RFR (2xS): 8181318: Allow C++ library headers on Solaris Studio

David Holmes david.holmes at oracle.com
Thu Jun 1 12:33:45 UTC 2017 

Hi Erik,

Just to be clear it is not the use of <limits> that I am concerned 
about, it is the -library=stlport4. It is the use of that flag that I 
would want to check in terms of having no affect on any existing code 
generation.

I'm finding the actual build situation very confusing. It seems to me in 
looking at the hotspot build files and the top-level build files that 
-xnolib is used for C++ compilation & linking whereas -library=%none is 
used for C compilation & linking. But the change is being applied to 
$2JVM_CFLAGS which one would think is for C compilation but we don't 
have $2JVM_CXXFLAGS, so it seems to be used for both!

David

On 1/06/2017 7:36 PM, Erik Österlund wrote:
> Hi David,
> 
> On 2017-06-01 08:09, David Holmes wrote:
>> Hi Kim,
>>
>> On 1/06/2017 3:51 PM, Kim Barrett wrote:
>>>> On May 31, 2017, at 9:22 PM, David Holmes <david.holmes at oracle.com> 
>>>> wrote:
>>>>
>>>> Hi Erik,
>>>>
>>>> A small change with big questions :)
>>>>
>>>> On 31/05/2017 11:45 PM, Erik Österlund wrote:
>>>>> Hi,
>>>>> It would be desirable to be able to use harmless C++ standard 
>>>>> library headers like <limits> in the code as long as it does not 
>>>>> add any link-time dependencies to the standard library.
>>>>
>>>> What does a 'harmless' C++ standard library header look like?
>>>
>>> Header-only (doesn't require linking), doesn't run afoul of our
>>> [vm]assert macro, and provides functionality we presently lack (or
>>> only handle poorly) and would not be easy to reproduce.
>>
>> And how does one establish those properties exist for a given header 
>> file? Just use it and if no link errors then all is good?
> 
> Objects from headers that are not ODR-used such as constant folded 
> expressions are not imposing link-time dependencies to C++ libraries. 
> The -xnolib that we already have in the LDFLAGS will catch any 
> accidental ODR-uses of C++ objects, and the JVM will not build if that 
> happens.
> 
> As for external headers being included and not playing nicely with 
> macros, this has to be evaluated on a case by case basis. Note that this 
> is a problem that occurs when using system headers (that we are already 
> using), as it is for using C++ standard library headers. We even run 
> into that in our own JVM when e.g. the min/max macros occasionally slaps 
> us gently in the face from time to time.
> 
>>
>>> The instigator for this is Erik and I are working on a project that
>>> needs information that is present in std::numeric_limits<> (provided
>>> by the <limits> header).  Reproducing that functionality ourselves
>>> would require platform-specific code (with all the complexity that can
>>> imply).  We'd really rather not re-discover and maintain information
>>> that is trivially accessible in every standard library.
>>
>> Understood. I have no issue with using <limits> but am concerned by 
>> the state of stlport4. Can you use <limits> without changing 
>> -library=%none?
> 
> No, that is precisely why we are here.
> 
>>
>>>>> This is possible on all supported platforms except the ones using 
>>>>> the solaris studio compiler where we enforce -library=%none in both 
>>>>> CFLAGS and LDFLAGS.
>>>>> I propose to remove the restriction from CFLAGS but keep it on 
>>>>> LDFLAGS.
>>>>> I have consulted with the studio folks, and they think this is 
>>>>> absolutely fine and thought that the choice of -library=stlport4 
>>>>> should be fine for our CFLAGS and is indeed what is already used in 
>>>>> the gtest launcher.
>>>>
>>>> So what exactly does this mean? IIUC this allows you to use headers 
>>>> for, and compile against "STLport’s Standard Library implementation 
>>>> version 4.5.3 instead of the default libCstd". But how do you then 
>>>> not need to link against libstlport.so ??
>>>>
>>>> https://docs.oracle.com/cd/E19205-01/819-5267/bkakg/index.html
>>>>
>>>> "STLport is binary incompatible with the default libCstd. If you use 
>>>> the STLport implementation of the standard library, then you must 
>>>> compile and link all files, including third-party libraries, with 
>>>> the option -library=stlport4”
>>>
>>> It means we can only use header-only parts of the standard library.
>>> This was confirmed / suggested by the Studio folks Erik consulted,
>>> providing such limited access while continuing to constrain our
