Compiling a standardized MPI shared object - linker

I have written a package in C++ that is parallelized using MPI. However, since there are so many complicated templates, it will take about half an hour for it to get compiled. Naturally many users have complained about this.
To resolve this I thought the best course of action would be to precompile all of the code. Then bundle all of the shared objects into one, but WITHOUT linking to MPI. The end user will get the shared object according to their platform and link it to implementation of MPI on their system.
However after doing so and taking "nm -B --undefined-only". I saw that there are symbols that are defined only in openmpi (the header files I used were from openmpi), for example ompi_mpi_op_sum or ompi_mpi_int.
Although I use the standard definitions in the header files there are macros that do this:
MPI_Sum-> ompi_mpi_op_sum
MPI_Int-> ompi_mpi_int
I looked around the and I could not find anything on the subject. Is there anyway to overcome this? or another solution to this problem?

As #ptb pointed out in the comments to question, there is a similar thread that deals with similar subject:
How to write an MPI wrapper for dynamic loading
After reading the thread and doing further research on Bridge design pattern and etc. I found the solution: static PIMPL (see https://www.gamedev.net/articles/programming/general-and-gameplay-programming/static-zero-overhead-probably-pimpl-in-c-r4338/)
I have created a github repository for anyone who has the same problem. It is by no means perfect but I am willing to expand/improve/generalize it upon request.
https://github.com/sinamoeini/mpi_bridge

Related

How to circumvent dlopen() caching?

According to its man page, dlopen() will not load the same library twice:
If the same shared object is loaded again with dlopen(), the same
object handle is returned. The dynamic linker maintains reference
counts for object handles, so a dynamically loaded shared object is
not deallocated until dlclose() has been called on it as many times
as dlopen() has succeeded on it. Any initialization returns (see
below) are called just once. However, a subsequent dlopen() call
that loads the same shared object with RTLD_NOW may force symbol
resolution for a shared object earlier loaded with RTLD_LAZY.
(emphasis mine).
But what actually determines the identity of shared objects? I tried to look into the code, but did not come very far. Is it:
some form of normalized path name (e.g. realpath?)
the inode ?
the contents of the libray?
I am pretty sure that I can rule out this last point, since an actual filesystem copy yields two different handles.
To explain the motivation behind this question: I am working with some code that has static global variables. I need multiple instances of that code to run in a thread-safe manner. My current approach is to compile and link said code into a dynamic library and load that library multiple times. With some linker magic, it appears to create several copies of the globals and resolve access in each library to its own copies. The only problem is that my prototype copies the generated library n times for n concurrent uses. This is not only somewhat ugly but I also suspect that it might break on a different platform.
So what is the exact behaviour of dlopen() according to the POSIX standard?
edit: Because it came up in a comment and an answer, no refactoring the code is definitely not an option. It would involve months or even years of work and potentially sacrifice all benefits of using the code in the first place. There exists an ongoing research project that might solve this problem in a much cleaner way, but it is actual research and might fail. I need a solution now.
edit2: Because people still seem to not believe the usecase is actually valid. I am working on a pure functional language, that shall be embedded into a larger C/C++ application. Because I need a prototype with a garbage collector, a proven typechecker, and reasonable performance ASAP, I used OCaml as intermediate code. Right now, I am compiling a source module into an OCaml module, link the generated object code (including startup etc.) into a shared library with the OCaml runtime and dlopen() that shared library. Every .so has its own copy of the runtime, including several global variabels (e.g. the pointer to the young generation) and that is, or rather should be, totally fine. The library exposes exactly two functions: An initializer and a single export that does whatever the original module is intended to do. No symbols of the OCaml runtime are exported/shared. when I load the library, its internal symbols are relocated as expected, the only issue I have right now is that I actually need to copy the .so file for each instance of the job at runtime.
Regarding thread-local-storage: That is actually an interesting idea, as the modification to the runtime is indeed rather simple. But the problem is the machine code generated by the OCaml compiler, as it cannot emit loading instructions for tls symbols (yet?).
POSIX says:
Only a single copy of an object file is brought into the address space, even if dlopen() is invoked multiple times in reference to the file, and even if different pathnames are used to reference the file.
So the answer is "inode". Copying the library file "should work", but hard links won't. Except. Since they will expose the same global symbols and when that happens all (portability) bets are off. You're in the middle of weakly defined behavior that has evolved through bug fixes rather than good design.
Don't dig deeper when you're in a hole. The approach to add additional horrible hacks to make a fundamentally broken library work just leads to additional breakage. Just spend a few hours to fix the library to not use globals instead of spending days to hack around dynamic linking (which will be unportable at best).

