[patch][graphite] Fix PR43306, fix leslie3d and gamess compile time, fix documentation

Richard Guenther richard.guenther@gmail.com
Wed Mar 10 16:55:00 GMT 2010


On Wed, Mar 10, 2010 at 5:22 PM, Sebastian Pop <sebpop@gmail.com> wrote:
> On Tue, Mar 9, 2010 at 13:53, Sebastian Pop <sebpop@gmail.com> wrote:
>> Hi,
>> I committed these patches to the Graphite branch for further testing.
>>
>> For 43306, we had a problem in the analysis of data reference
>> accesses: basically we were allowing {0, +, {0, +, 4}_1}_2 as a simple
>> access function, but the translation of this scev into a polyhedral
>> representation fails as this is not a linear access function: the
>> stride in loop_2 is not constant, it has an evolution in loop_1.  Note
>> that {{0, +, 4}_1, +, 1}_2 is a linear access function as the strides
>> in loop_1 and in loop_2 are constant.
>>
>> To fix leslie3d and gamess from the CPU2006 benchmarks, I added a
>> limitation to the number of parameters that are allowed in a scop.  From
>> now on the scops with more than 10 parameters will be discarded.  This
>> limit can be changed using a parameter.  We should also improve the
>> detection of scop parameters to detect relations between parameters.
>>
>> I also have added a parameter for the number of basic blocks per
>> function that we currently have hardly fixed to 100.
>>
>> I added the documentation for the undocumented loop blocking factor.
>>
>
> With these changes we have all the SPEC2006 benchmarks passing
> compile and runtime tests.  See the automatic test results on:
> http://groups.google.com/group/gcc-graphite-test

Great!

Richard.

> Sebastian
>



More information about the Gcc-patches mailing list