C++ Locales under GCC

The C++ locales is a cool feature for software internationalization and everybody certainly knows about or have heard of it. But when trying to use it on anything but a Windows or Linux system one may find out a crippled implementation. This is the case when GCC is ported to a different platform such as Solaris.

At first I thought it could be a simple matter of getting a more updated version of GCC. Then I've tried to document the potential solution on a previous post. But while digging into the matter I've sadly realized it's a more complicated issue.

As a matter of fact the issue lies in GNU libstdc++ which unfortunately does not provide a complete implementation of Locales. Its implementation is target just for systems already possessing a working implementation of the GNU C Library (glibc) which offers a key function usually found at newlocale.c source module that exposes the "special property" of being thread safe" at the C++ level". Solaris, as a standard UNIX system, uses libc which is thread safe just at the C level. Hence, GNU works around the issue by providing what is called a "general stub" only supporting "C"  (= "POSIX") locale and otherwise throwing an exception. This has been made clear as per the following archived message: generic locale.

It seems that the main issue is an underlying libc limitation of not providing necessary API points for building up a multithreaded implementation of libstdc++ on top of it. As a consequence one entirely lacks a key C++ feature because of an unresolved multithreading issue. At this point it turns out easier to understand why there's a GNU C Library on the first place and why GNU libstdc++ relies on it instead of a particular system's libc.

But wondering about this multithreading issue between the C and C++ libraries when dealing with other locale dependent calls, one could realized that in most cases all that's necessary is to simply set the locale once on start-up code or while entering the main thread. From that point on, as long as a program remains pure C++ at the locale-dependent high-level I/O calls all should work pretty well.
 
I mean that in general a program isn't expected to be switching locales anymore after having properly intialized itself. I know I could be missing many reasons why the contrary should hold, but it seems they all would narrow down to some processing that should be carried out multiple times (in parallel) for different locales, such as processing some sort of localized I/O to several branch offices under a single process.

Therefore due to strictness, an all or nothing situation, one ends up with virtually nothing at all under Solaris and other less popular platforms for which nobody else carried about on providing a slightly more decent implementation of the C++ standard :-(
 

Shared Libraries - Example #1

Let's try a first example of a fictitious dynamic application (one that links to dynamic objects ‒ a shared library in this case). The key aspects I wish to address on this post are how to build these artifacts under Solaris 11.3 using GCC 4.8.2 and properly adhering to the versioning naming convention. I would like to emphasize that in the name of straightforward software engineering there will be no excessively elaborated interfaces strategies, for instance by using crazy things such as GUIDs (or so) or any other related / similar interface infrastructure (if you like GUIDs you may perhaps have some fun playing with a GUID generator). My strategy is pure C++ resulting in a much more powerful, efficient and higher-quality artifact than some sort of software crazy and buggy Frankenstein monster such as COM and related. I'll be using C++ namespaces and a main C++ class on each library module.

The software engineering implementation/deployment depicted on this example consists (for the sake of simplicity) on a single (do-nothing) shared library and a (stub) dynamic application linking to it just for testing and demonstration. I'll build 2 major versions of the library (libgreeting.so.1 and libgreeting.so.2) the first of them with some different minor versions and releases (for demonstration purposes). For each major version of the library I'll show a backward compatibility tehcnique for avoiding or minimizing the rebuild of the application (myapp-1). Accompanying each new major version of the library there will be (for the sake of simplicity) a new / updated header-file which as I said will provide for backward compatibility with previous major version(s). The header-files shouldn't be modified within minor versions and releases as this would be a clear deviation of whole strategy. For demonstration purposes, the implementation of each interface will consist on multiple source-code files. Besides the interface mechanism the library will be able to "report" its hard-coded versioning information (as a kind of reflection) which will always correspond to the name of the shared library version final binary object, although (again for the sake of simplicity) there will be no provisions to enforce this correspondence.

I can only hope that the software engineering aspects I'll be exploring side-by-side do not hinder the main goals of this post about building and dealing with shared libraries!

GNU - Build preparation

OUT OF DATE

It's no doubt very convenient and trustworthy to install new software through IPS or even the legacy SVr4 packages. But when those options are not available fortunately there's still a chance for the GNU build automation, which although not as convenient can also perform a good job.

The well-known essentials of GNU build automation end-user operation work-flow was the subject of another post, in which I hopefully confirmed to work reasonably with a few adjustments under Solaris 11.3 and in which I successfully integrated some well-known ZFS essential advantages, such as source-tree snapshots.

In this post I hope to somewhat complement what's been discussed by presenting a suggestion of Shell script to hopefully simplify and avoid some common mistakes in the preparation of the GNU build automation work-flow environment.

For instance, assume that an application source code (tarball app-1.0.0.tar.gz) is to be built for both 32-bit and 64-bit targets using GNU tools and that the build operation is to take place at a local staff /software/build with the results installed locally under /software/prototype.

NOTE

I assume that the source code tarball is usually named under the following convention: name-version.tar.compression.

But there are cases this convention varies, for instance. node.js prepends a v before version, which slight variation, in this particular case, is a minor annoyance and fortunately won't break the following script. After the script is run, if desired, one can rename the datasets accordingly in order to normalize things, by removing the v before version number.

But there are other cases such as maven (apache-maven-3.5.3-src.tar.gz) where the tarball name would break the script. I could attempt fixing it, but as there are so many unpredictable possibilities this would certainly become a unmanageable.
 
The general solution is to conform the tarballs' names to my original assumption, usually with a simple manual procedure (as root) of extracting the archive and then storing everything back under a conformant archive name. For instance:
 
# cd /tmp
# gtar xf .../maven/3.5.3/apache-maven-3.5.3-src.tar.gz
# gtar czf maven-3.5.3.tar.gz maven-3.5.3
# cp maven-3.5.3.tar.gz .../maven/3.5.3/
 
As an alternative while recreating the tarball, one can attempt higher compression rates with j (for a .bz2) or J (for a .xz) instead of just z (for a more common .gz compressed file). On this particular example the sizes become: 2.6M (.gz), 2.2M (.bz2) and 1.8M (.xz).

GNU - Build Automation

Solaris 11.3 includes many modern packages, among them some essential GNU build automation ones. The problem is that the pace of update of these third-party packages is much faster than what get updated through support repository, needless to say the public release repository. In order to better keep up with this third-party pace, there's no other better alternative than to get the sources and build ourselves what's needed, in the hope (and unfortunately sometimes it's no more than just a hope) that the sources and the build tools are compatible and available to Solaris.