Hi all,
Just thought I'd toss in my two cents on this:
Statically linking to other libraries when you are building your own
library is not considered a very good idea on Windows. The reason for this
is that Microsoft implements the C runtime (CRT) functionality on a
per-library basis. i.e., each version of the CRT (debug vs. release, 10.0
vs. 9.0, etc.) has its own private version of the heap, file handles, and
other CRT objects. This state is absolutely not shared among the dlls.
This means that you have to be VERY careful to ensure that CRT objects are
not manipulated by the wrong library (e.g., calling 'free' on a buffer
allocated in a different version of the CRT). Not paying careful attention
to this will result in subtle, difficult to find bugs down the road. When
you statically link to a version of the CRT, you aren't fixing the problem,
you are actually multiplying it. Each static link gives you a separate CRT
state, just as if you mixed dll versions.
In general, we try to be very good about this in the HDF5 library and it's
not really an issue in normal use, though I'm pretty sure our help desk
still fields questions that are resolved by ensuring all libraries and the
executable are using the same CRT.
Even if it's not that big of a problem for us, the official word from
Microsoft is:
"Because a DLL built by linking to a static CRT will have its own CRT
state, it is not recommended to link statically to the CRT in a DLL unless
the consequences of this are specifically desired and understood."
http://msdn.microsoft.com/en-us/library/abx4dbyh(VS.80).aspx
Yeah, the static link makes it super easy to build your software, but at
the risk of corrupt applications. Personally, I'd rather solve the "where
are my dlls" problem instead of the "why is my software mysteriously
crashing or corrupting data" problem. The first one is a LOT more
tractable.
If anyone is sitting on a pile of money they need to spend, this might
actually be a fixable problem. The Win32 libraries do not have this
linkage issue, so if we extended our platform-independence layer so that
our C calls mapped to Win32 calls, the problem should go away, with almost
no change to the bulk of the source code.
Personally, I'm not keen on making it easy for people to statically link to
the CRT since it's a documented bad idea. If we're going to spend any
money on Windows library issues, I'd rather see us properly distribute
debug libraries.
Cheers,
Dana
On Sat, Jun 8, 2013 at 3:48 PM, Pedro Vicente <pvicente@xxxxxxx> wrote:
> **
>
> Hi Elena, John, Ward, Allen, Michael
>
> > Support for /MT(d) was removed from CMake because we do not have
> resources to support it with the HDF software (including HDF5).
> > If this is a desired feature, we will need someone to test it on a
> regular basis (daily with development and release HDF5 branches).
>
> > I had set up 4 nightly builds of hdf5, static and dynamic, with and
> > without fortran, and unfortunately, all goes wrong when you start
> > using ifort on windows because of the way the supplied fortran libs
> > are linked (once you link the intel fortran libraries your link errors
> > multiply and you have no control of those libraries so having an
> > option to change static/dynamic is no help!)
> > I would concur that
> > 1) Adding a version stamp to the dll/lib names is desirable (and
> easy to do)
> > 2) Adding a simple static/dynamic run time link option (and easy to
> do)
>
>
> OK, so, why not just add now the simplest configurations that are already
> working, and that Ward has already done for netCDF?
>
> The only minimal configuration that I think is certainly needed for
> everyone is item 2), the static/dynamic run time link option,
> the equivalent of GNU automake --enable-static/shared
>
> One of the things that I am interested in is to build a Cmake build system
> for NCO also, so what do you all think of a common CMake system for HDF,
> netCDF and NCO?
>
> I mean, things like
>
> 1) common syntax options like netCDF's
>
> -DBUILD_SHARED_LIBS=OFF
>
> 2) use of environment variables, with the same name for libraries and
> header files names;
> for example, I use on my projects for the location of the curl library,
> the name
>
> LIB_CURL
>
> and netCDF uses
>
> CURL_LIBRARY
>
> this would make things simpler for anyone that needs to build HDF, netCDF
> and NCO.
>
> Elena,
> if you need to discuss in detail, call me (same number still), or we can
> set Skype video/audio joint conference calls if needed.
