[LLVMdev] dragonegg polly support broken?

[Jack Howarth]

On Sun, Oct 21, 2012 at 03:52:39PM -0700, Tobias Grosser wrote:
> On 10/21/2012 03:46 PM, Jack Howarth wrote:
>> On Sun, Oct 21, 2012 at 02:31:50PM -0700, Tobias Grosser wrote:
>>> On 10/21/2012 12:47 PM, Jack Howarth wrote:
>>>> On Sun, Oct 21, 2012 at 11:58:43AM -0700, Tobias Grosser wrote:
>>>> ...
>>>>> Does attached patch work for you?
>>>>>
>>>>> Tobi
>>>>>
>>>>
>>>> Tobi,
>>>>     Yes, with this patch current llvm/polly svn builds and runs 'make polly-test'
>>>> without issues in finding LLVMPolly.so. Thanks for fixing this so quickly.
>>>
>>> Thanks for the nice bugreport. I committed a fix in 166402
>>>
>>>>     Do you have any comments on failing tests we are seeing on x86_64-apple-darwin11?
>>>>
>>>> Failing Tests (5):
>>>>       Polly :: CodeGen/do_pluto_matmult.ll
>>>>       Polly :: CodeGen/single_do_loop_int_max_iterations.ll
>>>>       Polly :: CodeGen/single_do_loop_scev_replace.ll
>>>
>>> I have some guesses for these, but it would be good to get run
>>>
>>> llvm-lit -v on these three test cases.
>>>
>>>>       Polly :: IslCodeGen/single_loop_param_less_equal.ll
>>>>       Polly :: IslCodeGen/single_loop_param_less_than.ll
>>
>> Tobias,
>>     I am unclear on how I should modify the 'make polly-test' run in order
>> for 'llvm-lit -v' to take effect and where to look for the additional information.
>> Thanks in advance for any clarifications.
>
> After you run 'make polly-test' once, you can just run
>
> $ llvm-lit -v ${POLLY_SRC}/test
>
> Like this you can easily pass the verbose flag.
>
>>> This is due to the use of an old isl version. I will change configure to
>>> fail if the isl version used is too old.
>>>
>>>> Are these darwin-specific or due to the fact that I am using isl 0.10 and
>>>> cloog 0.17 rather than their the upstream development svn? Let me know if
>>>> any of those merit a bugzilla of their own.
>>>
>>
>> It looks like MacPorts could easily switch their llvm-3.2 packaging to build polly
>> except that, unlike fink which uses cmake, they are still using configure for the
>> build. This fails against isl 0.10 due to the test...
> >
>>
>> checking for isl: isl/ast.h in /opt/local/include, and libisl in /opt/local/lib... configure: error: isl required but not found
>
> cmake should fail the same way (if it is reconfigured). I just committed  
> this changes to make ensure the isl version we are using
> contains all the features Polly is using.
>
>> Is there a schedule for releasing updated tarballs for isl and cloog?
>
> I am not aware of a concrete schedule, but I believe Sven (copied) was  
> planning to release a new version of isl at some point not too far away.
>
> Cheers
> Tobi

Tobi,
   I can confirm that your polly changes now prevents cmake from building against isl 0.10.
After updating my isl/cloog packages to the current git commit level, the failures are reduced
to...

Failing Tests (3):
    Polly :: CodeGen/do_pluto_matmult.ll
    Polly :: CodeGen/single_do_loop_int_max_iterations.ll
    Polly :: CodeGen/single_do_loop_scev_replace.ll

  Expected Passes    : 125
  Expected Failures  : 29
  Unsupported Tests  : 2
  Unexpected Failures: 3

The output of 'llvm-lit -v' is attached to the bugzilla I opened...

http://llvm.org/bugs/show_bug.cgi?id=14142

Let me know if I should have provided more than that from the 'llvm-lit -v' run.
          Jack