|
|
727bdf |
diff --git a/crypto/bn/asm/ppc64-mont-fixed.pl b/crypto/bn/asm/ppc64-mont-fixed.pl
|
|
|
727bdf |
index 56df89dc27da..e69de29bb2d1 100755
|
|
|
727bdf |
--- a/crypto/bn/asm/ppc64-mont-fixed.pl
|
|
|
727bdf |
+++ b/crypto/bn/asm/ppc64-mont-fixed.pl
|
|
|
727bdf |
@@ -1,581 +0,0 @@
|
|
|
727bdf |
-#! /usr/bin/env perl
|
|
|
727bdf |
-# Copyright 2021 The OpenSSL Project Authors. All Rights Reserved.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# Licensed under the Apache License 2.0 (the "License"). You may not use
|
|
|
727bdf |
-# this file except in compliance with the License. You can obtain a copy
|
|
|
727bdf |
-# in the file LICENSE in the source distribution or at
|
|
|
727bdf |
-# https://www.openssl.org/source/license.html
|
|
|
727bdf |
-
|
|
|
727bdf |
-# ====================================================================
|
|
|
727bdf |
-# Written by Amitay Isaacs <amitay@ozlabs.org>, Martin Schwenke
|
|
|
727bdf |
-# <martin@meltin.net> & Alastair D'Silva <alastair@d-silva.org> for
|
|
|
727bdf |
-# the OpenSSL project.
|
|
|
727bdf |
-# ====================================================================
|
|
|
727bdf |
-
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# Fixed length (n=6), unrolled PPC Montgomery Multiplication
|
|
|
727bdf |
-#
|
|
|
727bdf |
-
|
|
|
727bdf |
-# 2021
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# Although this is a generic implementation for unrolling Montgomery
|
|
|
727bdf |
-# Multiplication for arbitrary values of n, this is currently only
|
|
|
727bdf |
-# used for n = 6 to improve the performance of ECC p384.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# Unrolling allows intermediate results to be stored in registers,
|
|
|
727bdf |
-# rather than on the stack, improving performance by ~7% compared to
|
|
|
727bdf |
-# the existing PPC assembly code.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# The ISA 3.0 implementation uses combination multiply/add
|
|
|
727bdf |
-# instructions (maddld, maddhdu) to improve performance by an
|
|
|
727bdf |
-# additional ~10% on Power 9.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# Finally, saving non-volatile registers into volatile vector
|
|
|
727bdf |
-# registers instead of onto the stack saves a little more.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# On a Power 9 machine we see an overall improvement of ~18%.
|
|
|
727bdf |
-#
|
|
|
727bdf |
-
|
|
|
727bdf |
-use strict;
|
|
|
727bdf |
-use warnings;
|
|
|
727bdf |
-
|
|
|
727bdf |
-my ($flavour, $output, $dir, $xlate);
|
|
|
727bdf |
-
|
|
|
727bdf |
-# $output is the last argument if it looks like a file (it has an extension)
|
|
|
727bdf |
-# $flavour is the first argument if it doesn't look like a file
|
|
|
727bdf |
-$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
|
|
|
727bdf |
-$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
|
|
|
727bdf |
-
|
|
|
727bdf |
-$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
|
|
|
727bdf |
-( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
|
|
|
727bdf |
-( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
|
|
|
727bdf |
-die "can't locate ppc-xlate.pl";
|
|
|
727bdf |
-
|
|
|
727bdf |
-open STDOUT,"| $^X $xlate $flavour \"$output\""
|
|
|
727bdf |
- or die "can't call $xlate: $!";
|
|
|
727bdf |
-
|
|
|
727bdf |
-if ($flavour !~ /64/) {
|
|
|
727bdf |
- die "bad flavour ($flavour) - only ppc64 permitted";
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-my $SIZE_T= 8;
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Registers are global so the code is remotely readable
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Parameters for Montgomery multiplication
|
|
|
727bdf |
-my $sp = "r1";
|
|
|
727bdf |
