manx

@andreasfertig@mas.to May I suggest that you explain more thoroughly that std::start_lifetime_as still requires properly aligned memory. In almost all cases, it is also Undefined Behaviour just like reinterpret_cast is, if the memory comes from somewhere that does not guarantee proper alignment for the destination type. In particular, I consider your older post and example at completely broken. Please do not teach it this way.

@crystalmoon@chaos.social @sschoener@mastodon.gamedev.place While I agree with the sentiment that "POSIX-only" is really not "portable", in this particular case, the Makefile does not really do anything. It merely lists the files that need to be compiled. Trivial to "port" to any non-POSIX compiler. Frankly, "cl.exe src\*.c" does the trick.

@GyrosGeier@hachyderm.io How do I know that? git-svn does not handle .gitattributes, and if you have a rogue .gitattributes somewhere in the tree and use git-svn, affected files can be in modified state because the file stored in the repo is not equal to the file that git would commit to the repo.

@sam I see the primary use case for trackers in generating the stems.

@sam Spamming our forums at https://forum.openmpt.org/ would be fine too.

@stefan@front-end.social You started this thread by promoting nonsense LLM propaganda to random internet people.

@stefan@front-end.social Liking LLMs is turning your brains off. Do not promote this nonsense. This can trivially be seen by looking at the points the site mentions as being good uses of LLMs. ALL of these are themselves subject to hallucination. The author did not turn their brains on in the first place.

@kirb Once they have already scraped millions of URLs and put them in their todo queue, a captcha is not even a viable solution any more, because their hostile use of TCP connections is unaffected by the captcha. In some cases, I had to temporarily or permanently blackhole whole /8 subnets.

@andreasfertig@mas.to How is that still online without any correction at all?

@DanielaKEngert@hachyderm.io I think reducing the amount of UB in C++ is a good goal to strive for in general (within some limits), however in cases like this, all it can do in is changing UB to an assertion failure. From a perspective of the code author, the outcome is exactly the same: I need to fix a bug. The bug is still there, it might just manifest itself in less extreme ways (which is good). UB is not the important problem here. The problem is teaching writing completely broken code.

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to std::start_lifetime_as is probably fine for doing what it was designed to do, i.e. working around type confusion in legacy C APIs (I know of examples in the audio field), or C-style allocation functions. However, people apparently like to also teach it as some kind of crude de-serialization primitive. Even the original paper advertises this use case, which frankly horrifies me: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2590r2.pdf ...

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to ... You really cannot just take any std::byte* from somewhere and just "de-serialize" it with std::start_lifetime_as to some type that might have alignment requirements greater than 1 byte. Whether doing things wrong results in language level UB, or in assertion failure due to stdlib hardening, or really any other run-time decision that results in process termination, does really not matter at all - the bug lies in the intent, not in its manifestation.

@GyrosGeier @rygorous @aeva I am amazed how a lot of major filesystems still do not support data checksumming at all.

@videolan @sashk which you will forget, like the last time, or countless times before that. Hint: the screenshot is already 2 (TWO!) versions behind.

@videolan It feels very questionable when we see Linux distributions pushing important security updates, yet you are withholding these security fixes for Windows and/or Apple users. I honestly cannot remember the last time, the update check actually found updates for me. I always update by hand.

@stefan@front-end.social I am really not sure if linking this website is desirable at all. It literally says "I like LLMs.".

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to Honestly, it does not really matter in this case whether it is straight language Undefined Behaviour, or whether it optionally can trigger an assertion in some stdlib implementations. Both are bugs, and it should not be used or taught this way.

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to It still does not matter. It is not statically decidable in all cases anyway (unless someone adds pointer alignment to the type system, which P3100 does not talk about), so the Undefined Behaviour (or run-time assertion failure due to precondition violation) still remains. Ignoring alignment requirement in a call to std::start_lifetime_as is a bug, and it really must not be used or taught that way.

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to You missed "statically". If the compiler cannot tell me at compile-time that the code is wrong, it is rather useless. All you are suggesting is trading Undefined Behaviour for assertion failure. Both result in a Denial-Of-Service vulnerability due to a memory bug. Both are bugs I need to fix. Run-time hardening of the standard library cannot help here. P3100 cannot help here.

@DanielaKEngert@hachyderm.io @andreasfertig@mas.to https://andreasfertig.com/blog/2025/04/the-correct-way-to-do-type-punning-in-cpp-the-second-act/ BROKEN CODE: bool ProcessData(std::span bytes) { ... ConfigValues* cfgValues = std::start_lifetime_as(bytes.data());

@GyrosGeier@hachyderm.io Beware of .gitattributes. It can change what git would commit. If you do not honor that, git can get confused. Usually this never happens (as the source comes from git anyway), but if you want to handle potentially malicious sources, this can maybe get important. In particular text and eol attributes can cause git to normalize line endings.