Boost C++ Libraries: Ticket #9497: mpq_rational asserts when used in variant

John Maddock — Sat, 21 Dec 2013 17:12:12 GMT

I'm unable to reproduce here - but I do notice that gmp_rational, unlike the other classes, does not allow a move-copy from an already moved-from source object. I've committed a tentative fix (see https://github.com/boostorg/multiprecision/commit/f552968b21b2fd37c7ce69cb283d12d2c2a765bc), can you please try that out and report back?

Many thanks, John.

Sat, 21 Dec 2013 17:42:01 GMT

Everything looks good with the patch. It doesn't assert anymore and behaves as expected.

Thank you!

status changed; resolution set

John Maddock — Sat, 21 Dec 2013 18:08:35 GMT

status new → closed
resolution → fixed

That's a relief then!

Closing down.

status, component changed; resolution deleted

John Maddock — Sun, 22 Dec 2013 10:02:06 GMT

status closed → reopened
resolution fixed
component multiprecision → variant

I'm reopening and reassigning this as I don't really see why Variant would be move-copying from an already moved-from object. It might not be an issue, but it feels like a bug...

owner, status changed

John Maddock — Sun, 22 Dec 2013 10:02:26 GMT

owner changed from John Maddock to Antony Polukhin
status reopened → new

status changed

Antony Polukhin — Wed, 25 Dec 2013 08:58:26 GMT

status new → assigned

owner, status, component changed; cc set

Antony Polukhin — Wed, 25 Dec 2013 12:03:20 GMT

cc antoshkka@… added
owner changed from Antony Polukhin to John Maddock
status assigned → new
component variant → multiprecision

It's not a bug, boost::variant tries to satisfy the never-empty guarantee:

boost::multiprecision::mpq_rational has no noexcept default constructor because default constructor of gmp_rational is not marked with noexcept
Foo has no noexcept default constructor
because of previous two points variant can not use boost::multiprecision::mpq_rational or Foo as fallback types
copy constructor of Foo is not noexcept, there is no fallback type in variant, so it attempts to copy/move content of variant to backup storage before doing an assignment of Foo.
...

Looks like gmp_rational calls only pure C methods. Can the default constructor of gmp_rational be marked with noexcept?

status changed; resolution set

John Maddock — Tue, 11 Mar 2014 17:11:16 GMT

status new → closed
resolution → fixed

Looks like gmp_rational calls only pure C methods. Can the default constructor of gmp_rational be marked with noexcept?

Not really, as those methods perform memory allocation, and the allocator used can be a user-supplied custom one (which throws). In any case, if you're happy copying an already-moved-from-object and there's really no alternative, then that's fine. Closing down.

Antony Polukhin — Tue, 11 Mar 2014 19:31:03 GMT

Replying to johnmaddock:

In any case, if you're happy copying an already-moved-from-object and there's really no alternative, then that's fine. Closing down.

I'm not 100% happy with that, but the only solution I see is to totally disable move assignments for cases when there is no fallback type. This can break user's code where variant is used with move-only types and there's no fallback. I'm in doubt.