Do not run Fiber unit test on certain systems

[jpf91]

Systems which do use the C backend but the C library returns ENOSYS at runtime. (This happens on ARM with glibc < 2.17 for example).

We do not want to abort all testing in this case. We also don't want the static ctor to assert in this case whenever someone wants to use core.thread.

Note: This is part of a set of changes which are required to get test suite & unit tests passing on ARM GDC. All necessary changes for GDC are here for reference:
https://github.com/jpf91/GDC/commits/arm-old

Once these pull requests are merged upstream I'll merge them into GDC as well.

[MartinNowak]

Maybe we should rip out the ucontext_t fallback altogether?
Even if supported it adds a lot of unnecessary overhead to Fiber context switches and
the fiber_switchContext implementation is very similar to callWithStackShell. Both can easily be implemented in threadasm.S.

[MartinNowak]

Intersting http://www.boost.org/doc/libs/1_55_0/libs/context/doc/html/context/performance.html.

[jpf91]

OK we can remove ucontext_t if you want.
I've implemented the ASM for ARM now. It was more of an research task (ARM has many different fpu variants, do they have different registers? which registers must be saved? ...) but the actual implementation is really quite simple.

One remaining question: Would you say the floating point status register should be saved? It basically contains things like rounding mode, exception trap bits and exception 'occurred' bits:
http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0473c/Bcfhfbga.html

If we really want to micro-optimize fibers we probably have to enable inlining first. If the fiber_switchContext routines were implemented with extended inline ASM gdc should be able to inline those and hopefully the complete 'yield' function could be inlined.

[MartinNowak]

OK we can remove ucontext_t if you want.

I think that would make sense, it's a bad default and as you said an asm implementation is quite simple.

One remaining question: Would you say the floating point status register should be saved? It basically contains things like rounding mode, exception trap bits and exception 'occurred' bits:

We currently don't save/restore this. I tend to think of fiber context switching as a subroutine call.

If we really want to micro-optimize fibers we probably have to enable inlining first.

Could be useful, but I wouldn't expect much gain, because the function call overhead is minimal. Also having the asm code in an extra .S file allows us to share it among all compilers.

[MartinNowak]

Where is the ARM asm code btw?

[ibuclaw]

Does __builtin_unwind_init() not do the job?

I guess we can remove that version(GNU) code and put in x86 and x86_64 code paths into threadasm.S

[MartinNowak]

You'd still need code to switch the stack pointer.

[jpf91]

@MartinNowak it's not published yet. I want to test it on a system without FPU first.

@ibuclaw __builtin_unwind_init is fine for callWithStackShell but for Fibers the stack pointer must be switched and the registers have to be restored as well. AFAICS there's no builtin for that?

[ibuclaw]

@jpf91 oh right, I see it.

@MartinNowak would it be OK to copy D IASM code in fiber_switchContext to threadasm.S? That way I could add version (GNU) version = AsmExternal locally I guess...

[ibuclaw]

(And be one step ahead of removing ucontext_t)

[MartinNowak]

@MartinNowak would it be OK to copy D IASM code in fiber_switchContext to threadasm.S? That way I could add version (GNU) version = AsmExternal locally I guess...

Sure, I wouldn't even mind to switch from D IASM to external .S file altogether.
Maybe the IASM version is required for Windows though, because we don't have an assembler? Or do we have masm...?

[jpf91]

@MartinNowak the pull request for the ARM code is here #713