“They’ve done it again”, he exclaimed.
“The Committee has made a draft of what will become the next iteration of C++, and this time again they crammed a bunch of new language features but did not make my code faster in debug!”, he continued.
As the outrage reached its peak, he dared to ask what everybody had been secretly wondering all along: “What are those guys even paid for?!”
Every once in a while, a similar drama flares up on social media. I could go at lengths to explain that nobody pays people to attend ISO meetings except (sometimes) their own employers, or that WG21 doesn’t write compilers but this is not the article for it. For that you can read my previous entry.
Instead today I will argue about semantics.
The semantics of Debug
The recent episode reminded me that this is not the first time “C++ performance in debug” is brought up, and that again I wasn’t understanding the need. I almost never run a build of my game in what I’d consider “Debug”. The speed trade off is just not worth what it brings up for my day to day work. I only run unit tests in debug.
But let’s take a step back.
In his great talk, What Do You Mean?, Kevlin Henney made a good observation about the fact that semantics are quite important when you are trying to make a point, else your discussion isn’t gonna go far.
So when we say “Debug build”, what do we mean? As a quick experiment, I ran a poll on Twitter:
In C++, Debug means... — Mathieu Ropert (@MatRopert) July 30, 2019
As you can see from the 287 answers, there is some majority vote for debug symbols, but that’s not exactly what I’d call a consensus. Worse, out of all the options I offered, this is by far the one that has the least impact on runtime performance compared to full “Release” build!
There was also trick in my question: those 4 options are (almost) entirely orthogonal. You can mix and match them as you like while turning the others off. A better poll would have allowed for multiple answers but Twitter is quite limited there, so let’s not try to draw too many conclusions from this.
Four Axis of Debug
As conveniently laid out in my poll, a “classic” C++ build profile will offer (at least) 4 axis to play with on the debug spectrum:
/Odand the like control optimization. When disabled, the generated code will strongly resemble the C++ source code. No inlining will be performed, no instruction reordering, loop unrolling, variable elimination, nothing. This will have a huge impact on performance, as we usually write our code for humans first, and machines second.
/Ziand friends decide whether or not the compiler emits debug information. The process of compilation being very lossy, the debugger needs extra data to be able to link back machine code to source code when debugging. That database can be stored outside the actual binary, and has no impact on performance by itself.
/MDdwill switch the C++ runtime to a special debug version that contains extra checks. On Windows, the debug runtime will help catching heap corruptions and bad iterator accesses. It comes with a performance cost that is not nearly as bad as it looks if optimizations are enabled and the standard algorithms are used. Contrary to the two previous options it has ABI implications and should be turned on for all the objects in the binary (including dynamic libraries).
-UNDEBUG(or the absence of
NDEBUGin defines) ensures asserts are built in. The
assert()macro function is required to only produce runtime checks if
NDEBUGisn’t defined. The impact will greatly depend on the codebase usage of it. A badly placed assert in a tight loop can kill performance for good, whereas a sanity check on the inputs of a database API will probably not have a noticeable effect.
As previously mentioned, those 4 options can be mixed and matched depending on what’s needed and which tradeoffs are acceptable.
It’s important I think to notice that the most sought after feature, debug symbols, have no impact on performance and can even be stored outside the program for added isolation. For example it is customary for libraries on Windows to ship
.pdb files alongside
.dll files so that developers can get some modicum of information even in release mode.
I am genuinely curious as to why it came up as the most strongly associated with Debug while it is in fact probably the least tied to the “performance in debug” debate. A small suspicion I have is that it could be due to CMake defaults, in which the
Release profile doesn’t include debug symbols. In which case let me tell you, dear reader, that CMake defaults are wrong and you should override them. For example, if you create a project using MSVC’s wizard, it will only generate 2 profiles,
Release, both of which produce debug symbols.
Debug runtimes and ABI fun
The second one, the debug runtime (
/MTd) is a bit more complex. In my opinion its biggest impact is on compile time. Since it’s potentially an ABI breaker, all of the program (including 3rd parties) should be recompiled using that flag. In a classic workflow where the programmer suddenly realizes he needs the debug runtime to track a nasty bug, the thought of needing to recompile everything can be quite disheartening. And then, there’s the runtime performance cost of checked iterators.
The compilation issue can be partially mitigated by a good build system and package manager. Simply put, there is usually a lot of code in a project that doesn’t need recompiling often and could be cached. Third parties are the obvious example, but low level libraries, frameworks and other engines could also probably be built every night and shared across developers. Readers who want to know more about the state of package management in C++ can go and watch this guy talking about it at ACCU.
As for the performance impact, I ran my own benchmarks and discovered that it is indeed utterly terrible in tight loops… when used with custom algorithms. As it turns out, the optimizer has a hard time lifting iterator validity checks outside of the inner loop, leading to extremely redundant checks. This is why MSVC’s STL code has iterator debugging aware code in
<numeric> to keep performance decent.
