-fbounds-safety: Enforcing bounds safety for C
clang.llvm.org
clang.llvm.org
“-fbounds-safety is a language extension to enforce a strong bounds safety guarantee for C. Here is our original RFC.
We are thrilled to announce that the preview implementation of -fbounds-safety is publicly available at this fork of llvm-project. Please note that we are still actively working on incrementally open-sourcing this feature in the llvm.org/llvm-project . To date, we have landed only a small subset of our implementation, and the feature is not yet available for use there. However, the preview does contain the working feature. Here is a quick instruction on how to adopt it.”
“This fork” is https://github.com/swiftlang/llvm-project/tree/stable/202407..., Apple’s fork of LLVM. That branch is from a year ago.
I don’t know whether there’s a newer publicly available version.
There is a GSoC 2026 opportunity on upstreaming this into mainline LLVM (https://discourse.llvm.org/t/gsoc-2026-participating-in-upst...)
clang -g -Xclang -fbounds-safety program.c
stop reason = Bounds check failed: Dereferencing above boundsAt a certain point though, something's gotta give, the compiler can do guesswork, but it should do no more, if you have to add more metadata then so be it it's certainly less tedious than putting pragmas and _____ everywhere, some C code just looks like the writings of an insane person.
> At a certain point though, something's gotta give, the compiler can do guesswork, but it should do no more, if you have to add more metadata then so be it it's certainly less tedious than putting pragmas and _____ everywhere, some C code just looks like the writings of an insane person.
There is not even a single correct or factual statement in cited strings of words.
C optimisation is not «hacks» or «witchcraft»; it is built on decades of academic work and formal program analysis: optimisers use data-flow analysis over lattices and fixed points (abstract interpretation) and disciplined intermediate representations such as SSA, and there is academic work on proving that these transformations preserve semantics.
Modern C is also deliberately designed to permit optimisation under the as-if rule, with UB (undefined behaviour) and aliasing rules providing semantic latitude for aggressive transformations. The flip side is non-negotiable: compilers can't «guess» facts they can't prove, and many of the most valuable optimisations require guarantees about aliasing, alignment, loop independence, value ranges, and absence of UB that are often not derivable from arbitrary pointer-heavy C, especially under separate compilation.
That is why constructs such as «restrict», attributes and pragmas exist: they are not insanity, they are explicit semantic promises or cost-model steering that supply information the compiler otherwise must conservatively assume away.
«metadata instead» is the same trade-off in a different wrapper, unless you either trust it (changing the contract) or verify it (reintroducing the hard analysis problem).
Godbolt exists because these optimisations are systematic and comparable, not because optimisation is impossible.
Also, directives are not new, C-specific embarrassment: ALGOL-68 had «pragmats» (the direct ancestor of today’s «pragma» terminology), and PL/I had longstanding in-source compiler control directives, so this mechanism is decades older than and predates modern C tooling.
For people using Clang you can read more about libc++ hardening at https://libcxx.llvm.org/Hardening.html
OpenBSD maybe? or a fork of CheriBSD?
macOS clang has supported -fbounds-safety for a while, but I"m not sure how extensively it is used.
You first have to modify "all C code". It's not just a set and forget compiler flag.
Note that corresponding checks for C++ library containers can be enabled without modifying the source. Google measured some very small overhead (< 0.5% IIRC) so they turned it on in production. But I'd expect an OS distro to be mostly C.
And Android uses Clang for its Linux kernel.
-fbounds-safety is not yet available in upstream Clang though:
> NOTE: This is a design document and the feature is not available for users yet.
> The Linux kernel has always traditionally been compiled with GNU toolchains such as GCC and binutils. Ongoing work has allowed for Clang and LLVM utilities to be used as viable substitutes. Distributions such as Android, ChromeOS, OpenMandriva, and Chimera Linux use Clang built kernels. Google’s and Meta’s datacenter fleets also run kernels built with Clang.
https://docs.oracle.com/en/operating-systems/solaris/oracle-...
I still remember someone at Sun commented they treated warnings as errors. This is how software should be developed.
EDIT:
https://www.oracle.com/servers/sparc/
https://www.fujitsu.com/global/products/computing/servers/un...
Finally, it is up to Intel and AMD to come up with hardware memory tagging, so far they have messed up all attempts, with MPX being the last short lived one.
Some of us actually do read licenses before using products.
Also the FAANG are hardly any better only because they spew cool marketing stuff like do no evil.
One of these days the witch hunt against C will go away.
https://www.cisa.gov/sites/default/files/2023-12/CSAC_TAC_Re...
Surely your not suggesting that the Rust compiler never produces exploitable code?
