C++20 is back

Revision en3, by orz, 2024-03-27 12:51:58

In contests and in the custom invocation, after a pause, one can again find GNU G++20 64-bit. However, the slowdown issue because of which the language disappeared persists, there are still snippets of code that can slow down the execution on Codeforces servers by a factor of 100 or so.

So what is the official position of Codeforces headquarters and of our community on that?

  1. Are there general methods of constructing testcases that can exploit this GCC/Windows bug and therefore slow down solutions (like there are anti-hash tests for solutions using unordered_set)? Are there methods that slow down both contest mode and custom invocation mode (since 249807302 is only slow in the contest mode whereas 253289473 is only slow in the custom invocation)?

  2. If so, should we expect such tests in future contests, will such methods be used during the test preparation in contests?

  3. If this is deemed an unwanted bug, will participants be protected from unscrupulous hacks and tests? (For example, if the solution was hacked or FSTed because of a compiler bug rather than because of anything else, will the author be able to appeal against it and get their solution accepted?)

  4. Is there no bug in 32-bit version? If it's 64-bit only, maybe CF team could add a C++20 32-bit compiler?

  5. Is there this bug in Microsoft Visual C++ compilers (32 or 64 bits)? If so, is it the reason why we cannot see it among the options?

  6. Are there generic methods of protecting against the bug (except switching to 32 bits)? I read (and enjoyed) a blogpost by ToxicPie9 regarding this issue, he suggested to add several lines of code to the solution. However, it does not always help: 253290209 still gets TL (and the author said about that).

  7. Has anyone a clue what is going on, why do the aforementioned programs run slowly? Is this bug reported to the GNU or Microsoft developers who are responsible for this strange behavior?

  8. Why does behavior in custom invocation and in contest/gym mode differ?

  9. Is this bug reproducible on other platforms? For instance, on godbolt there are some windows compilers. Is there a way to slow down this code: https://godbolt.org/z/5G73dfezK (by cleverly choosing the value of evil or anything else)? UPD: probably no way since Godbolt uses UCRT rather than MSVCRT. UPDUPD: this code is slow on Godbolt: https://godbolt.org/z/WE4Mxv9cK.

  10. Is there surely no such bug on Linux? (Anyway, with regard to Codeforces it is probably a pointless question. I am quite certain that transition of all invokers to Linux is the last thing MikeMirzayanov would prefer to do since it is for sure really costly in time, effort, architecture changes, and/or money.)

Tags windows, 64 bit, slowdown, bug

History

 
 
 
 
Revisions
 
 
  Rev. Lang. By When Δ Comment
en5 English orz 2024-03-27 13:06:19 77
en4 English orz 2024-03-27 13:01:54 24 Tiny change: '/WE4Mxv9cK.\n\n9. Is' -> '/WE4Mxv9cK, so UCRT is also flawed.\n\n9. Is'
en3 English orz 2024-03-27 12:51:58 75
en2 English orz 2024-03-26 14:54:37 69
en1 English orz 2024-03-26 07:39:42 2863 Initial revision (published)