Incorrect == with __uint128t/uint64_t

[libbooze]

Following code:

template<typename Dec>
void mixed_compare()
{
    std::print("{}\n", boost::typeindex::type_id<Dec>().pretty_name());
    const Dec a{2, 1};
    const auto b = __uint128_t{std::numeric_limits<__uint128_t>::max() - std::numeric_limits<uint64_t>::max() + 20};
    std::print("a:{:.30e}\nb:{}\n", a , b);
    std::print("same {} \n\n", a  == b );
}

on my machine prints

boost::decimal::decimal32
a:2.000000000000000000000000000000e+01
b:340282366920938463444927863358058659860
same true 

boost::decimal::decimal64
a:2.000000000000000000000000000000e+01
b:340282366920938463444927863358058659860
same true 

boost::decimal::decimal128
a:2.000000000000000000000000000000e+01
b:340282366920938463444927863358058659860
same false 

I guess it is clear from test construction what I think the bug is(truncation). _fast code works fine for all sizes(probably because it dispatches to different equal_parts_impl).

edit:
For decimal32 bug persists if we "halve" the types used in b

    const auto b = uint64_t{std::numeric_limits<uint64_t>::max() - std::numeric_limits<uint32_t>::max() + 20};

[mborland]

We didn't design or test for math against non-standard extensions like __uint128_t. Really the only functionality that exists for them is conversion from or too. In the uint64_t case there's likely narrowing somewhere. It works in the decimal32_fast case because I'm willing to bet that your machine has uint_fast32_t as uint64_t

[libbooze]

I'm willing to bet that your machine has uint_fast32_t as uint64_t

I am on linux, this is what my IDE navigation resolves typedf to:
typedef unsigned long int uint_fast32_t;
so yes.