[llvm-commits] Cost Table, take 2

[Chandler Carruth]

At this point, I'm not really seeing the benefit of this patch.

It takes 2 structs and 2 functions, both of which are dead simple, and
replaces them with 3 class templates in an inheritance tree plus an inner
struct a constructor and one function that is significantly more complex.
It doesn't make things more efficient in the process. It doesn't really
make this a generic tool for table lookup, because the tables are not in
fact generic. I think ultimately I would prefer the original code here.
Maybe there is a massive simplification this provides to some future code
you're considering that I don't see yet? Dunno...


Sorry for being a bit of a rain cloud here. To try to give some
constructive ideas, I looked at the fundamental problem, and I see two
designs that seem more workable:

1) A TargetCostTable class template in include/llvm/Target which accepts a
T type parameter that can be either MVT, std::pair<MVT, MVT>, or
std::tuple<MVT, MVT, ...>. Then the ISD and the cost value can be concrete
things.

2) Teach tablegen to produce pre-sorted tables with lookup routines from a
.td file in the target.

My suspicion is that #1 will end up being nearly as much code and
complexity as the duplication between ARM and X86....

My suspicion is that *if* it is worth abstracting this stuff, we should
just do #2 and make it really powerful. Then we can quickly build very
large and detailed cost tables, and they will remain efficient and simple
to use in each target.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130124/2ca57fea/attachment.html>