How do you include standard CUDA libraries to link with NVRTC code?

Specifically, my issue is that I have CUDA code that needs <curand_kernel.h> to run. This isn't included by default in NVRTC. Presumably then when creating the program context (i.e. the call to nvrtcCreateProgram), I have to send in the name of the file (curand_kernel.h) and also the source code of curand_kernel.h? I feel like I shouldn't have to do that.
It's hard to tell; I haven't managed to find an example from NVIDIA of someone needing standard CUDA files like this as a source, so I really don't understand what the syntax is. Some issues: curand_kernel.h also has includes... Do I have to do the same for each of these? I am not even sure the NVRTC compiler will even run correctly on curand_kernel.h, because there are some language features it doesn't support, aren't there?
Next: if you've sent in the source code of a header file to nvrtcCreateProgram, do I still have to #include it in the code to be executed / will it cause an error if I do so?
A link to example code that does this or something like it would be appreciated much more than a straightforward answer; I really haven't managed to find any.
You have to send the "filename" and the source of each header separately.
When the preprocessor does its thing, it'll use any #include filenames as a key to find the source for the header, based on the collection that you provide.
I suspect that, in this case, the compiler (driver) doesn't have file system access, so you have to give it the source in much the same way that you would for shader includes in OpenGL.
So:
Include your header's name when calling nvrtcCreateProgram. The compiler will, internally, generate the equivalent of a std::map<string,string> containing the source of each header indexed by the given name.
In your kernel source, use #include "foo.cuh" as usual.
The compiler will use foo.cuh as an index or key into its internal map (created when you called nvrtcCreateProgram), and will retrieve the header source from that collection
Compilation proceeds as normal.
One of the reasons that nvrtc provides only a "subset" of features is that the compiler plays in a somewhat sandboxed environment, without necessarily having all of the supporting tools and utilities lying around that you have with offline compilation. So, you have to manually handle a lot of the stuff that the normal nvcc + (gcc | MSVC| clang) combination provides.
A possible, but non-ideal, solution would be to preprocess the file that you need in your IDE, save the result and then #include that. However, I bet there is a better way to do that. if you just want curand, consider diving into the library and extracting the part you need (blech) or using another GPU-friendly rand implementation. On older CUDA versions, I just generated a big array of random floats on the host, uploaded it to the GPU, and sampled it in the kernels.
This related link may be helpful.
You do not need to load curand_kernel.h yourself and add it to the include "aliases" mechanism.
Instead, you can simply add the CUDA include directory to your (set of) include paths, e.g. by adding --include-path=/usr/local/cuda/include to your NVRTC compiler options.
(I do this in my GPU-kernel-runner test harness, by default, to be on the safe side.)

How does PC-Lint (by Gimpel) look across multiple modules?

