[PATCH] D86147: [LangRef] WIP: Revise semantics of get.active.lane.mask

[Sam Parker via Phabricator via llvm-commits]

samparker added a comment.

Just for complete clarity in my mind of the problem: 'i' defined as the maximum vector width for the loop? So we would expect a loop body like this:

  loop:
    %base = phi i32 [ 0, %entry ], [ %base.next, %loop ]
    %count = phi i32 [ 0, %entry ], [ %count.next, %loop ]
    %mask = get.active.lane.mask(i32 %base, i32 %n)
    %base.next = add i32 %base, %vector.width
    %count.next = add nuw i32 %count, 1
    %cmp = icmp ne i32 %count.next, %vector.trip.count
    br i1 %cmp, label %loop, label %exit

And so now we need to prove that (%base + (%vector.trip.count * %vector.width)) doesn't overflow for (%vector.trip.count - 1) iterations? If so, that sounds like a task for AddRecExpr::evaluateAtIteration?

> icmp ule (%base + i), %n - 1

How about icmp ult (%base + i), %n?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D86147/new/

https://reviews.llvm.org/D86147