Comment by pizlonator

To say they are right is an overstatement, I think. There is insufficient discussion of the empirically measured performance benefits of specific forms of UB.

Some kinds of UB could be turned into something stricter, like reading a bad pointer. This either traps or returns some value. It won't format your hard drive unless you install a trap that formats your hard drive, but that's none of the spec's business. Traps can happen due to timers, so if arbitrary traps mean UB then every instruction is UB. Even if the spec punts on defining what a trap is, that's a progression over saying it's UB.

Same thing goes for division and modulo. In corner cases, it will either return some value or it will trap. It won't format your hard drive.

The most profitable "true" UB is stuff like TBAA, but smart people turn that off.

Do you know what the performance benefits are of other kinds of UB? Do you know how many of those perf benefits (like being maybe being able to take some shortcuts in SROA) can't be solved by changing the compiler (i.e. you'll get the same perf, but the compiler follows slightly different rules)? Maybe I'm not so well read, but I hardly ever hear of empirical evidence that proves the need for UB, only demos that show the existence of an optimisation in some compiler that would fail to kick in if the behaviour was defined.

Also, if there were perf benefits of the really gnarly kinds of UB, I would probably be happy to absorb the loss in most of the code I write. If I added up all of the time I've wasted fixing signed-unsigned comparison bugs and used that time to make WebKit faster, then I'd probably have made WebKit faster by a larger amount than the speed-up that WebKit gets from whatever corner-case optimisation the compiler can do by playing fast and loose with signed ints.

I suspect that UB is the way that it is because of politics - you can't get everyone to agree what will happen, nobody wants to lose some optimisation that they spent time writing, and so we punt on good semantics.