This has been an explicit design goal – based on the hindsight of the many-decades-old errors with include directories and header files.
With header files (except for generated ones) there's not the requirement of a build DAG, just out-of-date information. That's why we can generate the dependencies as part of the compilation step. With modules, at least with the current
physical model, we have to make sure the interface has been translated before it can be consumed. Incremental builds are going to break for a lot of people in subtle ways. And yes, this is very much like the past and using tools like `makedeps`.
What we learned then is that it's very very difficult to precisely reproduce the preprocessor and include searching. There's a lot of poorly specified and undefined behavior in the early phases of translation. If the tools disagree, build problems ensue.
I don't believe we have made the situation impossible. A tool should be able to proceed sufficiently with preprocessing in the face of missing named modules and be able to report all of the requirements of a TU. It wouldn't for header imports, but that's a more well understood problem, as those map to filenames in a well agreed on way.
What module an interface provides is also discoverable, but that may not be as useful in an open universe situation. The question of which instance of the primary module interface out of many is hard. We have that today with headers, of course, and we try to deal with it via ordering rules for search paths, but it's still a source of error. Duplicates and collisions are harder for named modules, even if figuring out what an interface provides is straight forward.
On Fri, Jun 18, 2021 at 2:12 PM David Blaikie via SG15 <email@example.com> wrote:
On Fri, Jun 18, 2021 at 11:01 AM Gabriel Dos Reis <firstname.lastname@example.org> wrote:
- but you cannot discover the graph with a simple lexer
Yes, and I have not found that to be an actual problem in practice. In fact there is really nothing new here (other than recognizing the ‘import’ directives and friends) in the non-module world: the ol’ C and C++ standards allow
where HEADER comes from macro expansions and all the other stuff that you mention about inability to use “simple lexer”, conditional inclusion, etc. As a PSA, I would like to disabuse anyone who ever thought that grep was sufficient to compute dependency of C or C++ source files pre-C++20.
Maybe it is my dinosaurus status showing bleeding out here but are we rehashing the same arguments from the past?
The one thing we can do for the community is to have a library readily available to anyone who ever thought that a simple lexer is sufficient to compute C++ source file dependency. That would be a major step forward.
I don't think Daniel is suggesting/has mentioned the idea that a simple lexer is sufficient for either headers or modules. But expected/would like it if we could simplify the situation further than module files alone currently allow in this regard.
I do lean towards "maybe it's simple enough, even if it doesn't allow a simple lexer to do this operation" - at least I don't think that's an ironclad proof that we need the dependency information in a more immediately legible format such as in the pkg-config files or equivalent. I wouldn't rule it out either, though.
Like, I wouldn't be averse to pkg-config or equivalent files supporting this as an optional feature - so that folks could explore whether it significantly improves performance. If it's more a matter of software complexity (since build tools that don't know anything about parsing C++ are going to want to discover this dependency graph) - yeah, maybe compilers could vend a command line mode that reports the dependencies as quickly as possible - ala clang-scan-deps & for library level uses (llvm-buildozer, background running build systems/IDE type situations) there could be a library for that.
But, going on past experience, those kinds of utilities/dependencies don't necessarily eventuate and I can understand the concern about proliferation/variation/lack of ubiquity in such things & so a preference to bake it into a common config file. I guess all I have on that is that I think it'd be unfortunate for developers to have to repeat this information in multiple places like that.
On Fri, Jun 18, 2021 at 1:20 PM Gabriel Dos Reis <email@example.com> wrote:
Also notice that while a dependency scanner needs to chase #included or imported files (via the appropriate implementation-defined mapping), it actually does not need to fully preprocess the source code found therein;
the spec has been carefully crafted to that effect.
It doesn't need to preprocess the complete file, but it needs to significantly preprocess the early sections of the file. While the spec has carefully prevented conditional specification of the keywords, it still allows macros to be used, which requires support for conditional evaluation, variable expansion, and string concatenation.
In other words, the only optimization granted by the standard today is on the discovery of whether or not a source file needs to participate on the construction of the DAG for processing modules, but you cannot discover the graph with a simple lexer, you need the full expression evaluation logic of the preprocessor.
SG15 mailing list