Hacker News Reader

Comment by rectang | Hacker News Reader

> But I also think a lot of discussions of this topic caricaturize compiler writers to a ridiculous degree.

The inflamed backlash should tell you just how damaging it is to impose silent failure on meticulously written, previously fine programs.

View on HN

> meticulously written, previously fine programs

With relatively few exceptions, if your program hits undefined behavior, then your program was already doing something pretty wrong to begin with. Signed overflow is a poignant example: in how many contexts is INT_MAX + 1 overflowing to INT_MIN actually sane semantics? Unless you're immediately attempting to check the result to see if it overflowed (which is extremely rare in the code I see), this overflow is almost certain to be unexpected, and a program which would have permitted this was not "previously fine" nor "meticulously written."

I feel compelled right now to point out that software development is a field where it is routine to tell users that it's their fault for expecting our products to work (that's what the big all-caps block of every software license and EULA says, translated into simple English).

Triggering signed overflow and testing whether it has occurred was how I protected the Delphi RTL memory allocation routines from signed overflow attacks, attacks which if not prevented, lead fairly directly to buffer overruns.

The code was written in Pascal and assembly, though, so it was safe from a C compiler.

I am quite thankfull that I learned systems programing via BASIC, Pascal and Assembly before coming to C.

As such my vision of low level coding isn't tainted by the ways of C.

> in how many contexts is INT_MAX + 1 overflowing to INT_MIN actually sane semantics? Unless you're immediately attempting to check the result to see if it overflowed (which is extremely rare in the code I see)

It's not that rare - I know that postgres got bit by particularly that issue, and several other other projects as well. Particularly painful because that obviously can cause security issues.

Isn't checking for overflow depending on UB? i.e. in

    int a;
    // lots of code
    int b = a + 1;
    // check for overflow
    if (b <= a)
        abort();

the compiler is allowed to remove the check because in the absence of UB a + 1 > a, therefore the conditional is always false.

As brilliantly said further in this thread by someone else, maybe check that something is OK before doing it.

In that case, something like: [...] if (a > INT_MAX - 1) abort(); int b = a + 1; [...]

> As brilliantly said further in this thread by someone else, maybe check that something is OK before doing it.

Much easier said than done:

https://github.com/postgres/postgres/blob/master/src/include...

Good point that checking for overflow of a multiplication beforehand is hard. But I don't think checking after the fact would be much easier. So thus a C modification to remove the undefined behavior probably wouldn't be very useful still.

The thing is that it's necessary, and (x86, not sure about arm) asm tells you via the d register if not the overflow flag. So why doesn't the C standard arguably provide a misuse-resistant way of doing so? Instead every project having to reinvent it and get it wrong if they do the "obvious" (but wrong) thing.

What would the misuse-resistant way be? Would you want the multiplication operator (*) to return 2 values: the number and a bool telling you whether it overflowed?

Yeah, a standard library function that does this would be good. But many people would just use * instead of this function, and so the problem would partially remain.

mul_would_overflow(), mul_wrap(), mul_saturate(), mul_do_whatever_the_hardware_does_just_please_dont_break_my_code_by_assuming_ub. Can't help people who just use * and don't care about overflow without breaking compatibility, but at least the people who do care won't have to reinvent safe arithmetic.

Yeah, I agree with you. But it would be hard due to there being so many integer types. A binary operation needs to consider 3 types: the types of each argument and the type of the result.

In c++ it would be a little simpler due to templates, so the types of the arguments to the function can be derived. But the type of the result can still confuse programmers. Although maybe it's not so bad, because an overflow that happens during multiplication (or other math operations) is undefined-behavior, but an overflow that happens during assignment of the multiplication result to a variable that can't hold it is only implementation-defined behavior, not undefined behavior.

> With relatively few exceptions, if your program hits undefined behavior, then your program was already doing something pretty wrong to begin with.

The author claims: “We have the absurd situation that C, specifically constructed to write the UNIX kernel, cannot be used to write operating systems. In fact, Linux and other operating systems are written in an unstable dialect of C that is produced by using a number of special flags that turn off compiler transformations based on undefined behavior (with no guarantees about future “optimizations”). The Postgres database also needs some of these flags as does the libsodium encryption library and even the machine learning tensor-flow package.”

So basically the programers of the most used C programs consider the C standard so broken that they force the compiler to deviate from it. (Or they are not able to do things right?)

Somewhat similar to sql situation. (difference being, i am not sure if anybody implements ansi-sql fully)

Standard is just common denominator agreed by actual competitors. (Standard takes into account the chipset where incrementing max_int produces a beep instead of min_int). If user wants to use more advantages of the product (dbms/compiler/hardware), one must sacrifice portability and use cnonstandard extensions.

> extremely rare in the code I see

Well, I learned of the change in compiler behavior some years back because I had written loop code with a sanity check which depended on signed integer overflow wrapping, along with a test case to prove that the sanity check worked, and that test case started failing:

   not ok 2 - catch overflow in token position calculation

   Failed test 'catch overflow in token position calculation'
   at t/152-inversion.t line 70.

