[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: Legalize atomicrmw fmin/fmax (PR #97048)

[Matt Arsenault via llvm-branch-commits]

================
@@ -1670,10 +1670,22 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   if (ST.hasAtomicFlatPkAdd16Insts())
     Atomic.legalFor({{V2F16, FlatPtr}, {V2BF16, FlatPtr}});
 
-  // FIXME: Handle flat, global and buffer cases.
-  getActionDefinitionsBuilder({G_ATOMICRMW_FMIN, G_ATOMICRMW_FMAX})
+
+  // Most of the legalization work here is done by AtomicExpand. We could
+  // probably use a simpler legality rule that just assumes anything is OK.
+  auto &AtomicFMinFMax =
+    getActionDefinitionsBuilder({G_ATOMICRMW_FMIN, G_ATOMICRMW_FMAX})
     .legalFor({{F32, LocalPtr}, {F64, LocalPtr}});
 
+  if (ST.hasAtomicFMinFMaxF32GlobalInsts())
+    AtomicFMinFMax.legalFor({{F32, GlobalPtr},{F32, BufferFatPtr}});
----------------
arsenm wrote:

Yes, it works for the fmin/fmax case. The fadd had some has-global-no-buffer cases 

https://github.com/llvm/llvm-project/pull/97048