md5: add assembly implementation for loongarch64

This change can improve md5 performance by using a hand-optimized
assembly implementation of the inner loop of md5 calculation.
This implementation refered to md5-x86_64.pl and made more effort
to reorder instructions for separating data dependencies as much
as possible.

Test with:
$ openssl speed md5

3A5000
type             16 bytes    64 bytes     256 bytes    1024 bytes   8192 bytes   16384 bytes
md5              45061.04k   130440.75k   291105.28k   421101.23k   484639.27k   488320.43k
md5-modified     47179.95k   139015.57k   308836.69k   445963.26k   512540.67k   518215.00k
                   +5%         +7%          +6%          +6%          +6%          +6%

3A6000
type             16 bytes    64 bytes     256 bytes    1024 bytes   8192 bytes   16384 bytes
md5              60070.06k   161822.76k   325817.60k   438017.02k   486864.21k   492243.31k
md5-modified     62827.74k   170294.04k   343795.03k   463324.50k   515831.13k   520060.93k
                   +5%         +5%          +6%          +6%          +6%          +6%

Signed-off-by: Min Zhou <zhoumin@loongson.cn>
Co-authored-by: Xi Ruoyao <xry111@xry111.site>

Reviewed-by: Shane Lontis <shane.lontis@oracle.com>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/21704)

