As for Jai, etc... right. I was conflating two things. Jai for keeping the spirit of C wrt "the programmer is in control." The other idea being FP languages with hardware-level capabilities.
This won't happen until FP implementors work on bridging the gap between typed lambda calculi (ideal source languages) and the C machine model (ideal target machine). This needs:
(0) A limited form of dependent types, tailored to express invariants of data structures, rather than arbitrary mathematical propositions.
(1) Substructural types, so that use-once objects can be reclaimed eagerly and deterministically.
(2) Designing abstractions so that their implementations have no overhead with respect to what a C programmer could have done manually.
But FP people in general don't seem terribly interested in any of these. Instead they write compilers targeting JavaScript.
I've been under the impression that there is some research in this area...
What I have in mind is one way to hopefully achieve that “significant usability improvement”:
(0) Have dependent types, but restrict them to be first-order. This is good enough for programming purposes: invariants of data structures can almost always be expressed as first-order formulas.
(1) Use known and proven type inference technology: generate a system of type-level equations, and solve them via first-order unification, which is known to be decidable and tractable in practice.
But I don't see anyone in the FP community exploring this possibility.
"(0) Have dependent types, but restrict them to be first-order. This is good enough for programming purposes: invariants of data structures can almost always be expressed as first-order formulas.
(1) Use known and proven type inference technology: generate a system of type-level equations, and solve them via first-order unification, which is known to be decidable and tractable in practice."
I'll add those to the list given that sounds worth a good try. Especially if combined with automated approaches to generating & solving for common attributes like in SPARK and AutoCorres. Maybe something will come of it.
The undefined behavior is there primarily to account for the variation in platforms. Higher-level languages avoid this by hiding it in a platform-specific compiler or VM, which in turn limits the number of platforms the code can actually be run on.
> Empirically, virtually every single large C program in existence depends on undefined behavior to continue working.
Citation needed. I get that undefined behavior is hard to spot, but having spent years writing cross-platform C I cannot accept that every large program depends on it without some evidence.
It may have started this way in the committee (though that isn't true for e.g. type based alias rules), but nowadays it's depended on for performance. You lose a lot of ability to reason about loop trip count if signed overflow is defined, for example.
> Citation needed. I get that undefined behavior is hard to spot, but having spent years writing cross-platform C I cannot accept that every large program depends on it without some evidence.
The probability that a program depends on any of (nonexhaustive list) (a) inspecting the value of a dangling pointer; (b) taking the offset from NULL; (c) type-punning through a union; (d) type-punning through any type other than char; (e) signed overflow; (f) order of operations between sequence points quickly approaches 1.
Here's a very interesting survey investigating this: https://www.cl.cam.ac.uk/~pes20/cerberus/notes50-survey-disc...
I agree it is depended on in many cases, and should not be. That is on the programmers though, not the language designers. The behavior does not exist to cover deficiencies in the language but those of its users.
> The probability that a program depends on any of (nonexhaustive list) (a) inspecting the value of a dangling pointer; (b) taking the offset from NULL; (c) type-punning through a union; (d) type-punning through any type other than char; (e) signed overflow; (f) order of operations between sequence points quickly approaches 1.
Just on your short list I can tell you the probably of any of our C code depending on that approached zero:
> inspecting the value of a dangling pointer
Our code was malloc once, free never. Dangling pointers were by definition impossible.
> taking the offset from NULL > type-punning through a union > type-punning through any type other than char > signed overflow > order of operations between sequence points
Our coding standards covered all of those, so their existence would be a code review failure.
We weren't specifically writing to MISRA C, but I will wager that your statement is not accurate for any project that follows it or any similar standard. Yes, it is a lot of work that you don't need to do with a more controlled and less flexible language. I'm not going to challenge that. I'm just challenging the idea that nobody is writing C that avoids undefined behavior.
Dangling pointers can arise in other ways: use an uninitialized pointer variable, take the wrong offset from an otherwise valid pointer, etc.
