Backport of ppc64le Montgomery multiply enhancement

Resolves: rhbz#2130708
epel8
Dmitry Belyavskiy 2 years ago
parent c29e183891
commit 657265459d

@ -0,0 +1,703 @@
From 33ffd36afa7594aeb958a925f521cb287ca850c8 Mon Sep 17 00:00:00 2001
From: Rohan McLure <rohanmclure@linux.ibm.com>
Date: Mon, 27 Jun 2022 12:14:55 +1000
Subject: [PATCH 1/2] Revert "Revert "bn: Add fixed length (n=6), unrolled PPC
Montgomery Multiplication""
This reverts commit 712d9cc90e355b2c98a959d4e9398610d2269c9e.
---
crypto/bn/asm/ppc64-mont-fixed.pl | 581 ++++++++++++++++++++++++++++++
crypto/bn/bn_ppc.c | 15 +
crypto/bn/build.info | 3 +-
3 files changed, 598 insertions(+), 1 deletion(-)
diff --git a/crypto/bn/asm/ppc64-mont-fixed.pl b/crypto/bn/asm/ppc64-mont-fixed.pl
index e69de29bb2d1..0fb397bc5f12 100755
--- a/crypto/bn/asm/ppc64-mont-fixed.pl
+++ b/crypto/bn/asm/ppc64-mont-fixed.pl
@@ -0,0 +1,581 @@
+#! /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
+
+# ====================================================================
+# Written by Amitay Isaacs <amitay@ozlabs.org>, Martin Schwenke
+# <martin@meltin.net> & Alastair D'Silva <alastair@d-silva.org> for
+# the OpenSSL project.
+# ====================================================================
+
+#
+# Fixed length (n=6), unrolled PPC Montgomery Multiplication
+#
+
+# 2021
+#
+# Although this is a generic implementation for unrolling Montgomery
+# Multiplication for arbitrary values of n, this is currently only
+# used for n = 6 to improve the performance of ECC p384.
+#
+# Unrolling allows intermediate results to be stored in registers,
+# rather than on the stack, improving performance by ~7% compared to
+# the existing PPC assembly code.
+#
+# The ISA 3.0 implementation uses combination multiply/add
+# instructions (maddld, maddhdu) to improve performance by an
+# additional ~10% on Power 9.
+#
+# Finally, saving non-volatile registers into volatile vector
+# registers instead of onto the stack saves a little more.
+#
+# On a Power 9 machine we see an overall improvement of ~18%.
+#
+
+use strict;
+use warnings;
+
+my ($flavour, $output, $dir, $xlate);
+
+# $output is the last argument if it looks like a file (it has an extension)
+# $flavour is the first argument if it doesn't look like a file
+$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
+$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
+die "can't locate ppc-xlate.pl";
+
+open STDOUT,"| $^X $xlate $flavour \"$output\""
+ or die "can't call $xlate: $!";
+
+if ($flavour !~ /64/) {
+ die "bad flavour ($flavour) - only ppc64 permitted";
+}
+
+my $SIZE_T= 8;
+
+# Registers are global so the code is remotely readable
+
+# Parameters for Montgomery multiplication
+my $sp = "r1";
+my $toc = "r2";
+my $rp = "r3";
+my $ap = "r4";
+my $bp = "r5";
+my $np = "r6";
+my $n0 = "r7";
+my $num = "r8";
+
+my $i = "r9";
+my $c0 = "r10";
+my $bp0 = "r11";
+my $bpi = "r11";
+my $bpj = "r11";
+my $tj = "r12";
+my $apj = "r12";
+my $npj = "r12";
+my $lo = "r14";
+my $c1 = "r14";
+
+# Non-volatile registers used for tp[i]
+#
+# 12 registers are available but the limit on unrolling is 10,
+# since registers from $tp[0] to $tp[$n+1] are used.