-my $toc = "r2";
|
|
|
727bdf |
-my $rp = "r3";
|
|
|
727bdf |
-my $ap = "r4";
|
|
|
727bdf |
-my $bp = "r5";
|
|
|
727bdf |
-my $np = "r6";
|
|
|
727bdf |
-my $n0 = "r7";
|
|
|
727bdf |
-my $num = "r8";
|
|
|
727bdf |
-
|
|
|
727bdf |
-my $i = "r9";
|
|
|
727bdf |
-my $c0 = "r10";
|
|
|
727bdf |
-my $bp0 = "r11";
|
|
|
727bdf |
-my $bpi = "r11";
|
|
|
727bdf |
-my $bpj = "r11";
|
|
|
727bdf |
-my $tj = "r12";
|
|
|
727bdf |
-my $apj = "r12";
|
|
|
727bdf |
-my $npj = "r12";
|
|
|
727bdf |
-my $lo = "r14";
|
|
|
727bdf |
-my $c1 = "r14";
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Non-volatile registers used for tp[i]
|
|
|
727bdf |
-#
|
|
|
727bdf |
-# 12 registers are available but the limit on unrolling is 10,
|
|
|
727bdf |
-# since registers from $tp[0] to $tp[$n+1] are used.
|
|
|
727bdf |
-my @tp = ("r20" .. "r31");
|
|
|
727bdf |
-
|
|
|
727bdf |
-# volatile VSRs for saving non-volatile GPRs - faster than stack
|
|
|
727bdf |
-my @vsrs = ("v32" .. "v46");
|
|
|
727bdf |
-
|
|
|
727bdf |
-package Mont;
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub new($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($class, $n) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- if ($n > 10) {
|
|
|
727bdf |
- die "Can't unroll for BN length ${n} (maximum 10)"
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- my $self = {
|
|
|
727bdf |
- code => "",
|
|
|
727bdf |
- n => $n,
|
|
|
727bdf |
- };
|
|
|
727bdf |
- bless $self, $class;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return $self;
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub add_code($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->{code} .= $c;
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_code($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return $self->{code};
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_function_name($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return "bn_mul_mont_fixed_n" . $self->{n};
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_label($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $l) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return "L" . $l . "_" . $self->{n};
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_labels($@)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, @labels) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my %out = ();
|
|
|
727bdf |
-
|
|
|
727bdf |
- foreach my $l (@labels) {
|
|
|
727bdf |
- $out{"$l"} = $self->get_label("$l");
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- return \%out;
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub nl($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code("\n");
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub copy_result($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my ($n) = $self->{n};
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 0; $j < $n; $j++) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- std $tp[$j],`$j*$SIZE_T`($rp)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub mul_mont_fixed($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my ($n) = $self->{n};
|
|
|
727bdf |
- my $fname = $self->get_function_name();
|
|
|
727bdf |
- my $label = $self->get_labels("outer", "enter", "sub", "copy", "end");
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
-
|
|
|
727bdf |
-.globl .${fname}
|
|
|
727bdf |
-.align 5
|
|
|
727bdf |
-.${fname}:
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->save_registers();
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $n0,0($n0)
|
|
|
727bdf |
-
|
|
|
727bdf |
- ld $bp0,0($bp)
|
|
|
727bdf |
-
|
|
|
727bdf |
- ld $apj,0($ap)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->mul_c_0($tp[0], $apj, $bp0, $c0);
