[Qemu-devel] gcc as a backend to generate better code

[Qemu-devel] gcc as a backend to generate better code

[Johan Rydberg]

The TODO file states the following:

- use gcc as a backend to generate better code (easy to do by using
  op-i386.c operations as local inline functions).

How would this be done, actually?

-- 
Johan Rydberg, Free Software Developer, Sweden
http://rtmk.sf.net | http://www.nongnu.org/guss/

Playing david kitt - saturdays

Re: [Qemu-devel] gcc as a backend to generate better code

[Fabrice Bellard]

Johan Rydberg wrote:
> The TODO file states the following:
> 
> - use gcc as a backend to generate better code (easy to do by using
>   op-i386.c operations as local inline functions).
> 
> How would this be done, actually?

The executable is launched at least once with the standard QEMU to 
record translation block statistics, location and code reference (from 
which file it comes, at which offset).

Then an off-line tool is launched to convert recorded translation blocks 
to a .so file using gcc by generating C code from each micro operation.

When QEMU is launched again, it can scan a directory containing .so 
files for all statically compiled programs. Then it can load the 
corresponding .so file and use the translated blocks stored inside.

With this method, only heavily used translations blocks are compiled 
with gcc. The rest is translated with QEMU.

Fabrice.

Re: [Qemu-devel] gcc as a backend to generate better code

[Chad Page]

	Sounds interesting.  Perhaps the off-line tool could be run
directly from qemu to facilitate ease of use.  Then to the end user it
would appear to run faster each time you use it.

	Also if the off-line tool could successfully dissaemble enough of
the program without running it, it could be precompiled before the first
invocation.  It might be too difficult to follow all code paths, though,
but you'ld know that :)

	- Chad


On Wed, 15 Oct 2003, Fabrice Bellard wrote:

> Johan Rydberg wrote:
> > The TODO file states the following:
> > 
> > - use gcc as a backend to generate better code (easy to do by using
> >   op-i386.c operations as local inline functions).
> > 
> > How would this be done, actually?
> 
> The executable is launched at least once with the standard QEMU to 
> record translation block statistics, location and code reference (from 
> which file it comes, at which offset).
> 
> Then an off-line tool is launched to convert recorded translation blocks 
> to a .so file using gcc by generating C code from each micro operation.
> 
> When QEMU is launched again, it can scan a directory containing .so 
> files for all statically compiled programs. Then it can load the 
> corresponding .so file and use the translated blocks stored inside.
> 
> With this method, only heavily used translations blocks are compiled 
> with gcc. The rest is translated with QEMU.
> 
> Fabrice.
> 
> 
> 
> _______________________________________________
> Qemu-devel mailing list
> Qemu-devel@nongnu.org
> http://mail.nongnu.org/mailman/listinfo/qemu-devel
>