[Qemu-devel] performance of qemu

[Qemu-devel] performance of qemu

[Johan Rydberg]

Hi,

has anyone of you guys (Fabrice et al) done any performance measurements of
QEMU?  I'm particular intrested in the performance of functions that
set the condition codes.  

If I'm not mistaken, the condition codes is simulated using two values and
an operator specifying what type of insn that generated the condition codes.

When is the full eflags register generated?  In the epilogue of a translated
block, or when ever an insn/callback reads the eflag register?

Is any optmization done in the case that two insns that kill the condition 
codes? (example: two insns that kills the condition codes, but no insn between 
them that uses it).

brgds,
Johan.

Re: [Qemu-devel] performance of qemu

[Fabrice Bellard]

Johan Rydberg wrote:
> Hi,
> 
> has anyone of you guys (Fabrice et al) done any performance measurements of
> QEMU?  I'm particular intrested in the performance of functions that
> set the condition codes.  

I plan to publish soon some benchmarks against bochs and native execution.

> If I'm not mistaken, the condition codes is simulated using two values and
> an operator specifying what type of insn that generated the condition codes.

OK.

> When is the full eflags register generated?  In the epilogue of a translated
> block, or when ever an insn/callback reads the eflag register?

Only when an instruction really needs the full eflags register (such as 
pushf).

> Is any optmization done in the case that two insns that kill the condition 
> codes? (example: two insns that kills the condition codes, but no insn between 
> them that uses it).

Yes.

Fabrice.

Re: [Qemu-devel] performance of qemu

[Johan Rydberg]

On Wed, 04 Jun 2003 10:20:24 +0200
Fabrice Bellard <fabrice.bellard@free.fr> wrote:

: I plan to publish soon some benchmarks against bochs and native execution.
: 
: > If I'm not mistaken, the condition codes is simulated using two values and
: > an operator specifying what type of insn that generated the condition codes.
: 
: OK.

Good.
 
: > When is the full eflags register generated?  In the epilogue of a translated
: > block, or when ever an insn/callback reads the eflag register?
: 
: Only when an instruction really needs the full eflags register (such as 
: pushf).

And at the end of the translated block?  Or does the CC_SRC, CC_DST and CC_OP
live between blocks?

Re: [Qemu-devel] performance of qemu

[Fabrice Bellard]

Johan Rydberg wrote:
> : > When is the full eflags register generated?  In the epilogue of a translated
> : > block, or when ever an insn/callback reads the eflag register?
> : 
> : Only when an instruction really needs the full eflags register (such as 
> : pushf).
> 
> And at the end of the translated block?  Or does the CC_SRC, CC_DST and CC_OP
> live between blocks?

CC_SRC, CC_DST and CC_OP live between blocks (I think computing the 
eflags would be too expensive).

Fabrice.

Re: [Qemu-devel] performance of qemu

[Johan Rydberg]

On Wed, 04 Jun 2003 18:01:25 +0200
Fabrice Bellard <fabrice.bellard@free.fr> wrote:

: CC_SRC, CC_DST and CC_OP live between blocks (I think computing the 
: eflags would be too expensive).

It most probably would.  How do callbacks, say for example in the
signal handing code of QEMU, retreive the EFLAGS value from the CC_*
variables?  Do it call some function that calculates the value, or
does the main loop calculate the full EFLAGS value whenever it breaks
out of a sequence of translated blocks to handle signals?
 
: Fabrice.

Johan