std.math: use most inline x86/x64 assembly only if real is defined as x87

[kinke]

Now that real == double on Win64. Useful in combination with ldc-developers/druntime#13 and MSVC x64 2013+.

[redstar]

version = INTRINSICS_FROM_305; is missing here.

[kinke]

No intrinsics introduced in LLVM 3.5/3.6 are used, that's why I limited the INTRINSICS_FROM_* to those which are actually required.

[redstar]

Does not build on Linux:

/home/travis/build/ldc-developers/ldc/runtime/phobos/std/math.d(90): Error: version INLINE_YL2X defined after use

See https://travis-ci.org/ldc-developers/ldc/jobs/38571952

[kinke]

Hmm, no idea what's causing this error. I've simply moved the version = INLINE_YL2X; statements into a static if (real.sizeof > double.sizeof) { ... } block.
Additionally, the problem seems to arise while compiling std.mathspecial, which publicly imports std.math. And according to Travis output, that phobos std.mathspecial seems to be compiled while building druntime (?!), specifically, rt.sections_linux and/or rt.sections_android...

[redstar]

The Travis output is not ordered because several processes are running in parallel.

[redstar]

I think the problem here is that the version blocks are evaluated first. This is done because the code of a not satisfied version is not compiled into the program. In contrast, the static if is always evaluated at compile time. The result is that version = INLINE_YL2X; is evaluated after all version guards have been evaluated. The TDPL rule is: You can set a version only before any read of that version.

It works in the Win64 case because the body of the static if is excluded.

[kinke]

Should be fixed in #4.

[dnadlinger]

I don't think duplicating the INTRINSICS_FROM_… logic here is a good idea. Instead, use __traits(compiles, <… call to the intrinsic in question…>).

[redstar]

Good hint.