PATCH: Add SSE4.2 support

[Uros Bizjak]

On 5/30/07, H. J. Lu <hjl@lucon.org> wrote:

> > > BTW: If it is not too much trouble, could string/text processing
> > > intrinsic be split out into separate patch? The first patch would then
> > > implement only SSE4.2 flags handling and logic/CRC operations that we
> > > are all somehow familiar with, and the second will add string
> > > processing.
> > >
> >
> > I prefer to use one patch for SSE4.2 if possible at all. But I will
> > try to use 2 if there is no way around it.
> >
>
> Here is the updated patch. I added OPTION_MASK_ISA_XXX_UNSET so
> that we only need to change one macro when we add a new ISA.  Tested
> on Linux/Intel64.

H.J.,

The reason for separate text/string processing patch is, that the
implementation looks fundametally wrong to me (I'll discuss this in a
separate mail), and text/string processing _could_ be moved in a
separate,  orthogonal patch. It is OK to include smmintrin.h and
nmmintrin.h unmodified, because unsupported __builtin_ia32_pcmp*
functions will be emmitted as a normal call to __builtin_ia32_pcmp*()
function; that is - at the moment, they won't be expanded to special
RTL sequences. Also, the documentation can be added as is in the
patch, including currently unimplemented __builtin_* string/text
functions. So, for now, just remove all pcmpstr* handling from
i386-modes.def (new modes), i386.c, i386.h (new modes), predicates.md
(new modes handling) and sse.md

> +  def_builtin (OPTION_MASK_ISA_SSE4_2, "__builtin_ia32_crc32qi",
> +              ftype, IX86_BUILTIN_CRC32QI);

These def_builtin() functions should be defined in one line, to
maintain some consistency with other def_builtin() calls.

> +  /* Only SSE4.1/SSE4.2 supports V2DImode.  */
> +  if (mode == V2DImode)
> +    {
> +      switch (code)
> +	{
> +	case EQ:
> +	  /* SSE4.1 supports EQ.  */
> +	  if (!TARGET_SSE4_1)
> +	    return false;
> +	  break;
> +
> +	case GT:
> +	case GTU:
> +	  /* SSE4.2 supports GT/GTU.  */
> +	  if (!TARGET_SSE4_2)
> +	    return false;
> +	  break;

You have to add supporting code to convert V2DI GTU into GT in the
code just below the chunk you added. Something similar to existing
V4SI mode handling.

Uros.