From: Andy Polyakov <appro@openssl.org>
Date: Fri, 1 Dec 2017 14:57:43 +0000 (+0100)
Subject: modes/asm/ghashv8-armx.pl: optimize modulo-scheduled loop.
X-Git-Tag: OpenSSL_1_1_1-pre1~374
X-Git-Url: https://git.openssl.org/gitweb/?p=openssl.git;a=commitdiff_plain;h=aa7bf316980259a11dcbaf6128ed86d33dc24b97

modes/asm/ghashv8-armx.pl: optimize modulo-scheduled loop.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4830)
---

diff --git a/crypto/modes/asm/ghashv8-armx.pl b/crypto/modes/asm/ghashv8-armx.pl
index 2498a1d646..8cb2e452cd 100644
--- a/crypto/modes/asm/ghashv8-armx.pl
+++ b/crypto/modes/asm/ghashv8-armx.pl
@@ -30,17 +30,18 @@
 #
 # November 2017
 #
-# AArch64 register bank to "accommodate" 4x aggregated reduction...
+# AArch64 register bank to "accommodate" 4x aggregated reduction and
+# improve performance by 20-70% depending on processor.
 #
 # Current performance in cycles per processed byte:
 #
 #		64-bit PMULL	32-bit PMULL	32-bit NEON(*)
-# Apple A7			0.92		5.62
-# Cortex-A53			1.01		8.39
-# Cortex-A57			1.17		7.61
-# Denver			0.71		6.02
-# Mongoose			1.10		8.06
-# Kryo				1.16		8.00
+# Apple A7	0.58		0.92		5.62
+# Cortex-A53	0.85		1.01		8.39
+# Cortex-A57	0.73		1.17		7.61
+# Denver	0.51		0.65		6.02
+# Mongoose	0.65		1.10		8.06
+# Kryo		0.76		1.16		8.00
 #
 # (*)	presented for reference/comparison purposes;
 
@@ -466,65 +467,61 @@ gcm_ghash_v8_4x:
 .align	4
 .Loop4x:
 	veor		$t0,$I0,$Xl
+	 vld1.64	{$I0-$j3},[$inp],#64
 	vext.8		$IN,$t0,$t0,#8
+#ifndef __ARMEB__
+	 vrev64.8	$j1,$j1
+	 vrev64.8	$j2,$j2
+	 vrev64.8	$j3,$j3
+	 vrev64.8	$I0,$I0
+#endif
 
 	vpmull.p64	$Xl,$H4,$IN		@ H^4·(Xi+Ii)
 	veor		$t0,$t0,$IN
 	vpmull2.p64	$Xh,$H4,$IN
+	 vext.8		$I3,$j3,$j3,#8
 	vpmull2.p64	$Xm,$H34,$t0
 
 	veor		$Xl,$Xl,$Yl
 	veor		$Xh,$Xh,$Yh
+	 vext.8		$I2,$j2,$j2,#8
 	veor		$Xm,$Xm,$Ym
+	 vext.8		$I1,$j1,$j1,#8
 
 	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
 	veor		$t2,$Xl,$Xh
+	 vpmull.p64	$Yl,$H,$I3		@ H·Ii+3
+	 veor		$j3,$j3,$I3
 	veor		$Xm,$Xm,$t1
+	 vpmull2.p64	$Yh,$H,$I3
 	veor		$Xm,$Xm,$t2
+	 vpmull.p64	$Ym,$Hhl,$j3
 
 	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
 	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
 	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
-	veor		$Xl,$Xm,$t2
-
-	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
-	vpmull.p64	$Xl,$Xl,$xC2
-	veor		$t2,$t2,$Xh
-	veor		$Xl,$Xl,$t2
-	vext.8		$Xl,$Xl,$Xl,#8
-
-	 vld1.64	{$I0-$j3},[$inp],#64
-#ifndef __ARMEB__
-	 vrev64.8	$j1,$j1
-	 vrev64.8	$j2,$j2
-	 vrev64.8	$j3,$j3
-	 vrev64.8	$I0,$I0
-#endif
-	 vext.8		$I3,$j3,$j3,#8
-	 vext.8		$I2,$j2,$j2,#8
-	 vext.8		$I1,$j1,$j1,#8
-
-	 vpmull.p64	$Yl,$H,$I3		@ H·Ii+3
-	 veor		$j3,$j3,$I3
-	 vpmull2.p64	$Yh,$H,$I3
-	 vpmull.p64	$Ym,$Hhl,$j3
-
 	 vpmull.p64	$t0,$H2,$I2		@ H^2·Ii+2
 	 veor		$j2,$j2,$I2
 	 vpmull2.p64	$I2,$H2,$I2
+	veor		$Xl,$Xm,$t2
 	 vpmull2.p64	$j2,$Hhl,$j2
 
 	 veor		$Yl,$Yl,$t0
 	 veor		$Yh,$Yh,$I2
 	 veor		$Ym,$Ym,$j2
 
+	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
+	vpmull.p64	$Xl,$Xl,$xC2
 	 vpmull.p64	$j3,$H3,$I1		@ H^3·Ii+1
 	 veor		$j1,$j1,$I1
+	veor		$t2,$t2,$Xh
 	 vpmull2.p64	$I1,$H3,$I1
 	 vpmull.p64	$j1,$H34,$j1
 
+	veor		$Xl,$Xl,$t2
 	 veor		$Yl,$Yl,$j3
 	 veor		$Yh,$Yh,$I1
+	vext.8		$Xl,$Xl,$Xl,#8
 	 veor		$Ym,$Ym,$j1
 
 	subs		$len,$len,#64