To the extent I can be, I'm done with C. I leave it to people who think that silently optimizing away previously functional sanity checks is an acceptable engineering tradeoff, and who disparage those of us who have been bitten.

> people who think that silently optimizing away previously functional sanity checks is an acceptable engineering tradeoff

To be fair, as several people and TFA have pointed out, this isn't a problem with C, but with defective/malicous C compilers. Admittedly, that's not much help if you can't find a compiler that isn't defective/malicous, though, so I can only wish you the best of luck.

I don't get what you or the comment you're responding to are wishing for.

Do you want compilers to stop adding optimizations while staying within the bounds defined by the spec? That they somehow guess that a given piece of code that may trigger UB is too important for them to optimize it based on the assumption that the developer knew what she was doing and ensured that it wouldn't?

The case of a compiler update breaking the test sounds like desirable to me. It pinpoints a critical piece of code that was relying on a specific implementation's behavior that is not specified. This could have been triggered by a compiler or architecture change. If this is something that is a hassle to fix immediately, you can temporarily downgrade to the version of the compiler used earlier or disable some optimizations.

I fail to see why one would consider a compiler evolving while conforming to language specification defective or malicious (however, with that definition I fear that finding one that isn't may indeed be difficult).

I want default behavior which does not surprise me. I have come to understand that this probably means I want a different language, because the C spec essentially requires compiler authors to make optimizations which introduce surprising semantics in order to compete on performance with other languages.

It would be fine if I could opt into new semantics — something like Rust's "editions" would resolve my objections about these compiler optimizations. But that doesn't seem to be on offer in the C ecosystem.

Only moving away to other languages will do it.

C culture and to certain extent Objective-C and C++ ones are tainted by microptimizaitons while typing, where the compilers are the worst examples.

Unfortunely UNIX and C go together, so those that want to keep UNIX like platforms around bettter fix C somehow.

That sucks. I can imagine next generation operating systems improving on Unix technically (standing on the shoulders of giants, yo), but establishing standards a la POSIX will be extremely difficult. It's hard to fight against the interest of platform vendors to lock their users in.

Microsoft fought for several years against C, trying to migrate everyone to C++ as the future of Windows systems programming.

Yet Azure Sphere, despite its security message, uses C only SDK.

Meanwhile Visual Studio now supports C11 and C17.

Market pressure, their customers weren't willing to buy into it, and the new Microsoft also wants all those POSIX FOSS packages written in C running on Windows.

I think that's unfair. The taint is 100% on the C++ side of things.

Hopefully the newer compiled languages will kill C++.

Nope, the tait lies 100% on the copy-paste compatibilty with a C subset.

All you've done here is show that with C the concept of the l-user is still around.

> I don't get what you or the comment you're responding to are wishing for.

Quoting from another of my comments:

> > [What are you objecting to?]

> Inferring any propositional statement about the program (eg "this pointer is not null") from the fact that its negation would imply undefined behaviour.

That is what the problem is. Undefined behaviour is a licence to implement operations without regard for unusual corner cases, not to infer the absence of said corner cases from those operations and then apply that 'knowledge' elsewhere.

> I fail to see why one would consider a compiler evolving while conforming to language specification defective or malicious

It's https://en.wikipedia.org/wiki/Malicious_compliance in near-textbook form.

> Inferring any propositional statement about the program (eg "this pointer is not null") from the fact that its negation would imply undefined behaviour.

I'm not completely sure I understand that correctly, but do you mean that statements that are constant unless considering a possible (and "credible"?) implementation of UB shouldn't be fair-game for the compiler to optimize out? EDIT: I think I see a more tricky case that may be one of those you're referring to. Dereferencing a pointer further in the code shouldn't be a valid justification for optimizing out previous tests of it being null. I can relate with that but I suspect that it would prevent many classes of branch pruning.

I get what you suggest while describing malicious compliance but I can imagine that it could be a false impression resulting from trade-offs that favor optimization opportunities to "out-of-spec but canonical/natural" implementations.

> Dereferencing a pointer further in the code shouldn't be a valid justification for optimizing out previous tests of it being null.

Not further. Anywhere. If the implementation wishes to rewrite pointer dereferences from `use(*p)` to:

  if(!p) abort();
  use(*p);

it may do so (undefined behaviour!), but if it chooses not to do so, it may not later pretend that it did, and remove a explict `if(!p)` that the programmer wrote. Given:

  use(*p);
  if(!p) return NOPE;

this is fine:

  if(!p) abort();
  use(*p);
  //if(!p) return NOPE; // unreachable because of if, not because of use

but not:

  use(*p);
  //if(!p) return NOPE; // CVE-20XX-XXXXX

The implementation can optimize based on information (like "p is not null") that is actually true (even if that's because it made it true), but not based on information it assumed was true on the basis that it counterfactually could have made it true (but didn't).

