[cfe-dev] [Modules] Silent textual inclusion when files not found.

[Richard Smith]

On Thu, Jun 4, 2015 at 1:49 PM, Sean Silva <chisophugis at gmail.com> wrote:

> On Thu, Jun 4, 2015 at 1:27 PM, Richard Smith <richard at metafoo.co.uk>
> wrote:
>
>> On Wed, Jun 3, 2015 at 6:24 PM, Sean Silva <chisophugis at gmail.com> wrote:
>>
>>> On Wed, Jun 3, 2015 at 8:40 AM, Ben Langmuir <blangmuir at apple.com>
>>> wrote:
>>>
>>>>
>>>> On Jun 2, 2015, at 7:57 PM, Sean Silva <chisophugis at gmail.com> wrote:
>>>>
>>>> Hi all,
>>>>
>>>> Currently we do not seem to issue any form of diagnostic when there are
>>>> missing headers named in a module map. See the attached test case. Even
>>>> worse, we will just treat all headers in the module as textual. Hilarity
>>>> ensues.
>>>>
>>>>
>>>> Yep, this is llvm.org/PR20507
>>>>
>>>>
>>>> Some prior art in this area:
>>>> Daniel - r197485
>>>> Ben - r206664
>>>>
>>>> Based on these commits, it ostensibly seems that clang does *some* sort
>>>> of checking for missing files in a module map, but I can't seem to coax
>>>> clang into doing this in C++ language mode.
>>>>
>>>> Looking at the source code, it seems like we end up conflating
>>>> "unavailable" due to a failed `requires` with "a header is missing". It
>>>> sounds like we essentially need two notions "unsatisfied `requires`" and
>>>> "necessary header is missing" (a header guarded by an unsatisfied
>>>> `requires` does not count as "necessary"). Haven't dug in deep yet, but my
>>>> hypothesis is that somewhere along the way we silently treat a module with
>>>> missing headers as though it had an unsatisfied `requires`, leading to us
>>>> silently neglecting that it was ever in a module at all.
>>>>
>>>>
>>>> Yes, they both come out as the module being “unvavailable”, which is
>>>> why we don’t import it.  I think the right answer is that attempting to
>>>> import any unavailable module (even via an auto-import) should be an
>>>> error.  That is:
>>>>
>>>> module MissingHeader {
>>>>     header “exists.h"
>>>>     header “doesnt_exist.h”
>>>> }
>>>>
>>>> #include <exists.h> // error, this would import MissingHeader, which is
>>>> unavailable because we're missing header “doesnt_exist.h"
>>>>
>>>> module Top {
>>>>     header “Top.h”
>>>>     module A {
>>>>         requires non_existent
>>>>        header “A.h”
>>>>     }
>>>>     module B {
>>>>         header “B.h”
>>>>     }
>>>> }
>>>>
>>>> #include <Top.h> // OK
>>>> #include <A.h> // error, this would import Top.A, which is unavailable
>>>> because of the missing requirment
>>>> #include <B.h> // OK
>>>>
>>>> But handling the first case is more important I think.
>>>>
>>>>
>>>> I'm glad to put some effort into fixing this; I spent a good part of
>>>> today with a bizarre error that I traced back to this and I don't want my
>>>> users to have to deal with the same. Any pointers would be appreciated.
>>>>
>>>>
>>>> That would be great! I was coincidentally hoping to fix this myself in
>>>> the next few weeks, but if you beat me to it even better.
>>>>
>>>>
>>> I'm still acclimating myself to the modules code, do you have any
>>> concrete thoughts on how to do this? I was just planning on looking through
>>> for everywhere we use "unavailable" happens and figure out a pattern, but
>>> if you already have an idea, it might save me some time.
>>>
>>
>> Something like:
>>  * make ModuleMap::findModuleForHeader able to produce an unavailable
>> module, but treat it as being worse than any available module
>>  * make Preprocessor::HandleIncludeDirective issue an error if its call
>> to LookupFile produces a SuggestedModule that is unavailable
>>
>> This should probably happen even if modules is disabled (that is, when we
>> are parsing module maps but not actually building / using precompiled
>> module files).
>>
>
> Thanks for the pointers. I'll take a look at doing this.
>
>
>>
>> Also, any preferred bikeshed for the names for "unvailable because of
>>> unsatisfied requires" and "necessary header is missing" concepts in the
>>> source?
>>>
>>
>> If the response to both is the same (as Ben suggested, and I agree) then
>> I'm not sure that we need to distinguish them anywhere other than the place
>> where we produce the diagnostic.
>>
>
> But aren't they different when it comes time for clang to build the
> module? We want to build a top-level module even if it has a submodule with
> an unsatisfied "requires" (e.g. _Builtin_intrinsics has submodules with
> contradictory `requires`), taking appropriate action to exclude any headers
> with unsatisfied `requires`. On the other hand (modulo submodules already
> removed from consideration due to unsatisfied `requires`), a missing header
> should be treated as a hard error.
>

I think you're right, but I think that will be the emergent behavior from
treating the two cases the same: if a module's header is missing, the
corresponding *top-level* module (and all its submodules) gets marked
unavailable, so we should never try to implicitly build it.

There might be a file system race here, where the main build process thinks
the module is available, but the module build finds a header is missing and
builds an empty module because all the submodules are unavailable. That's
probably worth fixing, but again the fix doesn't really depend on the
different kinds of availability: if the top-level module is unavailable, we
should produce an error if we try to build a module file for it. (This is
probably easy to test by building an unavailable module with -cc1
-emit-module.)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20150604/b90e057c/attachment.html>