I'm using Gimpel's PC-Lint v8.00 on a C codebase and am looking to understand how it traverses modules. The PC-lint manual only goes as far as to say that PC-Lint "looks across multiple modules". How does it do this? For example, does it start with one module and combine all related include files and source files into one large piece of code to analyze? How deep does it search in order to understand the program flow?
In a second related question, I have a use case where it is beneficial for me to lint one C module from the codebase at a time instead of providing every C module in a long list to PC-Lint. However, if I only provide one C module, will it automatically find the other C modules which it depends on, and use those to understand the program flow of the specified C module?
PC Lint creates some sort of run-time database when it parses your source files, noting things like global variables, extern-declarations, etc.
When it has processed all compilation units (C files with all included files, recursively), it does what a linker does to generate your output, but in stead of generating code, it reports on certain types of errors, for instance: An extern-declaration that has not been used, an unused prototype without implementation, unused global functions. These are issues not always reported by the linker, since the code generation is very well possible: The items have never been used anywhere!
The search depth can be influenced by the option -passes, which enables a far better value-tracking at the cost of execution time. Refer to seciton 10.2.2.4 in the PDF manual (for version 9.x).
To your second question, no, if you only provide one (or a few) source (C) file name(s) on your Lint command line, PC Lint will process only that file - and all include files used, recursively. You may want to use the option -u for "unit-checkout" to tell PC Lint that it only processes a part of a full project. Lint will then suppress certain kinds of warnings not useful for a partial project.
I think in principle you're asking about LINT OBJECT MODULES, see Chapter 9 of Lint Manual PDF.
Using say lint -u a1.c -oo procudes the a1.lob, when then again can be linked together using lint *.lob to produce the inter-module messages.
Also you asked a related, specific questions ( Any tips for speeding up static analysis tool PC-Lint? Any experiences using .LOB files?) but I'm not sure if I understand your concern with "How much would you say it affected linting time?", because I would say it depends. What is your current lint-time / speed? You posted some years ago now, how about running the job on a novel machine, new cpu then? KR

How to ensure unused symbols are not linked into the final executable?

First of all my apologies to those of you who would have followed my questions posted in the last few days. This might sound a little repetitive as I had been asking questions related to -ffunction-sections & -fdata-sections and this one is on the same line. Those questions and their answers didn't solve my problem, so I realized it is best for me to state the full problem here and let SO experts ponder about it. Sorry for not doing so earlier.
So, here goes my problem:
I build a set of static libraries which provide a lot of functionalities. These static libraries will be provided to many products. Not all products will use all of the functionalities provided by my libs. The problem is that the library sizes are quite big and the products want it to be reduced. The main goal is to reduce the final executable size and not the library size itself.
Now, I did some research and found out that, if there are 4 functions in a source file and only one function of that is used by the application, the linker will still include the rest of the 3 functions into the final executable as they all belong to the same object file. I further analyzed and found that -ffunction-sections, -fdata-sections and -gc-sections(this one is a linker option) will ensure only that one function gets linked.
But, these options for some reasons beyond my control cannot be used now.
Is there any other way in which I can ensure that the linker will link only the function which is strictly required and exclude all other functions even if they are in the same object file?
Are there any other ways of dealing with the problem?
Note: Reorganizing my code is almost ruled out as it is a legacy code and big.
I am dealing mainly with VxWorks & GCC here.
Thanks for any help!
Ultimately, the only way to ensure that only the functions you want are linked is to ensure that each source (object) file in the library only exports one function symbol - one (visible) function per file. Typically, there are some files which export several functions which are always all used together - the initialization and finalization functions for a package, for example. Also, there are often functions used by the exported function that do not need to be visible outside the source (object) file - make sure they are static.
If you looked at Plauger's "The Standard C Library", you'll find that every function is implemented in a separate file, even if the file ends up 4 lines long (one header, one function line, an open brace, one line of code, and a close brace).
Jay asked:
In the case of a big project, doesn't it become difficult to manage with so many files? Also, I don't find many open source projects following this model. OpenSSL is one example.
I didn't say it was widely used - it isn't. But it is the way to make sure that binaries are minimized. The compiler (linker) won't do the minimization for you - at least, I'm not aware of any that do. On a large project, you design the source files so that closely related functions that will normally all be used together are grouped in single source files. Functions that are only occasionally used should be placed in separate files. Ideally, the rarely used functions should each be in their own file; failing that, group small numbers of them into small (but non-minimal) files. That way, if one of the rarely used functions is used, you only get a limited amount of extra unused code linked.
As to number of files - yes, the technique espoused does mean a lot of files. You have to weigh the workload of managing (naming) lots of files against the benefit of minimal code size. Automatic build systems remove most of the pain; VCS systems handle lots of files.
Another alternative is to put the library code into a shared object - or dynamic link library (DLL). The programs then link with the shared object, which is loaded into memory just once and shared between programs using it. The (non-constant) data is replicated for each process. This reduces the size of the programs on disk, at the cost of fixups during the load process. However, you then don't need to worry about executable size; the executables do not include the shared objects. And you can update the library (if you're careful) without recompiling the main programs that use it. The reduced size of the executables is one reason shared libraries are popular.

