ossl_md5_block_asm_data_order
.type   ossl_md5_block_asm_data_order,\@function
ossl_md5_block_asm_data_order:
        // Save all callee-saved registers
        stp     x19,x20,[sp,#-80]!
        stp     x21,x22,[sp,#16]
        stp     x23,x24,[sp,#32]
        stp     x25,x26,[sp,#48]
        stp     x27,x28,[sp,#64]

        ldp w10, w11, [x0, #0]        // Load MD5 state->A and state->B
        ldp w12, w13, [x0, #8]        // Load MD5 state->C and state->D
.align 5
ossl_md5_blocks_loop:
        eor x17, x12, x13             // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        and x16, x17, x11             // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        ldp x15, x3, [x1]             // Load 4 words of input data0 M[0]/0
        eor x14, x16, x13             // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x9, #0xa478              // Load lower half of constant 0xd76aa478
        movk x9, #0xd76a, lsl #16     // Load upper half of constant 0xd76aa478
        add w8, w10, w15              // Add dest value
        add w7, w8, w9                // Add constant 0xd76aa478
        add w6, w7, w14               // Add aux function result
        ror w6, w6, #25               // Rotate left s=7 bits
        eor x5, x11, x12              // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w4, w11, w6               // Add X parameter round 1 A=FF(A, B, C, D, 0xd76aa478, s=7, M[0])
        and x8, x5, x4                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x17, x8, x12              // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x16, #0xb756             // Load lower half of constant 0xe8c7b756
        movk x16, #0xe8c7, lsl #16    // Load upper half of constant 0xe8c7b756
        lsr x20, x15, #32             // Right shift high input value containing M[1]
        add w9, w13, w20              // Add dest value
        add w7, w9, w16               // Add constant 0xe8c7b756
        add w14, w7, w17              // Add aux function result
        ror w14, w14, #20             // Rotate left s=12 bits
        eor x6, x4, x11               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w5, w4, w14               // Add X parameter round 1 D=FF(D, A, B, C, 0xe8c7b756, s=12, M[1])
        and x8, x6, x5                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x9, x8, x11               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x16, #0x70db             // Load lower half of constant 0x242070db
        movk x16, #0x2420, lsl #16    // Load upper half of constant 0x242070db
        add w7, w12, w3               // Add dest value
        add w17, w7, w16              // Add constant 0x242070db
        add w14, w17, w9              // Add aux function result
        ror w14, w14, #15             // Rotate left s=17 bits
        eor x6, x5, x4                // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w8, w5, w14               // Add X parameter round 1 C=FF(C, D, A, B, 0x242070db, s=17, M[2])
        and x7, x6, x8                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x16, x7, x4               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x9, #0xceee              // Load lower half of constant 0xc1bdceee
        movk x9, #0xc1bd, lsl #16     // Load upper half of constant 0xc1bdceee
        lsr x21, x3, #32              // Right shift high input value containing M[3]
        add w14, w11, w21             // Add dest value
        add w6, w14, w9               // Add constant 0xc1bdceee
        add w7, w6, w16               // Add aux function result
        ror w7, w7, #10               // Rotate left s=22 bits
        eor x17, x8, x5               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w9, w8, w7                // Add X parameter round 1 B=FF(B, C, D, A, 0xc1bdceee, s=22, M[3])
        ldp x14, x7, [x1, #16]        // Load 4 words of input data0 M[4]/0w
        and x16, x17, x9              // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x16, x5               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x16, #0xfaf              // Load lower half of constant 0xf57c0faf
        movk x16, #0xf57c, lsl #16    // Load upper half of constant 0xf57c0faf
        add w17, w4, w14              // Add dest value
        add w16, w17, w16             // Add constant 0xf57c0faf
        add w4, w16, w6               // Add aux function result
        ror w4, w4, #25               // Rotate left s=7 bits
        eor x16, x9, x8               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w17, w9, w4               // Add X parameter round 1 A=FF(A, B, C, D, 0xf57c0faf, s=7, M[4])
