[llvm] [InstSimplify] Implement simple folds for `ucmp`/`scmp` intrinsics (PR #95601)

[Nikita Popov via llvm-commits]

================
@@ -6505,6 +6505,33 @@ Value *llvm::simplifyBinaryIntrinsic(Intrinsic::ID IID, Type *ReturnType,
 
     break;
   }
+  case Intrinsic::scmp:
+  case Intrinsic::ucmp: {
+    // Fold cmp x, x -> 0
+    if (Op0 == Op1)
+      return Constant::getNullValue(ReturnType);
----------------
nikic wrote:

The rule of thumb is that you only need to optimize the case where things do *not* fold, because not folding is the common case. Unless a substantial fraction of inputs contain cmp(x, x), having a special case for it will make no difference.

If we wanted to optimize this, then we could drop the ICMP_EQ query instead, and argue that cases where the values are non-trivially equal should be handled by GVN instead. But I think it's fine to keep the general implementation for now.

https://github.com/llvm/llvm-project/pull/95601