If your attitude is that getting exploited doesn’t matter because your software is unprivileged, you need some part of your stack to be unexploitable. That’s a tall order if everything is C.
You can get exploitable code out of any compiler. But you’re far more likely to get it from real-world C than real-world Rust.
Kernel level process isolation is extremely robust.
> If your attitude is that getting exploited doesn’t matter because your software is unprivileged
It’s not that exploits doesn’t matter. It’s that process architecture is a stronger form of guarantee than anything provided by a language runtime.
I agree that the place where rust is most beneficial is for programs that must be privileged and that are likely to face attack - such as a web server.
But the idea that you can’t securely use a C program in your stack or that rust magically makes process isolation irrelevant is incorrect.
You have a process receiving untrusted, potentially malicious input from the outside. If there’s an exploit then an attacker can potentially take control of the process. Your process is isolated, that’s good. But it can still communicate with other parts of your system. It can make syscalls. Now you’re in the same situation where you have a program receiving untrusted, potentially malicious input from the outside, but now “the outside” is your subverted process, and “a program” is the kernel. The same factors that make your program difficult to secure from exploits if it’s written in C also apply to the kernel.
I’m not sure where those ideas as the end of your comment came from. I certainly didn’t say them.
Please learn more about this topic. You don't understand OS security models.
>Pizlix is LFS (Linux From Scratch) 12.2 with some added components, where userland is compiled with Fil-C. This means you get the most memory safe Linux-like OS currently available.
The author, @pizlonator, is active on HN.
Note also that it uses fil-c rather than clang with -fbounds-safety. I believe fil-c requires fewer code changes than -fbounds-safety.
I'm skeptical this is workable... it's pretty common in systems code to take the address of a local variable and pass it somewhere. Many event libraries implement waiting for an event that way: push a pointer to a futex on the stack to a global list, and block on it.
They address it explicitly later:
> Although simply modifying types of a local variable doesn’t normally impact the ABI, taking the address of such a modified type could create a pointer type that has an ABI mismatch
That breaks a lot of stuff.
The explicit annotations seem like they could have real value for libraries, especially since they can be ifdef'd away. But the general stack variable thing is going to break too much real world code.
The problem is that as long as it's something where the calling function checks it immediately after the function exits and never looks again (something like an error code or choosing a code path based on the result) they often get away with it, especially in single threaded code.
I'm running into this at this very moment as I'm trying to make my application run cleanly, but some of the libraries are chock full of this pattern. One big offender is the Unix port of Microsoft's ODBC library, at least the Postgre integration piece.
I also blame the Unix standard library for almost having this pattern but not quite. Functions that return some kind of internal state that the programmer is told not to touch. Later they had to add a bunch of _r variants that were thread safe. The standard library functions don't actually have this flaw due to how they define their variables, but from the outside it looks like they do. It makes beginning programmers think that is how the functions should work and write their code in a similar manner.
Sure, that's true, but nobody is suggesting returning the address of a local variable anywhere in this thread.
I'm describing putting a pointer to a local variable in a global data structure, which is safe so long as the function doing it is somehow guaranteed not to return until the pointer is removed from the global data structure.
So long as the thread is guaranteed not to exit while blocked, you know its stack, and therefore the object allocated on it, must continue to exist. So, as long as there is no way to wake the thread except by kicking that object, the memory backing it is guaranteed to continue to exist until that object is kicked. You do have to somehow serialize the global data structure lookup (e.g. lock/dequeue/unlock/kick), if multiple threads can find and kick the object concurrently that's unsafe (the thread might exit between the first and subsequent kicks).
Generally that's true, even in pthread userspace: while there are some obvious artificial counterexamples one can construct, real world code very rarely does things like that.
Boundary checking for array indexing is table stakes.
Trains shouldn't collide if the driver is correctly observing the signals, that's table stakes too. But rather than exclusively focussing on improving track to reduce derailments we also install train protection systems that automatically intervene when the driver does miss a signal. Cause that happens a lot more than a derailment. Even though "pay attention, see red signal? stop!" is conceptually super easy.
This brings C one tiny step closer to the state of the art, which is commendable, but I don't believe codebases which start using this will reduce their published vulnerability count significantly. Making use of this requires effort and diligence, and I believe most codebases that can expend such effort already have a pretty good security track record.
And in terms of how easy it is to convert a memory safety issue into an exploit, it’s not meaningfully much harder. The harder pieces are when sandboxing comes into play so that for example exploiting V8 doesn’t give you arbitrary broader access if the compromised process is itself sandboxed.
[1] https://projectzero.google/2022/04/the-more-you-know-more-yo...
