This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 0/3] Configure support for official CLooG versions.

 On 08/12/2010 03:39 PM, Jack Howarth wrote:
> On Thu, Aug 12, 2010 at 12:25:21AM +0200, Andreas Simbuerger wrote:
>>  On 08/11/2010 11:55 PM, Joseph S. Myers wrote:
>>> On Wed, 11 Aug 2010, Andreas Simbürger wrote:
>>>> this patch series adds configure support for upstream CLooG versions.
>>>> The official CLooG is able to use three different backends during
>>>> code generation. An ISL backend, a PPL backend and a PolyLib backend.
>>> Does ISL involve another library that would be linked against instead of 
>>> PPL, that needs installing separately from CLooG?  If so, this patch 
>>> series appears incomplete, as it doesn't include any addition of 
>>> information about ISL to install.texi, or --with-isl etc. options 
>>> analogous to the --with-ppl options, or configure code that adds -lisl or 
>>> similar (as opposed to -lcloog-isl).  All these things are needed for 
>>> making GCC link with a new host-side library.  In addition it will be 
>>> necessary to check the portability of ISL to different hosts, similarly to 
>>> how people tested PPL on various hosts before the original Graphite merge, 
>>> including cases such as building it with cross compilers.
>> The Integer Set Library (ISL) is used within CLooG's backend instead
>> of the PPL (Another backend using the PolyLib is available too).
>> The official CLooG version provides it's own ISL bundle which
>> get's installed alongside CLooG itself. The ISL is exclusively used by
>> CLooG for now.
> Andreas,
>     Are there any benchmarks comparing the performance of graphite with
> cloog-isl, cloog-polylib and cloog-ppl? In particular, how do the three
> compare in effectiveness of properly finding loops and are there large
> differences in the time they require to process the data? I had assumed
> that cloog-ppl built against ppl-0.11 would be the preferred option
> for graphite in gcc 4.6.
>                     Jack

Tests will follow. We wanted to discuss the configure changes in the
first place. After that we apply them to the graphite branch and
evaluate the performance of the backends. We assume
that the CLooG-ISL will be superior over the alternatives in most cases.

The CLooG-PPL fork that is used now should be deprecated, as it is
not maintained from upstream at all.

I will post benchmark results as i get the patches applied to the
graphite branch.

>> This patch series only prepare GCC for the transition from the CLooG-PPL
>> fork,
>> maintained by Graphite to the CLooG versions maintained by upstream.
>> Regards,
>> Andreas

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]