On Mon, 5 Nov 2018 at 14:36, Jamie Clarkson <jnc7...@gmail.com> wrote:
>
> Yes that does help, looks like a really good solution. In your opinion
> would the resulting code be worthwhile (including all the dictionaries,
> stubs etc.) when compared to the C++ code expansion method? I can see that
> it would save for the basic types and types derived from them but once you
> start getting any sort of structs you're getting towards needing one
> function per type/combination of types?
>
That is of course the danger, and it's more acute the more type parameters
there are.
It's possible that by-value struct types would be less common than pointer
types, and if you're aware of the underlying implementation strategy, you
can decide to use a certain representation (e.g. pointer) to reduce the
number of generated instances.
For some kinds of code, it may be possible to generate a shared instance
for values of different types that doesn't incur large indirection costs
(the internal map implementation is an example, of course), but the
compiler analysis to enable that might be quite costly.
In general, the compiler has the freedom to choose any of the options, from
pure code expansion (you'd probably do that as a matter of course if
there's only one instance) through one-instance-per-size/layout and on to
more generic but less efficient code, but there's no perfect solution!
I wonder if there's a study on what kind of values tend to be used as
generic type parameters in languages that distinguish between pointers and
values (C++, C#?).
> (I realise yours is a hypothetical implementation so not expected to be
>>> complete! Just interested in what your preferred solution is)
>>>
>>
>> Ah
>>
>>>
>>>
>>>
> More complete than it looked on the surface :)
>
Thanks :)
You might be interested: I ran the heap benchmark against my "generic"
implementation (instantiated for v8a8) and compared it against the same
benchmark in container/heap, and it does OK, so the overhead of passing the
extra instance pointer and an indirect function call is at least comparable
to calling methods on an interface.
benchmark old ns/op new ns/op delta
BenchmarkDup-4 355164 265610 -25.21%
cheers,
rog.
--
You received this message because you are subscribed to the Google Groups
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to golang-nuts+unsubscr...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