Occasionally an out-of-bounds access pops up, but they're generally so blindingly obvious and easy to fix that it's never been the slow part of bug fixing.
My last memory error in C code in production was in 2018. Prior to that it I had a memory error in C code in production in 2007 or 2008.
In C++, I eventually gave up trying to ship the same level of quality and left the language altogether.
As for why your experience may be different, my hunch is that either your code was super simple OR you didn’t test it thoroughly enough against malicious/unexpected inputs OR you never connected the code to untrusted I/O.
Keep in mind the data for this comes from popular projects that have enough attention to warrant active exploit research by a wide population. This is different from a project you wrote that doesn’t have the same level of attention.
You are misremembering the various reports - the reports were not that 80%[1] of issues were due to memory errors, but more along the lines of 80% of exploits were due to memory errors.
You could have 1000 bugs, with 10 of them being vulnerabilities, and 8 of those 10 being due to memory errors, and that would still be in line with the reports.
> As for why your experience may be different, my hunch is that either your code was super simple OR you didn’t test it thoroughly enough against malicious/unexpected inputs OR you never connected the code to untrusted I/O.
Payments processing, telecoms and munitions control software.
Of those, your explanation only applies to Telecoms; payments processing (EMV) was basically a constant stream of daily attacks, while munitions are live, in the field, with real explosives. We would've noticed any bugs, not just memory error bugs with the munitions one.
--------------------
[1] The number wasn't 80% IIRC, more like 70%?
For operating systems kernels, browsers etc, vulnerabilities have a much, much bigger impact than logic errors: vulnerabilities need to be fixed immediately, and released immediately. Most logic errors don't need to be fixed immediately (sure, it depends on the issue, and on the type of software.)
I would probably say "for memory unsafe languages, 80% of the _impact_ is due to memory vulnerabilities"
For EMV you had C connected directly to the network under a steady stream of attacks and only had an issue once? I find that hard to believe. What’s more likely is a Java websever frontend talking to some C processing / crypto code in which case again you’re less likely to encounter bugs in your code because it’s difficult to find a path to injecting unsanitized input.
For munitions there’s not generally I/O with uncontrolled input so it’s less likely you’d find cases where you didn’t properly sanitize inputs and relied on an untrusted length to access a buffer. As a famous quote states, it’s ok if your code has an uptime of 3 minutes until the first bug if the bomb explodes in 2
EMV terminals. No Java involved.
> As a famous quote states, it’s ok if your code has an uptime of 3 minutes until the first bug if the bomb explodes in 2
Look, first you commented that it's not possible for nontrivial or non-networked devices, now you're trivialising code that, if wrong, directly killed people!
All through the 80s, 90s and 2000s (and even now, believe it or not), the world was filled with millions and millions of devices programmed in C, and yet you did not live a life where all the devices around you routinely crashed.
Crs, Microwaves, security systems... they didn't routinely crash even though they were written in C.
I’m not trivializing the safety of munitions. I’m attempting to highlight that safety and stability in a munitions context is very different and memory safety issues could easily exist without you realizing. My overall point is that you are silently making the argument that C programmers (or programmers in general) used to be better, which is a wild argument to be making about a culture in which fuzzing didn’t even exist as a concept. You’re also confusing memory safety with implying a crash. That simply isn’t the case - it’s more often exploitable as a security vulnerability than an immediate crash by violating assumptions made that weren’t in the happy path of those microwaves and security systems. That millions of devices were and still are routinely exploitable.
You’re also making a fallacious line of reasoning that the C today is the same C that was in use in the 80s, 90s, and 2000s. It’s not and has gotten harder and more dangerous because a) multi threading became more of a thing and b) compiler authors started exploiting “undefined” behaviors for extra optimization.
It’s just wild for me to encounter someone who believes C is a safe language and is suitable to connect to I/O too when there’s so many anecdotal and wide statistical evidence gathered that that’s not the case. Even SQLite, the darling of the C community, is not safe if asked to open arbitrary SQLite files - there’s various security attacks known and possible.
They are under daily attack - in public, at tills, operated by minimum-wage earners.
> You probably had loads of vulnerabilities.
Sure. Hundreds of thousands of terminals sitting in the field, networked, under the control of minimum wage employees, each holding credit card details for hundreds of cards at a time...
Yeah, you're right, not a target at all!
> But also depending on when you did it, all you had to process was a bar code which is also isn’t some super complicated task.
You are hopelessly naive. Even in the magstripe era, certification was not easy.
> It’s just wild for me to encounter someone who believes C is a safe language
When did you meet this person?
