[Mlir-commits] [llvm] [mlir] [OMPIRBuilder] Restore the debug location after the reduction alloca (PR #221253)

[Abid Qadeer]

abidh wrote:

The `openmp-s390x-linux` failure that blamed this commit ([build 26727](https://lab.llvm.org/buildbot/#/builders/88/builds/26727)) is unrelated to it.

The failing test, `libomp :: tasking/issue-94260-2.c`, is a known-flaky test tracked in #176451 and already disabled on x86 Linux by b217e5c5a035 (#176452). The `UNSUPPORTED: linux && x86-target-arch` line added there does not cover the `systemz-1` worker, so that bot still runs it.

The same test failed twice on the same worker shortly before this landed, with unrelated blamelists: [build 26718](https://lab.llvm.org/buildbot/#/builders/88/builds/26718) and [build 26720](https://lab.llvm.org/buildbot/#/builders/88/builds/26720), the latter blaming ce27ca449c56, a SCEV test-coverage change. Both recovered on the next build without anything being reverted. That makes three failures of this one test in the last 25 builds on that worker.

Independently of the flakiness, this change cannot affect that test. It only touches `OpenMPIRBuilder::createReductions`, which clang never calls -- its only OMPIRBuilder reduction entry point is `createReductionsGPU` in `CGOpenMPRuntimeGPU.cpp`, and the only in-tree caller of `createReductions` is the MLIR OpenMP translation. The test is C compiled with plain `-fopenmp`, with no `-g` and no offloading, and it fails with a SIGSEGV at runtime inside libomp. Restoring a debug location in the compiler cannot change the behaviour of a program built without debug info.


https://github.com/llvm/llvm-project/pull/221253