[llvm-dev] stale info in the assumption cache

[Dmitriev, Serguei N via llvm-dev]

I have uploaded a patch with the straightforward fix (clearing function's assumption cache after extracting blocks from it) - https://reviews.llvm.org/D57215. Please take a look.

Thanks,
Sergey

From: Finkel, Hal J. [mailto:hfinkel at anl.gov]
Sent: Thursday, January 17, 2019 9:45 AM
To: Dmitriev, Serguei N <serguei.n.dmitriev at intel.com>; llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] stale info in the assumption cache

On 1/17/19 11:19 AM, Dmitriev, Serguei N wrote:
Hi Hal,

On 1/16/19 6:59 PM, Dmitriev, Serguei N via llvm-dev wrote:
Hi all,

We have recently encountered a problem with AssumptionCache that it does not get updated when a block with llvm.assume calls gets outlined by the CodeExtractor. As a result we end up with stale references to the llvm.assume calls that were moved to the outlined function in the parent function's cache.

The problem can be reproduced on the attached file as follows (many thanks to Andy Kaylor for creating this reproducer)

$ opt -slp-vectorizer -hotcoldsplit -slp-vectorizer -S -o - extract.ll
opt: .../llvm/lib/Analysis/CodeMetrics.cpp:107: static void llvm::CodeMetrics::collectEphemeralValues(const llvm::Function*, llvm::AssumptionCache*, llvm::SmallPtrSetImpl<const llvm::Value*>&): Assertion `I->getParent()->getParent() == F && "Found assumption for the wrong function!"' failed.
...

The comments in the AssumptionCache seem to indicate that we should never get into this situation because AssumptionCache is supposed to be self-updating, but the case when a block with llvm.assume call is moved from one function to another is not correctly handled now. I wonder what the right way to fix this problem is.

I guess a straightforward solution would be just changing CodeExtractor to clear function's assumption cache right after extracting blocks from it, but that would require adding assumption analysis as a dependency to every pass that uses CodeExtractor.



I don't, on first impression, have a better idea.


Ok, I will probably implement this solution so far.

Or maybe there could be other solutions like for example somehow extending callback mechanism to handle updates of the basic block parent. Can you please suggest any ideas?



I don't think that our current ValueHandle callbacks will give us this functionality. We'd need a new callback?


I also think that existing callbacks cannot handle the case when block's parent is changed, so something new will be needed. AssumptionCache as I understand was designed to be self-updating, so maybe it would be nice to maintain this property.



I agree. Unfortunately, when writing the code I didn't think about this use case. Also, we need to weigh the cost of adding the callback vs. updating the more-rare transformation code. If the callback would appear on an otherwise hot path, it might not be the right trade off.

 -Hal



Thanks,
Sergey


Thanks again,

Hal



Thanks,
Sergey




_______________________________________________

LLVM Developers mailing list

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory

--

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/20190205/23ffd875/attachment.html>