[llvm-dev] stale info in the assumption cache

Finkel, Hal J. via llvm-dev llvm-dev at lists.llvm.org
Thu Jan 17 00:00:29 PST 2019

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.

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?

Thanks again,



LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

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/20190117/0827f64d/attachment.html>

More information about the llvm-dev mailing list