I think the real problem here is that people underestimate the complexity of C and C++. It often feels like it would be easy to guess what the assembly would look like for some snippet of C/C++ code (especially when it's not using any C++ features, such as the loop in the article) when we really should be thinking in terms of the abstract machine that the standard defines.
I just don't think having 'gotcha' rules the compiler exploits to do whatever it wants regardless of programmer intent is a good idea, and that's how C/C++ treats UB. UB should be a compiler error and force the programmer to specify what they want the result to actually be so there are no assumptions on the programmer end and no surprises from the compiler.
> UB should be a compiler error and force the programmer to specify what they want the result to actually be so there are no assumptions on the programmer end and no surprises from the compiler.
Many types of UB are not easily detected at compile time, or cannot be detected in advance at all. For instance, prompting the user for an integer and incrementing it by 1 is UB when the user enters the maximum value of an integer. So should the compiler forbid adding two numbers unless you explicitly guard against overflow? Perhaps it should, but the resulting language definitely wouldn't look anything like C.
Besides, it's not that the compiler is going out of its way to screw you over with a 'gotcha' rule, it's just optimizing using the assumption that your program is free of UB. As a programmer it is your responsibility that there is no UB. If you feel there are too many pitfalls you could use a different programming language, lobby the standards organization for a standard that has less UB, or use compiler options to detect it (e.g. ubsan for clang).
That's a fair point, but it still feels wrong to me that the compiler can throw away large chunks of code and change the outcome of the function silently.
I think it's easier to define in a standard what's conforming behavior than it is to exactly and precisely define all the possible optimizations a compiler may (or may not) perform. So "undefined behavior" is a decent starting point. Exceptions can be added later, but if you try to spec all allowed optimizations in advance, the spec is going to be massive and also a massive burden.
Yeah, my problem isn't that optimization is changing the code per-se, it's that it is introducing a non-obvious bug by doing so. Ostensibly that loop either returns true or never returns. What if it were provided a number that would never return (assume this is possible given the algorithm provided), but because of the "optimization" now returns true? It literally gives us the wrong answer.
On the other hand, it is not hard for me to incorporate into my mental model of a language that a program should eventually "do something" and if it doesn't, it is literally a wrong program.
In fact, that's exactly what I thought when I saw the code in question. Before even realizing what it is about: I read it like a normal program, I saw that it can only ever return true, therefore that code can do only one thing, and all that fluff in there does not impact the end result. So I don't think it's particularly non-obvious, but it does require something from the reader. I'm ok with that, there's a lot in C and C++ you can't just assume.
I call this a "weird program." It's trying to brute-force an answer out of the computer, but it's the kind of an answer that may or may not ever come. Hardly unlike a program that tries to answer the halting problem. Not necessarily an illegitimate program (bruteforcing is a legitimate way to find answers), but not the kind of thing you'd put inside a normal application. So I'm not concerned about normal applications running into this issue much. If a normal application could go into an infinite do-nothing loop, that is almost definitely a bug -- a literally wrong program. The optimization here only changes how that wrongness manifests, which is expected for UB & optimizations.
I do like that C gives you the escape hatch of using a constant controlling expression if you really want a true "potentially-infinite" loop.
Consider this hypothetical: the loop is only potentially infinite because it is written incorrectly. In testing, I would notice an infinite loop and debugging where that is happening is pretty trivial. When the compiler optimizes that away to just always return true, it has made the source of the bug much more difficult to pin down because the code I asked the compiler to make produces a completely different result than what the compiler compiled.
Some bugs end up being extremely difficult to debug. I accept that as a fact of life; in exchange, we get pretty decent performance.
Guess what I do when there's a bug I have a hard time pinning down? Recompile with optimizations turned off.
By the way, I don't think I have ever witnessed a bug caused by an accidentally-infinite loop getting optimized out. So making that almost-never-happens situation easier to fix is not worth trading any optimizations for, imo. I'd rather just get a warning from compiler/static analyzer (if the code isn't the way it is due to macro expansion).
What should your program do if you overflow a 32 bit signed integer? What if you overflow a 64 bit signed integer? Should it do the "expected" thing and wrap according to two's complement? How do you intend to do that efficiently in a portable manner?
This rabbit hole can be chased forever until you are left with every C++ implementation being a full blown interpreter.
This is especially true given that (a) there is no standard compliant way of checking if you had overflow, and (b) the most pedantically correct way of looping over arrays and such uses an unsigned type (size_t), negating the usual claims about optimization opportunities.
Thus the word "efficiently" in my comment. By requiring weird and almost certainly buggy behavior to be defined, you require the program to behave more like an interpreted program, with associated performance costs.