We have special checked types you can use if you want checked arithmetic. The default is to not check, because CPUs currently make it expensive to check (although I would love it if that could change--we need hardware support though).
As you say, detecting the overflow is easy, but efficiently handling it is not. It adds a branch to every single arithmetic operation, and it makes it much harder for the compiler to optimise things e.g. it is hard to vectorise a loop summing an array, if every + has a conditional branch on the overflow flag.
(Also, I believe it introduces a lot of data dependencies, getting in the way of the out-of-order execution of modern CPUs.)
To handle the overflow on the places where you want to handle them, modern processor doesn't have any problem with an additional jump instruction. Also, modern compilers could optimize the checks away if they aren't used. In effect, implementing overflow checks definitely won't turn your C speed (1s) code in a Python speed (40s) code. I estimate it wouldn't be even two times slower in most of the use cases. It's certainly not the problem of the CPU's.
It can be the problem of the certain compilers if they don't have the infrastructure to reason about overflow flags though. But it's not a hardware problem.
> I estimate it wouldn't be even two times slower in most of the use cases.
In his "We Need Hardware Traps for Integer Overflow"[0], Regher quotes 5% to 100% overhead for languages such as JS or Racket, and that a "highly tuned" checker would likely be in the 5% range. Playing with arithmetics-heavy programs and Rust's checked_* (which are backed by LLVM's overflow intrinsics[1]) I got anywhere from 5 to 40% performance loss IIRC.
That's not a lot, but at the same time when you're competing with languages specifically not paying those 5%, a 5% hit on all computations is not going to get you much love.
Which is why Rust currently lets you do that (via num::Checked* and num::Saturating) but uses overflowing default semantics.
> To handle the overflow on the places where you want to handle them, modern processor doesn't have any problem with an additional jump instruction
This isn't just a jump, it is a branch. Especially when the body of the loop is 6 instructions, adding an extra branch is going to be noticable.
> Also, modern compilers could optimize the checks away if they aren't used
This is equivalent to the halting problem, and most code will not be able to optimise them away. Suggesting otherwise is invoking "sufficiently smart compiler", which is invalid.
In any case, you haven't addressed the problem of missed optimisations (especially vectorisation) caused by having to maintain semantics.
> It's certainly not the problem of the CPU's.
Yes, it partly is: the data dependencies and linearisation caused by checking the CPU flags is bad.
See why for loops are tricky? :)