Fixup X86_TUNE_DEEP_BRANCH_PREDICTION

[Jan Hubicka]

> This fixes up X86_TUNE_DEEP_BRANCH_PREDICTION on nocona, core2 and generic.
> 
> This one should be reviewed by people that really know this stuff
> cold, it is easy to get lost.

Does this really mean that Nocona/Core can work out the unpaired call
instruction? I don't seem to be able to find any exception to the
pairing in optimization guides.

For Generic we quite definitly want to keep the returns in sync for AMD
systems.

Honza
> 
> Ok?
> 
> 2007-04-20  Mike Stump  <mrs@apple.com>
> 
> 	* config/i386/i386.c (ix86_tune_features
> 	[X86_TUNE_DEEP_BRANCH_PREDICTION]: Prefer call over thunks on
> 	nocona, core2 and generic.
> 
> $ cat deep.diffs
> Doing diffs in config:
> --- config/i386/i386.c.~1~	2007-04-13 10:05:47.000000000 -0700
> +++ config/i386/i386.c	2007-04-20 15:21:16.000000000 -0700
> @@ -1026,8 +1026,7 @@ unsigned int ix86_tune_features[X86_TUNE
>    m_486 | m_PENT | m_PPRO | m_ATHLON_K8_AMDFAM10 | m_K6 | m_CORE2 | m_GENERIC,
>  
>    /* X86_TUNE_DEEP_BRANCH_PREDICTION */
> -  m_PPRO | m_K6_GEODE | m_ATHLON_K8_AMDFAM10 | m_PENT4
> -  | m_NOCONA | m_CORE2 | m_GENERIC,
> +  m_PPRO | m_K6_GEODE | m_ATHLON_K8_AMDFAM10 | m_PENT4,
>  
>    /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
>       on simulation result. But after P4 was made, no performance benefit
> --------------