libexpr/print: Require decimal point on float printing.

[crertel]

Motivation

Fixes #3077 .

Context

Tracking down the weird behavior of current nix (per the issue), we see:

I looked around for a while and realized this was probably happening over in the repl code, so I looked for CYAN and found printFloat over in src/libexpr/print.cc.

The original code:

    void printFloat(Value & v)
    {
        if (options.ansiColors)
            output << ANSI_CYAN;
        output << v.fpoint();
        if (options.ansiColors)
            output << ANSI_NORMAL;
    }

Now, iostreams by default don't show the decimal point, so we have to force it (after saving the flags):

    void printFloat(Value & v)
    {
        const std::ios_base::fmtflags flags = output.flags();

        if (options.ansiColors)
            output << ANSI_CYAN;

        output.setf(std::ios::showpoint);
        output << v.fpoint();
        output.flags(flags);

        if (options.ansiColors)
            output << ANSI_NORMAL;
    }

This has the desired effect, mostly:

Unfortunately, the question of "how do we get this to show the correct decimal precision?" is a bit open-ended.

The open question is:

• Should we always, for floats with no fractional parts, just do a single decimal point? E.g., 4.0000 renders as 4.0.
• Should floats always render with default precision? E.g., 4.0 renders as 4.00000...but so does 4.000001.
• Should floats try to display as much as they can out to machine precision? E.g., 4.00000000000004 displays as 4.00000000000004 but some larger one eventually rounds. If it rounds to all zeros, should that be truncated as well?

There's some test problems as well, but those are less important than this core formatting question.

---

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.