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 1; $j < $n - 1; $j++) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $apj,`$j*$SIZE_T`($ap)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- $self->mul($tp[$j], $apj, $bp0, $c0);
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $apj,`($n-1)*$SIZE_T`($ap)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->mul_last($tp[$n-1], $tp[$n], $apj, $bp0, $c0);
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- li $tp[$n+1],0
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- li $i,0
|
|
|
727bdf |
- mtctr $num
|
|
|
727bdf |
- b $label->{"enter"}
|
|
|
727bdf |
-
|
|
|
727bdf |
-.align 4
|
|
|
727bdf |
-$label->{"outer"}:
|
|
|
727bdf |
- ldx $bpi,$bp,$i
|
|
|
727bdf |
-
|
|
|
727bdf |
- ld $apj,0($ap)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->mul_add_c_0($tp[0], $tp[0], $apj, $bpi, $c0);
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 1; $j < $n; $j++) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $apj,`$j*$SIZE_T`($ap)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- $self->mul_add($tp[$j], $tp[$j], $apj, $bpi, $c0);
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- addc $tp[$n],$tp[$n],$c0
|
|
|
727bdf |
- addze $tp[$n+1],$tp[$n+1]
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
-.align 4
|
|
|
727bdf |
-$label->{"enter"}:
|
|
|
727bdf |
- mulld $bpi,$tp[0],$n0
|
|
|
727bdf |
-
|
|
|
727bdf |
- ld $npj,0($np)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->mul_add_c_0($lo, $tp[0], $bpi, $npj, $c0);
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 1; $j < $n; $j++) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $npj,`$j*$SIZE_T`($np)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- $self->mul_add($tp[$j-1], $tp[$j], $npj, $bpi, $c0);
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- addc $tp[$n-1],$tp[$n],$c0
|
|
|
727bdf |
- addze $tp[$n],$tp[$n+1]
|
|
|
727bdf |
-
|
|
|
727bdf |
- addi $i,$i,$SIZE_T
|
|
|
727bdf |
- bdnz $label->{"outer"}
|
|
|
727bdf |
-
|
|
|
727bdf |
- and. $tp[$n],$tp[$n],$tp[$n]
|
|
|
727bdf |
- bne $label->{"sub"}
|
|
|
727bdf |
-
|
|
|
727bdf |
- cmpld $tp[$n-1],$npj
|
|
|
727bdf |
- blt $label->{"copy"}
|
|
|
727bdf |
-
|
|
|
727bdf |
-$label->{"sub"}:
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- #
|
|
|
727bdf |
- # Reduction
|
|
|
727bdf |
- #
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $bpj,`0*$SIZE_T`($np)
|
|
|
727bdf |
- subfc $c1,$bpj,$tp[0]
|
|
|
727bdf |
- std $c1,`0*$SIZE_T`($rp)
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
- for (my $j = 1; $j < $n - 1; $j++) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $bpj,`$j*$SIZE_T`($np)
|
|
|
727bdf |
- subfe $c1,$bpj,$tp[$j]
|
|
|
727bdf |
- std $c1,`$j*$SIZE_T`($rp)
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- subfe $c1,$npj,$tp[$n-1]
|
|
|
727bdf |
- std $c1,`($n-1)*$SIZE_T`($rp)
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- addme. $tp[$n],$tp[$n]
|
|
|
727bdf |
- beq $label->{"end"}
|
|
|
727bdf |
-
|
|
|
727bdf |
-$label->{"copy"}:
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->copy_result();
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
-
|
|
|
727bdf |
-$label->{"end"}:
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->restore_registers();
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- li r3,1
|
|
|
727bdf |
- blr
|
|
|
727bdf |
-.size .${fname},.-.${fname}
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-package Mont::GPR;
|
|
|
727bdf |
-
|
|
|
727bdf |
-our @ISA = ('Mont');
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub new($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($class, $n) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return $class->SUPER::new($n);