+my @tp = ("r20" .. "r31");
+
+# volatile VSRs for saving non-volatile GPRs - faster than stack
+my @vsrs = ("v32" .. "v46");
+
+package Mont;
+
+sub new($$)
+{
+ my ($class, $n) = @_;
+
+ if ($n > 10) {
+ die "Can't unroll for BN length ${n} (maximum 10)"
+ }
+
+ my $self = {
+ code => "",
+ n => $n,
+ };
+ bless $self, $class;
+
+ return $self;
+}
+
+sub add_code($$)
+{
+ my ($self, $c) = @_;
+
+ $self->{code} .= $c;
+}
+
+sub get_code($)
+{
+ my ($self) = @_;
+
+ return $self->{code};
+}
+
+sub get_function_name($)
+{
+ my ($self) = @_;
+
+ return "bn_mul_mont_fixed_n" . $self->{n};
+}
+
+sub get_label($$)
+{
+ my ($self, $l) = @_;
+
+ return "L" . $l . "_" . $self->{n};
+}
+
+sub get_labels($@)
+{
+ my ($self, @labels) = @_;
+
+ my %out = ();
+
+ foreach my $l (@labels) {
+ $out{"$l"} = $self->get_label("$l");
+ }
+
+ return \%out;
+}
+
+sub nl($)
+{
+ my ($self) = @_;
+
+ $self->add_code("\n");
+}
+
+sub copy_result($)
+{
+ my ($self) = @_;
+
+ my ($n) = $self->{n};
+
+ for (my $j = 0; $j < $n; $j++) {
+ $self->add_code(<<___);
+ std $tp[$j],`$j*$SIZE_T`($rp)
+___
+ }
+
+}
+
+sub mul_mont_fixed($)
+{
+ my ($self) = @_;
+
+ my ($n) = $self->{n};
+ my $fname = $self->get_function_name();
+ my $label = $self->get_labels("outer", "enter", "sub", "copy", "end");
+
+ $self->add_code(<<___);
+
+.globl .${fname}
+.align 5
+.${fname}:
+
+___
+
+ $self->save_registers();
+
+ $self->add_code(<<___);
+ ld $n0,0($n0)
+
+ ld $bp0,0($bp)
+
+ ld $apj,0($ap)
+___
+
+ $self->mul_c_0($tp[0], $apj, $bp0, $c0);
+
+ for (my $j = 1; $j < $n - 1; $j++) {
+ $self->add_code(<<___);
+ ld $apj,`$j*$SIZE_T`($ap)
+___
+ $self->mul($tp[$j], $apj, $bp0, $c0);
+ }
+
+ $self->add_code(<<___);
+ ld $apj,`($n-1)*$SIZE_T`($ap)
+___
+
+ $self->mul_last($tp[$n-1], $tp[$n], $apj, $bp0, $c0);
+
+ $self->add_code(<<___);
+ li $tp[$n+1],0
+
+___
+
+ $self->add_code(<<___);
+ li $i,0
+ mtctr $num
+ b $label->{"enter"}
+
+.align 4
+$label->{"outer"}:
+ ldx $bpi,$bp,$i
+
+ ld $apj,0($ap)
+___
+
+ $self->mul_add_c_0($tp[0], $tp[0], $apj, $bpi, $c0);
+
+ for (my $j = 1; $j < $n; $j++) {
+ $self->add_code(<<___);
+ ld $apj,`$j*$SIZE_T`($ap)
+___
+ $self->mul_add($tp[$j], $tp[$j], $apj, $bpi, $c0);
+ }
+
+ $self->add_code(<<___);
+ addc $tp[$n],$tp[$n],$c0
+ addze $tp[$n+1],$tp[$n+1]
+___
+
+ $self->add_code(<<___);
+.align 4
+$label->{"enter"}:
+ mulld $bpi,$tp[0],$n0
+
+ ld $npj,0($np)
+___
+
+ $self->mul_add_c_0($lo, $tp[0], $bpi, $npj, $c0);
+
+ for (my $j = 1; $j < $n; $j++) {
+ $self->add_code(<<___);
+ ld $npj,`$j*$SIZE_T`($np)
+___
+ $self->mul_add($tp[$j-1], $tp[$j], $npj, $bpi, $c0);
+ }
+
+ $self->add_code(<<___);
+ addc $tp[$n-1],$tp[$n],$c0
+ addze $tp[$n],$tp[$n+1]
+
+ addi $i,$i,$SIZE_T
+ bdnz $label->{"outer"}
+
+ and. $tp[$n],$tp[$n],$tp[$n]
+ bne $label->{"sub"}
+
+ cmpld $tp[$n-1],$npj
+ blt $label->{"copy"}
+
+$label->{"sub"}:
+___
+
+ #
+ # Reduction
+ #
+
+ $self->add_code(<<___);
+ ld $bpj,`0*$SIZE_T`($np)
+ subfc $c1,$bpj,$tp[0]
+ std $c1,`0*$SIZE_T`($rp)
+
+___
+ for (my $j = 1; $j < $n - 1; $j++) {
+ $self->add_code(<<___);
+ ld $bpj,`$j*$SIZE_T`($np)
+ subfe $c1,$bpj,$tp[$j]
+ std $c1,`$j*$SIZE_T`($rp)
+
+___
+ }
+
+ $self->add_code(<<___);
+ subfe $c1,$npj,$tp[$n-1]
+ std $c1,`($n-1)*$SIZE_T`($rp)
+
+___
+
+ $self->add_code(<<___);
