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

Hal Finkel via Phabricator via Openmp-commits openmp-commits at lists.llvm.org
Tue Jun 11 15:59:59 PDT 2019

hfinkel added a comment.

In D62393#1539014 <https://reviews.llvm.org/D62393#1539014>, @ABataev wrote:

> In D62393#1539012 <https://reviews.llvm.org/D62393#1539012>, @jfb wrote:
> > FWIW we already support `-fms-volatile`, so there's precedent if you wanted `-fnv-volatile` (however terrible that is).
> Most probably, we don't need this option, clang should emit correct code for volatile vars in Cuda mode automatically.


That having been said, maybe this is a very simple change because it is the same as (or very similar to) what MSVolatile does?

  /// An LValue is a candidate for having its loads and stores be made atomic if
  /// we are operating under /volatile:ms *and* the LValue itself is volatile and
  /// performing such an operation can be performed without a libcall.
  bool CodeGenFunction::LValueIsSuitableForInlineAtomic(LValue LV) {
    if (!CGM.getCodeGenOpts().MSVolatile) return false;
    AtomicInfo AI(*this, LV);
    bool IsVolatile = LV.isVolatile() || hasVolatileMember(LV.getType());
    // An atomic is inline if we don't need to use a libcall.
    bool AtomicIsInline = !AI.shouldUseLibcall();
    // MSVC doesn't seem to do this for types wider than a pointer.
    if (getContext().getTypeSize(LV.getType()) >
      return false;
    return IsVolatile && AtomicIsInline;

Making this return true for when LangOpts.CUDAIsDevice might essentially do what we need?

  rOMP OpenMP



More information about the Openmp-commits mailing list