[llvm-dev] [RFC] Abstracting over SSA form IRs to implement generic analyses

Mon Dec 28 01:09:16 PST 2020

Hi Nuno,

On Sun, Dec 27, 2020 at 4:58 PM Nuno Lopes <nunoplopes at sapo.pt> wrote:

> I’m quite sympathetic with the idea of abstracting away details of IRs and
> having a single implementation of analysis and optimizations over these
> abstractions. IMHO, that’s the main feature missing in MLIR.
> That said, the analyses you consider in the document seem a bit limited.
> The first two, at least, related with dominators are just a few lines of
> codes. Which makes me wonder if the trouble, extra complexity, etc, of
> going this generic way is worth it. Isn’t it simple to just duplicate them
> in MLIR?
>

The main analysis that motivates this work is divergence analysis, which is
a lot more than just a few lines of code. It's mentioned in the document
but not pasted in there in full precisely because it's so large and has
complexities that are very much orthogonal to the proposal itself :)

Those complexities also mean that subtle bugs have occurred in the past and
we have to assume that we might find more in the future, so duplication of
code seems unwise to me: keeping all copies in sync manually wouldn't be
fun.

It would also be a triplication of effort, actually, not just duplication,
since we need to support MachineIR in order to make GlobalISel work
properly for the AMDGPU target.

I hope this clarifies our motivation and why we think this is a good
trade-off :)

Cheers,
Nicolai

>
> I would be very supportive of a generic alias analysis, for example.
> That’s a serious amount of work and potentially complicated algorithms that
> could be reused across LLVM and various MLIR dialects. Plus LLVM could use
> a new alias analysis algorithm. But that requires going a bit further than
> your proposal. One needs an abstraction that can cover several IRs and
> provide enough information to the AA algorithm in an efficient way. Also,
> it likely needs to know about undefined behavior. But MLIR has been
> avoiding that issue for now, AFAICT. So anything on that front would likely
> be wrong, as there hasn’t been much thought about UB in MLIR. (AFAIK; I
> don’t follow MLIR closely)
> MLIR’s ODS doesn’t seem sufficient either, as it only allows (at least
> now) to attach shallow semantics attributes to instructions, similar to
> LLVM ones. Maybe that could be extended somehow?
>
> TL;DR: While I’d love to see LLVM/MLIR go into the direction you are
> proposing, I’m not sure the set of considered analyses is sufficiently
> complicated to award the investment in complicated infrastructure.
>
> Nuno
>
> ______________________________________________________
> From: Nicolai Hähnle
> Sent: 17 December 2020 19:17
> To: llvm-dev
> Subject: [llvm-dev] [RFC] Abstracting over SSA form IRs to implement
> generic analyses
>
> Hi LLVM community,
>
> Earlier this year I first proposed a new way of writing analyses that can
> be applied to multiple IRs, for example, applying the same analysis to both
> LLVM IR and MachineIR.
>
> LLVM already has some analyses like that; for example, dominator tree
> construction and loop info. However, they're all limited to looking at the
> control flow graph: basic blocks and lists of predecessors and successors.
> We want to push the envelope with a divergence analysis that is also aware
> of instructions and values, and ran into severe limitations in what we
> could do with the techniques that are commonly used in LLVM today. Some
> limitations are around which concepts are exposed generically at all,
> though the bulk of limitations revolves around readability and
> maintainability of the resulting generic code.
>
> After more evolution of the ideas and many discussions over the last few
> months, I want to raise this proposal once more -- this time with an
> extensive document:
> https://docs.google.com/document/d/1sbeGw5uNGFV0ZPVk6h8Q5_dRhk4qFnKHa-uZ-O3c4UY/edit?usp=sharing
>
> Feel free to comment on the document, though high-level discussion is
> probably best kept in this email thread.
>
> The concrete proposal is to enable 4 tools for use by generic analyses:
>
> - type erasure
> - an SsaContext context class concept with a fairly small surface area
> - dynamic polymorphism via per-analysis adapters
> - dynamic polymorphism via an LLVM-wide adapter of SsaContext
>
> The document goes to some length to explain what precisely is meant by
> each of those bullets, including code examples, as well as describing a few
> other options that we _don't_ propose, based on their relative merits.
>
> There are concrete patches that go along with the proposal and you can
> refer to for additional context. In logical sequence, they are:
> - https://reviews.llvm.org/D92924: Introduce opaque handles for type
> erasure
> - https://reviews.llvm.org/D83089: Based on the handle infrastructure,
> refactor the dominator tree with type-erased base classes that can be used
> by generic algorithms
> - https://reviews.llvm.org/D92925: Introduce an SsaContext context class
> concept for static polymorphism
> - https://reviews.llvm.org/D92926: Introduce an ISsaContext “global”
> interface class for dynamic polymorphism built on top of SsaContext and
> opaque handles
> - https://reviews.llvm.org/D83094: A new analysis (cycle info) written
> generically as non-template code using opaque handles, ISsaContext, and
> analysis-specific dynamic polymorphism via the ICycleInfoSsaContext
> interface added in the patch
>
> I would like us to get to general agreement on this thread that this is a
> direction we want to go in and that we can proceed with the proposed code
> changes.
>
> Thanks,
> Nicolai
>
>

-- 
Lerne, wie die Welt wirklich ist,
aber vergiss niemals, wie sie sein sollte.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201228/74c7698a/attachment.html>