[llvm] [NFC][LAA] Minor stylistic/comments improvements (PR #185510)

[Andrei Elovikov via llvm-commits]

================
@@ -2575,12 +2577,13 @@ bool LoopAccessInfo::analyzeLoop(AAResults *AA, const LoopInfo *LI,
           HasConvergentOp = true;
       }
 
-      // With both a non-vectorizable memory instruction and a convergent
-      // operation, found in this loop, no reason to continue the search.
+      // Unsafe to vectorize and we already found a convergent operation, can
+      // early return now.
----------------
eas wrote:

I don't believe that's true, imagine loop trip count varying between threads, something like `4 + 4 * (%thread_id % 2)` (in RT, not statically visible). If we vectorize that with VF=8 and a remainder then we'd split first 4 "scalar" iterations into two different loops - vector for even threads and scalar for odd. That is illegal to split. a convergent operation like that, IIUC.

I actually wanted to add such a comment/test but we drop out in legality for an unknown intrinsic so I couldn't craft a test.

https://github.com/llvm/llvm-project/pull/185510