Any good reason to #include source (*.c *.cpp) files?

i've been working for some time with an opensource library ("fast artificial neural network"). I'm using it's source in my static library. When i compile it however, i get hundreds of linker warnings which are probably caused by the fact that the library includes it's *.c files in other *.c files (as i'm only including some headers i need and i did not touch the code of the lib itself).
My question: Is there a good reason why the developers of the library used this approach, which is strongly discouraged? (Or at least i've been told all my life that this is bad and from my own experience i believe it IS bad). Or is it just bad design and there is no gain in this approach?
I'm aware of this related question but it does not answer my question. I'm looking for reasons that might justify this.
A bonus question: Is there a way how to fix this without touching the library code too much? I have a lot of work of my own and don't want to create more ;)
As far as I see (grep '#include .*\.c'), they only do this in doublefann.c, fixedfann.c, and floatfann.c, and each time include the reason:
/* Easy way to allow for build of multiple binaries */
This exact use of the preprocessor for simple copy-pasting is indeed the only valid use of including implementation (*.c) files, and relatively rare. (If you want to include some code for another reason, just give it a different name, like *.h or *.inc.) An alternative is to specify configuration in macros given to the compiler (e.g. -DFANN_DOUBLE, -DFANN_FIXED, or -DFANN_FLOAT), but they didn't use this method. (Each approach has drawbacks, so I'm not saying they're necessarily wrong, I'd have to look at that project in depth to determine that.)
They provide makefiles and MSVS projects which should already not link doublefann.o (from doublefann.c) with either fann.o (from fann.c) or fixedfann.o (from fixedfann.c) and so on, and either their files are screwed up or something similar has gone wrong.
Did you try to create a project from scratch (or use your existing project) and add all the files to it? If you did, what is happening is each implementation file is being compiled independently and the resulting object files contain conflicting definitions. This is the standard way to deal with implementation files and many tools assume it. The only possible solution is to fix the project settings to not link these together. (Okay, you could drastically change their source too, but that's not really a solution.)
While you're at it, if you continue without using their project settings, you can likely skip compiling fann.c, et. al. and possibly just removing those from the project is enough – then they won't be compiled and linked. You'll want to choose exactly one of double-/fixed-/floatfann to use, otherwise you'll get the same link errors. (I haven't looked at their instructions, but would not be surprised to see this summary explained a bit more in-depth there.)
Including C/C++ code leads to all the code being stuck together in one translation unit. With a good compiler, this can lead to a massive speed boost (as stuff can be inlined and function calls optimized away).
If actual code is going to be included like this, though, it should have static in most of its declarations, or it will cause the warnings you're seeing.
If you ever declare a single global variable or function in that .c file, it cannot be included in two places which both compile to the same binary, or the two definitions will collide. If it is included in even one place, it cannot also be compiled on its own while still being linked into the same binary as its user.
If the file is only included in one place, why not just make it a discrete compilation unit (and use its globals via extern declarations)? Why bother having it included at all?
If your C files declare no global variables or functions, they are header files and should be named as such.
Therefore, by exhaustive search, I can say that the only time you would ever potentially want to include C files is if the same C code is used in building multiple different binaries. And even there, you're increasing your compile time for no real gain.
This is assuming that functions which should be inlined are marked inline and that you have a decent compiler and linker.
I don't know of a quick way to fix this.
I don't know that library, but as you describe it, it is either bad practice or your understanding of how to use it is not good enough.
A C project that wants to be included by others should always provide well structured .h files for others and then the compiled library for linking. If it wants to include function definitions in header files it should either mark them as static (old fashioned) or as inline (possible since C99).
I haven't looked at the code, but it's possible that the .c or .cpp files being included actually contain code that works in a header. For example, a template or an inline function. If that is the case, then the warnings would be spurious.
I'm doing this at the moment at home because I'm a relative newcomer to C++ on Linux and don't want to get bogged down in difficulties with the linker. But I wouldn't recommend it for proper work.
(I also once had to include a header.dat into a C++ program, because Rational Rose didn't allow headers to be part of the issued software and we needed that particular source file on the running system (for arcane reasons).)

Resources