LLVM's loop unroller & llvm.loop.parallel_accesses

Hi, in our backend, which is unfortunately not upstreamed, we are relying on llvm.loop.parallel_accesses metadata for certain passes like software pipelining so we can re-order instructions. Ideally, we would want the loop unroller to support the notion of the loop’s parallelism in its pre-unrolled version. This probably should happen by propagating !alias.scope and !alias metadata. Is there any plan or open patch for supporting this?

Simplified example:

for.body:

%0 = load […]

store %0 […]
br label %for.cond, !llvm.loop !2

!1 = distinct !{}

!2 = distinct !{!2, !3, !4, !5, !6, !7}
!3 = !{!“llvm.loop.parallel_accesses”, !1}
!4 = !{!“llvm.loop.vectorize.width”, i32 1}
!5 = !{!“llvm.loop.interleave.count”, i32 1}
!6 = !{!“llvm.loop.vectorize.enable”, i1 true}

!7 = !{!“llvm.loop.vectorize.followup_all”, !8}

!8 = !{!“llvm.loop.unroll.count”, i32 2}

(unroll by 2) =>

for.body:

%0 = load […] !alias.scope !9 !noalias !11

store %0 […] !alias.scope !9 !noalias !11
%1 = load […] !alias.scope !10 !noalias !12

store %1 […] !alias.scope !10 !noalias !12

br label %for.cond, !llvm.loop !2

[…]

!9 = distinct !{!9, !“iteration0”}

!10 = distinct !{!10, !“iteration1”}

!11 = !{!10}

!12 = !{!9}

Thanks, Hendrik

Hi, Hendrik,

A couple of thoughts:

  1. This might be related (but, perhaps, not in the way you’d prefer): https://bugs.llvm.org/show_bug.cgi?id=39282
  2. We’re scheduling a call to discuss further improvements in this area, and if you might be able to join, please fill out the Doodle poll: https://doodle.com/poll/evhwr2eyfvcf8ib3

-Hal

llvm.loop.parallel_accesses does not imply that these accesses from
different iterations are not aliasing. Examples where an access are
parallel are that the accesses are atomic or read-only from a specific
location.

The LoopUnrollPass might deduce that non-atomic stores are necessarily
not aliasing (when not using transactional memory), but I don't think
we can do this for all the read accesses. Would that be sufficiently
useful?

Michael

This is interesting! So are you saying that loop.parallel_accesses strictly loop parallel, and says nothing about aliasing? I see, I guess we may have been “abusing” the hint and re-purposed it. But isn’t llvm’s vectorizer using loop.parallel_accesses to vectorize loops including vectorize memory accesses that if you ignore loop-carried dependencies, usually means effectively re-ordering the accesses? I guess this still does not imply “noalias”? What about icc/gcc’s #pragma ivdep? Again here, it means no loop-carried dependencies, yet still doesn’t say anything about noalias? Another way indeed would be to propagate noalias data and indeed rely on the future fix that Hal mentions above.

Actually, I guess Hal’s fix won’t help us. It seems like we may need a completely the metadata node llvm.loop.parallel_noalias with noalias semantics, that then would get propagated in the unroller as originally stated…?

Trivial example:

#pragma clang loop vectorize(assume_safety)
for (int i = 0; i < n; i+=1) {
  (void)A[0];
}

I hope it is obvious that the loop is parallel and can be vectorized,
but A[0] from iteration 0 will alias with A[0] from iteration 1.
Replace `0` by `i*c` where c is a variable that can be 0 at runtime to
make the fact non-obvious to the compiler.

We had discussions about implementing "#pragma ivdep", but it's
semantics are not defined independently of the implementation. Anyway,
even with #pragma omp ivdep, a compiler is not required to vectorize
the loop.

In LLVM, runtime/partial unrolling only takes place after
vectorization, so there is less of an issue there.

Michael

Would you guys be open to supporting a new hint with the right semantics, like e.g. llvm.loop.noalias_accesses?! I would need to find support in clang however and the main point of support would be the loop unroller behaving as stated in the OP.

What would be its semantics? When would clang attach that attribute?

Michael

Skipping the clang question for now, this had to be a loop pragma of some kind. One step back: what we really need is a way to express that memory accesses between iterations can be re-ordered. The code that’s being compiled is noalias, but we don’t have to use noalias semantics, e.g. loop parallel semantics are sufficient. What’s missing is a way to express that past the llvm unroller. When the unroller merges iterations, loop parallel no longer depicts the original iterations. So the obvious idea was using noalias scope metadata for this, and llvm.loop.noalias_acccesses would cause the unroller to propagate different scopes for each iteration. Thinkable is also to keep the llvm.loop.parallel_accesses, and the unroller propagates a new type of metadata analog to noalias scope, but loop_parallel scope or something like that. We have methods to achieve this with intrinsics, but I am looking for something more robust that also works with clang.

You should know that LLVM's alias infrastructure+metadata current does
not handle cross-iteration aliasing. The current AliasAnalysis
interface, when comparing two accesses, assumes you mean executions in
the same iteration. This is a common source of bugs such as
http://lists.llvm.org/pipermail/llvm-dev/2019-May/132725.html and the
aforementioned https://bugs.llvm.org/show_bug.cgi?id=39282.

It might be difficult to add metadata for which there isn't a concept
of in LLVM. Of course, we are looking into improving the situation,
we'd enjoy if you could join the conference call:
http://lists.llvm.org/pipermail/llvm-dev/2020-May/141653.html

At the moment, the way to find out whether accesses from different
iterations might alias is DependenceAnalysis (which has its own wrong
assumptions about AliasAnalysis, see
https://bugs.llvm.org/show_bug.cgi?id=42143). A LoopUnroll could query
DependenceAnalysis for the iterations before unrolling and add noalias
metadata if DA found that they do not alias. However, there is
currently no "cross-iteration" analogon to alias.scope/alias.metadata
that could assist DA, which is where your idea could come into play.

Note that DA generally is a computationally expensive analysis, there
could be questions whether the gain in additional noalias information
is worth the additional compile-time cost (at least up to -O2) of an
otherwise "simple" transformation like unrolling.

Michael

Do you really need to query DependenceAnalysis though if instead the unroller could propagate some kind of “parallel scope” metadata simply based on the loop’s llvm.loop.parallel_accesses?

Maybe not, in a lot of cases BasicAA might handle it just well. For
instance, A[i] and A[i+1] can be analyzed just fine. The issue I am
raising is that alias analysis within an iteration is quite different
to an analysis cross-iteration and unrolling blurs the line.

A "llvm.loop.noalias_acccesses" as you suggested might lists of
accesses that will never alias from different iterations from that
loop. That would translate to a domain for !noalias/!alias.scope
metadata for each unrolled instance. Unfortunately, also a quadratic
(in the unroll factor) number of MDOperands.

Michael

Michael