runtime: yield with _mm_pause instead of __builtin_ia32...

runtime: yield with inline asm instead of __builtin_ia32_pause

The former is more portable than the latter, being supported by at
least GCC and Clang.

[iant, 2014-05-29 00:37:52]

The problem with calling _mm_pause is that the default 32-bit x86 target does
not support SSE.  Of course _mm_pause does not actually require SSE support.  On
the other hand <xmmintrin.h> should only be used for targets with SSE support. 
So it's not clear whether this is a bug or not.

What is clear is that the <ia32intrin.h> header file defines __pause, which has
the same effect and does work.  Does it work with clang to #include
<ia32intrin.h> and use __pause?

[pcc, 2014-05-29 03:55:22]

On 2014/05/29 00:37:52, iant wrote:
> The problem with calling _mm_pause is that the default 32-bit x86 target does
> not support SSE.  Of course _mm_pause does not actually require SSE support. 
On
> the other hand <xmmintrin.h> should only be used for targets with SSE support.

> So it's not clear whether this is a bug or not.
> 
> What is clear is that the <ia32intrin.h> header file defines __pause, which
has
> the same effect and does work.  Does it work with clang to #include
> <ia32intrin.h> and use __pause?

Clang does not support __pause. Maybe the simplest thing to do is to use inline
asm, which both compilers should support.

[pcc, 2014-05-29 22:27:15]

On 2014/05/29 03:55:22, pcc wrote:
> On 2014/05/29 00:37:52, iant wrote:
> > The problem with calling _mm_pause is that the default 32-bit x86 target
does
> > not support SSE.  Of course _mm_pause does not actually require SSE support.

> On
> > the other hand <xmmintrin.h> should only be used for targets with SSE
support.
> 
> > So it's not clear whether this is a bug or not.
> > 
> > What is clear is that the <ia32intrin.h> header file defines __pause, which
> has
> > the same effect and does work.  Does it work with clang to #include
> > <ia32intrin.h> and use __pause?
> 
> Clang does not support __pause. Maybe the simplest thing to do is to use
inline
> asm, which both compilers should support.

Done.

[iant, 2014-05-29 22:52:59]

I'm working on getting GCC to support _mm_pause while building libgo for older
32-bit processors (https://gcc.gnu.org/ml/gcc-patches/2014-05/msg02560.html). 
Let's see what happens with that.

[iant, 2014-05-30 13:54:32]

I've committed a version of this.  You can abandon this CL (hg change -d
100810044).