[llvm-branch-commits] [llvm] [AMDGPU][SIMemoryLegalizer] Lower single-wave workgroup scope to wavefront in absence of LDSDMA (PR #224065)

[Pierre van Houtryve via llvm-branch-commits]

================
@@ -207,6 +208,17 @@ class SIMemOpInfo final {
     // AGENT scope as a conservatively correct alternative.
     if (this->Scope == SIAtomicScope::CLUSTER && !ST.hasClusters())
       this->Scope = SIAtomicScope::AGENT;
+
+    // When max flat work-group size is at most the wavefront size, the
+    // work-group fits in a single wave, so LLVM workgroup scope matches
+    // wavefront scope. Demote workgroup → wavefront here for fences and for
+    // atomics with ordering stronger than monotonic.
+    if (CanDemoteWorkgroupToWavefront &&
+        this->Scope == SIAtomicScope::WORKGROUP &&
+        (llvm::isStrongerThan(this->Ordering, AtomicOrdering::Monotonic) ||
----------------
Pierre-vh wrote:

Yeah just do this uniformly, and if there is any doubt you can do it uniformly, then it's a big red flag and means this is not researched well enough IMO

https://github.com/llvm/llvm-project/pull/224065