[llvm-dev] loop invariant code hoisting

[Johannes Doerfert via llvm-dev]

On 12/7/20 9:52 AM, Florian Hahn via llvm-dev wrote:
> Hi,
>
>> On Dec 7, 2020, at 13:25, Sjoerd Meijer via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>>
>> Hello,
>>
>> I would like to restrict LICM in the number of memory operations it hoists out of a loop. Blindly hoisting out all instructions out could result in a lot of spilling/reloading, which is what I'd like to avoid. We see this for example on (inner)loops with small constant bounds that are unrolled. I am drafting something in https://reviews.llvm.org/D92488 <https://reviews.llvm.org/D92488>, which contains a reduced version of my motivating example.
>> It was brought to my attention that LICM might not have any restrictions by design because hoisting out all instructions could be some sort of canonical form that other passes depend on, so sinking back defs closer to their user(s) might be a back-end problem. I was wondering if there are any opinions on this.
>>
> I think one key benefit of hoisting as much as possible is that it enables/simplifies subsequent optimizations in the middle-end (especially for memory operations). Limiting hoisting is likely to have a bad knock-on effect on other transformations.
>
> Accurately estimating the number of spills in LICM is probably going to be tricky and/or brittle. Another thing to consider is that there are plenty of other transformations that extend live-ranges of values, so they would also need to also need updating.
>
> Sinking in the backend would catch those cases naturally, while making our lives easier in the middle-end. We already have plenty of infrastructure to reason about register pressure & co in CodeGen. As Bardia mentioned, currently there is limited support for re-materialzing instructions.
>
> Besides that, I don’t think there are dedicated passes to move defs to uses to reduce register pressure, beside the MachineScheduler. But the MachineScheduler currently only operates on sub-regions of basic blocks.
>
> I think there would be potential for having a dedicated pass to move defs across basic blocks to reduce register pressure. I think we have most building blocks (RegisterPressureTracker, MachineTraceMetrics for selecting a likely trace through the function). But reasoning about moving memory operations is probably going to be more difficult in the backend than in the middle-end though unfortunately.

+1

Assuming the transformation is not irreversible, I would continue to 
perform normalization steps like LICM and opt (pun intended) for a 
dedicated, target-specific pass for register pressure minimization.
Years ago I tried to estimate register pressure/usage in the vectorizer, 
in addition to the heuristic the vectorizer uses, and it turned out that 
you might get close but won't get it right at that level anyway... at 
least that is my experience.

~ Johannes

P.S.
   I know I'm late and people seem to have already converged on this 
conclusion anyway.


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