[PATCH] D103656: [ScalarEvolution] Ensure backedge-taken counts are not pointers.

[Philip Reames via Phabricator via llvm-commits]

reames accepted this revision.
reames added a comment.
This revision is now accepted and ready to land.

LGTM.

For the record, some of the test diffs are a bit concerning code quality wise, but I think we may need to accept them to achieve the goal.  At least a couple of them have obvious possibilities for improvement in follow ups if needed, so I'm willing to have this go in.  We may end up having to revert, fix a few things, then reapply though.  We'll see.

Also, mostly for myself when reading through this later, this may regression exit counts for non-integral pointer tested loops, but that's - I think - okay because they're not idiomatic.  Once (if?), we have a generic answer for pointer subtraction we can relax some of these and restore power while maintaining integer results.



================
Comment at: llvm/lib/Analysis/ScalarEvolution.cpp:1084
+  // theoretically teach SCEV to truncate wider pointers, but such pointer
+  // widths are generally used for non-integral pointers anyway.
   if (getDataLayout().getTypeSizeInBits(getEffectiveSCEVType(Op->getType())) !=
----------------
Just for my context, they are?  This is news to me.  Do you have an example?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103656/new/

https://reviews.llvm.org/D103656