        and x16, x16, x17             // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x16, x8               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x4, #0xc62a              // Load lower half of constant 0x4787c62a
        movk x4, #0x4787, lsl #16     // Load upper half of constant 0x4787c62a
        lsr x22, x14, #32             // Right shift high input value containing M[5]
        add w16, w5, w22              // Add dest value
        add w16, w16, w4              // Add constant 0x4787c62a
        add w5, w16, w6               // Add aux function result
        ror w5, w5, #20               // Rotate left s=12 bits
        eor x4, x17, x9               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w19, w17, w5              // Add X parameter round 1 D=FF(D, A, B, C, 0x4787c62a, s=12, M[5])
        and x6, x4, x19               // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x5, x6, x9                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x4, #0x4613              // Load lower half of constant 0xa8304613
        movk x4, #0xa830, lsl #16     // Load upper half of constant 0xa8304613
        add w6, w8, w7                // Add dest value
        add w8, w6, w4                // Add constant 0xa8304613
        add w4, w8, w5                // Add aux function result
        ror w4, w4, #15               // Rotate left s=17 bits
        eor x6, x19, x17              // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w8, w19, w4               // Add X parameter round 1 C=FF(C, D, A, B, 0xa8304613, s=17, M[6])
        and x5, x6, x8                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x4, x5, x17               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x6, #0x9501              // Load lower half of constant 0xfd469501
        movk x6, #0xfd46, lsl #16     // Load upper half of constant 0xfd469501
        lsr x23, x7, #32              // Right shift high input value containing M[7]
        add w9, w9, w23               // Add dest value
        add w5, w9, w6                // Add constant 0xfd469501
        add w9, w5, w4                // Add aux function result
        ror w9, w9, #10               // Rotate left s=22 bits
        eor x6, x8, x19               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w4, w8, w9                // Add X parameter round 1 B=FF(B, C, D, A, 0xfd469501, s=22, M[7])
        ldp x5, x16, [x1, #32]        // Load 4 words of input data0 M[8]/0
        and x9, x6, x4                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x9, x19               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x9, #0x98d8              // Load lower half of constant 0x698098d8
        movk x9, #0x6980, lsl #16     // Load upper half of constant 0x698098d8
        add w17, w17, w5              // Add dest value
        add w9, w17, w9               // Add constant 0x698098d8
        add w17, w9, w6               // Add aux function result
        ror w17, w17, #25             // Rotate left s=7 bits
        eor x9, x4, x8                // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w6, w4, w17               // Add X parameter round 1 A=FF(A, B, C, D, 0x698098d8, s=7, M[8])
        and x17, x9, x6               // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x9, x17, x8               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x17, #0xf7af             // Load lower half of constant 0x8b44f7af
        movk x17, #0x8b44, lsl #16    // Load upper half of constant 0x8b44f7af
        lsr x24, x5, #32              // Right shift high input value containing M[9]
        add w19, w19, w24             // Add dest value
        add w17, w19, w17             // Add constant 0x8b44f7af
        add w19, w17, w9              // Add aux function result
        ror w19, w19, #20             // Rotate left s=12 bits
        eor x9, x6, x4                // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w17, w6, w19              // Add X parameter round 1 D=FF(D, A, B, C, 0x8b44f7af, s=12, M[9])
        and x9, x9, x17               // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x9, x9, x4                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x11, #0x5bb1             // Load lower half of constant 0xffff5bb1
        movk x11, #0xffff, lsl #16    // Load upper half of constant 0xffff5bb1
        add w8, w8, w16               // Add dest value
        add w8, w8, w11               // Add constant 0xffff5bb1
        add w8, w8, w9                // Add aux function result
        ror w8, w8, #15               // Rotate left s=17 bits
        eor x9, x17, x6               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w8, w17, w8               // Add X parameter round 1 C=FF(C, D, A, B, 0xffff5bb1, s=17, M[10])
