[llvm] [SCEV] Consider IVMayOverflow to use cheap BE-count formula. (PR #218694)

[Nikita Popov via llvm-commits]

================
@@ -13564,9 +13566,20 @@ ScalarEvolution::howManyLessThans(const SCEV *LHS, const SCEV *RHS,
       //   So dividing that by Stride gives zero.
       //
       // * For RHS > Start (End is RHS), the backedge count must be
-      //   "RHS-Start /uceil Stride". Together with the precondition
-      //   "Start - Stride < Start", we have "RHS > Start > Start - Stride".
-      //   As such RHS - (Start - Stride) - 1 does not overflow.
+      //   "RHS-Start /uceil Stride", so it is sufficient to show that the
+      //   numerator "((RHS - 1) - (Start - Stride))" does not overflow.
+      //
+      //   If "Start - Stride < Start" holds, we have
+      //   "RHS > Start > Start - Stride". As such
+      //   "RHS - (Start - Stride) - 1" does not overflow, which is the
+      //   reassociated numerator.
+      //
+      //   Otherwise !IVMayOverflow guarantees "RHS + (Stride - 1) <= MaxV ",
----------------
nikic wrote:

```suggestion
      //   Otherwise !IVMayOverflow guarantees "RHS + (Stride - 1) <= MaxV",
```

https://github.com/llvm/llvm-project/pull/218694