[Openmp-commits] [openmp] [OpenMP][NFC] Remove `DelayedBinDesc` (PR #74360)

[Jon Chesterfield via Openmp-commits]

JonChesterfield wrote:

The multiple threads comment was in response to you claiming this code was racy. I said it wasn't, as you can't get here from multiple threads. Nothing to do with the original bug fix.

So the original patch hacked around using these objects in a partially initialised state and also sidestepped HSA 5.3ish doing dubious things. The partially initialised state is now gone, so we don't need that.

There is a missing piece which is I believe this original problem only occurred using dynamic_hsa, though I can believe other things fell over when linking it normally. Given you've already got this set up, could you toggle to using the dynamically loaded hsa and find out whether 5.3 segv's or not?

I'd rather we not keep a workaround in tree for 5.3, given I didn't want to introduce one in the first place, but it seems worth achieving the closure of knowing whether dynamic hsa which does the dubious address space walking would segv after this patch.

https://github.com/llvm/llvm-project/pull/74360