FPTemporary update

[9il]

Do not use x87 FPU stack when SSE2 is available.

[mihails-strasuns]

Need someone more knowledgeable with FP details to verify this change.

[dnadlinger]

Pinging @donc @don-clugston-sociomantic. I'd need to think about this some more to figure out whether this is really the correct strategy.

Some quick thoughts: version (X86) seems troublesome, as the ABI might still mandate x87 stack returns and so on, even if the target supports SSE math. Also, making this depend on vector support seems a bit odd. For example, you can instruct GCC to emit x87 code on x86_64 and so on.

[9il]

Agreed. Is it possible to add version(x87code) or something like that?
The reason of this PR is optimization of std.complex that use FPTemporary. Also one algorithm in std.numeric use this feature.

Would it better to reorganize std.complex to not use FPTemporary?

[don-clugston-sociomantic]

I'm not really sure about the concept of FPTemporary. A few observations:

1. Using 'does it compile' as a test is a bit doubtful. Some 32-bit Intel CPUs support SSE2, even though AMD never did. Conceivably, some compilers could allow that code to compile. We could forbid compilers from doing that, but it would be a funky rule.
2. Another architecture which has this behaviour is Itanium. It isn't x87.
3. The other issue with intermediate precision being higher than final precision is support for FMA. That can give you effectively 128 bit precision in certain cases. This means that the concept of FPTemporary is a bit simplistic.
4. Walter has advocated vague semantics, where a local variable of type 'float' or 'double' can be implemented as an 80 bit real on x87. That is, you may still get 'real' semantics even if you haven't specified it. In fact, to some extent there is no other option -- unless you do the silly thing that Java does, where you cripple performance by reducing precision. I have an ongoing dispute with Walter about point 4 (because you still need some way of enforcing extra precision to be abandoned) but certainly the C/C++ concept of 'sequence points' doesn't actually work.

Do you actually need to use FPTemporary? Because of Point 4, it might not be necessary. Because of point 3, it might not be sufficient!

[9il]

I don't need FPTemporary. In other hand I need std.complex. Can I make PR to remove FPTemporary from std.complex?

[don-clugston-sociomantic]

I think you can probably remove FPTemporary from std.complex. See if it passes the test suite when you do.

[9il]

Thanks!