> it would prevent many classes of branch pruning.

Yes, that's the general idea.

What is the meaning of use(*p) in case p is null? what should the compiler emit?

> What is the meaning of use(*p) in case p is null?

It dereferences a null pointer, invoking undefined behaviour, then calls the function `use` with the resulting value.

> what should the compiler emit?

Probably something to the effect of:

  ld r0 [sp+.p]  # if p is not already in a register
  ld r0 [r0]  # *p
  jsr use

but it would be fine to emit something like:

  ld r0 [sp+.p]  # if p is not already in a register
  jz r0 .panic
  ld r0 [r0]  # *p
  jsr use

because the jz can only be taken when undefined behaviour happens.

If p is statically null, it should emit compile time error. If not, it should emit machine instruction to deference p.

In all of the examples above it will emit a machine instruction to dereference p. What your grandparent is complaining about is that it will later remove an "if (!p)" test.

> I fail to see why one would consider a compiler evolving while conforming to language specification defective or malicious

Conforming to the spec is not a virtue. We want the compiler to be reasonable, regardless of whether the spec is. When the spec is malicious, conforming to the spec is malicious behavior.

For example, the Java spec says that the << operator (bit shift left) will accept any right operand, but performs a modulo-32 operation on the right operand before doing the shift. So `a << 4` is `a` shifted 4 places left, `a << 20` is `a` shifted 20 places left, and `a << 36` is `a` shifted 4 places left again.

This is absurd, and I'm comfortable calling it a bug in the spec. `a << 40` needs to have 0 in the lowest 40 bits. It does not need to have random values in bits 8-31.

This behavior is documented, but that doesn't make things better, it makes them worse.

But the philosophy that says "if it's documented, then it's OK" doesn't even allow for the concept of a bug in the spec.

I recall a bug-report discussion that I sadly have never been able to find. It contains a pretty bad side-effect of this.

It had code like:

    int *p;
    // lots of code
    
    if (p != NULL)
        return 1;
    // use p

Then a later refactor wrongly added a single line before the if statement:

    int *p;
    // lots of code
    
    int a = *p;
    if (p != NULL)
        return 1;
    // use p

This meant the null check was optimized away, since de referencing a null pointer is undefined behavior, so the if-statement can be assumed to be always false. This then lead to actual errors (perhaps even an exploit, I do not recall) arising from the removed null-check.

I think in general the "sanity check" cases are the worst. It is hard to determine whether an expression causes undefined behavior if you cannot try and evaluate it. Perhaps a compiler intrinsic that checks (at runtime) whether an expression causes undefined behavior could be useful here. Though I can imagine such an intrinsic being essentially impossible to implement.

This was in the Linux kernel, which is compiled with special kernel flags which make dereferencing null pointers legal. In the context of that code, dereferencing a pointer and later checking it for null was absoutely meaningful. Optimizing the later null check was a compiler bug that was acknowledged and fixed. The compiler here didn't respect the semantics it had promised to kernel code. This was entirely uncontroversial; it was not a case of "unwanted optimization based on undefined behavior", it was a case of "compiler bug breaking well-defined code". Again, all this in a kernel context.

In user code GCC will still happily remove the null check because in user code this is an actual bug in the user's code.

Actually, there's a even worse version:

  struct foo { ...; bar_t bar[NBAR]; };
  struct foo* p = ...;
  bar_t* q = &p->bar[0]; // add rq, rp, #foo_bar_offs
  // other declarations
  
  if(!p) return NOPE; // optimized out
  // use p and q

Not even any dereferencing, just pointer arithmetic.

Could you post a complete example please and tell us which compiler optimizes out this check? Clang and GCC don't: https://gcc.godbolt.org/z/dbTadEcra

Yes, that one is especially nasty as &p->bar[0] is a constant expression completely solvable at compile time. It's equivalent to the offsetof() macro of stddef.h

https://lwn.net/Articles/342330/

https://lwn.net/Articles/342420/

If the program was previously "fine" on version x.y.z of some compiler, then it is most likely still fine on it. That's the target that the program was written for.

There's some disagreement on whether you can call a program "fine" that breaks after switching to a newer version, or a different compiler.

I see a lot of programmers out there that unfortunately use the behavior of their code on whatever compiler they're using at the moment as a proxy for what the language actually guarantees.

Well, I can imagine a program having something where new C comments ( // ) makes it not divide, like:

a = 5 //* junk here */ 2 ;

I'm sure there are ioccc or underhanded C contest entries doing this to make code work differently on compilers based on if // is starting a comment line or not.

Sure it is an ugly way of writing stuff and you'd be hard pressed to find lots of real world traps like this, but when/if you did have code that "suddenly" miscompiles you might actually think your old code with an old compiler did work, and a new compiler for "the same" language breaks your program. I don't think everyone code base should need full rewrites ever time a new compiler comes out.

Reply on news.ycombinator.com