[LoopVectorize] Use available masked vector functions when required

[huntergr]

Summary

This patch builds on D132458 (use masked vector functions with a dummy mask when no unmasked variant is available) and D134422 (scalarize function calls when masking is required), adding support for actually using the masked functions when masking is required, either due to the control flow in the scalar loop or because we're performing tail folding.

[huntergr]

Rebased on top of the latest version of D132458, which incorporates the vector function into the vplan recipe instead of looking it up again.

[tschuett]

llvm/include/llvm/Analysis/VectorUtils.h

291

if (Info.isMasked())
  return true;


return false;

[tschuett]

llvm/include/llvm/Analysis/VectorUtils.h

291

  if (!VF || Info.Shape.VF == *VF)
    if (Info.isMasked())


return false;

[huntergr]

Tidied up hasMaskedVariant as requested.