On Thu, Nov 4, 2021 at 7:47 PM Andres Freund <andres@anarazel.de> wrote:
> The immediate goal is to be able to generate JITed code/LLVM-IR that doesn't
> contain any absolute pointer values. If the generated code doesn't change
> regardless of any of the other contents of ExprEvalStep, we can still cache
> the JIT optimization / code emission steps - which are the expensive bits.
I'm not sure why that requires all of this relative pointer stuff,
honestly. Under that problem statement, we don't need everything to be
one contiguous allocation. We just need it to have the same lifespan
as the JITted code. If you introduced no relative pointers at all,
you could still solve this problem: create a new memory context that
contains all of the EvalExprSteps and all of the allocations upon
which they depend, make sure everything you care about is allocated in
that context, and don't destroy any of it until you destroy it all. Or
another option would be: instead of having one giant allocation in
which we have to place data of every different type, have one
allocation per kind of thing. Figure out how many FunctionCallInfo
objects we need and make an array of them. Figure out how many
NullableDatum objects we need and make a separate array of those. And
so on. Then just use pointers.
I think that part of your motivation here is unrelated caching the JIT
results: you also want to improve performance by increasing memory
locality. That's a good goal as far as it goes, but maybe there's a
way to be a little less ambitious and still get most of the benefit.
--
Robert Haas
EDB: http://www.enterprisedb.com
