From 9833757b5d8683cc2e92ab45115794bc2bd9e66c Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Mon, 14 Nov 2011 20:47:22 +0000
Subject: s390x assembler pack update from HEAD.

---
 crypto/rc4/asm/rc4-s390x.pl | 49 ++++++++++++++++++++++++++++++++++++---------
 1 file changed, 39 insertions(+), 10 deletions(-)

(limited to 'crypto/rc4')

diff --git a/crypto/rc4/asm/rc4-s390x.pl b/crypto/rc4/asm/rc4-s390x.pl
index c7ed59510b..1aa754820c 100644
--- a/crypto/rc4/asm/rc4-s390x.pl
+++ b/crypto/rc4/asm/rc4-s390x.pl
@@ -13,6 +13,29 @@
 # "cluster" Address Generation Interlocks, so that one pipeline stall
 # resolves several dependencies.
 
+# 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 z990 it was measured to perform
+# 50% better than code generated by gcc 4.3.
+
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+	$SIZE_T=4;
+	$g="";
+} else {
+	$SIZE_T=8;
+	$g="g";
+}
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
+open STDOUT,">$output";
+
 $rp="%r14";
 $sp="%r15";
 $code=<<___;
@@ -39,7 +62,12 @@ $code.=<<___;
 .type	RC4,\@function
 .align	64
 RC4:
-	stmg	%r6,%r11,48($sp)
+	stm${g}	%r6,%r11,6*$SIZE_T($sp)
+___
+$code.=<<___ if ($flavour =~ /3[12]/);
+	llgfr	$len,$len
+___
+$code.=<<___;
 	llgc	$XX[0],0($key)
 	llgc	$YY,1($key)
 	la	$XX[0],1($XX[0])
@@ -90,7 +118,7 @@ $code.=<<___;
 	xgr	$acc,$TX[1]
 	stg	$acc,0($out)
 	la	$out,8($out)
-	brct	$cnt,.Loop8
+	brctg	$cnt,.Loop8
 
 .Lshort:
 	lghi	$acc,7
@@ -122,7 +150,7 @@ $code.=<<___;
 	ahi	$XX[0],-1
 	stc	$XX[0],0($key)
 	stc	$YY,1($key)
-	lmg	%r6,%r11,48($sp)
+	lm${g}	%r6,%r11,6*$SIZE_T($sp)
 	br	$rp
 .size	RC4,.-RC4
 .string	"RC4 for s390x, CRYPTOGAMS by <appro\@openssl.org>"
@@ -130,7 +158,7 @@ $code.=<<___;
 ___
 }
 
-# void private_RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
+# void RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
 {
 $cnt="%r0";
 $idx="%r1";
@@ -143,11 +171,11 @@ $ikey="%r7";
 $iinp="%r8";
 
 $code.=<<___;
-.globl	private_RC4_set_key
-.type	private_RC4_set_key,\@function
+.globl	RC4_set_key
+.type	RC4_set_key,\@function
 .align	64
-private_RC4_set_key:
-	stmg	%r6,%r8,48($sp)
+RC4_set_key:
+	stm${g}	%r6,%r8,6*$SIZE_T($sp)
 	lhi	$cnt,256
 	la	$idx,0(%r0)
 	sth	$idx,0($key)
@@ -180,9 +208,9 @@ private_RC4_set_key:
 	la	$iinp,0(%r0)
 	j	.L2ndloop
 .Ldone:
-	lmg	%r6,%r8,48($sp)
+	lm${g}	%r6,%r8,6*$SIZE_T($sp)
 	br	$rp
-.size	private_RC4_set_key,.-private_RC4_set_key
+.size	RC4_set_key,.-RC4_set_key
 
 ___
 }
@@ -203,3 +231,4 @@ RC4_options:
 ___
 
 print $code;
+close STDOUT;	# force flush
-- 
cgit v1.2.3