On Tue, 2025-02-18 at 08:54 -0800, Stephen Brennan wrote: > Alan Maguire <alan.maguire@xxxxxxxxxx> writes: Hi Stephen, Alan, [...] > I wonder if there's a hidden catch here... DECL_TAGs can refer to VARs, > but as far as I can tell, the semantics are that the DECL_TAG actually > *modifies* the VAR - it says "this variable declaration had this > annotation / tag / whatever". This is fundamentally because the > var_secinfo doesn't point at a chain of DECL_TAGs, it points directly at > the VAR. (Compare that to, e.g. const/volatile/restrict qualifiers) > > So, in a hypothetical situation where there are two variables of the > same name & type, but one of them has a DECL_TAG referring to it, would > the deduplication keep these separate? It seems to me that would be the > correct behavior. This is something too look out for. E.g. this is certainly true for decl tags attached to structs. [...] > > Anyway I'd be interested to hear what others think about whether solving > > this in BTF dedup itself or pahole makes most sense. Thanks! > > Thanks for taking a look into this! Hoping to hear from other > experienced voices on this question as well. [...] Deduplication for variables would be a bit of a weird one. As Alan notes in the previous email, for each data section there is a: - 'btf_type' entry of kind BTF_KIND_DATASEC where 'vlen' specifies the number of variables in the section; - followed by a number of 'btf_var_secinfo' entries, with 'type' referring to 'btf_type' entries of kind 'BTF_KIND_VAR, and 'offset' specifying offset of the variable within section. So, we are talking about deduplication of these 'btf_var_secinfo' entries. Nothing in existing dedup algorithm handles deduplication of "addendum" items like 'btf_var_secinfo', the subject of deduplication always has an ID. A new step, walking each BTF_KIND_DATASEC and collapsing identical 'btf_var_secinfo' entries can of-course be added. (Which also raises a question, whether this has to be done in a loop with an attempt to collapse duplicate BTF_KIND_DATASEC entries?). A case can be made both for and against this, but given the "weirdness" described above and the fact that Stephen already has the code, I think performing this deduplication at the pahole level is Ok at the moment. Thanks, Eduard
