EricWF marked an inline comment as done.
EricWF added inline comments.
================
Comment at: lib/CodeGen/CGExprAgg.cpp:971
+ auto EmitCmpRes = [&](const VarDecl *VD) {
+ return CGF.CGM.GetAddrOfGlobalVar(VD);
+ };
----------------
rsmith wrote:
> Perhaps directly emit the constant value here rather than the address of the
> global? I think we should consider what IR we want to see coming out of
> Clang, and I don't think that IR should contain loads from globals to get the
> small constant integer that is the value of the conversion result.
>
> I think it would be reasonable for us to say that we require the standard
> library types to contain exactly one non-static data member of integral type,
> and for us to form a select between the relevant integer values here. We
> really have no need to support all possible implementations of these types,
> and we can revisit this if some other standard library implementation ships
> types that don't follow that pattern. (If we find such a standard library, we
> could emit multiple selects, or a first-class aggregate select, or whatever
> generates the best code at -O0.)
I agree emitting the value would be better, and that most STL implementations
should implement the types using only one non-static member.
However, note that the specification for `partial_ordering` is described in
terms of two non-static data members, so it seems possible an STL
implementation might implement in that way.
Would it be appropriate to do this as a smaller follow up patch?
https://reviews.llvm.org/D45476
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