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub save_registers($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my $n = $self->{n};
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- std $lo,-8($sp)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 0; $j <= $n+1; $j++) {
|
|
|
727bdf |
- $self->{code}.=<<___;
|
|
|
727bdf |
- std $tp[$j],-`($j+2)*8`($sp)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub restore_registers($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my $n = $self->{n};
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- ld $lo,-8($sp)
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- for (my $j = 0; $j <= $n+1; $j++) {
|
|
|
727bdf |
- $self->{code}.=<<___;
|
|
|
727bdf |
- ld $tp[$j],-`($j+2)*8`($sp)
|
|
|
727bdf |
-___
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->{code} .=<<___;
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Direct translation of C mul()
|
|
|
727bdf |
-sub mul($$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $lo,$a,$w
|
|
|
727bdf |
- addc $r,$lo,$c
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
- addze $c,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like mul() but $c is ignored as an input - an optimisation to save a
|
|
|
727bdf |
-# preliminary instruction that would set input $c to 0
|
|
|
727bdf |
-sub mul_c_0($$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $r,$a,$w
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like mul() but does not to the final addition of CA into $c - an
|
|
|
727bdf |
-# optimisation to save an instruction
|
|
|
727bdf |
-sub mul_last($$$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r1, $r2, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $lo,$a,$w
|
|
|
727bdf |
- addc $r1,$lo,$c
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
-
|
|
|
727bdf |
- addze $r2,$c
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like C mul_add() but allow $r_out and $r_in to be different
|
|
|
727bdf |
-sub mul_add($$$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r_out, $r_in, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $lo,$a,$w
|
|
|
727bdf |
- addc $lo,$lo,$c
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
- addze $c,$c
|
|
|
727bdf |
- addc $r_out,$r_in,$lo
|
|
|
727bdf |
- addze $c,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like mul_add() but $c is ignored as an input - an optimisation to save a
|
|
|
727bdf |
-# preliminary instruction that would set input $c to 0
|
|
|
727bdf |
-sub mul_add_c_0($$$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r_out, $r_in, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $lo,$a,$w
|
|
|
727bdf |
- addc $r_out,$r_in,$lo
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
- addze $c,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-package Mont::GPR_300;
|
|
|
727bdf |
-
|
|
|
727bdf |
-our @ISA = ('Mont::GPR');
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub new($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($class, $n) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- my $mont = $class->SUPER::new($n);
|
|
|
727bdf |
-
|
|
|
727bdf |
- return $mont;
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_function_name($)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return "bn_mul_mont_300_fixed_n" . $self->{n};
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-sub get_label($$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $l) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- return "L" . $l . "_300_" . $self->{n};
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Direct translation of C mul()
|
|
|
727bdf |
-sub mul($$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r, $a, $w, $c, $last) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- maddld $r,$a,$w,$c
