sha1-armv4-large.pl: reschedule instructions for dual-issue pipeline.

crypto/sha/asm/sha1-armv4-large.pl

@@ -41,6 +41,13 @@
 #	issue Cortex A8 core was measured to process input block in
 #	~990 cycles.
 
+# August 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 13% improvement on
+# Cortex A8 core and in absolute terms ~870 cycles per input block
+# [or 13.6 cycles per byte].
+
+
 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
@@ -60,43 +67,22 @@ $t3="r12";
 $Xi="r14";
 @V=($a,$b,$c,$d,$e);
 
-# One can optimize this for aligned access on big-endian architecture,
-# but code's endian neutrality makes it too pretty:-)
-sub Xload {
-my ($a,$b,$c,$d,$e)=@_;
-$code.=<<___;
-	ldrb	$t0,[$inp],#4
-	ldrb	$t1,[$inp,#-3]
-	ldrb	$t2,[$inp,#-2]
-	ldrb	$t3,[$inp,#-1]
-	add	$e,$K,$e,ror#2			@ E+=K_00_19
-	orr	$t0,$t1,$t0,lsl#8
-	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
-	orr	$t0,$t2,$t0,lsl#8
-	eor	$t1,$c,$d			@ F_xx_xx
-	orr	$t0,$t3,$t0,lsl#8
-	add	$e,$e,$t0			@ E+=X[i]
-	str	$t0,[$Xi,#-4]!
-___
-}
 sub Xupdate {
-my ($a,$b,$c,$d,$e,$flag)=@_;
+my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_;
 $code.=<<___;
 	ldr	$t0,[$Xi,#15*4]
 	ldr	$t1,[$Xi,#13*4]
 	ldr	$t2,[$Xi,#7*4]
-	ldr	$t3,[$Xi,#2*4]
 	add	$e,$K,$e,ror#2			@ E+=K_xx_xx
+	ldr	$t3,[$Xi,#2*4]
 	eor	$t0,$t0,$t1
 	eor	$t2,$t2,$t3
-	eor	$t0,$t0,$t2
-	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
-___
-$code.=<<___ if (!defined($flag));
-	eor	$t1,$c,$d			@ F_xx_xx, but not in 40_59
-___
-$code.=<<___;
+	eor	$t1,$c,$d			@ F_xx_xx
 	mov	$t0,$t0,ror#31
+	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
+	eor	$t0,$t0,$t2,ror#31
+	$opt1					@ F_xx_xx
+	$opt2					@ F_xx_xx
 	add	$e,$e,$t0			@ E+=X[i]
 	str	$t0,[$Xi,#-4]!
 ___
@@ -104,19 +90,29 @@ ___
 
 sub BODY_00_15 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xload(@_);
 $code.=<<___;
+	ldrb	$t0,[$inp],#4
+	ldrb	$t1,[$inp,#-1]
+	ldrb	$t2,[$inp,#-2]
+	add	$e,$K,$e,ror#2			@ E+=K_00_19
+	ldrb	$t3,[$inp,#-3]
+	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
+	orr	$t0,$t1,$t0,lsl#24
+	eor	$t1,$c,$d			@ F_xx_xx
+	orr	$t0,$t0,$t2,lsl#8
+	orr	$t0,$t0,$t3,lsl#16
 	and	$t1,$b,$t1,ror#2
+	add	$e,$e,$t0			@ E+=X[i]
 	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
+	str	$t0,[$Xi,#-4]!
 	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
 ___
 }
 
 sub BODY_16_19 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xupdate(@_);
+	&Xupdate(@_,"and $t1,$b,$t1,ror#2");
 $code.=<<___;
-	and	$t1,$b,$t1,ror#2
 	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
 	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
 ___
@@ -124,34 +120,20 @@ ___
 
 sub BODY_20_39 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xupdate(@_);
+	&Xupdate(@_,"eor $t1,$b,$t1,ror#2");
 $code.=<<___;
-	eor	$t1,$b,$t1,ror#2		@ F_20_39(B,C,D)
 	add	$e,$e,$t1			@ E+=F_20_39(B,C,D)
 ___
 }
 
 sub BODY_40_59 {
 my ($a,$b,$c,$d,$e)=@_;
-if (1) {
-	&Xupdate(@_);
+	&Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
 $code.=<<___;
-	and	$t2,$c,$d
-	and	$t1,$b,$t1,ror#2
-	add	$e,$e,$t2,ror#2
-	add	$e,$e,$t1			@ E+=F_40_59(B,C,D)
-___
-} else {
-	&Xupdate(@_,1);
-$code.=<<___;
-	and	$t1,$b,$c,ror#2
-	orr	$t2,$b,$c,ror#2
-	and	$t2,$t2,$d,ror#2
-	orr	$t1,$t1,$t2			@ F_40_59(B,C,D)
 	add	$e,$e,$t1			@ E+=F_40_59(B,C,D)
+	add	$e,$e,$t2,ror#2
 ___
 }
-}
 
 $code=<<___;
 .text