RC4 for s390x.

2 files changed, 208 insertions, 0 deletions

diff --git a/crypto/rc4/Makefile b/crypto/rc4/Makefile
index f932a89035..264451a213 100644
--- a/crypto/rc4/Makefile
+++ b/crypto/rc4/Makefile
@@ -50,6 +50,9 @@ rc4-x86_64.s: asm/rc4-x86_64.pl
 rc4-ia64.S: asm/rc4-ia64.pl
 	$(PERL) asm/rc4-ia64.pl $(CFLAGS) > $@
 
+rc4-s390x.s:	asm/rc4-s390x.pl
+	$(PERL) asm/rc4-s390x.pl > $@
+
 rc4-ia64.s: rc4-ia64.S
 	@case `awk '/^#define RC4_INT/{print$$NF}' $(TOP)/include/openssl/opensslconf.h` in \
 	int)	set -x; $(CC) $(CFLAGS) -DSZ=4 -E rc4-ia64.S > $@ ;; \
diff --git a/crypto/rc4/asm/rc4-s390x.pl b/crypto/rc4/asm/rc4-s390x.pl
new file mode 100644
index 0000000000..4366c4fc1a
--- /dev/null
+++ b/crypto/rc4/asm/rc4-s390x.pl
@@ -0,0 +1,205 @@
+#!/usr/bin/env perl
+#
+# ====================================================================
+# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# February 2009
+#
+# Performance is 2x of gcc 3.4.6 on z10. Coding "secret" is to avoid
+# arithmetic instructions, but adhere to load and load address in
+# order to minimize Address Generation Interlock.
+
+$rp="%r14";
+$sp="%r15";
+$code=<<___;
+.text
+
+___
+
+# void RC4(RC4_KEY *key,size_t len,const void *inp,void *out)
+{
+$acc="%r0";
+$cnt="%r1";
+$key="%r2";
+$len="%r3";
+$inp="%r4";
+$out="%r5";
+
+@XX=("%r6","%r7");
+@TX=("%r8","%r9");
+$YY="%r10";
+$TY="%r11";
+
+$code.=<<___;
+.globl	RC4
+.type	RC4,\@function
+.align	64
+RC4:
+	stmg	%r6,%r11,48($sp)
+	llgc	$XX[0],0($key)
+	llgc	$YY,1($key)
+	la	$XX[0],1($XX[0])
+	llgcr	$XX[0],$XX[0]
+	llgc	$TX[0],2($XX[0],$key)
+	srlg	$cnt,$len,3
+	ltgr	$cnt,$cnt
+	jz	.Lshort
+	j	.Loop8
+
+.align	64
+.Loop8:
+___
+for ($i=0;$i<8;$i++) {
+$code.=<<___;
+	la	$YY,0($YY,$TX[0])	# $i
+	llgcr	$YY,$YY
+	la	$XX[1],1($XX[0])
+	llgcr	$XX[1],$XX[1]
+___
+$code.=<<___ if ($i>1);
+	sllg	$acc,$acc,8
+	ic	$acc,2($TY,$key)
+___
+$code.=<<___ if ($i==1);
+	llgc	$acc,2($TY,$key)
+___
+$code.=<<___;
+	llgc	$TY,2($YY,$key)
+	stc	$TX[0],2($YY,$key)
+	llgc	$TX[1],2($XX[1],$key)
+	stc	$TY,2($XX[0],$key)
+	cr	$XX[1],$YY
+	jne	.Lcmov$i
+	la	$TX[1],0($TX[0])
+.Lcmov$i:
+	la	$TY,0($TY,$TX[0])
+	llgcr	$TY,$TY
+___
+push(@TX,shift(@TX)); push(@XX,shift(@XX));     # "rotate" registers
+}
+
+$code.=<<___;
+	lg	$TX[1],0($inp)
+	sllg	$acc,$acc,8
+	la	$inp,8($inp)
+	ic	$acc,2($TY,$key)
+	xgr	$acc,$TX[1]
+	stg	$acc,0($out)
+	la	$out,8($out)
+	brct	$cnt,.Loop8
+
+.Lshort:
+	lghi	$acc,7
+	ngr	$len,$acc
+	jz	.Lexit
+	j	.Loop1
+
+.align	16
+.Loop1:
+	la	$YY,0($YY,$TX[0])
+	llgcr	$YY,$YY
+	llgc	$TY,2($YY,$key)
+	stc	$TX[0],2($YY,$key)
+	stc	$TY,2($XX[0],$key)
+	la	$TY,0($TY,$TX[0])
+	llgcr	$TY,$TY
+	la	$XX[0],1($XX[0])
+	llgcr	$XX[0],$XX[0]
+	llgc	$TY,2($TY,$key)
+	llgc	$TX[0],2($XX[0],$key)
+	llgc	$acc,0($inp)
+	la	$inp,1($inp)
+	xr	$acc,$TY
+	stc	$acc,0($out)
+	la	$out,1($out)
+	brct	$len,.Loop1
+
+.Lexit:
+	ahi	$XX[0],-1
+	stc	$XX[0],0($key)
+	stc	$YY,1($key)
+	lmg	%r6,%r11,48($sp)
+	br	$rp
+.size	RC4,.-RC4
+.string	"RC4 for s390x, CRYPTOGAMS by <appro\@openssl.org>"
+
+___
+}
+
+# void RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
+{
+$cnt="%r0";
+$idx="%r1";
+$key="%r2";
+$len="%r3";
+$inp="%r4";
+$acc="%r5";
+$dat="%r6";
+$ikey="%r7";
+$iinp="%r8";
+
+$code.=<<___;
+.globl	RC4_set_key
+.type	RC4_set_key,\@function
+.align	64
+RC4_set_key:
+	stmg	%r6,%r8,48($sp)
+	lhi	$cnt,256
+	la	$idx,0(%r0)
+	sth	$idx,0($key)
+.align	4
+.L1stloop:
+	stc	$idx,2($idx,$key)
+	la	$idx,1($idx)
+	brct	$cnt,.L1stloop
+
+	lghi	$ikey,-256
+	lr	$cnt,$len
+	la	$iinp,0(%r0)
+	la	$idx,0(%r0)
+.align	16
+.L2ndloop:
+	llgc	$acc,2+256($ikey,$key)
+	llgc	$dat,0($iinp,$inp)
+	la	$idx,0($idx,$acc)
+	la	$ikey,1($ikey)
+	la	$idx,0($idx,$dat)
+	la	$iinp,1($iinp)
+	llgcr	$idx,$idx
+	tml	$ikey,255
+	llgc	$dat,2($idx,$key)
+	stc	$dat,2+256-1($ikey,$key)
+	stc	$acc,2($idx,$key)
+	jz	.Ldone
+	brct	$cnt,.L2ndloop
+	lr	$cnt,$len
+	la	$iinp,0(%r0)
+	j	.L2ndloop
+.Ldone:
+	lmg	%r6,%r8,48($sp)
+	br	$rp
+.size	RC4_set_key,.-RC4_set_key
+
+___
+}
+
+# const char *RC4_options()
+$code.=<<___;
+.globl	RC4_options
+.type	RC4_options,\@function
+.align	16
+RC4_options:
+	larl	%r2,.Loptions
+	br	%r14
+.size	RC4_options,.-RC4_options
+.section	.rodata
+.Loptions:
+.align	8
+.string	"rc4(8x,char)"
+___
+
+print $code;