[WIP] Set real == double on Win64 MSVC.

[redstar]

The MS C Runtime does not support 80bit reals. Drop support in D, too.

[redstar]

@Trass3r @kinke Maybe you want to try this version?

[kinke]

Thanks Kai, that's most likely a sane decision for MSVC and may as well increase math performance (Phobos' std.math seems to encourage the use of reals, doesn't it?). Will try when I can - I'm currently using a core.stdc.math backported to LDC from dlang/druntime#968, for which I had to comment out ldc.longdouble to get rid of ambiguous symbols when linking against MSVC 2013.

[Trass3r]

That whole real by default attitude in phobos is just wrong anyway, esp. for x64 and non-x86.
I use stdc.tgmath though I wish calls like sin,exp,... would automatically be turned into llvm intrinsics.

[kinke]

+1 to that! Even worse, std.math implicitly assumes real to be 80-bit for x64 (version(D_InlineAsm_X86_64) guarding inline assembly). Imho, std.math should rely on core.stdc.math for the float and double overloads instead of forwarding to the phobos real implementation; almost all supported C runtimes provide all required functions anyway (even MS since 2013).
I'm not sure as to where we should place the LLVM intrinsics, i.e., std.math exclusively (where they currently are, at least those for reals) or in both core.stdc.math and std.math (only for the real implementations, the float/double functions would be aliases to core.stdc.math, as in core.stdc.tgmath). Yeah well, probably in both places, I guess.

[Trass3r]

Note that there's also the crippled core.math.
The reason for std.math reimplementing functions was given as the C versions being unreliable or something like that. Still real shouldn't be the default let alone forced on users like it currently is. But I think Walter needs to be convinced first IIRC.

[redstar]

This version caused an assertion failure in std.math. I have to fix this before I can merge this version.
If we are going this route (and IMHO it is wrong to fight against the C runtime) then we have to cleanup the math stuff.
BTW: On Linux/PPC I dropped the support for doubledouble type, too. It worked very well.

[redstar]

I found the root cause of the assertion failure. If the Travis build goes green then I merge this pull request.

[Trass3r]

Potential regression: #761