        and x9, x9, x8                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x9, x9, x6                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x11, #0xd7be             // Load lower half of constant 0x895cd7be
        movk x11, #0x895c, lsl #16    // Load upper half of constant 0x895cd7be
        lsr x25, x16, #32             // Right shift high input value containing M[11]
        add w4, w4, w25               // Add dest value
        add w4, w4, w11               // Add constant 0x895cd7be
        add w9, w4, w9                // Add aux function result
        ror w9, w9, #10               // Rotate left s=22 bits
        eor x4, x8, x17               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w9, w8, w9                // Add X parameter round 1 B=FF(B, C, D, A, 0x895cd7be, s=22, M[11])
        ldp x11, x12, [x1, #48]       // Load 4 words of input data0 M[12]/0
        and x4, x4, x9                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x4, x4, x17               // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x19, #0x1122             // Load lower half of constant 0x6b901122
        movk x19, #0x6b90, lsl #16    // Load upper half of constant 0x6b901122
        add w6, w6, w11               // Add dest value
        add w6, w6, w19               // Add constant 0x6b901122
        add w4, w6, w4                // Add aux function result
        ror w4, w4, #25               // Rotate left s=7 bits
        eor x6, x9, x8                // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w4, w9, w4                // Add X parameter round 1 A=FF(A, B, C, D, 0x6b901122, s=7, M[12])
        and x6, x6, x4                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x6, x8                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x19, #0x7193             // Load lower half of constant 0xfd987193
        movk x19, #0xfd98, lsl #16    // Load upper half of constant 0xfd987193
        lsr x26, x11, #32             // Right shift high input value containing M[13]
        add w17, w17, w26             // Add dest value
        add w17, w17, w19             // Add constant 0xfd987193
        add w17, w17, w6              // Add aux function result
        ror w17, w17, #20             // Rotate left s=12 bits
        eor x6, x4, x9                // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w17, w4, w17              // Add X parameter round 1 D=FF(D, A, B, C, 0xfd987193, s=12, M[13])
        and x6, x6, x17               // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x6, x9                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x13, #0x438e             // Load lower half of constant 0xa679438e
        movk x13, #0xa679, lsl #16    // Load upper half of constant 0xa679438e
        add w8, w8, w12               // Add dest value
        add w8, w8, w13               // Add constant 0xa679438e
        add w8, w8, w6                // Add aux function result
        ror w8, w8, #15               // Rotate left s=17 bits
        eor x6, x17, x4               // Begin aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        add w8, w17, w8               // Add X parameter round 1 C=FF(C, D, A, B, 0xa679438e, s=17, M[14])
        and x6, x6, x8                // Continue aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        eor x6, x6, x4                // End aux function round 1 F(x,y,z)=(((y^z)&x)^z)
        movz x13, #0x821              // Load lower half of constant 0x49b40821
        movk x13, #0x49b4, lsl #16    // Load upper half of constant 0x49b40821
        lsr x27, x12, #32             // Right shift high input value containing M[15]
        add w9, w9, w27               // Add dest value
        add w9, w9, w13               // Add constant 0x49b40821
        add w9, w9, w6                // Add aux function result
        ror w9, w9, #10               // Rotate left s=22 bits
        bic x6, x8, x17               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w9, w8, w9                // Add X parameter round 1 B=FF(B, C, D, A, 0x49b40821, s=22, M[15])
        and x13, x9, x17              // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0x2562             // Load lower half of constant 0xf61e2562
        movk x13, #0xf61e, lsl #16    // Load upper half of constant 0xf61e2562
        add w4, w4, w20               // Add dest value
        add w4, w4, w13               // Add constant 0xf61e2562
        add w4, w4, w6                // Add aux function result
        ror w4, w4, #27               // Rotate left s=5 bits