crypto/md5/asm/md5-loongarch64.pl

@@ -0,0 +1,297 @@
+#! /usr/bin/env perl
+# Author: Min Zhou <zhoumin@loongson.cn>
+# Copyright 2023 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the OpenSSL license (the "License").  You may not use
+# this file except in compliance with the License.  You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
+# Reference to crypto/md5/asm/md5-x86_64.pl
+# MD5 optimized for LoongArch.
+
+use strict;
+
+my $code;
+
+my ($zero,$ra,$tp,$sp,$fp)=map("\$r$_",(0..3,22));
+my ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$r$_",(4..11));
+my ($t0,$t1,$t2,$t3,$t4,$t5,$t6,$t7,$t8,$x)=map("\$r$_",(12..21));
+
+my $output;
+for (@ARGV) {	$output=$_ if (/\w[\w\-]*\.\w+$/);	}
+open STDOUT,">$output";
+
+# round1_step() does:
+#   dst = x + ((dst + F(x,y,z) + X[k] + T_i) <<< s)
+#   $t1 = y ^ z
+#   $t2 = dst + X[k_next]
+sub round1_step
+{
+    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
+    my $T_i_h = ($T_i & 0xfffff000) >> 12;
+    my $T_i_l = $T_i & 0xfff;
+
+# In LoongArch we have to use two instructions of lu12i.w and ori to load a
+# 32-bit immediate into a general register. Meanwhile, the instruction lu12i.w
+# treats the 20-bit immediate as a signed number. So if the T_i_h is greater
+# than or equal to (1<<19), we need provide lu12i.w a corresponding negative
+# number whose complement equals to the sign extension of T_i_h.
+
+# The details of the instruction lu12i.w can be found as following:
+# https://loongson.github.io/LoongArch-Documentation/LoongArch-Vol1-EN.html#_lu12i_w_lu32i_d_lu52i_d
+
+    $T_i_h = -((1<<32) - (0xfff00000 | $T_i_h)) if ($T_i_h >= (1<<19));
+
+    $code .= "	ld.w	$t0,$a1,0		/* (NEXT STEP) X[0] */\n" if ($pos == -1);
+    $code .= "	xor	$t1,$y,$z		/* y ^ z */\n" if ($pos == -1);
+    $code .= "	add.w	$t2,$dst,$t0		/* dst + X[k] */\n" if ($pos == -1);
+    $code .= <<EOF;
+	lu12i.w	$t8,$T_i_h			/* load bits [31:12] of constant */
+	and     $t1,$x,$t1			/* x & ... */
+	ori	$t8,$t8,$T_i_l			/* load bits [11:0] of constant */
+	xor     $t1,$z,$t1			/* z ^ ... */
+	add.w   $t7,$t2,$t8			/* dst + X[k] + Const */
+	ld.w	$t0,$a1,$k_next*4		/* (NEXT STEP) X[$k_next] */
+	add.w	$dst,$t7,$t1			/* dst += ... */
+	add.w	$t2,$z,$t0			/* (NEXT STEP) dst + X[$k_next] */
+EOF
+
+    $code .= "	rotri.w	$dst,$dst,32-$s		/* dst <<< s */\n";
+    if ($pos != 1) {
+        $code .= "	xor	$t1,$x,$y	/* (NEXT STEP) y ^ z */\n";
+    } else {
+        $code .= "	move	$t0,$a7		/* (NEXT ROUND) $t0 = z' (copy of z) */\n";
+        $code .= "	nor	$t1,$zero,$a7	/* (NEXT ROUND) $t1 = not z' (copy of not z) */\n";
+    }
+    $code .= "	add.w   $dst,$dst,$x		/* dst += x */\n";
+}
+
+# round2_step() does:
+#   dst = x + ((dst + G(x,y,z) + X[k] + T_i) <<< s)
+#   $t0 = z' (copy of z for the next step)
+#   $t1 = not z' (copy of not z for the next step)
+#   $t2 = dst + X[k_next]
+sub round2_step
+{
+    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
+    my $T_i_h = ($T_i & 0xfffff000) >> 12;
+    my $T_i_l = $T_i & 0xfff;
+    $T_i_h = -((1<<32) - (0xfff00000 | $T_i_h)) if ($T_i_h >= (1<<19));
+
+    $code .= <<EOF;
+	lu12i.w	$t8,$T_i_h			/* load bits [31:12] of Constant */
+	and	$t0,$x,$t0			/* x & z */
+	ori	$t8,$t8,$T_i_l			/* load bits [11:0] of Constant */
+	and	$t1,$y,$t1			/* y & (not z) */
+	add.w	$t7,$t2,$t8			/* dst + X[k] + Const */
+	or	$t1,$t0,$t1			/* (y & (not z)) | (x & z) */
+	ld.w	$t0,$a1,$k_next*4		/* (NEXT STEP) X[$k_next] */
+	add.w	$dst,$t7,$t1			/* dst += ... */
+	add.w	$t2,$z,$t0			/* (NEXT STEP) dst + X[$k_next] */
+EOF
+
+    $code .= "	rotri.w $dst,$dst,32-$s		/* dst <<< s */\n";
+    if ($pos != 1) {
+        $code .= "	move	$t0,$y		/* (NEXT STEP) z' = $y */\n";
+        $code .= "	nor	$t1,$zero,$y	/* (NEXT STEP) not z' = not $y */\n";
+    } else {
+        $code .= "	xor	$t1,$a6,$a7	/* (NEXT ROUND) $t1 = y ^ z */\n";
