[llvm-dev] Alias semantic

[Hal Finkel via llvm-dev]

On 03/09/2017 03:14 PM, Alexandre Isoard wrote:
>
>
> On Thu, Mar 9, 2017 at 8:03 PM, Hal Finkel <hfinkel at anl.gov 
> <mailto:hfinkel at anl.gov>> wrote:
>
>
>     On 03/09/2017 01:39 PM, Alexandre Isoard via llvm-dev wrote:
>>     Hello fellow LLVM developers,
>>
>>     I was wondering about the semantic of aliasing.
>>     Here are some examples where I am not sure if the two stores aliases:
>>
>>     Example 1:
>>
>>     for (int i = 2; i < n; ++i) {
>>       store A[i];
>>       store A[i-2];
>>     }
>>
>>     Example 2:
>>
>>     for (int i = 2; i < n; ++i)
>>       store A[i];
>>     for (int i = 2; i < n; ++i)
>>       store A[i-2];
>>
>>     In the example 1, they do not alias in a single iteration, but
>>     they do alias at some point if n > 4.
>
>     If you ask LLVM's AA whether 'store A[i]' and 'store A[i-2]'
>     alias, it should say NoAlias. When two memory accesses have
>     addresses that depend on common SSA values (the PHI node %i in
>     this case), then it is interpreted as having the same value in
>     both instructions.
>
>     In example 2, you should get a MayAlias result, the PHI in both
>     loops will be distinct SSA values (even though you've given them
>     the same name in your example), and we can't prove that the memory
>     accessed by the first store is disjoint from the memory accessed
>     by the second.
>
>     In any case, please don't confuse our AA for a loop dependence
>     analysis. Are you looking for a loop dependence analysis?
>
>
> That is most certainly what I am looking for.
> I was wrongly assuming that AliasAnalysis would be a cheap 
> over-approximation, I was wrong.

Not in that sense. If you construct a MemoryLocation object from both 
accesses, set both sizes to UnknownSize, and then check AA->alias(L1, 
L2) != NoAlias && AA->alias(L2, L1) != NoAlias, then you can essentially 
get this kind of cheap over-approximation.

You can also use the dependence analysis if you'd like 
(LoopAccessAnalysis, which is what the loop vectorizer uses). There is 
also an DependenceAnalysis analysis, which is more powerful in some ways.

  -Hal

>
> Thanks.
>
>      -Hal
>
>>     That is, there is an execution path from one instance of the
>>     store to an instance of the other store such that they access the
>>     same memory cell.
>>
>>     In example 2 it is easier to see the difference.
>>
>>     I am somewhat in the fog here.
>>
>>     Best regard.
>>
>>     -- 
>>     *Alexandre Isoard*
>>
>>
>>     _______________________________________________
>>     LLVM Developers mailing list
>>     llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
>>     http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>     <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
>
>     -- 
>     Hal Finkel
>     Lead, Compiler Technology and Programming Languages
>     Leadership Computing Facility
>     Argonne National Laboratory
>
> -- 
> *Alexandre Isoard*
-- 
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170309/d5114974/attachment.html>