Plus, the cases you highlighted are programming errors, not reliance on undefined behavior.
(b) True. I think this one should be defined and standardized.
(c) It seems that C99 defined this behaviour. https://stackoverflow.com/questions/11639947/is-type-punning...
(d) I am not knowledgeable enough. Could you give an example?
Honest question: Doesn't implementation-defined behavior suffice to address variations across platforms? Undefined behavior has far more drastic consequences: compilers are allowed to assume it simply won't happen.
False. Squeak Smalltalk runs with a bit-identical model on a ridiculous number of platforms. (which is to say OS/ISA combinations) Though to be fair, it accomplished this trick by having the VM written in a Smalltalk subset that gets translated into C, which is then compiled by local compilers.
I'm not finding an academic citation, but perhaps a respected academic's blog counts? I think John Regehr is the likely source of these claims. In his words:
My guess is that most or all executions of GCC
and LLVM (and basically all other large programs)
would be undefined.
He starts the post by stating that the "majority of the programs in SPEC CINT 2006 execute undefined behavior during the course of a 'reportable' run". A majority is definitely not "all", but the SPEC programs are also not "large":
Let’s consider the SPEC benchmark programs. Making them fast is something that most compiler vendors would like to do. Using IOC we found that a majority of the programs in SPEC CINT 2006 execute undefined behavior during the course of a “reportable” run.
For large programs, he then looks at the GCC and Clang compilers themselves, and concludes:
If we consider them to be written in C99, LLVM/Clang 3.1 and GCC (SVN head from July 14 2012), running on Linux on x86-64, do not appear to admit any non-trivial, non-erroneous executions. I expect the same is true of most or all previous versions of these compilers. IOC tells us that these compilers execute undefined behaviors even when compiling an empty C or C++ program with optimizations turned off.
Lest you think that the undefined behavior is caused by inappropriately forcing C99 on earlier code, he goes on to explain that the problems probably exist there too, but that the lack of tooling for earlier dialects prevents proper analysis:
My guess is that most or all executions of GCC and LLVM (and basically all other large programs) would be undefined by C89 (and not just C99) but we can’t tell for sure without using a comprehensive undefined behavior checker for C89. This tool does not exist. I confess to not personally having the gumption necessary for cramming GCC or LLVM through the best available dynamic undefined behavior checkers: KCC and Frama-C.
His overall message (which I think is true and merits significant reflection) is that C's utility as a real world language is based on unspoken agreements with compiler writers to use some degree of judgment and restraint when making optimizations based on undefined behavior. Rather than a standardized language defined by the written specifications, we are dependent on an "undocumented dialect of C where undefined behavior can sometimes be exploited and sometimes not":
This means that a correct C/C++ compiler could generate code that is almost infinitely faster than the code generated by current compilers. The SPEC harness would of course complain that the compiled programs produce the wrong results, but this is actually a bug in the SPEC harness, not a bug in the compilers. If we insist that a compiler turns these SPEC programs into executables that produce the specified results, then we are effectively mandating that this compiler implements a bizarre, undocumented dialect of C where undefined behavior can sometimes be exploited and sometimes not.
I think the amount of undefined behavior in C--and that undefined behavior is necessary to compensate for its design mistakes (e.g. signed overflow)--disqualifies it from being a "great language". Empirically, virtually every single large C program in existence depends on undefined behavior to continue working. And that UB is very difficult to spot [1].
> It'd be nice to be able to define how memory should be handled for a particular application at the granularity of the target platform while still getting the safety guarantees.
> I like where Blow is going with his Jai language. It'd be neat to see more research in this area.
You said that you wanted safety guarantees, but Jai is explicitly not about safety guarantees. In fact, being memory unsafe is one of the key design criteria behind the language. That's what differentiates it from for example Rust.
[1]: https://news.ycombinator.com/item?id=11728324