[PATCH] D134267: [C++] [Modules] Support one phase compilation model for named modules

[Iain Sandoe via Phabricator via cfe-commits]

iains added a comment.

In D134267#3869678 <https://reviews.llvm.org/D134267#3869678>, @dblaikie wrote:

> In D134267#3869643 <https://reviews.llvm.org/D134267#3869643>, @iains wrote:
>
>> In D134267#3869614 <https://reviews.llvm.org/D134267#3869614>, @dblaikie wrote:
>>
>>> In D134267#3869162 <https://reviews.llvm.org/D134267#3869162>, @iains wrote:
>>>
>>>> In D134267#3868830 <https://reviews.llvm.org/D134267#3868830>, @dblaikie wrote:
>>>>
>>>>> I'm OK with sticking with the existing `-fmodule-file` if that works for everyone. Yeah, it's short and ambiguous in a space with many concepts of what a "module file" is, but also the fact that it's a `-f` flag might help disambiguate it a bit - it's probably not the way anyone would think/expect to be passing source files, those just get passed without flags on the command line. And any use of it will show the .pcm extension or whatever that should make it clear enough what's going on.
>>>>
>>>> hmm (I realise I mentioned this, and hope it has not complicated things) ..
>>>>
>>>> .. I was thinking of the `-fmodule-file=<name>=filename` variant.  The problem with using it without (the <name>) is that -fmodule-file= can (and does) appear multiple times on the command line to specify dependent modules - so that one would have to specify which (named) module was related to the filename.
>>
>>
>>
>>> I /think/ currently `-fmodule-file` works with Clang header modules presumably because the module files contain within them enough information to enable the compiler to substitute the module for an include where necessary? I'm not sure the `<name>=` part is necessary, as such, if the BMIs can reasonably contain enough identifying info to make their use clear to the consumer.
>>
>> `-fmodule-file=` works with both `clang` modules  and standard C++ ones.  What it does (as of now) is cause the FE to open the named file and to pre-load the module name - so that when the FE tries to find a module the lookup mechanism can associate the module name with the module data.  So, if we need 3 dependent modules to be available when building a new source .. we would specify the PCM files containing those modules.
>>
>> At present, if you try to specify a non-existent PCM using `-fmodule-file=` the compile will fail early on (it is expecting this to indicate a source module, not a destination one).
>>
>> However, the second syntax `-fmodule-file=<name>=filename` I think should be able to work around this (since it says that the named module <name> is associated with the PCM `filename` which would allow us to cater for that file being missing (when we are creating it).
>
> Would the intent be that this might write out multiple different module files? I would hope we don't need to support that - that we only generate the module file for the single source module passed to the compiler. I guess maybe to support `clang++ x.cppm y.cppm` you might have multiple output files? But in that case you don't get to use `-o` so maybe you don't get to specify the module paths either and just get the default "same as .o but with .pcm suffix" behavior?
>
>> Does that clarify at all ?
>>
>> It would be great not to add more modules options flags, there are already way to many :/ - but if this seems too complex then one of the spellings suggested by @ChuanqiXu would work.
>
> Oh, right, super sorry - this is about how to specify the output filename. OK.
>
> (I'd still sort of lean towards "make it the same as the .o, but with the .pcm suffix" and if the build system wants to put things in other places, it can move them around - but I understand why that's not something everyone's on board with)

Actually, I would think that to be a great default (and put it in the CWD just like the object),
.. but it seems (in addition to the default) we do need a way to place specific module files - and I could imagine needing to make different places for different command line option sets.  Asking the driver or build system to move the file seems like spawning more processes and we'd still need a user-visible way to say where we wanted the file to be put.

> Yeah, if the `-fmodule-file` syntax is currently only for input I don't think we need to try to overload it for output as well.

My understanding [which could be flawed] is that the second syntax (which already exists, it's not an invention of this discussion) is intended to allow specification of output files.

> Perhaps we could extend the --precompile flag to take a filename? (but then that'd potentially confuse things when you want to produce both PCM and .o, passing `--precompile` currently only produces the .pcm on the main output path... ).

yeah, --precompile is well-defined to produce only the module (and I have a patch to alias it to -fmodule-only to match the GCC equivalent), it would not help to conflate that with the current objective.

> Anyone got a list of other compiler flags that take output file names? (I guess profile generation, maybe some other flags do this? Might be worth comparing/contrasting/seeking inspiration from those? I guess `-save-temps` uses the "same path, different suffix" strategy)

I'll see if I can find my master list of modules options (tomorrow .. erm later today).


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D134267/new/

https://reviews.llvm.org/D134267