+    }
+    $code .= "	add.w	$dst,$dst,$x		/* dst += x */\n";
+}
+
+# round3_step() does:
+#   dst = x + ((dst + H(x,y,z) + X[k] + T_i) <<< s)
+#   $t1 = y ^ z
+#   $t2 = dst + X[k_next]
+sub round3_step
+{
+    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
+    my $T_i_h = ($T_i & 0xfffff000) >> 12;
+    my $T_i_l = $T_i & 0xfff;
+    $T_i_h = -((1<<32) - (0xfff00000 | $T_i_h)) if ($T_i_h >= (1<<19));
+
+    $code .= <<EOF;
+	lu12i.w	$t8,$T_i_h			/* load bits [31:12] of Constant */
+	xor	$t1,$x,$t1			/* x ^ ... */
+	ori	$t8,$t8,$T_i_l			/* load bits [11:0] of Constant */
+	add.w	$t7,$t2,$t8			/* dst + X[k] + Const */
+	ld.w	$t0,$a1,$k_next*4		/* (NEXT STEP) X[$k_next] */
+	add.w	$dst,$t7,$t1			/* dst += ... */
+	add.w	$t2,$z,$t0			/* (NEXT STEP) dst + X[$k_next] */
+EOF
+
+    $code .= "	rotri.w $dst,$dst,32-$s		/* dst <<< s */\n";
+    if ($pos != 1) {
+	$code .= "	xor	$t1,$x,$y	/* (NEXT STEP) y ^ z */\n";
+    } else {
+        $code .= "	nor	$t1,$zero,$a7	/* (NEXT ROUND) $t1 = not z */\n";
+    }
+    $code .= "	add.w	$dst,$dst,$x		/* dst += x */\n";
+}
+
+# round4_step() does:
+#   dst = x + ((dst + I(x,y,z) + X[k] + T_i) <<< s)
+#   $t1 = not z' (copy of not z for the next step)
+#   $t2 = dst + X[k_next]
+sub round4_step
+{
+    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
+    my $T_i_h = ($T_i & 0xfffff000) >> 12;
+    my $T_i_l = $T_i & 0xfff;
+    $T_i_h = -((1<<32) - (0xfff00000 | $T_i_h)) if ($T_i_h >= (1<<19));
+
+    $code .= <<EOF;
+	lu12i.w	$t8,$T_i_h			/* load bits [31:12] of Constant */
+	or	$t1,$x,$t1			/* x | ... */
+	ori	$t8,$t8,$T_i_l			/* load bits [11:0] of Constant */
+	xor	$t1,$y,$t1			/* y ^ ... */
+	add.w	$t7,$t2,$t8			/* dst + X[k] + Const */
+EOF
+
+    if ($pos != 1) {
+        $code .= "	ld.w	$t0,$a1,$k_next*4		/* (NEXT STEP) X[$k_next] */\n";
+        $code .= "	add.w	$dst,$t7,$t1			/* dst += ... */\n";
+        $code .= "	add.w	$t2,$z,$t0			/* (NEXT STEP) dst + X[$k_next] */\n";
+        $code .= "	rotri.w $dst,$dst,32-$s			/* dst <<< s */\n";
+        $code .= "	nor	$t1,$zero,$y			/* (NEXT STEP) not z' = not $y */\n";
+        $code .= "	add.w	$dst,$dst,$x			/* dst += x */\n";
+    } else {
+        $code .= "	add.w	$a4,$t3,$a4			/* (NEXT LOOP) add old value of A */\n";
+        $code .= "	add.w	$dst,$t7,$t1			/* dst += ... */\n";
+        $code .= "	add.w	$a7,$t6,$a7			/* (NEXT LOOP) add old value of D */\n";
+        $code .= "	rotri.w $dst,$dst,32-$s			/* dst <<< s */\n";
+        $code .= "	addi.d	$a1,$a1,64			/* (NEXT LOOP) ptr += 64 */\n";
+        $code .= "	add.w	$dst,$dst,$x			/* dst += x */\n";
+    }
+}
+
+$code .= <<EOF;
+.text
+
+.globl ossl_md5_block_asm_data_order
+.type ossl_md5_block_asm_data_order function
+ossl_md5_block_asm_data_order:
+	# $a0 = arg #1 (ctx, MD5_CTX pointer)
+	# $a1 = arg #2 (ptr, data pointer)
+	# $a2 = arg #3 (nbr, number of 16-word blocks to process)
+	beqz	$a2,.Lend		# cmp nbr with 0, jmp if nbr == 0
+
+	# ptr is '$a1'
+	# end is '$a3'
+	slli.d	$t0,$a2,6
+	add.d	$a3,$a1,$t0
+
+	# A is '$a4'
+	# B is '$a5'
+	# C is '$a6'
+	# D is '$a7'
+	ld.w	$a4,$a0,0	# a4 = ctx->A
+	ld.w	$a5,$a0,4	# a5 = ctx->B
+	ld.w	$a6,$a0,8	# a6 = ctx->C
+	ld.w	$a7,$a0,12	# a7 = ctx->D
+
+# BEGIN of loop over 16-word blocks
+.align 6
+.Lloop:
+	# save old values of A, B, C, D
+	move	$t3,$a4
+	move	$t4,$a5
+	move	$t5,$a6
+	move	$t6,$a7
+
+	preld	0,$a1,0
+	preld	0,$a1,64
+EOF
+
+round1_step(-1, $a4, $a5, $a6, $a7, '1', 0xd76aa478, '7');
+round1_step(0, $a7, $a4, $a5, $a6, '2', 0xe8c7b756, '12');
+round1_step(0, $a6, $a7, $a4, $a5, '3', 0x242070db, '17');
+round1_step(0, $a5, $a6, $a7, $a4, '4', 0xc1bdceee, '22');
+round1_step(0, $a4, $a5, $a6, $a7, '5', 0xf57c0faf, '7');
+round1_step(0, $a7, $a4, $a5, $a6, '6', 0x4787c62a, '12');
+round1_step(0, $a6, $a7, $a4, $a5, '7', 0xa8304613, '17');
