Validate atomics for the GCN target, and work around two LLVM atomic bugs

[maleadt]

Atomics reach the GCN back-end from AMDGPU.jl, UnsafeAtomics and Atomix, Enzyme and Julia's own intrinsics, and some of them don't compile or crash the compiler without pointing at the Julia code. Like #996 does for PTX, this validates them on the IR, and works around two LLVM bugs.

Validation rejects, with an InvalidIRError and a Julia backtrace:

• atomics wider than 64 bits and under-aligned atomics. Julia's LLVM compiles them to libatomic calls (__atomic_compare_exchange_16), which only fail when the code object is loaded (hipErrorNoBinaryForGpu); the external back-end rejects them without Julia frames;
• atomics on memory other than flat, global, local, private and buffer fat pointers, which the back-end rejects or aborts on;
• unknown synchronization scopes. cluster is accepted when the LLVM that generates code knows it (LLVM 22+), which for the external back-end is AMDGPU_LLVM_Backend_jll's version rather than Julia's.

The workarounds expand atomics to compare-exchange loops after optimization, also when the IR isn't validated:

• since LLVM 22, the back-end selects atomicrmw usub_sat (UnsafeAtomics' saturating subtraction) natively on gfx10.3+, but has no pattern for flat memory, and for local memory only on gfx12, so it fails with "Cannot select" ([AMDGPU] Cannot select AtomicLoadUSubSat on LDS (gfx10.3/gfx11) and flat memory (gfx12, or with no.fine.grained.memory) llvm/llvm-project#229442). AMDGPU_LLVM_Backend_jll 23.1.3 carries the fix (AMDGPU: Fix selection of atomicrmw usub_sat for LDS and flat llvm/llvm-project#229471), so this requires that version and only expands 32-bit usub_sat with Julia's own LLVM (22+, currently nightly): on flat memory, and on local memory for gfx10.3 and gfx11. With the fixed back-end, gfx12 flat memory now uses the native flat_atomic_sub_clamp_u32.
• before LLVM 21, the atomic optimizer emits an illegal readfirstlane for 8- and 16-bit read-modify-writes whose result is used (AMDGPU: Fix creating illegally typed readfirstlane in atomic optimizer llvm/llvm-project#128388), which aborts or crashes Julia 1.12 and 1.13 with the in-process back-end, e.g. for an Atomix += on an Int8 array. The optimizer can only be turned off process-wide, so this expands the operations it handles when their result is used.

The tests check each rejection and, for the workarounds, the compile failure or crash without them (in a subprocess) and that the cases LLVM handles stay native. The GCN tests pass on Julia 1.10 through 1.13, and the full suite on 1.13 (at d1d5546). The AMDGPU_LLVM_Backend_jll 23.1.3 change (37e1d20) was checked with the GCN tests on Julia 1.13 and nightly (LLVM 22, which exercises the in-process workaround).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.73%. Comparing base (e6ff2e6) to head (d1d5546).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #997      +/-   ##
==========================================
+ Coverage   88.62%   88.73%   +0.11%     
==========================================
  Files          30       30              
  Lines        6443     6498      +55     
==========================================
+ Hits         5710     5766      +56     
+ Misses        733      732       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.