Re: [kvm-ppc-devel] First detailed look at instruction emulation using kvm_stat instruction extension

[Hollis B. <ho...@us...>]

On Monday 14 April 2008 08:00:47 Christian Ehrhardt wrote:
>
> A short check of the ratio between the different sprn's used gave that
> listing: grep sprn boot_halt_v4.instr.log | awk '{ printf("%s
> %s\n",$11,$12) }' | sort | uniq -c| sort -rn 2461 0x01b SRR1
>    2256 0x110 SPRG0
>    2245 0x111 SPRG1
>    2004 0x01a SRR0
>    1538 0x113 SPRG3
>    1127 0x114 SPRG4
>    1034 0x3b2 MMUCR
>     979 0x016 DEC
>     597 0x030 PID
>     509 0x115 SPRG5
>     499 0x117 SPRG7
>     475 0x150 TSR
>     358 0x03d DEAR
>     106 0x03e ESR
>       3 0x11f PVR
>       1 0x19f IVOR15
>       1 0x19e IVOR14
>       1 0x19d IVOR13
>       1 0x19c IVOR12
>       1 0x19b IVOR11
>       1 0x19a IVOR10
>       1 0x199 IVOR9
>       1 0x198 IVOR8
>       1 0x197 IVOR7
>       1 0x196 IVOR6
>       1 0x195 IVOR5
>       1 0x194 IVOR4
>       1 0x193 IVOR3
>       1 0x192 IVOR2
>       1 0x191 IVOR1
>       1 0x190 IVOR0
>       1 0x03f IVPR

Good stuff.

> While the instruction statistics for all instructions in that log is:
> $awk '{ printf("%s\n",$8) }' boot_halt_v4.instr.log | sort | uniq -c | sort
> -rn 9157 mtspr

What I find interesting is that the top 7 SPR traps are for registers that 
have no immediate effect on execution, and that means they could easily be 
replaced with memory references (in both mtspr and mfspr cases).

>    7494 wrteei
>    7051 mfspr
>    6073 mfmsr

Another good candidate for rewriting as a memory reference.

>    4532 wrtee

Hmm, wrtee(i) would be more tricky to replace with memory references, at least 
without nop padding in the guest...

>    1693 rfi
>    1647 tlbwe
>     943 mtmsr
>      89 lbz
>      84 stb
>      81 tlbsx
>      22 mfdcr
>      21 lbzx
>      20 mtdcr
>       7 stbx
>       3 sthbrx
>       3 iccci

-- 
Hollis Blanchard
IBM Linux Technology Center