The reason compress_compatability_test() fails in some cases (like when compiling for ARM with Clang with -O2) appears to be that shifts in vector_compress (the VECTOR24/VECTOR32 cases) are sometimes made like x >> y where x is unsigned int and y is greater than std::numeric_limits<unsigned int>::digits, which means the behavior is undefined. So if
i1 = float_istoobig (*p1)? ~0u : (bitget (*p1, 0, 23) | bitput (1, 23, 24)) >> (1 + max - bitget (*p1, 23, 31));
i2 = float_istoobig (*p2)? ~0u : (bitget (*p2, 0, 23) | bitput (1, 23, 24)) >> (1 + max - bitget (*p2, 23, 31));
i3 = float_istoobig (*p3)? ~0u : (bitget (*p3, 0, 23) | bitput (1, 23, 24)) >> (1 + max - bitget (*p3, 23, 31));
is replaced by
i1 = float_istoobig (*p1)? ~0u : (bitget (*p1, 0, 23) | bitput (1, 23, 24)) >> ((1 + max - bitget (*p1, 23, 31)) % 32);
i2 = float_istoobig (*p2)? ~0u : (bitget (*p2, 0, 23) | bitput (1, 23, 24)) >> ((1 + max - bitget (*p2, 23, 31)) % 32);
i3 = float_istoobig (*p3)? ~0u : (bitget (*p3, 0, 23) | bitput (1, 23, 24)) >> ((1 + max - bitget (*p3, 23, 31)) % 32);
I don't think the compatibility test is needed
The reason
compress_compatability_test()fails in some cases (like when compiling for ARM with Clang with -O2) appears to be that shifts invector_compress(theVECTOR24/VECTOR32cases) are sometimes made likex >> ywherexisunsigned intandyis greater thanstd::numeric_limits<unsigned int>::digits, which means the behavior is undefined. So ifis replaced by
I don't think the compatibility test is needed