+ addme. $tp[$n],$tp[$n]
+ beq $label->{"end"}
+
+$label->{"copy"}:
+___
+
+ $self->copy_result();
+
+ $self->add_code(<<___);
+
+$label->{"end"}:
+___
+
+ $self->restore_registers();
+
+ $self->add_code(<<___);
+ li r3,1
+ blr
+.size .${fname},.-.${fname}
+___
+
+}
+
+package Mont::GPR;
+
+our @ISA = ('Mont');
+
+sub new($$)
+{
+ my ($class, $n) = @_;
+
+ return $class->SUPER::new($n);
+}
+
+sub save_registers($)
+{
+ my ($self) = @_;
+
+ my $n = $self->{n};
+
+ $self->add_code(<<___);
+ std $lo,-8($sp)
+___
+
+ for (my $j = 0; $j <= $n+1; $j++) {
+ $self->{code}.=<<___;
+ std $tp[$j],-`($j+2)*8`($sp)
+___
+ }
+
+ $self->add_code(<<___);
+
+___
+}
+
+sub restore_registers($)
+{
+ my ($self) = @_;
+
+ my $n = $self->{n};
+
+ $self->add_code(<<___);
+ ld $lo,-8($sp)
+___
+
+ for (my $j = 0; $j <= $n+1; $j++) {
+ $self->{code}.=<<___;
+ ld $tp[$j],-`($j+2)*8`($sp)
+___
+ }
+
+ $self->{code} .=<<___;
+
+___
+}
+
+# Direct translation of C mul()
+sub mul($$$$$)
+{
+ my ($self, $r, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $lo,$a,$w
+ addc $r,$lo,$c
+ mulhdu $c,$a,$w
+ addze $c,$c
+
+___
+}
+
+# Like mul() but $c is ignored as an input - an optimisation to save a
+# preliminary instruction that would set input $c to 0
+sub mul_c_0($$$$$)
+{
+ my ($self, $r, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $r,$a,$w
+ mulhdu $c,$a,$w
+
+___
+}
+
+# Like mul() but does not to the final addition of CA into $c - an
+# optimisation to save an instruction
+sub mul_last($$$$$$)
+{
+ my ($self, $r1, $r2, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $lo,$a,$w
+ addc $r1,$lo,$c
+ mulhdu $c,$a,$w
+
+ addze $r2,$c
+___
+}
+
+# Like C mul_add() but allow $r_out and $r_in to be different
+sub mul_add($$$$$$)
+{
+ my ($self, $r_out, $r_in, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $lo,$a,$w
+ addc $lo,$lo,$c
+ mulhdu $c,$a,$w
+ addze $c,$c
+ addc $r_out,$r_in,$lo
+ addze $c,$c
+
+___
+}
+
+# Like mul_add() but $c is ignored as an input - an optimisation to save a
+# preliminary instruction that would set input $c to 0
+sub mul_add_c_0($$$$$$)
+{
+ my ($self, $r_out, $r_in, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $lo,$a,$w
+ addc $r_out,$r_in,$lo
+ mulhdu $c,$a,$w
+ addze $c,$c
+
+___
+}
+
+package Mont::GPR_300;
+
+our @ISA = ('Mont::GPR');
+
+sub new($$)
+{
+ my ($class, $n) = @_;
+
+ my $mont = $class->SUPER::new($n);
+
+ return $mont;
+}
+
+sub get_function_name($)
+{
+ my ($self) = @_;
+
+ return "bn_mul_mont_300_fixed_n" . $self->{n};
+}
+
+sub get_label($$)
+{
+ my ($self, $l) = @_;
+
+ return "L" . $l . "_300_" . $self->{n};
+}
+
+# Direct translation of C mul()
+sub mul($$$$$)
+{
+ my ($self, $r, $a, $w, $c, $last) = @_;
+
+ $self->add_code(<<___);
+ maddld $r,$a,$w,$c
+ maddhdu $c,$a,$w,$c
+
+___
+}
+
+# Save the last carry as the final entry
+sub mul_last($$$$$)
+{
+ my ($self, $r1, $r2, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ maddld $r1,$a,$w,$c
+ maddhdu $r2,$a,$w,$c
+
+___
+}
+
+# Like mul() but $c is ignored as an input - an optimisation to save a
+# preliminary instruction that would set input $c to 0
+sub mul_c_0($$$$$)
+{
+ my ($self, $r, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ mulld $r,$a,$w
+ mulhdu $c,$a,$w
+
+___
+}
+
