[PATCH] D23636: [ADT] Allocate memory less often by increase inline storage

Sanjoy Das via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 30 21:50:11 PDT 2016


sanjoy added a comment.

I'm not sure if this is the right approach -- I think it is better to avoid a change of this scale if possible.

Have you considered changing the access to `pVal` in APInt.cpp to go through an accessor that returns `&VAL[0]` if the bitwidth is <= 128?  The union could be changed to:

  union {
    uint64_t LocalArray[2];
    uint64_t VAL;
    uint64_t *pVal;
  };

to avoid breaking existing uses of `VAL`, and most of the existing code (except perhaps things like `AssignSlowCase`) should work as they do now.  This approach will also easily generalize to a `LocalArray[N]`.

The downside of what I've suggested above is that the "fast path" for some of the two-word operations will not be visible but a) they may not actually matter, and b) we can change them one by one as needed.


https://reviews.llvm.org/D23636





More information about the llvm-commits mailing list