>
>
> Pedro
> ------
> Pedro Vicente, Earth System Science
> University of California, Irvine
> http://www.ess.uci.edu/
>
>
>
> ----- Original Message -----
> *From:* Biddiscombe, John A. <biddisco@xxxxxxx>
> *To:* HDF Users Discussion List <hdf-forum@xxxxxxxxxxxxxxxxxx>
> *Cc:* help@xxxxxxxxxxxx ; netcdfgroup@xxxxxxxxxxxxxxxx
> *Sent:* Friday, June 07, 2013 1:05 AM
> *Subject:* Re: [netcdfgroup] [Hdf-forum] Make the Cmake Windows build
> staticplease !
>
> I’d just like to second Elena’s comment below. As one of the people
> responsible for the cmake implementation of hdf5, I had a really hard time
> with this static/dynamic CRT link issue.****
>
> I had set up 4 nightly builds of hdf5, static and dynamic, with and
> without fortran, and unfortunately, all goes wrong when you start using
> ifort on windows because of the way the supplied fortran libs are linked
> (once you link the intel fortran libraries your link errors multiply and
> you have no control of those libraries so having an option to change
> static/dynamic is no help!)****
>
> So there was not a simple combination that was easy for all setups and we
> basically opted to remove most of the options to get a combination that
> worked - It took up so much of my time that I didn’t ‘finish the job
> cleanly’ and allow the user to decide for all cases.****
>
> ** **
>
> I would concur that****
>
> **1) **Adding a version stamp to the dll/lib names is desirable (and
> easy to do)****
>
> **2) **Adding a simple static/dynamic run time link option (and easy
> to do)****
>
> ** **
>
> I’m happy to test patches if people create them, but all my stuff works
> fine, so I’m not volunteering!****
>
> ** **
>
> JB****
>
> ** **
>
> [I would also personally favour dynamic linking in 99% of cases. When you
> have plugin based architectures where objects are created and destroyed in
> different modules, static linking is a no go. I do not suffer from DLL hell
> and actually find it much much much easier to distribute binaries to people
> on windows than linux]****
>
> ** **
>
> ** **
>
> -- ****
>
> John Biddiscombe, email:biddisco @.at.@ cscs.ch****
>
> http://www.cscs.ch/****
>
> CSCS, Swiss National Supercomputing Centre | Tel: +41 (91) 610.82.07****
>
> Via Trevano 131, 6900 Lugano, Switzerland | Fax: +41 (91) 610.82.82****
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> ** **
>
> *From:* Hdf-forum [mailto:hdf-forum-bounces@xxxxxxxxxxxxxxxxxx] *On
> Behalf Of *Elena Pourmal
> *Sent:* 06 June 2013 17:01
> *To:* HDF Users Discussion List
> *Cc:* help@xxxxxxxxxxxx; netcdfgroup@xxxxxxxxxxxxxxxx
> *Subject:* Re: [Hdf-forum] Make the Cmake Windows build static please !***
> *
>
> ** **
>
> All,****
>
> ** **
>
> Support for /MT(d) was removed from CMake because we do not have resources
> to support it with the HDF software (including HDF5).****
>
> ** **
>
> If this is a desired feature, we will need someone to test it on a regular
> basis (daily with development and release HDF5 branches). If there are
> volunteers, please contact Allen. We will go from there.****
>
> ** **
>
> Thank you!****
>
> ** **
>
> Elena****
>
> ** **
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Elena Pourmal
> Director of Technical Services and Operations
> The HDF Group
> 1800 So. Oak St., Suite 203,
> Champaign, IL 61820
> www.hdfgroup.org
> (217)531-6112 (office)
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
>
>
> ****
>
> ** **
>
> On Jun 6, 2013, at 3:42 AM, Pedro Vicente wrote:****
>
>
>
> ****
>
>
> Hey , Allen & Ward, next time I send an email to *both* of your group
> lists, can you just do a "reply-all" ? :-) ... otherwise , we get half of
> the answers.****
>
> ****
>
>
> >>Building NetCDF statically is already an option, by passing
> -DBUILD_SHARED_LIBS=OFF during configuration.
> >>This will build the netcdf libraries and utilities statically, avoiding
> the direct dependency on MSVCR100.dll.
