[PATCH] D109500: [openmp] Fix 51647, corrupt bitcode on amdgpu

[Dan Palermo via Phabricator via llvm-commits]

dpalermo added a comment.

I've reenabled the use of DataLayout::getAllocaAddrSpace as that does indeed still work as intended in the larger runnable test cases that hit this problem...and is the right thing to do since not all target architectures would necessarily have alloca in the local address space.

This means the lit test is once again broken.  I've determined that the problem with the lit test is that:

- there is no datalayout specified so we get a default that doesn't put alloca in the local address space (so it doesn't add the cast)...and the test "fails"
- I tried just adding a datalayout, but statements end up moving around causing all sorts of differences (checking is currently very tight)....and the test fails

So the options I can think of are:

1. just update the lit test without the addrcast so it passes, but it isn't testing what is being done in this change
2. regenerate the test case & checkers using a datalayout that has alloca as local
3. or do #1 and add another test that exercises this change (e.g. using the reduced test from the JIRA)

Let me know what your preference is.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109500/new/

https://reviews.llvm.org/D109500