Yes, this is yet another post in the internet talking about using exceptions versus error returns. The topic has been flaming up at my workplace for quite some time now, and I felt that writing a blog post about it during the week-end would help me focus my thoughts and give me time to explain my point with the due care. In case you didn't know, I'm against using exceptions for error handling (maybe having spent many years working with Qt has had an effect on this); that does not mean that I never write code using exceptions: I certainly do my good share of try ... catch when dealing with third-party code (including the STL), but you won't find a throw in my programs.

I'm not going to write here all the reasons why I refrain myself from implementing error handling using exceptions; I'd rather like to focus on the one I consider to be the major one, and which I rarely see being given the due weight in the debate.

And please note that this post is about C++ only; it may be that exception handling in other languages is designed in such a way that all my concerns are addressed (either by the language itself, or by common error handling policies).

Code safety

I was about to title this “Code readability”, but this is more about code verifiability, that is making sure that the code is correct and, ultimately, safe. As we all know, code is written once but read many times, and even if it's code you've written yourself, chances are that in a few weeks time you'll have forgotten several details about it; error cases and error handling are one typical thing that doesn't stick in our memory for long.

When I look at a small piece of code, such as the one that can fit into my screen, or which I can read from a merge request diff, I want to be able to ascertain that the code I'm looking at is correct. Let's look at some examples.

A throw-free project

I just made this up, so please bear with me if it doesn't make any sense. What I want to show is that code like the above has very few fault risks, if found in a project which bans throwing errors as exceptions: if we exclude out-of-memory errors, that are generally not handled to let the application crash (though you can always catch them if you like), the reader can easily verify that this code is safe. Coding style policies and naming conventions can guarantee that setMaximumSpeed() and setName() won't have a return value that needs to be checked, and all other method calls either return an error that our code is properly handling, or return some value. Of course, by just looking at this piece of code we cannot know if the engineTemperature() method has some other overloaded sibling which accepts passing a reference to a boolean and which could be used to detect an error; so, it may be that our code could be improved in that respect, if we had a look at the header files for the Car class — but this does deny the fact that a simple glance at this snippet tells us exactly what errors are handled and what could be going wrong.

Let's look at this code instead:

If we continue on the assumption that we are working on a project which bans throwing exceptions, we can immediately say that this code is not safe: we don't know if the car successfully executed a lap on the track, and our process will crash if boundingRectArea() is invoked on a null object.

Enter the exception

In a project where exceptions are actively used, the code from the second snippet is not obviously wrong anymore: maybe executeLap() cannot throw any exceptions, or, if does, the caller of this snippet is catching the exception? In order to figure out whether this code is correct, I need to see the declaration of the executeLap() method, and hope that there's a nice noexcept in there; if there isn't, I have to look at its implementation, and recursively descend through all the methods it calls — at which point the safest attitude is just to assume that it can throw. But that's only half of the story, because once I accept the fact that executeLap() can throw, I need to check whether the exception is properly handled: I have to check the implementation of all the callers of my method, and if I don't find a catch there, I'll have to recursively walk up the tree of their callers.

And indeed even the first snippet, which looked so harmless when exception throwing was banned, suddenly becomes not obviously correct anymore: what if executeLap() or getPath() also throw an exception? You might say that it would be quite a silly thing to do, and I'd certainly agree; but it may be that indeed they don't throw any exceptions in their implementation, but some of the methods they call does.

A compromise: catch early, catch often

The obvious solution to the above issue is having a policy of handling exceptions right away, and explicitly rethrowing them (or even better, rethrow a different, more appropriate exception) up the stack:

What I can tell from the above snippet is that the code is handling errors, and this is somehow a relief. I'm sure some of you would suggest using a more specific catch clause, but for the sake of this example let's assume that this one is fine.

(Quick note: the above example does not catch std::exception, because that would also catch the std::bad_alloc exception which is typically thrown in out-of-memory situations; my advice is not handle it at all, unless you know what you are doing)

In real life, though, you might find that try-ing on a rather large block of operations is not enough: suppose that the Car methods all emit the same exception type, and that you need to handle them differently depending on when they occur. Then you'd need to split up the try into smaller blocks, and at that point your code won't look any cleaner than the equivalent code which uses ifs on return values. Of course if you own the Car class you could modify it to throw different exceptions, in order to keep more operations inside the try block and have specific catches at the end.