        bic x6, x9, x8                // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w4, w9, w4                // Add X parameter round 2 A=GG(A, B, C, D, 0xf61e2562, s=5, M[1])
        and x13, x4, x8               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0xb340             // Load lower half of constant 0xc040b340
        movk x13, #0xc040, lsl #16    // Load upper half of constant 0xc040b340
        add w17, w17, w7              // Add dest value
        add w17, w17, w13             // Add constant 0xc040b340
        add w17, w17, w6              // Add aux function result
        ror w17, w17, #23             // Rotate left s=9 bits
        bic x6, x4, x9                // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w17, w4, w17              // Add X parameter round 2 D=GG(D, A, B, C, 0xc040b340, s=9, M[6])
        and x13, x17, x9              // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0x5a51             // Load lower half of constant 0x265e5a51
        movk x13, #0x265e, lsl #16    // Load upper half of constant 0x265e5a51
        add w8, w8, w25               // Add dest value
        add w8, w8, w13               // Add constant 0x265e5a51
        add w8, w8, w6                // Add aux function result
        ror w8, w8, #18               // Rotate left s=14 bits
        bic x6, x17, x4               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w8, w17, w8               // Add X parameter round 2 C=GG(C, D, A, B, 0x265e5a51, s=14, M[11])
        and x13, x8, x4               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0xc7aa             // Load lower half of constant 0xe9b6c7aa
        movk x13, #0xe9b6, lsl #16    // Load upper half of constant 0xe9b6c7aa
        add w9, w9, w15               // Add dest value
        add w9, w9, w13               // Add constant 0xe9b6c7aa
        add w9, w9, w6                // Add aux function result
        ror w9, w9, #12               // Rotate left s=20 bits
        bic x6, x8, x17               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w9, w8, w9                // Add X parameter round 2 B=GG(B, C, D, A, 0xe9b6c7aa, s=20, M[0])
        and x13, x9, x17              // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0x105d             // Load lower half of constant 0xd62f105d
        movk x13, #0xd62f, lsl #16    // Load upper half of constant 0xd62f105d
        add w4, w4, w22               // Add dest value
        add w4, w4, w13               // Add constant 0xd62f105d
        add w4, w4, w6                // Add aux function result
        ror w4, w4, #27               // Rotate left s=5 bits
        bic x6, x9, x8                // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w4, w9, w4                // Add X parameter round 2 A=GG(A, B, C, D, 0xd62f105d, s=5, M[5])
        and x13, x4, x8               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0x1453             // Load lower half of constant 0x2441453
        movk x13, #0x244, lsl #16     // Load upper half of constant 0x2441453
        add w17, w17, w16             // Add dest value
        add w17, w17, w13             // Add constant 0x2441453
        add w17, w17, w6              // Add aux function result
        ror w17, w17, #23             // Rotate left s=9 bits
        bic x6, x4, x9                // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w17, w4, w17              // Add X parameter round 2 D=GG(D, A, B, C, 0x2441453, s=9, M[10])
        and x13, x17, x9              // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0xe681             // Load lower half of constant 0xd8a1e681
        movk x13, #0xd8a1, lsl #16    // Load upper half of constant 0xd8a1e681
        add w8, w8, w27               // Add dest value
        add w8, w8, w13               // Add constant 0xd8a1e681
        add w8, w8, w6                // Add aux function result
        ror w8, w8, #18               // Rotate left s=14 bits
        bic x6, x17, x4               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        add w8, w17, w8               // Add X parameter round 2 C=GG(C, D, A, B, 0xd8a1e681, s=14, M[15])
        and x13, x8, x4               // Aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        orr x6, x6, x13               // End aux function round 2 G(x,y,z)=((x&z)|(~z&y))
        movz x13, #0xfbc8             // Load lower half of constant 0xe7d3fbc8
        movk x13, #0xe7d3, lsl #16    // Load upper half of constant 0xe7d3fbc8
        add w9, w9, w14               // Add dest value
        add w9, w9, w13               // Add constant 0xe7d3fbc8
        add w9, w9, w6                // Add aux function result
        ror w9, w9, #12               // Rotate left s=20 bits<