[PATCH][Polly] Re: 0001-Fixup-assert-fails-caused-by-incorrect-LoopInfo-update

[Tobias Grosser]

On 09/01/2013 08:01 PM, Star Tan wrote:
> At 2013-08-25 02:10:02,"Tobias Grosser" <tobias at grosser.es> wrote:
>> On 08/23/2013 08:25 AM, Star Tan wrote:
>>>   Hi all,
>>>
>>>
>>>   I have attached a patch file to fix up the CodeGen assert fail, which is also reported in a recent bug  http://llvm.org/bugs/show_bug.cgi?id=16944
>>
>> Hi Star Tan,
>>
>> thanks for looking into this bug. The patch itself looks good and the
>> testcase is nicely small.
>>
>> However, when running the test with 'opt' I get:
>>
>> PHINode should have one entry for each predecessor of its parent basic
>> block!
>>     %indvar = phi i64 [ 0, %entry ], [ %0, %for.inc ]
>>
>> So the IR in the test case itself is already incorrect.
>
> Sorry for the stupid mistake. I have attached a new patch file!

Great. This one looks correct.

>Also, I am a little unhappy with the fact that we need to run
>
>> -polly-opt-isl to reproduce this bug. I would very much prefer to find a
>> test case that only runs 'opt %loadPolly -polly-codegen  < %s'. This ensures
>> that changes in -polly-opt-isl can not accidentally stop this test case
>> from testing the code path you fixed in this patch.
>
> Yes, you are right. But it seems difficult to construct a testcase that will show this bug using only "opt %loadPolly -polly-codegen".
> Our testcase fails with the following command:
> $ opt %loadPolly -basicaa -polly-opt-isl -polly-codegen test.ll
>
> but it works well with the following commands:
> $ opt %loadPolly -basicaa -polly-opt-isl -S -o test_1.ll test.ll
> $ opt %loadPolly -basicaa -polly-codegen test_1.ll
>
> The key of producing the bug is to construct a stmt_guard statment, so it will call codegen((clast_guard *)). Preliminary investigation shows that the  'stmt_guard' occurs only if we run opt-isl before codegen.  Sorry I have not found such a testcase by now.

I just looked into this and I could also not find a trivial test case 
that yields a guard statement. However, you can use

opt -polly-opt-isl -polly-export-jscop

to get a jscop file and the use opt -polly-import-jscop to read in this
jscop file instead of running polly-opt-isl. This at least removes the 
dependency on the isl scheduler.

test/Cloog/CodeGen/simple_vec_stride_one.ll

The above test case should give an example how to use the import/export 
of jscops.

Thanks,
Tobias