diff options
author | Andy Polyakov <appro@openssl.org> | 2010-07-08 07:47:35 +0000 |
---|---|---|
committer | Andy Polyakov <appro@openssl.org> | 2010-07-08 07:47:35 +0000 |
commit | 7c5889bf7a4ca21138e88d428fa3fa4ccc94789b (patch) | |
tree | 798e24f66fb3856ff972293ba89f331f9b71cd3f /crypto/sparccpuid.S | |
parent | dc53a037b0ce091f2ade0d067ce9e88dd1c2ca21 (diff) |
sparcv9cap.c: reiterate CPU detection logic.
Diffstat (limited to 'crypto/sparccpuid.S')
-rw-r--r-- | crypto/sparccpuid.S | 52 |
1 files changed, 49 insertions, 3 deletions
diff --git a/crypto/sparccpuid.S b/crypto/sparccpuid.S index e53c7fad61..1e0756a173 100644 --- a/crypto/sparccpuid.S +++ b/crypto/sparccpuid.S @@ -232,13 +232,59 @@ _sparcv9_rdtick: .type _sparcv9_rdtick,#function .size _sparcv9_rdtick,.-_sparcv9_rdtick +! Probe and instrument VIS1 instruction. Output is number of cycles it +! takes to execute rdtick and pair of VIS1 instructions. US-Tx VIS unit +! is slow (documented to be 6 cycles on T2) and the core is in-order +! single-issue, it should be possible to distinguish Tx reliably... +! Observed return values are: +! +! UltraSPARC IIi 7 +! UltraSPARC III 7 +! UltraSPARC T1 24 +! +! Numbers for T2 and SPARC64 V-VII are more than welcomed. +! +! It would be possible to detect specifically US-T1 by instrumenting +! fmul8ulx16, which is emulated on T1 and as such accounts for quite +! a lot of %tick-s, couple of thousand on Linux... .global _sparcv9_vis1_probe .align 8 _sparcv9_vis1_probe: - .word 0x81b00c20 !fzeros %f0 - add %sp,BIAS+2,%o0 + .word 0x91410000 !rd %tick,%o0 + .word 0x81b00d80 !fxor %f0,%f0,%f0 + .word 0x85b08d82 !fxor %f2,%f2,%f2 + .word 0x93410000 !rd %tick,%o1 + .word 0x81b00d80 !fxor %f0,%f0,%f0 + .word 0x85b08d82 !fxor %f2,%f2,%f2 + .word 0x95410000 !rd %tick,%o2 + .word 0x81b00d80 !fxor %f0,%f0,%f0 + .word 0x85b08d82 !fxor %f2,%f2,%f2 + .word 0x97410000 !rd %tick,%o3 + .word 0x81b00d80 !fxor %f0,%f0,%f0 + .word 0x85b08d82 !fxor %f2,%f2,%f2 + .word 0x99410000 !rd %tick,%o4 + + ! calculate intervals + sub %o1,%o0,%o0 + sub %o2,%o1,%o1 + sub %o3,%o2,%o2 + sub %o4,%o3,%o3 + + ! find minumum value + cmp %o0,%o1 + .word 0x38680002 !bgu,a %xcc,.+8 + mov %o1,%o0 + cmp %o0,%o2 + .word 0x38680002 !bgu,a %xcc,.+8 + mov %o2,%o0 + cmp %o0,%o3 + .word 0x38680002 !bgu,a %xcc,.+8 + mov %o3,%o0 + + ! check for ASI_FP16_P is redundant... + add %sp,BIAS+2,%o1 retl - .word 0xc19a1a40 !ldda [%o0]ASI_FP16_P,%f0 + .word 0xc19a5a40 !ldda [%o1]ASI_FP16_P,%f0 .type _sparcv9_vis1_probe,#function .size _sparcv9_vis1_probe,.-_sparcv9_vis1_probe |