The big catch (pun intended)

Even once you've refactored your methods to get the best out of exceptions (where "best" is highly subjective, but let's assume that it just means that you are happy with your exception-throwing code), there's something that still bothers me, and that's exactly the same thing that proponents of exceptions use as a “pro” in their argumentations: the business logic of your code gets separated from the error handling. You get a nice block of pure logic, not cluttered with error checking, and a catch section (which I call “the big catch”) where error cases are handled.

I really don't see how that makes the code any more readable or safe: sure, the logic is not intertwined with error handling and might help focus on the expected flow of the operations (though, really, I do not think that normal brains have a problem skipping over if blocks), but that's hardly what I'm interested in when I want to check that the code is correct. Most of program errors and bugs lie in handling the edge cases and the abnormal situations, the seldomly taken code paths, and that's where I need to focus my attention.

When I see code like this one, I need to mentally build a mapping of “operationX() → possible exceptions” (which, unless exception naming is making this obvious, requires me to look at the implementation of the operationX() functions), and then mentally reconstruct the possible code paths in case operationX() fails, for each line of the try block.

Not seeing the errors right there, right away makes the correctness verification harder, which in turns means that the code becomes less safe. It will make you focus on the best case scenario, while ignoring all those annoying edge cases — too bad that 90% of the bugs are there.

Reading through the ISO C++ propaganda FAQ

I've been given a link to the C++ FAQ about exceptions, and unfortunately I read it. While there isn't much to argue on the technical side of it, it also carries some misleading statements, which might be true in absolute terms but don't let you see the big picture by not mentioning all that you need to know (which is the fundamental technique behind propaganda). An example is when they mention that eliminating ifs makes for more robust code, without mentioning that the same applies to all code branches, including exceptions.

Another argument that bothered me when I read it is the one about error propagation; this is the example they make:

The claim is that this code is more readable than the one with explicit error handling, because all the f2(), f3, …, f9() functions don't have to handle the error occurring in f10(). It is indeed a convincing argument, when presented in these terms, but is this really how our code looks like? In real life, you'll hardly have a chain of 1-liner functions, all defined next to each other in the same file. The moment that you realize that each one of these fn() functions might be twenty or thirty lines long, and that they might be scattered over different files, and be called not just by fn-1() but by any other function in the codebase, the picture does not look so rosy anymore: we get back to my main point of pain, that is that looking at the code of, say, f5(), I will not be able to tell if the errors thrown by it, or by any of the methods invoked by it, are properly handled.

Exceptions in APIs

A side note about projects using exceptions. I'm not really bothered when a library I need to use is throwing exceptions: having to write

is not less readable or less safe than the code I'd write if Foo::fetch() returned an error code. I still do have a little complaint, because the library author has given himself the right to decide that a failure in his library should be considered a critical fault, whereas it may be that in my program it is an expected failure and using exceptions imposes a penalty which could have been avoided. But I digress.

As long as the library documents which exceptions are thrown, it is used by many people (which hopefully means that it has few bugs) and it is a library that I don't need to contribute to, wrapping some of its methods in try blocks is something I can live with.

One situation where I actually wish that libraries threw an exception is in out-of-memory situations; in that case, of course, I'd expect them to throw nothing else than std::bad_alloc, which is the exception emitted by the standard library in such situations. That allows the caller to decide whether to ignore the exception and have the process terminated (which is what I usually do, at least in desktop applications) or try their luck and handle the failure — the latter is not easy, but it can certainly be done.

This is one case where error returns can be problematic, because it's likely that your code would look something like

and in this case there's actually a risk that your code is going to trigger an out-of-memory error in logging the message; this shouldn't be a concern in most cases, but I can imagine some situations where one might want to know which was the exact operation that first incurred in the out-of-memory failure.

So, I'm actually fine with new throwing. As for my code, my throw statement is actually spelt as return.

Mappero Geotagger has now moved from its previous page from this site to a new, separate website built with the awesome Nikola static website generator.

The main reason for this change is that I didn't have an online space where to host the application binaries, and I wanted to experiment with a different selling method. Now, downloads are (poorly) hidden behind a payment page, whereas in multiple places of the website I also mention that I can provide the application for free to whomever asks for it. While it might seem weird at first, I do honestly believe that this will not stop people from buying it: first of all, many people just think it's fair to pay for a software applications, and secondly, for some people writing an e-mail and establishing a personal contact with a stranger is actually harder than paying a small amount of money. And in all sincerity, the majority of the income I've had so far for Mappero Geotagger came from donations, rather than purchases; so, not much to lose here.