>>> dependency on the library.  Figuring out what can be used will need to
>>> be determined on a case-by-case basis.  Maybe we could just link with
>>> a standard library on Solaris too.  So far as I can tell, Solaris is
>>> the only platform where we don't do that.  But Erik is trying to be
>>> conservative.
>>
>> Okay, but the docs don't seem to acknowledge the ability to use, but 
>> not link to, stlport4.
> 
> Not ODR-used objects do not require linkage. 
> (http://en.cppreference.com/w/cpp/language/definition)
> I have confirmed directly with the studio folks to be certain that 
> accidental linkage would fail by keeping our existing guards in the 
> LDFLAGS rather than the CFLAGS.
> This is also reasonably well documented already 
> (https://docs.oracle.com/cd/E19205-01/819-5267/bkbeq/index.html).
> 
>>
>>>> There are lots of other comments in that document regarding STLport 
>>>> that makes me think that using it may be introducing a fragile 
>>>> dependency into the OpenJDK code!
>>>>
>>>> "STLport is an open source product and does not guarantee 
>>>> compatibility across different releases. In other words, compiling 
>>>> with a future version of STLport may break applications compiled 
>>>> with STLport 4.5.3. It also might not be possible to link binaries 
>>>> compiled using STLport 4.5.3 with binaries compiled using a future 
>>>> version of STLport."
>>>>
>>>> "Future releases of the compiler might not include STLport4. They 
>>>> might include only a later version of STLport. The compiler option 
>>>> -library=stlport4 might not be available in future releases, but 
>>>> could be replaced by an option referring to a later STLport version."
>>>>
>>>> None of that sounds very good to me.
>>>
>>> I don't see how this is any different from any other part of the
>>> process for using a different version of Solaris Studio.
>>
>> Well we'd discover the problem when testing the compiler change, but 
>> my point was more to the fact that they don't seem very committed to 
>> this library - very much a "use at own risk" disclaimer.
> 
> If we eventually need to use something more modern for features that 
> have not been around for a decade, like C++11 features, then we can 
> change standard library when that day comes.
> 
>>
>>> stlport4 is one of the three standard libraries that are presently
>>> included with Solaris Studio (libCstd, stlport4, gcc).  Erik asked the
>>> Studio folks which to use (for the purposes of our present project, we
>>> don't have any particular preference, so long as it works), and
>>> stlport4 seemed the right choice (libCstd was, I think, described as
>>> "ancient").  Perhaps more importantly, we already use stlport4,
>>> including linking against it, for gtest builds.  Mixing two different
>>> standard libraries seems like a bad idea...
>>
>> So we have the choice of "ancient", "unsupported" or gcc :)
>>
>> My confidence in this has not increased :)
> 
> I trust that e.g. std::numeric_limits<T>::is_signed in the standard 
> libraries has more mileage than whatever simplified rewrite of that we 
> try to replicate in the JVM. So it is not obvious to me that we should 
> have less confidence in the same functionality from a standard library 
> shipped together with the compiler we are using and that has already 
> been used and tested in a variety of C++ applications for over a decade 
> compared to the alternative of reinventing it ourselves.
> 
>> What we do in gtest doesn't necessarily make things okay to do in the 
>> product.
>>
>> If this were part of a compiler upgrade process we'd be comparing 
>> binaries with old flag and new to ensure there are no unexpected 
>> consequences.
> 
> I would not compare including <limits> to a compiler upgrade process as 
> we are not changing the compiler and hence not the way code is 
> generated, but rather compare it to including a new system header that 
> has previously not been included to use a constant folded expression 
> from that header that has been used and tested for a decade. At least 
> that is how I think of it.
> 
> Thanks,
> /Erik
> 
>>
>> Cheers,
>> David
>>
>>>>
>>>> Cheers,
>>>> David
>>>>
>>>>
>>>>> Webrev for jdk10-hs top level repository:
>>>>> http://cr.openjdk.java.net/~eosterlund/8181318/webrev.00/
>>>>> Webrev for jdk10-hs hotspot repository:
>>>>> http://cr.openjdk.java.net/~eosterlund/8181318/webrev.01/
>>>>> Testing: JPRT.
>>>>> Will need a sponsor.
>>>>> Thanks,
>>>>> /Erik
>>>
>>>
>

More information about the build-dev mailing list