[llvm] AMDGPU: Do not bitcast atomic load in IR (PR #90060)

[Matt Arsenault via llvm-commits]

================
@@ -230,6 +230,12 @@ class AMDGPUTargetLowering : public TargetLowering {
   bool isCheapToSpeculateCtlz(Type *Ty) const override;
 
   bool isSDNodeAlwaysUniform(const SDNode *N) const override;
+
+  // FIXME: This hook should not exist
----------------
arsenm wrote:

Yes, I'm most of the way through getting amdgpu to only bitcast in the DAG. Then theoretically it should be easy to remove it. The atomicrmw case is giving me a bit of trouble, since it seems there's some value in keeping bitcasts out of the cmpxchg loop 

https://github.com/llvm/llvm-project/pull/90060