+round1_step(0, $a5, $a6, $a7, $a4, '8', 0xfd469501, '22');
+round1_step(0, $a4, $a5, $a6, $a7, '9', 0x698098d8, '7');
+round1_step(0, $a7, $a4, $a5, $a6, '10', 0x8b44f7af, '12');
+round1_step(0, $a6, $a7, $a4, $a5, '11', 0xffff5bb1, '17');
+round1_step(0, $a5, $a6, $a7, $a4, '12', 0x895cd7be, '22');
+round1_step(0, $a4, $a5, $a6, $a7, '13', 0x6b901122, '7');
+round1_step(0, $a7, $a4, $a5, $a6, '14', 0xfd987193, '12');
+round1_step(0, $a6, $a7, $a4, $a5, '15', 0xa679438e, '17');
+round1_step(1, $a5, $a6, $a7, $a4, '1', 0x49b40821, '22');
+
+round2_step(-1, $a4, $a5, $a6, $a7, '6', 0xf61e2562, '5');
+round2_step(0, $a7, $a4, $a5, $a6, '11', 0xc040b340, '9');
+round2_step(0, $a6, $a7, $a4, $a5, '0', 0x265e5a51, '14');
+round2_step(0, $a5, $a6, $a7, $a4, '5', 0xe9b6c7aa, '20');
+round2_step(0, $a4, $a5, $a6, $a7, '10', 0xd62f105d, '5');
+round2_step(0, $a7, $a4, $a5, $a6, '15', 0x2441453, '9');
+round2_step(0, $a6, $a7, $a4, $a5, '4', 0xd8a1e681, '14');
+round2_step(0, $a5, $a6, $a7, $a4, '9', 0xe7d3fbc8, '20');
+round2_step(0, $a4, $a5, $a6, $a7, '14', 0x21e1cde6, '5');
+round2_step(0, $a7, $a4, $a5, $a6, '3', 0xc33707d6, '9');
+round2_step(0, $a6, $a7, $a4, $a5, '8', 0xf4d50d87, '14');
+round2_step(0, $a5, $a6, $a7, $a4, '13', 0x455a14ed, '20');
+round2_step(0, $a4, $a5, $a6, $a7, '2', 0xa9e3e905, '5');
+round2_step(0, $a7, $a4, $a5, $a6, '7', 0xfcefa3f8, '9');
+round2_step(0, $a6, $a7, $a4, $a5, '12', 0x676f02d9, '14');
+round2_step(1, $a5, $a6, $a7, $a4, '5', 0x8d2a4c8a, '20');
+
+round3_step(-1, $a4, $a5, $a6, $a7, '8', 0xfffa3942, '4');
+round3_step(0, $a7, $a4, $a5, $a6, '11', 0x8771f681, '11');
+round3_step(0, $a6, $a7, $a4, $a5, '14', 0x6d9d6122, '16');
+round3_step(0, $a5, $a6, $a7, $a4, '1', 0xfde5380c, '23');
+round3_step(0, $a4, $a5, $a6, $a7, '4', 0xa4beea44, '4');
+round3_step(0, $a7, $a4, $a5, $a6, '7', 0x4bdecfa9, '11');
+round3_step(0, $a6, $a7, $a4, $a5, '10', 0xf6bb4b60, '16');
+round3_step(0, $a5, $a6, $a7, $a4, '13', 0xbebfbc70, '23');
+round3_step(0, $a4, $a5, $a6, $a7, '0', 0x289b7ec6, '4');
+round3_step(0, $a7, $a4, $a5, $a6, '3', 0xeaa127fa, '11');
+round3_step(0, $a6, $a7, $a4, $a5, '6', 0xd4ef3085, '16');
+round3_step(0, $a5, $a6, $a7, $a4, '9', 0x4881d05, '23');
+round3_step(0, $a4, $a5, $a6, $a7, '12', 0xd9d4d039, '4');
+round3_step(0, $a7, $a4, $a5, $a6, '15', 0xe6db99e5, '11');
+round3_step(0, $a6, $a7, $a4, $a5, '2', 0x1fa27cf8, '16');
+round3_step(1, $a5, $a6, $a7, $a4, '0', 0xc4ac5665, '23');
+
+round4_step(-1, $a4, $a5, $a6, $a7, '7', 0xf4292244, '6');
+round4_step(0, $a7, $a4, $a5, $a6, '14', 0x432aff97, '10');
+round4_step(0, $a6, $a7, $a4, $a5, '5', 0xab9423a7, '15');
+round4_step(0, $a5, $a6, $a7, $a4, '12', 0xfc93a039, '21');
+round4_step(0, $a4, $a5, $a6, $a7, '3', 0x655b59c3, '6');
+round4_step(0, $a7, $a4, $a5, $a6, '10', 0x8f0ccc92, '10');
+round4_step(0, $a6, $a7, $a4, $a5, '1', 0xffeff47d, '15');
+round4_step(0, $a5, $a6, $a7, $a4, '8', 0x85845dd1, '21');
+round4_step(0, $a4, $a5, $a6, $a7, '15', 0x6fa87e4f, '6');
+round4_step(0, $a7, $a4, $a5, $a6, '6', 0xfe2ce6e0, '10');
+round4_step(0, $a6, $a7, $a4, $a5, '13', 0xa3014314, '15');
+round4_step(0, $a5, $a6, $a7, $a4, '4', 0x4e0811a1, '21');
+round4_step(0, $a4, $a5, $a6, $a7, '11', 0xf7537e82, '6');
+round4_step(0, $a7, $a4, $a5, $a6, '2', 0xbd3af235, '10');
+round4_step(0, $a6, $a7, $a4, $a5, '9', 0x2ad7d2bb, '15');
+round4_step(1, $a5, $a6, $a7, $a4, '0', 0xeb86d391, '21');
+
+$code .= <<EOF;
+	# add old values of B, C
+	add.w	$a5,$t4,$a5
+	add.w	$a6,$t5,$a6
+
+	bltu	$a1,$a3,.Lloop	# jmp if ptr < end
+
+	st.w	$a4,$a0,0	# ctx->A = A
+	st.w	$a5,$a0,4	# ctx->B = B
+	st.w	$a6,$a0,8	# ctx->C = C
+	st.w	$a7,$a0,12	# ctx->D = D
+
+.Lend:
+	jr	$ra
+.size ossl_md5_block_asm_data_order,.-ossl_md5_block_asm_data_order
+EOF
+
+$code =~ s/\`([^\`]*)\`/eval($1)/gem;
+
+print $code;
+
+close STDOUT;