A brief glance Visual C++ 2017’s implementation of
std::accumulate() for example will reveal that an initial check is made on the validity of
last parameters, and then unwrapped iterators are used in the tight
while() loop to ensure vectorization can still happen. C++11 range-for loops also seem to be optimized properly, meaning the big thing to keep an eye about is hand-made specific code that runs tight loops using STL iterators.
Help me, /Ob1 Kenobi
Disabling optimizations in C++ is, of course, a serious tradeoff. Sure, no inlining or vectorizing makes the step through code process easier when debugging, but it comes at the price of most of the zero cost abstraction principle.
Fortunately, compiler vendors do not consider optimization an “all or nothing” deal and offer ways for the user to determine which tradeoff they are willing to accept. Clang and GCC, for instance, have the
-Og flag to enable all optimizations that don’t interfere with debugging. MSVC doesn’t have the equivalent but the compiler has an “optimized debug” feature (
/Zo) that is on by default and include extra debug info to track local variables and inline functions. As MSVC users might tell, results may vary…
Another option for MSVC is to disable optimization (
/Od) but keep a degree of inlining so that the STL stays reasonably fast:
/Ob1. What it does is allow inlining of functions declared
inline or using the
__forceinline macro. Incidentally the STL is tagged accordingly so that most of it can still be inlined. CMake’s default
RelWithDebInfo build profile keeps the optimization flag (
/O2) but lowers down the inlining from
/Ob2 (inline all the things) to
It’s important to note that optimization flags do not break ABI, meaning it’s perfectly fine to build 3rd parties and other “stable” code with full optimizations, while keeping the user’s project with less aggressive optimization settings. Again here a package manager would help, as it should allow the user to use a release toolchain for dependencies while keeping with a debug toolchain for the local project (as long as they only differ through optimization flags).
Another option would be to make external code inlined, while user’s code isn’t. A simple way would be similar to the
dllexport preprocessor trick: a macro that expands to either
__forceinline if no particular define is set (used from the outside) or to nothing when that define is set (when that library is being built for debug). This way the
/Ob1 flag would make the compiler inline code from external sources only.
A more reliable (and less intrusive) solution would require compiler support for an inlining level in which code found through
-isystem inclusion (or module
import) would be candidate for inlining, but code found through
-I search path wouldn’t. One would then just need to tell CMake to translate
INCLUDE_DIRECTORIES properties to
INTERFACE_INCLUDE_DIRECTORIES would be pulled though
-isystem (or equivalent). A package manager could also ensure that dependencies include directories are passed through
-isystem. Both solutions should be discussed with compiler and build systems vendors and not through WG21 as inlining isn’t defined in the C++ Standard (well the keyword is, but only in the context of linkage).
Oh and I’d like to quickly mention
-fomit-frame-pointer, the dreaded optimization that breaks debuggers. For one it doesn’t, as the C++ standard requires the runtime to be able to unwind local variables if an exception is thrown, meaning the debugger will also be able to find them. Secondly, omitting the frame pointer is a dubious optimization anyway as the x86 and x86_64 define
RBP as non-volatile, meaning a function that uses it will have to restore it so it only really saves a
push and a
pop in select cases, for total of probably less than 10 x86 cycles (as the stack is probably in L1 cache anyway). Benchmarking is advised before adding this flag to a project build profile.
assert(fibonacci(10000) == 3e10^2089))
assert() macro is disabled by setting the
NDEBUG preprocessor define. The standard requires it. What it doesn’t require is for optimizations to disable it.
Contrary to popular belief, passing
/O2 does not imply
/DNDEBUG. The two are orthogonal, but build systems usually put one with the other when generating pre-defined profiles. For example CMake defines
NDEBUG in both
Which means, again, that one can perfectly enable asserts in a build profile other than
Debug. I’m not advocating for enabling asserts in release (I don’t think anyone does?) but there’s no reason why a tailor-made “development” profile couldn’t have a modicum of optimization while keeping asserts (and debug symbols) on.
When it comes to usage, the only things to really watch out for are dubious asserts that have side effects, heavy computations or that are run in tight critical loops. In my experience the reason they may happen is precisely because so many default profiles disable them, leading to them been rarely seen.
As we have seen, build profiles are not a boolean that can be either
Release. Compiler vendors know that this is a complex problem that is unlikely to have one solution and offer a myriad a settings to play with.
For the same reason the C++ Standard will probably never specify what
Debug means or how fast it should run.
And as my regular readers probably know by now, there are many things we complain about regularly that end up being a build system issue more than anything. This one is just another in the list. There is a much to be done by figuring which build profile is good for one’s use case, formalizing it as a toolchain file, and making sure one’s build can correctly use it. So go out there and clean those