[llvm] [X86] Extend `combinei64TruncSrlAdd` to handle patterns with `or` and `xor` (PR #128435)

[João Gouveia via llvm-commits]

================
@@ -53810,11 +53827,15 @@ static SDValue combineTruncatedArithmetic(SDNode *N, SelectionDAG &DAG,
   if (!Src.hasOneUse())
     return SDValue();
 
-  if (SDValue R = combinei64TruncSrlAdd(Src, VT, DAG, DL))
-    return R;
+  if (VT == MVT::i32 && SrcVT == MVT::i64 && SrcOpcode == ISD::SRL &&
+      Src.getOperand(0).getNumOperands() == 2 &&
----------------
joaotgouveia wrote:

Removing that check causes LLVM to crash on a few regression tests (`CodeGen/X86/arg-cast.ll`, `CodeGen/X86/pr49162.ll`, `CodeGen/X86/vector-reduce-xor-bool.ll`, and `ExecutionEngine/MCJIT/non-extern-addend.ll`). I might be mistaken, but at the time `combineTruncatedArithmetic` is called, we have no information about whether the first operand of `Src` is actually a binary operation. Calling `Src.getOperand(0).getOperand(1)` might cause us to attempt to access an invalid operand.

https://github.com/llvm/llvm-project/pull/128435