diff options
| author | Andy Polyakov <appro@openssl.org> | 2011-03-04 13:09:16 +0000 |
|---|---|---|
| committer | Andy Polyakov <appro@openssl.org> | 2011-03-04 13:09:16 +0000 |
| commit | 0ab8fd58e1c3077c1e5e9057b3d73b73b99a843e (patch) | |
| tree | 80995d4501716ea1b90283dce4a140e3fe2b9465 /crypto/modes/asm/ghash-s390x.pl | |
| parent | 8aa6cff40fa392c993929a1661cc4ca66f3f5a8f (diff) | |
s390x assembler pack: tune-up and support for new z196 hardware.
Diffstat (limited to 'crypto/modes/asm/ghash-s390x.pl')
| -rw-r--r-- | crypto/modes/asm/ghash-s390x.pl | 13 |
1 files changed, 11 insertions, 2 deletions
diff --git a/crypto/modes/asm/ghash-s390x.pl b/crypto/modes/asm/ghash-s390x.pl index 16ad034fc1..48cb08d338 100644 --- a/crypto/modes/asm/ghash-s390x.pl +++ b/crypto/modes/asm/ghash-s390x.pl @@ -28,6 +28,15 @@ # remains z/Architecture specific. On z990 it was measured to perform # 2.8x better than 32-bit code generated by gcc 4.3. +# March 2011. +# +# Support for hardware KIMD-GHASH is verified to produce correct +# result and therefore is engaged. On z196 it was measured to process +# 8KB buffer ~7 faster than software implementation. It's not as +# impressive for smaller buffer sizes and for smallest 16-bytes buffer +# it's actually almost 2 times slower. Which is the reason why +# KIMD-GHASH is not used in gcm_gmult_4bit. + $flavour = shift; if ($flavour =~ /3[12]/) { @@ -41,7 +50,7 @@ if ($flavour =~ /3[12]/) { while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {} open STDOUT,">$output"; -$softonly=1; # disable hardware support for now +$softonly=0; $Zhi="%r0"; $Zlo="%r1"; @@ -70,7 +79,7 @@ $code.=<<___; .align 32 gcm_gmult_4bit: ___ -$code.=<<___ if(!$softonly); +$code.=<<___ if(!$softonly && 0); # hardware is slow for single block... larl %r1,OPENSSL_s390xcap_P lg %r0,0(%r1) tmhl %r0,0x4000 # check for message-security-assist |