sha/asm/sha512-armv8.pl: add NEON version of SHA256.

This provides up to 30% better performance on some of recent processors.

Reviewed-by: Richard Levitte <levitte@openssl.org>

1 files changed, 313 insertions, 4 deletions

diff --git a/crypto/sha/asm/sha512-armv8.pl b/crypto/sha/asm/sha512-armv8.pl
index ffd80d6be3..ffae5f23bc 100644
--- a/crypto/sha/asm/sha512-armv8.pl
+++ b/crypto/sha/asm/sha512-armv8.pl
@@ -37,6 +37,20 @@
 #	indication of some compiler "pathology", most notably code
 #	generated with -mgeneral-regs-only is significanty faster
 #	and the gap is only 40-90%.
+#
+# October 2016.
+#
+# Originally it was reckoned that it makes no sense to implement NEON
+# version of SHA256 for 64-bit processors. This is because performance
+# improvement on most wide-spread Cortex-A5x processors was observed
+# to be marginal, same on Cortex-A53 and ~10% on A57. But then it was
+# observed that 32-bit NEON SHA256 performs significantly better than
+# 64-bit scalar version on *some* of the more recent processors. As
+# result 64-bit NEON version of SHA256 was added to provide best
+# all-round performance. For example it executes ~30% faster on X-Gene
+# and Mongoose. [For reference, NEON version of SHA512 is bound to
+# deliver much less improvement, likely *negative* on Cortex-A5x.
+# Which is why NEON support is limited to SHA256.]
 
 $output=pop;
 $flavour=pop;
@@ -195,6 +209,8 @@ $code.=<<___	if ($SZ==4);
 	ldr	w16,[x16]
 	tst	w16,#ARMV8_SHA256
 	b.ne	.Lv8_entry
+	tst	w16,#ARMV7_NEON
+	b.ne	.Lneon_entry
 #endif
 ___
 $code.=<<___;
@@ -425,6 +441,296 @@ $code.=<<___;
 ___
 }
 
