[llvm] AMDGPU/GlobalISel: RegBankLegalize for atomic fmin/fmax num (PR #185504)

[Petar Avramovic via llvm-commits]

================
@@ -1505,6 +1505,13 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
   addRulesForIOpcs({amdgcn_global_atomic_ordered_add_b64})
       .Any({{DivS64}, {{Vgpr64}, {IntrId, VgprP1, Vgpr64}}});
 
+  addRulesForIOpcs(
+      {amdgcn_global_atomic_fmin_num, amdgcn_global_atomic_fmax_num})
+      .Any({{DivS32}, {{Vgpr32}, {IntrId, VgprP1, Vgpr32}}});
+
+  addRulesForIOpcs({amdgcn_flat_atomic_fmin_num, amdgcn_flat_atomic_fmax_num})
+      .Any({{DivS32}, {{Vgpr32}, {IntrId, VgprP0, Vgpr32}}});
----------------
petar-avramovic wrote:

should use Div(S32 instead

https://github.com/llvm/llvm-project/pull/185504