Number class shows degradation in math operations performance.

[gregtatcam]

Issue Description

Number class shows degradation in math operations performance as compared to similar operations
implemented in STAmount.

Steps to Reproduce

Here is a simple unit test, which demonstrates the issue:

void
testSwap1()
{
    testcase("Swap1");
    using namespace jtx;

    std::uint16_t tfee = 1000;
    auto const gw = Account("gw");
    auto const USD = gw["USD"];
    auto const GBP = gw["GBP"];
    auto const issueIn = USD;
    auto const issueOut = GBP;
    auto const in = STAmount{USD, 1000};
    auto const out = STAmount{GBP, 1000};
    auto const assetIn = STAmount{USD, 1};
    auto const assetOut = STAmount{GBP, 1};

    auto start = std::chrono::high_resolution_clock::now();
    for (int i = 0; i < 100; ++i)
    {
        auto const fm = 1 - Number{tfee} / 100000;
        auto const res = in - (in * out) / (in + assetIn * fm);
    }
    auto elapsed = std::chrono::high_resolution_clock::now() - start;
    std::cout << "Number(swapIn) math: " << std::chrono::duration_cast<std::chrono::microseconds>(
                                        elapsed).count() << std::endl;

    start = std::chrono::high_resolution_clock::now();
    for (int i = 0; i < 100; ++i)
    {
        auto const n1 = STAmount{1};
        auto const feeMult = n1 - divide(STAmount{tfee}, STAmount{100000}, n1.issue());
        auto const en = multiply(in, out, issueOut);
        auto const den = in + multiply(assetIn, feeMult, issueIn);
        auto const res = out - divide(en, den, issueOut);
    }
    elapsed = std::chrono::high_resolution_clock::now() - start;
    std::cout << "STAmount(swapIn) math: " << std::chrono::duration_cast<std::chrono::microseconds>(
        elapsed).count() << std::endl;
    BEAST_EXPECT(true);
}

Expected Result

Math operations using Number and STAmount function should produce similar results.

Actual Result

The issue manifested itself in the payment engine updated for AMM. Swap in formula used
to figure out initial AMM offer size resulted in the throughput degradation.

Environment

The benchmark test ran on 11 "Ubuntu 20.04.4 LTS" servers (five validators, four clients, two signing) in AWS.
The unit test ran on OSX Ventura.
AMM PR was used to build rippled: #4294

[HowardHinnant]

I believe this commit fully addresses this issue: 16f1057

This commit is part of the NumberMerge PR: #4192

[intelliot]

Fixed by 16f1057