Phobos: Enable x87 asm versions for std.math.{tan,expi}

[kinke]

Following #2854.

[kinke]

Hmm, calling another function by name directly doesn't work in naked DMD-style asm if compiled as PIC, so the return-real.nan-workaround by calling a helper function isn't viable either...

[pineapplemachine]

I haven't got an inline asm tangent or sincos implementation to work with LDC yet. However, I did have some apparent success with this particular part of the problem by declaring immutable nan = real.nan for DMD and static immutable nan = real.nan for LDC (each failed to compile for the other) and then writing e.g. fld nan[RBP]; ret; inside the asm block instead of return real.nan; outside of it.

[kinke]

static immutable nan = real.nan

That should have the same problem when compiled as PIC (-relocation-model=pic), the global won't be accessible directly.

[kinke]

Revived this with a viable workaround for PIC (pass real.nan as extra default parameter); the x87 asm version is more than 3x faster than the generic 'software' one on my Ivy Bridge, tested on Linux x64 with -O and:

void main()
{
    enum MAXITER = 1 << 20;

    import std.datetime.stopwatch;
    import std.math;
    import std.stdio;

    real sum = 0;

    auto sw = StopWatch(AutoStart.yes);
    for(uint i; i < MAXITER; i++)
        sum += tan(real(i)); // also tested with `real(i) / MAXITER`
    sw.stop();

    writeln(sw.peek.total!"msecs", "ms\t", sum);
}

[kinke]

The x87 asm for std.math.expi() is faster by ~5x (expi(real(i) / MAXITER)) to ~12x (expi(real(i))) on my box (edit: and still 3-5x times faster than a llvm_cos(x) + llvm_sin(x)*1i implementation with an additional -ffast-math).