+# Like C mul_add() but allow $r_out and $r_in to be different
+sub mul_add($$$$$$)
+{
+ my ($self, $r_out, $r_in, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ maddld $lo,$a,$w,$c
+ maddhdu $c,$a,$w,$c
+ addc $r_out,$r_in,$lo
+ addze $c,$c
+
+___
+}
+
+# Like mul_add() but $c is ignored as an input - an optimisation to save a
+# preliminary instruction that would set input $c to 0
+sub mul_add_c_0($$$$$$)
+{
+ my ($self, $r_out, $r_in, $a, $w, $c) = @_;
+
+ $self->add_code(<<___);
+ maddld $lo,$a,$w,$r_in
+ maddhdu $c,$a,$w,$r_in
+___
+
+ if ($r_out ne $lo) {
+ $self->add_code(<<___);
+ mr $r_out,$lo
+___
+ }
+
+ $self->nl();
+}
+
+
+package main;
+
+my $code;
+
+$code.=<<___;
+.machine "any"
+.text
+___
+
+my $mont;
+
+$mont = new Mont::GPR(6);
+$mont->mul_mont_fixed();
+$code .= $mont->get_code();
+
+$mont = new Mont::GPR_300(6);
+$mont->mul_mont_fixed();
+$code .= $mont->get_code();
+
+$code =~ s/\`([^\`]*)\`/eval $1/gem;
+
+$code.=<<___;
+.asciz "Montgomery Multiplication for PPC by <amitay\@ozlabs.org>, <alastair\@d-silva.org>"
+___
+
+print $code;
+close STDOUT or die "error closing STDOUT: $!";
diff --git a/crypto/bn/bn_ppc.c b/crypto/bn/bn_ppc.c
index 3ee76ea96574..1e9421bee213 100644
--- a/crypto/bn/bn_ppc.c
+++ b/crypto/bn/bn_ppc.c
@@ -19,6 +19,12 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
const BN_ULONG *np, const BN_ULONG *n0, int num);
int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
const BN_ULONG *np, const BN_ULONG *n0, int num);
+ int bn_mul_mont_fixed_n6(BN_ULONG *rp, const BN_ULONG *ap,
+ const BN_ULONG *bp, const BN_ULONG *np,
+ const BN_ULONG *n0, int num);
+ int bn_mul_mont_300_fixed_n6(BN_ULONG *rp, const BN_ULONG *ap,
+ const BN_ULONG *bp, const BN_ULONG *np,
+ const BN_ULONG *n0, int num);
if (num < 4)
return 0;
@@ -34,5 +40,14 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
* no opportunity to figure it out...
*/
+#if defined(_ARCH_PPC64) && !defined(__ILP32__)
+ if (num == 6) {
+ if (OPENSSL_ppccap_P & PPC_MADD300)
+ return bn_mul_mont_300_fixed_n6(rp, ap, bp, np, n0, num);
+ else
+ return bn_mul_mont_fixed_n6(rp, ap, bp, np, n0, num);
+ }
+#endif
+
return bn_mul_mont_int(rp, ap, bp, np, n0, num);
}
diff --git a/crypto/bn/build.info b/crypto/bn/build.info
index 4f8d0689b5ea..987a70ae263b 100644
--- a/crypto/bn/build.info
+++ b/crypto/bn/build.info
@@ -79,7 +79,7 @@ IF[{- !$disabled{asm} -}]
$BNASM_ppc32=bn_ppc.c bn-ppc.s ppc-mont.s
$BNDEF_ppc32=OPENSSL_BN_ASM_MONT
- $BNASM_ppc64=$BNASM_ppc32
+ $BNASM_ppc64=$BNASM_ppc32 ppc64-mont-fixed.s
$BNDEF_ppc64=$BNDEF_ppc32
$BNASM_c64xplus=asm/bn-c64xplus.asm
@@ -173,6 +173,7 @@ GENERATE[parisc-mont.s]=asm/parisc-mont.pl
GENERATE[bn-ppc.s]=asm/ppc.pl
GENERATE[ppc-mont.s]=asm/ppc-mont.pl
GENERATE[ppc64-mont.s]=asm/ppc64-mont.pl
+GENERATE[ppc64-mont-fixed.s]=asm/ppc64-mont-fixed.pl
GENERATE[alpha-mont.S]=asm/alpha-mont.pl
From 01ebad0d6e3a09bc9e32350b402901471610a3dc Mon Sep 17 00:00:00 2001
From: Rohan McLure <rohanmclure@linux.ibm.com>
Date: Thu, 30 Jun 2022 16:21:06 +1000
Subject: [PATCH 2/2] Fix unrolled montgomery multiplication for POWER9
In the reference C implementation in bn_asm.c, tp[num + 1] contains the
carry bit for accumulations into tp[num]. tp[num + 1] is only ever
assigned, never itself incremented.
