[LLVMdev] Basic AliasAnalysis: Can GEPs with the same base but different constant indices into a struct alias?

[Ahmed Bougacha]

On Tue, Jan 20, 2015 at 12:57 PM, Daniel Berlin <dberlin at dberlin.org> wrote:
> Are you counting unions, or literally just structs?

Just structs.

-Ahmed

>
> On Tue Jan 20 2015 at 12:34:38 PM Ahmed Bougacha <ahmed.bougacha at gmail.com>
> wrote:
>>
>> Hi all,
>>
>> This is covered by (struct-path aware) TBAA, but BasicAA disagrees.
>> See the attached testcase, where it prevents us from removing the
>> redundant load.
>> For arbitrary GEPs, we can't decide based on constant indices (because
>> of e.g., &A[0][1] and &A[1][0], with *A a one-element array).  BasicAA
>> has some logic to "try to distinguish something like &A[i][1] against
>> &A[42][0]".  For the testcase, it works for instance when the struct
>> has an even number of floats.
>>
>> However, I think we can safely say GEPs with different constant
>> indices into a struct can't alias, at least when:
>> - the GEP is inbounds
>> - the access sizes are such that there is no overlap
>> - the struct index is the final one
>>
>> That is, these can't alias:
>>     gep i1, j1, k1, 0
>>     gep i2, j2, k2, 1
>> If this is a struct pointer:
>>     gep i1, j1, k1
>>
>> I couldn't come up with a counterexample that would prevent us from
>> doing this in BasicAA.
>> I was surprised it wasn't; surely I missed something?
>>
>> Thanks!
>>
>> -Ahmed
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev