[Mono-devel-list] proposal for a different JIT

[Varga Zoltan]

                                           Hi,

   I thought about using KSI for an AOT compiler some time
ago, but
there are some problems: the main problem is that KSI is not an
official part of gcc, and will never be, because that would
enable
the creation of possibly non-free fronts ends to gcc, which
is agains
FSF policy. So if we wanted to use KSI, we would have to
distribute
our own extended version of gcc, which is a lot of work. For an
example of a project which choose this route, see:

http://www.cs.mu.oz.au/research/mercury/download/gcc-backend.html

                    bye

                           Zoltan


Emanuele Ruffaldi <pit at sssup.it> írta:

> 
> Hello there,
> 
> I propose an alternative way for the implementation of the
JIT in the 
> Mono system. The idea is to use the back end of the GCC
compiler, but 
> instead of creating a full front-end I propose to use the
KSI front-end, 
> that is a direct mapping of the GCC back-end tree (for
details see 
> http://ftp.pld-linux.org/people/malekith/ksi/)
> 
> Whenever a part of an assembly has to be jitted the JIT
should generate 
> a KSI source code that is quite platform indepentent and
then the GCC 
> back-end can compile it with different levels of
optimizations. During 
> the generation of the KSI code all the references to
internal functions 
> and variables can be fixed as constants improving the
performance.
> 
> I've tested this solution in a Cygwin (sorry) environment
with GCC 
> 3.3.1, where a custom program generates the KSI code on
the fly, calls 
> the GCC tools (ks1, as and ld) and then loads the shared
library created 
> by the linker.
> There is no need to have a full GCC running but only 
these files 
> (almost 5.5MB):
> 
> ksi1.exe
> ld.exe
> as.exe
> cygwin1.dll
> 
> I think that working on Linux, and with some optimizations
the memory 
> requirements can be reduced.
> 
> This solution should be used only on the platforms where
doesn't exist a 
> custom written JIT (like now for the x86), because a
custom JIT is much 
> faster. I realize that there are many problems in the
adoption of this 
> solution but it should be faster than the simple
intepretation, and can 
> be adapted to the AOT too.
> 
> Thanks,
> Emanuele Ruffaldi
> 
> 
> 
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list
> 
> 
>