[llvm-dev] [cfe-dev] Who wants faster LLVM/Clang builds?

[Philip Reames via llvm-dev]

On 12/12/2017 01:38 PM, Mikhail Zolotukhin via llvm-dev wrote:
>
>> On Dec 12, 2017, at 12:57 PM, James Y Knight <jyknight at google.com 
>> <mailto:jyknight at google.com>> wrote:
>>
>>
>>
>> On Mon, Dec 11, 2017 at 3:37 PM, Mikhail Zolotukhin via cfe-dev 
>> <cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>> wrote:
>>
>>     Hi Kim,
>>
>>>     On Dec 10, 2017, at 7:39 AM, Kim Gräsman <kim.grasman at gmail.com
>>>     <mailto:kim.grasman at gmail.com>> wrote:
>>>
>>>     Hi Michael,
>>>
>>>     On Thu, Dec 7, 2017 at 3:16 AM, Michael Zolotukhin
>>>     <mzolotukhin at apple.com <mailto:mzolotukhin at apple.com>> wrote:
>>>>
>>>>     Nice to IWYU developers here:) I wonder how hard it would be to
>>>>     run IWYU on
>>>>     LLVM/Clang (or, if it’s supposed to work, I wonder what I did
>>>>     wrong).
>>>
>>>     There are known problems with running IWYU over LLVM/Clang --
>>>     Zachary
>>>     Turner made an attempt a while back to get it up and running. Since
>>>     the LLVM tree uses all sorts of modern and moderately complex
>>>     patterns, we're struggling to keep up.
>>     I see.
>>>
>>>>     If we also can tweak it a bit to make it choose more human-like
>>>>     (~more
>>>>     conservative) decisions, we would be able to just apply what it
>>>>     suggests!
>>>
>>>     Different humans appear to have different preferences :)
>>     True, what I meant hear is to make the changes more conservative:
>>     e.g. if we can replace
>>     #include "MyClass.h"
>>     with
>>     class MyClass;
>>     then this change is probably desirable in every way: it documents
>>     the code better, it decreases coupling, it improves compile time.
>>
>>
>> This is not a transform which is clearly "desirable in every way", 
>> because it _increases_ coupling between the user of the class and the 
>> implementation. The owner of the class can't add optional template 
>> arguments, change a class into a typedef, change the namespace, or 
>> other such refactorings. It also introduces the possibility of code 
>> which changes behavior depending on whether the full or forward decl 
>> are available (which, then, may be an ODR-violation).
>>
>> Effectively the same reasons why the standard forbids users from 
>> forward-declaring std:: names apply to doing so to user-defined names.
>>
>> https://google.github.io/styleguide/cppguide.html#Forward_Declarations 
>> lists some of the issues, and a recommendation not to do so.
>>
>> Of course you do have the upside is that it can improve compile time. 
>> Which is certainly of value, and perhaps that's a worthwhile 
>> trade-off when the implementation and forward-declare are both within 
>> a single project and thus easy to coordinate. But, it's not by any 
>> means a _pure_ win.
> That's correct. I was speaking about the LLVM codebase though (I 
> should've stated that clearer), and in LLVM I don't remember many 
> occasions of refactorings you mentioned. For LLVM forward declaration 
> is recommended by the style guide:
> http://llvm.org/docs/CodingStandards.html#minimal-list-of-includes
>
As an aside, there is a standard idiom used in some code bases which 
might be applicable.

For each header "Header.h" which contains a class MyClass, you introduce 
a header called "Header.FwdDecls.h" which contains forward decls for all 
of the classes declared in the header.  There are also variants which 
declare "MyClass.FwdDecl.h".  Both of these schemes have the advantage 
of only putting the forward decl in one place while reducing the 
transitive include set greatly.  I've seen environments where these 
forward decl headers are automatically generated by the build system.

Philip
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171214/a19e5176/attachment.html>