+if ($SZ==4) {	######################################### NEON stuff #
+# You'll surely note a lot of similarities with sha256-armv4 module,
+# and of course it's not a coincidence. sha256-armv4 was used as
+# initial template, but was adapted for ARMv8 instruction set and
+# extensively re-tuned for all-round performance.
+
+my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
+my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
+my $Ktbl="x16";
+my $Xfer="x17";
+my @X = map("q$_",(0..3));
+my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
+my $j=0;
+
+sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
+{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
+  my $arg = pop;
+    $arg = "#$arg" if ($arg*1 eq $arg);
+    $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
+}
+
+sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
+sub Dlo     { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
+sub Dhi     { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
+
+sub Xupdate()
+{ use integer;
+  my $body = shift;
+  my @insns = (&$body,&$body,&$body,&$body);
+  my ($a,$b,$c,$d,$e,$f,$g,$h);
+
+	&ext_8		($T0,@X[0],@X[1],4);	# X[1..4]
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ext_8		($T3,@X[2],@X[3],4);	# X[9..12]
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&mov		(&Dscalar($T7),&Dhi(@X[3]));	# X[14..15]
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ushr_32	($T2,$T0,$sigma0[0]);
+	 eval(shift(@insns));
+	&ushr_32	($T1,$T0,$sigma0[2]);
+	 eval(shift(@insns));
+	&add_32 	(@X[0],@X[0],$T3);	# X[0..3] += X[9..12]
+	 eval(shift(@insns));
+	&sli_32		($T2,$T0,32-$sigma0[0]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ushr_32	($T3,$T0,$sigma0[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&eor_8		($T1,$T1,$T2);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&sli_32		($T3,$T0,32-$sigma0[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &ushr_32	($T4,$T7,$sigma1[0]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&eor_8		($T1,$T1,$T3);		# sigma0(X[1..4])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &sli_32	($T4,$T7,32-$sigma1[0]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &ushr_32	($T5,$T7,$sigma1[2]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &ushr_32	($T3,$T7,$sigma1[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&add_32		(@X[0],@X[0],$T1);	# X[0..3] += sigma0(X[1..4])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &sli_u32	($T3,$T7,32-$sigma1[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &eor_8	($T5,$T5,$T4);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &eor_8	($T5,$T5,$T3);		# sigma1(X[14..15])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&add_32		(@X[0],@X[0],$T5);	# X[0..1] += sigma1(X[14..15])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &ushr_32	($T6,@X[0],$sigma1[0]);
+	 eval(shift(@insns));
+	  &ushr_32	($T7,@X[0],$sigma1[2]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &sli_32	($T6,@X[0],32-$sigma1[0]);
+	 eval(shift(@insns));
+	  &ushr_32	($T5,@X[0],$sigma1[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &eor_8	($T7,$T7,$T6);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	  &sli_32	($T5,@X[0],32-$sigma1[1]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ld1_32		("{$T0}","[$Ktbl], #16");
+	 eval(shift(@insns));
+	  &eor_8	($T7,$T7,$T5);		# sigma1(X[16..17])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&eor_8		($T5,$T5,$T5);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&mov		(&Dhi($T5), &Dlo($T7));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&add_32		(@X[0],@X[0],$T5);	# X[2..3] += sigma1(X[16..17])
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&add_32		($T0,$T0,@X[0]);
+	 while($#insns>=1) { eval(shift(@insns)); }
+	&st1_32		("{$T0}","[$Xfer], #16");
+	 eval(shift(@insns));
+
+	push(@X,shift(@X));		# "rotate" X[]
+}
+
+sub Xpreload()
+{ use integer;
+  my $body = shift;
+  my @insns = (&$body,&$body,&$body,&$body);
+  my ($a,$b,$c,$d,$e,$f,$g,$h);
+
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ld1_8		("{@X[0]}","[$inp],#16");
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&ld1_32		("{$T0}","[$Ktbl],#16");