Look, the bottom line is, the errors due to memory safety in programs written in C is so small it's a rounding error. It's not even statistical noise. You spent your life surrounded by these programs that, if they went wrong, would kill you, and yet here you are, not only arguing from a place of ignorance, you are reveling in it.
Just out of interest, have you ever used an LLM to write code for you?
As for certification and it being difficult, what does that have to do with the process of bread in Paris? Unless you’re somehow equating certification with a stamp of vulnerability imperviousness in which case you’re seeing your own naivete instead of in others. Btw, Target was fully certified and fully had their payment system breached. Not through the terminals but through the PoS backend. And as for “but you’re here living and breathing”, there’s constant security breaches through whatever hole, memory safety or otherwise. Persistent access into the network is generally only obtainable through credential compromise or memory safety.
> When did you meet this person?
You. You’re here claiming that memory safety issues are statistical noise yet every cloud software I’ve seen deployed regularly had them in the field, sometimes even letting a bad one through to canary. And memory safety issues persisted despite repeated attempts to fix issues and you couldn’t even know if it was legitimately an issue or just a HW flaw due to being deployed at scale enough that you were observing bad components. It’s a real problem and claiming it’s statistical noise ignores the consequences of even one such issue being easily accessible.
Claiming that the exploit rate percentage is statistical noise is different from claiming that it's a safe language.
Looks like you have a premade argument to argue.
You haven't answered my question, though: Have you used LLMs to generate any code for yourself?
Bounds checking with fat pointers existed as a set of patches for GCC in the early 2000's. (C front end only).
Would be nice if the annotations could also be applied to structure fields.
struct bytes {
size_t count;
unsigned char * __counted_by(count) pointer;
};
void work_with(struct bytes);It's really cool that the kernel is using this. The compiler must be generating simple bounds checking code with traps instead of crazy stuff involving magical C standard library functions. Perfect for freestanding nostdlib projects.
template <typename T>
struct Slice {
T* data = nullptr;
size_t size = nullptr;
T& operator[](size_t index) {
if (index >= size) crash_the_program();
return data[index];
}
};
Your comment started out with "just."
As if there are never any compelling reasons to want to make existing C code better.
But instead of taking that as an opportunity to reflect on when various tools might be appropriate,
> as I'm advocating to use the more advanced feature-rich tool over the compiler-specific-hacks one.
You've simply doubled down.
However, you still need something like -fbounds-safety in C++, due to the copy-paste compatibility with C, and too many people writing Orthodox C++, C with Classes, Better C, kind of code, that we cannot get rid of.
I find it a bit hard to justify using the STL when a single <unordered_map> include costs 250ms compile time per compile unit.
The fact that I don't have to step through this in the debugger is also a bonus:
template <size_t _Offset, size_t _Count = dynamic_extent>
[[nodiscard]] _LIBCPP_HIDE_FROM_ABI constexpr auto subspan() const noexcept
-> span<element_type, _Count != dynamic_extent ? _Count : _Extent - _Offset> {
static_assert(_Offset <= _Extent, "span<T, N>::subspan<Offset, Count>(): Offset out of range");
static_assert(_Count == dynamic_extent || _Count <= _Extent - _Offset,
"span<T, N>::subspan<Offset, Count>(): Offset + Count out of range");
using _ReturnType = span<element_type, _Count != dynamic_extent ? _Count : _Extent - _Offset>;
return _ReturnType{data() + _Offset, _Count == dynamic_extent ? size() - _Offset : _Count};
}As someone that enjoys C++ since 1993, alongside other ecosystems, many pain points on using C++ complaints are self inflicted, by avoiding using modern tools.
Heck, C++ had nice .NET and Java alike frameworks, with bounds checking even, before those two systems came to exist, and nowadays all those frameworks are mostly gone with exception of Qt and C++ Builder ones, due to bias.
#define span(T) struct span_##T { size_t len; T *data; }
#define span_access(T, x, i) (*({ \
span(T) *_v = (x); \
auto _i = (i); \
if (((size_t)_i) >= _v->len) abort(); \
&_v->data[_i]; \
}))
I don't use much C but if you add them to the standard they'll probably trickle down to C++ compilers by 2045 and I'll have a good 10 years to use them before I retire.
But in all seriousness, I want this:
struct Result { int value; int err; };
#define TRY(res) \
({ \
Result _res = res; \
if (failed(res)) return res; \
res.value; \
})
Result f1(...) { ... }
Result f2() {
int res = TRY(f1(...)); // <<<<
...
return success();
}
size_t size = nullptr;
The incremental path matters more than the theoretical coverage. I'd love to see benchmarks on a real project — how many annotations per KLOC, and what % of OOB bugs it actually catches in practice vs. what ASAN already finds in CI.