QBS and MXE

Anyway, since this is primarily a technical blog, I want to share my experiences with cross-building from Linux to Windows. As you might remember, some time ago I switched the build system of Mappero from qmake to QBS, and I haven't regretted it at all. I've managed to build the application in Linux (of course), macOS, as a Debian package on the Ubuntu PPA builders, on Windows with AppVeyor and, last but not least, on Linux for Windows using the mingw setup provided by the MXE project.

QBS worked surprisingly well also in this case, though I had to fight with a small bug on the toolchain detection, which is hopefully going to be fixed soon. For the few of you who are interested in achieving something similar, here's the steps I ran to configure QBS for mingw:

Sorry for using that many environment variables ☺. After qbs is configured, it's just a matter of running

to build the application. You will get a nice window binary and, once you collect all the needed library dependencies, you'll be able to run it on Windows. Or WINE ☺.

As part of this effort, I also had to build libraw, so I didn't miss the occasion to contribute its recipe to MXE. I'm also trying to get a change accepted, that would make MXE support the dynamic OpenGL selection available since Qt 5.4.

Today I've released Imaginario 0.9. The big feature coming with this new release is a face tagging flow which I believe will be the fastest and simplest you've ever used, despite it being all manual. I even sat down and spent some quality time with Blender to prepare a video to show it off:

While some people might actually think that I spent more time for making the video than for implementing the face tagging feature itself, this couldn't be farther from the truth: the face tagging branch has been being worked on for at least three months (of course, that's my spare time — so it's actually less than one hour per day) and consisted of more than 40 commits (after squashing all the fixups), whereas for the video I spent no more than a couple of hours.

I would appreciate if the curious could go and try it out, and let me know about any issues you should find: there are built packages for Linux (AppImage), macOS and Windows. I do also have an Ubuntu PPA where nightly images are built, but I'm not sure if I can recommend that one, since I've not been using it myself and have no idea whether those packages actually even start. But you are welcome to try :-)

Your feedback will help me do better, so please don't be shy!

I've recently found some use for a first-generation Raspberry Pi (Pi 1 model B) which had been lying in a drawer since many years. A few days ago I've installed the Raspbian distribution in it, and was about to install motion on it, but I stopped as soon as I noticed that apt was suggesting to bring in 1 GB worth of dependencies. Adding --no-install-recommends reduced the proposal a bit, but it was still around 700 MB -- a bit too much for my taste. I figured out that the motion package for Debian (and Raspbian) depends on MySQL, PostgreSQL, FFmpeg and what not; so, I decided that I could probably just recompile it and disable all the stuff I didn't need at configure time.

But I didn't want to install all the build dependencies and the cross-compiler in my machine; containers exist for a reason, after all. So I had a look at the crossbuilder tool that we use in UBports: this is a nice little shell program that uses an Ubuntu-based LXD image to cross-compile a package after automatically fetching all its dependencies, and installs it into an UBports device. It does some magic with adb and stuff, but I thought that the basic functionality should work with minor modifications on any Debian-based distribution.

And indeed, some hours later, I got a branch where I can use crossbuilder to build packages for the Raspberry Pi. Assuming that you have LXD properly setup, the command

crossbuilder --raspbian source motion

will cause crossbuilder to create a container and download the Debian source package for motion; at this point you can modify the source code as you see fit, and rebuild it. I only changed the debian/rules file to add a few flags such as --without-mysql, --without-ffmpeg, etc. And

crossbuilder --raspbian

is the command to run in order to perform the build. This will first download all the dependencies (according to the debian/control file), build the package, and create a tar archive containing all the generated .deb files. This archive can then be copied into the target device and unpacked there.

Now, there's a small problem in that Raspbian claims to be armhf, while in fact its floating-point processor is somehow not compliant with the armhf architecture. So, you generally cannot use an armhf package from Debian or Ubuntu on the Raspberry Pi. Given that I didn't have the time to prepare a proper Raspbian image for LXD, I used the Debian distribution as a base instead, and I chose to target the armel architecture: this might impose some penalties on the performance (at least for floating-point code), but it seems to work fine on the Raspberry Pi. Unfortunately, this means that you cannot just install the generated packages o the Pi, as dpkg will complain about the architecture mismatch (your package is armel, while the distro claims to be armhf). But unpacking the debian package with

dpkg -x motion_<...>.deb tmp

will indeed give you package that you can use.

I will eventually get back to this and make the script work better with Raspbian, but this is a quick start.