[clang] [llvm] [AMDGPU][True16][MC] true16 for v_alignbyte_b32 (PR #119750)

[Matt Arsenault via llvm-commits]

================
@@ -2353,8 +2353,8 @@ def int_amdgcn_writelane :
   [IntrNoMem, IntrConvergent, IntrWillReturn, IntrNoCallback, IntrNoFree]
 >;
 
-def int_amdgcn_alignbyte : ClangBuiltin<"__builtin_amdgcn_alignbyte">,
-  DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
+def int_amdgcn_alignbyte : DefaultAttrsIntrinsic<[llvm_i32_ty],
+  [llvm_i32_ty, llvm_i32_ty, llvm_anyint_ty],
----------------
arsenm wrote:

It doesn't need to. We can just pattern match extract of high bits. We generally shouldn't expose the direct op_sel reads 

https://github.com/llvm/llvm-project/pull/119750