[LLVMdev] How to make Polly ignore some non-affine memory accesses

[Tobias Grosser]

On 11/03/2011 09:56 AM, Tobias Grosser wrote:
> On 11/02/2011 11:17 AM, Marcello Maggioni wrote:
>> Mmm I found out a very strange behavior (to me) of the SCEV analysis
>> of the loop bound of the external loop I posted.
>> When in ScopDetection it gets the SCEV of the external loop bound in
>> the "isValidLoop()" function with:
>> const SCEV *LoopCount = SE->getBackedgeTakenCount(L);
>>
>> It returns a SCEVCouldNotCompute, but if I change the "if" block
>> inside the loop from:
>> if (i+j> 1000)
>> to:
>> if (i> 1000)
>>
>> that SCEV results valid.
>>
>> Later in the ScopInfo pass it crashes analyzing the SCEV of the
>> comparison expression (in buildConditionSets ) . It's like if it
>> recognizes that "i" is a recurring expression that depends on the
>> execution of a loop, but can't derive that loop, and segfaults in
>> getLoopDepth ...
>>
>> Seems like a bug of the SCEV engine to me., but I'm not sure :(
>
> Hi Marcello,
>
> sorry for taking so long to reply. I just had a look at your test case
> and found the problem. What happens is that Polly fails to accept the
> outer loop (an unrelated issue), such that only the inner loop is
> detected as a scop (you can verify this with -view-scops). Now when
> building the polyhedral representation we analyze the SCEV expressions
> in the inner loop (the scop). Here the condition has an expression that
> looks similar to this:
>
> {1, +, 1024}<i> + {1,+,1024}<j>
>
> We translate this expression. When reaching {1, +, 1024}<i> we do not
> check if it is part of the scop and always handle it as a loop index.
> This is incorrect and fails when we try to find the corresponding loop
> in the scop. The right thing to do is to handle this expression as a
> parameter.
>
> Consequently this is no SCEV bug, but an ordinary Polly bug. It most
> probably slipped in when I recently refactored the SCEV parsing. I
> attached a hackish patch that should fix the issue for exactly this
> test case. Can you check it works for you?
>
> I will commit a complete fix later on.

The recent rewrite of the SCEV validation fixed this bug.

Tobi