---
crypto/bn/asm/ppc64-mont-fixed.pl | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/crypto/bn/asm/ppc64-mont-fixed.pl b/crypto/bn/asm/ppc64-mont-fixed.pl
index 0fb397bc5f12..e27d0ad93d85 100755
--- a/crypto/bn/asm/ppc64-mont-fixed.pl
+++ b/crypto/bn/asm/ppc64-mont-fixed.pl
@@ -63,6 +63,7 @@
# Registers are global so the code is remotely readable
# Parameters for Montgomery multiplication
+my $ze = "r0";
my $sp = "r1";
my $toc = "r2";
my $rp = "r3";
@@ -192,6 +193,7 @@ ($)
$self->save_registers();
$self->add_code(<<___);
+ li $ze,0
ld $n0,0($n0)
ld $bp0,0($bp)
@@ -242,7 +244,7 @@ ($)
$self->add_code(<<___);
addc $tp[$n],$tp[$n],$c0
- addze $tp[$n+1],$tp[$n+1]
+ addze $tp[$n+1],$ze
___
$self->add_code(<<___);
@@ -272,7 +274,7 @@ ($)
and. $tp[$n],$tp[$n],$tp[$n]
bne $label->{"sub"}
- cmpld $tp[$n-1],$npj
+ cmpld $tp[$n-1],$npj
blt $label->{"copy"}
$label->{"sub"}:

@ -109,6 +109,9 @@ Patch60: 0060-FIPS-KAT-signature-tests.patch
# https://bugzilla.redhat.com/show_bug.cgi?id=2087147 # https://bugzilla.redhat.com/show_bug.cgi?id=2087147
Patch61: 0061-Deny-SHA-1-signature-verification-in-FIPS-provider.patch Patch61: 0061-Deny-SHA-1-signature-verification-in-FIPS-provider.patch
Patch62: 0062-fips-Expose-a-FIPS-indicator.patch Patch62: 0062-fips-Expose-a-FIPS-indicator.patch
# https://bugzilla.redhat.com/show_bug.cgi?id=2130708
# https://github.com/openssl/openssl/pull/18883
Patch67: 0067-ppc64le-Montgomery-multiply.patch
# https://github.com/openssl/openssl/commit/44a563dde1584cd9284e80b6e45ee5019be8d36c # https://github.com/openssl/openssl/commit/44a563dde1584cd9284e80b6e45ee5019be8d36c
# https://github.com/openssl/openssl/commit/345c99b6654b8313c792d54f829943068911ddbd # https://github.com/openssl/openssl/commit/345c99b6654b8313c792d54f829943068911ddbd
Patch71: 0071-AES-GCM-performance-optimization.patch Patch71: 0071-AES-GCM-performance-optimization.patch
@ -488,6 +491,8 @@ install -m644 %{SOURCE9} \
Resolves: rhbz#2093804 Resolves: rhbz#2093804
- Adjusting include for the FIPS_mode macro - Adjusting include for the FIPS_mode macro
Resolves: rhbz#2083879 Resolves: rhbz#2083879
- Backport of ppc64le Montgomery multiply enhancement
Resolves: rhbz#2130708
* Tue Nov 22 2022 Dmitry Belyavskiy <dbelyavs@redhat.com> - 1:3.0.7-1 * Tue Nov 22 2022 Dmitry Belyavskiy <dbelyavs@redhat.com> - 1:3.0.7-1
- Rebasing to OpenSSL 3.0.7 - Rebasing to OpenSSL 3.0.7

Loading…
Cancel
Save