Python cryptography, Rust, and Gentoo [LWN.net]

Python cryptography, Rust, and Gentoo

Posted Feb 15, 2021 19:44 UTC (Mon) by Wol (subscriber, #4433)
In reply to: Python cryptography, Rust, and Gentoo by anton
Parent article: Python cryptography, Rust, and Gentoo

Thing is, C *is* a bad fit for modern architectures. It has a whole bunch of features that are undefined, or implementation-defined, which are MEANT to be low-level "match the hardware". Except that they aren't.

Let's just get rid of all these so-called "low level" cock-ups, accept that C is now a high-level language and that undefined and implemetation-specific behaviours shouldn't exist, and move on.

Someone brought up retpolines - that monstrosity that tries to make sure that both the hardware and the software agree on the imaginary hardware interface that's where the CPU microcode and language macrocode meet ... wtf are we doing!

Cheers,
Wol

to post comments

Python cryptography, Rust, and Gentoo

Posted Feb 16, 2021 9:27 UTC (Tue) by anton (subscriber, #25547) [Link] (2 responses)

I did not mean the language lawyer version of C. That version is a bad fit for any architecture (including the PDP-11). However, it's great for adversarial compiler maintainers who want to do whatever they want (e.g., produce good benchmark results, grudgingly cater to requests by paying customers and tell other users that their bug reports are invalid), because this version allows them to always blame the programmer for something or other. After all, no terminating C program is a "strictly conformant program", and whenever someone mentions "conformant program" (the only other conformance level for program defined in the C standard), the adversaries produce advocacy why we should consider "conformant programs" as meaningless (interestingly, they claim that we should take the rest of the C standard at face value).

I mean C as used in many programs, which has a pretty simple correspondence to architectural features (and you see it easily in contexts where optimization does not set in, e.g., when you separately compile a function that performs just one thing).

The adversaries want us to consider C as a high-level language with no correspondence to the bare metal; that makes it easier to blame the programmers and absolve compiler maintainers of responsibility. The question is why any programmer would want that. We have plenty of high-level languages, often better than C in that capacity, but not that many low-level languages; basically C is the only popular one.

Concerning a totally defined C: I think that is at odds with a low-level language for multiple architectures, but as most (all?) C compilers have demonstrated for the first quarter-century of the language, that's no hindrance for implementing C in a benign rather than adversarial way. And for those who don't know how to do that, I have written a paper (which also explains why I consider totally defined C impractical).

I don't know what retpolines have to do with any of that. They are a workaround for a vulnerability in some microarchitectures and they cannot be implemented in C (there are limitations to C's low-level nature). The vulnerability should be fixed at the microarchitecture level, and I expect that the hardware manufacturers will come out with microarchitectures that do not have this vulnerability.

Python cryptography, Rust, and Gentoo

Posted Feb 16, 2021 13:18 UTC (Tue) by mathstuf (subscriber, #69389) [Link] (1 responses)

> I did not mean the language lawyer version of C.

What version of C are you talking about? The ISO standard? The image of C that you have in your head? My head? The C that GCC 2.95 accepted and worked with?

Let's imagine a world where C compilers magically stop doing "magic optimization" steps that tend to break code. What's going to happen is that C programmers that don't know this stuff already is going to have their code be pessimized and, presumably, slower in practice. What are they going to do? Start writing their C in such a way that the compiler was doing internal transformations during optimization passes anyways. They'll learn C more (and, I imagine, less satisfied with it), hopefully be using linters and tooling to tell them where their NULL checks are inverted with uses and such.

Rereading that, maybe it wouldn't be so bad. Maybe folks would migrate to better languages. Others might actually learn more about how loose C is in practice. The optimization passes could be migrated to the linters rather than the compiler to explain "hey, you could reorder your code to This Way and gain some performance". Maybe these passes would then gain some prose explaining what and why of them.

Then again, I have no idea such a C would be specified at ISO to disallow these optimizations while still allowing for architectures to not be forced into twos-complement representations or the like because "it's faster/easier for them".

Python cryptography, Rust, and Gentoo

Posted Feb 16, 2021 15:28 UTC (Tue) by anton (subscriber, #25547) [Link]

As I wrote: "I mean C as used in many programs", and I actually point to a paper where I explain this in more detail. As for "pessimizing", it's certainly the case that advocates of adversarial C compilers claim that the adversarial behaviour is good for performance, invariably without giving any numbers to support these claims; maybe they think that repeating these claims and wishful thinking makes them true.

Wang et al. checked that for gcc-4.7 and clang-3.1 and found that the adversarial "optimizations" produced a minimal speedup on SPECint 2006, and that speedup could also be achieved by small changes to the source code in two places.

Yes, a performance advisor that points out places where changing the source code may improve performance would be a more productive way to spend both the compiler maintainer's and the compiler user's time than writing "optimizations" that break existing code, "sanitiziers" to find the places where the "optimizations" break it, and, on the programmer's side, "sanitizing" their code to withstand the latest attacks by "optimizers" (but not the next one). Moreover, such an advisor could point out optimizations that a programmer can do but a conformant compiler cannot (e.g., because there is an alias that even language lawyering cannot explain away). Of course, unlike maintained programs, benchmarks would not benefit from such a performance advisor, that's why no work goes into performance advisors; and conversely, "optimizations" don't break benchmarks (the compiler maintainers revert the "optimization" in that case), unlike other programs, and that's why we see "optimizations".

But what's more, by insisting on a very limited interpretation of what C means, the language lawyers remove opportunities for optimizations that programmers can make in the source code. I have discussed this at length.

I, too am skeptical that trying to change the C standard is the way to get rid of adversarial C compilers (not the least because you won't be able to achieve consensus with the implementors of these compilers on the committee), and I guess that's why advocates of adversarial compilers like to direct the blame for the misdeeds of these compilers at the standard, rather than at the compiler maintainers. It's not the standard that requires them to miscompile existing, tested and working, programs, it's the compiler maintainers' choice, so they alone are responsible.

Concerning architectures with other than two's complement representation of signed numbers, the last new such architecture was introduced over 50 years ago, and descendants of such architectures exist only in very few places and run select programs. There are far fewer of these machines than of the architectures (all two's-complement) that are not LLVM targets and that have brought about the parent article. And coding in a way that makes use of knowledge about the representation is one of the things that you can do for performance in a low-level language (and compilers do not perform all of these optimizations).