AlisdairM wrote:

I finally committed support for zero length arrays to the boost array library, and it is showing up what I believe are test case errors in the TR1 'tricky' tests.

[tuple interface remains available only through John's TR1 implementation right now]

The following tests from test_array_trick.cpp lines should be reconsidered:

line 51: empty will return true if the array is length zero. verify_return_type(ca.empty(), false);

It only checks that the return *type* is a bool: it doesn't care what the value is, so IMO it's correct.

lines 73-75: are all compile-time fails for array length zero BOOST_STATIC_ASSERT((::boost::is_same< typename std::tr1::tuple_element<0,T>::type, value_type>::value)); verify_return_type(&std::tr1::get<0>(a), static_cast<value_type*>(0)); verify_return_type(&std::tr1::get<0>(ca), static_cast<const value_type*>(0));

Yup, Fixed.

I thought about the tests for operator[]/at/front/back and figure these are OK, even the the results of calling those functions are unspecified (but not undefined) as: i/ you don't actually call the functions! ii/ they are required to be present, so this code MUST compile iii/ I believe even if the effects of calling the fns is unspecified, the actual return type is fixed.

So those tests are good, indeed they must be kept!

With the above mods, I believe we should now pass the TR1 tricky tests for array, and I will leave back-porting the tuple interface to namespace boost for 1.35.

Unfortunately that doesn't quite get the test case compiling: your stub implementations of front() back() etc all need to return a value to actually compile (this is with VC++ anyway), patch follows below, let me know if it's OK to commit. John. Index: boost/array.hpp =================================================================== RCS file: /cvsroot/boost/boost/boost/array.hpp,v retrieving revision 1.25 diff -u -r1.25 array.hpp --- boost/array.hpp 4 Jun 2006 11:01:59 -0000 1.25 +++ boost/array.hpp 4 Jun 2006 15:46:30 -0000 @@ -30,6 +30,7 @@ // Handles broken standard libraries better than <iterator> #include <boost/detail/iterator.hpp> +#include <boost/throw_exception.hpp> #include <algorithm> // FIXES for broken compilers @@ -212,37 +213,43 @@ { BOOST_ASSERT( "out of range" ); failed_rangecheck(); + return null_item(); } const_reference operator[](size_type i) const { BOOST_ASSERT( "out of range" ); failed_rangecheck(); + return null_item(); } // at() with range check - reference at(size_type i) { failed_rangecheck(); } - const_reference at(size_type i) const { failed_rangecheck(); } + reference at(size_type i) { failed_rangecheck(); return null_item(); } + const_reference at(size_type i) const { failed_rangecheck(); return null_item(); } // front() and back() reference front() { failed_rangecheck(); + return null_item(); } const_reference front() const { failed_rangecheck(); + return null_item(); } reference back() { failed_rangecheck(); + return null_item(); } const_reference back() const { failed_rangecheck(); + return null_item(); } // size is constant @@ -272,8 +279,19 @@ // check range (may be private because it is static) static void failed_rangecheck () { - throw std::range_error("attempt to access element of an empty a rray"); + std::range_error e("attempt to access element of an empty array "); + boost::throw_exception(e); } + static reference null_item() + { + // + // This function must exist to allow our various interfaces to + // actually compile, however it will never be called, because + // an exception will always be thrown before we get here. + // + static T placeholder; + return placeholder; + } }; #endif