[PATCH] D61809: [BPF] Preserve original struct/union type name/access index and array subscripts

[Eli Friedman via Phabricator via cfe-commits]

efriedma added a comment.

> The size you allocate here will presumably need to vary as the struct layout changes, and you have no way of knowing which allocas will need their sizes to be changed.

Your example is just a pointer; the size of a pointer won't change.

That said, yes, address computation isn't the only place C cares about the layout of a struct.  In particular variables (local or global) of struct type, sizeof(), and offsetof() need different handling.  But I'm not sure how much any of those matter for BPF.

> If this call is supposed to somehow represent that the %6 GEP is computed from %5 by way of a struct field access, this will not be robust against any kind of optimization: optimizations will happily rewrite uses of %6 to use %5 or some other form directly, and you will have lost track of the pointers you need to update.

I think actually, the idea is that the intrinsic represents the computation of some additional adjustment, and so the following instructions use %7, the result of the call, not %6.  So I think the existing formulation is sound.

That said, it's awkward to have both an instruction and an intrinsic representing parts of the same offset computation; it probably makes sense to emit just the intrinsic, instead of both the intrinsic and the GEP, like you suggest.


Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D61809/new/

https://reviews.llvm.org/D61809