On 08/31/2018 07:53 AM, Nathan Sidwell wrote:
> On 08/31/2018 07:09 AM, Boris Kolpackov wrote:
>> Message-ID: <boris.20180831104601_at_[hidden]>
>>
>> Tom Honermann <tom_at_[hidden]> writes:
>
>>> 2. A method of specifying a path to search for module description
>>> files, similar to existing include paths.
>>
>> I would argue against any kind of "search paths" approach (whether for
>> modules or description files themselves). We've used them for includes
>> and I think it has proven to be brittle (I am talking about the "header
>> doesn't exist where you expect it to exist but the compiler found you
>> another one" kind of situtions) and not toolable (where shoudl I
>> generate
>> this non-existent header?)
>
> A search path seems the obvious choice, because of the similarity to
> include paths. It is the route I originally went down. It's a trap!
> As Boris says, it's brittle. It continues the performance problems of
> searching a bunch of filesystem locations for each import.
>
> I abandoned that approach for an interface allowing the compiler to be
> agnostic, and other tools to plug into that. Expect a paper for San
> Diego.

Are you referring to the module mapper approach documented at
https://gcc.gnu.org/wiki/cxx-modules?

If so, my concern with that approach is that it effectively requires a
build system. Perhaps the default module mapper does not (I'm not sure
exactly what it does at present. My brief tests indicate it requires a
<module>.nms file to exist in gcc's current working directory). What
I'd like is for the default module mapper to use a common tool agnostic
approach that does not require build system integration, nor precompiled
artifacts. Repeating a point I'm trying to make, use of module
artifacts should be an optimization opportunity, not a requirement.

Tom.