|
|
|
727bdf |
- maddhdu $c,$a,$w,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Save the last carry as the final entry
|
|
|
727bdf |
-sub mul_last($$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r1, $r2, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- maddld $r1,$a,$w,$c
|
|
|
727bdf |
- maddhdu $r2,$a,$w,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like mul() but $c is ignored as an input - an optimisation to save a
|
|
|
727bdf |
-# preliminary instruction that would set input $c to 0
|
|
|
727bdf |
-sub mul_c_0($$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mulld $r,$a,$w
|
|
|
727bdf |
- mulhdu $c,$a,$w
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like C mul_add() but allow $r_out and $r_in to be different
|
|
|
727bdf |
-sub mul_add($$$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r_out, $r_in, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- maddld $lo,$a,$w,$c
|
|
|
727bdf |
- maddhdu $c,$a,$w,$c
|
|
|
727bdf |
- addc $r_out,$r_in,$lo
|
|
|
727bdf |
- addze $c,$c
|
|
|
727bdf |
-
|
|
|
727bdf |
-___
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-# Like mul_add() but $c is ignored as an input - an optimisation to save a
|
|
|
727bdf |
-# preliminary instruction that would set input $c to 0
|
|
|
727bdf |
-sub mul_add_c_0($$$$$$)
|
|
|
727bdf |
-{
|
|
|
727bdf |
- my ($self, $r_out, $r_in, $a, $w, $c) = @_;
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- maddld $lo,$a,$w,$r_in
|
|
|
727bdf |
- maddhdu $c,$a,$w,$r_in
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
- if ($r_out ne $lo) {
|
|
|
727bdf |
- $self->add_code(<<___);
|
|
|
727bdf |
- mr $r_out,$lo
|
|
|
727bdf |
-___
|
|
|
727bdf |
- }
|
|
|
727bdf |
-
|
|
|
727bdf |
- $self->nl();
|
|
|
727bdf |
-}
|
|
|
727bdf |
-
|
|
|
727bdf |
-
|
|
|
727bdf |
-package main;
|
|
|
727bdf |
-
|
|
|
727bdf |
-my $code;
|
|
|
727bdf |
-
|
|
|
727bdf |
-$code.=<<___;
|
|
|
727bdf |
-.machine "any"
|
|
|
727bdf |
-.text
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
-my $mont;
|
|
|
727bdf |
-
|
|
|
727bdf |
-$mont = new Mont::GPR(6);
|
|
|
727bdf |
-$mont->mul_mont_fixed();
|
|
|
727bdf |
-$code .= $mont->get_code();
|
|
|
727bdf |
-
|
|
|
727bdf |
-$mont = new Mont::GPR_300(6);
|
|
|
727bdf |
-$mont->mul_mont_fixed();
|
|
|
727bdf |
-$code .= $mont->get_code();
|
|
|
727bdf |
-
|
|
|
727bdf |
-$code =~ s/\`([^\`]*)\`/eval $1/gem;
|
|
|
727bdf |
-
|
|
|
727bdf |
-$code.=<<___;
|
|
|
727bdf |
-.asciz "Montgomery Multiplication for PPC by <amitay\@ozlabs.org>, <alastair\@d-silva.org>"
|
|
|
727bdf |
-___
|
|
|
727bdf |
-
|
|
|
727bdf |
-print $code;
|
|
|
727bdf |
-close STDOUT or die "error closing STDOUT: $!";
|
|
|
727bdf |
diff --git a/crypto/bn/bn_ppc.c b/crypto/bn/bn_ppc.c
|
|
|
727bdf |
index 1e9421bee213..3ee76ea96574 100644
|
|
|
727bdf |
--- a/crypto/bn/bn_ppc.c
|
|
|
727bdf |
+++ b/crypto/bn/bn_ppc.c
|
|
|
727bdf |
@@ -19,12 +19,6 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
|
|
|
727bdf |
const BN_ULONG *np, const BN_ULONG *n0, int num);
|
|
|
727bdf |
int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
|
|
|
727bdf |
const BN_ULONG *np, const BN_ULONG *n0, int num);
|
|
|
727bdf |
- int bn_mul_mont_fixed_n6(BN_ULONG *rp, const BN_ULONG *ap,
|
|
|
727bdf |
- const BN_ULONG *bp, const BN_ULONG *np,
|
|
|
727bdf |
- const BN_ULONG *n0, int num);
|
|
|
727bdf |
- int bn_mul_mont_300_fixed_n6(BN_ULONG *rp, const BN_ULONG *ap,
|
|
|
727bdf |
- const BN_ULONG *bp, const BN_ULONG *np,
|
|
|
727bdf |
- const BN_ULONG *n0, int num);
|
|
|
727bdf |
|
|
|
727bdf |
if (num < 4)
|
|
|
727bdf |
return 0;
|
|
|
727bdf |
@@ -40,14 +34,5 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
|
|
|
727bdf |
* no opportunity to figure it out...
