sha1-armv4-large.pl: reschedule instructions for dual-issue pipeline [from HEAD]
authorAndy Polyakov <appro@openssl.org>
Tue, 3 Aug 2010 15:36:52 +0000 (15:36 +0000)
committerAndy Polyakov <appro@openssl.org>
Tue, 3 Aug 2010 15:36:52 +0000 (15:36 +0000)
.

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

index 88861af..6e65fe3 100644 (file)
 #      modes are limited. As result it takes more instructions to do
 #      the same job in Thumb, therefore the code is never twice as
 #      small and always slower.
-# [***]        which is also ~35% better than compiler generated code.
+# [***]        which is also ~35% better than compiler generated code. Dual-
+#      issue Cortex A8 core was measured to process input block in
+#      ~990 cycles.
 
-$output=shift;
+# August 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 13% improvement on
+# Cortex A8 core and in absolute terms ~870 cycles per input block
+# [or 13.6 cycles per byte].
+
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 $ctx="r0";
@@ -58,43 +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     $t0,$t0,$t2
-       eor     $t0,$t0,$t3
-       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     $t2,$t2,$t3
+       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]!
 ___
@@ -102,19 +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)
 ___
@@ -122,22 +120,18 @@ ___
 
 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)=@_;
-       &Xupdate(@_,1);
+       &Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
 $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
 ___
 }