[LLVMdev] [Polly] Update lit config for Cloog

Tobias Grosser tobias at grosser.es
Mon Sep 2 09:12:56 PDT 2013

On 09/02/2013 07:44 AM, Star Tan wrote:
> At 2013-09-02 16:22:28,"Tobias Grosser" <tobias at grosser.es> wrote:
>> On 09/01/2013 08:02 PM, Star Tan wrote:
>>> Hi all,
>>> Attached patch file to update lit config for Cloog. Without it, Polly always skips Cloog testings when we run "make check-polly".
>> Dear Star Tan,
>> thanks a lot for the patch. It looks very reasonable, but I am wondering
>> why it was not needed before or what problem it fixes exactly. Could you
>> add some information about this to the commit message.
> I am not sure why it was not needed before; maybe it has never worked well before. The problem is that Polly never executes Cloog specific testcases no matther whether Cloog is found or not. I find this problem because I put a new testcase in test/Cloog/CodeGen/, but it is never executed.
> @Sebastian, you added the Cloog directory in r169159, including all testcases and the lit.local.cfg. The lit.local.cfg ensures that  cloog specific testcases are executed only with CLOOG_FOUND by adding the following lit commands in test/Cloog/lit.local.cfg:
>          cloog = config.root.cloog_found
>          if cloog not in ['TRUE', 'true'] :
>              config.unsupported = True
> However, there are two problems:
> First, since the cloog_found is set as "@CLOOG_FOUND@",  I think the following "sed" command should be added into Makefile:
>          sed -e "s#@CLOOG_FOUN@#$(CLOOG_FOUND)#g"
> Unfortunately such command is missed in current Polly.  If it is not needed, I am curious how could Polly determine the value of @CLOOG_FOUND@ ?
> Second, even if we add the "sed" command in Makefile, the config.root.cloog_found would be set as "yes" or null, but Polly currently compares it with "TRUE" or "true", which thus always fails and the Cloog testcases will never be executed.
> @Sebastian, could you do me a favor to have a review of this problem?  FYI, I have re-attached the patch file.

This is surprising. Even without your patch my test/lit.site.cfg file 

config.enable_gpgpu_codegen = ""
config.cloog_found = "TRUE"

(With CLOOG enabled and GPGPU_codegen disabled). Also, the CLooG test 
are run (and are failing) as expected.

I think it would be good to understand why this different behaviour can 
be observed. Sebastian, any ideas?


More information about the llvm-dev mailing list