crypto/md5/build.info

@@ -5,6 +5,7 @@ IF[{- !$disabled{asm} -}]
   $MD5ASM_x86=md5-586.S
   $MD5ASM_x86_64=md5-x86_64.s
   $MD5ASM_aarch64=md5-aarch64.S
+  $MD5ASM_loongarch64=md5-loongarch64.S
   $MD5ASM_sparcv9=md5-sparcv9.S
 
   # Now that we have defined all the arch specific variables, use the
@@ -39,5 +40,8 @@ GENERATE[md5-x86_64.s]=asm/md5-x86_64.pl
 GENERATE[md5-aarch64.S]=asm/md5-aarch64.pl
 INCLUDE[md5-aarch64.o]=..
 
+GENERATE[md5-loongarch64.S]=asm/md5-loongarch64.pl
+INCLUDE[md5-loongarch64.o]=..
+
 GENERATE[md5-sparcv9.S]=asm/md5-sparcv9.pl
 INCLUDE[md5-sparcv9.o]=..

crypto/md5/md5_local.h

@@ -15,7 +15,8 @@
 #ifdef MD5_ASM
 # if defined(__i386) || defined(__i386__) || defined(_M_IX86) || \
      defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || \
-     defined(_M_X64) || defined(__aarch64__)
+     defined(_M_X64) || defined(__aarch64__) || \
+     (defined(__loongarch__) && __loongarch_grlen == 64)
 #  define md5_block_data_order ossl_md5_block_asm_data_order
 # elif defined(__ia64) || defined(__ia64__) || defined(_M_IA64)
 #  define md5_block_data_order ossl_md5_block_asm_data_order