[PATCH] D79175: [ARM][MVE] Tail-Predication: use @llvm.get.active.lane.mask to get the BTC

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/lib/Target/ARM/MVETailPredication.cpp:382
   // TODO: This add likely already exists in the loop.
   Value *Remaining = Builder.CreateSub(Processed, Factor);
   Processed->addIncoming(Remaining, L->getLoopLatch());
----------------
efriedma wrote:
> samparker wrote:
> > efriedma wrote:
> > > This subtraction can also overflow.
> > But that's okay, right? This predication is only really useful when wrapping happens and the intrinsic reflects that overflow can/will happen.
> The problem here is that the original get_active_lane_mask can do something like this:
> 
> Iteration 1: get_active_lane_mask(0, 5) -> all-true
> Iteration 2: get_active_lane_mask(4, 5) -> <true, false, false, false>
> Iteration 3: get_active_lane_mask(8, 5) -> all-false
> 
> In the rewritten code, you end up with this:
> 
> Iteration 1: vctp(5) -> all-true
> Iteration 2: vctp(1)- > <true, false, false, false>
> Iteration 3: vctp(-3)- > all-true
> 
> There are a couple ways you could deal with this:
> 
> 1. Use saturating subtraction
> 2. Prove the trip count is small enough that this never happens, i.e. `ceil(ElementCount / VectorWidth) >= TripCount`.
Using saturating subtraction would require proving that the induction variable used as the first argument to llvm.get.active.mask doesn't overflow.  But I guess you have to check that anyway.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79175/new/

https://reviews.llvm.org/D79175