diff options
| author | Andy Polyakov <appro@openssl.org> | 2007-05-14 23:35:25 +0200 |
|---|---|---|
| committer | Andy Polyakov <appro@openssl.org> | 2007-05-14 23:35:25 +0200 |
| commit | b2dba9bf1f8f73376b9c1f0904a86996c728b236 (patch) | |
| tree | 8848a4f5efdf02d841b1ebea4969d879d6e61eb5 /crypto/x86cpuid.pl | |
| parent | x86_64 assembler updates. (diff) | |
| download | openssl-b2dba9bf1f8f73376b9c1f0904a86996c728b236.tar.xz openssl-b2dba9bf1f8f73376b9c1f0904a86996c728b236.zip | |
Profiling revealed that OPENSSL_cleanse consumes *more* CPU time than
sha1_block_data_order when hashing short messages. Move OPENSSL_cleanse
to "cpuid" assembler module and gain 2x.
Diffstat (limited to 'crypto/x86cpuid.pl')
| -rw-r--r-- | crypto/x86cpuid.pl | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/crypto/x86cpuid.pl b/crypto/x86cpuid.pl index 7d924a60b7..13828d5633 100644 --- a/crypto/x86cpuid.pl +++ b/crypto/x86cpuid.pl @@ -216,6 +216,37 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); } } &function_end_B("OPENSSL_indirect_call"); +&function_begin_B("OPENSSL_cleanse"); + &mov ("edx",&wparam(0)); + &mov ("ecx",&wparam(1)); + &xor ("eax","eax"); + &cmp ("ecx",7); + &jae (&label("lot")); +&set_label("little"); + &mov (&BP(0,"edx"),"al"); + &sub ("ecx",1); + &lea ("edx",&DWP(1,"edx")); + &jnz (&label("little")); + &ret (); + +&set_label("lot",16); + &test ("edx",3); + &jz (&label("aligned")); + &mov (&BP(0,"edx"),"al"); + &lea ("ecx",&DWP(-1,"ecx")); + &lea ("edx",&DWP(1,"edx")); + &jmp (&label("lot")); +&set_label("aligned"); + &mov (&DWP(0,"edx"),"eax"); + &lea ("ecx",&DWP(-4,"ecx")); + &test ("ecx",-4); + &lea ("edx",&DWP(4,"edx")); + &jnz (&label("aligned")); + &cmp ("ecx",0); + &jne (&label("little")); + &ret (); +&function_end_B("OPENSSL_cleanse"); + &initseg("OPENSSL_cpuid_setup"); &asm_finish(); |