> >>They will, however, still inherit any downstream .dll dependencies from
> the curl, hdf and zlib libraries. ****
>
> ****
>
> Exactly, Ward, netCDF inherits HDF, that 's why I am bugging Allen here :-)
> ****
>
> ****
>
> And, yes, the solution is to build *all* downstream code in the *same*
> way, either all static or all dynamic;
> mixing both will get you linking errors.****
>
> ****
>
> >>> You might be able to work around this by downloading (or building)
> static versions of these libraries. ****
>
> ****
>
> Yes, Ward, this can be done by editing the HDF Cmake generated VS
> projects, and changing all of them to static, which I just did.****
>
> ****
>
> In the curl case, you can do this by editing the file MakefileBuild.vc and
> change all /MD (dynamic) to /MT (static)****
>
> ****
>
> Runtime library configuration
> !IF "$(RTLIBCFG)"=="static"
> #RTLIB = /MT
> RTLIB = /MTd
> RTLIB_DEBUG = /MTd
> !ELSE
> #RTLIB = /MD
> #RTLIB_DEBUG = /MDd****
>
> ****
>
>
> >>Our binaries redistribute the MS CRT dlls that are used to build the
> binaries.****
>
> ****
>
> Hey, Allen, I was not asking for the *binaries* !****
>
> ****
>
> I was asking for an *option* in your CMake system that allows
> anyone who wants to generate the Visual Studio projects have them
> *already* "static ready" if they wish...****
>
> ****
>
> So that, next time I need to generate everything from scratch I don't have
> to go each to each one of them (50 + projects ) and click the static button
> 50 times ...****
>
> ****
>
>
> >>Because of the danger of using different CRTs (HDF lib uses one and user
>
> >>program uses different one) and the possible memory corruption with
> >>allocations, we build with /MD and provide the CRT with our binary.****
>
> ****
>
> I don't pretend to know everything, but when I don't (often :-) ) I search
> ****
>
> ****
>
> read this****
>
> ****
>
>
> http://stackoverflow.com/questions/14749662/microsoft-visual-studio-c-c-runtime-library-static-dynamic-linking
> ****
>
> ****
>
> quote****
>
> ****
>
> " i personally prefer statically linked. i hate scrambling around looking
> for the right redist/dll/etc."****
>
> ****
>
> that's what I do, I set to all static , and then move on to do more
> interesting things, like writing software, than dealing with DLL
> idiosyncracies.****
>
> ****
>
> quote****
>
> ****
>
> "Using /MT is risky if you create DLLs as well as an EXE. You'll end up
> with multiple copies of the CRT in your program.
> This was especially a problem with earlier versions of VS where each CRT
> would get its own heap, "****
>
> ****
>
>
> Was this the problem you meant ?****
>
> ****
>
> This might be true if you are distributing *binaries* with both the EXE
> and DLL. Or if you are linking your code against a 3rd party library
> *without* the code,
> someone gave you a LIB and a DLL only.****
>
> ****
>
> In the HDF, netCDF and NCO worlds none of this is true: all sources are
> available , no secrets here :-)****
>
> ****
>
> And you are lucky, Allen , because you only have the downstream ZLIB, and
> netCDF only has curl****
>
> ****
>
> Here's a list of all NCO dependencies****
>
> ****
>
> zlib ,
> HDF5,
> curl,
> netCDF, including OpenDAP, thank you for that :-)
> ANTLR, a parser generator for ncap2
> GSL, the GNU scientific library
> UDUnits, Unidata units (Not yet available for Windows)
> Regular Expressions (Not yet available for Windows, tough one this one )**
> **
>
> ****
>
> I have Visual Studio projects for all these (except UDUnits and Regular
> Expressions)
> , because I need to build the *source* , as you can see many projects to
> change to static/dynamic/32/64bit/debug/release combinations :-)****
>
> ****
>
> Pedro****
>
> ****
>
> ****
>
> ------
> Pedro Vicente, Earth System Science
> University of California, Irvine
> http://www.ess.uci.edu/****
>
> ****
>
> ****
>
> ****
>
> ----- Original Message -----****
>
> From: "Allen Byrne" <byrn@xxxxxxxxxxxx>****
>
> To: <hdf-forum@xxxxxxxxxxxxxxxxxx>****
>
> Sent: Wednesday, June 05, 2013 6:18 AM****
>
> Subject: Re: [Hdf-forum] Make the Cmake Windows build static please !****
>
> ** **
>
> > Our binaries redistribute the MS CRT dlls that are used to build the
> binaries.
> > Because of the danger of using different CRTs (HDF lib uses one and user
>
> > program uses different one) and the possible memory corruption with
> > allocations, we build with /MD and provide the CRT with our binary.
