[llvm-dev] How to best deal with undesirable Induction Variable Simplification?

[Philip Reames via llvm-dev]

On 8/9/19 8:27 AM, Danila Malyutin via llvm-dev wrote:
>
> Hi Hal,
>
>  
>
> I see. So LSR could theoretically counteract undesirable Ind Var
> transformations but it’s not implemented at the moment?
>
> I think I’ve managed to come up with a small reproducer that can also
> exhibit similar problem on x86, here it is: https://godbolt.org/z/_wxzut
>
>  
>
> As you can see, when rewriteLoopExitValues is not disabled Clang
> generates worse code due to additional spills, because Ind Vars
> rewrites all exit values of ‘a’ to recompute it’s value instead of
> reusing the value from the loop body. This requires extra registers
> for the new “a after the loop” value (since it’s not simply reused)
> and also to store the new “offset”, which leads to the extra spills
> since they all live across big loop body. When exit values are not
> rewritten ‘a’ stays in it’s `r15d` register with no extra costs.
>
This hits on a point I've thought some about, but haven't tried to
implement.

I think there might be room for a late pass which undoes the exit value
rewriting.  As an analogy, we have MachineLICM which sometimes undoes
the transforms performed by LICM, but we still want the IR form to hoist
aggressively for ease of optimization and analysis.  

Maybe this should be part of LSR, or maybe separate.  Haven't thought
about that part extensively.

It's worth noting that the SCEVs for the exit value of the value inside
the loop and the rewritten exit value should be identical.  So
recognizing the case for potential rewriting is quite straight-forward. 
The profitability reasoning might be more involved, but the legality
part should essentially be handled by SCEV, and should be able to reuse
exactly the same code as RLEV. 

>  
>
> --
>
> Danila
>
>  
>
> *From:* Finkel, Hal J. [mailto:hfinkel at anl.gov]
> *Sent:* Thursday, August 8, 2019 21:24
> *To:* Danila Malyutin <Danila.Malyutin at synopsys.com>
> *Subject:* Re: [llvm-dev] How to best deal with undesirable Induction
> Variable Simplification?
>
>  
>
> Hi, Danila,
>
>  
>
> Regarding the first case, this is certainly a problem that has come up
> before. As I recall, and I believe this is still
> true, LoopStrengthReduce, where we reason about induction variables
> while accounting for register pressure, won't currently add new PHIs.
> People have talked about extending LSR to consider adding new PHIs in
> the past.
>
>  
>
> Regarding the second case, could you post a more-detailed description?
> I don't quite understand the issue.
>
>  
>
>  -Hal
>
>  
>
> Hal Finkel
> Lead, Compiler Technology and Programming Languages
> Leadership Computing Facility
> Argonne National Laboratory
>
>  
>
> ------------------------------------------------------------------------
>
> *From:*llvm-dev <llvm-dev-bounces at lists.llvm.org
> <mailto:llvm-dev-bounces at lists.llvm.org>> on behalf of Danila Malyutin
> via llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
> *Sent:* Thursday, August 8, 2019 12:36 PM
> *To:* llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
> <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
> *Subject:* [llvm-dev] How to best deal with undesirable Induction
> Variable Simplification?
>
>  
>
> Hello,
> Recently I’ve come across two instances where Induction Variable
> Simplification lead to noticable performance regressions.
>
> In one case, the removal of extra IV lead to the inability to
> reschedule instructions in a tight loop to reduce stalls. In that
> case, there were enough registers to spare, so using extra register
> for extra induction variable was preferable since it reduced
> dependencies in the loop.
> In the second case, there was a big nested loop made even bigger after
> unswitching. However, the inner loop body was rather simple, of the form:
>
> loop {
>
>   p+=n;
>
> …
>
>   p+=n;
>
> …
>
> }
> use p.
>
>  
>
> Due to unswitching there were several such loops each with the
> different number of p+=n ops, so when the IndVars pass rewrote all
> exit values, it added a lot of slightly different offsets to the main
> loop header that couldn’t fit in the available registers which lead to
> unnecessary spills/reloads.
>
> I am wondering what is the usual strategy for dealing with such
> “pessimizations”? Is it possible to somehow modify the IndVarSimplify
> pass to take those issues into account (for example, tell it that
> adding offset computation + gep is potentially more expensive than
> simply reusing last var from the loop) or should it be recovered in
> some later pass? If so, is there an easy way to revert IV elimination?
> Have anyone dealt with similar issues before?
>
>  
>
> --
>
> Danila
>
>  
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20190809/520de0c5/attachment.html>