[llvm] [AMDGPU] Custom lower get_fpmode for demanded-bits based read narrowing (PR #218137)

[Matt Arsenault via llvm-commits]

================
@@ -14935,6 +14966,45 @@ SDValue SITargetLowering::performAndCombine(SDNode *N,
       }
     }
 
+    if (CRHS && LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN && LHS.hasOneUse()) {
----------------
arsenm wrote:

This isn't custom lowering, it's a combine. This can't replace the code you deleted, this would need to do a setOperationCustom and handle through LowerOperation

Also does this mean that global isel is broken now? The test seems to be missing -global-isel run lines 

https://github.com/llvm/llvm-project/pull/218137