[Openmp-commits] [PATCH] D62393: [OPENMP][NVPTX]Mark parallel level counter as volatile.

Tue Oct 8 12:03:22 PDT 2019

jlebar added a comment.

I spoke offline with tra@ about this; he was confused why I was coming down so hard here.

For lack of a better word, I think I have some PTSD from customers hitting by ptxas bugs and then expecting me to fix/work around them.  I know that even very simple CUDA code can hit bugs in old ptxas.  There are even data-dependent bugs which only manifest with certain pointer values.  This makes it difficult to have confidence that any particular function works properly, even if you've tested it thoroughly.  However awful you think it is, it's worse.

We've been working on this for a long time and have found that the only thing which consistently works is using the newest ptxas.  Upgrading is painful, but using a buggy optimizing assembler is worse.

I think the root reason I came down hard here is because I was/am afraid that this change would be used as justification for trying to work around even a subset of these bugs in LLVM itself.  That would be a Sisyphean task and so I don't want us to sign up for it.

If you all are willing to say that this patch won't be used as precedent to argue in favor of making changes to clang or LLVM for compatibility with ptxas 8.0, then I withdraw my objections here.

Repository:
  rOMP OpenMP

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D62393/new/

https://reviews.llvm.org/D62393