Fix issue 16555 - Generate correct code for pushing scalar parameters.

[LemonBoy]

This is my try at fixing this issue, bear with me as it's my first time playing around with the dmd internals.
The snippet below shows a major codegen problem, the first 8 parameters are correctly passed into XMM0-XMM7 but for the last one a push rax is generated, which is clearly wrong.
The patch corrects this issue by exiting the huge switch and using the code that's already there to correctly push fp values and vectors (the vector path is untested, I just assumed we'd hit the same problem).
There's a further change that attempts to use the XMM registers whenever possible to avoid the using the x87 instructions, it does work fine in this case but I'm not 100% sure about this.

import std.stdio;

void main()
{
    void inner(double x)
    {
        if(x == 999) // to show that x itself isn't corrupt
        {
            writefln("x=%s a=%s b=%s c=%s d=%s e=%s f=%s g=%s h=%s", x, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0);
            //outer(x, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0);
        }
    }

    inner(999.0);
}

Cheers

[dlang-bot]

Fix	Bugzilla	Description
✓	16555	Stack corruption when calling multi-parameter outer function from nested function

[ghost]

unit test needed for this one. Take alook at https://github.com/dlang/dmd/tree/master/test.

[WalterBright]

This definitely needs a test case. How about adapting the example in the bug report?

[JinShil]

Revived at #7890