[RFC] Clean up SSE variable shifts

[Lu, Hongjiu]

[hjl@gnu-6 strcspn]$ size sse4-strcspn-vshft1.o sse4-strcspn-vshft2.o sse4-strcspn-vshft3.o sse4-strcspn-vshft4.o sse4-strcspn-vshft5.o varshift?.o
   text	   data	    bss	    dec	    hex	filename
    684	      0	      0	    684	    2ac	sse4-strcspn-vshft1.o
    591	      0	      0	    591	    24f	sse4-strcspn-vshft2.o
    335	      0	      0	    335	    14f	sse4-strcspn-vshft3.o
    335	      0	      0	    335	    14f	sse4-strcspn-vshft4.o
    324	      0	      0	    324	    144	sse4-strcspn-vshft5.o
    174	      0	      0	    174	     ae	varshift3.o
    256	      0	      0	    256	    100	varshift4.o
     31	      0	      0	     31	     1f	varshift5.o

The order of size with the smallest first is

Replace palignr with unaligned load, replace intrinsic with pshufb + unaligned load
Replace palignr with unaligned load, replace intrinsic with a function call
Replace palignr with unaligned load, replace intrinsic with pshufb
Replace palignr with unaligned load, replace intrinsic with asm statement
Replace palignr with unaligned load

H.J.


> -----Original Message-----
> From: Ulrich Drepper [mailto:drepper@redhat.com]
> Sent: Monday, August 23, 2010 10:05 PM
> To: Lu, Hongjiu
> Cc: libc-alpha@sourceware.org; Richard Henderson
> Subject: Re: [RFC] Clean up SSE variable shifts
> 
> ----- "Hongjiu Lu" <hongjiu.lu@intel.com> wrote:
> > Here are TSC deltas between different implementations. It is
> > hard to tell which one is faster.
> 
> Agreed.  This is, though, the best result.  It means the
> implementations really don't differ at all in micro-benchmarks.
> THerefore we can use the one with has the minimum resource use, in
> code and data.  Can you post that data, too?  I mean the 'size' for
> the various sections.
> 
> --
> ➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