#
# sha512_block_data_order is ~70% faster than gcc 3.3 generated code.
+# January 2009.
+#
+# Add support for hardware SHA512 and reschedule instructions to
+# favour dual-issue z10 pipeline. Hardware SHA256/512 is ~4.7x faster
+# than software.
+
+# November 2010.
+#
+# Adapt for -m31 build. If kernel supports what's called "highgprs"
+# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
+# instructions and achieve "64-bit" performance even in 31-bit legacy
+# application context. The feature is not specific to any particular
+# processor, as long as it's "z-CPU". Latter implies that the code
+# remains z/Architecture specific. On z900 SHA256 was measured to
+# perform 2.4x and SHA512 - 13x better than code generated by gcc 4.3.
+
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+ $SIZE_T=4;
+ $g="";
+} else {
+ $SIZE_T=8;
+ $g="g";
+}
+
$t0="%r0";
$t1="%r1";
-$ctx="%r2";
+$ctx="%r2"; $t2="%r2";
$inp="%r3";
$len="%r4"; # used as index in inner loop
$T1="%r14";
$sp="%r15";
-$output=shift;
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
open STDOUT,">$output";
if ($output =~ /512/) {
@sigma0=(56,63, 7);
@sigma1=( 3,45, 6);
$rounds=80;
- $kimdfunc=0; # 0 means unknown/unsupported/unimplemented
+ $kimdfunc=3; # 0 means unknown/unsupported/unimplemented/disabled
} else {
$label="256";
$SZ=4;
}
$Func="sha${label}_block_data_order";
$Table="K${label}";
-$frame=160+16*$SZ;
+$stdframe=16*$SIZE_T+4*8;
+$frame=$stdframe+16*$SZ;
sub BODY_00_15 {
my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
$code.=<<___;
$ROT $t0,$e,$Sigma1[0]
$ROT $t1,$e,$Sigma1[1]
+ lgr $t2,$f
xgr $t0,$t1
$ROT $t1,$t1,`$Sigma1[2]-$Sigma1[1]`
+ xgr $t2,$g
+ $ST $T1,`$stdframe+$SZ*($i%16)`($sp)
xgr $t0,$t1 # Sigma1(e)
- $ST $T1,`160+$SZ*($i%16)`($sp)
- algr $T1,$t0 # T1+=Sigma1(e)
algr $T1,$h # T1+=h
- $ADD $T1,`$i*$SZ`($len,$tbl) # T1+=K[i]
- lgr $t0,$f
- xgr $t0,$g
- ngr $t0,$e
- xgr $t0,$g # Ch(e,f,g)
- algr $T1,$t0 # T1+=Ch(e,f,g)
+ ngr $t2,$e
+ lgr $t1,$a
+ algr $T1,$t0 # T1+=Sigma1(e)
$ROT $h,$a,$Sigma0[0]
+ xgr $t2,$g # Ch(e,f,g)
+ $ADD $T1,`$i*$SZ`($len,$tbl) # T1+=K[i]
$ROT $t0,$a,$Sigma0[1]
+ algr $T1,$t2 # T1+=Ch(e,f,g)
+ ogr $t1,$b
xgr $h,$t0
+ lgr $t2,$a
+ ngr $t1,$c
$ROT $t0,$t0,`$Sigma0[2]-$Sigma0[1]`
xgr $h,$t0 # h=Sigma0(a)
- lgr $t0,$a
- ogr $t0,$b
- ngr $t0,$c
- lgr $t1,$a
- ngr $t1,$b
- ogr $t0,$t1 # Maj(a,b,c)
- algr $h,$t0 # h+=Maj(a,b,c)
- algr $d,$T1 # d+=T1
+ ngr $t2,$b
algr $h,$T1 # h+=T1
+ ogr $t2,$t1 # Maj(a,b,c)
