The abstraction penalty for wide integer math on x86-64

Back in November 2018, following a thread on the std-proposals mailing list, I sat down and wrote out a really simple “wide integer” library.

“Wide integers” are integers of fixed-but-bigger-than-word-length size. For example, the __uint128_t supported by GCC and Clang would be considered a “wide” integer. Wide integers are not to be confused with “big integers,” i.e., integers of dynamic (heap-allocated) size.

I can also imagine “wide floating-point,” “wide fixed-point,” “wide rational” (that is, a number expressed and stored as the ratio of two wide integers), and “big rational.” I doubt that “big fixed-point” or “big floating-point” make philosophical sense.


My library provides a single class template: Wider<T>. The T can be either uint64_t, or some specialization of Wider. Each Wider doubles the bit-width of the unsigned integer represented; so we can say

using Uint128 = Wider<uint64_t>;
using Uint256 = Wider<Uint128>;
using Uint512 = Wider<Uint256>;
using Uint1024 = Wider<Uint512>;

I looked at Clang’s codegen for __uint128_t to see what might be the most optimal codegen, and then tweaked my code to try to achieve the same codegen. It was actually pretty easy, on Clang/LLVM! On GCC, not so much.

The idea is that we can write plain old C++ code — maybe with a few intrinsics to help out at the lowest level — and then let the compiler do all the hard work of optimizing it. Here’s the code for wide-integer addition:

inline bool producecarry(uint64_t& x, uint64_t y) {
    x += y;
    return (x < y);
}

inline bool addcarry(bool cf, uint64_t& x, uint64_t y) {
    return _addcarry_u64(cf, x, y, (unsigned long long*)&x);
}

template<class Int64>
struct Wider {
    Int64 lo;
    Int64 hi;

    friend bool producecarry(Wider& x, const Wider& y) {
        return addcarry(producecarry(x.lo, y.lo), x.hi, y.hi);
    }

    friend bool addcarry(bool cf, Wider& x, const Wider& y) {
        cf = addcarry(cf, x.lo, y.lo);
        cf = addcarry(cf, x.hi, y.hi);
        return cf;
    }

    friend Wider& operator+=(Wider& x, const Wider& y) { (void)producecarry(x, y); return x; }
    friend Wider operator+(Wider x, const Wider& y) { x += y; return x; }
};

Compile this with Clang trunk and we see that the compiler produces the exact same code for

Uint128 example(const Uint128& x, const Uint128& y)
{
    return x + y;
}

__uint128_t example(const __uint128_t& x, const __uint128_t& y)
{
    return x + y;
}

And that code is optimal:

movq (%rsi), %rax
addq (%rdi), %rax
movq 8(%rsi), %rdx
adcq 8(%rdi), %rdx
retq

GCC trunk, on the other hand, performs abysmally.


Anywhere there’s a gap between the performance of Wider<T> and the performance of a built-in type like __uint128_t, that’s an opportunity for some compiler writer to go improve the codegen. In this way, WideIntProofOfConcept is kind of like the “Stepanov Abstraction Penalty Benchmark”; it points to places where the compiler could do better at recognizing peephole-level idioms.

Back in 2018–2019 I filed a bunch of bugs against LLVM inspired by this benchmark: 39968, 40090, 24545, 31754, 40486, 40825. They were all fixed and closed quite expeditiously!

Bug 40908, an ICE on using __int128 as an NTTP, I actually discovered while writing “Is __int128 integral? A survey” (2019-02-28). It remains open.

To see the current “penalties” for using Wider<T> versus the built-in types, see the README on GitHub. There’s also a Python script you can download to regenerate the table using Godbolt’s API, if you want to try other (Godbolt-supported) compilers, or just to see whether my numbers are up-to-date.


Finally, thanks to Niall Douglas for (two years ago) motivating this whole thing by pointing out how absymally awful Boost.Multiprecision is! Here is a reduced version of Niall’s benchmark for a + b, showing boost::multiprecision::uint128_t alongside native unsigned __int128; Wider<uint64_t>; and the new kid on the block, Abseil’s absl::uint128.

Abseil’s absl::uint128 seems to be high-quality. I do detect some places where it gives worse codegen than Wider<uint64_t> — for example, try the expression a + -b — but ye gods, Abseil is worlds better than Boost.Multiprecision!

Posted 2020-02-13