Comment by muvlon
19 hours ago
It's not even about optimizing some big tech codebase by 0.5%. The progress guarantees in particular are in place s.t. Nvidia can choose a certain implementation strategy in Cuda C++ that has "surprising" consequences for users (one thread getting stuck in an infinite loop that never yields can livelock its entire warp) but still get to claim "full C++ standards compliance".
So let it livelock the entire warp when someone writes an infinite loop. Should we start replacing integer division by zero with INT_MAX so that people aren't "surprised" by their program crashing?
I mean that's what they did, and that's why there's that UB. All I'm saying is that this is the "weird platform behaviors exist and must be legalized by the standard" kind of UB and not the "we want a 0.5% win for benchmaxxing" kind of UB (the standard has plenty of both).
No, they didn't. UB is a cop out and inserting yield is just plain bad. Locking up one or more threads in an implementation defined manner would be the outcome of least surprise (I already know it's going to lock up at least the one thread).
4 replies →
I don't see the issue. Just let wrong code do wrong things But let it do the expected wrong thing, rather than changing the code to something unexpected.
"No." <-- the C++ committee.
Isn't that the strategy for most of the stuff in C++? It's the common denominator of a wide variety of platforms. That's why numbers didn't have to be two's complement and characters didn't have to be ASCII for ages.