clang optimization ignores UB (overflow) + standard library bug

[kelbon]

https://godbolt.org/z/r8Yd7nhzc

Not sure standard requires this behavior or not, but it is undefined behavior in IMPLICI T conversion, in naive code, which is expected to be working. Huge disappointment in std::chrono

#include <chrono>
#include <iostream>
int main() {
    std::chrono::nanoseconds ns = std::chrono::milliseconds::max();
    std::cout << ns.count() << "|" << ns.max().count();
}

Output:
/opt/compiler-explorer/gcc-13.1.0/include/c++/13.1.0/bits/chrono.h:225:38: runtime error: signed integer overflow: 9223372036854775807 * 1000000 cannot be represented in type 'long int'
-1000000|9223372036854775807

Expected: max count of nanoseconds

Additionaly: clang (when writed as one line) ignores undefined behavior while constant folding

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (kelbon)

https://godbolt.org/z/r8Yd7nhzc

Not sure standard requires this behavior or not, but it is undefined behavior in IMPLICI T conversion, in naive code, which is expected to be working. Huge disappointmentin in std::chrono

#include <chrono>
#include <iostream>
int main() {
    std::chrono::nanoseconds ns = std::chrono::milliseconds::max();
    std::cout << ns.count() << "|" << ns.max().count();
}

Output:
/opt/compiler-explorer/gcc-13.1.0/include/c++/13.1.0/bits/chrono.h:225:38: runtime error: signed integer overflow: 9223372036854775807 * 1000000 cannot be represented in type 'long int'
-1000000|9223372036854775807

Expected: max count of nanoseconds

Additionaly: clang (when writed as one line) ignores undefined behavior while constant folding

[frederick-vs-ja]

Not sure standard requires this behavior or not, but it is undefined behavior in IMPLICI T conversion, in naive code, which is expected to be working. Huge disappointmentin in std::chrono

#include <chrono>
#include <iostream>
int main() {
    std::chrono::nanoseconds ns = std::chrono::milliseconds::max();
    std::cout << ns.count() << "|" << ns.max().count();
}

Output: /opt/compiler-explorer/gcc-13.1.0/include/c++/13.1.0/bits/chrono.h:225:38: runtime error: signed integer overflow: 9223372036854775807 * 1000000 cannot be represented in type 'long int'
-1000000|9223372036854775807

I think the UB is specified by the standard, as (But this seems wrong):

• [time.syn] requires the representation types of std::chrono::nanoseconds and std::chrono::milliseconds to be some large enough signed integer types,
• [time.duration.cons]/4 requires duration_cast (but the Constraints should work), and
• [time.duration.cast]/2 requires multiplication.

So UB can be raised from the conversion.

Perhaps we can do something helpful, e.g., adding some hardening checks?

Additionaly: clang (when writed as one line) ignores undefined behavior while constant folding

This part is duplicate of #109239.

[kelbon]

some hardening checks?

It would be useful. Also would be useful to have atleast some way to do it without UB in standard. And make it default for implicit conversion

[frederick-vs-ja]

Also would be useful to have atleast some way to do it without UB in standard. And make it default for implicit conversion

Oh, my previous analyzation seemed wrong. [time.duration.cons]/3 seems to reject the conversion, and I need to check why the constraints didn't work.

[frederick-vs-ja]

[time.duration.cons]/3 seems to reject the conversion, and I need to check why the constraints didn't work.

The standard requirements seem somehow broken. See also LWG3090. But the currently proposed resolution would remove the "No overflow is induced in the conversion" part, which, unfortunately, means that UB would be confirmed.

[shafik]

If we make ns a constexpr variable then it fails: https://godbolt.org/z/GGrxvnGo7

[kelbon]

You may disagree, but I feel like chrono types should have 'never' (+-infinity) value, so ms -> secs -> ms will not truncate and will behave as expected. Also it is undefined behavior, i think concrete implementation may choose such way to handle that, may be agreement with other standard library and standarize that way...