Re: [boost] Regression test relying on undefined compiler behavior (c++11)

Subject: Re: [boost] Regression test relying on undefined compiler behavior (c++11)
From: Chris Cooper (chris.cooper_at_[hidden])
Date: 2014-04-23 14:20:58

Actually, I came up with a better idea:

#ifndef BOOST_NO_LIMITS

template <class Number> struct complement_traits

{

    BOOST_STATIC_CONSTANT(Number, min = std::numeric_limits<Number>::min());

    BOOST_STATIC_CONSTANT(Number, max = std::numeric_limits<Number>::max());

};

#else

[SNIP] All the other template stuff for calculating complement_traits for various types

#endif

From: Chris Cooper <chris.cooper_at_[hidden]<mailto:chris.cooper_at_[hidden]>>
Date: Wednesday, April 23, 2014 at 11:59 AM
To: "boost_at_[hidden]<mailto:boost_at_[hidden]>" <boost_at_[hidden]<mailto:boost_at_[hidden]>>
Subject: Regression test relying on undefined compiler behavior (c++11)

libs/utility/numeric_traits_test.cpp generates values complement_traits<Number>::min using a clever recursive template, but that template relies on left-shifting a negative value, and according to the C++11 standard that’s a no-no (“the behavior is undefined”) which means it’s not a constant expression, which means it can’t be calculated at compile time, which means the BOOST_STATIC_ASSERT in line 332 won’t compile, saying “static_assert expression is not an integral constant expression” (I’m using clang++).

See discussion here: http://stackoverflow.com/questions/23250651/strange-behavior-with-c-recursive-templates-when-c11-is-enabled

In addition, it looks like the existing code also relies on the compiler filling the vacated bits with one’s, which I don’t think is what actually happens.

The solution I’ve come up with, is to do the shifting using unsigned values, thus avoiding the compiler no-no, and then putting the result back in the specified type.  In other words, replacing line 73

BOOST_STATIC_CONSTANT(Number, min = Number(Number(prev::min) << CHAR_BIT));

With

BOOST_STATIC_CONSTANT(Number, min = boost::detail::is_signed<Number>::value ? (Number)(((unsignedlong)(prev::min) << 8) | 0x00FF) : Number(Number(prev::min) << CHAR_BIT) );

I’m using constants ‘8’ and ‘0x00FF’ rather than CHAR_BIT since the code would get really ugly if CHAR_BIT != 8.

Thoughts?

Thanks,
Chris