+ algr $d,$T1 # d+=T1
+ algr $h,$t2 # h+=Maj(a,b,c)
___
}
my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
$code.=<<___;
- $LD $T1,`160+$SZ*(($i+1)%16)`($sp) ### $i
- $LD $t1,`160+$SZ*(($i+14)%16)`($sp)
+ $LD $T1,`$stdframe+$SZ*(($i+1)%16)`($sp) ### $i
+ $LD $t1,`$stdframe+$SZ*(($i+14)%16)`($sp)
$ROT $t0,$T1,$sigma0[0]
$SHR $T1,$sigma0[2]
+ $ROT $t2,$t0,`$sigma0[1]-$sigma0[0]`
xgr $T1,$t0
- $ROT $t0,$t0,`$sigma0[1]-$sigma0[0]`
- xgr $T1,$t0 # sigma0(X[i+1])
$ROT $t0,$t1,$sigma1[0]
- $ADD $T1,`160+$SZ*($i%16)`($sp) # +=X[i]
+ xgr $T1,$t2 # sigma0(X[i+1])
$SHR $t1,$sigma1[2]
+ $ADD $T1,`$stdframe+$SZ*($i%16)`($sp) # +=X[i]
xgr $t1,$t0
- $ADD $T1,`160+$SZ*(($i+9)%16)`($sp) # +=X[i+9]
$ROT $t0,$t0,`$sigma1[1]-$sigma1[0]`
+ $ADD $T1,`$stdframe+$SZ*(($i+9)%16)`($sp) # +=X[i+9]
xgr $t1,$t0 # sigma1(X[i+14])
algr $T1,$t1 # +=sigma1(X[i+14])
___
.globl $Func
.type $Func,\@function
$Func:
+ sllg $len,$len,`log(16*$SZ)/log(2)`
___
$code.=<<___ if ($kimdfunc);
+ larl %r1,OPENSSL_s390xcap_P
+ lg %r0,0(%r1)
+ tmhl %r0,0x4000 # check for message-security assist
+ jz .Lsoftware
lghi %r0,0
- la %r1,16($sp)
+ la %r1,`2*$SIZE_T`($sp)
.long 0xb93e0002 # kimd %r0,%r2
- lg %r0,16($sp)
+ lg %r0,`2*$SIZE_T`($sp)
tmhh %r0,`0x8000>>$kimdfunc`
jz .Lsoftware
lghi %r0,$kimdfunc
lgr %r1,$ctx
lgr %r2,$inp
- sllg %r3,$len,`log(16*$SZ)/log(2)`
+ lgr %r3,$len
.long 0xb93e0002 # kimd %r0,%r2
brc 1,.-4 # pay attention to "partial completion"
br %r14
+.align 16
.Lsoftware:
___
$code.=<<___;
- sllg $len,$len,`log(16*$SZ)/log(2)`
- la $len,0($inp,$len)
- stmg $len,%r15,32($sp)
+ lghi %r1,-$frame
+ la $len,0($len,$inp)
+ stm${g} $ctx,%r15,`2*$SIZE_T`($sp)
lgr %r0,$sp
- aghi $sp,-$frame
- stg %r0,0($sp)
-
- bras $tbl,.Lpic
-.Lpic: aghi $tbl,$Table-.Lpic
+ la $sp,0(%r1,$sp)
+ st${g} %r0,0($sp)
+ larl $tbl,$Table
$LD $A,`0*$SZ`($ctx)
$LD $B,`1*$SZ`($ctx)
$LD $C,`2*$SZ`($ctx)
clgr $len,$t0
jne .Lrounds_16_xx
+ l${g} $ctx,`$frame+2*$SIZE_T`($sp)
+ la $inp,`16*$SZ`($inp)
$ADD $A,`0*$SZ`($ctx)
$ADD $B,`1*$SZ`($ctx)
$ADD $C,`2*$SZ`($ctx)
$ST $F,`5*$SZ`($ctx)
$ST $G,`6*$SZ`($ctx)
$ST $H,`7*$SZ`($ctx)
- la $inp,`16*$SZ`($inp)
- clg $inp,`$frame+32`($sp)
+ cl${g} $inp,`$frame+4*$SIZE_T`($sp)
jne .Lloop
- lmg %r6,%r15,`$frame+48`($sp)
+ lm${g} %r6,%r15,`$frame+6*$SIZE_T`($sp)
br %r14
.size $Func,.-$Func
.string "SHA${label} block transform for s390x, CRYPTOGAMS by <appro\@openssl.org>"
+.comm OPENSSL_s390xcap_P,16,8
___
$code =~ s/\`([^\`]*)\`/eval $1/gem;