This patch moves/implements all logic (multiplication, division, GetCompact, SetCompact) that main used from CBigNum to base_uint or uint256, and replaces its call sites by equivalent uint256 operations.
The division and multiplication operations are very straightforward, but not very efficient. None of their call sites require that, though.
Together with #3965 and #4048, this will allow up to drop bignum.h entirely.
Some benchmarks:
Added another commit that removes some duplicated code.
Also replaced the bits() implementation with a faster one, making division 10% faster.
ACK. (reviewed the code, tested a mainnet and a testnet resync, and added some errors and checked that they were caught— I do note that triggering divide by zero in— say— block validation results in the exception being caught which surprised me momentarily).
The division can be made another 10% faster by using CLZ instead of the inner loop, but further optimizations do not need to happen now. (We call the divide in GetBlockWork, which we do for every block, and it appears to make a difference in the very early sync for me; though I consider these measurements somewhat suspect.)