Comparison to Boost.Multiprecision

The two types are not the same shape. boost::multiprecision::uint256_t is number<cpp_int_backend<256, 256, unsigned_magnitude, unchecked, void>>, which stores a limb count and a sign next to the limbs, so sizeof is larger than 32 bytes and the object representation is not simply the value. uint256 is exactly 32 bytes, trivially copyable, and standard layout, so it can be memcpy-ed, written into a packet or a file format, stored in an array that a hardware instruction can load, or handed to a GPU kernel.

The signed comparison is sharper still. boost::multiprecision::int256_t is number<cpp_int_backend<256, 256, signed_magnitude, checked, void>>: a sign-magnitude representation with checked overflow, whose range is +-(2^256 - 1), one more negative value’s worth wider on the positive side and one narrower on the negative side than a 256-bit two’s-complement type, and whose default arithmetic throws std::overflow_error rather than wrapping. int256 is two’s complement over the same four std::uint64_t words[4] uint256 uses, with range [-2255, 2255 - 1], and it wraps on overflow rather than throwing, matching a built-in signed integer’s is_modulo behavior. boost::mp::int256_t is still the benchmark baseline (see int256 Benchmarks) because no platform has a hardware 256-bit type to compare against instead, not because the two types model the same thing.

The goal is for this library to be extremely lightweight. It has no Boost dependencies at all, whereas Boost.Multiprecision has a module weight of 25.

In Boost.Multiprecision every type is built on the high-level number template so that all of the backends interoperate. uint256 is a single concrete struct with individually implemented operators, designed to work with and act like the built-in integer types.

uint256 is usable in constant expressions from C++14 onward, and on CUDA and SYCL devices, neither of which a cpp_int backed type supports.