This is useful when doing several asserts in one loop, for instance: there
isn't always a way to deduce the loop index from the assert failure, so
these macros allow the user to give more context information.

we do not care about.

CPPUNIT_ASSERT_EQUAL(a, b) “expects” the value “a” and fails if “b” is
different. The logic seems backwards to me, but matching CppUnit’s
output is important so I’m doing the same in LolUnit.

and slightly tweak the API syntax.

CppUnit is really too large for our needs, would require to be precompiled
for many platforms, in all combinations of debug and release modes, with
additional variants for libstdc++ versions. We replace it with our own
lightweight implementation, LolUnit. Good riddance.

so that it builds properly.

quad application.

"make bench" to work when cross-compiling for Windows.

This is at the cost of one additional multiply, but performance increases
by more than 11%, because the PS3 pipeline is a lot happier now.

version of sinf().

Instead of dividing by 2 and rounding with magic number 2^52, we round with
magic number 2^53, which gives us the parity with at least one mul less.

This version is faster (about 25% on x86_64) and works when extreme
optimisation options such as -ffast-math are used.

We know we will not have denormal floats on the PS3, but we should still
create denormal halves in case the other end (maybe the GPU?) knows how
to handle them.

Apparently g++ 4.5.2 on native mingw32 inserts references to __Unwind_Resume
when exceptions are supported. This function is not present in the g++ 4.4
cross-compiler found in Debian, leading to linker errors when cross-
compiling. So we use the latter to build CppUnit itself.

Another problem arises with g++ 4.4, which is the incomplete stripping of
exceptions when -fno-exceptions is used. Since ExceptionTestCaseDecorator.h
has such code, we protect it with #ifdef __EXCEPTIONS in the meantime.

used on the PS3 platform.

number conversions.