llvm dev - Nov 2019 - Alias analysis only throwing mayAlias for something that seems should be identifiable as mustAlias

I have a global 2-D array ARRAY[N][M] and I am accessing it inside the for
loop like this:

for (i...)
  for (j ...)
    ARRAY[i][j] ...

So nothing really weird is happening. If I look at the generated IR, it is
also fairly straight forward.

@ARRAY0 = dso_local global [32 x [32 x i32]] zeroinitializer, section
".slow_mem", align 32, !dbg !84
...
%45 = getelementptr inbounds [32 x [32 x i32]], [32 x [32 x i32]]* @ARRAY0,
i32 0, i32 %22, i32 7, !dbg !180
store volatile i32 %43, i32* %45, align 4, !dbg !181, !tbaa !148

As you can see, the GEPInst exactly knows it is the ARRAY0, and it knows
statically that it is accessing the 7th element.
However, when I try to call alias analysis in the LLVM pass, as in

AliasAnalysis *AA = ...
AA->alias(a, b)

where a is the ARRAY0 global variable and b is the GEPInst (%45), I only
get mayAlias.
When reading the basicaa description, it seems like it should know that the
two is must alias. What am I missing?

I did not know how I should post code in the thread, so I apologize for the
messy format.
Thank you,
Best Regards,
Kiwan
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20191105/7da47957/attachment-0001.html>

Hi, Kiwan,

What sizes are you passing to AA when you make the query (i.e., how are you
constructing the MemoryLocation)? I would expect an access to the eighth element
and the first element to be NoAlias, although maybe MayAlias if something is
confusing the analysis.

 -Hal

On 11/5/19 6:05 PM, 맹기완 via llvm-dev wrote:
I have a global 2-D array ARRAY[N][M] and I am accessing it inside the for loop
like this:

for (i...)
  for (j ...)
    ARRAY[i][j] ...

So nothing really weird is happening. If I look at the generated IR, it is also
fairly straight forward.

@ARRAY0 = dso_local global [32 x [32 x i32]] zeroinitializer, section
".slow_mem", align 32, !dbg !84
...
%45 = getelementptr inbounds [32 x [32 x i32]], [32 x [32 x i32]]* @ARRAY0, i32
0, i32 %22, i32 7, !dbg !180
store volatile i32 %43, i32* %45, align 4, !dbg !181, !tbaa !148

As you can see, the GEPInst exactly knows it is the ARRAY0, and it knows
statically that it is accessing the 7th element.
However, when I try to call alias analysis in the LLVM pass, as in

AliasAnalysis *AA = ...
AA->alias(a, b)

where a is the ARRAY0 global variable and b is the GEPInst (%45), I only get
mayAlias.
When reading the basicaa description, it seems like it should know that the two
is must alias. What am I missing?

I did not know how I should post code in the thread, so I apologize for the
messy format.
Thank you,
Best Regards,
Kiwan



_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev


--
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/20191109/4a1a4a83/attachment.html>

Dear Hal,
Thanks for your reply.
I was initially not giving the LocationSize, which seems to have been the
problem.
I tried giving the size of the load/store for the first argument and size
of the entire array for the second argument and it is now giving MustAlias!
(Although I will have to check more if this is doing a correct thing in
general)

e.g., for future reference,
I have 'a' which is the argument of a load/store and 'b' which
is the
global array. I am doing

const auto &DL = m->getDataLayout();
LocationSize LS0 = LocationSize::precise(
                                        DL.getTypeStoreSize(a->getType()));
LocationSize LS1 = LocationSize::precise(
                                        DL.getTypeStoreSize(b->getType())
*ARRAY_SIZE );
AA->alias(a, LS0, b, LS1);

Thank you for your help,
Best Regards,
Kiwan

2019년 11월 9일 (토) 오전 4:40, Finkel, Hal J. <hfinkel at anl.gov>님이 작성:
> Hi, Kiwan,
>
> What sizes are you passing to AA when you make the query (i.e., how are
> you constructing the MemoryLocation)? I would expect an access to the
> eighth element and the first element to be NoAlias, although maybe MayAlias
> if something is confusing the analysis.
>
>  -Hal
> On 11/5/19 6:05 PM, 맹기완 via llvm-dev wrote:
>
> I have a global 2-D array ARRAY[N][M] and I am accessing it inside the for
> loop like this:
>
> for (i...)
>   for (j ...)
>     ARRAY[i][j] ...
>
> So nothing really weird is happening. If I look at the generated IR, it is
> also fairly straight forward.
>
> @ARRAY0 = dso_local global [32 x [32 x i32]] zeroinitializer, section
> ".slow_mem", align 32, !dbg !84
> ...
> %45 = getelementptr inbounds [32 x [32 x i32]], [32 x [32 x i32]]*
> @ARRAY0, i32 0, i32 %22, i32 7, !dbg !180
> store volatile i32 %43, i32* %45, align 4, !dbg !181, !tbaa !148
>
> As you can see, the GEPInst exactly knows it is the ARRAY0, and it knows
> statically that it is accessing the 7th element.
> However, when I try to call alias analysis in the LLVM pass, as in
>
> AliasAnalysis *AA = ...
> AA->alias(a, b)
>
> where a is the ARRAY0 global variable and b is the GEPInst (%45), I only
> get mayAlias.
> When reading the basicaa description, it seems like it should know that
> the two is must alias. What am I missing?
>
> I did not know how I should post code in the thread, so I apologize for
> the messy format.
> Thank you,
> Best Regards,
> Kiwan
>
> _______________________________________________
> LLVM Developers mailing listllvm-dev at
lists.llvm.orghttps://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
> --
> 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/20191109/39b5805e/attachment.html>