> >
> > Hopefully, as more people use CMake and build a common knowledge of using
>
> > CMake, those wishing to build alternative versions of HDF will share
> their
> > code changes.
> >
> > Allen
> >
> > On Wednesday, June 05, 2013 09:34:37 AM Michael Jackson wrote:
> >> Funny, I actually _prefer_ the DLL builds because I distribute the
> runtime
> >> C/C++ libraries as allowed by MS. Depending on how one is using the HDF5
> >> executables/libraries having each library linked statically against
> their
> >> own C/C++ libraries can also lead to problems because of how memory is
> >> allocated/deallocated in each library version. There are 2 evils here
> and
> >> the idea is to pick the least of them. If anything I would petition the
> >> HDFGroup to provide BOTH a dynamically linked AND a statically linked
> >> runtime version.
> >>
> >> Just my 2 cents. Then again, I build my own HDF5 for our project
> precisely
> >> because of issues like this.
> >> ___________________________________________________________
> >> Mike Jackson Principal Software Engineer
> >> BlueQuartz Software Dayton, Ohio
> >> mike.jackson@xxxxxxxxxxxxxx www.bluequartz.net
> >>
> >> On Jun 5, 2013, at 8:24 AM, Pedro Vicente <pvicente@xxxxxxx> wrote:
> >> > Hi Allen, Ward
> >> >
> >> > I have a request regarding your new CMake Windows build system, could
> you
> >> > add an option to make the build static regarding the Microsoft
> libraries
> >> > (MSVCR100D.dll) ?
> >> >
> >> > Starting with version 4.3.1, NCO
> >> >
> >> > http://nco.sourceforge.net/
> >> >
> >> > uses the HDF5 and netCDF Windows libraries made with your CMake
> system,
> >> > and this is causing problems for NCO users, as explained here
> >> >
> >> > https://sourceforge.net/projects/nco/forums/forum/9830/topic/8345151
> >> >
> >> > and here
> >> >
> >> > https://sourceforge.net/projects/nco/forums/forum/9829/topic/8417103
> >> >
> >> >
> >> > This is just a matter of changing the compiler flag to /MT(d)
> >> >
> >> > http://msdn.microsoft.com/en-us/library/2kzt1wy3.aspx
> >> >
> >> > Using a dynamic build is just a bad idea, because of these DLL issues.
> >> >
> >> > I have some Windows executables from code I did in the early 90's ,
> that
> >> > unfortunately I cannot run today, just because I linked them with
> DLLs,
> >> > with the DLLs from the Visual Studio from that time, that do not exist
> >> > anymore (it seems every new version they change the Visual Studio
> Dlls).
> >> >
> >> > Because of this I do not use Dlls, I know that eventually something
> will
> >> > go wrong :-)
> >> >
> >> > Pedro
> >> >
> >> > ------
> >> > Pedro Vicente, Earth System Science
> >> > University of California, Irvine
> >> > http://www.ess.uci.edu/
> >> >
> >> >
> >> > _______________________________________________
> >> > Hdf-forum is for HDF software users discussion.
> >> > Hdf-forum@xxxxxxxxxxxxxxxxxx
> >> >
> http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
> >>
> >> _______________________________________________
> >> Hdf-forum is for HDF software users discussion.
> >> Hdf-forum@xxxxxxxxxxxxxxxxxx
> >> http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
> >
> > _______________________________________________
> > Hdf-forum is for HDF software users discussion.
> > Hdf-forum@xxxxxxxxxxxxxxxxxx
> > http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
> >
> > _______________________________________________
> Hdf-forum is for HDF software users discussion.
> Hdf-forum@xxxxxxxxxxxxxxxxxx
> http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org***
> *
>
> ** **
>
> ------------------------------
>
> _______________________________________________
> netcdfgroup mailing list
> netcdfgroup@xxxxxxxxxxxxxxxx
> For list information or to unsubscribe, visit:
> http://www.unidata.ucar.edu/mailing_lists/
>
>
> _______________________________________________
> Hdf-forum is for HDF software users discussion.
> Hdf-forum@xxxxxxxxxxxxxxxxxx
> http://lists.hdfgroup.org/mailman/listinfo/hdf-forum_lists.hdfgroup.org
>
>
