> But are you saying there are usually never any issues with integer arithmetic and overflow vs. optimizations (reordering, common subexpressions etc)?
I'm saying that it is uncommon. For example, the example you gave works fine! Most examples work fine since you can do most of the same type of transformations. The exceptions are largely operations where the upper bits affect the lower bits. Division is one, and right shift is another, but nearly all the other bitwise and math operations do not have this property.
The other place where signed-wrapping-is-UB is often used for optimization is in things like loop bounds. Given something like:
for (int i = start; i
Due to the UB of signed wrapping, the compiler can assume that initially i In Java, the compiler couldn't take advantage of that. However, the impact isn't as big since in any loop that accesses arrays based on the index, bounds checks have to be made, and a typical pattern is to do some up front bounds checks [1] which can guarantee than the main body of the loop can then be run without additional checks: and this subsumes the checks that would be needed due to wrapping signed values anyways. So basically you have to do those checks anyways in many cases.
> A branch like “if a+1 Sure, but these cases aren't very interesting for comparing the effectiveness of the signed overflow optimizations. It's a case where the optimization breaks the (wrongly expressed) intent of the programmer, so the difference is only between a fast, broken program and a slower, perhaps correct one.
Presumably in the signed-overflow-is-UB world, such a check will have to re-written in a different way (which might even end up slower).
> Seems that not checking for overflow and not being able to assume there is no overflow, would give the worst of both worlds (slower because of lack of some optimizations but still not safe against overflow like C#’s “checked”).
Not exactly - it's just a different point on the spectrum. On the one side you have overflow is UB, which leads to some (fairly limited) optimization opportunities, but also more unexpected results and broken programs, and on the far other size you have overflow is an always-caught checked error. Java is somewhere in the middle: overflow is well-defined (it wraps), which gets you most of the speed of the UB approach (no checking needed, only a few relatively minor optimizations are lost) without the unexpected results of UB - but you still have broken programs when overflow actually occurs (unless it was unexpected). See also Rust's strategy here which is interesting.
> I thought a deref of a possibly overflown value was what could risk security, ie so long as all array indices and similar are range checked then nothing bad can happen?
When "nothing bad can happen" from the point of view of the JVM, i.e., the security sandbox isn't broken, you can't access arbitrary memory, violate the type system, interfere with unrelated code, etc. Of course, overflowing an index and then accessing into the array could still do plenty of "bad things" depending on the higher level semantics of the program, since you are now executing an unexpected code path. You might return sensitive data to an attacker, whatever.
---
[1] More details: https://wiki.openjdk.java.net/display/HotSpot/RangeCheckElim...