[clang] [llvm] [Offload][Driver] Split frontend CUDA/HIP language assumptions from backend device (PR #212373)

[via cfe-commits]

oltolm wrote:

https://lab.llvm.org/buildbot/#/builders/234/builds/2360

Codex says this about the failing test:

> The immediate defect is in the test: cudaMalloc does not initialize device memory, yet the test assumes the atomic counter starts at zero. A separate pending branch adds synchronization, but synchronization alone cannot make an uninitialized counter valid.

> Yes, the bug is still present on current main (2eeb5f10ff2).
The test calls cudaMalloc, then immediately performs atomic increments without initializing the allocated integer to zero. It still expects 42, so the test remains nondeterministic.

https://github.com/llvm/llvm-project/pull/212373