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&rev32		(@X[0],@X[0]);
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	 eval(shift(@insns));
+	&add_32		($T0,$T0,@X[0]);
+	 foreach (@insns) { eval; }	# remaining instructions
+	&st1_32		("{$T0}","[$Xfer], #16");
+
+	push(@X,shift(@X));		# "rotate" X[]
+}
+
+sub body_00_15 () {
+	(
+	'($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
+	'&add	($h,$h,$t1)',			# h+=X[i]+K[i]
+	'&add	($a,$a,$t4);'.			# h+=Sigma0(a) from the past
+	'&and	($t1,$f,$e)',
+	'&bic	($t4,$g,$e)',
+	'&eor	($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
+	'&add	($a,$a,$t2)',			# h+=Maj(a,b,c) from the past
+	'&orr	($t1,$t1,$t4)',			# Ch(e,f,g)
+	'&eor	($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))',	# Sigma1(e)
+	'&eor	($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
+	'&add	($h,$h,$t1)',			# h+=Ch(e,f,g)
+	'&ror	($t0,$t0,"#$Sigma1[0]")',
+	'&eor	($t2,$a,$b)',			# a^b, b^c in next round
+	'&eor	($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))',	# Sigma0(a)
+	'&add	($h,$h,$t0)',			# h+=Sigma1(e)
+	'&ldr	($t1,sprintf "[sp,#%d]",4*(($j+1)&15))	if (($j&15)!=15);'.
+	'&ldr	($t1,"[$Ktbl]")				if ($j==15);'.
+	'&and	($t3,$t3,$t2)',			# (b^c)&=(a^b)
+	'&ror	($t4,$t4,"#$Sigma0[0]")',
+	'&add	($d,$d,$h)',			# d+=h
+	'&eor	($t3,$t3,$b)',			# Maj(a,b,c)
+	'$j++;	unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
+	)
+}
+
+$code.=<<___;
+#ifdef	__KERNEL__
+.globl	sha256_block_neon
+#endif
+.type	sha256_block_neon,%function
+.align	4
+sha256_block_neon:
+.Lneon_entry:
+	stp	x29, x30, [sp, #-16]!
+	mov	x29, sp
+	sub	sp,sp,#16*4
+
+	adr	$Ktbl,.LK256
+	add	$num,$inp,$num,lsl#6	// len to point at the end of inp
+
+	ld1.8	{@X[0]},[$inp], #16
+	ld1.8	{@X[1]},[$inp], #16
+	ld1.8	{@X[2]},[$inp], #16
+	ld1.8	{@X[3]},[$inp], #16
+	ld1.32	{$T0},[$Ktbl], #16
+	ld1.32	{$T1},[$Ktbl], #16
+	ld1.32	{$T2},[$Ktbl], #16
+	ld1.32	{$T3},[$Ktbl], #16
+	rev32	@X[0],@X[0]		// yes, even on
+	rev32	@X[1],@X[1]		// big-endian
+	rev32	@X[2],@X[2]
+	rev32	@X[3],@X[3]
+	mov	$Xfer,sp
+	add.32	$T0,$T0,@X[0]
+	add.32	$T1,$T1,@X[1]
+	add.32	$T2,$T2,@X[2]
+	st1.32	{$T0-$T1},[$Xfer], #32
+	add.32	$T3,$T3,@X[3]
+	st1.32	{$T2-$T3},[$Xfer]
+	sub	$Xfer,$Xfer,#32
+
+	ldp	$A,$B,[$ctx]
+	ldp	$C,$D,[$ctx,#8]
+	ldp	$E,$F,[$ctx,#16]
+	ldp	$G,$H,[$ctx,#24]
+	ldr	$t1,[sp,#0]
+	mov	$t2,wzr
+	eor	$t3,$B,$C
+	mov	$t4,wzr
+	b	.L_00_48
+
+.align	4
+.L_00_48:
+___
+	&Xupdate(\&body_00_15);
+	&Xupdate(\&body_00_15);
+	&Xupdate(\&body_00_15);
+	&Xupdate(\&body_00_15);
+$code.=<<___;
+	cmp	$t1,#0				// check for K256 terminator
+	ldr	$t1,[sp,#0]
+	sub	$Xfer,$Xfer,#64
+	bne	.L_00_48
+
+	sub	$Ktbl,$Ktbl,#256		// rewind $Ktbl
+	cmp	$inp,$num
+	mov	$Xfer, #64
+	csel	$Xfer, $Xfer, xzr, eq
+	sub	$inp,$inp,$Xfer			// avoid SEGV
+	mov	$Xfer,sp
+___
+	&Xpreload(\&body_00_15);
+	&Xpreload(\&body_00_15);
+	&Xpreload(\&body_00_15);
+	&Xpreload(\&body_00_15);
+$code.=<<___;
+	add	$A,$A,$t4			// h+=Sigma0(a) from the past
+	ldp	$t0,$t1,[$ctx,#0]
+	add	$A,$A,$t2			// h+=Maj(a,b,c) from the past
+	ldp	$t2,$t3,[$ctx,#8]
+	add	$A,$A,$t0			// accumulate
+	add	$B,$B,$t1
+	ldp	$t0,$t1,[$ctx,#16]
+	add	$C,$C,$t2
+	add	$D,$D,$t3
+	ldp	$t2,$t3,[$ctx,#24]
+	add	$E,$E,$t0
+	add	$F,$F,$t1
+	 ldr	$t1,[sp,#0]
+	stp	$A,$B,[$ctx,#0]
+	add	$G,$G,$t2
+	 mov	$t2,wzr
+	stp	$C,$D,[$ctx,#8]
+	add	$H,$H,$t3
+	stp	$E,$F,[$ctx,#16]
+	 eor	$t3,$B,$C
+	stp	$G,$H,[$ctx,#24]
+	 mov	$t4,wzr
+	 mov	$Xfer,sp
+	b.ne	.L_00_48
+
+	ldr	x29,[x29]
+	add	sp,sp,#16*4+16
+	ret
+.size	sha256_block_neon,.-sha256_block_neon
+___
+}
+
 $code.=<<___;
 #ifndef	__KERNEL__
 .comm	OPENSSL_armcap_P,4,4
@@ -456,12 +762,15 @@ close SELF;
 
 foreach(split("\n",$code)) {
 
-	s/\`([^\`]*)\`/eval($1)/geo;
+	s/\`([^\`]*)\`/eval($1)/ge;
+
+	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
 
-	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
+	s/\bq([0-9]+)\b/v$1.16b/g;		# old->new registers
 
-	s/\.\w?32\b//o		and s/\.16b/\.4s/go;
-	m/(ld|st)1[^\[]+\[0\]/o	and s/\.4s/\.s/go;
+	s/\.[ui]?8(\s)/$1/;
+	s/\.\w?32\b//		and s/\.16b/\.4s/g;
+	m/(ld|st)1[^\[]+\[0\]/	and s/\.4s/\.s/g;
 
 	print $_,"\n";
 }