[llvm] [AMDGPU][DAG] Remove AssertZext before some intrinsics (PR #146052)

[Pierre van Houtryve via llvm-commits]

================
@@ -3985,6 +3985,24 @@ SDValue AMDGPUTargetLowering::performAssertSZExtCombine(SDNode *N,
     }
   }
 
+  // AssertZext in front of these intrinsics is not necessary, the lowering of
+  // the intrinsics into a register read will insert one if it is needed.
----------------
Pierre-vh wrote:

My bad, I must have made a mistake in the implementation the first time I tried this, which caused the codegen issues I saw.
I rewrote it cleanly and now it works as expected, we can indeed optimize `assert_zext (and)` out

https://github.com/llvm/llvm-project/pull/146052