[PATCH] D44523: Change calculation of MaxVectorSize

[Diego Caballero via Phabricator via llvm-commits]

dcaballe added a comment.

> Would it be possible to add a TTI callback that forces the minimum VF?

This sounds reasonable to me as long as the cost for the minimum VF is still better than the scalar version.

> I think Michael Kuperstein was trying to enable double pumping (could be one year or more ago). I think it's best to dig it up. There may be something usable in there.

Given Krzysztof's experiment, pumping doesn't seem to introduce stability issues in LV. `calculateRegisterUsage` seems to be handling some cost modeling for pumping (not sure if it's enough):

  // A lambda that gets the register usage for the given type and VF.
  auto GetRegUsage = [&DL, WidestRegister](Type *Ty, unsigned VF) {
    if (Ty->isTokenTy())
      return 0U;
    unsigned TypeSize = DL.getTypeSizeInBits(Ty->getScalarType());
    return std::max<unsigned>(1, VF * TypeSize / WidestRegister);
  };

Maybe it's just a question of changing `MaxVectorSize` as suggested, and refine the cost modeling, if necessary.
Let's wait for Michael.


Repository:
  rL LLVM

https://reviews.llvm.org/D44523