[RFC] Proposed update to convert two 64-bit attribute bitmasks to std::bitset

There are two 64-bit bitmasks maintained in AttributeImpl.h:

  • AvailableFunctionAttrs is part of the AttributeListImpl class, and
  • AvailableAttrs is part of the AttributeSetNode class

Both of these assume that the number of available enum attributes is limited to 64. In fact, a static_assert in Attributes.cpp enforces that the number of enum attributes stays at 64 or below. However, the bitcode writer and reader don’t communicate enum attributes via bitmask anymore. Enum attributes are encoded in attribute groups.

The AvailableFunctionAttrs and AvailableAttrs bitmasks are leftovers that need to be updated to remove the limitation on the number of enum attributes that can be defined in the Attribute::AttrKind enum.

Per a suggestion that I received a while ago from Reid Kleckner (on the llvm-dev list), I propose to implement both of these data members as std::bitset objects.

Here are the details for this proposed change:

llvm/lib/IR/AttributeImpl.h:

  • Define AvailableAttrs and AvailableFunctionAttrs with type std::bitset instead of uint64_t
  • Update AttributeSetNode::hasAttribute() to use the std::bitset test function on AvailableAttrs to check if an enum attribute is present
  • Update AttributeListImpl::hasFnAttribute() to use the std::bitset test function on AvailableFunctionAttrs to check if an enum attribute is present

llvm/lib/IR/Attributes.cpp:

  • Update AttributeSetNode::AttributeSetNode() constructor to use the std::bitset’s set function to initialize AvailableAttrs
  • Update AttributeListImpl::AttributeListImpl() constructor to use the std::bitset’s set function to initialize AvailableFunctionAttrs and remove the static_assert that enforces an upper limit of 64 on the number of enum attributes allowed

The representation change seems fine per se, but we still have a concern with the 65th attribute increasing the size of the objects in question. Do you have a plan to minimize that impact?

To be clear, I think the representation change is fine regardless. As one example, we have a bunch of downstream attributes, and would happily pay the extra word to integrate them cleanly with the built in variety in our downstream code. I’m sure we’re not the only ones.

Philip

Hi Philip,

I’m not sure I understand your concerns exactly.

My proposal is to construct AvailableFunctionAttrs, for example, with an initial size that comprehends the upper limit of the number of enum attribute slots in Attribute:AttrKind …

class AttributeListImpl final

: public FoldingSetNode,

private TrailingObjects<AttributeListImpl, AttributeSet> {

friend class AttributeList;

friend TrailingObjects;

private:

/// Bitset with a bit for each available attribute Attribute::AttrKind.

std::bitsetAttribute::EndAttrKinds AvailableFunctionAttrs;

};

Yes, I suspect that the footprint of a bitset implementation is bigger than 1 or 2 uint64_t words, but I’m wary of dictating an arbitrary limit on the number of enum attributes allowed based on how we implement the API for keeping track of whether a function or data object has an attribute or not.

If I’m misunderstanding what you mean by “pay[ing] an extra word to integrate them cleanly,” then can you please clarify?

~ Todd