[PATCH] D22377: [SCEV] trip count calculation for loops with unknown stride

[Pankaj Chawla via llvm-commits]

pankajchawla added inline comments.

================
Comment at: llvm/lib/Analysis/ScalarEvolution.cpp:8731
@@ -8704,2 +8730,3 @@
   // Avoid negative or zero stride values
-  if (!isKnownPositive(Stride))
+  // We can assume stride to be positive if NoWrap is true and loop does not
+  // have abnormal exits or side effects.
----------------
sanjoy wrote:
> As we discussed, it isn't that given the conditions you're testing  we can prove that the stride is non-negative, but that the math below is correct even if the stride is negative.
> 
> Also, I'd separate out the `!NoWrap || !loopHasNoAbnormalExits(L) || !loopHasNoSideEffects(L)` check into it's own `if` and have a small comment describing why they're necessary / sufficient (the example you had in the llvm-dev thread would be nice to add too).
The check for abnormal exits seems redundant now because loopHasNoSideEffects() is more conservative than it. I am removing this check as well.


https://reviews.llvm.org/D22377