SSE conversion optimization

[H.J. Lu]

On Sun, Sep 09, 2007 at 07:49:04PM +0200, Jan Hubicka wrote:
> Hi,
> this is variant of patch I comitted.  It change only amdfam10 codegen so
> far until we figure out what would be best setting for generic and
> core2.
> For AMDFAM10 it is also better to offload to memory operand of
> DImode->SF/DFmode conversions I will try to do next. 
> 
> Honza
> 
> Index: ChangeLog
> ===================================================================
> --- ChangeLog	(revision 128300)
> +++ ChangeLog	(working copy)
> @@ -1,3 +1,19 @@
> +2007-09-09  Jan Hubicka  <jh@suse.cz>
> +	     Dwarakanath Rajagopal <dwarak.rajagopal@amd.com>
> +
> +	* i386.h (ix86_tune_indices): Add X86_USE_VECTOR_CONVERTS.
> +	(TARGET_USE_VECTOR_CONVERTS): New.
> +	* i386.md: New post-reload splitters for converting SF to DF and DF to
> +	SF.
> +	(floatsi* expander): Special case vector conversions.
> +	(floatsisf2_mixed_vector, floatsisf2_sse_vector_nointernunit,
> +	floatsisf2_sse_vector_internunit, floatsisf2_sse_vector,
> +	floatsidf2_mixed_vector, floatsidf2_sse_vector): New.
> +	(floatsisf2_mixed, floatsisf2_sse, floatsidf2_mixed, floatsidf2_sse):
> +	Disable when doing vector converts.
> +	(floatsi<mode>2_i387): Disable when
> +	* sse.md (vec_dupv2df): Export.
> +	* i386.c (ix86_tune_features): Enable SSE conversions.
> +
>  2007-09-09  Richard Guenther  <rguenther@suse.de>
>  
>  	* tree-ssa-operands.c (add_virtual_operand): Only mark
> Index: config/i386/i386.h
> ===================================================================
> --- config/i386/i386.h	(revision 128300)
> +++ config/i386/i386.h	(working copy)
> @@ -257,6 +257,7 @@ enum ix86_tune_indices {
>    X86_TUNE_MOVE_M1_VIA_OR,
>    X86_TUNE_NOT_UNPAIRABLE,
>    X86_TUNE_NOT_VECTORMODE,
> +  X86_USE_VECTOR_CONVERTS,

Would you mind using X86_TUNE_USE_VECTOR_CONVERTS to make it
consistent with other tuning options?

>  
>    X86_TUNE_LAST
>  };

Thanks.


H.J.
H.J.