|
|
|
727bdf |
*/
|
|
|
727bdf |
|
|
|
727bdf |
-#if defined(_ARCH_PPC64)
|
|
|
727bdf |
- if (num == 6) {
|
|
|
727bdf |
- if (OPENSSL_ppccap_P & PPC_MADD300)
|
|
|
727bdf |
- return bn_mul_mont_300_fixed_n6(rp, ap, bp, np, n0, num);
|
|
|
727bdf |
- else
|
|
|
727bdf |
- return bn_mul_mont_fixed_n6(rp, ap, bp, np, n0, num);
|
|
|
727bdf |
- }
|
|
|
727bdf |
-#endif
|
|
|
727bdf |
-
|
|
|
727bdf |
return bn_mul_mont_int(rp, ap, bp, np, n0, num);
|
|
|
727bdf |
}
|
|
|
727bdf |
diff --git a/crypto/bn/build.info b/crypto/bn/build.info
|
|
|
727bdf |
index 987a70ae263b..4f8d0689b5ea 100644
|
|
|
727bdf |
--- a/crypto/bn/build.info
|
|
|
727bdf |
+++ b/crypto/bn/build.info
|
|
|
727bdf |
@@ -79,7 +79,7 @@ IF[{- !$disabled{asm} -}]
|
|
|
727bdf |
|
|
|
727bdf |
$BNASM_ppc32=bn_ppc.c bn-ppc.s ppc-mont.s
|
|
|
727bdf |
$BNDEF_ppc32=OPENSSL_BN_ASM_MONT
|
|
|
727bdf |
- $BNASM_ppc64=$BNASM_ppc32 ppc64-mont-fixed.s
|
|
|
727bdf |
+ $BNASM_ppc64=$BNASM_ppc32
|
|
|
727bdf |
$BNDEF_ppc64=$BNDEF_ppc32
|
|
|
727bdf |
|
|
|
727bdf |
$BNASM_c64xplus=asm/bn-c64xplus.asm
|
|
|
727bdf |
@@ -173,7 +173,6 @@ GENERATE[parisc-mont.s]=asm/parisc-mont.pl
|
|
|
727bdf |
GENERATE[bn-ppc.s]=asm/ppc.pl
|
|
|
727bdf |
GENERATE[ppc-mont.s]=asm/ppc-mont.pl
|
|
|
727bdf |
GENERATE[ppc64-mont.s]=asm/ppc64-mont.pl
|
|
|
727bdf |
-GENERATE[ppc64-mont-fixed.s]=asm/ppc64-mont-fixed.pl
|
|
|
727bdf |
|
|
|
727bdf |
GENERATE[alpha-mont.S]=asm/alpha-mont.pl
|
|
|
727bdf |
|
|
|
727bdf |
diff --git a/test/recipes/30-test_evp_data/evppkey_ecdsa.txt b/test/recipes/30-test_evp_data/evppkey_ecdsa.txt
|
|
|
727bdf |
index f36982845db4..1543ed9f7534 100644
|
|
|
727bdf |
--- a/test/recipes/30-test_evp_data/evppkey_ecdsa.txt
|
|
|
727bdf |
+++ b/test/recipes/30-test_evp_data/evppkey_ecdsa.txt
|
|
|
727bdf |
@@ -97,6 +97,18 @@ Key = P-256-PUBLIC
|
|
|
727bdf |
Input = "Hello World"
|
|
|
727bdf |
Output = 3046022100e7515177ec3817b77a4a94066ab3070817b7aa9d44a8a09f040da250116e8972022100ba59b0f631258e59a9026be5d84f60685f4cf22b9165a0c2736d5c21c8ec1862
|
|
|
727bdf |
|
|
|
727bdf |
+PublicKey=P-384-PUBLIC
|
|
|
727bdf |
+-----BEGIN PUBLIC KEY-----
|
|
|
727bdf |
+MHYwEAYHKoZIzj0CAQYFK4EEACIDYgAES/TlL5WEJ+u1kV+4yVlVUbTTo/2rZ7rd
|
|
|
727bdf |
+nWwwk/QlukNjDfcfQvDrfOqpTZ9kSKhd0wMxWIJJ/S/cCzCex+2EgbwW8ngAwT19
|
|
|
727bdf |
+twD8guGxyFRaoMDTtW47/nifwYqRaIfC
|
|
|
727bdf |
+-----END PUBLIC KEY-----
|
|
|
727bdf |
+
|
|
|
727bdf |
+DigestVerify = SHA384
|
|
|
727bdf |
+Key = P-384-PUBLIC
|
|
|
727bdf |
+Input = "123400"
|
|
|
727bdf |
+Output = 304d0218389cb27e0bc8d21fa7e5f24cb74f58851313e696333ad68b023100ffffffffffffffffffffffffffffffffffffffffffffffffc7634d81f4372ddf581a0db248b0a77aecec196accc52970
|
|
|
727bdf |
+
|
|
|
727bdf |
# Oneshot tests
|
|
|
727bdf |
OneShotDigestVerify = SHA256
|
|
|
727bdf |
Key = P-256-PUBLIC
|