diff options
| author | Andy Polyakov <appro@openssl.org> | 2007-05-14 21:35:25 +0000 |
|---|---|---|
| committer | Andy Polyakov <appro@openssl.org> | 2007-05-14 21:35:25 +0000 |
| commit | b2dba9bf1f8f73376b9c1f0904a86996c728b236 (patch) | |
| tree | 8848a4f5efdf02d841b1ebea4969d879d6e61eb5 /crypto/ia64cpuid.S | |
| parent | 932cc129ee61f5b72636eee6a7c3268e23967f7b (diff) | |
| download | openssl-new-b2dba9bf1f8f73376b9c1f0904a86996c728b236.tar.gz | |
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/ia64cpuid.S')
| -rw-r--r-- | crypto/ia64cpuid.S | 36 |
1 files changed, 36 insertions, 0 deletions
diff --git a/crypto/ia64cpuid.S b/crypto/ia64cpuid.S index 5836565abb..818e2d1e1d 100644 --- a/crypto/ia64cpuid.S +++ b/crypto/ia64cpuid.S @@ -1,11 +1,13 @@ // Works on all IA-64 platforms: Linux, HP-UX, Win64i... // On Win64i compile with ias.exe. .text + .global OPENSSL_cpuid_setup# .proc OPENSSL_cpuid_setup# OPENSSL_cpuid_setup: { .mib; br.ret.sptk.many b0 };; .endp OPENSSL_cpuid_setup# + .global OPENSSL_rdtsc# .proc OPENSSL_rdtsc# OPENSSL_rdtsc: @@ -124,3 +126,37 @@ OPENSSL_wipe_cpu: mov ar.lc=r3 br.ret.sptk b0 };; .endp OPENSSL_wipe_cpu# + +.global OPENSSL_cleanse# +.proc OPENSSL_cleanse# +OPENSSL_cleanse: +{ .mib; and r2=7,r32 + cmp.leu p6,p0=15,r33 // len>=15 +(p6) br.cond.dptk .Lot };; + +.Little: +{ .mib; st1 [r32]=r0,1 + cmp.ltu p6,p7=1,r33 } // len>1 +{ .mbb; add r33=-1,r33 // len-- +(p6) br.cond.dptk .Little +(p7) br.ret.sptk.many b0 };; + +.Lot: +{ .mib; cmp.eq p6,p0=0,r2 +(p6) br.cond.dptk .Laligned };; +{ .mmi; st1 [r32]=r0,1;; + and r2=7,r32 } +{ .mib; add r33=-1,r33 + br .Lot };; + +.Laligned: +{ .mmi; st8 [r32]=r0,8 + and r2=-8,r33 // len&~7 + add r33=-8,r33 };; // len-=8 +{ .mib; cmp.ltu p6,p0=8,r2 // ((len+8)&~7)>8 +(p6) br.cond.dptk .Laligned };; + +{ .mbb; cmp.eq p6,p7=r0,r33 +(p7) br.cond.dpnt .Little +(p6) br.ret.sptk.many b0 };; +.endp OPENSSL_cleanse# |